[<prev] [next>] [<thread-prev] [thread-next>] [day] [month] [year] [list]
Date: Thu, 7 Jul 2022 11:22:15 -0700
From: Boqun Feng <boqun.feng@...il.com>
To: Waiman Long <longman@...hat.com>
Cc: Peter Zijlstra <peterz@...radead.org>,
Ingo Molnar <mingo@...hat.com>, Will Deacon <will@...nel.org>,
linux-kernel@...r.kernel.org, Thomas Gleixner <tglx@...utronix.de>,
Sebastian Andrzej Siewior <bigeasy@...utronix.de>,
Juri Lelli <juri.lelli@...hat.com>,
Mike Stowell <mstowell@...hat.com>
Subject: Re: [PATCH v3] locking/rtmutex: Limit # of lock stealing for non-RT
waiters
On Wed, Jul 06, 2022 at 10:03:10AM -0400, Waiman Long wrote:
> On 7/6/22 09:59, Waiman Long wrote:
> > Commit 48eb3f4fcfd3 ("locking/rtmutex: Implement equal priority lock
> > stealing") allows unlimited number of lock stealing's for non-RT
> > tasks. That can lead to lock starvation of non-RT top waiter tasks if
> > there is a constant incoming stream of non-RT lockers. This can cause
> > rcu_preempt self-detected stall or even task lockup in PREEMPT_RT kernel.
> > For example,
> >
> > [77107.424943] rcu: INFO: rcu_preempt self-detected stall on CPU
> > [ 1249.921363] INFO: task systemd:2178 blocked for more than 622 seconds.
> >
> > Avoiding this problem and ensuring forward progress by limiting the
> > number of times that a lock can be stolen from each waiter. This patch
> > sets a threshold of 32. That number is arbitrary and can be changed
> > if needed.
> >
> > Fixes: 48eb3f4fcfd3 ("locking/rtmutex: Implement equal priority lock stealing")
> > Signed-off-by: Waiman Long <longman@...hat.com>
> > ---
> > kernel/locking/rtmutex.c | 9 ++++++---
> > kernel/locking/rtmutex_common.h | 8 ++++++++
> > 2 files changed, 14 insertions(+), 3 deletions(-)
> >
> > [v3: Increase threshold to 32 and add rcu_preempt self-detected stall]
>
> Note that I decided to increase the threshold to 32 from 10 to reduce the
> potential performance impact of this change, if any. We also found out that
> this patch can fix some of the rcu_preempt self-detected stall problems that
> we saw with the PREEMPT_RT kernel. So I added that information in the patch
> description.
>
Have you considered (and tested) whether we can set the threshold
directly proportional to nr_cpu_ids? Because IIUC, the favorable case
for lock stealing is that every CPU gets a chance to steal once. If one
CPU can steal twice, 1) either there is a context switch between two
tasks, which costs similarly as waking up the waiter, or 2) a task drops
and re-graps a lock, which means the task wants to yield to other
waiters of the lock.
Regards,
Boqun
> Cheers,
> Longman
>
Powered by blists - more mailing lists