Re: ZONE_PADDING wastes 4 bytes of the new cacheline

[Andrea Arcangeli <andrea@novell.com>]

On Thu, Oct 21, 2004 at 10:49:36PM -0500, Jesse Barnes wrote:
> On Thursday, October 21, 2004 10:38 pm, Nick Piggin wrote:
> > That problem shouldn't exist any more, so your one zone per node (?)
> > NUMA systems, incremental min won't have any effect at all.
> 
> Well, it used to affect us, since as the allocator iterated over nodes, the 
> incremental min would increase, and so by the time we hit the 3rd or so node, 
> we were leaving quite a bit of memory unused.  I just don't want to return to 
> the bad old days.

yes, but all you care about is to turn off the incremental min, you
don't really care about lowmem_reserved, because you don't have floppies
that do ZONE_DMA allocations on ia64, x86-64 would oom-kill tasks just
because insmod floppy.o was running (and this is one of the showstopper
bugs in my queue, I had to disable forced the oom killer to workaroaund
it, apparently the vm can then later on free some page after many loops
despite without swap enabled).