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
_is_cpu_device(struct device * dev)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
em_debug_create_ps(struct em_perf_state * ps,struct dentry * pd)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
em_debug_cpus_show(struct seq_file * s,void * unused)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
em_debug_create_pd(struct device * dev)55 static void em_debug_create_pd(struct device *dev)
56 {
57 struct dentry *d;
58 int i;
59
60 /* Create the directory of the performance domain */
61 d = debugfs_create_dir(dev_name(dev), rootdir);
62
63 if (_is_cpu_device(dev))
64 debugfs_create_file("cpus", 0444, d, dev->em_pd->cpus,
65 &em_debug_cpus_fops);
66
67 /* Create a sub-directory for each performance state */
68 for (i = 0; i < dev->em_pd->nr_perf_states; i++)
69 em_debug_create_ps(&dev->em_pd->table[i], d);
70
71 }
72
em_debug_remove_pd(struct device * dev)73 static void em_debug_remove_pd(struct device *dev)
74 {
75 struct dentry *debug_dir;
76
77 debug_dir = debugfs_lookup(dev_name(dev), rootdir);
78 debugfs_remove_recursive(debug_dir);
79 }
80
em_debug_init(void)81 static int __init em_debug_init(void)
82 {
83 /* Create /sys/kernel/debug/energy_model directory */
84 rootdir = debugfs_create_dir("energy_model", NULL);
85
86 return 0;
87 }
88 fs_initcall(em_debug_init);
89 #else /* CONFIG_DEBUG_FS */
em_debug_create_pd(struct device * dev)90 static void em_debug_create_pd(struct device *dev) {}
em_debug_remove_pd(struct device * dev)91 static void em_debug_remove_pd(struct device *dev) {}
92 #endif
93
em_create_perf_table(struct device * dev,struct em_perf_domain * pd,int nr_states,struct em_data_callback * cb)94 static int em_create_perf_table(struct device *dev, struct em_perf_domain *pd,
95 int nr_states, struct em_data_callback *cb)
96 {
97 unsigned long power, freq, prev_freq = 0, prev_cost = ULONG_MAX;
98 struct em_perf_state *table;
99 int i, ret;
100 u64 fmax;
101
102 table = kcalloc(nr_states, sizeof(*table), GFP_KERNEL);
103 if (!table)
104 return -ENOMEM;
105
106 /* Build the list of performance states for this performance domain */
107 for (i = 0, freq = 0; i < nr_states; i++, freq++) {
108 /*
109 * active_power() is a driver callback which ceils 'freq' to
110 * lowest performance state of 'dev' above 'freq' and updates
111 * 'power' and 'freq' accordingly.
112 */
113 ret = cb->active_power(&power, &freq, dev);
114 if (ret) {
115 dev_err(dev, "EM: invalid perf. state: %d\n",
116 ret);
117 goto free_ps_table;
118 }
119
120 /*
121 * We expect the driver callback to increase the frequency for
122 * higher performance states.
123 */
124 if (freq <= prev_freq) {
125 dev_err(dev, "EM: non-increasing freq: %lu\n",
126 freq);
127 goto free_ps_table;
128 }
129
130 /*
131 * The power returned by active_state() is expected to be
132 * positive, in milli-watts and to fit into 16 bits.
133 */
134 if (!power || power > EM_MAX_POWER) {
135 dev_err(dev, "EM: invalid power: %lu\n",
136 power);
137 goto free_ps_table;
138 }
139
140 table[i].power = power;
141 table[i].frequency = prev_freq = freq;
142 }
143
144 /* Compute the cost of each performance state. */
145 fmax = (u64) table[nr_states - 1].frequency;
146 for (i = nr_states - 1; i >= 0; i--) {
147 unsigned long power_res = em_scale_power(table[i].power);
148
149 table[i].cost = div64_u64(fmax * power_res,
150 table[i].frequency);
151 if (table[i].cost >= prev_cost) {
152 dev_dbg(dev, "EM: OPP:%lu is inefficient\n",
153 table[i].frequency);
154 } else {
155 prev_cost = table[i].cost;
156 }
157 }
158
159 pd->table = table;
160 pd->nr_perf_states = nr_states;
161
162 return 0;
163
164 free_ps_table:
165 kfree(table);
166 return -EINVAL;
167 }
168
em_create_pd(struct device * dev,int nr_states,struct em_data_callback * cb,cpumask_t * cpus)169 static int em_create_pd(struct device *dev, int nr_states,
170 struct em_data_callback *cb, cpumask_t *cpus)
171 {
172 struct em_perf_domain *pd;
173 struct device *cpu_dev;
174 int cpu, ret;
175
176 if (_is_cpu_device(dev)) {
177 pd = kzalloc(sizeof(*pd) + cpumask_size(), GFP_KERNEL);
178 if (!pd)
179 return -ENOMEM;
180
181 cpumask_copy(em_span_cpus(pd), cpus);
182 } else {
183 pd = kzalloc(sizeof(*pd), GFP_KERNEL);
184 if (!pd)
185 return -ENOMEM;
186 }
187
188 ret = em_create_perf_table(dev, pd, nr_states, cb);
189 if (ret) {
190 kfree(pd);
191 return ret;
192 }
193
194 if (_is_cpu_device(dev))
195 for_each_cpu(cpu, cpus) {
196 cpu_dev = get_cpu_device(cpu);
197 cpu_dev->em_pd = pd;
198 }
199
200 dev->em_pd = pd;
201
202 return 0;
203 }
204
205 /**
206 * em_pd_get() - Return the performance domain for a device
207 * @dev : Device to find the performance domain for
208 *
209 * Returns the performance domain to which @dev belongs, or NULL if it doesn't
210 * exist.
