[RFC PATCH 1/1] fair.c: Add/Export find_idlest_perfer_cpu API

Add/Export a new API for per-cpu thread model networking device driver
to choose a preferred idlest cpu within allowed cpumask.

The receiving CPUs of a networking device are not under cgroup controls.
Normally the receiving work will be scheduled on the cpu on which the
interrupts are received. When such a networking device uses per-cpu
thread model, the cpu which is chose to process the packets might not be
part of cgroup cpusets without using such an API here. 

On NUMA system, by using the preferred cpumask from the same NUMA node
would help to reduce expensive cross memory access to/from the other
NUMA node.

KVM per-cpu vhost will be the first one to use this API. Any other
device driver which uses per-cpu thread model and has cgroup cpuset
control will use this API later.

Signed-off-by: Shirley Ma <redacted>
---
 include/linux/sched.h |    2 ++
 kernel/sched/fair.c   |   41 +++++++++++++++++++++++++++++++++++++++++
 2 files changed, 43 insertions(+), 0 deletions(-)

diff --git a/include/linux/sched.h b/include/linux/sched.h
index 64d9df5..46cc4a7 100644
--- a/include/linux/sched.h
+++ b/include/linux/sched.h

@@ -2806,4 +2806,6 @@ static inline unsigned long rlimit_max(unsigned int limit)
 
 #endif /* __KERNEL__ */
 
+extern int find_idlest_prefer_cpu(struct cpumask *prefer,
+				 struct cpumask *allowed, int prev_cpu);
 #endif

diff --git a/kernel/sched/fair.c b/kernel/sched/fair.c
index c099cc6..d3da151 100644
--- a/kernel/sched/fair.c
+++ b/kernel/sched/fair.c

@@ -26,6 +26,7 @@
 #include <linux/slab.h>
 #include <linux/profile.h>
 #include <linux/interrupt.h>
+#include <linux/export.h>
 
 #include <trace/events/sched.h>
 

@@ -2809,6 +2810,46 @@ unlock:
 
 	return new_cpu;
 }
+
+/*
+ * This API is used to find the most idle cpu from both preferred and
+ * allowed cpuset.
+ *
+ * allowed: The allowed cpumask is the caller's task allowed cpuset, which could * be from cgroup cpuset control.
+ * prefer: The perfer cpumask is caller's perferred cpuset choice, which could
+ * be a cpuset of a NUMA node for better performance.
+ *
+ * It helps per-cpu thread model to choose the preferred cpu in the allowed
+ * cpuset to be scheduled when the work doesn't want to be scheduled on the
+ * same cpu on which the work is received for better performance. For example
+ * the network work doesn't want to be on the same cpu on which the
+ * interrupt is received.
+ *
+ * If these two cpusets have intersects, the cpu is chose from the intersects;
+ * if there is no intersects, then the cpu is chose from the allowed cpuset.
+ *
+ * prev_cpu helps to better local cache when prev_cpu is not busy.
+ */
+int find_idlest_prefer_cpu(struct cpumask *prefer, struct cpumask *allowed,
+			  int prev_cpu)
+{
+	unsigned long load, min_load = ULONG_MAX;
+	int check, i, idlest = -1;
+
+	check = cpumask_intersects(prefer, allowed);
+	/* Traverse only the allowed CPUs */
+	if (check == 0)
+		prefer = allowed;
+	for_each_cpu_and(i, prefer, allowed) {
+		load = weighted_cpuload(i);
+		if (load < min_load || (load == min_load && i == prev_cpu)) {
+			min_load = load;
+			idlest = i;
+		}
+	}
+	return idlest;
+}
+EXPORT_SYMBOL(find_idlest_prefer_cpu);
 #endif /* CONFIG_SMP */
 
 static unsigned long

thanks
Shirley

On Fri, 2012-08-17 at 12:46 -0700, Shirley Ma wrote:

Add/Export a new API for per-cpu thread model networking device driver
to choose a preferred idlest cpu within allowed cpumask.

The receiving CPUs of a networking device are not under cgroup controls.
Normally the receiving work will be scheduled on the cpu on which the
interrupts are received. When such a networking device uses per-cpu
thread model, the cpu which is chose to process the packets might not be
part of cgroup cpusets without using such an API here. 

On NUMA system, by using the preferred cpumask from the same NUMA node
would help to reduce expensive cross memory access to/from the other
NUMA node.

KVM per-cpu vhost will be the first one to use this API. Any other
device driver which uses per-cpu thread model and has cgroup cpuset
control will use this API later.

How often will this be called and how do you obtain the cpumasks
provided to the function?

On Mon, 2012-08-20 at 14:00 +0200, Peter Zijlstra wrote:

On Fri, 2012-08-17 at 12:46 -0700, Shirley Ma wrote:

quoted

Add/Export a new API for per-cpu thread model networking device

driver

quoted

to choose a preferred idlest cpu within allowed cpumask.

The receiving CPUs of a networking device are not under cgroup

controls.

quoted

Normally the receiving work will be scheduled on the cpu on which

the

quoted

interrupts are received. When such a networking device uses per-cpu
thread model, the cpu which is chose to process the packets might

not be

quoted

part of cgroup cpusets without using such an API here. 

On NUMA system, by using the preferred cpumask from the same NUMA

node

quoted

would help to reduce expensive cross memory access to/from the other
NUMA node.

