cpufreq_ondemand.c 13 KB
Newer Older
1
// SPDX-License-Identifier: GPL-2.0-only
Linus Torvalds's avatar
Linus Torvalds committed
2 3 4 5 6 7 8 9
/*
 *  drivers/cpufreq/cpufreq_ondemand.c
 *
 *  Copyright (C)  2001 Russell King
 *            (C)  2003 Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>.
 *                      Jun Nakajima <jun.nakajima@intel.com>
 */

10 11
#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt

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"
Linus Torvalds's avatar
Linus Torvalds committed
19

20
/* On-demand governor macros */
Linus Torvalds's avatar
Linus Torvalds committed
21
#define DEF_FREQUENCY_UP_THRESHOLD		(80)
22 23
#define DEF_SAMPLING_DOWN_FACTOR		(1)
#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)
Linus Torvalds's avatar
Linus Torvalds committed
27 28
#define MAX_FREQUENCY_UP_THRESHOLD		(100)

29 30
static struct od_ops od_ops;

31 32
static unsigned int default_powersave_bias;

33 34 35 36 37
/*
 * Not all CPUs want IO time to be accounted as busy; this depends on how
 * efficient idling at a higher frequency/voltage is.
 * Pavel Machek says this is not so for various generations of AMD and old
 * Intel systems.
38
 * Mike Chan (android.com) claims this is also not true for ARM.
39 40 41 42 43 44 45
 * Because of this, whitelist specific known (series) of CPUs by default, and
 * leave all others up to the user.
 */
static int should_io_be_busy(void)
{
#if defined(CONFIG_X86)
	/*
46
	 * For Intel, Core 2 (model 15) and later have an efficient idle.
47 48 49 50 51 52 53
	 */
	if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL &&
			boot_cpu_data.x86 == 6 &&
			boot_cpu_data.x86_model >= 15)
		return 1;
#endif
	return 0;
54 55
}

56 57
/*
 * Find right freq to be set now with powersave_bias on.
58 59
 * Returns the freq_hi to be used right now and will set freq_hi_delay_us,
 * freq_lo, and freq_lo_delay_us in percpu area for averaging freqs.
60
 */
61
static unsigned int generic_powersave_bias_target(struct cpufreq_policy *policy,
62
		unsigned int freq_next, unsigned int relation)
63 64 65
{
	unsigned int freq_req, freq_reduc, freq_avg;
	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 78 79
		return freq_next;
	}

80
	index = cpufreq_frequency_table_target(policy, freq_next, relation);
81
	freq_req = freq_table[index].frequency;
82
	freq_reduc = freq_req * od_tuners->powersave_bias / 1000;
83 84 85
	freq_avg = freq_req - freq_reduc;

	/* Find freq bounds for freq_avg in freq_table */
86
	index = cpufreq_table_find_index_h(policy, freq_avg);
87
	freq_lo = freq_table[index].frequency;
88
	index = cpufreq_table_find_index_l(policy, freq_avg);
89
	freq_hi = freq_table[index].frequency;
90 91 92 93

	/* Find out how long we have to be in hi and lo freqs */
	if (freq_hi == freq_lo) {
		dbs_info->freq_lo = 0;
94
		dbs_info->freq_lo_delay_us = 0;
95 96
		return freq_lo;
	}
97 98 99 100
	delay_hi_us = (freq_avg - freq_lo) * dbs_data->sampling_rate;
	delay_hi_us += (freq_hi - freq_lo) / 2;
	delay_hi_us /= freq_hi - freq_lo;
	dbs_info->freq_hi_delay_us = delay_hi_us;
101
	dbs_info->freq_lo = freq_lo;
102
	dbs_info->freq_lo_delay_us = dbs_data->sampling_rate - delay_hi_us;
103 104 105
	return freq_hi;
}

