xref: /linux/kernel/power/energy_model.c (revision c532de5a67a70f8533d495f8f2aaa9a0491c3ad0)
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 void em_cpufreq_update_efficiencies(struct device *dev,
27 					   struct em_perf_state *table);
28 static void em_check_capacity_update(void);
29 static void em_update_workfn(struct work_struct *work);
30 static DECLARE_DELAYED_WORK(em_update_work, em_update_workfn);
31 
32 static bool _is_cpu_device(struct device *dev)
33 {
34 	return (dev->bus == &cpu_subsys);
35 }
36 
37 #ifdef CONFIG_DEBUG_FS
38 static struct dentry *rootdir;
39 
40 struct em_dbg_info {
41 	struct em_perf_domain *pd;
42 	int ps_id;
43 };
44 
45 #define DEFINE_EM_DBG_SHOW(name, fname)					\
46 static int em_debug_##fname##_show(struct seq_file *s, void *unused)	\
47 {									\
48 	struct em_dbg_info *em_dbg = s->private;			\
49 	struct em_perf_state *table;					\
50 	unsigned long val;						\
51 									\
52 	rcu_read_lock();						\
53 	table = em_perf_state_from_pd(em_dbg->pd);			\
54 	val = table[em_dbg->ps_id].name;				\
55 	rcu_read_unlock();						\
56 									\
57 	seq_printf(s, "%lu\n", val);					\
58 	return 0;							\
59 }									\
60 DEFINE_SHOW_ATTRIBUTE(em_debug_##fname)
61 
62 DEFINE_EM_DBG_SHOW(frequency, frequency);
63 DEFINE_EM_DBG_SHOW(power, power);
64 DEFINE_EM_DBG_SHOW(cost, cost);
65 DEFINE_EM_DBG_SHOW(performance, performance);
66 DEFINE_EM_DBG_SHOW(flags, inefficiency);
67 
68 static void em_debug_create_ps(struct em_perf_domain *em_pd,
69 			       struct em_dbg_info *em_dbg, int i,
70 			       struct dentry *pd)
71 {
72 	struct em_perf_state *table;
73 	unsigned long freq;
74 	struct dentry *d;
75 	char name[24];
76 
77 	em_dbg[i].pd = em_pd;
78 	em_dbg[i].ps_id = i;
79 
80 	rcu_read_lock();
81 	table = em_perf_state_from_pd(em_pd);
82 	freq = table[i].frequency;
83 	rcu_read_unlock();
84 
85 	snprintf(name, sizeof(name), "ps:%lu", freq);
86 
87 	/* Create per-ps directory */
88 	d = debugfs_create_dir(name, pd);
89 	debugfs_create_file("frequency", 0444, d, &em_dbg[i],
90 			    &em_debug_frequency_fops);
91 	debugfs_create_file("power", 0444, d, &em_dbg[i],
92 			    &em_debug_power_fops);
93 	debugfs_create_file("cost", 0444, d, &em_dbg[i],
94 			    &em_debug_cost_fops);
95 	debugfs_create_file("performance", 0444, d, &em_dbg[i],
96 			    &em_debug_performance_fops);
97 	debugfs_create_file("inefficient", 0444, d, &em_dbg[i],
98 			    &em_debug_inefficiency_fops);
99 }
100 
101 static int em_debug_cpus_show(struct seq_file *s, void *unused)
102 {
103 	seq_printf(s, "%*pbl\n", cpumask_pr_args(to_cpumask(s->private)));
104 
105 	return 0;
106 }
107 DEFINE_SHOW_ATTRIBUTE(em_debug_cpus);
108 
109 static int em_debug_flags_show(struct seq_file *s, void *unused)
110 {
111 	struct em_perf_domain *pd = s->private;
112 
113 	seq_printf(s, "%#lx\n", pd->flags);
114 
115 	return 0;
116 }
117 DEFINE_SHOW_ATTRIBUTE(em_debug_flags);
118 
119 static void em_debug_create_pd(struct device *dev)
120 {
121 	struct em_dbg_info *em_dbg;
122 	struct dentry *d;
123 	int i;
124 
125 	/* Create the directory of the performance domain */
