Andrew Morton wrote:
> Matthew Dobson <colpatch@us.ibm.com> wrote:
> 
>>Okee dokee...  Here's the real core of the patch.
> 
> 
> Looks saneish to me.  I'd like to see thorough benchmark results when it is
> complete.  And it would be nice to make address_space.binding go away if
> !CONFIG_NUMA.
Yeah... I figured the one pointer wouldn't be a big deal, and it would 
save a few more ifdefs, but I guess consistency is better.  If I ifdef 
most of it, may as well ifdef it all!

> The explicit knowledge of ZONE_DMA/ZONE_NORMAL/ZONE_HIGHMEM in get_zonetype()
> should not be necessary - you don't want it to explode if ZONE_DMA32 is
> added.  It should be indexing into node_zonelists in some manner.
> 
> Will this code work if all memory is in ZONE_DMA, as some architectures do?

Well, what I do for now is use the gfp_mask in the address_space that is 
already there for the shm segments.  This means I have to look at that 
to decide if they want DMA, NORMAL, or HIGHMEM.  So it already would 
need to be changed if we added a new zone.  If the thing we're creating 
a binding for only can use DMA, for example, I need to make sure I use 
only the DMA zones.  I can't easily see a way to make sure I add only 
the appropriate zones to the zonelist, and also don't look at the 
gfp_flag to determine the zones they want.  I'll ponder it more deeply 
tonight, though.

New patch, sans binding pointer attatched.

Cheers!

-Matt