Re: [PATCH] Revert "x86/module: Detect and skip invalid relocations"

[Josh Poimboeuf]

On Mon, Jun 24, 2019 at 12:00:33PM +0200, Miroslav Benes wrote:
> On Sat, 22 Jun 2019, Thomas Gleixner wrote:
> 
> > Miroslav,
> > 
> > On Thu, 20 Jun 2019, Miroslav Benes wrote:
> > > On Thu, 20 Jun 2019, Cheng Jian wrote:
> > > 
> > > > This reverts commit eda9cec4c9a12208a6f69fbe68f72a6311d50032.
> > > > 
> > > > Since commit (eda9cec4c9a1 'x86/module: Detect and skip invalid
> > > > relocations') add some sanity check in apply_relocate_add, borke
> > > > re-insmod a kernel module which has been patched before,
> > > > 
> > > > The relocation informations of the livepatch module have been
> > > > overwritten since first patched, so if we rmmod and insmod the
> > > > kernel module, these values are not zero anymore, when
> > > > klp_module_coming doing, and that commit marks them as invalid
> > > > invalid_relocation.
> > > > 
> > > > Then the following error occurs:
> > > > 
> > > > 	module: x86/modules: Skipping invalid relocation target, existing value is nonzero for type 2, loc (____ptrval____), val ffffffffc000236c
> > > > 	livepatch: failed to initialize patch 'livepatch_0001_test' for module 'test' (-8)
> > > > 	livepatch: patch 'livepatch_0001_test' failed for module 'test', refusing to load module 'test'
> > > 
> > > Oh yeah. First reported here 20180602161151.apuhs2dygsexmcg2@treble (LP ML 
> > > only and there is no archive on lore.kernel.org yet. Sorry about that.). I 
> > > posted v1 here 
> > > https://lore.kernel.org/lkml/20180607092949.1706-1-mbenes@xxxxxxx/ and 
> > > even started to work on v2 in March with arch-specific nullifying, but 
> > > then I got sidetracked again. I'll move it up my todo list a bit.
> > 
> > so we need to revert it for now, right?
> 
> Not necessarily.
> 
> Quoting Josh from the original bug report:
> "Possible ways to fix it:
> 
> 1) Remove the error check in apply_relocate_add().  I don't think we
>    should do this, because the error is actually useful for detecting
>    corrupt modules.  And also, powerpc has the similar error so this
>    wouldn't be a universal solution.
> 
> 2) In klp_unpatch_object(), call an arch-specific arch_unpatch_object()
>    which reverses any arch-specific patching: on x86, clearing all
>    relocation targets to zero; on powerpc, converting the instructions
>    after relative link branches to nops.  I don't think we should do
>    this because it's not a global solution and requires fidgety
>    arch-specific patching code.
> 
> 3) Don't allow patched modules to be removed.  I think this makes the
>    most sense.  Nobody needs this functionality anyway (right?).
> "
> 
> 1 would be the revert. We decided against it. The scenario (rmmod a 
> module) is (supposedly) not that common in practice. Even the current bug 
> report was triggered just in testing if I am not mistaken. Moreover, you 
> need kpatch-build to properly set up relocation records. Upstream 
> livepatch does not offer it as of now. That's why (I think) Josh thought 
> the benefits of the check outweighed the disadvantage.
> 
> Then I tried to implement 3, but there were problems with it too. 2 
> remains to be finished and then we can decide what the best approach is.
> 
> That being said... I am not against the reverting the commit per se, but 
> we lived with it or quite a long time and no one has met it so far in 
> "real life". I don't think it is the classic "we broke something, we have 
> to revert" scenario.
> 
> Josh, any comment? I think your opinion matters here much more than mine.

Agreed, as far as I know the problem is purely theoretical and we
haven't seen any real-world bug reports, because people aren't reloading
patched modules in the real world.

If we were to revert the error checks in apply_relocate_add() then it
could expose us to real-world regressions (which we have actually seen
in the past).

So I would vote to leave the error checks in place, at least until it
becomes a real-world issue.  And in the meantime hopefully you can
finish implementing #2 or #3 soon :-)

-- 
Josh