inode allocation bug.

[David Higgen]

Clarifying note. 

(Summary: SGI filesystem is different: any problems seen
on SGI machines are NOT the "well known sysV inode problem").

There have been a number of discussions lately about an inode allocation bug
on sysV systems: a filesystem will appear to be out of inodes even though
all inodes are not really allocated (as shown by the fact that an fsck of
the filesystem restores the "missing" inodes).

The mechanism of this on the Bell filesystem appears to be understood.
However, the SGI filesystem (EFS) does not contain any sysV code, and its
inode allocation method is different; there is no free i-list in the
superblock.

Nonetheless, there is evidence that a symptomatically similar problem has been
seen on SGI systems. We are anxious to find and fix this, however to date
we don't have a working hypothesis nor a test case. Tests designed to elicit
the Bell filesystem bug fail, not unexpectedly, to produce any problems on
SGI systems.

I would like to hear from anyone who has seen this type of problem on an
SGI system, with as much information as you can give me. A reproducible
test case will buy you Eternal Gratitude, of course!

Thanks,

			Dave Higgen (daveh at xtenk.asd.sgi.com)