Re: [PATCH v3] ubsan: Reintroduce signed overflow sanitizer

[Marco Elver]

On Mon, 5 Feb 2024 at 13:59, Kees Cook <keescook@xxxxxxxxxxxx> wrote:
>
> On Mon, Feb 05, 2024 at 01:54:24PM +0100, Andrey Ryabinin wrote:
> >
> >
> > On 2/5/24 10:37, Kees Cook wrote:
> >
> > > ---
> > >  include/linux/compiler_types.h |  9 ++++-
> > >  lib/Kconfig.ubsan              | 14 +++++++
> > >  lib/test_ubsan.c               | 37 ++++++++++++++++++
> > >  lib/ubsan.c                    | 68 ++++++++++++++++++++++++++++++++++
> > >  lib/ubsan.h                    |  4 ++
> > >  scripts/Makefile.lib           |  3 ++
> > >  scripts/Makefile.ubsan         |  3 ++
> > >  7 files changed, 137 insertions(+), 1 deletion(-)
> > >
> > > diff --git a/include/linux/compiler_types.h b/include/linux/compiler_types.h
> > > index 6f1ca49306d2..ee9d272008a5 100644
> > > --- a/include/linux/compiler_types.h
> > > +++ b/include/linux/compiler_types.h
> > > @@ -282,11 +282,18 @@ struct ftrace_likely_data {
> > >  #define __no_sanitize_or_inline __always_inline
> > >  #endif
> > >
> > > +/* Do not trap wrapping arithmetic within an annotated function. */
> > > +#ifdef CONFIG_UBSAN_SIGNED_WRAP
> > > +# define __signed_wrap __attribute__((no_sanitize("signed-integer-overflow")))
> > > +#else
> > > +# define __signed_wrap
> > > +#endif
> > > +
> > >  /* Section for code which can't be instrumented at all */
> > >  #define __noinstr_section(section)                                 \
> > >     noinline notrace __attribute((__section__(section)))            \
> > >     __no_kcsan __no_sanitize_address __no_profile __no_sanitize_coverage \
> > > -   __no_sanitize_memory
> > > +   __no_sanitize_memory __signed_wrap
> > >
> >
> > Given this disables all kinds of code instrumentations,
> > shouldn't we just add __no_sanitize_undefined here?
>
> Yeah, that's a very good point.
>
> > I suspect that ubsan's instrumentation usually doesn't cause problems
> > because it calls __ubsan_* functions with all heavy stuff (printk, locks etc)
> > only if code has an UB. So the answer to the question above depends on
> > whether we want to ignore UBs in "noinstr" code or to get some weird side effect,
> > possibly without proper UBSAN report in dmesg.
>
> I think my preference would be to fail safe (i.e. leave in the
> instrumentation), but the intent of noinstr is pretty clear. :P I wonder
> if, instead, we could adjust objtool to yell about cases where calls are
> made in noinstr functions (like it does for UACCESS)... maybe it already
> does?

It already does, see CONFIG_NOINSTR_VALIDATION (yes by default on x86).