Date: Tue, 5 Sep 2017 21:37:03 +0200 (CEST)
From: Thomas Gleixner <tglx@linutronix.de>
To: "chengjian (D)" <cj.chengjian@huawei.com>
cc: huawei.libin@huawei.com, mingo@redhat.com, peterz@infradead.org,
        dvhart@infradead.org, linux-kernel@vger.kernel.org
Subject: Re: a competition when some threads acquire futex
In-Reply-To: <3e497508-876f-6474-3f3d-acab06a63b55@huawei.com>
Message-ID: <alpine.DEB.2.20.1709052131130.2393@nanos>
References: <3e497508-876f-6474-3f3d-acab06a63b55@huawei.com>
User-Agent: Alpine 2.20 (DEB 67 2015-01-07)
MIME-Version: 1.0
Content-Type: text/plain; charset=US-ASCII
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 1118
Lines: 40

On Tue, 5 Sep 2017, chengjian (D) wrote:

> int main(int argc, char** argv)
> {
>         pthread_t id1;
>         pthread_t id2;
> 
>         printf("main pid = %d\n", getpid( ));
> 
>         pthread_mutex_init(&mutex, NULL);

So this is a plain mutex, which maps to a plain futex.

> use perf ftrace to shows the graph of the function calls. We found that the
> process 17327 auquire the lock quickly after call futex_wake( ), so the
> process 17328 futex_wait( ) all the time.

The observed functions look correct for that futex type.

> diff --git a/kernel/futex.c b/kernel/futex.c
> index 3d38eaf..0b2d17a 100644
> --- a/kernel/futex.c
> +++ b/kernel/futex.c
> @@ -1545,6 +1545,7 @@ static int wake_futex_pi(u32 __user *uaddr, u32 uval,
> struct futex_pi_state *pi_
>   	spin_unlock(&hb->lock);
>  	wake_up_q(&wake_q);
> +        _cond_resched( );

What's less correct is the placement of the cond_resched() which you patch
into the function:

       wake_futex_pi() 

wake_futex_pi() is not even remotely involved in this problem, so I have a
hard time to understand how this patch 'solves' it.

Thanks,

	tglx