Re: [PATCH] UDF: fix deadlock on inode being dropped

[Jan Kara <jack@suse.cz>]

On Thu 07-06-07 17:54:58, Cyrill Gorcunov wrote:
> [Jan Kara - Thu, Jun 07, 2007 at 11:36:07AM +0200]
> |   Hi Cyrill!
> | 
> | On Wed 06-06-07 21:53:51, Cyrill Gorcunov wrote:
> | > This patch prevents from deadlock on inode being dropped.
> | > The deadlock is caused by inderect call of mark_inode_dirty()
> | > within udf_drop_inode() but inode lock is already kept
> | > by the kernel. So moving code from udf_drop_inode() to
> | > udf_delete_inode() we save its functionality and avoid
> | > deadlock.
> |   The patch is wrong. You cannot truncate the extent just in delete_inode.
> | That would lead to inodes with untruncated last extent on disk after
> | unmounting, which is forbidden in the specification. You need to truncate
> | the last extent whenever inode is being removed from memory or something
> | like that... I'm already thinking how to do it and avoid calling
> | mark_inode_dirty()...
> | 
> 
> Arh, thanks... Jan, actually the reason I've moved the code into
> 'delete' section was that I found no reasonable difference for our
> case between 'drop' and 'delete'. Moreover, by seeing into VFS code
> the only diff between 'drop' and 'delete' is that
> inside generic_delete_inode() a few inode structure elements
> are being destroyed and then our udf_drop_inode is called. So assuming,
> that you're right in drop_inode I've code just moved to 'delete' section.
  The difference is that udf_delete_inode() is called only when inode has
i_nlink == 0 and thus it's being deleted on disk. udf_drop_inode() is
called whenever inode is removed from memory which is what we want.
  I'm already testing a patch which should fix the problem...

								Honza
-- 
Jan Kara <jack@suse.cz>
SuSE CR Labs