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 | ||
4471a34f VK |
13 | #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt |
14 | ||
1da177e4 | 15 | #include <linux/cpufreq.h> |
4471a34f VK |
16 | #include <linux/init.h> |
17 | #include <linux/kernel.h> | |
1da177e4 | 18 | #include <linux/kernel_stat.h> |
4471a34f VK |
19 | #include <linux/kobject.h> |
20 | #include <linux/module.h> | |
3fc54d37 | 21 | #include <linux/mutex.h> |
4471a34f | 22 | #include <linux/percpu-defs.h> |
4d5dcc42 | 23 | #include <linux/slab.h> |
4471a34f | 24 | #include <linux/sysfs.h> |
80800913 | 25 | #include <linux/tick.h> |
4471a34f | 26 | #include <linux/types.h> |
fb30809e | 27 | #include <linux/cpu.h> |
1da177e4 | 28 | |
4471a34f | 29 | #include "cpufreq_governor.h" |
1da177e4 | 30 | |
06eb09d1 | 31 | /* On-demand governor macros */ |
e9d95bf7 | 32 | #define DEF_FREQUENCY_DOWN_DIFFERENTIAL (10) |
1da177e4 | 33 | #define DEF_FREQUENCY_UP_THRESHOLD (80) |
3f78a9f7 DN |
34 | #define DEF_SAMPLING_DOWN_FACTOR (1) |
35 | #define MAX_SAMPLING_DOWN_FACTOR (100000) | |
80800913 | 36 | #define MICRO_FREQUENCY_DOWN_DIFFERENTIAL (3) |
37 | #define MICRO_FREQUENCY_UP_THRESHOLD (95) | |
cef9615a | 38 | #define MICRO_FREQUENCY_MIN_SAMPLE_RATE (10000) |
c29f1403 | 39 | #define MIN_FREQUENCY_UP_THRESHOLD (11) |
1da177e4 LT |
40 | #define MAX_FREQUENCY_UP_THRESHOLD (100) |
41 | ||
4471a34f | 42 | static DEFINE_PER_CPU(struct od_cpu_dbs_info_s, od_cpu_dbs_info); |
1da177e4 | 43 | |
fb30809e JS |
44 | static struct od_ops od_ops; |
45 | ||
3e33ee9e FB |
46 | #ifndef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND |
47 | static struct cpufreq_governor cpufreq_gov_ondemand; | |
48 | #endif | |
49 | ||
c2837558 JS |
50 | static unsigned int default_powersave_bias; |
51 | ||
4471a34f | 52 | static void ondemand_powersave_bias_init_cpu(int cpu) |
6b8fcd90 | 53 | { |
4471a34f | 54 | struct od_cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info, cpu); |
6b8fcd90 | 55 | |
4471a34f VK |
56 | dbs_info->freq_table = cpufreq_frequency_get_table(cpu); |
57 | dbs_info->freq_lo = 0; | |
58 | } | |
6b8fcd90 | 59 | |
4471a34f VK |
60 | /* |
61 | * Not all CPUs want IO time to be accounted as busy; this depends on how | |
62 | * efficient idling at a higher frequency/voltage is. | |
63 | * Pavel Machek says this is not so for various generations of AMD and old | |
64 | * Intel systems. | |
06eb09d1 | 65 | * Mike Chan (android.com) claims this is also not true for ARM. |
4471a34f VK |
66 | * Because of this, whitelist specific known (series) of CPUs by default, and |
67 | * leave all others up to the user. | |
68 | */ | |
69 | static int should_io_be_busy(void) | |
70 | { | |
71 | #if defined(CONFIG_X86) | |
72 | /* | |
06eb09d1 | 73 | * For Intel, Core 2 (model 15) and later have an efficient idle. |
4471a34f VK |
74 | */ |
75 | if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL && | |
76 | boot_cpu_data.x86 == 6 && | |
77 | boot_cpu_data.x86_model >= 15) | |
78 | return 1; | |
79 | #endif | |
80 | return 0; | |
6b8fcd90 AV |
81 | } |
82 | ||
05ca0350 AS |
83 | /* |
84 | * Find right freq to be set now with powersave_bias on. | |
85 | * Returns the freq_hi to be used right now and will set freq_hi_jiffies, | |
86 | * freq_lo, and freq_lo_jiffies in percpu area for averaging freqs. | |
87 | */ | |
fb30809e | 88 | static unsigned int generic_powersave_bias_target(struct cpufreq_policy *policy, |
