On Wed, Jul 26, 2017 at 03:45:40PM -0700, David Miller wrote:

From: "Paul E. McKenney" <redacted>
Date: Wed, 26 Jul 2017 15:36:58 -0700

quoted

And without CONFIG_SOFTLOCKUP_DETECTOR, I see five runs of 24 with RCU
CPU stall warnings.  So it seems likely that CONFIG_SOFTLOCKUP_DETECTOR
really is having an effect.

Thanks for all of the info Paul, I'll digest this and scan over the
code myself.

Just out of curiousity, what x86 idle method is your machine using?
The mwait one or the one which simply uses 'halt'?  The mwait variant
might mask this bug, and halt would be a lot closer to how sparc64 and
Jonathan's system operates.

My kernel builds with CONFIG_INTEL_IDLE=n, which I believe means that
I am not using the mwait one.  Here is a grep for IDLE in my .config:

	CONFIG_NO_HZ_IDLE=y
	CONFIG_GENERIC_SMP_IDLE_THREAD=y
	# CONFIG_IDLE_PAGE_TRACKING is not set
	CONFIG_ACPI_PROCESSOR_IDLE=y
	CONFIG_CPU_IDLE=y
	# CONFIG_CPU_IDLE_GOV_LADDER is not set
	CONFIG_CPU_IDLE_GOV_MENU=y
	# CONFIG_ARCH_NEEDS_CPU_IDLE_COUPLED is not set
	# CONFIG_INTEL_IDLE is not set

On sparc64 the cpu yield we do in the idle loop sleeps the cpu.  It's
local TICK register keeps advancing, and the local timer therefore
will still trigger.  Also, any externally generated interrupts
(including cross calls) will wake up the cpu as well.

The tick-sched code is really tricky wrt. NO_HZ even in the NO_HZ_IDLE
case.  One of my running theories is that we miss scheduling a tick
due to a race.  That would be consistent with the behavior we see
in the RCU dumps, I think.

But wouldn't you have to miss a -lot- of ticks to get an RCU CPU stall
warning?  By default, your grace period needs to extend for more than
21 seconds (more than one-third of a -minute-) to get one.  Or do
you mean that the ticks get shut off now and forever, as opposed to
just losing one of them?

Anyways, just a theory, and that's why I keep mentioning that commit
about the revert of the revert (specifically
411fe24e6b7c283c3a1911450cdba6dd3aaea56e).

:-)

I am running an overnight test in preparation for attempting to push
some fixes for regressions into 4.12, but will try reverting this
and enabling CONFIG_HZ_PERIODIC tomorrow.

Jonathan, might the commit that Dave points out above be what reduces
the probability of occurrence as you test older releases?

							Thanx, Paul