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 | |
94 | */ | |
9acef487 | 95 | static DEFINE_MUTEX(dbs_mutex); |
b9170836 | 96 | |
8e677ce8 AC |
97 | static struct workqueue_struct *kconservative_wq; |
98 | ||
99 | static struct dbs_tuners { | |
18a7247d DJ |
100 | unsigned int sampling_rate; |
101 | unsigned int sampling_down_factor; | |
102 | unsigned int up_threshold; | |
103 | unsigned int down_threshold; | |
104 | unsigned int ignore_nice; | |
105 | unsigned int freq_step; | |
8e677ce8 | 106 | } dbs_tuners_ins = { |
18a7247d DJ |
107 | .up_threshold = DEF_FREQUENCY_UP_THRESHOLD, |
108 | .down_threshold = DEF_FREQUENCY_DOWN_THRESHOLD, | |
109 | .sampling_down_factor = DEF_SAMPLING_DOWN_FACTOR, | |
110 | .ignore_nice = 0, | |
111 | .freq_step = 5, | |
b9170836 DJ |
112 | }; |
113 | ||
8e677ce8 AC |
114 | static inline cputime64_t get_cpu_idle_time_jiffy(unsigned int cpu, |
115 | cputime64_t *wall) | |
dac1c1a5 | 116 | { |
8e677ce8 AC |
117 | cputime64_t idle_time; |
118 | cputime64_t cur_wall_time; | |
119 | cputime64_t busy_time; | |
120 | ||
121 | cur_wall_time = jiffies64_to_cputime64(get_jiffies_64()); | |
122 | busy_time = cputime64_add(kstat_cpu(cpu).cpustat.user, | |
123 | kstat_cpu(cpu).cpustat.system); | |
e08f5f5b | 124 | |
8e677ce8 AC |
125 | busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.irq); |
126 | busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.softirq); | |
127 | busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.steal); | |
128 | busy_time = cputime64_add(busy_time, kstat_cpu(cpu).cpustat.nice); | |
e08f5f5b | 129 | |
8e677ce8 AC |
130 | idle_time = cputime64_sub(cur_wall_time, busy_time); |
131 | if (wall) | |
132 | *wall = cur_wall_time; | |
e08f5f5b | 133 | |
8e677ce8 AC |
134 | return idle_time; |
135 | } | |
136 | ||
137 | static inline cputime64_t get_cpu_idle_time(unsigned int cpu, cputime64_t *wall) | |
138 | { | |
139 | u64 idle_time = get_cpu_idle_time_us(cpu, wall); | |
140 | ||
141 | if (idle_time == -1ULL) | |
142 | return get_cpu_idle_time_jiffy(cpu, wall); | |
143 | ||
144 | return idle_time; | |
dac1c1a5 DJ |
145 | } |
146 | ||
a8d7c3bc EO |
147 | /* keep track of frequency transitions */ |
148 | static int | |
149 | dbs_cpufreq_notifier(struct notifier_block *nb, unsigned long val, | |
150 | void *data) | |
151 | { | |
152 | struct cpufreq_freqs *freq = data; | |
153 | struct cpu_dbs_info_s *this_dbs_info = &per_cpu(cpu_dbs_info, | |
154 | freq->cpu); | |
155 | ||
f407a08b AC |
156 | struct cpufreq_policy *policy; |
157 | ||
a8d7c3bc EO |
158 | if (!this_dbs_info->enable) |
159 | return 0; | |
160 | ||
f407a08b AC |
161 | policy = this_dbs_info->cur_policy; |
162 | ||
163 | /* | |
164 | * we only care if our internally tracked freq moves outside | |
165 | * the 'valid' ranges of freqency available to us otherwise | |
166 | * we do not change it | |
167 | */ | |
168 | if (this_dbs_info->requested_freq > policy->max | |
169 | || this_dbs_info->requested_freq < policy->min) | |
170 | this_dbs_info->requested_freq = freq->new; | |
a8d7c3bc EO |
171 | |
172 | return 0; | |
173 | } | |
174 | ||
175 | static struct notifier_block dbs_cpufreq_notifier_block = { | |
176 | .notifier_call = dbs_cpufreq_notifier | |
177 | }; | |
178 | ||
b9170836 DJ |
