On 07-06-18, 10:32, Peter Zijlstra wrote: > On Thu, Jun 07, 2018 at 10:18:27AM +0200, Daniel Lezcano wrote: > > So IIUC, neither atomic or WRITE|READ_ONCE are necessary in this code > > because of the wake_up_process() barrier is enough, right ? > > I didn't look hard enough; if there ever is a time where the loads and > stores happen concurrently, you need READ/WRITE_ONCE(). If there is no > concurrency on the variables, you don't need anything. > > Neither atomic_read/set() nor REAd/WRITE_ONCE() will help with ordering, > which is what the wake_up_process() would provide here, different things > entirely.
Right and you still need the READ/WRITE_ONCE() thing as idle_injection_set_duration() may run in parallel with the idle_injection_fn() thread. And I don't think the purpose of atomic_read/write was ever to take care of the ordering issues in this code, it was always about parallel loads/stores. -- viresh
