Received-SPF: pass (google.com: domain of linux-kernel-owner@vger.kernel.org designates 23.128.96.18 as permitted sender) client-ip=23.128.96.18;
MIME-Version: 1.0
References: <20211109164549.1724710-1-ardb@kernel.org> <20211109164549.1724710-2-ardb@kernel.org>
 <YYrAHMy8xmPXj4ql@hirez.programming.kicks-ass.net>
In-Reply-To: <YYrAHMy8xmPXj4ql@hirez.programming.kicks-ass.net>
From:   Ard Biesheuvel <ardb@kernel.org>
Date:   Tue, 9 Nov 2021 19:41:03 +0100
Message-ID: <CAMj1kXGdqDQ-WKLBcLhGNofmXpf-Xt8wAE18EBEGeTe9wNqZog@mail.gmail.com>
Subject: Re: [RFC PATCH 1/7] static_call: get rid of static_call_cond()
To:     Peter Zijlstra <peterz@infradead.org>
Cc:     Linux Kernel Mailing List <linux-kernel@vger.kernel.org>,
        Linux ARM <linux-arm-kernel@lists.infradead.org>,
        Josh Poimboeuf <jpoimboe@redhat.com>,
        Jason Baron <jbaron@akamai.com>,
        Steven Rostedt <rostedt@goodmis.org>,
        Mark Rutland <mark.rutland@arm.com>,
        Kees Cook <keescook@chromium.org>
Content-Type: text/plain; charset="UTF-8"
Precedence: bulk

On Tue, 9 Nov 2021 at 19:38, Peter Zijlstra <peterz@infradead.org> wrote:
>
> On Tue, Nov 09, 2021 at 05:45:43PM +0100, Ard Biesheuvel wrote:
> > The main reason for the existence of static_call_cond() seems to be that
> > in theory, when using the generic implementation of static calls, it may
> > be possible for a compiler to elide the indirect call entirely if the
> > target is NULL, while still guaranteeing that all side effects of
> > argument evaluation occur as expected.
> >
> > This is rather optimistic: as documented by an existing code comment,
> > both GCC and Clang (version 10) get this wrong, and even if they ever
> > get it right, this is far too subtle to rely on for a code path that is
> > expected to be used only by the 'remaining' architectures once all the
> > best supported ones implement either the out-of-line or inline optimized
> > variety of static calls.
> >
> > Given that having static_call_cond() clutters up the API, and puts the
> > burden on the caller to go and check what kind of static call they are
> > dealing with, let's just get rid of the distinction.
>
> No, static_call_cond() signifies the function can be NULL. Both gcc and
> clang generate correct (but wildly ineffecient) code for this. Without
> static_call_cond() the generic implementation will do a NULL deref.
>
> That is, static_call_cond() does properly encapuslate:
>
>         func = READ_ONCE(key.func);
>         if (func)
>                 func(ARGS);
>
> You can't take that out.

I actually address that in the patch.

AIUI, the compiler generates an indirect call to __static_call_nop(),
right? So why not simply set .func to the address of
__static_call_nop() when NULL is passed to update / the initializer?