]> git.proxmox.com Git - mirror_ubuntu-bionic-kernel.git/blame - drivers/cpufreq/cpufreq_ondemand.c
[CPUFREQ] powernow-k8: Set transition latency to 1 if ACPI tables export 0
[mirror_ubuntu-bionic-kernel.git] / drivers / cpufreq / cpufreq_ondemand.c
CommitLineData
1da177e4
LT
1/*
2 * drivers/cpufreq/cpufreq_ondemand.c
3 *
4 * Copyright (C) 2001 Russell King
5 * (C) 2003 Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>.
6 * Jun Nakajima <jun.nakajima@intel.com>
7 *
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.
11 */
12
13#include <linux/kernel.h>
14#include <linux/module.h>
1da177e4 15#include <linux/init.h>
1da177e4 16#include <linux/cpufreq.h>
138a0128 17#include <linux/cpu.h>
1da177e4
LT
18#include <linux/jiffies.h>
19#include <linux/kernel_stat.h>
3fc54d37 20#include <linux/mutex.h>
80800913 21#include <linux/hrtimer.h>
22#include <linux/tick.h>
23#include <linux/ktime.h>
9411b4ef 24#include <linux/sched.h>
1da177e4
LT
25
26/*
27 * dbs is used in this file as a shortform for demandbased switching
28 * It helps to keep variable names smaller, simpler
29 */
30
e9d95bf7 31#define DEF_FREQUENCY_DOWN_DIFFERENTIAL (10)
1da177e4 32#define DEF_FREQUENCY_UP_THRESHOLD (80)
80800913 33#define MICRO_FREQUENCY_DOWN_DIFFERENTIAL (3)
34#define MICRO_FREQUENCY_UP_THRESHOLD (95)
cef9615a 35#define MICRO_FREQUENCY_MIN_SAMPLE_RATE (10000)
c29f1403 36#define MIN_FREQUENCY_UP_THRESHOLD (11)
1da177e4
LT
37#define MAX_FREQUENCY_UP_THRESHOLD (100)
38
32ee8c3e
DJ
39/*
40 * The polling frequency of this governor depends on the capability of
1da177e4 41 * the processor. Default polling frequency is 1000 times the transition
32ee8c3e
DJ
42 * latency of the processor. The governor will work on any processor with
43 * transition latency <= 10mS, using appropriate sampling
1da177e4
LT
44 * rate.
45 * For CPUs with transition latency > 10mS (mostly drivers with CPUFREQ_ETERNAL)
46 * this governor will not work.
47 * All times here are in uS.
48 */
df8b59be 49#define MIN_SAMPLING_RATE_RATIO (2)
112124ab 50
cef9615a
TR
51static unsigned int min_sampling_rate;
52
112124ab 53#define LATENCY_MULTIPLIER (1000)
cef9615a 54#define MIN_LATENCY_MULTIPLIER (100)
1c256245 55#define TRANSITION_LATENCY_LIMIT (10 * 1000 * 1000)
1da177e4 56
c4028958
DH
57static void do_dbs_timer(struct work_struct *work);
58
59/* Sampling types */
529af7a1 60enum {DBS_NORMAL_SAMPLE, DBS_SUB_SAMPLE};
1da177e4
LT
61
62struct cpu_dbs_info_s {
ccb2fe20
VP
63 cputime64_t prev_cpu_idle;
64 cputime64_t prev_cpu_wall;
80800913 65 cputime64_t prev_cpu_nice;
32ee8c3e 66 struct cpufreq_policy *cur_policy;
2b03f891 67 struct delayed_work work;
05ca0350
AS
68 struct cpufreq_frequency_table *freq_table;
69 unsigned int freq_lo;
70 unsigned int freq_lo_jiffies;
71 unsigned int freq_hi_jiffies;
529af7a1
VP
72 int cpu;
73 unsigned int enable:1,
2b03f891 74 sample_type:1;
1da177e4
LT
75};
76static DEFINE_PER_CPU(struct cpu_dbs_info_s, cpu_dbs_info);
77
78static unsigned int dbs_enable; /* number of CPUs using this policy */
79
4ec223d0
VP
80/*
81 * DEADLOCK ALERT! There is a ordering requirement between cpu_hotplug
82 * lock and dbs_mutex. cpu_hotplug lock should always be held before
83 * dbs_mutex. If any function that can potentially take cpu_hotplug lock
84 * (like __cpufreq_driver_target()) is being called with dbs_mutex taken, then
85 * cpu_hotplug lock should be taken before that. Note that cpu_hotplug lock
86 * is recursive for the same process. -Venki
b14893a6
MD
87 * DEADLOCK ALERT! (2) : do_dbs_timer() must not take the dbs_mutex, because it
88 * would deadlock with cancel_delayed_work_sync(), which is needed for proper
89 * raceless workqueue teardown.
