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) | |
c29f1403 | 35 | #define MIN_FREQUENCY_UP_THRESHOLD (11) |
1da177e4 LT |
36 | #define MAX_FREQUENCY_UP_THRESHOLD (100) |
37 | ||
32ee8c3e DJ |
38 | /* |
39 | * The polling frequency of this governor depends on the capability of | |
1da177e4 | 40 | * the processor. Default polling frequency is 1000 times the transition |
32ee8c3e DJ |
41 | * latency of the processor. The governor will work on any processor with |
42 | * transition latency <= 10mS, using appropriate sampling | |
1da177e4 LT |
43 | * rate. |
44 | * For CPUs with transition latency > 10mS (mostly drivers with CPUFREQ_ETERNAL) | |
45 | * this governor will not work. | |
46 | * All times here are in uS. | |
47 | */ | |
32ee8c3e | 48 | static unsigned int def_sampling_rate; |
df8b59be DJ |
49 | #define MIN_SAMPLING_RATE_RATIO (2) |
50 | /* for correct statistics, we need at least 10 ticks between each measure */ | |
e08f5f5b GS |
51 | #define MIN_STAT_SAMPLING_RATE \ |
52 | (MIN_SAMPLING_RATE_RATIO * jiffies_to_usecs(10)) | |
53 | #define MIN_SAMPLING_RATE \ | |
54 | (def_sampling_rate / MIN_SAMPLING_RATE_RATIO) | |
112124ab TR |
55 | /* Above MIN_SAMPLING_RATE will vanish with its sysfs file soon |
56 | * Define the minimal settable sampling rate to the greater of: | |
57 | * - "HW transition latency" * 100 (same as default sampling / 10) | |
58 | * - MIN_STAT_SAMPLING_RATE | |
59 | * To avoid that userspace shoots itself. | |
60 | */ | |
61 | static unsigned int minimum_sampling_rate(void) | |
62 | { | |
63 | return max(def_sampling_rate / 10, MIN_STAT_SAMPLING_RATE); | |
64 | } | |
65 | ||
66 | /* This will also vanish soon with removing sampling_rate_max */ | |
1da177e4 | 67 | #define MAX_SAMPLING_RATE (500 * def_sampling_rate) |
112124ab | 68 | #define LATENCY_MULTIPLIER (1000) |
1c256245 | 69 | #define TRANSITION_LATENCY_LIMIT (10 * 1000 * 1000) |
1da177e4 | 70 | |
c4028958 DH |
71 | static void do_dbs_timer(struct work_struct *work); |
72 | ||
73 | /* Sampling types */ | |
529af7a1 | 74 | enum {DBS_NORMAL_SAMPLE, DBS_SUB_SAMPLE}; |
1da177e4 LT |
75 | |
76 | struct cpu_dbs_info_s { | |
ccb2fe20 VP |
77 | cputime64_t prev_cpu_idle; |
78 | cputime64_t prev_cpu_wall; | |
80800913 | 79 | cputime64_t prev_cpu_nice; |
32ee8c3e | 80 | struct cpufreq_policy *cur_policy; |
2b03f891 | 81 | struct delayed_work work; |
05ca0350 AS |
82 | struct cpufreq_frequency_table *freq_table; |
83 | unsigned int freq_lo; | |
84 | unsigned int freq_lo_jiffies; | |
85 | unsigned int freq_hi_jiffies; | |
529af7a1 VP |
86 | int cpu; |
87 | unsigned int enable:1, | |
2b03f891 | 88 | sample_type:1; |
1da177e4 LT |
89 | }; |
90 | static DEFINE_PER_CPU(struct cpu_dbs_info_s, cpu_dbs_info); | |
91 | ||
92 | static unsigned int dbs_enable; /* number of CPUs using this policy */ | |
93 | ||
4ec223d0 VP |
94 | /* |
95 | * DEADLOCK ALERT! There is a ordering requirement between cpu_hotplug | |
96 | * lock and dbs_mutex. cpu_hotplug lock should always be held before | |
97 | * dbs_mutex. If any function that can potentially take cpu_hotplug lock | |
98 | * (like __cpufreq_driver_target()) is being called with dbs_mutex taken, then | |
99 | * cpu_hotplug lock should be taken before that. Note that cpu_hotplug lock | |
100 | * is recursive for the same process. -Venki | |
