On Mon, 2018-03-19 at 08:21 -0700, tj@xxxxxxxxxx wrote: > On Mon, Mar 19, 2018 at 03:16:07PM +0000, Bart Van Assche wrote: > > As explained in the comment in scsi_device_quiesce(), the effect of > > blk_set_preempt_only() must be visible for percpu_ref_tryget() calls that > > occur after the queue unfreeze triggered by scsi_device_quiesce(). Hence the > > RCU read lock calls in blk_queue_enter(). Since these RCU read lock calls > > surround a function call that uses rcu_read_lock_sched(), namely > > percpu_ref_tryget_live(), we have to use sched-RCU in both blk_queue_enter() > > and scsi_device_quiesce(). > > Let's please not depend on rcu_read_lock_sched() in > percpu_ref_tryget_live(). That's an implementation detail. If the > code needs RCU based synchronization, it should perform them > explicitly. Hello Tejun, The algorithm explained above does not depend on sched-rcu. But because percpu_ref_tryget_live() uses sched-rcu and because we need to add an RCU lock around that call we are forced to use sched-rcu. I hope this makes it clear. Bart.