Loading...
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}
1/*
2 * drivers/cpufreq/cpufreq_stats.c
3 *
4 * Copyright (C) 2003-2004 Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>.
5 * (C) 2004 Zou Nan hai <nanhai.zou@intel.com>.
6 *
7 * This program is free software; you can redistribute it and/or modify
8 * it under the terms of the GNU General Public License version 2 as
9 * published by the Free Software Foundation.
10 */
11
12#include <linux/kernel.h>
13#include <linux/slab.h>
14#include <linux/cpu.h>
15#include <linux/sysfs.h>
16#include <linux/cpufreq.h>
17#include <linux/module.h>
18#include <linux/jiffies.h>
19#include <linux/percpu.h>
20#include <linux/kobject.h>
21#include <linux/spinlock.h>
22#include <linux/notifier.h>
23#include <asm/cputime.h>
24
25static spinlock_t cpufreq_stats_lock;
26
27#define CPUFREQ_STATDEVICE_ATTR(_name, _mode, _show) \
28static struct freq_attr _attr_##_name = {\
29 .attr = {.name = __stringify(_name), .mode = _mode, }, \
30 .show = _show,\
31};
32
33struct cpufreq_stats {
34 unsigned int cpu;
35 unsigned int total_trans;
36 unsigned long long last_time;
37 unsigned int max_state;
38 unsigned int state_num;
39 unsigned int last_index;
40 cputime64_t *time_in_state;
41 unsigned int *freq_table;
42#ifdef CONFIG_CPU_FREQ_STAT_DETAILS
43 unsigned int *trans_table;
44#endif
45};
46
47static DEFINE_PER_CPU(struct cpufreq_stats *, cpufreq_stats_table);
48
49struct cpufreq_stats_attribute {
50 struct attribute attr;
51 ssize_t(*show) (struct cpufreq_stats *, char *);
52};
53
54static int cpufreq_stats_update(unsigned int cpu)
55{
56 struct cpufreq_stats *stat;
57 unsigned long long cur_time;
58
59 cur_time = get_jiffies_64();
60 spin_lock(&cpufreq_stats_lock);
61 stat = per_cpu(cpufreq_stats_table, cpu);
62 if (stat->time_in_state)
63 stat->time_in_state[stat->last_index] +=
64 cur_time - stat->last_time;
65 stat->last_time = cur_time;
66 spin_unlock(&cpufreq_stats_lock);
67 return 0;
68}
69
70static ssize_t show_total_trans(struct cpufreq_policy *policy, char *buf)
71{
72 struct cpufreq_stats *stat = per_cpu(cpufreq_stats_table, policy->cpu);
73 if (!stat)
74 return 0;
75 return sprintf(buf, "%d\n",
76 per_cpu(cpufreq_stats_table, stat->cpu)->total_trans);
77}
78
79static ssize_t show_time_in_state(struct cpufreq_policy *policy, char *buf)
80{
81 ssize_t len = 0;
82 int i;
83 struct cpufreq_stats *stat = per_cpu(cpufreq_stats_table, policy->cpu);
84 if (!stat)
85 return 0;
86 cpufreq_stats_update(stat->cpu);
87 for (i = 0; i < stat->state_num; i++) {
88 len += sprintf(buf + len, "%u %llu\n", stat->freq_table[i],
89 (unsigned long long)
90 cputime64_to_clock_t(stat->time_in_state[i]));
91 }
92 return len;
93}
94
95#ifdef CONFIG_CPU_FREQ_STAT_DETAILS
96static ssize_t show_trans_table(struct cpufreq_policy *policy, char *buf)
97{
98 ssize_t len = 0;
99 int i, j;
100
101 struct cpufreq_stats *stat = per_cpu(cpufreq_stats_table, policy->cpu);
102 if (!stat)
103 return 0;
104 cpufreq_stats_update(stat->cpu);
105 len += snprintf(buf + len, PAGE_SIZE - len, " From : To\n");
106 len += snprintf(buf + len, PAGE_SIZE - len, " : ");
107 for (i = 0; i < stat->state_num; i++) {
108 if (len >= PAGE_SIZE)
109 break;
110 len += snprintf(buf + len, PAGE_SIZE - len, "%9u ",
111 stat->freq_table[i]);
112 }
113 if (len >= PAGE_SIZE)
114 return PAGE_SIZE;
115
116 len += snprintf(buf + len, PAGE_SIZE - len, "\n");
117
118 for (i = 0; i < stat->state_num; i++) {
119 if (len >= PAGE_SIZE)
120 break;
121
122 len += snprintf(buf + len, PAGE_SIZE - len, "%9u: ",
123 stat->freq_table[i]);
124
125 for (j = 0; j < stat->state_num; j++) {
126 if (len >= PAGE_SIZE)
127 break;
128 len += snprintf(buf + len, PAGE_SIZE - len, "%9u ",
129 stat->trans_table[i*stat->max_state+j]);
130 }
131 if (len >= PAGE_SIZE)
132 break;
133 len += snprintf(buf + len, PAGE_SIZE - len, "\n");
134 }
135 if (len >= PAGE_SIZE)
136 return PAGE_SIZE;
137 return len;
138}
139CPUFREQ_STATDEVICE_ATTR(trans_table, 0444, show_trans_table);
140#endif
141
142CPUFREQ_STATDEVICE_ATTR(total_trans, 0444, show_total_trans);
143CPUFREQ_STATDEVICE_ATTR(time_in_state, 0444, show_time_in_state);
144
145static struct attribute *default_attrs[] = {
146 &_attr_total_trans.attr,
147 &_attr_time_in_state.attr,
148#ifdef CONFIG_CPU_FREQ_STAT_DETAILS
149 &_attr_trans_table.attr,
150#endif
151 NULL
152};
153static struct attribute_group stats_attr_group = {
154 .attrs = default_attrs,
155 .name = "stats"
156};
157
158static int freq_table_get_index(struct cpufreq_stats *stat, unsigned int freq)
159{
160 int index;
161 for (index = 0; index < stat->max_state; index++)
162 if (stat->freq_table[index] == freq)
163 return index;
164 return -1;
165}
166
167/* should be called late in the CPU removal sequence so that the stats
168 * memory is still available in case someone tries to use it.
