Re: Terminate process that fails on a constrained allocation

[Paul Jackson <pj@sgi.com>]

If the oom killer is of any use on a numa system using set_mempolicy,
mbind and/or cpusets, then it would be of use in the bootcpuset -- that
cpuset which is often setup to hold the classic Unix daemons.  I'm not
sure we want to do that.

If we changed the mm/oom_kill.c select_bad_process() constraint from

                /* If p's nodes don't overlap ours, it won't help to kill p. */
                if (!cpuset_excl_nodes_overlap(p))
                        continue;

which kills any task that has any overlap with nodes with the current task:

 1) to the much tighter limit of only killing tasks that are on the same or
    a subset of the same nodes, and

 2) changed it to look at -both- the cpuset and MPOL_BIND constraints,

then oom killer would work in a bootcpuset rather as it does now on
simple UMA systems.

The new logic would be -- only kill tasks that are constrained to
the same or a subset of the same nodes as the current task.

With this, the problem you describe with the oom killer and a task that
has used mbind or cpuset to just allow a single node would be fixed --
only tasks constrained to just that node, no more, would be considered
for killing.

At the same time, a typical bootcpuset would have oom killer behaviour
resembling what people have become accustomed to.

If we are going to neuter or eliminate the oom killer, it seems like
we should do it across the board, not just on numa boxes using
some form of memory constraint (mempolicy or cpuset).

-- 
                  I won't rest till it's the best ...
                  Programmer, Linux Scalability
                  Paul Jackson <pj@sgi.com> 1.925.600.0401
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html
Please read the FAQ at  http://www.tux.org/lkml/