Re: [PATCH] oom killer (Core)

[William Lee Irwin III <wli@holomorphy.com>]

On Fri, Dec 03, 2004 at 03:28:54AM +0100, Andrea Arcangeli wrote:
> Ok, I expect this patch to fix the problem completely. The biggest
> difference is that it doesn't affect the exit fast path and it doesn't need
> notification. It's not even slower than your patch since you were
> really polling too. This schedule properly, if you get any PREEMPT=n
> problem I'd really like to know. This is on top of my previous patch so
> it does the checks for the watermarks correctly (those are not obsoleted
> by whatever thing we do in out_of_memory()). This still make use of the
> PF_MEMDIE info but not in kernel/, only in mm/oom_kill.c where it born
> and dies there, so the race is somewhat hidden in there. As you said
> converting PF_MEMDIE to a set_tsk_thread_flag or some other non-racy
> thing is a due change but I'm not doing it in the below patch.

I see some real bugfixes in here. I'll start up a fresh thread with some
of those broken out and cc: the relevant people shortly. It's not really
any "new material", just reorganizing things into smaller patches, in
particular the ones I see needing merging ASAP that aren't controversial.


-- wli