Re: inode_unused list corruption in 2.4.26 - spin_lock problem?

[Marcelo Tosatti <marcelo.tosatti@cyclades.com>]

Hi Chris,

I've seen your previous post -- should have answered you earlier.

On Fri, Jun 18, 2004 at 05:47:05PM -0700, Chris Caputo wrote:
> In 2.4.26 on two different dual-proc x86 machines (one dual-P4 Xeon based,
> the other dual-PIII) I am seeing crashes which are the result of the
> inode_unused doubly linked list in fs/inode.c becoming corrupted.

What steps are required to reproduce the problem?

> A particular instance of the corruption I have isolated is in a call from
> iput() to __refile_inode().  To try to diagnose this further I placed list
> verification code before and after the list_del() and list_add() calls in
> __refile_inode() and observed a healthy list become corrupted after the
> del/add was completed.

Can you show us this data in more detail?

> It would seem to me that list corruption on otherwise healthy machines
> would only be the result of the inode_lock spinlock not being properly
> locked prior to the call to __refile_inode(), but as far as I can tell,
> the call to atomic_dec_and_lock() in iput() is doing that properly.
> 
> So I am at a loss.  Has anyone else seen this or does anyone have any idea
> what routes I should be exploring to fix this problem?

The changes between 2.4.25->2.4.26 (which introduce __refile_inode() and 
the unused_pagecache list) must have something to do with this. 

David, Rik, can you give some help here?