Re: [PATCH v2 3/9] rcu/sync: Remove custom check for reader-section

[Joel Fernandes]

On Sat, Jul 13, 2019 at 08:50:10AM -0700, Paul E. McKenney wrote:
> On Sat, Jul 13, 2019 at 11:36:06AM -0400, Joel Fernandes wrote:
> > On Sat, Jul 13, 2019 at 07:41:08AM -0700, Paul E. McKenney wrote:
> > > On Sat, Jul 13, 2019 at 09:30:49AM -0400, Joel Fernandes wrote:
> > > > On Sat, Jul 13, 2019 at 01:21:14AM -0700, Paul E. McKenney wrote:
> > > > > On Fri, Jul 12, 2019 at 11:10:08PM -0400, Joel Fernandes wrote:
> > > > > > On Fri, Jul 12, 2019 at 11:01:50PM -0400, Joel Fernandes wrote:
> > > > > > > On Fri, Jul 12, 2019 at 04:32:06PM -0700, Paul E. McKenney wrote:
> > > > > > > > On Fri, Jul 12, 2019 at 05:35:59PM -0400, Joel Fernandes wrote:
> > > > > > > > > On Fri, Jul 12, 2019 at 01:00:18PM -0400, Joel Fernandes (Google) wrote:
> > > > > > > > > > The rcu/sync code was doing its own check whether we are in a reader
> > > > > > > > > > section. With RCU consolidating flavors and the generic helper added in
> > > > > > > > > > this series, this is no longer need. We can just use the generic helper
> > > > > > > > > > and it results in a nice cleanup.
> > > > > > > > > > 
> > > > > > > > > > Cc: Oleg Nesterov <oleg@xxxxxxxxxx>
> > > > > > > > > > Signed-off-by: Joel Fernandes (Google) <joel@xxxxxxxxxxxxxxxxx>
> > > > > > > > > 
> > > > > > > > > Hi Oleg,
> > > > > > > > > Slightly unrelated to the patch,
> > > > > > > > > I tried hard to understand this comment below in percpu_down_read() but no dice.
> > > > > > > > > 
> > > > > > > > > I do understand how rcu sync and percpu rwsem works, however the comment
> > > > > > > > > below didn't make much sense to me. For one, there's no readers_fast anymore
> > > > > > > > > so I did not follow what readers_fast means. Could the comment be updated to
> > > > > > > > > reflect latest changes?
> > > > > > > > > Also could you help understand how is a writer not able to change
> > > > > > > > > sem->state and count the per-cpu read counters at the same time as the
> > > > > > > > > comment tries to say?
> > > > > > > > > 
> > > > > > > > > 	/*
> > > > > > > > > 	 * We are in an RCU-sched read-side critical section, so the writer
> > > > > > > > > 	 * cannot both change sem->state from readers_fast and start checking
> > > > > > > > > 	 * counters while we are here. So if we see !sem->state, we know that
> > > > > > > > > 	 * the writer won't be checking until we're past the preempt_enable()
> > > > > > > > > 	 * and that once the synchronize_rcu() is done, the writer will see
> > > > > > > > > 	 * anything we did within this RCU-sched read-size critical section.
> > > > > > > > > 	 */
> > > > > > > > > 
> > > > > > > > > Also,
> > > > > > > > > I guess we could get rid of all of the gp_ops struct stuff now that since all
> > > > > > > > > the callbacks are the same now. I will post that as a follow-up patch to this
> > > > > > > > > series.
> > > > > > > > 
> > > > > > > > Hello, Joel,
> > > > > > > > 
> > > > > > > > Oleg has a set of patches updating this code that just hit mainline
> > > > > > > > this week.  These patches get rid of the code that previously handled
> > > > > > > > RCU's multiple flavors.  Or are you looking at current mainline and
> > > > > > > > me just missing your point?
> > > > > > > > 
> > > > > > > 
> > > > > > > Hi Paul,
> > > > > > > You are right on point. I have a bad habit of not rebasing my trees. In this
