1 // SPDX-License-Identifier: GPL-2.0 2 /* 3 * Energy Model of devices 4 * 5 * Copyright (c) 2018-2021, Arm ltd. 6 * Written by: Quentin Perret, Arm ltd. 7 * Improvements provided by: Lukasz Luba, Arm ltd. 8 */ 9 10 #define pr_fmt(fmt) "energy_model: " fmt 11 12 #include <linux/cpu.h> 13 #include <linux/cpufreq.h> 14 #include <linux/cpumask.h> 15 #include <linux/debugfs.h> 16 #include <linux/energy_model.h> 17 #include <linux/sched/topology.h> 18 #include <linux/slab.h> 19 20 /* 21 * Mutex serializing the registrations of performance domains and letting 22 * callbacks defined by drivers sleep. 23 */ 24 static DEFINE_MUTEX(em_pd_mutex); 25 26 static bool _is_cpu_device(struct device *dev) 27 { 28 return (dev->bus == &cpu_subsys); 29 } 30 31 #ifdef CONFIG_DEBUG_FS 32 static struct dentry *rootdir; 33 34 static void em_debug_create_ps(struct em_perf_state *ps, struct dentry *pd) 35 { 36 struct dentry *d; 37 char name[24]; 38 39 snprintf(name, sizeof(name), "ps:%lu", ps->frequency); 40 41 /* Create per-ps directory */ 42 d = debugfs_create_dir(name, pd); 43 debugfs_create_ulong("frequency", 0444, d, &ps->frequency); 44 debugfs_create_ulong("power", 0444, d, &ps->power); 45 debugfs_create_ulong("cost", 0444, d, &ps->cost); 46 debugfs_create_ulong("inefficient", 0444, d, &ps->flags); 47 } 48 49 static int em_debug_cpus_show(struct seq_file *s, void *unused) 50 { 51 seq_printf(s, "%*pbl\n", cpumask_pr_args(to_cpumask(s->private))); 52 53 return 0; 54 } 55 DEFINE_SHOW_ATTRIBUTE(em_debug_cpus); 56 57 static int em_debug_flags_show(struct seq_file *s, void *unused) 58 { 59 struct em_perf_domain *pd = s->private; 60 61 seq_printf(s, "%#lx\n", pd->flags); 62 63 return 0; 64 } 65 DEFINE_SHOW_ATTRIBUTE(em_debug_flags); 66 67 static void em_debug_create_pd(struct device *dev) 68 { 69 struct dentry *d; 70 int i; 71 72 /* Create the directory of the performance domain */ 73 d = debugfs_create_dir(dev_name(dev), rootdir); 74 75 if (_is_cpu_device(dev)) 76 debugfs_create_file("cpus", 0444, d, dev->em_pd->cpus, 77 &em_debug_cpus_fops); 78 79 debugfs_create_file("flags", 0444, d, dev->em_pd, 80 &em_debug_flags_fops); 81 82 /* Create a sub-directory for each performance state */ 83 for (i = 0; i < dev->em_pd->nr_perf_states; i++) 84 em_debug_create_ps(&dev->em_pd->table[i], d); 85 86 } 87 88 static void em_debug_remove_pd(struct device *dev) 89 { 90 struct dentry *debug_dir; 91 92 debug_dir = debugfs_lookup(dev_name(dev), rootdir); 93 debugfs_remove_recursive(debug_dir); 94 } 95 96 static int __init em_debug_init(void) 97 { 98 /* Create /sys/kernel/debug/energy_model directory */ 99 rootdir = debugfs_create_dir("energy_model", NULL); 100 101 return 0; 102 } 103 fs_initcall(em_debug_init); 104 #else /* CONFIG_DEBUG_FS */ 105 static void em_debug_create_pd(struct device *dev) {} 106 static void em_debug_remove_pd(struct device *dev) {} 107 #endif 108 109 static int em_create_perf_table(struct device *dev, struct em_perf_domain *pd, 110 int nr_states, struct em_data_callback *cb, 111 unsigned long flags) 112 { 113 unsigned long power, freq, prev_freq = 0, prev_cost = ULONG_MAX; 114 struct em_perf_state *table; 115 int i, ret; 116 u64 fmax; 117 118 table = kcalloc(nr_states, sizeof(*table), GFP_KERNEL); 119 if (!table) 120 return -ENOMEM; 121 122 /* Build the list of performance states for this performance domain */ 123 for (i = 0, freq = 0; i < nr_states; i++, freq++) { 124 /* 125 * active_power() is a driver callback which ceils 'freq' to 126 * lowest performance state of 'dev' above 'freq' and updates 127 * 'power' and 'freq' accordingly. 128 */ 129 ret = cb->active_power(dev, &power, &freq); 130 if (ret) { 131 dev_err(dev, "EM: invalid perf. state: %d\n", 132 ret); 133 goto free_ps_table; 134 } 135 136 /* 137 * We expect the driver callback to increase the frequency for 138 * higher performance states. 