[PATCH V2 0/2] sched: Cleanups,fixes in nohz_kick_needed()

Changes from V1:https://lkml.org/lkml/2013/10/21/248

1. Swapped the order of PATCH1 and PATCH2 in V1 so as to not mess with the
nr_busy_cpus parameter computation during asymmetric balancing, while fixing
it.

2. nohz_busy_cpus parameter is to be updated and queried at only one level of
the sched domain-sd_busy where it is relevant.

3. Introduce sd_asym to represent the sched domain where asymmetric load
balancing has to be done.
---

Preeti U Murthy (1):
      sched: Remove un-necessary iteration over sched domains to update nr_busy_cpus

Vaidyanathan Srinivasan (1):
      sched: Fix asymmetric scheduling for POWER7


 kernel/sched/core.c  |    6 ++++++
 kernel/sched/fair.c  |   38 ++++++++++++++++++++------------------
 kernel/sched/sched.h |    2 ++
 3 files changed, 28 insertions(+), 18 deletions(-)

From: Vaidyanathan Srinivasan <redacted>

Asymmetric scheduling within a core is a scheduler loadbalancing
feature that is triggered when SD_ASYM_PACKING flag is set.  The goal
for the load balancer is to move tasks to lower order idle SMT threads
within a core on a POWER7 system.

In nohz_kick_needed(), we intend to check if our sched domain (core)
is completely busy or we have idle cpu.

The following check for SD_ASYM_PACKING:

    (cpumask_first_and(nohz.idle_cpus_mask, sched_domain_span(sd)) < cpu)

already covers the case of checking if the domain has an idle cpu,
because cpumask_first_and() will not yield any set bits if this domain
has no idle cpu.

Hence, nr_busy check against group weight can be removed.

Reported-by: Michael Neuling <redacted>
Signed-off-by: Vaidyanathan Srinivasan <redacted>
Signed-off-by: Preeti U Murthy <redacted>
Tested-by: Michael Neuling <redacted>
---

 kernel/sched/fair.c |    2 +-
 1 file changed, 1 insertion(+), 1 deletion(-)

diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
index 813dd61..e9c9549 100644
--- a/kernel/sched/fair.c
+++ b/kernel/sched/fair.c

@@ -6781,7 +6781,7 @@ static inline int nohz_kick_needed(struct rq *rq, int cpu)
 		if (sd->flags & SD_SHARE_PKG_RESOURCES && nr_busy > 1)
 			goto need_kick_unlock;
 
-		if (sd->flags & SD_ASYM_PACKING && nr_busy != sg->group_weight
+		if (sd->flags & SD_ASYM_PACKING
 		    && (cpumask_first_and(nohz.idle_cpus_mask,
 					  sched_domain_span(sd)) < cpu))
 			goto need_kick_unlock;

nr_busy_cpus parameter is used by nohz_kick_needed() to find out the number
of busy cpus in a sched domain which has SD_SHARE_PKG_RESOURCES flag set.
Therefore instead of updating nr_busy_cpus at every level of sched domain,
since it is irrelevant, we can update this parameter only at the parent
domain of the sd which has this flag set. Introduce a per-cpu parameter
sd_busy which represents this parent domain.

In nohz_kick_needed() we directly query the nr_busy_cpus parameter
associated with the groups of sd_busy.

By associating sd_busy with the highest domain which has
SD_SHARE_PKG_RESOURCES flag set, we cover all lower level domains which could
have this flag set and trigger nohz_idle_balancing if any of the levels have
more than one busy cpu.

sd_busy is irrelevant for asymmetric load balancing.

While we are at it, we might as well change the nohz_idle parameter to be
updated at the sd_busy domain level alone and not the base domain level of a CPU.
This will unify the concept of busy cpus at just one level of sched domain
where it is currently used.