4ec223d0 90 */
ffac80e9 91static DEFINE_MUTEX(dbs_mutex);
1da177e4 92
2f8a835c 93static struct workqueue_struct *kondemand_wq;
6810b548 94
05ca0350 95static struct dbs_tuners {
32ee8c3e 96 unsigned int sampling_rate;
32ee8c3e 97 unsigned int up_threshold;
e9d95bf7 98 unsigned int down_differential;
32ee8c3e 99 unsigned int ignore_nice;
05ca0350
AS
100 unsigned int powersave_bias;
101} dbs_tuners_ins = {
32ee8c3e 102 .up_threshold = DEF_FREQUENCY_UP_THRESHOLD,
e9d95bf7 103 .down_differential = DEF_FREQUENCY_DOWN_DIFFERENTIAL,
9cbad61b 104 .ignore_nice = 0,
05ca0350 105 .powersave_bias = 0,
1da177e4
LT
106};
107
80800913 108static inline cputime64_t get_cpu_idle_time_jiffy(unsigned int cpu,
109 cputime64_t *wall)
dac1c1a5 110{
ea487615 111 cputime64_t idle_time;
3430502d 112 cputime64_t cur_wall_time;
ea487615 113 cputime64_t busy_time;
ccb2fe20 114
3430502d 115 cur_wall_time = jiffies64_to_cputime64(get_jiffies_64());
ea487615
VP
116 busy_time = cputime64_add(kstat_cpu(cpu).cpustat.user,
117 kstat_cpu(cpu).cpustat.system);
ccb2fe20 118
ea487615
VP
119 busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.irq);
120 busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.softirq);
121 busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.steal);
1ca3abdb 122 busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.nice);
ea487615 123
3430502d 124 idle_time = cputime64_sub(cur_wall_time, busy_time);
125 if (wall)
126 *wall = cur_wall_time;
127
ea487615 128 return idle_time;
dac1c1a5
DJ
129}
130
80800913 131static inline cputime64_t get_cpu_idle_time(unsigned int cpu, cputime64_t *wall)
132{
133 u64 idle_time = get_cpu_idle_time_us(cpu, wall);
134
135 if (idle_time == -1ULL)
136 return get_cpu_idle_time_jiffy(cpu, wall);
137
80800913 138 return idle_time;
139}
140
05ca0350
AS
141/*
142 * Find right freq to be set now with powersave_bias on.
143 * Returns the freq_hi to be used right now and will set freq_hi_jiffies,
144 * freq_lo, and freq_lo_jiffies in percpu area for averaging freqs.
