Commit | Line | Data |
---|---|---|
b9170836 DJ |
1 | /* |
2 | * drivers/cpufreq/cpufreq_conservative.c | |
3 | * | |
4 | * Copyright (C) 2001 Russell King | |
5 | * (C) 2003 Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>. | |
6 | * Jun Nakajima <jun.nakajima@intel.com> | |
11a80a9c | 7 | * (C) 2009 Alexander Clouter <alex@digriz.org.uk> |
b9170836 DJ |
8 | * |
9 | * This program is free software; you can redistribute it and/or modify | |
10 | * it under the terms of the GNU General Public License version 2 as | |
11 | * published by the Free Software Foundation. | |
12 | */ | |
13 | ||
14 | #include <linux/kernel.h> | |
15 | #include <linux/module.h> | |
b9170836 | 16 | #include <linux/init.h> |
b9170836 | 17 | #include <linux/cpufreq.h> |
138a0128 | 18 | #include <linux/cpu.h> |
b9170836 DJ |
19 | #include <linux/jiffies.h> |
20 | #include <linux/kernel_stat.h> | |
3fc54d37 | 21 | #include <linux/mutex.h> |
8e677ce8 AC |
22 | #include <linux/hrtimer.h> |
23 | #include <linux/tick.h> | |
24 | #include <linux/ktime.h> | |
25 | #include <linux/sched.h> | |
26 | ||
b9170836 DJ |
27 | /* |
28 | * dbs is used in this file as a shortform for demandbased switching | |
29 | * It helps to keep variable names smaller, simpler | |
30 | */ | |
31 | ||
32 | #define DEF_FREQUENCY_UP_THRESHOLD (80) | |
b9170836 | 33 | #define DEF_FREQUENCY_DOWN_THRESHOLD (20) |
b9170836 | 34 | |
18a7247d DJ |
35 | /* |
36 | * The polling frequency of this governor depends on the capability of | |
b9170836 | 37 | * the processor. Default polling frequency is 1000 times the transition |
18a7247d DJ |
38 | * latency of the processor. The governor will work on any processor with |
39 | * transition latency <= 10mS, using appropriate sampling | |
b9170836 | 40 | * rate. |
8e677ce8 AC |
41 | * For CPUs with transition latency > 10mS (mostly drivers with CPUFREQ_ETERNAL) |
42 | * this governor will not work. | |
b9170836 DJ |
43 | * All times here are in uS. |
44 | */ | |
18a7247d | 45 | static unsigned int def_sampling_rate; |
2c906b31 AC |
46 | #define MIN_SAMPLING_RATE_RATIO (2) |
47 | /* for correct statistics, we need at least 10 ticks between each measure */ | |
8e677ce8 | 48 | #define MIN_STAT_SAMPLING_RATE \ |
e08f5f5b GS |
49 | (MIN_SAMPLING_RATE_RATIO * jiffies_to_usecs(10)) |
50 | #define MIN_SAMPLING_RATE \ | |
51 | (def_sampling_rate / MIN_SAMPLING_RATE_RATIO) | |
112124ab TR |
52 | /* Above MIN_SAMPLING_RATE will vanish with its sysfs file soon |
53 | * Define the minimal settable sampling rate to the greater of: | |
54 | * - "HW transition latency" * 100 (same as default sampling / 10) | |
55 | * - MIN_STAT_SAMPLING_RATE | |
56 | * To avoid that userspace shoots itself. | |
57 | */ | |
58 | static unsigned int minimum_sampling_rate(void) | |
59 | { | |
60 | return max(def_sampling_rate / 10, MIN_STAT_SAMPLING_RATE); | |
61 | } | |
62 | ||
63 | /* This will also vanish soon with removing sampling_rate_max */ | |
b9170836 | 64 | #define MAX_SAMPLING_RATE (500 * def_sampling_rate) |
112124ab | 65 | #define LATENCY_MULTIPLIER (1000) |
2c906b31 AC |
66 | #define DEF_SAMPLING_DOWN_FACTOR (1) |
67 | #define MAX_SAMPLING_DOWN_FACTOR (10) | |
1c256245 | 68 | #define TRANSITION_LATENCY_LIMIT (10 * 1000 * 1000) |
b9170836 | 69 | |
c4028958 | 70 | static void do_dbs_timer(struct work_struct *work); |
b9170836 DJ |
