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/pm_qos.h>
20 #include <linux/slab.h>
21 #include <linux/scmi_protocol.h>
22 #include <linux/types.h>
23 #include <linux/units.h>
24 #include <trace/hooks/cpufreq.h>
25
26 struct scmi_data {
27 int domain_id;
28 int nr_opp;
29 struct device *cpu_dev;
30 cpumask_var_t opp_shared_cpus;
31 struct notifier_block limit_notify_nb;
32 struct freq_qos_request limits_freq_req;
33 };
34
35 static struct scmi_protocol_handle *ph;
36 static const struct scmi_perf_proto_ops *perf_ops;
37 static struct cpufreq_driver scmi_cpufreq_driver;
38
scmi_cpufreq_get_rate(unsigned int cpu)39 static unsigned int scmi_cpufreq_get_rate(unsigned int cpu)
40 {
41 struct cpufreq_policy *policy;
42 struct scmi_data *priv;
43 unsigned long rate;
44 int ret;
45
46 policy = cpufreq_cpu_get_raw(cpu);
47 if (unlikely(!policy))
48 return 0;
49
50 priv = policy->driver_data;
51
52 ret = perf_ops->freq_get(ph, priv->domain_id, &rate, false);
53 if (ret)
54 return 0;
55 return rate / 1000;
56 }
57
58 /*
59 * perf_ops->freq_set is not a synchronous, the actual OPP change will
60 * happen asynchronously and can get notified if the events are
61 * subscribed for by the SCMI firmware
62 */
63 static int
scmi_cpufreq_set_target(struct cpufreq_policy * policy,unsigned int index)64 scmi_cpufreq_set_target(struct cpufreq_policy *policy, unsigned int index)
65 {
66 struct scmi_data *priv = policy->driver_data;
67 u64 freq = policy->freq_table[index].frequency;
68
69 return perf_ops->freq_set(ph, priv->domain_id, freq * 1000, false);
70 }
71
scmi_cpufreq_fast_switch(struct cpufreq_policy * policy,unsigned int target_freq)72 static unsigned int scmi_cpufreq_fast_switch(struct cpufreq_policy *policy,
73 unsigned int target_freq)
74 {
75 struct scmi_data *priv = policy->driver_data;
76 unsigned long freq = target_freq;
77
78 if (!perf_ops->freq_set(ph, priv->domain_id, freq * 1000, true))
79 return target_freq;
80
81 return 0;
82 }
83
scmi_cpu_domain_id(struct device * cpu_dev)84 static int scmi_cpu_domain_id(struct device *cpu_dev)
85 {
86 struct device_node *np = cpu_dev->of_node;
87 struct of_phandle_args domain_id;
88 int index;
89
90 if (of_parse_phandle_with_args(np, "clocks", "#clock-cells", 0,
91 &domain_id)) {
92 /* Find the corresponding index for power-domain "perf". */
93 index = of_property_match_string(np, "power-domain-names",
94 "perf");
95 if (index < 0)
96 return -EINVAL;
97
98 if (of_parse_phandle_with_args(np, "power-domains",
99 "#power-domain-cells", index,
100 &domain_id))
101 return -EINVAL;
102 }
103
104 return domain_id.args[0];
105 }
106
107 static int
scmi_get_sharing_cpus(struct device * cpu_dev,int domain,struct cpumask * cpumask)108 scmi_get_sharing_cpus(struct device *cpu_dev, int domain,
109 struct cpumask *cpumask)
110 {
111 int cpu, tdomain;
112 struct device *tcpu_dev;
113
114 for_each_possible_cpu(cpu) {
115 if (cpu == cpu_dev->id)
116 continue;
117
118 tcpu_dev = get_cpu_device(cpu);
119 if (!tcpu_dev)
120 continue;
121
122 tdomain = scmi_cpu_domain_id(tcpu_dev);
123 if (tdomain == domain)
124 cpumask_set_cpu(cpu, cpumask);
125 }
126
127 return 0;
128 }
129
130 static int __maybe_unused
scmi_get_cpu_power(struct device * cpu_dev,unsigned long * power,unsigned long * KHz)131 scmi_get_cpu_power(struct device *cpu_dev, unsigned long *power,
132 unsigned long *KHz)
133 {
134 enum scmi_power_scale power_scale = perf_ops->power_scale_get(ph);
135 unsigned long Hz;
136 int ret, domain;
137
138 domain = scmi_cpu_domain_id(cpu_dev);
139 if (domain < 0)
140 return domain;
141
142 /* Get the power cost of the performance domain. */
143 Hz = *KHz * 1000;
144 ret = perf_ops->est_power_get(ph, domain, &Hz, power);
145 if (ret)
146 return ret;
147
148 /* Convert the power to uW if it is mW (ignore bogoW) */
149 if (power_scale == SCMI_POWER_MILLIWATTS)
150 *power *= MICROWATT_PER_MILLIWATT;
151
152 /* The EM framework specifies the frequency in KHz. */
153 *KHz = Hz / 1000;
154
155 return 0;
156 }
157
158 static int
scmi_get_rate_limit(u32 domain,bool has_fast_switch)159 scmi_get_rate_limit(u32 domain, bool has_fast_switch)
160 {
161 int ret, rate_limit;
162
163 if (has_fast_switch) {
164 /*
165 * Fast channels are used whenever available,
166 * so use their rate_limit value if populated.