145 */
b5ecf60f
AB
146static unsigned int powersave_bias_target(struct cpufreq_policy *policy,
147 unsigned int freq_next,
148 unsigned int relation)
05ca0350
AS
149{
150 unsigned int freq_req, freq_reduc, freq_avg;
151 unsigned int freq_hi, freq_lo;
152 unsigned int index = 0;
153 unsigned int jiffies_total, jiffies_hi, jiffies_lo;
154 struct cpu_dbs_info_s *dbs_info = &per_cpu(cpu_dbs_info, policy->cpu);
155
156 if (!dbs_info->freq_table) {
157 dbs_info->freq_lo = 0;
158 dbs_info->freq_lo_jiffies = 0;
159 return freq_next;
160 }
161
162 cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_next,
163 relation, &index);
164 freq_req = dbs_info->freq_table[index].frequency;
165 freq_reduc = freq_req * dbs_tuners_ins.powersave_bias / 1000;
166 freq_avg = freq_req - freq_reduc;
167
168 /* Find freq bounds for freq_avg in freq_table */
169 index = 0;
170 cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_avg,
171 CPUFREQ_RELATION_H, &index);
172 freq_lo = dbs_info->freq_table[index].frequency;
173 index = 0;
174 cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_avg,
175 CPUFREQ_RELATION_L, &index);
176 freq_hi = dbs_info->freq_table[index].frequency;
177
178 /* Find out how long we have to be in hi and lo freqs */
179 if (freq_hi == freq_lo) {
180 dbs_info->freq_lo = 0;
181 dbs_info->freq_lo_jiffies = 0;
182 return freq_lo;
183 }
184 jiffies_total = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
185 jiffies_hi = (freq_avg - freq_lo) * jiffies_total;
186 jiffies_hi += ((freq_hi - freq_lo) / 2);
187 jiffies_hi /= (freq_hi - freq_lo);
188 jiffies_lo = jiffies_total - jiffies_hi;
189 dbs_info->freq_lo = freq_lo;
190 dbs_info->freq_lo_jiffies = jiffies_lo;
191 dbs_info->freq_hi_jiffies = jiffies_hi;
192 return freq_hi;
193}
194
195static void ondemand_powersave_bias_init(void)
196{
197 int i;
198 for_each_online_cpu(i) {
199 struct cpu_dbs_info_s *dbs_info = &per_cpu(cpu_dbs_info, i);
200 dbs_info->freq_table = cpufreq_frequency_get_table(i);
201 dbs_info->freq_lo = 0;
202 }
203}
204
1da177e4
LT
205/************************** sysfs interface ************************/
206static ssize_t show_sampling_rate_max(struct cpufreq_policy *policy, char *buf)
207{
9411b4ef
TR
208 static int print_once;
209
210 if (!print_once) {
211 printk(KERN_INFO "CPUFREQ: ondemand sampling_rate_max "
212 "sysfs file is deprecated - used by: %s\n",
213 current->comm);
214 print_once = 1;
215 }
cef9615a 216 return sprintf(buf, "%u\n", -1U);
1da177e4
LT
217}
218
219static ssize_t show_sampling_rate_min(struct cpufreq_policy *policy, char *buf)
220{
9411b4ef
TR
221 static int print_once;
222
223 if (!print_once) {
224 printk(KERN_INFO "CPUFREQ: ondemand sampling_rate_min "
225 "sysfs file is deprecated - used by: %s\n",
226 current->comm);
227 print_once = 1;
228 }
cef9615a 229 return sprintf(buf, "%u\n", min_sampling_rate);
1da177e4
LT
230}
231
32ee8c3e
DJ
232#define define_one_ro(_name) \
233static struct freq_attr _name = \
1da177e4
LT
234__ATTR(_name, 0444, show_##_name, NULL)
235
236define_one_ro(sampling_rate_max);
237define_one_ro(sampling_rate_min);
238
239/* cpufreq_ondemand Governor Tunables */
240#define show_one(file_name, object) \
241static ssize_t show_##file_name \
242(struct cpufreq_policy *unused, char *buf) \
243{ \
244 return sprintf(buf, "%u\n", dbs_tuners_ins.object); \
245}
246show_one(sampling_rate, sampling_rate);
1da177e4 247show_one(up_threshold, up_threshold);
001893cd 248show_one(ignore_nice_load, ignore_nice);
05ca0350 249show_one(powersave_bias, powersave_bias);
1da177e4 250