71 | |
72 | struct cpu_dbs_info_s { | |
8e677ce8 AC |
73 | cputime64_t prev_cpu_idle; |
74 | cputime64_t prev_cpu_wall; | |
75 | cputime64_t prev_cpu_nice; | |
18a7247d | 76 | struct cpufreq_policy *cur_policy; |
8e677ce8 | 77 | struct delayed_work work; |
18a7247d DJ |
78 | unsigned int down_skip; |
79 | unsigned int requested_freq; | |
8e677ce8 AC |
80 | int cpu; |
81 | unsigned int enable:1; | |
b9170836 DJ |
82 | }; |
83 | static DEFINE_PER_CPU(struct cpu_dbs_info_s, cpu_dbs_info); | |
84 | ||
85 | static unsigned int dbs_enable; /* number of CPUs using this policy */ | |
86 | ||
4ec223d0 VP |
87 | /* |
88 | * DEADLOCK ALERT! There is a ordering requirement between cpu_hotplug | |
89 | * lock and dbs_mutex. cpu_hotplug lock should always be held before | |
90 | * dbs_mutex. If any function that can potentially take cpu_hotplug lock | |
91 | * (like __cpufreq_driver_target()) is being called with dbs_mutex taken, then | |
92 | * cpu_hotplug lock should be taken before that. Note that cpu_hotplug lock | |
93 | * is recursive for the same process. -Venki | |
b253d2b2 MD |
94 | * DEADLOCK ALERT! (2) : do_dbs_timer() must not take the dbs_mutex, because it |
95 | * would deadlock with cancel_delayed_work_sync(), which is needed for proper | |
96 | * raceless workqueue teardown. | |
4ec223d0 | 97 | */ |
9acef487 | 98 | static DEFINE_MUTEX(dbs_mutex); |
b9170836 | 99 | |
8e677ce8 AC |
100 | static struct workqueue_struct *kconservative_wq; |
101 | ||
102 | static struct dbs_tuners { | |
18a7247d DJ |
103 | unsigned int sampling_rate; |
104 | unsigned int sampling_down_factor; | |
105 | unsigned int up_threshold; | |
106 | unsigned int down_threshold; | |
107 | unsigned int ignore_nice; | |
108 | unsigned int freq_step; | |
8e677ce8 | 109 | } dbs_tuners_ins = { |
18a7247d DJ |
110 | .up_threshold = DEF_FREQUENCY_UP_THRESHOLD, |
111 | .down_threshold = DEF_FREQUENCY_DOWN_THRESHOLD, | |
112 | .sampling_down_factor = DEF_SAMPLING_DOWN_FACTOR, | |
113 | .ignore_nice = 0, | |
114 | .freq_step = 5, | |
b9170836 DJ |
115 | }; |
116 | ||
8e677ce8 AC |
117 | static inline cputime64_t get_cpu_idle_time_jiffy(unsigned int cpu, |
118 | cputime64_t *wall) | |
dac1c1a5 | 119 | { |
8e677ce8 AC |
120 | cputime64_t idle_time; |
121 | cputime64_t cur_wall_time; | |
122 | cputime64_t busy_time; | |
123 | ||
124 | cur_wall_time = jiffies64_to_cputime64(get_jiffies_64()); | |
125 | busy_time = cputime64_add(kstat_cpu(cpu).cpustat.user, | |
126 | kstat_cpu(cpu).cpustat.system); | |
e08f5f5b | 127 | |
8e677ce8 AC |
128 | busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.irq); |
129 | busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.softirq); | |
130 | busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.steal); | |
131 | busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.nice); | |
e08f5f5b | 132 | |
8e677ce8 AC |
133 | idle_time = cputime64_sub(cur_wall_time, busy_time); |
134 | if (wall) | |
135 | *wall = cur_wall_time; | |
e08f5f5b | 136 | |
8e677ce8 AC |
137 | return idle_time; |
138 | } | |
139 | ||
140 | static inline cputime64_t get_cpu_idle_time(unsigned int cpu, cputime64_t *wall) | |
141 | { | |
142 | u64 idle_time = get_cpu_idle_time_us(cpu, wall); | |
143 | ||
144 | if (idle_time == -1ULL) | |
145 | return get_cpu_idle_time_jiffy(cpu, wall); | |
146 | ||
147 | return idle_time; | |
