kernel/power/energy_model.c

   1 // SPDX-License-Identifier: GPL-2.0
   2 /*
   3  * Energy Model of devices
   4  *
   5  * Copyright (c) 2018-2020, 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/cpumask.h>
  14 #include <linux/debugfs.h>
  15 #include <linux/energy_model.h>
  16 #include <linux/sched/topology.h>
  17 #include <linux/slab.h>
  18
  19 /*
  20  * Mutex serializing the registrations of performance domains and letting
  21  * callbacks defined by drivers sleep.
  22  */
  23 static DEFINE_MUTEX(em_pd_mutex);
  24
  25 static bool _is_cpu_device(struct device *dev)
  26 {
  27         return (dev->bus == &cpu_subsys);
  28 }
  29
  30 #ifdef CONFIG_DEBUG_FS
  31 static struct dentry *rootdir;
  32
  33 static void em_debug_create_ps(struct em_perf_state *ps, struct dentry *pd)
  34 {
  35         struct dentry *d;
  36         char name[24];
  37
  38         snprintf(name, sizeof(name), "ps:%lu", ps->frequency);
  39
  40         /* Create per-ps directory */
  41         d = debugfs_create_dir(name, pd);
  42         debugfs_create_ulong("frequency", 0444, d, &ps->frequency);
  43         debugfs_create_ulong("power", 0444, d, &ps->power);
  44         debugfs_create_ulong("cost", 0444, d, &ps->cost);
  45 }
  46
  47 static int em_debug_cpus_show(struct seq_file *s, void *unused)
  48 {
  49         seq_printf(s, "%*pbl\n", cpumask_pr_args(to_cpumask(s->private)));
  50
  51         return 0;
  52 }
  53 DEFINE_SHOW_ATTRIBUTE(em_debug_cpus);
  54
  55 static int em_debug_units_show(struct seq_file *s, void *unused)
  56 {
  57         struct em_perf_domain *pd = s->private;
  58         char *units = pd->milliwatts ? "milliWatts" : "bogoWatts";
  59
  60         seq_printf(s, "%s\n", units);
  61
  62         return 0;
  63 }
  64 DEFINE_SHOW_ATTRIBUTE(em_debug_units);
  65
  66 static void em_debug_create_pd(struct device *dev)
  67 {
  68         struct dentry *d;
  69         int i;
  70
  71         /* Create the directory of the performance domain */
  72         d = debugfs_create_dir(dev_name(dev), rootdir);
  73
  74         if (_is_cpu_device(dev))
  75                 debugfs_create_file("cpus", 0444, d, dev->em_pd->cpus,
  76                                     &em_debug_cpus_fops);
  77
  78         debugfs_create_file("units", 0444, d, dev->em_pd, &em_debug_units_fops);
  79
  80         /* Create a sub-directory for each performance state */
  81         for (i = 0; i < dev->em_pd->nr_perf_states; i++)
  82                 em_debug_create_ps(&dev->em_pd->table[i], d);
  83
  84 }
  85
  86 static void em_debug_remove_pd(struct device *dev)
  87 {
  88         struct dentry *debug_dir;
  89
  90         debug_dir = debugfs_lookup(dev_name(dev), rootdir);
  91         debugfs_remove_recursive(debug_dir);
  92 }
  93
  94 static int __init em_debug_init(void)
  95 {
  96         /* Create /sys/kernel/debug/energy_model directory */
  97         rootdir = debugfs_create_dir("energy_model", NULL);
  98
  99         return 0;
 100 }
 101 core_initcall(em_debug_init);
 102 #else /* CONFIG_DEBUG_FS */
 103 static void em_debug_create_pd(struct device *dev) {}
 104 static void em_debug_remove_pd(struct device *dev) {}
 105 #endif
 106
 107 static int em_create_perf_table(struct device *dev, struct em_perf_domain *pd,
 108                                 int nr_states, struct em_data_callback *cb)
 109 {
 110         unsigned long opp_eff, prev_opp_eff = ULONG_MAX;
 111         unsigned long power, freq, prev_freq = 0;
 112         struct em_perf_state *table;
 113         int i, ret;
 114         u64 fmax;
 115
 116         table = kcalloc(nr_states, sizeof(*table), GFP_KERNEL);
 117         if (!table)
 118                 return -ENOMEM;
 119
 120         /* Build the list of performance states for this performance domain */
 121         for (i = 0, freq = 0; i < nr_states; i++, freq++) {
 122                 /*
 123                  * active_power() is a driver callback which ceils 'freq' to
 124                  * lowest performance state of 'dev' above 'freq' and updates
 125                  * 'power' and 'freq' accordingly.
