Re: [RFC v1 2/2] rcu/tree: Remove dynticks_nmi_nesting counter

[Joel Fernandes]

On Wed, Aug 28, 2019 at 02:19:04PM -0700, Paul E. McKenney wrote:
> On Wed, Aug 28, 2019 at 05:05:25PM -0400, Joel Fernandes wrote:
> > On Wed, Aug 28, 2019 at 01:23:30PM -0700, Paul E. McKenney wrote:
> > > On Mon, Aug 26, 2019 at 09:33:54PM -0400, Joel Fernandes (Google) wrote:
> > > > The dynticks_nmi_nesting counter serves 4 purposes:
> > > > 
> > > >       (a) rcu_is_cpu_rrupt_from_idle() needs to be able to detect first
> > > >           interrupt nesting level.
> > > > 
> > > >       (b) We need to detect half-interrupts till we are sure they're not an
> > > >           issue. However, change the comparison to DYNTICK_IRQ_NONIDLE with 0.
> > > > 
> > > >       (c) When a quiescent state report is needed from a nohz_full CPU.
> > > >           The nesting counter detects we are a first level interrupt.
> > > > 
> > > > For (a) we can just use dyntick_nesting == 1 to determine this. Only the
> > > > outermost interrupt that interrupted an RCU-idle state can set it to 1.
> > > > 
> > > > For (b), this warning condition has not occurred for several kernel
> > > > releases.  But we still keep the warning but change it to use
> > > > in_interrupt() instead of the nesting counter. In a later year, we can
> > > > remove the warning.
> > > > 
> > > > For (c), the nest check is not really necessary since forced_tick would
> > > > have been set to true in the outermost interrupt, so the nested/NMI
> > > > interrupts will check forced_tick anyway, and bail.
> > > 
> > > Skipping the commit log and documentation for this pass.
> > [snip] 
> > > > diff --git a/kernel/rcu/tree.c b/kernel/rcu/tree.c
> > > > index 255cd6835526..1465a3e406f8 100644
> > > > --- a/kernel/rcu/tree.c
> > > > +++ b/kernel/rcu/tree.c
> > > > @@ -81,7 +81,6 @@
> > > >  
> > > >  static DEFINE_PER_CPU_SHARED_ALIGNED(struct rcu_data, rcu_data) = {
> > > >  	.dynticks_nesting = 1,
> > > > -	.dynticks_nmi_nesting = 0,
> > > 
> > > This should be in the previous patch, give or take naming.
> > 
> > Done.
> > 
> > > >  	.dynticks = ATOMIC_INIT(RCU_DYNTICK_CTRL_CTR),
> > > >  };
> > > >  struct rcu_state rcu_state = {
> > > > @@ -392,15 +391,9 @@ static int rcu_is_cpu_rrupt_from_idle(void)
> > > >  	/* Check for counter underflows */
> > > >  	RCU_LOCKDEP_WARN(__this_cpu_read(rcu_data.dynticks_nesting) < 0,
> > > >  			 "RCU dynticks_nesting counter underflow!");
> > > > -	RCU_LOCKDEP_WARN(__this_cpu_read(rcu_data.dynticks_nmi_nesting) <= 0,
> > > > -			 "RCU dynticks_nmi_nesting counter underflow/zero!");
> > > >  
> > > > -	/* Are we at first interrupt nesting level? */
> > > > -	if (__this_cpu_read(rcu_data.dynticks_nmi_nesting) != 1)
> > > > -		return false;
> > > > -
> > > > -	/* Does CPU appear to be idle from an RCU standpoint? */
> > > > -	return __this_cpu_read(rcu_data.dynticks_nesting) == 0;
> > > > +	/* Are we the outermost interrupt that arrived when RCU was idle? */
> > > > +	return __this_cpu_read(rcu_data.dynticks_nesting) == 1;
> > > >  }
> > > >  
> > > >  #define DEFAULT_RCU_BLIMIT 10     /* Maximum callbacks per rcu_do_batch ... */
> > > > @@ -564,11 +557,10 @@ static void rcu_eqs_enter(bool user)
> > > >  	struct rcu_data *rdp = this_cpu_ptr(&rcu_data);
> > > >  
> > > >  	/* Entering usermode/idle from interrupt is not handled. These would
> > > > -	 * mean usermode upcalls or idle entry happened from interrupts. But,
> > > > -	 * reset the counter if we warn.
> > > > +	 * mean usermode upcalls or idle exit happened from interrupts. Remove
> > > > +	 * the warning by 2020.
> > > >  	 */
> > > > -	if (WARN_ON_ONCE(rdp->dynticks_nmi_nesting != 0))
> > > > -		WRITE_ONCE(rdp->dynticks_nmi_nesting, 0);
> > > > +	WARN_ON_ONCE(in_interrupt());
> > > 
> > > And this is a red flag.  Bad things happen should some common code
> > > that disables BH be invoked from the idle loop.  This might not be
> > > happening now, but we need to avoid this sort of constraint.
> > > How about instead merging ->dyntick_nesting into the low-order bits
> > > of ->dyntick_nmi_nesting?
> > > 
> > > Yes, this assumes that we don't enter process level twice, but it should
> > > be easy to add a WARN_ON() to test for that.  Except that we don't have
> > > to because there is already this near the end of rcu_eqs_exit():
> > > 
> > > 	WARN_ON_ONCE(rdp->dynticks_nmi_nesting);
> > > 
> > > So the low-order bit of the combined counter could indicate process-level
> > > non-idle, the next three bits could be unused to make interpretation
> > > of hex printouts easier, and then the rest of the bits could be used in
> > > the same way as currently.
> > > 
> > > This would allow a single read to see the full state, so that 0x1 means
> > > at process level in the kernel, 0x11 is interrupt (or NMI) from process
> > > level, 0x10 is interrupt/NMI from idle/user, and so on.
> > > 
> > > What am I missing here?  Why wouldn't this work, and without adding yet
> > > another RCU-imposed constraint on some other subsystem?
> > 
> > What about replacing the warning with a WARN_ON_ONCE(in_irq()), would that
> > address your concern?
> > 
> > Also, considering this warning condition is most likely never occurring as we
> > know it, and we are considering deleting it soon enough, is it really worth
> > reimplementing the whole mechanism with a complex bit-sharing scheme just
> > because of the BH-disable condition you mentioned, which likely doesn't
> > happen today? In my implementation, this is just a simple counter. I feel
> > combining bits in the same counter will just introduce more complexity that
> > this patch tries to address/avoid.
> > 
> > OTOH, I also don't mind with just deleting the warning altogether if you are
> > Ok with that.
> 
> The big advantage of combining the counters is that all of the state is
> explicit and visible in one place.  Plus it can be accessed atomically.
> And it avoids setting a time bomb for some poor guys just trying to get
> their idle-loop jobs done some time in the dim distant future.

I could try the approach you're suggesting but I didn't actually see an issue
with the patch in its current state other than the WARN_ON_ONCE which I could
change to WARN_ON_ONCE(in_irq()) to remove the concern. AFAICS, we don't
detect "half soft-interrupts" in this code in anyway.

I do feel the approach you're suggesting can be a follow up, these 2 patches
just focus on deleting dynticks_nmi_nesting counter and we can test this
approach thoroughly for a release or so.

> Besides, this pair of patches already makes a large change from a
> conceptual viewpoint.  If we are going to make a large change, let's
> get our money's worth out of that change!

IMHO, most of the changes are to code comments, the actual code change is
very little and is just removal of dynticks_nmi_nesting and simplification;
its not really an introduction of a new mechanism.

thanks,

 - Joel