dac1c1a5 DJ |
148 | } |
149 | ||
a8d7c3bc EO |
150 | /* keep track of frequency transitions */ |
151 | static int | |
152 | dbs_cpufreq_notifier(struct notifier_block *nb, unsigned long val, | |
153 | void *data) | |
154 | { | |
155 | struct cpufreq_freqs *freq = data; | |
156 | struct cpu_dbs_info_s *this_dbs_info = &per_cpu(cpu_dbs_info, | |
157 | freq->cpu); | |
158 | ||
f407a08b AC |
159 | struct cpufreq_policy *policy; |
160 | ||
a8d7c3bc EO |
161 | if (!this_dbs_info->enable) |
162 | return 0; | |
163 | ||
f407a08b AC |
164 | policy = this_dbs_info->cur_policy; |
165 | ||
166 | /* | |
167 | * we only care if our internally tracked freq moves outside | |
168 | * the 'valid' ranges of freqency available to us otherwise | |
169 | * we do not change it | |
170 | */ | |
171 | if (this_dbs_info->requested_freq > policy->max | |
172 | || this_dbs_info->requested_freq < policy->min) | |
173 | this_dbs_info->requested_freq = freq->new; | |
a8d7c3bc EO |
174 | |
175 | return 0; | |
176 | } | |
177 | ||
178 | static struct notifier_block dbs_cpufreq_notifier_block = { | |
179 | .notifier_call = dbs_cpufreq_notifier | |
180 | }; | |
181 | ||
b9170836 DJ |
182 | /************************** sysfs interface ************************/ |
183 | static ssize_t show_sampling_rate_max(struct cpufreq_policy *policy, char *buf) | |
184 | { | |
9411b4ef TR |
185 | static int print_once; |
186 | ||
187 | if (!print_once) { | |
188 | printk(KERN_INFO "CPUFREQ: conservative sampling_rate_max " | |
189 | "sysfs file is deprecated - used by: %s\n", | |
190 | current->comm); | |
191 | print_once = 1; | |
192 | } | |
9acef487 | 193 | return sprintf(buf, "%u\n", MAX_SAMPLING_RATE); |
b9170836 DJ |
194 | } |
195 | ||
196 | static ssize_t show_sampling_rate_min(struct cpufreq_policy *policy, char *buf) | |
197 | { | |
9411b4ef TR |
198 | static int print_once; |
199 | ||
200 | if (!print_once) { | |
201 | printk(KERN_INFO "CPUFREQ: conservative sampling_rate_max " | |
202 | "sysfs file is deprecated - used by: %s\n", current->comm); | |
203 | print_once = 1; | |
204 | } | |
9acef487 | 205 | return sprintf(buf, "%u\n", MIN_SAMPLING_RATE); |
b9170836 DJ |
206 | } |
207 | ||
8e677ce8 AC |
208 | #define define_one_ro(_name) \ |
209 | static struct freq_attr _name = \ | |
b9170836 DJ |
210 | __ATTR(_name, 0444, show_##_name, NULL) |
211 | ||
212 | define_one_ro(sampling_rate_max); | |
213 | define_one_ro(sampling_rate_min); | |
214 | ||
215 | /* cpufreq_conservative Governor Tunables */ | |
216 | #define show_one(file_name, object) \ | |
217 | static ssize_t show_##file_name \ | |
218 | (struct cpufreq_policy *unused, char *buf) \ | |
219 | { \ | |
220 | return sprintf(buf, "%u\n", dbs_tuners_ins.object); \ | |
221 | } | |
222 | show_one(sampling_rate, sampling_rate); | |
223 | show_one(sampling_down_factor, sampling_down_factor); | |
224 | show_one(up_threshold, up_threshold); | |
225 | show_one(down_threshold, down_threshold); | |
001893cd | 226 | show_one(ignore_nice_load, ignore_nice); |
b9170836 DJ |
227 | show_one(freq_step, freq_step); |
228 | ||
18a7247d | 229 | static ssize_t store_sampling_down_factor(struct cpufreq_policy *unused, |
b9170836 DJ |
230 | const char *buf, size_t count) |
231 | { | |
232 | unsigned int input; | |
233 | int ret; | |
9acef487 | 234 | ret = sscanf(buf, "%u", &input); |
8e677ce8 | 235 | |