126 	d = debugfs_create_dir(dev_name(dev), rootdir);
127 
128 	if (_is_cpu_device(dev))
129 		debugfs_create_file("cpus", 0444, d, dev->em_pd->cpus,
130 				    &em_debug_cpus_fops);
131 
132 	debugfs_create_file("flags", 0444, d, dev->em_pd,
133 			    &em_debug_flags_fops);
134 
135 	em_dbg = devm_kcalloc(dev, dev->em_pd->nr_perf_states,
136 			      sizeof(*em_dbg), GFP_KERNEL);
137 	if (!em_dbg)
138 		return;
139 
140 	/* Create a sub-directory for each performance state */
141 	for (i = 0; i < dev->em_pd->nr_perf_states; i++)
142 		em_debug_create_ps(dev->em_pd, em_dbg, i, d);
143 
144 }
145 
146 static void em_debug_remove_pd(struct device *dev)
147 {
148 	debugfs_lookup_and_remove(dev_name(dev), rootdir);
149 }
150 
151 static int __init em_debug_init(void)
152 {
153 	/* Create /sys/kernel/debug/energy_model directory */
154 	rootdir = debugfs_create_dir("energy_model", NULL);
155 
156 	return 0;
157 }
158 fs_initcall(em_debug_init);
159 #else /* CONFIG_DEBUG_FS */
160 static void em_debug_create_pd(struct device *dev) {}
161 static void em_debug_remove_pd(struct device *dev) {}
162 #endif
163 
164 static void em_destroy_table_rcu(struct rcu_head *rp)
165 {
166 	struct em_perf_table __rcu *table;
167 
168 	table = container_of(rp, struct em_perf_table, rcu);
169 	kfree(table);
170 }
171 
172 static void em_release_table_kref(struct kref *kref)
173 {
174 	struct em_perf_table __rcu *table;
175 
176 	/* It was the last owner of this table so we can free */
177 	table = container_of(kref, struct em_perf_table, kref);
178 
179 	call_rcu(&table->rcu, em_destroy_table_rcu);
180 }
181 
182 /**
183  * em_table_free() - Handles safe free of the EM table when needed
184  * @table : EM table which is going to be freed
185  *
186  * No return values.
187  */
188 void em_table_free(struct em_perf_table __rcu *table)
189 {
190 	kref_put(&table->kref, em_release_table_kref);
191 }
192 
193 /**
194  * em_table_alloc() - Allocate a new EM table
195  * @pd		: EM performance domain for which this must be done
196  *
197  * Allocate a new EM table and initialize its kref to indicate that it
198  * has a user.
199  * Returns allocated table or NULL.
200  */
201 struct em_perf_table __rcu *em_table_alloc(struct em_perf_domain *pd)
202 {
203 	struct em_perf_table __rcu *table;
204 	int table_size;
205 
206 	table_size = sizeof(struct em_perf_state) * pd->nr_perf_states;
207 
208 	table = kzalloc(sizeof(*table) + table_size, GFP_KERNEL);
209 	if (!table)
210 		return NULL;
211 
212 	kref_init(&table->kref);
213 
214 	return table;
215 }
216 
217 static void em_init_performance(struct device *dev, struct em_perf_domain *pd,
218 				struct em_perf_state *table, int nr_states)
219 {
220 	u64 fmax, max_cap;
221 	int i, cpu;
222 
223 	/* This is needed only for CPUs and EAS skip other devices */
224 	if (!_is_cpu_device(dev))
225 		return;
226 
227 	cpu = cpumask_first(em_span_cpus(pd));
228 
229 	/*
230 	 * Calculate the performance value for each frequency with
231 	 * linear relationship. The final CPU capacity might not be ready at
232 	 * boot time, but the EM will be updated a bit later with correct one.
