On Tue 07-12-21 12:08:57, David Hildenbrand wrote: > On 07.12.21 11:54, Michal Hocko wrote: > > Hi, > > I didn't have much time to dive into this deeper and I have hit some > > problems handling this in an arch specific code so I have tried to play > > with this instead: > > > > diff --git a/mm/page_alloc.c b/mm/page_alloc.c > > index c5952749ad40..4d71759d0d9b 100644 > > --- a/mm/page_alloc.c > > +++ b/mm/page_alloc.c > > @@ -8032,8 +8032,16 @@ void __init free_area_init(unsigned long *max_zone_pfn) > > /* Initialise every node */ > > mminit_verify_pageflags_layout(); > > setup_nr_node_ids(); > > - for_each_online_node(nid) { > > + for_each_node(nid) { > > pg_data_t *pgdat = NODE_DATA(nid); > > + > > + if (!node_online(nid)) { > > + pr_warn("Node %d uninitialized by the platform. Please report with memory map.\n"); > > + alloc_node_data(nid); > > That's x86 specific and not exposed to generic code -- at least in my > code base. I think we'd want an arch_alloc_nodedata() variant that > allocates via memblock -- and initializes all fields to 0. So > essentially a generic alloc_node_data(). you are right > > > + free_area_init_memoryless_node(nid); > > That's really just free_area_init_node() below, I do wonder what value > free_area_init_memoryless_node() has as of today. I am not sure there is any real value in having this special name for this but I have kept is sync with what x86 does currently. If we want to remove the wrapper then just do it everywhere. I can do that on top. > > + continue; > > + } > > + > > free_area_init_node(nid); > > > > /* Any memory on that node */ > > > > Could you give it a try? I do not have any machine which would exhibit > > the problem so I cannot really test this out. I hope build_zone_info > > will not choke on this. I assume the node distance table is > > uninitialized for these nodes and IIUC this should lead to an assumption > > that all other nodes are close. But who knows that can blow up there. > > > > Btw. does this make any sense at all to others? > > > > __build_all_zonelists() has to update the zonelists of all nodes I think. I am not sure what you mean. This should be achieved by this patch because the boot time build_all_zonelists will go over all online nodes (i.e. with pgdat). free_area_init happens before that. I am just worried that the arch specific node_distance() will generate a complete garbage or blow up for some reason. -- Michal Hocko SUSE Labs