106
static void ondemand_powersave_bias_init(struct cpufreq_policy *policy)
107
{
108
	struct od_policy_dbs_info *dbs_info = to_dbs_info(policy->governor_data);
109 110

	dbs_info->freq_lo = 0;
111 112
}

113
static void dbs_freq_increase(struct cpufreq_policy *policy, unsigned int freq)
114
{
115 116
	struct policy_dbs_info *policy_dbs = policy->governor_data;
	struct dbs_data *dbs_data = policy_dbs->dbs_data;
117 118 119
	struct od_dbs_tuners *od_tuners = dbs_data->tuners;

	if (od_tuners->powersave_bias)
120
		freq = od_ops.powersave_bias_target(policy, freq,
121
				CPUFREQ_RELATION_H);
122
	else if (policy->cur == policy->max)
123
		return;
124

125
	__cpufreq_driver_target(policy, freq, od_tuners->powersave_bias ?
126 127 128 129 130
			CPUFREQ_RELATION_L : CPUFREQ_RELATION_H);
}

/*
 * Every sampling_rate, we check, if current idle time is less than 20%
131 132
 * (default), then we try to increase frequency. Else, we adjust the frequency
 * proportional to load.
133
 */
134
static void od_update(struct cpufreq_policy *policy)
Linus Torvalds's avatar
Linus Torvalds committed
135
{
136 137
	struct policy_dbs_info *policy_dbs = policy->governor_data;
	struct od_policy_dbs_info *dbs_info = to_dbs_info(policy_dbs);
138
	struct dbs_data *dbs_data = policy_dbs->dbs_data;
139
	struct od_dbs_tuners *od_tuners = dbs_data->tuners;
140
	unsigned int load = dbs_update(policy);
141 142 143 144

	dbs_info->freq_lo = 0;

	/* Check for frequency increase */
145
	if (load > dbs_data->up_threshold) {
146 147
		/* If switching to max speed, apply sampling_down_factor */
		if (policy->cur < policy->max)
148
			policy_dbs->rate_mult = dbs_data->sampling_down_factor;
149
		dbs_freq_increase(policy, policy->max);
150 151
	} else {
		/* Calculate the next frequency proportional to load */
152 153 154 155 156
		unsigned int freq_next, min_f, max_f;

		min_f = policy->cpuinfo.min_freq;
		max_f = policy->cpuinfo.max_freq;
		freq_next = min_f + load * (max_f - min_f) / 100;
157 158

		/* No longer fully busy, reset rate_mult */
159
		policy_dbs->rate_mult = 1;
160

161 162 163 164 165
		if (od_tuners->powersave_bias)
			freq_next = od_ops.powersave_bias_target(policy,
								 freq_next,
								 CPUFREQ_RELATION_L);

166
		__cpufreq_driver_target(policy, freq_next, CPUFREQ_RELATION_C);
167
	}
Linus Torvalds's avatar
Linus Torvalds committed
168 169
}

170
static unsigned int od_dbs_update(struct cpufreq_policy *policy)
171
{
172 173
	struct policy_dbs_info *policy_dbs = policy->governor_data;
	struct dbs_data *dbs_data = policy_dbs->dbs_data;
174
	struct od_policy_dbs_info *dbs_info = to_dbs_info(policy_dbs);
175
	int sample_type = dbs_info->sample_type;
176

177
	/* Common NORMAL_SAMPLE setup */
178
	dbs_info->sample_type = OD_NORMAL_SAMPLE;
179 180 181 182 183
	/*
	 * OD_SUB_SAMPLE doesn't make sense if sample_delay_ns is 0, so ignore
	 * it then.
	 */
	if (sample_type == OD_SUB_SAMPLE && policy_dbs->sample_delay_ns > 0) {
184
		__cpufreq_driver_target(policy, dbs_info->freq_lo,
185
					CPUFREQ_RELATION_H);
186
		return dbs_info->freq_lo_delay_us;
187 188 189 190 191
	}

	od_update(policy);

	if (dbs_info->freq_lo) {
192
		/* Setup SUB_SAMPLE */
193
		dbs_info->sample_type = OD_SUB_SAMPLE;
194
		return dbs_info->freq_hi_delay_us;
195 196
	}