4471a34f | 89 | unsigned int freq_next, unsigned int relation) |
05ca0350 AS |
90 | { |
91 | unsigned int freq_req, freq_reduc, freq_avg; | |
92 | unsigned int freq_hi, freq_lo; | |
93 | unsigned int index = 0; | |
94 | unsigned int jiffies_total, jiffies_hi, jiffies_lo; | |
4471a34f | 95 | struct od_cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info, |
245b2e70 | 96 | policy->cpu); |
4d5dcc42 VK |
97 | struct dbs_data *dbs_data = policy->governor_data; |
98 | struct od_dbs_tuners *od_tuners = dbs_data->tuners; | |
05ca0350 AS |
99 | |
100 | if (!dbs_info->freq_table) { | |
101 | dbs_info->freq_lo = 0; | |
102 | dbs_info->freq_lo_jiffies = 0; | |
103 | return freq_next; | |
104 | } | |
105 | ||
106 | cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_next, | |
107 | relation, &index); | |
108 | freq_req = dbs_info->freq_table[index].frequency; | |
4d5dcc42 | 109 | freq_reduc = freq_req * od_tuners->powersave_bias / 1000; |
05ca0350 AS |
110 | freq_avg = freq_req - freq_reduc; |
111 | ||
112 | /* Find freq bounds for freq_avg in freq_table */ | |
113 | index = 0; | |
114 | cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_avg, | |
115 | CPUFREQ_RELATION_H, &index); | |
116 | freq_lo = dbs_info->freq_table[index].frequency; | |
117 | index = 0; | |
118 | cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_avg, | |
119 | CPUFREQ_RELATION_L, &index); | |
120 | freq_hi = dbs_info->freq_table[index].frequency; | |
121 | ||
122 | /* Find out how long we have to be in hi and lo freqs */ | |
123 | if (freq_hi == freq_lo) { | |
124 | dbs_info->freq_lo = 0; | |
125 | dbs_info->freq_lo_jiffies = 0; | |
126 | return freq_lo; | |
127 | } | |
4d5dcc42 | 128 | jiffies_total = usecs_to_jiffies(od_tuners->sampling_rate); |
05ca0350 AS |
129 | jiffies_hi = (freq_avg - freq_lo) * jiffies_total; |
130 | jiffies_hi += ((freq_hi - freq_lo) / 2); | |
131 | jiffies_hi /= (freq_hi - freq_lo); | |
132 | jiffies_lo = jiffies_total - jiffies_hi; | |
133 | dbs_info->freq_lo = freq_lo; | |
134 | dbs_info->freq_lo_jiffies = jiffies_lo; | |
135 | dbs_info->freq_hi_jiffies = jiffies_hi; | |
136 | return freq_hi; | |
137 | } | |
138 | ||
139 | static void ondemand_powersave_bias_init(void) | |
140 | { | |
141 | int i; | |
142 | for_each_online_cpu(i) { | |
5a75c828 | 143 | ondemand_powersave_bias_init_cpu(i); |
05ca0350 AS |
144 | } |
145 | } | |
146 | ||
4471a34f VK |
147 | static void dbs_freq_increase(struct cpufreq_policy *p, unsigned int freq) |
148 | { | |
4d5dcc42 VK |
149 | struct dbs_data *dbs_data = p->governor_data; |
150 | struct od_dbs_tuners *od_tuners = dbs_data->tuners; | |
151 | ||
152 | if (od_tuners->powersave_bias) | |
fb30809e JS |
153 | freq = od_ops.powersave_bias_target(p, freq, |
154 | CPUFREQ_RELATION_H); | |
4471a34f VK |
155 | else if (p->cur == p->max) |
156 | return; | |
0e625ac1 | 157 | |
4d5dcc42 | 158 | __cpufreq_driver_target(p, freq, od_tuners->powersave_bias ? |
4471a34f VK |
159 | CPUFREQ_RELATION_L : CPUFREQ_RELATION_H); |
160 | } | |
161 | ||
162 | /* | |
163 | * Every sampling_rate, we check, if current idle time is less than 20% | |
06eb09d1 SK |
164 | * (default), then we try to increase frequency. Every sampling_rate, we look |
165 | * for the lowest frequency which can sustain the load while keeping idle time | |
4471a34f VK |
166 | * over 30%. If such a frequency exist, we try to decrease to this frequency. |
