On Wed, Nov 29, 2017 at 02:07:03PM -0800, Paul E. McKenney wrote:
> On Wed, Nov 29, 2017 at 11:08:19AM -0800, Paul E. McKenney wrote:
> > On Tue, Nov 28, 2017 at 01:08:10PM -0800, Paul E. McKenney wrote:
> > > On Tue, Nov 28, 2017 at 12:46:19PM -0800, Paul E. McKenney wrote:
> > > > On Tue, Nov 28, 2017 at 09:35:54AM -0800, Paul E. McKenney wrote:
> > > > > On Tue, Nov 28, 2017 at 06:10:08PM +0100, Thomas Gleixner wrote:
> > > > > > On Tue, 28 Nov 2017, Paul E. McKenney wrote:
> > > > > > > On Tue, Nov 28, 2017 at 05:47:35PM +0100, Thomas Gleixner wrote:
> > > > > > > diff --git a/kernel/time/timer.c b/kernel/time/timer.c
> > > > > > > index db774b0f217e..a3321bb565db 100644
> > > > > > > --- a/kernel/time/timer.c
> > > > > > > +++ b/kernel/time/timer.c
> > > > > > > @@ -1803,7 +1803,7 @@ signed long __sched schedule_timeout(signed
> > > > > > > long timeout)
> > > > > > > idx = timer_get_idx(&timer.timer);
> > > > > > > idx_now = calc_wheel_index(j, base->clk);
> > > > > > > raw_spin_unlock_irqrestore(&base->lock, flags);
> > > > > > > - pr_info("%s: Waylayed timer idx: %u idx_now: %u\n",
> > > > > > > __func__, idx, idx_now);
> > > > > > > + pr_info("%s: Waylayed timer base->clk: %#lx jiffies:
> > > > > > > %#lx base->next_expiry: %#lx timer->flags: %#x timer->expires
> > > > > > > %#lx idx: %u idx_now: %u\n", __func__, base->clk, j,
> > > > > > > base->next_expiry, timer.timer.flags, timer.timer.expires, idx,
> > > > > > > idx_now);
> > > > > >
> > > > > > Please print idx and idx_now as hex values. It's simpler to decode
> > > > > > that way.
> > > > >
> > > > > Here you go! Starting tests at this end, focusing on TREE01 and
> > > > > TREE04.
> > > > > BTW, TREE04 doesn't do any CPU hotplug, providing a counterexample to
> > > > > my long-held assumption that this only happened in the presence of CPU
> > > > > hotplug operations.
> > > >
> > > > And here is output with changes discussed on IRC. TREE04 managed to
> > > > have not one but two overlapping RCU CPU stall warnings, one for RCU-bh
> > > > and the second for RCU-sched. TREE04 and TREE04. HZ=1000.
> > >
> > > And here is the full patch, in all its lack of aesthetic appeal.
> >
> > And here is the list of waylaid timers from last night's testing. The big
> > pile of them from TREE01 at the end is due to wakeups from kthread_stop(),
> > I am guessing. The TREE04 run only had two of them, but they seem reliable
> > enough that I just might be able to bisect. I will try that.
>
> And it converged to 5c4991e24c69 ("sched/isolation: Split out new
> CONFIG_CPU_ISOLATION=y config from CONFIG_NO_HZ_FULL"), which is a bit
> hard to believe. Please see below for the log. I will be retesting
> some of the allegedly good commits, just in case.
And the bisection really did converge here. It appears that splitting
CONFIG_CPU_ISOLATION from CONFIG_NO_HZ_FULL isn't fully cooked.
Adding CONFIG_CPU_ISOLATION=y to my tests causes them to pass, but I
get the eternal wait for a three-jiffy timeout given CONFIG_NO_HZ_FULL=y
and CONFIG_CPU_ISOLATION=n. Both cases use CONFIG_NO_HZ_FULL_ALL=y.
Adding Frederic for his perspective.
Thanx, Paul