139 */ 140 if (freq <= prev_freq) { 141 dev_err(dev, "EM: non-increasing freq: %lu\n", 142 freq); 143 goto free_ps_table; 144 } 145 146 /* 147 * The power returned by active_state() is expected to be 148 * positive and to fit into 16 bits. 149 */ 150 if (!power || power > EM_MAX_POWER) { 151 dev_err(dev, "EM: invalid power: %lu\n", 152 power); 153 goto free_ps_table; 154 } 155 156 table[i].power = power; 157 table[i].frequency = prev_freq = freq; 158 } 159 160 /* Compute the cost of each performance state. */ 161 fmax = (u64) table[nr_states - 1].frequency; 162 for (i = nr_states - 1; i >= 0; i--) { 163 unsigned long power_res, cost; 164 165 if (flags & EM_PERF_DOMAIN_ARTIFICIAL) { 166 ret = cb->get_cost(dev, table[i].frequency, &cost); 167 if (ret || !cost || cost > EM_MAX_POWER) { 168 dev_err(dev, "EM: invalid cost %lu %d\n", 169 cost, ret); 170 goto free_ps_table; 171 } 172 } else { 173 power_res = em_scale_power(table[i].power); 174 cost = div64_u64(fmax * power_res, table[i].frequency); 175 } 176 177 table[i].cost = cost; 178 179 if (table[i].cost >= prev_cost) { 180 table[i].flags = EM_PERF_STATE_INEFFICIENT; 181 dev_dbg(dev, "EM: OPP:%lu is inefficient\n", 182 table[i].frequency); 183 } else { 184 prev_cost = table[i].cost; 185 } 186 } 187 188 pd->table = table; 189 pd->nr_perf_states = nr_states; 190 191 return 0; 192 193 free_ps_table: 194 kfree(table); 195 return -EINVAL; 196 } 197 198 static int em_create_pd(struct device *dev, int nr_states, 199 struct em_data_callback *cb, cpumask_t *cpus, 200 unsigned long flags) 201 { 202 struct em_perf_domain *pd; 203 struct device *cpu_dev; 204 int cpu, ret; 205 206 if (_is_cpu_device(dev)) { 207 pd = kzalloc(sizeof(*pd) + cpumask_size(), GFP_KERNEL); 208 if (!pd) 209 return -ENOMEM; 210 211 cpumask_copy(em_span_cpus(pd), cpus); 212 } else { 213 pd = kzalloc(sizeof(*pd), GFP_KERNEL); 214 if (!pd) 215 return -ENOMEM; 216 } 217 218 ret = em_create_perf_table(dev, pd, nr_states, cb, flags); 219 if (ret) { 220 kfree(pd); 221 return ret; 222 } 223 224 if (_is_cpu_device(dev)) 225 for_each_cpu(cpu, cpus) { 226 cpu_dev = get_cpu_device(cpu); 227 cpu_dev->em_pd = pd; 228 } 229 230 dev->em_pd = pd; 231 232 return 0; 233 } 234 235 static void em_cpufreq_update_efficiencies(struct device *dev) 236 { 237 struct em_perf_domain *pd = dev->em_pd; 238 struct em_perf_state *table; 239 struct cpufreq_policy *policy; 240 int found = 0; 241 int i; 242 243 if (!_is_cpu_device(dev) || !pd) 244 return; 245 246 policy = cpufreq_cpu_get(cpumask_first(em_span_cpus(pd))); 247 if (!policy) { 248 dev_warn(dev, "EM: Access to CPUFreq policy failed"); 249 return; 250 } 251 252 table = pd->table; 253 254 for (i = 0; i < pd->nr_perf_states; i++) { 255 if (!(table[i].flags & EM_PERF_STATE_INEFFICIENT)) 256 continue; 257 258 if (!cpufreq_table_set_inefficient(policy, table[i].frequency)) 259 found++; 260 } 261 262 cpufreq_cpu_put(policy); 263 264 if (!found) 265 return; 266 267 /* 268 * Efficiencies have been installed in CPUFreq, inefficient frequencies 269 * will be skipped. The EM can do the same. 270 */ 271 pd->flags |= EM_PERF_DOMAIN_SKIP_INEFFICIENCIES; 272 } 273 274 /** 275 * em_pd_get() - Return the performance domain for a device 276 * @dev : Device to find the performance domain for 277 * 278 * Returns the performance domain to which @dev belongs, or NULL if it doesn't 279 * exist. 280 */ 281 struct em_perf_domain *em_pd_get(struct device *dev) 282 { 283 if (IS_ERR_OR_NULL(dev)) 284 return NULL; 285 286 return dev->em_pd; 287 } 288 EXPORT_SYMBOL_GPL(em_pd_get); 289 290 /** 291 * em_cpu_get() - Return the performance domain for a CPU 292 * @cpu : CPU to find the performance domain for 293 * 294 * Returns the performance domain to which @cpu belongs, or NULL if it doesn't 295 * exist. 