On Fri 12-02-21 11:16:28, David Hildenbrand wrote:
> On 12.02.21 11:11, Michal Hocko wrote:
> > On Fri 12-02-21 10:56:19, David Hildenbrand wrote:
> > > On 12.02.21 10:55, David Hildenbrand wrote:
> > > > On 08.02.21 12:08, Mike Rapoport wrote:
> > [...]
> > > > > @@ -6519,8 +6581,19 @@ void __init get_pfn_range_for_nid(unsigned int nid,
> > > > > *end_pfn = max(*end_pfn, this_end_pfn);
> > > > > }
> > > > > - if (*start_pfn == -1UL)
> > > > > + if (*start_pfn == -1UL) {
> > > > > *start_pfn = 0;
> > > > > + return;
> > > > > + }
> > > > > +
> > > > > +#ifdef CONFIG_SPARSEMEM
> > > > > + /*
> > > > > + * Sections in the memory map may not match actual populated
> > > > > + * memory, extend the node span to cover the entire section.
> > > > > + */
> > > > > + *start_pfn = round_down(*start_pfn, PAGES_PER_SECTION);
> > > > > + *end_pfn = round_up(*end_pfn, PAGES_PER_SECTION);
> > > >
> > > > Does that mean that we might create overlapping zones when one node
> > >
> > > s/overlapping zones/overlapping nodes/
> >
> > I didn't get to review the patch yet. Just wanted to note that we can
> > interleave nodes/zone. Or what kind of concern do you have in mind?
>
> I know that we can have it after boot, when hotplugging memory. How about
> during boot?
I have seen systems where NUMA nodes are interleaved.
> For example, which node will a PFN then actually be assigned to?
I would have to double check all the details but I do remember there was
a non trivial work to handle those systems.
--
Michal Hocko
SUSE Labs
