Ceph » CephFS

We have a known damage in a production cluster, https://tracker.ceph.com/issues/60986#note-15

We know that damage is irreparable as there has been a data loss.

I am able to reproduce it easily by some hack.

1. Create a directory e.g, /dir
2. Create a file inside /dir , e.g. /dir/x.txt
3. Create a link e.g. /y.txt of /dir/x.txt
4. remove the associdated inode object of /dir from the meta pool(Just to create the damage forcefully)
5. Try to stat the /y.txt and damage will occur.
6. Remove the damage from the damage list
7. Scrub the root directory, damage list will remain empty.

I found in the ScrubStack.cc code,

      if (dnl->is_primary()) {
    _enqueue(dnl->get_inode(), header, false);
      } else if (dnl->is_remote()) {
    // TODO: check remote linkage
      }

If dnl->is_remote() it is doing nothing, there is TODO marked. What should be ideal in this case as doing scrub will make redundant scrub operations on a single inode. In that case can we just try to open the inode using MDCache::open_remote_dentry(CDentry *dn, bool projected, MDSContext *fin, bool want_xlocked)(This is they way used for finding damage while path traversal)?