On 22/06/11 10:44, Mel Gorman wrote:

I haven't started looking at this properly yet (stuck with other
bugs unfortunately) but I glanced at the sysrq message and on a 2G
64-bit machine, you have a tiny Normal zone! This is very unexpected.
Can you boot with mminit_loglevel=4 loglevel=9 and post your full
dmesg please? I want to see what the memory layout of this thing
looks like to see in the future if there is a correlation between
this type of bug and a tiny highest zone.

Note this machine has 3G RAM
dmesg attached

Broadly speaking though from seeing that, it reminds me of a
similar bug where small zones could keep kswapd alive for high-order
allocations reclaiming slab constantly. I suspect on your machine
that the Normal zone cannot be balanced for order-0 allocations and
is keeping kswapd awake.

Can you try booting with mem=1792M and if the Normal zone disappears,
try reproducing the bug?

I tried mem=1792M but grub gave an ENOSPC error
Maybe I need to supply a memmap= too?

cheers,
Padraig.