On Mon, 19 Jul 2021 17:38:45 +0200 Paolo Bonzini wrote: >On 18/07/21 14:42, Hillf Danton wrote: >>> It's caused by the missing wakeup, i.e. eventfd_signal not really >>> signaling anything. >> >> Can you please point me to the waiters in the mainline? > >It's irqfd_wakeup. Thanks for your light. With PREEMPT_RT put aside, the race looks like the following. CPU0 CPU1 CPU2 ---- ---- ---- lock waitqueue wake up waiters unlock waitqueue lock waitqueue no waiter unlock waitqueue lock waitqueue add waiter unlock waitqueue If the missing wakeup on CPU1 is bloody critical to the waiter added on CPU2 then the size of race window is supposed to play magic. The race window can be simulated by giving up wakeup if trylock fails. With PREEMPT_RT before your patch, eventfd_wake_count prevents the preempting waker from acquiring the waitqueue lock and ends up with the race window widened because of the certainty of missing wakeup. CPU0 CPU1 ---- ---- lock waitqueue wake up <-- preempt waiters unlock waitqueue lock waitqueue add waiter unlock waitqueue With PREEMPT_RT after your patch, the race window goes back to without PREEMPT_RT because of no occurence of preemption. But the preempting waker can not make sense without the waiter who is bloody special. Why is it so in the first place? Or it is not at all but the race existing from Monday to Friday.
