Re: [PATCH Resend] cpufreq: remove sysfs files for CPU which failed to come back after resume

["Bjørn Mork" <bjorn@mork.no>]

Viresh Kumar <viresh.kumar@linaro.org> writes:
> On 23 December 2013 11:25, Bjørn Mork <bjorn@mork.no> wrote:
>> I can confirm that it fixes the major regression.  With this branch, the
>> cpufreq directory is completely removed after a cancelled userspace
>> hibernate (with the acpi-cpufreq problem causing failure).  So it is
>> possible to restore cpufreq by manually offlining and onlining non-boot
>> cores.  No more leftover sysfs attributes.
>
> Thanks for giving it a try once again :)
>
>> But there is still a minor regression compared to the old (v3.11)
>> behaviour: Previously the cpufreq functionality would be automatically
>> restored by any completed hibernate or suspend cycle, since it would
>> effectively do the CPU offline/online. This automatix fixup won't happen
>> with the current pm-cpufreq branch.
>
> I didn't understood it completely, sorry :)
>
> As far as I can see from 3.11 code we simply used to fail with any failure
> resulting with a call to ->init() or some other call..
>
> And so cpufreq wouldn't have added any directories at all in that case.
> And so I think we still required an offline/online sequence to guarantee
> things..

That's correct.  The immediate result of the failure is exactly the
same.

The difference is that a subsequent resume would restore the cpufreq
device whether it existed or not.  That made a complete suspend/resume
fix up any missing cpufreq device, e.g. one that was removed by a
previous error.

One effect of saving state on suspend is that a missing device isn't
added on resume.  You can of course see that as a feature.  But to me
it's a regression, because:
 - it didn't use to work that way, and
 - the addition of missing devices on resume is always wanted AFAICS.


Bjørn