drivers/cpufreq/scmi-cpufreq.c

0001 // SPDX-License-Identifier: GPL-2.0
0002 /*
0003  * System Control and Power Interface (SCMI) based CPUFreq Interface driver
0004  *
0005  * Copyright (C) 2018-2021 ARM Ltd.
0006  * Sudeep Holla <sudeep.holla@arm.com>
0007  */
0008
0009 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
0010
0011 #include <linux/clk-provider.h>
0012 #include <linux/cpu.h>
0013 #include <linux/cpufreq.h>
0014 #include <linux/cpumask.h>
0015 #include <linux/energy_model.h>
0016 #include <linux/export.h>
0017 #include <linux/module.h>
0018 #include <linux/pm_opp.h>
0019 #include <linux/slab.h>
0020 #include <linux/scmi_protocol.h>
0021 #include <linux/types.h>
0022 #include <linux/units.h>
0023
0024 struct scmi_data {
0025     int domain_id;
0026     int nr_opp;
0027     struct device *cpu_dev;
0028     cpumask_var_t opp_shared_cpus;
0029 };
0030
0031 static struct scmi_protocol_handle *ph;
0032 static const struct scmi_perf_proto_ops *perf_ops;
0033
0034 static unsigned int scmi_cpufreq_get_rate(unsigned int cpu)
0035 {
0036     struct cpufreq_policy *policy = cpufreq_cpu_get_raw(cpu);
0037     struct scmi_data *priv = policy->driver_data;
0038     unsigned long rate;
0039     int ret;
0040
0041     ret = perf_ops->freq_get(ph, priv->domain_id, &rate, false);
0042     if (ret)
0043         return 0;
0044     return rate / 1000;
0045 }
0046
0047 /*
0048  * perf_ops->freq_set is not a synchronous, the actual OPP change will
0049  * happen asynchronously and can get notified if the events are
0050  * subscribed for by the SCMI firmware
0051  */
0052 static int
0053 scmi_cpufreq_set_target(struct cpufreq_policy *policy, unsigned int index)
0054 {
0055     struct scmi_data *priv = policy->driver_data;
0056     u64 freq = policy->freq_table[index].frequency;
0057
0058     return perf_ops->freq_set(ph, priv->domain_id, freq * 1000, false);
0059 }
0060
0061 static unsigned int scmi_cpufreq_fast_switch(struct cpufreq_policy *policy,
0062                          unsigned int target_freq)
0063 {
0064     struct scmi_data *priv = policy->driver_data;
0065
0066     if (!perf_ops->freq_set(ph, priv->domain_id,
0067                 target_freq * 1000, true))
0068         return target_freq;
0069
0070     return 0;
0071 }
0072
0073 static int
0074 scmi_get_sharing_cpus(struct device *cpu_dev, struct cpumask *cpumask)
0075 {
0076     int cpu, domain, tdomain;
0077     struct device *tcpu_dev;
0078
0079     domain = perf_ops->device_domain_id(cpu_dev);
0080     if (domain < 0)
0081         return domain;
0082
0083     for_each_possible_cpu(cpu) {
0084         if (cpu == cpu_dev->id)
0085             continue;
0086
0087         tcpu_dev = get_cpu_device(cpu);
0088         if (!tcpu_dev)
0089             continue;
0090
0091         tdomain = perf_ops->device_domain_id(tcpu_dev);
0092         if (tdomain == domain)
0093             cpumask_set_cpu(cpu, cpumask);
0094     }
0095
0096     return 0;
0097 }
0098
0099 static int __maybe_unused
0100 scmi_get_cpu_power(struct device *cpu_dev, unsigned long *power,
0101            unsigned long *KHz)
0102 {
0103     enum scmi_power_scale power_scale = perf_ops->power_scale_get(ph);
0104     unsigned long Hz;
0105     int ret, domain;
0106
0107     domain = perf_ops->device_domain_id(cpu_dev);
0108     if (domain < 0)
0109         return domain;
0110
0111     /* Get the power cost of the performance domain. */
0112     Hz = *KHz * 1000;
0113     ret = perf_ops->est_power_get(ph, domain, &Hz, power);
0114     if (ret)
0115         return ret;
0116
0117     /* Convert the power to uW if it is mW (ignore bogoW) */
0118     if (power_scale == SCMI_POWER_MILLIWATTS)
0119         *power *= MICROWATT_PER_MILLIWATT;
0120
0121     /* The EM framework specifies the frequency in KHz. */
0122     *KHz = Hz / 1000;
0123
0124     return 0;
0125 }
0126
0127 static int scmi_cpufreq_init(struct cpufreq_policy *policy)
0128 {
0129     int ret, nr_opp;
0130     unsigned int latency;
0131     struct device *cpu_dev;
0132     struct scmi_data *priv;
0133     struct cpufreq_frequency_table *freq_table;
0134
0135     cpu_dev = get_cpu_device(policy->cpu);
0136     if (!cpu_dev) {
0137         pr_err("failed to get cpu%d device\n", policy->cpu);
0138         return -ENODEV;
0139     }
0140
0141     priv = kzalloc(sizeof(*priv), GFP_KERNEL);
0142     if (!priv)
0143         return -ENOMEM;
0144
0145     if (!zalloc_cpumask_var(&priv->opp_shared_cpus, GFP_KERNEL)) {
0146         ret = -ENOMEM;
0147         goto out_free_priv;
0148     }
0149
0150     /* Obtain CPUs that share SCMI performance controls */
0151     ret = scmi_get_sharing_cpus(cpu_dev, policy->cpus);
0152     if (ret) {
0153         dev_warn(cpu_dev, "failed to get sharing cpumask\n");
0154         goto out_free_cpumask;
0155     }
0156
0157     /*
0158      * Obtain CPUs that share performance levels.
