fdio: fix fd offset handling with `FICLONE`

When using `FICLONE`, the kernel does a full file clone and disregards the
fd offsets. Users of this API however assume that it *is*
offset-sensitive. So we need to verify that the fd offsets are at the
start of the files before we call `FICLONE`. This was done in systemd also
in:

https://github.com/systemd/systemd/commit/c622fbdb8d37

The commit message does not explain this but `ioctl_ficlone(2)` says:

    The `FICLONE` ioctl clones entire files.

(Compare with `FICLONERANGE`, which takes a struct with offsets and the
length).

Similarly, we need to seek to the end of the file descriptors on success
so that we're consistent with the behaviour of the other backends
available (`copy_file_range`, `sendfile` and manual copying). This also
matches what systemd does nowadays:

https://github.com/systemd/systemd/blob/80f967311ac5/src/shared/copy.c#L199

1 files changed, 13 insertions, 2 deletions

diff --git a/glnx-fdio.c b/glnx-fdio.c
index ee69aa1..b1af679 100644
--- a/glnx-fdio.c
+++ b/glnx-fdio.c
@@ -794,10 +794,21 @@ glnx_regfile_copy_bytes (int fdf, int fdt, off_t max_bytes)
 
   /* If we've requested to copy the whole range, try a full-file clone first.
    */
-  if (max_bytes == (off_t) -1)
+  if (max_bytes == (off_t) -1 &&
+      lseek (fdf, 0, SEEK_CUR) == 0 &&
+      lseek (fdt, 0, SEEK_CUR) == 0)
     {
       if (ioctl (fdt, FICLONE, fdf) == 0)
-        return 0;
+        {
+          /* All the other methods advance the fds. Do it here too for consistency. */
+          if (lseek (fdf, 0, SEEK_END) < 0)
+            return -1;
+          if (lseek (fdt, 0, SEEK_END) < 0)
+            return -1;
+
+          return 0;
+        }
+
       /* Fall through */
       struct stat stbuf;