Re: [RFC PATCH 08/11] x86: Add support for finer grained KASLR

[Peter Zijlstra]

On Thu, Feb 06, 2020 at 04:06:17AM -0800, Kees Cook wrote:
> On Thu, Feb 06, 2020 at 11:38:30AM +0100, Peter Zijlstra wrote:
> > On Wed, Feb 05, 2020 at 02:39:47PM -0800, Kristen Carlson Accardi wrote:
> > > +static long __start___ex_table_addr;
> > > +static long __stop___ex_table_addr;
> > > +static long _stext;
> > > +static long _etext;
> > > +static long _sinittext;
> > > +static long _einittext;
> > 
> > Should you not also adjust __jump_table, __mcount_loc,
> > __kprobe_blacklist and possibly others that include text addresses?
> 
> These don't appear to be sorted at build time. 

The ORC tables are though:

  57fa18994285 ("scripts/sorttable: Implement build-time ORC unwind table sorting")

> AIUI, the problem with
> ex_table and kallsyms is that they're preprocessed at build time and
> opaque to the linker's relocation generation.

I was under the impression these tables no longer had relocation data;
that since they're part of the main kernel, the final link stage could
completely resolve them.

That said, I now see we actually have .rela__extable .rela.orc_unwind_ip
etc.

> For example, looking at __jump_table, it gets sorted at runtime:
> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/kernel/jump_label.c#n474

For now, yes. Depends a bit on how hard people are pushing on getting
jump_labels working earlier and ealier in boot.

> As you're likely aware, we have a number of "special"
> sections like this, currently collected manually, see *_TEXT:
> https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git/tree/arch/x86/kernel/vmlinux.lds.S#n128

Right.

> I think we can actually add (most of) these to fg-kaslr's awareness (at
> which point their order will be shuffled respective to other sections,
> but with their content order unchanged), but it'll require a bit of
> linker work. I'll mention this series's dependency on the linker's
> orphaned section handling in another thread...

I have some patches pending where we rely on link script order. That's
data sections though, so I suppose that's safe for the moment.