167 | * | |
168 | * Any frequency increase takes it to the maximum frequency. Frequency reduction | |
169 | * happens at minimum steps of 5% (default) of current frequency | |
170 | */ | |
171 | static void od_check_cpu(int cpu, unsigned int load_freq) | |
1da177e4 | 172 | { |
4471a34f VK |
173 | struct od_cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info, cpu); |
174 | struct cpufreq_policy *policy = dbs_info->cdbs.cur_policy; | |
4d5dcc42 VK |
175 | struct dbs_data *dbs_data = policy->governor_data; |
176 | struct od_dbs_tuners *od_tuners = dbs_data->tuners; | |
4471a34f VK |
177 | |
178 | dbs_info->freq_lo = 0; | |
179 | ||
180 | /* Check for frequency increase */ | |
4d5dcc42 | 181 | if (load_freq > od_tuners->up_threshold * policy->cur) { |
4471a34f VK |
182 | /* If switching to max speed, apply sampling_down_factor */ |
183 | if (policy->cur < policy->max) | |
184 | dbs_info->rate_mult = | |
4d5dcc42 | 185 | od_tuners->sampling_down_factor; |
4471a34f VK |
186 | dbs_freq_increase(policy, policy->max); |
187 | return; | |
188 | } | |
189 | ||
190 | /* Check for frequency decrease */ | |
191 | /* if we cannot reduce the frequency anymore, break out early */ | |
192 | if (policy->cur == policy->min) | |
193 | return; | |
194 | ||
195 | /* | |
196 | * The optimal frequency is the frequency that is the lowest that can | |
197 | * support the current CPU usage without triggering the up policy. To be | |
198 | * safe, we focus 10 points under the threshold. | |
199 | */ | |
4d5dcc42 VK |
200 | if (load_freq < od_tuners->adj_up_threshold |
201 | * policy->cur) { | |
4471a34f | 202 | unsigned int freq_next; |
4d5dcc42 | 203 | freq_next = load_freq / od_tuners->adj_up_threshold; |
4471a34f VK |
204 | |
205 | /* No longer fully busy, reset rate_mult */ | |
206 | dbs_info->rate_mult = 1; | |
207 | ||
208 | if (freq_next < policy->min) | |
209 | freq_next = policy->min; | |
210 | ||
4d5dcc42 | 211 | if (!od_tuners->powersave_bias) { |
4471a34f VK |
212 | __cpufreq_driver_target(policy, freq_next, |
213 | CPUFREQ_RELATION_L); | |
fb30809e | 214 | return; |
4471a34f | 215 | } |
fb30809e JS |
216 | |
217 | freq_next = od_ops.powersave_bias_target(policy, freq_next, | |
218 | CPUFREQ_RELATION_L); | |
219 | __cpufreq_driver_target(policy, freq_next, CPUFREQ_RELATION_L); | |
4471a34f | 220 | } |
1da177e4 LT |
221 | } |
222 | ||
4447266b | 223 | static void od_dbs_timer(struct work_struct *work) |
4471a34f | 224 | { |
4447266b VK |
225 | struct od_cpu_dbs_info_s *dbs_info = |
226 | container_of(work, struct od_cpu_dbs_info_s, cdbs.work.work); | |
09dca5ae | 227 | unsigned int cpu = dbs_info->cdbs.cur_policy->cpu; |
4447266b VK |
228 | struct od_cpu_dbs_info_s *core_dbs_info = &per_cpu(od_cpu_dbs_info, |
229 | cpu); | |
4d5dcc42 VK |
230 | struct dbs_data *dbs_data = dbs_info->cdbs.cur_policy->governor_data; |
231 | struct od_dbs_tuners *od_tuners = dbs_data->tuners; | |
9d445920 | 232 | int delay = 0, sample_type = core_dbs_info->sample_type; |
031299b3 | 233 | bool modify_all = true; |
4447266b VK |
234 | |
235 | mutex_lock(&core_dbs_info->cdbs.timer_mutex); | |
031299b3 VK |
236 | if (!need_load_eval(&core_dbs_info->cdbs, od_tuners->sampling_rate)) { |
237 | modify_all = false; | |
9d445920 | 238 | goto max_delay; |
031299b3 | 239 | } |
1da177e4 | 240 | |
4471a34f | 241 | /* Common NORMAL_SAMPLE setup */ |
4447266b | 242 | core_dbs_info->sample_type = OD_NORMAL_SAMPLE; |
