Re: [clocksource] 8901ecc231: stress-ng.lockbus.ops_per_sec -9.5% regression

From: Matthew Wilcox <willy@infradead.org>
To: "Paul E. McKenney" <paulmck@kernel.org>
Cc: Andi Kleen <ak@linux.intel.com>, Feng Tang <feng.tang@intel.com>,
	kernel test robot <oliver.sang@intel.com>,
	John Stultz <john.stultz@linaro.org>,
	Thomas Gleixner <tglx@linutronix.de>,
	Stephen Boyd <sboyd@kernel.org>, Jonathan Corbet <corbet@lwn.net>,
	Mark Rutland <Mark.Rutland@arm.com>,
	Marc Zyngier <maz@kernel.org>,
	Xing Zhengjun <zhengjun.xing@linux.intel.com>,
	Chris Mason <clm@fb.com>, LKML <linux-kernel@vger.kernel.org>,
	Linux Memory Management List <linux-mm@kvack.org>,
	lkp@lists.01.org, lkp@intel.com, ying.huang@intel.com,
	zhengjun.xing@intel.com
Subject: Re: [clocksource] 8901ecc231: stress-ng.lockbus.ops_per_sec -9.5% regression
Date: Thu, 27 May 2021 20:29:32 +0100	[thread overview]
Message-ID: <YK/zHMPSZSKrmXC6@casper.infradead.org> (raw)
In-Reply-To: <20210527191923.GD4397@paulmck-ThinkPad-P17-Gen-1>

On Thu, May 27, 2021 at 12:19:23PM -0700, Paul E. McKenney wrote:
> On Thu, May 27, 2021 at 12:01:23PM -0700, Andi Kleen wrote:
> > 
> > >      Nevertheless, it is quite possible that real-world use will result in
> > >      some situation requiring that high-stress workloads run on hardware
> > >      not designed to accommodate them, and also requiring that the kernel
> > >      refrain from marking clocksources unstable.
> > >      Therefore, provide an out-of-tree patch that reacts to this situation
> > 
> > out-of-tree means it will not be submitted?
> > 
> > I think it would make sense upstream, but perhaps guarded with some option.
> 
> The reason I do not intend to immediately upstream this patch is that
> it increases the probability that a real clocksource read-latency issue
> will be ignored, for example, during hardware bringup.  Furthermore,
> the only known need from it comes from hardware that is, in the words
> of the stress-ng man page, "poorly designed".  And the timing of this
> email thread leads me to believe that such hardware is not easy to obtain.

I think you're placing a little too much weight on the documentation
here.  It seems that a continuous stream of locked operations executed
in userspace on a single CPU can cause this problem to occur.  If that's
true all the way out to one guest in a hypervisor can cause problems
for the hypervisor itself, I think cloud providers everywhere are
going to want this patch?

> My thought is therefore to keep this patch out of tree for now.
> If it becomes clear that long-latency clocksource reads really are
> a significant issue in their own right (as opposed to merely being a
> symptom of a hardware or firmware bug), then this patch is available to
> immediately respond to that issue.
> 
> And there would then be strong evidence in favor of me biting the bullet,
> adding the complexity and the additional option (with your Suggested-by),
> and getting that upstream and into -stable.
> 
> Seem reasonable?
> 
> 							Thanx, Paul
> 