2c906b31 | 236 | if (ret != 1 || input > MAX_SAMPLING_DOWN_FACTOR || input < 1) |
b9170836 DJ |
237 | return -EINVAL; |
238 | ||
3fc54d37 | 239 | mutex_lock(&dbs_mutex); |
b9170836 | 240 | dbs_tuners_ins.sampling_down_factor = input; |
3fc54d37 | 241 | mutex_unlock(&dbs_mutex); |
b9170836 DJ |
242 | |
243 | return count; | |
244 | } | |
245 | ||
18a7247d | 246 | static ssize_t store_sampling_rate(struct cpufreq_policy *unused, |
b9170836 DJ |
247 | const char *buf, size_t count) |
248 | { | |
249 | unsigned int input; | |
250 | int ret; | |
9acef487 | 251 | ret = sscanf(buf, "%u", &input); |
b9170836 | 252 | |
8e677ce8 | 253 | if (ret != 1) |
b9170836 | 254 | return -EINVAL; |
8e677ce8 AC |
255 | |
256 | mutex_lock(&dbs_mutex); | |
112124ab | 257 | dbs_tuners_ins.sampling_rate = max(input, minimum_sampling_rate()); |
3fc54d37 | 258 | mutex_unlock(&dbs_mutex); |
b9170836 DJ |
259 | |
260 | return count; | |
261 | } | |
262 | ||
18a7247d | 263 | static ssize_t store_up_threshold(struct cpufreq_policy *unused, |
b9170836 DJ |
264 | const char *buf, size_t count) |
265 | { | |
266 | unsigned int input; | |
267 | int ret; | |
9acef487 | 268 | ret = sscanf(buf, "%u", &input); |
b9170836 | 269 | |
3fc54d37 | 270 | mutex_lock(&dbs_mutex); |
9acef487 | 271 | if (ret != 1 || input > 100 || |
8e677ce8 | 272 | input <= dbs_tuners_ins.down_threshold) { |
3fc54d37 | 273 | mutex_unlock(&dbs_mutex); |
b9170836 DJ |
274 | return -EINVAL; |
275 | } | |
276 | ||
277 | dbs_tuners_ins.up_threshold = input; | |
3fc54d37 | 278 | mutex_unlock(&dbs_mutex); |
b9170836 DJ |
279 | |
280 | return count; | |
281 | } | |
282 | ||
18a7247d | 283 | static ssize_t store_down_threshold(struct cpufreq_policy *unused, |
b9170836 DJ |
284 | const char *buf, size_t count) |
285 | { | |
286 | unsigned int input; | |
287 | int ret; | |
9acef487 | 288 | ret = sscanf(buf, "%u", &input); |
b9170836 | 289 | |
3fc54d37 | 290 | mutex_lock(&dbs_mutex); |
8e677ce8 AC |
291 | /* cannot be lower than 11 otherwise freq will not fall */ |
292 | if (ret != 1 || input < 11 || input > 100 || | |
293 | input >= dbs_tuners_ins.up_threshold) { | |
3fc54d37 | 294 | mutex_unlock(&dbs_mutex); |
b9170836 DJ |
295 | return -EINVAL; |
296 | } | |
297 | ||
298 | dbs_tuners_ins.down_threshold = input; | |
3fc54d37 | 299 | mutex_unlock(&dbs_mutex); |
b9170836 DJ |
300 | |
301 | return count; | |
302 | } | |
303 | ||
001893cd | 304 | static ssize_t store_ignore_nice_load(struct cpufreq_policy *policy, |
b9170836 DJ |
305 | const char *buf, size_t count) |
306 | { | |
307 | unsigned int input; | |
308 | int ret; | |
309 | ||
310 | unsigned int j; | |
18a7247d DJ |
311 | |
312 | ret = sscanf(buf, "%u", &input); | |
313 | if (ret != 1) | |
b9170836 DJ |
314 | return -EINVAL; |
315 | ||
18a7247d | 316 | if (input > 1) |
b9170836 | 317 | input = 1; |
18a7247d | 318 | |
3fc54d37 | 319 | mutex_lock(&dbs_mutex); |
18a7247d | 320 | if (input == dbs_tuners_ins.ignore_nice) { /* nothing to do */ |
3fc54d37 | 321 | mutex_unlock(&dbs_mutex); |
b9170836 DJ |
322 | return count; |
323 | } | |
324 | dbs_tuners_ins.ignore_nice = input; | |
325 | ||
8e677ce8 | 326 | /* we need to re-evaluate prev_cpu_idle */ |
dac1c1a5 | 327 | for_each_online_cpu(j) { |
8e677ce8 AC |
328 | struct cpu_dbs_info_s *dbs_info; |
329 | dbs_info = &per_cpu(cpu_dbs_info, j); | |