167 */
168 ret = perf_ops->fast_switch_rate_limit(ph, domain,
169 &rate_limit);
170 if (!ret && rate_limit)
171 return rate_limit;
172 }
173
174 ret = perf_ops->rate_limit_get(ph, domain, &rate_limit);
175 if (ret)
176 return 0;
177
178 return rate_limit;
179 }
180
181 static struct freq_attr *scmi_cpufreq_hw_attr[] = {
182 &cpufreq_freq_attr_scaling_available_freqs,
183 NULL,
184 NULL,
185 };
186
scmi_limit_notify_cb(struct notifier_block * nb,unsigned long event,void * data)187 static int scmi_limit_notify_cb(struct notifier_block *nb, unsigned long event, void *data)
188 {
189 struct scmi_data *priv = container_of(nb, struct scmi_data, limit_notify_nb);
190 struct scmi_perf_limits_report *limit_notify = data;
191 unsigned int limit_freq_khz;
192 bool done = false;
193 int ret = 0;
194
195 limit_freq_khz = limit_notify->range_max_freq / HZ_PER_KHZ;
196
197 trace_android_rvh_scmi_limit_notify_cb(&done, priv->cpu_dev, limit_freq_khz);
198 if (done)
199 return NOTIFY_OK;
200
201 if (ret < 0)
202 pr_warn("failed to update freq constraint: %d\n", ret);
203
204 return NOTIFY_OK;
205 }
206
scmi_cpufreq_init(struct cpufreq_policy * policy)207 static int scmi_cpufreq_init(struct cpufreq_policy *policy)
208 {
209 int ret, nr_opp, domain;
210 unsigned int latency;
211 struct device *cpu_dev;
212 struct scmi_data *priv;
213 struct cpufreq_frequency_table *freq_table;
214 struct scmi_device *sdev = cpufreq_get_driver_data();
215
216 cpu_dev = get_cpu_device(policy->cpu);
217 if (!cpu_dev) {
218 pr_err("failed to get cpu%d device\n", policy->cpu);
219 return -ENODEV;
220 }
221
222 domain = scmi_cpu_domain_id(cpu_dev);
223 if (domain < 0)
224 return domain;
225
226 priv = kzalloc(sizeof(*priv), GFP_KERNEL);
227 if (!priv)
228 return -ENOMEM;
229
230 if (!zalloc_cpumask_var(&priv->opp_shared_cpus, GFP_KERNEL)) {
231 ret = -ENOMEM;
232 goto out_free_priv;
233 }
234
235 /* Obtain CPUs that share SCMI performance controls */
236 ret = scmi_get_sharing_cpus(cpu_dev, domain, policy->cpus);
237 if (ret) {
238 dev_warn(cpu_dev, "failed to get sharing cpumask\n");
239 goto out_free_cpumask;
240 }
241
242 /*
243 * Obtain CPUs that share performance levels.
244 * The OPP 'sharing cpus' info may come from DT through an empty opp
245 * table and opp-shared.
246 */
247 ret = dev_pm_opp_of_get_sharing_cpus(cpu_dev, priv->opp_shared_cpus);
248 if (ret || cpumask_empty(priv->opp_shared_cpus)) {
249 /*
250 * Either opp-table is not set or no opp-shared was found.
251 * Use the CPU mask from SCMI to designate CPUs sharing an OPP
252 * table.
253 */
254 cpumask_copy(priv->opp_shared_cpus, policy->cpus);
255 }
256
257 /*
258 * A previous CPU may have marked OPPs as shared for a few CPUs, based on
259 * what OPP core provided. If the current CPU is part of those few, then
260 * there is no need to add OPPs again.
