Re: lib/crypto/chacha.c:65:1: warning: the frame size of 1604 bytes is larger than 1024 bytes

[Ard Biesheuvel <ardb@kernel.org>]

On Thu, 27 Aug 2020 at 10:33, Arnd Bergmann <arnd@arndb.de> wrote:
>
> On Thu, Aug 27, 2020 at 10:10 AM Ard Biesheuvel <ardb@kernel.org> wrote:
> > On Thu, 27 Aug 2020 at 10:06, Herbert Xu <herbert@gondor.apana.org.au> wrote:
> > >
> > > On Thu, Aug 27, 2020 at 11:52:50AM +0800, kernel test robot wrote:
> > > >
> > > > First bad commit (maybe != root cause):
> > > >
> > > > tree:   https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git master
> > > > head:   15bc20c6af4ceee97a1f90b43c0e386643c071b4
> > > > commit: 5fb8ef25803ef33e2eb60b626435828b937bed75 crypto: chacha - move existing library code into lib/crypto
> > > > date:   9 months ago
> > > > config: i386-randconfig-r015-20200827 (attached as .config)
> > > > compiler: gcc-9 (Debian 9.3.0-15) 9.3.0
> > > > reproduce (this is a W=1 build):
> > > >         git checkout 5fb8ef25803ef33e2eb60b626435828b937bed75
> > > >         # save the attached .config to linux build tree
> > > >         make W=1 ARCH=i386
> > > >
> > > > If you fix the issue, kindly add following tag as appropriate
> > > > Reported-by: kernel test robot <lkp@intel.com>
> > > >
> > > > All warnings (new ones prefixed by >>):
> > > >
> > > >    lib/crypto/chacha.c: In function 'chacha_permute':
> > > > >> lib/crypto/chacha.c:65:1: warning: the frame size of 1604 bytes is larger than 1024 bytes [-Wframe-larger-than=]
> > > >       65 | }
> > > >          | ^
> > >
> > > This doesn't happen with a normal configuration.  To recreate
> > > this warning, you need to enable both GCOV_KERNEL and UBSAN.
> > >
> > > This is the minimal gcc command-line to recreate it:
> > >
> > > gcc -Wframe-larger-than=1024 -fprofile-arcs -fsanitize=object-size -c -O2 chacha.c
> > >
> > > If you take away either profile-arcs or sanitize=object-size then
> > > the problem goes away.
> > >
> > > Any suggestions?
> > >
> >
> > Is it really worth it to obsess about this? Special compiler
> > instrumentation simply leads to a larger stack footprint in many
> > cases, which is why we use a larger stack to begin with (at least we
> > do so for Kasan, so if we don't for Ubsan, we should consider it)
> >
> > Past experience also shows that this is highly dependent on the exact
> > compiler version, so issues like these are often moving targets.
>
> Yes, I do think it is important to address these in some form,
> for multiple reasons:
>
> * With the limited amount of stack space in normal uninstrumented
>   kernels, I do think it is vital to have a fairly low default warning
>   limit in order to catch all cases that do something dangerously
>   stupid, either because of code bugs or compiler bugs.
>
> * I also think we do want the warning enabled in other configurations,
>   in particular because the compiler tends to make increasingly stupid
>   decisions when combining less common instrumentations, which
>   again can lead to instant exploitable bugs, e.g. when a function's
>   stack frame grows beyond the total stack size. In many cases the
>   gcc and clang developers are both able to address these quickly
>   when we send a good bug report (which unfortunately can be a lot of
>   work).
>
> * The crypto cipher code unfortunately is particularly prone to running
>   into these issues because each new compiler version tries to
>   find more clever tricks to optimize code that in turn implements
>   an algorithm that intentionally tries to not have any clever shortcuts.
>   In many cases the stack size warning that we see in some
>   configurations is an indicator for the compiler running into a false
>   optimization even on the non-instrumented code that leads to lower
>   performance from unnecessary register spilling that should be
>   avoided.
>

In that case, I suppose we should simply disable instrumentation for
chacha_permute()? It is a straight-forward arithmetic transformation
on a u32[16] array, where ubsan has limited value afaict.