4471a34f | 243 | if (sample_type == OD_SUB_SAMPLE) { |
4447266b | 244 | delay = core_dbs_info->freq_lo_jiffies; |
9d445920 VK |
245 | __cpufreq_driver_target(core_dbs_info->cdbs.cur_policy, |
246 | core_dbs_info->freq_lo, CPUFREQ_RELATION_H); | |
4471a34f | 247 | } else { |
9d445920 | 248 | dbs_check_cpu(dbs_data, cpu); |
4447266b | 249 | if (core_dbs_info->freq_lo) { |
4471a34f | 250 | /* Setup timer for SUB_SAMPLE */ |
4447266b VK |
251 | core_dbs_info->sample_type = OD_SUB_SAMPLE; |
252 | delay = core_dbs_info->freq_hi_jiffies; | |
4471a34f VK |
253 | } |
254 | } | |
255 | ||
9d445920 VK |
256 | max_delay: |
257 | if (!delay) | |
258 | delay = delay_for_sampling_rate(od_tuners->sampling_rate | |
259 | * core_dbs_info->rate_mult); | |
260 | ||
031299b3 | 261 | gov_queue_work(dbs_data, dbs_info->cdbs.cur_policy, delay, modify_all); |
4447266b | 262 | mutex_unlock(&core_dbs_info->cdbs.timer_mutex); |
da53d61e FB |
263 | } |
264 | ||
4471a34f | 265 | /************************** sysfs interface ************************/ |
4d5dcc42 | 266 | static struct common_dbs_data od_dbs_cdata; |
1da177e4 | 267 | |
fd0ef7a0 MH |
268 | /** |
269 | * update_sampling_rate - update sampling rate effective immediately if needed. | |
270 | * @new_rate: new sampling rate | |
271 | * | |
06eb09d1 | 272 | * If new rate is smaller than the old, simply updating |
4471a34f VK |
273 | * dbs_tuners_int.sampling_rate might not be appropriate. For example, if the |
274 | * original sampling_rate was 1 second and the requested new sampling rate is 10 | |
275 | * ms because the user needs immediate reaction from ondemand governor, but not | |
276 | * sure if higher frequency will be required or not, then, the governor may | |
277 | * change the sampling rate too late; up to 1 second later. Thus, if we are | |
278 | * reducing the sampling rate, we need to make the new value effective | |
279 | * immediately. | |
fd0ef7a0 | 280 | */ |
4d5dcc42 VK |
281 | static void update_sampling_rate(struct dbs_data *dbs_data, |
282 | unsigned int new_rate) | |
fd0ef7a0 | 283 | { |
4d5dcc42 | 284 | struct od_dbs_tuners *od_tuners = dbs_data->tuners; |
fd0ef7a0 MH |
285 | int cpu; |
286 | ||
4d5dcc42 VK |
287 | od_tuners->sampling_rate = new_rate = max(new_rate, |
288 | dbs_data->min_sampling_rate); | |
fd0ef7a0 MH |
289 | |
290 | for_each_online_cpu(cpu) { | |
291 | struct cpufreq_policy *policy; | |
4471a34f | 292 | struct od_cpu_dbs_info_s *dbs_info; |
fd0ef7a0 MH |
293 | unsigned long next_sampling, appointed_at; |
294 | ||
295 | policy = cpufreq_cpu_get(cpu); | |
296 | if (!policy) | |
297 | continue; | |
3e33ee9e FB |
298 | if (policy->governor != &cpufreq_gov_ondemand) { |
299 | cpufreq_cpu_put(policy); | |
300 | continue; | |
301 | } | |
8ee2ec51 | 302 | dbs_info = &per_cpu(od_cpu_dbs_info, cpu); |
fd0ef7a0 MH |
303 | cpufreq_cpu_put(policy); |
304 | ||
4471a34f | 305 | mutex_lock(&dbs_info->cdbs.timer_mutex); |
fd0ef7a0 | 306 | |
4471a34f VK |
307 | if (!delayed_work_pending(&dbs_info->cdbs.work)) { |
308 | mutex_unlock(&dbs_info->cdbs.timer_mutex); | |
fd0ef7a0 MH |
309 | continue; |
310 | } | |
311 | ||
4471a34f VK |
312 | next_sampling = jiffies + usecs_to_jiffies(new_rate); |
313 | appointed_at = dbs_info->cdbs.work.timer.expires; | |
fd0ef7a0 MH |
314 | |
315 | if (time_before(next_sampling, appointed_at)) { | |
316 | ||
4471a34f VK |
317 | mutex_unlock(&dbs_info->cdbs.timer_mutex); |
