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/of.h> 19 #include <linux/pm_opp.h> 20 #include <linux/pm_qos.h> 21 #include <linux/slab.h> 22 #include <linux/scmi_protocol.h> 23 #include <linux/types.h> 24 #include <linux/units.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 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 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 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 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 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_present_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 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 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 int scmi_limit_notify_cb(struct notifier_block *nb, unsigned long event, void *data) 182 { 183 struct scmi_data *priv = container_of(nb, struct scmi_data, limit_notify_nb); 184 struct scmi_perf_limits_report *limit_notify = data; 185 unsigned int limit_freq_khz; 186 int ret; 187 188 limit_freq_khz = limit_notify->range_max_freq / HZ_PER_KHZ; 189 190 ret = freq_qos_update_request(&priv->limits_freq_req, limit_freq_khz); 191 if (ret < 0) 192 pr_warn("failed to update freq constraint: %d\n", ret); 193 194 return NOTIFY_OK; 195 } 196 197 static int scmi_cpufreq_init(struct cpufreq_policy *policy) 198 { 199 int ret, nr_opp, domain; 200 unsigned int latency; 201 struct device *cpu_dev; 202 struct scmi_data *priv; 203 struct cpufreq_frequency_table *freq_table; 204 struct scmi_device *sdev = cpufreq_get_driver_data(); 205 206 cpu_dev = get_cpu_device(policy->cpu); 207 if (!cpu_dev) { 208 pr_err("failed to get cpu%d device\n", policy->cpu); 209 return -ENODEV; 210 } 211 212 domain = scmi_cpu_domain_id(cpu_dev); 213 if (domain < 0) 214 return domain; 215 216 priv = kzalloc(sizeof(*priv), GFP_KERNEL); 217 if (!priv) 218 return -ENOMEM; 219 220 if (!zalloc_cpumask_var(&priv->opp_shared_cpus, GFP_KERNEL)) { 221 ret = -ENOMEM; 222 goto out_free_priv; 223 } 224 225 /* Obtain CPUs that share SCMI performance controls */ 226 ret = scmi_get_sharing_cpus(cpu_dev, domain, policy->cpus); 227 if (ret) { 228 dev_warn(cpu_dev, "failed to get sharing cpumask\n"); 229 goto out_free_cpumask; 230 } 231 232 /* 233 * Obtain CPUs that share performance levels. 234 * The OPP 'sharing cpus' info may come from DT through an empty opp 235 * table and opp-shared. 236 */ 237 ret = dev_pm_opp_of_get_sharing_cpus(cpu_dev, priv->opp_shared_cpus); 238 if (ret || cpumask_empty(priv->opp_shared_cpus)) { 239 /* 240 * Either opp-table is not set or no opp-shared was found. 241 * Use the CPU mask from SCMI to designate CPUs sharing an OPP 242 * table. 243 */ 244 cpumask_copy(priv->opp_shared_cpus, policy->cpus); 245 } 246 247 /* 248 * A previous CPU may have marked OPPs as shared for a few CPUs, based on 249 * what OPP core provided. If the current CPU is part of those few, then 250 * there is no need to add OPPs again. 