Re: [PATCH 1/9] mm: fix 100% CPU kswapd busyloop on unreclaimable nodes

[Mel Gorman <mgorman@suse.de>]

On Tue, Mar 07, 2017 at 11:56:31AM -0500, Johannes Weiner wrote:
> On Tue, Mar 07, 2017 at 11:17:02AM +0100, Michal Hocko wrote:
> > On Mon 06-03-17 11:24:10, Johannes Weiner wrote:
> > > @@ -3271,7 +3271,8 @@ static int balance_pgdat(pg_data_t *pgdat, int order, int classzone_idx)
> > >  		 * Raise priority if scanning rate is too low or there was no
> > >  		 * progress in reclaiming pages
> > >  		 */
> > > -		if (raise_priority || !sc.nr_reclaimed)
> > > +		nr_reclaimed = sc.nr_reclaimed - nr_reclaimed;
> > > +		if (raise_priority || !nr_reclaimed)
> > >  			sc.priority--;
> > >  	} while (sc.priority >= 1);
> > >  
> > 
> > I would rather not play with the sc state here. From a quick look at
> > least 
> > 	/*
> > 	 * Fragmentation may mean that the system cannot be rebalanced for
> > 	 * high-order allocations. If twice the allocation size has been
> > 	 * reclaimed then recheck watermarks only at order-0 to prevent
> > 	 * excessive reclaim. Assume that a process requested a high-order
> > 	 * can direct reclaim/compact.
> > 	 */
> > 	if (sc->order && sc->nr_reclaimed >= compact_gap(sc->order))
> > 		sc->order = 0;
> > 
> > does rely on the value. Wouldn't something like the following be safer?
> 
> Well, what behavior is correct, though? This check looks like an
> argument *against* resetting sc.nr_reclaimed.
> 
> If kswapd is woken up for a higher order, this check sets a reclaim
> cutoff beyond which it should give up on the order and balance for 0.
> 
> That's on the scope of the kswapd invocation. Applying this threshold
> to the outcome of just the preceeding priority seems like a mistake.
> 
> Mel? Vlastimil?

I cannot say which is definitely the correct behaviour. The current
behaviour is conservative due to the historical concerns about kswapd
reclaiming the world. The hazard as I see it is that resetting it *may*
lead to more aggressive reclaim for high-order allocations. That may be a
welcome outcome to some that really want high-order pages and be unwelcome
to others that prefer pages to remain resident.

However, in this case it's a tight window and problems would be tricky to
detect. THP allocations won't trigger the behaviour and with vmalloc'd
stack, I'd expect that only SLUB-intensive workloads using high-order
pages would trigger any adverse behaviour. While I'm mildly concerned, I
would be a little surprised if it actually caused runaway reclaim.

-- 
Mel Gorman
SUSE Labs