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 AS |
111 | unsigned int powersave_bias; |
112 | } dbs_tuners_ins = { | |
32ee8c3e | 113 | .up_threshold = DEF_FREQUENCY_UP_THRESHOLD, |
e9d95bf7 | 114 | .down_differential = DEF_FREQUENCY_DOWN_DIFFERENTIAL, |
9cbad61b | 115 | .ignore_nice = 0, |
05ca0350 | 116 | .powersave_bias = 0, |
1da177e4 LT |
117 | }; |
118 | ||
80800913 | 119 | static inline cputime64_t get_cpu_idle_time_jiffy(unsigned int cpu, |
120 | cputime64_t *wall) | |
dac1c1a5 | 121 | { |
ea487615 | 122 | cputime64_t idle_time; |
3430502d | 123 | cputime64_t cur_wall_time; |
ea487615 | 124 | cputime64_t busy_time; |
ccb2fe20 | 125 | |
3430502d | 126 | cur_wall_time = jiffies64_to_cputime64(get_jiffies_64()); |
ea487615 VP |
127 | busy_time = cputime64_add(kstat_cpu(cpu).cpustat.user, |
128 | kstat_cpu(cpu).cpustat.system); | |
ccb2fe20 | 129 | |
ea487615 VP |
130 | busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.irq); |
131 | busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.softirq); | |
132 | busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.steal); | |
1ca3abdb | 133 | busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.nice); |
ea487615 | 134 | |
3430502d | 135 | idle_time = cputime64_sub(cur_wall_time, busy_time); |
136 | if (wall) | |
137 | *wall = cur_wall_time; | |
138 | ||
ea487615 | 139 | return idle_time; |
dac1c1a5 DJ |
140 | } |
141 | ||
80800913 | 142 | static inline cputime64_t get_cpu_idle_time(unsigned int cpu, cputime64_t *wall) |
143 | { | |
144 | u64 idle_time = get_cpu_idle_time_us(cpu, wall); | |
145 | ||
146 | if (idle_time == -1ULL) | |
147 | return get_cpu_idle_time_jiffy(cpu, wall); | |
148 | ||
80800913 | 149 | return idle_time; |
150 | } | |
151 | ||
05ca0350 AS |
152 | /* |
153 | * Find right freq to be set now with powersave_bias on. | |
154 | * Returns the freq_hi to be used right now and will set freq_hi_jiffies, | |
155 | * freq_lo, and freq_lo_jiffies in percpu area for averaging freqs. | |
156 | */ | |
b5ecf60f AB |
157 | static unsigned int powersave_bias_target(struct cpufreq_policy *policy, |
158 | unsigned int freq_next, | |
159 | unsigned int relation) | |
05ca0350 AS |
160 | { |
161 | unsigned int freq_req, freq_reduc, freq_avg; | |
162 | unsigned int freq_hi, freq_lo; | |
163 | unsigned int index = 0; | |
164 | unsigned int jiffies_total, jiffies_hi, jiffies_lo; | |
165 | struct cpu_dbs_info_s *dbs_info = &per_cpu(cpu_dbs_info, policy->cpu); | |
166 | ||
167 | if (!dbs_info->freq_table) { | |
168 | dbs_info->freq_lo = 0; | |
169 | dbs_info->freq_lo_jiffies = 0; | |
170 | return freq_next; | |
171 | } | |
172 | ||
173 | cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_next, | |
174 | relation, &index); | |
175 | freq_req = dbs_info->freq_table[index].frequency; | |
176 | freq_reduc = freq_req * dbs_tuners_ins.powersave_bias / 1000; | |
177 | freq_avg = freq_req - freq_reduc; | |
178 | ||
179 | /* Find freq bounds for freq_avg in freq_table */ | |
180 | index = 0; | |
181 | cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_avg, | |
182 | CPUFREQ_RELATION_H, &index); | |
183 | freq_lo = dbs_info->freq_table[index].frequency; | |
184 | index = 0; | |
185 | cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_avg, | |
186 | CPUFREQ_RELATION_L, &index); | |
187 | freq_hi = dbs_info->freq_table[index].frequency; | |