179 | /************************** sysfs interface ************************/ |
180 | static ssize_t show_sampling_rate_max(struct cpufreq_policy *policy, char *buf) | |
181 | { | |
9411b4ef TR |
182 | static int print_once; |
183 | ||
184 | if (!print_once) { | |
185 | printk(KERN_INFO "CPUFREQ: conservative sampling_rate_max " | |
186 | "sysfs file is deprecated - used by: %s\n", | |
187 | current->comm); | |
188 | print_once = 1; | |
189 | } | |
9acef487 | 190 | return sprintf(buf, "%u\n", MAX_SAMPLING_RATE); |
b9170836 DJ |
191 | } |
192 | ||
193 | static ssize_t show_sampling_rate_min(struct cpufreq_policy *policy, char *buf) | |
194 | { | |
9411b4ef TR |
195 | static int print_once; |
196 | ||
197 | if (!print_once) { | |
198 | printk(KERN_INFO "CPUFREQ: conservative sampling_rate_max " | |
199 | "sysfs file is deprecated - used by: %s\n", current->comm); | |
200 | print_once = 1; | |
201 | } | |
9acef487 | 202 | return sprintf(buf, "%u\n", MIN_SAMPLING_RATE); |
b9170836 DJ |
203 | } |
204 | ||
8e677ce8 AC |
205 | #define define_one_ro(_name) \ |
206 | static struct freq_attr _name = \ | |
b9170836 DJ |
207 | __ATTR(_name, 0444, show_##_name, NULL) |
208 | ||
209 | define_one_ro(sampling_rate_max); | |
210 | define_one_ro(sampling_rate_min); | |
211 | ||
212 | /* cpufreq_conservative Governor Tunables */ | |
213 | #define show_one(file_name, object) \ | |
214 | static ssize_t show_##file_name \ | |
215 | (struct cpufreq_policy *unused, char *buf) \ | |
216 | { \ | |
217 | return sprintf(buf, "%u\n", dbs_tuners_ins.object); \ | |
218 | } | |
219 | show_one(sampling_rate, sampling_rate); | |
220 | show_one(sampling_down_factor, sampling_down_factor); | |
221 | show_one(up_threshold, up_threshold); | |
222 | show_one(down_threshold, down_threshold); | |
001893cd | 223 | show_one(ignore_nice_load, ignore_nice); |
b9170836 DJ |
224 | show_one(freq_step, freq_step); |
225 | ||
18a7247d | 226 | static ssize_t store_sampling_down_factor(struct cpufreq_policy *unused, |
b9170836 DJ |
227 | const char *buf, size_t count) |
228 | { | |
229 | unsigned int input; | |
230 | int ret; | |
9acef487 | 231 | ret = sscanf(buf, "%u", &input); |
8e677ce8 | 232 | |
2c906b31 | 233 | if (ret != 1 || input > MAX_SAMPLING_DOWN_FACTOR || input < 1) |
b9170836 DJ |
234 | return -EINVAL; |
235 | ||
3fc54d37 | 236 | mutex_lock(&dbs_mutex); |
b9170836 | 237 | dbs_tuners_ins.sampling_down_factor = input; |
3fc54d37 | 238 | mutex_unlock(&dbs_mutex); |
b9170836 DJ |
239 | |
240 | return count; | |
241 | } | |
242 | ||
18a7247d | 243 | static ssize_t store_sampling_rate(struct cpufreq_policy *unused, |
b9170836 DJ |
244 | const char *buf, size_t count) |
245 | { | |
246 | unsigned int input; | |
247 | int ret; | |
9acef487 | 248 | ret = sscanf(buf, "%u", &input); |
b9170836 | 249 | |
8e677ce8 | 250 | if (ret != 1) |
b9170836 | 251 | return -EINVAL; |
8e677ce8 AC |
252 | |
253 | mutex_lock(&dbs_mutex); | |
112124ab | 254 | dbs_tuners_ins.sampling_rate = max(input, minimum_sampling_rate()); |
3fc54d37 | 255 | mutex_unlock(&dbs_mutex); |
b9170836 DJ |
256 | |
257 | return count; | |
258 | } | |
259 | ||
18a7247d | 260 | static ssize_t store_up_threshold(struct cpufreq_policy *unused, |
b9170836 DJ |
261 | const char *buf, size_t count) |
