Re: [PATCH RFC 03/33] locking: Introduce <linux/thread_safety.h>

[Marco Elver]

On Fri, 7 Feb 2025 at 23:34, Bart Van Assche <bvanassche@xxxxxxx> wrote:
[...]
> > Those are all nasty shouting names, without and good prefixing.
> >
> > But more importantly ACQUIRE() and RELEASE() seems to duplicate the
> > existing __acquires/__releases annotations from sparse.  We really need
> > to find away to unify them instead of duplicating the annotations.
>
> I think that we are better off to drop support for the sparse locking
> annotations. Linus added the macro __cond_acquires() two years ago in

+1 to dropping Sparse support -- although we can retain the same keywords.

[...]
> * The argument of __acquire() and __release() can be a "capability" or
>    the address of a synchronization object. A few examples where the
>    argument represents a capability:

Not a problem, and can be solved by introducing "token" instances...

>    __acquire(RCU);
>    __acquire(RCU_BH);
>    __acquire(RCU_SCHED);

For RCU: https://lore.kernel.org/lkml/20250206181711.1902989-16-elver@xxxxxxxxxx/

>    __acquire(bitlock);

For bitlock: https://lore.kernel.org/lkml/20250206181711.1902989-15-elver@xxxxxxxxxx/

>    An example where the argument represents the address of a
>    synchronization object:
>
>    static inline void do_raw_spin_lock(raw_spinlock_t *lock)
>          __acquires(lock)
>    {
>         __acquire(lock);
>         arch_spin_lock(&lock->raw_lock);
>         mmiowb_spin_lock();
>    }
>
>    If __acquire() and __release() would have to be supported for both
>    sparse and Clang, that would probably involve modifying all instances
>    of these two macros across the entire kernel tree and making it
>    explicit whether the argument is a capability or the address of a
>    synchronization object.

This is not a problem as I have demonstrated with my approach. Using
token instances is also the way we'd be able to support things like
preempt-disable-enable, irq-disable-enable, etc.

In general I'm in favor of dropping Sparse support, but the majority
of existing annotations can be reused, simplifying our work.