]>
Commit | Line | Data |
---|---|---|
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 |
51 | static 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 |
57 | static void do_dbs_timer(struct work_struct *work); |
58 | ||
59 | /* Sampling types */ | |
529af7a1 | 60 | enum {DBS_NORMAL_SAMPLE, DBS_SUB_SAMPLE}; |
1da177e4 LT |
61 | |
62 | struct 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 | }; |
76 | static DEFINE_PER_CPU(struct cpu_dbs_info_s, cpu_dbs_info); | |
77 | ||
78 | static unsigned int dbs_enable; /* number of CPUs using this policy */ | |
79 | ||
4ec223d0 | 80 | /* |
7d26e2d5 | 81 | * dbs_mutex protects data in dbs_tuners_ins from concurrent changes on |
82 | * different CPUs. It protects dbs_enable in governor start/stop. It also | |
83 | * serializes governor limit_change with do_dbs_timer. We do not want | |
84 | * do_dbs_timer to run when user is changing the governor or limits. | |
4ec223d0 | 85 | */ |
ffac80e9 | 86 | static DEFINE_MUTEX(dbs_mutex); |
1da177e4 | 87 | |
2f8a835c | 88 | static struct workqueue_struct *kondemand_wq; |
6810b548 | 89 | |
05ca0350 | 90 | static struct dbs_tuners { |
32ee8c3e | 91 | unsigned int sampling_rate; |
32ee8c3e | 92 | unsigned int up_threshold; |
e9d95bf7 | 93 | unsigned int down_differential; |
32ee8c3e | 94 | unsigned int ignore_nice; |
05ca0350 AS |
95 | unsigned int powersave_bias; |
96 | } dbs_tuners_ins = { | |
32ee8c3e | 97 | .up_threshold = DEF_FREQUENCY_UP_THRESHOLD, |
e9d95bf7 | 98 | .down_differential = DEF_FREQUENCY_DOWN_DIFFERENTIAL, |
9cbad61b | 99 | .ignore_nice = 0, |
05ca0350 | 100 | .powersave_bias = 0, |
1da177e4 LT |
101 | }; |
102 | ||
80800913 | 103 | static inline cputime64_t get_cpu_idle_time_jiffy(unsigned int cpu, |
104 | cputime64_t *wall) | |
dac1c1a5 | 105 | { |
ea487615 | 106 | cputime64_t idle_time; |
3430502d | 107 | cputime64_t cur_wall_time; |
ea487615 | 108 | cputime64_t busy_time; |
ccb2fe20 | 109 | |
3430502d | 110 | cur_wall_time = jiffies64_to_cputime64(get_jiffies_64()); |
ea487615 VP |
111 | busy_time = cputime64_add(kstat_cpu(cpu).cpustat.user, |
112 | kstat_cpu(cpu).cpustat.system); | |
ccb2fe20 | 113 | |
ea487615 VP |
114 | busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.irq); |
115 | busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.softirq); | |
116 | busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.steal); | |
1ca3abdb | 117 | busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.nice); |
ea487615 | 118 | |
3430502d | 119 | idle_time = cputime64_sub(cur_wall_time, busy_time); |
120 | if (wall) | |
121 | *wall = cur_wall_time; | |
122 | ||
ea487615 | 123 | return idle_time; |
dac1c1a5 DJ |
124 | } |
125 | ||
80800913 | 126 | static inline cputime64_t get_cpu_idle_time(unsigned int cpu, cputime64_t *wall) |
127 | { | |
128 | u64 idle_time = get_cpu_idle_time_us(cpu, wall); | |
129 | ||
130 | if (idle_time == -1ULL) | |
131 | return get_cpu_idle_time_jiffy(cpu, wall); | |
132 | ||
80800913 | 133 | return idle_time; |
134 | } | |
135 | ||
05ca0350 AS |
136 | /* |
137 | * Find right freq to be set now with powersave_bias on. | |
