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> |
1da177e4 LT |
21 | |
22 | /* | |
23 | * dbs is used in this file as a shortform for demandbased switching | |
24 | * It helps to keep variable names smaller, simpler | |
25 | */ | |
26 | ||
27 | #define DEF_FREQUENCY_UP_THRESHOLD (80) | |
c29f1403 | 28 | #define MIN_FREQUENCY_UP_THRESHOLD (11) |
1da177e4 LT |
29 | #define MAX_FREQUENCY_UP_THRESHOLD (100) |
30 | ||
32ee8c3e DJ |
31 | /* |
32 | * The polling frequency of this governor depends on the capability of | |
1da177e4 | 33 | * the processor. Default polling frequency is 1000 times the transition |
32ee8c3e DJ |
34 | * latency of the processor. The governor will work on any processor with |
35 | * transition latency <= 10mS, using appropriate sampling | |
1da177e4 LT |
36 | * rate. |
37 | * For CPUs with transition latency > 10mS (mostly drivers with CPUFREQ_ETERNAL) | |
38 | * this governor will not work. | |
39 | * All times here are in uS. | |
40 | */ | |
32ee8c3e | 41 | static unsigned int def_sampling_rate; |
df8b59be DJ |
42 | #define MIN_SAMPLING_RATE_RATIO (2) |
43 | /* for correct statistics, we need at least 10 ticks between each measure */ | |
e08f5f5b GS |
44 | #define MIN_STAT_SAMPLING_RATE \ |
45 | (MIN_SAMPLING_RATE_RATIO * jiffies_to_usecs(10)) | |
46 | #define MIN_SAMPLING_RATE \ | |
47 | (def_sampling_rate / MIN_SAMPLING_RATE_RATIO) | |
1da177e4 LT |
48 | #define MAX_SAMPLING_RATE (500 * def_sampling_rate) |
49 | #define DEF_SAMPLING_RATE_LATENCY_MULTIPLIER (1000) | |
1da177e4 | 50 | #define TRANSITION_LATENCY_LIMIT (10 * 1000) |
1da177e4 LT |
51 | |
52 | static void do_dbs_timer(void *data); | |
53 | ||
54 | struct cpu_dbs_info_s { | |
ccb2fe20 VP |
55 | cputime64_t prev_cpu_idle; |
56 | cputime64_t prev_cpu_wall; | |
32ee8c3e | 57 | struct cpufreq_policy *cur_policy; |
2f8a835c | 58 | struct work_struct work; |
32ee8c3e | 59 | unsigned int enable; |
05ca0350 AS |
60 | struct cpufreq_frequency_table *freq_table; |
61 | unsigned int freq_lo; | |
62 | unsigned int freq_lo_jiffies; | |
63 | unsigned int freq_hi_jiffies; | |
1da177e4 LT |
64 | }; |
65 | static DEFINE_PER_CPU(struct cpu_dbs_info_s, cpu_dbs_info); | |
66 | ||
67 | static unsigned int dbs_enable; /* number of CPUs using this policy */ | |
68 | ||
4ec223d0 VP |
69 | /* |
70 | * DEADLOCK ALERT! There is a ordering requirement between cpu_hotplug | |
71 | * lock and dbs_mutex. cpu_hotplug lock should always be held before | |
72 | * dbs_mutex. If any function that can potentially take cpu_hotplug lock | |
73 | * (like __cpufreq_driver_target()) is being called with dbs_mutex taken, then | |
74 | * cpu_hotplug lock should be taken before that. Note that cpu_hotplug lock | |
75 | * is recursive for the same process. -Venki | |
76 | */ | |
ffac80e9 | 77 | static DEFINE_MUTEX(dbs_mutex); |
1da177e4 | 78 | |
2f8a835c | 79 | static struct workqueue_struct *kondemand_wq; |