188 | ||
189 | /* Find out how long we have to be in hi and lo freqs */ | |
190 | if (freq_hi == freq_lo) { | |
191 | dbs_info->freq_lo = 0; | |
192 | dbs_info->freq_lo_jiffies = 0; | |
193 | return freq_lo; | |
194 | } | |
195 | jiffies_total = usecs_to_jiffies(dbs_tuners_ins.sampling_rate); | |
196 | jiffies_hi = (freq_avg - freq_lo) * jiffies_total; | |
197 | jiffies_hi += ((freq_hi - freq_lo) / 2); | |
198 | jiffies_hi /= (freq_hi - freq_lo); | |
199 | jiffies_lo = jiffies_total - jiffies_hi; | |
200 | dbs_info->freq_lo = freq_lo; | |
201 | dbs_info->freq_lo_jiffies = jiffies_lo; | |
202 | dbs_info->freq_hi_jiffies = jiffies_hi; | |
203 | return freq_hi; | |
204 | } | |
205 | ||
206 | static void ondemand_powersave_bias_init(void) | |
207 | { | |
208 | int i; | |
209 | for_each_online_cpu(i) { | |
210 | struct cpu_dbs_info_s *dbs_info = &per_cpu(cpu_dbs_info, i); | |
211 | dbs_info->freq_table = cpufreq_frequency_get_table(i); | |
212 | dbs_info->freq_lo = 0; | |
213 | } | |
214 | } | |
215 | ||
1da177e4 LT |
216 | /************************** sysfs interface ************************/ |
217 | static ssize_t show_sampling_rate_max(struct cpufreq_policy *policy, char *buf) | |
218 | { | |
9411b4ef TR |
219 | static int print_once; |
220 | ||
221 | if (!print_once) { | |
222 | printk(KERN_INFO "CPUFREQ: ondemand sampling_rate_max " | |
223 | "sysfs file is deprecated - used by: %s\n", | |
224 | current->comm); | |
225 | print_once = 1; | |
226 | } | |
2b03f891 | 227 | return sprintf(buf, "%u\n", MAX_SAMPLING_RATE); |
1da177e4 LT |
228 | } |
229 | ||
230 | static ssize_t show_sampling_rate_min(struct cpufreq_policy *policy, char *buf) | |
231 | { | |
9411b4ef TR |
232 | static int print_once; |
233 | ||
234 | if (!print_once) { | |
235 | printk(KERN_INFO "CPUFREQ: ondemand sampling_rate_min " | |
236 | "sysfs file is deprecated - used by: %s\n", | |
237 | current->comm); | |
238 | print_once = 1; | |
239 | } | |
2b03f891 | 240 | return sprintf(buf, "%u\n", MIN_SAMPLING_RATE); |
1da177e4 LT |
241 | } |
242 | ||
32ee8c3e DJ |
243 | #define define_one_ro(_name) \ |
244 | static struct freq_attr _name = \ | |
1da177e4 LT |
245 | __ATTR(_name, 0444, show_##_name, NULL) |
246 | ||
247 | define_one_ro(sampling_rate_max); | |
248 | define_one_ro(sampling_rate_min); | |
249 | ||
250 | /* cpufreq_ondemand Governor Tunables */ | |
251 | #define show_one(file_name, object) \ | |
252 | static ssize_t show_##file_name \ | |
253 | (struct cpufreq_policy *unused, char *buf) \ | |
254 | { \ | |
255 | return sprintf(buf, "%u\n", dbs_tuners_ins.object); \ | |
256 | } | |
257 | show_one(sampling_rate, sampling_rate); | |
1da177e4 | 258 | show_one(up_threshold, up_threshold); |
001893cd | 259 | show_one(ignore_nice_load, ignore_nice); |
05ca0350 | 260 | show_one(powersave_bias, powersave_bias); |
1da177e4 | 261 | |
32ee8c3e | 262 | static ssize_t store_sampling_rate(struct cpufreq_policy *unused, |
1da177e4 LT |
263 | const char *buf, size_t count) |
264 | { | |
265 | unsigned int input; | |
266 | int ret; | |
ffac80e9 | 267 | ret = sscanf(buf, "%u", &input); |
1da177e4 | 268 | |
3fc54d37 | 269 | mutex_lock(&dbs_mutex); |
112124ab | 270 | if (ret != 1) { |
3fc54d37 | 271 | mutex_unlock(&dbs_mutex); |
1da177e4 LT |
272 | return -EINVAL; |
273 | } | |
112124ab | 274 | dbs_tuners_ins.sampling_rate = max(input, minimum_sampling_rate()); |
