Re: [PATCH v2] Staging: dgnc: release the lock before testing for nullity

[Dan Carpenter <dan.carpenter@oracle.com>]

On Wed, Mar 18, 2015 at 02:43:01PM +0100, Quentin Lambert wrote:
> 
> 
> On 18/03/2015 14:36, Dan Carpenter wrote:
> >This changelog still doesn't make sense so I took a look at the code.
> >
> >tty_ldisc_deref() is an unlock function.  So this is a lock ordering
> >bug.  What makes you think the original ordering was correct?  Who
> >reported this bug?  What are the effects of this bug?
> I was the one who introduced the ordering change in the first place.
> I am just trying to fix it because although nobody complained I am not
> sure of the impact and restoring the previous control flow seems to be the
> right thing to do.

Your changelog should tell me this stuff.

The original code is wrong.  We take "spin_lock_irqsave(&ch->ch_lock,
flags);" before we do "ld = tty_ldisc_ref(tp);" so we should deref
before we unlock.

It's normally:

lock_outer();
lock_inner();
unlock_inner();
unlock_outer();

On the success path we unlock first then deref and that is a mistake.

This kind of change is a bit dangerous though so it requires testing.

regards,
dan carpenter