Signed-off-by: Preeti U Murthy<redacted>
---

 kernel/sched/core.c  |    6 ++++++
 kernel/sched/fair.c  |   38 ++++++++++++++++++++------------------
 kernel/sched/sched.h |    2 ++
 3 files changed, 28 insertions(+), 18 deletions(-)

diff --git a/kernel/sched/core.c b/kernel/sched/core.c
index c06b8d3..e6a6244 100644
--- a/kernel/sched/core.c
+++ b/kernel/sched/core.c

@@ -5271,6 +5271,8 @@ DEFINE_PER_CPU(struct sched_domain *, sd_llc);
 DEFINE_PER_CPU(int, sd_llc_size);
 DEFINE_PER_CPU(int, sd_llc_id);
 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
+DEFINE_PER_CPU(struct sched_domain *, sd_busy);
+DEFINE_PER_CPU(struct sched_domain *, sd_asym);
 
 static void update_top_cache_domain(int cpu)
 {

@@ -5282,6 +5284,7 @@ static void update_top_cache_domain(int cpu)
 	if (sd) {
 		id = cpumask_first(sched_domain_span(sd));
 		size = cpumask_weight(sched_domain_span(sd));
+		rcu_assign_pointer(per_cpu(sd_busy, cpu), sd->parent);
 	}
 
 	rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);

@@ -5290,6 +5293,9 @@ static void update_top_cache_domain(int cpu)
 
 	sd = lowest_flag_domain(cpu, SD_NUMA);
 	rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
+
+	sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
+	rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
 }
 
 /*

diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
index e9c9549..8602b2c 100644
--- a/kernel/sched/fair.c
+++ b/kernel/sched/fair.c

@@ -6515,16 +6515,16 @@ static inline void nohz_balance_exit_idle(int cpu)
 static inline void set_cpu_sd_state_busy(void)
 {
 	struct sched_domain *sd;
+	int cpu = smp_processor_id();
 
 	rcu_read_lock();
-	sd = rcu_dereference_check_sched_domain(this_rq()->sd);
+	sd = rcu_dereference(per_cpu(sd_busy, cpu));
 
 	if (!sd || !sd->nohz_idle)
 		goto unlock;
 	sd->nohz_idle = 0;
 
-	for (; sd; sd = sd->parent)
-		atomic_inc(&sd->groups->sgp->nr_busy_cpus);
+	atomic_inc(&sd->groups->sgp->nr_busy_cpus);
 unlock:
 	rcu_read_unlock();
 }

@@ -6532,16 +6532,16 @@ unlock:
 void set_cpu_sd_state_idle(void)
 {
 	struct sched_domain *sd;
+	int cpu = smp_processor_id();
 
 	rcu_read_lock();
-	sd = rcu_dereference_check_sched_domain(this_rq()->sd);
+	sd = rcu_dereference(per_cpu(sd_busy, cpu));
 
 	if (!sd || sd->nohz_idle)
 		goto unlock;
 	sd->nohz_idle = 1;
 
-	for (; sd; sd = sd->parent)
-		atomic_dec(&sd->groups->sgp->nr_busy_cpus);
+	atomic_dec(&sd->groups->sgp->nr_busy_cpus);
 unlock:
 	rcu_read_unlock();
 }

@@ -6748,6 +6748,8 @@ static inline int nohz_kick_needed(struct rq *rq, int cpu)
 {
 	unsigned long now = jiffies;
 	struct sched_domain *sd;
+	struct sched_group_power *sgp;
+	int nr_busy;
 
 	if (unlikely(idle_cpu(cpu)))
 		return 0;

@@ -6773,22 +6775,22 @@ static inline int nohz_kick_needed(struct rq *rq, int cpu)
 		goto need_kick;
 
 	rcu_read_lock();
-	for_each_domain(cpu, sd) {
-		struct sched_group *sg = sd->groups;
-		struct sched_group_power *sgp = sg->sgp;
-		int nr_busy = atomic_read(&sgp->nr_busy_cpus);
+	sd = rcu_dereference(per_cpu(sd_busy, cpu));
 
-		if (sd->flags & SD_SHARE_PKG_RESOURCES && nr_busy > 1)
-			goto need_kick_unlock;
+	if (sd) {
+		sgp = sd->groups->sgp;
+		nr_busy = atomic_read(&sgp->nr_busy_cpus);
 
-		if (sd->flags & SD_ASYM_PACKING
-		    && (cpumask_first_and(nohz.idle_cpus_mask,
-					  sched_domain_span(sd)) < cpu))
+		if (nr_busy > 1)
 			goto need_kick_unlock;
-
-		if (!(sd->flags & (SD_SHARE_PKG_RESOURCES | SD_ASYM_PACKING)))
-			break;
 	}
+
+	sd = rcu_dereference(per_cpu(sd_asym, cpu));
+
+	if (sd && (cpumask_first_and(nohz.idle_cpus_mask,
+				  sched_domain_span(sd)) < cpu))
+		goto need_kick_unlock;
+
 	rcu_read_unlock();
 	return 0;
 

diff --git a/kernel/sched/sched.h b/kernel/sched/sched.h
index ffc7087..c8cb145 100644
--- a/kernel/sched/sched.h
+++ b/kernel/sched/sched.h

