]>
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) |
3f78a9f7 DN |
33 | #define DEF_SAMPLING_DOWN_FACTOR (1) |
34 | #define MAX_SAMPLING_DOWN_FACTOR (100000) | |
80800913 | 35 | #define MICRO_FREQUENCY_DOWN_DIFFERENTIAL (3) |
36 | #define MICRO_FREQUENCY_UP_THRESHOLD (95) | |
cef9615a | 37 | #define MICRO_FREQUENCY_MIN_SAMPLE_RATE (10000) |
c29f1403 | 38 | #define MIN_FREQUENCY_UP_THRESHOLD (11) |
1da177e4 LT |
39 | #define MAX_FREQUENCY_UP_THRESHOLD (100) |
40 | ||
32ee8c3e DJ |
41 | /* |
42 | * The polling frequency of this governor depends on the capability of | |
1da177e4 | 43 | * the processor. Default polling frequency is 1000 times the transition |
32ee8c3e DJ |
44 | * latency of the processor. The governor will work on any processor with |
45 | * transition latency <= 10mS, using appropriate sampling | |
1da177e4 LT |
46 | * rate. |
47 | * For CPUs with transition latency > 10mS (mostly drivers with CPUFREQ_ETERNAL) | |
48 | * this governor will not work. | |
49 | * All times here are in uS. | |
50 | */ | |
df8b59be | 51 | #define MIN_SAMPLING_RATE_RATIO (2) |
112124ab | 52 | |
cef9615a TR |
53 | static unsigned int min_sampling_rate; |
54 | ||
112124ab | 55 | #define LATENCY_MULTIPLIER (1000) |
cef9615a | 56 | #define MIN_LATENCY_MULTIPLIER (100) |
1c256245 | 57 | #define TRANSITION_LATENCY_LIMIT (10 * 1000 * 1000) |
1da177e4 | 58 | |
c4028958 | 59 | static void do_dbs_timer(struct work_struct *work); |
0e625ac1 TR |
60 | static int cpufreq_governor_dbs(struct cpufreq_policy *policy, |
61 | unsigned int event); | |
62 | ||
63 | #ifndef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND | |
64 | static | |
65 | #endif | |
66 | struct cpufreq_governor cpufreq_gov_ondemand = { | |
67 | .name = "ondemand", | |
68 | .governor = cpufreq_governor_dbs, | |
69 | .max_transition_latency = TRANSITION_LATENCY_LIMIT, | |
70 | .owner = THIS_MODULE, | |
71 | }; | |
c4028958 DH |
72 | |
73 | /* Sampling types */ | |
529af7a1 | 74 | enum {DBS_NORMAL_SAMPLE, DBS_SUB_SAMPLE}; |
1da177e4 LT |
75 | |
76 | struct cpu_dbs_info_s { | |
ccb2fe20 | 77 | cputime64_t prev_cpu_idle; |
6b8fcd90 | 78 | cputime64_t prev_cpu_iowait; |
ccb2fe20 | 79 | cputime64_t prev_cpu_wall; |
80800913 | 80 | cputime64_t prev_cpu_nice; |
32ee8c3e | 81 | struct cpufreq_policy *cur_policy; |
2b03f891 | 82 | struct delayed_work work; |
05ca0350 AS |
83 | struct cpufreq_frequency_table *freq_table; |
84 | unsigned int freq_lo; | |
85 | unsigned int freq_lo_jiffies; | |
86 | unsigned int freq_hi_jiffies; | |
3f78a9f7 | 87 | unsigned int rate_mult; |
529af7a1 | 88 | int cpu; |
5a75c828 | 89 | unsigned int sample_type:1; |
90 | /* | |
91 | * percpu mutex that serializes governor limit change with | |
92 | * do_dbs_timer invocation. We do not want do_dbs_timer to run | |
93 | * when user is changing the governor or limits. | |
94 | */ | |
95 | struct mutex timer_mutex; | |
1da177e4 | 96 | }; |
245b2e70 | 97 | static DEFINE_PER_CPU(struct cpu_dbs_info_s, od_cpu_dbs_info); |
1da177e4 LT |
98 | |
99 | static unsigned int dbs_enable; /* number of CPUs using this policy */ | |
100 | ||
4ec223d0 | 101 | /* |
7d26e2d5 | 102 | * dbs_mutex protects data in dbs_tuners_ins from concurrent changes on |
5a75c828 | 103 | * different CPUs. It protects dbs_enable in governor start/stop. |
4ec223d0 | 104 | */ |
ffac80e9 | 105 | static DEFINE_MUTEX(dbs_mutex); |
1da177e4 | 106 | |
2f8a835c | 107 | static struct workqueue_struct *kondemand_wq; |
6810b548 | 108 | |
05ca0350 | 109 | static struct dbs_tuners { |
32ee8c3e | 110 | unsigned int sampling_rate; |
32ee8c3e | 111 | unsigned int up_threshold; |
e9d95bf7 | 112 | unsigned int down_differential; |
32ee8c3e | 113 | unsigned int ignore_nice; |
3f78a9f7 | 114 | unsigned int sampling_down_factor; |
05ca0350 | 115 | unsigned int powersave_bias; |
19379b11 | 116 | unsigned int io_is_busy; |
05ca0350 | 117 | } dbs_tuners_ins = { |
32ee8c3e | 118 | .up_threshold = DEF_FREQUENCY_UP_THRESHOLD, |
3f78a9f7 | 119 | .sampling_down_factor = DEF_SAMPLING_DOWN_FACTOR, |
e9d95bf7 | 120 | .down_differential = DEF_FREQUENCY_DOWN_DIFFERENTIAL, |
9cbad61b | 121 | .ignore_nice = 0, |
05ca0350 | 122 | .powersave_bias = 0, |
1da177e4 LT |
123 | }; |
124 | ||
80800913 | 125 | static inline cputime64_t get_cpu_idle_time_jiffy(unsigned int cpu, |
