diff options
Diffstat (limited to 'drivers/cpufreq/cpufreq_conservative.c')
-rw-r--r-- | drivers/cpufreq/cpufreq_conservative.c | 404 |
1 files changed, 244 insertions, 160 deletions
diff --git a/drivers/cpufreq/cpufreq_conservative.c b/drivers/cpufreq/cpufreq_conservative.c index 0320962c4ec5..2ecd95e4ab1a 100644 --- a/drivers/cpufreq/cpufreq_conservative.c +++ b/drivers/cpufreq/cpufreq_conservative.c | |||
@@ -4,7 +4,7 @@ | |||
4 | * Copyright (C) 2001 Russell King | 4 | * Copyright (C) 2001 Russell King |
5 | * (C) 2003 Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>. | 5 | * (C) 2003 Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>. |
6 | * Jun Nakajima <jun.nakajima@intel.com> | 6 | * Jun Nakajima <jun.nakajima@intel.com> |
7 | * (C) 2004 Alexander Clouter <alex-kernel@digriz.org.uk> | 7 | * (C) 2009 Alexander Clouter <alex@digriz.org.uk> |
8 | * | 8 | * |
9 | * This program is free software; you can redistribute it and/or modify | 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 | 10 | * it under the terms of the GNU General Public License version 2 as |
@@ -13,22 +13,17 @@ | |||
13 | 13 | ||
14 | #include <linux/kernel.h> | 14 | #include <linux/kernel.h> |
15 | #include <linux/module.h> | 15 | #include <linux/module.h> |
16 | #include <linux/smp.h> | ||
17 | #include <linux/init.h> | 16 | #include <linux/init.h> |
18 | #include <linux/interrupt.h> | ||
19 | #include <linux/ctype.h> | ||
20 | #include <linux/cpufreq.h> | 17 | #include <linux/cpufreq.h> |
21 | #include <linux/sysctl.h> | ||
22 | #include <linux/types.h> | ||
23 | #include <linux/fs.h> | ||
24 | #include <linux/sysfs.h> | ||
25 | #include <linux/cpu.h> | 18 | #include <linux/cpu.h> |
26 | #include <linux/kmod.h> | ||
27 | #include <linux/workqueue.h> | ||
28 | #include <linux/jiffies.h> | 19 | #include <linux/jiffies.h> |
29 | #include <linux/kernel_stat.h> | 20 | #include <linux/kernel_stat.h> |
30 | #include <linux/percpu.h> | ||
31 | #include <linux/mutex.h> | 21 | #include <linux/mutex.h> |
22 | #include <linux/hrtimer.h> | ||
23 | #include <linux/tick.h> | ||
24 | #include <linux/ktime.h> | ||
25 | #include <linux/sched.h> | ||
26 | |||
32 | /* | 27 | /* |
33 | * dbs is used in this file as a shortform for demandbased switching | 28 | * dbs is used in this file as a shortform for demandbased switching |
34 | * It helps to keep variable names smaller, simpler | 29 | * It helps to keep variable names smaller, simpler |
@@ -43,19 +38,31 @@ | |||
43 | * latency of the processor. The governor will work on any processor with | 38 | * latency of the processor. The governor will work on any processor with |
44 | * transition latency <= 10mS, using appropriate sampling | 39 | * transition latency <= 10mS, using appropriate sampling |
45 | * rate. | 40 | * rate. |
46 | * For CPUs with transition latency > 10mS (mostly drivers | 41 | * For CPUs with transition latency > 10mS (mostly drivers with CPUFREQ_ETERNAL) |
47 | * with CPUFREQ_ETERNAL), this governor will not work. | 42 | * this governor will not work. |
48 | * All times here are in uS. | 43 | * All times here are in uS. |
49 | */ | 44 | */ |
50 | static unsigned int def_sampling_rate; | 45 | static unsigned int def_sampling_rate; |
51 | #define MIN_SAMPLING_RATE_RATIO (2) | 46 | #define MIN_SAMPLING_RATE_RATIO (2) |
52 | /* for correct statistics, we need at least 10 ticks between each measure */ | 47 | /* for correct statistics, we need at least 10 ticks between each measure */ |