262 | { | |
263 | unsigned int input; | |
264 | int ret; | |
9acef487 | 265 | ret = sscanf(buf, "%u", &input); |
b9170836 | 266 | |
3fc54d37 | 267 | mutex_lock(&dbs_mutex); |
9acef487 | 268 | if (ret != 1 || input > 100 || |
8e677ce8 | 269 | input <= dbs_tuners_ins.down_threshold) { |
3fc54d37 | 270 | mutex_unlock(&dbs_mutex); |
b9170836 DJ |
271 | return -EINVAL; |
272 | } | |
273 | ||
274 | dbs_tuners_ins.up_threshold = input; | |
3fc54d37 | 275 | mutex_unlock(&dbs_mutex); |
b9170836 DJ |
276 | |
277 | return count; | |
278 | } | |
279 | ||
18a7247d | 280 | static ssize_t store_down_threshold(struct cpufreq_policy *unused, |
b9170836 DJ |
281 | const char *buf, size_t count) |
282 | { | |
283 | unsigned int input; | |
284 | int ret; | |
9acef487 | 285 | ret = sscanf(buf, "%u", &input); |
b9170836 | 286 | |
3fc54d37 | 287 | mutex_lock(&dbs_mutex); |
8e677ce8 AC |
288 | /* cannot be lower than 11 otherwise freq will not fall */ |
289 | if (ret != 1 || input < 11 || input > 100 || | |
290 | input >= dbs_tuners_ins.up_threshold) { | |
3fc54d37 | 291 | mutex_unlock(&dbs_mutex); |
b9170836 DJ |
292 | return -EINVAL; |
293 | } | |
294 | ||
295 | dbs_tuners_ins.down_threshold = input; | |
3fc54d37 | 296 | mutex_unlock(&dbs_mutex); |
b9170836 DJ |
297 | |
298 | return count; | |
299 | } | |
300 | ||
001893cd | 301 | static ssize_t store_ignore_nice_load(struct cpufreq_policy *policy, |
b9170836 DJ |
302 | const char *buf, size_t count) |
303 | { | |
304 | unsigned int input; | |
305 | int ret; | |
306 | ||
307 | unsigned int j; | |
18a7247d DJ |
308 | |
309 | ret = sscanf(buf, "%u", &input); | |
310 | if (ret != 1) | |
b9170836 DJ |
311 | return -EINVAL; |
312 | ||
18a7247d | 313 | if (input > 1) |
b9170836 | 314 | input = 1; |
18a7247d | 315 | |
3fc54d37 | 316 | mutex_lock(&dbs_mutex); |
18a7247d | 317 | if (input == dbs_tuners_ins.ignore_nice) { /* nothing to do */ |
3fc54d37 | 318 | mutex_unlock(&dbs_mutex); |
b9170836 DJ |
319 | return count; |
320 | } | |
321 | dbs_tuners_ins.ignore_nice = input; | |
322 | ||
8e677ce8 | 323 | /* we need to re-evaluate prev_cpu_idle */ |
dac1c1a5 | 324 | for_each_online_cpu(j) { |
8e677ce8 AC |
325 | struct cpu_dbs_info_s *dbs_info; |
326 | dbs_info = &per_cpu(cpu_dbs_info, j); | |
327 | dbs_info->prev_cpu_idle = get_cpu_idle_time(j, | |
328 | &dbs_info->prev_cpu_wall); | |
329 | if (dbs_tuners_ins.ignore_nice) | |
330 | dbs_info->prev_cpu_nice = kstat_cpu(j).cpustat.nice; | |
b9170836 | 331 | } |
3fc54d37 | 332 | mutex_unlock(&dbs_mutex); |
b9170836 DJ |
333 | |
334 | return count; | |
335 | } | |
336 | ||
337 | static ssize_t store_freq_step(struct cpufreq_policy *policy, | |
338 | const char *buf, size_t count) | |
339 | { | |
340 | unsigned int input; | |
341 | int ret; | |
18a7247d | 342 | ret = sscanf(buf, "%u", &input); |
b9170836 | 343 | |
18a7247d | 344 | if (ret != 1) |
b9170836 DJ |
345 | return -EINVAL; |
346 | ||
18a7247d | 347 | if (input > 100) |
b9170836 | 348 | input = 100; |
18a7247d | 349 | |
b9170836 DJ |
350 | /* no need to test here if freq_step is zero as the user might actually |
351 | * want this, they would be crazy though :) */ | |
3fc54d37 | 352 | mutex_lock(&dbs_mutex); |
b9170836 | 353 | dbs_tuners_ins.freq_step = input; |
