1// SPDX-License-Identifier: GPL-2.0-only
2/*
3 * drivers/cpufreq/cpufreq_ondemand.c
4 *
5 * Copyright (C) 2001 Russell King
6 * (C) 2003 Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>.
7 * Jun Nakajima <jun.nakajima@intel.com>
8 */
9
10#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
11
12#include <linux/cpu.h>
13#include <linux/percpu-defs.h>
14#include <linux/slab.h>
15#include <linux/tick.h>
16#include <linux/sched/cpufreq.h>
17
18#include "cpufreq_ondemand.h"
19
20/* On-demand governor macros */
21#define DEF_FREQUENCY_UP_THRESHOLD (80)
22#define DEF_SAMPLING_DOWN_FACTOR (1)
23#define MAX_SAMPLING_DOWN_FACTOR (100000)
24#define MICRO_FREQUENCY_UP_THRESHOLD (95)
25#define MICRO_FREQUENCY_MIN_SAMPLE_RATE (10000)
26#define MIN_FREQUENCY_UP_THRESHOLD (1)
27#define MAX_FREQUENCY_UP_THRESHOLD (100)
28
29static struct od_ops od_ops;
30
31static unsigned int default_powersave_bias;
32
33/*
34 * Not all CPUs want IO time to be accounted as busy; this depends on how
35 * efficient idling at a higher frequency/voltage is.
36 * Pavel Machek says this is not so for various generations of AMD and old
37 * Intel systems.
38 * Mike Chan (android.com) claims this is also not true for ARM.
39 * Because of this, whitelist specific known (series) of CPUs by default, and
40 * leave all others up to the user.
41 */
42static int should_io_be_busy(void)
43{
44#if defined(CONFIG_X86)
45 /*
46 * For Intel, Core 2 (model 15) and later have an efficient idle.
47 */
48 if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL &&
49 boot_cpu_data.x86 == 6 &&
50 boot_cpu_data.x86_model >= 15)
51 return 1;
52#endif
53 return 0;
54}
55
56/*
57 * Find right freq to be set now with powersave_bias on.
58 * Returns the freq_hi to be used right now and will set freq_hi_delay_us,
59 * freq_lo, and freq_lo_delay_us in percpu area for averaging freqs.
60 */
61static unsigned int generic_powersave_bias_target(struct cpufreq_policy *policy,
62 unsigned int freq_next, unsigned int relation)
63{
64 unsigned int freq_req, freq_reduc, freq_avg;
65 unsigned int freq_hi, freq_lo;
66 unsigned int index;
67 unsigned int delay_hi_us;
68 struct policy_dbs_info *policy_dbs = policy->governor_data;
69 struct od_policy_dbs_info *dbs_info = to_dbs_info(policy_dbs);
70 struct dbs_data *dbs_data = policy_dbs->dbs_data;
71 struct od_dbs_tuners *od_tuners = dbs_data->tuners;
72 struct cpufreq_frequency_table *freq_table = policy->freq_table;
73
74 if (!freq_table) {
75 dbs_info->freq_lo = 0;
76 dbs_info->freq_lo_delay_us = 0;
77 return freq_next;
78 }
79
80 index = cpufreq_frequency_table_target(policy, target_freq: freq_next, relation);
81 freq_req = freq_table[index].frequency;
82 freq_reduc = freq_req * od_tuners->powersave_bias / 1000;
83 freq_avg = freq_req - freq_reduc;
84
85 /* Find freq bounds for freq_avg in freq_table */
86 index = cpufreq_table_find_index_h(policy, target_freq: freq_avg,
87 efficiencies: relation & CPUFREQ_RELATION_E);
88 freq_lo = freq_table[index].frequency;
89 index = cpufreq_table_find_index_l(policy, target_freq: freq_avg,
90 efficiencies: relation & CPUFREQ_RELATION_E);
91 freq_hi = freq_table[index].frequency;
92
93 /* Find out how long we have to be in hi and lo freqs */
94 if (freq_hi == freq_lo) {
95 dbs_info->freq_lo = 0;
96 dbs_info->freq_lo_delay_us = 0;
97 return freq_lo;
98 }
99 delay_hi_us = (freq_avg - freq_lo) * dbs_data->sampling_rate;
100 delay_hi_us += (freq_hi - freq_lo) / 2;
101 delay_hi_us /= freq_hi - freq_lo;
102 dbs_info->freq_hi_delay_us = delay_hi_us;
103 dbs_info->freq_lo = freq_lo;
104 dbs_info->freq_lo_delay_us = dbs_data->sampling_rate - delay_hi_us;
105 return freq_hi;
106}
107
108static void ondemand_powersave_bias_init(struct cpufreq_policy *policy)
109{
110 struct od_policy_dbs_info *dbs_info = to_dbs_info(policy_dbs: policy->governor_data);
111
112 dbs_info->freq_lo = 0;
113}
114
115static void dbs_freq_increase(struct cpufreq_policy *policy, unsigned int freq)
116{
117 struct policy_dbs_info *policy_dbs = policy->governor_data;
118 struct dbs_data *dbs_data = policy_dbs->dbs_data;
119 struct od_dbs_tuners *od_tuners = dbs_data->tuners;
120
121 if (od_tuners->powersave_bias)
122 freq = od_ops.powersave_bias_target(policy, freq,
123 CPUFREQ_RELATION_HE);
124 else if (policy->cur == policy->max)
125 return;
126
127 __cpufreq_driver_target(policy, target_freq: freq, relation: od_tuners->powersave_bias ?
