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_units_show(struct seq_file *s, void *unused) 58 { 59 struct em_perf_domain *pd = s->private; 60 char *units = (pd->flags & EM_PERF_DOMAIN_MILLIWATTS) ? 61 "milliWatts" : "bogoWatts"; 62 63 seq_printf(s, "%s\n", units); 64 65 return 0; 66 } 67 DEFINE_SHOW_ATTRIBUTE(em_debug_units); 68 69 static int em_debug_skip_inefficiencies_show(struct seq_file *s, void *unused) 70 { 71 struct em_perf_domain *pd = s->private; 72 int enabled = (pd->flags & EM_PERF_DOMAIN_SKIP_INEFFICIENCIES) ? 1 : 0; 73 74 seq_printf(s, "%d\n", enabled); 75 76 return 0; 77 } 78 DEFINE_SHOW_ATTRIBUTE(em_debug_skip_inefficiencies); 79 80 static void em_debug_create_pd(struct device *dev) 81 { 82 struct dentry *d; 83 int i; 84 85 /* Create the directory of the performance domain */ 86 d = debugfs_create_dir(dev_name(dev), rootdir); 87 88 if (_is_cpu_device(dev)) 89 debugfs_create_file("cpus", 0444, d, dev->em_pd->cpus, 90 &em_debug_cpus_fops); 91 92 debugfs_create_file("units", 0444, d, dev->em_pd, &em_debug_units_fops); 93 debugfs_create_file("skip-inefficiencies", 0444, d, dev->em_pd, 94 &em_debug_skip_inefficiencies_fops); 95 96 /* Create a sub-directory for each performance state */ 97 for (i = 0; i < dev->em_pd->nr_perf_states; i++) 98 em_debug_create_ps(&dev->em_pd->table[i], d); 99 100 } 101 102 static void em_debug_remove_pd(struct device *dev) 103 { 104 struct dentry *debug_dir; 105 106 debug_dir = debugfs_lookup(dev_name(dev), rootdir); 107 debugfs_remove_recursive(debug_dir); 108 } 109 110 static int __init em_debug_init(void) 111 { 112 /* Create /sys/kernel/debug/energy_model directory */ 113 rootdir = debugfs_create_dir("energy_model", NULL); 114 115 return 0; 116 } 117 fs_initcall(em_debug_init); 118 #else /* CONFIG_DEBUG_FS */ 119 static void em_debug_create_pd(struct device *dev) {} 120 static void em_debug_remove_pd(struct device *dev) {} 121 #endif 122 123 static int em_create_perf_table(struct device *dev, struct em_perf_domain *pd, 124 int nr_states, struct em_data_callback *cb, 125 unsigned long flags) 126 { 127 unsigned long power, freq, prev_freq = 0, prev_cost = ULONG_MAX; 128 struct em_perf_state *table; 129 int i, ret; 130 u64 fmax; 131 132 table = kcalloc(nr_states, sizeof(*table), GFP_KERNEL); 133 if (!table) 134 return -ENOMEM; 135 136 /* Build the list of performance states for this performance domain */ 137 for (i = 0, freq = 0; i < nr_states; i++, freq++) { 138 /* 139 * active_power() is a driver callback which ceils 'freq' to 140 * lowest performance state of 'dev' above 'freq' and updates 141 * 'power' and 'freq' accordingly. 142 */ 143 ret = cb->active_power(dev, &power, &freq); 144 if (ret) { 145 dev_err(dev, "EM: invalid perf. state: %d\n", 146 ret); 147 goto free_ps_table; 148 } 149 150 /* 151 * We expect the driver callback to increase the frequency for 152 * higher performance states. 153 */ 154 if (freq <= prev_freq) { 155 dev_err(dev, "EM: non-increasing freq: %lu\n", 156 freq); 157 goto free_ps_table; 158 } 159 160 /* 161 * The power returned by active_state() is expected to be 162 * positive and to fit into 16 bits. 163 */ 164 if (!power || power > EM_MAX_POWER) { 165 dev_err(dev, "EM: invalid power: %lu\n", 166 power); 167 goto free_ps_table; 168 } 169 170 table[i].power = power; 171 table[i].frequency = prev_freq = freq; 172 } 173 174 /* Compute the cost of each performance state. */ 175 fmax = (u64) table[nr_states - 1].frequency; 176 for (i = nr_states - 1; i >= 0; i--) { 