Re: [PATCH v4 15/16] module: Move where we mark modules RO,X

[Peter Zijlstra]

On Fri, Oct 25, 2019 at 08:44:56AM +0200, Petr Mladek wrote:
> On Thu 2019-10-24 15:16:34, Peter Zijlstra wrote:
> > On Wed, Oct 23, 2019 at 12:00:25PM -0500, Josh Poimboeuf wrote:
> > 
> > > > This then raises a number of questions:
> > > > 
> > > >  1) why is that RELA (that obviously does not depend on any module)
> > > >     applied so late?
> > > 
> > > Good question.  The 'pv_ops' symbol is exported by the core kernel, so I
> > > can't see any reason why we'd need to apply that rela late.  In theory,
> > > kpatch-build isn't supposed to convert that to a klp rela.  Maybe
> > > something went wrong in the patch creation code.
> > > 
> > > I'm also questioning why we even need to apply the parainstructions
> > > section late.  Maybe we can remove that apply_paravirt() call
> > > altogether, along with .klp.arch.parainstruction sections.
> 
> Hmm, the original bug report against livepatching was actually about
> paravirt ops, see below.

Yes, I found that.

> > > I'm not sure about alternatives, but maybe we can enforce such
> > > limitations with tooling and/or kernel checks.
> > 
> > Right, so on IRC you implied you might have some additional details on
> > how alternatives were affected; did you manage to dig that up?
> 
> I am not sure what Josh had in mind. But the problem with livepatches,
> paravort ops, and alternatives was described in the related patchset, see
> https://lkml.kernel.org/r/1471481911-5003-1-git-send-email-jeyu@xxxxxxxxxx

Yes, and my complaint there is that that thread is void of useful
content.

> The original bug report is
> https://lkml.kernel.org/r/20160329120518.GA21252@xxxxxxxxxxxxx

I found the github (*groan*) link in the thread above.

>From all that I could only make that the paravirt stuff is just doing it
wrong (see earlier emails, core-kernel RELAs really should be applied at
the time of patch-module load, there's no excuse for them to be delayed
to the .klp.rela. section) at which point paravirt will also magically
work.

But none of that explains why apply_alternatives() is also delayed.

So I'm very tempted to just revert that patchset for doing it all
wrong.