Re: [PATCH v2 1/1] mm/vmscan.c: Prevent allocating shrinker_info on offlined nodes

[Andrew Morton]

On Tue, 7 Dec 2021 19:25:25 -0500 Nico Pache <npache@xxxxxxxxxx> wrote:

> 
> 
> On 12/7/21 18:34, Matthew Wilcox wrote:
> > On Tue, Dec 07, 2021 at 05:40:13PM -0500, Nico Pache wrote:
> >> +++ b/mm/vmscan.c
> >> @@ -222,13 +222,16 @@ static int expand_one_shrinker_info(struct mem_cgroup *memcg,
> >>  	int size = map_size + defer_size;
> >>  
> >>  	for_each_node(nid) {
> >> +		int tmp = nid;
> >>  		pn = memcg->nodeinfo[nid];
> >>  		old = shrinker_info_protected(memcg, nid);
> >>  		/* Not yet online memcg */
> >>  		if (!old)
> >>  			return 0;
> >>  
> >> -		new = kvmalloc_node(sizeof(*new) + size, GFP_KERNEL, nid);
> >> +		if(!node_online(nid))
> >> +			tmp = numa_mem_id();
> >> +		new = kvmalloc_node(sizeof(*new) + size, GFP_KERNEL, tmp);
> >>  		if (!new)
> > 
> > Why should this be fixed here and not in, say, kvmalloc_node()?
> 
> according to Michal, the caller should be responsible for making sure it is
> allocating on a correct node. This avoids adding branches to hot-paths and
> wasting cycles. Im not opposed to moving it to kvmalloc_node, but it may result
> in masking other issues from other callers.
> > 

Yes, kvmalloc_node(nid) should allocate on `nid', or should fail.

A new kvmalloc_try_node() or whatever would express this idea.