x86/boot/64: Move 5-level paging global variable assignments back
[sfrench/cifs-2.6.git] / drivers / cpufreq / scmi-cpufreq.c
1 // SPDX-License-Identifier: GPL-2.0
2 /*
3  * System Control and Power Interface (SCMI) based CPUFreq Interface driver
4  *
5  * Copyright (C) 2018-2021 ARM Ltd.
6  * Sudeep Holla <sudeep.holla@arm.com>
7  */
8
9 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
10
11 #include <linux/clk-provider.h>
12 #include <linux/cpu.h>
13 #include <linux/cpufreq.h>
14 #include <linux/cpumask.h>
15 #include <linux/energy_model.h>
16 #include <linux/export.h>
17 #include <linux/module.h>
18 #include <linux/pm_opp.h>
19 #include <linux/slab.h>
20 #include <linux/scmi_protocol.h>
21 #include <linux/types.h>
22 #include <linux/units.h>
23
24 struct scmi_data {
25         int domain_id;
26         int nr_opp;
27         struct device *cpu_dev;
28         cpumask_var_t opp_shared_cpus;
29 };
30
31 static struct scmi_protocol_handle *ph;
32 static const struct scmi_perf_proto_ops *perf_ops;
33
34 static unsigned int scmi_cpufreq_get_rate(unsigned int cpu)
35 {
36         struct cpufreq_policy *policy = cpufreq_cpu_get_raw(cpu);
37         struct scmi_data *priv = policy->driver_data;
38         unsigned long rate;
39         int ret;
40
41         ret = perf_ops->freq_get(ph, priv->domain_id, &rate, false);
42         if (ret)
43                 return 0;
44         return rate / 1000;
45 }
46
47 /*
48  * perf_ops->freq_set is not a synchronous, the actual OPP change will
49  * happen asynchronously and can get notified if the events are
50  * subscribed for by the SCMI firmware
51  */
52 static int
53 scmi_cpufreq_set_target(struct cpufreq_policy *policy, unsigned int index)
54 {
55         struct scmi_data *priv = policy->driver_data;
56         u64 freq = policy->freq_table[index].frequency;
57
58         return perf_ops->freq_set(ph, priv->domain_id, freq * 1000, false);
59 }
60
61 static unsigned int scmi_cpufreq_fast_switch(struct cpufreq_policy *policy,
62                                              unsigned int target_freq)
63 {
64         struct scmi_data *priv = policy->driver_data;
65
66         if (!perf_ops->freq_set(ph, priv->domain_id,
67                                 target_freq * 1000, true))
68                 return target_freq;
69
70         return 0;
71 }
72
73 static int scmi_cpu_domain_id(struct device *cpu_dev)
74 {
75         struct device_node *np = cpu_dev->of_node;
76         struct of_phandle_args domain_id;
77         int index;
78
79         if (of_parse_phandle_with_args(np, "clocks", "#clock-cells", 0,
80                                        &domain_id)) {
81                 /* Find the corresponding index for power-domain "perf". */
82                 index = of_property_match_string(np, "power-domain-names",
83                                                  "perf");
84                 if (index < 0)
85                         return -EINVAL;
86
87                 if (of_parse_phandle_with_args(np, "power-domains",
88                                                "#power-domain-cells", index,
89                                                &domain_id))
90                         return -EINVAL;
91         }
92
93         return domain_id.args[0];
94 }
95
96 static int
97 scmi_get_sharing_cpus(struct device *cpu_dev, int domain,
98                       struct cpumask *cpumask)
99 {
100         int cpu, tdomain;
101         struct device *tcpu_dev;
102
103         for_each_possible_cpu(cpu) {
104                 if (cpu == cpu_dev->id)
105                         continue;
106
107                 tcpu_dev = get_cpu_device(cpu);
108                 if (!tcpu_dev)
109                         continue;
110
111                 tdomain = scmi_cpu_domain_id(tcpu_dev);
112                 if (tdomain == domain)
113                         cpumask_set_cpu(cpu, cpumask);
114         }
115
116         return 0;
117 }
118
119 static int __maybe_unused
120 scmi_get_cpu_power(struct device *cpu_dev, unsigned long *power,
121                    unsigned long *KHz)
122 {
123         enum scmi_power_scale power_scale = perf_ops->power_scale_get(ph);
124         unsigned long Hz;
125         int ret, domain;
126
127         domain = scmi_cpu_domain_id(cpu_dev);
128         if (domain < 0)
129                 return domain;
130
131         /* Get the power cost of the performance domain. */
132         Hz = *KHz * 1000;
133         ret = perf_ops->est_power_get(ph, domain, &Hz, power);
134         if (ret)
135                 return ret;
136
137         /* Convert the power to uW if it is mW (ignore bogoW) */
138         if (power_scale == SCMI_POWER_MILLIWATTS)
139                 *power *= MICROWATT_PER_MILLIWATT;
140
141         /* The EM framework specifies the frequency in KHz. */
142         *KHz = Hz / 1000;
143
144         return 0;
145 }
146
147 static int
148 scmi_get_rate_limit(u32 domain, bool has_fast_switch)
149 {
150         int ret, rate_limit;
151
152         if (has_fast_switch) {
153                 /*
154                  * Fast channels are used whenever available,
155                  * so use their rate_limit value if populated.
