Re: [Lhms-devel] Re: [RFC] buddy allocator without bitmap [3/4]

[Hiroyuki KAMEZAWA <kamezawa.hiroyu@jp.fujitsu.com>]

Dave Hansen wrote:

>>+                       if (zone->nr_mem_map > 1) {
>>+                               /*
>>+                                * there may be hole in zone's memmap &&
>>+                                * hole is not aligned in this order.
>>+                                * currently, I think CONFIG_VIRTUAL_MEM_MAP
>>+                                * case is only case to reach here.
>>+                                * Is there any other case ?
>>+                                */
>>+                               /*
>>+                                * Is there better call than pfn_valid ?
>>+                                */
>>+                               if (!pfn_valid(zone->zone_start_pfn
>>+                                              + (page_idx ^ (1 << order))))
>>+                                       break;
>>+                       }
> 
> 
> Nice try.  How about putting the ia64 code in a macro or header function
> that you can #ifdef out on all the other architectures?  We used to be
> able to see that entire while loop on one screen.  That's a bit harder
> now.  
> 
Currently, I think zone->nr_mem_map itself is very vague.
I'm now looking for another way to remove this part entirely.

I think mem_section approarch may be helpful to remove this part,
but to implement full feature of CONFIG_NONLINEAR,
I'll need lots of different kind of patches.
(If mem_map is guaranteed to be contiguous in one mem_section)

1. Now, I think some small parts, some essence of mem_section which
   makes pfn_valid() faster may be good.

And another way,

2. A method which enables page -> page's max_order calculation
   may be good and consistent way in this no-bitmap approach.

But this problem would be my week-end homework :).

--Kame

-- 
--the clue is these footmarks leading to the door.--
KAMEZAWA Hiroyuki <kamezawa.hiroyu@jp.fujitsu.com>