177 unsigned long power_res, cost; 178 179 if (flags & EM_PERF_DOMAIN_ARTIFICIAL) { 180 ret = cb->get_cost(dev, table[i].frequency, &cost); 181 if (ret || !cost || cost > EM_MAX_POWER) { 182 dev_err(dev, "EM: invalid cost %lu %d\n", 183 cost, ret); 184 goto free_ps_table; 185 } 186 } else { 187 power_res = em_scale_power(table[i].power); 188 cost = div64_u64(fmax * power_res, table[i].frequency); 189 } 190 191 table[i].cost = cost; 192 193 if (table[i].cost >= prev_cost) { 194 table[i].flags = EM_PERF_STATE_INEFFICIENT; 195 dev_dbg(dev, "EM: OPP:%lu is inefficient\n", 196 table[i].frequency); 197 } else { 198 prev_cost = table[i].cost; 199 } 200 } 201 202 pd->table = table; 203 pd->nr_perf_states = nr_states; 204 205 return 0; 206 207 free_ps_table: 208 kfree(table); 209 return -EINVAL; 210 } 211 212 static int em_create_pd(struct device *dev, int nr_states, 213 struct em_data_callback *cb, cpumask_t *cpus, 214 unsigned long flags) 215 { 216 struct em_perf_domain *pd; 217 struct device *cpu_dev; 218 int cpu, ret; 219 220 if (_is_cpu_device(dev)) { 221 pd = kzalloc(sizeof(*pd) + cpumask_size(), GFP_KERNEL); 222 if (!pd) 223 return -ENOMEM; 224 225 cpumask_copy(em_span_cpus(pd), cpus); 226 } else { 227 pd = kzalloc(sizeof(*pd), GFP_KERNEL); 228 if (!pd) 229 return -ENOMEM; 230 } 231 232 ret = em_create_perf_table(dev, pd, nr_states, cb, flags); 233 if (ret) { 234 kfree(pd); 235 return ret; 236 } 237 238 if (_is_cpu_device(dev)) 239 for_each_cpu(cpu, cpus) { 240 cpu_dev = get_cpu_device(cpu); 241 cpu_dev->em_pd = pd; 242 } 243 244 dev->em_pd = pd; 245 246 return 0; 247 } 248 249 static void em_cpufreq_update_efficiencies(struct device *dev) 250 { 251 struct em_perf_domain *pd = dev->em_pd; 252 struct em_perf_state *table; 253 struct cpufreq_policy *policy; 254 int found = 0; 255 int i; 256 257 if (!_is_cpu_device(dev) || !pd) 258 return; 259 260 policy = cpufreq_cpu_get(cpumask_first(em_span_cpus(pd))); 261 if (!policy) { 262 dev_warn(dev, "EM: Access to CPUFreq policy failed"); 263 return; 264 } 265 266 table = pd->table; 267 268 for (i = 0; i < pd->nr_perf_states; i++) { 269 if (!(table[i].flags & EM_PERF_STATE_INEFFICIENT)) 270 continue; 271 272 if (!cpufreq_table_set_inefficient(policy, table[i].frequency)) 273 found++; 274 } 275 276 if (!found) 277 return; 278 279 /* 280 * Efficiencies have been installed in CPUFreq, inefficient frequencies 281 * will be skipped. The EM can do the same. 282 */ 283 pd->flags |= EM_PERF_DOMAIN_SKIP_INEFFICIENCIES; 284 } 285 286 /** 287 * em_pd_get() - Return the performance domain for a device 288 * @dev : Device to find the performance domain for 289 * 290 * Returns the performance domain to which @dev belongs, or NULL if it doesn't 291 * exist. 292 */ 293 struct em_perf_domain *em_pd_get(struct device *dev) 294 { 295 if (IS_ERR_OR_NULL(dev)) 296 return NULL; 297 298 return dev->em_pd; 299 } 300 EXPORT_SYMBOL_GPL(em_pd_get); 301 302 /** 303 * em_cpu_get() - Return the performance domain for a CPU 304 * @cpu : CPU to find the performance domain for 305 * 306 * Returns the performance domain to which @cpu belongs, or NULL if it doesn't 307 * exist. 308 */ 309 struct em_perf_domain *em_cpu_get(int cpu) 310 { 311 struct device *cpu_dev; 312 313 cpu_dev = get_cpu_device(cpu); 314 if (!cpu_dev) 315 return NULL; 316 317 return em_pd_get(cpu_dev); 318 } 319 EXPORT_SYMBOL_GPL(em_cpu_get); 320 321 /** 322 * em_dev_register_perf_domain() - Register the Energy Model (EM) for a device 323 * @dev : Device for which the EM is to register 324 * @nr_states : Number of performance states to register 325 * @cb : Callback functions providing the data of the Energy Model 326 * @cpus : Pointer to cpumask_t, which in case of a CPU device is 327 * obligatory. It can be taken from i.e. 