138 | * Returns the freq_hi to be used right now and will set freq_hi_jiffies, | |
139 | * freq_lo, and freq_lo_jiffies in percpu area for averaging freqs. | |
140 | */ | |
b5ecf60f AB |
141 | static unsigned int powersave_bias_target(struct cpufreq_policy *policy, |
142 | unsigned int freq_next, | |
143 | unsigned int relation) | |
05ca0350 AS |
144 | { |
145 | unsigned int freq_req, freq_reduc, freq_avg; | |
146 | unsigned int freq_hi, freq_lo; | |
147 | unsigned int index = 0; | |
148 | unsigned int jiffies_total, jiffies_hi, jiffies_lo; | |
149 | struct cpu_dbs_info_s *dbs_info = &per_cpu(cpu_dbs_info, policy->cpu); | |
150 | ||
151 | if (!dbs_info->freq_table) { | |
152 | dbs_info->freq_lo = 0; | |
153 | dbs_info->freq_lo_jiffies = 0; | |
154 | return freq_next; | |
155 | } | |
156 | ||
157 | cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_next, | |
158 | relation, &index); | |
159 | freq_req = dbs_info->freq_table[index].frequency; | |
160 | freq_reduc = freq_req * dbs_tuners_ins.powersave_bias / 1000; | |
161 | freq_avg = freq_req - freq_reduc; | |
162 | ||
163 | /* Find freq bounds for freq_avg in freq_table */ | |
164 | index = 0; | |
165 | cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_avg, | |
166 | CPUFREQ_RELATION_H, &index); | |
167 | freq_lo = dbs_info->freq_table[index].frequency; | |
168 | index = 0; | |
169 | cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_avg, | |
170 | CPUFREQ_RELATION_L, &index); | |
171 | freq_hi = dbs_info->freq_table[index].frequency; | |
172 | ||
173 | /* Find out how long we have to be in hi and lo freqs */ | |
174 | if (freq_hi == freq_lo) { | |
175 | dbs_info->freq_lo = 0; | |
176 | dbs_info->freq_lo_jiffies = 0; | |
177 | return freq_lo; | |
178 | } | |
179 | jiffies_total = usecs_to_jiffies(dbs_tuners_ins.sampling_rate); | |
180 | jiffies_hi = (freq_avg - freq_lo) * jiffies_total; | |
181 | jiffies_hi += ((freq_hi - freq_lo) / 2); | |
182 | jiffies_hi /= (freq_hi - freq_lo); | |
183 | jiffies_lo = jiffies_total - jiffies_hi; | |
184 | dbs_info->freq_lo = freq_lo; | |
185 | dbs_info->freq_lo_jiffies = jiffies_lo; | |
186 | dbs_info->freq_hi_jiffies = jiffies_hi; | |
187 | return freq_hi; | |
188 | } | |
189 | ||
190 | static void ondemand_powersave_bias_init(void) | |
191 | { | |
192 | int i; | |
193 | for_each_online_cpu(i) { | |
194 | struct cpu_dbs_info_s *dbs_info = &per_cpu(cpu_dbs_info, i); | |
195 | dbs_info->freq_table = cpufreq_frequency_get_table(i); | |
196 | dbs_info->freq_lo = 0; | |
197 | } | |
198 | } | |
199 | ||
1da177e4 LT |
200 | /************************** sysfs interface ************************/ |
201 | static ssize_t show_sampling_rate_max(struct cpufreq_policy *policy, char *buf) | |
202 | { | |
4f4d1ad6 TR |
203 | printk_once(KERN_INFO "CPUFREQ: ondemand sampling_rate_max " |
204 | "sysfs file is deprecated - used by: %s\n", current->comm); | |
cef9615a | 205 | return sprintf(buf, "%u\n", -1U); |
1da177e4 LT |
206 | } |
207 | ||
208 | static ssize_t show_sampling_rate_min(struct cpufreq_policy *policy, char *buf) | |
209 | { | |
