Re: perf/tracepoint: another fuzzer generated lockup

[Frederic Weisbecker <fweisbec@gmail.com>]

On Sat, Nov 09, 2013 at 03:10:39PM +0100, Peter Zijlstra wrote:
> On Fri, Nov 08, 2013 at 11:36:58PM +0100, Frederic Weisbecker wrote:
> > [  237.627769] perf samples too long (3397569 > 2500), lowering kernel.perf_event_max_sample_rate to 50000
> > [  237.637124] INFO: NMI handler (perf_event_nmi_handler) took too long to run: 444.233 msecs
> > 
> > 444 msecs is huge.
> 
> Be glad your system lived to tell about it ;-) Calling printk() from NMI
> context is Russian roulette; I'm still waiting for the first report it
> actually locked up :-)
> 
> That said, I'm not sure what kernel you're running, but there were some
> issues with time-keeping hereabouts, but more importantly that second
> timing includes the printk() call of the first -- so that's always going
> to be fucked.

So, an idea of what may be happening: an event overflows while FASYNC flag is set so it triggers an irq work
to send the signal (kill_fasync).
After the irq work triggers, it generates an irq_work_exit event, which in turn overflows and,
if it has FASYNC, triggers a new irq work. The irq work triggers and generates an irq work exit event which
has FASYNC flag, etc...

Looks like a nice way to deadlock with an infinite loop of irq work.