timer: make deferrable cpu unbound timers really not bound to a cpu

kernel/time/timer.c struct tvec_base { · line 98

Quoted 3 lines · Joonwoo Park

kernel/time/timer.c +3 -0

+#ifdef CONFIG_SMP
+static struct tvec_base *tvec_base_deferral = &boot_tvec_bases;
+#endif

In principle I like the idea of a deferrable wheel, but this implementation is going to go nowhere.

First of all making it SMP only is silly. The deferrable stuff is a pain in other places as well.

But whats way worse is:

Quoted 4 lines

kernel/time/timer.c +1 -0

+static inline void __run_timers(struct tvec_base *base, bool try)
 {
 	struct timer_list *timer;

kernel/time/timer.c static void call_timer_fn(struct timer_list *timer, void (*fn)(unsigned long), · line 1192

Quoted 5 lines

kernel/time/timer.c +4 -1

-	spin_lock_irq(&base->lock);
+	if (!try)
+		spin_lock_irq(&base->lock);
+	else if (!spin_trylock_irq(&base->lock))
+		return;

Yuck. All cpus fighting about a single spinlock roughly at the same time? You just created a proper thundering herd cacheline bouncing issue.

No way. We have already mechanisms in place to deal with such problems, you just have to use them.

Thanks,

tglx

kernel/time/timer.c struct tvec_base { · line 98

Quoted 3 lines · Joonwoo Park

kernel/time/timer.c +3 -0

+#ifdef CONFIG_SMP
+static struct tvec_base *tvec_base_deferral = &boot_tvec_bases;
+#endif

In principle I like the idea of a deferrable wheel, but this implementation is going to go nowhere.

First of all making it SMP only is silly. The deferrable stuff is a pain in other places as well.

But whats way worse is:

Quoted 4 lines

+static inline void __run_timers(struct tvec_base *base, bool try)
{
struct timer_list *timer;

kernel/time/timer.c static void call_timer_fn(struct timer_list *timer, void (*fn)(unsigned long), · line 1192

Quoted 5 lines

kernel/time/timer.c +4 -1

-	spin_lock_irq(&base->lock);
+	if (!try)
+		spin_lock_irq(&base->lock);
+	else if (!spin_trylock_irq(&base->lock))
+		return;

Yuck. All cpus fighting about a single spinlock roughly at the same time? You just created a proper thundering herd cacheline bouncing issue.

No way. We have already mechanisms in place to deal with such problems, you just have to use them.

Quoted 7 lines · Thomas Gleixner

kernel/time/timer.c struct tvec_base { · line 98

Quoted 3 lines · Joonwoo Park

kernel/time/timer.c +3 -0

+#ifdef CONFIG_SMP
+static struct tvec_base *tvec_base_deferral = &boot_tvec_bases;
+#endif

In principle I like the idea of a deferrable wheel, but this implementation is going to go nowhere.

Hi Thomas,

To give some more context:

This bug is a serious pain in the a** for anyone using deferrable timers or deferrable workqueues today for some periodic work and don't care for which CPU the code runs in.

Couple of examples of such issues in existing code:

1) In a SMP system, CPUfreq governors (ondemand and conservative) end up queueing a deferrable work on every single CPU and the first one to run the deferrable workqueue cancels the work on all other CPUS, runs the work and then sets up a workqueue on all the CPUs again for the next sampling point.

2) Devfreq actually doesn't work well today because it doesn't do this nasty hack like CPUfreq. So, if the devfreq work happens to run on a CPU that's typically idle, then it doesn't run for a long time. I've actually seen logs where the devfreq polling interval is set to 20ms, but ends up running 800ms later.

Don't know how many other drivers are suffering from this bug. Yes, IMHO, this is a bug. When the timer is CPU unbound, anyone who hasn't looked at the actual timer code would assume that it'll run as long as any CPU is busy.

Quoted 15 lines

First of all making it SMP only is silly. The deferrable stuff is a pain in other places as well.

But whats way worse is:

kernel/time/timer.c static void call_timer_fn(struct timer_list *timer, void (*fn)(unsigned long), · line 1192

Quoted 9 lines

kernel/time/timer.c +5 -1

+static inline void __run_timers(struct tvec_base *base, bool try)
  {
      struct timer_list *timer;
 
-    spin_lock_irq(&base->lock);
+    if (!try)
+        spin_lock_irq(&base->lock);
+    else if (!spin_trylock_irq(&base->lock))
+        return;

Quoted 3 lines

Yuck. All cpus fighting about a single spinlock roughly at the same time? You just created a proper thundering herd cacheline bouncing issue.

