Re: [PATCH] oom killer (Core)

[Andrea Arcangeli <andrea@suse.de>]

On Fri, Dec 10, 2004 at 05:52:33PM +0100, Thomas Gleixner wrote:
> I'm wondering why he did not post the final version. Andrea ???

I already posted the final version since it had no bugs and I asked to
get it merged twice. The only bugs are obviously in the drivers (or the
callers) and they needs urgent fixing and additionally the
might_sleep_if must stop checking if the system is running so these bugs
can see the light of the day. Not being allowed to schedule in
alloc_pages with __GFP_WAIT set is a mistake.

Your patch was orthogonal to mine, so I didn't merge it. Go figure that
every time I post something it gets splitted into trivial pieces, so
it's a waste of time to try to merge any additional patch and post a
final one since it'll never be final anyway.

I am about to merge the things together for some other tree (not
mainline), that is a worthwhile effort but with the split behaviour of
mainline, for mainline it'd be a waste of time.

One last thing worth discussing on my side is if we should worry about
the tiny race between the watermark checks and the entering of the oom
killing. In theory we could wrap the thing around a semaphore and close
the race completely, though current code is simpler and as you find
it works fine in practice.