0159      * The OPP 'sharing cpus' info may come from DT through an empty opp
0160      * table and opp-shared.
0161      */
0162     ret = dev_pm_opp_of_get_sharing_cpus(cpu_dev, priv->opp_shared_cpus);
0163     if (ret || cpumask_empty(priv->opp_shared_cpus)) {
0164         /*
0165          * Either opp-table is not set or no opp-shared was found.
0166          * Use the CPU mask from SCMI to designate CPUs sharing an OPP
0167          * table.
0168          */
0169         cpumask_copy(priv->opp_shared_cpus, policy->cpus);
0170     }
0171
0172      /*
0173       * A previous CPU may have marked OPPs as shared for a few CPUs, based on
0174       * what OPP core provided. If the current CPU is part of those few, then
0175       * there is no need to add OPPs again.
0176       */
0177     nr_opp = dev_pm_opp_get_opp_count(cpu_dev);
0178     if (nr_opp <= 0) {
0179         ret = perf_ops->device_opps_add(ph, cpu_dev);
0180         if (ret) {
0181             dev_warn(cpu_dev, "failed to add opps to the device\n");
0182             goto out_free_cpumask;
0183         }
0184
0185         nr_opp = dev_pm_opp_get_opp_count(cpu_dev);
0186         if (nr_opp <= 0) {
0187             dev_err(cpu_dev, "%s: No OPPs for this device: %d\n",
0188                 __func__, nr_opp);
0189
0190             ret = -ENODEV;
0191             goto out_free_opp;
0192         }
0193
0194         ret = dev_pm_opp_set_sharing_cpus(cpu_dev, priv->opp_shared_cpus);
0195         if (ret) {
0196             dev_err(cpu_dev, "%s: failed to mark OPPs as shared: %d\n",
0197                 __func__, ret);
0198
0199             goto out_free_opp;
0200         }
0201
0202         priv->nr_opp = nr_opp;
0203     }
0204
0205     ret = dev_pm_opp_init_cpufreq_table(cpu_dev, &freq_table);
0206     if (ret) {
0207         dev_err(cpu_dev, "failed to init cpufreq table: %d\n", ret);
0208         goto out_free_opp;
0209     }
0210
0211     priv->cpu_dev = cpu_dev;
0212     priv->domain_id = perf_ops->device_domain_id(cpu_dev);
0213
0214     policy->driver_data = priv;
0215     policy->freq_table = freq_table;
0216
0217     /* SCMI allows DVFS request for any domain from any CPU */
0218     policy->dvfs_possible_from_any_cpu = true;
0219
0220     latency = perf_ops->transition_latency_get(ph, cpu_dev);
0221     if (!latency)
0222         latency = CPUFREQ_ETERNAL;
0223
0224     policy->cpuinfo.transition_latency = latency;
0225
0226     policy->fast_switch_possible =
0227         perf_ops->fast_switch_possible(ph, cpu_dev);
0228
0229     return 0;
0230
0231 out_free_opp:
0232     dev_pm_opp_remove_all_dynamic(cpu_dev);
0233
0234 out_free_cpumask:
0235     free_cpumask_var(priv->opp_shared_cpus);
0236
0237 out_free_priv:
0238     kfree(priv);
0239
0240     return ret;
0241 }
0242
0243 static int scmi_cpufreq_exit(struct cpufreq_policy *policy)
0244 {
0245     struct scmi_data *priv = policy->driver_data;
0246
0247     dev_pm_opp_free_cpufreq_table(priv->cpu_dev, &policy->freq_table);
0248     dev_pm_opp_remove_all_dynamic(priv->cpu_dev);
0249     free_cpumask_var(priv->opp_shared_cpus);
0250     kfree(priv);
0251
0252     return 0;
0253 }
0254
0255 static void scmi_cpufreq_register_em(struct cpufreq_policy *policy)
0256 {
0257     struct em_data_callback em_cb = EM_DATA_CB(scmi_get_cpu_power);
0258     enum scmi_power_scale power_scale = perf_ops->power_scale_get(ph);
0259     struct scmi_data *priv = policy->driver_data;
0260     bool em_power_scale = false;
0261
0262     /*
0263      * This callback will be called for each policy, but we don't need to
0264      * register with EM every time. Despite not being part of the same
0265      * policy, some CPUs may still share their perf-domains, and a CPU from
0266      * another policy may already have registered with EM on behalf of CPUs
0267      * of this policy.
