Re: [PATCH] cpuacct: VIRT_CPU_ACCOUNTING don't prevent percpu cputime count

[Balbir Singh <balbir@linux.vnet.ibm.com>]

* KOSAKI Motohiro <kosaki.motohiro@jp.fujitsu.com> [2009-04-29 11:26:30]:

> Hi
> 
> Thanks for reviewing.
> 
> > >  /* return cpu accounting group corresponding to this container */
> > >  static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
> > > @@ -10146,6 +10147,9 @@ static struct cgroup_subsys_state *cpuac
> > >  	if (!ca->cpuusage)
> > >  		goto out_free_ca;
> > > 
> > > +	if (!cpuacct_batch)
> > > +		cpuacct_batch = jiffies_to_cputime(percpu_counter_batch);
> > > +
> > 
> > I expect cpuacct_batch to be a large number
> 
> Yes.
> 
> 
> > 
> > >  	for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
> > >  		if (percpu_counter_init(&ca->cpustat[i], 0))
> > >  			goto out_free_counters;
> > > @@ -10342,7 +10346,7 @@ static void cpuacct_update_stats(struct 
> > >  	ca = task_ca(tsk);
> > > 
> > >  	do {
> > > -		percpu_counter_add(&ca->cpustat[idx], val);
> > > +		__percpu_counter_add(&ca->cpustat[idx], val, cpuacct_batch);
> > 
> > This will make the end result very off the real value due to large
> > batch value per cpu. If we are going to go this route, we should
> > probably consider using __percpu_counter_sum so that the batch value
> > does not show data that is way off.
> 
> No problem.
> 
> end-user don't see cputime itself. they see converted time.
> cpuacct_stats_show() use cputime64_to_clock_t. it mean
> the value less than 10msec don't display.
>

Yes, I know, I reviewed Bharata's patch and suggested converting to
clock_t for consistency with other metrics.
 
> --------------------------------------------------------
> static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
>                 struct cgroup_map_cb *cb)
> {
>         struct cpuacct *ca = cgroup_ca(cgrp);
>         int i;
> 
>         for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
>                 s64 val = percpu_counter_read(&ca->cpustat[i]);

My point is, this should probably be percpu_counter_sum(), but that
can be expensive and we were willing to tollerate some inaccuracy due
to batch value, I think your patch adds to the inaccuracy even more,
even though it fixes a genuine problem.


>                 val = cputime64_to_clock_t(val);
>                 cb->fill(cb, cpuacct_stat_desc[i], val);
>         }
>         return 0;
> }
> --------------------------------------------------------
> 
> 
> 
> 
> 

-- 
	Balbir