xref: /openbmc/linux/kernel/power/energy_model.c (revision 63705da3)
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 {
126 	unsigned long power, freq, prev_freq = 0, prev_cost = ULONG_MAX;
127 	struct em_perf_state *table;
128 	int i, ret;
129 	u64 fmax;
130 
131 	table = kcalloc(nr_states, sizeof(*table), GFP_KERNEL);
132 	if (!table)
133 		return -ENOMEM;
134 
135 	/* Build the list of performance states for this performance domain */
136 	for (i = 0, freq = 0; i < nr_states; i++, freq++) {
137 		/*
138 		 * active_power() is a driver callback which ceils 'freq' to
139 		 * lowest performance state of 'dev' above 'freq' and updates
140 		 * 'power' and 'freq' accordingly.
141 		 */
142 		ret = cb->active_power(&power, &freq, dev);
143 		if (ret) {
144 			dev_err(dev, "EM: invalid perf. state: %d\n",
145 				ret);
146 			goto free_ps_table;
147 		}
148 
149 		/*
150 		 * We expect the driver callback to increase the frequency for
151 		 * higher performance states.
152 		 */
153 		if (freq <= prev_freq) {
154 			dev_err(dev, "EM: non-increasing freq: %lu\n",
155 				freq);
156 			goto free_ps_table;
157 		}
158 
159 		/*
160 		 * The power returned by active_state() is expected to be
161 		 * positive and to fit into 16 bits.
162 		 */
163 		if (!power || power > EM_MAX_POWER) {
164 			dev_err(dev, "EM: invalid power: %lu\n",
165 				power);
166 			goto free_ps_table;
167 		}
168 
169 		table[i].power = power;
170 		table[i].frequency = prev_freq = freq;
171 	}
172 
173 	/* Compute the cost of each performance state. */
174 	fmax = (u64) table[nr_states - 1].frequency;
175 	for (i = nr_states - 1; i >= 0; i--) {
176 		unsigned long power_res = em_scale_power(table[i].power);
177 
178 		table[i].cost = div64_u64(fmax * power_res,
179 					  table[i].frequency);
180 		if (table[i].cost >= prev_cost) {
181 			table[i].flags = EM_PERF_STATE_INEFFICIENT;
182 			dev_dbg(dev, "EM: OPP:%lu is inefficient\n",
183 				table[i].frequency);
184 		} else {
185 			prev_cost = table[i].cost;
186 		}
187 	}
188 
189 	pd->table = table;
190 	pd->nr_perf_states = nr_states;
191 
192 	return 0;
193 
194 free_ps_table:
195 	kfree(table);
196 	return -EINVAL;
197 }
198 
199 static int em_create_pd(struct device *dev, int nr_states,
200 			struct em_data_callback *cb, cpumask_t *cpus)
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);
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 	if (!found)
263 		return;
264 
265 	/*
266 	 * Efficiencies have been installed in CPUFreq, inefficient frequencies
267 	 * will be skipped. The EM can do the same.
268 	 */
269 	pd->flags |= EM_PERF_DOMAIN_SKIP_INEFFICIENCIES;
270 }
271 
272 /**
273  * em_pd_get() - Return the performance domain for a device
274  * @dev : Device to find the performance domain for
275  *
276  * Returns the performance domain to which @dev belongs, or NULL if it doesn't
277  * exist.
278  */
279 struct em_perf_domain *em_pd_get(struct device *dev)
280 {
281 	if (IS_ERR_OR_NULL(dev))
282 		return NULL;
283 
284 	return dev->em_pd;
285 }
286 EXPORT_SYMBOL_GPL(em_pd_get);
287 
288 /**
289  * em_cpu_get() - Return the performance domain for a CPU
290  * @cpu : CPU to find the performance domain for
291  *
292  * Returns the performance domain to which @cpu belongs, or NULL if it doesn't
293  * exist.
294  */
295 struct em_perf_domain *em_cpu_get(int cpu)
296 {
297 	struct device *cpu_dev;
298 
299 	cpu_dev = get_cpu_device(cpu);
300 	if (!cpu_dev)
301 		return NULL;
302 
303 	return em_pd_get(cpu_dev);
304 }
305 EXPORT_SYMBOL_GPL(em_cpu_get);
306 
307 /**
308  * em_dev_register_perf_domain() - Register the Energy Model (EM) for a device
309  * @dev		: Device for which the EM is to register
310  * @nr_states	: Number of performance states to register
311  * @cb		: Callback functions providing the data of the Energy Model
312  * @cpus	: Pointer to cpumask_t, which in case of a CPU device is
313  *		obligatory. It can be taken from i.e. 'policy->cpus'. For other
314  *		type of devices this should be set to NULL.
