On Wed, Oct 19, 2016 at 09:19:43PM +0200, Jiri Olsa wrote:
> I think the reason here is that presume pmu devices are always added,
> but we add them only if pmu_bus_running (in perf_event_sysfs_init)
> is set which might happen after uncore initcall
>
> attached patch fixes the issue for me
Right, we never expected to be unloaded before userspace runs.
Strictly speaking we should only read pmu_bus_running while holding
pmus_lock, that way we're serialized against perf_event_sysfs_init()
flipping it while we're being removed etc..
With the current setup the introduced race is harmless, but who knows
what other crazy these device people will come up with ;-)
> ---
> diff --git a/kernel/events/core.c b/kernel/events/core.c
> index c6e47e97b33f..c2099b799d16 100644
> --- a/kernel/events/core.c
> +++ b/kernel/events/core.c
> @@ -8871,8 +8871,10 @@ void perf_pmu_unregister(struct pmu *pmu)
> idr_remove(&pmu_idr, pmu->type);
> if (pmu->nr_addr_filters)
> device_remove_file(pmu->dev, &dev_attr_nr_addr_filters);
> - device_del(pmu->dev);
> - put_device(pmu->dev);
> + if (pmu_bus_running) {
> + device_del(pmu->dev);
> + put_device(pmu->dev);
> + }
> free_pmu_context(pmu);
> }
> EXPORT_SYMBOL_GPL(perf_pmu_unregister);
