Re: [PATCH v2] cpufreq: powernv: Add support of frequency domain

[Abhishek]

On 12/20/2017 12:20 PM, Viresh Kumar wrote:

On 20-12-17, 12:12, Abhishek Goel wrote:

diff --git a/drivers/cpufreq/powernv-cpufreq.c 
b/drivers/cpufreq/powernv-cpufreq.c
index b6d7c4c..fd642bc 100644
--- a/drivers/cpufreq/powernv-cpufreq.c
+++ b/drivers/cpufreq/powernv-cpufreq.c
@@ -37,6 +37,7 @@
  #include  /* Required for cpu_sibling_mask() in UP configs */
  #include 
  #include 
+#include 
  
  #define POWERNV_MAX_PSTATES	256

  #define PMSR_PSAFE_ENABLE (1UL << 30)
@@ -130,6 +131,9 @@ static struct chip {
  static int nr_chips;
  static DEFINE_PER_CPU(struct chip *, chip_info);
  
+static u32 freq_domain_indicator;

+static u32 flag;

I wouldn't name it as flag, its unreadable. Maybe its better to name
it based on the quirk you are trying to workaround with ?


+
  /*
   * Note:
   * The set of pstates consists of contiguous integers.
@@ -194,6 +198,38 @@ static inline void reset_gpstates(struct cpufreq_policy 
*policy)
gpstates->last_gpstate_idx = 0;
  }
  
+#define SIZE NR_CPUS

+#define ORDER_FREQ_MAP ilog2(SIZE)
+
+static DEFINE_HASHTABLE(freq_domain_map, ORDER_FREQ_MAP);
+
+struct hashmap {
+   cpumask_t mask;
+   int chip_id;
+   u32 pir_key;
+   struct hlist_node hash_node;
+};
+
+static void insert(u32 key, int cpu)
+{
+   struct hashmap *data;
+
+   hash_for_each_possible(freq_domain_map, data, hash_node, key%SIZE) {
+   if (data->chip_id == cpu_to_chip_id(cpu) &&
+   data->pir_key == key) {
+   cpumask_set_cpu(cpu, >mask);
+   return;
+   }
+   }
+
+   data = kzalloc(sizeof(*data), GFP_KERNEL);
+   hash_add(freq_domain_map, >hash_node, key%SIZE);
+   cpumask_set_cpu(cpu, >mask);
+   data->chip_id = cpu_to_chip_id(cpu);
+   data->pir_key = key;
+
+}
+
  /*
   * Initialize the freq table based on data obtained
   * from the firmware passed via device-tree
@@ -206,7 +242,9 @@ static int init_powernv_pstates(void)
u32 len_ids, len_freqs;
u32 pstate_min, pstate_max, pstate_nominal;
u32 pstate_turbo, pstate_ultra_turbo;
+   u32 key;
  
+	flag = 0;

Isn't flag already 0 (global-uninitialized) ?


power_mgt = of_find_node_by_path("/ibm,opal/power-mgt");
if (!power_mgt) {
pr_warn("power-mgt node not found\n");
@@ -229,6 +267,17 @@ static int init_powernv_pstates(void)
return -ENODEV;
}
  
+	if (of_device_is_compatible(power_mgt, "freq-domain-v1") &&

+   of_property_read_u32(power_mgt, "ibm,freq-domain-indicator",
+_domain_indicator)) {
+   pr_warn("ibm,freq-domain-indicator not found\n");
+   freq_domain_indicator = 0;

You shouldn't be required to set it to 0 here.


+   }
+
+   if (of_device_is_compatible(power_mgt, "P9-occ-quirk")) {
+   flag = 1;
+   }

Remove {} and a better name like p9_occ_quirk would be good for flag.
Also making it a bool may be better ?


+
if (of_property_read_u32(power_mgt, "ibm,pstate-ultra-turbo",
 _ultra_turbo)) {
powernv_pstate_info.wof_enabled = false;
@@ -249,6 +298,7 @@ static int init_powernv_pstates(void)
  next:
pr_info("cpufreq pstate min %d nominal %d max %d\n", pstate_min,
pstate_nominal, pstate_max);
+   pr_info("frequency domain indicator %d", freq_domain_indicator);
pr_info("Workload Optimized Frequency is %s in the platform\n",
(powernv_pstate_info.wof_enabled) ? "enabled" : "disabled");
  
@@ -276,6 +326,15 @@ static int init_powernv_pstates(void)

return -ENODEV;
}
  
+	if (freq_domain_indicator) {

+   hash_init(freq_domain_map);
+   for_each_possible_cpu(i) {
+   key = ((u32) get_hard_smp_processor_id(i) &
+   freq_domain_indicator);

Maybe break it like:

key = (u32) get_hard_smp_processor_id(i);
 key &= freq_domain_indicator;

to make it easily readable ?


