Re: [PATCH v3] livepatch: Prevent to apply the patch once coming module notifier fails

[Josh Poimboeuf <jpoimboe@redhat.com>]

On Mon, May 18, 2015 at 05:50:52PM +0200, Petr Mladek wrote:
> On Mon 2015-05-18 10:22:21, Josh Poimboeuf wrote:
> > On Mon, May 18, 2015 at 02:08:06PM +0200, Petr Mladek wrote:
> > > On Wed 2015-05-13 09:14:15, Josh Poimboeuf wrote:
> > > > On Tue, May 12, 2015 at 10:04:44PM +0800, Minfei Huang wrote:
> > > > > @@ -930,6 +932,7 @@ disabled:
> > > > >  static int klp_module_notify(struct notifier_block *nb, unsigned long action,
> > > > >  			     void *data)
> > > > >  {
> > > > > +	int ret;
> > > > >  	struct module *mod = data;
> > > > >  	struct klp_patch *patch;
> > > > >  	struct klp_object *obj;
> > > > > @@ -955,7 +958,13 @@ static int klp_module_notify(struct notifier_block *nb, unsigned long action,
> > > > >  
> > > > >  			if (action == MODULE_STATE_COMING) {
> > > > >  				obj->mod = mod;
> > > > > -				klp_module_notify_coming(patch, obj);
> > > > > +				ret = klp_module_notify_coming(patch, obj);
> > > > > +				if (ret) {
> > > > > +					obj->mod = NULL;
> > > > > +					pr_warn("patch '%s' is dead, remove it "
> > > > > +						"or re-install the module '%s'\n",
> > > > > +						patch->mod->name, obj->name);
> > > > > +				}
> > > > 
> > > > The patch isn't necessarily dead, since it might also include previously
> > > > enabled changes for vmlinux or other modules.  It can actually be a
> > > > dangerous condition if there's a mismatch between old code in the module
> > > > and new code elsewhere.  How about something like:
> > > > 
> > > > "patch '%s' is in an inconsistent state!\n"
> > > 
> > > It must not be dangerous, otherwise the patch could not get applied
> > > immediately.
> > >
> > > I would omit this message completely. It would just duplicate the
> > > warning printed by klp_module_notify_coming().
> > 
> > This error path doesn't mean that the entire patch isn't applied.  It
> > only affects the subset of the patch which applies to the coming module.
> > So you can have a dangerous mismatch in the case of a patch which
> > patches multiple objects.
> 
> We apply the patch immediately. This simple consistency model allows
> to call patched function from an upatched one and vice versa. It means
> that there must _not_ be any dependency between patched functions.
> And it means that it must be safe to keep the module unpatched.
> 
> The situation will change after we introduce a more complex
> consistency model. Then we will need to patch the module
> directly in load_module() and refuse loading in case of error.
> By other words, we will not and must not allow any dangerous state.
> 
> Does it make sense? Or did I miss anything, please?

Yeah, ok, that makes sense.  Given the simple consistency model, it's
not dangerous.  It's still inconsistent, and something the user should
know about, but perhaps a single warning in klp_module_notify_coming()
is enough.

I also agree that, once we have a better consistency model, failing to
load the module would be a better way to handle this error.

-- 
Josh