6810b548 | 80 | |
05ca0350 | 81 | static struct dbs_tuners { |
32ee8c3e | 82 | unsigned int sampling_rate; |
32ee8c3e DJ |
83 | unsigned int up_threshold; |
84 | unsigned int ignore_nice; | |
05ca0350 AS |
85 | unsigned int powersave_bias; |
86 | } dbs_tuners_ins = { | |
32ee8c3e | 87 | .up_threshold = DEF_FREQUENCY_UP_THRESHOLD, |
9cbad61b | 88 | .ignore_nice = 0, |
05ca0350 | 89 | .powersave_bias = 0, |
1da177e4 LT |
90 | }; |
91 | ||
ccb2fe20 | 92 | static inline cputime64_t get_cpu_idle_time(unsigned int cpu) |
dac1c1a5 | 93 | { |
ccb2fe20 VP |
94 | cputime64_t retval; |
95 | ||
96 | retval = cputime64_add(kstat_cpu(cpu).cpustat.idle, | |
97 | kstat_cpu(cpu).cpustat.iowait); | |
98 | ||
99 | if (dbs_tuners_ins.ignore_nice) | |
100 | retval = cputime64_add(retval, kstat_cpu(cpu).cpustat.nice); | |
101 | ||
102 | return retval; | |
dac1c1a5 DJ |
103 | } |
104 | ||
05ca0350 AS |
105 | /* |
106 | * Find right freq to be set now with powersave_bias on. | |
107 | * Returns the freq_hi to be used right now and will set freq_hi_jiffies, | |
108 | * freq_lo, and freq_lo_jiffies in percpu area for averaging freqs. | |
109 | */ | |
b5ecf60f AB |
110 | static unsigned int powersave_bias_target(struct cpufreq_policy *policy, |
111 | unsigned int freq_next, | |
112 | unsigned int relation) | |
05ca0350 AS |
113 | { |
114 | unsigned int freq_req, freq_reduc, freq_avg; | |
115 | unsigned int freq_hi, freq_lo; | |
116 | unsigned int index = 0; | |
117 | unsigned int jiffies_total, jiffies_hi, jiffies_lo; | |
118 | struct cpu_dbs_info_s *dbs_info = &per_cpu(cpu_dbs_info, policy->cpu); | |
119 | ||
120 | if (!dbs_info->freq_table) { | |
121 | dbs_info->freq_lo = 0; | |
122 | dbs_info->freq_lo_jiffies = 0; | |
123 | return freq_next; | |
124 | } | |
125 | ||
126 | cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_next, | |
127 | relation, &index); | |
128 | freq_req = dbs_info->freq_table[index].frequency; | |
129 | freq_reduc = freq_req * dbs_tuners_ins.powersave_bias / 1000; | |
130 | freq_avg = freq_req - freq_reduc; | |
131 | ||
132 | /* Find freq bounds for freq_avg in freq_table */ | |
133 | index = 0; | |
134 | cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_avg, | |
135 | CPUFREQ_RELATION_H, &index); | |
136 | freq_lo = dbs_info->freq_table[index].frequency; | |
137 | index = 0; | |
138 | cpufreq_frequency_table_target(policy, dbs_info->freq_table, freq_avg, | |
139 | CPUFREQ_RELATION_L, &index); | |
140 | freq_hi = dbs_info->freq_table[index].frequency; | |
141 | ||
142 | /* Find out how long we have to be in hi and lo freqs */ | |
143 | if (freq_hi == freq_lo) { | |
144 | dbs_info->freq_lo = 0; | |
145 | dbs_info->freq_lo_jiffies = 0; | |
146 | return freq_lo; | |
147 | } | |
148 | jiffies_total = usecs_to_jiffies(dbs_tuners_ins.sampling_rate); | |
149 | jiffies_hi = (freq_avg - freq_lo) * jiffies_total; | |
