drivers base/arch_topology: enforce SCHED_CAPACITY_SCALE as highest CPU capacity

The default CPU capacity is SCHED_CAPACITY_SCALE (1024).

On a heterogeneous system (hmp) this value can be smaller for some cpus.
The CPU capacity parsing code normalizes the capacity-dmips-mhz
properties w.r.t. the highest value found while parsing the DT to
SCHED_CAPACITY_SCALE.

CPU capacity can also be changed by writing to
/sys/devices/system/cpu/cpu*/cpu_capacity.

To make sure that a subset of all online cpus still has a CPU capacity
value of SCHED_CAPACITY_SCALE enforce in the appropriate sysfs attribute
store function cpu_capacity_store().

This will avoid weird setup's like transforming an hmp into an smp
system with a CPU capacity < SCHED_CAPACITY_SCALE for all cpus.

The current cpu_capacity_store() assumes that all cpus of a cluster have
the same CPU capacity value which is true for existing hmp systems (e.g.
big.LITTLE). This assumption is also used by this patch.
If the new CPU capacity value for a cpu is smaller than
SCHED_CAPACITY_SCALE we iterate over the cpus which do not belong to the
cpu's cluster and check that there is still a cpu with CPU capacity
equal SCHED_CAPACITY_SCALE.

The use of &cpu_topology[this_cpu].core_sibling is replaced by
topology_core_cpumask(this_cpu).

Signed-off-by: Dietmar Eggemann <dietmar.eggemann@arm.com>
Signed-off-by: Lukasz Luba <l.luba@partner.samsung.com>

diff --git a/drivers/base/arch_topology.c b/drivers/base/arch_topology.c
index e6341ac7b3945e272552b0db9b884b1d1f946a5e..473f67c40a3cedb2da1ee2afc728a120f53b0043 100644 (file)
--- a/drivers/base/arch_topology.c
+++ b/drivers/base/arch_topology.c
@@ -67,6 +67,7 @@ static ssize_t cpu_capacity_store(struct device *dev,
        int i;
        unsigned long new_capacity;
        ssize_t ret;
+       cpumask_var_t mask;
 
        if (!count)
                return 0;
@@ -78,7 +79,35 @@ static ssize_t cpu_capacity_store(struct device *dev,
                return -EINVAL;
 
        mutex_lock(&cpu_scale_mutex);
-       for_each_cpu(i, &cpu_topology[this_cpu].core_sibling)
+
+       if (new_capacity < SCHED_CAPACITY_SCALE) {
+               int highest_score_cpu = 0;
+
+               if (!alloc_cpumask_var(&mask, GFP_KERNEL)) {
+                       mutex_unlock(&cpu_scale_mutex);
+                       return -ENOMEM;
+               }
+
+               cpumask_andnot(mask, cpu_online_mask,
+                               topology_core_cpumask(this_cpu));
+
+               for_each_cpu(i, mask) {
+                       if (topology_get_cpu_scale(NULL, i) ==
+                                       SCHED_CAPACITY_SCALE) {
+                               highest_score_cpu = 1;
+                               break;
+                       }
+               }
+
+               free_cpumask_var(mask);
+
+               if (!highest_score_cpu) {
+                       mutex_unlock(&cpu_scale_mutex);
+                       return -EINVAL;
+               }
+       }
+
+       for_each_cpu(i, topology_core_cpumask(this_cpu))
                topology_set_cpu_scale(i, new_capacity);
        mutex_unlock(&cpu_scale_mutex);