Kernel panic due to page migration accessing memory holes

[KAMEZAWA Hiroyuki]

On Thu, 18 Feb 2010 21:48:37 -0800
Michael Bohan <mbohan at codeaurora.org> wrote:

> On 2/18/2010 6:00 PM, KAMEZAWA Hiroyuki wrote:
> > memmap for memory holes should be marked as PG_reserved and never be freed
> > by free_bootmem(). Then, memmap for memory holes will not be in buddy allocator.
> >
> > Again, pfn_valid() just show "there is memmap", not for "there is a valid page"
> >    
> 
> ARM seems to have been freeing the memmap holes for a long time.
Ouch.

> I'm pretty sure there would be a lot of pushback if we tried to change 
> that.  For example, in my memory map running FLATMEM, I would be 
> consuming an extra ~7 MB of memory if these structures were not freed.
> 
> As a compromise, perhaps we could free everything except the first 
> 'pageblock_nr_pages' in a hole?  This would guarantee that 
> move_freepages() doesn't deference any memory that doesn't belong to the 
> memmap -- but still only waste a relatively small amount of memory.  For 
> a 4 MB page block, it should only consume an extra 32 KB per hole in the 
> memory map.
> 
No. You have to implement pfn_valid() to return correct value as
"pfn_valid() returnes true if there is memmap." even if you do that.
Otherwise, many things will go bad.

You have 2 or 3 ways.

1. re-implement pfn_valid() which returns correct value.
   maybe not difficult. but please take care of defining CONFIG_HOLES_IN_....
   etc.

2. use DISCONTIGMEM and treat each bank and NUMA node.
   There will be no waste for memmap. But other complication of CONFIG_NUMA.
   
3. use SPARSEMEM.
   You have even 2 choisce here. 
   a - Set your MAX_ORDER and SECTION_SIZE to be proper value.
   b - waste some amount of memory for memmap on the edge of section.
       (and don't free memmap for the edge.)
      
Thanks,
-Kame