On Tue, 2017-09-19 at 13:37 +0800, Boqun Feng wrote: > On Mon, Sep 18, 2017 at 09:04:56PM -0700, Paul E. McKenney wrote: > > On Tue, Sep 19, 2017 at 11:48:22AM +0900, Byungchul Park wrote: > > > On Mon, Sep 18, 2017 at 07:33:29PM -0700, Paul E. McKenney wrote: > > > > > > Hello Paul and Steven, > > > > > > > > So I think this is another false positive, and the reason is we use > st->done for multiple purposes. > > > > > > > This is saying: > > > > > > > > > > > > Thread A > > > > > > -------- > > > > > > takedown_cpu() > > > > > > irq_lock_sparse() > > > > > > wait_for_completion(&st->done) // Wait for completion of B > > Thread A wait for the idle task on the outgoing to set the st->state to > CPUHP_AP_IDLE_DEAD(i.e. the corresponding complete() is the one in > cpuhp_complete_idle_dead()), and it happens when we try to _offline_ a > cpu. > > > > > > > irq_unlock_sparse() > > > > > > > > > > > > Thread B > > > > > > -------- > > > > > > cpuhp_invoke_callback() > > > > > > irq_lock_sparse() // Wait for A to irq_unlock_sparse() > > irq_affinity_online_cpu() is called here, so it happens when we try to > _online_ a cpu. > > > > > > > (on the way going to complete(&st->done)) > > and we are going to complete(&st->done) in a hotplug thread context to > indicate the hotplug thread has finished its job(i.e. this complete() is > the one in cpuhp_thread_fun()). > > > So even though the &st->done are the same instance, the deadlock could > not happen, I think, as we could not up/down a same cpu at the same > time? > > If I'm not missing something subtle. To fix this we can either > > 1) have dedicated completion instances for different wait purposes > in cpuhp_cpu_state. > > or > > 2) extend crossrelease to have the "subclass" concept, so that > callsite of complete() and wait_for_completion() for the same > completion instance but with different purposes could be > differed by lockdep. > > Thoughts?
https://lkml.org/lkml/2017/9/5/184 Peter's patches worked for me, but per tglx, additional (non- grasshopper level) hotplug-fu is required. -Mike
