Re: [RFC PATCH] mm: exclude memory less nodes from zone_reclaim

[Michal Hocko]

On Wed 19-02-14 09:16:28, Nishanth Aravamudan wrote:
> On 19.02.2014 [18:03:03 +0100], Michal Hocko wrote:
> > We had a report about strange OOM killer strikes on a PPC machine
> > although there was a lot of swap free and a tons of anonymous memory
> > which could be swapped out. In the end it turned out that the OOM was
> > a side effect of zone reclaim which wasn't doesn't unmap and swapp out
> > and so the system was pushed to the OOM. Although this sounds like a bug
> > somewhere in the kswapd vs. zone reclaim vs. direct reclaim interaction
> > numactl on the said hardware suggests that the zone reclaim should
> > have been set in the first place:
> > node 0 cpus: 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
> > node 0 size: 0 MB
> > node 0 free: 0 MB
> > node 2 cpus:
> > node 2 size: 7168 MB
> > node 2 free: 6019 MB
> > node distances:
> > node   0   2
> > 0:  10  40
> > 2:  40  10
> > 
> > So all the CPUs are associated with Node0 which doesn't have any memory
> > while Node2 contains all the available memory. Node distances cause an
> > automatic zone_reclaim_mode enabling.
> > 
> > Zone reclaim is intended to keep the allocations local but this doesn't
> > make any sense on the memory less nodes. So let's exlcude such nodes
> > for init_zone_allows_reclaim which evaluates zone reclaim behavior and
> > suitable reclaim_nodes.
> > 
> > Signed-off-by: Michal Hocko <mhocko@xxxxxxx>
> > ---
> > I haven't got to testing this so I am sending this as an RFC for now.
> > But does this look reasonable?
> > 
> >  mm/page_alloc.c | 5 +++--
> >  1 file changed, 3 insertions(+), 2 deletions(-)
> > 
> > diff --git a/mm/page_alloc.c b/mm/page_alloc.c
> > index 3e953f07edb0..4a44bdc7a8cf 100644
> > --- a/mm/page_alloc.c
> > +++ b/mm/page_alloc.c
> > @@ -1855,7 +1855,7 @@ static void __paginginit init_zone_allows_reclaim(int nid)
> >  {
> >  	int i;
> > 
> > -	for_each_online_node(i)
> > +	for_each_node_state(i, N_HIGH_MEMORY)
> >  		if (node_distance(nid, i) <= RECLAIM_DISTANCE)
> >  			node_set(i, NODE_DATA(nid)->reclaim_nodes);
> >  		else
> > @@ -4901,7 +4901,8 @@ void __paginginit free_area_init_node(int nid, unsigned long *zones_size,
> > 
> >  	pgdat->node_id = nid;
> >  	pgdat->node_start_pfn = node_start_pfn;
> > -	init_zone_allows_reclaim(nid);
> > +	if (node_state(nid, N_HIGH_MEMORY))
> > +		init_zone_allows_reclaim(nid);
> 
> I don't think this will work, because what sets N_HIGH_MEMORY (and
> shouldn't it be N_MEMORY?)

This should be the same thing AFAIU.

> is check_for_memory() (free_area_init_nodes() for N_MEMORY), which is
> run *after* init_zone_allows_reclaim().

early_calculate_totalpages sets the memory state before we get here.

> Further, the for_each_node_state() loop doesn't make sense at this
> point, becuase we are actually setting up the nids as we go. So node
> 0, will only see node 0 in the N_HIGH_MEMORY mask (if any). Node 1,
> will only see nodes 0 and 1, etc.

I am not sure I understand what you are saying here. Why would se
consider distance to a memoryless node in the first place.
for_each_node_state just makes sure we are comparing only to a node
which has some memory.

> I'm working on testing a patch that reorders some of this in hopefully a
> safe way.

Although it might make some sense to reorganize the code (it's a mess if
you ask me), but I do not think it is necessary.

-- 
Michal Hocko
SUSE Labs
--
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/