Messages in this thread | | | Date | Thu, 25 Jun 2020 15:20:43 +0100 | From | Qais Yousef <> | Subject | Re: [PATCH] cros_ec_spi: Even though we're RT priority, don't bump cpu freq |
| |
On 06/24/20 11:29, Doug Anderson wrote: > Hi, > > On Wed, Jun 24, 2020 at 10:55 AM Joel Fernandes <joelaf@google.com> wrote: > > > > On Wed, Jun 24, 2020 at 1:52 PM Qais Yousef <qais.yousef@arm.com> wrote: > > > > > > On 06/24/20 13:35, Joel Fernandes wrote: > > > > > > [...] > > > > > > > > Doing the in-kernel opt-out via API should be fine, I think. But this will > > > > > need to be discussed in the wider circle. It will already clash with this for > > > > > example > > > > > > > > > > https://lore.kernel.org/lkml/20200619172011.5810-1-qais.yousef@arm.com/ > > > > > > > > Have not yet looked closer at that patch, but are you saying this > > > > patch clashes with that work? Sorry I am operating on 2 hours of sleep > > > > here. > > > > > > The series is an optimization to remove the uclamp overhead from the scheduler > > > fastpath until the userspace uses it. It introduces a static key that is > > > disabled by default and will cause uclamp logic not to execute in the fast > > > path. Once the userspace starts using util clamp, which we detect by either > > > > > > 1. Changing uclamp value of a task with sched_setattr() > > > 2. Modifying the default sysctl_sched_util_clamp_{min, max} > > > 3. Modifying the default cpu.uclamp.{min, max} value in cgroup > > > > > > If we start having in-kernel users changing uclamp value this means drivers > > > will cause the system to opt-in into uclamp automatically even if the > > > userspace doesn't actually use it. > > > > > > I think we can solve this by providing a special API to opt-out safely. Which > > > is the right thing to do anyway even if we didn't have this clash. > > > > Makes sense, thanks. > > OK, so I think the summary is: > > 1. There are enough external dependencies that are currently in the > works that it makes sense for those to land first without trying to > cram my patch to cros_ec in.
+1
> > 2. Maybe, as part of the work that's already going on, someone will > add an API that I can use. If so then I can write my patch once that > lands.
I won't be adding this API. Mainly because I can't argue for it personally as I'm still not convinced it's a valid way of handling RT default boosting behavior from within the kernel. But it's a valid discussion to have if you want to drive it.
> > 3. If nobody adds an API then I could look at adding the API myself > once everything else is settled. > > 4. It looks as if the patch you mentioned originally [1] that allows > userspace to just fully disable uclamp for RT tasks will land > eventually (if we're stuck for a short term solution we can pick the > existing patch). I believe Chrome OS will use that to just fully > disable the default boosting for RT tasks across our system and (if > needed) add boosts on a case-by-case basis. Once we do that, the > incentive to land a patch for cros_ec will be mostly gone and probably > we could consider my patch abandoned. While it would technically > still be sane to land it won't have any real-world benefit.
I think this is the best way forward. I'm interesting in hearing what difficulties you encounter while doing this work.
Regarding the patch [1], I need to tweak the way it is implemented and send v6, but there are no objection to the idea and interface, so hopefully once I send v6 it'd be accepted.
Thanks
-- Qais Yousef
> [1] https://lore.kernel.org/r/20200511154053.7822-1-qais.yousef@arm.com
| |