Re: [PATCH 01/62] kvm: Make pi_enable_wakeup_handler() easier to analyze

[Marco Elver]

On Thu, 26 Feb 2026 at 18:48, Sean Christopherson <seanjc@xxxxxxxxxx> wrote:
> On Tue, Feb 24, 2026, Bart Van Assche wrote:
[...]
> > Regarding why the above patch is necessary, I don't think that it is
> > fair to blame the compiler in this case. The macros that implement
> > per_cpu() make it impossible for the compiler to conclude that the
> > pointers passed to the raw_spin_lock_nested() and raw_spin_unlock()
> > calls are identical:
>
> Well rats, that pretty much makes it infeasible to solve the underlying problem.
>
> > /*
> >  * Add an offset to a pointer.  Use RELOC_HIDE() to prevent the compiler
> >  * from making incorrect assumptions about the pointer value.
> >  */
> > #define SHIFT_PERCPU_PTR(__p, __offset)                               \
> >       RELOC_HIDE(PERCPU_PTR(__p), (__offset))
> >
> > #define RELOC_HIDE(ptr, off)                                  \
> > ({                                                            \
> >       unsigned long __ptr;                                    \
> >       __asm__ ("" : "=r"(__ptr) : "0"(ptr));                  \
> >       (typeof(ptr)) (__ptr + (off));                          \
> > })

There's a slim chance we can "fix" this with a similar approach as in:
https://lore.kernel.org/all/20260216142436.2207937-2-elver@xxxxxxxxxx/
(specifically see patch 2/2)

The goal of RELOC_HIDE is to make the optimizer be less aggressive.
But the Thread Safety Analysis's alias analysis happens during
semantic analysis and is completely detached from the optimizer, and
we could potentially construct an expression that (a) lets Thread
Safety Analysis figure out that __ptr is an alias to ptr, while (b)
still hiding it from the optimizer. But I think we're sufficiently
scared of breaking (b) that I'm not sure if this is feasible in a
clean enough way that won't have other side-effects (e.g. worse
codegen).

If I find time I'll have a think unless someone beats me to it.

Thanks,
-- Marco