128 CPUFREQ_RELATION_LE : CPUFREQ_RELATION_HE);
129}
130
131/*
132 * Every sampling_rate, we check, if current idle time is less than 20%
133 * (default), then we try to increase frequency. Else, we adjust the frequency
134 * proportional to load.
135 */
136static void od_update(struct cpufreq_policy *policy)
137{
138 struct policy_dbs_info *policy_dbs = policy->governor_data;
139 struct od_policy_dbs_info *dbs_info = to_dbs_info(policy_dbs);
140 struct dbs_data *dbs_data = policy_dbs->dbs_data;
141 struct od_dbs_tuners *od_tuners = dbs_data->tuners;
142 unsigned int load = dbs_update(policy);
143
144 dbs_info->freq_lo = 0;
145
146 /* Check for frequency increase */
147 if (load > dbs_data->up_threshold) {
148 /* If switching to max speed, apply sampling_down_factor */
149 if (policy->cur < policy->max)
150 policy_dbs->rate_mult = dbs_data->sampling_down_factor;
151 dbs_freq_increase(policy, freq: policy->max);
152 } else {
153 /* Calculate the next frequency proportional to load */
154 unsigned int freq_next, min_f, max_f;
155
156 min_f = policy->cpuinfo.min_freq;
157 max_f = policy->cpuinfo.max_freq;
158 freq_next = min_f + load * (max_f - min_f) / 100;
159
160 /* No longer fully busy, reset rate_mult */
161 policy_dbs->rate_mult = 1;
162
163 if (od_tuners->powersave_bias)
164 freq_next = od_ops.powersave_bias_target(policy,
165 freq_next,
166 CPUFREQ_RELATION_LE);
167
168 __cpufreq_driver_target(policy, target_freq: freq_next, CPUFREQ_RELATION_CE);
169 }
170}
171
172static unsigned int od_dbs_update(struct cpufreq_policy *policy)
173{
174 struct policy_dbs_info *policy_dbs = policy->governor_data;
175 struct dbs_data *dbs_data = policy_dbs->dbs_data;
176 struct od_policy_dbs_info *dbs_info = to_dbs_info(policy_dbs);
177 int sample_type = dbs_info->sample_type;
178
179 /* Common NORMAL_SAMPLE setup */
180 dbs_info->sample_type = OD_NORMAL_SAMPLE;
181 /*
182 * OD_SUB_SAMPLE doesn't make sense if sample_delay_ns is 0, so ignore
183 * it then.
184 */
185 if (sample_type == OD_SUB_SAMPLE && policy_dbs->sample_delay_ns > 0) {
186 __cpufreq_driver_target(policy, target_freq: dbs_info->freq_lo,
187 CPUFREQ_RELATION_HE);
188 return dbs_info->freq_lo_delay_us;
189 }
190
191 od_update(policy);
192
193 if (dbs_info->freq_lo) {
194 /* Setup SUB_SAMPLE */
195 dbs_info->sample_type = OD_SUB_SAMPLE;
196 return dbs_info->freq_hi_delay_us;
197 }
198
199 return dbs_data->sampling_rate * policy_dbs->rate_mult;
200}
201
202/************************** sysfs interface ************************/
203static struct dbs_governor od_dbs_gov;
204
205static ssize_t io_is_busy_store(struct gov_attr_set *attr_set, const char *buf,
206 size_t count)
207{
208 struct dbs_data *dbs_data = to_dbs_data(attr_set);
209 unsigned int input;
210 int ret;
211
212 ret = sscanf(buf, "%u", &input);
213 if (ret != 1)
214 return -EINVAL;
215 dbs_data->io_is_busy = !!input;
216
217 /* we need to re-evaluate prev_cpu_idle */
218 gov_update_cpu_data(dbs_data);
219
220 return count;
221}
222
223static ssize_t up_threshold_store(struct gov_attr_set *attr_set,
224 const char *buf, size_t count)
225{
226 struct dbs_data *dbs_data = to_dbs_data(attr_set);
227 unsigned int input;
228 int ret;
229 ret = sscanf(buf, "%u", &input);
230
231 if (ret != 1 || input > MAX_FREQUENCY_UP_THRESHOLD ||
232 input < MIN_FREQUENCY_UP_THRESHOLD) {
233 return -EINVAL;
234 }
235
236 dbs_data->up_threshold = input;
237 return count;
238}
239
240static ssize_t sampling_down_factor_store(struct gov_attr_set *attr_set,
241 const char *buf, size_t count)
242{
243 struct dbs_data *dbs_data = to_dbs_data(attr_set);
244 struct policy_dbs_info *policy_dbs;
245 unsigned int input;
246 int ret;
247 ret = sscanf(buf, "%u", &input);
248
249 if (ret != 1 || input > MAX_SAMPLING_DOWN_FACTOR || input < 1)
250 return -EINVAL;
251
252 dbs_data->sampling_down_factor = input;
253
254 /* Reset down sampling multiplier in case it was active */
255 list_for_each_entry(policy_dbs, &attr_set->policy_list, list) {
256 /*
257 * Doing this without locking might lead to using different
258 * rate_mult values in od_update() and od_dbs_update().
