Re: [PATCH v2 2.6.38-rc8-tip 7/20] 7: uprobes: store/restore original instruction.

[Peter Zijlstra <peterz@infradead.org>]

On Tue, 2011-03-15 at 14:52 +0530, Srikar Dronamraju wrote:
> * Stephen Wilson <wilsons@start.ca> [2011-03-14 14:09:14]:
> 
> > On Mon, Mar 14, 2011 at 07:05:22PM +0530, Srikar Dronamraju wrote:
> > >  static int install_uprobe(struct mm_struct *mm, struct uprobe *uprobe)
> > >  {
> > > -	int ret = 0;
> > > +	struct task_struct *tsk;
> > > +	int ret = -EINVAL;
> > >  
> > > -	/*TODO: install breakpoint */
> > > -	if (!ret)
> > > +	get_task_struct(mm->owner);
> > > +	tsk = mm->owner;
> > > +	if (!tsk)
> > > +		return ret;
> > 
> > I think you need to check that tsk != NULL before calling
> > get_task_struct()...
> > 
> 
> Guess checking for tsk != NULL would only help if and only if we are doing
> within rcu.  i.e we have to change to something like this
> 
> 	rcu_read_lock()
> 	if (mm->owner) {
> 		get_task_struct(mm->owner)
> 		tsk = mm->owner;
> 	}
> 	rcu_read_unlock()
> 	if (!tsk)
> 		return ret;

so the whole mm->owner semantics seem vague, memcontrol.c doesn't seem
consistent in itself, one site uses rcu_dereference() the other site
doesn't.

Also, the assignments in kernel/fork.c and kernel/exit.c don't use
rcu_assign_pointer() and therefore lack the needed write barrier.

Git blames Balbir for this.