Re: [PATCH 7/8] cpufreq: Preserve policy structure across suspend/resume

[Viresh Kumar <viresh.kumar@linaro.org>]

On 15 July 2013 15:35, Srivatsa S. Bhat
<srivatsa.bhat@linux.vnet.ibm.com> wrote:
> Actually even I was wondering about this while writing the patch and
> I even tested shutdown after multiple suspend/resume cycles, to verify that
> the refcount is messed up. But surprisingly, things worked just fine.
>
> Logically there should've been a refcount mismatch and things should have
> failed, but everything worked fine during my tests. Apart from suspend/resume
> and shutdown tests, I even tried mixing a few regular CPU hotplug operations
> (echo 0/1 to sysfs online files), but nothing stood out.
>
> Sorry, I forgot to document this in the patch. Either the patch is wrong
> or something else is silently fixing this up. Not sure what is the exact
> situation.

To understand it I actually applied your patches to get better view of the code.
(Haven't tested it though).. And found that your code is doing the right thing
and we shouldn't get a mismatch.. This is the sequence of events I can draw:

- __cpu_add_dev() for first cpu. sets the refcount to 'x', where x are
the no. of
cpus in its clock domain.
- _cpu_add_dev() for other cpus: doesn't change anything in refcount

- Suspend:
 - cpu_remove_dev() for all cpus, due to frozen flag we don't touch the value
of count
- Resume:
 - cpu_add_dev() for all cpus, due to frozen flag we don't touch the
value of count.

And so things work as expected. That's why your code isn't breaking anything I
believe.

But can no. of cpus change inbetween suspend/resume? Then count would be
tricky as we are using the same policy structure.