Messages in this thread | | | From | Wei Wang <> | Date | Fri, 24 Jan 2020 11:12:58 -0800 | Subject | Re: [PATCH] [RFC] sched: restrict iowait boost for boosted task only |
| |
On Fri, Jan 24, 2020 at 3:01 AM Valentin Schneider <valentin.schneider@arm.com> wrote: > > On 24/01/2020 09:51, Quentin Perret wrote: > >>> +static inline bool iowait_boosted(struct task_struct *p) > >>> +{ > >>> + return p->in_iowait && uclamp_eff_value(p, UCLAMP_MIN) > 0; > >> > >> I think this is overloading the usage of util clamp. You're basically using > >> cpu.uclamp.min to temporarily switch iowait boost on/off. > >> > >> Isn't it better to add a new cgroup attribute to toggle this feature? > >> > >> The problem does seem generic enough and could benefit other battery-powered > >> devices outside of the Android world. I don't think the dependency on uclamp && > >> energy model are necessary to solve this. > > > > I think using uclamp is not a bad idea here, but perhaps we could do > > things differently. As of today the iowait boost escapes the clamping > > mechanism, so one option would be to change that. That would let us set > > a low max clamp in the 'background' cgroup, which in turns would limit > > the frequency request for those tasks even if they're IO-intensive. > >
Something we see e.g. is f2fs's gc thread, and my thought on this is instead of chasing everything down, it is a lot easier for us to only boost what we know in foreground (and on Android we sort of have that information on hand from framework). I was hesitant to introduce a new switch ( e.g. Android's older EAS kernel prefer_idle toggle ) but would be happy to do that if someone supports that idea.
> > > > So I'm pretty sure we *do* want tasks with the default clamps to get iowait > boost'd. What we don't want are background tasks driving up the frequency, > and that should be via uclamp.max (as Quentin is suggesting) rather than > uclamp.min (as is suggested in the patch). > > Now, whether that is overloading the usage of uclamp... I'm not sure. > One of the argument for uclamp was actually frequency selection, so if > we just make iowait boost respect that, IOW not boost further than > uclamp.max (which is a bit better than a simple on/off switch), that > wouldn't be too crazy I think. > > > > That'll have to be done at the RQ level, but figuring out what's the > > current max clamp on the rq should be doable from sugov I think. > > > > Wei: would that work for your use case ? > > > > Also, the iowait boost really should be per-task and not per-cpu, so it > > can be taken into account during wake-up balance on big.LITTLE. That > > might also help on SMP if a task doing a lot of IO migrates, as the > > boost would migrate with it. But that's perhaps for later ... > > > > Thanks, > > Quentin > >
| |