Re: [RFC v2] rcu/tree: Try to invoke_rcu_core() if in_irq() during unlock

[Joel Fernandes <joel@joelfernandes.org>]

On Sun, Aug 18, 2019 at 06:21:53PM -0700, Paul E. McKenney wrote:
> On Sun, Aug 18, 2019 at 07:38:39PM -0400, Joel Fernandes wrote:
> > On Sun, Aug 18, 2019 at 04:31:35PM -0700, Paul E. McKenney wrote:
> > > On Sun, Aug 18, 2019 at 06:35:11PM -0400, Joel Fernandes wrote:
> > > > On Sun, Aug 18, 2019 at 06:32:30PM -0400, Joel Fernandes wrote:
> > > > > On Sun, Aug 18, 2019 at 03:12:10PM -0700, Paul E. McKenney wrote:
> > > > > > On Sun, Aug 18, 2019 at 05:49:48PM -0400, Joel Fernandes (Google) wrote:
> > > > > > > When we're in hard interrupt context in rcu_read_unlock_special(), we
> > > > > > > can still benefit from invoke_rcu_core() doing wake ups of rcuc
> > > > > > > threads when the !use_softirq parameter is passed.  This is safe
> > > > > > > to do so because:
> > > > > > > 
> > > > > > > 1. We avoid the scheduler deadlock issues thanks to the deferred_qs bit
> > > > > > > introduced in commit 23634ebc1d94 ("rcu: Check for wakeup-safe
> > > > > > > conditions in rcu_read_unlock_special()") by checking for the same in
> > > > > > > this patch.
> > > > > > > 
> > > > > > > 2. in_irq() implies in_interrupt() which implies raising softirq will
> > > > > > > not do any wake ups.
> > > > > > > 
> > > > > > > The rcuc thread which is awakened will run when the interrupt returns.
> > > > > > > 
> > > > > > > We also honor 25102de ("rcu: Only do rcu_read_unlock_special() wakeups
> > > > > > > if expedited") thus doing the rcuc awakening only when none of the
> > > > > > > following are true:
> > > > > > >   1. Critical section is blocking an expedited GP.
> > > > > > >   2. A nohz_full CPU.
> > > > > > > If neither of these cases are true (exp == false), then the "else" block
> > > > > > > will run to do the irq_work stuff.
> > > > > > > 
> > > > > > > This commit is based on a partial revert of d143b3d1cd89 ("rcu: Simplify
> > > > > > > rcu_read_unlock_special() deferred wakeups") with an additional in_irq()
> > > > > > > check added.
> > > > > > > 
> > > > > > > Signed-off-by: Joel Fernandes (Google) <joel@joelfernandes.org>
> > > > > > 
> > > > > > OK, I will bite...  If it is safe to wake up an rcuc kthread, why
> > > > > > is it not safe to do raise_softirq()?
> > > > > 
> > > > > Because raise_softirq should not be done and/or doesn't do anything
> > > > > if use_softirq == false. In fact, RCU_SOFTIRQ doesn't even existing if
> > > > > use_softirq == false. The "else if" condition of this patch uses for
> > > > > use_softirq.
> > > > > 
> > > > > Or, did I miss your point?
> > > 
> > > I am concerned that added "else if" condition might not be sufficient
> > > to eliminate all possible cases of the caller holding a scheduler lock,
> > > which could result in deadlock in the ensuing wakeup.  Might be me missing
> > > something, but such deadlocks have been a recurring problem in the past.
> > 
> > I thought that was the whole point of the
> > rcu_read_unlock_special.b.deferred_qs bit that was introduced in
> > 23634ebc1d94. We are checking that bit in the "else if" here as well. So this
> > should be no less immune to scheduler deadlocks any more than the preceding
> > "else if" where we are checking this bit.
> 
> I would have much more confidence in a line of reasoning that led to
> "immune to scheduler deadlocks" than one that led to "no less immune to
> scheduler deadlocks".  ;-)

That is fair :-D But let me explain,

What I meant is, if we are saying that this solution has a scheduler
deadlock, then that would almost certainly imply that the existing code has
scheduler deadlock issue. Since the existing code uses the same technique
(using the deferred_qs bit in the union) to prevent the deadlock we were
discussing a few months back. If that is indeed the case, it is good to be
discussing this since we can discuss if the existing code needs any fixing as
well.

> > > Also, your commit log's point #2 is "in_irq() implies in_interrupt()
> > > which implies raising softirq will not do any wake ups."  This mention
> > > of softirq seems a bit odd, given that we are going to wake up a rcuc
> > > kthread.  Of course, this did nothing to quell my suspicions.  ;-)
> > 
> > Yes, I should delete this #2 from the changelog since it is not very relevant
> > (I feel now). My point with #2 was that even if were to raise a softirq
> > (which we are not), a scheduler wakeup of ksoftirqd is impossible in this
> > path anyway since in_irq() implies in_interrupt().
> 
> Please!  Could you also add a first-principles explanation of why
> the added condition is immune from scheduler deadlocks?

Sure I can add an example in the change log, however I was thinking of this
example which you mentioned:
https://lore.kernel.org/lkml/20190627173831.GW26519@linux.ibm.com/

	previous_reader()
	{
		rcu_read_lock();
		do_something(); /* Preemption happened here. */
		local_irq_disable(); /* Cannot be the scheduler! */
		do_something_else();
		rcu_read_unlock();  /* Must defer QS, task still queued. */
		do_some_other_thing();
		local_irq_enable();
	}

	current_reader() /* QS from previous_reader() is still deferred. */
	{
		local_irq_disable();  /* Might be the scheduler. */
		do_whatever();
		rcu_read_lock();
		do_whatever_else();
		rcu_read_unlock();  /* Must still defer reporting QS. */
		do_whatever_comes_to_mind();
		local_irq_enable();
	}

One modification of the example could be, previous_reader() could also do:
	previous_reader()
	{
		rcu_read_lock();
		do_something_that_takes_really_long(); /* causes need_qs in
							  the unlock_special_union to be set */
		local_irq_disable(); /* Cannot be the scheduler! */
		do_something_else();
		rcu_read_unlock();  /* Must defer QS, task still queued. */
		do_some_other_thing();
		local_irq_enable();
	}


thanks!

 - Joel