Re: [PATCH 4/4] livepatch: Remove the redundant enabled flag in struct klp_patch

[Joe Lawrence]

On 1/22/19 5:06 AM, Miroslav Benes wrote:
> On Wed, 16 Jan 2019, Petr Mladek wrote:
>
> > diff --git a/kernel/livepatch/core.c b/kernel/livepatch/core.c
> > index 684766d306ad..8e644837e668 100644
> > --- a/kernel/livepatch/core.c
> > +++ b/kernel/livepatch/core.c
> > @@ -59,6 +59,17 @@ static bool klp_is_module(struct klp_object *obj)
> >   	return obj->name;
> >   }
> >   
> > +static bool klp_patch_enabled(struct klp_patch *patch)
> > +{
> > +	if (patch == klp_transition_patch) {
> > +		WARN_ON_ONCE(klp_target_state == KLP_UNDEFINED);
>
> I think we'd have a race in the code then. enabled_show() does not take
> klp_mutex() when calling klp_patch_enabled().
>
> A patch sysfs attributes are added quite early during its enablement.
> klp_init_transition() first sets klp_transition_patch, then
> klp_target_state. It means one can call enabled_show() with patch ==
> klp_transition_patch and klp_target_state == KLP_UNDEFINED. No?
>
> The similar applies the disablement. klp_complete_transition() first
> clears klp_target_state (sets it to KLP_UNDEFINED), then it clears
> klp_transition_patch.
>
> We could add locking to enabled_show() or swap the assignments with some
> barriers on top.

Taking the mutex as enabled_store() does would be simplest, no?

> Or we could remove WARN_ON_ONCE() and live with false results in
> enabled_show(). It does not matter much, I think. All the other call sites
> of klp_patch_enabled() should be fine.

Hmm, the self-tests and the kpatch tool inspect the sysfs files, but as 
long as the false result is a stale value, I think they would be okay. 
Those tools poll sysfs and don't depend on a one-shot-read value to make 
an enabled/disabled determination.

> > +		return klp_target_state == KLP_PATCHED;
> > +	}
> > +
> > +	return !list_empty(&patch->list);
> > +}
>
> Shouldn't we also change list_del(&patch->list) in klp_free_patch_start()
> to list_del_init(&patch->list)?

Right, we should do that if klp_patch_enabled() is going to subsequently 
check that list.

> > @@ -955,7 +964,7 @@ static int __klp_enable_patch(struct klp_patch *patch)
> >   	if (klp_transition_patch)
> >   		return -EBUSY;
> >   
> > -	if (WARN_ON(patch->enabled))
> > +	if (list_empty(&patch->list))
> >   		return -EINVAL;
>
> I wanted to ask why there is list_empty() and not klp_patch_enabled(), so
> just to be sure... the patch was added to klp_patches list, so patch->list
> is not empty (should not be). We could achieve the same by calling
> !klp_patch_enabled() given its implementation, but it would look
> counter-intuitive here.
>
> The rest looks fine.
>
> However, I am not sure if the outcome is better than what we have. Yes,
> patch->enabled is not technically necessary and we can live with that (as
> the patch proves). On the other hand, it gives the reader clear guidance
> about the patch's state. klp_patch_enabled() is not a complete
> replacement. We have to call list_empty() in __klp_enable_patch() or check
> the original klp_target_state in klp_try_complete_transition().
>
> I am not against the change, I am glad to see it is achievable, but I am
> not sure if the code is better with it. Joe acked it. What do the others
> think?

Let me qualify my ack -- I think minimizing the number of state 
variables like patch->enabled can help readability... on the other hand, 
deducing the same information from other properties like list-empty can 
be confusing, ie, klp_patch_enabled() is generally a lot clearer than 
list_empty(&patch->list)).

So I like this idea and would be interested to hear what folks think 
about the exception cases you pointed out.

-- Joe