Re: [RFC] sched: Limit idle_balance() when it is being used too frequently

[Jason Low <jason.low2@hp.com>]

On Fri, 2013-07-19 at 20:37 +0200, Peter Zijlstra wrote:
> On Thu, Jul 18, 2013 at 12:06:39PM -0700, Jason Low wrote:
> 
> > N = 1
> > -----
> > 19.21%  reaim  [k] __read_lock_failed                     
> > 14.79%  reaim  [k] mspin_lock                             
> > 12.19%  reaim  [k] __write_lock_failed                    
> > 7.87%   reaim  [k] _raw_spin_lock                          
> > 2.03%   reaim  [k] start_this_handle                       
> > 1.98%   reaim  [k] update_sd_lb_stats                      
> > 1.92%   reaim  [k] mutex_spin_on_owner                     
> > 1.86%   reaim  [k] update_cfs_rq_blocked_load              
> > 1.14%   swapper  [k] intel_idle                              
> > 1.10%   reaim  [.] add_long                                
> > 1.09%   reaim  [.] add_int                                 
> > 1.08%   reaim  [k] load_balance                            
> 
> But but but but.. wth is causing this? The only thing we do more of with
> N=1 is idle_balance(); where would that cause __{read,write}_lock_failed
> and or mspin_lock() contention like that.
> 
> There shouldn't be a rwlock_t in the entire scheduler; those things suck
> worse than quicksand.
> 
> If, as Rik thought, we'd have more rq->lock contention, then I'd
> expected _raw_spin_lock to be up highest.

For this particular fserver workload, that mutex was acquired in the
function calls from ext4_orphan_add() and ext4_orphan_del(). Those read
and write lock calls were from start_this_handle(). 

Although these functions are not called within the idle_balance() code
path, update_sd_lb_stats(), tg_load_down(), idle_cpu(), spin_lock(),
ect... increases the time spent in the kernel and that appears to be
indirectly causing more time to be spent acquiring those other kernel
locks.

Thanks,
Jason