197
	return dbs_data->sampling_rate * policy_dbs->rate_mult;
198 199
}

200
/************************** sysfs interface ************************/
201
static struct dbs_governor od_dbs_gov;
Linus Torvalds's avatar
Linus Torvalds committed
202

203 204
static ssize_t store_io_is_busy(struct gov_attr_set *attr_set, const char *buf,
				size_t count)
205
{
206
	struct dbs_data *dbs_data = to_dbs_data(attr_set);
207 208 209 210 211 212
	unsigned int input;
	int ret;

	ret = sscanf(buf, "%u", &input);
	if (ret != 1)
		return -EINVAL;
213
	dbs_data->io_is_busy = !!input;
214 215

	/* we need to re-evaluate prev_cpu_idle */
216
	gov_update_cpu_data(dbs_data);
217

218 219 220
	return count;
}

221 222
static ssize_t store_up_threshold(struct gov_attr_set *attr_set,
				  const char *buf, size_t count)
Linus Torvalds's avatar
Linus Torvalds committed
223
{
224
	struct dbs_data *dbs_data = to_dbs_data(attr_set);
Linus Torvalds's avatar
Linus Torvalds committed
225 226
	unsigned int input;
	int ret;
227
	ret = sscanf(buf, "%u", &input);
Linus Torvalds's avatar
Linus Torvalds committed
228

229
	if (ret != 1 || input > MAX_FREQUENCY_UP_THRESHOLD ||
230
			input < MIN_FREQUENCY_UP_THRESHOLD) {
Linus Torvalds's avatar
Linus Torvalds committed
231 232
		return -EINVAL;
	}
233

234
	dbs_data->up_threshold = input;
Linus Torvalds's avatar
Linus Torvalds committed
235 236 237
	return count;
}

238 239
static ssize_t store_sampling_down_factor(struct gov_attr_set *attr_set,
					  const char *buf, size_t count)
240
{
241
	struct dbs_data *dbs_data = to_dbs_data(attr_set);
242 243
	struct policy_dbs_info *policy_dbs;
	unsigned int input;
244 245 246 247 248
	int ret;
	ret = sscanf(buf, "%u", &input);

	if (ret != 1 || input > MAX_SAMPLING_DOWN_FACTOR || input < 1)
		return -EINVAL;
249

250
	dbs_data->sampling_down_factor = input;
251 252

	/* Reset down sampling multiplier in case it was active */
253
	list_for_each_entry(policy_dbs, &attr_set->policy_list, list) {
254 255
		/*
		 * Doing this without locking might lead to using different
256
		 * rate_mult values in od_update() and od_dbs_update().
257
		 */
258
		mutex_lock(&policy_dbs->update_mutex);
259
		policy_dbs->rate_mult = 1;
260
		mutex_unlock(&policy_dbs->update_mutex);
261
	}
262

263 264 265
	return count;
}

266 267
static ssize_t store_ignore_nice_load(struct gov_attr_set *attr_set,
				      const char *buf, size_t count)
268
{
269
	struct dbs_data *dbs_data = to_dbs_data(attr_set);
270 271 272
	unsigned int input;
	int ret;

273
	ret = sscanf(buf, "%u", &input);
274
	if (ret != 1)
275 276
		return -EINVAL;

277
	if (input > 1)
278
		input = 1;
279

280
	if (input == dbs_data->ignore_nice_load) { /* nothing to do */
281 282
		return count;
	}
283
	dbs_data->ignore_nice_load = input;
284

285
	/* we need to re-evaluate prev_cpu_idle */
286
	gov_update_cpu_data(dbs_data);
287

