Re: Out of Memory: Killed process 16498 (java).

[Jens Axboe <axboe@suse.de>]

On Fri, Jan 20 2006, Andrew Morton wrote:
> Jens Axboe <axboe@suse.de> wrote:
> >
> > On Thu, Jan 19 2006, Andrew Morton wrote:
> > > Dave Jones <davej@redhat.com> wrote:
> > > >
> > > > On Thu, Jan 19, 2006 at 03:11:45PM -0000, Andy Chittenden wrote:
> > > >  > DMA free:20kB min:24kB low:28kB high:36kB active:0kB inactive:0kB
> > > >  > present:12740kB pages_scanned:4 all_unreclaimable? yes
> > > > 
> > > > Note we only scanned 4 pages before we gave up.
> > > > Larry Woodman came up with this patch below that clears all_unreclaimable
> > > > when in two places where we've made progress at freeing up some pages
> > > > which has helped oom situations for some of our users.
> > > 
> > > That won't help - there are exactly zero pages on ZONE_DMA's LRU.
> > > 
> > > The problem appears to be that all of the DMA zone has been gobbled up by
> > > the BIO layer.  It seems quite inappropriate that a modern 64-bit machine
> > > is allocating tons of disk I/O pages from the teeny ZONE_DMA.  I'm
> > > suspecting that someone has gone and set a queue's ->bounce_gfp to the wrong
> > > thing.
> > > 
> > > Jens, would you have time to investigate please?
> > 
> > Certainly, I'll get this tested and fixed this afternoon.
> 
> Wow ;)
>
> You may find it's an x86_64 glitch - setting max_[low_]pfn wrong down in
> the bowels of the arch mm init code, something like that.
> 
> I thought it might have been a regression which came in when we added
> ZONE_DMA32 but the RH reporter is based on 2.6.14-<redhat stuff>, and he
> didn't have ZONE_DMA32.

Sorry, spoke too soon, I thought this was the 'bio/scsi leaks' which
most likely is a scsi leak that also results in the bios not getting
freed.

This DMA32 zone shortage looks like a vm short coming, you're likely the
better candidate to fix that :-)

-- 
Jens Axboe