211 */
em_pd_get(struct device * dev)212 struct em_perf_domain *em_pd_get(struct device *dev)
213 {
214 if (IS_ERR_OR_NULL(dev))
215 return NULL;
216
217 return dev->em_pd;
218 }
219 EXPORT_SYMBOL_GPL(em_pd_get);
220
221 /**
222 * em_cpu_get() - Return the performance domain for a CPU
223 * @cpu : CPU to find the performance domain for
224 *
225 * Returns the performance domain to which @cpu belongs, or NULL if it doesn't
226 * exist.
227 */
em_cpu_get(int cpu)228 struct em_perf_domain *em_cpu_get(int cpu)
229 {
230 struct device *cpu_dev;
231
232 cpu_dev = get_cpu_device(cpu);
233 if (!cpu_dev)
234 return NULL;
235
236 return em_pd_get(cpu_dev);
237 }
238 EXPORT_SYMBOL_GPL(em_cpu_get);
239
240 /**
241 * em_dev_register_perf_domain() - Register the Energy Model (EM) for a device
242 * @dev : Device for which the EM is to register
243 * @nr_states : Number of performance states to register
244 * @cb : Callback functions providing the data of the Energy Model
245 * @cpus : Pointer to cpumask_t, which in case of a CPU device is
246 * obligatory. It can be taken from i.e. 'policy->cpus'. For other
247 * type of devices this should be set to NULL.
248 *
249 * Create Energy Model tables for a performance domain using the callbacks
250 * defined in cb.
251 *
252 * If multiple clients register the same performance domain, all but the first
253 * registration will be ignored.
254 *
255 * Return 0 on success
256 */
em_dev_register_perf_domain(struct device * dev,unsigned int nr_states,struct em_data_callback * cb,cpumask_t * cpus)257 int em_dev_register_perf_domain(struct device *dev, unsigned int nr_states,
258 struct em_data_callback *cb, cpumask_t *cpus)
259 {
260 unsigned long cap, prev_cap = 0;
261 int cpu, ret;
262
263 if (!dev || !nr_states || !cb)
264 return -EINVAL;
265
266 /*
267 * Use a mutex to serialize the registration of performance domains and
268 * let the driver-defined callback functions sleep.
269 */
270 mutex_lock(&em_pd_mutex);
271
272 if (dev->em_pd) {
273 ret = -EEXIST;
274 goto unlock;
275 }
276
277 if (_is_cpu_device(dev)) {
278 if (!cpus) {
279 dev_err(dev, "EM: invalid CPU mask\n");
280 ret = -EINVAL;
281 goto unlock;
282 }
283
284 for_each_cpu(cpu, cpus) {
285 if (em_cpu_get(cpu)) {
286 dev_err(dev, "EM: exists for CPU%d\n", cpu);
287 ret = -EEXIST;
288 goto unlock;
289 }
290 /*
291 * All CPUs of a domain must have the same
292 * micro-architecture since they all share the same
293 * table.
294 */
295 cap = arch_scale_cpu_capacity(cpu);
296 if (prev_cap && prev_cap != cap) {
297 dev_err(dev, "EM: CPUs of %*pbl must have the same capacity\n",
298 cpumask_pr_args(cpus));
299
300 ret = -EINVAL;
301 goto unlock;
302 }
303 prev_cap = cap;
304 }
305 }
306
307 ret = em_create_pd(dev, nr_states, cb, cpus);
308 if (ret)
309 goto unlock;
310
311 em_debug_create_pd(dev);
312 dev_info(dev, "EM: created perf domain\n");
313
314 unlock:
315 mutex_unlock(&em_pd_mutex);
316 return ret;
317 }
318 EXPORT_SYMBOL_GPL(em_dev_register_perf_domain);
319
320 /**
321 * em_dev_unregister_perf_domain() - Unregister Energy Model (EM) for a device
322 * @dev : Device for which the EM is registered
323 *
324 * Unregister the EM for the specified @dev (but not a CPU device).
325 */
em_dev_unregister_perf_domain(struct device * dev)326 void em_dev_unregister_perf_domain(struct device *dev)
327 {
328 if (IS_ERR_OR_NULL(dev) || !dev->em_pd)
329 return;
330
331 if (_is_cpu_device(dev))
332 return;
333
334 /*
335 * The mutex separates all register/unregister requests and protects
336 * from potential clean-up/setup issues in the debugfs directories.
337 * The debugfs directory name is the same as device's name.
338 */
339 mutex_lock(&em_pd_mutex);
340 em_debug_remove_pd(dev);
341
342 kfree(dev->em_pd->table);
343 kfree(dev->em_pd);
344 dev->em_pd = NULL;
345 mutex_unlock(&em_pd_mutex);
346 }
347 EXPORT_SYMBOL_GPL(em_dev_unregister_perf_domain);
348