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 <[email protected]>
---
drivers/cpufreq/powernv-cpufreq.c | 95 +++++++++++++++++++++++++++++++++++----
1 file changed, 87 insertions(+), 8 deletions(-)
diff --git a/drivers/cpufreq/powernv-cpufreq.c b/drivers/cpufreq/powernv-cpufreq.c
index b6d7c4c..9384110 100644
--- a/drivers/cpufreq/powernv-cpufreq.c
+++ b/drivers/cpufreq/powernv-cpufreq.c
@@ -37,6 +37,7 @@
#include <asm/smp.h> /* Required for cpu_sibling_mask() in UP configs */
#include <asm/opal.h>
#include <linux/timer.h>
+#include <linux/hashtable.h>
#define POWERNV_MAX_PSTATES 256
#define PMSR_PSAFE_ENABLE (1UL << 30)
@@ -130,6 +131,8 @@ static struct chip {
static int nr_chips;
static DEFINE_PER_CPU(struct chip *, chip_info);
+static u32 freq_domain_indicator;
+
/*
* Note:
* The set of pstates consists of contiguous integers.
@@ -194,6 +197,38 @@ 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, &data->mask);
+ return;
+ }
+ }
+
+ data = kzalloc(sizeof(*data), GFP_KERNEL);
+ hash_add(freq_domain_map, &data->hash_node, key%SIZE);
+ cpumask_set_cpu(cpu, &data->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
@@ -206,6 +241,7 @@ 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;
power_mgt = of_find_node_by_path("/ibm,opal/power-mgt");
if (!power_mgt) {
@@ -229,6 +265,13 @@ 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",
+ &freq_domain_indicator)) {
+ pr_warn("ibm,freq-domain-indicator not found\n");
+ freq_domain_indicator = 0;
+ }
+
if (of_property_read_u32(power_mgt, "ibm,pstate-ultra-turbo",
&pstate_ultra_turbo)) {
powernv_pstate_info.wof_enabled = false;
@@ -249,6 +292,7 @@ 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 +320,15 @@ 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 +746,8 @@ static int powernv_cpufreq_target_index(struct cpufreq_policy *policy,
{
struct powernv_smp_call_data freq_data;
unsigned int cur_msec, gpstate_idx;
+ cpumask_t temp;
+ u32 cpu;
struct global_pstate_info *gpstates = policy->driver_data;
if (unlikely(rebooting) && new_index != get_nominal_index())
@@ -761,24 +816,48 @@ static int powernv_cpufreq_target_index(struct cpufreq_policy *policy,
spin_unlock(&gpstates->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 policy->cpus (core)
+ * Use smp_call_function to send IPI and execute the mtspr on CPU.
+ * This needs to be done on every core of the policy
*/
- smp_call_function_any(policy->cpus, set_pstate, &freq_data, 1);
+ cpumask_copy(&temp, policy->cpus);
+
+ while (!cpumask_empty(&temp)) {
+ cpu = cpumask_first(&temp);
+ smp_call_function_any(cpu_sibling_mask(cpu),
+ set_pstate, &freq_data, 1);
+ cpumask_andnot(&temp, &temp, cpu_sibling_mask(cpu));
+ }
+
return 0;
}
static int powernv_cpufreq_cpu_init(struct cpufreq_policy *policy)
{
- int base, i, ret;
+ int ret;
struct kernfs_node *kn;
struct global_pstate_info *gpstates;
- base = cpu_first_thread_sibling(policy->cpu);
+ if (!freq_domain_indicator) {
+ int base, i;
- for (i = 0; i < threads_per_core; i++)
- cpumask_set_cpu(base + i, policy->cpus);
+ base = cpu_first_thread_sibling(policy->cpu);
+ for (i = 0; i < threads_per_core; i++)
+ cpumask_set_cpu(base + i, policy->cpus);
+ } else {
+ u32 key;
+ struct hashmap *data;
+
+ key = ((u32) get_hard_smp_processor_id(policy->cpu) &
+ freq_domain_indicator);
+ hash_for_each_possible(freq_domain_map, data, hash_node,
+ key%SIZE) {
+ if (data->chip_id == cpu_to_chip_id(policy->cpu) &&
+ data->pir_key == key) {
+ cpumask_copy(policy->cpus, &data->mask);
+ break;
+ }
+ }
+ }
kn = kernfs_find_and_get(policy->kobj.sd, throttle_attr_grp.name);
if (!kn) {
--
2.9.3
+ Gautham,
@Gautham: Can you please help reviewing this one ?