126 | cputime64_t *wall) | |
dac1c1a5 | 127 | { |
ea487615 | 128 | cputime64_t idle_time; |
3430502d | 129 | cputime64_t cur_wall_time; |
ea487615 | 130 | cputime64_t busy_time; |
ccb2fe20 | 131 | |
3430502d | 132 | cur_wall_time = jiffies64_to_cputime64(get_jiffies_64()); |
ea487615 VP |
133 | busy_time = cputime64_add(kstat_cpu(cpu).cpustat.user, |
134 | kstat_cpu(cpu).cpustat.system); | |
ccb2fe20 | 135 | |
ea487615 VP |
136 | busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.irq); |
137 | busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.softirq); | |
138 | busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.steal); | |
1ca3abdb | 139 | busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.nice); |
ea487615 | 140 | |
3430502d | 141 | idle_time = cputime64_sub(cur_wall_time, busy_time); |
142 | if (wall) | |
54c9a35d | 143 | *wall = (cputime64_t)jiffies_to_usecs(cur_wall_time); |
3430502d | 144 | |
54c9a35d | 145 | return (cputime64_t)jiffies_to_usecs(idle_time); |
dac1c1a5 DJ |
146 | } |
147 | ||
80800913 | 148 | static inline cputime64_t get_cpu_idle_time(unsigned int cpu, cputime64_t *wall) |
149 | { | |
150 | u64 idle_time = get_cpu_idle_time_us(cpu, wall); | |
151 | ||
152 | if (idle_time == -1ULL) | |
153 | return get_cpu_idle_time_jiffy(cpu, wall); | |
154 | ||
80800913 | 155 | return idle_time; |
156 | } | |
157 | ||
6b8fcd90 AV |
158 | static inline cputime64_t get_cpu_iowait_time(unsigned int cpu, cputime64_t *wall) |
159 | { | |
160 | u64 iowait_time = get_cpu_iowait_time_us(cpu, wall); | |
161 | ||
162 | if (iowait_time == -1ULL) | |
163 | return 0; | |
164 | ||
165 | return iowait_time; | |
166 | } | |
167 | ||
05ca0350 AS |
168 | /* |
169 | * Find right freq to be set now with powersave_bias on. | |
170 | * Returns the freq_hi to be used right now and will set freq_hi_jiffies, | |
171 | * freq_lo, and freq_lo_jiffies in percpu area for averaging freqs. | |
172 | */ | |
b5ecf60f AB |
173 | static unsigned int powersave_bias_target(struct cpufreq_policy *policy, |
174 | unsigned int freq_next, | |
175 | unsigned int relation) | |
05ca0350 AS |
176 | { |
177 | unsigned int freq_req, freq_reduc, freq_avg; | |
178 | unsigned int freq_hi, freq_lo; | |
179 | unsigned int index = 0; | |
180 | unsigned int jiffies_total, jiffies_hi, jiffies_lo; | |
245b2e70 TH |
181 | struct cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info, |
182 | policy->cpu); | |
05ca0350 AS |
183 | |
184 | if (!dbs_info->freq_table) { | |
185 | dbs_info->freq_lo = 0; | |
186 | dbs_info->freq_lo_jiffies = 0; | |
187 | return freq_next; | |
188 | } | |
189 | ||
190 | cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_next, | |
191 | relation, &index); | |
192 | freq_req = dbs_info->freq_table[index].frequency; | |
193 | freq_reduc = freq_req * dbs_tuners_ins.powersave_bias / 1000; | |
194 | freq_avg = freq_req - freq_reduc; | |
195 | ||
196 | /* Find freq bounds for freq_avg in freq_table */ | |
197 | index = 0; | |
198 | cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_avg, | |
199 | CPUFREQ_RELATION_H, &index); | |
200 | freq_lo = dbs_info->freq_table[index].frequency; | |
201 | index = 0; | |
202 | cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_avg, | |
203 | CPUFREQ_RELATION_L, &index); | |
204 | freq_hi = dbs_info->freq_table[index].frequency; | |
205 | ||
206 | /* Find out how long we have to be in hi and lo freqs */ | |
207 | if (freq_hi == freq_lo) { | |
208 | dbs_info->freq_lo = 0; | |
209 | dbs_info->freq_lo_jiffies = 0; | |
210 | return freq_lo; | |
211 | } | |
212 | jiffies_total = usecs_to_jiffies(dbs_tuners_ins.sampling_rate); | |
213 | jiffies_hi = (freq_avg - freq_lo) * jiffies_total; | |
214 | jiffies_hi += ((freq_hi - freq_lo) / 2); | |
215 | jiffies_hi /= (freq_hi - freq_lo); | |
216 | jiffies_lo = jiffies_total - jiffies_hi; | |
217 | dbs_info->freq_lo = freq_lo; | |
218 | dbs_info->freq_lo_jiffies = jiffies_lo; | |
219 | dbs_info->freq_hi_jiffies = jiffies_hi; | |
220 | return freq_hi; | |
221 | } | |
222 | ||
5a75c828 | 223 | static void ondemand_powersave_bias_init_cpu(int cpu) |
224 | { | |
384be2b1 | 225 | struct cpu_dbs_info_s *dbs_info = &per_cpu(od_cpu_dbs_info, cpu); |
