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

[Pingfan Liu <kernelfans@gmail.com>]

On Thu, Dec 6, 2018 at 8:11 PM Michal Hocko <mhocko@kernel.org> wrote:
>
> On Thu 06-12-18 18:44:03, Pingfan Liu wrote:
> > On Thu, Dec 6, 2018 at 6:03 PM Pingfan Liu <kernelfans@gmail.com> wrote:
> [...]
> > > Which commit is this patch applied on? I can not apply it on latest linux tree.
> > >
> > I applied it by manual, will see the test result. I think it should
> > work since you instance all the node.
> > But there are two things worth to consider:
> > -1st. why x86 do not bring up all nodes by default, apparently it will
> > be more simple by that way
>
> What do you mean? Why it didn't bring up before? Or do you see some

Yes, this is what I mean. But maybe the author does not consider about
the nr_cpus, otherwise, using:
+       for_each_node(node)
+               if (!node_online(node))
+                       init_memory_less_node(node);
in init_cpu_to_node() is more simple.

> nodes not being brought up after this patch?
>
> > -2nd. there are other archs, do they obey the rules?
>
> I am afraid that each arch does its own initialization.

Then it is arguable whether to fix this issue in memory core or let
each archs to fix this issue. I check the powerpc code, it should also
need a fix, it maybe the same in arm and mips ..
 BTW, your patch can not work for normal bootup, and the kernel hang
without any kernel message.
I think it is due to the bug in the patch:
                alloc_node_data(nid);
+               if (!end)
+                       init_memory_less_node(nid); //which calls
alloc_node_data(nid) also.
How about the following:
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);

        for_each_possible_cpu(cpu) {
-               int node = numa_cpu_node(cpu);
+               node = numa_cpu_node(cpu);

                if (node == NUMA_NO_NODE)
                        continue;

-               if (!node_online(node))
-                       init_memory_less_node(node);
-
                numa_set_node(cpu, node);
        }
 }

Although it works, I hesitate about the idea, due to the semantic of
online-node, does the online-node require either cpu or memory inside
the node to be online?
In a short word, the fix method should consider about the two factors:
semantic of online-node and the effect on all archs

Thanks,
Pingfan

> --
> Michal Hocko
> SUSE Labs