@@ -623,6 +623,8 @@ DECLARE_PER_CPU(struct sched_domain *, sd_llc);
 DECLARE_PER_CPU(int, sd_llc_size);
 DECLARE_PER_CPU(int, sd_llc_id);
 DECLARE_PER_CPU(struct sched_domain *, sd_numa);
+DECLARE_PER_CPU(struct sched_domain *, sd_busy);
+DECLARE_PER_CPU(struct sched_domain *, sd_asym);
 
 struct sched_group_power {
 	atomic_t ref;

The changelog has missed mentioning the introduction of sd_asym per_cpu sched domain.
Apologies for this. The patch with the changelog including mention of sd_asym is
pasted below.

Regards
Preeti U Murthy

---------------

sched: Remove un-necessary iteration over sched domains to update nr_busy_cpus

From: Preeti U Murthy <redacted>

nr_busy_cpus parameter is used by nohz_kick_needed() to find out the number
of busy cpus in a sched domain which has SD_SHARE_PKG_RESOURCES flag set.
Therefore instead of updating nr_busy_cpus at every level of sched domain,
since it is irrelevant, we can update this parameter only at the parent
domain of the sd which has this flag set. Introduce a per-cpu parameter
sd_busy which represents this parent domain.

In nohz_kick_needed() we directly query the nr_busy_cpus parameter
associated with the groups of sd_busy.

By associating sd_busy with the highest domain which has
SD_SHARE_PKG_RESOURCES flag set, we cover all lower level domains which could
have this flag set and trigger nohz_idle_balancing if any of the levels have
more than one busy cpu.

sd_busy is irrelevant for asymmetric load balancing. However sd_asym has been
introduced to represent the highest sched domain which has SD_ASYM_PACKING flag set
so that it can be queried directly when required.

While we are at it, we might as well change the nohz_idle parameter to be
updated at the sd_busy domain level alone and not the base domain level of a CPU.
This will unify the concept of busy cpus at just one level of sched domain
where it is currently used.

Signed-off-by: Preeti U Murthy<redacted>
---
 kernel/sched/core.c  |    6 ++++++
 kernel/sched/fair.c  |   38 ++++++++++++++++++++------------------
 kernel/sched/sched.h |    2 ++
 3 files changed, 28 insertions(+), 18 deletions(-)

diff --git a/kernel/sched/core.c b/kernel/sched/core.c
index c06b8d3..e6a6244 100644
--- a/kernel/sched/core.c
+++ b/kernel/sched/core.c

@@ -5271,6 +5271,8 @@ DEFINE_PER_CPU(struct sched_domain *, sd_llc);
 DEFINE_PER_CPU(int, sd_llc_size);
 DEFINE_PER_CPU(int, sd_llc_id);
 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
+DEFINE_PER_CPU(struct sched_domain *, sd_busy);
+DEFINE_PER_CPU(struct sched_domain *, sd_asym);
 
 static void update_top_cache_domain(int cpu)
 {

@@ -5282,6 +5284,7 @@ static void update_top_cache_domain(int cpu)
 	if (sd) {
 		id = cpumask_first(sched_domain_span(sd));
 		size = cpumask_weight(sched_domain_span(sd));
+		rcu_assign_pointer(per_cpu(sd_busy, cpu), sd->parent);
 	}
 
 	rcu_assign_pointer(per_cpu(sd_llc, cpu), sd);

@@ -5290,6 +5293,9 @@ static void update_top_cache_domain(int cpu)
 
 	sd = lowest_flag_domain(cpu, SD_NUMA);
 	rcu_assign_pointer(per_cpu(sd_numa, cpu), sd);
+
+	sd = highest_flag_domain(cpu, SD_ASYM_PACKING);
+	rcu_assign_pointer(per_cpu(sd_asym, cpu), sd);
 }
 
 /*

diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
index e9c9549..8602b2c 100644
--- a/kernel/sched/fair.c
+++ b/kernel/sched/fair.c

