From: Steven Whitehouse Date: Tue, 12 Dec 2006 09:13:02 +0000 Subject: [Cluster-devel] [RFC 3/3] GFS2 rename race - core changes In-Reply-To: <457E65F9.50609@redhat.com> References: <457E65F9.50609@redhat.com> Message-ID: <1165914782.3752.1038.camel@quoit.chygwyn.com> List-Id: To: cluster-devel.redhat.com MIME-Version: 1.0 Content-Type: text/plain; charset="us-ascii" Content-Transfer-Encoding: 7bit Hi, On Tue, 2006-12-12 at 03:19 -0500, Wendy Cheng wrote: > Upon failure of gfs2_unlink_ok(), we discard the old dentry, create a > new dentry, and then fill in the correct inode contents. This is to > avoid gfs_rename() returns ENOENT for destination file changes. > > I think that you might have to look at the set/reset of FS_RENAME_DOES_D_MOVE again. This flag is intended to be constant and set as part of the struct file_system_type. Also it needs to be set on both of the filesystem types since its required on the metafs as well. I think its a better plan for GFS2 to do the d_move in all cases to avoid the races which are otherwise introduced by trying to change the flag. Otherwise I think the patch looks good, Steve.