Re: [PATCH] mm/alloc: fallback to first node if the wanted node offline

From: Pingfan Liu
Date: Tue Dec 04 2018 - 02:16:58 EST


On Tue, Dec 4, 2018 at 11:53 AM David Rientjes <rientjes@xxxxxxxxxx> wrote:
>
> On Tue, 4 Dec 2018, Pingfan Liu wrote:
>
> > diff --git a/include/linux/gfp.h b/include/linux/gfp.h
> > index 76f8db0..8324953 100644
> > --- a/include/linux/gfp.h
> > +++ b/include/linux/gfp.h
> > @@ -453,6 +453,8 @@ static inline int gfp_zonelist(gfp_t flags)
> > */
> > static inline struct zonelist *node_zonelist(int nid, gfp_t flags)
> > {
> > + if (unlikely(!node_online(nid)))
> > + nid = first_online_node;
> > return NODE_DATA(nid)->node_zonelists + gfp_zonelist(flags);
> > }
> >
>
> So we're passing the node id from dev_to_node() to kmalloc which
> interprets that as the preferred node and then does node_zonelist() to
> find the zonelist at allocation time.
>
> What happens if we fix this in alloc_dr()? Does anything else cause
> problems?
>
I think it is better to fix it mm, since it can protect any new
similar bug in future. While fixing in alloc_dr() just work at present

> And rather than using first_online_node, would next_online_node() work?
>
What is the gain? Is it for memory pressure on node0?

Thanks,
Pingfan

> I'm thinking about this:
>
> diff --git a/drivers/base/devres.c b/drivers/base/devres.c
> --- a/drivers/base/devres.c
> +++ b/drivers/base/devres.c
> @@ -100,6 +100,8 @@ static __always_inline struct devres * alloc_dr(dr_release_t release,
> &tot_size)))
> return NULL;
>
> + if (unlikely(!node_online(nid)))
> + nid = next_online_node(nid);
> dr = kmalloc_node_track_caller(tot_size, gfp, nid);
> if (unlikely(!dr))
> return NULL;