From: Olivier Galibert Subject: Race condition with 4.1.4 and a symlink Date: Wed, 17 Jan 2007 17:16:57 +0100 Message-ID: <20070117161657.GA66984@dspnet.fr.eu.org> Mime-Version: 1.0 Content-Type: text/plain; charset="us-ascii" Content-Transfer-Encoding: 7bit Return-path: Content-Disposition: inline List-Id: List-Unsubscribe: , List-Archive: List-Post: List-Help: List-Subscribe: , Sender: autofs-bounces@linux.kernel.org Errors-To: autofs-bounces@linux.kernel.org To: autofs@linux.kernel.org Fedora Core 5 on x86_64, autofs-4.1.4-29, kernel-2.6.18-1.2200.fc5 The situation: - /people is on autofs, /people/gadda points to m179:/disk05/disk11/gadda - /hosts in on autofs so that /hosts/host/disk mounts host:/disk - m179:/disk05/disk11/gadda/normalisation (aka /people/gadda/normalisation) is a symlink to /hosts/m179/disk03/gadda/lemonde A script in /people/gadda/normalisation that opens and closes files repeatedly in normalisation has random ENOENT failures. Crossing a strace and a tcpdump gave me this trace: 01:49:17.429638 open("/people/gadda/normalisation/bugs2.norm") 01:49:17.429703 getattr(bugs2.norm) 01:49:17.430163 getattr ok 01:49:17.432830 open ok 01:49:17.443308 open("/people/gadda/normalisation/bugs2.norm") 01:49:17.443360 getattr(bugs2.norm) 01:49:17.443798 getattr ok 01:49:17.446893 open ok 01:49:17.451674 open("/people/gadda/normalisation/bugs2") 01:49:17.451752 getattr(bugs2) 01:49:17.452232 getattr ok 01:49:17.452268 open ok 01:49:17.457657 open("/people/gadda/normalisation/bugs2.norm") 01:49:17.457706 getattr(bugs2.norm) 01:49:17.458152 getattr ok 01:49:17.465475 open ok 01:49:17.473659 open("/people/gadda/normalisation/bugs2") 01:49:17.473707 getattr(bugs2) 01:49:17.474156 getattr ok 01:49:17.474189 open ok 01:49:17.478682 open("/people/gadda/normalisation/bugs2.norm") 01:49:17.478728 getattr(bugs2.norm) 01:49:17.479180 getattr ok 01:49:17.479407 open ok 01:49:17.480524 open("/people/gadda/normalisation/bugs2") 01:49:17.480569 getattr(bugs2) 01:49:17.481018 getattr ok 01:49:17.481513 open ok 01:49:17.486468 start TCP to sunrpc to do the getport 01:49:17.487203 open("/people/gadda/normalisation/bugs2") 01:49:17.488679 getport(MOUNT) 01:49:17.491467 getport reply 01:49:17.492038 mount_null 01:49:17.492822 mount_null reply 01:49:17.496420 umount("/disk05/disk11/gadda") (mounted on /people/gadda) 01:49:17.499833 umount ok 01:49:17.525547 getattr(general.log) 01:49:17.525596 open fail ENOENT <---- script failure here 01:49:17.526026 getattr ok 01:49:17.526201 write(general.log) 01:49:17.526421 getattr(general.log) 01:49:17.527115 getattr ok 01:49:17.529122 write ok 01:49:17.529170 write(general.log) 01:49:17.531651 write ok 01:49:17.575045 getattr(general3.cmd) 01:49:17.575515 getattr ok 01:49:17.620203 getport(NFS) 01:49:17.620728 getport reply 01:49:17.620747 nfs_null 01:49:17.621167 nfs_null ok 01:49:17.621184 getport(MOUNT) 01:49:17.621682 getport reply 01:49:17.621699 mount_null 01:49:17.622182 mount_null ok 01:49:17.622209 mount("/disk05/disk11/gadda") 01:49:17.626987 mount ok [...] 01:49:17.685213 readlink("normalisation") 01:49:17.685678 readlink reply "/hosts/m179/disk03/gadda/lemonde" Looks like having the link looked up all the time does not count as being busy for a mountpoint (not that surprising), and for that reason autofs unmounts on timeout. It asks for the remount fast but there is a window between unmounting and remounting where opens through the link fail. Is there something I can do? OG.