Re: [rcu] [ INFO: suspicious RCU usage. ]

[Krzysztof Kozlowski]

 On Åro, 2015-02-04 at 07:10 -0800, Paul E. McKenney wrote:
> On Wed, Feb 04, 2015 at 03:16:27PM +0100, Krzysztof Kozlowski wrote:
> > On Åro, 2015-02-04 at 05:14 -0800, Paul E. McKenney wrote:
> > > On Wed, Feb 04, 2015 at 01:00:18PM +0000, Russell King - ARM Linux wrote:
> > > > On Wed, Feb 04, 2015 at 12:39:07PM +0100, Krzysztof Kozlowski wrote:
> > > > > +Cc some ARM people
> > > > 
> > > > I wish that people would CC this list with problems seen on ARM.  I'm
> > > > minded to just ignore this message because of this in the hope that by
> > > > doing so, people will learn something...
> > > > 
> > > > > > Another thing I could do would be to have an arch-specific Kconfig
> > > > > > variable that made ARM responsible for informing RCU that the CPU
> > > > > > was departing, which would allow a call to as follows to be placed
> > > > > > immediately after the complete():
> > > > > > 
> > > > > > rcu_cpu_notify(NULL, CPU_DYING_IDLE, (void *)(long)smp_processor_id());
> > > > > > 
> > > > > > Note:  This absolutely requires that the rcu_cpu_notify() -always-
> > > > > > be allowed to execute!!!  This will not work if there is -any- possibility
> > > > > > of __cpu_die() powering off the outgoing CPU before the call to
> > > > > > rcu_cpu_notify() returns.
> > > > 
> > > > Exactly, so that's not going to be possible.  The completion at that
> > > > point marks the point at which power _could_ be removed from the CPU
> > > > going down.
> > > 
> > > OK, sounds like a polling loop is required.
> > 
> > I thought about using wait_on_bit() in __cpu_die() (the waiting thread)
> > and clearing the bit on CPU being powered down. What do you think about
> > such idea?
> 
> Hmmm...  It looks to me that wait_on_bit() calls out_of_line_wait_on_bit(),
> which in turn calls __wait_on_bit(), which calls prepare_to_wait() and
> finish_wait().  These are in the scheduler, but this is being called from
> the CPU that remains online, so that should be OK.
> 
> But what do you invoke on the outgoing CPU?  Can you get away with
> simply clearing the bit, or do you also have to do a wakeup?  It looks
> to me like a wakeup is required, which would be illegal on the outgoing
> CPU, which is at a point where it cannot legally invoke the scheduler.
> Or am I missing something?

Actually the timeout versions but I think that doesn't matter.
The wait_on_bit will busy-loop with testing for the bit. Inside the loop
it calls the 'action' which in my case will be bit_wait_io_timeout().
This calls schedule_timeout().

See proof of concept in attachment. One observed issue: hot unplug from
commandline takes a lot more time. About 7 seconds instead of ~0.5.
Probably I did something wrong.

> 
> You know, this situation is giving me a bad case of nostalgia for the
> old Sequent Symmetry and NUMA-Q hardware.  On those platforms, the
> outgoing CPU could turn itself off, and thus didn't need to tell some
> other CPU when it was ready to be turned off.  Seems to me that this
> self-turn-off capability would be a great feature for future systems!

There are a lot more issues with hotplug on ARM...

Patch/RFC attached.