150 | jiffies_hi += ((freq_hi - freq_lo) / 2); | |
151 | jiffies_hi /= (freq_hi - freq_lo); | |
152 | jiffies_lo = jiffies_total - jiffies_hi; | |
153 | dbs_info->freq_lo = freq_lo; | |
154 | dbs_info->freq_lo_jiffies = jiffies_lo; | |
155 | dbs_info->freq_hi_jiffies = jiffies_hi; | |
156 | return freq_hi; | |
157 | } | |
158 | ||
159 | static void ondemand_powersave_bias_init(void) | |
160 | { | |
161 | int i; | |
162 | for_each_online_cpu(i) { | |
163 | struct cpu_dbs_info_s *dbs_info = &per_cpu(cpu_dbs_info, i); | |
164 | dbs_info->freq_table = cpufreq_frequency_get_table(i); | |
165 | dbs_info->freq_lo = 0; | |
166 | } | |
167 | } | |
168 | ||
1da177e4 LT |
169 | /************************** sysfs interface ************************/ |
170 | static ssize_t show_sampling_rate_max(struct cpufreq_policy *policy, char *buf) | |
171 | { | |
172 | return sprintf (buf, "%u\n", MAX_SAMPLING_RATE); | |
173 | } | |
174 | ||
175 | static ssize_t show_sampling_rate_min(struct cpufreq_policy *policy, char *buf) | |
176 | { | |
177 | return sprintf (buf, "%u\n", MIN_SAMPLING_RATE); | |
178 | } | |
179 | ||
32ee8c3e DJ |
180 | #define define_one_ro(_name) \ |
181 | static struct freq_attr _name = \ | |
1da177e4 LT |
182 | __ATTR(_name, 0444, show_##_name, NULL) |
183 | ||
184 | define_one_ro(sampling_rate_max); | |
185 | define_one_ro(sampling_rate_min); | |
186 | ||
187 | /* cpufreq_ondemand Governor Tunables */ | |
188 | #define show_one(file_name, object) \ | |
189 | static ssize_t show_##file_name \ | |
190 | (struct cpufreq_policy *unused, char *buf) \ | |
191 | { \ | |
192 | return sprintf(buf, "%u\n", dbs_tuners_ins.object); \ | |
193 | } | |
194 | show_one(sampling_rate, sampling_rate); | |
1da177e4 | 195 | show_one(up_threshold, up_threshold); |
001893cd | 196 | show_one(ignore_nice_load, ignore_nice); |
05ca0350 | 197 | show_one(powersave_bias, powersave_bias); |
1da177e4 | 198 | |
32ee8c3e | 199 | static ssize_t store_sampling_rate(struct cpufreq_policy *unused, |
1da177e4 LT |
200 | const char *buf, size_t count) |
201 | { | |
202 | unsigned int input; | |
203 | int ret; | |
ffac80e9 | 204 | ret = sscanf(buf, "%u", &input); |
1da177e4 | 205 | |
3fc54d37 | 206 | mutex_lock(&dbs_mutex); |
e08f5f5b GS |
207 | if (ret != 1 || input > MAX_SAMPLING_RATE |
208 | || input < MIN_SAMPLING_RATE) { | |
3fc54d37 | 209 | mutex_unlock(&dbs_mutex); |
1da177e4 LT |
210 | return -EINVAL; |
211 | } | |
212 | ||
213 | dbs_tuners_ins.sampling_rate = input; | |
3fc54d37 | 214 | mutex_unlock(&dbs_mutex); |
1da177e4 LT |
215 | |
216 | return count; | |
217 | } | |
218 | ||
32ee8c3e | 219 | static ssize_t store_up_threshold(struct cpufreq_policy *unused, |
1da177e4 LT |
220 | const char *buf, size_t count) |
221 | { | |
222 | unsigned int input; | |
223 | int ret; | |
ffac80e9 | 224 | ret = sscanf(buf, "%u", &input); |
1da177e4 | 225 | |
3fc54d37 | 226 | mutex_lock(&dbs_mutex); |
