On Thu, Oct 15, 2020 at 09:15:01AM -0700, Paul E. McKenney wrote: > On Thu, Oct 15, 2020 at 11:49:26AM +0200, Peter Zijlstra wrote: > > On Wed, Oct 14, 2020 at 08:41:28PM -0700, Paul E. McKenney wrote: [ . . . ] > > --- a/kernel/rcu/tree.c > > +++ b/kernel/rcu/tree.c > > @@ -1764,8 +1764,7 @@ static bool rcu_gp_init(void) > > smp_mb(); // Pair with barriers used when updating ->ofl_seq to odd values. > > firstseq = READ_ONCE(rnp->ofl_seq); > > if (firstseq & 0x1) > > - while (firstseq == smp_load_acquire(&rnp->ofl_seq)) > > - schedule_timeout_idle(1); // Can't wake unless RCU is watching. > > + smp_cond_load_relaxed(&rnp->ofl_seq, VAL == firstseq); > > smp_mb(); // Pair with barriers used when updating ->ofl_seq to even values. > > raw_spin_lock(&rcu_state.ofl_lock); > > raw_spin_lock_irq_rcu_node(rnp); > > This would work, and would be absolutely necessary if grace periods > took only (say) 500 nanoseconds to complete. But given that they take > multiple milliseconds at best, and given that this race is extremely > unlikely, and given the heavy use of virtualization, I have to stick > with the schedule_timeout_idle(). > > In fact, I have on my list to force this race to happen on the grounds > that if it ain't tested, it don't work... And it only too about 1000 seconds of TREE03 to make this happen, so we should be good just relying on rcutorture. ;-) Thanx, Paul