From mboxrd@z Thu Jan 1 00:00:00 1970 From: Stephan von Krawczynski Subject: Re: [Bug #13648] nfsd: page allocation failure Date: Fri, 31 Jul 2009 13:48:43 +0200 Message-ID: <20090731134843.6ea9d3dd.skraw@ithnet.com> References: <20090727130420.1514cf5c.skraw@ithnet.com> Mime-Version: 1.0 Content-Transfer-Encoding: 7bit Return-path: In-Reply-To: Sender: kernel-testers-owner-u79uwXL29TY76Z2rM5mHXA@public.gmane.org List-ID: Content-Type: text/plain; charset="us-ascii" To: David Rientjes Cc: "Rafael J. Wysocki" , Linux Kernel Mailing List , Kernel Testers List , Justin Piszcz , Jeff Kirsher , Jesse Brandeburg , "David S. Miller" On Thu, 30 Jul 2009 14:30:42 -0700 (PDT) David Rientjes wrote: > On Mon, 27 Jul 2009, Stephan von Krawczynski wrote: > > > This is no regression between 2.6.29 and 2.6.30. > > In fact we could reproduce the problem with kernel versions: > > > > 2.6.27.26 < X <= 2.6.30.3 > > > > (Meaning 2.6.27.26 is the last one _not_ showing the problem). > > > > And 2.6.28.10 is showing the exact same problem as initially reported, > right? Yes, that is correct. > I noticed your /var/log/messages is showing you're using slub as opposed > to slab (which Justin was using, and causing order-0 allocations errors). > SLUB uses order-1 allocations for this cache growth and it's failing > because of memory fragmentation, not because you're truly oom. Originally I used slab, and as someone wanted me to test slub I tried. The results looked pretty much the same to me. > The only thing that is immediately apparent that changed in this path over > these kernel versions (there were significant changes to e1000e) is the > CRC stripping. If it's loaded as a module, perhaps you could try > > modprobe e1000e CrcStripping=0,0 > > (assuming you have two adapters). I will try that. > I've cc'd some relevant e1000e driver people in the hopes they'll be able > to diagnose this problem. Memory fragmentation as the result of page > group changes wouldn't affect order-0 allocations such as this on slab, so > it's doubtful the VM regressed if you can reproduce the problem with > CONFIG_SLAB. I can, as said before, the problem first showed up with slab. -- Regards, Stephan