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

[Josh Poimboeuf <jpoimboe@redhat.com>]

On Wed, Feb 11, 2015 at 05:28:13PM +0100, Miroslav Benes wrote:
> On Tue, 10 Feb 2015, Josh Poimboeuf wrote:
> 
> > On Tue, Feb 10, 2015 at 04:59:17PM +0100, Miroslav Benes wrote:
> > > 
> > > On Mon, 9 Feb 2015, Josh Poimboeuf wrote:
> > > 
> > > > Add a basic per-task consistency model.  This is the foundation which
> > > > will eventually enable us to patch those ~10% of security patches which
> > > > change function prototypes and/or data semantics.
> > > > 
> > > > When a patch is enabled, livepatch enters into a transition state where
> > > > tasks are converging from the old universe to the new universe.  If a
> > > > given task isn't using any of the patched functions, it's switched to
> > > > the new universe.  Once all the tasks have been converged to the new
> > > > universe, patching is complete.
> > > > 
> > > > The same sequence occurs when a patch is disabled, except the tasks
> > > > converge from the new universe to the old universe.
> > > > 
> > > > The /sys/kernel/livepatch/<patch>/transition file shows whether a patch
> > > > is in transition.  Only a single patch (the topmost patch on the stack)
> > > > can be in transition at a given time.  A patch can remain in the
> > > > transition state indefinitely, if any of the tasks are stuck in the
> > > > previous universe.
> > > > 
> > > > A transition can be reversed and effectively canceled by writing the
> > > > opposite value to the /sys/kernel/livepatch/<patch>/enabled file while
> > > > the transition is in progress.  Then all the tasks will attempt to
> > > > converge back to the original universe.
> > > 
> > > Hi Josh,
> > > 
> > > first, thanks a lot for great work. I'm starting to go through it and it's 
> > > gonna take me some time to do and send a complete review.
> > 
> > I know there are a lot of details to look at, please take your time.  I
> > really appreciate your review.  (And everybody else's, for that matter
> > :-)
> > 
> > > > +	/* success! unpatch obsolete functions and do some cleanup */
> > > > +
> > > > +	if (klp_universe_goal == KLP_UNIVERSE_OLD) {
> > > > +		klp_unpatch_objects(klp_transition_patch);
> > > > +
> > > > +		/* prevent ftrace handler from reading old func->transition */
> > > > +		synchronize_rcu();
> > > > +	}
> > > > +
> > > > +	pr_notice("'%s': %s complete\n", klp_transition_patch->mod->name,
> > > > +		  klp_universe_goal == KLP_UNIVERSE_NEW ? "patching" :
> > > > +							  "unpatching");
> > > > +
> > > > +	klp_complete_transition();
> > > > +}
> > > 
> > > ...synchronize_rcu() could be insufficient. There still can be some  
> > > process in our ftrace handler after the call.
> > > 
> > > Consider the following scenario:
> > > 
> > > When synchronize_rcu is called some process could have been preempted on 
> > > some other cpu somewhere at the start of the ftrace handler before  
> > > rcu_read_lock. synchronize_rcu waits for the grace period to pass, but that 
> > > does not mean anything for our process in the handler, because it is not 
> > > in rcu critical section. There is no guarantee that after synchronize_rcu 
> > > the process would be away from the handler. 
> > > 
> > > "Meanwhile" klp_try_complete_transition continues and calls 
> > > klp_complete_transition. This clears func->transition flags. Now the 
> > > process in the handler could be scheduled again. It reads the wrong value 
> > > of func->transition and redirection to the wrong function is done.
> > > 
> > > What do you think? I hope I made myself clear.
> > 
> > You really made me think.  But I don't think there's a race here.
> > 
> > Consider the two separate cases, patching and unpatching:
> > 
> > 1. patching has completed: klp_universe_goal and all tasks'
> >    klp_universes are at KLP_UNIVERSE_NEW.  In this case, the value of
> >    func->transition doesn't matter, because we want to use the func at
> >    the top of the stack, and if klp_universe is NEW, the ftrace handler
> >    will do that, regardless of the value of func->transition.  This is
> >    why I didn't do the rcu_synchronize() in this case.  But maybe you're
> >    not worried about this case anyway, I just described it for the sake
> >    of completeness :-)
> 
> Yes, this case shouldn't be a problem :)
> 
> > 2. unpatching has completed: klp_universe_goal and all tasks'
> >    klp_universes are at KLP_UNIVERSE_OLD.  In this case, the value of
> >    func->transition _does_ matter.  However, notice that
> >    klp_unpatch_objects() is called before rcu_synchronize().  That
> >    removes the "new" func from the klp_ops stack.  Since the ftrace
> >    handler accesses the list _after_ calling rcu_read_lock(), it will
> >    never see the "new" func, and thus func->transition will never be
> >    set.
> 
> Hm, so indeed I messed it up. Let me rework the scenario a bit. We have a 
> function foo(), which has been already patched with foo_1() from patch_1 
> and foo_2() from patch_2. Now we would like to unpatch patch_2. It is 
> successfully completed and klp_try_complete_transition calls 
> klp_unpatch_objects and synchronize_rcu. Thus foo_2() is removed from the 
> RCU list in ops. 
> 
> Now to the funny part. After synchronize_rcu() and before 
> klp_complete_transition some process might get to the ftrace handler (it 
> is still there because of the patch_1 still being present). It gets foo_1 
> from the list_first_or_null_rcu, sees that func->transition is 1 (it 
> hasn't been cleared yet)

Same answer as the other email, foo_1()'s func->transition will be 0 :-)

When patching, only the new klp_func gets transition set to 1.

When unpatching, only the klp_func being removed gets transition set to
1.

> , current->klp_universe is KLP_UNIVERSE_OLD... so 
> it tries to get previous function. There is none and foo() is called. This 
> is incorrect.
> 
> It is very similar scenario to the one in my other email earlier this day. 
> I think we need to clear func->transition before calling 
> klp_unpatch_objects. More or less.

-- 
Josh