On Tue, Aug 27, 2019 at 11:23:33AM +0200, Sebastian Andrzej Siewior wrote: > On 2019-08-26 09:29:45 [-0700], Paul E. McKenney wrote: > > > The mechanism that is used here may change in future. I just wanted to > > > make sure that from RCU's side it is okay to schedule here. > > > > Good point. > > > > The effect from RCU's viewpoint will be to split any non-rcu_read_lock() > > RCU read-side critical section at this point. This alrady happens in a > > few places, for example, rcu_note_context_switch() constitutes an RCU > > quiescent state despite being invoked with interrupts disabled (as is > > required!). The __schedule() function just needs to understand (and does > > understand) that the RCU read-side critical section that would otherwise > > span that call to rcu_node_context_switch() is split in two by that call. > > Okay. So I read this as invoking schedule() at this point is okay. As long as no one is relying on a non-rcu_read_lock() RCU read-side critical section (local_bh_disable(), preempt_disable(), local_irq_disable(), ...) spanning this call. But that depends on the calling code and on other code it interacts with it, not on any specific need on the part of RCU itself. > Looking at this again, this could also happen on a PREEMPT=y kernel if > the kernel decides to preempt a task within a rcu_read_lock() section > and put it back later on another CPU. This is an rcu_read_lock() critical section, so yes, on a PREEMPT=y kernel, executing schedule() will cause the corresponding RCU read-side critical section to persist, following the preempted tasks. Give or take lockdep complaints. On a PREEMPT=n kernel, schedule() within an RCU read-side critical section instead results in that critical section being split in two. And this will also results in lockdep complaints. > > However, if this was instead an rcu_read_lock() critical section within > > a PREEMPT=y kernel, then if a schedule() occured within stop_one_task(), > > RCU would consider that critical section to be preempted. This means > > that any RCU grace period that is blocked by this RCU read-side critical > > section would remain blocked until stop_one_cpu() resumed, returned, > > and so on until the matching rcu_read_unlock() was reached. In other > > words, RCU would consider that RCU read-side critical section to span > > the call to stop_one_cpu() even if stop_one_cpu() invoked schedule(). > > Isn't that my example from above and what we do in RT? My understanding > is that this is the reason why we need BOOST on RT otherwise the RCU > critical section could remain blocked for some time. At this point, I must confess that I have lost track of whose example it is. It was first reported in 2006, if I remember correctly. ;-) But yes, you are correct, the point of RCU priority boosting is to cause tasks that have been preempted while within RCU read-side critical sections to be scheduled so that they can reach their rcu_read_unlock() calls, thus allowing the current grace period to end. > > On the other hand, within a PREEMPT=n kernel, the call to schedule() > > would split even an rcu_read_lock() critical section. Which is why I > > asked earlier if sleeping_lock_inc() and sleeping_lock_dec() are no-ops > > in !PREEMPT_RT_BASE kernels. We would after all want the usual lockdep > > complaints in that case. > > sleeping_lock_inc() +dec() is only RT specific. It is part of RT's > spin_lock() implementation and used by RCU (rcu_note_context_switch()) > to not complain if invoked within a critical section. Then this is being called when we have something like this, correct? DEFINE_SPINLOCK(mylock); // As opposed to DEFINE_RAW_SPINLOCK(). ... rcu_read_lock(); do_something(); spin_lock(&mylock); // Can block in -rt, thus needs sleeping_lock_inc() ... rcu_read_unlock(); Without sleeping_lock_inc(), lockdep would complain about a voluntary schedule within an RCU read-side critical section. But in -rt, voluntary schedules due to sleeping on a "spinlock" are OK. Am I understanding this correctly? Thanx, Paul