32ee8c3e 251static ssize_t store_sampling_rate(struct cpufreq_policy *unused,
1da177e4
LT
252 const char *buf, size_t count)
253{
254 unsigned int input;
255 int ret;
ffac80e9 256 ret = sscanf(buf, "%u", &input);
1da177e4 257
3fc54d37 258 mutex_lock(&dbs_mutex);
112124ab 259 if (ret != 1) {
3fc54d37 260 mutex_unlock(&dbs_mutex);
1da177e4
LT
261 return -EINVAL;
262 }
cef9615a 263 dbs_tuners_ins.sampling_rate = max(input, min_sampling_rate);
3fc54d37 264 mutex_unlock(&dbs_mutex);
1da177e4
LT
265
266 return count;
267}
268
32ee8c3e 269static ssize_t store_up_threshold(struct cpufreq_policy *unused,
1da177e4
LT
270 const char *buf, size_t count)
271{
272 unsigned int input;
273 int ret;
ffac80e9 274 ret = sscanf(buf, "%u", &input);
1da177e4 275
3fc54d37 276 mutex_lock(&dbs_mutex);
32ee8c3e 277 if (ret != 1 || input > MAX_FREQUENCY_UP_THRESHOLD ||
c29f1403 278 input < MIN_FREQUENCY_UP_THRESHOLD) {
3fc54d37 279 mutex_unlock(&dbs_mutex);
1da177e4
LT
280 return -EINVAL;
281 }
282
283 dbs_tuners_ins.up_threshold = input;
3fc54d37 284 mutex_unlock(&dbs_mutex);
1da177e4
LT
285
286 return count;
287}
288
001893cd 289static ssize_t store_ignore_nice_load(struct cpufreq_policy *policy,
3d5ee9e5
DJ
290 const char *buf, size_t count)
291{
292 unsigned int input;
293 int ret;
294
295 unsigned int j;
32ee8c3e 296
ffac80e9 297 ret = sscanf(buf, "%u", &input);
2b03f891 298 if (ret != 1)
3d5ee9e5
DJ
299 return -EINVAL;
300
2b03f891 301 if (input > 1)
3d5ee9e5 302 input = 1;
32ee8c3e 303
3fc54d37 304 mutex_lock(&dbs_mutex);
2b03f891 305 if (input == dbs_tuners_ins.ignore_nice) { /* nothing to do */
3fc54d37 306 mutex_unlock(&dbs_mutex);
3d5ee9e5
DJ
307 return count;
308 }
309 dbs_tuners_ins.ignore_nice = input;
310
ccb2fe20 311 /* we need to re-evaluate prev_cpu_idle */
dac1c1a5 312 for_each_online_cpu(j) {
ccb2fe20
VP
313 struct cpu_dbs_info_s *dbs_info;
314 dbs_info = &per_cpu(cpu_dbs_info, j);
3430502d 315 dbs_info->prev_cpu_idle = get_cpu_idle_time(j,
316 &dbs_info->prev_cpu_wall);
1ca3abdb
VP
317 if (dbs_tuners_ins.ignore_nice)
318 dbs_info->prev_cpu_nice = kstat_cpu(j).cpustat.nice;
319
3d5ee9e5 320 }
3fc54d37 321 mutex_unlock(&dbs_mutex);
3d5ee9e5
DJ
322
323 return count;
324}
325
05ca0350
AS
326static ssize_t store_powersave_bias(struct cpufreq_policy *unused,
327 const char *buf, size_t count)
328{
329 unsigned int input;
330 int ret;
331 ret = sscanf(buf, "%u", &input);
332
333 if (ret != 1)
334 return -EINVAL;
335
336 if (input > 1000)
337 input = 1000;
338
339 mutex_lock(&dbs_mutex);
340 dbs_tuners_ins.powersave_bias = input;
341 ondemand_powersave_bias_init();
342 mutex_unlock(&dbs_mutex);
343
344 return count;
345}
346
1da177e4
LT
347#define define_one_rw(_name) \
348static struct freq_attr _name = \
349__ATTR(_name, 0644, show_##_name, store_##_name)
350
351define_one_rw(sampling_rate);
1da177e4 352define_one_rw(up_threshold);
001893cd 353define_one_rw(ignore_nice_load);
05ca0350 354define_one_rw(powersave_bias);
1da177e4 355
2b03f891 356static struct attribute *dbs_attributes[] = {
1da177e4
LT
357 &sampling_rate_max.attr,
358 &sampling_rate_min.attr,
359 &sampling_rate.attr,
1da177e4 360 &up_threshold.attr,
001893cd 361 &ignore_nice_load.attr,
05ca0350 362 &powersave_bias.attr,
1da177e4
LT
363 NULL
364};
365
366static struct attribute_group dbs_attr_group = {
367 .attrs = dbs_attributes,
368 .name = "ondemand",
369};
370
371/************************** sysfs end ************************/
372
2f8a835c 373static void dbs_check_cpu(struct cpu_dbs_info_s *this_dbs_info)
1da177e4 374{
