drivers/powercap/dtpm_cpu.c

0001 // SPDX-License-Identifier: GPL-2.0-only
0002 /*
0003  * Copyright 2020 Linaro Limited
0004  *
0005  * Author: Daniel Lezcano <daniel.lezcano@linaro.org>
0006  *
0007  * The DTPM CPU is based on the energy model. It hooks the CPU in the
0008  * DTPM tree which in turns update the power number by propagating the
0009  * power number from the CPU energy model information to the parents.
0010  *
0011  * The association between the power and the performance state, allows
0012  * to set the power of the CPU at the OPP granularity.
0013  *
0014  * The CPU hotplug is supported and the power numbers will be updated
0015  * if a CPU is hot plugged / unplugged.
0016  */
0017 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
0018
0019 #include <linux/cpumask.h>
0020 #include <linux/cpufreq.h>
0021 #include <linux/cpuhotplug.h>
0022 #include <linux/dtpm.h>
0023 #include <linux/energy_model.h>
0024 #include <linux/of.h>
0025 #include <linux/pm_qos.h>
0026 #include <linux/slab.h>
0027 #include <linux/units.h>
0028
0029 struct dtpm_cpu {
0030     struct dtpm dtpm;
0031     struct freq_qos_request qos_req;
0032     int cpu;
0033 };
0034
0035 static DEFINE_PER_CPU(struct dtpm_cpu *, dtpm_per_cpu);
0036
0037 static struct dtpm_cpu *to_dtpm_cpu(struct dtpm *dtpm)
0038 {
0039     return container_of(dtpm, struct dtpm_cpu, dtpm);
0040 }
0041
0042 static u64 set_pd_power_limit(struct dtpm *dtpm, u64 power_limit)
0043 {
0044     struct dtpm_cpu *dtpm_cpu = to_dtpm_cpu(dtpm);
0045     struct em_perf_domain *pd = em_cpu_get(dtpm_cpu->cpu);
0046     struct cpumask cpus;
0047     unsigned long freq;
0048     u64 power;
0049     int i, nr_cpus;
0050
0051     cpumask_and(&cpus, cpu_online_mask, to_cpumask(pd->cpus));
0052     nr_cpus = cpumask_weight(&cpus);
0053
0054     for (i = 0; i < pd->nr_perf_states; i++) {
0055
0056         power = pd->table[i].power * nr_cpus;
0057
0058         if (power > power_limit)
0059             break;
0060     }
0061
0062     freq = pd->table[i - 1].frequency;
0063
0064     freq_qos_update_request(&dtpm_cpu->qos_req, freq);
0065
0066     power_limit = pd->table[i - 1].power * nr_cpus;
0067
0068     return power_limit;
0069 }
0070
0071 static u64 scale_pd_power_uw(struct cpumask *pd_mask, u64 power)
0072 {
0073     unsigned long max, sum_util = 0;
0074     int cpu;
0075
0076     /*
0077      * The capacity is the same for all CPUs belonging to
0078      * the same perf domain.
