Re: [PATCH v2] ipc/msg: Implement lockless pipelined wakeups

[Peter Zijlstra <peterz@infradead.org>]

On Tue, Nov 03, 2015 at 04:03:29PM +0100, Sebastian Andrzej Siewior wrote:

> -	struct msg_msg		*volatile r_msg;
> +	struct msg_msg		*r_msg;

> +				wake_q_add(wake_q, msr->r_tsk);
>  				msr->r_msg = msg;
> +				/*
> +				 * Rely on the implicit cmpxchg barrier from
> +				 * wake_q_add such that we can ensure that

Davidlohr, didn't you want to make that cmpxchg_relaxed() or
cmpxchg_release() ?

> +				 * updating msr->r_msg is the last write
> +				 * operation: As once set, the receiver can
> +				 * continue, and if we don't have the reference

Which seems to suggest you want to at least make that WRITE_ONCE()

> +				 * count from the wake_q, yet, at that point we
> +				 * can later have a use-after-free condition and
> +				 * bogus wakeup.
> +				 */

And I agree with DAvidlohr that the comment is placed oddly, one would
expect it between wake_q_add() and the store in question.