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
28struct dtpm_cpu {
29	struct dtpm dtpm;
30	struct freq_qos_request qos_req;
31	int cpu;
32};
33
34static DEFINE_PER_CPU(struct dtpm_cpu *, dtpm_per_cpu);
35
36static struct dtpm_cpu *to_dtpm_cpu(struct dtpm *dtpm)
37{
38	return container_of(dtpm, struct dtpm_cpu, dtpm);
39}
40
41static 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	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	rcu_read_lock();
55	table = em_perf_state_from_pd(pd);
56	for (i = 0; i < pd->nr_perf_states; i++) {
57
58		power = table[i].power * nr_cpus;
59
60		if (power > power_limit)
61			break;
62	}
63
64	freq = table[i - 1].frequency;
65	power_limit = table[i - 1].power * nr_cpus;
66	rcu_read_unlock();
67
68	freq_qos_update_request(&dtpm_cpu->qos_req, freq);
69
70	return power_limit;
71}
72
73static u64 scale_pd_power_uw(struct cpumask *pd_mask, u64 power)
74{
75	unsigned long max, sum_util = 0;
76	int cpu;
77
78	/*
79	 * The capacity is the same for all CPUs belonging to
80	 * the same perf domain.
81	 */
82	max = arch_scale_cpu_capacity(cpumask_first(pd_mask));
83
84	for_each_cpu_and(cpu, pd_mask, cpu_online_mask)
85		sum_util += sched_cpu_util(cpu);
86
87	return (power * ((sum_util << 10) / max)) >> 10;
88}
89
90static u64 get_pd_power_uw(struct dtpm *dtpm)
91{
92	struct dtpm_cpu *dtpm_cpu = to_dtpm_cpu(dtpm);
93	struct em_perf_state *table;
94	struct em_perf_domain *pd;
95	struct cpumask *pd_mask;
96	unsigned long freq;
97	u64 power = 0;
98	int i;
99
100	pd = em_cpu_get(dtpm_cpu->cpu);
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
121static 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	struct cpumask cpus;
127	int nr_cpus;
128
129	cpumask_and(&cpus, cpu_online_mask, to_cpumask(em->cpus));
130	nr_cpus = cpumask_weight(&cpus);
131
132	rcu_read_lock();
133	table = em_perf_state_from_pd(em);
134
135	dtpm->power_min = table[0].power;
136	dtpm->power_min *= nr_cpus;
137
138	dtpm->power_max = table[em->nr_perf_states - 1].power;
139	dtpm->power_max *= nr_cpus;
140
141	rcu_read_unlock();
142
143	return 0;
144}
145
146static void pd_release(struct dtpm *dtpm)
147{
148	struct dtpm_cpu *dtpm_cpu = to_dtpm_cpu(dtpm);
149	struct cpufreq_policy *policy;
150
151	if (freq_qos_request_active(&dtpm_cpu->qos_req))
152		freq_qos_remove_request(&dtpm_cpu->qos_req);
153
154	policy = cpufreq_cpu_get(dtpm_cpu->cpu);
155	if (policy) {
156		for_each_cpu(dtpm_cpu->cpu, policy->related_cpus)
157			per_cpu(dtpm_per_cpu, dtpm_cpu->cpu) = NULL;
158
159		cpufreq_cpu_put(policy);
160	}
161
162	kfree(dtpm_cpu);
163}
164
165static struct dtpm_ops dtpm_ops = {
166	.set_power_uw	 = set_pd_power_limit,
167	.get_power_uw	 = get_pd_power_uw,
168	.update_power_uw = update_pd_power_uw,
169	.release	 = pd_release,
170};
171
172static int cpuhp_dtpm_cpu_offline(unsigned int cpu)
173{
174	struct dtpm_cpu *dtpm_cpu;
175
176	dtpm_cpu = per_cpu(dtpm_per_cpu, cpu);
177	if (dtpm_cpu)
178		dtpm_update_power(&dtpm_cpu->dtpm);
179
180	return 0;
181}
182
183static int cpuhp_dtpm_cpu_online(unsigned int cpu)
184{
185	struct dtpm_cpu *dtpm_cpu;
186
187	dtpm_cpu = per_cpu(dtpm_per_cpu, cpu);
188	if (dtpm_cpu)
189		return dtpm_update_power(&dtpm_cpu->dtpm);
190
191	return 0;
192}
193
194static int __dtpm_cpu_setup(int cpu, struct dtpm *parent)
195{
196	struct dtpm_cpu *dtpm_cpu;
