Re: [RFC PATCH] mm/vmscan: fix high cpu usage of kswapd if there

[hejianet]

Hi Michal

On 22/02/2017 11:48 PM, Michal Hocko wrote:
> On Wed 22-02-17 22:31:50, hejianet wrote:
> > Hi Michal
> >
> > On 22/02/2017 7:41 PM, Michal Hocko wrote:
> > > On Wed 22-02-17 17:04:48, Jia He wrote:
> > > > When I try to dynamically allocate the hugepages more than system total
> > > > free memory:
> > > > e.g. echo 4000 >/proc/sys/vm/nr_hugepages
> > >
> > > I assume that the command has terminated with less huge pages allocated
> > > than requested but
> > Yes, at last the allocated hugepages are less than 4000
> > HugePages_Total:    1864
> > HugePages_Free:     1864
> > HugePages_Rsvd:        0
> > HugePages_Surp:        0
> > Hugepagesize:      16384 kB
> >
> > In the bad case, although kswapd takes 100% cpu, the number of
> > HugePages_Total is not increase at all.
> >
> > > > Node 3, zone      DMA
> > > [...]
> > > >   pages free     2951
> > > >         min      2821
> > > >         low      3526
> > > >         high     4231
> > >
> > > it left the zone below high watermark with
> > >
> > > >    node_scanned  0
> > > >         spanned  245760
> > > >         present  245760
> > > >         managed  245388
> > > >       nr_free_pages 2951
> > > >       nr_zone_inactive_anon 0
> > > >       nr_zone_active_anon 0
> > > >       nr_zone_inactive_file 0
> > > >       nr_zone_active_file 0
> > >
> > > no pages reclaimable, so kswapd will not go to sleep. It would be quite
> > > easy and comfortable to call it a misconfiguration but it seems that
> > > it might be quite easy to hit with NUMA machines which have large
> > > differences in the node sizes. I guess it makes sense to back off
> > > the kswapd rather than burning CPU without any way to make forward
> > > progress.
> >
> > agree.
>
> please make sure that this information is in the changelog
>
> [...]
> > > > @@ -3502,6 +3503,7 @@ void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
> > > >  {
> > > >  	pg_data_t *pgdat;
> > > >  	int z;
> > > > +	int node_has_relaimable_pages = 0;
> > > >
> > > >  	if (!managed_zone(zone))
> > > >  		return;
> > > > @@ -3522,8 +3524,15 @@ void wakeup_kswapd(struct zone *zone, int order, enum zone_type classzone_idx)
> > > >
> > > >  		if (zone_balanced(zone, order, classzone_idx))
> > > >  			return;
> > > > +
> > > > +		if (!zone_reclaimable_pages(zone))
> > > > +			node_has_relaimable_pages = 1;
> > >
> > > What, this doesn't make any sense? Did you mean if (zone_reclaimable_pages)?
> >
> > I mean, if any one zone has reclaimable pages, then this zone's *node* has
> > reclaimable pages. Thus, the kswapN for this node should be waken up.
> > e.g. node 1 has 2 zones.
> > zone A has no reclaimable pages but zone B has.
> > Thus node 1 has reclaimable pages, and kswapd1 will be waken up.
> > I use node_has_relaimable_pages in the loop to check all the zones' reclaimable
> > pages number. So I prefer the name node_has_relaimable_pages instead of
> > zone_has_relaimable_pages
>
> I still do not understand. This code starts with
> node_has_relaimable_pages = 0. If you see a zone with no reclaimable
> pages then you make it node_has_relaimable_pages = 1 which means that
>
> > > > +	/* Dont wake kswapd if no reclaimable pages */
> > > > +	if (!node_has_relaimable_pages)
> > > > +		return;
>
> this will not hold and we will wake up the kswapd. I believe what
> you want instead, is to skip the wake up if _all_ zones have
> !zone_reclaimable_pages() Or I am missing your point. This means that
> you want
> 	if (zone_reclaimable_pages(zone)
> 		node_has_relaimable_pages = 1;	
You are right, will send v2 soon after testing it
B.R.
Jia
> > > > +
> > > >  	trace_mm_vmscan_wakeup_kswapd(pgdat->node_id, zone_idx(zone), order);
> > > >  	wake_up_interruptible(&pgdat->kswapd_wait);