5a75c828 | 226 | dbs_info->freq_table = cpufreq_frequency_get_table(cpu); |
227 | dbs_info->freq_lo = 0; | |
228 | } | |
229 | ||
05ca0350 AS |
230 | static void ondemand_powersave_bias_init(void) |
231 | { | |
232 | int i; | |
233 | for_each_online_cpu(i) { | |
5a75c828 | 234 | ondemand_powersave_bias_init_cpu(i); |
05ca0350 AS |
235 | } |
236 | } | |
237 | ||
1da177e4 | 238 | /************************** sysfs interface ************************/ |
0e625ac1 TR |
239 | |
240 | static ssize_t show_sampling_rate_max(struct kobject *kobj, | |
241 | struct attribute *attr, char *buf) | |
1da177e4 | 242 | { |
4f4d1ad6 TR |
243 | printk_once(KERN_INFO "CPUFREQ: ondemand sampling_rate_max " |
244 | "sysfs file is deprecated - used by: %s\n", current->comm); | |
cef9615a | 245 | return sprintf(buf, "%u\n", -1U); |
1da177e4 LT |
246 | } |
247 | ||
0e625ac1 TR |
248 | static ssize_t show_sampling_rate_min(struct kobject *kobj, |
249 | struct attribute *attr, char *buf) | |
1da177e4 | 250 | { |
cef9615a | 251 | return sprintf(buf, "%u\n", min_sampling_rate); |
1da177e4 LT |
252 | } |
253 | ||
6dad2a29 BP |
254 | define_one_global_ro(sampling_rate_max); |
255 | define_one_global_ro(sampling_rate_min); | |
1da177e4 LT |
256 | |
257 | /* cpufreq_ondemand Governor Tunables */ | |
258 | #define show_one(file_name, object) \ | |
259 | static ssize_t show_##file_name \ | |
0e625ac1 | 260 | (struct kobject *kobj, struct attribute *attr, char *buf) \ |
1da177e4 LT |
261 | { \ |
262 | return sprintf(buf, "%u\n", dbs_tuners_ins.object); \ | |
263 | } | |
264 | show_one(sampling_rate, sampling_rate); | |
19379b11 | 265 | show_one(io_is_busy, io_is_busy); |
1da177e4 | 266 | show_one(up_threshold, up_threshold); |
3f78a9f7 | 267 | show_one(sampling_down_factor, sampling_down_factor); |
001893cd | 268 | show_one(ignore_nice_load, ignore_nice); |
05ca0350 | 269 | show_one(powersave_bias, powersave_bias); |
1da177e4 | 270 | |
0e625ac1 TR |
271 | /*** delete after deprecation time ***/ |
272 | ||
273 | #define DEPRECATION_MSG(file_name) \ | |
274 | printk_once(KERN_INFO "CPUFREQ: Per core ondemand sysfs " \ | |
275 | "interface is deprecated - " #file_name "\n"); | |
276 | ||
277 | #define show_one_old(file_name) \ | |
278 | static ssize_t show_##file_name##_old \ | |
279 | (struct cpufreq_policy *unused, char *buf) \ | |
280 | { \ | |
281 | printk_once(KERN_INFO "CPUFREQ: Per core ondemand sysfs " \ | |
282 | "interface is deprecated - " #file_name "\n"); \ | |
283 | return show_##file_name(NULL, NULL, buf); \ | |
284 | } | |
285 | show_one_old(sampling_rate); | |
286 | show_one_old(up_threshold); | |
287 | show_one_old(ignore_nice_load); | |
288 | show_one_old(powersave_bias); | |
289 | show_one_old(sampling_rate_min); | |
290 | show_one_old(sampling_rate_max); | |
291 | ||
6dad2a29 BP |
292 | cpufreq_freq_attr_ro_old(sampling_rate_min); |
293 | cpufreq_freq_attr_ro_old(sampling_rate_max); | |
0e625ac1 TR |
294 | |
295 | /*** delete after deprecation time ***/ | |
296 | ||
297 | static ssize_t store_sampling_rate(struct kobject *a, struct attribute *b, | |
298 | const char *buf, size_t count) | |
1da177e4 LT |
299 | { |
300 | unsigned int input; | |
301 | int ret; | |
ffac80e9 | 302 | ret = sscanf(buf, "%u", &input); |
5a75c828 | 303 | if (ret != 1) |
304 | return -EINVAL; | |
1da177e4 | 305 | |
3fc54d37 | 306 | mutex_lock(&dbs_mutex); |
cef9615a | 307 | dbs_tuners_ins.sampling_rate = max(input, min_sampling_rate); |
3fc54d37 | 308 | mutex_unlock(&dbs_mutex); |
1da177e4 LT |
309 | |
310 | return count; | |
311 | } | |
312 | ||
19379b11 AV |
313 | static ssize_t store_io_is_busy(struct kobject *a, struct attribute *b, |
314 | const char *buf, size_t count) | |
315 | { | |
316 | unsigned int input; | |
317 | int ret; | |
318 | ||
319 | ret = sscanf(buf, "%u", &input); | |
320 | if (ret != 1) | |
321 | return -EINVAL; | |
322 | ||
323 | mutex_lock(&dbs_mutex); | |
324 | dbs_tuners_ins.io_is_busy = !!input; | |
325 | mutex_unlock(&dbs_mutex); | |
326 | ||
327 | return count; | |
328 | } | |
329 | ||
0e625ac1 TR |
330 | static ssize_t store_up_threshold(struct kobject *a, struct attribute *b, |
331 | const char *buf, size_t count) | |
1da177e4 LT |
332 | { |
333 | unsigned int input; | |
334 | int ret; | |