3fc54d37 | 275 | mutex_unlock(&dbs_mutex); |
1da177e4 LT |
276 | |
277 | return count; | |
278 | } | |
279 | ||
32ee8c3e | 280 | static ssize_t store_up_threshold(struct cpufreq_policy *unused, |
1da177e4 LT |
281 | const char *buf, size_t count) |
282 | { | |
283 | unsigned int input; | |
284 | int ret; | |
ffac80e9 | 285 | ret = sscanf(buf, "%u", &input); |
1da177e4 | 286 | |
3fc54d37 | 287 | mutex_lock(&dbs_mutex); |
32ee8c3e | 288 | if (ret != 1 || input > MAX_FREQUENCY_UP_THRESHOLD || |
c29f1403 | 289 | input < MIN_FREQUENCY_UP_THRESHOLD) { |
3fc54d37 | 290 | mutex_unlock(&dbs_mutex); |
1da177e4 LT |
291 | return -EINVAL; |
292 | } | |
293 | ||
294 | dbs_tuners_ins.up_threshold = input; | |
3fc54d37 | 295 | mutex_unlock(&dbs_mutex); |
1da177e4 LT |
296 | |
297 | return count; | |
298 | } | |
299 | ||
001893cd | 300 | static ssize_t store_ignore_nice_load(struct cpufreq_policy *policy, |
3d5ee9e5 DJ |
301 | const char *buf, size_t count) |
302 | { | |
303 | unsigned int input; | |
304 | int ret; | |
305 | ||
306 | unsigned int j; | |
32ee8c3e | 307 | |
ffac80e9 | 308 | ret = sscanf(buf, "%u", &input); |
2b03f891 | 309 | if (ret != 1) |
3d5ee9e5 DJ |
310 | return -EINVAL; |
311 | ||
2b03f891 | 312 | if (input > 1) |
3d5ee9e5 | 313 | input = 1; |
32ee8c3e | 314 | |
3fc54d37 | 315 | mutex_lock(&dbs_mutex); |
2b03f891 | 316 | if (input == dbs_tuners_ins.ignore_nice) { /* nothing to do */ |
3fc54d37 | 317 | mutex_unlock(&dbs_mutex); |
3d5ee9e5 DJ |
318 | return count; |
319 | } | |
320 | dbs_tuners_ins.ignore_nice = input; | |
321 | ||
ccb2fe20 | 322 | /* we need to re-evaluate prev_cpu_idle */ |
dac1c1a5 | 323 | for_each_online_cpu(j) { |
ccb2fe20 VP |
324 | struct cpu_dbs_info_s *dbs_info; |
325 | dbs_info = &per_cpu(cpu_dbs_info, j); | |
3430502d | 326 | dbs_info->prev_cpu_idle = get_cpu_idle_time(j, |
327 | &dbs_info->prev_cpu_wall); | |
1ca3abdb VP |
328 | if (dbs_tuners_ins.ignore_nice) |
329 | dbs_info->prev_cpu_nice = kstat_cpu(j).cpustat.nice; | |
330 | ||
3d5ee9e5 | 331 | } |
3fc54d37 | 332 | mutex_unlock(&dbs_mutex); |
3d5ee9e5 DJ |
333 | |
334 | return count; | |
335 | } | |
336 | ||
05ca0350 AS |
337 | static ssize_t store_powersave_bias(struct cpufreq_policy *unused, |
338 | const char *buf, size_t count) | |
339 | { | |
340 | unsigned int input; | |
341 | int ret; | |
342 | ret = sscanf(buf, "%u", &input); | |
343 | ||
344 | if (ret != 1) | |
345 | return -EINVAL; | |
346 | ||
347 | if (input > 1000) | |
348 | input = 1000; | |
349 | ||
350 | mutex_lock(&dbs_mutex); | |
351 | dbs_tuners_ins.powersave_bias = input; | |
352 | ondemand_powersave_bias_init(); | |
353 | mutex_unlock(&dbs_mutex); | |
354 | ||
355 | return count; | |
356 | } | |
357 | ||
1da177e4 LT |
358 | #define define_one_rw(_name) \ |
359 | static struct freq_attr _name = \ | |
360 | __ATTR(_name, 0644, show_##_name, store_##_name) | |
361 | ||
362 | define_one_rw(sampling_rate); | |
1da177e4 | 363 | define_one_rw(up_threshold); |
001893cd | 364 | define_one_rw(ignore_nice_load); |
05ca0350 | 365 | define_one_rw(powersave_bias); |
1da177e4 | 366 | |
2b03f891 | 367 | static struct attribute *dbs_attributes[] = { |
1da177e4 LT |
368 | &sampling_rate_max.attr, |
369 | &sampling_rate_min.attr, | |
370 | &sampling_rate.attr, | |
1da177e4 | 371 | &up_threshold.attr, |
