On Wed, 16 Aug 2017 09:27:31 -0700
"Paul E. McKenney" [off-list ref] wrote:

On Wed, Aug 16, 2017 at 05:56:17AM -0700, Paul E. McKenney wrote:

Thomas, John, am I misinterpreting the timer trace event messages?

So I did some digging, and what you find is that rcu_sched seems to do a
simple scheudle_timeout(1) and just goes out to lunch for many seconds.
The process_timeout timer never fires (when it finally does wake after
one of these events, it usually removes the timer with del_timer_sync).

So this patch seems to fix it. Testing, comments welcome.

Thanks,
Nick

[PATCH] timers: Fix excessive granularity of new timers after a nohz idle

When a timer base is idle, it is forwarded when a new timer is added to
ensure that granularity does not become excessive. When not idle, the
timer tick is expected to increment the base.

However there is a window after a timer is restarted from nohz, when it
is marked not-idle, and before the timer tick on this CPU, where a timer
may be added on an ancient base that does not get forwarded (beacause
the timer appears not-idle).

This results in excessive granularity. So much so that a 1 jiffy timeout
has blown out to 10s of seconds and triggered the RCU stall warning
detector.

Fix this by always forwarding the base when adding a new timer if it is
more than 1 jiffy behind. Another approach I looked at first was to note
if the base was idle but not yet run or forwarded, however this just
seemed to add more branches and complexity when it seems we can just
cover it with this test.

Also add a comment noting a case where we could get an unexpectedly
large granularity for a timer. I debugged this problem by adding
warnings for such cases, but it seems we can't add them in general due
to this corner case.

Signed-off-by: Nicholas Piggin <npiggin@gmail.com>
---
 kernel/time/timer.c | 13 ++++++++++---
 1 file changed, 10 insertions(+), 3 deletions(-)

diff --git a/kernel/time/timer.c b/kernel/time/timer.c
index 8f5d1bf18854..8f69b3105b8f 100644
--- a/kernel/time/timer.c
+++ b/kernel/time/timer.c

@@ -859,10 +859,10 @@ static inline void forward_timer_base(struct timer_base *base)
 	unsigned long jnow = READ_ONCE(jiffies);
 
 	/*
-	 * We only forward the base when it's idle and we have a delta between
-	 * base clock and jiffies.
+	 * We only forward the base when we have a delta between base clock
+	 * and jiffies. In the common case, run_timers will take care of it.
 	 */
-	if (!base->is_idle || (long) (jnow - base->clk) < 2)
+	if ((long) (jnow - base->clk) < 2)
 		return;
 
 	/*

@@ -938,6 +938,13 @@ __mod_timer(struct timer_list *timer, unsigned long expires, bool pending_only)
 	 * same array bucket then just return:
 	 */
 	if (timer_pending(timer)) {
+		/*
+		 * The downside of this optimization is that it can result in
+		 * larger granularity than you would get from adding a new
+		 * timer with this expiry. Would a timer flag for networking
+		 * be appropriate, then we can try to keep expiry of general
+		 * timers within ~1/8th of their interval?
+		 */
 		if (timer->expires == expires)
 			return 1;
 

-- 
2.13.3