ffac80e9 | 335 | ret = sscanf(buf, "%u", &input); |
1da177e4 | 336 | |
32ee8c3e | 337 | if (ret != 1 || input > MAX_FREQUENCY_UP_THRESHOLD || |
c29f1403 | 338 | input < MIN_FREQUENCY_UP_THRESHOLD) { |
1da177e4 LT |
339 | return -EINVAL; |
340 | } | |
341 | ||
5a75c828 | 342 | mutex_lock(&dbs_mutex); |
1da177e4 | 343 | dbs_tuners_ins.up_threshold = input; |
3fc54d37 | 344 | mutex_unlock(&dbs_mutex); |
1da177e4 LT |
345 | |
346 | return count; | |
347 | } | |
348 | ||
3f78a9f7 DN |
349 | static ssize_t store_sampling_down_factor(struct kobject *a, |
350 | struct attribute *b, const char *buf, size_t count) | |
351 | { | |
352 | unsigned int input, j; | |
353 | int ret; | |
354 | ret = sscanf(buf, "%u", &input); | |
355 | ||
356 | if (ret != 1 || input > MAX_SAMPLING_DOWN_FACTOR || input < 1) | |
357 | return -EINVAL; | |
358 | mutex_lock(&dbs_mutex); | |
359 | dbs_tuners_ins.sampling_down_factor = input; | |
360 | ||
361 | /* Reset down sampling multiplier in case it was active */ | |
362 | for_each_online_cpu(j) { | |
363 | struct cpu_dbs_info_s *dbs_info; | |
364 | dbs_info = &per_cpu(od_cpu_dbs_info, j); | |
365 | dbs_info->rate_mult = 1; | |
366 | } | |
367 | mutex_unlock(&dbs_mutex); | |
368 | ||
369 | return count; | |
370 | } | |
371 | ||
0e625ac1 TR |
372 | static ssize_t store_ignore_nice_load(struct kobject *a, struct attribute *b, |
373 | const char *buf, size_t count) | |
3d5ee9e5 DJ |
374 | { |
375 | unsigned int input; | |
376 | int ret; | |
377 | ||
378 | unsigned int j; | |
32ee8c3e | 379 | |
ffac80e9 | 380 | ret = sscanf(buf, "%u", &input); |
2b03f891 | 381 | if (ret != 1) |
3d5ee9e5 DJ |
382 | return -EINVAL; |
383 | ||
2b03f891 | 384 | if (input > 1) |
3d5ee9e5 | 385 | input = 1; |
32ee8c3e | 386 | |
3fc54d37 | 387 | mutex_lock(&dbs_mutex); |
2b03f891 | 388 | if (input == dbs_tuners_ins.ignore_nice) { /* nothing to do */ |
3fc54d37 | 389 | mutex_unlock(&dbs_mutex); |
3d5ee9e5 DJ |
390 | return count; |
391 | } | |
392 | dbs_tuners_ins.ignore_nice = input; | |
393 | ||
ccb2fe20 | 394 | /* we need to re-evaluate prev_cpu_idle */ |
dac1c1a5 | 395 | for_each_online_cpu(j) { |
ccb2fe20 | 396 | struct cpu_dbs_info_s *dbs_info; |
245b2e70 | 397 | dbs_info = &per_cpu(od_cpu_dbs_info, j); |
3430502d | 398 | dbs_info->prev_cpu_idle = get_cpu_idle_time(j, |
399 | &dbs_info->prev_cpu_wall); | |
1ca3abdb VP |
400 | if (dbs_tuners_ins.ignore_nice) |
401 | dbs_info->prev_cpu_nice = kstat_cpu(j).cpustat.nice; | |
402 | ||
3d5ee9e5 | 403 | } |
3fc54d37 | 404 | mutex_unlock(&dbs_mutex); |
3d5ee9e5 DJ |
405 | |
406 | return count; | |
407 | } | |
408 | ||
0e625ac1 TR |
409 | static ssize_t store_powersave_bias(struct kobject *a, struct attribute *b, |
410 | const char *buf, size_t count) | |
05ca0350 AS |
411 | { |
412 | unsigned int input; | |
413 | int ret; | |
414 | ret = sscanf(buf, "%u", &input); | |
415 | ||
416 | if (ret != 1) | |
417 | return -EINVAL; | |
418 | ||
419 | if (input > 1000) | |
420 | input = 1000; | |
421 | ||
422 | mutex_lock(&dbs_mutex); | |
423 | dbs_tuners_ins.powersave_bias = input; | |
424 | ondemand_powersave_bias_init(); | |
425 | mutex_unlock(&dbs_mutex); | |
426 | ||
427 | return count; | |
428 | } | |
429 | ||
6dad2a29 | 430 | define_one_global_rw(sampling_rate); |
07d77759 | 431 | define_one_global_rw(io_is_busy); |
6dad2a29 | 432 | define_one_global_rw(up_threshold); |
3f78a9f7 | 433 | define_one_global_rw(sampling_down_factor); |
6dad2a29 BP |
434 | define_one_global_rw(ignore_nice_load); |
435 | define_one_global_rw(powersave_bias); | |
1da177e4 | 436 | |
2b03f891 | 437 | static struct attribute *dbs_attributes[] = { |
1da177e4 LT |
438 | &sampling_rate_max.attr, |
439 | &sampling_rate_min.attr, | |
440 | &sampling_rate.attr, | |
1da177e4 | 441 | &up_threshold.attr, |
3f78a9f7 | 442 | &sampling_down_factor.attr, |
001893cd | 443 | &ignore_nice_load.attr, |
05ca0350 | 444 | &powersave_bias.attr, |
19379b11 | 445 | &io_is_busy.attr, |
1da177e4 LT |
446 | NULL |
447 | }; | |
448 | ||
449 | static struct attribute_group dbs_attr_group = { | |
450 | .attrs = dbs_attributes, | |
451 | .name = "ondemand", | |
452 | }; | |
453 | ||
0e625ac1 TR |
454 | /*** delete after deprecation time ***/ |
455 | ||
456 | #define write_one_old(file_name) \ | |