259 */
260 mutex_lock(&policy_dbs->update_mutex);
261 policy_dbs->rate_mult = 1;
262 mutex_unlock(lock: &policy_dbs->update_mutex);
263 }
264
265 return count;
266}
267
268static ssize_t ignore_nice_load_store(struct gov_attr_set *attr_set,
269 const char *buf, size_t count)
270{
271 struct dbs_data *dbs_data = to_dbs_data(attr_set);
272 unsigned int input;
273 int ret;
274
275 ret = sscanf(buf, "%u", &input);
276 if (ret != 1)
277 return -EINVAL;
278
279 if (input > 1)
280 input = 1;
281
282 if (input == dbs_data->ignore_nice_load) { /* nothing to do */
283 return count;
284 }
285 dbs_data->ignore_nice_load = input;
286
287 /* we need to re-evaluate prev_cpu_idle */
288 gov_update_cpu_data(dbs_data);
289
290 return count;
291}
292
293static ssize_t powersave_bias_store(struct gov_attr_set *attr_set,
294 const char *buf, size_t count)
295{
296 struct dbs_data *dbs_data = to_dbs_data(attr_set);
297 struct od_dbs_tuners *od_tuners = dbs_data->tuners;
298 struct policy_dbs_info *policy_dbs;
299 unsigned int input;
300 int ret;
301 ret = sscanf(buf, "%u", &input);
302
303 if (ret != 1)
304 return -EINVAL;
305
306 if (input > 1000)
307 input = 1000;
308
309 od_tuners->powersave_bias = input;
310
311 list_for_each_entry(policy_dbs, &attr_set->policy_list, list)
312 ondemand_powersave_bias_init(policy: policy_dbs->policy);
313
314 return count;
315}
316
317gov_show_one_common(sampling_rate);
318gov_show_one_common(up_threshold);
319gov_show_one_common(sampling_down_factor);
320gov_show_one_common(ignore_nice_load);
321gov_show_one_common(io_is_busy);
322gov_show_one(od, powersave_bias);
323
324gov_attr_rw(sampling_rate);
325gov_attr_rw(io_is_busy);
326gov_attr_rw(up_threshold);
327gov_attr_rw(sampling_down_factor);
328gov_attr_rw(ignore_nice_load);
329gov_attr_rw(powersave_bias);
330
331static struct attribute *od_attrs[] = {
332 &sampling_rate.attr,
333 &up_threshold.attr,
334 &sampling_down_factor.attr,
335 &ignore_nice_load.attr,
336 &powersave_bias.attr,
337 &io_is_busy.attr,
338 NULL
339};
340ATTRIBUTE_GROUPS(od);
341
342/************************** sysfs end ************************/
343
344static struct policy_dbs_info *od_alloc(void)
345{
346 struct od_policy_dbs_info *dbs_info;
347
348 dbs_info = kzalloc(size: sizeof(*dbs_info), GFP_KERNEL);
349 return dbs_info ? &dbs_info->policy_dbs : NULL;
350}
351
352static void od_free(struct policy_dbs_info *policy_dbs)
353{
354 kfree(objp: to_dbs_info(policy_dbs));
355}
356
357static int od_init(struct dbs_data *dbs_data)
358{
359 struct od_dbs_tuners *tuners;
360 u64 idle_time;
361 int cpu;
362
363 tuners = kzalloc(size: sizeof(*tuners), GFP_KERNEL);
364 if (!tuners)
365 return -ENOMEM;
366
367 cpu = get_cpu();
368 idle_time = get_cpu_idle_time_us(cpu, NULL);
369 put_cpu();
370 if (idle_time != -1ULL) {
371 /* Idle micro accounting is supported. Use finer thresholds */
372 dbs_data->up_threshold = MICRO_FREQUENCY_UP_THRESHOLD;
373 } else {
374 dbs_data->up_threshold = DEF_FREQUENCY_UP_THRESHOLD;
375 }
376
377 dbs_data->sampling_down_factor = DEF_SAMPLING_DOWN_FACTOR;
378 dbs_data->ignore_nice_load = 0;
379 tuners->powersave_bias = default_powersave_bias;
