1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 *  drivers/cpufreq/cpufreq_stats.c
4 *
5 *  Copyright (C) 2003-2004 Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>.
6 *  (C) 2004 Zou Nan hai <nanhai.zou@intel.com>.
7 */
8
9#include <linux/cpu.h>
10#include <linux/cpufreq.h>
11#include <linux/module.h>
12#include <linux/sched/clock.h>
13#include <linux/slab.h>
14
15struct cpufreq_stats {
16	unsigned int total_trans;
17	unsigned long long last_time;
18	unsigned int max_state;
19	unsigned int state_num;
20	unsigned int last_index;
21	u64 *time_in_state;
22	unsigned int *freq_table;
23	unsigned int *trans_table;
24
25	/* Deferred reset */
26	unsigned int reset_pending;
27	unsigned long long reset_time;
28};
29
30static void cpufreq_stats_update(struct cpufreq_stats *stats,
31				 unsigned long long time)
32{
33	unsigned long long cur_time = local_clock();
34
35	stats->time_in_state[stats->last_index] += cur_time - time;
36	stats->last_time = cur_time;
37}
38
39static void cpufreq_stats_reset_table(struct cpufreq_stats *stats)
40{
41	unsigned int count = stats->max_state;
42
43	memset(stats->time_in_state, 0, count * sizeof(u64));
44	memset(stats->trans_table, 0, count * count * sizeof(int));
45	stats->last_time = local_clock();
46	stats->total_trans = 0;
47
48	/* Adjust for the time elapsed since reset was requested */
49	WRITE_ONCE(stats->reset_pending, 0);
50	/*
51	 * Prevent the reset_time read from being reordered before the
52	 * reset_pending accesses in cpufreq_stats_record_transition().
53	 */
54	smp_rmb();
55	cpufreq_stats_update(stats, READ_ONCE(stats->reset_time));
56}
57
58static ssize_t show_total_trans(struct cpufreq_policy *policy, char *buf)
59{
60	struct cpufreq_stats *stats = policy->stats;
61
62	if (READ_ONCE(stats->reset_pending))
63		return sprintf(buf, "%d\n", 0);
64	else
65		return sprintf(buf, "%u\n", stats->total_trans);
66}
67cpufreq_freq_attr_ro(total_trans);
68
69static ssize_t show_time_in_state(struct cpufreq_policy *policy, char *buf)
70{
71	struct cpufreq_stats *stats = policy->stats;
72	bool pending = READ_ONCE(stats->reset_pending);
73	unsigned long long time;
74	ssize_t len = 0;
75	int i;
76
77	for (i = 0; i < stats->state_num; i++) {
78		if (pending) {
79			if (i == stats->last_index) {
80				/*
81				 * Prevent the reset_time read from occurring
82				 * before the reset_pending read above.
83				 */
84				smp_rmb();
85				time = local_clock() - READ_ONCE(stats->reset_time);
86			} else {
87				time = 0;
88			}
89		} else {
90			time = stats->time_in_state[i];
91			if (i == stats->last_index)
92				time += local_clock() - stats->last_time;
93		}
94
95		len += sprintf(buf + len, "%u %llu\n", stats->freq_table[i],
96			       nsec_to_clock_t(time));
97	}
98	return len;
99}
100cpufreq_freq_attr_ro(time_in_state);
101
102/* We don't care what is written to the attribute */
103static ssize_t store_reset(struct cpufreq_policy *policy, const char *buf,
104			   size_t count)
105{
106	struct cpufreq_stats *stats = policy->stats;
107
108	/*
109	 * Defer resetting of stats to cpufreq_stats_record_transition() to
110	 * avoid races.
111	 */
112	WRITE_ONCE(stats->reset_time, local_clock());
113	/*
114	 * The memory barrier below is to prevent the readers of reset_time from
115	 * seeing a stale or partially updated value.
