Re: [PATCH 0/2] Fix issue with alternatives/paravirt patches

[Josh Poimboeuf]

On Thu, Jul 07, 2016 at 05:56:33PM +0200, Petr Mladek wrote:
> On Tue 2016-07-05 22:34:58, Jessica Yu wrote:
> > Hi,
> > 
> > A few months ago, Chris Arges reported a bug involving alternatives/paravirt
> > patching that was discussed here [1] and here [2]. To briefly summarize the
> > bug, patch modules that contained .altinstructions or .parainstructions
> > sections would break because these alternative/paravirt patches would be
> > applied first by the module loader (see x86 module_finalize()), then
> > livepatch would later clobber these patches when applying per-object
> > relocations. This lead to crashes and unpredictable behavior.
> > 
> > One conclusion we reached from our last discussion was that we will
> > need to introduce some arch-specific code to address this problem.
> > This patchset presents a possible fix for the bug by adding a new
> > arch-specific arch_klp_init_object_loaded() function that by default
> > does nothing but can be overridden by different arches.
> > 
> > To fix this issue for x86, since we can access a patch module's Elf
> > sections through mod->klp_info, we can simply delay the calls to
> > apply_paravirt() and apply_alternatives() to arch_klp_init_object_loaded(),
> > which is called after relocations have been written for an object.
> > In addition, for patch modules, .parainstructions and .altinstructions are
> > prefixed by ".klp.arch.${objname}" so that the module loader ignores them
> > and livepatch can apply them manually.
> 
> The solution looks correct to me. The fun will be how to generate
> the sections. If I get this correctly, it is not enough to rename
> the existing ones. Instead, we need to split .parainstructions
> and .altinstructions sections into per-object ones.
> 
> I wonder if there is a plan for this. Especially I am interested
> into the patches created from sources ;-) I wonder if we could add
> a tag somewhere and improve the build infrastructure.

Yeah.  I'd like to reiterate[1] that this would all be a lot easier if
we weren't circumventing module dependencies.

[1] https://lkml.kernel.org/r/20160404161428.3qap2i4vpgda66iw@xxxxxxxxxxxxxxxxx

-- 
Josh