drivers/powercap/dtpm_cpu.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * Copyright 2020 Linaro Limited
   4  *
   5  * Author: Daniel Lezcano <daniel.lezcano@linaro.org>
   6  *
   7  * The DTPM CPU is based on the energy model. It hooks the CPU in the
   8  * DTPM tree which in turns update the power number by propagating the
   9  * power number from the CPU energy model information to the parents.
  10  *
  11  * The association between the power and the performance state, allows
  12  * to set the power of the CPU at the OPP granularity.
  13  *
  14  * The CPU hotplug is supported and the power numbers will be updated
  15  * if a CPU is hot plugged / unplugged.
  16  */
  17 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  18
  19 #include <linux/cpumask.h>
  20 #include <linux/cpufreq.h>
  21 #include <linux/cpuhotplug.h>
  22 #include <linux/dtpm.h>
  23 #include <linux/energy_model.h>
  24 #include <linux/of.h>
  25 #include <linux/pm_qos.h>
  26 #include <linux/slab.h>
  27 #include <linux/units.h>
  28
  29 struct dtpm_cpu {
  30         struct dtpm dtpm;
  31         struct freq_qos_request qos_req;
  32         int cpu;
  33 };
  34
  35 static DEFINE_PER_CPU(struct dtpm_cpu *, dtpm_per_cpu);
  36
  37 static struct dtpm_cpu *to_dtpm_cpu(struct dtpm *dtpm)
  38 {
  39         return container_of(dtpm, struct dtpm_cpu, dtpm);
  40 }
  41
  42 static u64 set_pd_power_limit(struct dtpm *dtpm, u64 power_limit)
  43 {
  44         struct dtpm_cpu *dtpm_cpu = to_dtpm_cpu(dtpm);
  45         struct em_perf_domain *pd = em_cpu_get(dtpm_cpu->cpu);
  46         struct cpumask cpus;
  47         unsigned long freq;
  48         u64 power;
  49         int i, nr_cpus;
  50
  51         cpumask_and(&cpus, cpu_online_mask, to_cpumask(pd->cpus));
  52         nr_cpus = cpumask_weight(&cpus);
  53
  54         for (i = 0; i < pd->nr_perf_states; i++) {
  55
  56                 power = pd->table[i].power * nr_cpus;
  57
  58                 if (power > power_limit)
  59                         break;
  60         }
  61
  62         freq = pd->table[i - 1].frequency;
  63
  64         freq_qos_update_request(&dtpm_cpu->qos_req, freq);
  65
  66         power_limit = pd->table[i - 1].power * nr_cpus;
  67
  68         return power_limit;
  69 }
  70
  71 static u64 scale_pd_power_uw(struct cpumask *pd_mask, u64 power)
  72 {
  73         unsigned long max, sum_util = 0;
  74         int cpu;
  75
  76         /*
  77          * The capacity is the same for all CPUs belonging to
  78          * the same perf domain.