+   insert(key, i);
+   }
+   }
+
powernv_pstate_info.nr_pstates = nr_pstates;
pr_debug("NR PStates %d\n", nr_pstates);
for (i = 0; i < nr_pstates; i++) {
@@ -693,6 +752,7 @@ static int powernv_cpufreq_target_index(struct 
cpufreq_policy *policy,
  {
struct powernv_smp_call_data freq_data;
unsigned int cur_msec, gpstate_idx;
+

:(


struct global_pstate_info *gpstates = policy->driver_data;
  
  	if (unlikely(rebooting) && new_index != get_nominal_index())

@@ -760,25 +820,55 @@ static int powernv_cpufreq_target_index(struct 
cpufreq_policy *policy,
  
  	spin_unlock(>gpstate_lock);
  
-	/*

-* Use smp_call_function to send IPI and execute the
-* mtspr on target CPU.  We could do that without IPI
-* if current CPU is within 

Re: [PATCH v2] cpufreq: powernv: Add support of frequency domain

[Viresh Kumar]

On 20-12-17, 12:12, Abhishek Goel wrote:
> diff --git a/drivers/cpufreq/powernv-cpufreq.c 
> b/drivers/cpufreq/powernv-cpufreq.c
> index b6d7c4c..fd642bc 100644
> --- a/drivers/cpufreq/powernv-cpufreq.c
> +++ b/drivers/cpufreq/powernv-cpufreq.c
> @@ -37,6 +37,7 @@
>  #include  /* Required for cpu_sibling_mask() in UP configs */
>  #include 
>  #include 
> +#include 
>  
>  #define POWERNV_MAX_PSTATES  256
>  #define PMSR_PSAFE_ENABLE(1UL << 30)
> @@ -130,6 +131,9 @@ static struct chip {
>  static int nr_chips;
>  static DEFINE_PER_CPU(struct chip *, chip_info);
>  
> +static u32 freq_domain_indicator;
> +static u32 flag;

I wouldn't name it as flag, its unreadable. Maybe its better to name
it based on the quirk you are trying to workaround with ?

> +
>  /*
>   * Note:
>   * The set of pstates consists of contiguous integers.
> @@ -194,6 +198,38 @@ static inline void reset_gpstates(struct cpufreq_policy 
> *policy)
>   gpstates->last_gpstate_idx = 0;
>  }
>  
> +#define SIZE NR_CPUS
> +#define ORDER_FREQ_MAP ilog2(SIZE)
> +
> +static DEFINE_HASHTABLE(freq_domain_map, ORDER_FREQ_MAP);
> +
> +struct hashmap {
> + cpumask_t mask;
> + int chip_id;
> + u32 pir_key;
> + struct hlist_node hash_node;
> +};
> +
> +static void insert(u32 key, int cpu)
> +{
> + struct hashmap *data;
> +
> + hash_for_each_possible(freq_domain_map, data, hash_node, key%SIZE) {
> + if (data->chip_id == cpu_to_chip_id(cpu) &&
> + data->pir_key == key) {
> + cpumask_set_cpu(cpu, >mask);
> + return;
> + }
> + }
> +
> + data = kzalloc(sizeof(*data), GFP_KERNEL);
> + hash_add(freq_domain_map, >hash_node, key%SIZE);
> + cpumask_set_cpu(cpu, >mask);
> + data->chip_id = cpu_to_chip_id(cpu);
> + data->pir_key = key;
> +
> +}
> +
>  /*
>   * Initialize the freq table based on data obtained
>   * from the firmware passed via device-tree
> @@ -206,7 +242,9 @@ static int init_powernv_pstates(void)
>   u32 len_ids, len_freqs;
>   u32 pstate_min, pstate_max, pstate_nominal;
>   u32 pstate_turbo, pstate_ultra_turbo;
> + u32 key;
>  
> + flag = 0;

Isn't flag already 0 (global-uninitialized) ?

