xref: /linux/drivers/powercap/dtpm_cpu.c (revision 53edfecef66bfa65882ae065ed1a52f466c88979)
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 
to_dtpm_cpu(struct dtpm * dtpm)36 static struct dtpm_cpu *to_dtpm_cpu(struct dtpm *dtpm)
37 {
38 	return container_of(dtpm, struct dtpm_cpu, dtpm);
39 }
40 
set_pd_power_limit(struct dtpm * dtpm,u64 power_limit)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 
scale_pd_power_uw(struct cpumask * pd_mask,u64 power)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 
get_pd_power_uw(struct dtpm * dtpm)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 	if (!pd)
100 		return 0;
101 
102 	pd_mask = em_span_cpus(pd);
103 
104 	freq = cpufreq_quick_get(dtpm_cpu->cpu);
105 
106 	rcu_read_lock();
107 	table = em_perf_state_from_pd(pd);
108 	for (i = 0; i < pd->nr_perf_states; i++) {
109 
110 		if (table[i].frequency < freq)
111 			continue;
112 
113 		power = scale_pd_power_uw(pd_mask, table[i].power);
114 		break;
115 	}
116 	rcu_read_unlock();
117 
118 	return power;
119 }
120 
update_pd_power_uw(struct dtpm * dtpm)121 static int update_pd_power_uw(struct dtpm *dtpm)
122 {
123 	struct dtpm_cpu *dtpm_cpu = to_dtpm_cpu(dtpm);
124 	struct em_perf_domain *em = em_cpu_get(dtpm_cpu->cpu);
125 	struct em_perf_state *table;
126 	int nr_cpus;
127 
128 	nr_cpus = cpumask_weight_and(cpu_online_mask, to_cpumask(em->cpus));
129 
130 	rcu_read_lock();
131 	table = em_perf_state_from_pd(em);
132 
133 	dtpm->power_min = table[0].power;
134 	dtpm->power_min *= nr_cpus;
135 
136 	dtpm->power_max = table[em->nr_perf_states - 1].power;
137 	dtpm->power_max *= nr_cpus;
138 
139 	rcu_read_unlock();
140 
141 	return 0;
142 }
143 
pd_release(struct dtpm * dtpm)144 static void pd_release(struct dtpm *dtpm)
145 {
146 	struct dtpm_cpu *dtpm_cpu = to_dtpm_cpu(dtpm);
147 	struct cpufreq_policy *policy;
148 
149 	if (freq_qos_request_active(&dtpm_cpu->qos_req))
150 		freq_qos_remove_request(&dtpm_cpu->qos_req);
151 
152 	policy = cpufreq_cpu_get(dtpm_cpu->cpu);
153 	if (policy) {
154 		for_each_cpu(dtpm_cpu->cpu, policy->related_cpus)
155 			per_cpu(dtpm_per_cpu, dtpm_cpu->cpu) = NULL;
156 
157 		cpufreq_cpu_put(policy);
158 	}
159 
160 	kfree(dtpm_cpu);
161 }
162 
163 static struct dtpm_ops dtpm_ops = {
164 	.set_power_uw	 = set_pd_power_limit,
165 	.get_power_uw	 = get_pd_power_uw,
166 	.update_power_uw = update_pd_power_uw,
167 	.release	 = pd_release,
168 };
169 
cpuhp_dtpm_cpu_offline(unsigned int cpu)170 static int cpuhp_dtpm_cpu_offline(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 		dtpm_update_power(&dtpm_cpu->dtpm);
177 
178 	return 0;
179 }
180 
cpuhp_dtpm_cpu_online(unsigned int cpu)181 static int cpuhp_dtpm_cpu_online(unsigned int cpu)
182 {
183 	struct dtpm_cpu *dtpm_cpu;
184 
185 	dtpm_cpu = per_cpu(dtpm_per_cpu, cpu);
186 	if (dtpm_cpu)
187 		return dtpm_update_power(&dtpm_cpu->dtpm);
188 
189 	return 0;
190 }
191 
__dtpm_cpu_setup(int cpu,struct dtpm * parent)192 static int __dtpm_cpu_setup(int cpu, struct dtpm *parent)
193 {
194 	struct dtpm_cpu *dtpm_cpu;
195 	struct cpufreq_policy *policy;
196 	struct em_perf_state *table;
197 	struct em_perf_domain *pd;
