xref: /openbmc/linux/drivers/powercap/dtpm_cpu.c (revision 1f012283)
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/pm_qos.h>
25 #include <linux/slab.h>
26 #include <linux/units.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 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 * MICROWATT_PER_MILLIWATT * 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 *
66 		MICROWATT_PER_MILLIWATT * nr_cpus;
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 = 0, sum_util = 0;
74 	int cpu;
75 
76 	for_each_cpu_and(cpu, pd_mask, cpu_online_mask) {
77 
78 		/*
79 		 * The capacity is the same for all CPUs belonging to
80 		 * the same perf domain, so a single call to
81 		 * arch_scale_cpu_capacity() is enough. However, we
82 		 * need the CPU parameter to be initialized by the
83 		 * loop, so the call ends up in this block.
84 		 *
85 		 * We can initialize 'max' with a cpumask_first() call
86 		 * before the loop but the bits computation is not
87 		 * worth given the arch_scale_cpu_capacity() just
88 		 * returns a value where the resulting assembly code
89 		 * will be optimized by the compiler.
90 		 */
91 		max = arch_scale_cpu_capacity(cpu);
92 		sum_util += sched_cpu_util(cpu, max);
93 	}
94 
95 	/*
96 	 * In the improbable case where all the CPUs of the perf
97 	 * domain are offline, 'max' will be zero and will lead to an
98 	 * illegal operation with a zero division.
99 	 */
100 	return max ? (power * ((sum_util << 10) / max)) >> 10 : 0;
101 }
102 
103 static u64 get_pd_power_uw(struct dtpm *dtpm)
104 {
105 	struct dtpm_cpu *dtpm_cpu = to_dtpm_cpu(dtpm);
106 	struct em_perf_domain *pd;
107 	struct cpumask *pd_mask;
108 	unsigned long freq;
109 	int i;
110 
111 	pd = em_cpu_get(dtpm_cpu->cpu);
112 
113 	pd_mask = em_span_cpus(pd);
114 
115 	freq = cpufreq_quick_get(dtpm_cpu->cpu);
116 
117 	for (i = 0; i < pd->nr_perf_states; i++) {
118 
119 		if (pd->table[i].frequency < freq)
120 			continue;
121 
122 		return scale_pd_power_uw(pd_mask, pd->table[i].power *
123 					 MICROWATT_PER_MILLIWATT);
124 	}
125 
126 	return 0;
127 }
128 
129 static int update_pd_power_uw(struct dtpm *dtpm)
130 {
131 	struct dtpm_cpu *dtpm_cpu = to_dtpm_cpu(dtpm);
132 	struct em_perf_domain *em = em_cpu_get(dtpm_cpu->cpu);
133 	struct cpumask cpus;
134 	int nr_cpus;
135 
136 	cpumask_and(&cpus, cpu_online_mask, to_cpumask(em->cpus));
137 	nr_cpus = cpumask_weight(&cpus);
138 
139 	dtpm->power_min = em->table[0].power;
140 	dtpm->power_min *= MICROWATT_PER_MILLIWATT;
141 	dtpm->power_min *= nr_cpus;
142 
143 	dtpm->power_max = em->table[em->nr_perf_states - 1].power;
144 	dtpm->power_max *= MICROWATT_PER_MILLIWATT;
145 	dtpm->power_max *= nr_cpus;
146 
147 	return 0;
148 }
149 
150 static void pd_release(struct dtpm *dtpm)
151 {
152 	struct dtpm_cpu *dtpm_cpu = to_dtpm_cpu(dtpm);
153 
154 	if (freq_qos_request_active(&dtpm_cpu->qos_req))
155 		freq_qos_remove_request(&dtpm_cpu->qos_req);
156 
157 	kfree(dtpm_cpu);
158 }
159 
160 static struct dtpm_ops dtpm_ops = {
161 	.set_power_uw	 = set_pd_power_limit,
162 	.get_power_uw	 = get_pd_power_uw,
163 	.update_power_uw = update_pd_power_uw,
