Date: Tue, 9 Jan 2018 14:26:02 +0100
From: Peter Zijlstra <peterz@infradead.org>
To: Vince Weaver <vincent.weaver@maine.edu>
Cc: Ingo Molnar <mingo@kernel.org>, linux-kernel@vger.kernel.org,
        Ingo Molnar <mingo@redhat.com>,
        Arnaldo Carvalho de Melo <acme@kernel.org>,
        Thomas Gleixner <tglx@linutronix.de>
Subject: Re: perf: perf_fuzzer quickly locks up on 4.15-rc7
Message-ID: <20180109132602.GA2369@hirez.programming.kicks-ass.net>
References: <alpine.DEB.2.21.1801081105570.319@macbook-air>
 <20180108173005.lkglqrixb2ota6g2@gmail.com>
 <alpine.DEB.2.21.1801081528110.14364@macbook-air>
 <20180109102507.GG6176@hirez.programming.kicks-ass.net>
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20180109102507.GG6176@hirez.programming.kicks-ass.net>
User-Agent: Mutt/1.9.2 (2017-12-15)
Sender: linux-kernel-owner@vger.kernel.org

On Tue, Jan 09, 2018 at 11:25:07AM +0100, Peter Zijlstra wrote:
> On Mon, Jan 08, 2018 at 03:29:42PM -0500, Vince Weaver wrote:
> > On Mon, 8 Jan 2018, Ingo Molnar wrote:
> > > 
> > > Note that the page table isolation (PTI) feature has a number of effects on perf 
> > > and on NMI handlers, so one of the things to try would be to disable PTI.
> > 
> > Yes, it seems to be a KPTI issue.
> > 
> > With KPTI disabled I can fuzz for a few hours, no problems.
> > 
> > With KPTI enabled the perf_fuzzer on this haswell machine will lock up 
> > within minutes.
> 
> I'll go take a look...

OK, so I'm running on an IVB-EP with PTI enabled. I insta triggered a
lockdep splat, after fixing that I get NMI watchdog triggers that look
like false positives; and I suspect:

[  290.791675] perf: Dynamic interrupt throttling disabled, can hang your system!

is responsible for that. That is, subsequent stacktraces for the same
CPU are unrelated, which make it look like the CPU actually makes
progress.

So remind me again, how are you running that fuzzer? I'm running
./fast_repro99.sh as root.