Mysterious File System Curdling (4.2bsd)

[urban at trwspp.UUCP]

We are running 4.2bsd on an 11/780 with System Industries
9900 disk controllers (using the BRL device driver).
Periodically, /usr (which resides on our RM05-type drive)
gets trashed due to a duplicate block in different files.
When fsck tries to clean up the mess, a mysterious thing
happens: it finds dups in the cylinder groups, and claims
to be correcting the problem.  But if we run fsck again, it
finds a different cylinder group problem, and tries to
correct it... this process can continue indefinitely, with
the file system never getting fixed, and the number of free
blocks and fragment count never appearing to be the same
thing twice.  Kirk McKusick says he never heard of a
problem like this; the fsck code for finding free blocks
and ensuring they're not in files seems pretty
straightforward.  Right now, our only solution to the
problem is to use dump/newfs/restore.  This is getting
annoying; this morning the losing file was the directory
/usr/spool/news/net/legal.  Has anyone out there
experienced anything like this before?


	Mike