Return-Path: Received: from mail-pd0-f181.google.com ([209.85.192.181]:34921 "EHLO mail-pd0-f181.google.com" rhost-flags-OK-OK-OK-OK) by vger.kernel.org with ESMTP id S1756782AbbEEIdg (ORCPT ); Tue, 5 May 2015 04:33:36 -0400 Received: by pdbqd1 with SMTP id qd1so188674077pdb.2 for ; Tue, 05 May 2015 01:33:35 -0700 (PDT) Message-ID: <55488006.6050502@gmail.com> Date: Tue, 05 May 2015 16:32:06 +0800 From: Kinglong Mee MIME-Version: 1.0 To: NeilBrown CC: "J. Bruce Fields" , "linux-nfs@vger.kernel.org" , Al Viro , Steve Dickson , kinglongmee@gmail.com Subject: Re: [PATCH RFC] NFSD: fix cannot umounting mount points under pseudo root References: <20150421215417.GE13782@fieldses.org> <553781E2.1000900@gmail.com> <20150422150703.GA1247@fieldses.org> <20150423094431.1a8aa68b@notabene.brown> <5538EB18.7080802@gmail.com> <20150424130045.6bbdb2f9@notabene.brown> <553E2784.6020906@gmail.com> <20150429125728.69ddfc6c@notabene.brown> <20150429191934.GA23980@fieldses.org> <20150430075225.21a71056@notabene.brown> <20150430213602.GB9509@fieldses.org> <55483EB7.5060104@gmail.com> <20150505141957.2aef920e@notabene.brown> In-Reply-To: <20150505141957.2aef920e@notabene.brown> Content-Type: text/plain; charset=windows-1252 Sender: linux-nfs-owner@vger.kernel.org List-ID: On 5/5/2015 12:19 PM, NeilBrown wrote: > On Tue, 05 May 2015 11:53:27 +0800 Kinglong Mee wrote: > >> Cc Steve, Viro, >> >> On 5/1/2015 5:36 AM, J. Bruce Fields wrote: >>> On Thu, Apr 30, 2015 at 07:52:25AM +1000, NeilBrown wrote: >>>> On Wed, 29 Apr 2015 15:19:34 -0400 "J. Bruce Fields" >>>> wrote: >>>>> Maybe drop the locking from nfsd_buffered_readdir and *just* take the >>>>> i_mutex around lookup_one_len(), if that's the only place we need it? >> >> As description in other thread, before the upcall to rpc.mountd, >> nfsd have call lookup_one_len() for the file, but why rpc.mountd >> also blocked in lookup ? >> >> There is a bug in rpc.mountd when checking sub-directory, >> it sets bad patch length for child. >> >> If parent if "/nfs/xfs" and child is "/nfs/test", the child name >> will be truncated to "/nfs/tes" for strlen(parent), "/nfs/test" >> have exist in kernel's cache for the lookup_one_len(), but >> "/nfs/tes" is a bad path, which needs lookup_slow(), so blocked. > > Testing for "/nfs/tes" certain seems like a wrong thing to do. > >> >> static int is_subdirectory(char *child, char *parent) >> { >> /* Check is child is strictly a subdirectory of >> * parent or a more distant descendant. >> */ >> size_t l = strlen(parent); >> >> if (strcmp(parent, "/") == 0 && child[1] != 0) >> return 1; >> >> return (same_path(child, parent, l) && child[l] == '/'); > > I guess this should be: > > child[l] == '/' && same_path(child, parent, l) > > That way there would be no risk of truncating anything. > > Can you please test if that one-line change removes the problem? Yes, that's OK. I have think about the "/nfs/nfs" and "/nfs/123" of the first patch, yes, that's redundant. > >> } >> >> The following path makes a correct path, not a truncated path. >> Have be tested, everything is OK. >> >> thanks, >> Kinglong Mee >> >> ----------------------------------------------------------------------------------- >> >From 70b9d1d93a24db8a7837998cb7eb0ff4e98480a6 Mon Sep 17 00:00:00 2001 >> From: Kinglong Mee >> Date: Tue, 5 May 2015 11:47:20 +0800 >> Subject: [PATCH] mountd: Case-insensitive path length must equal to parent >> >> Commit 6091c0a4c4 (mountd: add support for case-insensitive file names) >> introdues a bug cause mutex race when looking bad path. > > I think we should be clear that the mutex race is already present. > I think you are right that there is a bug here which is making it easy to > trigger, but it isn't exactly "causing" the bug. Thanks for the comments. thanks, Kinglong Mee -----------------------------------------------------------------------------------