001893cd | 372 | &ignore_nice_load.attr, |
05ca0350 | 373 | &powersave_bias.attr, |
1da177e4 LT |
374 | NULL |
375 | }; | |
376 | ||
377 | static struct attribute_group dbs_attr_group = { | |
378 | .attrs = dbs_attributes, | |
379 | .name = "ondemand", | |
380 | }; | |
381 | ||
382 | /************************** sysfs end ************************/ | |
383 | ||
2f8a835c | 384 | static void dbs_check_cpu(struct cpu_dbs_info_s *this_dbs_info) |
1da177e4 | 385 | { |
c43aa3bd | 386 | unsigned int max_load_freq; |
1da177e4 LT |
387 | |
388 | struct cpufreq_policy *policy; | |
389 | unsigned int j; | |
390 | ||
1da177e4 LT |
391 | if (!this_dbs_info->enable) |
392 | return; | |
393 | ||
05ca0350 | 394 | this_dbs_info->freq_lo = 0; |
1da177e4 | 395 | policy = this_dbs_info->cur_policy; |
ea487615 | 396 | |
32ee8c3e | 397 | /* |
c29f1403 DJ |
398 | * Every sampling_rate, we check, if current idle time is less |
399 | * than 20% (default), then we try to increase frequency | |
ccb2fe20 | 400 | * Every sampling_rate, we look for a the lowest |
c29f1403 DJ |
401 | * frequency which can sustain the load while keeping idle time over |
402 | * 30%. If such a frequency exist, we try to decrease to this frequency. | |
1da177e4 | 403 | * |
32ee8c3e DJ |
404 | * Any frequency increase takes it to the maximum frequency. |
405 | * Frequency reduction happens at minimum steps of | |
406 | * 5% (default) of current frequency | |
1da177e4 LT |
407 | */ |
408 | ||
c43aa3bd | 409 | /* Get Absolute Load - in terms of freq */ |
410 | max_load_freq = 0; | |
411 | ||
835481d9 | 412 | for_each_cpu(j, policy->cpus) { |
1da177e4 | 413 | struct cpu_dbs_info_s *j_dbs_info; |
c43aa3bd | 414 | cputime64_t cur_wall_time, cur_idle_time; |
415 | unsigned int idle_time, wall_time; | |
416 | unsigned int load, load_freq; | |
417 | int freq_avg; | |
1da177e4 | 418 | |
1da177e4 | 419 | j_dbs_info = &per_cpu(cpu_dbs_info, j); |
3430502d | 420 | |
421 | cur_idle_time = get_cpu_idle_time(j, &cur_wall_time); | |
422 | ||
c43aa3bd | 423 | wall_time = (unsigned int) cputime64_sub(cur_wall_time, |
424 | j_dbs_info->prev_cpu_wall); | |
425 | j_dbs_info->prev_cpu_wall = cur_wall_time; | |
426 | ||
c43aa3bd | 427 | idle_time = (unsigned int) cputime64_sub(cur_idle_time, |
ccb2fe20 | 428 | j_dbs_info->prev_cpu_idle); |
c43aa3bd | 429 | j_dbs_info->prev_cpu_idle = cur_idle_time; |
1da177e4 | 430 | |
1ca3abdb VP |
431 | if (dbs_tuners_ins.ignore_nice) { |
432 | cputime64_t cur_nice; | |
433 | unsigned long cur_nice_jiffies; | |
434 | ||
435 | cur_nice = cputime64_sub(kstat_cpu(j).cpustat.nice, | |
436 | j_dbs_info->prev_cpu_nice); | |
437 | /* | |
438 | * Assumption: nice time between sampling periods will | |
439 | * be less than 2^32 jiffies for 32 bit sys | |
440 | */ | |
441 | cur_nice_jiffies = (unsigned long) | |
442 | cputime64_to_jiffies64(cur_nice); | |
443 | ||
444 | j_dbs_info->prev_cpu_nice = kstat_cpu(j).cpustat.nice; | |
445 | idle_time += jiffies_to_usecs(cur_nice_jiffies); | |
446 | } | |
447 | ||
3430502d | 448 | if (unlikely(!wall_time || wall_time < idle_time)) |
c43aa3bd | 449 | continue; |
c43aa3bd | 450 | |
451 | load = 100 * (wall_time - idle_time) / wall_time; | |
452 | ||
453 | freq_avg = __cpufreq_driver_getavg(policy, j); | |
454 | if (freq_avg <= 0) | |
455 | freq_avg = policy->cur; | |
456 | ||
457 | load_freq = load * freq_avg; | |
