Re: [PATCH] livepatch: Fix ftrace module text permissions race

[Josh Poimboeuf <jpoimboe@redhat.com>]

On Thu, May 30, 2019 at 03:54:14PM +0200, Petr Mladek wrote:
> On Wed 2019-05-29 14:02:24, Josh Poimboeuf wrote:
> > The above panic occurs when loading two modules at the same time with
> > ftrace enabled, where at least one of the modules is a livepatch module:
> > 
> > CPU0					CPU1
> > klp_enable_patch()
> >   klp_init_object_loaded()
> >     module_disable_ro()
> >     					ftrace_module_enable()
> > 					  ftrace_arch_code_modify_post_process()
> > 				    	    set_all_modules_text_ro()
> >       klp_write_object_relocations()
> >         apply_relocate_add()
> > 	  *patches read-only code* - BOOM
> 
> This patch looks fine and fixes the race:
> 
> Reviewed-by: Petr Mladek <pmladek@suse.com>
> 
> 
> That said, the semantic of text_mutex is a bit unclear:
> 
>    + It serializes RO/RW setting but not NX

True.  module_enable_nx() is a static function which is only called
internally.  I should probably rename it to __module_enable_nx() so the
locking semantics match the others.

>    + Nothing prevents manipulation of the access rights
>      by external code before the module is ready-enough.
>      I mean before the sections are set RO by the module
>      loader itself.
> 
>      Most sections are ready in MODULE_STATE_COMMING state.
>      Only ro_after_init sections need to stay RW longer,
>      see my question below.
> 
> 
> > diff --git a/kernel/module.c b/kernel/module.c
> > index 6e6712b3aaf5..3c056b56aefa 100644
> > --- a/kernel/module.c
> > +++ b/kernel/module.c
> > @@ -3519,7 +3534,7 @@ static noinline int do_init_module(struct module *mod)
> >  	/* Switch to core kallsyms now init is done: kallsyms may be walking! */
> >  	rcu_assign_pointer(mod->kallsyms, &mod->core_kallsyms);
> >  #endif
> > -	module_enable_ro(mod, true);
> > +	__module_enable_ro(mod, true);
> 
> The "true" parameter causes that also ro_after_init section is
> set read only. What is the purpose of this section, please?
> 
> I ask because module_enable_ro(mod, true) can be called
> earlier from klp_init_object_loaded() from do_one_initcall().
> 
> For example, it could some MODULE_STATE_LIVE notifier
> when it requires write access to ro_after_init section.

Hm, I think you're right.  klp_init_object_loaded() should change the
module_enable_ro() argument to false when it's called from a patch
module's initcall.

Maybe we can instead remove __module_enable_ro()'s after_init argument
and just make it smarter?  It should only do ro_after_init frobbing if
the module state is MODULE_STATE_LIVE.

> Anyway, the above is a separate problem. This patch looks
> fine for the original problem.

Thanks for the review.  I'll post another version, with the above
changes and with the patches split up like Miroslav suggested.

-- 
Josh