While writing the on-disk-format upgrade tool for the next fhgfs version, I also added a sanity check for the hash-directories. And while testing it it reported a problem:

* Migrating inodes... (Processing 8192 hash directories...)
* Progress: 
Warning entry is not a directory: /fhgfs/meta/meta.dir//entries/A7 type: 8
Warning entry is not a directory: /fhgfs/meta/meta.dir//entries/177 type: 8
Warning entry is not a directory: /fhgfs/meta/meta.dir//entries/739 type: 8
Warning entry is not a directory: /fhgfs/meta/meta.dir//entries/835 type: 8
Warning entry is not a directory: /fhgfs/meta/meta.dir//entries/D84 type: 8
Warning entry is not a directory: /fhgfs/meta/meta.dir//entries/111C type: 8
Warning entry is not a directory: /fhgfs/meta/meta.dir//entries/13C1 type: 8
Warning entry is not a directory: /fhgfs/meta/meta.dir//entries/15B7 type: 8
Warning entry is not a directory: /fhgfs/meta/meta.dir//entries/17EB type: 8
Warning entry is not a directory: /fhgfs/meta/meta.dir//entries/19CA type: 8
Warning entry is not a directory: /fhgfs/meta/meta.dir//entries/1FD4 type: 8

So somehow those few directories out of 8192 (for dentries and inodes, each, so two times 8192) seem to have a wrong directory type. 
So I went ahead and checked if e2fsck would correct it. But e2fsck reported no errors. So I went further ahead and added some debug code to e2fsck and indeed, e2fsck sees the correct type:

inode: 6293106 nameLen: 4 entryName: 17EB fileType: 2 should_be: 2

(fprintf line in check_filetype() ).

For verification, the stat() output also looks right:

schubert@squeeze2 ~>stat /fhgfs/meta/meta.dir//entries/17EB
  File: `/fhgfs/meta/meta.dir//entries/17EB'
  Size: 4096            Blocks: 8          IO Block: 4096   directory
Device: fe10h/65040d    Inode: 6293106     Links: 2
Access: (0755/drwxr-xr-x)  Uid: ( 5741/schubert)   Gid: (  100/   users)
Access: 2012-11-07 13:37:00.280689000 +0100
Modify: 2012-10-31 16:27:14.024689006 +0100
Change: 2012-10-31 16:27:14.024689006 +0100


So somehow either libc or kernel report a wrong file_type field for those dentries. I will do some in-kernel debugging later on.