Re: [RFC 2/2] rcutree: Add checks for dynticks counters in rcu_is_cpu_rrupt_from_idle

[Joel Fernandes]

On Mon, Mar 25, 2019 at 08:53:41AM -0700, Paul E. McKenney wrote:
> On Mon, Mar 25, 2019 at 09:36:46AM -0400, Joel Fernandes wrote:
> > On Sun, Mar 24, 2019 at 04:43:51PM -0700, Paul E. McKenney wrote:
> > > On Fri, Mar 22, 2019 at 11:02:51PM -0400, Joel Fernandes wrote:
> > > > On Fri, Mar 22, 2019 at 09:29:39PM -0400, Joel Fernandes (Google) wrote:
> > > > > In the future we would like to combine the dynticks and dynticks_nesting
> > > > > counters thus leading to simplifying the code. At the moment we cannot
> > > > > do that due to concerns about usermode upcalls appearing to RCU as half
> > > > > of an interrupt. Byungchul tried to do it in [1] but the
> > > > > "half-interrupt" concern was raised. It is half because, what RCU
> > > > > expects is rcu_irq_enter() and rcu_irq_exit() pairs when the usermode
> > > > > exception happens. However, only rcu_irq_enter() is observed. This
> > > > > concern may not be valid anymore, but at least it used to be the case.
> > > > > 
> > > > > Out of abundance of caution, Paul added warnings [2] in the RCU code
> > > > > which if not fired by 2021 may allow us to assume that such
> > > > > half-interrupt scenario cannot happen any more, which can lead to
> > > > > simplification of this code.
> > > > > 
> > > > > Summary of the changes are the following:
> > > > > 
> > > > > (1) In preparation for this combination of counters in the future, we
> > > > > first need to first be sure that rcu_rrupt_from_idle cannot be called
> > > > > from anywhere but a hard-interrupt because previously, the comments
> > > > > suggested otherwise so let us be sure. We discussed this here [3]. We
> > > > > use the services of lockdep to accomplish this.
> > > > > 
> > > > > (2) Further rcu_rrupt_from_idle() is not explicit about how it is using
> > > > > the counters which can lead to weird future bugs. This patch therefore
> > > > > makes it more explicit about the specific counter values being tested
> > > > > 
> > > > > (3) Lastly, we check for counter underflows just to be sure these are
> > > > > not happening, because the previous code in rcu_rrupt_from_idle() was
> > > > > allowing the case where the counters can underflow, and the function
> > > > > would still return true. Now we are checking for specific values so let
> > > > > us be confident by additional checking, that such underflows don't
> > > > > happen. Any case, if they do, we should fix them and the screaming
> > > > > warning is appropriate. All these checks checks are NOOPs if PROVE_RCU
> > > > > and PROVE_LOCKING are disabled.
> > > > > 
> > > > > [1] https://lore.kernel.org/patchwork/patch/952349/
> > > > > [2] Commit e11ec65cc8d6 ("rcu: Add warning to detect half-interrupts")
> > > > > [3] https://lore.kernel.org/lkml/20190312150514.GB249405@xxxxxxxxxx/
> > > > > 
> > > > > Cc: byungchul.park@xxxxxxx
> > > > > Cc: kernel-team@xxxxxxxxxxx
> > > > > Cc: rcu@xxxxxxxxxxxxxxx
> > > > > Signed-off-by: Joel Fernandes (Google) <joel@xxxxxxxxxxxxxxxxx>
> > > > > ---
> > > > >  kernel/rcu/tree.c | 21 +++++++++++++++++----
> > > > >  1 file changed, 17 insertions(+), 4 deletions(-)
> > > > > 
> > > > > diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c
> > > > > index 9180158756d2..d94c8ed29f6b 100644
> > > > > --- a/kernel/rcu/tree.c
> > > > > +++ b/kernel/rcu/tree.c
> > > > > @@ -381,16 +381,29 @@ static void __maybe_unused rcu_momentary_dyntick_idle(void)
> > > > >  }
> > > > >  
> > > > >  /**
> > > > > - * rcu_is_cpu_rrupt_from_idle - see if idle or immediately interrupted from idle
> > > > > + * rcu_is_cpu_rrupt_from_idle - see if interrupted from idle
> > > > >   *
> > > > > - * If the current CPU is idle or running at a first-level (not nested)
> > > > > + * If the current CPU is idle and running at a first-level (not nested)