>   power_mgt = of_find_node_by_path("/ibm,opal/power-mgt");
>   if (!power_mgt) {
>   pr_warn("power-mgt node not found\n");
> @@ -229,6 +267,17 @@ static int init_powernv_pstates(void)
>   return -ENODEV;
>   }
>  
> + if (of_device_is_compatible(power_mgt, "freq-domain-v1") &&
> + of_property_read_u32(power_mgt, "ibm,freq-domain-indicator",
> +  _domain_indicator)) {
> + pr_warn("ibm,freq-domain-indicator not found\n");
> + freq_domain_indicator = 0;

You shouldn't be required to set it to 0 here.

> + }
> +
> + if (of_device_is_compatible(power_mgt, "P9-occ-quirk")) {
> + flag = 1;
> + }

Remove {} and a better name like p9_occ_quirk would be good for flag.
Also making it a bool may be better ?

> +
>   if (of_property_read_u32(power_mgt, "ibm,pstate-ultra-turbo",
>_ultra_turbo)) {
>   powernv_pstate_info.wof_enabled = false;
> @@ -249,6 +298,7 @@ static int init_powernv_pstates(void)
>  next:
>   pr_info("cpufreq pstate min %d nominal %d max %d\n", pstate_min,
>   pstate_nominal, pstate_max);
> + pr_info("frequency domain indicator %d", freq_domain_indicator);
>   pr_info("Workload Optimized Frequency is %s in the platform\n",
>   (powernv_pstate_info.wof_enabled) ? "enabled" : "disabled");
>  
> @@ -276,6 +326,15 @@ static int init_powernv_pstates(void)
>   return -ENODEV;
>   }
>  
> + if (freq_domain_indicator) {
> + hash_init(freq_domain_map);
> + for_each_possible_cpu(i) {
> + key = ((u32) get_hard_smp_processor_id(i) &
> + freq_domain_indicator);

Maybe break it like:

key = (u32) get_hard_smp_processor_id(i);
key &= freq_domain_indicator;

to make it easily readable ?

> + insert(key, i);
> + }
> + }
> +
>   powernv_pstate_info.nr_pstates = nr_pstates;
>   pr_debug("NR PStates %d\n", nr_pstates);
>   for (i = 0; i < nr_pstates; i++) {
> @@ -693,6 +752,7 @@ static int powernv_cpufreq_target_index(struct 
> cpufreq_policy *policy,
>  {
>   struct powernv_smp_call_data freq_data;
>   unsigned int cur_msec, gpstate_idx;
> +

:(

>   struct global_pstate_info *gpstates = policy->driver_data;
>  
>   if (unlikely(rebooting) && new_index != get_nominal_index())
> @@ -760,25 +820,55 @@ static int powernv_cpufreq_target_index(struct 
> cpufreq_policy *policy,
>  
>   spin_unlock(>gpstate_lock);
>  
> - /*
> -  * Use smp_call_function to send IPI and execute the
> -  * mtspr on 

[PATCH v2] cpufreq: powernv: Add support of frequency domain

[Abhishek Goel]

Frequency-domain indicates group of CPUs that would share same frequency.
It is detected using device-tree node "frequency-domain-indicator".
frequency-domain-indicator is a bitmask which will have different value
depending upon the generation of the processor.

CPUs of the same chip for which the result of a bitwise AND between
their PIR and the frequency-domain-indicator is the same share the same
frequency.

In this patch, we define hash-table indexed by the aforementioned
bitwise ANDed value to store the cpumask of the CPUs sharing the same
frequency domain. Further, the cpufreq policy will be created per
frequency-domain

So for POWER9, a cpufreq policy is created per quad while for POWER8 it
is created per core. Governor decides frequency for each policy but
multiple cores may come under same policy. In such case frequency needs
to be set on each core sharing that policy.

