Re: CONFIG_PREEMPT_RCU in next/mmotm

["Paul E. McKenney" <paulmck@linux.vnet.ibm.com>]

On Wed, Aug 12, 2009 at 11:22:50AM +0200, Ingo Molnar wrote:
> 
> * Paul E. McKenney <paulmck@linux.vnet.ibm.com> wrote:
> 
> > Should these tests pass...

Things are working much better, but I can still cause failures by
hotplugging CPUs with zero wait time between them while concurrently
modprobe-ing and rmmod-ing rcutorture repeatedly while running
CONFIG_PREEMPT_RCU.  I increased the kernel's lifespan by an order of
magnitude or so by simplifying rcupreempt's hotplug code path.

And I just -know- that I will deeply regret having described my test
process, given that my life is much easier when my RCU testing is more
rigorous than anyone else's.  ;-)

The remaining (known) problem appears to be due to a kthread_stop()
deadlock between the migration threads and a few of rcutorture's kthreads.
In this deadlock, rcutorture is waiting for one of the fakewriters
to stop (via kthread_stop()), while the fakewriter is waiting for
synchronize_rcu() to complete.  The migration thread's CPU-hotplug
notifier is blocked in kthread_stop() because rcutorture holds the
kthread_stop() mutex.

I could argue that CPU hotplug should allow RCU grace periods to
proceed regardless, but I believe that the problem is that some thread
is preempted in the middle of an RCU read-side critical section, but
cannot be migrated to a CPU that could run it due to the fact that the
migration kthread is stuck in its CPU-hotplug notifier.  RCU being what
it is, it seems reasonable for me to instead arrange so that rcutorture
never invokes kthread_stop() unless it knows that the target thread cannot
possibly be in the midst of synchronize_rcu().  That said, there is the
concern that this general pattern might rear its ugly head elsewhere.

> > Unless someone tells me otherwise, I will make a patch series 
> > intended to replace tip/core/rcu commits 7fe616c5d ("Simplify RCU 
> > CPU-hotplug notification"), 04b06256c ("Fix RCU & CPU hotplug 
> > hang"), and 7256cf0e83b ("Add diagnostic check for a possible 
> > CPU-hotplug race"), re-run all tests on that patchset, and submit 
> > the series.  I expect the resulting patch set to have three 
> > patches, one to split out boot-time initialization for RCU_TREE, a 
> > second to create the cpu_notifier() API, and the third to make RCU 
> > use it.

While thinking this over, I am rebasing as described above, and doing
full-up testing at each step.  No more Mr. Nice Guy!!!  ;-)

In the meantime, can anyone tell me why we only let one kthread stop at
a time?

> Sure - we can reasonably rebase portions of that stack of commits:
> 
>  earth4:~/tip> gll linus..core/rcu
>  7256cf0: rcu: Add diagnostic check for a possible CPU-hotplug race
>  04b0625: rcu: Fix RCU & CPU hotplug hang
>  7fe616c: rcu: Simplify RCU CPU-hotplug notification
>  240ebbf: rcu: Add synchronize_sched_expedited() rcutorture doc + updates
>  0acc512: rcu: Add synchronize_sched_expedited() torture tests
>  03b042b: rcu: Add synchronize_sched_expedited() primitive
>  c17ef45: rcu: Remove Classic RCU
> 
> Please mention the magic words "please reset core/rcu to 240ebbf 
> before applying these patches" in the mail to me, should i forget in 
> the days to come.

Will do!

> (hm, what was i supposed to not forget? Weird.)

I can't remember.  ;-)

							Thanx, Paul