 126                  */
 127                 ret = cb->active_power(&power, &freq, dev);
 128                 if (ret) {
 129                         dev_err(dev, "EM: invalid perf. state: %d\n",
 130                                 ret);
 131                         goto free_ps_table;
 132                 }
 133
 134                 /*
 135                  * We expect the driver callback to increase the frequency for
 136                  * higher performance states.
 137                  */
 138                 if (freq <= prev_freq) {
 139                         dev_err(dev, "EM: non-increasing freq: %lu\n",
 140                                 freq);
 141                         goto free_ps_table;
 142                 }
 143
 144                 /*
 145                  * The power returned by active_state() is expected to be
 146                  * positive and to fit into 16 bits.
 147                  */
 148                 if (!power || power > EM_MAX_POWER) {
 149                         dev_err(dev, "EM: invalid power: %lu\n",
 150                                 power);
 151                         goto free_ps_table;
 152                 }
 153
 154                 table[i].power = power;
 155                 table[i].frequency = prev_freq = freq;
 156
 157                 /*
 158                  * The hertz/watts efficiency ratio should decrease as the
 159                  * frequency grows on sane platforms. But this isn't always
 160                  * true in practice so warn the user if a higher OPP is more
 161                  * power efficient than a lower one.
 162                  */
 163                 opp_eff = freq / power;
 164                 if (opp_eff >= prev_opp_eff)
 165                         dev_dbg(dev, "EM: hertz/watts ratio non-monotonically decreasing: em_perf_state %d >= em_perf_state%d\n",
 166                                         i, i - 1);
 167                 prev_opp_eff = opp_eff;
 168         }
 169
 170         /* Compute the cost of each performance state. */
 171         fmax = (u64) table[nr_states - 1].frequency;
 172         for (i = 0; i < nr_states; i++) {
 173                 table[i].cost = div64_u64(fmax * table[i].power,
 174                                           table[i].frequency);
 175         }
 176
 177         pd->table = table;
 178         pd->nr_perf_states = nr_states;
 179
 180         return 0;
 181
 182 free_ps_table:
 183         kfree(table);
 184         return -EINVAL;
 185 }
 186
 187 static int em_create_pd(struct device *dev, int nr_states,
 188                         struct em_data_callback *cb, cpumask_t *cpus)
 189 {
 190         struct em_perf_domain *pd;
 191         struct device *cpu_dev;
 192         int cpu, ret;
 193
 194         if (_is_cpu_device(dev)) {
 195                 pd = kzalloc(sizeof(*pd) + cpumask_size(), GFP_KERNEL);
 196                 if (!pd)
 197                         return -ENOMEM;
 198
 199                 cpumask_copy(em_span_cpus(pd), cpus);
 200         } else {
 201                 pd = kzalloc(sizeof(*pd), GFP_KERNEL);
 202                 if (!pd)
 203                         return -ENOMEM;
 204         }
 205
 206         ret = em_create_perf_table(dev, pd, nr_states, cb);
 207         if (ret) {
 208                 kfree(pd);
 209                 return ret;
 210         }
 211
 212         if (_is_cpu_device(dev))
 213                 for_each_cpu(cpu, cpus) {
 214                         cpu_dev = get_cpu_device(cpu);
 215                         cpu_dev->em_pd = pd;
 216                 }
 217
 218         dev->em_pd = pd;
 219
 220         return 0;
 221 }
 222
 223 /**
 224  * em_pd_get() - Return the performance domain for a device
 225  * @dev : Device to find the performance domain for
 226  *
 227  * Returns the performance domain to which @dev belongs, or NULL if it doesn't
 228  * exist.
 229  */
 230 struct em_perf_domain *em_pd_get(struct device *dev)
 231 {
 232         if (IS_ERR_OR_NULL(dev))
 233                 return NULL;
 234
 235         return dev->em_pd;
 236 }
 237 EXPORT_SYMBOL_GPL(em_pd_get);
 238
 239 /**
 240  * em_cpu_get() - Return the performance domain for a CPU
 241  * @cpu : CPU to find the performance domain for
 242  *
 243  * Returns the performance domain to which @cpu belongs, or NULL if it doesn't
 244  * exist.
