Re: [PATCH v5 1/3] livepatch: add (un)patch callbacks

[Petr Mladek]

On Tue 2017-09-26 15:01:52, Joe Lawrence wrote:
> On 09/26/2017 10:49 AM, Petr Mladek wrote:
> > On Thu 2017-08-31 10:53:51, Joe Lawrence wrote:
> >> diff --git a/kernel/livepatch/core.c b/kernel/livepatch/core.c
> >> index b9628e43c78f..aca62c4b8616 100644
> >> --- a/kernel/livepatch/core.c
> >> +++ b/kernel/livepatch/core.c
> >> @@ -54,11 +54,6 @@ static bool klp_is_module(struct klp_object *obj)
> >>  	return obj->name;
> >>  }
> >>  
> >> -static bool klp_is_object_loaded(struct klp_object *obj)
> >> -{
> >> -	return !obj->name || obj->mod;
> >> -}
> >> -
> >>  /* sets obj->mod if object is not vmlinux and module is found */
> >>  static void klp_find_object_module(struct klp_object *obj)
> >>  {
> >> @@ -285,6 +280,8 @@ static int klp_write_object_relocations(struct module *pmod,
> >>  
> >>  static int __klp_disable_patch(struct klp_patch *patch)
> >>  {
> >> +	struct klp_object *obj;
> >> +
> >>  	if (klp_transition_patch)
> >>  		return -EBUSY;
> >>  
> >> @@ -295,6 +292,10 @@ static int __klp_disable_patch(struct klp_patch *patch)
> >>  
> >>  	klp_init_transition(patch, KLP_UNPATCHED);
> >>  
> >> +	klp_for_each_object(patch, obj)
> >> +		if (patch->enabled && obj->patched)
> >> +			klp_pre_unpatch_callback(obj);
> >> +
> >>  	/*
> >>  	 * Enforce the order of the func->transition writes in
> >>  	 * klp_init_transition() and the TIF_PATCH_PENDING writes in
> >> @@ -388,13 +389,18 @@ static int __klp_enable_patch(struct klp_patch *patch)
> >>  		if (!klp_is_object_loaded(obj))
> >>  			continue;
> >>  
> >> -		ret = klp_patch_object(obj);
> >> +		ret = klp_pre_patch_callback(obj);
> >>  		if (ret) {
> >> -			pr_warn("failed to enable patch '%s'\n",
> >> -				patch->mod->name);
> >> +			pr_warn("pre-patch callback failed for object '%s'\n",
> >> +				klp_is_module(obj) ? obj->name : "vmlinux");
> >> +			goto err;
> >> +		}
> >>  
> >> -			klp_cancel_transition();
> >> -			return ret;
> >> +		ret = klp_patch_object(obj);
> >> +		if (ret) {
> >> +			pr_warn("failed to patch object '%s'\n",
> >> +				klp_is_module(obj) ? obj->name : "vmlinux");
> > 
> > We should call klp_post_unpatch_callback(obj) here to make it
> > synchronous.
> 
> Are you talking about the error path?  As its coded here,
> klp_cancel_transition() will call klp_complete_transition() with
> klp_target_state = KLP_UNPATCHED and then klp_complete_transition()'s
> done: code will call klp_post_unpatch_callback() on all the necessary
> kobj's.  Is there something asynchronous about that?

Ah, I have missed it. It is a bit tricky ;-)


> > Well, what about calling:
> > 
> >       klp_pre_patch_callback() inside klp_patch_object() and
> >       klp_post_unpatch_callback() inside klp_unpatch_object()
> 
> v1 started out that way, but we migrated to placing these around the
> callers of klp_(un)patch_object() to try and better line up the
> locations of the pre- hooks  with the post- hook locations.

I guess that the move was mainly motivated by introducing 4 callbacks
instead of only two of them.

On one hand, it is fine to see a symmetric code like, for example,
in klp_module_going():

	klp_pre_unpatch_callback(obj);
	klp_unpatch_object(obj);
	klp_post_unpatch_callback(obj);


On the other hand, it adds yet another asymmetry between
__klp_enable_patch()/__klp_disable_patch() and
klp_finish_transition(), see my confusion above.

I know that that the asymmetry was already there because of the
klp_patch_object() and klp_unpatch_object().

I mean that klp_patch_object() calls klp_unpatch_object() in case
of errors. But this handles only the current object. We still
rely on calling klp_cancel_transition()->klp_complete_transition()
to call klp_unpatch_object() for the other already proceed objects.


> I can take a second look at reversing this decision, but that may take a
> little time while I page all the testing corner cases back into my brain :)

I am sorry for the late reply. Heh, I needed to refresh a lot of
things as well. The advantage is that one could see things from
new perspective when the head was cleaned in between ;-)


> > By other words, we would do the two operations. It would have
> > two advantages:
> > 
> >    + error handling for free
> >    + no need for the strange callbacks_enabled flag
> 
> Indeed, it would be nice to ditch that callbacks_enabled wart.

Yup, I hope that in this case the less states would mean
the easier logic. And handling of klp_patch_object()/klp_unpatch()
object is already tricky enough. It would be lovely to just reuse
it if we can.


> I think the only other outstanding issue before rolling a v6 is the one
> that Miroslav raised about the error path in klp_module_coming():
> 
>   https://marc.info/?l=linux-kernel&m=150590635602784&w=2
>   https://marc.info/?l=linux-kernel&m=150592065007463&w=2

I am going to look at it.

Best Regards,
Petr