318 | cancel_delayed_work_sync(&dbs_info->cdbs.work); | |
319 | mutex_lock(&dbs_info->cdbs.timer_mutex); | |
fd0ef7a0 | 320 | |
031299b3 VK |
321 | gov_queue_work(dbs_data, dbs_info->cdbs.cur_policy, |
322 | usecs_to_jiffies(new_rate), true); | |
fd0ef7a0 MH |
323 | |
324 | } | |
4471a34f | 325 | mutex_unlock(&dbs_info->cdbs.timer_mutex); |
fd0ef7a0 MH |
326 | } |
327 | } | |
328 | ||
4d5dcc42 VK |
329 | static ssize_t store_sampling_rate(struct dbs_data *dbs_data, const char *buf, |
330 | size_t count) | |
1da177e4 LT |
331 | { |
332 | unsigned int input; | |
333 | int ret; | |
ffac80e9 | 334 | ret = sscanf(buf, "%u", &input); |
5a75c828 | 335 | if (ret != 1) |
336 | return -EINVAL; | |
4d5dcc42 VK |
337 | |
338 | update_sampling_rate(dbs_data, input); | |
1da177e4 LT |
339 | return count; |
340 | } | |
341 | ||
4d5dcc42 VK |
342 | static ssize_t store_io_is_busy(struct dbs_data *dbs_data, const char *buf, |
343 | size_t count) | |
19379b11 | 344 | { |
4d5dcc42 | 345 | struct od_dbs_tuners *od_tuners = dbs_data->tuners; |
19379b11 AV |
346 | unsigned int input; |
347 | int ret; | |
9366d840 | 348 | unsigned int j; |
19379b11 AV |
349 | |
350 | ret = sscanf(buf, "%u", &input); | |
351 | if (ret != 1) | |
352 | return -EINVAL; | |
4d5dcc42 | 353 | od_tuners->io_is_busy = !!input; |
9366d840 SK |
354 | |
355 | /* we need to re-evaluate prev_cpu_idle */ | |
356 | for_each_online_cpu(j) { | |
357 | struct od_cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info, | |
358 | j); | |
359 | dbs_info->cdbs.prev_cpu_idle = get_cpu_idle_time(j, | |
360 | &dbs_info->cdbs.prev_cpu_wall, od_tuners->io_is_busy); | |
361 | } | |
19379b11 AV |
362 | return count; |
363 | } | |
364 | ||
4d5dcc42 VK |
365 | static ssize_t store_up_threshold(struct dbs_data *dbs_data, const char *buf, |
366 | size_t count) | |
1da177e4 | 367 | { |
4d5dcc42 | 368 | struct od_dbs_tuners *od_tuners = dbs_data->tuners; |
1da177e4 LT |
369 | unsigned int input; |
370 | int ret; | |
ffac80e9 | 371 | ret = sscanf(buf, "%u", &input); |
1da177e4 | 372 | |
32ee8c3e | 373 | if (ret != 1 || input > MAX_FREQUENCY_UP_THRESHOLD || |
c29f1403 | 374 | input < MIN_FREQUENCY_UP_THRESHOLD) { |
1da177e4 LT |
375 | return -EINVAL; |
376 | } | |
4bd4e428 | 377 | /* Calculate the new adj_up_threshold */ |
4d5dcc42 VK |
378 | od_tuners->adj_up_threshold += input; |
379 | od_tuners->adj_up_threshold -= od_tuners->up_threshold; | |
4bd4e428 | 380 | |
4d5dcc42 | 381 | od_tuners->up_threshold = input; |
1da177e4 LT |
382 | return count; |
383 | } | |
384 | ||
4d5dcc42 VK |
385 | static ssize_t store_sampling_down_factor(struct dbs_data *dbs_data, |
386 | const char *buf, size_t count) | |
3f78a9f7 | 387 | { |
4d5dcc42 | 388 | struct od_dbs_tuners *od_tuners = dbs_data->tuners; |
3f78a9f7 DN |
389 | unsigned int input, j; |
390 | int ret; | |
391 | ret = sscanf(buf, "%u", &input); | |
392 | ||
393 | if (ret != 1 || input > MAX_SAMPLING_DOWN_FACTOR || input < 1) | |
394 | return -EINVAL; | |
4d5dcc42 | 395 | od_tuners->sampling_down_factor = input; |
3f78a9f7 DN |
396 | |
397 | /* Reset down sampling multiplier in case it was active */ | |
398 | for_each_online_cpu(j) { | |
4471a34f VK |
399 | struct od_cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info, |
400 | j); | |
3f78a9f7 DN |
401 | dbs_info->rate_mult = 1; |
402 | } | |
3f78a9f7 DN |
403 | return count; |
404 | } | |
405 | ||
