Re: [PATCH] Avoid null-pointer access in w1/slaves/w1_therm

[David Fries <david@fries.net>]

It has not been solved.  Evgeniy would like to make use of the sysfs
device management instead of the current reference counting, however I
haven't heard any volunteers to do that work.  I posted a quick fix
patch, it was very easy to crash without this patch, it doesn't
completely solve the race conditions, and I don't think it can be
solved in just a slave driver change.

Are you up for the challenge?

On Wed, Apr 15, 2015 at 09:52:27AM +0200, Jonathan ALIBERT wrote:
> Hi,
> 
> Do you know if the problem has been solved ?
> 
> Cheers,
> 
> *Jonathan ALIBERT*
> *06 32 26 59 12*
> *265, route de Saint Haon*
> *42 370 RENAISON*
> 
> 
> 2015-03-19 1:09 GMT+01:00 David Fries <david@fries.net>:
> 
> > On Wed, Mar 18, 2015 at 06:18:53PM +0300, Evgeniy Polyakov wrote:
> > > Hi
> > >
> > > 18.03.2015, 07:20, "David Fries" <david@fries.net>:
> > > >  static void w1_therm_remove_slave(struct w1_slave *sl)
> > > >  {
> > > > + int refcnt = atomic_sub_return(1, THERM_REFCNT(sl->family_data));
> > > > + while(refcnt) {
> > > > + msleep(1000);
> > > > + refcnt = atomic_read(THERM_REFCNT(sl->family_data));
> > > > + }
> > > >          kfree(sl->family_data);
> > > >          sl->family_data = NULL;
> > > >  }
> > >
> > > Can we replace this whole atomic manipulations with kref_t and free
> > family data in the place
> > > which actually drops reference counter to zero?
> > >
> > > I.e. we return from remove_slave() function potentially leaving family
> > data floating around, it will be freed
> > > when the last user drops the reference. There is still a race between
> > increasing reference when starting
> > > reading and removing slave device, i.e. one starts reading, while
> > attached slave device is being removed,
> > > but that's a different problem.
> >
> > With the two while loops I posted, I see with two clients reading
> > w1_slave, the other command to remove a slave gets permanently stuck
> > in w1_therm_remove_slave, which keeps the slave around while the
> > clients continue to read.  I wouldn't predict things going better by
> > keeping family_data around longer, the slave data would still go away
> > with readers around.
> >
> > --
> > David Fries <david@fries.net>    PGP pub CB1EE8F0
> > http://fries.net/~david/
> >

-- 
David Fries <david@fries.net>    PGP pub CB1EE8F0
http://fries.net/~david/