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

["Paul E. McKenney" <paulmck@kernel.org>]

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.

> Signed-off-by: Joel Fernandes (Google) <joel@joelfernandes.org>
> ---
>  .../Data-Structures/Data-Structures.rst       | 31 +++------
>  Documentation/RCU/stallwarn.txt               |  6 +-
>  kernel/rcu/tree.c                             | 64 +++++++------------
>  kernel/rcu/tree.h                             |  4 +-
>  kernel/rcu/tree_stall.h                       |  4 +-
>  5 files changed, 41 insertions(+), 68 deletions(-)

[ . . . ]

> 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.

>  	.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?

							Thanx, Paul