We have a per-cpu variable for managing which cluster does a CPU belong to. Currently physical_cluster is set only for the policy->cpu. And that results in following on some SoC's:
- There are two clusters: - Cluster 0 has four ARM Cortex A7 CPUs (slower ones): 0,1,2,3 - Cluster 1 has four ARM Cortex A15 CPUs (faster ones): 4,5,6,7 - CPUs are booted in order 0,1..7 and so initially policy->cpu for A7 cluster would be 0 and for A15 cluster would be 4. - Now CPU4 (i.e. A15_0) is hotplugged out and so policy->cpu for A15 cluster becomes 5 (i.e. A15_1). - But physical cluster is only set for CPU0 and CPU4 in ARM big LITTLE driver and isn't updated. - Now freq change request comes for A15 cluster and we would try to update freq of physical_cluster of CPU5, i.e. A15_1. And it is currently set to zero (default value of uninitialized global variables). - And so we actually try to change freq of A7 cluster instead of A15. - This also results in kernel crash as sometimes we might request freq above A7's limit and CPU may behave badly..
Fix this by initializing physical_cluster for all CPUs of a policy.
(Adding signed-off by Xin as he reported this issue and provided this diff)
Signed-off-by: Xin Wang xin.wang@arm.com Signed-off-by: Viresh Kumar viresh.kumar@linaro.org --- drivers/cpufreq/arm_big_little.c | 5 ++++- 1 file changed, 4 insertions(+), 1 deletion(-)
diff --git a/drivers/cpufreq/arm_big_little.c b/drivers/cpufreq/arm_big_little.c index 3d87078..bad2ed3 100644 --- a/drivers/cpufreq/arm_big_little.c +++ b/drivers/cpufreq/arm_big_little.c @@ -446,9 +446,12 @@ static int bL_cpufreq_init(struct cpufreq_policy *policy) }
if (cur_cluster < MAX_CLUSTERS) { + int cpu; + cpumask_copy(policy->cpus, topology_core_cpumask(policy->cpu));
- per_cpu(physical_cluster, policy->cpu) = cur_cluster; + for_each_cpu(cpu, policy->cpus) + per_cpu(physical_cluster, cpu) = cur_cluster; } else { /* Assumption: during init, we are always running on A15 */ per_cpu(physical_cluster, policy->cpu) = A15_CLUSTER;