c43aa3bd 375 unsigned int max_load_freq;
1da177e4
LT
376
377 struct cpufreq_policy *policy;
378 unsigned int j;
379
1da177e4
LT
380 if (!this_dbs_info->enable)
381 return;
382
05ca0350 383 this_dbs_info->freq_lo = 0;
1da177e4 384 policy = this_dbs_info->cur_policy;
ea487615 385
32ee8c3e 386 /*
c29f1403
DJ
387 * Every sampling_rate, we check, if current idle time is less
388 * than 20% (default), then we try to increase frequency
ccb2fe20 389 * Every sampling_rate, we look for a the lowest
c29f1403
DJ
390 * frequency which can sustain the load while keeping idle time over
391 * 30%. If such a frequency exist, we try to decrease to this frequency.
1da177e4 392 *
32ee8c3e
DJ
393 * Any frequency increase takes it to the maximum frequency.
394 * Frequency reduction happens at minimum steps of
395 * 5% (default) of current frequency
1da177e4
LT
396 */
397
c43aa3bd 398 /* Get Absolute Load - in terms of freq */
399 max_load_freq = 0;
400
835481d9 401 for_each_cpu(j, policy->cpus) {
1da177e4 402 struct cpu_dbs_info_s *j_dbs_info;
c43aa3bd 403 cputime64_t cur_wall_time, cur_idle_time;
404 unsigned int idle_time, wall_time;
405 unsigned int load, load_freq;
406 int freq_avg;
1da177e4 407
1da177e4 408 j_dbs_info = &per_cpu(cpu_dbs_info, j);
3430502d 409
410 cur_idle_time = get_cpu_idle_time(j, &cur_wall_time);
411
c43aa3bd 412 wall_time = (unsigned int) cputime64_sub(cur_wall_time,
413 j_dbs_info->prev_cpu_wall);
414 j_dbs_info->prev_cpu_wall = cur_wall_time;
415
c43aa3bd 416 idle_time = (unsigned int) cputime64_sub(cur_idle_time,
ccb2fe20 417 j_dbs_info->prev_cpu_idle);
c43aa3bd 418 j_dbs_info->prev_cpu_idle = cur_idle_time;
1da177e4 419
1ca3abdb
VP
420 if (dbs_tuners_ins.ignore_nice) {
421 cputime64_t cur_nice;
422 unsigned long cur_nice_jiffies;
423
424 cur_nice = cputime64_sub(kstat_cpu(j).cpustat.nice,
425 j_dbs_info->prev_cpu_nice);
426 /*
427 * Assumption: nice time between sampling periods will
428 * be less than 2^32 jiffies for 32 bit sys
429 */
430 cur_nice_jiffies = (unsigned long)
431 cputime64_to_jiffies64(cur_nice);
432
433 j_dbs_info->prev_cpu_nice = kstat_cpu(j).cpustat.nice;
434 idle_time += jiffies_to_usecs(cur_nice_jiffies);
435 }
436
3430502d 437 if (unlikely(!wall_time || wall_time < idle_time))
c43aa3bd 438 continue;
c43aa3bd 439
440 load = 100 * (wall_time - idle_time) / wall_time;
441
442 freq_avg = __cpufreq_driver_getavg(policy, j);
443 if (freq_avg <= 0)
444 freq_avg = policy->cur;
445
446 load_freq = load * freq_avg;
447 if (load_freq > max_load_freq)
448 max_load_freq = load_freq;
1da177e4
LT
449 }
450
ccb2fe20 451 /* Check for frequency increase */
c43aa3bd 452 if (max_load_freq > dbs_tuners_ins.up_threshold * policy->cur) {
c11420a6 453 /* if we are already at full speed then break out early */
05ca0350
AS
454 if (!dbs_tuners_ins.powersave_bias) {
455 if (policy->cur == policy->max)
456 return;
457
458 __cpufreq_driver_target(policy, policy->max,
459 CPUFREQ_RELATION_H);
460 } else {
461 int freq = powersave_bias_target(policy, policy->max,
462 CPUFREQ_RELATION_H);
463 __cpufreq_driver_target(policy, freq,
464 CPUFREQ_RELATION_L);
465 }
1da177e4
LT
466 return;
467 }
468
469 /* Check for frequency decrease */
c29f1403
DJ
470 /* if we cannot reduce the frequency anymore, break out early */
471 if (policy->cur == policy->min)
472 return;
1da177e4 473
c29f1403
DJ
474 /*
475 * The optimal frequency is the frequency that is the lowest that
476 * can support the current CPU usage without triggering the up
477 * policy. To be safe, we focus 10 points under the threshold.