0079      */
0080     max = arch_scale_cpu_capacity(cpumask_first(pd_mask));
0081
0082     for_each_cpu_and(cpu, pd_mask, cpu_online_mask)
0083         sum_util += sched_cpu_util(cpu);
0084
0085     return (power * ((sum_util << 10) / max)) >> 10;
0086 }
0087
0088 static u64 get_pd_power_uw(struct dtpm *dtpm)
0089 {
0090     struct dtpm_cpu *dtpm_cpu = to_dtpm_cpu(dtpm);
0091     struct em_perf_domain *pd;
0092     struct cpumask *pd_mask;
0093     unsigned long freq;
0094     int i;
0095
0096     pd = em_cpu_get(dtpm_cpu->cpu);
0097
0098     pd_mask = em_span_cpus(pd);
0099
0100     freq = cpufreq_quick_get(dtpm_cpu->cpu);
0101
0102     for (i = 0; i < pd->nr_perf_states; i++) {
0103
0104         if (pd->table[i].frequency < freq)
0105             continue;
0106
0107         return scale_pd_power_uw(pd_mask, pd->table[i].power *
0108                      MICROWATT_PER_MILLIWATT);
0109     }
0110
0111     return 0;
0112 }
0113
0114 static int update_pd_power_uw(struct dtpm *dtpm)
0115 {
0116     struct dtpm_cpu *dtpm_cpu = to_dtpm_cpu(dtpm);
0117     struct em_perf_domain *em = em_cpu_get(dtpm_cpu->cpu);
0118     struct cpumask cpus;
0119     int nr_cpus;
0120
0121     cpumask_and(&cpus, cpu_online_mask, to_cpumask(em->cpus));
0122     nr_cpus = cpumask_weight(&cpus);
0123
0124     dtpm->power_min = em->table[0].power;
0125     dtpm->power_min *= MICROWATT_PER_MILLIWATT;
0126     dtpm->power_min *= nr_cpus;
0127
0128     dtpm->power_max = em->table[em->nr_perf_states - 1].power;
0129     dtpm->power_max *= MICROWATT_PER_MILLIWATT;
0130     dtpm->power_max *= nr_cpus;
0131
0132     return 0;
0133 }
0134
0135 static void pd_release(struct dtpm *dtpm)
0136 {
0137     struct dtpm_cpu *dtpm_cpu = to_dtpm_cpu(dtpm);
0138     struct cpufreq_policy *policy;
0139
0140     if (freq_qos_request_active(&dtpm_cpu->qos_req))
0141         freq_qos_remove_request(&dtpm_cpu->qos_req);
0142
0143     policy = cpufreq_cpu_get(dtpm_cpu->cpu);
0144     if (policy) {
0145         for_each_cpu(dtpm_cpu->cpu, policy->related_cpus)
0146             per_cpu(dtpm_per_cpu, dtpm_cpu->cpu) = NULL;
0147     }
0148
0149     kfree(dtpm_cpu);
0150 }
0151
0152 static struct dtpm_ops dtpm_ops = {
0153     .set_power_uw    = set_pd_power_limit,
0154     .get_power_uw    = get_pd_power_uw,
0155     .update_power_uw = update_pd_power_uw,
0156     .release     = pd_release,
0157 };
0158
0159 static int cpuhp_dtpm_cpu_offline(unsigned int cpu)
0160 {
0161     struct dtpm_cpu *dtpm_cpu;
0162
0163     dtpm_cpu = per_cpu(dtpm_per_cpu, cpu);
0164     if (dtpm_cpu)
0165         dtpm_update_power(&dtpm_cpu->dtpm);
0166
0167     return 0;
0168 }
0169
0170 static int cpuhp_dtpm_cpu_online(unsigned int cpu)
0171 {
0172     struct dtpm_cpu *dtpm_cpu;
0173
0174     dtpm_cpu = per_cpu(dtpm_per_cpu, cpu);
0175     if (dtpm_cpu)
0176         return dtpm_update_power(&dtpm_cpu->dtpm);
0177
0178     return 0;
0179 }
0180
0181 static int __dtpm_cpu_setup(int cpu, struct dtpm *parent)
0182 {
0183     struct dtpm_cpu *dtpm_cpu;
0184     struct cpufreq_policy *policy;
0185     struct em_perf_domain *pd;
0186     char name[CPUFREQ_NAME_LEN];
0187     int ret = -ENOMEM;
0188
0189     dtpm_cpu = per_cpu(dtpm_per_cpu, cpu);
0190     if (dtpm_cpu)
0191         return 0;
0192
0193     policy = cpufreq_cpu_get(cpu);
