Date:   Fri, 10 Mar 2023 21:59:26 +0100
From:   Peter Zijlstra <peterz@infradead.org>
To:     Josh Poimboeuf <jpoimboe@kernel.org>
Cc:     x86@kernel.org, linux-kernel@vger.kernel.org,
        Mark Rutland <mark.rutland@arm.com>,
        Jason Baron <jbaron@akamai.com>,
        Steven Rostedt <rostedt@goodmis.org>,
        Ard Biesheuvel <ardb@kernel.org>,
        Christophe Leroy <christophe.leroy@csgroup.eu>,
        Paolo Bonzini <pbonzini@redhat.com>,
        Sean Christopherson <seanjc@google.com>
Subject: Re: [RFC][PATCH 1/5] static_call: Make NULL static calls consistent
Message-ID: <20230310205926.GB1605437@hirez.programming.kicks-ass.net>
References: <cover.1678474914.git.jpoimboe@kernel.org>
 <016c1e9cbdf726a885a406ff6baed85087ad1213.1678474914.git.jpoimboe@kernel.org>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <016c1e9cbdf726a885a406ff6baed85087ad1213.1678474914.git.jpoimboe@kernel.org>
Precedence: bulk

On Fri, Mar 10, 2023 at 12:31:13PM -0800, Josh Poimboeuf wrote:

> -/*
> - * This horrific hack takes care of two things:
> - *
> - *  - it ensures the compiler will only load the function pointer ONCE,
> - *    which avoids a reload race.
> - *
> - *  - it ensures the argument evaluation is unconditional, similar
> - *    to the HAVE_STATIC_CALL variant.
> - *
> - * Sadly current GCC/Clang (10 for both) do not optimize this properly
> - * and will emit an indirect call for the NULL case :-(
> - */
> -#define __static_call_cond(name)					\
> -({									\
> -	void *func = READ_ONCE(STATIC_CALL_KEY(name).func);		\
> -	if (!func)							\
> -		func = &__static_call_nop;				\
> -	(typeof(STATIC_CALL_TRAMP(name))*)func;				\
> -})

So a sufficiently clever compiler can optimize the above to avoid the
actual indirect call (and resulting CFI violation, see below), because
__static_call_nop() is inline and hence visible as an empty stub
function. Currently none of the compilers are that clever :/

> -#define static_call_cond(name)	(void)__static_call_cond(name)
> +#define static_call_cond(name) (void)static_call(name)
>  
>  static inline
>  void __static_call_update(struct static_call_key *key, void *tramp, void *func)
>  {
> -	WRITE_ONCE(key->func, func);
> +	WRITE_ONCE(key->func, func ? : (void *)__static_call_nop);
>  }

This will break ARM64 I think, they don't HAVE_STATIC_CALL but do have
CLANG_CFI, which means the above will end up being a runtime indirect
call to a non-matching signature function.

Now, I suppose we don't actually have this happen in current code by the
simple expedient of not actually having any static_call_cond() usage
outside of arch code.

(/me git-grep's some and *arrrggh* trusted-keys)

I really don't think we can do this though, must not promote CFI
violations.