Re: [RFC] [PATCH] memory controller background reclamation

[yamamoto-jCdQPDEk3idL9jVzuh4AOg@public.gmane.org (YAMAMOTO Takashi)]

hi,

> > --- linux-2.6.24-rc2-mm1-kame-pd/kernel/res_counter.c.BACKUP	2007-11-14 16:05:52.000000000 +0900
> > +++ linux-2.6.24-rc2-mm1-kame-pd/kernel/res_counter.c	2007-11-22 15:14:32.000000000 +0900
> > @@ -17,6 +17,8 @@ void res_counter_init(struct res_counter
> >  {
> >  	spin_lock_init(&counter->lock);
> >  	counter->limit = (unsigned long long)LLONG_MAX;
> > +	counter->high_watermark = (unsigned long long)LLONG_MAX;
> > +	counter->low_watermark = (unsigned long long)LLONG_MAX;
> 
> Should low watermark also be LLONG_MAX?

what else do you suggest?  0?
currently it doesn't matter much because low_watermark is not used at all
as far as high_watermark is LLONG_MAX.

> > +static void
> > +mem_cgroup_reclaim(struct work_struct *work)
> > +{
> > +	struct mem_cgroup * const mem =
> > +	    container_of(work, struct mem_cgroup, reclaim_work);
> > +	int batch_count = 128; /* XXX arbitrary */
> 
> Could we define and use something like MEM_CGROUP_BATCH_COUNT for now?
> Later we could consider and see if it needs to be tunable. numbers are
> hard to read in code.

although i don't think it makes sense, i can do so if you prefer.

> > +
> > +	for (; batch_count > 0; batch_count--) {
> > +		if (res_counter_below_low_watermark(&mem->res))
> > +			break;
> 
> Shouldn't we also check to see that we start reclaim in background only
> when we are above the high watermark?

i don't understand what you mean.  can you explain?
highwatermark is checked by mem_cgroup_charge_common before waking
these threads.

> I'll start some tests on these patches.

thanks.

YAMAMOTO Takashi