I agree, This is not good. I think a simple way to fix this is to have only the CPU that's the jiffy owner to run through this deferrable timer list.

That should address any unnecessary cache bouncing issues. Would that be an acceptable solution to this?

Quoted 3 lines

No way. We have already mechanisms in place to deal with such problems, you just have to use them.

I'm not sure which problem you are referring to here. Or what the already existing solutions are.

I don't think you were referring to the "deferrable timer getting delayed for long periods despite CPUs being active" problem, because I don't think we have a better solution than this patch (with the jiffy owner CPU fix). Asking every driver that doesn't care which CPU the work runs on to queue a work or set up a timer on every CPU is definitely not a good solution -- that's spreading the complexity to every other driver instead of fixing it in one place. And that causes unnecessary cache pollution too.

Thoughts? I would really like to see a fix for this go in soon so that we can simplify cpufreq and have devfreq and other drivers work correctly.

Thanks,
Saravana

From 0c91f82a0b43b247f1ed310212ef3aada7ccc9f7 Mon Sep 17 00:00:00 2001

Quoted 78 lines · Saravana Kannan

Quoted 7 lines · Thomas Gleixner

Review

Quoted 3 lines · Joonwoo Park

(quoted diff) +3 -0

+#ifdef CONFIG_SMP
+static struct tvec_base *tvec_base_deferral = &boot_tvec_bases;
+#endif

In principle I like the idea of a deferrable wheel, but this implementation is going to go nowhere.

Hi Thomas,

To give some more context:

This bug is a serious pain in the a** for anyone using deferrable timers or deferrable workqueues today for some periodic work and don't care for which CPU the code runs in.

Couple of examples of such issues in existing code:

1) In a SMP system, CPUfreq governors (ondemand and conservative) end up queueing a deferrable work on every single CPU and the first one to run the deferrable workqueue cancels the work on all other CPUS, runs the work and then sets up a workqueue on all the CPUs again for the next sampling point.

2) Devfreq actually doesn't work well today because it doesn't do this nasty hack like CPUfreq. So, if the devfreq work happens to run on a CPU that's typically idle, then it doesn't run for a long time. I've actually seen logs where the devfreq polling interval is set to 20ms, but ends up running 800ms later.

Don't know how many other drivers are suffering from this bug. Yes, IMHO, this is a bug. When the timer is CPU unbound, anyone who hasn't looked at the actual timer code would assume that it'll run as long as any CPU is busy.

Quoted 15 lines

First of all making it SMP only is silly. The deferrable stuff is a pain in other places as well.

But whats way worse is:

Review

Quoted 9 lines

(quoted diff) +5 -1

+static inline void __run_timers(struct tvec_base *base, bool try)
 {
     struct timer_list *timer;
 
-    spin_lock_irq(&base->lock);
+    if (!try)
+        spin_lock_irq(&base->lock);
+    else if (!spin_trylock_irq(&base->lock))
+        return;

Quoted 3 lines

Yuck. All cpus fighting about a single spinlock roughly at the same time? You just created a proper thundering herd cacheline bouncing issue.

I agree, This is not good. I think a simple way to fix this is to have only the CPU that's the jiffy owner to run through this deferrable timer list.

That should address any unnecessary cache bouncing issues. Would that be an acceptable solution to this?

Quoted 3 lines

No way. We have already mechanisms in place to deal with such problems, you just have to use them.

I'm not sure which problem you are referring to here. Or what the already existing solutions are.

I don't think you were referring to the "deferrable timer getting delayed for long periods despite CPUs being active" problem, because I don't think we have a better solution than this patch (with the jiffy owner CPU fix). Asking every driver that doesn't care which CPU the work runs on to queue a work or set up a timer on every CPU is definitely not a good solution -- that's spreading the complexity to every other driver instead of fixing it in one place. And that causes unnecessary cache pollution too.

Quoted 6 lines

Thoughts? I would really like to see a fix for this go in soon so that we can simplify cpufreq and have devfreq and other drivers work correctly.

Thanks,
Saravana

Thomas,

Bump.

-Saravana

Signature · 2 lines

The Qualcomm Innovation Center, Inc. is a member of Code Aurora Forum,
a Linux Foundation Collaborative Project

When a deferrable work (INIT_DEFERRABLE_WORK, etc.) is queued via queue_delayed_work() it's probably intended to run the work item on any CPU that isn't idle. However, we queue the work to run at a later time by starting a deferrable timer that binds to whatever CPU the work is queued on which is same with queue_delayed_work_on(smp_processor_id()) effectively.