315  * @milliwatts	: Flag indicating that the power values are in milliWatts or
316  *		in some other scale. It must be set properly.
317  *
318  * Create Energy Model tables for a performance domain using the callbacks
319  * defined in cb.
320  *
321  * The @milliwatts is important to set with correct value. Some kernel
322  * sub-systems might rely on this flag and check if all devices in the EM are
323  * using the same scale.
324  *
325  * If multiple clients register the same performance domain, all but the first
326  * registration will be ignored.
327  *
328  * Return 0 on success
329  */
330 int em_dev_register_perf_domain(struct device *dev, unsigned int nr_states,
331 				struct em_data_callback *cb, cpumask_t *cpus,
332 				bool milliwatts)
333 {
334 	unsigned long cap, prev_cap = 0;
335 	int cpu, ret;
336 
337 	if (!dev || !nr_states || !cb)
338 		return -EINVAL;
339 
340 	/*
341 	 * Use a mutex to serialize the registration of performance domains and
342 	 * let the driver-defined callback functions sleep.
343 	 */
344 	mutex_lock(&em_pd_mutex);
345 
346 	if (dev->em_pd) {
347 		ret = -EEXIST;
348 		goto unlock;
349 	}
350 
351 	if (_is_cpu_device(dev)) {
352 		if (!cpus) {
353 			dev_err(dev, "EM: invalid CPU mask\n");
354 			ret = -EINVAL;
355 			goto unlock;
356 		}
357 
358 		for_each_cpu(cpu, cpus) {
359 			if (em_cpu_get(cpu)) {
360 				dev_err(dev, "EM: exists for CPU%d\n", cpu);
361 				ret = -EEXIST;
362 				goto unlock;
363 			}
364 			/*
365 			 * All CPUs of a domain must have the same
366 			 * micro-architecture since they all share the same
367 			 * table.
368 			 */
369 			cap = arch_scale_cpu_capacity(cpu);
370 			if (prev_cap && prev_cap != cap) {
371 				dev_err(dev, "EM: CPUs of %*pbl must have the same capacity\n",
372 					cpumask_pr_args(cpus));
373 
374 				ret = -EINVAL;
375 				goto unlock;
376 			}
377 			prev_cap = cap;
378 		}
379 	}
380 
381 	ret = em_create_pd(dev, nr_states, cb, cpus);
382 	if (ret)
383 		goto unlock;
384 
385 	if (milliwatts)
386 		dev->em_pd->flags |= EM_PERF_DOMAIN_MILLIWATTS;
387 
388 	em_cpufreq_update_efficiencies(dev);
389 
390 	em_debug_create_pd(dev);
391 	dev_info(dev, "EM: created perf domain\n");
392 
393 unlock:
394 	mutex_unlock(&em_pd_mutex);
395 	return ret;
396 }
397 EXPORT_SYMBOL_GPL(em_dev_register_perf_domain);
398 
399 /**
400  * em_dev_unregister_perf_domain() - Unregister Energy Model (EM) for a device
401  * @dev		: Device for which the EM is registered
402  *
403  * Unregister the EM for the specified @dev (but not a CPU device).
404  */
405 void em_dev_unregister_perf_domain(struct device *dev)
406 {
407 	if (IS_ERR_OR_NULL(dev) || !dev->em_pd)
408 		return;
409 
410 	if (_is_cpu_device(dev))
411 		return;
412 
413 	/*
414 	 * The mutex separates all register/unregister requests and protects
415 	 * from potential clean-up/setup issues in the debugfs directories.
416 	 * The debugfs directory name is the same as device's name.
417 	 */
418 	mutex_lock(&em_pd_mutex);
419 	em_debug_remove_pd(dev);
420 
421 	kfree(dev->em_pd->table);
422 	kfree(dev->em_pd);
423 	dev->em_pd = NULL;
424 	mutex_unlock(&em_pd_mutex);
425 }
426 EXPORT_SYMBOL_GPL(em_dev_unregister_perf_domain);
427