Re: [PATCH 3/3] rcu: Assume rcu_report_dead() always deals with local CPU

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

On Thu, May 20, 2021 at 02:54:26AM +0200, Frederic Weisbecker wrote:
> On Wed, May 19, 2021 at 11:51:07AM -0700, Paul E. McKenney wrote:
> > On Wed, May 19, 2021 at 02:09:30AM +0200, Frederic Weisbecker wrote:
> > > rcu_report_dead() is always called locally from the idle path. Passing
> > > a CPU number to it suggests otherwise and is rather error-prone as the
> > > code inside relies on locality.
> > > 
> > > Robustify the function prototype and refine the name along the way.
> > > 
> > > Signed-off-by: Frederic Weisbecker <frederic@kernel.org>
> > 
> > Makes a lot of sense, thank you!
> > 
> > On the function name, here is the list:
> > 
> > int rcutree_prepare_cpu(unsigned int cpu) -- notifier from any CPU.
> > void rcu_cpu_starting(unsigned int cpu) -- direct call on incoming CPU.
> > int rcutree_online_cpu(unsigned int cpu) -- notifier from any CPU.
> > 
> > int rcutree_offline_cpu(unsigned int cpu) -- notifier from any CPU.
> > void rcu_report_dead(unsigned int cpu) -- direct call on outgoing CPU.
> > void rcutree_migrate_callbacks(int cpu) -- direct call from surviving CPU.
> > int rcutree_dead_cpu(unsigned int cpu) -- notifier from any CPU.
> > 
> > Note that rcu_report_dead() can also be invoked from cpu_die_early() on
> > other CPU when onlining a CPU fails.  This happens on arm64.  Which might
> > be an arm64 bug, but unless I am missing something it is a case where
> > rcu_report_dead() is called non-locally.
> 
> Hmm, I see it only called with smp_processor_id() from cpu_die_early().

I clearly should have looked at the initialization of that "cpu"
local variable.  Once again, right you are!

> > And the naming is currently a bit random, isn't it?  :-/
> > 
> > Maybe rcutree_*_cpu() if there is a CPU parameter and rcutree_*_self()
> > if all calls run on the CPU in question?
> 
> Makes sense. Or rcutree_*_curr_cpu() but it's going to produce long names.

Either way, it should be a separate patch.

> > I cannot immediately think of a reason to make names reflect whether
> > the corresponding functions are directly called or are called via notifier.
> > Thoughts?
> 
> No indeed, let's wait for some convention to ever emerge :)

Sounds like a long wait, but that works for me!

							Thanx, Paul