458 | if (load_freq > max_load_freq) | |
459 | max_load_freq = load_freq; | |
1da177e4 LT |
460 | } |
461 | ||
ccb2fe20 | 462 | /* Check for frequency increase */ |
c43aa3bd | 463 | if (max_load_freq > dbs_tuners_ins.up_threshold * policy->cur) { |
c11420a6 | 464 | /* if we are already at full speed then break out early */ |
05ca0350 AS |
465 | if (!dbs_tuners_ins.powersave_bias) { |
466 | if (policy->cur == policy->max) | |
467 | return; | |
468 | ||
469 | __cpufreq_driver_target(policy, policy->max, | |
470 | CPUFREQ_RELATION_H); | |
471 | } else { | |
472 | int freq = powersave_bias_target(policy, policy->max, | |
473 | CPUFREQ_RELATION_H); | |
474 | __cpufreq_driver_target(policy, freq, | |
475 | CPUFREQ_RELATION_L); | |
476 | } | |
1da177e4 LT |
477 | return; |
478 | } | |
479 | ||
480 | /* Check for frequency decrease */ | |
c29f1403 DJ |
481 | /* if we cannot reduce the frequency anymore, break out early */ |
482 | if (policy->cur == policy->min) | |
483 | return; | |
1da177e4 | 484 | |
c29f1403 DJ |
485 | /* |
486 | * The optimal frequency is the frequency that is the lowest that | |
487 | * can support the current CPU usage without triggering the up | |
488 | * policy. To be safe, we focus 10 points under the threshold. | |
489 | */ | |
e9d95bf7 | 490 | if (max_load_freq < |
491 | (dbs_tuners_ins.up_threshold - dbs_tuners_ins.down_differential) * | |
492 | policy->cur) { | |
c43aa3bd | 493 | unsigned int freq_next; |
e9d95bf7 | 494 | freq_next = max_load_freq / |
495 | (dbs_tuners_ins.up_threshold - | |
496 | dbs_tuners_ins.down_differential); | |
dfde5d62 | 497 | |
05ca0350 AS |
498 | if (!dbs_tuners_ins.powersave_bias) { |
499 | __cpufreq_driver_target(policy, freq_next, | |
500 | CPUFREQ_RELATION_L); | |
501 | } else { | |
502 | int freq = powersave_bias_target(policy, freq_next, | |
503 | CPUFREQ_RELATION_L); | |
504 | __cpufreq_driver_target(policy, freq, | |
505 | CPUFREQ_RELATION_L); | |
506 | } | |
ccb2fe20 | 507 | } |
1da177e4 LT |
508 | } |
509 | ||
c4028958 | 510 | static void do_dbs_timer(struct work_struct *work) |
32ee8c3e | 511 | { |
529af7a1 VP |
512 | struct cpu_dbs_info_s *dbs_info = |
513 | container_of(work, struct cpu_dbs_info_s, work.work); | |
514 | unsigned int cpu = dbs_info->cpu; | |
515 | int sample_type = dbs_info->sample_type; | |
516 | ||
1ce28d6b AS |
517 | /* We want all CPUs to do sampling nearly on same jiffy */ |
518 | int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate); | |
c4028958 | 519 | |
1ce28d6b | 520 | delay -= jiffies % delay; |
2f8a835c | 521 | |
56463b78 | 522 | if (lock_policy_rwsem_write(cpu) < 0) |
2cd7cbdf | 523 | return; |
56463b78 VP |
524 | |
525 | if (!dbs_info->enable) { | |
526 | unlock_policy_rwsem_write(cpu); | |
527 | return; | |
528 | } | |
529 | ||
05ca0350 | 530 | /* Common NORMAL_SAMPLE setup */ |
c4028958 | 531 | dbs_info->sample_type = DBS_NORMAL_SAMPLE; |
05ca0350 | 532 | if (!dbs_tuners_ins.powersave_bias || |
c4028958 | 533 | sample_type == DBS_NORMAL_SAMPLE) { |
05ca0350 | 534 | dbs_check_cpu(dbs_info); |
05ca0350 AS |
535 | if (dbs_info->freq_lo) { |
536 | /* Setup timer for SUB_SAMPLE */ | |
c4028958 | 537 | dbs_info->sample_type = DBS_SUB_SAMPLE; |
05ca0350 AS |
538 | delay = dbs_info->freq_hi_jiffies; |
539 | } | |
540 | } else { | |
541 | __cpufreq_driver_target(dbs_info->cur_policy, | |
