-
- /* We are not actually interested in the file handles, but
- * name_to_handle_at() also passes us the mount ID, hence use
- * it but throw the handle away */
-
- r = name_to_handle_at(fd, "", &h.handle, &mount_id, AT_EMPTY_PATH);
+ assert(filename);
+
+ /* First we will try the name_to_handle_at() syscall, which
+ * tells us the mount id and an opaque file "handle". It is
+ * not supported everywhere though (kernel compile-time
+ * option, not all file systems are hooked up). If it works
+ * the mount id is usually good enough to tell us whether
+ * something is a mount point.
+ *
+ * If that didn't work we will try to read the mount id from
+ * /proc/self/fdinfo/<fd>. This is almost as good as
+ * name_to_handle_at(), however, does not return the
+ * opaque file handle. The opaque file handle is pretty useful
+ * to detect the root directory, which we should always
+ * consider a mount point. Hence we use this only as
+ * fallback. Exporting the mnt_id in fdinfo is a pretty recent
+ * kernel addition.
+ *
+ * As last fallback we do traditional fstat() based st_dev
+ * comparisons. This is how things were traditionally done,
+ * but unionfs breaks breaks this since it exposes file
+ * systems with a variety of st_dev reported. Also, btrfs
+ * subvolumes have different st_dev, even though they aren't
+ * real mounts of their own. */
+
+ r = name_to_handle_at(fd, filename, &h.handle, &mount_id, flags);