Re: [RFC 1/2] slub: Avoid trying to allocate memory on offline nodes

[Jeremy Linton]

Hi,

On 08/02/2018 09:23 AM, Christopher Lameter wrote:
> On Wed, 1 Aug 2018, Jeremy Linton wrote:
>
> > diff --git a/mm/slub.c b/mm/slub.c
> > index 51258eff4178..e03719bac1e2 100644
> > --- a/mm/slub.c
> > +++ b/mm/slub.c
> > @@ -2519,6 +2519,8 @@ static void *___slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
> >   		if (unlikely(!node_match(page, searchnode))) {
> >   			stat(s, ALLOC_NODE_MISMATCH);
> >   			deactivate_slab(s, page, c->freelist, c);
> > +			if (!node_online(searchnode))
> > +				node = NUMA_NO_NODE;
> >   			goto new_slab;
> >   		}
> >   	}
>
> Would it not be better to implement this check in the page allocator?
> There is also the issue of how to fallback to the nearest node.

Possibly? Falling back to the nearest node though, should be handled if 
memory-less nodes is enabled, which in the problematic case isn't.

> NUMA_NO_NODE should fallback to the current memory allocation policy but
> it seems by inserting it here you would end up just with the default node
> for the processor.

I picked this spot (compared to 2/2) because a number of paths are 
funneling through here, and in this case it shouldn't be a very hot path.