6c4640c3 VK |
406 | static ssize_t store_ignore_nice_load(struct dbs_data *dbs_data, |
407 | const char *buf, size_t count) | |
3d5ee9e5 | 408 | { |
4d5dcc42 | 409 | struct od_dbs_tuners *od_tuners = dbs_data->tuners; |
3d5ee9e5 DJ |
410 | unsigned int input; |
411 | int ret; | |
412 | ||
413 | unsigned int j; | |
32ee8c3e | 414 | |
ffac80e9 | 415 | ret = sscanf(buf, "%u", &input); |
2b03f891 | 416 | if (ret != 1) |
3d5ee9e5 DJ |
417 | return -EINVAL; |
418 | ||
2b03f891 | 419 | if (input > 1) |
3d5ee9e5 | 420 | input = 1; |
32ee8c3e | 421 | |
6c4640c3 | 422 | if (input == od_tuners->ignore_nice_load) { /* nothing to do */ |
3d5ee9e5 DJ |
423 | return count; |
424 | } | |
6c4640c3 | 425 | od_tuners->ignore_nice_load = input; |
3d5ee9e5 | 426 | |
ccb2fe20 | 427 | /* we need to re-evaluate prev_cpu_idle */ |
dac1c1a5 | 428 | for_each_online_cpu(j) { |
4471a34f | 429 | struct od_cpu_dbs_info_s *dbs_info; |
245b2e70 | 430 | dbs_info = &per_cpu(od_cpu_dbs_info, j); |
4471a34f | 431 | dbs_info->cdbs.prev_cpu_idle = get_cpu_idle_time(j, |
9366d840 | 432 | &dbs_info->cdbs.prev_cpu_wall, od_tuners->io_is_busy); |
6c4640c3 | 433 | if (od_tuners->ignore_nice_load) |
4471a34f VK |
434 | dbs_info->cdbs.prev_cpu_nice = |
435 | kcpustat_cpu(j).cpustat[CPUTIME_NICE]; | |
1ca3abdb | 436 | |
3d5ee9e5 | 437 | } |
3d5ee9e5 DJ |
438 | return count; |
439 | } | |
440 | ||
4d5dcc42 VK |
441 | static ssize_t store_powersave_bias(struct dbs_data *dbs_data, const char *buf, |
442 | size_t count) | |
05ca0350 | 443 | { |
4d5dcc42 | 444 | struct od_dbs_tuners *od_tuners = dbs_data->tuners; |
05ca0350 AS |
445 | unsigned int input; |
446 | int ret; | |
447 | ret = sscanf(buf, "%u", &input); | |
448 | ||
449 | if (ret != 1) | |
450 | return -EINVAL; | |
451 | ||
452 | if (input > 1000) | |
453 | input = 1000; | |
454 | ||
4d5dcc42 | 455 | od_tuners->powersave_bias = input; |
05ca0350 | 456 | ondemand_powersave_bias_init(); |
05ca0350 AS |
457 | return count; |
458 | } | |
459 | ||
4d5dcc42 VK |
460 | show_store_one(od, sampling_rate); |
461 | show_store_one(od, io_is_busy); | |
462 | show_store_one(od, up_threshold); | |
463 | show_store_one(od, sampling_down_factor); | |
6c4640c3 | 464 | show_store_one(od, ignore_nice_load); |
4d5dcc42 VK |
465 | show_store_one(od, powersave_bias); |
466 | declare_show_sampling_rate_min(od); | |
467 | ||
468 | gov_sys_pol_attr_rw(sampling_rate); | |
469 | gov_sys_pol_attr_rw(io_is_busy); | |
470 | gov_sys_pol_attr_rw(up_threshold); | |
471 | gov_sys_pol_attr_rw(sampling_down_factor); | |
6c4640c3 | 472 | gov_sys_pol_attr_rw(ignore_nice_load); |
4d5dcc42 VK |
473 | gov_sys_pol_attr_rw(powersave_bias); |
474 | gov_sys_pol_attr_ro(sampling_rate_min); | |
475 | ||
476 | static struct attribute *dbs_attributes_gov_sys[] = { | |
477 | &sampling_rate_min_gov_sys.attr, | |
478 | &sampling_rate_gov_sys.attr, | |
479 | &up_threshold_gov_sys.attr, | |
480 | &sampling_down_factor_gov_sys.attr, | |
6c4640c3 | 481 | &ignore_nice_load_gov_sys.attr, |
4d5dcc42 VK |
482 | &powersave_bias_gov_sys.attr, |
483 | &io_is_busy_gov_sys.attr, | |
1da177e4 LT |
484 | NULL |
485 | }; | |
486 | ||
4d5dcc42 VK |
487 | static struct attribute_group od_attr_group_gov_sys = { |
488 | .attrs = dbs_attributes_gov_sys, | |
489 | .name = "ondemand", | |
490 | }; | |
491 | ||
492 | static struct attribute *dbs_attributes_gov_pol[] = { | |
493 | &sampling_rate_min_gov_pol.attr, | |
