Merge tag 'for_linus' of git://git.kernel.org/pub/scm/linux/kernel/git/mst/vhost
[cris-mirror.git] / drivers / cpufreq / cpufreq_ondemand.c
blob6b423eebfd5dbefccc2f20e52825271d0b2a7114
1 /*
2 * drivers/cpufreq/cpufreq_ondemand.c
4 * Copyright (C) 2001 Russell King
5 * (C) 2003 Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>.
6 * Jun Nakajima <jun.nakajima@intel.com>
8 * This program is free software; you can redistribute it and/or modify
9 * it under the terms of the GNU General Public License version 2 as
10 * published by the Free Software Foundation.
13 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
15 #include <linux/cpu.h>
16 #include <linux/percpu-defs.h>
17 #include <linux/slab.h>
18 #include <linux/tick.h>
19 #include <linux/sched/cpufreq.h>
21 #include "cpufreq_ondemand.h"
23 /* On-demand governor macros */
24 #define DEF_FREQUENCY_UP_THRESHOLD (80)
25 #define DEF_SAMPLING_DOWN_FACTOR (1)
26 #define MAX_SAMPLING_DOWN_FACTOR (100000)
27 #define MICRO_FREQUENCY_UP_THRESHOLD (95)
28 #define MICRO_FREQUENCY_MIN_SAMPLE_RATE (10000)
29 #define MIN_FREQUENCY_UP_THRESHOLD (1)
30 #define MAX_FREQUENCY_UP_THRESHOLD (100)
32 static struct od_ops od_ops;
34 static unsigned int default_powersave_bias;
37 * Not all CPUs want IO time to be accounted as busy; this depends on how
38 * efficient idling at a higher frequency/voltage is.
39 * Pavel Machek says this is not so for various generations of AMD and old
40 * Intel systems.
41 * Mike Chan (android.com) claims this is also not true for ARM.
42 * Because of this, whitelist specific known (series) of CPUs by default, and
43 * leave all others up to the user.
45 static int should_io_be_busy(void)
47 #if defined(CONFIG_X86)
49 * For Intel, Core 2 (model 15) and later have an efficient idle.
51 if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL &&
52 boot_cpu_data.x86 == 6 &&
53 boot_cpu_data.x86_model >= 15)
54 return 1;
55 #endif
56 return 0;
60 * Find right freq to be set now with powersave_bias on.
61 * Returns the freq_hi to be used right now and will set freq_hi_delay_us,
62 * freq_lo, and freq_lo_delay_us in percpu area for averaging freqs.
64 static unsigned int generic_powersave_bias_target(struct cpufreq_policy *policy,
65 unsigned int freq_next, unsigned int relation)
67 unsigned int freq_req, freq_reduc, freq_avg;
68 unsigned int freq_hi, freq_lo;
69 unsigned int index;
70 unsigned int delay_hi_us;
71 struct policy_dbs_info *policy_dbs = policy->governor_data;
72 struct od_policy_dbs_info *dbs_info = to_dbs_info(policy_dbs);
73 struct dbs_data *dbs_data = policy_dbs->dbs_data;
74 struct od_dbs_tuners *od_tuners = dbs_data->tuners;
75 struct cpufreq_frequency_table *freq_table = policy->freq_table;
77 if (!freq_table) {
78 dbs_info->freq_lo = 0;
79 dbs_info->freq_lo_delay_us = 0;
80 return freq_next;
83 index = cpufreq_frequency_table_target(policy, freq_next, relation);
84 freq_req = freq_table[index].frequency;
85 freq_reduc = freq_req * od_tuners->powersave_bias / 1000;
86 freq_avg = freq_req - freq_reduc;
88 /* Find freq bounds for freq_avg in freq_table */
89 index = cpufreq_table_find_index_h(policy, freq_avg);
90 freq_lo = freq_table[index].frequency;
91 index = cpufreq_table_find_index_l(policy, freq_avg);
92 freq_hi = freq_table[index].frequency;
94 /* Find out how long we have to be in hi and lo freqs */
95 if (freq_hi == freq_lo) {
96 dbs_info->freq_lo = 0;
97 dbs_info->freq_lo_delay_us = 0;
98 return freq_lo;
100 delay_hi_us = (freq_avg - freq_lo) * dbs_data->sampling_rate;
101 delay_hi_us += (freq_hi - freq_lo) / 2;
102 delay_hi_us /= freq_hi - freq_lo;
103 dbs_info->freq_hi_delay_us = delay_hi_us;
104 dbs_info->freq_lo = freq_lo;
105 dbs_info->freq_lo_delay_us = dbs_data->sampling_rate - delay_hi_us;
106 return freq_hi;
109 static void ondemand_powersave_bias_init(struct cpufreq_policy *policy)
111 struct od_policy_dbs_info *dbs_info = to_dbs_info(policy->governor_data);
113 dbs_info->freq_lo = 0;
116 static void dbs_freq_increase(struct cpufreq_policy *policy, unsigned int freq)
118 struct policy_dbs_info *policy_dbs = policy->governor_data;
119 struct dbs_data *dbs_data = policy_dbs->dbs_data;
120 struct od_dbs_tuners *od_tuners = dbs_data->tuners;
122 if (od_tuners->powersave_bias)
123 freq = od_ops.powersave_bias_target(policy, freq,
124 CPUFREQ_RELATION_H);
125 else if (policy->cur == policy->max)
126 return;
128 __cpufreq_driver_target(policy, freq, od_tuners->powersave_bias ?
