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

[Jeremy Linton]

Hi,

On 08/02/2018 04:15 AM, Michal Hocko wrote:
> On Wed 01-08-18 15:04:17, Jeremy Linton wrote:
> [...]
> > @@ -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;
>
> This is inherently racy. Numa node can get offline at any point after
> you check it here. Making it race free would involve some sort of
> locking and I am not really convinced this is a good idea.

I spent some time looking/thinking about this, and i'm pretty sure its 
not creating any new problems. But OTOH, I think the node_online() check 
is probably a bit misleading as what we really want to assure is that 
node<MAX_NUMNODES and that there is going to be a valid entry in 
NODE_DATA() so we don't deference null.


> >   		}
> >   	}
> > --
> > 2.14.3