457 | static ssize_t store_##file_name##_old \ | |
458 | (struct cpufreq_policy *unused, const char *buf, size_t count) \ | |
459 | { \ | |
460 | printk_once(KERN_INFO "CPUFREQ: Per core ondemand sysfs " \ | |
461 | "interface is deprecated - " #file_name "\n"); \ | |
462 | return store_##file_name(NULL, NULL, buf, count); \ | |
463 | } | |
464 | write_one_old(sampling_rate); | |
465 | write_one_old(up_threshold); | |
466 | write_one_old(ignore_nice_load); | |
467 | write_one_old(powersave_bias); | |
468 | ||
6dad2a29 BP |
469 | cpufreq_freq_attr_rw_old(sampling_rate); |
470 | cpufreq_freq_attr_rw_old(up_threshold); | |
471 | cpufreq_freq_attr_rw_old(ignore_nice_load); | |
472 | cpufreq_freq_attr_rw_old(powersave_bias); | |
0e625ac1 TR |
473 | |
474 | static struct attribute *dbs_attributes_old[] = { | |
475 | &sampling_rate_max_old.attr, | |
476 | &sampling_rate_min_old.attr, | |
477 | &sampling_rate_old.attr, | |
478 | &up_threshold_old.attr, | |
479 | &ignore_nice_load_old.attr, | |
480 | &powersave_bias_old.attr, | |
481 | NULL | |
482 | }; | |
483 | ||
484 | static struct attribute_group dbs_attr_group_old = { | |
485 | .attrs = dbs_attributes_old, | |
486 | .name = "ondemand", | |
487 | }; | |
488 | ||
489 | /*** delete after deprecation time ***/ | |
490 | ||
1da177e4 LT |
491 | /************************** sysfs end ************************/ |
492 | ||
00e299ff MC |
493 | static void dbs_freq_increase(struct cpufreq_policy *p, unsigned int freq) |
494 | { | |
495 | if (dbs_tuners_ins.powersave_bias) | |
496 | freq = powersave_bias_target(p, freq, CPUFREQ_RELATION_H); | |
497 | else if (p->cur == p->max) | |
498 | return; | |
499 | ||
500 | __cpufreq_driver_target(p, freq, dbs_tuners_ins.powersave_bias ? | |
501 | CPUFREQ_RELATION_L : CPUFREQ_RELATION_H); | |
502 | } | |
503 | ||
2f8a835c | 504 | static void dbs_check_cpu(struct cpu_dbs_info_s *this_dbs_info) |
1da177e4 | 505 | { |
c43aa3bd | 506 | unsigned int max_load_freq; |
1da177e4 LT |
507 | |
508 | struct cpufreq_policy *policy; | |
509 | unsigned int j; | |
510 | ||
05ca0350 | 511 | this_dbs_info->freq_lo = 0; |
1da177e4 | 512 | policy = this_dbs_info->cur_policy; |
ea487615 | 513 | |
32ee8c3e | 514 | /* |
c29f1403 DJ |
515 | * Every sampling_rate, we check, if current idle time is less |
516 | * than 20% (default), then we try to increase frequency | |
ccb2fe20 | 517 | * Every sampling_rate, we look for a the lowest |
c29f1403 DJ |
518 | * frequency which can sustain the load while keeping idle time over |
519 | * 30%. If such a frequency exist, we try to decrease to this frequency. | |
1da177e4 | 520 | * |
32ee8c3e DJ |
521 | * Any frequency increase takes it to the maximum frequency. |
522 | * Frequency reduction happens at minimum steps of | |
523 | * 5% (default) of current frequency | |
1da177e4 LT |
524 | */ |
525 | ||
c43aa3bd | 526 | /* Get Absolute Load - in terms of freq */ |
527 | max_load_freq = 0; | |
528 | ||
835481d9 | 529 | for_each_cpu(j, policy->cpus) { |
1da177e4 | 530 | struct cpu_dbs_info_s *j_dbs_info; |
6b8fcd90 AV |
531 | cputime64_t cur_wall_time, cur_idle_time, cur_iowait_time; |
532 | unsigned int idle_time, wall_time, iowait_time; | |
c43aa3bd | 533 | unsigned int load, load_freq; |
534 | int freq_avg; | |
1da177e4 | 535 | |
245b2e70 | 536 | j_dbs_info = &per_cpu(od_cpu_dbs_info, j); |
3430502d | 537 | |
538 | cur_idle_time = get_cpu_idle_time(j, &cur_wall_time); | |
6b8fcd90 | 539 | cur_iowait_time = get_cpu_iowait_time(j, &cur_wall_time); |
3430502d | 540 | |
c43aa3bd | 541 | wall_time = (unsigned int) cputime64_sub(cur_wall_time, |
542 | j_dbs_info->prev_cpu_wall); | |
543 | j_dbs_info->prev_cpu_wall = cur_wall_time; | |
544 | ||
c43aa3bd | 545 | idle_time = (unsigned int) cputime64_sub(cur_idle_time, |
ccb2fe20 | 546 | j_dbs_info->prev_cpu_idle); |
c43aa3bd | 547 | j_dbs_info->prev_cpu_idle = cur_idle_time; |
1da177e4 | 548 | |
6b8fcd90 AV |
549 | iowait_time = (unsigned int) cputime64_sub(cur_iowait_time, |
550 | j_dbs_info->prev_cpu_iowait); | |
551 | j_dbs_info->prev_cpu_iowait = cur_iowait_time; | |
552 | ||
1ca3abdb VP |
553 | if (dbs_tuners_ins.ignore_nice) { |
554 | cputime64_t cur_nice; | |
555 | unsigned long cur_nice_jiffies; | |
556 | ||
557 | cur_nice = cputime64_sub(kstat_cpu(j).cpustat.nice, | |