cef9615a | 210 | return sprintf(buf, "%u\n", min_sampling_rate); |
1da177e4 LT |
211 | } |
212 | ||
32ee8c3e DJ |
213 | #define define_one_ro(_name) \ |
214 | static struct freq_attr _name = \ | |
1da177e4 LT |
215 | __ATTR(_name, 0444, show_##_name, NULL) |
216 | ||
217 | define_one_ro(sampling_rate_max); | |
218 | define_one_ro(sampling_rate_min); | |
219 | ||
220 | /* cpufreq_ondemand Governor Tunables */ | |
221 | #define show_one(file_name, object) \ | |
222 | static ssize_t show_##file_name \ | |
223 | (struct cpufreq_policy *unused, char *buf) \ | |
224 | { \ | |
225 | return sprintf(buf, "%u\n", dbs_tuners_ins.object); \ | |
226 | } | |
227 | show_one(sampling_rate, sampling_rate); | |
1da177e4 | 228 | show_one(up_threshold, up_threshold); |
001893cd | 229 | show_one(ignore_nice_load, ignore_nice); |
05ca0350 | 230 | show_one(powersave_bias, powersave_bias); |
1da177e4 | 231 | |
32ee8c3e | 232 | static ssize_t store_sampling_rate(struct cpufreq_policy *unused, |
1da177e4 LT |
233 | const char *buf, size_t count) |
234 | { | |
235 | unsigned int input; | |
236 | int ret; | |
ffac80e9 | 237 | ret = sscanf(buf, "%u", &input); |
1da177e4 | 238 | |
3fc54d37 | 239 | mutex_lock(&dbs_mutex); |
112124ab | 240 | if (ret != 1) { |
3fc54d37 | 241 | mutex_unlock(&dbs_mutex); |
1da177e4 LT |
242 | return -EINVAL; |
243 | } | |
cef9615a | 244 | dbs_tuners_ins.sampling_rate = max(input, min_sampling_rate); |
3fc54d37 | 245 | mutex_unlock(&dbs_mutex); |
1da177e4 LT |
246 | |
247 | return count; | |
248 | } | |
249 | ||
32ee8c3e | 250 | static ssize_t store_up_threshold(struct cpufreq_policy *unused, |
1da177e4 LT |
251 | const char *buf, size_t count) |
252 | { | |
253 | unsigned int input; | |
254 | int ret; | |
ffac80e9 | 255 | ret = sscanf(buf, "%u", &input); |
1da177e4 | 256 | |
3fc54d37 | 257 | mutex_lock(&dbs_mutex); |
32ee8c3e | 258 | if (ret != 1 || input > MAX_FREQUENCY_UP_THRESHOLD || |
c29f1403 | 259 | input < MIN_FREQUENCY_UP_THRESHOLD) { |
3fc54d37 | 260 | mutex_unlock(&dbs_mutex); |
1da177e4 LT |
261 | return -EINVAL; |
262 | } | |
263 | ||
264 | dbs_tuners_ins.up_threshold = input; | |
3fc54d37 | 265 | mutex_unlock(&dbs_mutex); |
1da177e4 LT |
266 | |
267 | return count; | |
268 | } | |
269 | ||
001893cd | 270 | static ssize_t store_ignore_nice_load(struct cpufreq_policy *policy, |
3d5ee9e5 DJ |
271 | const char *buf, size_t count) |
272 | { | |
273 | unsigned int input; | |
274 | int ret; | |
275 | ||
276 | unsigned int j; | |
32ee8c3e | 277 | |
ffac80e9 | 278 | ret = sscanf(buf, "%u", &input); |
2b03f891 | 279 | if (ret != 1) |
3d5ee9e5 DJ |
280 | return -EINVAL; |
281 | ||
2b03f891 | 282 | if (input > 1) |
3d5ee9e5 | 283 | input = 1; |
32ee8c3e | 284 | |
3fc54d37 | 285 | mutex_lock(&dbs_mutex); |
2b03f891 | 286 | if (input == dbs_tuners_ins.ignore_nice) { /* nothing to do */ |
3fc54d37 | 287 | mutex_unlock(&dbs_mutex); |
3d5ee9e5 DJ |
288 | return count; |
289 | } | |
290 | dbs_tuners_ins.ignore_nice = input; | |
291 | ||
ccb2fe20 | 292 | /* we need to re-evaluate prev_cpu_idle */ |
dac1c1a5 | 293 | for_each_online_cpu(j) { |
ccb2fe20 VP |