On 13-12-17, 13:49, Abhishek Goel wrote:
> @@ -693,6 +746,8 @@ static int powernv_cpufreq_target_index(struct cpufreq_policy *policy,
> {
> struct powernv_smp_call_data freq_data;
> unsigned int cur_msec, gpstate_idx;
> + cpumask_t temp;
> + u32 cpu;
> struct global_pstate_info *gpstates = policy->driver_data;
>
> if (unlikely(rebooting) && new_index != get_nominal_index())
> @@ -761,24 +816,48 @@ static int powernv_cpufreq_target_index(struct cpufreq_policy *policy,
> spin_unlock(&gpstates->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 policy->cpus (core)
> + * Use smp_call_function to send IPI and execute the mtspr on CPU.
> + * This needs to be done on every core of the policy
Why on each CPU ?
> */
> - smp_call_function_any(policy->cpus, set_pstate, &freq_data, 1);
> + cpumask_copy(&temp, policy->cpus);
> +
> + while (!cpumask_empty(&temp)) {
> + cpu = cpumask_first(&temp);
> + smp_call_function_any(cpu_sibling_mask(cpu),
> + set_pstate, &freq_data, 1);
> + cpumask_andnot(&temp, &temp, cpu_sibling_mask(cpu));
> + }
> +
> return 0;
> }
--
viresh
On 12/14/2017 10:12 AM, Viresh Kumar wrote:
> + Gautham,
>
> @Gautham: Can you please help reviewing this one ?
>
> On 13-12-17, 13:49, Abhishek Goel wrote:
>> @@ -693,6 +746,8 @@ static int powernv_cpufreq_target_index(struct cpufreq_policy *policy,
>> {
>> struct powernv_smp_call_data freq_data;
>> unsigned int cur_msec, gpstate_idx;
>> + cpumask_t temp;
>> + u32 cpu;
>> struct global_pstate_info *gpstates = policy->driver_data;
>>
>> if (unlikely(rebooting) && new_index != get_nominal_index())
>> @@ -761,24 +816,48 @@ static int powernv_cpufreq_target_index(struct cpufreq_policy *policy,
>> spin_unlock(&gpstates->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 policy->cpus (core)
>> + * Use smp_call_function to send IPI and execute the mtspr on CPU.
>> + * This needs to be done on every core of the policy
> Why on each CPU ?
We need to do it in this way as the current implementation takes the max
of the PMSR of the cores. Thus, when the frequency is required to be
ramped up, it suffices to write to just the local PMSR, but when the
frequency is to be ramped down, if we don't send the IPI it breaks the
compatibility with P8.
>
>> */
>> - smp_call_function_any(policy->cpus, set_pstate, &freq_data, 1);
>> + cpumask_copy(&temp, policy->cpus);
>> +
>> + while (!cpumask_empty(&temp)) {
>> + cpu = cpumask_first(&temp);
>> + smp_call_function_any(cpu_sibling_mask(cpu),
>> + set_pstate, &freq_data, 1);
>> + cpumask_andnot(&temp, &temp, cpu_sibling_mask(cpu));
>> + }
>> +
>> return 0;
>> }
On 18-12-17, 10:41, Abhishek wrote:
> We need to do it in this way as the current implementation takes the max of
> the PMSR of the cores. Thus, when the frequency is required to be ramped up,
> it suffices to write to just the local PMSR, but when the frequency is to be
> ramped down, if we don't send the IPI it breaks the compatibility with P8.
Looks strange really that you have to program this differently for speeding up
or down. These CPUs are part of one cpufreq policy and so I would normally
expect changes to any CPU should reflect for other CPUs as well.
@Goutham: Do you know why it is so ?
--
viresh
Hi Viresh,
On Mon, Dec 18, 2017 at 01:59:35PM +0530, Viresh Kumar wrote:
> On 18-12-17, 10:41, Abhishek wrote:
> > We need to do it in this way as the current implementation takes the max of
> > the PMSR of the cores. Thus, when the frequency is required to be ramped up,
> > it suffices to write to just the local PMSR, but when the frequency is to be
> > ramped down, if we don't send the IPI it breaks the compatibility with P8.
>
> Looks strange really that you have to program this differently for speeding up
> or down. These CPUs are part of one cpufreq policy and so I would normally
> expect changes to any CPU should reflect for other CPUs as well.
>
> @Goutham: Do you know why it is so ?