558 | j_dbs_info->prev_cpu_nice); | |
559 | /* | |
560 | * Assumption: nice time between sampling periods will | |
561 | * be less than 2^32 jiffies for 32 bit sys | |
562 | */ | |
563 | cur_nice_jiffies = (unsigned long) | |
564 | cputime64_to_jiffies64(cur_nice); | |
565 | ||
566 | j_dbs_info->prev_cpu_nice = kstat_cpu(j).cpustat.nice; | |
567 | idle_time += jiffies_to_usecs(cur_nice_jiffies); | |
568 | } | |
569 | ||
6b8fcd90 AV |
570 | /* |
571 | * For the purpose of ondemand, waiting for disk IO is an | |
572 | * indication that you're performance critical, and not that | |
573 | * the system is actually idle. So subtract the iowait time | |
574 | * from the cpu idle time. | |
575 | */ | |
576 | ||
19379b11 | 577 | if (dbs_tuners_ins.io_is_busy && idle_time >= iowait_time) |
6b8fcd90 AV |
578 | idle_time -= iowait_time; |
579 | ||
3430502d | 580 | if (unlikely(!wall_time || wall_time < idle_time)) |
c43aa3bd | 581 | continue; |
c43aa3bd | 582 | |
583 | load = 100 * (wall_time - idle_time) / wall_time; | |
584 | ||
585 | freq_avg = __cpufreq_driver_getavg(policy, j); | |
586 | if (freq_avg <= 0) | |
587 | freq_avg = policy->cur; | |
588 | ||
589 | load_freq = load * freq_avg; | |
590 | if (load_freq > max_load_freq) | |
591 | max_load_freq = load_freq; | |
1da177e4 LT |
592 | } |
593 | ||
ccb2fe20 | 594 | /* Check for frequency increase */ |
c43aa3bd | 595 | if (max_load_freq > dbs_tuners_ins.up_threshold * policy->cur) { |
3f78a9f7 DN |
596 | /* If switching to max speed, apply sampling_down_factor */ |
597 | if (policy->cur < policy->max) | |
598 | this_dbs_info->rate_mult = | |
599 | dbs_tuners_ins.sampling_down_factor; | |
00e299ff | 600 | dbs_freq_increase(policy, policy->max); |
1da177e4 LT |
601 | return; |
602 | } | |
603 | ||
604 | /* Check for frequency decrease */ | |
c29f1403 DJ |
605 | /* if we cannot reduce the frequency anymore, break out early */ |
606 | if (policy->cur == policy->min) | |
607 | return; | |
1da177e4 | 608 | |
c29f1403 DJ |
609 | /* |
610 | * The optimal frequency is the frequency that is the lowest that | |
611 | * can support the current CPU usage without triggering the up | |
612 | * policy. To be safe, we focus 10 points under the threshold. | |
613 | */ | |
e9d95bf7 | 614 | if (max_load_freq < |
615 | (dbs_tuners_ins.up_threshold - dbs_tuners_ins.down_differential) * | |
616 | policy->cur) { | |
c43aa3bd | 617 | unsigned int freq_next; |
e9d95bf7 | 618 | freq_next = max_load_freq / |
619 | (dbs_tuners_ins.up_threshold - | |
620 | dbs_tuners_ins.down_differential); | |
dfde5d62 | 621 | |
3f78a9f7 DN |
622 | /* No longer fully busy, reset rate_mult */ |
623 | this_dbs_info->rate_mult = 1; | |
624 | ||
1dbf5888 NC |
625 | if (freq_next < policy->min) |
626 | freq_next = policy->min; | |
627 | ||
05ca0350 AS |
628 | if (!dbs_tuners_ins.powersave_bias) { |
629 | __cpufreq_driver_target(policy, freq_next, | |
630 | CPUFREQ_RELATION_L); | |
631 | } else { | |
632 | int freq = powersave_bias_target(policy, freq_next, | |
633 | CPUFREQ_RELATION_L); | |
634 | __cpufreq_driver_target(policy, freq, | |
635 | CPUFREQ_RELATION_L); | |
636 | } | |
ccb2fe20 | 637 | } |
1da177e4 LT |
638 | } |
639 | ||
c4028958 | 640 | static void do_dbs_timer(struct work_struct *work) |
32ee8c3e | 641 | { |
529af7a1 VP |
642 | struct cpu_dbs_info_s *dbs_info = |
643 | container_of(work, struct cpu_dbs_info_s, work.work); | |
644 | unsigned int cpu = dbs_info->cpu; | |
645 | int sample_type = dbs_info->sample_type; | |
646 | ||
1ce28d6b | 647 | /* We want all CPUs to do sampling nearly on same jiffy */ |
3f78a9f7 DN |
648 | int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate |
649 | * dbs_info->rate_mult); | |
c4028958 | 650 | |
a665df9d JF |
651 | if (num_online_cpus() > 1) |
652 | delay -= jiffies % delay; | |
653 | ||
5a75c828 | 654 | mutex_lock(&dbs_info->timer_mutex); |
56463b78 | 655 | |
05ca0350 | 656 | /* Common NORMAL_SAMPLE setup */ |
c4028958 | 657 | dbs_info->sample_type = DBS_NORMAL_SAMPLE; |
05ca0350 | 658 | if (!dbs_tuners_ins.powersave_bias || |
c4028958 | 659 | sample_type == DBS_NORMAL_SAMPLE) { |
05ca0350 | 660 | dbs_check_cpu(dbs_info); |
05ca0350 AS |
661 | if (dbs_info->freq_lo) { |
662 | /* Setup timer for SUB_SAMPLE */ | |
