[PATCH v2 0/5] mm: Enable CONFIG_NODES_SPAN_OTHER_NODES by default for NUMA

[Mike Rapoport]

On Wed, Jul 31, 2019 at 01:40:16PM +0200, Michal Hocko wrote:
> On Wed 31-07-19 14:14:22, Mike Rapoport wrote:
> > On Wed, Jul 31, 2019 at 10:03:09AM +0200, Michal Hocko wrote:
> > > On Wed 31-07-19 09:24:21, Mike Rapoport wrote:
> > > > [ sorry for a late reply too, somehow I missed this thread before ]
> > > > 
> > > > On Tue, Jul 30, 2019 at 10:14:15AM +0200, Michal Hocko wrote:
> > > > > [Sorry for a late reply]
> > > > > 
> > > > > On Mon 15-07-19 17:55:07, Hoan Tran OS wrote:
> > > > > > Hi,
> > > > > > 
> > > > > > On 7/12/19 10:00 PM, Michal Hocko wrote:
> > > > > [...]
> > > > > > > Hmm, I thought this was selectable. But I am obviously wrong here.
> > > > > > > Looking more closely, it seems that this is indeed only about
> > > > > > > __early_pfn_to_nid and as such not something that should add a config
> > > > > > > symbol. This should have been called out in the changelog though.
> > > > > > 
> > > > > > Yes, do you have any other comments about my patch?
> > > > > 
> > > > > Not really. Just make sure to explicitly state that
> > > > > CONFIG_NODES_SPAN_OTHER_NODES is only about __early_pfn_to_nid and that
> > > > > doesn't really deserve it's own config and can be pulled under NUMA.
> > > > > 
> > > > > > > Also while at it, does HAVE_MEMBLOCK_NODE_MAP fall into a similar
> > > > > > > bucket? Do we have any NUMA architecture that doesn't enable it?
> > > > > > > 
> > > > 
> > > > HAVE_MEMBLOCK_NODE_MAP makes huge difference in node/zone initialization
> > > > sequence so it's not only about a singe function.
> > > 
> > > The question is whether we want to have this a config option or enable
> > > it unconditionally for each NUMA system.
> > 
> > We can make it 'default NUMA', but we can't drop it completely because
> > microblaze uses sparse_memory_present_with_active_regions() which is
> > unavailable when HAVE_MEMBLOCK_NODE_MAP=n.
> 
> I suppose you mean that microblaze is using
> sparse_memory_present_with_active_regions even without CONFIG_NUMA,
> right?

Yes.

> I have to confess I do not understand that code. What is the deal
> with setting node id there?

The sparse_memory_present_with_active_regions() iterates over
memblock.memory regions and uses the node id of each region as the
parameter to memory_present(). The assumption here is that sometime before
each region was assigned a proper non-negative node id. 

microblaze uses device tree for memory enumeration and the current FDT code
does memblock_add() that implicitly sets nid in memblock.memory regions to -1.

So in order to have proper node id passed to memory_present() microblaze
has to call memblock_set_node() before it can use
sparse_memory_present_with_active_regions().

> -- 
> Michal Hocko
> SUSE Labs

-- 
Sincerely yours,
Mike.