Re: modversions: redefine kcrctab entries as 32-bit values

[Jessica Yu]

+++ Jessica Yu [02/02/17 22:54 -0500]:
> +++ Ard Biesheuvel [24/01/17 16:16 +0000]:
> > This v4 is a followup to [0] 'modversions: redefine kcrctab entries as
> > relative CRC pointers', but since relative CRC pointers do not work in
> > modules, and are actually only needed by powerpc with CONFIG_RELOCATABLE=y,
> > I have made it a Kconfig selectable feature instead.
> >
> > Patch #1 introduces the MODULE_REL_CRCS Kconfig symbol, and adds the kbuild
> > handling of it, i.e., modpost, genksyms and kallsyms.
> >
> > Patch #2 switches all architectures to 32-bit CRC entries in kcrctab, where
> > all architectures except powerpc with CONFIG_RELOCATABLE=y use absolute ELF
> > symbol references as before.
> >
> > v4: make relative CRCs kconfig selectable
> >   use absolute CRC symbols in modules regardless of kconfig selection
> >   split into two patches
>
> This asymmetry threw me off a bit, especially the Kconfig naming (only
> vmlinux crcs get the relative offsets, and only on powerpc atm, but all
> modules keep the absolute syms, but it is called MODULE_REL_CRCS...),
> if we keep this asymmetric crc treatment, it would be really nice to
> note this discrepancy this somewhere, perhaps in the Kconfig, to keep
> our heads from spinning :-)
>
> I'm still catching up on the previous discussion threads, but can you
> explain a bit more why you switched away from full blown relative crcs
> from your last patchset [1]? I had lightly tested your v3 on ppc64le
> previously, and relative offsets with modules seemed to worked very
> well. I'm probably missing something very obvious.

Ah, I just saw your other comment about other arches not having support
for the rel32 offsets :-/

The asymmetry still bothers me though. Can't we have something that just
switches relative crcs on and off, that applies to *both* vmlinux and modules?
Then we can get rid of the crc_owner check in check_version() and just have
something like:

  if (IS_ENABLED(CONFIG_RELATIVE_CRCS))
      crcval = resolve_crc(crc);

Also we could get rid of the '&& !defined(MODULE)' checks scattered in
export.h. Then the arches that want relative crcs and that *do* have rel32
relocation support can turn relative crcs on, and powerpc can enable it, right?

Would that work or is there another reason this won't work with modules
(assuming that the arches that select this option support the relative
offsets)?

Jessica