330 | dbs_info->prev_cpu_idle = get_cpu_idle_time(j, | |
331 | &dbs_info->prev_cpu_wall); | |
332 | if (dbs_tuners_ins.ignore_nice) | |
333 | dbs_info->prev_cpu_nice = kstat_cpu(j).cpustat.nice; | |
b9170836 | 334 | } |
3fc54d37 | 335 | mutex_unlock(&dbs_mutex); |
b9170836 DJ |
336 | |
337 | return count; | |
338 | } | |
339 | ||
340 | static ssize_t store_freq_step(struct cpufreq_policy *policy, | |
341 | const char *buf, size_t count) | |
342 | { | |
343 | unsigned int input; | |
344 | int ret; | |
18a7247d | 345 | ret = sscanf(buf, "%u", &input); |
b9170836 | 346 | |
18a7247d | 347 | if (ret != 1) |
b9170836 DJ |
348 | return -EINVAL; |
349 | ||
18a7247d | 350 | if (input > 100) |
b9170836 | 351 | input = 100; |
18a7247d | 352 | |
b9170836 DJ |
353 | /* no need to test here if freq_step is zero as the user might actually |
354 | * want this, they would be crazy though :) */ | |
3fc54d37 | 355 | mutex_lock(&dbs_mutex); |
b9170836 | 356 | dbs_tuners_ins.freq_step = input; |
3fc54d37 | 357 | mutex_unlock(&dbs_mutex); |
b9170836 DJ |
358 | |
359 | return count; | |
360 | } | |
361 | ||
362 | #define define_one_rw(_name) \ | |
363 | static struct freq_attr _name = \ | |
364 | __ATTR(_name, 0644, show_##_name, store_##_name) | |
365 | ||
366 | define_one_rw(sampling_rate); | |
367 | define_one_rw(sampling_down_factor); | |
368 | define_one_rw(up_threshold); | |
369 | define_one_rw(down_threshold); | |
001893cd | 370 | define_one_rw(ignore_nice_load); |
b9170836 DJ |
371 | define_one_rw(freq_step); |
372 | ||
9acef487 | 373 | static struct attribute *dbs_attributes[] = { |
b9170836 DJ |
374 | &sampling_rate_max.attr, |
375 | &sampling_rate_min.attr, | |
376 | &sampling_rate.attr, | |
377 | &sampling_down_factor.attr, | |
378 | &up_threshold.attr, | |
379 | &down_threshold.attr, | |
001893cd | 380 | &ignore_nice_load.attr, |
b9170836 DJ |
381 | &freq_step.attr, |
382 | NULL | |
383 | }; | |
384 | ||
385 | static struct attribute_group dbs_attr_group = { | |
386 | .attrs = dbs_attributes, | |
387 | .name = "conservative", | |
388 | }; | |
389 | ||
390 | /************************** sysfs end ************************/ | |
391 | ||
8e677ce8 | 392 | static void dbs_check_cpu(struct cpu_dbs_info_s *this_dbs_info) |
b9170836 | 393 | { |
8e677ce8 | 394 | unsigned int load = 0; |
f068c04b | 395 | unsigned int freq_target; |
b9170836 | 396 | |
8e677ce8 AC |
397 | struct cpufreq_policy *policy; |
398 | unsigned int j; | |
b9170836 | 399 | |
08a28e2e AC |
400 | policy = this_dbs_info->cur_policy; |
401 | ||
18a7247d | 402 | /* |
8e677ce8 AC |
403 | * Every sampling_rate, we check, if current idle time is less |
404 | * than 20% (default), then we try to increase frequency | |
405 | * Every sampling_rate*sampling_down_factor, we check, if current | |
406 | * idle time is more than 80%, then we try to decrease frequency | |
b9170836 | 407 | * |
18a7247d DJ |
408 | * Any frequency increase takes it to the maximum frequency. |
409 | * Frequency reduction happens at minimum steps of | |
8e677ce8 | 410 | * 5% (default) of maximum frequency |
b9170836 DJ |
411 | */ |
412 | ||
8e677ce8 AC |
413 | /* Get Absolute Load */ |
414 | for_each_cpu(j, policy->cpus) { | |
415 | struct cpu_dbs_info_s *j_dbs_info; | |
416 | cputime64_t cur_wall_time, cur_idle_time; | |
417 | unsigned int idle_time, wall_time; | |
b9170836 | 418 | |
8e677ce8 AC |