169 */
170static void cpufreq_stats_free_table(unsigned int cpu)
171{
172 struct cpufreq_stats *stat = per_cpu(cpufreq_stats_table, cpu);
173 if (stat) {
174 kfree(stat->time_in_state);
175 kfree(stat);
176 }
177 per_cpu(cpufreq_stats_table, cpu) = NULL;
178}
179
180/* must be called early in the CPU removal sequence (before
181 * cpufreq_remove_dev) so that policy is still valid.
182 */
183static void cpufreq_stats_free_sysfs(unsigned int cpu)
184{
185 struct cpufreq_policy *policy = cpufreq_cpu_get(cpu);
186 if (policy && policy->cpu == cpu)
187 sysfs_remove_group(&policy->kobj, &stats_attr_group);
188 if (policy)
189 cpufreq_cpu_put(policy);
190}
191
192static int cpufreq_stats_create_table(struct cpufreq_policy *policy,
193 struct cpufreq_frequency_table *table)
194{
195 unsigned int i, j, count = 0, ret = 0;
196 struct cpufreq_stats *stat;
197 struct cpufreq_policy *data;
198 unsigned int alloc_size;
199 unsigned int cpu = policy->cpu;
200 if (per_cpu(cpufreq_stats_table, cpu))
201 return -EBUSY;
202 stat = kzalloc(sizeof(struct cpufreq_stats), GFP_KERNEL);
203 if ((stat) == NULL)
204 return -ENOMEM;
205
206 data = cpufreq_cpu_get(cpu);
207 if (data == NULL) {
208 ret = -EINVAL;
209 goto error_get_fail;
210 }
211
212 ret = sysfs_create_group(&data->kobj, &stats_attr_group);
213 if (ret)
214 goto error_out;
215
216 stat->cpu = cpu;
217 per_cpu(cpufreq_stats_table, cpu) = stat;
218
219 for (i = 0; table[i].frequency != CPUFREQ_TABLE_END; i++) {
220 unsigned int freq = table[i].frequency;
221 if (freq == CPUFREQ_ENTRY_INVALID)
222 continue;
223 count++;
224 }
225
226 alloc_size = count * sizeof(int) + count * sizeof(cputime64_t);
227
228#ifdef CONFIG_CPU_FREQ_STAT_DETAILS
229 alloc_size += count * count * sizeof(int);
230#endif
231 stat->max_state = count;
232 stat->time_in_state = kzalloc(alloc_size, GFP_KERNEL);
233 if (!stat->time_in_state) {
234 ret = -ENOMEM;
235 goto error_out;
236 }
237 stat->freq_table = (unsigned int *)(stat->time_in_state + count);
238
239#ifdef CONFIG_CPU_FREQ_STAT_DETAILS
240 stat->trans_table = stat->freq_table + count;
241#endif
242 j = 0;
243 for (i = 0; table[i].frequency != CPUFREQ_TABLE_END; i++) {
244 unsigned int freq = table[i].frequency;
245 if (freq == CPUFREQ_ENTRY_INVALID)
246 continue;
247 if (freq_table_get_index(stat, freq) == -1)
248 stat->freq_table[j++] = freq;
249 }
250 stat->state_num = j;
251 spin_lock(&cpufreq_stats_lock);
252 stat->last_time = get_jiffies_64();
253 stat->last_index = freq_table_get_index(stat, policy->cur);
254 spin_unlock(&cpufreq_stats_lock);
255 cpufreq_cpu_put(data);
256 return 0;
257error_out:
258 cpufreq_cpu_put(data);
259error_get_fail:
260 kfree(stat);
261 per_cpu(cpufreq_stats_table, cpu) = NULL;
262 return ret;
263}
264
265static int cpufreq_stat_notifier_policy(struct notifier_block *nb,
266 unsigned long val, void *data)
267{
268 int ret;
269 struct cpufreq_policy *policy = data;
270 struct cpufreq_frequency_table *table;
271 unsigned int cpu = policy->cpu;
272 if (val != CPUFREQ_NOTIFY)
273 return 0;
274 table = cpufreq_frequency_get_table(cpu);
275 if (!table)
276 return 0;
277 ret = cpufreq_stats_create_table(policy, table);
278 if (ret)
279 return ret;
