On Thu, 2016-06-23 at 14:37 -0500, Reza Arbab wrote:

On Thu, Jun 23, 2016 at 10:47:20PM +0530, Aneesh Kumar K.V wrote:

quoted

Reza Arbab [off-list ref] writes:

quoted

These functions are making direct calls to the hash table APIs,
leading to a BUG() on systems using radix.

Switch them to the vmemmap_{create,remove}_mapping() wrappers, and
move to the __meminit section.

They are really not the same. They can possibly end up using different
base page size. Also vmemmap is available only with SPARSEMEM_VMEMMAP
enabled. Does hotplug depend on sparsemem vmemmap ?

I'm not sure. Maybe it's best if I back up a step and explain what lead 
me to this patch. During hotplug, you get

...
	arch_add_memory
		create_section_mapping
			htab_bolt_mapping
				BUG_ON(!ppc_md.hpte_insert);

So it seemed to me that I needed a radix equivalent of 
create_section_mapping().

After some digging, I found hash__vmemmap_create_mapping() and 
radix__vmemmap_create_mapping() did what I needed. I did not notice the 
#ifdef SPARSEMEM_VMEMMAP around them.

I think that's more by luck than design. The vmemmap routines use
mmu_vmemmap_psize which is probably but not definitely the same as
mmu_linear_psize.

Could it be that the functions just need to be renamed 
hash__create_mapping()/radix__create_mapping() and moved out of #ifdef 
SPARSEMEM_VMEMMAP?

No, you need to use mmu_linear_psize for the hotplug case.

But you can probably factor out a common routine that both cases use, and hide
the hash vs radix check in that.

And probably send me a patch to make MEMORY_HOTPLUG depend on !RADIX for v4.7?

cheers