* If there is an error, the caller will reset the flags via
  * configfs_detach_rollback().
  */
-static int configfs_detach_prep(struct dentry *dentry)
+static int configfs_detach_prep(struct dentry *dentry, struct mutex **wait_mutex)
 {
        struct configfs_dirent *parent_sd = dentry->d_fsdata;
        struct configfs_dirent *sd;
                if (sd->s_type & CONFIGFS_NOT_PINNED)
                        continue;
                if (sd->s_type & CONFIGFS_USET_DEFAULT) {
+                       /* Abort if racing with mkdir() */
+                       if (sd->s_type & CONFIGFS_USET_IN_MKDIR) {
+                               if (wait_mutex)
+                                       *wait_mutex = &sd->s_dentry->d_inode->i_mutex;
+                               return -EAGAIN;
+                       }
                        /* Mark that we're trying to drop the group */
                        sd->s_type |= CONFIGFS_USET_DROPPING;
 
                         * Yup, recursive.  If there's a problem, blame
                         * deep nesting of default_groups
                         */
-                       ret = configfs_detach_prep(sd->s_dentry);
+                       ret = configfs_detach_prep(sd->s_dentry, wait_mutex);
                        if (!ret)
                                continue;
                } else
         */
        module_got = 1;
 
+       /*
+        * Make racing rmdir() fail if it did not tag parent with
+        * CONFIGFS_USET_DROPPING
+        * Note: if CONFIGFS_USET_DROPPING is already set, attach_group() will
+        * fail and let rmdir() terminate correctly
+        */
+       spin_lock(&configfs_dirent_lock);
+       /* This will make configfs_detach_prep() fail */
+       sd->s_type |= CONFIGFS_USET_IN_MKDIR;
+       spin_unlock(&configfs_dirent_lock);
+
        if (group)
                ret = configfs_attach_group(parent_item, item, dentry);
        else
                ret = configfs_attach_item(parent_item, item, dentry);
 
+       spin_lock(&configfs_dirent_lock);
+       sd->s_type &= ~CONFIGFS_USET_IN_MKDIR;
+       spin_unlock(&configfs_dirent_lock);
+
 out_unlink:
        if (ret) {
                /* Tear down everything we built up */
        }
 
        spin_lock(&configfs_dirent_lock);
-       ret = configfs_detach_prep(dentry);
-       if (ret) {
-               configfs_detach_rollback(dentry);
-               spin_unlock(&configfs_dirent_lock);
-               config_item_put(parent_item);
-               return ret;
-       }
+       do {
+               struct mutex *wait_mutex;
+
+               ret = configfs_detach_prep(dentry, &wait_mutex);
+               if (ret) {
+                       configfs_detach_rollback(dentry);
+                       spin_unlock(&configfs_dirent_lock);
+                       if (ret != -EAGAIN) {
+                               config_item_put(parent_item);
+                               return ret;
+                       }
+
+                       /* Wait until the racing operation terminates */
+                       mutex_lock(wait_mutex);
+                       mutex_unlock(wait_mutex);
+
+                       spin_lock(&configfs_dirent_lock);
+               }
+       } while (ret == -EAGAIN);
        spin_unlock(&configfs_dirent_lock);
 
        /* Get a working ref for the duration of this function */
                          I_MUTEX_PARENT);
        mutex_lock_nested(&dentry->d_inode->i_mutex, I_MUTEX_CHILD);
        spin_lock(&configfs_dirent_lock);
-       if (configfs_detach_prep(dentry)) {
+       if (configfs_detach_prep(dentry, NULL)) {
                printk(KERN_ERR "configfs: Tried to unregister non-empty subsystem!\n");
        }
        spin_unlock(&configfs_dirent_lock);