Re: [RFC 1/2] refcount: introduce generic lockptr funcs

[Peter Zijlstra]

On Fri, Nov 03, 2023 at 03:20:08PM -0400, Alexander Aring wrote:
> Hi,
> 
> On Fri, Nov 3, 2023 at 2:54 PM Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
> >
> > On Fri, Nov 03, 2023 at 12:16:34PM -0400, Alexander Aring wrote:
> >
> > > diff --git a/lib/refcount.c b/lib/refcount.c
> > > index a207a8f22b3c..e28678f0f473 100644
> > > --- a/lib/refcount.c
> > > +++ b/lib/refcount.c
> > > @@ -94,6 +94,34 @@ bool refcount_dec_not_one(refcount_t *r)
> > >  }
> > >  EXPORT_SYMBOL(refcount_dec_not_one);
> > >
> > > +bool refcount_dec_and_lockptr(refcount_t *r, void (*lock)(void *lockptr),
> > > +                           void (*unlock)(void *lockptr),  void *lockptr)
> > > +{
> > > +     if (refcount_dec_not_one(r))
> > > +             return false;
> > > +
> > > +     lock(lockptr);
> > > +     if (!refcount_dec_and_test(r)) {
> > > +             unlock(lockptr);
> > > +             return false;
> > > +     }
> > > +
> > > +     return true;
> > > +}
> > > +EXPORT_SYMBOL(refcount_dec_and_lockptr);
> >
> > This is terrible, you're forcing indirect calls on everything.
> >
> 
> Okay, I see. How about introducing a macro producing all the code at
> preprocessor time?

__always_inline should work, then you get constant propagation for the
function pointer.

But indeed, perhaps a macro is more convenient vs the irq flags
argument. You'll then end up with something like:

#define __refcount_dec_and_lock(_ref, _lock, _unlock) \
({ 	bool _ret = false; \
	if (!refcount_dec_not_one(_ref)) { \
		_lock; \
		if (!refcount_dec_and_test(_ref)) { \
			_unlock; \
		} else { \
			_ret = true; \
		} \
	} \
	_ret; \
})


bool refcount_dec_and_spinlock_irqsave(refcount_t *r, spinlock_t *lock, 
				       unsigned long *flags)
{
	return __refcount_dec_and_lock(r, spin_lock_irqsave(*lock, *flags),
					  spin_unlock_irqrestore(*lock, *flags));
}