Ceph » CephFS

Looks like the problem is caused by linking b/bar to b/foo. The server response to goes through insert_dentry_inode via insert_trace, and then it incorrectly(?) relinks the inode from b/foo to b/bar. The relink function also removes b/foo from the list of directory contents in the local cache, which is what causes cfuse to try and delete the not-empty dir. I think the relink is the incorrect step, but am checking, and then need to figure out how to differentiate between good and bad relinking.

Okay, a proper fix for this is going to require a bit of work, since right now Inodes can only have one parent dentry. A hacky fix is to remove the I_COMPLETE flag on olddir in the relink function; this will make the client go to the MDS for the complete dir contents, though it'll slow stuff down. Hopefully the client will quickly re-establish the I_COMPLETE once it does go to the MDS.
This does fix the issue and I'm continuing to spot-check it on other workunits just to make sure this doesn't have any odd and unexpected consequences. Pushed in commit:c43455cee4b7b45de6bd04454a40bc7016f2d6d1

Bulk updating project=ceph category=ceph-fuse issues to move to fs project so that we can remove the ceph-fuse category from the ceph project