261 */
262 nr_opp = dev_pm_opp_get_opp_count(cpu_dev);
263 if (nr_opp <= 0) {
264 ret = perf_ops->device_opps_add(ph, cpu_dev, domain);
265 if (ret) {
266 dev_warn(cpu_dev, "failed to add opps to the device\n");
267 goto out_free_cpumask;
268 }
269
270 nr_opp = dev_pm_opp_get_opp_count(cpu_dev);
271 if (nr_opp <= 0) {
272 dev_err(cpu_dev, "%s: No OPPs for this device: %d\n",
273 __func__, nr_opp);
274
275 ret = -ENODEV;
276 goto out_free_opp;
277 }
278
279 ret = dev_pm_opp_set_sharing_cpus(cpu_dev, priv->opp_shared_cpus);
280 if (ret) {
281 dev_err(cpu_dev, "%s: failed to mark OPPs as shared: %d\n",
282 __func__, ret);
283
284 goto out_free_opp;
285 }
286
287 priv->nr_opp = nr_opp;
288 }
289
290 ret = dev_pm_opp_init_cpufreq_table(cpu_dev, &freq_table);
291 if (ret) {
292 dev_err(cpu_dev, "failed to init cpufreq table: %d\n", ret);
293 goto out_free_opp;
294 }
295
296 priv->cpu_dev = cpu_dev;
297 priv->domain_id = domain;
298
299 policy->driver_data = priv;
300 policy->freq_table = freq_table;
301
302 /* SCMI allows DVFS request for any domain from any CPU */
303 policy->dvfs_possible_from_any_cpu = true;
304
305 latency = perf_ops->transition_latency_get(ph, domain);
306 if (!latency)
307 latency = CPUFREQ_ETERNAL;
308
309 policy->cpuinfo.transition_latency = latency;
310
311 policy->fast_switch_possible =
312 perf_ops->fast_switch_possible(ph, domain);
313
314 policy->transition_delay_us =
315 scmi_get_rate_limit(domain, policy->fast_switch_possible);
316
317 if (policy_has_boost_freq(policy)) {
318 ret = cpufreq_enable_boost_support();
319 if (ret) {
320 dev_warn(cpu_dev, "failed to enable boost: %d\n", ret);
321 goto out_free_table;
322 } else {
323 scmi_cpufreq_hw_attr[1] = &cpufreq_freq_attr_scaling_boost_freqs;
324 scmi_cpufreq_driver.boost_enabled = true;
325 }
326 }
327
328 ret = freq_qos_add_request(&policy->constraints, &priv->limits_freq_req, FREQ_QOS_MAX,
329 FREQ_QOS_MAX_DEFAULT_VALUE);
330 if (ret < 0) {
331 dev_err(cpu_dev, "failed to add qos limits request: %d\n", ret);
332 goto out_free_table;
333 }
334
335 priv->limit_notify_nb.notifier_call = scmi_limit_notify_cb;
336 ret = sdev->handle->notify_ops->event_notifier_register(sdev->handle, SCMI_PROTOCOL_PERF,
337 SCMI_EVENT_PERFORMANCE_LIMITS_CHANGED,
338 &priv->domain_id,
339 &priv->limit_notify_nb);
340 if (ret)
341 dev_warn(&sdev->dev,
342 "failed to register for limits change notifier for domain %d\n",
343 priv->domain_id);
344
345 return 0;
346
347 out_free_table:
348 dev_pm_opp_free_cpufreq_table(cpu_dev, &freq_table);
349 out_free_opp:
350 dev_pm_opp_remove_all_dynamic(cpu_dev);
351
352 out_free_cpumask:
353 free_cpumask_var(priv->opp_shared_cpus);
354
355 out_free_priv:
356 kfree(priv);
357
358 return ret;
359 }
360
scmi_cpufreq_exit(struct cpufreq_policy * policy)361 static void scmi_cpufreq_exit(struct cpufreq_policy *policy)
362 {
363 struct scmi_data *priv = policy->driver_data;
364 struct scmi_device *sdev = cpufreq_get_driver_data();
365
366 sdev->handle->notify_ops->event_notifier_unregister(sdev->handle, SCMI_PROTOCOL_PERF,
367 SCMI_EVENT_PERFORMANCE_LIMITS_CHANGED,
368 &priv->domain_id,
369 &priv->limit_notify_nb);
370 freq_qos_remove_request(&priv->limits_freq_req);
371 dev_pm_opp_free_cpufreq_table(priv->cpu_dev, &policy->freq_table);
372 dev_pm_opp_remove_all_dynamic(priv->cpu_dev);
373 free_cpumask_var(priv->opp_shared_cpus);
374 kfree(priv);
375 }
376
scmi_cpufreq_register_em(struct cpufreq_policy * policy)377 static void scmi_cpufreq_register_em(struct cpufreq_policy *policy)
378 {
379 struct em_data_callback em_cb = EM_DATA_CB(scmi_get_cpu_power);
380 enum scmi_power_scale power_scale = perf_ops->power_scale_get(ph);
381 struct scmi_data *priv = policy->driver_data;
382 bool em_power_scale = false;
383
384 /*
385 * This callback will be called for each policy, but we don't need to
386 * register with EM every time. Despite not being part of the same
387 * policy, some CPUs may still share their perf-domains, and a CPU from
388 * another policy may already have registered with EM on behalf of CPUs
389 * of this policy.