294 | struct cpu_dbs_info_s *dbs_info; |
295 | dbs_info = &per_cpu(cpu_dbs_info, j); | |
3430502d | 296 | dbs_info->prev_cpu_idle = get_cpu_idle_time(j, |
297 | &dbs_info->prev_cpu_wall); | |
1ca3abdb VP |
298 | if (dbs_tuners_ins.ignore_nice) |
299 | dbs_info->prev_cpu_nice = kstat_cpu(j).cpustat.nice; | |
300 | ||
3d5ee9e5 | 301 | } |
3fc54d37 | 302 | mutex_unlock(&dbs_mutex); |
3d5ee9e5 DJ |
303 | |
304 | return count; | |
305 | } | |
306 | ||
05ca0350 AS |
307 | static ssize_t store_powersave_bias(struct cpufreq_policy *unused, |
308 | const char *buf, size_t count) | |
309 | { | |
310 | unsigned int input; | |
311 | int ret; | |
312 | ret = sscanf(buf, "%u", &input); | |
313 | ||
314 | if (ret != 1) | |
315 | return -EINVAL; | |
316 | ||
317 | if (input > 1000) | |
318 | input = 1000; | |
319 | ||
320 | mutex_lock(&dbs_mutex); | |
321 | dbs_tuners_ins.powersave_bias = input; | |
322 | ondemand_powersave_bias_init(); | |
323 | mutex_unlock(&dbs_mutex); | |
324 | ||
325 | return count; | |
326 | } | |
327 | ||
1da177e4 LT |
328 | #define define_one_rw(_name) \ |
329 | static struct freq_attr _name = \ | |
330 | __ATTR(_name, 0644, show_##_name, store_##_name) | |
331 | ||
332 | define_one_rw(sampling_rate); | |
1da177e4 | 333 | define_one_rw(up_threshold); |
001893cd | 334 | define_one_rw(ignore_nice_load); |
05ca0350 | 335 | define_one_rw(powersave_bias); |
1da177e4 | 336 | |
2b03f891 | 337 | static struct attribute *dbs_attributes[] = { |
1da177e4 LT |
338 | &sampling_rate_max.attr, |
339 | &sampling_rate_min.attr, | |
340 | &sampling_rate.attr, | |
1da177e4 | 341 | &up_threshold.attr, |
001893cd | 342 | &ignore_nice_load.attr, |
05ca0350 | 343 | &powersave_bias.attr, |
1da177e4 LT |
344 | NULL |
345 | }; | |
346 | ||
347 | static struct attribute_group dbs_attr_group = { | |
348 | .attrs = dbs_attributes, | |
349 | .name = "ondemand", | |
350 | }; | |
351 | ||
352 | /************************** sysfs end ************************/ | |
353 | ||
2f8a835c | 354 | static void dbs_check_cpu(struct cpu_dbs_info_s *this_dbs_info) |
1da177e4 | 355 | { |
c43aa3bd | 356 | unsigned int max_load_freq; |
1da177e4 LT |
357 | |
358 | struct cpufreq_policy *policy; | |
359 | unsigned int j; | |
360 | ||
1da177e4 LT |
361 | if (!this_dbs_info->enable) |
362 | return; | |
363 | ||
05ca0350 | 364 | this_dbs_info->freq_lo = 0; |
1da177e4 | 365 | policy = this_dbs_info->cur_policy; |
ea487615 | 366 | |
32ee8c3e | 367 | /* |
c29f1403 DJ |
368 | * Every sampling_rate, we check, if current idle time is less |
369 | * than 20% (default), then we try to increase frequency | |
ccb2fe20 | 370 | * Every sampling_rate, we look for a the lowest |
c29f1403 DJ |
371 | * frequency which can sustain the load while keeping idle time over |
372 | * 30%. If such a frequency exist, we try to decrease to this frequency. | |
1da177e4 | 373 | * |
32ee8c3e DJ |
374 | * Any frequency increase takes it to the maximum frequency. |
375 | * Frequency reduction happens at minimum steps of | |
376 | * 5% (default) of current frequency | |
1da177e4 LT |
377 | */ |
378 | ||