3fc54d37 | 354 | mutex_unlock(&dbs_mutex); |
b9170836 DJ |
355 | |
356 | return count; | |
357 | } | |
358 | ||
359 | #define define_one_rw(_name) \ | |
360 | static struct freq_attr _name = \ | |
361 | __ATTR(_name, 0644, show_##_name, store_##_name) | |
362 | ||
363 | define_one_rw(sampling_rate); | |
364 | define_one_rw(sampling_down_factor); | |
365 | define_one_rw(up_threshold); | |
366 | define_one_rw(down_threshold); | |
001893cd | 367 | define_one_rw(ignore_nice_load); |
b9170836 DJ |
368 | define_one_rw(freq_step); |
369 | ||
9acef487 | 370 | static struct attribute *dbs_attributes[] = { |
b9170836 DJ |
371 | &sampling_rate_max.attr, |
372 | &sampling_rate_min.attr, | |
373 | &sampling_rate.attr, | |
374 | &sampling_down_factor.attr, | |
375 | &up_threshold.attr, | |
376 | &down_threshold.attr, | |
001893cd | 377 | &ignore_nice_load.attr, |
b9170836 DJ |
378 | &freq_step.attr, |
379 | NULL | |
380 | }; | |
381 | ||
382 | static struct attribute_group dbs_attr_group = { | |
383 | .attrs = dbs_attributes, | |
384 | .name = "conservative", | |
385 | }; | |
386 | ||
387 | /************************** sysfs end ************************/ | |
388 | ||
8e677ce8 | 389 | static void dbs_check_cpu(struct cpu_dbs_info_s *this_dbs_info) |
b9170836 | 390 | { |
8e677ce8 | 391 | unsigned int load = 0; |
f068c04b | 392 | unsigned int freq_target; |
b9170836 | 393 | |
8e677ce8 AC |
394 | struct cpufreq_policy *policy; |
395 | unsigned int j; | |
b9170836 | 396 | |
08a28e2e AC |
397 | policy = this_dbs_info->cur_policy; |
398 | ||
18a7247d | 399 | /* |
8e677ce8 AC |
400 | * Every sampling_rate, we check, if current idle time is less |
401 | * than 20% (default), then we try to increase frequency | |
402 | * Every sampling_rate*sampling_down_factor, we check, if current | |
403 | * idle time is more than 80%, then we try to decrease frequency | |
b9170836 | 404 | * |
18a7247d DJ |
405 | * Any frequency increase takes it to the maximum frequency. |
406 | * Frequency reduction happens at minimum steps of | |
8e677ce8 | 407 | * 5% (default) of maximum frequency |
b9170836 DJ |
408 | */ |
409 | ||
8e677ce8 AC |
410 | /* Get Absolute Load */ |
411 | for_each_cpu(j, policy->cpus) { | |
412 | struct cpu_dbs_info_s *j_dbs_info; | |
413 | cputime64_t cur_wall_time, cur_idle_time; | |
414 | unsigned int idle_time, wall_time; | |
b9170836 | 415 | |
8e677ce8 AC |
416 | j_dbs_info = &per_cpu(cpu_dbs_info, j); |
417 | ||
418 | cur_idle_time = get_cpu_idle_time(j, &cur_wall_time); | |
419 | ||
420 | wall_time = (unsigned int) cputime64_sub(cur_wall_time, | |
421 | j_dbs_info->prev_cpu_wall); | |
422 | j_dbs_info->prev_cpu_wall = cur_wall_time; | |
08a28e2e | 423 | |
8e677ce8 AC |
424 | idle_time = (unsigned int) cputime64_sub(cur_idle_time, |
425 | j_dbs_info->prev_cpu_idle); | |
426 | j_dbs_info->prev_cpu_idle = cur_idle_time; | |
b9170836 | 427 | |
8e677ce8 AC |
428 | if (dbs_tuners_ins.ignore_nice) { |
429 | cputime64_t cur_nice; | |
430 | unsigned long cur_nice_jiffies; | |
431 | ||
432 | cur_nice = cputime64_sub(kstat_cpu(j).cpustat.nice, | |
433 | j_dbs_info->prev_cpu_nice); | |
434 | /* | |
435 | * Assumption: nice time between sampling periods will | |
436 | * be less than 2^32 jiffies for 32 bit sys | |
437 | */ | |
