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 cpumask cpus;
 46	unsigned long freq;
 47	u64 power;
 48	int i, nr_cpus;
 49
 50	cpumask_and(&cpus, cpu_online_mask, to_cpumask(pd->cpus));
 51	nr_cpus = cpumask_weight(&cpus);
 52
 53	for (i = 0; i < pd->nr_perf_states; i++) {
 54
 55		power = pd->table[i].power * nr_cpus;
 56
 57		if (power > power_limit)
 58			break;
 59	}
 60
 61	freq = pd->table[i - 1].frequency;
 62
 63	freq_qos_update_request(&dtpm_cpu->qos_req, freq);
 64
 65	power_limit = pd->table[i - 1].power * nr_cpus;
 66
 67	return power_limit;
 68}
 69
 70static u64 scale_pd_power_uw(struct cpumask *pd_mask, u64 power)
 71{
 72	unsigned long max, sum_util = 0;
 73	int cpu;
 74
 75	/*
 76	 * The capacity is the same for all CPUs belonging to
 77	 * the same perf domain.
 78	 */
 79	max = arch_scale_cpu_capacity(cpumask_first(pd_mask));
 80
 81	for_each_cpu_and(cpu, pd_mask, cpu_online_mask)
 82		sum_util += sched_cpu_util(cpu);
 83
 84	return (power * ((sum_util << 10) / max)) >> 10;
 85}
 86
 87static u64 get_pd_power_uw(struct dtpm *dtpm)
 88{
 89	struct dtpm_cpu *dtpm_cpu = to_dtpm_cpu(dtpm);
 90	struct em_perf_domain *pd;
 91	struct cpumask *pd_mask;
 92	unsigned long freq;
 93	int i;
 94
 95	pd = em_cpu_get(dtpm_cpu->cpu);
 96
 97	pd_mask = em_span_cpus(pd);
 98
 99	freq = cpufreq_quick_get(dtpm_cpu->cpu);
100
101	for (i = 0; i < pd->nr_perf_states; i++) {
102
103		if (pd->table[i].frequency < freq)
104			continue;
105
106		return scale_pd_power_uw(pd_mask, pd->table[i].power);
107	}
108
109	return 0;
110}
111
112static int update_pd_power_uw(struct dtpm *dtpm)
113{
114	struct dtpm_cpu *dtpm_cpu = to_dtpm_cpu(dtpm);
115	struct em_perf_domain *em = em_cpu_get(dtpm_cpu->cpu);
116	struct cpumask cpus;
117	int nr_cpus;
118
119	cpumask_and(&cpus, cpu_online_mask, to_cpumask(em->cpus));
120	nr_cpus = cpumask_weight(&cpus);
121
122	dtpm->power_min = em->table[0].power;
123	dtpm->power_min *= nr_cpus;
124
125	dtpm->power_max = em->table[em->nr_perf_states - 1].power;
126	dtpm->power_max *= nr_cpus;
127
128	return 0;
129}
130
131static void pd_release(struct dtpm *dtpm)
132{
133	struct dtpm_cpu *dtpm_cpu = to_dtpm_cpu(dtpm);
134	struct cpufreq_policy *policy;
135
136	if (freq_qos_request_active(&dtpm_cpu->qos_req))
137		freq_qos_remove_request(&dtpm_cpu->qos_req);
138
139	policy = cpufreq_cpu_get(dtpm_cpu->cpu);
140	if (policy) {
141		for_each_cpu(dtpm_cpu->cpu, policy->related_cpus)
142			per_cpu(dtpm_per_cpu, dtpm_cpu->cpu) = NULL;
143
144		cpufreq_cpu_put(policy);
145	}
146	
147	kfree(dtpm_cpu);
148}
149
150static struct dtpm_ops dtpm_ops = {
151	.set_power_uw	 = set_pd_power_limit,
152	.get_power_uw	 = get_pd_power_uw,
153	.update_power_uw = update_pd_power_uw,
154	.release	 = pd_release,
155};
156
157static int cpuhp_dtpm_cpu_offline(unsigned int cpu)
158{
159	struct dtpm_cpu *dtpm_cpu;
160
161	dtpm_cpu = per_cpu(dtpm_per_cpu, cpu);
162	if (dtpm_cpu)
163		dtpm_update_power(&dtpm_cpu->dtpm);
164
165	return 0;
166}
167
168static int cpuhp_dtpm_cpu_online(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		return dtpm_update_power(&dtpm_cpu->dtpm);
175
176	return 0;
177}
178
179static int __dtpm_cpu_setup(int cpu, struct dtpm *parent)
180{