>
These are due to some implementation quirks where the platform has
provided a PMCR per-core to be backward compatible with POWER8, but
controls the frequency at a quad-level, by taking the maximum of the
four PMCR values instead of the latest one. So, changes to any CPU in
the core will reflect on all the cores if the frequency is higher than
the current frequency, but not necessarily if the requested frequency
is lower than the current frequency.
Without sending the extra IPIs, we will be breaking the ABI since if
we set userspace governor, and change the frequency of a core by
lowering it, then it will not reflect on the CPUs of the cores in the
quad.
Abhishek,
I think we can rework this by sending the extra IPIs only in the
presence of the quirk which can be indicated through a device-tree
parameter. If the future implementation fix this, then we won't need
the extra IPIs.
> --
> viresh
>
On Tue, Dec 19, 2017 at 8:20 PM, Gautham R Shenoy
<[email protected]> wrote:
> Hi Viresh,
> On Mon, Dec 18, 2017 at 01:59:35PM +0530, Viresh Kumar wrote:
>> On 18-12-17, 10:41, Abhishek wrote:
>> > We need to do it in this way as the current implementation takes the max of
>> > the PMSR of the cores. Thus, when the frequency is required to be ramped up,
>> > it suffices to write to just the local PMSR, but when the frequency is to be
>> > ramped down, if we don't send the IPI it breaks the compatibility with P8.
>>
>> Looks strange really that you have to program this differently for speeding up
>> or down. These CPUs are part of one cpufreq policy and so I would normally
>> expect changes to any CPU should reflect for other CPUs as well.
>>
>> @Goutham: Do you know why it is so ?
>>
>
> These are due to some implementation quirks where the platform has
> provided a PMCR per-core to be backward compatible with POWER8, but
> controls the frequency at a quad-level, by taking the maximum of the
> four PMCR values instead of the latest one. So, changes to any CPU in
> the core will reflect on all the cores if the frequency is higher than
> the current frequency, but not necessarily if the requested frequency
> is lower than the current frequency.
>
> Without sending the extra IPIs, we will be breaking the ABI since if
> we set userspace governor, and change the frequency of a core by
> lowering it, then it will not reflect on the CPUs of the cores in the
> quad.
What about cpufreq_policy->cpus/related_cpus? Am I missing something?
>
> Abhishek,
> I think we can rework this by sending the extra IPIs only in the
> presence of the quirk which can be indicated through a device-tree
> parameter. If the future implementation fix this, then we won't need
> the extra IPIs.
Balbir Singh.
On Tue, Dec 19, 2017 at 09:21:52PM +1100, Balbir Singh wrote:
> On Tue, Dec 19, 2017 at 8:20 PM, Gautham R Shenoy
> <[email protected]> wrote:
> > Hi Viresh,
> > On Mon, Dec 18, 2017 at 01:59:35PM +0530, Viresh Kumar wrote:
> >> On 18-12-17, 10:41, Abhishek wrote:
> >> > We need to do it in this way as the current implementation takes the max of
> >> > the PMSR of the cores. Thus, when the frequency is required to be ramped up,
> >> > it suffices to write to just the local PMSR, but when the frequency is to be
> >> > ramped down, if we don't send the IPI it breaks the compatibility with P8.
> >>
> >> Looks strange really that you have to program this differently for speeding up
> >> or down. These CPUs are part of one cpufreq policy and so I would normally
> >> expect changes to any CPU should reflect for other CPUs as well.
> >>
> >> @Goutham: Do you know why it is so ?
> >>
> >
> > These are due to some implementation quirks where the platform has
> > provided a PMCR per-core to be backward compatible with POWER8, but
> > controls the frequency at a quad-level, by taking the maximum of the
> > four PMCR values instead of the latest one. So, changes to any CPU in
> > the core will reflect on all the cores if the frequency is higher than
> > the current frequency, but not necessarily if the requested frequency
> > is lower than the current frequency.
> >
> > Without sending the extra IPIs, we will be breaking the ABI since if
> > we set userspace governor, and change the frequency of a core by
> > lowering it, then it will not reflect on the CPUs of the cores in the
> > quad.
>
>
> What about cpufreq_policy->cpus/related_cpus? Am I missing something?
The frequency indicator passed via the device tree is used to derive
the mask corresponding to the set of CPUs that share the same
frequency. It is this mask that is set to
cpufreq_policy->cpus/related_cpus.
>
> >
> > Abhishek,
> > I think we can rework this by sending the extra IPIs only in the
> > presence of the quirk which can be indicated through a device-tree
> > parameter. If the future implementation fix this, then we won't need
> > the extra IPIs.
>
> Balbir Singh.
>