Re: [PATCH -rcu/kcsan 23/23] objtool, kcsan: Remove memory barrier instrumentation from noinstr

[Marco Elver <elver@google.com>]

On Tue, 5 Oct 2021 at 17:13, Marco Elver <elver@google.com> wrote:
> On Tue, Oct 05, 2021 at 04:37PM +0200, Peter Zijlstra wrote:
> > On Tue, Oct 05, 2021 at 12:59:05PM +0200, Marco Elver wrote:
> > > Teach objtool to turn instrumentation required for memory barrier
> > > modeling into nops in noinstr text.
> > >
> > > The __tsan_func_entry/exit calls are still emitted by compilers even
> > > with the __no_sanitize_thread attribute. The memory barrier
> > > instrumentation will be inserted explicitly (without compiler help), and
> > > thus needs to also explicitly be removed.
> >
> > How is arm64 and others using kernel/entry + noinstr going to fix this?
> >
> > ISTR they fully rely on the compilers not emitting instrumentation,
> > since they don't have objtool to fix up stray issues like this.
>
> So this is where I'd like to hear if the approach of:
>
>  | #if !defined(CONFIG_ARCH_WANTS_NO_INSTR) || defined(CONFIG_STACK_VALIDATION)
>  | ...
>  | #else
>  | #define kcsan_noinstr noinstr
>  | static __always_inline bool within_noinstr(unsigned long ip)
>  | {
>  |      return (unsigned long)__noinstr_text_start <= ip &&
>  |             ip < (unsigned long)__noinstr_text_end;
>  | }
>  | #endif
>
> and then (using the !STACK_VALIDATION definitions)
>
>  | kcsan_noinstr void instrumentation_may_appear_in_noinstr(void)
>  | {
>  |      if (within_noinstr(_RET_IP_))
>  |              return;
>
> works for the non-x86 arches that select ARCH_WANTS_NO_INSTR.
>
> If it doesn't I can easily just remove kcsan_noinstr/within_noinstr, and
> add a "depends on !ARCH_WANTS_NO_INSTR || STACK_VALIDATION" to the
> KCSAN_WEAK_MEMORY option.
>
> Looking at a previous discussion [1], however, I was under the
> impression that this would work.
>
> [1] https://lkml.kernel.org/r/CANpmjNMAZiW-Er=2QDgGP+_3hg1LOvPYcbfGSPMv=aR6MVTB-g@mail.gmail.com

I'll send v2 of this series after 5.16-rc1. So far I think we haven't
been able to say the above doesn't work, which means I'll assume it
works on non-x86 architectures with ARCH_WANTS_NO_INSTR until we get
evidence of the opposite.