156                  */
157                 ret = perf_ops->fast_switch_rate_limit(ph, domain,
158                                                        &rate_limit);
159                 if (!ret && rate_limit)
160                         return rate_limit;
161         }
162
163         ret = perf_ops->rate_limit_get(ph, domain, &rate_limit);
164         if (ret)
165                 return 0;
166
167         return rate_limit;
168 }
169
170 static int scmi_cpufreq_init(struct cpufreq_policy *policy)
171 {
172         int ret, nr_opp, domain;
173         unsigned int latency;
174         struct device *cpu_dev;
175         struct scmi_data *priv;
176         struct cpufreq_frequency_table *freq_table;
177
178         cpu_dev = get_cpu_device(policy->cpu);
179         if (!cpu_dev) {
180                 pr_err("failed to get cpu%d device\n", policy->cpu);
181                 return -ENODEV;
182         }
183
184         domain = scmi_cpu_domain_id(cpu_dev);
185         if (domain < 0)
186                 return domain;
187
188         priv = kzalloc(sizeof(*priv), GFP_KERNEL);
189         if (!priv)
190                 return -ENOMEM;
191
192         if (!zalloc_cpumask_var(&priv->opp_shared_cpus, GFP_KERNEL)) {
193                 ret = -ENOMEM;
194                 goto out_free_priv;
195         }
196
197         /* Obtain CPUs that share SCMI performance controls */
198         ret = scmi_get_sharing_cpus(cpu_dev, domain, policy->cpus);
199         if (ret) {
200                 dev_warn(cpu_dev, "failed to get sharing cpumask\n");
201                 goto out_free_cpumask;
202         }
203
204         /*
205          * Obtain CPUs that share performance levels.
206          * The OPP 'sharing cpus' info may come from DT through an empty opp
207          * table and opp-shared.
208          */
209         ret = dev_pm_opp_of_get_sharing_cpus(cpu_dev, priv->opp_shared_cpus);
210         if (ret || cpumask_empty(priv->opp_shared_cpus)) {
211                 /*
212                  * Either opp-table is not set or no opp-shared was found.
213                  * Use the CPU mask from SCMI to designate CPUs sharing an OPP
214                  * table.
215                  */
216                 cpumask_copy(priv->opp_shared_cpus, policy->cpus);
217         }
218
219          /*
220           * A previous CPU may have marked OPPs as shared for a few CPUs, based on
221           * what OPP core provided. If the current CPU is part of those few, then
222           * there is no need to add OPPs again.
223           */
224         nr_opp = dev_pm_opp_get_opp_count(cpu_dev);
225         if (nr_opp <= 0) {
226                 ret = perf_ops->device_opps_add(ph, cpu_dev, domain);
227                 if (ret) {
228                         dev_warn(cpu_dev, "failed to add opps to the device\n");
229                         goto out_free_cpumask;
230                 }
231
232                 nr_opp = dev_pm_opp_get_opp_count(cpu_dev);
233                 if (nr_opp <= 0) {
234                         dev_err(cpu_dev, "%s: No OPPs for this device: %d\n",
235                                 __func__, nr_opp);
236
237                         ret = -ENODEV;
238                         goto out_free_opp;
239                 }
240
241                 ret = dev_pm_opp_set_sharing_cpus(cpu_dev, priv->opp_shared_cpus);
242                 if (ret) {
243                         dev_err(cpu_dev, "%s: failed to mark OPPs as shared: %d\n",
244                                 __func__, ret);
245
246                         goto out_free_opp;
247                 }
248
249                 priv->nr_opp = nr_opp;
250         }
251
252         ret = dev_pm_opp_init_cpufreq_table(cpu_dev, &freq_table);
253         if (ret) {
254                 dev_err(cpu_dev, "failed to init cpufreq table: %d\n", ret);
255                 goto out_free_opp;
256         }
257
258         priv->cpu_dev = cpu_dev;
259         priv->domain_id = domain;
260
261         policy->driver_data = priv;
262         policy->freq_table = freq_table;
263
264         /* SCMI allows DVFS request for any domain from any CPU */
265         policy->dvfs_possible_from_any_cpu = true;
266
267         latency = perf_ops->transition_latency_get(ph, domain);
268         if (!latency)
269                 latency = CPUFREQ_ETERNAL;
270
271         policy->cpuinfo.transition_latency = latency;
272
273         policy->fast_switch_possible =
274                 perf_ops->fast_switch_possible(ph, domain);
275
276         policy->transition_delay_us =
277                 scmi_get_rate_limit(domain, policy->fast_switch_possible);
278
279         return 0;
280
281 out_free_opp:
282         dev_pm_opp_remove_all_dynamic(cpu_dev);
283
284 out_free_cpumask:
285         free_cpumask_var(priv->opp_shared_cpus);
286
287 out_free_priv:
288         kfree(priv);
289
290         return ret;
291 }
292
293 static int scmi_cpufreq_exit(struct cpufreq_policy *policy)
294 {
295         struct scmi_data *priv = policy->driver_data;
296
297         dev_pm_opp_free_cpufreq_table(priv->cpu_dev, &policy->freq_table);
298         dev_pm_opp_remove_all_dynamic(priv->cpu_dev);
299         free_cpumask_var(priv->opp_shared_cpus);
300         kfree(priv);
301
302         return 0;
303 }
304
305 static void scmi_cpufreq_register_em(struct cpufreq_policy *policy)
306 {
307         struct em_data_callback em_cb = EM_DATA_CB(scmi_get_cpu_power);
308         enum scmi_power_scale power_scale = perf_ops->power_scale_get(ph);
309         struct scmi_data *priv = policy->driver_data;
310         bool em_power_scale = false;
311
312         /*
313          * This callback will be called for each policy, but we don't need to
314          * register with EM every time. Despite not being part of the same
315          * policy, some CPUs may still share their perf-domains, and a CPU from
316          * another policy may already have registered with EM on behalf of CPUs
317          * of this policy.