494 | &sampling_rate_gov_pol.attr, | |
495 | &up_threshold_gov_pol.attr, | |
496 | &sampling_down_factor_gov_pol.attr, | |
6c4640c3 | 497 | &ignore_nice_load_gov_pol.attr, |
4d5dcc42 VK |
498 | &powersave_bias_gov_pol.attr, |
499 | &io_is_busy_gov_pol.attr, | |
500 | NULL | |
501 | }; | |
502 | ||
503 | static struct attribute_group od_attr_group_gov_pol = { | |
504 | .attrs = dbs_attributes_gov_pol, | |
1da177e4 LT |
505 | .name = "ondemand", |
506 | }; | |
507 | ||
508 | /************************** sysfs end ************************/ | |
509 | ||
4d5dcc42 VK |
510 | static int od_init(struct dbs_data *dbs_data) |
511 | { | |
512 | struct od_dbs_tuners *tuners; | |
513 | u64 idle_time; | |
514 | int cpu; | |
515 | ||
516 | tuners = kzalloc(sizeof(struct od_dbs_tuners), GFP_KERNEL); | |
517 | if (!tuners) { | |
518 | pr_err("%s: kzalloc failed\n", __func__); | |
519 | return -ENOMEM; | |
520 | } | |
521 | ||
522 | cpu = get_cpu(); | |
523 | idle_time = get_cpu_idle_time_us(cpu, NULL); | |
524 | put_cpu(); | |
525 | if (idle_time != -1ULL) { | |
526 | /* Idle micro accounting is supported. Use finer thresholds */ | |
527 | tuners->up_threshold = MICRO_FREQUENCY_UP_THRESHOLD; | |
528 | tuners->adj_up_threshold = MICRO_FREQUENCY_UP_THRESHOLD - | |
529 | MICRO_FREQUENCY_DOWN_DIFFERENTIAL; | |
530 | /* | |
531 | * In nohz/micro accounting case we set the minimum frequency | |
532 | * not depending on HZ, but fixed (very low). The deferred | |
533 | * timer might skip some samples if idle/sleeping as needed. | |
534 | */ | |
535 | dbs_data->min_sampling_rate = MICRO_FREQUENCY_MIN_SAMPLE_RATE; | |
536 | } else { | |
537 | tuners->up_threshold = DEF_FREQUENCY_UP_THRESHOLD; | |
538 | tuners->adj_up_threshold = DEF_FREQUENCY_UP_THRESHOLD - | |
539 | DEF_FREQUENCY_DOWN_DIFFERENTIAL; | |
540 | ||
541 | /* For correct statistics, we need 10 ticks for each measure */ | |
542 | dbs_data->min_sampling_rate = MIN_SAMPLING_RATE_RATIO * | |
543 | jiffies_to_usecs(10); | |
544 | } | |
545 | ||
546 | tuners->sampling_down_factor = DEF_SAMPLING_DOWN_FACTOR; | |
6c4640c3 | 547 | tuners->ignore_nice_load = 0; |
c2837558 | 548 | tuners->powersave_bias = default_powersave_bias; |
4d5dcc42 VK |
549 | tuners->io_is_busy = should_io_be_busy(); |
550 | ||
551 | dbs_data->tuners = tuners; | |
4d5dcc42 VK |
552 | mutex_init(&dbs_data->mutex); |
553 | return 0; | |
554 | } | |
555 | ||
556 | static void od_exit(struct dbs_data *dbs_data) | |
557 | { | |
558 | kfree(dbs_data->tuners); | |
559 | } | |
560 | ||
4471a34f | 561 | define_get_cpu_dbs_routines(od_cpu_dbs_info); |
6b8fcd90 | 562 | |
4471a34f | 563 | static struct od_ops od_ops = { |
4471a34f | 564 | .powersave_bias_init_cpu = ondemand_powersave_bias_init_cpu, |
fb30809e | 565 | .powersave_bias_target = generic_powersave_bias_target, |
4471a34f VK |
566 | .freq_increase = dbs_freq_increase, |
567 | }; | |
2f8a835c | 568 | |
4d5dcc42 | 569 | static struct common_dbs_data od_dbs_cdata = { |
4471a34f | 570 | .governor = GOV_ONDEMAND, |
4d5dcc42 VK |
571 | .attr_group_gov_sys = &od_attr_group_gov_sys, |
572 | .attr_group_gov_pol = &od_attr_group_gov_pol, | |
4471a34f VK |
573 | .get_cpu_cdbs = get_cpu_cdbs, |
574 | .get_cpu_dbs_info_s = get_cpu_dbs_info_s, | |
575 | .gov_dbs_timer = od_dbs_timer, | |
576 | .gov_check_cpu = od_check_cpu, | |
577 | .gov_ops = &od_ops, | |
4d5dcc42 VK |
578 | .init = od_init, |
579 | .exit = od_exit, | |