c4028958 | 663 | dbs_info->sample_type = DBS_SUB_SAMPLE; |
05ca0350 AS |
664 | delay = dbs_info->freq_hi_jiffies; |
665 | } | |
666 | } else { | |
667 | __cpufreq_driver_target(dbs_info->cur_policy, | |
2b03f891 | 668 | dbs_info->freq_lo, CPUFREQ_RELATION_H); |
05ca0350 | 669 | } |
1ce28d6b | 670 | queue_delayed_work_on(cpu, kondemand_wq, &dbs_info->work, delay); |
5a75c828 | 671 | mutex_unlock(&dbs_info->timer_mutex); |
32ee8c3e | 672 | } |
1da177e4 | 673 | |
529af7a1 | 674 | static inline void dbs_timer_init(struct cpu_dbs_info_s *dbs_info) |
1da177e4 | 675 | { |
1ce28d6b AS |
676 | /* We want all CPUs to do sampling nearly on same jiffy */ |
677 | int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate); | |
a665df9d JF |
678 | |
679 | if (num_online_cpus() > 1) | |
680 | delay -= jiffies % delay; | |
2f8a835c | 681 | |
c4028958 | 682 | dbs_info->sample_type = DBS_NORMAL_SAMPLE; |
28287033 | 683 | INIT_DELAYED_WORK_DEFERRABLE(&dbs_info->work, do_dbs_timer); |
529af7a1 | 684 | queue_delayed_work_on(dbs_info->cpu, kondemand_wq, &dbs_info->work, |
2b03f891 | 685 | delay); |
1da177e4 LT |
686 | } |
687 | ||
2cd7cbdf | 688 | static inline void dbs_timer_exit(struct cpu_dbs_info_s *dbs_info) |
1da177e4 | 689 | { |
b14893a6 | 690 | cancel_delayed_work_sync(&dbs_info->work); |
1da177e4 LT |
691 | } |
692 | ||
19379b11 AV |
693 | /* |
694 | * Not all CPUs want IO time to be accounted as busy; this dependson how | |
695 | * efficient idling at a higher frequency/voltage is. | |
696 | * Pavel Machek says this is not so for various generations of AMD and old | |
697 | * Intel systems. | |
698 | * Mike Chan (androidlcom) calis this is also not true for ARM. | |
699 | * Because of this, whitelist specific known (series) of CPUs by default, and | |
700 | * leave all others up to the user. | |
701 | */ | |
702 | static int should_io_be_busy(void) | |
703 | { | |
704 | #if defined(CONFIG_X86) | |
705 | /* | |
706 | * For Intel, Core 2 (model 15) andl later have an efficient idle. | |
707 | */ | |
708 | if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL && | |
709 | boot_cpu_data.x86 == 6 && | |
710 | boot_cpu_data.x86_model >= 15) | |
711 | return 1; | |
712 | #endif | |
713 | return 0; | |
714 | } | |
715 | ||
1da177e4 LT |
716 | static int cpufreq_governor_dbs(struct cpufreq_policy *policy, |
717 | unsigned int event) | |
718 | { | |
719 | unsigned int cpu = policy->cpu; | |
720 | struct cpu_dbs_info_s *this_dbs_info; | |
721 | unsigned int j; | |
914f7c31 | 722 | int rc; |
1da177e4 | 723 | |
245b2e70 | 724 | this_dbs_info = &per_cpu(od_cpu_dbs_info, cpu); |
1da177e4 LT |
725 | |
726 | switch (event) { | |
727 | case CPUFREQ_GOV_START: | |
ffac80e9 | 728 | if ((!cpu_online(cpu)) || (!policy->cur)) |
1da177e4 LT |
729 | return -EINVAL; |
730 | ||
3fc54d37 | 731 | mutex_lock(&dbs_mutex); |
914f7c31 | 732 | |
0e625ac1 | 733 | rc = sysfs_create_group(&policy->kobj, &dbs_attr_group_old); |
914f7c31 | 734 | if (rc) { |
914f7c31 JG |
735 | mutex_unlock(&dbs_mutex); |
736 | return rc; | |
737 | } | |
738 | ||
5a75c828 | 739 | dbs_enable++; |
835481d9 | 740 | for_each_cpu(j, policy->cpus) { |
1da177e4 | 741 | struct cpu_dbs_info_s *j_dbs_info; |
245b2e70 | 742 | j_dbs_info = &per_cpu(od_cpu_dbs_info, j); |
1da177e4 | 743 | j_dbs_info->cur_policy = policy; |
32ee8c3e | 744 | |
3430502d | 745 | j_dbs_info->prev_cpu_idle = get_cpu_idle_time(j, |
746 | &j_dbs_info->prev_cpu_wall); | |
1ca3abdb VP |
747 | if (dbs_tuners_ins.ignore_nice) { |
748 | j_dbs_info->prev_cpu_nice = | |
749 | kstat_cpu(j).cpustat.nice; | |
750 | } | |
1da177e4 | 751 | } |
529af7a1 | 752 | this_dbs_info->cpu = cpu; |
3f78a9f7 | 753 | this_dbs_info->rate_mult = 1; |
5a75c828 | 754 | ondemand_powersave_bias_init_cpu(cpu); |
1da177e4 LT |
755 | /* |
756 | * Start the timerschedule work, when this governor | |
757 | * is used for first time | |
758 | */ | |
759 | if (dbs_enable == 1) { | |
760 | unsigned int latency; | |
0e625ac1 TR |
761 | |
762 | rc = sysfs_create_group(cpufreq_global_kobject, | |
763 | &dbs_attr_group); | |
764 | if (rc) { | |
765 | mutex_unlock(&dbs_mutex); | |
766 | return rc; | |
767 | } | |
768 | ||
1da177e4 | 769 | /* policy latency is in nS. Convert it to uS first */ |
df8b59be DJ |
