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 28 struct dtpm_cpu { 29 struct dtpm dtpm; 30 struct freq_qos_request qos_req; 31 int cpu; 32 }; 33 34 static DEFINE_PER_CPU(struct dtpm_cpu *, dtpm_per_cpu); 35 36 static struct dtpm_cpu *to_dtpm_cpu(struct dtpm *dtpm) 37 { 38 return container_of(dtpm, struct dtpm_cpu, dtpm); 39 } 40 41 static u64 set_pd_power_limit(struct dtpm *dtpm, u64 power_limit) 42 { 43 struct dtpm_cpu *dtpm_cpu = to_dtpm_cpu(dtpm); 44 struct em_perf_domain *pd = em_cpu_get(dtpm_cpu->cpu); 45 struct em_perf_state *table; 46 unsigned long freq; 47 u64 power; 48 int i, nr_cpus; 49 50 nr_cpus = cpumask_weight_and(cpu_online_mask, to_cpumask(pd->cpus)); 51 52 rcu_read_lock(); 53 table = em_perf_state_from_pd(pd); 54 for (i = 0; i < pd->nr_perf_states; i++) { 55 56 power = table[i].power * nr_cpus; 57 58 if (power > power_limit) 59 break; 60 } 61 62 freq = table[i - 1].frequency; 63 power_limit = table[i - 1].power * nr_cpus; 64 rcu_read_unlock(); 65 66 freq_qos_update_request(&dtpm_cpu->qos_req, freq); 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_state *table; 92 struct em_perf_domain *pd; 93 struct cpumask *pd_mask; 94 unsigned long freq; 95 u64 power = 0; 96 int i; 97 98 pd = em_cpu_get(dtpm_cpu->cpu); 99 100 pd_mask = em_span_cpus(pd); 101 102 freq = cpufreq_quick_get(dtpm_cpu->cpu); 103 104 rcu_read_lock(); 105 table = em_perf_state_from_pd(pd); 106 for (i = 0; i < pd->nr_perf_states; i++) { 107 108 if (table[i].frequency < freq) 109 continue; 110 111 power = scale_pd_power_uw(pd_mask, table[i].power); 112 break; 113 } 114 rcu_read_unlock(); 115 116 return power; 117 } 118 119 static int update_pd_power_uw(struct dtpm *dtpm) 120 { 121 struct dtpm_cpu *dtpm_cpu = to_dtpm_cpu(dtpm); 122 struct em_perf_domain *em = em_cpu_get(dtpm_cpu->cpu); 123 struct em_perf_state *table; 124 int nr_cpus; 125 126 nr_cpus = cpumask_weight_and(cpu_online_mask, to_cpumask(em->cpus)); 127 128 rcu_read_lock(); 129 table = em_perf_state_from_pd(em); 130 131 dtpm->power_min = table[0].power; 132 dtpm->power_min *= nr_cpus; 133 134 dtpm->power_max = table[em->nr_perf_states - 1].power; 135 dtpm->power_max *= nr_cpus; 136 137 rcu_read_unlock(); 138 139 return 0; 140 } 141 142 static void pd_release(struct dtpm *dtpm) 143 { 144 struct dtpm_cpu *dtpm_cpu = to_dtpm_cpu(dtpm); 145 struct cpufreq_policy *policy; 146 147 if (freq_qos_request_active(&dtpm_cpu->qos_req)) 148 freq_qos_remove_request(&dtpm_cpu->qos_req); 149 150 policy = cpufreq_cpu_get(dtpm_cpu->cpu); 151 if (policy) { 152 for_each_cpu(dtpm_cpu->cpu, policy->related_cpus) 153 per_cpu(dtpm_per_cpu, dtpm_cpu->cpu) = NULL; 154 155 cpufreq_cpu_put(policy); 156 } 157 158 kfree(dtpm_cpu); 159 } 160 161 static struct dtpm_ops dtpm_ops = { 162 .set_power_uw = set_pd_power_limit, 163 .get_power_uw = get_pd_power_uw, 164 .update_power_uw = update_pd_power_uw, 165 .release = pd_release, 166 }; 167 168 static int cpuhp_dtpm_cpu_offline(unsigned int cpu) 169 { 170 struct dtpm_cpu *dtpm_cpu; 171 172 dtpm_cpu = per_cpu(dtpm_per_cpu, cpu); 173 if (dtpm_cpu) 174 dtpm_update_power(&dtpm_cpu->dtpm); 175 176 return 0; 177 } 178 179 static int cpuhp_dtpm_cpu_online(unsigned int cpu) 180 { 181 struct dtpm_cpu *dtpm_cpu; 182 183 dtpm_cpu = per_cpu(dtpm_per_cpu, cpu); 184 if (dtpm_cpu) 185 return dtpm_update_power(&dtpm_cpu->dtpm); 186 187 return 0; 188 } 189 190 static int __dtpm_cpu_setup(int cpu, struct dtpm *parent) 191 { 192 struct dtpm_cpu *dtpm_cpu; 193 struct cpufreq_policy *policy; 194 struct em_perf_state *table; 195 struct em_perf_domain *pd; 196 char name[CPUFREQ_NAME_LEN]; 197 int ret = -ENOMEM; 198 199 dtpm_cpu = per_cpu(dtpm_per_cpu, cpu); 200 if (dtpm_cpu) 201 return 0; 202 203 policy = cpufreq_cpu_get(cpu); 204 if (!policy) 205 return 0; 206 207 pd = em_cpu_get(cpu); 208 if (!pd || em_is_artificial(pd)) { 209 ret = -EINVAL; 210 goto release_policy; 211 } 212 213 dtpm_cpu = kzalloc(sizeof(*dtpm_cpu), GFP_KERNEL); 214 if (!dtpm_cpu) { 215 ret = -ENOMEM; 216 goto release_policy; 217 } 218 219 dtpm_init(&dtpm_cpu->dtpm, &dtpm_ops); 220 dtpm_cpu->cpu = cpu; 221 222 for_each_cpu(cpu, policy->related_cpus) 223 per_cpu(dtpm_per_cpu, cpu) = dtpm_cpu; 224 225 snprintf(name, sizeof(name), "cpu%d-cpufreq", dtpm_cpu->cpu); 226 227 ret = dtpm_register(name, &dtpm_cpu->dtpm, parent); 228 if (ret) 229 goto out_kfree_dtpm_cpu; 230 231 rcu_read_lock(); 232 table = em_perf_state_from_pd(pd); 233 ret = freq_qos_add_request(&policy->constraints, 234 &dtpm_cpu->qos_req, FREQ_QOS_MAX, 235 table[pd->nr_perf_states - 1].frequency); 236 rcu_read_unlock(); 237 if (ret < 0) 238 goto out_dtpm_unregister; 239 240 cpufreq_cpu_put(policy); 241 return 0; 242 243 out_dtpm_unregister: 244 dtpm_unregister(&dtpm_cpu->dtpm); 245 dtpm_cpu = NULL; 246 247 out_kfree_dtpm_cpu: 248 for_each_cpu(cpu, policy->related_cpus) 249 per_cpu(dtpm_per_cpu, cpu) = NULL; 250 kfree(dtpm_cpu); 251 252 release_policy: 253 cpufreq_cpu_put(policy); 254 return ret; 255 } 256 257 static int dtpm_cpu_setup(struct dtpm *dtpm, struct device_node *np) 258 { 259 int cpu; 260 261 cpu = of_cpu_node_to_id(np); 262 if (cpu < 0) 263 return 0; 264 265 return __dtpm_cpu_setup(cpu, dtpm); 266 } 267 268 static int dtpm_cpu_init(void) 269 { 270 int ret; 271 272 /* 273 * The callbacks at CPU hotplug time are calling 274 * dtpm_update_power() which in turns calls update_pd_power(). 275 * 276 * The function update_pd_power() uses the online mask to 277 * figure out the power consumption limits. 278 * 279 * At CPUHP_AP_ONLINE_DYN, the CPU is present in the CPU 280 * online mask when the cpuhp_dtpm_cpu_online function is 281 * called, but the CPU is still in the online mask for the 282 * tear down callback. So the power can not be updated when 283 * the CPU is unplugged. 284 * 285 * At CPUHP_AP_DTPM_CPU_DEAD, the situation is the opposite as 286 * above. The CPU online mask is not up to date when the CPU 287 * is plugged in. 288 * 289 * For this reason, we need to call the online and offline 290 * callbacks at different moments when the CPU online mask is 291 * consistent with the power numbers we want to update. 292 */ 293 ret = cpuhp_setup_state(CPUHP_AP_DTPM_CPU_DEAD, "dtpm_cpu:offline", 294 NULL, cpuhp_dtpm_cpu_offline); 295 if (ret < 0) 296 return ret; 297 298 ret = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "dtpm_cpu:online", 299 cpuhp_dtpm_cpu_online, NULL); 300 if (ret < 0) 301 return ret; 302 303 return 0; 304 } 305 306 static void dtpm_cpu_exit(void) 307 { 308 cpuhp_remove_state_nocalls(CPUHP_AP_ONLINE_DYN); 309 cpuhp_remove_state_nocalls(CPUHP_AP_DTPM_CPU_DEAD); 310 } 311 312 struct dtpm_subsys_ops dtpm_cpu_ops = { 313 .name = KBUILD_MODNAME, 314 .init = dtpm_cpu_init, 315 .exit = dtpm_cpu_exit, 316 .setup = dtpm_cpu_setup, 317 }; 318