Re: [RFC PATCH 6/9] livepatch: create per-task consistency model

[Josh Poimboeuf <jpoimboe@redhat.com>]

On Wed, Feb 11, 2015 at 11:21:51AM +0100, Miroslav Benes wrote:
> 
> On Mon, 9 Feb 2015, Josh Poimboeuf wrote:
> 
> [...]
> 
> > @@ -38,14 +39,34 @@ static void notrace klp_ftrace_handler(unsigned long ip,
> >  	ops = container_of(fops, struct klp_ops, fops);
> >  
> >  	rcu_read_lock();
> > +
> >  	func = list_first_or_null_rcu(&ops->func_stack, struct klp_func,
> >  				      stack_node);
> > -	rcu_read_unlock();
> >  
> >  	if (WARN_ON_ONCE(!func))
> > -		return;
> > +		goto unlock;
> > +
> > +	if (unlikely(func->transition)) {
> > +		/* corresponding smp_wmb() is in klp_init_transition() */
> > +		smp_rmb();
> > +
> > +		if (current->klp_universe == KLP_UNIVERSE_OLD) {
> > +			/*
> > +			 * Use the previously patched version of the function.
> > +			 * If no previous patches exist, use the original
> > +			 * function.
> > +			 */
> > +			func = list_entry_rcu(func->stack_node.next,
> > +					      struct klp_func, stack_node);
> > +
> > +			if (&func->stack_node == &ops->func_stack)
> > +				goto unlock;
> > +		}
> > +	}
> >  
> >  	klp_arch_set_pc(regs, (unsigned long)func->new_func);
> > +unlock:
> > +	rcu_read_unlock();
> >  }
> 
> I decided to understand the code more before answering the email about the 
> race and found another problem. I think.
> 
> Imagine we patched some function foo() with foo_1() from patch_1 and now 
> we'd like to patch it again with foo_2() in patch_2. __klp_enable_patch 
> calls klp_init_transition which sets klp_universe for all processes to 
> KLP_UNIVERSE_OLD and marks the foo_2() for transition (it is gonna be 1). 
> Then __klp_enable_patch adds foo_2() to the RCU-protected list for foo(). 
> BUT what if somebody calls foo() right between klp_init_transition and 
> the loop in __klp_enable_patch? The ftrace handler first returns the 
> first entry in the list which is foo_1() (foo_2() is still not present), 
> then it checks for func->transition. It is 1.

No, actually foo_1()'s func->transition will be 0.  Only foo_2()'s
func->transition will be 1.

> It checks for 
> current->klp_universe which is KLP_UNIVERSE_OLD and so the next entry is 
> retrieved. There is no such and therefore foo() is called. This is 
> obviously wrong because foo_1() was expected.
> 
> Everything would work fine if one would call foo() before 
> klp_start_transition and after the loop in __klp_enable_patch. The 
> solution might be to move the setting of func->transition to 
> klp_start_transition, but this could break something different. I don't 
> know yet.
> 
> Am I wrong?
> 
> Miroslav

-- 
Josh