Re: [PATCH v2 3/3] livepatch: Cleanup module page permission changes

[Josh Poimboeuf <jpoimboe@redhat.com>]

On Fri, Nov 06, 2015 at 11:40:55AM +0100, Petr Mladek wrote:
> On Thu 2015-11-05 15:18:05, Josh Poimboeuf wrote:
> > Calling set_memory_rw() and set_memory_ro() for every iteration of the
> > loop in klp_write_object_relocations() is messy, inefficient, and
> > error-prone.
> > 
> > Change all the read-only pages to read-write before the loop and convert
> > them back to read-only again afterwards.
> > 
> > The {un}set_module_core_ro_nx() functions are used to change the
> > page permissions.  Toggling NX isn't necessary in this case, but it's
> > not highly performance sensitive code so it should be fine.
> 
> Hmm, the name (un)set_module_core_ro_nx() still sounds a bit strange,
> especially the "ro_nx" suffix.

> Alternative solution would be to create
> 
>    set_module_text_rw()
>    set_module_text_ro()
> 
> There already exists
> 
>    set_all_modules_text_rw()
>    set_all_modules_text_ro()
> 
> They modify only the ro/rw flags. IMHO, the name is more descriptive
> They are used by ftrace for very similar purpose.

That wouldn't be enough.  Relocations can occur not only in text, but
also in data.  That includes read-only data.

The (un)set_module_core_ro_nx() naming was taken from the names of
existing module functions (unset_module_{core,init}_ro_nx()).  They
enable/disable the CONFIG_DEBUG_SET_MODULE_RONX feature on the core part
of the module.  The name makes sense to me, though I'm certainly open to
other ideas.

> They modify also the init section. But we might want to touch it
> as well. klp_module_notify() is called too late now. But once we
> have a more complex consistency model, we will need to reject
> the module when the patching fails. We will need to call the
> livepatch init earlier, close to ftrace_module_init(mod).
> Then the init section might be interesting as well.

Init section functions don't have the __fentry() call, so they can't be
patched.  If that were to change in the future, we could use the
(un)set_module_init_ro_nx() functions, which already exist.

-- 
Josh