181	struct dtpm_cpu *dtpm_cpu;
182	struct cpufreq_policy *policy;
183	struct em_perf_domain *pd;
184	char name[CPUFREQ_NAME_LEN];
185	int ret = -ENOMEM;
186
187	dtpm_cpu = per_cpu(dtpm_per_cpu, cpu);
188	if (dtpm_cpu)
189		return 0;
190
191	policy = cpufreq_cpu_get(cpu);
192	if (!policy)
193		return 0;
194
195	pd = em_cpu_get(cpu);
196	if (!pd || em_is_artificial(pd)) {
197		ret = -EINVAL;
198		goto release_policy;
199	}
200
201	dtpm_cpu = kzalloc(sizeof(*dtpm_cpu), GFP_KERNEL);
202	if (!dtpm_cpu) {
203		ret = -ENOMEM;
204		goto release_policy;
205	}
206
207	dtpm_init(&dtpm_cpu->dtpm, &dtpm_ops);
208	dtpm_cpu->cpu = cpu;
209
210	for_each_cpu(cpu, policy->related_cpus)
211		per_cpu(dtpm_per_cpu, cpu) = dtpm_cpu;
212
213	snprintf(name, sizeof(name), "cpu%d-cpufreq", dtpm_cpu->cpu);
214
215	ret = dtpm_register(name, &dtpm_cpu->dtpm, parent);
216	if (ret)
217		goto out_kfree_dtpm_cpu;
218
219	ret = freq_qos_add_request(&policy->constraints,
220				   &dtpm_cpu->qos_req, FREQ_QOS_MAX,
221				   pd->table[pd->nr_perf_states - 1].frequency);
222	if (ret)
223		goto out_dtpm_unregister;
224
225	cpufreq_cpu_put(policy);
226	return 0;
227
228out_dtpm_unregister:
229	dtpm_unregister(&dtpm_cpu->dtpm);
230	dtpm_cpu = NULL;
231
232out_kfree_dtpm_cpu:
233	for_each_cpu(cpu, policy->related_cpus)
234		per_cpu(dtpm_per_cpu, cpu) = NULL;
235	kfree(dtpm_cpu);
236
237release_policy:
238	cpufreq_cpu_put(policy);
239	return ret;
240}
241
242static int dtpm_cpu_setup(struct dtpm *dtpm, struct device_node *np)
243{
244	int cpu;
245
246	cpu = of_cpu_node_to_id(np);
247	if (cpu < 0)
248		return 0;
249
250	return __dtpm_cpu_setup(cpu, dtpm);
251}
252
253static int dtpm_cpu_init(void)
254{
255	int ret;
256
257	/*
258	 * The callbacks at CPU hotplug time are calling
259	 * dtpm_update_power() which in turns calls update_pd_power().
260	 *
261	 * The function update_pd_power() uses the online mask to
262	 * figure out the power consumption limits.
263	 *
264	 * At CPUHP_AP_ONLINE_DYN, the CPU is present in the CPU
265	 * online mask when the cpuhp_dtpm_cpu_online function is
266	 * called, but the CPU is still in the online mask for the
267	 * tear down callback. So the power can not be updated when
268	 * the CPU is unplugged.
269	 *
270	 * At CPUHP_AP_DTPM_CPU_DEAD, the situation is the opposite as
271	 * above. The CPU online mask is not up to date when the CPU
272	 * is plugged in.
273	 *
274	 * For this reason, we need to call the online and offline
275	 * callbacks at different moments when the CPU online mask is
276	 * consistent with the power numbers we want to update.
277	 */
278	ret = cpuhp_setup_state(CPUHP_AP_DTPM_CPU_DEAD, "dtpm_cpu:offline",
279				NULL, cpuhp_dtpm_cpu_offline);
280	if (ret < 0)
281		return ret;
282
283	ret = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "dtpm_cpu:online",
284				cpuhp_dtpm_cpu_online, NULL);
285	if (ret < 0)
286		return ret;
287
288	return 0;
289}
290
291static void dtpm_cpu_exit(void)
292{
293	cpuhp_remove_state_nocalls(CPUHP_AP_ONLINE_DYN);
294	cpuhp_remove_state_nocalls(CPUHP_AP_DTPM_CPU_DEAD);
295}
296
297struct dtpm_subsys_ops dtpm_cpu_ops = {
298	.name = KBUILD_MODNAME,
299	.init = dtpm_cpu_init,
300	.exit = dtpm_cpu_exit,
301	.setup = dtpm_cpu_setup,
302};