197	struct cpufreq_policy *policy;
198	struct em_perf_state *table;
199	struct em_perf_domain *pd;
200	char name[CPUFREQ_NAME_LEN];
201	int ret = -ENOMEM;
202
203	dtpm_cpu = per_cpu(dtpm_per_cpu, cpu);
204	if (dtpm_cpu)
205		return 0;
206
207	policy = cpufreq_cpu_get(cpu);
208	if (!policy)
209		return 0;
210
211	pd = em_cpu_get(cpu);
212	if (!pd || em_is_artificial(pd)) {
213		ret = -EINVAL;
214		goto release_policy;
215	}
216
217	dtpm_cpu = kzalloc(sizeof(*dtpm_cpu), GFP_KERNEL);
218	if (!dtpm_cpu) {
219		ret = -ENOMEM;
220		goto release_policy;
221	}
222
223	dtpm_init(&dtpm_cpu->dtpm, &dtpm_ops);
224	dtpm_cpu->cpu = cpu;
225
226	for_each_cpu(cpu, policy->related_cpus)
227		per_cpu(dtpm_per_cpu, cpu) = dtpm_cpu;
228
229	snprintf(name, sizeof(name), "cpu%d-cpufreq", dtpm_cpu->cpu);
230
231	ret = dtpm_register(name, &dtpm_cpu->dtpm, parent);
232	if (ret)
233		goto out_kfree_dtpm_cpu;
234
235	rcu_read_lock();
236	table = em_perf_state_from_pd(pd);
237	ret = freq_qos_add_request(&policy->constraints,
238				   &dtpm_cpu->qos_req, FREQ_QOS_MAX,
239				   table[pd->nr_perf_states - 1].frequency);
240	rcu_read_unlock();
241	if (ret < 0)
242		goto out_dtpm_unregister;
243
244	cpufreq_cpu_put(policy);
245	return 0;
246
247out_dtpm_unregister:
248	dtpm_unregister(&dtpm_cpu->dtpm);
249	dtpm_cpu = NULL;
250
251out_kfree_dtpm_cpu:
252	for_each_cpu(cpu, policy->related_cpus)
253		per_cpu(dtpm_per_cpu, cpu) = NULL;
254	kfree(dtpm_cpu);
255
256release_policy:
257	cpufreq_cpu_put(policy);
258	return ret;
259}
260
261static int dtpm_cpu_setup(struct dtpm *dtpm, struct device_node *np)
262{
263	int cpu;
264
265	cpu = of_cpu_node_to_id(np);
266	if (cpu < 0)
267		return 0;
268
269	return __dtpm_cpu_setup(cpu, dtpm);
270}
271
272static int dtpm_cpu_init(void)
273{
274	int ret;
275
276	/*
277	 * The callbacks at CPU hotplug time are calling
278	 * dtpm_update_power() which in turns calls update_pd_power().
279	 *
280	 * The function update_pd_power() uses the online mask to
281	 * figure out the power consumption limits.
282	 *
283	 * At CPUHP_AP_ONLINE_DYN, the CPU is present in the CPU
284	 * online mask when the cpuhp_dtpm_cpu_online function is
285	 * called, but the CPU is still in the online mask for the
286	 * tear down callback. So the power can not be updated when
287	 * the CPU is unplugged.
288	 *
289	 * At CPUHP_AP_DTPM_CPU_DEAD, the situation is the opposite as
290	 * above. The CPU online mask is not up to date when the CPU
291	 * is plugged in.
292	 *
293	 * For this reason, we need to call the online and offline
294	 * callbacks at different moments when the CPU online mask is
295	 * consistent with the power numbers we want to update.
296	 */
297	ret = cpuhp_setup_state(CPUHP_AP_DTPM_CPU_DEAD, "dtpm_cpu:offline",
298				NULL, cpuhp_dtpm_cpu_offline);
299	if (ret < 0)
300		return ret;
301
302	ret = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "dtpm_cpu:online",
303				cpuhp_dtpm_cpu_online, NULL);
304	if (ret < 0)
305		return ret;
306
307	return 0;
308}
309
310static void dtpm_cpu_exit(void)
311{
312	cpuhp_remove_state_nocalls(CPUHP_AP_ONLINE_DYN);
313	cpuhp_remove_state_nocalls(CPUHP_AP_DTPM_CPU_DEAD);
314}
315
316struct dtpm_subsys_ops dtpm_cpu_ops = {
317	.name = KBUILD_MODNAME,
318	.init = dtpm_cpu_init,
319	.exit = dtpm_cpu_exit,
320	.setup = dtpm_cpu_setup,
321};
322