Re: [PATCH 05/15] x86: Implement function_nocfi

From: Sami Tolvanen <samitolvanen@google.com>
To: Thomas Gleixner <tglx@linutronix.de>
Cc: Kees Cook <keescook@chromium.org>,
	Andy Lutomirski <luto@kernel.org>, Borislav Petkov <bp@alien8.de>,
	X86 ML <x86@kernel.org>, Josh Poimboeuf <jpoimboe@redhat.com>,
	Peter Zijlstra <peterz@infradead.org>,
	Nathan Chancellor <nathan@kernel.org>,
	Nick Desaulniers <ndesaulniers@google.com>,
	Sedat Dilek <sedat.dilek@gmail.com>,
	linux-hardening@vger.kernel.org,
	LKML <linux-kernel@vger.kernel.org>,
	clang-built-linux <clang-built-linux@googlegroups.com>
Subject: Re: [PATCH 05/15] x86: Implement function_nocfi
Date: Mon, 19 Apr 2021 08:13:47 -0700	[thread overview]
Message-ID: <CABCJKueKDX+6DJnoVns1K35B9htRXRHLdYUckJ25dDhbWUezFw@mail.gmail.com> (raw)
In-Reply-To: <875z0ltdv8.ffs@nanos.tec.linutronix.de>

On Sat, Apr 17, 2021 at 3:16 AM Thomas Gleixner <tglx@linutronix.de> wrote:
>
> On Sat, Apr 17 2021 at 01:02, Thomas Gleixner wrote:
> > On Fri, Apr 16 2021 at 15:37, Kees Cook wrote:
> >
> >> On Fri, Apr 16, 2021 at 03:20:17PM -0700, Andy Lutomirski wrote:
> >>> But obviously there is code that needs real function pointers.  How
> >>> about making this a first-class feature, or at least hacking around it
> >>> more cleanly.  For example, what does this do:
> >>>
> >>> char entry_whatever[];
> >>> wrmsrl(..., (unsigned long)entry_whatever);
> >>
> >> This is just casting. It'll still resolve to the jump table entry.
> >>
> >>> or, alternatively,
> >>>
> >>> extern void func() __attribute__((nocfi));
> >>
> >> __nocfi says func() should not perform checking of correct jump table
> >> membership for indirect calls.
> >>
> >> But we don't want a global marking for a function to be ignored by CFI;
> >> we don't want functions to escape CFI -- we want specific _users_ to
> >> either not check CFI for indirect calls (__nocfi) or we want specific
> >> passed addresses to avoid going through the jump table
> >> (function_nocfi()).
> >
> > And that's why you mark entire files to be exempt without any rationale
> > why it makes sense.
>
> The reason why you have to do that is because function_nocfi() is not
> provided by the compiler.
>
> So you need to hack around that with that macro which fails to work
> e.g. for the idt data arrays.
>
> Is there any fundamental reason why the compiler does not provide that
> in a form which allows to use it everywhere?

I'm not aware of a fundamental reason why the compiler couldn't
provide a built-in here. This series attempts to work with what's
available at the moment, and admittedly that's not quite ideal on x86.

> It's not too much asked from a tool which provides new functionality to
> provide it in a way which is usable.

Sure, that's reasonable. I'll talk to our compiler folks and see how
we can proceed here.

Sami