Re: [RFC PATCH 2/2] Add per-cgroup CPU controller statistics

[Bharata B Rao <bharata@linux.vnet.ibm.com>]

On Wed, Feb 25, 2009 at 12:04:10PM +0100, Peter Zijlstra wrote:
> On Wed, 2009-02-25 at 16:29 +0530, Bharata B Rao wrote:
> > From: Balaji Rao <balajirrao@gmail.com>
> > 
> > sched: Add cpu controller statistics
> > 
> > Add per-cgroup cpu controller statistics like system time and user time
> > consumed by groups of tasks.
> 
> Do we want this unconditionally?

Not sure if I understand you correctly here, but we are collecting
stats only when CONFIG_CGROUP_SCHED.

Or do you mean we need a config option to enable stats collection ?
If so, we have memory controller already providing stats unconditionally.

> >  
> > +#ifdef CONFIG_CGROUP_SCHED
> > +static void account_task_group_time(struct task_struct *p,
> > +			enum cpu_cgroup_stat_index idx, int val)
> > +{
> > +	struct task_group *tg = task_group(p);
> > +
> > +	if (likely(tg->stat))
> > +		percpu_counter_add(&tg->stat->cpustat[idx], val);
> > +}
> > +#else
> > +#define account_task_group_time(x, y, z)	{ 0; }
> 
> inline please, so we get argument validation.

Actually wanted to inline, but one of the arguments (cpu_cgroup_stat_index)
is defined only CONFIG_CGROUP_SCHED. Hence didn't want argument
validation here.

> > +
> > +	for (i = 0; i < CPU_CGROUP_STAT_NSTATS; i++)
> > +		percpu_counter_init(&tg->stat->cpustat[i], 0);
> 
> percpu_counter_init() can fail with -ENOMEM.

Ok, will fix.

Thanks for your review, Peter.

Regards,
Bharata.