On Thursday, August 01, 2013 10:54:08 PM Viresh Kumar wrote:
On 1 August 2013 20:54, Srivatsa S. Bhat srivatsa.bhat@linux.vnet.ibm.com wrote:
Well, I now agree that we don't have to keep the module refcount non-zero as long as CPUs are being managed (that was just my misunderstanding, sorry for the noise). However, I think the _get() and _put() used in the existing code is for synchronization: that is, to avoid races between trying to unload the cpufreq driver module and running a hotplug notifier (and calling the driver module's ->init() or ->exit() function).
With that being the case, I think we can retain the module refcounts and use them only for synchronization. And naturally the refcount should drop to zero after the critical section; no point keeping it incremented until the CPU is taken offline.
Or, am I confused again?
No, you aren't.
But, for synchronization we need some blocking stuff, so that when user tries to rmmod the module, it should wait until other critical sections are over and then unload the module. But here, we are simply returning from rmmod, saying that we are busy :)
So, for that kind of synchronization we better use locks available inside cpufreq core of if required another variable which can be used for refcount. But now this.
I suppose you wanted to say "But not this"?
I agree. That said the situation now is that for acpi-cpufreq the module usage count is 0 (as of linux-next from today), so you can actually rmmod it.
I don't see why it should be different in the case when there are multiple CPUs for the same policy and hence my patch. [I don't see the problem with it as I said to my reply to Srivatsa, so care to explain it to me?.]
Now it seems to me that the code *is* racy (I haven't verified that, though), so I agree that we should use some other synchronization framework to be able to rmmod and modprobe cpufreq drivers safely. Using the driver module usage counter for that is not quite correct in my view.
Thanks, Rafael