Re: 2.6.24-rc8: possible circular locking dependency detected

[David Chinner <dgc@sgi.com>]

On Fri, Jan 18, 2008 at 10:45:17PM +0100, Christian Kujau wrote:
> Hi,
> 
> just FYI, upgrading to -rc8 gave the following messages in kern.log in
> the morning hours, when the backups were run:
> 
> =======================================================
> [ INFO: possible circular locking dependency detected ]
> 2.6.24-rc8 #2
> -------------------------------------------------------
> rsync/23295 is trying to acquire lock:
>  (iprune_mutex){--..}, at: [<c017a552>] shrink_icache_memory+0x72/0x220
> 
> but task is already holding lock:
>  (&(&ip->i_iolock)->mr_lock){----}, at: [<c0275056>] xfs_ilock+0x96/0xb0
> 
> which lock already depends on the new lock.

memory reclaim can occur when an inode lock is held,
causing i_iolock -> iprune_mutex to occur. This is quite
common.

During reclaim, while holding iprune_mutex, we lock a
different inode to complete the cleaning up of it,
resulting in iprune_mutex -> i_iolock.

At this point, lockdep gets upset and blats out a warning.

But, there's no problem here as it is always safe for us
to take the i_iolock in inode reclaim because it can never
be the same as the i_iolock that we've taken prior to memory
reclaim being entered. Therefore false positive.

Lockdep folk - we really need an annotation to prevent this false
positive from being reported because we are getting reports at
least once a week....

Cheers,

Dave.
-- 
Dave Chinner
Principal Engineer
SGI Australian Software Group