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

[Paul E. McKenney]

On Thu, Aug 29, 2019 at 11:13:25AM -0400, Joel Fernandes wrote:
> On Thu, Aug 29, 2019 at 10:43:55AM -0400, Joel Fernandes wrote:
> > On Wed, Aug 28, 2019 at 08:43:36PM -0700, Paul E. McKenney wrote:
> > [snip]
> > > > > > > This change is not fixing a bug, so there is no need for an emergency fix,
> > > > > > > and thus no point in additional churn.  I understand that it is a bit
> > > > > > > annoying to code and test something and have your friendly maintainer say
> > > > > > > "sorry, wrong rocks", and the reason that I understand this is that I do
> > > > > > > that to myself rather often.
> > > > > > 
> > > > > > The motivation for me for this change is to avoid future bugs such as with
> > > > > > the following patch where "== 2" did not take the force write of
> > > > > > DYNTICK_IRQ_NONIDLE into account:
> > > > > > https://git.kernel.org/pub/scm/linux/kernel/git/paulmck/linux-rcu.git/commit/?h=dev&id=13c4b07593977d9288e5d0c21c89d9ba27e2ea1f
> > > > > 
> > > > > Yes, the current code does need some simplification.
> > > > > 
> > > > > > I still don't see it as pointless churn, it is also a maintenance cost in its
> > > > > > current form and the simplification is worth it IMHO both from a readability,
> > > > > > and maintenance stand point.
> > > > > > 
> > > > > > I still don't see what's technically wrong with the patch. I could perhaps
> > > > > > add the above "== 2" point in the patch?
> > > > > 
> > > > > I don't know of a crash or splat your patch would cause, if that is
> > > > > your question.  But that is also true of the current code, so the point
> > > > > is simplification, not bug fixing.  And from what I can see, there is an
> > > > > opportunity to simplify quite a bit further.  And with something like
> > > > > RCU, further simplification is worth -serious- consideration.
> > > > > 
> > > > > > We could also discuss f2f at LPC to see if we can agree about it?
> > > > > 
> > > > > That might make a lot of sense.
> > > > 
> > > > Sure. I am up for a further redesign / simplification. I will think more
> > > > about your suggestions and can also further discuss at LPC.
> > > 
> > > One question that might (or might not) help:  Given the compound counter,
> > > where the low-order hex digit indicates whether the corresponding CPU
> > > is running in a non-idle kernel task and the rest of the hex digits
> > > indicate the NMI-style nesting counter shifted up by four bits, what
> > > could rcu_is_cpu_rrupt_from_idle() be reduced to?
> > > 
> > > > And this patch is on LKML archives and is not going anywhere so there's no
> > > > rush I guess ;-)
> > > 
> > > True enough!  ;-)
> > 
> > Paul, do we also nuke rcu_eqs_special_set()?  Currently I don't see anyone
> > using it. And also remove the bottom most bit of dynticks?
> > 
> > Also what happens if a TLB flush broadcast is needed? Do we IPI nohz or idle
> > CPUs are the moment?
> > 
> > All of this was introduced in:
> > b8c17e6664c4 ("rcu: Maintain special bits at bottom of ->dynticks counter")
> 
> 
> Paul, also what what happens in the following scenario:
> 
> CPU0                                                 CPU1
> 
> A syscall causes rcu_eqs_exit()
> rcu_read_lock();
>                                                      ---> FQS loop waiting on
> 						           dyntick_snap
> usermode-upcall  entry -->causes rcu_eqs_enter();
> 
> usermode-upcall  exit  -->causes rcu_eqs_exit();
> 
>                                                      ---> FQS loop sees
> 						          dyntick snap
> 							  increment and
> 							  declares CPU0 is
> 							  in a QS state
> 							  before the
> 							  rcu_read_unlock!
> 
> rcu_read_unlock();
> ---
> 
> Does the context tracking not call rcu_user_enter() in this case, or did I
> really miss something?

Holding rcu_read_lock() across usermode execution (in this case,
the usermode upcall) is a bad idea.  Why is CPU 0 doing that?

							Thanx, Paul