Re: [PATCH] rcu: Report a quiescent state when it's exactly in the state

[Paul E. McKenney]

On Fri, May 11, 2018 at 11:30:37PM -0700, Joel Fernandes wrote:
> On Fri, May 11, 2018 at 10:08:24PM -0700, Paul E. McKenney wrote:
> > On Fri, May 11, 2018 at 03:41:38PM -0700, Joel Fernandes wrote:
> > > On Fri, May 11, 2018 at 09:17:46AM -0700, Paul E. McKenney wrote:
> > > > On Fri, May 11, 2018 at 09:57:54PM +0900, Byungchul Park wrote:
> > > > > Hello folks,
> > > > > 
> > > > > I think I wrote the title in a misleading way.
> > > > > 
> > > > > Please change the title to something else such as,
> > > > > "rcu: Report a quiescent state when it's in the state" or,
> > > > > "rcu: Add points reporting quiescent states where proper" or so on.
> > > > > 
> > > > > On 2018-05-11 ìí 5:30, Byungchul Park wrote:
> > > > > >We expect a quiescent state of TASKS_RCU when cond_resched_tasks_rcu_qs()
> > > > > >is called, no matter whether it actually be scheduled or not. However,
> > > > > >it currently doesn't report the quiescent state when the task enters
> > > > > >into __schedule() as it's called with preempt = true. So make it report
> > > > > >the quiescent state unconditionally when cond_resched_tasks_rcu_qs() is
> > > > > >called.
> > > > > >
> > > > > >And in TINY_RCU, even though the quiescent state of rcu_bh also should
> > > > > >be reported when the tick interrupt comes from user, it doesn't. So make
> > > > > >it reported.
> > > > > >
> > > > > >Lastly in TREE_RCU, rcu_note_voluntary_context_switch() should be
> > > > > >reported when the tick interrupt comes from not only user but also idle,
> > > > > >as an extended quiescent state.
> > > > > >
> > > > > >Signed-off-by: Byungchul Park <byungchul.park@xxxxxxx>
> > > > > >---
> > > > > >  include/linux/rcupdate.h | 4 ++--
> > > > > >  kernel/rcu/tiny.c        | 6 +++---
> > > > > >  kernel/rcu/tree.c        | 4 ++--
> > > > > >  3 files changed, 7 insertions(+), 7 deletions(-)
> > > > > >
> > > > > >diff --git a/include/linux/rcupdate.h b/include/linux/rcupdate.h
> > > > > >index ee8cf5fc..7432261 100644
> > > > > >--- a/include/linux/rcupdate.h
> > > > > >+++ b/include/linux/rcupdate.h
> > > > > >@@ -195,8 +195,8 @@ static inline void exit_tasks_rcu_finish(void) { }
> > > > > >   */
> > > > > >  #define cond_resched_tasks_rcu_qs() \
> > > > > >  do { \
> > > > > >-	if (!cond_resched()) \
> > > > > >-		rcu_note_voluntary_context_switch_lite(current); \
> > > > > >+	rcu_note_voluntary_context_switch_lite(current); \
> > > > > >+	cond_resched(); \
> > > > 
> > > > Ah, good point.
> > > > 
> > > > Peter, I have to ask...  Why is "cond_resched()" considered a preemption
> > > > while "schedule()" is not?
> > > 
> > > Infact something interesting I inferred from the __schedule loop related to
> > > your question:
> > > 
> > > switch_count can either be set to prev->invcsw or prev->nvcsw. If we can
> > > assume that switch_count reflects whether the context switch is involuntary
> > > or voluntary,
> > >                       
> > > task-running-state	preempt		switch_count
> > > 0 (running)		1		involuntary
> > > 0			0		involuntary
> > > 1			0		voluntary
> > > 1			1		involuntary
> > > 
> > > According to the above table, both the task's running state and the preempt
> > > parameter to __schedule should be used together to determine if the switch is
> > > a voluntary one or not.
> > > 
> > > So this code in rcu_note_context_switch should really be:
> > > if (!preempt && !(current->state & TASK_RUNNING))
> 
> I should have writte here- !preempt && current->state
> 
> > > 	rcu_note_voluntary_context_switch_lite(current);
> > > 
> > > According to the above table, cond_resched always classifies as an
