Re: [RFC 0/2] Delay initializing of large sections of memory

[Mike Travis <travis@sgi.com>]

On 6/25/2013 11:44 AM, H. Peter Anvin wrote:
> On 06/25/2013 11:40 AM, Yinghai Lu wrote:
>> On Tue, Jun 25, 2013 at 11:17 AM, H. Peter Anvin <hpa@zytor.com> wrote:
>>> On 06/25/2013 10:35 AM, Mike Travis wrote:
>>
>>> However, please consider Ingo's counterproposal of doing this via the
>>> buddy allocator, i.e. hugepages being broken on demand.  That is a
>>> *very* powerful model, although would require more infrastructure.
>>
>> Can you or Ingo elaborate more about the buddy allocator proposal?
>>
> 
> Start by initializing 1G hyperpages only, but mark them so that the
> allocator knows that if it needs to break them apart it has to
> initialize the page structures for the 2M subpages.
> 
> Same thing with 2M -> 4K.
> 
> 	-hpa
> 
> 

It is worth experimenting with but the big question would be,
if it still avoids the very expensive "memmap_init_zone" and
it's sub-functions using huge expanses of memory.  I'll do some
experimenting as soon as I can.  Our 32TB system is being
brought back to 16TB (we found a number of problems as we
get closer and closer to the 64TB limit), but that's still
a significant size.