 245  */
 246 struct em_perf_domain *em_cpu_get(int cpu)
 247 {
 248         struct device *cpu_dev;
 249
 250         cpu_dev = get_cpu_device(cpu);
 251         if (!cpu_dev)
 252                 return NULL;
 253
 254         return em_pd_get(cpu_dev);
 255 }
 256 EXPORT_SYMBOL_GPL(em_cpu_get);
 257
 258 /**
 259  * em_dev_register_perf_domain() - Register the Energy Model (EM) for a device
 260  * @dev         : Device for which the EM is to register
 261  * @nr_states   : Number of performance states to register
 262  * @cb          : Callback functions providing the data of the Energy Model
 263  * @cpus        : Pointer to cpumask_t, which in case of a CPU device is
 264  *              obligatory. It can be taken from i.e. 'policy->cpus'. For other
 265  *              type of devices this should be set to NULL.
 266  * @milliwatts  : Flag indicating that the power values are in milliWatts or
 267  *              in some other scale. It must be set properly.
 268  *
 269  * Create Energy Model tables for a performance domain using the callbacks
 270  * defined in cb.
 271  *
 272  * The @milliwatts is important to set with correct value. Some kernel
 273  * sub-systems might rely on this flag and check if all devices in the EM are
 274  * using the same scale.
 275  *
 276  * If multiple clients register the same performance domain, all but the first
 277  * registration will be ignored.
 278  *
 279  * Return 0 on success
 280  */
 281 int em_dev_register_perf_domain(struct device *dev, unsigned int nr_states,
 282                                 struct em_data_callback *cb, cpumask_t *cpus,
 283                                 bool milliwatts)
 284 {
 285         unsigned long cap, prev_cap = 0;
 286         int cpu, ret;
 287
 288         if (!dev || !nr_states || !cb)
 289                 return -EINVAL;
 290
 291         /*
 292          * Use a mutex to serialize the registration of performance domains and
 293          * let the driver-defined callback functions sleep.
 294          */
 295         mutex_lock(&em_pd_mutex);
 296
 297         if (dev->em_pd) {
 298                 ret = -EEXIST;
 299                 goto unlock;
 300         }
 301
 302         if (_is_cpu_device(dev)) {
 303                 if (!cpus) {
 304                         dev_err(dev, "EM: invalid CPU mask\n");
 305                         ret = -EINVAL;
 306                         goto unlock;
 307                 }
 308
 309                 for_each_cpu(cpu, cpus) {
 310                         if (em_cpu_get(cpu)) {
 311                                 dev_err(dev, "EM: exists for CPU%d\n", cpu);
 312                                 ret = -EEXIST;
 313                                 goto unlock;
 314                         }
 315                         /*
 316                          * All CPUs of a domain must have the same
 317                          * micro-architecture since they all share the same
 318                          * table.
 319                          */
 320                         cap = arch_scale_cpu_capacity(cpu);
 321                         if (prev_cap && prev_cap != cap) {
 322                                 dev_err(dev, "EM: CPUs of %*pbl must have the same capacity\n",
 323                                         cpumask_pr_args(cpus));
 324
 325                                 ret = -EINVAL;
 326                                 goto unlock;
 327                         }
 328                         prev_cap = cap;
 329                 }
 330         }
 331
 332         ret = em_create_pd(dev, nr_states, cb, cpus);
 333         if (ret)
 334                 goto unlock;
 335
 336         dev->em_pd->milliwatts = milliwatts;
 337
 338         em_debug_create_pd(dev);
 339         dev_info(dev, "EM: created perf domain\n");
 340
 341 unlock:
 342         mutex_unlock(&em_pd_mutex);
 343         return ret;
 344 }
 345 EXPORT_SYMBOL_GPL(em_dev_register_perf_domain);
 346
 347 /**
 348  * em_dev_unregister_perf_domain() - Unregister Energy Model (EM) for a device
 349  * @dev         : Device for which the EM is registered
 350  *
 351  * Unregister the EM for the specified @dev (but not a CPU device).
 352  */
 353 void em_dev_unregister_perf_domain(struct device *dev)
 354 {
 355         if (IS_ERR_OR_NULL(dev) || !dev->em_pd)
 356                 return;
 357
 358         if (_is_cpu_device(dev))
 359                 return;
 360
 361         /*
 362          * The mutex separates all register/unregister requests and protects
 363          * from potential clean-up/setup issues in the debugfs directories.
 364          * The debugfs directory name is the same as device's name.
 365          */
 366         mutex_lock(&em_pd_mutex);
 367         em_debug_remove_pd(dev);
 368
 369         kfree(dev->em_pd->table);
 370         kfree(dev->em_pd);
 371         dev->em_pd = NULL;
 372         mutex_unlock(&em_pd_mutex);
 373 }
 374 EXPORT_SYMBOL_GPL(em_dev_unregister_perf_domain);