233 	 */
234 	fmax = (u64) table[nr_states - 1].frequency;
235 	max_cap = (u64) arch_scale_cpu_capacity(cpu);
236 	for (i = 0; i < nr_states; i++)
237 		table[i].performance = div64_u64(max_cap * table[i].frequency,
238 						 fmax);
239 }
240 
241 static int em_compute_costs(struct device *dev, struct em_perf_state *table,
242 			    struct em_data_callback *cb, int nr_states,
243 			    unsigned long flags)
244 {
245 	unsigned long prev_cost = ULONG_MAX;
246 	int i, ret;
247 
248 	/* Compute the cost of each performance state. */
249 	for (i = nr_states - 1; i >= 0; i--) {
250 		unsigned long power_res, cost;
251 
252 		if ((flags & EM_PERF_DOMAIN_ARTIFICIAL) && cb->get_cost) {
253 			ret = cb->get_cost(dev, table[i].frequency, &cost);
254 			if (ret || !cost || cost > EM_MAX_POWER) {
255 				dev_err(dev, "EM: invalid cost %lu %d\n",
256 					cost, ret);
257 				return -EINVAL;
258 			}
259 		} else {
260 			/* increase resolution of 'cost' precision */
261 			power_res = table[i].power * 10;
262 			cost = power_res / table[i].performance;
263 		}
264 
265 		table[i].cost = cost;
266 
267 		if (table[i].cost >= prev_cost) {
268 			table[i].flags = EM_PERF_STATE_INEFFICIENT;
269 			dev_dbg(dev, "EM: OPP:%lu is inefficient\n",
270 				table[i].frequency);
271 		} else {
272 			prev_cost = table[i].cost;
273 		}
274 	}
275 
276 	return 0;
277 }
278 
279 /**
280  * em_dev_compute_costs() - Calculate cost values for new runtime EM table
281  * @dev		: Device for which the EM table is to be updated
282  * @table	: The new EM table that is going to get the costs calculated
283  * @nr_states	: Number of performance states
284  *
285  * Calculate the em_perf_state::cost values for new runtime EM table. The
286  * values are used for EAS during task placement. It also calculates and sets
287  * the efficiency flag for each performance state. When the function finish
288  * successfully the EM table is ready to be updated and used by EAS.
289  *
290  * Return 0 on success or a proper error in case of failure.
291  */
292 int em_dev_compute_costs(struct device *dev, struct em_perf_state *table,
293 			 int nr_states)
294 {
295 	return em_compute_costs(dev, table, NULL, nr_states, 0);
296 }
297 
298 /**
299  * em_dev_update_perf_domain() - Update runtime EM table for a device
300  * @dev		: Device for which the EM is to be updated
301  * @new_table	: The new EM table that is going to be used from now
302  *
303  * Update EM runtime modifiable table for the @dev using the provided @table.
304  *
305  * This function uses a mutex to serialize writers, so it must not be called
306  * from a non-sleeping context.
307  *
308  * Return 0 on success or an error code on failure.
309  */
310 int em_dev_update_perf_domain(struct device *dev,
311 			      struct em_perf_table __rcu *new_table)
312 {
313 	struct em_perf_table __rcu *old_table;
314 	struct em_perf_domain *pd;
315 
316 	if (!dev)
317 		return -EINVAL;
318 
319 	/* Serialize update/unregister or concurrent updates */
320 	mutex_lock(&em_pd_mutex);
321 
322 	if (!dev->em_pd) {
323 		mutex_unlock(&em_pd_mutex);
324 		return -EINVAL;
325 	}
326 	pd = dev->em_pd;
327 
328 	kref_get(&new_table->kref);
329 
330 	old_table = pd->em_table;
331 	rcu_assign_pointer(pd->em_table, new_table);
332 
333 	em_cpufreq_update_efficiencies(dev, new_table->state);
334 
335 	em_table_free(old_table);
336 
337 	mutex_unlock(&em_pd_mutex);
338 	return 0;
339 }
340 EXPORT_SYMBOL_GPL(em_dev_update_perf_domain);
341 
342 static int em_create_perf_table(struct device *dev, struct em_perf_domain *pd,
343 				struct em_perf_state *table,
344 				struct em_data_callback *cb,
345 				unsigned long flags)
346 {
347 	unsigned long power, freq, prev_freq = 0;
348 	int nr_states = pd->nr_perf_states;
349 	int i, ret;
350 
351 	/* Build the list of performance states for this performance domain */
352 	for (i = 0, freq = 0; i < nr_states; i++, freq++) {
353 		/*
354 		 * active_power() is a driver callback which ceils 'freq' to
355 		 * lowest performance state of 'dev' above 'freq' and updates
356 		 * 'power' and 'freq' accordingly.
357 		 */
358 		ret = cb->active_power(dev, &power, &freq);
359 		if (ret) {
360 			dev_err(dev, "EM: invalid perf. state: %d\n",
361 				ret);
362 			return -EINVAL;
363 		}
364 
365 		/*
366 		 * We expect the driver callback to increase the frequency for
367 		 * higher performance states.
