[PATCH RFC] mm: remove CONFIG_HAVE_MEMBLOCK_NODE_MAP (was: Re: [PATCH v3 0/5] mm: Enable CONFIG_NODES_SPAN_OTHER_NODES by default for NUMA)

[Baoquan He]

On 04/02/20 at 10:01am, Michal Hocko wrote:
> On Wed 01-04-20 10:51:55, Mike Rapoport wrote:
> > Hi,
> > 
> > On Wed, Apr 01, 2020 at 01:42:27PM +0800, Baoquan He wrote:
> [...]
> > > From above information, we can remove HAVE_MEMBLOCK_NODE_MAP, and
> > > replace it with CONFIG_NUMA. That sounds more sensible to store nid into
> > > memblock when NUMA support is enabled.
> >  
> > Replacing CONFIG_HAVE_MEMBLOCK_NODE_MAP with CONFIG_NUMA will work, but
> > this will not help cleaning up the whole node/zone initialization mess and
> > we'll be stuck with two implementations.
> 
> Yeah, this is far from optimal.
> 
> > The overhead of enabling HAVE_MEMBLOCK_NODE_MAP is only for init time as
> > most architectures will anyway discard the entire memblock, so having it in
> > a UMA arch won't be a problem. The only exception is arm that uses
> > memblock for pfn_valid(), here we may also think about a solution to
> > compensate the addition of nid to the memblock structures. 
> 
> Well, we can make memblock_region->nid defined only for CONFIG_NUMA.
> memblock_get_region_node would then unconditionally return 0 on UMA.
> Essentially the same way we do NUMA for other MM code. I only see few
> direct usage of region->nid.

Checked code again, seems HAVE_MEMBLOCK_NODE_MAP is selected directly in
all ARCHes which support it. Means HAVE_MEMBLOCK_NODE_MAP is enabled by
default on those ARCHes, and has no dependency on CONFIG_NUMA at all.
E.g on x86, it just calls free_area_init_nodes() in generic code path,
while free_area_init_nodes() is defined in CONFIG_HAVE_MEMBLOCK_NODE_MAP
ifdeffery scope. So I tend to agree with Mike to remove
HAVE_MEMBLOCK_NODE_MAP firstly on all ARCHes. We can check if it's worth
only defining memblock_region->nid for CONFIG_NUMA case after
HAVE_MEMBLOCK_NODE_MAP is removed.

config X86
        def_bool y
	...
	select HAVE_MEMBLOCK_NODE_MAP
	...