438 | cur_nice_jiffies = (unsigned long) | |
439 | cputime64_to_jiffies64(cur_nice); | |
440 | ||
441 | j_dbs_info->prev_cpu_nice = kstat_cpu(j).cpustat.nice; | |
442 | idle_time += jiffies_to_usecs(cur_nice_jiffies); | |
443 | } | |
444 | ||
445 | if (unlikely(!wall_time || wall_time < idle_time)) | |
446 | continue; | |
447 | ||
448 | load = 100 * (wall_time - idle_time) / wall_time; | |
449 | } | |
450 | ||
451 | /* | |
452 | * break out if we 'cannot' reduce the speed as the user might | |
453 | * want freq_step to be zero | |
454 | */ | |
455 | if (dbs_tuners_ins.freq_step == 0) | |
456 | return; | |
b9170836 | 457 | |
8e677ce8 AC |
458 | /* Check for frequency increase */ |
459 | if (load > dbs_tuners_ins.up_threshold) { | |
a159b827 | 460 | this_dbs_info->down_skip = 0; |
790d76fa | 461 | |
b9170836 | 462 | /* if we are already at full speed then break out early */ |
a159b827 | 463 | if (this_dbs_info->requested_freq == policy->max) |
b9170836 | 464 | return; |
18a7247d | 465 | |
f068c04b | 466 | freq_target = (dbs_tuners_ins.freq_step * policy->max) / 100; |
b9170836 DJ |
467 | |
468 | /* max freq cannot be less than 100. But who knows.... */ | |
f068c04b DJ |
469 | if (unlikely(freq_target == 0)) |
470 | freq_target = 5; | |
18a7247d | 471 | |
f068c04b | 472 | this_dbs_info->requested_freq += freq_target; |
a159b827 AC |
473 | if (this_dbs_info->requested_freq > policy->max) |
474 | this_dbs_info->requested_freq = policy->max; | |
b9170836 | 475 | |
a159b827 | 476 | __cpufreq_driver_target(policy, this_dbs_info->requested_freq, |
b9170836 | 477 | CPUFREQ_RELATION_H); |
b9170836 DJ |
478 | return; |
479 | } | |
480 | ||
8e677ce8 AC |
481 | /* |
482 | * The optimal frequency is the frequency that is the lowest that | |
483 | * can support the current CPU usage without triggering the up | |
484 | * policy. To be safe, we focus 10 points under the threshold. | |
485 | */ | |
486 | if (load < (dbs_tuners_ins.down_threshold - 10)) { | |
f068c04b | 487 | freq_target = (dbs_tuners_ins.freq_step * policy->max) / 100; |
b9170836 | 488 | |
f068c04b | 489 | this_dbs_info->requested_freq -= freq_target; |
a159b827 AC |
490 | if (this_dbs_info->requested_freq < policy->min) |
491 | this_dbs_info->requested_freq = policy->min; | |
b9170836 | 492 | |
8e677ce8 AC |
493 | /* |
494 | * if we cannot reduce the frequency anymore, break out early | |
495 | */ | |
496 | if (policy->cur == policy->min) | |
497 | return; | |
498 | ||
a159b827 | 499 | __cpufreq_driver_target(policy, this_dbs_info->requested_freq, |
2c906b31 | 500 | CPUFREQ_RELATION_H); |
b9170836 DJ |
501 | return; |
502 | } | |
503 | } | |
504 | ||
c4028958 | 505 | static void do_dbs_timer(struct work_struct *work) |
18a7247d | 506 | { |
8e677ce8 AC |
507 | struct cpu_dbs_info_s *dbs_info = |
508 | container_of(work, struct cpu_dbs_info_s, work.work); | |
509 | unsigned int cpu = dbs_info->cpu; | |
510 | ||
511 | /* We want all CPUs to do sampling nearly on same jiffy */ | |
512 | int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate); | |
513 | ||
514 | delay -= jiffies % delay; | |
515 | ||
516 | if (lock_policy_rwsem_write(cpu) < 0) | |
517 | return; | |
518 | ||
519 | if (!dbs_info->enable) { | |
520 | unlock_policy_rwsem_write(cpu); | |
521 | return; | |