419 | j_dbs_info = &per_cpu(cpu_dbs_info, j); |
420 | ||
421 | cur_idle_time = get_cpu_idle_time(j, &cur_wall_time); | |
422 | ||
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; | |
08a28e2e | 426 | |
8e677ce8 AC |
427 | idle_time = (unsigned int) cputime64_sub(cur_idle_time, |
428 | j_dbs_info->prev_cpu_idle); | |
429 | j_dbs_info->prev_cpu_idle = cur_idle_time; | |
b9170836 | 430 | |
8e677ce8 AC |
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 | ||
448 | if (unlikely(!wall_time || wall_time < idle_time)) | |
449 | continue; | |
450 | ||
451 | load = 100 * (wall_time - idle_time) / wall_time; | |
452 | } | |
453 | ||
454 | /* | |
455 | * break out if we 'cannot' reduce the speed as the user might | |
456 | * want freq_step to be zero | |
457 | */ | |
458 | if (dbs_tuners_ins.freq_step == 0) | |
459 | return; | |
b9170836 | 460 | |
8e677ce8 AC |
461 | /* Check for frequency increase */ |
462 | if (load > dbs_tuners_ins.up_threshold) { | |
a159b827 | 463 | this_dbs_info->down_skip = 0; |
790d76fa | 464 | |
b9170836 | 465 | /* if we are already at full speed then break out early */ |
a159b827 | 466 | if (this_dbs_info->requested_freq == policy->max) |
b9170836 | 467 | return; |
18a7247d | 468 | |
f068c04b | 469 | freq_target = (dbs_tuners_ins.freq_step * policy->max) / 100; |
b9170836 DJ |
470 | |
471 | /* max freq cannot be less than 100. But who knows.... */ | |
f068c04b DJ |
472 | if (unlikely(freq_target == 0)) |
473 | freq_target = 5; | |
18a7247d | 474 | |
f068c04b | 475 | this_dbs_info->requested_freq += freq_target; |
a159b827 AC |
476 | if (this_dbs_info->requested_freq > policy->max) |
477 | this_dbs_info->requested_freq = policy->max; | |
b9170836 | 478 | |
a159b827 | 479 | __cpufreq_driver_target(policy, this_dbs_info->requested_freq, |
b9170836 | 480 | CPUFREQ_RELATION_H); |
b9170836 DJ |
481 | return; |
482 | } | |
483 | ||
8e677ce8 AC |
484 | /* |
485 | * The optimal frequency is the frequency that is the lowest that | |
486 | * can support the current CPU usage without triggering the up | |
487 | * policy. To be safe, we focus 10 points under the threshold. | |
488 | */ | |
489 | if (load < (dbs_tuners_ins.down_threshold - 10)) { | |
f068c04b | 490 | freq_target = (dbs_tuners_ins.freq_step * policy->max) / 100; |
b9170836 | 491 | |
f068c04b | 492 | this_dbs_info->requested_freq -= freq_target; |
a159b827 AC |
493 | if (this_dbs_info->requested_freq < policy->min) |
494 | this_dbs_info->requested_freq = policy->min; | |
b9170836 | 495 | |
8e677ce8 AC |
496 | /* |
497 | * if we cannot reduce the frequency anymore, break out early | |
498 | */ | |
499 | if (policy->cur == policy->min) | |
500 | return; | |
501 | ||
a159b827 | 502 | __cpufreq_driver_target(policy, this_dbs_info->requested_freq, |
2c906b31 | 503 | CPUFREQ_RELATION_H); |
b9170836 DJ |
504 | return; |
505 | } | |
506 | } | |
507 | ||
c4028958 | 508 | static void do_dbs_timer(struct work_struct *work) |
18a7247d | 509 | { |
8e677ce8 AC |
510 | struct cpu_dbs_info_s *dbs_info = |
511 | container_of(work, struct cpu_dbs_info_s, work.work); | |
512 | unsigned int cpu = dbs_info->cpu; | |
513 | ||
514 | /* We want all CPUs to do sampling nearly on same jiffy */ | |
515 | int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate); | |
516 | ||
517 | delay -= jiffies % delay; | |
518 | ||
