Re: [PATCH v2] tracing: Make sure RCU is watching before calling a stack trace

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

On Thu, May 18, 2017 at 03:30:26PM -0400, Steven Rostedt wrote:
> On Thu, 18 May 2017 11:45:28 -0700
> "Paul E. McKenney" <paulmck@linux.vnet.ibm.com> wrote:
> 
> > On Thu, May 18, 2017 at 12:39:14PM -0400, Steven Rostedt wrote:
> > > On Thu, 18 May 2017 08:47:11 -0700
> > > "Paul E. McKenney" <paulmck@linux.vnet.ibm.com> wrote:
> > >   
> > > > On Thu, May 18, 2017 at 09:38:09AM -0400, Steven Rostedt wrote:  
> > > > > 
> > > > > From: "Steven Rostedt (VMware)" <rostedt@goodmis.org>
> > > > > 
> > > > > As stack tracing now requires "rcu watching", force RCU to be watching when
> > > > > recording a stack trace.
> > > > > 
> > > > > Link: http://lkml.kernel.org/r/20170512172449.879684501@goodmis.org
> > > > > 
> > > > > Cc: "Paul E. McKenney" <paulmck@linux.vnet.ibm.com>
> > > > > Signed-off-by: Steven Rostedt (VMware) <rostedt@goodmis.org>
> > > > > ---
> > > > > 
> > > > > Changes since v1:
> > > > > 
> > > > >    My testing discovered that the stack trace can be called with
> > > > >    interrupts enabled, which is a no no to have when calling
> > > > >    rcu_irq_enter(). When interrupts are enabled, as with being in an
> > > > >    NMI, RCU will also be watching.
> > > > > 
> > > > > kernel/trace/trace.c | 26 +++++++++++++++++++++++++-
> > > > >  1 file changed, 25 insertions(+), 1 deletion(-)
> > > > > 
> > > > > diff --git a/kernel/trace/trace.c b/kernel/trace/trace.c
> > > > > index fcc9a2d..34a98ba 100644
> > > > > --- a/kernel/trace/trace.c
> > > > > +++ b/kernel/trace/trace.c
> > > > > @@ -2568,7 +2568,31 @@ static inline void ftrace_trace_stack(struct trace_array *tr,
> > > > >  void __trace_stack(struct trace_array *tr, unsigned long flags, int skip,
> > > > >  		   int pc)
> > > > >  {
> > > > > -	__ftrace_trace_stack(tr->trace_buffer.buffer, flags, skip, pc, NULL);
> > > > > +	struct ring_buffer *buffer = tr->trace_buffer.buffer;
> > > > > +
> > > > > +	/*
> > > > > +	 * When an NMI triggers, RCU is enabled via rcu_nmi_enter()
> > > > > +	 * Also, RCU is always enabled when interrupts are.
> > > > > +	 */
> > > > > +	if (!irqs_disabled() || in_nmi()) {    
> > > > 
> > > > You lost me on this one.  RCU might not be watching if irqs are
> > > > enabled, for example, in the idle loop.  What am I missing here?
> > > >   
> > > 
> > > Hmm, no, maybe I'm missing something :-/
> > > 
> > > OK, so if we trace in the idle loop, rcu may not be watching, so I may
> > > need to byte the bullet and use the rcu_irq_enter_irqon() instead :-(  
> > 
> > If rcu_is_watching() returns false, yes.  Of course, if rcu_is_watching()
> > returns true, you are golden.
> > 
> 
> Hmm, so I can just replace the entire if with:
> 
> 	if (rcu_is_watching())
> 
> because it should be watching in NMI context, correct?

Yes, and that would be even better!  ;-)

However, this assumes that there is no possibility of this being called
during the time that in_nmi() returns true but RCU is not aware of the
NMI.  (Yes, we did go through this a bit earlier, just my usual paranoia
showing...)

							Thanx, Paul