Re: [PATCH 0/2] Proper ro_after_init implementation on s390

[Heiko Carstens]

On Tue, Jun 07, 2016 at 11:11:17AM -0700, Kees Cook wrote:
> On Tue, Jun 7, 2016 at 11:07 AM, Heiko Carstens
> <heiko.carstens@xxxxxxxxxx> wrote:
> > On Tue, Jun 07, 2016 at 08:49:14AM -0700, Kees Cook wrote:
> >> > Heiko Carstens (2):
> >> >   vmlinux.lds.h: allow arch specific handling of ro_after_init data section
> >> >   s390/mm: add proper __ro_after_init support
> >> >
> >> >  arch/s390/include/asm/cache.h     |  3 ---
> >> >  arch/s390/include/asm/sections.h  |  1 +
> >> >  arch/s390/kernel/vmlinux.lds.S    | 12 +++++++++++-
> >> >  arch/s390/mm/init.c               |  7 ++++---
> >> >  arch/s390/mm/vmem.c               |  7 +++----
> >> >  include/asm-generic/vmlinux.lds.h | 10 +++++++++-
> >> >  6 files changed, 28 insertions(+), 12 deletions(-)
> >>
> >> Awesome! This looks great to me! Have you had a chance to look through
> >> any of the arch/s390/ __init code for variables that should be marked
> >> __ro_after_init?
> >
> > Not yet, and actually this I'm a bit reluctant to do that, since any wrong
> > annotation will lead to kernel crashes sooner or later ;)
> > However I'll look into this as well.
> 
> Yup, though the good news is it's usually discovered very quickly. :)

Eventually it might make sense to add something like
DEBUG_SECTION_MISMATCH, which would only report on _write_ accesses from
non-init sections.

Not sure if this can be done easily and without the need of a new compiler
feature. The new problem class I'm afraid of is more or less the same that
we had when non-init code referenced (already freed) initdata objects.