I wrote:
> I'm attaching a patch which changes a semaphore in cpufreq into a 
> completion. With this patch, my system runs OK even with cpufreqd.
> 

Although the patch worked for me, it was probably bogus.
The real reason why cpufreq caused problems was that it does:
   init_MUTEX_LOCKED(&policy->lock);
and later:
   up(&policy->lock);
where policy->lock is declared as:
   struct semaphore        lock;

In PREEMPT_RT, the init_MUTEX_LOCKED is defined in include/linux/rt_lock.h :
   /*
    * No locked initialization for RT semaphores:
    */
   #define init_MUTEX_LOCKED(sem) compat_init_MUTEX_LOCKED(sem)
(BTW, I don't understand why we have init_MUTEX but no init_MUTEX_LOCKED 
for RT semaphores).

Therefore we have an inconsistency - the up() gets translated into 
rt_up() because the lock is of type struct semaphore. However, 
compat_init_MUTEX_LOCKED assumes that it has a struct compat_semaphore.
The compiler actually warns about incompatible pointers but I missed 
that warning at first.
So the fix is to change the lock type into compat_semaphore. I'm 
attaching the patch. It works for me with 2.6.12-rc5-RT-V0.7.47-12.

Regards
Michal