478 */
e9d95bf7 479 if (max_load_freq <
480 (dbs_tuners_ins.up_threshold - dbs_tuners_ins.down_differential) *
481 policy->cur) {
c43aa3bd 482 unsigned int freq_next;
e9d95bf7 483 freq_next = max_load_freq /
484 (dbs_tuners_ins.up_threshold -
485 dbs_tuners_ins.down_differential);
dfde5d62 486
05ca0350
AS
487 if (!dbs_tuners_ins.powersave_bias) {
488 __cpufreq_driver_target(policy, freq_next,
489 CPUFREQ_RELATION_L);
490 } else {
491 int freq = powersave_bias_target(policy, freq_next,
492 CPUFREQ_RELATION_L);
493 __cpufreq_driver_target(policy, freq,
494 CPUFREQ_RELATION_L);
495 }
ccb2fe20 496 }
1da177e4
LT
497}
498
c4028958 499static void do_dbs_timer(struct work_struct *work)
32ee8c3e 500{
529af7a1
VP
501 struct cpu_dbs_info_s *dbs_info =
502 container_of(work, struct cpu_dbs_info_s, work.work);
503 unsigned int cpu = dbs_info->cpu;
504 int sample_type = dbs_info->sample_type;
505
1ce28d6b
AS
506 /* We want all CPUs to do sampling nearly on same jiffy */
507 int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
c4028958 508
1ce28d6b 509 delay -= jiffies % delay;
2f8a835c 510
56463b78 511 if (lock_policy_rwsem_write(cpu) < 0)
2cd7cbdf 512 return;
56463b78
VP
513
514 if (!dbs_info->enable) {
515 unlock_policy_rwsem_write(cpu);
516 return;
517 }
518
05ca0350 519 /* Common NORMAL_SAMPLE setup */
c4028958 520 dbs_info->sample_type = DBS_NORMAL_SAMPLE;
05ca0350 521 if (!dbs_tuners_ins.powersave_bias ||
c4028958 522 sample_type == DBS_NORMAL_SAMPLE) {
05ca0350 523 dbs_check_cpu(dbs_info);
05ca0350
AS
524 if (dbs_info->freq_lo) {
525 /* Setup timer for SUB_SAMPLE */
c4028958 526 dbs_info->sample_type = DBS_SUB_SAMPLE;
05ca0350
AS
527 delay = dbs_info->freq_hi_jiffies;
528 }
529 } else {
530 __cpufreq_driver_target(dbs_info->cur_policy,
2b03f891 531 dbs_info->freq_lo, CPUFREQ_RELATION_H);
05ca0350 532 }
1ce28d6b 533 queue_delayed_work_on(cpu, kondemand_wq, &dbs_info->work, delay);
56463b78 534 unlock_policy_rwsem_write(cpu);
32ee8c3e 535}
1da177e4 536
529af7a1 537static inline void dbs_timer_init(struct cpu_dbs_info_s *dbs_info)
1da177e4 538{
1ce28d6b
AS
539 /* We want all CPUs to do sampling nearly on same jiffy */
540 int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate);
541 delay -= jiffies % delay;
2f8a835c 542
c18a1483 543 dbs_info->enable = 1;
05ca0350 544 ondemand_powersave_bias_init();
c4028958 545 dbs_info->sample_type = DBS_NORMAL_SAMPLE;
28287033 546 INIT_DELAYED_WORK_DEFERRABLE(&dbs_info->work, do_dbs_timer);
529af7a1 547 queue_delayed_work_on(dbs_info->cpu, kondemand_wq, &dbs_info->work,
2b03f891 548 delay);
1da177e4
LT
549}
550
2cd7cbdf 551static inline void dbs_timer_exit(struct cpu_dbs_info_s *dbs_info)