53 | #define MIN_STAT_SAMPLING_RATE \ | 48 | #define MIN_STAT_SAMPLING_RATE \ |
54 | (MIN_SAMPLING_RATE_RATIO * jiffies_to_usecs(10)) | 49 | (MIN_SAMPLING_RATE_RATIO * jiffies_to_usecs(10)) |
55 | #define MIN_SAMPLING_RATE \ | 50 | #define MIN_SAMPLING_RATE \ |
56 | (def_sampling_rate / MIN_SAMPLING_RATE_RATIO) | 51 | (def_sampling_rate / MIN_SAMPLING_RATE_RATIO) |
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 */ | ||
57 | #define MAX_SAMPLING_RATE (500 * def_sampling_rate) | 64 | #define MAX_SAMPLING_RATE (500 * def_sampling_rate) |
58 | #define DEF_SAMPLING_RATE_LATENCY_MULTIPLIER (1000) | 65 | #define LATENCY_MULTIPLIER (1000) |
59 | #define DEF_SAMPLING_DOWN_FACTOR (1) | 66 | #define DEF_SAMPLING_DOWN_FACTOR (1) |
60 | #define MAX_SAMPLING_DOWN_FACTOR (10) | 67 | #define MAX_SAMPLING_DOWN_FACTOR (10) |
61 | #define TRANSITION_LATENCY_LIMIT (10 * 1000 * 1000) | 68 | #define TRANSITION_LATENCY_LIMIT (10 * 1000 * 1000) |
@@ -63,12 +70,15 @@ static unsigned int def_sampling_rate; | |||
63 | static void do_dbs_timer(struct work_struct *work); | 70 | static void do_dbs_timer(struct work_struct *work); |
64 | 71 | ||
65 | struct cpu_dbs_info_s { | 72 | struct cpu_dbs_info_s { |
73 | cputime64_t prev_cpu_idle; | ||
74 | cputime64_t prev_cpu_wall; | ||
75 | cputime64_t prev_cpu_nice; | ||
66 | struct cpufreq_policy *cur_policy; | 76 | struct cpufreq_policy *cur_policy; |
67 | unsigned int prev_cpu_idle_up; | 77 | struct delayed_work work; |
68 | unsigned int prev_cpu_idle_down; | ||
69 | unsigned int enable; | ||
70 | unsigned int down_skip; | 78 | unsigned int down_skip; |
71 | unsigned int requested_freq; | 79 | unsigned int requested_freq; |
80 | int cpu; | ||
81 | unsigned int enable:1; | ||
72 | }; | 82 | }; |
73 | static DEFINE_PER_CPU(struct cpu_dbs_info_s, cpu_dbs_info); | 83 | static DEFINE_PER_CPU(struct cpu_dbs_info_s, cpu_dbs_info); |
74 | 84 | ||
@@ -82,19 +92,18 @@ static unsigned int dbs_enable; /* number of CPUs using this policy */ | |||
82 | * cpu_hotplug lock should be taken before that. Note that cpu_hotplug lock | 92 | * cpu_hotplug lock should be taken before that. Note that cpu_hotplug lock |
83 | * is recursive for the same process. -Venki | 93 | * is recursive for the same process. -Venki |
84 | */ | 94 | */ |
85 | static DEFINE_MUTEX (dbs_mutex); | 95 | static DEFINE_MUTEX(dbs_mutex); |
86 | static DECLARE_DELAYED_WORK(dbs_work, do_dbs_timer); | ||
87 | 96 | ||
88 | struct dbs_tuners { | 97 | static struct workqueue_struct *kconservative_wq; |
98 | |||
99 | static struct dbs_tuners { | ||
89 | unsigned int sampling_rate; | 100 | unsigned int sampling_rate; |
90 | unsigned int sampling_down_factor; | 101 | unsigned int sampling_down_factor; |
91 | unsigned int up_threshold; | 102 | unsigned int up_threshold; |
92 | unsigned int down_threshold; | 103 | unsigned int down_threshold; |
93 | unsigned int ignore_nice; | 104 | unsigned int ignore_nice; |
94 | unsigned int freq_step; | 105 | unsigned int freq_step; |
95 | }; | 106 | } dbs_tuners_ins = { |
96 | |||
97 | static struct dbs_tuners dbs_tuners_ins = { | ||
98 | .up_threshold = DEF_FREQUENCY_UP_THRESHOLD, | 107 | .up_threshold = DEF_FREQUENCY_UP_THRESHOLD, |
99 | .down_threshold = DEF_FREQUENCY_DOWN_THRESHOLD, | 108 | .down_threshold = DEF_FREQUENCY_DOWN_THRESHOLD, |
100 | .sampling_down_factor = DEF_SAMPLING_DOWN_FACTOR, | 109 | .sampling_down_factor = DEF_SAMPLING_DOWN_FACTOR, |