522 | } | |
523 | ||
524 | dbs_check_cpu(dbs_info); | |
525 | ||
526 | queue_delayed_work_on(cpu, kconservative_wq, &dbs_info->work, delay); | |
527 | unlock_policy_rwsem_write(cpu); | |
18a7247d | 528 | } |
b9170836 | 529 | |
8e677ce8 | 530 | static inline void dbs_timer_init(struct cpu_dbs_info_s *dbs_info) |
b9170836 | 531 | { |
8e677ce8 AC |
532 | /* We want all CPUs to do sampling nearly on same jiffy */ |
533 | int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate); | |
534 | delay -= jiffies % delay; | |
535 | ||
536 | dbs_info->enable = 1; | |
537 | INIT_DELAYED_WORK_DEFERRABLE(&dbs_info->work, do_dbs_timer); | |
538 | queue_delayed_work_on(dbs_info->cpu, kconservative_wq, &dbs_info->work, | |
539 | delay); | |
b9170836 DJ |
540 | } |
541 | ||
8e677ce8 | 542 | static inline void dbs_timer_exit(struct cpu_dbs_info_s *dbs_info) |
b9170836 | 543 | { |
8e677ce8 AC |
544 | dbs_info->enable = 0; |
545 | cancel_delayed_work(&dbs_info->work); | |
b9170836 DJ |
546 | } |
547 | ||
548 | static int cpufreq_governor_dbs(struct cpufreq_policy *policy, | |
549 | unsigned int event) | |
550 | { | |
551 | unsigned int cpu = policy->cpu; | |
552 | struct cpu_dbs_info_s *this_dbs_info; | |
553 | unsigned int j; | |
914f7c31 | 554 | int rc; |
b9170836 DJ |
555 | |
556 | this_dbs_info = &per_cpu(cpu_dbs_info, cpu); | |
557 | ||
558 | switch (event) { | |
559 | case CPUFREQ_GOV_START: | |
18a7247d | 560 | if ((!cpu_online(cpu)) || (!policy->cur)) |
b9170836 DJ |
561 | return -EINVAL; |
562 | ||
b9170836 DJ |
563 | if (this_dbs_info->enable) /* Already enabled */ |
564 | break; | |
18a7247d | 565 | |
3fc54d37 | 566 | mutex_lock(&dbs_mutex); |
914f7c31 JG |
567 | |
568 | rc = sysfs_create_group(&policy->kobj, &dbs_attr_group); | |
569 | if (rc) { | |
570 | mutex_unlock(&dbs_mutex); | |
571 | return rc; | |
572 | } | |
573 | ||
835481d9 | 574 | for_each_cpu(j, policy->cpus) { |
b9170836 DJ |
575 | struct cpu_dbs_info_s *j_dbs_info; |
576 | j_dbs_info = &per_cpu(cpu_dbs_info, j); | |
577 | j_dbs_info->cur_policy = policy; | |
18a7247d | 578 | |
8e677ce8 AC |
579 | j_dbs_info->prev_cpu_idle = get_cpu_idle_time(j, |
580 | &j_dbs_info->prev_cpu_wall); | |
581 | if (dbs_tuners_ins.ignore_nice) { | |
582 | j_dbs_info->prev_cpu_nice = | |
583 | kstat_cpu(j).cpustat.nice; | |
584 | } | |
b9170836 | 585 | } |
a159b827 AC |
586 | this_dbs_info->down_skip = 0; |
587 | this_dbs_info->requested_freq = policy->cur; | |
914f7c31 | 588 | |
b9170836 DJ |
589 | dbs_enable++; |
590 | /* | |
591 | * Start the timerschedule work, when this governor | |
592 | * is used for first time | |
593 | */ | |
594 | if (dbs_enable == 1) { | |
595 | unsigned int latency; | |
596 | /* policy latency is in nS. Convert it to uS first */ | |
2c906b31 AC |
597 | latency = policy->cpuinfo.transition_latency / 1000; |
598 | if (latency == 0) | |
599 | latency = 1; | |
b9170836 | 600 | |
112124ab | 601 | def_sampling_rate = |
a75603a0 | 602 | max(latency * LATENCY_MULTIPLIER, |
112124ab | 603 | MIN_STAT_SAMPLING_RATE); |
2c906b31 | 604 | |
b9170836 | 605 | dbs_tuners_ins.sampling_rate = def_sampling_rate; |
b9170836 | 606 | |
a8d7c3bc EO |
607 | cpufreq_register_notifier( |
608 | &dbs_cpufreq_notifier_block, | |
609 | CPUFREQ_TRANSITION_NOTIFIER); | |
