Re: Severe performance regression w/ 4.4+ on Android due to cgroup locking changes

[Paul E. McKenney]

On Wed, Jul 13, 2016 at 10:51:02PM +0200, Peter Zijlstra wrote:
> On Wed, Jul 13, 2016 at 04:39:44PM -0400, Tejun Heo wrote:
> 
> > > There is a synchronize_sched() in there, so sorta. That thing is heavily
> > > geared towards readers, as is the only 'sane' choice for global locks.
> > 
> > It used to use the expedited variant until 001dac627ff3
> > ("locking/percpu-rwsem: Make use of the rcu_sync infrastructure"), so
> > it might have been okay before then.
> 
> Right, but expedited stuff sprays IPIs around the entire system. That's
> stuff other people complain about.

Do anyone other than the non-NO_HZ_FULL low-latency guys and the -rt
guys care?

> > The options that I can see are
> > 
> > 1. Somehow make percpu_rwsem's write behavior more responsive in a way
> >    which is acceptable all use cases.  This would be great but
> >    probably impossible.
> > 
> > 2. Add a fast-writer option to percpu_rwsem so that users which care
> >    about write latency can opt in for higher processing overhead for
> >    lower latency.
> 
> So, IIRC, the trade-off is a full memory barrier in read_lock and
> read_unlock() vs sync_sched() in write.
> 
> Full memory barriers are expensive and while the combined cost might
> well exceed the cost of the sync_sched() it doesn't suffer the latency
> issues.
> 
> Not sure if we can frob the two in a single codebase, but I can have a
> poke if Oleg or Paul doesn't beat me to it.

Take the patch that I just sent out and make the choice of normal
vs. expedited depend on CONFIG_PREEMPT_RT or whatever the -rt guys are
calling it these days.  Is there a low-latency Kconfig option other
than CONFIG_NO_HZ_FULL?

The memory-barrier approach can definitely be made to work, but is
going to be more complex due to the need to wait for readers.

							Thanx, Paul