From: "J. Bruce Fields" Subject: Re: [PATCH 0/3] NFSD EOS deferral Date: Fri, 17 Oct 2008 16:36:29 -0400 Message-ID: <20081017203629.GB14960@fieldses.org> References: <1224104426-12293-1-git-send-email-andros@netapp.com> <20081017174454.GB11884@fieldses.org> Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Cc: Marc Eshel , andros@netapp.com, linux-nfs@vger.kernel.org To: "Talpey, Thomas" Return-path: Received: from mail.fieldses.org ([66.93.2.214]:56376 "EHLO fieldses.org" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1751892AbYJQUgj (ORCPT ); Fri, 17 Oct 2008 16:36:39 -0400 In-Reply-To: Sender: linux-nfs-owner@vger.kernel.org List-ID: On Fri, Oct 17, 2008 at 04:26:18PM -0400, Talpey, Thomas wrote: > At 02:59 PM 10/17/2008, Marc Eshel wrote: > >linux-nfs-owner@vger.kernel.org wrote on 10/17/2008 10:44:54 AM: > > > >> "J. Bruce Fields" > >> Requests longer than a page are still not deferred, so large writes that > >> trigger upcalls still get an ERR_DELAY. OK, probably no big deal. > >> > >> I don't think we can apply this until we have some way to track the > >> number and size of deferred requests outstanding and fall back on > >> ERR_DELAY if it's too much. > > > >But I thought that the problem here is that the Linux NFS client doesn't > >handle this return code properly. > > Definitely this is an issue. Early clients do one of two things, they either > pass the error back to the application, or they enter a buzz loop resending > the operation with no delay. Later clients back off, but for a constant > five seconds. I haven't tested it, but from fs/nfs/nfs4proc.c:nfs4_delay() it appears to start at a tenth of a second and then do exponential backoff (up to 15 seconds). Looks to me like the code's been that way since at least 2.6.19. --b. > Either way, the server is generally better off gritting its > teeth and completing the operation. > > Blocking server threads is drastic, but in effect it will stall the client > queues and "push back". The issue on Linux is the small number of > nfsd contexts involved. It could lead to significant issues possibly > including DOS attack. Dropping connections (judiciously) could be > used instead of blocking the last few threads, though even that will > have consequences. > > The easy way to test all this is decorate /etc/exports with lots of > names, then break the nameservice and start sending requests from > many new clients. It's very hard to get it all right. > > Tom. >