Re: Markers: multi-probe locking fun (was: Re: [PATCH 0/2] Markers Implementation for RCU Tracing - Ver II)

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

On Tue, Feb 19, 2008 at 03:33:26PM -0500, Mathieu Desnoyers wrote:
> * Jan Kiszka (jan.kiszka@siemens.com) wrote:
> > Paul E. McKenney wrote:
> > > On Mon, Feb 18, 2008 at 01:47:31PM +0100, Jan Kiszka wrote:
> > >> K. Prasad wrote:
> > >>> Hi Ingo,
> > >>> 	Please accept these patches into the rt tree which convert the
> > >>> existing RCU tracing mechanism for Preempt RCU and RCU Boost into
> > >>> markers.
> > >>>  
> > >>> These patches are based upon the 2.6.24-rc5-rt1 kernel tree.
> > >>>  
> > >>> Along with marker transition, the RCU Tracing infrastructure has also
> > >>> been modularised to be built as a kernel module, thereby enabling
> > >>> runtime changes to the RCU Tracing infrastructure.
> > >>>  
> > >>> Patch [1/2] - Patch that converts the Preempt RCU tracing in
> > >>> rcupreempt.c into markers.
> > >>>  
> > >>> Patch [1/2] - Patch that converts the Preempt RCU Boost tracing in
> > >>> rcupreempt-boost.c into markers.
> > >>>  
> > >> I have a technical problem with marker-based RCU tracing: It causes
> > >> nasty recursions with latest multi-probe marker patches (sorry, no link
> > >> at hand, can be found in latest LTTng, maybe also already in -mm). Those
> > >> patches introduce a marker probe trampoline like this:
> > >>
> > >> void marker_probe_cb(const struct marker *mdata, void *call_private,
> > >> 	const char *fmt, ...)
> > >> {
> > >> 	va_list args;
> > >> 	char ptype;
> > >>
> > >> 	/*
> > >> 	 * rcu_read_lock does two things : disabling preemption to make sure the
> > >> 	 * teardown of the callbacks can be done correctly when they are in
> > >> 	 * modules and they insure RCU read coherency.
> > >> 	 */
> > >> 	rcu_read_lock();
> > >> 	preempt_disable();
> > >> 	...
> > >>
> > >> Can we do multi-probe with pure preempt_disable/enable protection? I
> > >> guess it's fine with classic RCU, but what about preemptible RCU? Any
> > >> suggestion appreciated!
> > > 
> > > If you substitute synchronize_sched() for synchronize_rcu(), this should
> > > work fine.  Of course, this approach would cause RCU tracing to degrade
> > > latencies somewhat in -rt.
> > > 
> > > If tracing is using call_rcu(), we will need to add a call_sched()
> > > or some such.
> > 
> > You mean something like "#define call_sched call_rcu_classic"?
> > 
> > I just learned that there is another reason for killing
> > rcu_read_lock&friends from the marker probes: It can deadlock on -rt
> > with PREEMPT_RCU_BOOST (hit probe inside rq-lock protected region =>
> > rcu_read_unlock triggers unboost => stuck on rq_lock :( ).
> > 
> 
> Yep, ok, let's do this :
> 
> in include/linux/rcupdate.h
> 
> #ifndef PREEMPT_RT
> #define call_sched call_rcu
> #else
> #define call_sched call_rcu_classic
> #endif
> 
> And I'll adapt the markers accordingly.

Good point, this will indeed work for 2.6.24-rt1!

Will need to do a bit more for 2.6.25-rc1.  My current thought is to
provide a kernel thread that loops over the CPUs, advancing/invoking
per-CPU lists of callbacks as it does so.  Then call_sched() would simply
enqueue its callback on the current CPU's next list.

							Thanx, Paul