On Fri, Jun 30, 2017 at 11:10:33AM +0530, Viresh Kumar wrote: > On 30-06-17, 06:53, Dominik Brodowski wrote: > > On Fri, Jun 30, 2017 at 09:04:25AM +0530, Viresh Kumar wrote: > > > On 29-06-17, 20:01, Dominik Brodowski wrote: > > > > On Thu, Jun 29, 2017 at 04:29:06PM +0530, Viresh Kumar wrote: > > > > > The cpufreq core and governors aren't supposed to set a limit on how > > > > > fast we want to try changing the frequency. This is currently done for > > > > > the legacy governors with help of min_sampling_rate. > > > > > > > > > > At worst, we may end up setting the sampling rate to a value lower than > > > > > the rate at which frequency can be changed and then one of the CPUs in > > > > > the policy will be only changing frequency for ever. > > > > > > > > Is it safe to issue requests to change the CPU frequency so frequently, > > > > > > Well, I assumed so. I am not sure the hardware would break though. > > > Overheating ? > > > > > > > even > > > > on historic hardware such as speedstep-{ich,smi,centrino}? In the past, > > speedstep-smi is the only one which sets transition_latency to > CPUFREQ_ETERNAL and the others are putting some meaningful values. So > yes, they should be doing DVFS dynamically. > > > > > these checks more or less disallowed the running of dynamic frequency > > > > scaling at least on speedstep-smi[*], > > > > > > We must by doing dynamic freq scaling even without this patch. I don't > > > see why you say the above then. > > > > > > All we do here is that we get rid of the limit on how soon we can > > > change the freq again. > > > > Well, as I understand it, first generation "speedstep" was designed more or > > less to switch frequencies only when AC power was lost or restored. > > > > The Linux implementation merely said: "no on-the-fly changes", but switch > > frequencies whenever a user explicitly requested such a change (presumably > > only every once in an unspecified while). > > > > This same reasoning may be present in other drivers using CPUFREQ_ETERNAL. > > Thanks for the explanation here and I am convinced that this series > has at least done one thing wrong. And that is removal of > max_transition_latency from governors and allowing ondemand to run on > such platforms (which may end up breaking them). > > So I will actually modify that patch and set max_transition_latency to > CPUFREQ_ETERNAL for ondemand/conservative instead of 10ms. Also we > should do the same for schedutil as well, so that will also use the > max_transition_latency field. > > But I hope, this patch will still be fine. Right ? Indeed, I have no comments otherwise. Thanks! Best Dominik
Attachment:
signature.asc
Description: PGP signature