ceph: clear parent D_COMPLETE flag when on dentry prune

When the VFS prunes a dentry from the cache, clear the D_COMPLETE flag
on the parent dentry.  Do this for the live and snapshotted namespaces. Do
not bother for the .snap dir contents, since we do not cache that.

Signed-off-by: Sage Weil <sage@newdream.net>

fs/ceph/dir.c

@@ -1092,7 +1092,33 @@ static int ceph_snapdir_d_revalidate(struct dentry *dentry,
 	return 1;
 }
 
+/*
+ * When the VFS prunes a dentry from the cache, we need to clear the
+ * complete flag on the parent directory.
+ *
+ * Called under dentry->d_lock.
+ */
+static void ceph_d_prune(struct dentry *dentry)
+{
+	struct ceph_dentry_info *di;
+
+	dout("d_release %p\n", dentry);
+
+	/* do we have a valid parent? */
+	if (!dentry->d_parent || IS_ROOT(dentry))
+		return;
+
+	/* if we are not hashed, we don't affect D_COMPLETE */
+	if (d_unhashed(dentry))
+		return;
 
+	/*
+	 * we hold d_lock, so d_parent is stable, and d_fsdata is never
+	 * cleared until d_release
+	 */
+	di = ceph_dentry(dentry->d_parent);
+	clear_bit(CEPH_D_COMPLETE, &di->flags);
+}
 
 /*
  * read() on a dir.  This weird interface hack only works if mounted
@@ -1306,6 +1332,7 @@ const struct inode_operations ceph_dir_iops = {
 const struct dentry_operations ceph_dentry_ops = {
 	.d_revalidate = ceph_d_revalidate,
 	.d_release = ceph_d_release,
+	.d_prune = ceph_d_prune,
 };
 
 const struct dentry_operations ceph_snapdir_dentry_ops = {
@@ -1315,4 +1342,5 @@ const struct dentry_operations ceph_snapdir_dentry_ops = {
 
 const struct dentry_operations ceph_snap_dentry_ops = {
 	.d_release = ceph_d_release,
+	.d_prune = ceph_d_prune,
 };

fs/ceph/super.h

@@ -203,6 +203,7 @@ struct ceph_inode_xattr {
  * Ceph dentry state
  */
 struct ceph_dentry_info {
+	unsigned long flags;
 	struct ceph_mds_session *lease_session;
 	u32 lease_gen, lease_shared_gen;
 	u32 lease_seq;
@@ -213,6 +214,18 @@ struct ceph_dentry_info {
 	u64 offset;
 };
 
+/*
+ * dentry flags
+ *
+ * The locking for D_COMPLETE is a bit odd:
+ *  - we can clear it at almost any time (see ceph_d_prune)
+ *  - it is only meaningful if:
+ *    - we hold dir inode i_lock
+ *    - we hold dir FILE_SHARED caps
+ *    - the dentry D_COMPLETE is set
+ */
+#define CEPH_D_COMPLETE 1  /* if set, d_u.d_subdirs is complete directory */
+
 struct ceph_inode_xattrs_info {
 	/*
 	 * (still encoded) xattr blob. we avoid the overhead of parsing