On Tue, Jul 25, 2017 at 10:42:45PM +0800, Jonathan Cameron wrote:

On Tue, 25 Jul 2017 06:46:26 -0700
"Paul E. McKenney" [off-list ref] wrote:

quoted

On Tue, Jul 25, 2017 at 10:26:54PM +1000, Nicholas Piggin wrote:

quoted

On Tue, 25 Jul 2017 19:32:10 +0800
Jonathan Cameron [off-list ref] wrote:
  

quoted

Hi All,

We observed a regression on our d05 boards (but curiously not
the fairly similar but single socket / smaller core count
d03), initially seen with linux-next prior to the merge window
and still present in v4.13-rc2.

The symptom is:  

Adding Dave Miller and the sparclinux@vger.kernel.org email on CC, as
they have been seeing something similar, and you might well have saved
them the trouble of bisecting.

[ . . . ]

quoted

quoted

[ 1984.628602] rcu_preempt kthread starved for 5663 jiffies! g1566 c1565 f0x0 RCU_GP_WAIT_FQS(3) ->state=0x1  

This is the cause from an RCU perspective.  You had a lot of idle CPUs,
and RCU is not permitted to disturb them -- the battery-powered embedded
guys get very annoyed by that sort of thing.  What happens instead is
that each CPU updates a per-CPU state variable when entering or exiting
idle, and the grace-period kthread ("rcu_preempt kthread" in the above
message) checks these state variables, and if when sees an idle CPU,
it reports a quiescent state on that CPU's behalf.

But the grace-period kthread can only do this work if it gets a chance
to run.  And the message above says that this kthread hasn't had a chance
to run for a full 5,663 jiffies.  For completeness, the "g1566 c1565"
says that grace period #1566 is in progress, the "f0x0" says that no one
is needing another grace period #1567.  The "RCU_GP_WAIT_FQS(3)" says
that the grace-period kthread has fully initialized the current grace
period and is sleeping for a few jiffies waiting to scan for idle tasks.
Finally, the "->state=0x1" says that the grace-period kthread is in
TASK_INTERRUPTIBLE state, in other words, still sleeping.

Thanks for the explanation!

quoted

So my first question is "What did commit 05a4a9527 (kernel/watchdog:
split up config options) do to prevent the grace-period kthread from
getting a chance to run?" 

As far as we can tell it was a side effect of that patch.

The real cause is that patch changed the result of defconfigs to stop running
the softlockup detector - now CONFIG_SOFTLOCKUP_DETECTOR

Enabling that on 4.13-rc2 (and presumably everything in between)
means we don't see the problem any more.

quoted

I must confess that I don't see anything
obvious in that commit, so my second question is "Are we sure that
reverting this commit makes the problem go away?"

Simply enabling CONFIG_SOFTLOCKUP_DETECTOR seems to make it go away.
That detector fires up a thread on every cpu, which may be relevant.

Interesting...  Why should it be necessary to fire up a thread on every
CPU in order to make sure that RCU's grace-period kthreads get some
CPU time?  Especially give how many idle CPUs you had on your system.

So I have to ask if there is some other bug that the softlockup detector
is masking.

quoted

and my third is "Is
this an intermittent problem that led to a false bisection?"

Whilst it is a bit slow to occur, we verified with long runs on either
side of that patch and since with the option enabled on latest mainline.

Also can cause the issue before that patch by disabling the previous
relevant option on 4.12.

OK, thank you -- hard to argue with that!  ;-)

Except that I am still puzzled as to why per-CPU softlockup threads
are needed for RCU's kthreads to get their wakeups.  We really should
be able to disable softlockup and still have kthreads get wakeups and
access to CPU, after all.

quoted

[ . . . ]

quoted

quoted

Reducing the RCU CPU stall timeout makes it happen more often,
but we are seeing even with the default value of 24 seconds.

Tends to occur after a period or relatively low usage, but has
also been seen mid way through performance tests.

This was not seen with v4.12 so a bisection run later lead to
commit 05a4a9527 (kernel/watchdog: split up config options).

Which was odd until we discovered that a side effect of this patch
was to change whether the softlockup detector was enabled or not in
the arm64 defconfig.

On 4.13-rc2 enabling the softlockup detector indeed stopped us
seeing the rcu issue. Disabling the equivalent on 4.12 made the
issue occur there as well.

Clearly the softlockup detector results in a thread on every cpu,
which might be related but beyond that we are still looking into
the issue.

So the obvious question is whether anyone else is seeing this as
it might help us to focus in on where to look!  

Huh. Something similar has been seen very intermittently on powerpc
as well. We couldn't reproduce it reliably to bisect it already, so
this is a good help.

http://marc.info/?l=linuxppc-embedded&m=149872815523646&w=2

It looks like the watchdog patch has a similar effect on powerpc in
that it stops enabling the softlockup detector by default. Haven't
confirmed, but it looks like the same thing.

A bug in RCU stall detection?  

Well, if I am expected to make grace periods complete when my grace-period
kthreads aren't getting any CPU time, I will have to make some substantial
changes.  ;-)

One possibility is that the timer isn't firing and another is that the
timer's wakeup is being lost somehow.

So another thing to try is to boot with rcutree.rcu_kick_kthreads=1.
This will cause RCU to do redundant wakeups on the grace-period kthread
if the grace period is moving slowly.  This is of course a crude hack,
which is why this boot parameter will also cause a splat if it ever has
to do anything.

Running that now will let you know how it goes.  Not seen the issue yet
but might just be a 'lucky' run - will give it a few hours.

Thank you very much!

							Thanx, Paul

Jonathan

quoted

Does this help at all?

							Thanx, Paul

quoted

quoted

In the meantime we'll carry on digging.

Thanks,

Jonathan

p.s. As a more general question, do we want to have the
soft lockup detector enabledon arm64 by default?  

I've cc'ed Don. My patch should not have changed defconfigs, I
should have been more careful with that.

Thanks,
Nick