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