On Fri, 2004-11-19 at 06:09 -0200, Marcelo Tosatti wrote:

As Thomas Gleixner has investigated, the OOM killer selection is problematic.

When testing your ignore-page-referenced patch it first killed the memory hog
then shortly afterwards the shell I was running it on.

You've seen Thomas emails, he has nice description there.

I had another go on 2.6.10-rc2-mm2. 

The reentrancy blocking and the additional test of freepages in
out_of_memory() make all the ugly time and counter checks superfluid. 

I think they were neccecary to make the spurious kill triggering less
obvious. :)

Can somebody else check with his test cases, if the behaviour is
correct ?

tglx