Sys V inode bug

[t blakely]

We've been bitten several time in the last few months by the infamous
System V disappearing inode bug.  To make things worse, we have found
that fsck, instead of recovering the "lost" inodes, trashed the file
system.  The first time we run out of inodes, fsck appears to fix the
problem.  The next time, fsck starts deleting files and directories.
It never completes without dup table overflows and running out of space
in lost+found.  Subsequent runs of fsck destroy more and more of the
file system until eventually nothing is left.  fsck never completes
without errors after the first time.  This is our news partition, so
it gets lots of activity.  It's a full 147 MB partition on a SCSI
drive on a 3B2/500 running V.3.1.  I made the file system originally
with 60000 inodes because we _really_ ran out using the default
number (full news feed, two week retention).

If anyone has _any_ solution, workaround, whatever for this problem,
I'd love to hear it.  Is fsck buggy or am I causing the problem with
too many inodes (how does it ever work on _really_ big file systems?)
Maybe I should split the disk into 2 file systems, but that would
complicate news a bit.  I've only been working with SysV for about a
year now, but in several years on Berkeley systems I've never had any
problems like this.  I understand that the "disappearing inode" bug
has been in Sys V since before the beginning.  If so, what's wrong
with AT&T (aside from the obvious) that they can't fix something
like this?

Tom Blakely
Univ. of North Fla.
(904)646-2820
uflorida!unf7!tfb
tfb at unfvm.BITNET