Re: [PATCH] fix spurious OOM kills

[Marcelo Tosatti <marcelo.tosatti@cyclades.com>]

On Wed, Nov 17, 2004 at 04:08:52AM -0200, Marcelo Tosatti wrote:
> On Wed, Nov 17, 2004 at 04:06:48AM -0200, Marcelo Tosatti wrote:
> > On Wed, Nov 17, 2004 at 01:23:46AM -0800, Andrew Morton wrote:
> > > Chris Ross <chris@tebibyte.org> wrote:
> > > >
> > > > As I suspected, like a recalcitrant teenager it was sneakily waiting 
> > > >  until everyone was out then it threw a wild party with several ooms and 
> > > >  an oops. See below...
> > > 
> > > That's not an oops - it's just a stack trace.
> > > 
> > > >  This, obviously is still without Kame's patch, just the same tree as 
> > > >  before with the one change you asked for.
> > > 
> > > Please ignore the previous patch and try the below.  It looks like Rik's
> > > analysis is correct: when the caller doesn't have the swap token it just
> > > cannot reclaim referenced pages and scans its way into an oom.  Defeating
> > > that logic when we've hit the highest scanning priority does seem to fix
> > > the problem and those nice qsbench numbers which the thrashing control gave
> > > us appear to be unaffected.
> > 
> > Oh, this fixes my testcase, and was the reason for the hog slow speed.
> > 
> > Excellent, wasted several days in vain. :(
> 
> Before the swap token patches went in you remember spurious OOM reports  
> or things were working fine then?

Just went on through the archives and indeed the spurious OOM kills started
happening when the swap token code was added to the tree.

Next time I should be looking into the easy stuff before trying miraculous 
solutions. :(