251 */ 252 nr_opp = dev_pm_opp_get_opp_count(cpu_dev); 253 if (nr_opp <= 0) { 254 ret = perf_ops->device_opps_add(ph, cpu_dev, domain); 255 if (ret) { 256 dev_warn(cpu_dev, "failed to add opps to the device\n"); 257 goto out_free_cpumask; 258 } 259 260 nr_opp = dev_pm_opp_get_opp_count(cpu_dev); 261 if (nr_opp <= 0) { 262 dev_err(cpu_dev, "%s: No OPPs for this device: %d\n", 263 __func__, nr_opp); 264 265 ret = -ENODEV; 266 goto out_free_opp; 267 } 268 269 ret = dev_pm_opp_set_sharing_cpus(cpu_dev, priv->opp_shared_cpus); 270 if (ret) { 271 dev_err(cpu_dev, "%s: failed to mark OPPs as shared: %d\n", 272 __func__, ret); 273 274 goto out_free_opp; 275 } 276 277 priv->nr_opp = nr_opp; 278 } 279 280 ret = dev_pm_opp_init_cpufreq_table(cpu_dev, &freq_table); 281 if (ret) { 282 dev_err(cpu_dev, "failed to init cpufreq table: %d\n", ret); 283 goto out_free_opp; 284 } 285 286 priv->cpu_dev = cpu_dev; 287 priv->domain_id = domain; 288 289 policy->driver_data = priv; 290 policy->freq_table = freq_table; 291 292 /* SCMI allows DVFS request for any domain from any CPU */ 293 policy->dvfs_possible_from_any_cpu = true; 294 295 latency = perf_ops->transition_latency_get(ph, domain); 296 if (!latency) 297 latency = CPUFREQ_ETERNAL; 298 299 policy->cpuinfo.transition_latency = latency; 300 301 policy->fast_switch_possible = 302 perf_ops->fast_switch_possible(ph, domain); 303 304 policy->transition_delay_us = 305 scmi_get_rate_limit(domain, policy->fast_switch_possible); 306 307 ret = freq_qos_add_request(&policy->constraints, &priv->limits_freq_req, FREQ_QOS_MAX, 308 FREQ_QOS_MAX_DEFAULT_VALUE); 309 if (ret < 0) { 310 dev_err(cpu_dev, "failed to add qos limits request: %d\n", ret); 311 goto out_free_table; 312 } 313 314 priv->limit_notify_nb.notifier_call = scmi_limit_notify_cb; 315 ret = sdev->handle->notify_ops->event_notifier_register(sdev->handle, SCMI_PROTOCOL_PERF, 316 SCMI_EVENT_PERFORMANCE_LIMITS_CHANGED, 317 &priv->domain_id, 318 &priv->limit_notify_nb); 319 if (ret) 320 dev_warn(&sdev->dev, 321 "failed to register for limits change notifier for domain %d\n", 322 priv->domain_id); 323 324 return 0; 325 326 out_free_table: 327 dev_pm_opp_free_cpufreq_table(cpu_dev, &freq_table); 328 out_free_opp: 329 dev_pm_opp_remove_all_dynamic(cpu_dev); 330 331 out_free_cpumask: 332 free_cpumask_var(priv->opp_shared_cpus); 333 334 out_free_priv: 335 kfree(priv); 336 337 return ret; 338 } 339 340 static void scmi_cpufreq_exit(struct cpufreq_policy *policy) 341 { 342 struct scmi_data *priv = policy->driver_data; 343 struct scmi_device *sdev = cpufreq_get_driver_data(); 344 345 sdev->handle->notify_ops->event_notifier_unregister(sdev->handle, SCMI_PROTOCOL_PERF, 346 SCMI_EVENT_PERFORMANCE_LIMITS_CHANGED, 347 &priv->domain_id, 348 &priv->limit_notify_nb); 349 freq_qos_remove_request(&priv->limits_freq_req); 350 dev_pm_opp_free_cpufreq_table(priv->cpu_dev, &policy->freq_table); 351 dev_pm_opp_remove_all_dynamic(priv->cpu_dev); 352 free_cpumask_var(priv->opp_shared_cpus); 353 kfree(priv); 354 } 355 356 static void scmi_cpufreq_register_em(struct cpufreq_policy *policy) 357 { 358 struct em_data_callback em_cb = EM_DATA_CB(scmi_get_cpu_power); 359 enum scmi_power_scale power_scale = perf_ops->power_scale_get(ph); 360 struct scmi_data *priv = policy->driver_data; 361 bool em_power_scale = false; 362 363 /* 364 * This callback will be called for each policy, but we don't need to 365 * register with EM every time. Despite not being part of the same 366 * policy, some CPUs may still share their perf-domains, and a CPU from 367 * another policy may already have registered with EM on behalf of CPUs 368 * of this policy. 