129 CPUFREQ_RELATION_L : CPUFREQ_RELATION_H);
133 * Every sampling_rate, we check, if current idle time is less than 20%
134 * (default), then we try to increase frequency. Else, we adjust the frequency
135 * proportional to load.
137 static void od_update(struct cpufreq_policy *policy)
139 struct policy_dbs_info *policy_dbs = policy->governor_data;
140 struct od_policy_dbs_info *dbs_info = to_dbs_info(policy_dbs);
141 struct dbs_data *dbs_data = policy_dbs->dbs_data;
142 struct od_dbs_tuners *od_tuners = dbs_data->tuners;
143 unsigned int load = dbs_update(policy);
145 dbs_info->freq_lo = 0;
147 /* Check for frequency increase */
148 if (load > dbs_data->up_threshold) {
149 /* If switching to max speed, apply sampling_down_factor */
150 if (policy->cur < policy->max)
151 policy_dbs->rate_mult = dbs_data->sampling_down_factor;
152 dbs_freq_increase(policy, policy->max);
153 } else {
154 /* Calculate the next frequency proportional to load */
155 unsigned int freq_next, min_f, max_f;
157 min_f = policy->cpuinfo.min_freq;
158 max_f = policy->cpuinfo.max_freq;
159 freq_next = min_f + load * (max_f - min_f) / 100;
161 /* No longer fully busy, reset rate_mult */
162 policy_dbs->rate_mult = 1;
164 if (od_tuners->powersave_bias)
165 freq_next = od_ops.powersave_bias_target(policy,
166 freq_next,
167 CPUFREQ_RELATION_L);
169 __cpufreq_driver_target(policy, freq_next, CPUFREQ_RELATION_C);
173 static unsigned int od_dbs_update(struct cpufreq_policy *policy)
175 struct policy_dbs_info *policy_dbs = policy->governor_data;
176 struct dbs_data *dbs_data = policy_dbs->dbs_data;
177 struct od_policy_dbs_info *dbs_info = to_dbs_info(policy_dbs);
178 int sample_type = dbs_info->sample_type;
180 /* Common NORMAL_SAMPLE setup */
181 dbs_info->sample_type = OD_NORMAL_SAMPLE;
183 * OD_SUB_SAMPLE doesn't make sense if sample_delay_ns is 0, so ignore
184 * it then.