@@ -6515,16 +6515,16 @@ static inline void nohz_balance_exit_idle(int cpu)
 static inline void set_cpu_sd_state_busy(void)
 {
 	struct sched_domain *sd;
+	int cpu = smp_processor_id();
 
 	rcu_read_lock();
-	sd = rcu_dereference_check_sched_domain(this_rq()->sd);
+	sd = rcu_dereference(per_cpu(sd_busy, cpu));
 
 	if (!sd || !sd->nohz_idle)
 		goto unlock;
 	sd->nohz_idle = 0;
 
-	for (; sd; sd = sd->parent)
-		atomic_inc(&sd->groups->sgp->nr_busy_cpus);
+	atomic_inc(&sd->groups->sgp->nr_busy_cpus);
 unlock:
 	rcu_read_unlock();
 }

@@ -6532,16 +6532,16 @@ unlock:
 void set_cpu_sd_state_idle(void)
 {
 	struct sched_domain *sd;
+	int cpu = smp_processor_id();
 
 	rcu_read_lock();
-	sd = rcu_dereference_check_sched_domain(this_rq()->sd);
+	sd = rcu_dereference(per_cpu(sd_busy, cpu));
 
 	if (!sd || sd->nohz_idle)
 		goto unlock;
 	sd->nohz_idle = 1;
 
-	for (; sd; sd = sd->parent)
-		atomic_dec(&sd->groups->sgp->nr_busy_cpus);
+	atomic_dec(&sd->groups->sgp->nr_busy_cpus);
 unlock:
 	rcu_read_unlock();
 }

@@ -6748,6 +6748,8 @@ static inline int nohz_kick_needed(struct rq *rq, int cpu)
 {
 	unsigned long now = jiffies;
 	struct sched_domain *sd;
+	struct sched_group_power *sgp;
+	int nr_busy;
 
 	if (unlikely(idle_cpu(cpu)))
 		return 0;

@@ -6773,22 +6775,22 @@ static inline int nohz_kick_needed(struct rq *rq, int cpu)
 		goto need_kick;
 
 	rcu_read_lock();
-	for_each_domain(cpu, sd) {
-		struct sched_group *sg = sd->groups;
-		struct sched_group_power *sgp = sg->sgp;
-		int nr_busy = atomic_read(&sgp->nr_busy_cpus);
+	sd = rcu_dereference(per_cpu(sd_busy, cpu));
 
-		if (sd->flags & SD_SHARE_PKG_RESOURCES && nr_busy > 1)
-			goto need_kick_unlock;
+	if (sd) {
+		sgp = sd->groups->sgp;
+		nr_busy = atomic_read(&sgp->nr_busy_cpus);
 
-		if (sd->flags & SD_ASYM_PACKING
-		    && (cpumask_first_and(nohz.idle_cpus_mask,
-					  sched_domain_span(sd)) < cpu))
+		if (nr_busy > 1)
 			goto need_kick_unlock;
-
-		if (!(sd->flags & (SD_SHARE_PKG_RESOURCES | SD_ASYM_PACKING)))
-			break;
 	}
+
+	sd = rcu_dereference(per_cpu(sd_asym, cpu));
+
+	if (sd && (cpumask_first_and(nohz.idle_cpus_mask,
+				  sched_domain_span(sd)) < cpu))
+		goto need_kick_unlock;
+
 	rcu_read_unlock();
 	return 0;
 

diff --git a/kernel/sched/sched.h b/kernel/sched/sched.h
index ffc7087..c8cb145 100644
--- a/kernel/sched/sched.h
+++ b/kernel/sched/sched.h

@@ -623,6 +623,8 @@ DECLARE_PER_CPU(struct sched_domain *, sd_llc);
 DECLARE_PER_CPU(int, sd_llc_size);
 DECLARE_PER_CPU(int, sd_llc_id);
 DECLARE_PER_CPU(struct sched_domain *, sd_numa);
+DECLARE_PER_CPU(struct sched_domain *, sd_busy);
+DECLARE_PER_CPU(struct sched_domain *, sd_asym);
 
 struct sched_group_power {
 	atomic_t ref;

Hi Preeti,

consider a machine with single socket, dual core with HT enabled. The top most
domain is also the highest domain with SD_SHARE_PKG_RESOURCES flag set,
i.e MC domain (the machine toplogy consist of SIBLING and MC domain).