770 | latency = policy->cpuinfo.transition_latency / 1000; |
771 | if (latency == 0) | |
772 | latency = 1; | |
cef9615a TR |
773 | /* Bring kernel and HW constraints together */ |
774 | min_sampling_rate = max(min_sampling_rate, | |
775 | MIN_LATENCY_MULTIPLIER * latency); | |
776 | dbs_tuners_ins.sampling_rate = | |
777 | max(min_sampling_rate, | |
778 | latency * LATENCY_MULTIPLIER); | |
19379b11 | 779 | dbs_tuners_ins.io_is_busy = should_io_be_busy(); |
1da177e4 | 780 | } |
3fc54d37 | 781 | mutex_unlock(&dbs_mutex); |
7d26e2d5 | 782 | |
0e625ac1 | 783 | mutex_init(&this_dbs_info->timer_mutex); |
7d26e2d5 | 784 | dbs_timer_init(this_dbs_info); |
1da177e4 LT |
785 | break; |
786 | ||
787 | case CPUFREQ_GOV_STOP: | |
2cd7cbdf | 788 | dbs_timer_exit(this_dbs_info); |
7d26e2d5 | 789 | |
790 | mutex_lock(&dbs_mutex); | |
0e625ac1 | 791 | sysfs_remove_group(&policy->kobj, &dbs_attr_group_old); |
5a75c828 | 792 | mutex_destroy(&this_dbs_info->timer_mutex); |
1da177e4 | 793 | dbs_enable--; |
3fc54d37 | 794 | mutex_unlock(&dbs_mutex); |
0e625ac1 TR |
795 | if (!dbs_enable) |
796 | sysfs_remove_group(cpufreq_global_kobject, | |
797 | &dbs_attr_group); | |
1da177e4 LT |
798 | |
799 | break; | |
800 | ||
801 | case CPUFREQ_GOV_LIMITS: | |
5a75c828 | 802 | mutex_lock(&this_dbs_info->timer_mutex); |
1da177e4 | 803 | if (policy->max < this_dbs_info->cur_policy->cur) |
ffac80e9 | 804 | __cpufreq_driver_target(this_dbs_info->cur_policy, |
2b03f891 | 805 | policy->max, CPUFREQ_RELATION_H); |
1da177e4 | 806 | else if (policy->min > this_dbs_info->cur_policy->cur) |
ffac80e9 | 807 | __cpufreq_driver_target(this_dbs_info->cur_policy, |
2b03f891 | 808 | policy->min, CPUFREQ_RELATION_L); |
5a75c828 | 809 | mutex_unlock(&this_dbs_info->timer_mutex); |
1da177e4 LT |
810 | break; |
811 | } | |
812 | return 0; | |
813 | } | |
814 | ||
1da177e4 LT |
815 | static int __init cpufreq_gov_dbs_init(void) |
816 | { | |
888a794c | 817 | int err; |
80800913 | 818 | cputime64_t wall; |
4f6e6b9f AR |
819 | u64 idle_time; |
820 | int cpu = get_cpu(); | |
80800913 | 821 | |
4f6e6b9f AR |
822 | idle_time = get_cpu_idle_time_us(cpu, &wall); |
823 | put_cpu(); | |
80800913 | 824 | if (idle_time != -1ULL) { |
825 | /* Idle micro accounting is supported. Use finer thresholds */ | |
826 | dbs_tuners_ins.up_threshold = MICRO_FREQUENCY_UP_THRESHOLD; | |
827 | dbs_tuners_ins.down_differential = | |
828 | MICRO_FREQUENCY_DOWN_DIFFERENTIAL; | |
cef9615a TR |
829 | /* |
830 | * In no_hz/micro accounting case we set the minimum frequency | |
831 | * not depending on HZ, but fixed (very low). The deferred | |
832 | * timer might skip some samples if idle/sleeping as needed. | |
833 | */ | |
834 | min_sampling_rate = MICRO_FREQUENCY_MIN_SAMPLE_RATE; | |
835 | } else { | |
836 | /* For correct statistics, we need 10 ticks for each measure */ | |
837 | min_sampling_rate = | |
838 | MIN_SAMPLING_RATE_RATIO * jiffies_to_usecs(10); | |
80800913 | 839 | } |
888a794c | 840 | |
56463b78 VP |
841 | kondemand_wq = create_workqueue("kondemand"); |
842 | if (!kondemand_wq) { | |
843 | printk(KERN_ERR "Creation of kondemand failed\n"); | |
844 | return -EFAULT; | |
845 | } | |
888a794c AM |
846 | err = cpufreq_register_governor(&cpufreq_gov_ondemand); |
847 | if (err) | |
848 | destroy_workqueue(kondemand_wq); | |
849 | ||
850 | return err; | |
1da177e4 LT |
851 | } |
852 | ||
853 | static void __exit cpufreq_gov_dbs_exit(void) | |
854 | { | |
1c256245 | 855 | cpufreq_unregister_governor(&cpufreq_gov_ondemand); |
56463b78 | 856 | destroy_workqueue(kondemand_wq); |
1da177e4 LT |
857 | } |
858 | ||
859 | ||
ffac80e9 VP |
860 | MODULE_AUTHOR("Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>"); |
861 | MODULE_AUTHOR("Alexey Starikovskiy <alexey.y.starikovskiy@intel.com>"); | |
862 | MODULE_DESCRIPTION("'cpufreq_ondemand' - A dynamic cpufreq governor for " | |
2b03f891 | 863 | "Low Latency Frequency Transition capable processors"); |
ffac80e9 | 864 | MODULE_LICENSE("GPL"); |
1da177e4 | 865 | |
6915719b JW |
866 | #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND |
867 | fs_initcall(cpufreq_gov_dbs_init); | |
868 | #else | |
1da177e4 | 869 | module_init(cpufreq_gov_dbs_init); |
6915719b | 870 | #endif |
1da177e4 | 871 | module_exit(cpufreq_gov_dbs_exit); |