c43aa3bd | 379 | /* Get Absolute Load - in terms of freq */ |
380 | max_load_freq = 0; | |
381 | ||
835481d9 | 382 | for_each_cpu(j, policy->cpus) { |
1da177e4 | 383 | struct cpu_dbs_info_s *j_dbs_info; |
c43aa3bd | 384 | cputime64_t cur_wall_time, cur_idle_time; |
385 | unsigned int idle_time, wall_time; | |
386 | unsigned int load, load_freq; | |
387 | int freq_avg; | |
1da177e4 | 388 | |
1da177e4 | 389 | j_dbs_info = &per_cpu(cpu_dbs_info, j); |
3430502d | 390 | |
391 | cur_idle_time = get_cpu_idle_time(j, &cur_wall_time); | |
392 | ||
c43aa3bd | 393 | wall_time = (unsigned int) cputime64_sub(cur_wall_time, |
394 | j_dbs_info->prev_cpu_wall); | |
395 | j_dbs_info->prev_cpu_wall = cur_wall_time; | |
396 | ||
c43aa3bd | 397 | idle_time = (unsigned int) cputime64_sub(cur_idle_time, |
ccb2fe20 | 398 | j_dbs_info->prev_cpu_idle); |
c43aa3bd | 399 | j_dbs_info->prev_cpu_idle = cur_idle_time; |
1da177e4 | 400 | |
1ca3abdb VP |
401 | if (dbs_tuners_ins.ignore_nice) { |
402 | cputime64_t cur_nice; | |
403 | unsigned long cur_nice_jiffies; | |
404 | ||
405 | cur_nice = cputime64_sub(kstat_cpu(j).cpustat.nice, | |
406 | j_dbs_info->prev_cpu_nice); | |
407 | /* | |
408 | * Assumption: nice time between sampling periods will | |
409 | * be less than 2^32 jiffies for 32 bit sys | |
410 | */ | |
411 | cur_nice_jiffies = (unsigned long) | |
412 | cputime64_to_jiffies64(cur_nice); | |
413 | ||
414 | j_dbs_info->prev_cpu_nice = kstat_cpu(j).cpustat.nice; | |
415 | idle_time += jiffies_to_usecs(cur_nice_jiffies); | |
416 | } | |
417 | ||
3430502d | 418 | if (unlikely(!wall_time || wall_time < idle_time)) |
c43aa3bd | 419 | continue; |
c43aa3bd | 420 | |
421 | load = 100 * (wall_time - idle_time) / wall_time; | |
422 | ||
423 | freq_avg = __cpufreq_driver_getavg(policy, j); | |
424 | if (freq_avg <= 0) | |
425 | freq_avg = policy->cur; | |
426 | ||
427 | load_freq = load * freq_avg; | |
428 | if (load_freq > max_load_freq) | |
429 | max_load_freq = load_freq; | |
1da177e4 LT |
430 | } |
431 | ||
ccb2fe20 | 432 | /* Check for frequency increase */ |
c43aa3bd | 433 | if (max_load_freq > dbs_tuners_ins.up_threshold * policy->cur) { |
c11420a6 | 434 | /* if we are already at full speed then break out early */ |
05ca0350 AS |
435 | if (!dbs_tuners_ins.powersave_bias) { |
436 | if (policy->cur == policy->max) | |
437 | return; | |
438 | ||
439 | __cpufreq_driver_target(policy, policy->max, | |
440 | CPUFREQ_RELATION_H); | |
441 | } else { | |
442 | int freq = powersave_bias_target(policy, policy->max, | |
443 | CPUFREQ_RELATION_H); | |
444 | __cpufreq_driver_target(policy, freq, | |
445 | CPUFREQ_RELATION_L); | |
446 | } | |
1da177e4 LT |
447 | return; |
448 | } | |
449 | ||
450 | /* Check for frequency decrease */ | |
c29f1403 DJ |
451 | /* if we cannot reduce the frequency anymore, break out early */ |
452 | if (policy->cur == policy->min) | |
453 | return; | |
1da177e4 | 454 | |
c29f1403 DJ |
455 | /* |
456 | * The optimal frequency is the frequency that is the lowest that | |
457 | * can support the current CPU usage without triggering the up | |
458 | * policy. To be safe, we focus 10 points under the threshold. | |