32ee8c3e | 227 | if (ret != 1 || input > MAX_FREQUENCY_UP_THRESHOLD || |
c29f1403 | 228 | input < MIN_FREQUENCY_UP_THRESHOLD) { |
3fc54d37 | 229 | mutex_unlock(&dbs_mutex); |
1da177e4 LT |
230 | return -EINVAL; |
231 | } | |
232 | ||
233 | dbs_tuners_ins.up_threshold = input; | |
3fc54d37 | 234 | mutex_unlock(&dbs_mutex); |
1da177e4 LT |
235 | |
236 | return count; | |
237 | } | |
238 | ||
001893cd | 239 | static ssize_t store_ignore_nice_load(struct cpufreq_policy *policy, |
3d5ee9e5 DJ |
240 | const char *buf, size_t count) |
241 | { | |
242 | unsigned int input; | |
243 | int ret; | |
244 | ||
245 | unsigned int j; | |
32ee8c3e | 246 | |
ffac80e9 | 247 | ret = sscanf(buf, "%u", &input); |
3d5ee9e5 DJ |
248 | if ( ret != 1 ) |
249 | return -EINVAL; | |
250 | ||
251 | if ( input > 1 ) | |
252 | input = 1; | |
32ee8c3e | 253 | |
3fc54d37 | 254 | mutex_lock(&dbs_mutex); |
3d5ee9e5 | 255 | if ( input == dbs_tuners_ins.ignore_nice ) { /* nothing to do */ |
3fc54d37 | 256 | mutex_unlock(&dbs_mutex); |
3d5ee9e5 DJ |
257 | return count; |
258 | } | |
259 | dbs_tuners_ins.ignore_nice = input; | |
260 | ||
ccb2fe20 | 261 | /* we need to re-evaluate prev_cpu_idle */ |
dac1c1a5 | 262 | for_each_online_cpu(j) { |
ccb2fe20 VP |
263 | struct cpu_dbs_info_s *dbs_info; |
264 | dbs_info = &per_cpu(cpu_dbs_info, j); | |
265 | dbs_info->prev_cpu_idle = get_cpu_idle_time(j); | |
266 | dbs_info->prev_cpu_wall = get_jiffies_64(); | |
3d5ee9e5 | 267 | } |
3fc54d37 | 268 | mutex_unlock(&dbs_mutex); |
3d5ee9e5 DJ |
269 | |
270 | return count; | |
271 | } | |
272 | ||
05ca0350 AS |
273 | static ssize_t store_powersave_bias(struct cpufreq_policy *unused, |
274 | const char *buf, size_t count) | |
275 | { | |
276 | unsigned int input; | |
277 | int ret; | |
278 | ret = sscanf(buf, "%u", &input); | |
279 | ||
280 | if (ret != 1) | |
281 | return -EINVAL; | |
282 | ||
283 | if (input > 1000) | |
284 | input = 1000; | |
285 | ||
286 | mutex_lock(&dbs_mutex); | |
287 | dbs_tuners_ins.powersave_bias = input; | |
288 | ondemand_powersave_bias_init(); | |
289 | mutex_unlock(&dbs_mutex); | |
290 | ||
291 | return count; | |
292 | } | |
293 | ||
1da177e4 LT |
294 | #define define_one_rw(_name) \ |
295 | static struct freq_attr _name = \ | |
296 | __ATTR(_name, 0644, show_##_name, store_##_name) | |
297 | ||
298 | define_one_rw(sampling_rate); | |
1da177e4 | 299 | define_one_rw(up_threshold); |
001893cd | 300 | define_one_rw(ignore_nice_load); |
05ca0350 | 301 | define_one_rw(powersave_bias); |
1da177e4 LT |
302 | |
303 | static struct attribute * dbs_attributes[] = { | |
304 | &sampling_rate_max.attr, | |
305 | &sampling_rate_min.attr, | |
306 | &sampling_rate.attr, | |
1da177e4 | 307 | &up_threshold.attr, |
001893cd | 308 | &ignore_nice_load.attr, |
05ca0350 | 309 | &powersave_bias.attr, |
1da177e4 LT |
310 | NULL |
311 | }; | |
312 | ||
313 | static struct attribute_group dbs_attr_group = { | |
