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 * MICROWATT_PER_MILLIWATT * 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 *
  67                 MICROWATT_PER_MILLIWATT * nr_cpus;
  68
  69         return power_limit;
  70 }
  71
  72 static u64 scale_pd_power_uw(struct cpumask *pd_mask, u64 power)
  73 {
  74         unsigned long max = 0, sum_util = 0;
  75         int cpu;
  76
  77         for_each_cpu_and(cpu, pd_mask, cpu_online_mask) {
  78
  79                 /*
  80                  * The capacity is the same for all CPUs belonging to
  81                  * the same perf domain, so a single call to
  82                  * arch_scale_cpu_capacity() is enough. However, we
  83                  * need the CPU parameter to be initialized by the
  84                  * loop, so the call ends up in this block.
  85                  *
  86                  * We can initialize 'max' with a cpumask_first() call
  87                  * before the loop but the bits computation is not
  88                  * worth given the arch_scale_cpu_capacity() just
  89                  * returns a value where the resulting assembly code
  90                  * will be optimized by the compiler.
  91                  */
  92                 max = arch_scale_cpu_capacity(cpu);
  93                 sum_util += sched_cpu_util(cpu, max);
  94         }
  95
  96         /*
  97          * In the improbable case where all the CPUs of the perf
  98          * domain are offline, 'max' will be zero and will lead to an
  99          * illegal operation with a zero division.
 100          */
 101         return max ? (power * ((sum_util << 10) / max)) >> 10 : 0;
 102 }
 103
 104 static u64 get_pd_power_uw(struct dtpm *dtpm)
 105 {
 106         struct dtpm_cpu *dtpm_cpu = to_dtpm_cpu(dtpm);
 107         struct em_perf_domain *pd;
 108         struct cpumask *pd_mask;
 109         unsigned long freq;
 110         int i;
 111
 112         pd = em_cpu_get(dtpm_cpu->cpu);
 113
 114         pd_mask = em_span_cpus(pd);
 115
 116         freq = cpufreq_quick_get(dtpm_cpu->cpu);
 117
 118         for (i = 0; i < pd->nr_perf_states; i++) {
 119
 120                 if (pd->table[i].frequency < freq)
 121                         continue;
 122
 123                 return scale_pd_power_uw(pd_mask, pd->table[i].power *
 124                                          MICROWATT_PER_MILLIWATT);
 125         }
 126
 127         return 0;
 128 }
 129
 130 static int update_pd_power_uw(struct dtpm *dtpm)
 131 {
 132         struct dtpm_cpu *dtpm_cpu = to_dtpm_cpu(dtpm);
 133         struct em_perf_domain *em = em_cpu_get(dtpm_cpu->cpu);
 134         struct cpumask cpus;
 135         int nr_cpus;
 136
 137         cpumask_and(&cpus, cpu_online_mask, to_cpumask(em->cpus));
 138         nr_cpus = cpumask_weight(&cpus);
 139
 140         dtpm->power_min = em->table[0].power;
 141         dtpm->power_min *= MICROWATT_PER_MILLIWATT;
 142         dtpm->power_min *= nr_cpus;
 143
 144         dtpm->power_max = em->table[em->nr_perf_states - 1].power;
 145         dtpm->power_max *= MICROWATT_PER_MILLIWATT;
 146         dtpm->power_max *= nr_cpus;
 147
 148         return 0;
 149 }
 150
 151 static void pd_release(struct dtpm *dtpm)
 152 {
 153         struct dtpm_cpu *dtpm_cpu = to_dtpm_cpu(dtpm);
 154         struct cpufreq_policy *policy;
 155
 156         if (freq_qos_request_active(&dtpm_cpu->qos_req))
 157                 freq_qos_remove_request(&dtpm_cpu->qos_req);
 158
 159         policy = cpufreq_cpu_get(dtpm_cpu->cpu);
 160         if (policy) {
 161                 for_each_cpu(dtpm_cpu->cpu, policy->related_cpus)
 162                         per_cpu(dtpm_per_cpu, dtpm_cpu->cpu) = NULL;
 163         }
 164
 165         kfree(dtpm_cpu);
 166 }
 167
 168 static struct dtpm_ops dtpm_ops = {
 169         .set_power_uw    = set_pd_power_limit,
 170         .get_power_uw    = get_pd_power_uw,
 171         .update_power_uw = update_pd_power_uw,
 172         .release         = pd_release,
 173 };
 174
 175 static int cpuhp_dtpm_cpu_offline(unsigned int cpu)
 176 {
 177         struct dtpm_cpu *dtpm_cpu;
 178
 179         dtpm_cpu = per_cpu(dtpm_per_cpu, cpu);
 180         if (dtpm_cpu)
 181                 dtpm_update_power(&dtpm_cpu->dtpm);
 182
 183         return 0;
 184 }
 185