As a result WORK_CPU_UNBOUND work items aren't really cpu unbound now. In fact this is perfectly fine with UP kernel and also won't affect much a system without dyntick with SMP kernel too as every cpus run timers periodically. But on SMP systems with dyntick current implementation leads deferrable timers not very scalable because the timer's base which has queued the deferrable timer won't wake up till next non-deferrable timer expires even though there are possible other non idle cpus are running which are able to run expired deferrable timers.

The deferrable work is a good example of the current implementation's victim like below.

INIT_DEFERRABLE_WORK(&dwork, fn);
CPU 0 CPU 1

queue_delayed_work(wq, &dwork, HZ);
    queue_delayed_work_on(WORK_CPU_UNBOUND);
        ...
	__mod_timer() -> queues timer to the
			 current cpu's timer
			 base.
	...

tick_nohz_idle_enter() -> cpu enters idle.
A second later
cpu 0 is now in idle. cpu 1 exits idle or wasn't in idle so
now it's in active but won't
cpu 0 won't wake up till next handle cpu unbound deferrable timer
non-deferrable timer expires. as it's in cpu 0's timer base.

To make all cpu unbound deferrable timers are scalable, introduce a common timer base which is only for cpu unbound deferrable timers to make those are indeed cpu unbound so that can be scheduled by tick_do_timer_cpu. This common timer fixes scalability issue of delayed work and all other cpu unbound deferrable timer using implementations.

CC Thomas Gleixner <[email protected]>

CC John Stultz <[email protected]>

CC Tejun Heo <[email protected]>

Signed-off-by Joonwoo Park <[email protected]>

---
Changes in v3:
* Make only tick_do_timer_cpu to run deferral timer wheel to reduce cache bouncing.

Changes in v4:
* Kill CONFIG_SMP ifdefry.
* Allocate and initialize tvec_base_deferrable at compile time.
* Pin pinned deferrable timer.
* s/deferral/deferrable/

 include/linux/timer.h |  14 ++++++-
 kernel/time/timer.c   | 103 ++++++++++++++++++++++++++++++++++++++++----------
 2 files changed, 97 insertions(+), 20 deletions(-)

include/linux/timer.h +13 -1

index 8c5a197..45847ca 100644
@@ -34,6 +34,9 @@ struct timer_list {
 };
 
 extern struct tvec_base boot_tvec_bases;
+#ifdef CONFIG_SMP
+extern struct tvec_base tvec_base_deferrable;
+#endif
 
 #ifdef CONFIG_LOCKDEP
 /*

@@ -70,12 +73,21 @@ extern struct tvec_base boot_tvec_bases;
 
 #define TIMER_FLAG_MASK			0x3LU
 
+#ifdef CONFIG_SMP
+#define __TIMER_BASE(_flags) \
+	((_flags) & TIMER_DEFERRABLE ? \
+	 (unsigned long)&tvec_base_deferrable + (_flags) : \
+	 (unsigned long)&boot_tvec_bases + (_flags))
+#else
+#define __TIMER_BASE(_flags) ((unsigned long)&boot_tvec_bases + (_flags))
+#endif
+
 #define __TIMER_INITIALIZER(_function, _expires, _data, _flags) { \
 		.entry = { .prev = TIMER_ENTRY_STATIC },	\
 		.function = (_function),			\
 		.expires = (_expires),				\
 		.data = (_data),				\
-		.base = (void *)((unsigned long)&boot_tvec_bases + (_flags)), \
+		.base = (void *)(__TIMER_BASE(_flags)),		\
 		.slack = -1,					\
 		__TIMER_LOCKDEP_MAP_INITIALIZER(		\
 			__FILE__ ":" __stringify(__LINE__))	\

kernel/time/timer.c +84 -19

index e5d5733c..133e94a 100644
@@ -49,6 +49,8 @@
 #include <asm/timex.h>
 #include <asm/io.h>
 
+#include "tick-internal.h"
+
 #define CREATE_TRACE_POINTS
 #include <trace/events/timer.h>
 

@@ -103,6 +105,9 @@ struct tvec_base boot_tvec_bases;
 EXPORT_SYMBOL(boot_tvec_bases);
 
 static DEFINE_PER_CPU(struct tvec_base *, tvec_bases) = &boot_tvec_bases;
+#ifdef CONFIG_SMP
+struct tvec_base tvec_base_deferrable;
+#endif
 
 /* Functions below help us manage 'deferrable' flag */
 static inline unsigned int tbase_get_deferrable(struct tvec_base *base)

@@ -662,10 +667,63 @@ static inline void debug_assert_init(struct timer_list *timer)
 	debug_timer_assert_init(timer);
 }
 