314 | .attrs = dbs_attributes, | |
315 | .name = "ondemand", | |
316 | }; | |
317 | ||
318 | /************************** sysfs end ************************/ | |
319 | ||
2f8a835c | 320 | static void dbs_check_cpu(struct cpu_dbs_info_s *this_dbs_info) |
1da177e4 | 321 | { |
ccb2fe20 VP |
322 | unsigned int idle_ticks, total_ticks; |
323 | unsigned int load; | |
ccb2fe20 | 324 | cputime64_t cur_jiffies; |
1da177e4 LT |
325 | |
326 | struct cpufreq_policy *policy; | |
327 | unsigned int j; | |
328 | ||
1da177e4 LT |
329 | if (!this_dbs_info->enable) |
330 | return; | |
331 | ||
05ca0350 | 332 | this_dbs_info->freq_lo = 0; |
1da177e4 | 333 | policy = this_dbs_info->cur_policy; |
ccb2fe20 VP |
334 | cur_jiffies = jiffies64_to_cputime64(get_jiffies_64()); |
335 | total_ticks = (unsigned int) cputime64_sub(cur_jiffies, | |
336 | this_dbs_info->prev_cpu_wall); | |
337 | this_dbs_info->prev_cpu_wall = cur_jiffies; | |
2cd7cbdf LT |
338 | if (!total_ticks) |
339 | return; | |
32ee8c3e | 340 | /* |
c29f1403 DJ |
341 | * Every sampling_rate, we check, if current idle time is less |
342 | * than 20% (default), then we try to increase frequency | |
ccb2fe20 | 343 | * Every sampling_rate, we look for a the lowest |
c29f1403 DJ |
344 | * frequency which can sustain the load while keeping idle time over |
345 | * 30%. If such a frequency exist, we try to decrease to this frequency. | |
1da177e4 | 346 | * |
32ee8c3e DJ |
347 | * Any frequency increase takes it to the maximum frequency. |
348 | * Frequency reduction happens at minimum steps of | |
349 | * 5% (default) of current frequency | |
1da177e4 LT |
350 | */ |
351 | ||
ccb2fe20 | 352 | /* Get Idle Time */ |
9c7d269b | 353 | idle_ticks = UINT_MAX; |
1da177e4 | 354 | for_each_cpu_mask(j, policy->cpus) { |
ccb2fe20 VP |
355 | cputime64_t total_idle_ticks; |
356 | unsigned int tmp_idle_ticks; | |
1da177e4 LT |
357 | struct cpu_dbs_info_s *j_dbs_info; |
358 | ||
1da177e4 | 359 | j_dbs_info = &per_cpu(cpu_dbs_info, j); |
dac1c1a5 | 360 | total_idle_ticks = get_cpu_idle_time(j); |
ccb2fe20 VP |
361 | tmp_idle_ticks = (unsigned int) cputime64_sub(total_idle_ticks, |
362 | j_dbs_info->prev_cpu_idle); | |
363 | j_dbs_info->prev_cpu_idle = total_idle_ticks; | |
1da177e4 LT |
364 | |
365 | if (tmp_idle_ticks < idle_ticks) | |
366 | idle_ticks = tmp_idle_ticks; | |
367 | } | |
ccb2fe20 | 368 | load = (100 * (total_ticks - idle_ticks)) / total_ticks; |
1da177e4 | 369 | |
ccb2fe20 VP |
370 | /* Check for frequency increase */ |
371 | if (load > dbs_tuners_ins.up_threshold) { | |
c11420a6 | 372 | /* if we are already at full speed then break out early */ |
05ca0350 AS |
373 | if (!dbs_tuners_ins.powersave_bias) { |
374 | if (policy->cur == policy->max) | |
375 | return; | |
376 | ||
377 | __cpufreq_driver_target(policy, policy->max, | |
378 | CPUFREQ_RELATION_H); | |
379 | } else { | |