2b03f891 | 542 | dbs_info->freq_lo, CPUFREQ_RELATION_H); |
05ca0350 | 543 | } |
1ce28d6b | 544 | queue_delayed_work_on(cpu, kondemand_wq, &dbs_info->work, delay); |
56463b78 | 545 | unlock_policy_rwsem_write(cpu); |
32ee8c3e | 546 | } |
1da177e4 | 547 | |
529af7a1 | 548 | static inline void dbs_timer_init(struct cpu_dbs_info_s *dbs_info) |
1da177e4 | 549 | { |
1ce28d6b AS |
550 | /* We want all CPUs to do sampling nearly on same jiffy */ |
551 | int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate); | |
552 | delay -= jiffies % delay; | |
2f8a835c | 553 | |
c18a1483 | 554 | dbs_info->enable = 1; |
05ca0350 | 555 | ondemand_powersave_bias_init(); |
c4028958 | 556 | dbs_info->sample_type = DBS_NORMAL_SAMPLE; |
28287033 | 557 | INIT_DELAYED_WORK_DEFERRABLE(&dbs_info->work, do_dbs_timer); |
529af7a1 | 558 | queue_delayed_work_on(dbs_info->cpu, kondemand_wq, &dbs_info->work, |
2b03f891 | 559 | delay); |
1da177e4 LT |
560 | } |
561 | ||
2cd7cbdf | 562 | static inline void dbs_timer_exit(struct cpu_dbs_info_s *dbs_info) |
1da177e4 | 563 | { |
2cd7cbdf LT |
564 | dbs_info->enable = 0; |
565 | cancel_delayed_work(&dbs_info->work); | |
1da177e4 LT |
566 | } |
567 | ||
568 | static int cpufreq_governor_dbs(struct cpufreq_policy *policy, | |
569 | unsigned int event) | |
570 | { | |
571 | unsigned int cpu = policy->cpu; | |
572 | struct cpu_dbs_info_s *this_dbs_info; | |
573 | unsigned int j; | |
914f7c31 | 574 | int rc; |
1da177e4 LT |
575 | |
576 | this_dbs_info = &per_cpu(cpu_dbs_info, cpu); | |
577 | ||
578 | switch (event) { | |
579 | case CPUFREQ_GOV_START: | |
ffac80e9 | 580 | if ((!cpu_online(cpu)) || (!policy->cur)) |
1da177e4 LT |
581 | return -EINVAL; |
582 | ||
1da177e4 LT |
583 | if (this_dbs_info->enable) /* Already enabled */ |
584 | break; | |
32ee8c3e | 585 | |
3fc54d37 | 586 | mutex_lock(&dbs_mutex); |
2f8a835c | 587 | dbs_enable++; |
914f7c31 JG |
588 | |
589 | rc = sysfs_create_group(&policy->kobj, &dbs_attr_group); | |
590 | if (rc) { | |
914f7c31 JG |
591 | dbs_enable--; |
592 | mutex_unlock(&dbs_mutex); | |
593 | return rc; | |
594 | } | |
595 | ||
835481d9 | 596 | for_each_cpu(j, policy->cpus) { |
1da177e4 LT |
597 | struct cpu_dbs_info_s *j_dbs_info; |
598 | j_dbs_info = &per_cpu(cpu_dbs_info, j); | |
599 | j_dbs_info->cur_policy = policy; | |
32ee8c3e | 600 | |
3430502d | 601 | j_dbs_info->prev_cpu_idle = get_cpu_idle_time(j, |
602 | &j_dbs_info->prev_cpu_wall); | |
1ca3abdb VP |
603 | if (dbs_tuners_ins.ignore_nice) { |
604 | j_dbs_info->prev_cpu_nice = | |
605 | kstat_cpu(j).cpustat.nice; | |
606 | } | |
1da177e4 | 607 | } |
529af7a1 | 608 | this_dbs_info->cpu = cpu; |
1da177e4 LT |
609 | /* |
610 | * Start the timerschedule work, when this governor | |
611 | * is used for first time | |
612 | */ | |
613 | if (dbs_enable == 1) { | |
614 | unsigned int latency; | |
615 | /* policy latency is in nS. Convert it to uS first */ | |
df8b59be DJ |
616 | latency = policy->cpuinfo.transition_latency / 1000; |
617 | if (latency == 0) | |
618 | latency = 1; | |
1da177e4 | 619 | |
112124ab TR |
620 | def_sampling_rate = |
621 | max(latency * LATENCY_MULTIPLIER, | |
622 | MIN_STAT_SAMPLING_RATE); | |
df8b59be | 623 | |
1da177e4 | 624 | dbs_tuners_ins.sampling_rate = def_sampling_rate; |
1da177e4 | 625 | } |
529af7a1 | 626 | dbs_timer_init(this_dbs_info); |