  79          */
  80         max = arch_scale_cpu_capacity(cpumask_first(pd_mask));
  81
  82         for_each_cpu_and(cpu, pd_mask, cpu_online_mask)
  83                 sum_util += sched_cpu_util(cpu);
  84
  85         return (power * ((sum_util << 10) / max)) >> 10;
  86 }
  87
  88 static u64 get_pd_power_uw(struct dtpm *dtpm)
  89 {
  90         struct dtpm_cpu *dtpm_cpu = to_dtpm_cpu(dtpm);
  91         struct em_perf_domain *pd;
  92         struct cpumask *pd_mask;
  93         unsigned long freq;
  94         int i;
  95
  96         pd = em_cpu_get(dtpm_cpu->cpu);
  97
  98         pd_mask = em_span_cpus(pd);
  99
 100         freq = cpufreq_quick_get(dtpm_cpu->cpu);
 101
 102         for (i = 0; i < pd->nr_perf_states; i++) {
 103
 104                 if (pd->table[i].frequency < freq)
 105                         continue;
 106
 107                 return scale_pd_power_uw(pd_mask, pd->table[i].power *
 108                                          MICROWATT_PER_MILLIWATT);
 109         }
 110
 111         return 0;
 112 }
 113
 114 static int update_pd_power_uw(struct dtpm *dtpm)
 115 {
 116         struct dtpm_cpu *dtpm_cpu = to_dtpm_cpu(dtpm);
 117         struct em_perf_domain *em = em_cpu_get(dtpm_cpu->cpu);
 118         struct cpumask cpus;
 119         int nr_cpus;
 120
 121         cpumask_and(&cpus, cpu_online_mask, to_cpumask(em->cpus));
 122         nr_cpus = cpumask_weight(&cpus);
 123
 124         dtpm->power_min = em->table[0].power;
 125         dtpm->power_min *= MICROWATT_PER_MILLIWATT;
 126         dtpm->power_min *= nr_cpus;
 127
 128         dtpm->power_max = em->table[em->nr_perf_states - 1].power;
 129         dtpm->power_max *= MICROWATT_PER_MILLIWATT;
 130         dtpm->power_max *= nr_cpus;
 131
 132         return 0;
 133 }
 134
 135 static void pd_release(struct dtpm *dtpm)
 136 {
 137         struct dtpm_cpu *dtpm_cpu = to_dtpm_cpu(dtpm);
 138         struct cpufreq_policy *policy;
 139
 140         if (freq_qos_request_active(&dtpm_cpu->qos_req))
 141                 freq_qos_remove_request(&dtpm_cpu->qos_req);
 142
 143         policy = cpufreq_cpu_get(dtpm_cpu->cpu);
 144         if (policy) {
 145                 for_each_cpu(dtpm_cpu->cpu, policy->related_cpus)
 146                         per_cpu(dtpm_per_cpu, dtpm_cpu->cpu) = NULL;
 147         }
 148
 149         kfree(dtpm_cpu);
 150 }
 151
 152 static struct dtpm_ops dtpm_ops = {
 153         .set_power_uw    = set_pd_power_limit,
 154         .get_power_uw    = get_pd_power_uw,
 155         .update_power_uw = update_pd_power_uw,
 156         .release         = pd_release,
 157 };
 158
 159 static int cpuhp_dtpm_cpu_offline(unsigned int cpu)
 160 {
 161         struct dtpm_cpu *dtpm_cpu;
 162
 163         dtpm_cpu = per_cpu(dtpm_per_cpu, cpu);
 164         if (dtpm_cpu)
 165                 dtpm_update_power(&dtpm_cpu->dtpm);
 166
 167         return 0;
 168 }
 169
 170 static int cpuhp_dtpm_cpu_online(unsigned int cpu)
 171 {
 172         struct dtpm_cpu *dtpm_cpu;
 173
 174         dtpm_cpu = per_cpu(dtpm_per_cpu, cpu);
 175         if (dtpm_cpu)
 176                 return dtpm_update_power(&dtpm_cpu->dtpm);
 177
 178         return 0;
 179 }
 180
 181 static int __dtpm_cpu_setup(int cpu, struct dtpm *parent)
 182 {
 183         struct dtpm_cpu *dtpm_cpu;
 184         struct cpufreq_policy *policy;
 185         struct em_perf_domain *pd;
 186         char name[CPUFREQ_NAME_LEN];
 187         int ret = -ENOMEM;
 188
 189         dtpm_cpu = per_cpu(dtpm_per_cpu, cpu);
 190         if (dtpm_cpu)
 191                 return 0;
 192
 193         policy = cpufreq_cpu_get(cpu);