164 	.release	 = pd_release,
165 };
166 
167 static int cpuhp_dtpm_cpu_offline(unsigned int cpu)
168 {
169 	struct em_perf_domain *pd;
170 	struct dtpm_cpu *dtpm_cpu;
171 
172 	pd = em_cpu_get(cpu);
173 	if (!pd)
174 		return -EINVAL;
175 
176 	dtpm_cpu = per_cpu(dtpm_per_cpu, cpu);
177 
178 	return dtpm_update_power(&dtpm_cpu->dtpm);
179 }
180 
181 static int cpuhp_dtpm_cpu_online(unsigned int cpu)
182 {
183 	struct dtpm_cpu *dtpm_cpu;
184 	struct cpufreq_policy *policy;
185 	struct em_perf_domain *pd;
186 	char name[CPUFREQ_NAME_LEN];
187 	int ret = -ENOMEM;
188 
189 	policy = cpufreq_cpu_get(cpu);
190 	if (!policy)
191 		return 0;
192 
193 	pd = em_cpu_get(cpu);
194 	if (!pd)
195 		return -EINVAL;
196 
197 	dtpm_cpu = per_cpu(dtpm_per_cpu, cpu);
198 	if (dtpm_cpu)
199 		return dtpm_update_power(&dtpm_cpu->dtpm);
200 
201 	dtpm_cpu = kzalloc(sizeof(*dtpm_cpu), GFP_KERNEL);
202 	if (!dtpm_cpu)
203 		return -ENOMEM;
204 
205 	dtpm_init(&dtpm_cpu->dtpm, &dtpm_ops);
206 	dtpm_cpu->cpu = cpu;
207 
208 	for_each_cpu(cpu, policy->related_cpus)
209 		per_cpu(dtpm_per_cpu, cpu) = dtpm_cpu;
210 
211 	snprintf(name, sizeof(name), "cpu%d-cpufreq", dtpm_cpu->cpu);
212 
213 	ret = dtpm_register(name, &dtpm_cpu->dtpm, NULL);
214 	if (ret)
215 		goto out_kfree_dtpm_cpu;
216 
217 	ret = freq_qos_add_request(&policy->constraints,
218 				   &dtpm_cpu->qos_req, FREQ_QOS_MAX,
219 				   pd->table[pd->nr_perf_states - 1].frequency);
220 	if (ret)
221 		goto out_dtpm_unregister;
222 
223 	return 0;
224 
225 out_dtpm_unregister:
226 	dtpm_unregister(&dtpm_cpu->dtpm);
227 	dtpm_cpu = NULL;
228 
229 out_kfree_dtpm_cpu:
230 	for_each_cpu(cpu, policy->related_cpus)
231 		per_cpu(dtpm_per_cpu, cpu) = NULL;
232 	kfree(dtpm_cpu);
233 
234 	return ret;
235 }
236 
237 static int __init dtpm_cpu_init(void)
238 {
239 	int ret;
240 
241 	/*
242 	 * The callbacks at CPU hotplug time are calling
243 	 * dtpm_update_power() which in turns calls update_pd_power().
244 	 *
245 	 * The function update_pd_power() uses the online mask to
246 	 * figure out the power consumption limits.
247 	 *
248 	 * At CPUHP_AP_ONLINE_DYN, the CPU is present in the CPU
249 	 * online mask when the cpuhp_dtpm_cpu_online function is
250 	 * called, but the CPU is still in the online mask for the
251 	 * tear down callback. So the power can not be updated when
252 	 * the CPU is unplugged.
253 	 *
254 	 * At CPUHP_AP_DTPM_CPU_DEAD, the situation is the opposite as
255 	 * above. The CPU online mask is not up to date when the CPU
256 	 * is plugged in.
257 	 *
258 	 * For this reason, we need to call the online and offline
259 	 * callbacks at different moments when the CPU online mask is
260 	 * consistent with the power numbers we want to update.
261 	 */
262 	ret = cpuhp_setup_state(CPUHP_AP_DTPM_CPU_DEAD, "dtpm_cpu:offline",
263 				NULL, cpuhp_dtpm_cpu_offline);
264 	if (ret < 0)
265 		return ret;
266 
267 	ret = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "dtpm_cpu:online",
268 				cpuhp_dtpm_cpu_online, NULL);
269 	if (ret < 0)
270 		return ret;
271 
272 	return 0;
273 }
274 
275 DTPM_DECLARE(dtpm_cpu, dtpm_cpu_init);
276