On Thu, Jul 14, 2016 at 08:25:47PM +0200, Peter Zijlstra wrote: > @@ -19,6 +20,13 @@ int __percpu_init_rwsem(struct percpu_rw > > /* ->rw_sem represents the whole percpu_rw_semaphore for lockdep */ > rcu_sync_init(&sem->rss, RCU_SCHED_SYNC); > + if (bias == PERCPU_RWSEM_WRITER) { > + /* > + * Disable rcu_sync() and force slow path. > + */ > + sem->rss.gp_count++; > + sem->rss.gp_state = !0; > + } > __init_rwsem(&sem->rw_sem, name, rwsem_key); > init_waitqueue_head(&sem->writer); > sem->state = readers_slow;
So this seemed like a better deal than calling rcu_sync_enter(), because that would still incur a (pointless) synchronize_sched() at init time and people do tend to complain about things like that.