On Fri, Sep 01, 2017 at 09:29:07PM +0200, Thomas Gleixner wrote: > On Fri, 1 Sep 2017, Don Zickus wrote: > > On Thu, Aug 31, 2017 at 09:16:08AM +0200, Thomas Gleixner wrote: > > > The following deadlock is possible in the watchdog hotplug code: > > > > > > cpus_write_lock() > > > ... > > > takedown_cpu() > > > smpboot_park_threads() > > > smpboot_park_thread() > > > kthread_park() > > > ->park() := watchdog_disable() > > > watchdog_nmi_disable() > > > perf_event_release_kernel(); > > > put_event() > > > _free_event() > > > ->destroy() := hw_perf_event_destroy() > > > x86_release_hardware() > > > release_ds_buffers() > > > get_online_cpus() > > > > > > when a per cpu watchdog perf event is destroyed which drops the last > > > reference to the PMU hardware. The cleanup code there invokes > > > get_online_cpus() which instantly deadlocks because the hotplug percpu > > > rwsem is write locked. > > > > The main reason perf_event_release_kernel is in this path is because the > > oprofile folks complained they couldn't use the perf counters when the > > nmi_watchdog was disabled on the command line. > > If the nmi watchdog is disabled on the command line then there are no > counters claimed at all.
Ah, I see it now. When you park all the threads, you clear the cpumask which then calls the lockup detector_cleanup to release all perf_counters that are no longer used by the cpumask. Further reading this code, I see how the code covers various race conditions we tried solving in the past. Thanks! Cheers, Don