@@ -102,18 +111,37 @@ static struct dbs_tuners dbs_tuners_ins = { | |||
102 | .freq_step = 5, | 111 | .freq_step = 5, |
103 | }; | 112 | }; |
104 | 113 | ||
105 | static inline unsigned int get_cpu_idle_time(unsigned int cpu) | 114 | static inline cputime64_t get_cpu_idle_time_jiffy(unsigned int cpu, |
115 | cputime64_t *wall) | ||
106 | { | 116 | { |
107 | unsigned int add_nice = 0, ret; | 117 | cputime64_t idle_time; |
118 | cputime64_t cur_wall_time; | ||
119 | cputime64_t busy_time; | ||
108 | 120 | ||
109 | if (dbs_tuners_ins.ignore_nice) | 121 | cur_wall_time = jiffies64_to_cputime64(get_jiffies_64()); |
110 | add_nice = kstat_cpu(cpu).cpustat.nice; | 122 | busy_time = cputime64_add(kstat_cpu(cpu).cpustat.user, |
123 | kstat_cpu(cpu).cpustat.system); | ||
124 | |||
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); | ||
129 | |||
130 | idle_time = cputime64_sub(cur_wall_time, busy_time); | ||
131 | if (wall) | ||
132 | *wall = cur_wall_time; | ||
133 | |||
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); | ||
111 | 140 | ||
112 | ret = kstat_cpu(cpu).cpustat.idle + | 141 | if (idle_time == -1ULL) |
113 | kstat_cpu(cpu).cpustat.iowait + | 142 | return get_cpu_idle_time_jiffy(cpu, wall); |
114 | add_nice; | ||
115 | 143 | ||
116 | return ret; | 144 | return idle_time; |
117 | } | 145 | } |
118 | 146 | ||
119 | /* keep track of frequency transitions */ | 147 | /* keep track of frequency transitions */ |
@@ -125,10 +153,21 @@ dbs_cpufreq_notifier(struct notifier_block *nb, unsigned long val, | |||
125 | struct cpu_dbs_info_s *this_dbs_info = &per_cpu(cpu_dbs_info, | 153 | struct cpu_dbs_info_s *this_dbs_info = &per_cpu(cpu_dbs_info, |
126 | freq->cpu); | 154 | freq->cpu); |
127 | 155 | ||
156 | struct cpufreq_policy *policy; | ||
157 | |||
128 | if (!this_dbs_info->enable) | 158 | if (!this_dbs_info->enable) |
129 | return 0; | 159 | return 0; |
130 | 160 | ||
131 | this_dbs_info->requested_freq = freq->new; | 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; | ||
132 | 171 | ||
133 | return 0; | 172 | return 0; |
134 | } | 173 | } |
@@ -140,16 +179,31 @@ static struct notifier_block dbs_cpufreq_notifier_block = { | |||
140 | /************************** sysfs interface ************************/ | 179 | /************************** sysfs interface ************************/ |
141 | static ssize_t show_sampling_rate_max(struct cpufreq_policy *policy, char *buf) | 180 | static ssize_t show_sampling_rate_max(struct cpufreq_policy *policy, char *buf) |
142 | { | 181 | { |
143 | return sprintf (buf, "%u\n", MAX_SAMPLING_RATE); | 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 | } | ||
190 | return sprintf(buf, "%u\n", MAX_SAMPLING_RATE); | ||
144 | } | 191 | } |
145 | 192 | ||
146 | static ssize_t show_sampling_rate_min(struct cpufreq_policy *policy, char *buf) | 193 | static ssize_t show_sampling_rate_min(struct cpufreq_policy *policy, char *buf) |
147 | { | 194 | { |
148 | return sprintf (buf, "%u\n", MIN_SAMPLING_RATE); | 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 | } | ||
202 | return sprintf(buf, "%u\n", MIN_SAMPLING_RATE); | ||
149 | } | 203 | } |
150 | 204 | ||
151 | #define define_one_ro(_name) \ | 205 | #define define_one_ro(_name) \ |
152 | static struct freq_attr _name = \ | 206 | static struct freq_attr _name = \ |
153 | __ATTR(_name, 0444, show_##_name, NULL) | 207 | __ATTR(_name, 0444, show_##_name, NULL) |
