Re: [PATCH 01/14] locking/atomic: import atomic_dec_not_zero()

[Fabian Frederick]

> On 01 February 2017 at 10:25 Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote:
>
>
> On Tue, Jan 31, 2017 at 09:55:08PM +0100, Fabian Frederick wrote:
> > Once again it's just about readability:
>
> I feel APIs should be about common use-cases, not about sporadic weird cases.
>
> > "add -1 unless value is zero" looks more complex in code than "dec not zero"
> > but maybe it's just a matter of taste :) It it's not the case why would
> > there be
> > more sense about having
> > atomic_inc_not_zero() globally ?
>
> inc_not_zero() has a very strong use-case, its for lockless refcount
> increment. Incrementing a 0 reference count is bad because the object
> will be freed and you'll have a use-after-free.
>
> Arguably, once we move reference counting over to its own type, it would
> make sense to remove it from atomic, specifically to discourage that use
> case.

Do you know about some commit doing such conversion or could you give me 2
distinct models
so I could compare ? btw atomic_inc_not_zero() is also adviced in
Documentation/RCU/rcuref.txt

Regards,
Fabian