Re: [PATCH 05/10] nohz: New tick dependency mask

[Peter Zijlstra <peterz@infradead.org>]

On Fri, Jul 24, 2015 at 01:43:38PM -0400, Chris Metcalf wrote:
> On 07/24/2015 01:16 PM, Frederic Weisbecker wrote:
> >On Fri, Jul 24, 2015 at 12:55:35PM -0400, Chris Metcalf wrote:
> >>On 07/23/2015 12:42 PM, Frederic Weisbecker wrote:
> >>>+unsigned long __tick_nohz_set_tick_dependency(enum tick_dependency_bit bit,
> >>>+					      unsigned long *dep)
> >>>+{
> >>>+	unsigned long prev;
> >>>+	unsigned long old = *dep;
> >>>+	unsigned long mask = BIT_MASK(bit);
> >>>+
> >>>+	while ((prev = cmpxchg(dep, old, old | mask)) != old) {
> >>>+		old = prev;
> >>>+		cpu_relax();
> >>>+	}
> >>>+
> >>>+	return prev;
> >>>+}
> >>Why not use set_bit() here?  It is suitably atomic.
> >Because I don't want to send an IPI if the CPU already had bits set in
> >the dependency.
> >
> >Ideally I need something like test_and_set_bit() but which returns the
> >whole previous value and not just the previous value of the bit.
> 
> Ah, of course.  Peter, maybe we need atomic_or_return() as part
> of your new atomic_or/_and/_xor series?  Certainly on tilegx, and
> likely other architectures, we can do better than Frederic's
> cmpxchg() loop.

No, atomic_or_return() would return the new value and is entirely
pointless for the logic ops since they're not reversible (with the
exception of xor).

What you'd need is atomic_fetch_or(), but we don't have any fetch_$op
primitives at all. Introducing them might make sense, but it'll have to
be a separate series.

Note that I have a fetch_or() macro in kernel/sched/core.c.