Re: [PATCH 08/11 v2] nohz: Allow rcu extended quiescent state handling seperately from tick stop

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

On Sun, Oct 16, 2011 at 03:28:08PM +0200, Frederic Weisbecker wrote:
> On Fri, Oct 14, 2011 at 10:00:19AM -0700, Paul E. McKenney wrote:
> > On Fri, Oct 14, 2011 at 02:08:36PM +0200, Frederic Weisbecker wrote:
> > > On Thu, Oct 13, 2011 at 03:51:36PM -0700, Paul E. McKenney wrote:
> > > > On Thu, Oct 13, 2011 at 02:50:20PM +0200, Frederic Weisbecker wrote:
> > > > > On Thu, Oct 13, 2011 at 12:03:57AM -0700, Paul E. McKenney wrote:
> > > > > > On Wed, Oct 12, 2011 at 11:57:52PM -0700, Paul E. McKenney wrote:
> > > > > > > On Sat, Oct 08, 2011 at 04:01:00PM +0200, Frederic Weisbecker wrote:
> > > > > > > > It is assumed that rcu won't be used once we switch to tickless
> > > > > > > > mode and until we restart the tick. However this is not always
> > > > > > > > true, as in x86-64 where we dereference the idle notifiers after
> > > > > > > > the tick is stopped.
> > > > > > > > 
> > > > > > > > To prepare for fixing this, add two new APIs:
> > > > > > > > tick_nohz_idle_enter_norcu() and tick_nohz_idle_exit_norcu().
> > > > > > > > 
> > > > > > > > If no use of RCU is made in the idle loop between
> > > > > > > > tick_nohz_enter_idle() and tick_nohz_exit_idle() calls, the arch
> > > > > > > > must instead call the new *_norcu() version such that the arch doesn't
> > > > > > > > need to call rcu_idle_enter() and rcu_idle_exit().
> > > > > > > > 
> > > > > > > > Otherwise the arch must call tick_nohz_enter_idle() and
> > > > > > > > tick_nohz_exit_idle() and also call explicitly:
> > > > > > > > 
> > > > > > > > - rcu_idle_enter() after its last use of RCU before the CPU is put
> > > > > > > > to sleep.
> > > > > > > > - rcu_idle_exit() before the first use of RCU after the CPU is woken
> > > > > > > > up.
> > > > > > > 
> > > > > > > Thank you, Frederic!  I have queued this to replace the earlier
> > > > > > > version.  The set is available on branch rcu/dyntick of
> > > > > > > 
> > > > > > > 	https://github.com/paulmckrcu/linux
> > > > > > 
> > > > > > Which reminds me...  About the ultimate objective, getting tick-free
> > > > > > operation.  (Or, for the guys who want to eliminate the tick entirely,
> > > > > > shutting up the hrtimer stuff that they want to replace it with.)
> > > > > > 
> > > > > > I believe that you will then need to have two levels of not-in-dynticks
> > > > > > for processes, one for idle vs. not and another for when a process
> > > > > > switches from user-space to kernel execution.  Correct, or am I
> > > > > > confused?
> > > > > > 
> > > > > > The reason I ask is that commit e11f5981 currently only allows one
> > > > > > level of not-in-dynticks for processes.  It is easy to add another
> > > > > > level, but thought I should check beforehand.
> > > > > 
> > > > > Hmm, yeah looking at that patch, it's going to be hard to have a nesting
> > > > > that looks like:
> > > > > 
> > > > > rcu_irq_enter();
> > > > > rcu_user_enter();
> > > > > rcu_irq_exit(); <-- with effective extended quiescent state starting there
> > > > 
> > > > OK, so the idea here is that there has been two runnable processes on
> > > > the current CPU, but during the irq handler one of them moves or some
> > > > such?
> > > 
> > > No it happens when we have an irq in userspace and we stop the tick
> > > from that irq. Noticing we are in userspace, we want to be in extended
> > > quiescent state when we resume from the interrupt to userspace.
> > 
> > Ah, OK!
> > 
> > > > If so, how about a rcu_user_enter_fromirq() that sets the counter
> > > > to 1 so that the rcu_irq_exit() cleans up properly?  If need be, I could
> > > > of course provide an argument to allow you to specify the count offset.
> > > 
> > > Yeah I think that should work.
> > 
> > Very good.  I will start off with no argument, easy enough to add it
> > later if needed.
> 
> Yeah, but for now, this is out of tree code. The functions seem quite trivial
> to write so I guess I can sort it out in my tree when I'll rebase on 3.2
> (if the current queue goes to the next merge window)

I will be pushing rcu/next to get -next exposure later this week (once
I reach a geography that permits github access), and if all goes well...

> > > > > I also need to be able to call rcu_user_enter() from non-irq path.
> > > > 
> > > > Then rcu_user_enter_fromirq() would be for the irq path and
> > > > rcu_user_enter() from the non-irq path.
> > > > 
> > > > Would that work for you?
> > > 
> > > Yep!
> > 
> > Very good, I will take a whack at it.  BTW, testing is going quite
> > well thus far with your current patches combined with my paranoid
> > idle-count approach.  One test in particular that previously failed
> > reliably within minutes just successfully completed a ten-hour run.
> > So things are looking up!  (Famous last words...)
> 
> Great, I cross my fingers! :)

Good results for that one.  Here's hoping for continued test passing.

> > > > > I don't truly understand the problem of the usermode helpers that
> > > > > mess up the dynticks counts. May be we can somewhow fix it differently
> > > > > from the offending callsite?
> > > > 
> > > > I tried a few approaches along these lines, but there were way too
> > > > many opportunities for interruption and preemption along the way.
> > > > The problem is that unless the fixup happens under a no-preempt
> > > > region of code that includes the rcu_irq_enter() or rcu_irq_exit()
> > > > call (as the case may be), then you end up messing up the idle-depth
> > > > count of two CPUs rather than just one.  :-(
> > > > 
> > > > But maybe I am missing something -- suggestions more than welcome!
> > > 
> > > It's rather me missing everything :)
> > > It happens when we call call_usermodehelper()? If so how? We have a
> > > call to rcu_irq_enter() that lacks an rcu_irq_exit() ?
> > 
> > On powerpc, it executes the "sc" ("system call") instruction from
> > kernel mode, which results in an exception.  But from what I can see,
> > there is no corresponding return from exception, so my not-so-paranoid
> > counting scheme would lose count.  That said, please keep in mind that
> > I in no way fully understand that code.  It is also far from clear to
> > me why my earlier dyntick-idle code worked in this situation -- perhaps
> > the value of preempt_count() gets fixed up somehow -- I haven't really
> > studied all the assembly language involved in detail, so there is lots
> > of opportunity for such a fixup somewhere.
> > 
> > You asked!  ;-)
> 
> Haha! ;)

;-)

							Thanx, Paul