296 */ 297 struct em_perf_domain *em_cpu_get(int cpu) 298 { 299 struct device *cpu_dev; 300 301 cpu_dev = get_cpu_device(cpu); 302 if (!cpu_dev) 303 return NULL; 304 305 return em_pd_get(cpu_dev); 306 } 307 EXPORT_SYMBOL_GPL(em_cpu_get); 308 309 /** 310 * em_dev_register_perf_domain() - Register the Energy Model (EM) for a device 311 * @dev : Device for which the EM is to register 312 * @nr_states : Number of performance states to register 313 * @cb : Callback functions providing the data of the Energy Model 314 * @cpus : Pointer to cpumask_t, which in case of a CPU device is 315 * obligatory. It can be taken from i.e. 'policy->cpus'. For other 316 * type of devices this should be set to NULL. 317 * @milliwatts : Flag indicating that the power values are in milliWatts or 318 * in some other scale. It must be set properly. 319 * 320 * Create Energy Model tables for a performance domain using the callbacks 321 * defined in cb. 322 * 323 * The @milliwatts is important to set with correct value. Some kernel 324 * sub-systems might rely on this flag and check if all devices in the EM are 325 * using the same scale. 326 * 327 * If multiple clients register the same performance domain, all but the first 328 * registration will be ignored. 329 * 330 * Return 0 on success 331 */ 332 int em_dev_register_perf_domain(struct device *dev, unsigned int nr_states, 333 struct em_data_callback *cb, cpumask_t *cpus, 334 bool milliwatts) 335 { 336 unsigned long cap, prev_cap = 0; 337 unsigned long flags = 0; 338 int cpu, ret; 339 340 if (!dev || !nr_states || !cb) 341 return -EINVAL; 342 343 /* 344 * Use a mutex to serialize the registration of performance domains and 345 * let the driver-defined callback functions sleep. 346 */ 347 mutex_lock(&em_pd_mutex); 348 349 if (dev->em_pd) { 350 ret = -EEXIST; 351 goto unlock; 352 } 353 354 if (_is_cpu_device(dev)) { 355 if (!cpus) { 356 dev_err(dev, "EM: invalid CPU mask\n"); 357 ret = -EINVAL; 358 goto unlock; 359 } 360 361 for_each_cpu(cpu, cpus) { 362 if (em_cpu_get(cpu)) { 363 dev_err(dev, "EM: exists for CPU%d\n", cpu); 364 ret = -EEXIST; 365 goto unlock; 366 } 367 /* 368 * All CPUs of a domain must have the same 369 * micro-architecture since they all share the same 370 * table. 371 */ 372 cap = arch_scale_cpu_capacity(cpu); 373 if (prev_cap && prev_cap != cap) { 374 dev_err(dev, "EM: CPUs of %*pbl must have the same capacity\n", 375 cpumask_pr_args(cpus)); 376 377 ret = -EINVAL; 378 goto unlock; 379 } 380 prev_cap = cap; 381 } 382 } 383 384 if (milliwatts) 385 flags |= EM_PERF_DOMAIN_MILLIWATTS; 386 else if (cb->get_cost) 387 flags |= EM_PERF_DOMAIN_ARTIFICIAL; 388 389 ret = em_create_pd(dev, nr_states, cb, cpus, flags); 390 if (ret) 391 goto unlock; 392 393 dev->em_pd->flags |= flags; 394 395 em_cpufreq_update_efficiencies(dev); 396 397 em_debug_create_pd(dev); 398 dev_info(dev, "EM: created perf domain\n"); 399 400 unlock: 401 mutex_unlock(&em_pd_mutex); 402 return ret; 403 } 404 EXPORT_SYMBOL_GPL(em_dev_register_perf_domain); 405 406 /** 407 * em_dev_unregister_perf_domain() - Unregister Energy Model (EM) for a device 408 * @dev : Device for which the EM is registered 409 * 410 * Unregister the EM for the specified @dev (but not a CPU device). 411 */ 412 void em_dev_unregister_perf_domain(struct device *dev) 413 { 414 if (IS_ERR_OR_NULL(dev) || !dev->em_pd) 415 return; 416 417 if (_is_cpu_device(dev)) 418 return; 419 420 /* 421 * The mutex separates all register/unregister requests and protects 422 * from potential clean-up/setup issues in the debugfs directories. 423 * The debugfs directory name is the same as device's name. 424 */ 425 mutex_lock(&em_pd_mutex); 426 em_debug_remove_pd(dev); 427 428 kfree(dev->em_pd->table); 429 kfree(dev->em_pd); 430 dev->em_pd = NULL; 431 mutex_unlock(&em_pd_mutex); 432 } 433 EXPORT_SYMBOL_GPL(em_dev_unregister_perf_domain); 434