lkml.org 
[lkml]   [2019]   [Oct]   [11]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
Date
From
SubjectRe: [PATCH 1/4] arm64: add support for the AMU extension v1
On Fri, Oct 11, 2019 at 11:31:40AM +0100, Ionela Voinescu wrote:
> On 10/10/2019 18:20, Catalin Marinas wrote:
> > On Tue, Sep 17, 2019 at 02:42:25PM +0100, Ionela Voinescu wrote:
> >> +#ifdef CONFIG_ARM64_AMU_EXTN
> >> +
> >> +/*
> >> + * This per cpu variable only signals that the CPU implementation supports the
> >> + * AMU but does not provide information regarding all the events that it
> >> + * supports.
> >> + * When this amu_feat per CPU variable is true, the user of this feature can
> >> + * only rely on the presence of the 4 fixed counters. But this does not
> >> + * guarantee that the counters are enabled or access to these counters is
> >> + * provided by code executed at higher exception levels.
> >> + */
> >> +DEFINE_PER_CPU(bool, amu_feat) = false;
> >> +
> >> +static void cpu_amu_enable(struct arm64_cpu_capabilities const *cap)
> >> +{
> >> + if (has_cpuid_feature(cap, SCOPE_LOCAL_CPU)) {
> >> + pr_info("detected CPU%d: Activity Monitors extension\n",
> >> + smp_processor_id());
> >> + this_cpu_write(amu_feat, true);
> >> + }
> >> +}
> >
> > Sorry if I missed anything as I just skimmed through this series. I
> > can't see the amu_feat used anywhere in these patches, so on its own it
> > doesn't make much sense.
>
> No worries, you are correct, at the moment the per-cpu amu_feat is not
> yet used anywhere. But the intention is to use it to discover the
> feature at CPU level as some CPUs might implement AMU and some might
> not.
>
> I'll soon submit some patches using the counters for the scheduler's
> frequency invariance - to discover the frequency the CPUs are actually
> running at in case there is power or thermal mitigation happening
> outside of the OS.

Thanks for the explanation. I guess I'll wait for the other patches to
see how all fits together. In general I'm not keen on per-CPU variables
exposed to the rest of the kernel. For example, is it always read in a
non-preemptible context? I'd rather have an accessor function with the
corresponding WARN_ON_ONCE(preemptible()). This may come with the rest
of the patches.

> More practically, it's possible that we'll see big.LITTLE platforms
> where the big CPUs only will implement activity monitors and for those
> CPUs it will be useful to get more accurate information on the current
> frequency, for example, as power and thermal mitigation is more
> probable to happen in the power domain of the big CPUs.

As long as that's a realistic possibility (not just a theoretical one)
and the in-kernel code can handle such asymmetry, it's fine by me. This
could be another reason to never expose the AMU counters to user-space
or guests. You can control preemption in the kernel but can't run
user-space in a non-preemptible context.

--
Catalin

\
 
 \ /
  Last update: 2019-10-11 16:32    [W:0.570 / U:0.128 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site