'policy->cpus'. For other 328 * type of devices this should be set to NULL. 329 * @milliwatts : Flag indicating that the power values are in milliWatts or 330 * in some other scale. It must be set properly. 331 * 332 * Create Energy Model tables for a performance domain using the callbacks 333 * defined in cb. 334 * 335 * The @milliwatts is important to set with correct value. Some kernel 336 * sub-systems might rely on this flag and check if all devices in the EM are 337 * using the same scale. 338 * 339 * If multiple clients register the same performance domain, all but the first 340 * registration will be ignored. 341 * 342 * Return 0 on success 343 */ 344 int em_dev_register_perf_domain(struct device *dev, unsigned int nr_states, 345 struct em_data_callback *cb, cpumask_t *cpus, 346 bool milliwatts) 347 { 348 unsigned long cap, prev_cap = 0; 349 unsigned long flags = 0; 350 int cpu, ret; 351 352 if (!dev || !nr_states || !cb) 353 return -EINVAL; 354 355 /* 356 * Use a mutex to serialize the registration of performance domains and 357 * let the driver-defined callback functions sleep. 358 */ 359 mutex_lock(&em_pd_mutex); 360 361 if (dev->em_pd) { 362 ret = -EEXIST; 363 goto unlock; 364 } 365 366 if (_is_cpu_device(dev)) { 367 if (!cpus) { 368 dev_err(dev, "EM: invalid CPU mask\n"); 369 ret = -EINVAL; 370 goto unlock; 371 } 372 373 for_each_cpu(cpu, cpus) { 374 if (em_cpu_get(cpu)) { 375 dev_err(dev, "EM: exists for CPU%d\n", cpu); 376 ret = -EEXIST; 377 goto unlock; 378 } 379 /* 380 * All CPUs of a domain must have the same 381 * micro-architecture since they all share the same 382 * table. 383 */ 384 cap = arch_scale_cpu_capacity(cpu); 385 if (prev_cap && prev_cap != cap) { 386 dev_err(dev, "EM: CPUs of %*pbl must have the same capacity\n", 387 cpumask_pr_args(cpus)); 388 389 ret = -EINVAL; 390 goto unlock; 391 } 392 prev_cap = cap; 393 } 394 } 395 396 if (milliwatts) 397 flags |= EM_PERF_DOMAIN_MILLIWATTS; 398 else if (cb->get_cost) 399 flags |= EM_PERF_DOMAIN_ARTIFICIAL; 400 401 ret = em_create_pd(dev, nr_states, cb, cpus, flags); 402 if (ret) 403 goto unlock; 404 405 dev->em_pd->flags |= flags; 406 407 em_cpufreq_update_efficiencies(dev); 408 409 em_debug_create_pd(dev); 410 dev_info(dev, "EM: created perf domain\n"); 411 412 unlock: 413 mutex_unlock(&em_pd_mutex); 414 return ret; 415 } 416 EXPORT_SYMBOL_GPL(em_dev_register_perf_domain); 417 418 /** 419 * em_dev_unregister_perf_domain() - Unregister Energy Model (EM) for a device 420 * @dev : Device for which the EM is registered 421 * 422 * Unregister the EM for the specified @dev (but not a CPU device). 423 */ 424 void em_dev_unregister_perf_domain(struct device *dev) 425 { 426 if (IS_ERR_OR_NULL(dev) || !dev->em_pd) 427 return; 428 429 if (_is_cpu_device(dev)) 430 return; 431 432 /* 433 * The mutex separates all register/unregister requests and protects 434 * from potential clean-up/setup issues in the debugfs directories. 435 * The debugfs directory name is the same as device's name. 436 */ 437 mutex_lock(&em_pd_mutex); 438 em_debug_remove_pd(dev); 439 440 kfree(dev->em_pd->table); 441 kfree(dev->em_pd); 442 dev->em_pd = NULL; 443 mutex_unlock(&em_pd_mutex); 444 } 445 EXPORT_SYMBOL_GPL(em_dev_unregister_perf_domain); 446