368 		 */
369 		if (freq <= prev_freq) {
370 			dev_err(dev, "EM: non-increasing freq: %lu\n",
371 				freq);
372 			return -EINVAL;
373 		}
374 
375 		/*
376 		 * The power returned by active_state() is expected to be
377 		 * positive and be in range.
378 		 */
379 		if (!power || power > EM_MAX_POWER) {
380 			dev_err(dev, "EM: invalid power: %lu\n",
381 				power);
382 			return -EINVAL;
383 		}
384 
385 		table[i].power = power;
386 		table[i].frequency = prev_freq = freq;
387 	}
388 
389 	em_init_performance(dev, pd, table, nr_states);
390 
391 	ret = em_compute_costs(dev, table, cb, nr_states, flags);
392 	if (ret)
393 		return -EINVAL;
394 
395 	return 0;
396 }
397 
398 static int em_create_pd(struct device *dev, int nr_states,
399 			struct em_data_callback *cb, cpumask_t *cpus,
400 			unsigned long flags)
401 {
402 	struct em_perf_table __rcu *em_table;
403 	struct em_perf_domain *pd;
404 	struct device *cpu_dev;
405 	int cpu, ret, num_cpus;
406 
407 	if (_is_cpu_device(dev)) {
408 		num_cpus = cpumask_weight(cpus);
409 
410 		/* Prevent max possible energy calculation to not overflow */
411 		if (num_cpus > EM_MAX_NUM_CPUS) {
412 			dev_err(dev, "EM: too many CPUs, overflow possible\n");
413 			return -EINVAL;
414 		}
415 
416 		pd = kzalloc(sizeof(*pd) + cpumask_size(), GFP_KERNEL);
417 		if (!pd)
418 			return -ENOMEM;
419 
420 		cpumask_copy(em_span_cpus(pd), cpus);
421 	} else {
422 		pd = kzalloc(sizeof(*pd), GFP_KERNEL);
423 		if (!pd)
424 			return -ENOMEM;
425 	}
426 
427 	pd->nr_perf_states = nr_states;
428 
429 	em_table = em_table_alloc(pd);
430 	if (!em_table)
431 		goto free_pd;
432 
433 	ret = em_create_perf_table(dev, pd, em_table->state, cb, flags);
434 	if (ret)
435 		goto free_pd_table;
436 
437 	rcu_assign_pointer(pd->em_table, em_table);
438 
439 	if (_is_cpu_device(dev))
440 		for_each_cpu(cpu, cpus) {
441 			cpu_dev = get_cpu_device(cpu);
442 			cpu_dev->em_pd = pd;
443 		}
444 
445 	dev->em_pd = pd;
446 
447 	return 0;
448 
449 free_pd_table:
450 	kfree(em_table);
451 free_pd:
452 	kfree(pd);
453 	return -EINVAL;
454 }
455 
456 static void
457 em_cpufreq_update_efficiencies(struct device *dev, struct em_perf_state *table)
458 {
459 	struct em_perf_domain *pd = dev->em_pd;
460 	struct cpufreq_policy *policy;
461 	int found = 0;
462 	int i, cpu;
463 
464 	if (!_is_cpu_device(dev))
465 		return;
466 
467 	/* Try to get a CPU which is active and in this PD */
468 	cpu = cpumask_first_and(em_span_cpus(pd), cpu_active_mask);
469 	if (cpu >= nr_cpu_ids) {
470 		dev_warn(dev, "EM: No online CPU for CPUFreq policy\n");
471 		return;
472 	}
473 
474 	policy = cpufreq_cpu_get(cpu);
475 	if (!policy) {
476 		dev_warn(dev, "EM: Access to CPUFreq policy failed\n");
477 		return;
478 	}
479 
480 	for (i = 0; i < pd->nr_perf_states; i++) {
481 		if (!(table[i].flags & EM_PERF_STATE_INEFFICIENT))
482 			continue;
483 
484 		if (!cpufreq_table_set_inefficient(policy, table[i].frequency))
485 			found++;
486 	}
487 
488 	cpufreq_cpu_put(policy);
489 
490 	if (!found)
491 		return;
492 
493 	/*
494 	 * Efficiencies have been installed in CPUFreq, inefficient frequencies
495 	 * will be skipped. The EM can do the same.