380 | int freq = powersave_bias_target(policy, policy->max, | |
381 | CPUFREQ_RELATION_H); | |
382 | __cpufreq_driver_target(policy, freq, | |
383 | CPUFREQ_RELATION_L); | |
384 | } | |
1da177e4 LT |
385 | return; |
386 | } | |
387 | ||
388 | /* Check for frequency decrease */ | |
c29f1403 DJ |
389 | /* if we cannot reduce the frequency anymore, break out early */ |
390 | if (policy->cur == policy->min) | |
391 | return; | |
1da177e4 | 392 | |
c29f1403 DJ |
393 | /* |
394 | * The optimal frequency is the frequency that is the lowest that | |
395 | * can support the current CPU usage without triggering the up | |
396 | * policy. To be safe, we focus 10 points under the threshold. | |
397 | */ | |
ccb2fe20 | 398 | if (load < (dbs_tuners_ins.up_threshold - 10)) { |
dfde5d62 VP |
399 | unsigned int freq_next, freq_cur; |
400 | ||
401 | freq_cur = cpufreq_driver_getavg(policy); | |
402 | if (!freq_cur) | |
403 | freq_cur = policy->cur; | |
404 | ||
405 | freq_next = (freq_cur * load) / | |
c29f1403 | 406 | (dbs_tuners_ins.up_threshold - 10); |
dfde5d62 | 407 | |
05ca0350 AS |
408 | if (!dbs_tuners_ins.powersave_bias) { |
409 | __cpufreq_driver_target(policy, freq_next, | |
410 | CPUFREQ_RELATION_L); | |
411 | } else { | |
412 | int freq = powersave_bias_target(policy, freq_next, | |
413 | CPUFREQ_RELATION_L); | |
414 | __cpufreq_driver_target(policy, freq, | |
415 | CPUFREQ_RELATION_L); | |
416 | } | |
ccb2fe20 | 417 | } |
1da177e4 LT |
418 | } |
419 | ||
05ca0350 AS |
420 | /* Sampling types */ |
421 | enum {DBS_NORMAL_SAMPLE, DBS_SUB_SAMPLE}; | |
422 | ||
1da177e4 | 423 | static void do_dbs_timer(void *data) |
32ee8c3e | 424 | { |
2f8a835c VP |
425 | unsigned int cpu = smp_processor_id(); |
426 | struct cpu_dbs_info_s *dbs_info = &per_cpu(cpu_dbs_info, cpu); | |
1ce28d6b AS |
427 | /* We want all CPUs to do sampling nearly on same jiffy */ |
428 | int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate); | |
429 | delay -= jiffies % delay; | |
2f8a835c | 430 | |
2cd7cbdf LT |
431 | if (!dbs_info->enable) |
432 | return; | |
05ca0350 AS |
433 | /* Common NORMAL_SAMPLE setup */ |
434 | INIT_WORK(&dbs_info->work, do_dbs_timer, (void *)DBS_NORMAL_SAMPLE); | |
435 | if (!dbs_tuners_ins.powersave_bias || | |
436 | (unsigned long) data == DBS_NORMAL_SAMPLE) { | |
437 | lock_cpu_hotplug(); | |
438 | dbs_check_cpu(dbs_info); | |
439 | unlock_cpu_hotplug(); | |
440 | if (dbs_info->freq_lo) { | |
441 | /* Setup timer for SUB_SAMPLE */ | |
442 | INIT_WORK(&dbs_info->work, do_dbs_timer, | |
443 | (void *)DBS_SUB_SAMPLE); | |
444 | delay = dbs_info->freq_hi_jiffies; | |
445 | } | |
446 | } else { | |
447 | __cpufreq_driver_target(dbs_info->cur_policy, | |
448 | dbs_info->freq_lo, | |
449 | CPUFREQ_RELATION_H); | |
450 | } | |
1ce28d6b | 451 | queue_delayed_work_on(cpu, kondemand_wq, &dbs_info->work, delay); |
32ee8c3e | 452 | } |
1da177e4 | 453 | |
2f8a835c | 454 | static inline void dbs_timer_init(unsigned int cpu) |