1da177e4 552{
2cd7cbdf 553 dbs_info->enable = 0;
b14893a6 554 cancel_delayed_work_sync(&dbs_info->work);
1da177e4
LT
555}
556
557static int cpufreq_governor_dbs(struct cpufreq_policy *policy,
558 unsigned int event)
559{
560 unsigned int cpu = policy->cpu;
561 struct cpu_dbs_info_s *this_dbs_info;
562 unsigned int j;
914f7c31 563 int rc;
1da177e4
LT
564
565 this_dbs_info = &per_cpu(cpu_dbs_info, cpu);
566
567 switch (event) {
568 case CPUFREQ_GOV_START:
ffac80e9 569 if ((!cpu_online(cpu)) || (!policy->cur))
1da177e4
LT
570 return -EINVAL;
571
1da177e4
LT
572 if (this_dbs_info->enable) /* Already enabled */
573 break;
32ee8c3e 574
3fc54d37 575 mutex_lock(&dbs_mutex);
2f8a835c 576 dbs_enable++;
914f7c31
JG
577
578 rc = sysfs_create_group(&policy->kobj, &dbs_attr_group);
579 if (rc) {
914f7c31
JG
580 dbs_enable--;
581 mutex_unlock(&dbs_mutex);
582 return rc;
583 }
584
835481d9 585 for_each_cpu(j, policy->cpus) {
1da177e4
LT
586 struct cpu_dbs_info_s *j_dbs_info;
587 j_dbs_info = &per_cpu(cpu_dbs_info, j);
588 j_dbs_info->cur_policy = policy;
32ee8c3e 589
3430502d 590 j_dbs_info->prev_cpu_idle = get_cpu_idle_time(j,
591 &j_dbs_info->prev_cpu_wall);
1ca3abdb
VP
592 if (dbs_tuners_ins.ignore_nice) {
593 j_dbs_info->prev_cpu_nice =
594 kstat_cpu(j).cpustat.nice;
595 }
1da177e4 596 }
529af7a1 597 this_dbs_info->cpu = cpu;
1da177e4
LT
598 /*
599 * Start the timerschedule work, when this governor
600 * is used for first time
601 */
602 if (dbs_enable == 1) {
603 unsigned int latency;
604 /* policy latency is in nS. Convert it to uS first */
df8b59be
DJ
605 latency = policy->cpuinfo.transition_latency / 1000;
606 if (latency == 0)
607 latency = 1;
cef9615a
TR
608 /* Bring kernel and HW constraints together */
609 min_sampling_rate = max(min_sampling_rate,
610 MIN_LATENCY_MULTIPLIER * latency);
611 dbs_tuners_ins.sampling_rate =
612 max(min_sampling_rate,
613 latency * LATENCY_MULTIPLIER);
1da177e4 614 }
529af7a1 615 dbs_timer_init(this_dbs_info);
32ee8c3e 616
3fc54d37 617 mutex_unlock(&dbs_mutex);
1da177e4
LT
618 break;
619
620 case CPUFREQ_GOV_STOP:
3fc54d37 621 mutex_lock(&dbs_mutex);
2cd7cbdf 622 dbs_timer_exit(this_dbs_info);
1da177e4
LT
623 sysfs_remove_group(&policy->kobj, &dbs_attr_group);
624 dbs_enable--;
3fc54d37 625 mutex_unlock(&dbs_mutex);
1da177e4
LT
626
627 break;
628
629 case CPUFREQ_GOV_LIMITS:
3fc54d37 630 mutex_lock(&dbs_mutex);
1da177e4 631 if (policy->max < this_dbs_info->cur_policy->cur)
ffac80e9 632 __cpufreq_driver_target(this_dbs_info->cur_policy,
2b03f891 633 policy->max, CPUFREQ_RELATION_H);
1da177e4 634 else if (policy->min > this_dbs_info->cur_policy->cur)