496 	 */
497 	pd->flags |= EM_PERF_DOMAIN_SKIP_INEFFICIENCIES;
498 }
499 
500 /**
501  * em_pd_get() - Return the performance domain for a device
502  * @dev : Device to find the performance domain for
503  *
504  * Returns the performance domain to which @dev belongs, or NULL if it doesn't
505  * exist.
506  */
507 struct em_perf_domain *em_pd_get(struct device *dev)
508 {
509 	if (IS_ERR_OR_NULL(dev))
510 		return NULL;
511 
512 	return dev->em_pd;
513 }
514 EXPORT_SYMBOL_GPL(em_pd_get);
515 
516 /**
517  * em_cpu_get() - Return the performance domain for a CPU
518  * @cpu : CPU to find the performance domain for
519  *
520  * Returns the performance domain to which @cpu belongs, or NULL if it doesn't
521  * exist.
522  */
523 struct em_perf_domain *em_cpu_get(int cpu)
524 {
525 	struct device *cpu_dev;
526 
527 	cpu_dev = get_cpu_device(cpu);
528 	if (!cpu_dev)
529 		return NULL;
530 
531 	return em_pd_get(cpu_dev);
532 }
533 EXPORT_SYMBOL_GPL(em_cpu_get);
534 
535 /**
536  * em_dev_register_perf_domain() - Register the Energy Model (EM) for a device
537  * @dev		: Device for which the EM is to register
538  * @nr_states	: Number of performance states to register
539  * @cb		: Callback functions providing the data of the Energy Model
540  * @cpus	: Pointer to cpumask_t, which in case of a CPU device is
541  *		obligatory. It can be taken from i.e. 'policy->cpus'. For other
542  *		type of devices this should be set to NULL.
543  * @microwatts	: Flag indicating that the power values are in micro-Watts or
544  *		in some other scale. It must be set properly.
545  *
546  * Create Energy Model tables for a performance domain using the callbacks
547  * defined in cb.
548  *
549  * The @microwatts is important to set with correct value. Some kernel
550  * sub-systems might rely on this flag and check if all devices in the EM are
551  * using the same scale.
552  *
553  * If multiple clients register the same performance domain, all but the first
554  * registration will be ignored.
555  *
556  * Return 0 on success
557  */
558 int em_dev_register_perf_domain(struct device *dev, unsigned int nr_states,
559 				struct em_data_callback *cb, cpumask_t *cpus,
560 				bool microwatts)
561 {
562 	unsigned long cap, prev_cap = 0;
563 	unsigned long flags = 0;
564 	int cpu, ret;
565 
566 	if (!dev || !nr_states || !cb)
567 		return -EINVAL;
568 
569 	/*
570 	 * Use a mutex to serialize the registration of performance domains and
571 	 * let the driver-defined callback functions sleep.
572 	 */
573 	mutex_lock(&em_pd_mutex);
574 
575 	if (dev->em_pd) {
576 		ret = -EEXIST;
577 		goto unlock;
578 	}
579 
580 	if (_is_cpu_device(dev)) {
581 		if (!cpus) {
582 			dev_err(dev, "EM: invalid CPU mask\n");
583 			ret = -EINVAL;
584 			goto unlock;
585 		}
586 
587 		for_each_cpu(cpu, cpus) {
588 			if (em_cpu_get(cpu)) {
589 				dev_err(dev, "EM: exists for CPU%d\n", cpu);
590 				ret = -EEXIST;
591 				goto unlock;
592 			}
593 			/*
594 			 * All CPUs of a domain must have the same
595 			 * micro-architecture since they all share the same
596 			 * table.
597 			 */
598 			cap = arch_scale_cpu_capacity(cpu);
599 			if (prev_cap && prev_cap != cap) {
600 				dev_err(dev, "EM: CPUs of %*pbl must have the same capacity\n",
601 					cpumask_pr_args(cpus));
602 
603 				ret = -EINVAL;
604 				goto unlock;
605 			}
606 			prev_cap = cap;
607 		}
608 	}
609 
610 	if (microwatts)
611 		flags |= EM_PERF_DOMAIN_MICROWATTS;
612 	else if (cb->get_cost)
613 		flags |= EM_PERF_DOMAIN_ARTIFICIAL;
614 
615 	/*
616 	 * EM only supports uW (exception is artificial EM).