519 | if (lock_policy_rwsem_write(cpu) < 0) | |
520 | return; | |
521 | ||
522 | if (!dbs_info->enable) { | |
523 | unlock_policy_rwsem_write(cpu); | |
524 | return; | |
525 | } | |
526 | ||
527 | dbs_check_cpu(dbs_info); | |
528 | ||
529 | queue_delayed_work_on(cpu, kconservative_wq, &dbs_info->work, delay); | |
530 | unlock_policy_rwsem_write(cpu); | |
18a7247d | 531 | } |
b9170836 | 532 | |
8e677ce8 | 533 | static inline void dbs_timer_init(struct cpu_dbs_info_s *dbs_info) |
b9170836 | 534 | { |
8e677ce8 AC |
535 | /* We want all CPUs to do sampling nearly on same jiffy */ |
536 | int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate); | |
537 | delay -= jiffies % delay; | |
538 | ||
539 | dbs_info->enable = 1; | |
540 | INIT_DELAYED_WORK_DEFERRABLE(&dbs_info->work, do_dbs_timer); | |
541 | queue_delayed_work_on(dbs_info->cpu, kconservative_wq, &dbs_info->work, | |
542 | delay); | |
b9170836 DJ |
543 | } |
544 | ||
8e677ce8 | 545 | static inline void dbs_timer_exit(struct cpu_dbs_info_s *dbs_info) |
b9170836 | 546 | { |
8e677ce8 | 547 | dbs_info->enable = 0; |
b253d2b2 | 548 | cancel_delayed_work_sync(&dbs_info->work); |
b9170836 DJ |
549 | } |
550 | ||
551 | static int cpufreq_governor_dbs(struct cpufreq_policy *policy, | |
552 | unsigned int event) | |
553 | { | |
554 | unsigned int cpu = policy->cpu; | |
555 | struct cpu_dbs_info_s *this_dbs_info; | |
556 | unsigned int j; | |
914f7c31 | 557 | int rc; |
b9170836 DJ |
558 | |
559 | this_dbs_info = &per_cpu(cpu_dbs_info, cpu); | |
560 | ||
561 | switch (event) { | |
562 | case CPUFREQ_GOV_START: | |
18a7247d | 563 | if ((!cpu_online(cpu)) || (!policy->cur)) |
b9170836 DJ |
564 | return -EINVAL; |
565 | ||
b9170836 DJ |
566 | if (this_dbs_info->enable) /* Already enabled */ |
567 | break; | |
18a7247d | 568 | |
3fc54d37 | 569 | mutex_lock(&dbs_mutex); |
914f7c31 JG |
570 | |
571 | rc = sysfs_create_group(&policy->kobj, &dbs_attr_group); | |
572 | if (rc) { | |
573 | mutex_unlock(&dbs_mutex); | |
574 | return rc; | |
575 | } | |
576 | ||
835481d9 | 577 | for_each_cpu(j, policy->cpus) { |
b9170836 DJ |
578 | struct cpu_dbs_info_s *j_dbs_info; |
579 | j_dbs_info = &per_cpu(cpu_dbs_info, j); | |
580 | j_dbs_info->cur_policy = policy; | |
18a7247d | 581 | |
8e677ce8 AC |
582 | j_dbs_info->prev_cpu_idle = get_cpu_idle_time(j, |
583 | &j_dbs_info->prev_cpu_wall); | |
584 | if (dbs_tuners_ins.ignore_nice) { | |
585 | j_dbs_info->prev_cpu_nice = | |
586 | kstat_cpu(j).cpustat.nice; | |
587 | } | |
b9170836 | 588 | } |
a159b827 AC |
589 | this_dbs_info->down_skip = 0; |
590 | this_dbs_info->requested_freq = policy->cur; | |
914f7c31 | 591 | |
b9170836 DJ |
592 | dbs_enable++; |
593 | /* | |
594 | * Start the timerschedule work, when this governor | |
595 | * is used for first time | |
596 | */ | |
597 | if (dbs_enable == 1) { | |
598 | unsigned int latency; | |
599 | /* policy latency is in nS. Convert it to uS first */ | |
2c906b31 AC |
600 | latency = policy->cpuinfo.transition_latency / 1000; |
601 | if (latency == 0) | |
602 | latency = 1; | |
b9170836 | 603 | |
112124ab | 604 | def_sampling_rate = |
a75603a0 | 605 | max(latency * LATENCY_MULTIPLIER, |
112124ab | 606 | MIN_STAT_SAMPLING_RATE); |
2c906b31 | 607 | |
b9170836 | 608 | dbs_tuners_ins.sampling_rate = def_sampling_rate; |
b9170836 | 609 | |
a8d7c3bc EO |
610 | cpufreq_register_notifier( |