459 | */ | |
e9d95bf7 | 460 | if (max_load_freq < |
461 | (dbs_tuners_ins.up_threshold - dbs_tuners_ins.down_differential) * | |
462 | policy->cur) { | |
c43aa3bd | 463 | unsigned int freq_next; |
e9d95bf7 | 464 | freq_next = max_load_freq / |
465 | (dbs_tuners_ins.up_threshold - | |
466 | dbs_tuners_ins.down_differential); | |
dfde5d62 | 467 | |
05ca0350 AS |
468 | if (!dbs_tuners_ins.powersave_bias) { |
469 | __cpufreq_driver_target(policy, freq_next, | |
470 | CPUFREQ_RELATION_L); | |
471 | } else { | |
472 | int freq = powersave_bias_target(policy, freq_next, | |
473 | CPUFREQ_RELATION_L); | |
474 | __cpufreq_driver_target(policy, freq, | |
475 | CPUFREQ_RELATION_L); | |
476 | } | |
ccb2fe20 | 477 | } |
1da177e4 LT |
478 | } |
479 | ||
c4028958 | 480 | static void do_dbs_timer(struct work_struct *work) |
32ee8c3e | 481 | { |
529af7a1 VP |
482 | struct cpu_dbs_info_s *dbs_info = |
483 | container_of(work, struct cpu_dbs_info_s, work.work); | |
484 | unsigned int cpu = dbs_info->cpu; | |
485 | int sample_type = dbs_info->sample_type; | |
486 | ||
1ce28d6b AS |
487 | /* We want all CPUs to do sampling nearly on same jiffy */ |
488 | int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate); | |
c4028958 | 489 | |
1ce28d6b | 490 | delay -= jiffies % delay; |
2f8a835c | 491 | |
7d26e2d5 | 492 | mutex_lock(&dbs_mutex); |
56463b78 VP |
493 | |
494 | if (!dbs_info->enable) { | |
7d26e2d5 | 495 | mutex_unlock(&dbs_mutex); |
56463b78 VP |
496 | return; |
497 | } | |
498 | ||
05ca0350 | 499 | /* Common NORMAL_SAMPLE setup */ |
c4028958 | 500 | dbs_info->sample_type = DBS_NORMAL_SAMPLE; |
05ca0350 | 501 | if (!dbs_tuners_ins.powersave_bias || |
c4028958 | 502 | sample_type == DBS_NORMAL_SAMPLE) { |
05ca0350 | 503 | dbs_check_cpu(dbs_info); |
05ca0350 AS |
504 | if (dbs_info->freq_lo) { |
505 | /* Setup timer for SUB_SAMPLE */ | |
c4028958 | 506 | dbs_info->sample_type = DBS_SUB_SAMPLE; |
05ca0350 AS |
507 | delay = dbs_info->freq_hi_jiffies; |
508 | } | |
509 | } else { | |
510 | __cpufreq_driver_target(dbs_info->cur_policy, | |
2b03f891 | 511 | dbs_info->freq_lo, CPUFREQ_RELATION_H); |
05ca0350 | 512 | } |
1ce28d6b | 513 | queue_delayed_work_on(cpu, kondemand_wq, &dbs_info->work, delay); |
7d26e2d5 | 514 | mutex_unlock(&dbs_mutex); |
32ee8c3e | 515 | } |
1da177e4 | 516 | |
529af7a1 | 517 | static inline void dbs_timer_init(struct cpu_dbs_info_s *dbs_info) |
1da177e4 | 518 | { |
1ce28d6b AS |
519 | /* We want all CPUs to do sampling nearly on same jiffy */ |
520 | int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate); | |
521 | delay -= jiffies % delay; | |
2f8a835c | 522 | |
c18a1483 | 523 | dbs_info->enable = 1; |
05ca0350 | 524 | ondemand_powersave_bias_init(); |
c4028958 | 525 | dbs_info->sample_type = DBS_NORMAL_SAMPLE; |
28287033 | 526 | INIT_DELAYED_WORK_DEFERRABLE(&dbs_info->work, do_dbs_timer); |
529af7a1 | 527 | queue_delayed_work_on(dbs_info->cpu, kondemand_wq, &dbs_info->work, |
2b03f891 | 528 | delay); |
1da177e4 LT |
529 | } |
530 | ||
2cd7cbdf | 531 | static inline void dbs_timer_exit(struct cpu_dbs_info_s *dbs_info) |
