Date: Tue, 29 Mar 2016 18:36:54 +0200
From: Peter Zijlstra <peterz@infradead.org>
To: Waiman Long <Waiman.Long@hpe.com>
Cc: Ingo Molnar <mingo@redhat.com>, linux-kernel@vger.kernel.org,
        Linus Torvalds <torvalds@linux-foundation.org>,
        Ding Tianhong <dingtianhong@huawei.com>,
        Jason Low <jason.low2@hpe.com>, Davidlohr Bueso <dave@stgolabs.net>,
        "Paul E. McKenney" <paulmck@us.ibm.com>,
        Thomas Gleixner <tglx@linutronix.de>,
        Will Deacon <Will.Deacon@arm.com>,
        Tim Chen <tim.c.chen@linux.intel.com>,
        Waiman Long <Waiman.Long@hp.com>
Subject: Re: [PATCH v3 3/3] locking/mutex: Avoid missed wakeup of mutex waiter
Message-ID: <20160329163654.GM3408@twins.programming.kicks-ass.net>
References: <1458668804-10138-1-git-send-email-Waiman.Long@hpe.com>
 <1458668804-10138-4-git-send-email-Waiman.Long@hpe.com>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <1458668804-10138-4-git-send-email-Waiman.Long@hpe.com>
User-Agent: Mutt/1.5.21 (2012-12-30)
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 1024
Lines: 21

On Tue, Mar 22, 2016 at 01:46:44PM -0400, Waiman Long wrote:
> The current mutex code sets count to -1 and then sets the task
> state. This is the same sequence that the mutex unlock path is checking
> count and task state. That could lead to a missed wakeup even though
> the problem will be cleared when a new waiter enters the waiting queue.
> 
> This patch reverses the order in the locking slowpath so that the task
> state is set first before setting the count. This should eliminate
> the potential missed wakeup and improve latency.

Is it really a problem though?

So the 'race' is __mutex_lock_common() against
__mutex_fastpath_unlock(), and that is fully serialized as per the
atomic instructions. Either the fast unlock path does 1->0 and the lock
acquires, or the lock sets -1, at which the unlock fails and enters
__mutex_unlock_common_slowpath, which is fully serialised against
__mutex_lock_common by the lock->wait_lock.

I agree that the code is nicer after your patch, but I don't actually
see a problem.