The Ceph client is told by the server when it has the entire contents of a directory in cache, and is notified prior to any changes. However, the current VFS infrastructure does not allow the client to handle a lookup on a non-existent entry in a non-racy way. The first patch addes a new d_prune dentry_operation that is called before the VFS throws dentries out of cache (specifically, before the victim dentry is unhashed). The next two patches make the necessary changes in the Ceph fs code to safely clear a D_COMPLETE flag on the directory dentry when a child is pruned. The third patch specifically compensates for calls to dentry_unhash() in vfs_rmdir() and vfs_rename_dir(). The last patch adjusts the Ceph fs code to take advantage of the new flag. That change is pretty simple because most of the infrastructure is already in place (we were previously relying on d_release for notification of pruning in a racy way). Adding this interface would more or less codify the idea that the VFS shouldn't unhash random dentries without first calling d_prune. There are currently two places where the VFS currently unhashes: vfs_rmdir and vfs_rename_dir both call dentry_unhash(), which is there to make it easy for simple file systems to avoid races with directory removal and lookups. That could arguably be pushed down into those file systems, but it's a more delicate cleanup. Sage Weil (3): vfs: add d_prune dentry operation ceph: clear parent D_COMPLETE flag when on dentry prune ceph: switch I_COMPLETE to D_COMPLETE Documentation/filesystems/Locking | 1 + fs/ceph/caps.c | 8 +-- fs/ceph/dir.c | 87 +++++++++++++++++++++++++++++++++---- fs/ceph/inode.c | 9 ++-- fs/ceph/mds_client.c | 6 +- fs/ceph/super.h | 23 +++++++++- fs/dcache.c | 8 +++ include/linux/dcache.h | 3 + 8 files changed, 121 insertions(+), 24 deletions(-) -- To unsubscribe from this list: send the line "unsubscribe linux-fsdevel" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