> > > > > > > case the feature branch of mine in concern was based on v5.1. Needless to
> > > > > > > say, I need to rebase my tree.
> > > > > > > 
> > > > > > > Yes, this sync clean up patch does conflict when I rebase, but other patches
> > > > > > > rebase just fine.
> > > > > > > 
> > > > > > > The 2 options I see are:
> > > > > > > 1. Let us drop this patch for now and I resend it later.
> > > > > > > 2. I resend all patches based on Linus's master branch.
> > > > > > 
> > > > > > Below is the updated patch based on Linus master branch:
> > > > > > 
> > > > > > ---8<-----------------------
> > > > > > 
> > > > > > >From 5f40c9a07fcf3d6dafc2189599d0ba9443097d0f Mon Sep 17 00:00:00 2001
> > > > > > From: "Joel Fernandes (Google)" <joel@xxxxxxxxxxxxxxxxx>
> > > > > > Date: Fri, 12 Jul 2019 12:13:27 -0400
> > > > > > Subject: [PATCH v2.1 3/9] rcu/sync: Remove custom check for reader-section
> > > > > > 
> > > > > > The rcu/sync code was doing its own check whether we are in a reader
> > > > > > section. With RCU consolidating flavors and the generic helper added in
> > > > > > this series, this is no longer need. We can just use the generic helper
> > > > > > and it results in a nice cleanup.
> > > > > > 
> > > > > > Cc: Oleg Nesterov <oleg@xxxxxxxxxx>
> > > > > > Signed-off-by: Joel Fernandes (Google) <joel@xxxxxxxxxxxxxxxxx>
> > > > > > ---
> > > > > >  include/linux/rcu_sync.h | 4 +---
> > > > > >  1 file changed, 1 insertion(+), 3 deletions(-)
> > > > > > 
> > > > > > diff --git a/include/linux/rcu_sync.h b/include/linux/rcu_sync.h
> > > > > > index 9b83865d24f9..0027d4c8087c 100644
> > > > > > --- a/include/linux/rcu_sync.h
> > > > > > +++ b/include/linux/rcu_sync.h
> > > > > > @@ -31,9 +31,7 @@ struct rcu_sync {
> > > > > >   */
> > > > > >  static inline bool rcu_sync_is_idle(struct rcu_sync *rsp)
> > > > > >  {
> > > > > > -	RCU_LOCKDEP_WARN(!rcu_read_lock_held() &&
> > > > > > -			 !rcu_read_lock_bh_held() &&
> > > > > > -			 !rcu_read_lock_sched_held(),
> > > > > > +	RCU_LOCKDEP_WARN(!rcu_read_lock_any_held(),
> > > > > 
> > > > > I believe that replacing rcu_read_lock_sched_held() with preemptible()
> > > > > in a CONFIG_PREEMPT=n kernel will give you false-positive splats here.
> > > > > If you have not already done so, could you please give it a try?
> > > > 
> > > > Hi Paul,
> > > > I don't think it will cause splats for !CONFIG_PREEMPT.
> > > > 
> > > > Currently, rcu_read_lock_any_held() introduced in this patch returns true if
> > > > !preemptible(). This means that:
> > > > 
> > > > The following expression above:
> > > > RCU_LOCKDEP_WARN(!rcu_read_lock_any_held(),...)
> > > > 
> > > > Becomes:
> > > > RCU_LOCKDEP_WARN(preemptible(), ...)
> > > > 
> > > > For, CONFIG_PREEMPT=n kernels, this means:
> > > > RCU_LOCKDEP_WARN(0, ...)
> > > > 
> > > > Which would mean no splats. Or, did I miss the point?
> > > 
> > > I suggest trying it out on a CONFIG_PREEMPT=n kernel.
> > 
> > Sure, will do, sorry did not try it out yet because was busy with weekend
> > chores but will do soon, thanks!
> 
> I am not faulting you for taking the weekend off, actually.  ;-)

;-) 

I tried doing RCU_LOCKDEP_WARN(preemptible(), ...) in this code path and I
don't get any splats. I also disassembled the code and it seems to me
RCU_LOCKDEP_WARN() becomes a NOOP which also the above reasoning confirms.

thanks,

 - Joel