318          */
319         if (!priv->nr_opp)
320                 return;
321
322         if (power_scale == SCMI_POWER_MILLIWATTS
323             || power_scale == SCMI_POWER_MICROWATTS)
324                 em_power_scale = true;
325
326         em_dev_register_perf_domain(get_cpu_device(policy->cpu), priv->nr_opp,
327                                     &em_cb, priv->opp_shared_cpus,
328                                     em_power_scale);
329 }
330
331 static struct cpufreq_driver scmi_cpufreq_driver = {
332         .name   = "scmi",
333         .flags  = CPUFREQ_HAVE_GOVERNOR_PER_POLICY |
334                   CPUFREQ_NEED_INITIAL_FREQ_CHECK |
335                   CPUFREQ_IS_COOLING_DEV,
336         .verify = cpufreq_generic_frequency_table_verify,
337         .attr   = cpufreq_generic_attr,
338         .target_index   = scmi_cpufreq_set_target,
339         .fast_switch    = scmi_cpufreq_fast_switch,
340         .get    = scmi_cpufreq_get_rate,
341         .init   = scmi_cpufreq_init,
342         .exit   = scmi_cpufreq_exit,
343         .register_em    = scmi_cpufreq_register_em,
344 };
345
346 static int scmi_cpufreq_probe(struct scmi_device *sdev)
347 {
348         int ret;
349         struct device *dev = &sdev->dev;
350         const struct scmi_handle *handle;
351
352         handle = sdev->handle;
353
354         if (!handle)
355                 return -ENODEV;
356
357         perf_ops = handle->devm_protocol_get(sdev, SCMI_PROTOCOL_PERF, &ph);
358         if (IS_ERR(perf_ops))
359                 return PTR_ERR(perf_ops);
360
361 #ifdef CONFIG_COMMON_CLK
362         /* dummy clock provider as needed by OPP if clocks property is used */
363         if (of_property_present(dev->of_node, "#clock-cells")) {
364                 ret = devm_of_clk_add_hw_provider(dev, of_clk_hw_simple_get, NULL);
365                 if (ret)
366                         return dev_err_probe(dev, ret, "%s: registering clock provider failed\n", __func__);
367         }
368 #endif
369
370         ret = cpufreq_register_driver(&scmi_cpufreq_driver);
371         if (ret) {
372                 dev_err(dev, "%s: registering cpufreq failed, err: %d\n",
373                         __func__, ret);
374         }
375
376         return ret;
377 }
378
379 static void scmi_cpufreq_remove(struct scmi_device *sdev)
380 {
381         cpufreq_unregister_driver(&scmi_cpufreq_driver);
382 }
383
384 static const struct scmi_device_id scmi_id_table[] = {
385         { SCMI_PROTOCOL_PERF, "cpufreq" },
386         { },
387 };
388 MODULE_DEVICE_TABLE(scmi, scmi_id_table);
389
390 static struct scmi_driver scmi_cpufreq_drv = {
391         .name           = "scmi-cpufreq",
392         .probe          = scmi_cpufreq_probe,
393         .remove         = scmi_cpufreq_remove,
394         .id_table       = scmi_id_table,
395 };
396 module_scmi_driver(scmi_cpufreq_drv);
397
398 MODULE_AUTHOR("Sudeep Holla <sudeep.holla@arm.com>");
399 MODULE_DESCRIPTION("ARM SCMI CPUFreq interface driver");
400 MODULE_LICENSE("GPL v2");