369 */ 370 if (!priv->nr_opp) 371 return; 372 373 if (power_scale == SCMI_POWER_MILLIWATTS 374 || power_scale == SCMI_POWER_MICROWATTS) 375 em_power_scale = true; 376 377 em_dev_register_perf_domain(get_cpu_device(policy->cpu), priv->nr_opp, 378 &em_cb, priv->opp_shared_cpus, 379 em_power_scale); 380 } 381 382 static struct cpufreq_driver scmi_cpufreq_driver = { 383 .name = "scmi", 384 .flags = CPUFREQ_HAVE_GOVERNOR_PER_POLICY | 385 CPUFREQ_NEED_INITIAL_FREQ_CHECK | 386 CPUFREQ_IS_COOLING_DEV, 387 .verify = cpufreq_generic_frequency_table_verify, 388 .target_index = scmi_cpufreq_set_target, 389 .fast_switch = scmi_cpufreq_fast_switch, 390 .get = scmi_cpufreq_get_rate, 391 .init = scmi_cpufreq_init, 392 .exit = scmi_cpufreq_exit, 393 .register_em = scmi_cpufreq_register_em, 394 .set_boost = cpufreq_boost_set_sw, 395 }; 396 397 static bool scmi_dev_used_by_cpus(struct device *scmi_dev) 398 { 399 struct device_node *scmi_np = dev_of_node(scmi_dev); 400 struct device_node *cpu_np, *np; 401 struct device *cpu_dev; 402 int cpu, idx; 403 404 if (!scmi_np) 405 return false; 406 407 for_each_possible_cpu(cpu) { 408 cpu_dev = get_cpu_device(cpu); 409 if (!cpu_dev) 410 continue; 411 412 cpu_np = dev_of_node(cpu_dev); 413 414 np = of_parse_phandle(cpu_np, "clocks", 0); 415 of_node_put(np); 416 417 if (np == scmi_np) 418 return true; 419 420 idx = of_property_match_string(cpu_np, "power-domain-names", "perf"); 421 np = of_parse_phandle(cpu_np, "power-domains", idx); 422 of_node_put(np); 423 424 if (np == scmi_np) 425 return true; 426 } 427 428 /* 429 * Older Broadcom STB chips had a "clocks" property for CPU node(s) 430 * that did not match the SCMI performance protocol node, if we got 431 * there, it means we had such an older Device Tree, therefore return 432 * true to preserve backwards compatibility. 433 */ 434 if (of_machine_is_compatible("brcm,brcmstb")) 435 return true; 436 437 return false; 438 } 439 440 static int scmi_cpufreq_probe(struct scmi_device *sdev) 441 { 442 int ret; 443 struct device *dev = &sdev->dev; 444 const struct scmi_handle *handle; 445 446 handle = sdev->handle; 447 448 if (!handle || !scmi_dev_used_by_cpus(dev)) 449 return -ENODEV; 450 451 scmi_cpufreq_driver.driver_data = sdev; 452 453 perf_ops = handle->devm_protocol_get(sdev, SCMI_PROTOCOL_PERF, &ph); 454 if (IS_ERR(perf_ops)) 455 return PTR_ERR(perf_ops); 456 457 #ifdef CONFIG_COMMON_CLK 458 /* dummy clock provider as needed by OPP if clocks property is used */ 459 if (of_property_present(dev->of_node, "#clock-cells")) { 460 ret = devm_of_clk_add_hw_provider(dev, of_clk_hw_simple_get, NULL); 461 if (ret) 462 return dev_err_probe(dev, ret, "%s: registering clock provider failed\n", __func__); 463 } 464 #endif 465 466 ret = cpufreq_register_driver(&scmi_cpufreq_driver); 467 if (ret) { 468 dev_err(dev, "%s: registering cpufreq failed, err: %d\n", 469 __func__, ret); 470 } 471 472 return ret; 473 } 474 475 static void scmi_cpufreq_remove(struct scmi_device *sdev) 476 { 477 cpufreq_unregister_driver(&scmi_cpufreq_driver); 478 } 479 480 static const struct scmi_device_id scmi_id_table[] = { 481 { SCMI_PROTOCOL_PERF, "cpufreq" }, 482 { }, 483 }; 484 MODULE_DEVICE_TABLE(scmi, scmi_id_table); 485 486 static struct scmi_driver scmi_cpufreq_drv = { 487 .name = "scmi-cpufreq", 488 .probe = scmi_cpufreq_probe, 489 .remove = scmi_cpufreq_remove, 490 .id_table = scmi_id_table, 491 }; 492 module_scmi_driver(scmi_cpufreq_drv); 493 494 MODULE_AUTHOR("Sudeep Holla <sudeep.holla@arm.com>"); 495 MODULE_DESCRIPTION("ARM SCMI CPUFreq interface driver"); 496 MODULE_LICENSE("GPL v2"); 497