On Thu, Feb 25, 2016 at 10:08:48PM +0100, Rafael J. Wysocki wrote: > On Thursday, February 25, 2016 10:28:37 AM Peter Zijlstra wrote: > > Its vile though; one should not spray IPIs if one can avoid it. Such > > things are much better done with RCU. Sure sync_sched() takes a little > > longer, but this isn't a fast path by any measure. > > I see, thanks! > > BTW, when cpufreq_update_util() callbacks are removed, I use synchronize_rcu() > to wait for the running ones, but would it be better to use > synchronize_sched() > in there instead?
So I think we only call the callback with rq->lock held, in which case sync_sched() is good enough. It would allow you to get rid of the rcu_read_{,un}lock() calls as well. The down-side is that it all makes the code a little harder to get, because you're relying on caller context to DTRT.