Re: [RFC PATCH 2/2] rcu,debug_core: allow the kernel debugger to reset the rcu stall timer

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

On Mon, Aug 09, 2010 at 01:26:19PM -0500, Jason Wessel wrote:
> On 08/09/2010 12:43 PM, Paul E. McKenney wrote:
> > On Mon, Aug 09, 2010 at 12:12:12AM -0500, Jason Wessel wrote:
> >
> >> +void rcu_cpu_stall_reset(void)
> >> +{
> >> +    rcu_sched_state.jiffies_stall = 0;
> >> +    rcu_bh_state.jiffies_stall = 0;
> >> +}
> >> +
> >
> > OK, so you are suppressing RCU CPU stall warnings for rcu_sched and
> > rcu_bh, but not for preemptible RCU.  I believe that you want all of
> > them covered.
> 
> What is the state variable for the preemptible RCU I had not hit a
> warning in my testing so I must needs some more test cases. :-)

Well, you won't hit preemptible RCU unless you set TREE_PREEMPT_RCU.  ;-)

> > I have a number of recent patches that allow RCU CPU stall warnings to
> > be suppressed, one of which allows them to be suppressed using sysfs.
> > Would that work for you, or do you need an in-kernel interface?
> 
> We need an in-kernel interface for sure.

OK, good to know.

> > If you do need an in-kernel interface, I could export (and probably
> > rename) rcu_panic(), which is a static in 2.6.35.  This assumes that you
> > never want to re-enable RCU CPU stall warnings once you suppress them,
> > which is what your patch appears to do.
> >
> > So, if I export a suppress_rcu_cpu_stall() function that permanently
> > disabled RCU CPU stall warnings, would that work for you?  (They could
> > be manually re-enabled via sysfs.)
> 
> This is an RFC patch for a reason.  The intent behind the interface is
> to allow for one stall check cycle to go by after resuming kernel
> execution and after that the normal rules are in play.  Code flow
> wise, it looked like the easiest thing to do was set the jiffies_stall
> value to zero and then exit when the.  The patch I created was
> supposed to only ignore one stall cycle.
> 
> Here is the pseudo code.
> 
> /* before restarting kernel execution zero out the jiffies_stall  value.
> 
> __rcu_pending() {
> 
>     check_cpu_stall();  <- Here we check if the stall val is set to zero
> and just return
>     /* do all normal work */
> 
> }
> 
> In the normal flow of things rc_start_gp() will ultimately call
> record_gp_stall_check_time which updates the jiffies_stall back to non
> zero and the stall accounting is back in play.

Ah, I get it now.  Just out of curiosity, why not set the various
->jiffies_stall fields to jiffies + RCU_SECONDS_TILL_STALL_CHECK?
Is the value of jiffies likely to advance a lot after you call
rcu_cpu_stall_reset(), perhaps due to the system trying to catch up with
the passage of time?

							Thanx, Paul