Re: [PATCH v2] rcu/segcblist: Add debug checks for segment lengths

[Joel Fernandes <joel@joelfernandes.org>]

On Thu, Nov 19, 2020 at 12:16:15PM -0800, Paul E. McKenney wrote:
> On Thu, Nov 19, 2020 at 02:44:35PM -0500, Joel Fernandes wrote:
> > On Thu, Nov 19, 2020 at 2:22 PM Paul E. McKenney <paulmck@kernel.org> wrote:
> > > > > > > On Wed, Nov 18, 2020 at 11:15:41AM -0500, Joel Fernandes (Google) wrote:
> > > > > > > > After rcu_do_batch(), add a check for whether the seglen counts went to
> > > > > > > > zero if the list was indeed empty.
> > > > > > > >
> > > > > > > > Signed-off-by: Joel Fernandes (Google) <joel@joelfernandes.org>
> > > > > > >
> > > > > > > Queued for testing and further review, thank you!
> > > > > >
> > > > > > FYI, the second of the two checks triggered in all four one-hour runs of
> > > > > > TREE01, all four one-hour runs of TREE04, and one of the four one-hour
> > > > > > runs of TREE07.  This one:
> > > > > >
> > > > > > WARN_ON_ONCE(count != 0 && rcu_segcblist_n_segment_cbs(&rdp->cblist) == 0);
> > > > > >
> > > > > > That is, there are callbacks in the list, but the sum of the segment
> > > > > > counts is nevertheless zero.  The ->nocb_lock is held.
> > > > > >
> > > > > > Thoughts?
> > > > >
> > > > > FWIW, TREE01 reproduces it very quickly compared to the other two
> > > > > scenarios, on all four run, within five minutes.
> > > >
> > > > So far for TREE01, I traced it down to an rcu_barrier happening so it could
> > > > be related to some interaction with rcu_barrier() (Just a guess).
> > >
> > > Well, rcu_barrier() and srcu_barrier() are the only users of
> > > rcu_segcblist_entrain(), if that helps.  Your modification to that
> > > function looks plausible to me, but the system's opinion always overrules
> > > mine.  ;-)
> > 
> > Right. Does anything the bypass code standout? That happens during
> > rcu_barrier() as well, and it messes with the lengths.
> 
> In theory, rcu_barrier_func() flushes the bypass before doing the
> entrain, and does the rcu_segcblist_entrain() afterwards.
> 
> Ah, and that is the issue.  If ->cblist is empty and ->nocb_bypass
> is not, then ->cblist length will be nonzero, and none of the
> segments will be nonzero.
> 
> So you need something like this for that second WARN, correct?
> 
> 	WARN_ON_ONCE(!rcu_segcblist_empty(&rdp->cblist) &&
> 		     rcu_segcblist_n_segment_cbs(&rdp->cblist) == 0);
> 
> This is off the cuff, so should be taken with a grain of salt.  And
> there might well be other similar issues.

Ah, makes sense. Or maybe should be made like the other warning?
        WARN_ON_ONCE(!IS_ENABLED(CONFIG_RCU_NOCB_CPU) && count != 0
		&& rcu_segcblist_n_segment_cbs(&rdp->cblist) == 0);

Though your warning is better.

I will try these out and see if it goes away. I am afraid though that there
is an issue with !NOCB code since you had other configs that were failing
similarly.. :-\.

thanks, :-)

 - Joel