 186 static int cpuhp_dtpm_cpu_online(unsigned int cpu)
 187 {
 188         struct dtpm_cpu *dtpm_cpu;
 189
 190         dtpm_cpu = per_cpu(dtpm_per_cpu, cpu);
 191         if (dtpm_cpu)
 192                 return dtpm_update_power(&dtpm_cpu->dtpm);
 193
 194         return 0;
 195 }
 196
 197 static int __dtpm_cpu_setup(int cpu, struct dtpm *parent)
 198 {
 199         struct dtpm_cpu *dtpm_cpu;
 200         struct cpufreq_policy *policy;
 201         struct em_perf_domain *pd;
 202         char name[CPUFREQ_NAME_LEN];
 203         int ret = -ENOMEM;
 204
 205         dtpm_cpu = per_cpu(dtpm_per_cpu, cpu);
 206         if (dtpm_cpu)
 207                 return 0;
 208
 209         policy = cpufreq_cpu_get(cpu);
 210         if (!policy)
 211                 return 0;
 212
 213         pd = em_cpu_get(cpu);
 214         if (!pd || em_is_artificial(pd))
 215                 return -EINVAL;
 216
 217         dtpm_cpu = kzalloc(sizeof(*dtpm_cpu), GFP_KERNEL);
 218         if (!dtpm_cpu)
 219                 return -ENOMEM;
 220
 221         dtpm_init(&dtpm_cpu->dtpm, &dtpm_ops);
 222         dtpm_cpu->cpu = cpu;
 223
 224         for_each_cpu(cpu, policy->related_cpus)
 225                 per_cpu(dtpm_per_cpu, cpu) = dtpm_cpu;
 226
 227         snprintf(name, sizeof(name), "cpu%d-cpufreq", dtpm_cpu->cpu);
 228
 229         ret = dtpm_register(name, &dtpm_cpu->dtpm, parent);
 230         if (ret)
 231                 goto out_kfree_dtpm_cpu;
 232
 233         ret = freq_qos_add_request(&policy->constraints,
 234                                    &dtpm_cpu->qos_req, FREQ_QOS_MAX,
 235                                    pd->table[pd->nr_perf_states - 1].frequency);
 236         if (ret)
 237                 goto out_dtpm_unregister;
 238
 239         return 0;
 240
 241 out_dtpm_unregister:
 242         dtpm_unregister(&dtpm_cpu->dtpm);
 243         dtpm_cpu = NULL;
 244
 245 out_kfree_dtpm_cpu:
 246         for_each_cpu(cpu, policy->related_cpus)
 247                 per_cpu(dtpm_per_cpu, cpu) = NULL;
 248         kfree(dtpm_cpu);
 249
 250         return ret;
 251 }
 252
 253 static int dtpm_cpu_setup(struct dtpm *dtpm, struct device_node *np)
 254 {
 255         int cpu;
 256
 257         cpu = of_cpu_node_to_id(np);
 258         if (cpu < 0)
 259                 return 0;
 260
 261         return __dtpm_cpu_setup(cpu, dtpm);
 262 }
 263
 264 static int dtpm_cpu_init(void)
 265 {
 266         int ret;
 267
 268         /*
 269          * The callbacks at CPU hotplug time are calling
 270          * dtpm_update_power() which in turns calls update_pd_power().
 271          *
 272          * The function update_pd_power() uses the online mask to
 273          * figure out the power consumption limits.
 274          *
 275          * At CPUHP_AP_ONLINE_DYN, the CPU is present in the CPU
 276          * online mask when the cpuhp_dtpm_cpu_online function is
 277          * called, but the CPU is still in the online mask for the
 278          * tear down callback. So the power can not be updated when
 279          * the CPU is unplugged.
 280          *
 281          * At CPUHP_AP_DTPM_CPU_DEAD, the situation is the opposite as
 282          * above. The CPU online mask is not up to date when the CPU
 283          * is plugged in.
 284          *
 285          * For this reason, we need to call the online and offline
 286          * callbacks at different moments when the CPU online mask is
 287          * consistent with the power numbers we want to update.
 288          */
 289         ret = cpuhp_setup_state(CPUHP_AP_DTPM_CPU_DEAD, "dtpm_cpu:offline",
 290                                 NULL, cpuhp_dtpm_cpu_offline);
 291         if (ret < 0)
 292                 return ret;
 293
 294         ret = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "dtpm_cpu:online",
 295                                 cpuhp_dtpm_cpu_online, NULL);
 296         if (ret < 0)
 297                 return ret;
 298
 299         return 0;
 300 }
 301
 302 static void dtpm_cpu_exit(void)
 303 {
 304         cpuhp_remove_state_nocalls(CPUHP_AP_ONLINE_DYN);
 305         cpuhp_remove_state_nocalls(CPUHP_AP_DTPM_CPU_DEAD);
 306 }
 307
 308 struct dtpm_subsys_ops dtpm_cpu_ops = {
 309         .name = KBUILD_MODNAME,
 310         .init = dtpm_cpu_init,
 311         .exit = dtpm_cpu_exit,
 312         .setup = dtpm_cpu_setup,
 313 };