617 	 * Therefore, check and force the drivers to provide
618 	 * power in uW.
619 	 */
620 	if (!microwatts && !(flags & EM_PERF_DOMAIN_ARTIFICIAL)) {
621 		dev_err(dev, "EM: only supports uW power values\n");
622 		ret = -EINVAL;
623 		goto unlock;
624 	}
625 
626 	ret = em_create_pd(dev, nr_states, cb, cpus, flags);
627 	if (ret)
628 		goto unlock;
629 
630 	dev->em_pd->flags |= flags;
631 
632 	em_cpufreq_update_efficiencies(dev, dev->em_pd->em_table->state);
633 
634 	em_debug_create_pd(dev);
635 	dev_info(dev, "EM: created perf domain\n");
636 
637 unlock:
638 	mutex_unlock(&em_pd_mutex);
639 
640 	if (_is_cpu_device(dev))
641 		em_check_capacity_update();
642 
643 	return ret;
644 }
645 EXPORT_SYMBOL_GPL(em_dev_register_perf_domain);
646 
647 /**
648  * em_dev_unregister_perf_domain() - Unregister Energy Model (EM) for a device
649  * @dev		: Device for which the EM is registered
650  *
651  * Unregister the EM for the specified @dev (but not a CPU device).
652  */
653 void em_dev_unregister_perf_domain(struct device *dev)
654 {
655 	if (IS_ERR_OR_NULL(dev) || !dev->em_pd)
656 		return;
657 
658 	if (_is_cpu_device(dev))
659 		return;
660 
661 	/*
662 	 * The mutex separates all register/unregister requests and protects
663 	 * from potential clean-up/setup issues in the debugfs directories.
664 	 * The debugfs directory name is the same as device's name.
665 	 */
666 	mutex_lock(&em_pd_mutex);
667 	em_debug_remove_pd(dev);
668 
669 	em_table_free(dev->em_pd->em_table);
670 
671 	kfree(dev->em_pd);
672 	dev->em_pd = NULL;
673 	mutex_unlock(&em_pd_mutex);
674 }
675 EXPORT_SYMBOL_GPL(em_dev_unregister_perf_domain);
676 
677 static struct em_perf_table __rcu *em_table_dup(struct em_perf_domain *pd)
678 {
679 	struct em_perf_table __rcu *em_table;
680 	struct em_perf_state *ps, *new_ps;
681 	int ps_size;
682 
683 	em_table = em_table_alloc(pd);
684 	if (!em_table)
685 		return NULL;
686 
687 	new_ps = em_table->state;
688 
689 	rcu_read_lock();
690 	ps = em_perf_state_from_pd(pd);
691 	/* Initialize data based on old table */
692 	ps_size = sizeof(struct em_perf_state) * pd->nr_perf_states;
693 	memcpy(new_ps, ps, ps_size);
694 
695 	rcu_read_unlock();
696 
697 	return em_table;
698 }
699 
700 static int em_recalc_and_update(struct device *dev, struct em_perf_domain *pd,
701 				struct em_perf_table __rcu *em_table)
702 {
703 	int ret;
704 
705 	ret = em_compute_costs(dev, em_table->state, NULL, pd->nr_perf_states,
706 			       pd->flags);
707 	if (ret)
708 		goto free_em_table;
709 
710 	ret = em_dev_update_perf_domain(dev, em_table);
711 	if (ret)
712 		goto free_em_table;
713 
714 	/*
715 	 * This is one-time-update, so give up the ownership in this updater.
716 	 * The EM framework has incremented the usage counter and from now
717 	 * will keep the reference (then free the memory when needed).
718 	 */
719 free_em_table:
720 	em_table_free(em_table);
721 	return ret;
722 }
723 
724 /*
725  * Adjustment of CPU performance values after boot, when all CPUs capacites
726  * are correctly calculated.