186 if (sample_type == OD_SUB_SAMPLE && policy_dbs->sample_delay_ns > 0) {
187 __cpufreq_driver_target(policy, dbs_info->freq_lo,
188 CPUFREQ_RELATION_H);
189 return dbs_info->freq_lo_delay_us;
192 od_update(policy);
194 if (dbs_info->freq_lo) {
195 /* Setup SUB_SAMPLE */
196 dbs_info->sample_type = OD_SUB_SAMPLE;
197 return dbs_info->freq_hi_delay_us;
200 return dbs_data->sampling_rate * policy_dbs->rate_mult;
203 /************************** sysfs interface ************************/
204 static struct dbs_governor od_dbs_gov;
206 static ssize_t store_io_is_busy(struct gov_attr_set *attr_set, const char *buf,
207 size_t count)
209 struct dbs_data *dbs_data = to_dbs_data(attr_set);
210 unsigned int input;
211 int ret;
213 ret = sscanf(buf, "%u", &input);
214 if (ret != 1)
215 return -EINVAL;
216 dbs_data->io_is_busy = !!input;
218 /* we need to re-evaluate prev_cpu_idle */
219 gov_update_cpu_data(dbs_data);
221 return count;
224 static ssize_t store_up_threshold(struct gov_attr_set *attr_set,
225 const char *buf, size_t count)
227 struct dbs_data *dbs_data = to_dbs_data(attr_set);
228 unsigned int input;
229 int ret;
230 ret = sscanf(buf, "%u", &input);
232 if (ret != 1 || input > MAX_FREQUENCY_UP_THRESHOLD ||
233 input < MIN_FREQUENCY_UP_THRESHOLD) {
234 return -EINVAL;
237 dbs_data->up_threshold = input;
238 return count;
241 static ssize_t store_sampling_down_factor(struct gov_attr_set *attr_set,
242 const char *buf, size_t count)
244 struct dbs_data *dbs_data = to_dbs_data(attr_set);
245 struct policy_dbs_info *policy_dbs;
246 unsigned int input;
247 int ret;
248 ret = sscanf(buf, "%u", &input);
250 if (ret != 1 || input > MAX_SAMPLING_DOWN_FACTOR || input < 1)
251 return -EINVAL;
253 dbs_data->sampling_down_factor = input;
255 /* Reset down sampling multiplier in case it was active */
256 list_for_each_entry(policy_dbs, &attr_set->policy_list, list) {
258 * Doing this without locking might lead to using different
259 * rate_mult values in od_update() and od_dbs_update().
261 mutex_lock(&policy_dbs->update_mutex);
262 policy_dbs->rate_mult = 1;
263 mutex_unlock(&policy_dbs->update_mutex);
266 return count;
269 static ssize_t store_ignore_nice_load(struct gov_attr_set *attr_set,
270 const char *buf, size_t count)
272 struct dbs_data *dbs_data = to_dbs_data(attr_set);
273 unsigned int input;
274 int ret;
276 ret = sscanf(buf, "%u", &input);
277 if (ret != 1)
278 return -EINVAL;
280 if (input > 1)
281 input = 1;
283 if (input == dbs_data->ignore_nice_load) { /* nothing to do */
284 return count;
286 dbs_data->ignore_nice_load = input;
288 /* we need to re-evaluate prev_cpu_idle */
289 gov_update_cpu_data(dbs_data);
291 return count;
294 static ssize_t store_powersave_bias(struct gov_attr_set *attr_set,
295 const char *buf, size_t count)
297 struct dbs_data *dbs_data = to_dbs_data(attr_set);
298 struct od_dbs_tuners *od_tuners = dbs_data->tuners;
299 struct policy_dbs_info *policy_dbs;
300 unsigned int input;
301 int ret;
302 ret = sscanf(buf, "%u", &input);
304 if (ret != 1)
305 return -EINVAL;
307 if (input > 1000)
308 input = 1000;
310 od_tuners->powersave_bias = input;
312 list_for_each_entry(policy_dbs, &attr_set->policy_list, list)
313 ondemand_powersave_bias_init(policy_dbs->policy);
315 return count;
318 gov_show_one_common(sampling_rate);
319 gov_show_one_common(up_threshold);
320 gov_show_one_common(sampling_down_factor);
321 gov_show_one_common(ignore_nice_load);
322 gov_show_one_common(io_is_busy);
323 gov_show_one(od, powersave_bias);
325 gov_attr_rw(sampling_rate);
326 gov_attr_rw(io_is_busy);
327 gov_attr_rw(up_threshold);
328 gov_attr_rw(sampling_down_factor);
329 gov_attr_rw(ignore_nice_load);
330 gov_attr_rw(powersave_bias);