280 return 0;
281}
282
283static int cpufreq_stat_notifier_trans(struct notifier_block *nb,
284 unsigned long val, void *data)
285{
286 struct cpufreq_freqs *freq = data;
287 struct cpufreq_stats *stat;
288 int old_index, new_index;
289
290 if (val != CPUFREQ_POSTCHANGE)
291 return 0;
292
293 stat = per_cpu(cpufreq_stats_table, freq->cpu);
294 if (!stat)
295 return 0;
296
297 old_index = stat->last_index;
298 new_index = freq_table_get_index(stat, freq->new);
299
300 /* We can't do stat->time_in_state[-1]= .. */
301 if (old_index == -1 || new_index == -1)
302 return 0;
303
304 cpufreq_stats_update(freq->cpu);
305
306 if (old_index == new_index)
307 return 0;
308
309 spin_lock(&cpufreq_stats_lock);
310 stat->last_index = new_index;
311#ifdef CONFIG_CPU_FREQ_STAT_DETAILS
312 stat->trans_table[old_index * stat->max_state + new_index]++;
313#endif
314 stat->total_trans++;
315 spin_unlock(&cpufreq_stats_lock);
316 return 0;
317}
318
319static int __cpuinit cpufreq_stat_cpu_callback(struct notifier_block *nfb,
320 unsigned long action,
321 void *hcpu)
322{
323 unsigned int cpu = (unsigned long)hcpu;
324
325 switch (action) {
326 case CPU_ONLINE:
327 case CPU_ONLINE_FROZEN:
328 cpufreq_update_policy(cpu);
329 break;
330 case CPU_DOWN_PREPARE:
331 cpufreq_stats_free_sysfs(cpu);
332 break;
333 case CPU_DEAD:
334 case CPU_DEAD_FROZEN:
335 cpufreq_stats_free_table(cpu);
336 break;
337 }
338 return NOTIFY_OK;
339}
340
341/* priority=1 so this will get called before cpufreq_remove_dev */
342static struct notifier_block cpufreq_stat_cpu_notifier __refdata = {
343 .notifier_call = cpufreq_stat_cpu_callback,
344 .priority = 1,
345};
346
347static struct notifier_block notifier_policy_block = {
348 .notifier_call = cpufreq_stat_notifier_policy
349};
350
351static struct notifier_block notifier_trans_block = {
352 .notifier_call = cpufreq_stat_notifier_trans
353};
354
355static int __init cpufreq_stats_init(void)
356{
357 int ret;
358 unsigned int cpu;
359
360 spin_lock_init(&cpufreq_stats_lock);
361 ret = cpufreq_register_notifier(¬ifier_policy_block,
362 CPUFREQ_POLICY_NOTIFIER);
363 if (ret)
364 return ret;
365
366 ret = cpufreq_register_notifier(¬ifier_trans_block,
367 CPUFREQ_TRANSITION_NOTIFIER);
368 if (ret) {
369 cpufreq_unregister_notifier(¬ifier_policy_block,
370 CPUFREQ_POLICY_NOTIFIER);
371 return ret;
372 }
373
374 register_hotcpu_notifier(&cpufreq_stat_cpu_notifier);
375 for_each_online_cpu(cpu) {
376 cpufreq_update_policy(cpu);
377 }
378 return 0;
379}
380static void __exit cpufreq_stats_exit(void)
381{
382 unsigned int cpu;
383
384 cpufreq_unregister_notifier(¬ifier_policy_block,
385 CPUFREQ_POLICY_NOTIFIER);
386 cpufreq_unregister_notifier(¬ifier_trans_block,
387 CPUFREQ_TRANSITION_NOTIFIER);
388 unregister_hotcpu_notifier(&cpufreq_stat_cpu_notifier);
389 for_each_online_cpu(cpu) {
390 cpufreq_stats_free_table(cpu);
391 cpufreq_stats_free_sysfs(cpu);
392 }
393}
394
395MODULE_AUTHOR("Zou Nan hai <nanhai.zou@intel.com>");
396MODULE_DESCRIPTION("'cpufreq_stats' - A driver to export cpufreq stats "
397 "through sysfs filesystem");
398MODULE_LICENSE("GPL");
399
400module_init(cpufreq_stats_init);
401module_exit(cpufreq_stats_exit);