727  */
728 static void em_adjust_new_capacity(struct device *dev,
729 				   struct em_perf_domain *pd,
730 				   u64 max_cap)
731 {
732 	struct em_perf_table __rcu *em_table;
733 
734 	em_table = em_table_dup(pd);
735 	if (!em_table) {
736 		dev_warn(dev, "EM: allocation failed\n");
737 		return;
738 	}
739 
740 	em_init_performance(dev, pd, em_table->state, pd->nr_perf_states);
741 
742 	em_recalc_and_update(dev, pd, em_table);
743 }
744 
745 static void em_check_capacity_update(void)
746 {
747 	cpumask_var_t cpu_done_mask;
748 	struct em_perf_state *table;
749 	struct em_perf_domain *pd;
750 	unsigned long cpu_capacity;
751 	int cpu;
752 
753 	if (!zalloc_cpumask_var(&cpu_done_mask, GFP_KERNEL)) {
754 		pr_warn("no free memory\n");
755 		return;
756 	}
757 
758 	/* Check if CPUs capacity has changed than update EM */
759 	for_each_possible_cpu(cpu) {
760 		struct cpufreq_policy *policy;
761 		unsigned long em_max_perf;
762 		struct device *dev;
763 
764 		if (cpumask_test_cpu(cpu, cpu_done_mask))
765 			continue;
766 
767 		policy = cpufreq_cpu_get(cpu);
768 		if (!policy) {
769 			pr_debug("Accessing cpu%d policy failed\n", cpu);
770 			schedule_delayed_work(&em_update_work,
771 					      msecs_to_jiffies(1000));
772 			break;
773 		}
774 		cpufreq_cpu_put(policy);
775 
776 		pd = em_cpu_get(cpu);
777 		if (!pd || em_is_artificial(pd))
778 			continue;
779 
780 		cpumask_or(cpu_done_mask, cpu_done_mask,
781 			   em_span_cpus(pd));
782 
783 		cpu_capacity = arch_scale_cpu_capacity(cpu);
784 
785 		rcu_read_lock();
786 		table = em_perf_state_from_pd(pd);
787 		em_max_perf = table[pd->nr_perf_states - 1].performance;
788 		rcu_read_unlock();
789 
790 		/*
791 		 * Check if the CPU capacity has been adjusted during boot
792 		 * and trigger the update for new performance values.
793 		 */
794 		if (em_max_perf == cpu_capacity)
795 			continue;
796 
797 		pr_debug("updating cpu%d cpu_cap=%lu old capacity=%lu\n",
798 			 cpu, cpu_capacity, em_max_perf);
799 
800 		dev = get_cpu_device(cpu);
801 		em_adjust_new_capacity(dev, pd, cpu_capacity);
802 	}
803 
804 	free_cpumask_var(cpu_done_mask);
805 }
806 
807 static void em_update_workfn(struct work_struct *work)
808 {
809 	em_check_capacity_update();
810 }
811 
812 /**
813  * em_dev_update_chip_binning() - Update Energy Model after the new voltage
814  *				information is present in the OPPs.
815  * @dev		: Device for which the Energy Model has to be updated.
816  *
817  * This function allows to update easily the EM with new values available in
818  * the OPP framework and DT. It can be used after the chip has been properly
819  * verified by device drivers and the voltages adjusted for the 'chip binning'.
820  */
821 int em_dev_update_chip_binning(struct device *dev)
822 {
823 	struct em_perf_table __rcu *em_table;
824 	struct em_perf_domain *pd;
825 	int i, ret;
826 
827 	if (IS_ERR_OR_NULL(dev))
828 		return -EINVAL;
829 
830 	pd = em_pd_get(dev);
831 	if (!pd) {
832 		dev_warn(dev, "Couldn't find Energy Model\n");
833 		return -EINVAL;
834 	}
835 
836 	em_table = em_table_dup(pd);
837 	if (!em_table) {
838 		dev_warn(dev, "EM: allocation failed\n");
839 		return -ENOMEM;
840 	}
841 
842 	/* Update power values which might change due to new voltage in OPPs */
843 	for (i = 0; i < pd->nr_perf_states; i++) {
844 		unsigned long freq = em_table->state[i].frequency;
845 		unsigned long power;
846 
847 		ret = dev_pm_opp_calc_power(dev, &power, &freq);
848 		if (ret) {
849 			em_table_free(em_table);
850 			return ret;
851 		}
852 
853 		em_table->state[i].power = power;
854 	}
855 
856 	return em_recalc_and_update(dev, pd, em_table);
857 }
858 EXPORT_SYMBOL_GPL(em_dev_update_chip_binning);
859