116	 */
117	smp_wmb();
118	WRITE_ONCE(stats->reset_pending, 1);
119
120	return count;
121}
122cpufreq_freq_attr_wo(reset);
123
124static ssize_t show_trans_table(struct cpufreq_policy *policy, char *buf)
125{
126	struct cpufreq_stats *stats = policy->stats;
127	bool pending = READ_ONCE(stats->reset_pending);
128	ssize_t len = 0;
129	int i, j, count;
130
131	len += sysfs_emit_at(buf, len, "   From  :    To\n");
132	len += sysfs_emit_at(buf, len, "         : ");
133	for (i = 0; i < stats->state_num; i++) {
134		if (len >= PAGE_SIZE - 1)
135			break;
136		len += sysfs_emit_at(buf, len, "%9u ", stats->freq_table[i]);
137	}
138	if (len >= PAGE_SIZE - 1)
139		return PAGE_SIZE - 1;
140
141	len += sysfs_emit_at(buf, len, "\n");
142
143	for (i = 0; i < stats->state_num; i++) {
144		if (len >= PAGE_SIZE - 1)
145			break;
146
147		len += sysfs_emit_at(buf, len, "%9u: ", stats->freq_table[i]);
148
149		for (j = 0; j < stats->state_num; j++) {
150			if (len >= PAGE_SIZE - 1)
151				break;
152
153			if (pending)
154				count = 0;
155			else
156				count = stats->trans_table[i * stats->max_state + j];
157
158			len += sysfs_emit_at(buf, len, "%9u ", count);
159		}
160		if (len >= PAGE_SIZE - 1)
161			break;
162		len += sysfs_emit_at(buf, len, "\n");
163	}
164
165	if (len >= PAGE_SIZE - 1) {
166		pr_warn_once("cpufreq transition table exceeds PAGE_SIZE. Disabling\n");
167		return -EFBIG;
168	}
169	return len;
170}
171cpufreq_freq_attr_ro(trans_table);
172
173static struct attribute *default_attrs[] = {
174	&total_trans.attr,
175	&time_in_state.attr,
176	&reset.attr,
177	&trans_table.attr,
178	NULL
179};
180static const struct attribute_group stats_attr_group = {
181	.attrs = default_attrs,
182	.name = "stats"
183};
184
185static int freq_table_get_index(struct cpufreq_stats *stats, unsigned int freq)
186{
187	int index;
188	for (index = 0; index < stats->max_state; index++)
189		if (stats->freq_table[index] == freq)
190			return index;
191	return -1;
192}
193
194void cpufreq_stats_free_table(struct cpufreq_policy *policy)
195{
196	struct cpufreq_stats *stats = policy->stats;
197
198	/* Already freed */
199	if (!stats)
200		return;
201
202	pr_debug("%s: Free stats table\n", __func__);
203
204	sysfs_remove_group(&policy->kobj, &stats_attr_group);
205	kfree(stats->time_in_state);
206	kfree(stats);
207	policy->stats = NULL;
208}
209
210void cpufreq_stats_create_table(struct cpufreq_policy *policy)
211{
212	unsigned int i = 0, count;
213	struct cpufreq_stats *stats;
214	unsigned int alloc_size;
215	struct cpufreq_frequency_table *pos;
216
217	count = cpufreq_table_count_valid_entries(policy);
218	if (!count)
219		return;
220
221	/* stats already initialized */
222	if (policy->stats)
223		return;
224
225	stats = kzalloc(sizeof(*stats), GFP_KERNEL);
226	if (!stats)
227		return;
228
229	alloc_size = count * sizeof(int) + count * sizeof(u64);
230
231	alloc_size += count * count * sizeof(int);
232
233	/* Allocate memory for time_in_state/freq_table/trans_table in one go */
234	stats->time_in_state = kzalloc(alloc_size, GFP_KERNEL);
235	if (!stats->time_in_state)
236		goto free_stat;
237
238	stats->freq_table = (unsigned int *)(stats->time_in_state + count);
239
240	stats->trans_table = stats->freq_table + count;
241
242	stats->max_state = count;
243
244	/* Find valid-unique entries */
245	cpufreq_for_each_valid_entry(pos, policy->freq_table)
246		if (policy->freq_table_sorted != CPUFREQ_TABLE_UNSORTED ||
247		    freq_table_get_index(stats, pos->frequency) == -1)
248			stats->freq_table[i++] = pos->frequency;
249
250	stats->state_num = i;
251	stats->last_time = local_clock();
252	stats->last_index = freq_table_get_index(stats, policy->cur);
253
254	policy->stats = stats;
255	if (!sysfs_create_group(&policy->kobj, &stats_attr_group))
256		return;
257
258	/* We failed, release resources */
259	policy->stats = NULL;
260	kfree(stats->time_in_state);
261free_stat:
262	kfree(stats);
263}
264
265void cpufreq_stats_record_transition(struct cpufreq_policy *policy,
266				     unsigned int new_freq)
267{
268	struct cpufreq_stats *stats = policy->stats;
269	int old_index, new_index;
270
271	if (unlikely(!stats))
272		return;
273
274	if (unlikely(READ_ONCE(stats->reset_pending)))
275		cpufreq_stats_reset_table(stats);
276
277	old_index = stats->last_index;
278	new_index = freq_table_get_index(stats, new_freq);
279
280	/* We can't do stats->time_in_state[-1]= .. */
281	if (unlikely(old_index == -1 || new_index == -1 || old_index == new_index))
282		return;
283
284	cpufreq_stats_update(stats, stats->last_time);
285
286	stats->last_index = new_index;
287	stats->trans_table[old_index * stats->max_state + new_index]++;
288	stats->total_trans++;
289}
290