MIME-Version: 1.0
References: <20230207182135.2671106-1-revest@chromium.org> <20230207182135.2671106-5-revest@chromium.org>
 <20230315194334.58eb56ab@gandalf.local.home>
In-Reply-To: <20230315194334.58eb56ab@gandalf.local.home>
From:   Florent Revest <revest@chromium.org>
Date:   Thu, 16 Mar 2023 16:40:48 +0100
Message-ID: <CABRcYmLLbRGZXWwEpyLW1YFz87tTPA8pCL7oLd4K6Hp9Etr5LA@mail.gmail.com>
Subject: Re: [PATCH v2 04/10] ftrace: Store direct called addresses in their ops
To:     Steven Rostedt <rostedt@goodmis.org>
Cc:     linux-arm-kernel@lists.infradead.org, linux-kernel@vger.kernel.org,
        linux-trace-kernel@vger.kernel.org, bpf@vger.kernel.org,
        catalin.marinas@arm.com, will@kernel.org, mhiramat@kernel.org,
        mark.rutland@arm.com, ast@kernel.org, daniel@iogearbox.net,
        andrii@kernel.org, kpsingh@kernel.org, jolsa@kernel.org,
        xukuohai@huaweicloud.com, lihuafei1@huawei.com
Content-Type: text/plain; charset="UTF-8"
Content-Transfer-Encoding: quoted-printable
Precedence: bulk

On Thu, Mar 16, 2023 at 12:43=E2=80=AFAM Steven Rostedt <rostedt@goodmis.or=
g> wrote:
>
> On Tue,  7 Feb 2023 19:21:29 +0100
> Florent Revest <revest@chromium.org> wrote:
>
> > @@ -5445,6 +5445,7 @@ __modify_ftrace_direct(struct ftrace_ops *ops, un=
signed long addr)
> >       /* Enable the tmp_ops to have the same functions as the direct op=
s */
> >       ftrace_ops_init(&tmp_ops);
> >       tmp_ops.func_hash =3D ops->func_hash;
> > +     tmp_ops.direct_call =3D addr;
> >
> >       err =3D register_ftrace_function_nolock(&tmp_ops);
> >       if (err)
> > @@ -5466,6 +5467,7 @@ __modify_ftrace_direct(struct ftrace_ops *ops, un=
signed long addr)
> >                       entry->direct =3D addr;
> >               }
> >       }
> > +     WRITE_ONCE(ops->direct_call, addr);
>
> I'm curious about the use of WRITE_ONCE(). It should not go outside the
> mutex barrier.

This WRITE_ONCE was originally suggested by Mark here:
https://lore.kernel.org/all/Y9vW99htjOphDXqY@FVFF77S0Q05N.cambridge.arm.com=
/#t

My understanding is that it's not so much about avoiding re-ordering
but rather about avoiding store tearing since a ftrace_caller
trampoline could concurrently read ops->direct_call. Does that make
sense ?

> -- Steve
>
> >
> >       mutex_unlock(&ftrace_lock);
> >