Re: [PATCH v3 7/8] cpufreq: amd-pstate: add frequency dynamic boost sysfs control

From: Limonciello, Mario
Date: Mon Nov 07 2022 - 14:10:35 EST


On 11/7/2022 12:16, Limonciello, Mario wrote:
+wyes.karny@xxxxxxx

On 11/7/2022 11:57, Perry Yuan wrote:
Add one sysfs entry to control the CPU cores frequency boost state
The attribute file can allow user to set max performance boosted or
keeping at normal perf level.

Signed-off-by: Perry Yuan <Perry.Yuan@xxxxxxx>
---
  drivers/cpufreq/amd-pstate.c | 53 ++++++++++++++++++++++++++++++++++++
  1 file changed, 53 insertions(+)

Make sure that you update the documentation for this new sysfs file as well.


diff --git a/drivers/cpufreq/amd-pstate.c b/drivers/cpufreq/amd-pstate.c
index 841b1e2383b8..6958c5fd9e1c 100644
--- a/drivers/cpufreq/amd-pstate.c
+++ b/drivers/cpufreq/amd-pstate.c
@@ -786,12 +786,46 @@ static ssize_t show_energy_performance_preference(
          return  sprintf(buf, "%s\n", energy_perf_strings[preference]);
  }
+static void amd_pstate_update_policies(void)
+{
+    int cpu;
+
+    for_each_possible_cpu(cpu)
+        cpufreq_update_policy(cpu);
+}
+
+static ssize_t show_pstate_dynamic_boost(struct kobject *kobj,
+                struct kobj_attribute *attr, char *buf)
+{
+    return sprintf(buf, "%u\n", cppc_boost);
+}
+
+static ssize_t store_pstate_dynamic_boost(struct kobject *a,
+                       struct kobj_attribute *b,
+                       const char *buf, size_t count)
+{
+    unsigned int input;
+    int ret;
+
+    ret = kstrtouint(buf, 10, &input);
+    if (ret)
+        return ret;

To be more flexible maybe kstrtobool would be better here instead?

+
+    mutex_lock(&amd_pstate_driver_lock);
+    cppc_boost = !!input;
+    amd_pstate_update_policies();
+    mutex_unlock(&amd_pstate_driver_lock);
+
+    return count;
+}
+
  cpufreq_freq_attr_ro(amd_pstate_max_freq);
  cpufreq_freq_attr_ro(amd_pstate_lowest_nonlinear_freq);
  cpufreq_freq_attr_ro(amd_pstate_highest_perf);
  cpufreq_freq_attr_rw(energy_performance_preference);
  cpufreq_freq_attr_ro(energy_performance_available_preferences);
+define_one_global_rw(pstate_dynamic_boost);
  static struct freq_attr *amd_pstate_attr[] = {
      &amd_pstate_max_freq,
@@ -809,6 +843,15 @@ static struct freq_attr *amd_pstate_epp_attr[] = {
      NULL,
  };
+static struct attribute *pstate_global_attributes[] = {
+    &pstate_dynamic_boost.attr,
+    NULL
+};
+
+static const struct attribute_group amd_pstate_global_attr_group = {
+    .attrs = pstate_global_attributes,
+};
+
  static inline void update_boost_state(void)
  {
      u64 misc_en;
@@ -1416,6 +1459,16 @@ static int __init amd_pstate_init(void)
          pr_err("failed to register amd pstate driver with return %d\n",
                 ret);
+    amd_pstate_kobj = kobject_create_and_add("amd-pstate", &cpu_subsys.dev_root->kobj);
+    if (!amd_pstate_kobj)
+        pr_err("amd-pstate: Global sysfs registration failed.\n");
+
+    ret = sysfs_create_group(amd_pstate_kobj, &amd_pstate_global_attr_group);
+    if (ret) {
+        pr_err("amd-pstate: Sysfs attribute export failed with error %d.\n",
+               ret);

amd-pstate can currently be a module too.  So don't you need the matching cleanup path for this code too?

I noticed your patch 3/8 makes this change, so this specific comment might not be as important anymore.


Also, regarding the error messages you don't need to prefix with "amd-pstate: ", amd-pstate.c already sets `pr_fmt`.

Also, shouldn't failing to make these attributes fail the rest of amd-pstate init rather than just showing error messages?


+    }
+
      return ret;
  }
  device_initcall(amd_pstate_init);