Re: [PATCH] cpufreq: fix garbage kobj on errors during suspend/resume

["Srivatsa S. Bhat" <srivatsa.bhat@linux.vnet.ibm.com>]

On 12/06/2013 03:59 AM, Rafael J. Wysocki wrote:
> On Thursday, December 05, 2013 06:11:19 PM Srivatsa S. Bhat wrote:
>> On 12/05/2013 12:27 PM, Bjørn Mork wrote:
>>> Viresh Kumar <viresh.kumar@linaro.org> writes:
>>>
>>>> Sending from phone.. html.. so left list.
>>>>
>>>> Here is the old thread where we discussed this.. see if this helps..
>>>>
>>>> http://marc.info/?t=136845016900003&r=1&w=2
>>>
>>> Thanks.  That helped a lot.
>>>
>>> Unless I miss something, it looks like the permission problem *started*
>>> with fallout from special suspend code - surprising the user by not
>>> creating any offline/online event on suspend/resume.  Quoting from
>>> http://marc.info/?l=linux-pm&m=136847781510358 :
>>>
>>>   (And yes, even code-wise, we use a slightly different
>>>    path in the S3 code, to initiate hotplug. That's why the uevents
>>>    are by-passed.)
>>>
>>
>> I hope you didn't miss the main idea I was trying to convey in that
>> reply: 
>> "IMHO, using CPU hotplug (offline/online of CPUs) in the S3 path is
>> supposed to be totally internal to the suspend/resume code. It is not
>> intended for userspace to know that we are internally offlining/
>> onlining CPUs."
> 
> By the way, in the meantime I discussed this with Viresh in the context of
> a different (although related) fix and I suggested a different approach.
> 
> Namely, to split the CPU offline/online code into "core" and "add-ons"
> parts, where the core part will do just whatever is needed to offline/online
> CPU cores cleanly and the "add-ons" part will carry out the rest (e.g.
> removal/addition of sysfs attributes and so on).
> 
> Then, the system suspend/resume code path will only run the "core" part
> and whatever else CPU handling is needed during suspend/resume will be
> carried out by the device suspend/resume code (via driver callbacks or
> stuff similar to cpufreq_suspend() and cpufreq_resume() recently proposed
> by Viresh).
> 
> In turn, the "runtime" CPU offline/online will carry out both the core
> and the add-ons parts as it does today.
> 
> In my view this should address the problems we have with sysfs attributes,
> governors start/stop etc. during system suspend/resume.
> 

Hmm, yes that sounds like a good idea. Are you suggesting this "core" and
"add-on" split only for the cpufreq parts of CPU hotplug or for everything
involved in CPU hotplug code? IIUC you are suggesting the latter, which is
likely to be a significant undertaking, but very well worth it in the long
run, since it gives us an elegant solution for all these problems.

I guess the *_FROZEN CPU hotplug notifications were originally introduced
to provide us an infrastructure to do something like this, but obviously
that hasn't worked out very well. So I agree that a fundamental restructuring
is in order, to cure all the innumerable problems.

Regards,
Srivatsa S. Bhat