KVM per-cpu vhost will be the first one to use this API. Any other
device driver which uses per-cpu thread model and has cgroup cpuset
control will use this API later.

How often will this be called and how do you obtain the cpumasks
provided to the function? 

It depends. It might be called pretty often if the user keeps changing
cgroups control cpuset. It might be less called if the cgroups control
cpuset is stable, and the host scheduler always schedules the work on
the same NUMA node.

The preferred cpumasks are obtained from local numa node. The allowed
cpumasks are obtained from caller's task allowed cpumasks (cgroups
control cpuset).

Thanks
Shirley

On Mon, 2012-08-20 at 15:17 -0700, Shirley Ma wrote:

On Mon, 2012-08-20 at 14:00 +0200, Peter Zijlstra wrote:

quoted

On Fri, 2012-08-17 at 12:46 -0700, Shirley Ma wrote:

quoted

Add/Export a new API for per-cpu thread model networking device

driver

quoted

to choose a preferred idlest cpu within allowed cpumask.

The receiving CPUs of a networking device are not under cgroup

controls.

quoted

Normally the receiving work will be scheduled on the cpu on which

the

quoted

interrupts are received. When such a networking device uses per-cpu
thread model, the cpu which is chose to process the packets might

not be

quoted

part of cgroup cpusets without using such an API here. 

On NUMA system, by using the preferred cpumask from the same NUMA

node

quoted

would help to reduce expensive cross memory access to/from the other
NUMA node.

KVM per-cpu vhost will be the first one to use this API. Any other
device driver which uses per-cpu thread model and has cgroup cpuset
control will use this API later.

How often will this be called and how do you obtain the cpumasks
provided to the function? 

It depends. It might be called pretty often if the user keeps changing
cgroups control cpuset. It might be less called if the cgroups control
cpuset is stable, and the host scheduler always schedules the work on
the same NUMA node.

This just doesn't make any sense, you're scanning for the least loaded
cpu, this is unrelated to a change in cpuset. So tying the scan
frequency to changes in configuration is just broken.

The preferred cpumasks are obtained from local numa node.

So why pass it as argument at all? Also, who says the current node is
the right one? It might just be running there temporarily.

 The allowed
cpumasks are obtained from caller's task allowed cpumasks (cgroups
control cpuset).

task->cpus_allowed != cpusets.. Also, since you're using
task->cpus_allowed, pass a task_struct *, not a cpumask.

On Tue, 2012-08-21 at 09:07 +0200, Peter Zijlstra wrote:

On Mon, 2012-08-20 at 15:17 -0700, Shirley Ma wrote:

quoted

On Mon, 2012-08-20 at 14:00 +0200, Peter Zijlstra wrote:

quoted

On Fri, 2012-08-17 at 12:46 -0700, Shirley Ma wrote:

quoted

Add/Export a new API for per-cpu thread model networking device

driver

quoted

to choose a preferred idlest cpu within allowed cpumask.

The receiving CPUs of a networking device are not under cgroup

controls.

quoted

Normally the receiving work will be scheduled on the cpu on

which

quoted

quoted

the

quoted

interrupts are received. When such a networking device uses

per-cpu

quoted

quoted

quoted

thread model, the cpu which is chose to process the packets

might

quoted

quoted

not be

quoted

part of cgroup cpusets without using such an API here. 

On NUMA system, by using the preferred cpumask from the same

NUMA

quoted

quoted

node

quoted

would help to reduce expensive cross memory access to/from the

other

quoted

quoted

quoted

NUMA node.

KVM per-cpu vhost will be the first one to use this API. Any

other

quoted

quoted

quoted

device driver which uses per-cpu thread model and has cgroup

cpuset

quoted

quoted

quoted

control will use this API later.

How often will this be called and how do you obtain the cpumasks
provided to the function? 

It depends. It might be called pretty often if the user keeps

changing

quoted

cgroups control cpuset. It might be less called if the cgroups

control

quoted

cpuset is stable, and the host scheduler always schedules the work

on

quoted

the same NUMA node.

This just doesn't make any sense, you're scanning for the least loaded
cpu, this is unrelated to a change in cpuset. So tying the scan
frequency to changes in configuration is just broken.

Thanks for your review. I am just back from my vacation. 

Why not? the caller knows the cpuset changes, and pass the right NUMA
node to choose the idlest cpu from that NUMA node. Practically, the VMs
don't change the cgroups. So it will not frequency to change the
configuration.

quoted

The preferred cpumasks are obtained from local numa node.

So why pass it as argument at all? Also, who says the current node is
the right one? It might just be running there temporarily.

It leaves to the caller to make the right node choice. It tries to avoid
VMs running on the same cpu but on the same node with the host to
process the guest network packets.

quoted

 The allowed
cpumasks are obtained from caller's task allowed cpumasks (cgroups
control cpuset).

task->cpus_allowed != cpusets.. Also, since you're using
task->cpus_allowed, pass a task_struct *, not a cpumask. 

Based on the documentation I read before, I thought the cpus_allowed ==
cgroups control cpuset. If not, where are the cgroups control cpusets
saved?

task->cpus_allowed = tsk_cpus_allowed(task_struct *p), which is
cpumask_t.

I can change the argument from cpumask to task_struct *, and call
tsk_cpus_allowed() instead of using task->cpus_allowed.

Thanks
Shirley