Signed-off-by: Abhishek Goel 
---
 drivers/cpufreq/powernv-cpufreq.c | 110 ++
 1 file changed, 100 insertions(+), 10 deletions(-)

diff --git a/drivers/cpufreq/powernv-cpufreq.c 
b/drivers/cpufreq/powernv-cpufreq.c
index b6d7c4c..fd642bc 100644
--- a/drivers/cpufreq/powernv-cpufreq.c
+++ b/drivers/cpufreq/powernv-cpufreq.c
@@ -37,6 +37,7 @@
 #include  /* Required for cpu_sibling_mask() in UP configs */
 #include 
 #include 
+#include 
 
 #define POWERNV_MAX_PSTATES256
 #define PMSR_PSAFE_ENABLE  (1UL << 30)
@@ -130,6 +131,9 @@ static struct chip {
 static int nr_chips;
 static DEFINE_PER_CPU(struct chip *, chip_info);
 
+static u32 freq_domain_indicator;
+static u32 flag;
+
 /*
  * Note:
  * The set of pstates consists of contiguous integers.
@@ -194,6 +198,38 @@ static inline void reset_gpstates(struct cpufreq_policy 
*policy)
gpstates->last_gpstate_idx = 0;
 }
 
+#define SIZE NR_CPUS
+#define ORDER_FREQ_MAP ilog2(SIZE)
+
+static DEFINE_HASHTABLE(freq_domain_map, ORDER_FREQ_MAP);
+
+struct hashmap {
+   cpumask_t mask;
+   int chip_id;
+   u32 pir_key;
+   struct hlist_node hash_node;
+};
+
+static void insert(u32 key, int cpu)
+{
+   struct hashmap *data;
+
+   hash_for_each_possible(freq_domain_map, data, hash_node, key%SIZE) {
+   if (data->chip_id == cpu_to_chip_id(cpu) &&
+   data->pir_key == key) {
+   cpumask_set_cpu(cpu, >mask);
+   return;
+   }
+   }
+
+   data = kzalloc(sizeof(*data), GFP_KERNEL);
+   hash_add(freq_domain_map, >hash_node, key%SIZE);
+   cpumask_set_cpu(cpu, >mask);
+   data->chip_id = cpu_to_chip_id(cpu);
+   data->pir_key = key;
+
+}
+
 /*
  * Initialize the freq table based on data obtained
  * from the firmware passed via device-tree
@@ -206,7 +242,9 @@ static int init_powernv_pstates(void)
u32 len_ids, len_freqs;
u32 pstate_min, pstate_max, pstate_nominal;
u32 pstate_turbo, pstate_ultra_turbo;
+   u32 key;
 
+   flag = 0;
power_mgt = of_find_node_by_path("/ibm,opal/power-mgt");
if (!power_mgt) {
pr_warn("power-mgt node not found\n");
@@ -229,6 +267,17 @@ static int init_powernv_pstates(void)
return -ENODEV;
}
 
+   if (of_device_is_compatible(power_mgt, "freq-domain-v1") &&
+   of_property_read_u32(power_mgt, "ibm,freq-domain-indicator",
+_domain_indicator)) {
+   pr_warn("ibm,freq-domain-indicator not found\n");
+   freq_domain_indicator = 0;
+   }
+
+   if (of_device_is_compatible(power_mgt, "P9-occ-quirk")) {
+   flag = 1;
+   }
+
if (of_property_read_u32(power_mgt, "ibm,pstate-ultra-turbo",
 _ultra_turbo)) {
powernv_pstate_info.wof_enabled = false;
@@ -249,6 +298,7 @@ static int init_powernv_pstates(void)
 next:
pr_info("cpufreq pstate min %d nominal %d max %d\n", pstate_min,
pstate_nominal, pstate_max);
+   pr_info("frequency domain indicator %d", freq_domain_indicator);
pr_info("Workload Optimized Frequency is %s in the platform\n",
(powernv_pstate_info.wof_enabled) ? "enabled" : "disabled");
 
@@ -276,6 +326,15 @@ static int init_powernv_pstates(void)
return -ENODEV;
}
 
+   if (freq_domain_indicator) {
+   hash_init(freq_domain_map);
+   for_each_possible_cpu(i) {
+   key = ((u32) get_hard_smp_processor_id(i) &
+   freq_domain_indicator);
+   insert(key, i);
+   }
+   }
+
powernv_pstate_info.nr_pstates = nr_pstates;
pr_debug("NR PStates %d\n", nr_pstates);
for (i = 0; i < nr_pstates; i++) {
@@ -693,6 +752,7 @@ static int powernv_cpufreq_target_index(struct 
cpufreq_policy *policy,
 {
struct powernv_smp_call_data freq_data;