611 | &dbs_cpufreq_notifier_block, | |
612 | CPUFREQ_TRANSITION_NOTIFIER); | |
b9170836 | 613 | } |
8e677ce8 | 614 | dbs_timer_init(this_dbs_info); |
18a7247d | 615 | |
3fc54d37 | 616 | mutex_unlock(&dbs_mutex); |
8e677ce8 | 617 | |
b9170836 DJ |
618 | break; |
619 | ||
620 | case CPUFREQ_GOV_STOP: | |
3fc54d37 | 621 | mutex_lock(&dbs_mutex); |
8e677ce8 | 622 | dbs_timer_exit(this_dbs_info); |
b9170836 DJ |
623 | sysfs_remove_group(&policy->kobj, &dbs_attr_group); |
624 | dbs_enable--; | |
8e677ce8 | 625 | |
b9170836 DJ |
626 | /* |
627 | * Stop the timerschedule work, when this governor | |
628 | * is used for first time | |
629 | */ | |
8e677ce8 | 630 | if (dbs_enable == 0) |
a8d7c3bc EO |
631 | cpufreq_unregister_notifier( |
632 | &dbs_cpufreq_notifier_block, | |
633 | CPUFREQ_TRANSITION_NOTIFIER); | |
a8d7c3bc | 634 | |
3fc54d37 | 635 | mutex_unlock(&dbs_mutex); |
b9170836 DJ |
636 | |
637 | break; | |
638 | ||
639 | case CPUFREQ_GOV_LIMITS: | |
3fc54d37 | 640 | mutex_lock(&dbs_mutex); |
b9170836 DJ |
641 | if (policy->max < this_dbs_info->cur_policy->cur) |
642 | __cpufreq_driver_target( | |
643 | this_dbs_info->cur_policy, | |
18a7247d | 644 | policy->max, CPUFREQ_RELATION_H); |
b9170836 DJ |
645 | else if (policy->min > this_dbs_info->cur_policy->cur) |
646 | __cpufreq_driver_target( | |
647 | this_dbs_info->cur_policy, | |
18a7247d | 648 | policy->min, CPUFREQ_RELATION_L); |
3fc54d37 | 649 | mutex_unlock(&dbs_mutex); |
8e677ce8 | 650 | |
b9170836 DJ |
651 | break; |
652 | } | |
653 | return 0; | |
654 | } | |
655 | ||
c4d14bc0 SW |
656 | #ifndef CONFIG_CPU_FREQ_DEFAULT_GOV_CONSERVATIVE |
657 | static | |
658 | #endif | |
1c256245 TR |
659 | struct cpufreq_governor cpufreq_gov_conservative = { |
660 | .name = "conservative", | |
661 | .governor = cpufreq_governor_dbs, | |
662 | .max_transition_latency = TRANSITION_LATENCY_LIMIT, | |
663 | .owner = THIS_MODULE, | |
b9170836 DJ |
664 | }; |
665 | ||
666 | static int __init cpufreq_gov_dbs_init(void) | |
667 | { | |
8e677ce8 AC |
668 | int err; |
669 | ||
670 | kconservative_wq = create_workqueue("kconservative"); | |
671 | if (!kconservative_wq) { | |
672 | printk(KERN_ERR "Creation of kconservative failed\n"); | |
673 | return -EFAULT; | |
674 | } | |
675 | ||
676 | err = cpufreq_register_governor(&cpufreq_gov_conservative); | |
677 | if (err) | |
678 | destroy_workqueue(kconservative_wq); | |
679 | ||
680 | return err; | |
b9170836 DJ |
681 | } |
682 | ||
683 | static void __exit cpufreq_gov_dbs_exit(void) | |
684 | { | |
1c256245 | 685 | cpufreq_unregister_governor(&cpufreq_gov_conservative); |
8e677ce8 | 686 | destroy_workqueue(kconservative_wq); |
b9170836 DJ |
687 | } |
688 | ||
689 | ||
11a80a9c | 690 | MODULE_AUTHOR("Alexander Clouter <alex@digriz.org.uk>"); |
9acef487 | 691 | MODULE_DESCRIPTION("'cpufreq_conservative' - A dynamic cpufreq governor for " |
b9170836 DJ |
692 | "Low Latency Frequency Transition capable processors " |
693 | "optimised for use in a battery environment"); | |
9acef487 | 694 | MODULE_LICENSE("GPL"); |
b9170836 | 695 | |
6915719b JW |
696 | #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_CONSERVATIVE |
697 | fs_initcall(cpufreq_gov_dbs_init); | |
698 | #else | |
b9170836 | 699 | module_init(cpufreq_gov_dbs_init); |
6915719b | 700 | #endif |
b9170836 | 701 | module_exit(cpufreq_gov_dbs_exit); |