Re: x86-64: Maintain 16-byte stack alignment

[Josh Poimboeuf]

On Thu, Jan 12, 2017 at 08:46:01AM +0100, Ingo Molnar wrote:
> 
> * Herbert Xu <herbert@xxxxxxxxxxxxxxxxxxx> wrote:
> 
> > On Tue, Jan 10, 2017 at 09:05:28AM -0800, Linus Torvalds wrote:
> > >
> > > I'm pretty sure we have random asm code that may not maintain a
> > > 16-byte stack alignment when it calls other code (including, in some
> > > cases, calling C code).
> > > 
> > > So I'm not at all convinced that this is a good idea. We shouldn't
> > > expect 16-byte alignment to be something trustworthy.
> > 
> > So what if we audited all the x86 assembly code to fix this? Would
> > it then be acceptable to do a 16-byte aligned stack?
> 
> Audits for small but deadly details that isn't checked automatically by tooling 
> would inevitably bitrot again - and in this particular case there's a 50% chance 
> that a new, buggy change would test out to be 'fine' on a kernel developer's own 
> box - and break on different configs, different hw or with unrelated (and 
> innocent) kernel changes, sometime later - spreading the pain unnecessarily.
> 
> So my feeling is that we really need improved tooling for this (and yes, the GCC 
> toolchain should have handled this correctly).
> 
> But fortunately we have related tooling in the kernel: could objtool handle this? 
> My secret hope was always that objtool would grow into a kind of life insurance 
> against toolchain bogosities (which is a must for things like livepatching or a 
> DWARF unwinder - but I digress).

Are we talking about entry code, or other asm code?  Because objtool
audits *most* asm code, but entry code is way too specialized for
objtool to understand.

(I do have a pending objtool rewrite which would make it very easy to
ensure 16-byte stack alignment.  But again, objtool can only understand
callable C or asm functions, not entry code.)

Another approach would be to solve this problem with unwinder warnings,
*if* there's enough test coverage.

I recently made some changes to try to standardize the "end" of the
stack, so that the stack pointer is always a certain value before
calling into C code.  I also added some warnings to the unwinder to
ensure that it always reaches that point on the stack.  So if the "end"
of the stack were adjusted by a word by adding padding to pt_regs, the
unwinder warnings could help preserve that.

We could take that a step further by adding an unwinder check to ensure
that *every* frame is 16-byte aligned if -mpreferred-stack-boundary=3
isn't used.

Yet another step would be to add a debug feature which does stack sanity
checking from a periodic NMI, to flush out these unwinder warnings.

(Though I've found that current 0-day and fuzzing efforts, combined with
lockdep and perf's frequent unwinder usage, are already doing a great
job at flushing out unwinder warnings.)

The only question is if there would be enough test coverage,
particularly with those versions of gcc which don't have
-mpreferred-stack-boundary=3.

-- 
Josh