[PATCH 2/2] ARM: mm: keep rodata non-executable

[Jon Medhurst (Tixy)]

On Sun, 2014-03-23 at 16:21 -0600, Kees Cook wrote:
> On Sun, Mar 23, 2014 at 12:47 PM, Laura Abbott <lauraa at codeaurora.org> wrote:
> > On 2/17/2014 4:34 AM, Dave Martin wrote:
> >> On Fri, Feb 14, 2014 at 11:11:07AM -0800, Kees Cook wrote:
> >>> On Fri, Feb 14, 2014 at 8:22 AM, Dave Martin <Dave.Martin at arm.com> wrote:
> >>>> On Thu, Feb 13, 2014 at 05:04:10PM -0800, Kees Cook wrote:
> >>>>> Introduce "CONFIG_DEBUG_RODATA" to mostly match the x86 config, though
> >>>>> the behavior is different: it depends on STRICT_KERNMEM_PERMS, which
> >>>>> sets rodata read-only (but executable), where as this option additionally
> >>>>> splits rodata from the kernel text (resulting in potentially more memory
> >>>>> lost to padding) and sets it non-executable as well. The end result is
> >>>>> that on builds with CONFIG_DEBUG_RODATA=y (like x86) the rodata with be
> >>>>> marked purely read-only.
> >>>>
> >>>> This triggers an Oops in kexec, because we have a block of code in .text
> >>>> which is a template for generating baremetal code to relocate the new
> >>>> kernel, and some literal words are written into it before copying.
> >>>
> >>> You're writing into the text area? I would imagine that
> >>> CONFIG_ARM_KERNMEM_PERMS would break that. However, that's not the
> >>> right place to be building code -- shouldn't the module area be used
> >>> for that?
> >>>
> >>>> Possibly this should be in .rodata, not .text.
> >>>
> >>> Well, rodata should be neither writable nor executable.
> >>
> >> We're not writing into code exactly.
> >>
> >> This code is never executed in-place in vmlinux.  It gets copied, and
> >> only copies are ever executed.
> >>
> >> Some pointers and offsets get poked into the code to configure it.
> >>
> >> I think it would be better simply to put the code in .rodata, and
> >> poke paramaters into the copy, not the original -- but that's a bit
> >> more awkward to code up, since the values can't be poked simply by
> >> writing global variables.
> >>
> >>>
> >>>> There may be a few other instances of this kind of thing.
> >>>
> >>> This config will certainly find them! :) But, that's why it's behind a config.
> >>
> >> I haven't tested exhaustively, but it think this is sufficient for a
> >> Tested-by.  The patch does seem to be doing what it is intended to
> >> do, and doesn't seem to be triggering false positives all over the
> >> place.
> >>
> >>>
> >>>> Are you aware of similar situations on other arches?
> >>>
> >>> I think there were some problems a long time ago on x86 for rodata too.
> >>
> >> It would be good to get this kexec case fixed -- I'll try to hack up
> >> a separate patch.
> >>
> >
> > FWIW, we've hit issues not just with kexec but kprobes as well. The same
> > problems exist with this series:
> 
> For this stage, how about I make this "depends on KEXEC=n &&
> KPROBES=n"? 

There's also ftrace (CONFIG_DYNAMIC_FTRACE I believe) which modifies
kernel code with a call to probe_kernel_write(), which GDB uses as well.
 
And grepping for the patch_text() function also shows
__arch_jump_label_transform() modifies kernel code. Not sure how and
when that gets used.

-- 
Tixy