Message-Id: <20190821153957.GG28441@linux.ibm.com>
Date:   Wed, 21 Aug 2019 08:39:57 -0700
From:   "Paul E. McKenney" <paulmck@...ux.ibm.com>
To:     Joel Fernandes <joel@...lfernandes.org>
Cc:     linux-kernel@...r.kernel.org,
        Josh Triplett <josh@...htriplett.org>,
        Lai Jiangshan <jiangshanlai@...il.com>,
        Mathieu Desnoyers <mathieu.desnoyers@...icios.com>,
        rcu@...r.kernel.org, Steven Rostedt <rostedt@...dmis.org>
Subject: Re: [RFC v2] rcu/tree: Try to invoke_rcu_core() if in_irq() during
 unlock

On Wed, Aug 21, 2019 at 10:56:17AM -0400, Joel Fernandes wrote:
> On Wed, Aug 21, 2019 at 10:38:41AM -0400, Joel Fernandes wrote:
> > On Mon, Aug 19, 2019 at 08:41:43AM -0700, Paul E. McKenney wrote:
> > > On Mon, Aug 19, 2019 at 07:33:14AM -0700, Paul E. McKenney wrote:
> > > > On Mon, Aug 19, 2019 at 05:57:57AM -0700, Paul E. McKenney wrote:
> > > > > On Sun, Aug 18, 2019 at 07:29:27PM -0700, Paul E. McKenney wrote:
> > > > > > On Sun, Aug 18, 2019 at 09:46:23PM -0400, Joel Fernandes wrote:
> > > > > > > On Sun, Aug 18, 2019 at 09:41:43PM -0400, Joel Fernandes wrote:
> > > > > > > > On Sun, Aug 18, 2019 at 06:21:53PM -0700, Paul E. McKenney wrote:
> > > > > > > [snip]
> > > > > > > > > > > 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();
> > > > > > > > 	}
> > > > > > > 
> > > > > > > The point you were making in that thread being, current_reader() ->
> > > > > > > rcu_read_unlock() -> rcu_read_unlock_special() would not do any wakeups
> > > > > > > because previous_reader() sets the deferred_qs bit.
> > > > > > > 
> > > > > > > Anyway, I will add all of this into the changelog.
> > > > > > 
> > > > > > Examples are good, but what makes it so that there are no examples of
> > > > > > its being unsafe?
> > > > > > 
> > > > > > And a few questions along the way, some quick quiz, some more serious.
> > > > > > Would it be safe if it checked in_interrupt() instead of in_irq()?
> > > > > > If not, should the in_interrupt() in the "if" condition preceding the
> > > > > > added "else if" be changed to in_irq()?  Would it make sense to add an
> > > > > > "|| !irqs_were_disabled" do your new "else if" condition?  Would the
> > > > > > body of the "else if" actually be executed in current mainline?
> > > > > > 
> > > > > > In an attempt to be at least a little constructive, I am doing some
> > > > > > testing of this patch overnight, along with a WARN_ON_ONCE() to see if
> > > > > > that invoke_rcu_core() is ever reached.
> > > > > 
> > > > > And that WARN_ON_ONCE() never triggered in two-hour rcutorture runs of
> > > > > TREE01, TREE02, TREE03, and TREE09.  (These are the TREE variants in
> > > > > CFLIST that have CONFIG_PREEMPT=y.)
> > > > > 
> > > > > This of course raises other questions.  But first, do you see that code
> > > > > executing in your testing?
> > > > 
> > > > Never mind!  Idiot here forgot the "--bootargs rcutree.use_softirq"...
> > > 
> > > So this time I ran the test this way:
> > > 
> > > tools/testing/selftests/rcutorture/bin/kvm.sh --cpus 8 --duration 10 --configs "TREE01 TREE02 TREE03 TREE09" --bootargs "rcutree.use_softirq=0"
> > > 
> > > Still no splats.  Though only 10-minute runs instead of the two-hour runs
> > > I did last night.  (Got other stuff I need to do, sorry!)
> > > 
> > > My test version of your patch is shown below.  Please let me know if I messed
> > > something up.
> > 
> > I think you also need to pass rcutorture.irqreader=1 ?
> > 
> > Otherwise seems all readers happen in process context AFAICS.
> 
> Which is the default setting for that, so that's not the issue.

Yep!

> I think one reason could be, in_irq() is false when the timer callback
> executes, since the timer callback is executing after a grace-period. The
> stack is as follows:
> 
> Any reason why we cannot both test for call_rcu() and execute the RCU
> callback from the timer hardirq handler?
> 
> In fact, I guess on use_nosoftirq systems, the callback will not even run
> in softirq context.
> 
> [   20.553361]  => rcu_torture_timer_cb
> [   20.553361]  => rcu_do_batch
> [   20.553361]  => rcu_core
> [   20.553361]  => __do_softirq
> [   20.553361]  => do_softirq_own_stack
> [   20.553361]  => do_softirq.part.16
> [   20.553361]  => __local_bh_enable_ip
> [   20.553361]  => rcutorture_one_extend
> [   20.553361]  => rcu_torture_one_read
> [   20.553361]  => rcu_torture_reader
> [   20.553361]  => kthread
> [   20.553361]  => ret_from_fork

Well, it is rcu_read_lock() and rcu_read_unlock() that matters
for this case rather than the callback.  But yes, given in_irq(),
rcu_read_lock() and rcu_read_unlock() would need to have executed
from a hardware interrupt handler.  And would need to get one of the
->rcu_read_lock_special bits set somehow.

But you can use smp_call_function() to invoke a function that runs in
hardware interrupt handler context, and you can do this within either
rcuperf or rcutorture.

And yes, this line of reasoning did inform at least some of my skepticism
surrounding your initial patch, in case you were wondering about some
of my earlier questions.  ;-)

							Thanx, Paul

Powered by blists - more mailing lists