0194     if (!policy)
0195         return 0;
0196
0197     pd = em_cpu_get(cpu);
0198     if (!pd || em_is_artificial(pd))
0199         return -EINVAL;
0200
0201     dtpm_cpu = kzalloc(sizeof(*dtpm_cpu), GFP_KERNEL);
0202     if (!dtpm_cpu)
0203         return -ENOMEM;
0204
0205     dtpm_init(&dtpm_cpu->dtpm, &dtpm_ops);
0206     dtpm_cpu->cpu = cpu;
0207
0208     for_each_cpu(cpu, policy->related_cpus)
0209         per_cpu(dtpm_per_cpu, cpu) = dtpm_cpu;
0210
0211     snprintf(name, sizeof(name), "cpu%d-cpufreq", dtpm_cpu->cpu);
0212
0213     ret = dtpm_register(name, &dtpm_cpu->dtpm, parent);
0214     if (ret)
0215         goto out_kfree_dtpm_cpu;
0216
0217     ret = freq_qos_add_request(&policy->constraints,
0218                    &dtpm_cpu->qos_req, FREQ_QOS_MAX,
0219                    pd->table[pd->nr_perf_states - 1].frequency);
0220     if (ret)
0221         goto out_dtpm_unregister;
0222
0223     return 0;
0224
0225 out_dtpm_unregister:
0226     dtpm_unregister(&dtpm_cpu->dtpm);
0227     dtpm_cpu = NULL;
0228
0229 out_kfree_dtpm_cpu:
0230     for_each_cpu(cpu, policy->related_cpus)
0231         per_cpu(dtpm_per_cpu, cpu) = NULL;
0232     kfree(dtpm_cpu);
0233
0234     return ret;
0235 }
0236
0237 static int dtpm_cpu_setup(struct dtpm *dtpm, struct device_node *np)
0238 {
0239     int cpu;
0240
0241     cpu = of_cpu_node_to_id(np);
0242     if (cpu < 0)
0243         return 0;
0244
0245     return __dtpm_cpu_setup(cpu, dtpm);
0246 }
0247
0248 static int dtpm_cpu_init(void)
0249 {
0250     int ret;
0251
0252     /*
0253      * The callbacks at CPU hotplug time are calling
0254      * dtpm_update_power() which in turns calls update_pd_power().
0255      *
0256      * The function update_pd_power() uses the online mask to
0257      * figure out the power consumption limits.
0258      *
0259      * At CPUHP_AP_ONLINE_DYN, the CPU is present in the CPU
0260      * online mask when the cpuhp_dtpm_cpu_online function is
0261      * called, but the CPU is still in the online mask for the
0262      * tear down callback. So the power can not be updated when
0263      * the CPU is unplugged.
0264      *
0265      * At CPUHP_AP_DTPM_CPU_DEAD, the situation is the opposite as
0266      * above. The CPU online mask is not up to date when the CPU
0267      * is plugged in.
0268      *
0269      * For this reason, we need to call the online and offline
0270      * callbacks at different moments when the CPU online mask is
0271      * consistent with the power numbers we want to update.
0272      */
0273     ret = cpuhp_setup_state(CPUHP_AP_DTPM_CPU_DEAD, "dtpm_cpu:offline",
0274                 NULL, cpuhp_dtpm_cpu_offline);
0275     if (ret < 0)
0276         return ret;
0277
0278     ret = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "dtpm_cpu:online",
0279                 cpuhp_dtpm_cpu_online, NULL);
0280     if (ret < 0)
0281         return ret;
0282
0283     return 0;
0284 }
0285
0286 static void dtpm_cpu_exit(void)
0287 {
0288     cpuhp_remove_state_nocalls(CPUHP_AP_ONLINE_DYN);
0289     cpuhp_remove_state_nocalls(CPUHP_AP_DTPM_CPU_DEAD);
0290 }
0291
0292 struct dtpm_subsys_ops dtpm_cpu_ops = {
0293     .name = KBUILD_MODNAME,
0294     .init = dtpm_cpu_init,
0295     .exit = dtpm_cpu_exit,
0296     .setup = dtpm_cpu_setup,
0297 };