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

[Michal Hocko]

On Tue 11-12-18 16:05:58, Pingfan Liu wrote:
> On Mon, Dec 10, 2018 at 8:37 PM Michal Hocko <mhocko@xxxxxxxxxx> wrote:
> >
> > On Fri 07-12-18 16:56:27, Michal Hocko wrote:
> > > On Fri 07-12-18 22:27:13, Pingfan Liu wrote:
> > > [...]
> > > > diff --git a/arch/x86/mm/numa.c b/arch/x86/mm/numa.c
> > > > index 1308f54..4dc497d 100644
> > > > --- a/arch/x86/mm/numa.c
> > > > +++ b/arch/x86/mm/numa.c
> > > > @@ -754,18 +754,23 @@ void __init init_cpu_to_node(void)
> > > >  {
> > > >         int cpu;
> > > >         u16 *cpu_to_apicid = early_per_cpu_ptr(x86_cpu_to_apicid);
> > > > +       int node, nr;
> > > >
> > > >         BUG_ON(cpu_to_apicid == NULL);
> > > > +       nr = cpumask_weight(cpu_possible_mask);
> > > > +
> > > > +       /* bring up all possible node, since dev->numa_node */
> > > > +       //should check acpi works for node possible,
> > > > +       for_each_node(node)
> > > > +               if (!node_online(node))
> > > > +                       init_memory_less_node(node);
> > >
> > > I suspect there is no change if you replace for_each_node by
> > >       for_each_node_mask(nid, node_possible_map)
> > >
> > > here. If that is the case then we are probably calling
> > > free_area_init_node too early. I do not see it yet though.
> >
> > OK, so it is not about calling it late or soon. It is just that
> > node_possible_map is a misnomer and it has a different semantic than
> > I've expected. numa_nodemask_from_meminfo simply considers only nodes
> > with some memory. So my patch didn't really make any difference and the
> > node stayed uninialized.
> >
> > In other words. Does the following work? I am sorry to wildguess this
> > way but I am not able to recreate your setups to play with this myself.
> >
> No problem. Yeah, in order to debug the patch, you need a numa machine
> with a memory-less node. And unlucky, the patch can not work either by
> grub bootup or kexec -l boot. There is nothing, just silent.  I will
> dig into numa_register_memblks() to figure out the problem.

I do not have such a machine handy. Anyway, can you post the full serial
console log. Maybe I can infer something. It is quite weird that this
patch would make an existing situation any worse.
-- 
Michal Hocko
SUSE Labs