> > > > >   * interrupt from idle, return true.  The caller must have at least
> > > > >   * disabled preemption.
> > > > >   */
> > > > >  static int rcu_is_cpu_rrupt_from_idle(void)
> > > > >  {
> > > > > -	return __this_cpu_read(rcu_data.dynticks_nesting) <= 0 &&
> > > > > -	       __this_cpu_read(rcu_data.dynticks_nmi_nesting) <= 1;
> > > > > +	/* Called only from within the scheduling-clock interrupt */
> > > > > +	lockdep_assert_in_irq();
> > > > > +
> > > > > +	/* Check for counter underflows */
> > > > > +	RCU_LOCKDEP_WARN(
> > > > > +		(__this_cpu_read(rcu_data.dynticks_nesting) < 0) &&
> > > > > +		(__this_cpu_read(rcu_data.dynticks_nmi_nesting) < 0),
> > > > 
> > > >  
> > > > This condition for the warning is supposed to be || instead of &&. Sorry.
> > > > 
> > > > Or, I will just use 2 RCU_LOCKDEP_WARN(s) here, that's better.
> > > 
> > > Also, the dynticks_nmi_nesting being zero is a bug given that we know
> > > we are in an interrupt handler, right?  Or am I off by one again?
> > 
> > You are right, we can do additional checking for making sure its never zero.
> > I refreshed the patch as below, does this look Ok?
> > 
> > ---8<-----------------------
> > 
> > From: "Joel Fernandes (Google)" <joel@xxxxxxxxxxxxxxxxx>
> > Subject: [RFC v2] rcutree: Add checks for dynticks counters in
> > 
> > In the future we would like to combine the dynticks and dynticks_nesting
> > counters thus leading to simplifying the code. At the moment we cannot
> > do that due to concerns about usermode upcalls appearing to RCU as half
> > of an interrupt. Byungchul tried to do it in [1] but the
> > "half-interrupt" concern was raised. It is half because, what RCU
> > expects is rcu_irq_enter() and rcu_irq_exit() pairs when the usermode
> > exception happens. However, only rcu_irq_enter() is observed. This
> > concern may not be valid anymore, but at least it used to be the case.
> > 
> > Out of abundance of caution, Paul added warnings [2] in the RCU code
> > which if not fired by 2021 may allow us to assume that such
> > half-interrupt scenario cannot happen any more, which can lead to
> > simplification of this code.
> > 
> > Summary of the changes are the following:
> > 
> > (1) In preparation for this combination of counters in the future, we
> > first need to first be sure that rcu_rrupt_from_idle cannot be called
> > from anywhere but a hard-interrupt because previously, the comments
> > suggested otherwise so let us be sure. We discussed this here [3]. We
> > use the services of lockdep to accomplish this.
> > 
> > (2) Further rcu_rrupt_from_idle() is not explicit about how it is using
> > the counters which can lead to weird future bugs. This patch therefore
> > makes it more explicit about the specific counter values being tested
> > 
> > (3) Lastly, we check for counter underflows just to be sure these are
> > not happening, because the previous code in rcu_rrupt_from_idle() was
> > allowing the case where the counters can underflow, and the function
> > would still return true. Now we are checking for specific values so let
> > us be confident by additional checking, that such underflows don't
> > happen. Any case, if they do, we should fix them and the screaming
> > warning is appropriate. All these checks checks are NOOPs if PROVE_RCU
> > and PROVE_LOCKING are disabled.
> > 
> > [1] https://lore.kernel.org/patchwork/patch/952349/
> > [2] Commit e11ec65cc8d6 ("rcu: Add warning to detect half-interrupts")
> > [3] https://lore.kernel.org/lkml/20190312150514.GB249405@xxxxxxxxxx/
> > 
> > Cc: byungchul.park@xxxxxxx
> > Cc: kernel-team@xxxxxxxxxxx
> > Cc: rcu@xxxxxxxxxxxxxxx
> > Signed-off-by: Joel Fernandes (Google) <joel@xxxxxxxxxxxxxxxxx>
> 
> Looks better!  I have applied this and its predecessor, if in backwards
> order.  (Will fix, rebase coming up anyway.)
> 
> I do like your Cc-ing kernel-team@xxxxxxxxxxx -- one less thing for me
> to remember!  ;-)
 
Thanks a lot Paul! :)

 - Joel