On Mon, 2019-08-26 at 17:59 +0200, Sebastian Andrzej Siewior wrote: > On 2019-08-23 14:46:39 [-0500], Scott Wood wrote: > > > > Before consolidation, RT mapped rcu_read_lock_bh_held() to > > > > rcu_read_lock_bh() and called rcu_read_lock() from > > > > rcu_read_lock_bh(). This > > > > somehow got lost when rebasing on top of 5.0. > > > > > > so now rcu_read_lock_bh_held() is untouched and in_softirq() reports > > > 1. > > > So the problem is that we never hold RCU but report 1 like we do? > > > > Yes. > > I understand the part where "rcu_read_lock() becomes part of > local_bh_disable()". But why do you modify rcu_read_lock_bh_held() and > rcu_read_lock_bh()? Couldn't they remain as-is? Yes, it looks like they can. I recall having problems with rcu_read_lock_bh_held() in an earlier version which is what prompted the change, but looking back I don't see what the problem was. -Scott
