Received-SPF: pass (google.com: domain of linux-kernel+bounces-96154-linux.lists.archive=gmail.com@vger.kernel.org designates 147.75.48.161 as permitted sender) client-ip=147.75.48.161;
Precedence: bulk
MIME-Version: 1.0
References: <CAHk-=wgG6Dmt1JTXDbrbXh_6s2yLjL=9pHo7uv0==LHFD+aBtg@mail.gmail.com>
 <20240306135504.2b3872ef@gandalf.local.home> <CAHk-=wjbDgMKLgxbV+yK4LKZ+2Qj6zVL_sHeb+L9KDia980Q8Q@mail.gmail.com>
 <20240306142738.7b66a716@rorschach.local.home> <CAHk-=wgPAZ4KnCQergqAOUypwinYh=gZ0q4EQbwvuUcJ_8UK+Q@mail.gmail.com>
 <83b47424-e5e0-46de-aa63-d413a5aa6cec@paulmck-laptop> <CAHk-=wiX_zF5Mpt8kUm_LFQpYY-mshrXJPOe+wKNwiVhEUcU9g@mail.gmail.com>
 <851dc594-d2ea-4050-b7c6-e33a1cddf3a1@efficios.com> <72b14322-78c1-4479-9c4e-b0e11c1f0d53@paulmck-laptop>
 <bebbed4a-ced1-42c5-865c-dc9dc7857b6c@efficios.com> <c1bb35c4-29af-4a84-8ba7-81ba30639a69@paulmck-laptop>
In-Reply-To: <c1bb35c4-29af-4a84-8ba7-81ba30639a69@paulmck-laptop>
From: Linus Torvalds <torvalds@linux-foundation.org>
Date: Thu, 7 Mar 2024 12:00:44 -0800
Message-ID: <CAHk-=wia769uoyVz3P7yZURhO8NNB7xeOLX07ZM2vWf1nTLYkQ@mail.gmail.com>
Subject: Re: [PATCH] rcutorture: Fix rcu_torture_pipe_update_one()/rcu_torture_writer()
 data race and concurrency bug
To: paulmck@kernel.org
Cc: Mathieu Desnoyers <mathieu.desnoyers@efficios.com>, Steven Rostedt <rostedt@goodmis.org>, 
	linke li <lilinke99@qq.com>, joel@joelfernandes.org, boqun.feng@gmail.com, 
	dave@stgolabs.net, frederic@kernel.org, jiangshanlai@gmail.com, 
	josh@joshtriplett.org, linux-kernel@vger.kernel.org, 
	qiang.zhang1211@gmail.com, quic_neeraju@quicinc.com, rcu@vger.kernel.org
Content-Type: text/plain; charset="UTF-8"

On Thu, 7 Mar 2024 at 11:47, Paul E. McKenney <paulmck@kernel.org> wrote:
>
> > - The per-thread counter (Thread-Local Storage) incremented by a single
> >   thread, read by various threads concurrently, is a good target
> >   for WRITE_ONCE()/READ_ONCE() pairing. This is actually what we do in
> >   various liburcu implementations which track read-side critical sections
> >   per-thread.
>
> Agreed, but do any of these use WRITE_ONCE(x, READ_ONCE(x) + 1) or
> similar?

Absolutely not.

The READ_ONCE->WRITE_ONCE pattern is almost certainly a bug.

The valid reason to have a WRITE_ONCE() is that there's a _later_
READ_ONCE() on another CPU.

So WRITE_ONCE->READ_ONCE (across threads) is very valid. But
READ_ONCE->WRITE_ONCE (inside a thread) simply is not a valid
operation.

We do have things like "local_t", which allows for non-smp-safe local
thread atomic accesses, but they explicitly are *NOT* about some kind
of READ_ONCE -> WRITE_ONCE sequence that by definition cannot be
atomic unless you disable interrupts and disable preemption (at which
point they become pointless and only generate worse code).

But the point of "local_t" is that you can do things that aresafe if
there is no SMP issues. They are kind of an extension of the
percpu_add() kind of operations.

In fact, I think it might be interesting to catch those
READ_ONCE->WRITE_ONCE chains (perhaps with coccinelle?) because they
are a sign of bugs.

Now, there's certainly some possibility of "I really don't care about
some stats, I'm willing to do non-smp-safe and non-thread safe
operations if they are faster". So I'm not saying a
READ_ONCE->WRITE_ONCE data dependency is _always_ a bug, but I do
think it's a pattern that is very very close to being one.

             Linus