On Sat, Oct 05, 2024 at 03:01:45PM GMT, Linus Torvalds wrote:
> On Sat, 5 Oct 2024 at 14:42, Linus Torvalds
> <torvalds@xxxxxxxxxxxxxxxxxxxx> wrote:
> >
> > and I think that might work, although the zero count case worries me
> > (ie 'fput twice').
> >
> > Currently we avoid the fput twice because we use that
> > "inc_not_zero()". So that needs some thinking about.
>
> Actually, it's worse. Even the
>
> if (ret > 1)
>
> case is dangerous, because we could have two or more threads doing
> that atomic_inc_return() on a dead file descriptor at the same time.
>
> So that approach is just broken.
Iiuc, then we should retain the deadzone handling but should replace
atomic_long_add_negative() with atomic_long_add_negative_relaxed().
So I would add:
static inline __must_check bool rcuref_long_inc(rcuref_long_t *ref)
{
/*
* Unconditionally increase the reference count with full
* ordering. The saturation and dead zones provide enough
* tolerance for this.
*/
if (likely(!atomic_long_add_negative(1, &ref->refcnt)))
return true;
/* Handle the cases inside the saturation and dead zones */
return rcuref_long_get_slowpath(ref);
}
Or did I miss something else?
