NFS incoherency.

[Ian Dall]

We have a sun4 named crackle and a sun3 named basil. Both run SunOS 4.0.3.
Basil has a small disk, but most user files reside on a crackle partition
which is NFS mounted on basil.

If I update the file on crackle using emacs. I can then cat the file on
crackle and basil and get different results. This is not reproducable on
demand, but once the two machines have become inconsistent, it tends to
stay that way. 

Emacs renames the old file and creates a new one when you update a file.
For definiteness, consider a file called foo :-) When I update it, it is
renamed to foo~ and a new file foo is created.

An ls -i on both machines shows that the inode basil thinks is associated
with foo is actually the inode of foo~ on crackle. It seems that directory
entries are being cached incoherently!

BSD has some hash table to speed up directory lookups doesn't it? My bet
is that Sun thought through the disk block buffer consistency problem, but
forgot about this directory hash table!

Has anyone else seen this?

Ian Dall
   ACSnet: idall at augean.oz
 Internet: idall at augean.oz.au