On Tue, Jun 06, 2017 at 04:45:57PM +0200, Christian Borntraeger wrote: > A the same time, the implicit memory barrier of the atomic_inc should be > even cheaper. In contrast to x86, a full smp_mb seems to be almost for > free (looks like <= 1 cycle for a bcr 14,0 and no contention). So I > _think_ that this should be really fast enough. So there is a patch out there that changes the x86 smp_mb() implementation to do "LOCK ADD some_stack_location, 0" which is lots cheaper than the "MFENCE" instruction and provides similar guarantees. HPA was running that through some of the architects.. ping? (Also, I can imagine OoO CPUs collapsing back-to-back ordering stuff, but what do I know). > As a side note, I am asking myself, though, why we do need the > preempt_disable/enable for the cases where we use the opcodes > like lao (atomic load and or to a memory location) and friends. I suspect the real reason is CPU hotplug, because regular preemption should not matter. It would be the same as getting migrated the moment _after_ you do the $op. But preempt_disable() also holds off hotplug and thereby serializes against hotplug notifiers that want to, for instance, move the value of the per-cpu variable to a still online CPU. Without this serialization it would be possible for the $op to happen _after_ the hotplug notifier runs. -- To unsubscribe from this list: send the line "unsubscribe linux-s390" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
