Re: [PATCH 2/9] vfs: export do_splice_direct() to modules

[Dave Chinner <david@fromorbit.com>]

On Sun, Mar 17, 2013 at 01:06:59PM +0000, David Howells wrote:
> Miklos Szeredi <miklos@szeredi.hu> wrote:
> 
> > Export do_splice_direct() to modules.  Needed by overlay filesystem.
> 
> Apparently you cannot call this from any function that is holding an i_mutex
> if the target of the splice uses generic_file_splice_write().
> 
> The problem is a potential deadlock situation:
> 
> We have places already that do:
> 
> 	mnt_want_write()
> 	mutex_lock()
> 
> This can be found in do_last() for example.
> 
> However, mnt_want_write() calls sb_start_write() as does
> generic_file_splice_write().  So now in ovl_copy_up_locked() you're adding:
> 
> 	mutex_lock()
> 	sb_start_write()
> 
> which lockdep reports as a potential ABBA deadlock.
> 
> Now, looking at __sb_start_write(), I'm not entirely sure how the deadlock
> might operate, so it's possible that this is a false alarm.  Maybe Jan Kara can
> illuminate further, so I've added him to the cc list.
> 
> I've attached the report I got with unionmount.

There's plenty of problems with splice locking that can lead to
deadlocks. Here's another that's been known for ages:

http://oss.sgi.com/archives/xfs/2011-08/msg00168.html

Cheers,

Dave.
-- 
Dave Chinner
david@fromorbit.com