From: Neil Brown Subject: Re: Invisible files and input/output errors Date: Thu, 6 Dec 2007 10:31:17 +1100 Message-ID: <18263.13509.397157.489566@notabene.brown> References: <200712040037.36598.bartoschek@or.uni-bonn.de> <18260.40472.276614.78174@notabene.brown> <200712041106.56189.bartoschek@or.uni-bonn.de> Mime-Version: 1.0 Content-Type: text/plain; charset=us-ascii Cc: linux-nfs@vger.kernel.org To: Christoph Bartoschek Return-path: Received: from cantor2.suse.de ([195.135.220.15]:51645 "EHLO mx2.suse.de" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1751216AbXLEXbX (ORCPT ); Wed, 5 Dec 2007 18:31:23 -0500 In-Reply-To: message from Christoph Bartoschek on Tuesday December 4 Sender: linux-nfs-owner@vger.kernel.org List-ID: On Tuesday December 4, bartoschek-eefaeeawk+jFLWzOrOlZPg@public.gmane.org wrote: > > > Sounds a lot like directory cookie problems. > > > > Can you get a network trace of the NFS packets. > > Something like > > > > tcpdump -s 0 -w /tmp/trace host NAME-OF-SERVER > > > > then cause the failure while that is running. > > Either compress and attach the result to an Email, or if it is too > > big, stick it on a website somewhere. > > I have attached a trace of a "ls" that does not show some directories. There is one interesting feature of this trace. The 'cookie' assigned to each directory entry normally is 32 more than the previous entry, though in some places the gap is 64 or another large multiple of 32. The whole directory is fetched in three requests - each request sending the cookie from the last name in the previous reply. For the last request, the cookie for the first entry returned is 992 more than that last cookie in the previous reply. This is by far that largest gap, and that fact that it aligns with a break between two requests seems a little suspicious. It might be interesting to remove some of the files that appear early in the list, e.g. COUNT8SF manga.tcl test-run tanga.tcl flex sopec mopec thumper and see how that affects the returned list of files. In any case, it looks from the trace that the Linux client so doing the right thing based on the information returned, so I would suggested reporting this to whoever provided your NFS-DFS gateway. NeilBrown