I've got a 2TB 8-node system (256GB per NUMA node) that's behaving a bit
strangely (OOMs with GB of free memory).

Its watermarks look wonky, with a min watermark of 0 pages for DMA and
only 11 pages for DMA32:

> Node 0 DMA    free:7428kB    min:0kB    low:0kB    high:0kB    ...
> Node 0 DMA32  free:1024084kB min:44kB   low:52kB   high:64kB   ... present:1941936kB   managed:1862456kB
> Node 0 Normal free:4808kB    min:6348kB low:7932kB high:9520kB ... present:266338304kB managed:262138972kB

This looks to be caused by us trying to evenly distribute the
min_free_kbytes value across the zones, but with such a huge size
imbalance (16MB zone vs 2TB system), 1/131072th of the default
min_free_kbytes ends up <1 page.

Should we be setting up some absolute floors on the watermarks, like the
attached patch?

BTW, it seems to be this code:

> static void __setup_per_zone_wmarks(void)
> {
>         unsigned long pages_min = min_free_kbytes >> (PAGE_SHIFT - 10);
...
>         for_each_zone(zone) {
>                 u64 tmp;
> 
>                 spin_lock_irqsave(&zone->lock, flags);
>                 tmp = (u64)pages_min * zone->managed_pages;
>                 do_div(tmp, lowmem_pages);