0268      */
0269     if (!priv->nr_opp)
0270         return;
0271
0272     if (power_scale == SCMI_POWER_MILLIWATTS
0273         || power_scale == SCMI_POWER_MICROWATTS)
0274         em_power_scale = true;
0275
0276     em_dev_register_perf_domain(get_cpu_device(policy->cpu), priv->nr_opp,
0277                     &em_cb, priv->opp_shared_cpus,
0278                     em_power_scale);
0279 }
0280
0281 static struct cpufreq_driver scmi_cpufreq_driver = {
0282     .name   = "scmi",
0283     .flags  = CPUFREQ_HAVE_GOVERNOR_PER_POLICY |
0284           CPUFREQ_NEED_INITIAL_FREQ_CHECK |
0285           CPUFREQ_IS_COOLING_DEV,
0286     .verify = cpufreq_generic_frequency_table_verify,
0287     .attr   = cpufreq_generic_attr,
0288     .target_index   = scmi_cpufreq_set_target,
0289     .fast_switch    = scmi_cpufreq_fast_switch,
0290     .get    = scmi_cpufreq_get_rate,
0291     .init   = scmi_cpufreq_init,
0292     .exit   = scmi_cpufreq_exit,
0293     .register_em    = scmi_cpufreq_register_em,
0294 };
0295
0296 static int scmi_cpufreq_probe(struct scmi_device *sdev)
0297 {
0298     int ret;
0299     struct device *dev = &sdev->dev;
0300     const struct scmi_handle *handle;
0301
0302     handle = sdev->handle;
0303
0304     if (!handle)
0305         return -ENODEV;
0306
0307     perf_ops = handle->devm_protocol_get(sdev, SCMI_PROTOCOL_PERF, &ph);
0308     if (IS_ERR(perf_ops))
0309         return PTR_ERR(perf_ops);
0310
0311 #ifdef CONFIG_COMMON_CLK
0312     /* dummy clock provider as needed by OPP if clocks property is used */
0313     if (of_find_property(dev->of_node, "#clock-cells", NULL))
0314         devm_of_clk_add_hw_provider(dev, of_clk_hw_simple_get, NULL);
0315 #endif
0316
0317     ret = cpufreq_register_driver(&scmi_cpufreq_driver);
0318     if (ret) {
0319         dev_err(dev, "%s: registering cpufreq failed, err: %d\n",
0320             __func__, ret);
0321     }
0322
0323     return ret;
0324 }
0325
0326 static void scmi_cpufreq_remove(struct scmi_device *sdev)
0327 {
0328     cpufreq_unregister_driver(&scmi_cpufreq_driver);
0329 }
0330
0331 static const struct scmi_device_id scmi_id_table[] = {
0332     { SCMI_PROTOCOL_PERF, "cpufreq" },
0333     { },
0334 };
0335 MODULE_DEVICE_TABLE(scmi, scmi_id_table);
0336
0337 static struct scmi_driver scmi_cpufreq_drv = {
0338     .name       = "scmi-cpufreq",
0339     .probe      = scmi_cpufreq_probe,
0340     .remove     = scmi_cpufreq_remove,
0341     .id_table   = scmi_id_table,
0342 };
0343 module_scmi_driver(scmi_cpufreq_drv);
0344
0345 MODULE_AUTHOR("Sudeep Holla <sudeep.holla@arm.com>");
0346 MODULE_DESCRIPTION("ARM SCMI CPUFreq interface driver");
0347 MODULE_LICENSE("GPL v2");