# lstopo-no-graphics --no-bridges --no-io
Machine (7869MB) + Socket L#0 + L3 L#0 (3072KB)
  L2 L#0 (256KB) + L1d L#0 (32KB) + L1i L#0 (32KB) + Core L#0
    PU L#0 (P#0)
    PU L#1 (P#1)
  L2 L#1 (256KB) + L1d L#1 (32KB) + L1i L#1 (32KB) + Core L#1
    PU L#2 (P#2)
    PU L#3 (P#3)

With this approach parent of MC domain is NULL and given that sd_busy is NULL,
nr_busy_cpus of sched domain sd_busy will never be incremented/decremented.
Resulting is nohz_kick_needed returning 0.

Thanks,
Kamalesh.

Hi Kamalesh,

On 10/30/2013 02:53 PM, Kamalesh Babulal wrote:

Hi Preeti,

quoted

nr_busy_cpus parameter is used by nohz_kick_needed() to find out the number
of busy cpus in a sched domain which has SD_SHARE_PKG_RESOURCES flag set.
Therefore instead of updating nr_busy_cpus at every level of sched domain,
since it is irrelevant, we can update this parameter only at the parent
domain of the sd which has this flag set. Introduce a per-cpu parameter
sd_busy which represents this parent domain.

In nohz_kick_needed() we directly query the nr_busy_cpus parameter
associated with the groups of sd_busy.

By associating sd_busy with the highest domain which has
SD_SHARE_PKG_RESOURCES flag set, we cover all lower level domains which could
have this flag set and trigger nohz_idle_balancing if any of the levels have
more than one busy cpu.

sd_busy is irrelevant for asymmetric load balancing. However sd_asym has been
introduced to represent the highest sched domain which has SD_ASYM_PACKING flag set
so that it can be queried directly when required.

While we are at it, we might as well change the nohz_idle parameter to be
updated at the sd_busy domain level alone and not the base domain level of a CPU.
This will unify the concept of busy cpus at just one level of sched domain
where it is currently used.

Signed-off-by: Preeti U Murthy<redacted>
---
 kernel/sched/core.c  |    6 ++++++
 kernel/sched/fair.c  |   38 ++++++++++++++++++++------------------
 kernel/sched/sched.h |    2 ++
 3 files changed, 28 insertions(+), 18 deletions(-)

diff --git a/kernel/sched/core.c b/kernel/sched/core.c
index c06b8d3..e6a6244 100644
--- a/kernel/sched/core.c
+++ b/kernel/sched/core.c

@@ -5271,6 +5271,8 @@ DEFINE_PER_CPU(struct sched_domain *, sd_llc);
 DEFINE_PER_CPU(int, sd_llc_size);
 DEFINE_PER_CPU(int, sd_llc_id);
 DEFINE_PER_CPU(struct sched_domain *, sd_numa);
+DEFINE_PER_CPU(struct sched_domain *, sd_busy);
+DEFINE_PER_CPU(struct sched_domain *, sd_asym);

 static void update_top_cache_domain(int cpu)
 {

@@ -5282,6 +5284,7 @@ static void update_top_cache_domain(int cpu)
 	if (sd) {
 		id = cpumask_first(sched_domain_span(sd));
 		size = cpumask_weight(sched_domain_span(sd));
+		rcu_assign_pointer(per_cpu(sd_busy, cpu), sd->parent);
 	}

consider a machine with single socket, dual core with HT enabled. The top most
domain is also the highest domain with SD_SHARE_PKG_RESOURCES flag set,
i.e MC domain (the machine toplogy consist of SIBLING and MC domain).

# lstopo-no-graphics --no-bridges --no-io
Machine (7869MB) + Socket L#0 + L3 L#0 (3072KB)
  L2 L#0 (256KB) + L1d L#0 (32KB) + L1i L#0 (32KB) + Core L#0
    PU L#0 (P#0)
    PU L#1 (P#1)
  L2 L#1 (256KB) + L1d L#1 (32KB) + L1i L#1 (32KB) + Core L#1
    PU L#2 (P#2)
    PU L#3 (P#3)

With this approach parent of MC domain is NULL and given that sd_busy is NULL,
nr_busy_cpus of sched domain sd_busy will never be incremented/decremented.
Resulting is nohz_kick_needed returning 0.

Right and it *should* return 0. There is no sibling domain that can
offload tasks from it. Therefore there is no point kicking nohz idle
balance.

Regards
Preeti U Murthy

Thanks,
Kamalesh.