Re: [PATCH -tip] kasan: Emit different calls for instrumentable memintrinsics

From: Marco Elver <elver@google.com>
To: Jakub Jelinek <jakub@redhat.com>
Cc: Peter Zijlstra <peterz@infradead.org>,
	Masahiro Yamada <masahiroy@kernel.org>,
	Nathan Chancellor <nathan@kernel.org>,
	Nick Desaulniers <ndesaulniers@google.com>,
	Nicolas Schier <nicolas@fjasle.eu>,
	Andrey Ryabinin <ryabinin.a.a@gmail.com>,
	Alexander Potapenko <glider@google.com>,
	Andrey Konovalov <andreyknvl@gmail.com>,
	Dmitry Vyukov <dvyukov@google.com>,
	Vincenzo Frascino <vincenzo.frascino@arm.com>,
	linux-kbuild@vger.kernel.org, kasan-dev@googlegroups.com,
	linux-kernel@vger.kernel.org, Ingo Molnar <mingo@kernel.org>,
	Tony Lindgren <tony@atomide.com>,
	Ulf Hansson <ulf.hansson@linaro.org>,
	linux-toolchains@vger.kernel.org
Subject: Re: [PATCH -tip] kasan: Emit different calls for instrumentable memintrinsics
Date: Fri, 10 Feb 2023 21:07:14 +0100	[thread overview]
Message-ID: <CANpmjNO3w9h=QLQ9NRf0QZoR86S7aqJrnAEQ3i2L0L3axALzmw@mail.gmail.com> (raw)
In-Reply-To: <Y+aaDP32wrsd8GZq@tucnak>

On Fri, 10 Feb 2023 at 20:25, Jakub Jelinek <jakub@redhat.com> wrote:
>
> On Wed, Feb 08, 2023 at 07:42:03PM +0100, Marco Elver wrote:
> > Clang 15 will provide an option to prefix calls to memcpy/memset/memmove
> > with __asan_ in instrumented functions: https://reviews.llvm.org/D122724
> >
> > GCC does not yet have similar support.
>
> GCC has support to rename memcpy/memset etc. for years, say on
> following compiled with
> -fsanitize=kernel-address -O2 -mstringop-strategy=libcall
> (the last option just to make sure the compiler doesn't prefer to emit
> rep mov*/stos* or loop or something similar, of course kernel can keep
> whatever it uses) you'll get just __asan_memcpy/__asan_memset calls,
> no memcpy/memset, while without -fsanitize=kernel-address you get
> normally memcpy/memset.

> Or do you need the __asan_* functions only in asan instrumented functions
> and normal ones in non-instrumented functions in the same TU?

Yes, exactly that: __asan_ in instrumented, and normal ones in
no_sanitize functions; they can be mixed in the same TU. We can't
rename normal mem*() functions everywhere. In no_sanitize functions
(in particular noinstr), normal mem*() should be used. But in
instrumented code, it should be __asan_mem*(). Another longer
explanation I also just replied here:
https://lore.kernel.org/all/CANpmjNNH-O+38U6zRWJUCU-eJTfMhUosy==GWEOn1vcu=J2dcw@mail.gmail.com/

At least clang has had this behaviour for user space ASan forever:
https://godbolt.org/z/h5sWExzef - so it was easy to just add the flag
to make it behave like in user space for mem*() in the kernel. It
might also be worthwhile for GCC to emit __asan_ for user space, given
that the runtimes are shared and the user space runtime definitely has
__asan_. The kernel needs the param (asan-kernel-mem-intrinsic-prefix)
though, to not break older kernels.

Thanks,
-- Marco