1da177e4 | 532 | { |
2cd7cbdf | 533 | dbs_info->enable = 0; |
b14893a6 | 534 | cancel_delayed_work_sync(&dbs_info->work); |
1da177e4 LT |
535 | } |
536 | ||
537 | static int cpufreq_governor_dbs(struct cpufreq_policy *policy, | |
538 | unsigned int event) | |
539 | { | |
540 | unsigned int cpu = policy->cpu; | |
541 | struct cpu_dbs_info_s *this_dbs_info; | |
542 | unsigned int j; | |
914f7c31 | 543 | int rc; |
1da177e4 LT |
544 | |
545 | this_dbs_info = &per_cpu(cpu_dbs_info, cpu); | |
546 | ||
547 | switch (event) { | |
548 | case CPUFREQ_GOV_START: | |
ffac80e9 | 549 | if ((!cpu_online(cpu)) || (!policy->cur)) |
1da177e4 LT |
550 | return -EINVAL; |
551 | ||
1da177e4 LT |
552 | if (this_dbs_info->enable) /* Already enabled */ |
553 | break; | |
32ee8c3e | 554 | |
3fc54d37 | 555 | mutex_lock(&dbs_mutex); |
2f8a835c | 556 | dbs_enable++; |
914f7c31 JG |
557 | |
558 | rc = sysfs_create_group(&policy->kobj, &dbs_attr_group); | |
559 | if (rc) { | |
914f7c31 JG |
560 | dbs_enable--; |
561 | mutex_unlock(&dbs_mutex); | |
562 | return rc; | |
563 | } | |
564 | ||
835481d9 | 565 | for_each_cpu(j, policy->cpus) { |
1da177e4 LT |
566 | struct cpu_dbs_info_s *j_dbs_info; |
567 | j_dbs_info = &per_cpu(cpu_dbs_info, j); | |
568 | j_dbs_info->cur_policy = policy; | |
32ee8c3e | 569 | |
3430502d | 570 | j_dbs_info->prev_cpu_idle = get_cpu_idle_time(j, |
571 | &j_dbs_info->prev_cpu_wall); | |
1ca3abdb VP |
572 | if (dbs_tuners_ins.ignore_nice) { |
573 | j_dbs_info->prev_cpu_nice = | |
574 | kstat_cpu(j).cpustat.nice; | |
575 | } | |
1da177e4 | 576 | } |
529af7a1 | 577 | this_dbs_info->cpu = cpu; |
1da177e4 LT |
578 | /* |
579 | * Start the timerschedule work, when this governor | |
580 | * is used for first time | |
581 | */ | |
582 | if (dbs_enable == 1) { | |
583 | unsigned int latency; | |
584 | /* policy latency is in nS. Convert it to uS first */ | |
df8b59be DJ |
585 | latency = policy->cpuinfo.transition_latency / 1000; |
586 | if (latency == 0) | |
587 | latency = 1; | |
cef9615a TR |
588 | /* Bring kernel and HW constraints together */ |
589 | min_sampling_rate = max(min_sampling_rate, | |
590 | MIN_LATENCY_MULTIPLIER * latency); | |
591 | dbs_tuners_ins.sampling_rate = | |
592 | max(min_sampling_rate, | |
593 | latency * LATENCY_MULTIPLIER); | |
1da177e4 | 594 | } |
3fc54d37 | 595 | mutex_unlock(&dbs_mutex); |
7d26e2d5 | 596 | |
597 | dbs_timer_init(this_dbs_info); | |
1da177e4 LT |
598 | break; |
599 | ||
600 | case CPUFREQ_GOV_STOP: | |
2cd7cbdf | 601 | dbs_timer_exit(this_dbs_info); |
7d26e2d5 | 602 | |
603 | mutex_lock(&dbs_mutex); | |
1da177e4 LT |
604 | sysfs_remove_group(&policy->kobj, &dbs_attr_group); |
605 | dbs_enable--; | |
3fc54d37 | 606 | mutex_unlock(&dbs_mutex); |
1da177e4 LT |
607 | |
608 | break; | |
609 | ||
610 | case CPUFREQ_GOV_LIMITS: | |
3fc54d37 | 611 | mutex_lock(&dbs_mutex); |
1da177e4 | 612 | if (policy->max < this_dbs_info->cur_policy->cur) |
ffac80e9 | 613 | __cpufreq_driver_target(this_dbs_info->cur_policy, |
2b03f891 | 614 | policy->max, CPUFREQ_RELATION_H); |
1da177e4 | 615 | else if (policy->min > this_dbs_info->cur_policy->cur) |
