On Tue, 22 Jan 2008, Mel Gorman wrote:

quoted

quoted

Whatever this was a problem fixed in the past or not, it's broken again now
:( . It's possible that there is a __GFP_THISNODE that can be dropped early
at boot-time that would also fix this problem in a way that doesn't
affect runtime (like altering cache_grow in my patch does).

The dropping of GFP_THISNODE has the same effect as your patch. 

The dropping of it totally? If so, this patch might fix a boot but it'll
potentially be a performance regression on NUMA machines that only have
nodes with memory, right?

No the dropping during early allocations.,

o 0
o 2
Nodes with regular memory
o 2
Current running CPU 0 is associated with node 0
Current node is 0

So node 2 has regular memory but it's trying to use node 0 at a glance.
I've attached the patch I used against 2.6.24-rc8. It includes the revert.

We need the current processor to be attached to a node that has 
memory. We cannot fall back that early because the structures for the 
other nodes do not exist yet.

Online nodes
o 0
o 2
Nodes with regular memory
o 2
Current running CPU 0 is associated with node 0
Current node is 0
 o kmem_list3_init

This needs to be node 2.

[c0000000005c3b40] c0000000000dadec .cache_grow+0x7c/0x338
[c0000000005c3c00] c0000000000db54c .fallback_alloc+0x1c0/0x224

Fallback during bootstrap.