198 	char name[CPUFREQ_NAME_LEN];
199 	int ret = -ENOMEM;
200 
201 	dtpm_cpu = per_cpu(dtpm_per_cpu, cpu);
202 	if (dtpm_cpu)
203 		return 0;
204 
205 	policy = cpufreq_cpu_get(cpu);
206 	if (!policy)
207 		return 0;
208 
209 	pd = em_cpu_get(cpu);
210 	if (!pd || em_is_artificial(pd)) {
211 		ret = -EINVAL;
212 		goto release_policy;
213 	}
214 
215 	dtpm_cpu = kzalloc(sizeof(*dtpm_cpu), GFP_KERNEL);
216 	if (!dtpm_cpu) {
217 		ret = -ENOMEM;
218 		goto release_policy;
219 	}
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 	rcu_read_lock();
234 	table = em_perf_state_from_pd(pd);
235 	ret = freq_qos_add_request(&policy->constraints,
236 				   &dtpm_cpu->qos_req, FREQ_QOS_MAX,
237 				   table[pd->nr_perf_states - 1].frequency);
238 	rcu_read_unlock();
239 	if (ret < 0)
240 		goto out_dtpm_unregister;
241 
242 	cpufreq_cpu_put(policy);
243 	return 0;
244 
245 out_dtpm_unregister:
246 	dtpm_unregister(&dtpm_cpu->dtpm);
247 	dtpm_cpu = NULL;
248 
249 out_kfree_dtpm_cpu:
250 	for_each_cpu(cpu, policy->related_cpus)
251 		per_cpu(dtpm_per_cpu, cpu) = NULL;
252 	kfree(dtpm_cpu);
253 
254 release_policy:
255 	cpufreq_cpu_put(policy);
256 	return ret;
257 }
258 
dtpm_cpu_setup(struct dtpm * dtpm,struct device_node * np)259 static int dtpm_cpu_setup(struct dtpm *dtpm, struct device_node *np)
260 {
261 	int cpu;
262 
263 	cpu = of_cpu_node_to_id(np);
264 	if (cpu < 0)
265 		return 0;
266 
267 	return __dtpm_cpu_setup(cpu, dtpm);
268 }
269 
dtpm_cpu_init(void)270 static int dtpm_cpu_init(void)
271 {
272 	int ret;
273 
274 	/*
275 	 * The callbacks at CPU hotplug time are calling
276 	 * dtpm_update_power() which in turns calls update_pd_power().
277 	 *
278 	 * The function update_pd_power() uses the online mask to
279 	 * figure out the power consumption limits.
280 	 *
281 	 * At CPUHP_AP_ONLINE_DYN, the CPU is present in the CPU
282 	 * online mask when the cpuhp_dtpm_cpu_online function is
283 	 * called, but the CPU is still in the online mask for the
284 	 * tear down callback. So the power can not be updated when
285 	 * the CPU is unplugged.
286 	 *
287 	 * At CPUHP_AP_DTPM_CPU_DEAD, the situation is the opposite as
288 	 * above. The CPU online mask is not up to date when the CPU
289 	 * is plugged in.
290 	 *
291 	 * For this reason, we need to call the online and offline
292 	 * callbacks at different moments when the CPU online mask is
293 	 * consistent with the power numbers we want to update.
294 	 */
295 	ret = cpuhp_setup_state(CPUHP_AP_DTPM_CPU_DEAD, "dtpm_cpu:offline",
296 				NULL, cpuhp_dtpm_cpu_offline);
297 	if (ret < 0)
298 		return ret;
299 
300 	ret = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "dtpm_cpu:online",
301 				cpuhp_dtpm_cpu_online, NULL);
302 	if (ret < 0)
303 		return ret;
304 
305 	return 0;
306 }
307 
dtpm_cpu_exit(void)308 static void dtpm_cpu_exit(void)
309 {
310 	cpuhp_remove_state_nocalls(CPUHP_AP_ONLINE_DYN);
311 	cpuhp_remove_state_nocalls(CPUHP_AP_DTPM_CPU_DEAD);
312 }
313 
314 struct dtpm_subsys_ops dtpm_cpu_ops = {
315 	.name = KBUILD_MODNAME,
316 	.init = dtpm_cpu_init,
317 	.exit = dtpm_cpu_exit,
318 	.setup = dtpm_cpu_setup,
319 };
320