On 17-02-17, 13:48, Peter Zijlstra wrote: > On Fri, Feb 17, 2017 at 01:15:56PM +0100, Rafael J. Wysocki wrote: > > On Thursday, February 16, 2017 01:36:05 PM Peter Zijlstra wrote: > > > On Thu, Feb 16, 2017 at 03:42:10PM +0530, Viresh Kumar wrote: > > > > But when I discussed this with Vincent, he suggested that it may not be > > > > required > > > > at all as the scheduler (with the helped of "decayed") doesn't call into > > > > schedutil too often, i.e. at least 1 ms. And if the CPUs are stable > > > > enough (i.e. > > > > no interruptions to the running task), we wouldn't reevaluate before > > > > the next > > > > tick. > > > > > > There are still the attach/detach callers to cfs_rq_util_change() that > > > kick in for fork/exit and migration. > > > > > > But yes, barring those we shouldn't end up calling it at silly rates. > > > > OK > > > > Does this mean that running governor computations every time its callback > > is invoked by the scheduler would be fine? > > I'd say yes right up till the point someone reports a regression ;-)
@Rafael: Do you want me to send a V2 with the changes you suggested in commit log? -- viresh
