Hi Michal, What about dropping the change of the online definition of your patch, just do the following? diff --git a/arch/x86/mm/numa.c b/arch/x86/mm/numa.c index e6dad60..9c087c3 100644 --- a/arch/x86/mm/numa.c +++ b/arch/x86/mm/numa.c @@ -749,13 +749,12 @@ static void __init init_memory_less_node(int nid) */ void __init init_cpu_to_node(void) { - int cpu; + int cpu, node; u16 *cpu_to_apicid = early_per_cpu_ptr(x86_cpu_to_apicid); BUG_ON(cpu_to_apicid == NULL); - for_each_possible_cpu(cpu) { - int node = numa_cpu_node(cpu); + for_each_node_mask(node, numa_nodes_parsed) { if (node == NUMA_NO_NODE) continue; @@ -765,6 +764,10 @@ void __init init_cpu_to_node(void) numa_set_node(cpu, node); } + for_each_possible_cpu(cpu) { + int node = numa_cpu_node(cpu); + numa_set_node(cpu, node); + } } Thanks, Pingfan On Fri, Jun 21, 2019 at 9:55 PM Michal Hocko <mhocko@xxxxxxxxxx> wrote: > > On Fri 21-06-19 09:17:58, Qian Cai wrote: > > Sigh... > > > > I don't see any benefit to keep the broken commit, > > > > "x86, numa: always initialize all possible nodes" > > > > for so long in linux-next that just prevent x86 NUMA machines with any memory- > > less node from booting. > > > > Andrew, maybe it is time to drop this patch until Michal found some time to fix > > it properly. > > Yes, please drop the patch for now, Andrew. I thought I could get to > this but time is just scarce. > -- > Michal Hocko > SUSE Labs