1da177e4 | 455 | { |
2f8a835c | 456 | struct cpu_dbs_info_s *dbs_info = &per_cpu(cpu_dbs_info, cpu); |
1ce28d6b AS |
457 | /* We want all CPUs to do sampling nearly on same jiffy */ |
458 | int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate); | |
459 | delay -= jiffies % delay; | |
2f8a835c | 460 | |
05ca0350 | 461 | ondemand_powersave_bias_init(); |
3906f4ed | 462 | INIT_WORK(&dbs_info->work, do_dbs_timer, NULL); |
1ce28d6b | 463 | queue_delayed_work_on(cpu, kondemand_wq, &dbs_info->work, delay); |
1da177e4 LT |
464 | } |
465 | ||
2cd7cbdf | 466 | static inline void dbs_timer_exit(struct cpu_dbs_info_s *dbs_info) |
1da177e4 | 467 | { |
2cd7cbdf LT |
468 | dbs_info->enable = 0; |
469 | cancel_delayed_work(&dbs_info->work); | |
470 | flush_workqueue(kondemand_wq); | |
1da177e4 LT |
471 | } |
472 | ||
473 | static int cpufreq_governor_dbs(struct cpufreq_policy *policy, | |
474 | unsigned int event) | |
475 | { | |
476 | unsigned int cpu = policy->cpu; | |
477 | struct cpu_dbs_info_s *this_dbs_info; | |
478 | unsigned int j; | |
914f7c31 | 479 | int rc; |
1da177e4 LT |
480 | |
481 | this_dbs_info = &per_cpu(cpu_dbs_info, cpu); | |
482 | ||
483 | switch (event) { | |
484 | case CPUFREQ_GOV_START: | |
ffac80e9 | 485 | if ((!cpu_online(cpu)) || (!policy->cur)) |
1da177e4 LT |
486 | return -EINVAL; |
487 | ||
488 | if (policy->cpuinfo.transition_latency > | |
ff8c288d EP |
489 | (TRANSITION_LATENCY_LIMIT * 1000)) { |
490 | printk(KERN_WARNING "ondemand governor failed to load " | |
491 | "due to too long transition latency\n"); | |
1da177e4 | 492 | return -EINVAL; |
ff8c288d | 493 | } |
1da177e4 LT |
494 | if (this_dbs_info->enable) /* Already enabled */ |
495 | break; | |
32ee8c3e | 496 | |
3fc54d37 | 497 | mutex_lock(&dbs_mutex); |
2f8a835c VP |
498 | dbs_enable++; |
499 | if (dbs_enable == 1) { | |
500 | kondemand_wq = create_workqueue("kondemand"); | |
501 | if (!kondemand_wq) { | |
e08f5f5b GS |
502 | printk(KERN_ERR |
503 | "Creation of kondemand failed\n"); | |
2f8a835c VP |
504 | dbs_enable--; |
505 | mutex_unlock(&dbs_mutex); | |
506 | return -ENOSPC; | |
507 | } | |
508 | } | |
914f7c31 JG |
509 | |
510 | rc = sysfs_create_group(&policy->kobj, &dbs_attr_group); | |
511 | if (rc) { | |
512 | if (dbs_enable == 1) | |
513 | destroy_workqueue(kondemand_wq); | |
514 | dbs_enable--; | |
515 | mutex_unlock(&dbs_mutex); | |
516 | return rc; | |
517 | } | |
518 | ||
1da177e4 LT |
519 | for_each_cpu_mask(j, policy->cpus) { |
520 | struct cpu_dbs_info_s *j_dbs_info; | |
521 | j_dbs_info = &per_cpu(cpu_dbs_info, j); | |
522 | j_dbs_info->cur_policy = policy; | |
32ee8c3e | 523 | |
ccb2fe20 VP |
524 | j_dbs_info->prev_cpu_idle = get_cpu_idle_time(j); |
525 | j_dbs_info->prev_cpu_wall = get_jiffies_64(); | |
1da177e4 LT |
526 | } |
527 | this_dbs_info->enable = 1; | |
1da177e4 LT |
528 | /* |
