On 25 July 2014 18:11, Prarit Bhargava [off-list ref] wrote:

Okay -- I think I got it:  The above happens only with an *older* sysfs stack
which acquires a &buffer->mutex.  This no longer happens with the new sysfs
stack so the above change is still no longer necessary in mainline linux.git (at
least AFAICT).

Hmm, that's something new for me. Would it be possible to give some commits
for reference where the behavior changed? This would be quite helpful once
we face it again and need to come back to this thread :)

In any case, the change still appears to be incorrect in that it breaks the
rwsem locking scheme of the cpufreq policy.  I'll do some additional testing on
various systems and get back to you in a few days.

What about breaking rwsem locking scheme?