On Mon, Feb 19, 2018 at 06:44:13PM +0100, Peter Zijlstra wrote: > On Mon, Feb 19, 2018 at 12:14:45PM -0500, Alan Stern wrote: > > Note that operations like atomic_add_unless() already include memory > > barriers. > > It is valid for atomic_add_unless() to not imply any barriers when the > addition doesn't happen.
Agreed, given that atomic_add_unless() just returns 0 or 1, not the pointer being added. Of course, the __atomic_add_unless() function that it calls is another story, as it does return the old value. Sigh. And __atomic_add_unless() is called directly from some code. All of which looks to be counters rather than pointers, thankfully. So, do we want to rely on atomic_add_unless() always being invoked on counters rather than pointers, or does it need an smp_read_barrier_depends()? Thanx, Paul