288 289 290
	return count;
}

291 292
static ssize_t store_powersave_bias(struct gov_attr_set *attr_set,
				    const char *buf, size_t count)
293
{
294
	struct dbs_data *dbs_data = to_dbs_data(attr_set);
295
	struct od_dbs_tuners *od_tuners = dbs_data->tuners;
296
	struct policy_dbs_info *policy_dbs;
297 298 299 300 301 302 303 304 305 306
	unsigned int input;
	int ret;
	ret = sscanf(buf, "%u", &input);

	if (ret != 1)
		return -EINVAL;

	if (input > 1000)
		input = 1000;

307
	od_tuners->powersave_bias = input;
308

309
	list_for_each_entry(policy_dbs, &attr_set->policy_list, list)
310 311
		ondemand_powersave_bias_init(policy_dbs->policy);

312 313 314
	return count;
}

315 316 317 318
gov_show_one_common(sampling_rate);
gov_show_one_common(up_threshold);
gov_show_one_common(sampling_down_factor);
gov_show_one_common(ignore_nice_load);
319
gov_show_one_common(io_is_busy);
320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335
gov_show_one(od, powersave_bias);

gov_attr_rw(sampling_rate);
gov_attr_rw(io_is_busy);
gov_attr_rw(up_threshold);
gov_attr_rw(sampling_down_factor);
gov_attr_rw(ignore_nice_load);
gov_attr_rw(powersave_bias);

static struct attribute *od_attributes[] = {
	&sampling_rate.attr,
	&up_threshold.attr,
	&sampling_down_factor.attr,
	&ignore_nice_load.attr,
	&powersave_bias.attr,
	&io_is_busy.attr,
Linus Torvalds's avatar
Linus Torvalds committed
336 337 338 339 340
	NULL
};

/************************** sysfs end ************************/

341 342 343 344 345 346 347 348 349 350 351 352 353
static struct policy_dbs_info *od_alloc(void)
{
	struct od_policy_dbs_info *dbs_info;

	dbs_info = kzalloc(sizeof(*dbs_info), GFP_KERNEL);
	return dbs_info ? &dbs_info->policy_dbs : NULL;
}

static void od_free(struct policy_dbs_info *policy_dbs)
{
	kfree(to_dbs_info(policy_dbs));
}

354
static int od_init(struct dbs_data *dbs_data)
355 356 357 358 359
{
	struct od_dbs_tuners *tuners;
	u64 idle_time;
	int cpu;

360
	tuners = kzalloc(sizeof(*tuners), GFP_KERNEL);
361
	if (!tuners)
362 363 364 365 366 367 368
		return -ENOMEM;

	cpu = get_cpu();
	idle_time = get_cpu_idle_time_us(cpu, NULL);
	put_cpu();
	if (idle_time != -1ULL) {
		/* Idle micro accounting is supported. Use finer thresholds */
369
		dbs_data->up_threshold = MICRO_FREQUENCY_UP_THRESHOLD;
370
	} else {
371
		dbs_data->up_threshold = DEF_FREQUENCY_UP_THRESHOLD;
372 373
	}

374 375
	dbs_data->sampling_down_factor = DEF_SAMPLING_DOWN_FACTOR;
	dbs_data->ignore_nice_load = 0;
376
	tuners->powersave_bias = default_powersave_bias;
377
	dbs_data->io_is_busy = should_io_be_busy();
378 379 380 381 382

	dbs_data->tuners = tuners;
	return 0;
}

383
static void od_exit(struct dbs_data *dbs_data)
384 385 386 387
{
	kfree(dbs_data->tuners);
}

388 389
static void od_start(struct cpufreq_policy *policy)
{
390
	struct od_policy_dbs_info *dbs_info = to_dbs_info(policy->governor_data);
391 392

	dbs_info->sample_type = OD_NORMAL_SAMPLE;
393
	ondemand_powersave_bias_init(policy);
394 395
}

396
static struct od_ops od_ops = {
397
	.powersave_bias_target = generic_powersave_bias_target,
398
};
399

