Re: [PATCH] Inline double_unlock_balance()

[Peter Zijlstra <a.p.zijlstra@chello.nl>]

On Thu, 2008-11-06 at 08:32 +0100, Ingo Molnar wrote:
> * Peter Zijlstra <a.p.zijlstra@chello.nl> wrote:
> 
> > On Wed, 2008-11-05 at 18:57 +0530, Sripathi Kodi wrote:
> > > Hi,
> > > 
> > > We have a test case which measures the variation in the amount of time 
> > > needed to perform a fixed amount of work on the preempt_rt kernel. We 
> > > started seeing deterioration in it's performance recently. The test 
> > > should never take more than 10 microseconds, but we started 5-10% 
> > > failure rate. Using elimination method, we traced the problem to commit 
> > > 1b12bbc747560ea68bcc132c3d05699e52271da0 (lockdep: re-annotate 
> > > scheduler runqueues). When LOCKDEP is disabled, this patch only adds an 
> > > additional function call to double_unlock_balance(). Hence I inlined 
> > > double_unlock_balance() and the problem went away. Here is a patch to 
> > > make this change.
> > > 
> > > Thanks,
> > > Sripathi.
> > > 
> > > lockdep: Inline double_unlock_balance()
> > > 
> > > Additional function call for double_unlock_balance() causes latency 
> > > problems for some test cases on the preempt_rt kernel.
> > > 
> > > Signed-off-by: Sripathi Kodi <sripathik@in.ibm.com>
> > 
> > Acked-by; Peter Zijlstra <a.p.zijlstra@chello.nl>
> 
> hm, i'm not sure why it makes such a difference. Possibly cache 
> alignment or code generation details pushing the critical path just 
> beyond the L1 cache limit and causing thrashing?
> 
> Anyway, i've applied it to tip/sched/rt, as we generally want to 
> inline such short locking ops.

I'm thinking sripathi's gcc had a massive brainfart and did something
funny, maybe the extra register pressure from the calling convention
messed it up.

He failed to quantify the exact benefit, ie scheduling cost/latency
before and after and what platform. But still the patch is simple
enough.