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