Re: [PATCH v5 15/36] srcu: Support Clang's context analysis

On Mon, 26 Jan 2026 at 22:36, Peter Zijlstra <[email protected]> wrote:
>
> On Mon, Jan 26, 2026 at 10:54:56AM -0800, Bart Van Assche wrote:
>
> > Has it ever been considered to add support in the clang compiler for a
> > variant of __must_hold() that expresses that one of two capabilities
> > must be held by the caller? I think that would remove the need to
> > annotate SRCU update-side code with __acquire_shared(ssp) and
> > __release_shared(ssp).
>
> Right, I think I've asked for logical operators like that. Although I
> think it was in the __guarded_by() clause rather than the __must_hold().
> Both || and && would be nice to have ;-)

Some attributes take multiple arguments (__must_hold does), though
__guarded_by doesn't. Yet, && can still be had with adding it multiple
times e.g. '__guarded_by(pi_lock) __guarded_by(rq->__lock)'.

Only thing that doesn't exist is ||. I think the syntax you ask for
won't fly, but I can add it to the backlog to investigate an _any
variant of these attributes. Don't hold your breath though, given the
time it takes to land all that in a released Clang version.

> Specifically, I think I asked for something like:
>
>         cpumask_t       cpus_allowed __guarded_by(pi_lock && rq->__lock)
>                                      __guarded_shared_by(pi_lock || rq->__lock);
>
>
> I think Marco's suggestion was to use 'fake' locks to mimic those
> semantics.