ffac80e9 | 616 | __cpufreq_driver_target(this_dbs_info->cur_policy, |
2b03f891 | 617 | policy->min, CPUFREQ_RELATION_L); |
3fc54d37 | 618 | mutex_unlock(&dbs_mutex); |
1da177e4 LT |
619 | break; |
620 | } | |
621 | return 0; | |
622 | } | |
623 | ||
c4d14bc0 SW |
624 | #ifndef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND |
625 | static | |
626 | #endif | |
1c256245 TR |
627 | struct cpufreq_governor cpufreq_gov_ondemand = { |
628 | .name = "ondemand", | |
629 | .governor = cpufreq_governor_dbs, | |
630 | .max_transition_latency = TRANSITION_LATENCY_LIMIT, | |
631 | .owner = THIS_MODULE, | |
1da177e4 | 632 | }; |
1da177e4 LT |
633 | |
634 | static int __init cpufreq_gov_dbs_init(void) | |
635 | { | |
888a794c | 636 | int err; |
80800913 | 637 | cputime64_t wall; |
4f6e6b9f AR |
638 | u64 idle_time; |
639 | int cpu = get_cpu(); | |
80800913 | 640 | |
4f6e6b9f AR |
641 | idle_time = get_cpu_idle_time_us(cpu, &wall); |
642 | put_cpu(); | |
80800913 | 643 | if (idle_time != -1ULL) { |
644 | /* Idle micro accounting is supported. Use finer thresholds */ | |
645 | dbs_tuners_ins.up_threshold = MICRO_FREQUENCY_UP_THRESHOLD; | |
646 | dbs_tuners_ins.down_differential = | |
647 | MICRO_FREQUENCY_DOWN_DIFFERENTIAL; | |
cef9615a TR |
648 | /* |
649 | * In no_hz/micro accounting case we set the minimum frequency | |
650 | * not depending on HZ, but fixed (very low). The deferred | |
651 | * timer might skip some samples if idle/sleeping as needed. | |
652 | */ | |
653 | min_sampling_rate = MICRO_FREQUENCY_MIN_SAMPLE_RATE; | |
654 | } else { | |
655 | /* For correct statistics, we need 10 ticks for each measure */ | |
656 | min_sampling_rate = | |
657 | MIN_SAMPLING_RATE_RATIO * jiffies_to_usecs(10); | |
80800913 | 658 | } |
888a794c | 659 | |
56463b78 VP |
660 | kondemand_wq = create_workqueue("kondemand"); |
661 | if (!kondemand_wq) { | |
662 | printk(KERN_ERR "Creation of kondemand failed\n"); | |
663 | return -EFAULT; | |
664 | } | |
888a794c AM |
665 | err = cpufreq_register_governor(&cpufreq_gov_ondemand); |
666 | if (err) | |
667 | destroy_workqueue(kondemand_wq); | |
668 | ||
669 | return err; | |
1da177e4 LT |
670 | } |
671 | ||
672 | static void __exit cpufreq_gov_dbs_exit(void) | |
673 | { | |
1c256245 | 674 | cpufreq_unregister_governor(&cpufreq_gov_ondemand); |
56463b78 | 675 | destroy_workqueue(kondemand_wq); |
1da177e4 LT |
676 | } |
677 | ||
678 | ||
ffac80e9 VP |
679 | MODULE_AUTHOR("Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>"); |
680 | MODULE_AUTHOR("Alexey Starikovskiy <alexey.y.starikovskiy@intel.com>"); | |
681 | MODULE_DESCRIPTION("'cpufreq_ondemand' - A dynamic cpufreq governor for " | |
2b03f891 | 682 | "Low Latency Frequency Transition capable processors"); |
ffac80e9 | 683 | MODULE_LICENSE("GPL"); |
1da177e4 | 684 | |
6915719b JW |
685 | #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND |
686 | fs_initcall(cpufreq_gov_dbs_init); | |
687 | #else | |
1da177e4 | 688 | module_init(cpufreq_gov_dbs_init); |
6915719b | 689 | #endif |
1da177e4 | 690 | module_exit(cpufreq_gov_dbs_exit); |