b9170836 | 610 | } |
8e677ce8 | 611 | dbs_timer_init(this_dbs_info); |
18a7247d | 612 | |
3fc54d37 | 613 | mutex_unlock(&dbs_mutex); |
8e677ce8 | 614 | |
b9170836 DJ |
615 | break; |
616 | ||
617 | case CPUFREQ_GOV_STOP: | |
3fc54d37 | 618 | mutex_lock(&dbs_mutex); |
8e677ce8 | 619 | dbs_timer_exit(this_dbs_info); |
b9170836 DJ |
620 | sysfs_remove_group(&policy->kobj, &dbs_attr_group); |
621 | dbs_enable--; | |
8e677ce8 | 622 | |
b9170836 DJ |
623 | /* |
624 | * Stop the timerschedule work, when this governor | |
625 | * is used for first time | |
626 | */ | |
8e677ce8 | 627 | if (dbs_enable == 0) |
a8d7c3bc EO |
628 | cpufreq_unregister_notifier( |
629 | &dbs_cpufreq_notifier_block, | |
630 | CPUFREQ_TRANSITION_NOTIFIER); | |
a8d7c3bc | 631 | |
3fc54d37 | 632 | mutex_unlock(&dbs_mutex); |
b9170836 DJ |
633 | |
634 | break; | |
635 | ||
636 | case CPUFREQ_GOV_LIMITS: | |
3fc54d37 | 637 | mutex_lock(&dbs_mutex); |
b9170836 DJ |
638 | if (policy->max < this_dbs_info->cur_policy->cur) |
639 | __cpufreq_driver_target( | |
640 | this_dbs_info->cur_policy, | |
18a7247d | 641 | policy->max, CPUFREQ_RELATION_H); |
b9170836 DJ |
642 | else if (policy->min > this_dbs_info->cur_policy->cur) |
643 | __cpufreq_driver_target( | |
644 | this_dbs_info->cur_policy, | |
18a7247d | 645 | policy->min, CPUFREQ_RELATION_L); |
3fc54d37 | 646 | mutex_unlock(&dbs_mutex); |
8e677ce8 | 647 | |
b9170836 DJ |
648 | break; |
649 | } | |
650 | return 0; | |
651 | } | |
652 | ||
c4d14bc0 SW |
653 | #ifndef CONFIG_CPU_FREQ_DEFAULT_GOV_CONSERVATIVE |
654 | static | |
655 | #endif | |
1c256245 TR |
656 | struct cpufreq_governor cpufreq_gov_conservative = { |
657 | .name = "conservative", | |
658 | .governor = cpufreq_governor_dbs, | |
659 | .max_transition_latency = TRANSITION_LATENCY_LIMIT, | |
660 | .owner = THIS_MODULE, | |
b9170836 DJ |
661 | }; |
662 | ||
663 | static int __init cpufreq_gov_dbs_init(void) | |
664 | { | |
8e677ce8 AC |
665 | int err; |
666 | ||
667 | kconservative_wq = create_workqueue("kconservative"); | |
668 | if (!kconservative_wq) { | |
669 | printk(KERN_ERR "Creation of kconservative failed\n"); | |
670 | return -EFAULT; | |
671 | } | |
672 | ||
673 | err = cpufreq_register_governor(&cpufreq_gov_conservative); | |
674 | if (err) | |
675 | destroy_workqueue(kconservative_wq); | |
676 | ||
677 | return err; | |
b9170836 DJ |
678 | } |
679 | ||
680 | static void __exit cpufreq_gov_dbs_exit(void) | |
681 | { | |
1c256245 | 682 | cpufreq_unregister_governor(&cpufreq_gov_conservative); |
8e677ce8 | 683 | destroy_workqueue(kconservative_wq); |
b9170836 DJ |
684 | } |
685 | ||
686 | ||
11a80a9c | 687 | MODULE_AUTHOR("Alexander Clouter <alex@digriz.org.uk>"); |
9acef487 | 688 | MODULE_DESCRIPTION("'cpufreq_conservative' - A dynamic cpufreq governor for " |
b9170836 DJ |
689 | "Low Latency Frequency Transition capable processors " |
690 | "optimised for use in a battery environment"); | |
9acef487 | 691 | MODULE_LICENSE("GPL"); |
b9170836 | 692 | |
6915719b JW |
693 | #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_CONSERVATIVE |
694 | fs_initcall(cpufreq_gov_dbs_init); | |
695 | #else | |
b9170836 | 696 | module_init(cpufreq_gov_dbs_init); |
6915719b | 697 | #endif |
b9170836 | 698 | module_exit(cpufreq_gov_dbs_exit); |