Re: [PATCH 0/4] rcu: Unify a bit [non-]PREEMPT expedited quiescent state report

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

On Thu, Sep 16, 2021 at 02:10:44PM +0200, Frederic Weisbecker wrote:
> This eventually removes rcu_data.exp_deferred_qs to use
> rcu_data.cpu no_qs.b.exp instead.
> 
> For those like me who need a headlamp to walk there: https://ibb.co/3d06r0V
> 
> git://git.kernel.org/pub/scm/linux/kernel/git/frederic/linux-dynticks.git
> 	rcu/dev
> 
> HEAD: 7d9d8a0c6141f95cbac4367b12e755bfabb383ee

Applied and pushed with some additional marking and commit-log editing.
The marking is needed because rcu_qs() can be invoked with interrupts
enabled, which can result in a low-probability data race with the
expedited IPI handler rcu_exp_handler().  There was another access in
sync_sched_exp_online_cleanup() than needed marking, so I am queuing a
separate commit for that one.

I suspect that this problem predated your series.  This is a low
probability data race, so KCSAN might be hard-pressed to spot it.

Anyway, the result in much much nicer with the converged use of the
->cpu_no_qs.b.exp field, so thank you!!!

Might simplify a few of your future headlamp diagrams as well?  ;-)

						Thanx, Paul

> Thanks,
> 	Frederic
> ---
> 
> Frederic Weisbecker (4):
>       rcu: Ignore rdp.cpu_no_qs.b.exp on premptible RCU's rcu_qs()
>       rcu: Remove useless WRITE_ONCE() on rcu_data.exp_deferred_qs
>       rcu: Move rcu_data.cpu_no_qs.b.exp reset to rcu_export_exp_rdp()
>       rcu: Remove rcu_data.exp_deferred_qs and convert to rcu_data.cpu no_qs.b.exp
> 
> 
>  kernel/rcu/tree.h        |  1 -
>  kernel/rcu/tree_exp.h    |  6 +++---
>  kernel/rcu/tree_plugin.h | 24 +++++++++++++-----------
>  3 files changed, 16 insertions(+), 15 deletions(-)