 194         if (!policy)
 195                 return 0;
 196
 197         pd = em_cpu_get(cpu);
 198         if (!pd || em_is_artificial(pd))
 199                 return -EINVAL;
 200
 201         dtpm_cpu = kzalloc(sizeof(*dtpm_cpu), GFP_KERNEL);
 202         if (!dtpm_cpu)
 203                 return -ENOMEM;
 204
 205         dtpm_init(&dtpm_cpu->dtpm, &dtpm_ops);
 206         dtpm_cpu->cpu = cpu;
 207
 208         for_each_cpu(cpu, policy->related_cpus)
 209                 per_cpu(dtpm_per_cpu, cpu) = dtpm_cpu;
 210
 211         snprintf(name, sizeof(name), "cpu%d-cpufreq", dtpm_cpu->cpu);
 212
 213         ret = dtpm_register(name, &dtpm_cpu->dtpm, parent);
 214         if (ret)
 215                 goto out_kfree_dtpm_cpu;
 216
 217         ret = freq_qos_add_request(&policy->constraints,
 218                                    &dtpm_cpu->qos_req, FREQ_QOS_MAX,
 219                                    pd->table[pd->nr_perf_states - 1].frequency);
 220         if (ret)
 221                 goto out_dtpm_unregister;
 222
 223         return 0;
 224
 225 out_dtpm_unregister:
 226         dtpm_unregister(&dtpm_cpu->dtpm);
 227         dtpm_cpu = NULL;
 228
 229 out_kfree_dtpm_cpu:
 230         for_each_cpu(cpu, policy->related_cpus)
 231                 per_cpu(dtpm_per_cpu, cpu) = NULL;
 232         kfree(dtpm_cpu);
 233
 234         return ret;
 235 }
 236
 237 static int dtpm_cpu_setup(struct dtpm *dtpm, struct device_node *np)
 238 {
 239         int cpu;
 240
 241         cpu = of_cpu_node_to_id(np);
 242         if (cpu < 0)
 243                 return 0;
 244
 245         return __dtpm_cpu_setup(cpu, dtpm);
 246 }
 247
 248 static int dtpm_cpu_init(void)
 249 {
 250         int ret;
 251
 252         /*
 253          * The callbacks at CPU hotplug time are calling
 254          * dtpm_update_power() which in turns calls update_pd_power().
 255          *
 256          * The function update_pd_power() uses the online mask to
 257          * figure out the power consumption limits.
 258          *
 259          * At CPUHP_AP_ONLINE_DYN, the CPU is present in the CPU
 260          * online mask when the cpuhp_dtpm_cpu_online function is
 261          * called, but the CPU is still in the online mask for the
 262          * tear down callback. So the power can not be updated when
 263          * the CPU is unplugged.
 264          *
 265          * At CPUHP_AP_DTPM_CPU_DEAD, the situation is the opposite as
 266          * above. The CPU online mask is not up to date when the CPU
 267          * is plugged in.
 268          *
 269          * For this reason, we need to call the online and offline
 270          * callbacks at different moments when the CPU online mask is
 271          * consistent with the power numbers we want to update.
 272          */
 273         ret = cpuhp_setup_state(CPUHP_AP_DTPM_CPU_DEAD, "dtpm_cpu:offline",
 274                                 NULL, cpuhp_dtpm_cpu_offline);
 275         if (ret < 0)
 276                 return ret;
 277
 278         ret = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "dtpm_cpu:online",
 279                                 cpuhp_dtpm_cpu_online, NULL);
 280         if (ret < 0)
 281                 return ret;
 282
 283         return 0;
 284 }
 285
 286 static void dtpm_cpu_exit(void)
 287 {
 288         cpuhp_remove_state_nocalls(CPUHP_AP_ONLINE_DYN);
 289         cpuhp_remove_state_nocalls(CPUHP_AP_DTPM_CPU_DEAD);
 290 }
 291
 292 struct dtpm_subsys_ops dtpm_cpu_ops = {
 293         .name = KBUILD_MODNAME,
 294         .init = dtpm_cpu_init,
 295         .exit = dtpm_cpu_exit,
 296         .setup = dtpm_cpu_setup,
 297 };