390 */
391 if (!priv->nr_opp)
392 return;
393
394 if (power_scale == SCMI_POWER_MILLIWATTS
395 || power_scale == SCMI_POWER_MICROWATTS)
396 em_power_scale = true;
397
398 em_dev_register_perf_domain(get_cpu_device(policy->cpu), priv->nr_opp,
399 &em_cb, priv->opp_shared_cpus,
400 em_power_scale);
401 }
402
403 static struct cpufreq_driver scmi_cpufreq_driver = {
404 .name = "scmi",
405 .flags = CPUFREQ_HAVE_GOVERNOR_PER_POLICY |
406 CPUFREQ_NEED_INITIAL_FREQ_CHECK |
407 CPUFREQ_IS_COOLING_DEV,
408 .verify = cpufreq_generic_frequency_table_verify,
409 .attr = scmi_cpufreq_hw_attr,
410 .target_index = scmi_cpufreq_set_target,
411 .fast_switch = scmi_cpufreq_fast_switch,
412 .get = scmi_cpufreq_get_rate,
413 .init = scmi_cpufreq_init,
414 .exit = scmi_cpufreq_exit,
415 .register_em = scmi_cpufreq_register_em,
416 };
417
scmi_dev_used_by_cpus(struct device * scmi_dev)418 static bool scmi_dev_used_by_cpus(struct device *scmi_dev)
419 {
420 struct device_node *scmi_np = dev_of_node(scmi_dev);
421 struct device_node *cpu_np, *np;
422 struct device *cpu_dev;
423 int cpu, idx;
424
425 if (!scmi_np)
426 return false;
427
428 for_each_possible_cpu(cpu) {
429 cpu_dev = get_cpu_device(cpu);
430 if (!cpu_dev)
431 continue;
432
433 cpu_np = dev_of_node(cpu_dev);
434
435 np = of_parse_phandle(cpu_np, "clocks", 0);
436 of_node_put(np);
437
438 if (np == scmi_np)
439 return true;
440
441 idx = of_property_match_string(cpu_np, "power-domain-names", "perf");
442 np = of_parse_phandle(cpu_np, "power-domains", idx);
443 of_node_put(np);
444
445 if (np == scmi_np)
446 return true;
447 }
448
449 return false;
450 }
451
scmi_cpufreq_probe(struct scmi_device * sdev)452 static int scmi_cpufreq_probe(struct scmi_device *sdev)
453 {
454 int ret;
455 struct device *dev = &sdev->dev;
456 const struct scmi_handle *handle;
457
458 handle = sdev->handle;
459
460 if (!handle || !scmi_dev_used_by_cpus(dev))
461 return -ENODEV;
462
463 scmi_cpufreq_driver.driver_data = sdev;
464
465 perf_ops = handle->devm_protocol_get(sdev, SCMI_PROTOCOL_PERF, &ph);
466 if (IS_ERR(perf_ops))
467 return PTR_ERR(perf_ops);
468
469 #ifdef CONFIG_COMMON_CLK
470 /* dummy clock provider as needed by OPP if clocks property is used */
471 if (of_property_present(dev->of_node, "#clock-cells")) {
472 ret = devm_of_clk_add_hw_provider(dev, of_clk_hw_simple_get, NULL);
473 if (ret)
474 return dev_err_probe(dev, ret, "%s: registering clock provider failed\n", __func__);
475 }
476 #endif
477
478 ret = cpufreq_register_driver(&scmi_cpufreq_driver);
479 if (ret) {
480 dev_err(dev, "%s: registering cpufreq failed, err: %d\n",
481 __func__, ret);
482 }
483
484 return ret;
485 }
486
scmi_cpufreq_remove(struct scmi_device * sdev)487 static void scmi_cpufreq_remove(struct scmi_device *sdev)
488 {
489 cpufreq_unregister_driver(&scmi_cpufreq_driver);
490 }
491
492 static const struct scmi_device_id scmi_id_table[] = {
493 { SCMI_PROTOCOL_PERF, "cpufreq" },
494 { },
495 };
496 MODULE_DEVICE_TABLE(scmi, scmi_id_table);
497
498 static struct scmi_driver scmi_cpufreq_drv = {
499 .name = "scmi-cpufreq",
500 .probe = scmi_cpufreq_probe,
501 .remove = scmi_cpufreq_remove,
502 .id_table = scmi_id_table,
503 };
504 module_scmi_driver(scmi_cpufreq_drv);
505
506 MODULE_AUTHOR("Sudeep Holla <sudeep.holla@arm.com>");
507 MODULE_DESCRIPTION("ARM SCMI CPUFreq interface driver");
508 MODULE_LICENSE("GPL v2");
509