4471a34f | 580 | }; |
1da177e4 | 581 | |
fb30809e JS |
582 | static void od_set_powersave_bias(unsigned int powersave_bias) |
583 | { | |
584 | struct cpufreq_policy *policy; | |
585 | struct dbs_data *dbs_data; | |
586 | struct od_dbs_tuners *od_tuners; | |
587 | unsigned int cpu; | |
588 | cpumask_t done; | |
589 | ||
c2837558 | 590 | default_powersave_bias = powersave_bias; |
fb30809e JS |
591 | cpumask_clear(&done); |
592 | ||
593 | get_online_cpus(); | |
594 | for_each_online_cpu(cpu) { | |
595 | if (cpumask_test_cpu(cpu, &done)) | |
596 | continue; | |
597 | ||
598 | policy = per_cpu(od_cpu_dbs_info, cpu).cdbs.cur_policy; | |
c2837558 JS |
599 | if (!policy) |
600 | continue; | |
fb30809e JS |
601 | |
602 | cpumask_or(&done, &done, policy->cpus); | |
c2837558 JS |
603 | |
604 | if (policy->governor != &cpufreq_gov_ondemand) | |
605 | continue; | |
606 | ||
607 | dbs_data = policy->governor_data; | |
608 | od_tuners = dbs_data->tuners; | |
609 | od_tuners->powersave_bias = default_powersave_bias; | |
fb30809e JS |
610 | } |
611 | put_online_cpus(); | |
612 | } | |
613 | ||
614 | void od_register_powersave_bias_handler(unsigned int (*f) | |
615 | (struct cpufreq_policy *, unsigned int, unsigned int), | |
616 | unsigned int powersave_bias) | |
617 | { | |
618 | od_ops.powersave_bias_target = f; | |
619 | od_set_powersave_bias(powersave_bias); | |
620 | } | |
621 | EXPORT_SYMBOL_GPL(od_register_powersave_bias_handler); | |
622 | ||
623 | void od_unregister_powersave_bias_handler(void) | |
624 | { | |
625 | od_ops.powersave_bias_target = generic_powersave_bias_target; | |
626 | od_set_powersave_bias(0); | |
627 | } | |
628 | EXPORT_SYMBOL_GPL(od_unregister_powersave_bias_handler); | |
629 | ||
4471a34f VK |
630 | static int od_cpufreq_governor_dbs(struct cpufreq_policy *policy, |
631 | unsigned int event) | |
1da177e4 | 632 | { |
4d5dcc42 | 633 | return cpufreq_governor_dbs(policy, &od_dbs_cdata, event); |
1da177e4 LT |
634 | } |
635 | ||
4471a34f VK |
636 | #ifndef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND |
637 | static | |
19379b11 | 638 | #endif |
4471a34f VK |
639 | struct cpufreq_governor cpufreq_gov_ondemand = { |
640 | .name = "ondemand", | |
641 | .governor = od_cpufreq_governor_dbs, | |
642 | .max_transition_latency = TRANSITION_LATENCY_LIMIT, | |
643 | .owner = THIS_MODULE, | |
644 | }; | |
1da177e4 | 645 | |
1da177e4 LT |
646 | static int __init cpufreq_gov_dbs_init(void) |
647 | { | |
57df5573 | 648 | return cpufreq_register_governor(&cpufreq_gov_ondemand); |
1da177e4 LT |
649 | } |
650 | ||
651 | static void __exit cpufreq_gov_dbs_exit(void) | |
652 | { | |
1c256245 | 653 | cpufreq_unregister_governor(&cpufreq_gov_ondemand); |
1da177e4 LT |
654 | } |
655 | ||
ffac80e9 VP |
656 | MODULE_AUTHOR("Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>"); |
657 | MODULE_AUTHOR("Alexey Starikovskiy <alexey.y.starikovskiy@intel.com>"); | |
658 | MODULE_DESCRIPTION("'cpufreq_ondemand' - A dynamic cpufreq governor for " | |
2b03f891 | 659 | "Low Latency Frequency Transition capable processors"); |
ffac80e9 | 660 | MODULE_LICENSE("GPL"); |
1da177e4 | 661 | |
6915719b JW |
662 | #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND |
663 | fs_initcall(cpufreq_gov_dbs_init); | |
664 | #else | |
1da177e4 | 665 | module_init(cpufreq_gov_dbs_init); |
6915719b | 666 | #endif |
1da177e4 | 667 | module_exit(cpufreq_gov_dbs_exit); |