+#ifdef CONFIG_SMP
+static inline struct tvec_base *__get_timer_base(unsigned int flags)
+{
+	if (flags & TIMER_DEFERRABLE)
+		return &tvec_base_deferrable;
+	else
+		return raw_cpu_read(tvec_bases);
+}
+
+static inline bool is_deferrable_timer_base(struct tvec_base *base)
+{
+	return base == &tvec_base_deferrable;
+}
+
+static inline void __run_timers(struct tvec_base *base);
+static inline void __run_deferrable_timers(void)
+{
+	if (smp_processor_id() == tick_do_timer_cpu &&
+	    time_after_eq(jiffies, tvec_base_deferrable.timer_jiffies))
+		__run_timers(&tvec_base_deferrable);
+}
+
+static void __init init_timer_cpu(struct tvec_base *base, int cpu);
+static inline void init_deferrable_timer(void)
+{
+	init_timer_cpu(&tvec_base_deferrable, NR_CPUS);
+}
+#else
+static inline struct tvec_base *__get_timer_base(unsigned int flags)
+{
+	return raw_cpu_read(tvec_bases);
+}
+
+static inline bool is_deferrable_timer_base(struct tvec_base *base)
+{
+	return false;
+}
+
+static inline void __run_deferrable_timers(void)
+{
+}
+
+static inline void init_deferrable_timer(void)
+{
+	/*
+	 * initialize cpu unbound deferrable timer base only when CONFIG_SMP.
+	 * UP kernel handles the timers with cpu 0 timer base.
+	 */
+}
+#endif
+
 static void do_init_timer(struct timer_list *timer, unsigned int flags,
 			  const char *name, struct lock_class_key *key)
 {
-	struct tvec_base *base = raw_cpu_read(tvec_bases);
+	struct tvec_base *base;
+
+	base = __get_timer_base(flags);
 
 	timer->entry.next = NULL;
 	timer->base = (void *)((unsigned long)base | flags);

@@ -787,24 +845,26 @@ __mod_timer(struct timer_list *timer, unsigned long expires,
 
 	debug_activate(timer, expires);
 
-	cpu = get_nohz_timer_target(pinned);
-	new_base = per_cpu(tvec_bases, cpu);
+	if (!is_deferrable_timer_base(base) || pinned == TIMER_PINNED) {
+		cpu = get_nohz_timer_target(pinned);
+		new_base = per_cpu(tvec_bases, cpu);
 
-	if (base != new_base) {
-		/*
-		 * We are trying to schedule the timer on the local CPU.
-		 * However we can't change timer's base while it is running,
-		 * otherwise del_timer_sync() can't detect that the timer's
-		 * handler yet has not finished. This also guarantees that
-		 * the timer is serialized wrt itself.
-		 */
-		if (likely(base->running_timer != timer)) {
-			/* See the comment in lock_timer_base() */
-			timer_set_base(timer, NULL);
-			spin_unlock(&base->lock);
-			base = new_base;
-			spin_lock(&base->lock);
-			timer_set_base(timer, base);
+		if (base != new_base) {
+			/*
+			 * We are trying to schedule the timer on the local CPU.
+			 * However we can't change timer's base while it is
+			 * running, otherwise del_timer_sync() can't detect that
+			 * the timer's handler yet has not finished. This also
+			 * guarantees that the timer is serialized wrt itself.
+			 */
+			if (likely(base->running_timer != timer)) {
+				/* See the comment in lock_timer_base() */
+				timer_set_base(timer, NULL);
+				spin_unlock(&base->lock);
+				base = new_base;
+				spin_lock(&base->lock);
+				timer_set_base(timer, base);
+			}
 		}
 	}
 

@@ -1411,6 +1471,8 @@ static void run_timer_softirq(struct softirq_action *h)
 
 	hrtimer_run_pending();
 
+	__run_deferrable_timers();
+
 	if (time_after_eq(jiffies, base->timer_jiffies))
 		__run_timers(base);
 }

@@ -1623,7 +1685,8 @@ static void __init init_timer_cpu(struct tvec_base *base, int cpu)
 	BUG_ON(base != tbase_get_base(base));
 
 	base->cpu = cpu;
-	per_cpu(tvec_bases, cpu) = base;
+	if (cpu != NR_CPUS)
+		per_cpu(tvec_bases, cpu) = base;
 	spin_lock_init(&base->lock);
 
 	for (j = 0; j < TVN_SIZE; j++) {

@@ -1655,6 +1718,8 @@ static void __init init_timer_cpus(void)
 
 		init_timer_cpu(base, cpu);
 	}
+
+	init_deferrable_timer();
 }
 
 void __init init_timers(void)