Re: [PATCH v3 13/13] KVM: selftests: Add and use double-underscore versions of vmlaunch() and vmresume()

[Sean Christopherson]

On Thu, Aug 27, 2026, Yosry Ahmed wrote:
> On Thu, Aug 27, 2026 at 1:48 PM Sean Christopherson <seanjc@xxxxxxxxxx> wrote:
> >
> > On Thu, Aug 27, 2026, Yosry Ahmed wrote:
> > > On Thu, Aug 27, 2026 at 1:28 PM Sean Christopherson <seanjc@xxxxxxxxxx> wrote:
> > > > Yeah.  Another idea would be:
> > > >
> > > >   void vmenter(bool do_vmlaunch)
> > > >   {
> > > >         if (do_vmlaunch)
> > > >                 vmlaunch();
> > > >         else
> > > >                 vmresume();
> > > >   }
> > > >
> > > > To yield:
> > > >
> > > >         vmenter(vector == SS_VECTOR);
> > > >         vmenter(!i);
> > > >
> > > > Never mind, that's a terrible idea, the call sites are inscrutable.
> > >
> > > Yeah, but I do agree that vmlaunch() vs vmresume() handling in
> > > selftests is annoying (e.g. see run_l2() in [1]).
> > >
> > > If you wanna go scorched earth on the nested VMX infra, what if we
> > > track launch vs resume (e.g. set launch=true in prepare_vmcs() and
> > > lauch=false in vmlaunch()), then provide a single vmenter() API for
> > > tests (with no arguments)?
> >
> > For selftests, it would be much more difficult to do correctly, as compared to
> > KUT.  KUT has a "fixed" number of vCPUs, doesn't need to worry about pCPU:vCPU
> > bindings, and has a strong need for per-vCPU state for other reasons (because
> > it's more like per-pCPU state than per-vCPU state).
> >
> > Selftests doesn't have per-vCPU state, so making the tracking thread-safe would
> > be challenging.  And I don't really want to add per-vCPU state because making
> > GS (or FS) off-limits is actually problematic when the goal is often to do "bad"
> > things to vCPU state (I've debugged weird KUT failures due to GS getting clobbered
> > more than once).
> 
> Isn't vmx_pages effectively per-vCPU state (for nested)?

Yeah, but you said "with no arguments", there are tests that use helpers, and
IMO passing in a vmx_pages pointer is uglier than implicitly tracking launched
state.