Re: unable to load modules with CONFIG_MODVERSIONS=y after commit 8ab2ae655b

[Arnd Bergmann]

On Tuesday, December 6, 2016 8:39:53 AM CET Linus Torvalds wrote:
> On Tue, Dec 6, 2016 at 5:44 AM, Adam Borowski <kilobyte@xxxxxxxxxx> wrote:
> > On Tue, Dec 06, 2016 at 08:31:01AM -0500, Jan Stancek wrote:
> >> Starting with 4.9-rc8 / commit 8ab2ae655b ("default exported asm symbols to zero")
> >> I'm running into issue with kernel built with CONFIG_MODVERSIONS=y
> >> and (older) binutils (binutils-2.25.1-20.base.el7.ppc64le).
> >>
> >> Modules fail to load, for example:
> >>
> >> [    3.163646] Found checksum 0 vs module 4829A47E
> >> [    3.163787] dm_mod: disagrees about version of symbol memcpy
> >> [    3.163862] dm_mod: Unknown symbol memcpy (err -22)
> >>
> >> Bisect led me to 8ab2ae655b, reverting it allows boot to
> >> progress as before.
> >
> > powerpc happens to be the only arch that actually followed the plan and
> > implemented asm-prototypes.h (not including Debian which applied my patch to
> > do so on x86, that patch is not in mainline).
> >
> > Could you try reverting commits that add exports to that file?
> 
> Let's not do this. Let's just assume that "checksum=0" matches anything.
> 
> Because it's too late to play games with this any more, and it's too
> damn fragile.
> 
> In fact, I think I'll just revert Arnd's patch, and just rely on
> commit faaae2a58143 ("Re-enable CONFIG_MODVERSIONS in a slightly
> weaker form") for 4.9.
> 
> Because Arnd's patch did explain what was going on, but it also broke
> alpha, due to ".set" apparently meaning something else than "set
> value" there. So I like Arnd's patch even if it gets reverted, just
> because of the "explain things" part.
> 
> Or would people prefer just an explicit "zero in the crc tables means
> that it never got filled in correctly, so let's match it"?
> 
> Arnd, comments?

I'm not overly attached to my patch, and I intentionally marked
it RFC as I had done very little testing on it and I'm not surprised
it causes problems. Reverting it for v4.9 seems best if that
gets us a working kernel with modversions again on all
architectures.

In the long run, I'd still hope to to resolve those issues:

- without my patch, Nick's patch d8c1eb86e952 ("kbuild: modpost
  warn if export version crc is missing") in linux-next produces
  warnings for most architectures with CONFIG_MODVERSIONS=y
  and binutils-2.27+, and that means we don't get a clean
  allmodconfig build any more.

- I have been sick the past few days and haven't had a chance
  to investigate exactly what happens that prevents the modules
  from getting loaded. I assume the problem that Jan sees
  is the same that Ben Hutchings already reported on Saturday.
  This might be just another trivial fix, or something more
  fundamental, but finding this out would be helpful in case
  we want to bring it back.

- If we decide to keep the warning about broken toolchains
  in the long run and not also make sure that all symbols
  have a nonzero crc, then we should ask Alan Modra to revert
  the linker change.

	Arnd