32ee8c3e | 627 | |
3fc54d37 | 628 | mutex_unlock(&dbs_mutex); |
1da177e4 LT |
629 | break; |
630 | ||
631 | case CPUFREQ_GOV_STOP: | |
3fc54d37 | 632 | mutex_lock(&dbs_mutex); |
2cd7cbdf | 633 | dbs_timer_exit(this_dbs_info); |
1da177e4 LT |
634 | sysfs_remove_group(&policy->kobj, &dbs_attr_group); |
635 | dbs_enable--; | |
3fc54d37 | 636 | mutex_unlock(&dbs_mutex); |
1da177e4 LT |
637 | |
638 | break; | |
639 | ||
640 | case CPUFREQ_GOV_LIMITS: | |
3fc54d37 | 641 | mutex_lock(&dbs_mutex); |
1da177e4 | 642 | if (policy->max < this_dbs_info->cur_policy->cur) |
ffac80e9 | 643 | __cpufreq_driver_target(this_dbs_info->cur_policy, |
2b03f891 | 644 | policy->max, CPUFREQ_RELATION_H); |
1da177e4 | 645 | else if (policy->min > this_dbs_info->cur_policy->cur) |
ffac80e9 | 646 | __cpufreq_driver_target(this_dbs_info->cur_policy, |
2b03f891 | 647 | policy->min, CPUFREQ_RELATION_L); |
3fc54d37 | 648 | mutex_unlock(&dbs_mutex); |
1da177e4 LT |
649 | break; |
650 | } | |
651 | return 0; | |
652 | } | |
653 | ||
c4d14bc0 SW |
654 | #ifndef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND |
655 | static | |
656 | #endif | |
1c256245 TR |
657 | struct cpufreq_governor cpufreq_gov_ondemand = { |
658 | .name = "ondemand", | |
659 | .governor = cpufreq_governor_dbs, | |
660 | .max_transition_latency = TRANSITION_LATENCY_LIMIT, | |
661 | .owner = THIS_MODULE, | |
1da177e4 | 662 | }; |
1da177e4 LT |
663 | |
664 | static int __init cpufreq_gov_dbs_init(void) | |
665 | { | |
888a794c | 666 | int err; |
80800913 | 667 | cputime64_t wall; |
4f6e6b9f AR |
668 | u64 idle_time; |
669 | int cpu = get_cpu(); | |
80800913 | 670 | |
4f6e6b9f AR |
671 | idle_time = get_cpu_idle_time_us(cpu, &wall); |
672 | put_cpu(); | |
80800913 | 673 | if (idle_time != -1ULL) { |
674 | /* Idle micro accounting is supported. Use finer thresholds */ | |
675 | dbs_tuners_ins.up_threshold = MICRO_FREQUENCY_UP_THRESHOLD; | |
676 | dbs_tuners_ins.down_differential = | |
677 | MICRO_FREQUENCY_DOWN_DIFFERENTIAL; | |
678 | } | |
888a794c | 679 | |
56463b78 VP |
680 | kondemand_wq = create_workqueue("kondemand"); |
681 | if (!kondemand_wq) { | |
682 | printk(KERN_ERR "Creation of kondemand failed\n"); | |
683 | return -EFAULT; | |
684 | } | |
888a794c AM |
685 | err = cpufreq_register_governor(&cpufreq_gov_ondemand); |
686 | if (err) | |
687 | destroy_workqueue(kondemand_wq); | |
688 | ||
689 | return err; | |
1da177e4 LT |
690 | } |
691 | ||
692 | static void __exit cpufreq_gov_dbs_exit(void) | |
693 | { | |
1c256245 | 694 | cpufreq_unregister_governor(&cpufreq_gov_ondemand); |
56463b78 | 695 | destroy_workqueue(kondemand_wq); |
1da177e4 LT |
696 | } |
697 | ||
698 | ||
ffac80e9 VP |
699 | MODULE_AUTHOR("Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>"); |
700 | MODULE_AUTHOR("Alexey Starikovskiy <alexey.y.starikovskiy@intel.com>"); | |
701 | MODULE_DESCRIPTION("'cpufreq_ondemand' - A dynamic cpufreq governor for " | |
2b03f891 | 702 | "Low Latency Frequency Transition capable processors"); |
ffac80e9 | 703 | MODULE_LICENSE("GPL"); |
1da177e4 | 704 | |
6915719b JW |
705 | #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_ONDEMAND |
706 | fs_initcall(cpufreq_gov_dbs_init); | |
707 | #else | |
1da177e4 | 708 | module_init(cpufreq_gov_dbs_init); |
6915719b | 709 | #endif |
1da177e4 | 710 | module_exit(cpufreq_gov_dbs_exit); |