380 dbs_data->io_is_busy = should_io_be_busy();
381
382 dbs_data->tuners = tuners;
383 return 0;
384}
385
386static void od_exit(struct dbs_data *dbs_data)
387{
388 kfree(objp: dbs_data->tuners);
389}
390
391static void od_start(struct cpufreq_policy *policy)
392{
393 struct od_policy_dbs_info *dbs_info = to_dbs_info(policy_dbs: policy->governor_data);
394
395 dbs_info->sample_type = OD_NORMAL_SAMPLE;
396 ondemand_powersave_bias_init(policy);
397}
398
399static struct od_ops od_ops = {
400 .powersave_bias_target = generic_powersave_bias_target,
401};
402
403static struct dbs_governor od_dbs_gov = {
404 .gov = CPUFREQ_DBS_GOVERNOR_INITIALIZER("ondemand"),
405 .kobj_type = { .default_groups = od_groups },
406 .gov_dbs_update = od_dbs_update,
407 .alloc = od_alloc,
408 .free = od_free,
409 .init = od_init,
410 .exit = od_exit,
411 .start = od_start,
412};
413
414#define CPU_FREQ_GOV_ONDEMAND (od_dbs_gov.gov)
415
416static void od_set_powersave_bias(unsigned int powersave_bias)
417{
418 unsigned int cpu;
419 cpumask_var_t done;
420
421 if (!alloc_cpumask_var(mask: &done, GFP_KERNEL))
422 return;
423
424 default_powersave_bias = powersave_bias;
425 cpumask_clear(dstp: done);
426
427 cpus_read_lock();
428 for_each_online_cpu(cpu) {
429 struct cpufreq_policy *policy;
430 struct policy_dbs_info *policy_dbs;
431 struct dbs_data *dbs_data;
432 struct od_dbs_tuners *od_tuners;
433
434 if (cpumask_test_cpu(cpu, cpumask: done))
435 continue;
436
437 policy = cpufreq_cpu_get_raw(cpu);
438 if (!policy || policy->governor != &CPU_FREQ_GOV_ONDEMAND)
439 continue;
440
441 policy_dbs = policy->governor_data;
442 if (!policy_dbs)
443 continue;
444
445 cpumask_or(dstp: done, src1p: done, src2p: policy->cpus);
446
447 dbs_data = policy_dbs->dbs_data;
448 od_tuners = dbs_data->tuners;
449 od_tuners->powersave_bias = default_powersave_bias;
450 }
451 cpus_read_unlock();
452
453 free_cpumask_var(mask: done);
454}
455
456void od_register_powersave_bias_handler(unsigned int (*f)
457 (struct cpufreq_policy *, unsigned int, unsigned int),
458 unsigned int powersave_bias)
459{
460 od_ops.powersave_bias_target = f;
461 od_set_powersave_bias(powersave_bias);
462}
463EXPORT_SYMBOL_GPL(od_register_powersave_bias_handler);
464
465void od_unregister_powersave_bias_handler(void)
466{
467 od_ops.powersave_bias_target = generic_powersave_bias_target;
468 od_set_powersave_bias(powersave_bias: 0);
469}
470EXPORT_SYMBOL_GPL(od_unregister_powersave_bias_handler);
471
472MODULE_AUTHOR("Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>");
473MODULE_AUTHOR("Alexey Starikovskiy <alexey.y.starikovskiy@intel.com>");
474MODULE_DESCRIPTION("'cpufreq_ondemand' - A dynamic cpufreq governor for "
475 "Low Latency Frequency Transition capable processors");
476MODULE_LICENSE("GPL");
477
478#ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND
479struct cpufreq_governor *cpufreq_default_governor(void)
480{
481 return &CPU_FREQ_GOV_ONDEMAND;
482}
483#endif
484
485cpufreq_governor_init(CPU_FREQ_GOV_ONDEMAND);
486cpufreq_governor_exit(CPU_FREQ_GOV_ONDEMAND);
487

source code of linux/drivers/cpufreq/cpufreq_ondemand.c