529 | * Start the timerschedule work, when this governor | |
530 | * is used for first time | |
531 | */ | |
532 | if (dbs_enable == 1) { | |
533 | unsigned int latency; | |
534 | /* policy latency is in nS. Convert it to uS first */ | |
df8b59be DJ |
535 | latency = policy->cpuinfo.transition_latency / 1000; |
536 | if (latency == 0) | |
537 | latency = 1; | |
1da177e4 | 538 | |
df8b59be | 539 | def_sampling_rate = latency * |
1da177e4 | 540 | DEF_SAMPLING_RATE_LATENCY_MULTIPLIER; |
df8b59be DJ |
541 | |
542 | if (def_sampling_rate < MIN_STAT_SAMPLING_RATE) | |
543 | def_sampling_rate = MIN_STAT_SAMPLING_RATE; | |
544 | ||
1da177e4 | 545 | dbs_tuners_ins.sampling_rate = def_sampling_rate; |
1da177e4 | 546 | } |
2f8a835c | 547 | dbs_timer_init(policy->cpu); |
32ee8c3e | 548 | |
3fc54d37 | 549 | mutex_unlock(&dbs_mutex); |
1da177e4 LT |
550 | break; |
551 | ||
552 | case CPUFREQ_GOV_STOP: | |
3fc54d37 | 553 | mutex_lock(&dbs_mutex); |
2cd7cbdf | 554 | dbs_timer_exit(this_dbs_info); |
1da177e4 LT |
555 | sysfs_remove_group(&policy->kobj, &dbs_attr_group); |
556 | dbs_enable--; | |
32ee8c3e | 557 | if (dbs_enable == 0) |
2f8a835c | 558 | destroy_workqueue(kondemand_wq); |
32ee8c3e | 559 | |
3fc54d37 | 560 | mutex_unlock(&dbs_mutex); |
1da177e4 LT |
561 | |
562 | break; | |
563 | ||
564 | case CPUFREQ_GOV_LIMITS: | |
3fc54d37 | 565 | mutex_lock(&dbs_mutex); |
1da177e4 | 566 | if (policy->max < this_dbs_info->cur_policy->cur) |
ffac80e9 VP |
567 | __cpufreq_driver_target(this_dbs_info->cur_policy, |
568 | policy->max, | |
569 | CPUFREQ_RELATION_H); | |
1da177e4 | 570 | else if (policy->min > this_dbs_info->cur_policy->cur) |
ffac80e9 VP |
571 | __cpufreq_driver_target(this_dbs_info->cur_policy, |
572 | policy->min, | |
573 | CPUFREQ_RELATION_L); | |
3fc54d37 | 574 | mutex_unlock(&dbs_mutex); |
1da177e4 LT |
575 | break; |
576 | } | |
577 | return 0; | |
578 | } | |
579 | ||
7f335d4e | 580 | static struct cpufreq_governor cpufreq_gov_dbs = { |
ffac80e9 VP |
581 | .name = "ondemand", |
582 | .governor = cpufreq_governor_dbs, | |
583 | .owner = THIS_MODULE, | |
1da177e4 | 584 | }; |
1da177e4 LT |
585 | |
586 | static int __init cpufreq_gov_dbs_init(void) | |
587 | { | |
588 | return cpufreq_register_governor(&cpufreq_gov_dbs); | |
589 | } | |
590 | ||
591 | static void __exit cpufreq_gov_dbs_exit(void) | |
592 | { | |
1da177e4 LT |
593 | cpufreq_unregister_governor(&cpufreq_gov_dbs); |
594 | } | |
595 | ||
596 | ||
ffac80e9 VP |
597 | MODULE_AUTHOR("Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>"); |
598 | MODULE_AUTHOR("Alexey Starikovskiy <alexey.y.starikovskiy@intel.com>"); | |
599 | MODULE_DESCRIPTION("'cpufreq_ondemand' - A dynamic cpufreq governor for " | |
600 | "Low Latency Frequency Transition capable processors"); | |
601 | MODULE_LICENSE("GPL"); | |
1da177e4 LT |
602 | |
603 | module_init(cpufreq_gov_dbs_init); | |
604 | module_exit(cpufreq_gov_dbs_exit); |