> > > involuntary switch which makes sense to me. Even though cond_resched is
> > > explicitly called, its still sort of involuntary in the sense its not called
> > > into the scheduler for sleeping, but rather for seeing if something else can
> > > run instead (a preemption point). Infact none of the task deactivation in the
> > > __schedule loop will run if cond_resched is used.
> > > 
> > > I agree that if schedule was called directly but with TASK_RUNNING=1, then
> > > that could probably be classified an involuntary switch too...
> > > 
> > > Also since we're deciding to call rcu_note_voluntary_context_switch_lite
> > > unconditionally, then IMO this comment on that macro:
> > > 
> > > /*
> > >  * Note a voluntary context switch for RCU-tasks benefit.  This is a
> > >  * macro rather than an inline function to avoid #include hell.
> > >  */
> > >  #ifdef CONFIG_TASKS_RCU
> > >  #define rcu_note_voluntary_context_switch_lite(t)
> > > 
> > > Should be changed to:
> > > 
> > > /*
> > >  * Note a attempt to perform a voluntary context switch for RCU-tasks
> > >  * benefit.  This is called even in situations where a context switch
> > >  * didn't really happen even though it was requested. This is a
> > >  * macro rather than an inline function to avoid #include hell.
> > >  */
> > >  #ifdef CONFIG_TASKS_RCU
> > >  #define rcu_note_voluntary_context_switch_lite(t)
> > > 
> > > Right?
> > > 
> > > Correct me if I'm wrong about anything, thanks,
> > 
> > The starting point for me is that Tasks RCU is a special-purpose mechanism
> > for freeing trampolines in PREEMPT=y kernels.  The approach is to arrange
> > for the trampoline to be inaccessible to future execution, wait for a
> > tasks-RCU grace period, then free the trampoline.  So a tasks-RCU grace
> > period must wait until all tasks have spent at least some time outside
> > of a trampoline.  My understanding is that trampolines cannot contain
> > preemption points, such as cond_resched() and cond_resched_tasks_rcu_qs(),
> > so we want to count them as quiescent states regardless of whether or
> > not any associated context switch is counted as involuntary.
> > 
> > What situations lead to the second line of your table above?
> > The sched_yield() system call, but trampolines don't do system calls,
> > either, as far as I know.
> > 
> > So it looks to me like that test can leave out the TASK_RUNNING check.
> 
> I don't know much about tasks-RCU to comment more, sorry. Probably a few more
> reading nights for me to catch up with that. Its possible the check is not
> needed and tasks-RCU can survive without it, but I was thinking from a
> correctness and future-proofing stand point... I generally don't like
> inconsistencies. The check in the __schedule loop is as:
> 
> 	if (!preempt && prev->state) {
> 		....
> 		// switch_count = voluntary context switch counter pointer
> 		....
> 	} else {
> 		....
> 
> 		// switch_count = involuntary context switch counter pointer
> 		....
> 	}
> 
> 	// context switch really happening
> 	if (prev != next) {
> 		....
> 		++switch_count;
> 	}
> 
> The first conditional if (!preempt...) above is what I was referring to which
> also checks the state.
> 
> Also this issue aside, I was more trying to answer your question about why
> schedule() is or isn't a preemption point, by sharing the table but I
> possibly caused more confusion, sorry :-(. I'll let Peter and Steven chime in
> since they know more than me about that and will just shutup and listen
> instead of being more noisy.. :-D

Don't get me wrong, this discussion was quite useful to me.  We probably
need to at least change the comments, and perhaps the code as well.  But
I agree that we need input from Peter and Steven to make much more forward
progress.

								Thanx, Paul