ffac80e9 635 __cpufreq_driver_target(this_dbs_info->cur_policy,
2b03f891 636 policy->min, CPUFREQ_RELATION_L);
3fc54d37 637 mutex_unlock(&dbs_mutex);
1da177e4
LT
638 break;
639 }
640 return 0;
641}
642
c4d14bc0
SW
643#ifndef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND
644static
645#endif
1c256245
TR
646struct cpufreq_governor cpufreq_gov_ondemand = {
647 .name = "ondemand",
648 .governor = cpufreq_governor_dbs,
649 .max_transition_latency = TRANSITION_LATENCY_LIMIT,
650 .owner = THIS_MODULE,
1da177e4 651};
1da177e4
LT
652
653static int __init cpufreq_gov_dbs_init(void)
654{
888a794c 655 int err;
80800913 656 cputime64_t wall;
4f6e6b9f
AR
657 u64 idle_time;
658 int cpu = get_cpu();
80800913 659
4f6e6b9f
AR
660 idle_time = get_cpu_idle_time_us(cpu, &wall);
661 put_cpu();
80800913 662 if (idle_time != -1ULL) {
663 /* Idle micro accounting is supported. Use finer thresholds */
664 dbs_tuners_ins.up_threshold = MICRO_FREQUENCY_UP_THRESHOLD;
665 dbs_tuners_ins.down_differential =
666 MICRO_FREQUENCY_DOWN_DIFFERENTIAL;
cef9615a
TR
667 /*
668 * In no_hz/micro accounting case we set the minimum frequency
669 * not depending on HZ, but fixed (very low). The deferred
670 * timer might skip some samples if idle/sleeping as needed.
671 */
672 min_sampling_rate = MICRO_FREQUENCY_MIN_SAMPLE_RATE;
673 } else {
674 /* For correct statistics, we need 10 ticks for each measure */
675 min_sampling_rate =
676 MIN_SAMPLING_RATE_RATIO * jiffies_to_usecs(10);
80800913 677 }
888a794c 678
56463b78
VP
679 kondemand_wq = create_workqueue("kondemand");
680 if (!kondemand_wq) {
681 printk(KERN_ERR "Creation of kondemand failed\n");
682 return -EFAULT;
683 }
888a794c
AM
684 err = cpufreq_register_governor(&cpufreq_gov_ondemand);
685 if (err)
686 destroy_workqueue(kondemand_wq);
687
688 return err;
1da177e4
LT
689}
690
691static void __exit cpufreq_gov_dbs_exit(void)
692{
1c256245 693 cpufreq_unregister_governor(&cpufreq_gov_ondemand);
56463b78 694 destroy_workqueue(kondemand_wq);
1da177e4
LT
695}
696
697
ffac80e9
VP
698MODULE_AUTHOR("Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>");
699MODULE_AUTHOR("Alexey Starikovskiy <alexey.y.starikovskiy@intel.com>");
700MODULE_DESCRIPTION("'cpufreq_ondemand' - A dynamic cpufreq governor for "
2b03f891 701 "Low Latency Frequency Transition capable processors");
ffac80e9 702MODULE_LICENSE("GPL");
1da177e4 703
6915719b
JW
704#ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND
705fs_initcall(cpufreq_gov_dbs_init);
706#else
1da177e4 707module_init(cpufreq_gov_dbs_init);
6915719b 708#endif
1da177e4 709module_exit(cpufreq_gov_dbs_exit);