On Fri, 5 Jul 2019, Thomas Gleixner wrote:
> On Fri, 5 Jul 2019, Mathieu Desnoyers wrote:
> > ----- On Jul 5, 2019, at 4:49 AM, Ingo Molnar mi...@kernel.org wrote:
> > > * Mathieu Desnoyers <mathieu.desnoy...@efficios.com> wrote:
> > >> The semantic I am looking for here is C11's relaxed atomics.
> > >
> > > What does this mean?
> >
> > C11 states:
> >
> > "Atomic operations specifying memory_order_relaxed are relaxed only with
> > respect
> > to memory ordering. Implementations must still guarantee that any given
> > atomic access
> > to a particular atomic object be indivisible with respect to all other
> > atomic accesses
> > to that object."
> >
> > So I am concerned that num_online_cpus() as proposed in this patch
> > try to access __num_online_cpus non-atomically, and without using
> > READ_ONCE().
> >
> >
> > Similarly, the update-side should use WRITE_ONCE(). Protecting with a mutex
> > does not provide mutual exclusion against concurrent readers of that
> > variable.
>
> Again. This is nothing new. The current implementation of num_online_cpus()
> has no guarantees whatsoever.
>
> bitmap_hweight() can be hit by a concurrent update of the mask it is
> looking at.
>
> num_online_cpus() gives you only the correct number if you invoke it inside
> a cpuhp_lock held section. So why do we need that fuzz about atomicity now?
>
> It's racy and was racy forever and even if we add that READ/WRITE_ONCE muck
> then it still wont give you a reliable answer unless you hold cpuhp_lock at
> least for read. So fore me that READ/WRITE_ONCE is just a cosmetic and
> misleading reality distortion.
That said. If it makes everyone happy and feel better, I'm happy to add it
along with a bit fat comment which explains that it's just preventing a
theoretical store/load tearing issue and does not provide any guarantees
other than that.
Thanks,
tglx
