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 */
 23static DEFINE_MUTEX(em_pd_mutex);
 24
 25static bool _is_cpu_device(struct device *dev)
 26{
 27	return (dev->bus == &cpu_subsys);
 28}
 29
 30#ifdef CONFIG_DEBUG_FS
 31static struct dentry *rootdir;
 32
 33static 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
 47static 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}
 53DEFINE_SHOW_ATTRIBUTE(em_debug_cpus);
 54
 55static 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
 73static 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
 81static 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}
 88core_initcall(em_debug_init);
 89#else /* CONFIG_DEBUG_FS */
 90static void em_debug_create_pd(struct device *dev) {}
 91static void em_debug_remove_pd(struct device *dev) {}
 92#endif
 93
 94static 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 opp_eff, prev_opp_eff = ULONG_MAX;
 98	unsigned long power, freq, prev_freq = 0;
 99	struct em_perf_state *table;
100	int i, ret;
101	u64 fmax;
102
103	table = kcalloc(nr_states, sizeof(*table), GFP_KERNEL);
104	if (!table)
105		return -ENOMEM;
106
107	/* Build the list of performance states for this performance domain */
108	for (i = 0, freq = 0; i < nr_states; i++, freq++) {
109		/*
110		 * active_power() is a driver callback which ceils 'freq' to
111		 * lowest performance state of 'dev' above 'freq' and updates
112		 * 'power' and 'freq' accordingly.
113		 */
114		ret = cb->active_power(&power, &freq, dev);
115		if (ret) {
116			dev_err(dev, "EM: invalid perf. state: %d\n",
117				ret);
118			goto free_ps_table;
119		}
120
121		/*
122		 * We expect the driver callback to increase the frequency for
123		 * higher performance states.
124		 */
125		if (freq <= prev_freq) {
126			dev_err(dev, "EM: non-increasing freq: %lu\n",
127				freq);
128			goto free_ps_table;
129		}
130
131		/*
132		 * The power returned by active_state() is expected to be
133		 * positive, in milli-watts and to fit into 16 bits.
134		 */
135		if (!power || power > EM_MAX_POWER) {
136			dev_err(dev, "EM: invalid power: %lu\n",
137				power);
138			goto free_ps_table;
139		}
140
141		table[i].power = power;
142		table[i].frequency = prev_freq = freq;
143
144		/*
145		 * The hertz/watts efficiency ratio should decrease as the
146		 * frequency grows on sane platforms. But this isn't always
147		 * true in practice so warn the user if a higher OPP is more
148		 * power efficient than a lower one.
149		 */
150		opp_eff = freq / power;
151		if (opp_eff >= prev_opp_eff)
152			dev_dbg(dev, "EM: hertz/watts ratio non-monotonically decreasing: em_perf_state %d >= em_perf_state%d\n",
153					i, i - 1);
154		prev_opp_eff = opp_eff;
155	}
156
157	/* Compute the cost of each performance state. */
158	fmax = (u64) table[nr_states - 1].frequency;
159	for (i = 0; i < nr_states; i++) {
160		table[i].cost = div64_u64(fmax * table[i].power,
161					  table[i].frequency);
162	}
163
164	pd->table = table;
165	pd->nr_perf_states = nr_states;
166
167	return 0;
168
169free_ps_table:
170	kfree(table);
171	return -EINVAL;
172}
173
174static int em_create_pd(struct device *dev, int nr_states,
175			struct em_data_callback *cb, cpumask_t *cpus)
176{
177	struct em_perf_domain *pd;
178	struct device *cpu_dev;
179	int cpu, ret;
180
181	if (_is_cpu_device(dev)) {
182		pd = kzalloc(sizeof(*pd) + cpumask_size(), GFP_KERNEL);
183		if (!pd)
184			return -ENOMEM;
185
186		cpumask_copy(em_span_cpus(pd), cpus);
187	} else {
188		pd = kzalloc(sizeof(*pd), GFP_KERNEL);
189		if (!pd)
190			return -ENOMEM;
191	}
192
193	ret = em_create_perf_table(dev, pd, nr_states, cb);
194	if (ret) {
195		kfree(pd);
196		return ret;
197	}
198
199	if (_is_cpu_device(dev))
200		for_each_cpu(cpu, cpus) {
201			cpu_dev = get_cpu_device(cpu);
202			cpu_dev->em_pd = pd;
203		}
204
205	dev->em_pd = pd;
206
207	return 0;
208}
209
210/**
211 * em_pd_get() - Return the performance domain for a device
212 * @dev : Device to find the performance domain for
213 *
214 * Returns the performance domain to which @dev belongs, or NULL if it doesn't
215 * exist.
