On Thu, Feb 26, 2015 at 4:11 PM, Kees Cook [off-list ref] wrote:

On Thu, Feb 26, 2015 at 4:06 PM, Andrew Morton
[off-list ref] wrote:

quoted

On Thu, 26 Feb 2015 15:37:37 -0800 Kees Cook [off-list ref] wrote:

quoted

Agh, no, please let's avoid the CONFIG addition.

That is precisely how we do this.

quoted

Hector mentioned in private mail that he was looking at an alternative
that adds exec_base to struct mm which would avoid all this insanity.

Can't we do something like:

#ifndef mmap_rnd
# define mmap_rnd 0
#endif

Sure, and sprinkle

#define mmap_rnd mmap_rnd

in five arch header files where nobody thinks to look.

For better or for worse, we are consolidating such things into arch/*/Kconfig.

Okay, fair enough. Even with your configs (though shouldn't they be
ARCH_HAS or just HAVE?) I've now stumbled over the issue that we can't
put randomize_et_dyn in binfmt_elf because it conflicts with linking
against compat_binfmt_elf.

Instead of all this, how about we rework the existing CONFIG and just
change around how s390 does this to match the other architectures and
remove the ifdef in binfmt_elf.c at the same time? Let me work
something up...

-Kees

-- 
Kees Cook
Chrome OS Security