On Fri, Jul 20, 2018 at 4:52 AM Peter Zijlstra <peterz@xxxxxxxxxxxxx> wrote: > > On Thu, Jul 19, 2018 at 12:12:53PM -0700, Cong Wang wrote: > > hrtimer_cancel() busy-waits for the hrtimer callback to stop, > > pretty much like del_timer_sync(). This creates a possible deadlock > > scenario where we hold a spinlock before calling hrtimer_cancel() > > while in trying to acquire the same spinlock in the callback. > > Has this actually been observed? Without lockdep annotation, it is not easy to observe. > > > cpu_clock_event_init(): > > perf_swevent_init_hrtimer(): > > hwc->hrtimer.function = perf_swevent_hrtimer; > > > > perf_swevent_hrtimer(): > > __perf_event_overflow(): > > __perf_event_account_interrupt(): > > perf_adjust_period(): > > pmu->stop(): > > cpu_clock_event_stop(): > > perf_swevent_cancel(): > > hrtimer_cancel() > > Please explain how a hrtimer event ever gets to perf_adjust_period(). > Last I checked perf_swevent_init_hrtimer() results in attr.freq=0. Good point. I thought attr.freq is specified by user-space, but seems perf_swevent_init_hrtimer() clears it purposely and will not change after initialization, interesting... > > > Getting stuck in an hrtimer is a disaster: > > You'll get NMI watchdog splats. Getting stuck in NMI context is far more > 'interesting :-) Yes, I did see a stack trace in perf_swevent_hrtimer() which led me here. But I have to admit among those hundreds of soft lockup's, I only saw one showing swevent hrtimer backtrace. Previously I thought this is because of NMI handler race, but Jiri pointed out the race doesn't exist. > > > +#define PERF_EF_NO_WAIT 0x08 /* do not wait when stopping, for > > + * example, waiting for a timer > > + */ > > That's a broken comment style. It is picked by checkpatch.pl, not me, I chose a different one and got a complain. :) Thanks!