332 static struct attribute *od_attributes[] = {
333 &sampling_rate.attr,
334 &up_threshold.attr,
335 &sampling_down_factor.attr,
336 &ignore_nice_load.attr,
337 &powersave_bias.attr,
338 &io_is_busy.attr,
339 NULL
342 /************************** sysfs end ************************/
344 static struct policy_dbs_info *od_alloc(void)
346 struct od_policy_dbs_info *dbs_info;
348 dbs_info = kzalloc(sizeof(*dbs_info), GFP_KERNEL);
349 return dbs_info ? &dbs_info->policy_dbs : NULL;
352 static void od_free(struct policy_dbs_info *policy_dbs)
354 kfree(to_dbs_info(policy_dbs));
357 static int od_init(struct dbs_data *dbs_data)
359 struct od_dbs_tuners *tuners;
360 u64 idle_time;
361 int cpu;
363 tuners = kzalloc(sizeof(*tuners), GFP_KERNEL);
364 if (!tuners)
365 return -ENOMEM;
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;
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();
382 dbs_data->tuners = tuners;
383 return 0;
386 static void od_exit(struct dbs_data *dbs_data)
388 kfree(dbs_data->tuners);
391 static void od_start(struct cpufreq_policy *policy)
393 struct od_policy_dbs_info *dbs_info = to_dbs_info(policy->governor_data);
395 dbs_info->sample_type = OD_NORMAL_SAMPLE;
396 ondemand_powersave_bias_init(policy);
399 static struct od_ops od_ops = {
400 .powersave_bias_target = generic_powersave_bias_target,
403 static struct dbs_governor od_dbs_gov = {
404 .gov = CPUFREQ_DBS_GOVERNOR_INITIALIZER("ondemand"),
405 .kobj_type = { .default_attrs = od_attributes },
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,
414 #define CPU_FREQ_GOV_ONDEMAND (&od_dbs_gov.gov)
416 static void od_set_powersave_bias(unsigned int powersave_bias)
418 unsigned int cpu;
419 cpumask_t done;
421 default_powersave_bias = powersave_bias;
422 cpumask_clear(&done);
424 get_online_cpus();
425 for_each_online_cpu(cpu) {
426 struct cpufreq_policy *policy;
427 struct policy_dbs_info *policy_dbs;
428 struct dbs_data *dbs_data;
429 struct od_dbs_tuners *od_tuners;
431 if (cpumask_test_cpu(cpu, &done))
432 continue;
434 policy = cpufreq_cpu_get_raw(cpu);
435 if (!policy || policy->governor != CPU_FREQ_GOV_ONDEMAND)
436 continue;
438 policy_dbs = policy->governor_data;
439 if (!policy_dbs)
440 continue;
442 cpumask_or(&done, &done, policy->cpus);
444 dbs_data = policy_dbs->dbs_data;
445 od_tuners = dbs_data->tuners;
446 od_tuners->powersave_bias = default_powersave_bias;
448 put_online_cpus();
451 void od_register_powersave_bias_handler(unsigned int (*f)
452 (struct cpufreq_policy *, unsigned int, unsigned int),
453 unsigned int powersave_bias)
455 od_ops.powersave_bias_target = f;
456 od_set_powersave_bias(powersave_bias);
458 EXPORT_SYMBOL_GPL(od_register_powersave_bias_handler);
460 void od_unregister_powersave_bias_handler(void)
462 od_ops.powersave_bias_target = generic_powersave_bias_target;
463 od_set_powersave_bias(0);
465 EXPORT_SYMBOL_GPL(od_unregister_powersave_bias_handler);
467 static int __init cpufreq_gov_dbs_init(void)
469 return cpufreq_register_governor(CPU_FREQ_GOV_ONDEMAND);
472 static void __exit cpufreq_gov_dbs_exit(void)
474 cpufreq_unregister_governor(CPU_FREQ_GOV_ONDEMAND);
477 MODULE_AUTHOR("Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>");
478 MODULE_AUTHOR("Alexey Starikovskiy <alexey.y.starikovskiy@intel.com>");
479 MODULE_DESCRIPTION("'cpufreq_ondemand' - A dynamic cpufreq governor for "
480 "Low Latency Frequency Transition capable processors");
481 MODULE_LICENSE("GPL");
483 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND
484 struct cpufreq_governor *cpufreq_default_governor(void)
486 return CPU_FREQ_GOV_ONDEMAND;
489 fs_initcall(cpufreq_gov_dbs_init);
490 #else
491 module_init(cpufreq_gov_dbs_init);
492 #endif
493 module_exit(cpufreq_gov_dbs_exit);