Re: [PATCH linux-next v8] cpufreq: convert the cpufreq_driver to use the rcu

[Nathan Zimmer <nzimmer@sgi.com>]

On 04/29/2013 04:47 PM, Rafael J. Wysocki wrote:
> On Monday, April 29, 2013 02:37:28 PM Paul E. McKenney wrote:
>> On Mon, Apr 29, 2013 at 12:22:32AM +0200, Rafael J. Wysocki wrote:
>>> On Thursday, April 04, 2013 09:57:19 PM Viresh Kumar wrote:
>>>> On 4 April 2013 20:23, Nathan Zimmer <nzimmer@sgi.com> wrote:
>>>>> We eventually would like to remove the rwlock cpufreq_driver_lock or convert
>>>>> it back to a spinlock and protect the read sections with RCU.  The first step in
>>>>> that is moving the cpufreq_driver to use the rcu.
>>>>> I don't see an easy wasy to protect the cpufreq_cpu_data structure with the
>>>>> RCU, so I am leaving it with the rwlock for now since under certain configs
>>>>> __cpufreq_cpu_get is hot spot with 256+ cores.
>>>>>
>>>>> v5: Go a different way and split up the lock and use the rcu
>>>>> v6: use bools instead of checking function pointers
>>>>>      covert the cpufreq_data_lock to a rwlock
>>>>> v7: Rebase to use the already accepted half
>>>>> v8: Correct have_governor_per_policy
>>>>>      Reviewed location of rcu_read_(un)lock in several spots
>>>> Sorry for long delay or too many versions of this patch :)
>>>>
>>>> Acked-by: Viresh Kumar <viresh.kumar@linaro.org>
>>> Unfortunately, I had to revert this one, because it is obviously buggy.  Why?
>>> Because it adds rcu_read_lock()/rcu_read_unlock() around sysfs_create_file()
>>> which may sleep due to a GFP_KERNEL memory allocation.  Sorry for failing to
>>> notice that earlier.
>> One workaround might be to use SRCU, which allows sleeping in its
>> critical sections.
> Agreed, but at this point of the cycle I'd just preferred to do the revert and
> start over.
>
> Thanks,
> Rafael
>
>
Agreed, I think that would be cleanest.  I probably won't have time to 
get to it this week though.

Nate