400
static struct dbs_governor od_dbs_gov = {
401
	.gov = CPUFREQ_DBS_GOVERNOR_INITIALIZER("ondemand"),
402
	.kobj_type = { .default_attrs = od_attributes },
403
	.gov_dbs_update = od_dbs_update,
404 405
	.alloc = od_alloc,
	.free = od_free,
406 407
	.init = od_init,
	.exit = od_exit,
408
	.start = od_start,
409
};
Linus Torvalds's avatar
Linus Torvalds committed
410

411
#define CPU_FREQ_GOV_ONDEMAND	(&od_dbs_gov.gov)
412

413 414 415 416 417
static void od_set_powersave_bias(unsigned int powersave_bias)
{
	unsigned int cpu;
	cpumask_t done;

418
	default_powersave_bias = powersave_bias;
419 420 421 422
	cpumask_clear(&done);

	get_online_cpus();
	for_each_online_cpu(cpu) {
423
		struct cpufreq_policy *policy;
424
		struct policy_dbs_info *policy_dbs;
425 426
		struct dbs_data *dbs_data;
		struct od_dbs_tuners *od_tuners;
427

428 429 430
		if (cpumask_test_cpu(cpu, &done))
			continue;

431 432 433 434 435
		policy = cpufreq_cpu_get_raw(cpu);
		if (!policy || policy->governor != CPU_FREQ_GOV_ONDEMAND)
			continue;

		policy_dbs = policy->governor_data;
436
		if (!policy_dbs)
437
			continue;
438 439

		cpumask_or(&done, &done, policy->cpus);
440

441
		dbs_data = policy_dbs->dbs_data;
442 443
		od_tuners = dbs_data->tuners;
		od_tuners->powersave_bias = default_powersave_bias;
444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463
	}
	put_online_cpus();
}

void od_register_powersave_bias_handler(unsigned int (*f)
		(struct cpufreq_policy *, unsigned int, unsigned int),
		unsigned int powersave_bias)
{
	od_ops.powersave_bias_target = f;
	od_set_powersave_bias(powersave_bias);
}
EXPORT_SYMBOL_GPL(od_register_powersave_bias_handler);

void od_unregister_powersave_bias_handler(void)
{
	od_ops.powersave_bias_target = generic_powersave_bias_target;
	od_set_powersave_bias(0);
}
EXPORT_SYMBOL_GPL(od_unregister_powersave_bias_handler);

Linus Torvalds's avatar
Linus Torvalds committed
464 465
static int __init cpufreq_gov_dbs_init(void)
{
466
	return cpufreq_register_governor(CPU_FREQ_GOV_ONDEMAND);
Linus Torvalds's avatar
Linus Torvalds committed
467 468 469 470
}

static void __exit cpufreq_gov_dbs_exit(void)
{
471
	cpufreq_unregister_governor(CPU_FREQ_GOV_ONDEMAND);
Linus Torvalds's avatar
Linus Torvalds committed
472 473
}

474 475 476
MODULE_AUTHOR("Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>");
MODULE_AUTHOR("Alexey Starikovskiy <alexey.y.starikovskiy@intel.com>");
MODULE_DESCRIPTION("'cpufreq_ondemand' - A dynamic cpufreq governor for "
477
	"Low Latency Frequency Transition capable processors");
478
MODULE_LICENSE("GPL");
Linus Torvalds's avatar
Linus Torvalds committed
479

480
#ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND
481 482
struct cpufreq_governor *cpufreq_default_governor(void)
{
483
	return CPU_FREQ_GOV_ONDEMAND;
484 485
}

486 487
fs_initcall(cpufreq_gov_dbs_init);
#else
Linus Torvalds's avatar
Linus Torvalds committed
488
module_init(cpufreq_gov_dbs_init);
489
#endif
Linus Torvalds's avatar
Linus Torvalds committed
490
module_exit(cpufreq_gov_dbs_exit);