216 */
217struct em_perf_domain *em_pd_get(struct device *dev)
218{
219	if (IS_ERR_OR_NULL(dev))
220		return NULL;
221
222	return dev->em_pd;
223}
224EXPORT_SYMBOL_GPL(em_pd_get);
225
226/**
227 * em_cpu_get() - Return the performance domain for a CPU
228 * @cpu : CPU to find the performance domain for
229 *
230 * Returns the performance domain to which @cpu belongs, or NULL if it doesn't
231 * exist.
232 */
233struct em_perf_domain *em_cpu_get(int cpu)
234{
235	struct device *cpu_dev;
236
237	cpu_dev = get_cpu_device(cpu);
238	if (!cpu_dev)
239		return NULL;
240
241	return em_pd_get(cpu_dev);
242}
243EXPORT_SYMBOL_GPL(em_cpu_get);
244
245/**
246 * em_dev_register_perf_domain() - Register the Energy Model (EM) for a device
247 * @dev		: Device for which the EM is to register
248 * @nr_states	: Number of performance states to register
249 * @cb		: Callback functions providing the data of the Energy Model
250 * @cpus	: Pointer to cpumask_t, which in case of a CPU device is
251 *		obligatory. It can be taken from i.e. 'policy->cpus'. For other
252 *		type of devices this should be set to NULL.
253 *
254 * Create Energy Model tables for a performance domain using the callbacks
255 * defined in cb.
256 *
257 * If multiple clients register the same performance domain, all but the first
258 * registration will be ignored.
259 *
260 * Return 0 on success
261 */
262int em_dev_register_perf_domain(struct device *dev, unsigned int nr_states,
263				struct em_data_callback *cb, cpumask_t *cpus)
264{
265	unsigned long cap, prev_cap = 0;
266	int cpu, ret;
267
268	if (!dev || !nr_states || !cb)
269		return -EINVAL;
270
271	/*
272	 * Use a mutex to serialize the registration of performance domains and
273	 * let the driver-defined callback functions sleep.
274	 */
275	mutex_lock(&em_pd_mutex);
276
277	if (dev->em_pd) {
278		ret = -EEXIST;
279		goto unlock;
280	}
281
282	if (_is_cpu_device(dev)) {
283		if (!cpus) {
284			dev_err(dev, "EM: invalid CPU mask\n");
285			ret = -EINVAL;
286			goto unlock;
287		}
288
289		for_each_cpu(cpu, cpus) {
290			if (em_cpu_get(cpu)) {
291				dev_err(dev, "EM: exists for CPU%d\n", cpu);
292				ret = -EEXIST;
293				goto unlock;
294			}
295			/*
296			 * All CPUs of a domain must have the same
297			 * micro-architecture since they all share the same
298			 * table.
299			 */
300			cap = arch_scale_cpu_capacity(cpu);
301			if (prev_cap && prev_cap != cap) {
302				dev_err(dev, "EM: CPUs of %*pbl must have the same capacity\n",
303					cpumask_pr_args(cpus));
304
305				ret = -EINVAL;
306				goto unlock;
307			}
308			prev_cap = cap;
309		}
310	}
311
312	ret = em_create_pd(dev, nr_states, cb, cpus);
313	if (ret)
314		goto unlock;
315
316	em_debug_create_pd(dev);
317	dev_info(dev, "EM: created perf domain\n");
318
319unlock:
320	mutex_unlock(&em_pd_mutex);
321	return ret;
322}
323EXPORT_SYMBOL_GPL(em_dev_register_perf_domain);
324
325/**
326 * em_dev_unregister_perf_domain() - Unregister Energy Model (EM) for a device
327 * @dev		: Device for which the EM is registered
328 *
329 * Unregister the EM for the specified @dev (but not a CPU device).
330 */
331void em_dev_unregister_perf_domain(struct device *dev)
332{
333	if (IS_ERR_OR_NULL(dev) || !dev->em_pd)
334		return;
335
336	if (_is_cpu_device(dev))
337		return;
338
339	/*
340	 * The mutex separates all register/unregister requests and protects
341	 * from potential clean-up/setup issues in the debugfs directories.
342	 * The debugfs directory name is the same as device's name.
343	 */
344	mutex_lock(&em_pd_mutex);
345	em_debug_remove_pd(dev);
346
347	kfree(dev->em_pd->table);
348	kfree(dev->em_pd);
349	dev->em_pd = NULL;
350	mutex_unlock(&em_pd_mutex);
351}
352EXPORT_SYMBOL_GPL(em_dev_unregister_perf_domain);