Re: [PATCH tip/core/rcu 17/86] rcu: fix boost-tracing bug and update tracing documentation

["Paul E. McKenney" <paulmck@linux.vnet.ibm.com>]

On Mon, May 02, 2011 at 09:18:24AM -0700, Paul E. McKenney wrote:
> On Mon, May 02, 2011 at 01:33:01AM -0700, Paul E. McKenney wrote:
> > On Sun, May 01, 2011 at 08:43:10AM -0700, Josh Triplett wrote:
> > > The bugfix here should really go in a separate commit from the updated
> > > documentation.  Also:
> > > 
> > > On Sun, May 01, 2011 at 06:20:57AM -0700, Paul E. McKenney wrote:
> > > > --- a/Documentation/RCU/trace.txt
> > > > +++ b/Documentation/RCU/trace.txt
> > > > @@ -10,34 +10,37 @@ for rcutree and next for rcutiny.
> > > >  
> > > >  CONFIG_TREE_RCU and CONFIG_TREE_PREEMPT_RCU debugfs Files and Formats
> > > >  
> > > > -These implementations of RCU provides five debugfs files under the
> > > > +These implementations of RCU provides seven debugfs files under the
> > > 
> > > s/seven/several/, so this doesn't continue to need updates as you add
> > > more tracing?  It doesn't seem entirely accurate even now, given that
> > > the set of files available now depends on CONFIG_RCU_BOOST.
> > 
> > Good point, will fix.
> > 
> > > >  top-level directory RCU: rcu/rcudata (which displays fields in struct
> > > 
> > > RCU or rcu?
> > 
> > Indeed, "rcu".  Will fix.
> > 
> > > >  rcu_data), rcu/rcudata.csv (which is a .csv spreadsheet version of
> > > >  rcu/rcudata), rcu/rcugp (which displays grace-period counters),
> > > > -rcu/rcuhier (which displays the struct rcu_node hierarchy), and
> > > > +rcu/rcuhier (which displays the struct rcu_node hierarchy),
> > > >  rcu/rcu_pending (which displays counts of the reasons that the
> > > > -rcu_pending() function decided that there was core RCU work to do).
> > > > +rcu_pending() function decided that there was core RCU work to do),
> > > > +rcu/rcutorture (which displays rcutorture test progress), and, if the
> > > > +kernel is built with CONFIG_RCU_BOOST, rcu/rcuboost (which displays RCU
> > > > +boosting statistics).
> > > 
> > > I think this would work better as a list:
> > > 
> > > rcu/rcudata:
> > >     Displays fields in struct rcu_data.
> > > rcu/rcudata.csv:
> > >     Comma-separated values spreadsheet version of rcudata.
> > > rcu/rcugp:
> > >     Displays grace-period counters.
> > > rcu/rcuhier
> > >     Displays the struct rcu_node hierarchy.
> > > rcu/rcu_pending:
> > >     Displays counts of the reasons rcu_pending() decided that RCU had
> > >     work to do.
> > > rcu/rcutorture:
> > >     Displays rcutorture test progress.
> > > rcu/rcuboost
> > >     Displays RCU boosting statistics. Only present if
> > >     CONFIG_RCU_BOOST=y.
> > > 
> > > Feel free to put the descriptions on the same lines as the filenames if
> > > you prefer; however, I thought this style worked better for the
> > > descriptions longer than one line.
> > 
> > Makes sense, will do a list.
> > 
> > > Also, does rcu/rcutorture appear with rcutorture not loaded?  Hopefully
> > > not, since rcutorture shouldn't take up any additional resources when
> > > not loaded.
> > 
> > Not much choice, unfortunately.  If I put the info into rcutorture, then
> > I lose it when rcutorture is unloaded.  This is problematic in automated
> > tests that repeatedly load and unload rcutorture.
> > 
> > But the amount of data is small, and I set things up so that TINY_RCU
> > can omit this.  TREE_RCU is big enough that it doesn't really matter.
> > 
> > > > --- a/kernel/rcutree_plugin.h
> > > > +++ b/kernel/rcutree_plugin.h
> > > > @@ -1086,7 +1086,7 @@ static void rcu_initiate_boost_trace(struct rcu_node *rnp)
> > > >  	else if (rnp->gp_tasks != NULL && rnp->qsmask != 0)
> > > >  		rnp->n_balk_notblocked++;
> > > >  	else if (rnp->gp_tasks != NULL &&
> > > > -		 ULONG_CMP_GE(jiffies, rnp->boost_time))
> > > > +		 ULONG_CMP_LT(jiffies, rnp->boost_time))
> > > >  		rnp->n_balk_notyet++;
> > > >  	else
> > > >  		rnp->n_balk_nos++;
> > > 
> > > This bugfix should go in a separate commit.
> > 
> > Excellent point, will fix.
> 
> Here is the updated documentation update.

And here is the bug fix.

							Thanx, Paul

------------------------------------------------------------------------

rcu: fix tracing bug thinko on boost-balk attribution

The rcu_initiate_boost_trace() function mis-attributed refusals to
initiate RCU priority boosting that were in fact due to its not yet
being time to boost.  This patch fixes the faulty comparison.

Signed-off-by: Paul E. McKenney <paul.mckenney@linaro.org>
Signed-off-by: Paul E. McKenney <paulmck@linux.vnet.ibm.com>

diff --git a/kernel/rcutree_plugin.h b/kernel/rcutree_plugin.h
index 22a6a46..a21413d 100644
--- a/kernel/rcutree_plugin.h
+++ b/kernel/rcutree_plugin.h
@@ -1086,7 +1086,7 @@ static void rcu_initiate_boost_trace(struct rcu_node *rnp)
 	else if (rnp->gp_tasks != NULL && rnp->qsmask != 0)
 		rnp->n_balk_notblocked++;
 	else if (rnp->gp_tasks != NULL &&
-		 ULONG_CMP_GE(jiffies, rnp->boost_time))
+		 ULONG_CMP_LT(jiffies, rnp->boost_time))
 		rnp->n_balk_notyet++;
 	else
 		rnp->n_balk_nos++;