Re: [PATCH 2/6] drivers base/arch_topology: frequency-invariant load-tracking support

[Dietmar Eggemann <dietmar.eggemann@arm.com>]

On 08/06/17 08:55, Dietmar Eggemann wrote:
> Implements an arch-specific frequency-scaling function
> topology_get_freq_scale() which provides the following frequency
> scaling factor:
> 
>   current_freq(cpu) << SCHED_CAPACITY_SHIFT / max_supported_freq(cpu)

[...]

Frequency and cpu-invariant load tracking are part of the task
schedulers hot path:

e.g.:
 
 __update_load_avg_se()-> ___update_load_avg() -> accumulate_sum()

That's why function calls should be avoided here.

I would like to fold the following changes into patch 2/6 in v2:

commit 1397770fe47ce5d34511e7062bd3a8bc96a74590
Author: Dietmar Eggemann <dietmar.eggemann@arm.com>
Date:   Sat Jun 24 16:46:45 2017 +0100

    drivers base/arch_topology: eliminate function call for cpu and
    frequency-invariant accounting
    
    topology_get_cpu_scale() and topology_get_freq_scale() are the arm/arm64
    architecture specific implementations to provide cpu-invariant and
    frequency-invariant accounting support up to the task scheduler.
    
    Define them as static inline functions to allow cpu-invariant and
    frequency-invariant accounting to happen without an extra function call
    involved.
    
    Test results on JUNO (arm64):
    
    root@juno:~# grep
    "__update_load_avg_\|update_group_capacity\|topology_get"
    available_filter_functions > set_ftrace_filter
    
    root@juno:~# echo function_graph > current_tracer
    
    root@juno:~# cat trace | tail -50
    
    w/ this patch:
    
     ...
     3)   0.700 us    |  __update_load_avg_se.isra.5();
     ...
     3)   0.750 us    |  __update_load_avg_cfs_rq();
     ...
     3)   0.780 us    |  update_group_capacity();
     ...
    
    w/o this patch:
    
     4)               |  __update_load_avg_cfs_rq() {
     4)   0.380 us    |    topology_get_freq_scale();
     4)   0.340 us    |    topology_get_cpu_scale();
     4)   6.420 us    |  }
     ...
     4)               |  __update_load_avg_se.isra.4() {
     4)   0.300 us    |    topology_get_freq_scale();
     4)   0.260 us    |    topology_get_cpu_scale();
     4)   5.800 us    |  }
     ...
     4)               |  update_group_capacity() {
     4)   0.260 us    |    topology_get_cpu_scale();
     4)   3.540 us    |  }
     ...
    
    So these extra function calls cost ~2.5us each (on Cortex A53,
    cpu0,3,4,5). Since this happens in the task scheduler hot-path,
    they have to be avoided.
    
    Signed-off-by: Dietmar Eggemann <dietmar.eggemann@arm.com>

diff --git a/drivers/base/arch_topology.c b/drivers/base/arch_topology.c
index d7e130c268fb..8dfa4c3dbfc2 100644
--- a/drivers/base/arch_topology.c
+++ b/drivers/base/arch_topology.c
@@ -23,18 +23,8 @@
 #include <linux/sched/topology.h>
 
 static DEFINE_MUTEX(cpu_scale_mutex);
-static DEFINE_PER_CPU(unsigned long, cpu_scale) = SCHED_CAPACITY_SCALE;
-static DEFINE_PER_CPU(unsigned long, freq_scale) = SCHED_CAPACITY_SCALE;
-
-unsigned long topology_get_cpu_scale(struct sched_domain *sd, int cpu)
-{
-       return per_cpu(cpu_scale, cpu);
-}
-
-unsigned long topology_get_freq_scale(struct sched_domain *sd, int cpu)
-{
-       return per_cpu(freq_scale, cpu);
-}
+DEFINE_PER_CPU(unsigned long, cpu_scale) = SCHED_CAPACITY_SCALE;
+DEFINE_PER_CPU(unsigned long, freq_scale) = SCHED_CAPACITY_SCALE;
 
 void topology_set_cpu_scale(unsigned int cpu, unsigned long capacity)
 {
diff --git a/include/linux/arch_topology.h b/include/linux/arch_topology.h
index 3fb4d8ccb179..cf22631e6765 100644
--- a/include/linux/arch_topology.h
+++ b/include/linux/arch_topology.h
@@ -9,10 +9,21 @@ void topology_normalize_cpu_scale(void);
 struct device_node;
 int topology_parse_cpu_capacity(struct device_node *cpu_node, int cpu);
 
+DECLARE_PER_CPU(unsigned long, cpu_scale);
+DECLARE_PER_CPU(unsigned long, freq_scale);
+
 struct sched_domain;
-unsigned long topology_get_cpu_scale(struct sched_domain *sd, int cpu);
+static inline
+unsigned long topology_get_cpu_scale(struct sched_domain *sd, int cpu)
+{
+       return per_cpu(cpu_scale, cpu);
+}
 
-unsigned long topology_get_freq_scale(struct sched_domain *sd, int cpu);
+static inline
+unsigned long topology_get_freq_scale(struct sched_domain *sd, int cpu)
+{
+       return per_cpu(freq_scale, cpu);
+}
 
 void topology_set_cpu_scale(unsigned int cpu, unsigned long capacity);

[...]