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

[Michal Hocko]

On Tue 04-12-18 16:20:32, Pingfan Liu wrote:
> On Tue, Dec 4, 2018 at 3:22 PM Michal Hocko <mhocko@xxxxxxxxxx> wrote:
> >
> > On Tue 04-12-18 11:05:57, Pingfan Liu wrote:
> > > During my test on some AMD machine, with kexec -l nr_cpus=x option, the
> > > kernel failed to bootup, because some node's data struct can not be allocated,
> > > e.g, on x86, initialized by init_cpu_to_node()->init_memory_less_node(). But
> > > device->numa_node info is used as preferred_nid param for
> > > __alloc_pages_nodemask(), which causes NULL reference
> > >   ac->zonelist = node_zonelist(preferred_nid, gfp_mask);
> > > This patch tries to fix the issue by falling back to the first online node,
> > > when encountering such corner case.
> >
> > We have seen similar issues already and the bug was usually that the
> > zonelists were not initialized yet or the node is completely bogus.
> > Zonelists should be initialized by build_all_zonelists quite early so I
> > am wondering whether the later is the case. What is the actual node
> > number the device is associated with?
> >
> The device's node num is 2. And in my case, I used nr_cpus param. Due
> to init_cpu_to_node() initialize all the possible node.  It is hard
> for me to figure out without this param, how zonelists is accessed
> before page allocator works.

I believe we should focus on this. Why does the node have no zonelist
even though all zonelists should be initialized already? Maybe this is
nr_cpus pecularity and we do not initialize all the existing numa nodes.
Or maybe the device is associated to a non-existing node with that
setup. A full dmesg might help us here.

> > Your patch is not correct btw, because we want to fallback into the node in
> > the distance order rather into the first online node.
> > --
> What about this:
> +extern int find_next_best_node(int node, nodemask_t *used_node_mask);
> +
>  /*
>   * We get the zone list from the current node and the gfp_mask.
>   * This zone list contains a maximum of MAXNODES*MAX_NR_ZONES zones.
> @@ -453,6 +455,11 @@ static inline int gfp_zonelist(gfp_t flags)
>   */
>  static inline struct zonelist *node_zonelist(int nid, gfp_t flags)
>  {
> +       if (unlikely(!node_online(nid))) {
> +               nodemask_t used_mask;
> +               nodes_complement(used_mask, node_online_map);
> +               nid = find_next_best_node(nid, &used_mask);
> +       }
>         return NODE_DATA(nid)->node_zonelists + gfp_zonelist(flags);
>  }
> 
> I just finished the compiling, not test it yet, since the machine is
> not on hand yet. It needs some time to get it again.

This is clearly a no-go. nodemask_t can be giant and you cannot have it
on the stack for allocation paths which might be called from a deep
stack already. Also this is called from the allocator hot paths and each
branch counts.

-- 
Michal Hocko
SUSE Labs