154 | 208 | ||
155 | define_one_ro(sampling_rate_max); | 209 | define_one_ro(sampling_rate_max); |
@@ -174,7 +228,8 @@ static ssize_t store_sampling_down_factor(struct cpufreq_policy *unused, | |||
174 | { | 228 | { |
175 | unsigned int input; | 229 | unsigned int input; |
176 | int ret; | 230 | int ret; |
177 | ret = sscanf (buf, "%u", &input); | 231 | ret = sscanf(buf, "%u", &input); |
232 | |||
178 | if (ret != 1 || input > MAX_SAMPLING_DOWN_FACTOR || input < 1) | 233 | if (ret != 1 || input > MAX_SAMPLING_DOWN_FACTOR || input < 1) |
179 | return -EINVAL; | 234 | return -EINVAL; |
180 | 235 | ||
@@ -190,15 +245,13 @@ static ssize_t store_sampling_rate(struct cpufreq_policy *unused, | |||
190 | { | 245 | { |
191 | unsigned int input; | 246 | unsigned int input; |
192 | int ret; | 247 | int ret; |
193 | ret = sscanf (buf, "%u", &input); | 248 | ret = sscanf(buf, "%u", &input); |
194 | 249 | ||
195 | mutex_lock(&dbs_mutex); | 250 | if (ret != 1) |
196 | if (ret != 1 || input > MAX_SAMPLING_RATE || input < MIN_SAMPLING_RATE) { | ||
197 | mutex_unlock(&dbs_mutex); | ||
198 | return -EINVAL; | 251 | return -EINVAL; |
199 | } | ||
200 | 252 | ||
201 | dbs_tuners_ins.sampling_rate = input; | 253 | mutex_lock(&dbs_mutex); |
254 | dbs_tuners_ins.sampling_rate = max(input, minimum_sampling_rate()); | ||
202 | mutex_unlock(&dbs_mutex); | 255 | mutex_unlock(&dbs_mutex); |
203 | 256 | ||
204 | return count; | 257 | return count; |
@@ -209,10 +262,11 @@ static ssize_t store_up_threshold(struct cpufreq_policy *unused, | |||
209 | { | 262 | { |
210 | unsigned int input; | 263 | unsigned int input; |
211 | int ret; | 264 | int ret; |
212 | ret = sscanf (buf, "%u", &input); | 265 | ret = sscanf(buf, "%u", &input); |
213 | 266 | ||
214 | mutex_lock(&dbs_mutex); | 267 | mutex_lock(&dbs_mutex); |
215 | if (ret != 1 || input > 100 || input <= dbs_tuners_ins.down_threshold) { | 268 | if (ret != 1 || input > 100 || |
269 | input <= dbs_tuners_ins.down_threshold) { | ||
216 | mutex_unlock(&dbs_mutex); | 270 | mutex_unlock(&dbs_mutex); |
217 | return -EINVAL; | 271 | return -EINVAL; |
218 | } | 272 | } |
@@ -228,10 +282,12 @@ static ssize_t store_down_threshold(struct cpufreq_policy *unused, | |||
228 | { | 282 | { |
229 | unsigned int input; | 283 | unsigned int input; |
230 | int ret; | 284 | int ret; |
231 | ret = sscanf (buf, "%u", &input); | 285 | ret = sscanf(buf, "%u", &input); |
232 | 286 | ||
233 | mutex_lock(&dbs_mutex); | 287 | mutex_lock(&dbs_mutex); |
234 | if (ret != 1 || input > 100 || input >= dbs_tuners_ins.up_threshold) { | 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) { | ||
235 | mutex_unlock(&dbs_mutex); | 291 | mutex_unlock(&dbs_mutex); |
236 | return -EINVAL; | 292 | return -EINVAL; |
237 | } | 293 | } |
@@ -264,12 +320,14 @@ static ssize_t store_ignore_nice_load(struct cpufreq_policy *policy, | |||
264 | } | 320 | } |
265 | dbs_tuners_ins.ignore_nice = input; | 321 | dbs_tuners_ins.ignore_nice = input; |
266 | 322 | ||
267 | /* we need to re-evaluate prev_cpu_idle_up and prev_cpu_idle_down */ | 323 | /* we need to re-evaluate prev_cpu_idle */ |
268 | for_each_online_cpu(j) { | 324 | for_each_online_cpu(j) { |
269 | struct cpu_dbs_info_s *j_dbs_info; | 325 | struct cpu_dbs_info_s *dbs_info; |
270 | j_dbs_info = &per_cpu(cpu_dbs_info, j); | 326 | dbs_info = &per_cpu(cpu_dbs_info, j); |
271 | j_dbs_info->prev_cpu_idle_up = get_cpu_idle_time(j); | 327 | dbs_info->prev_cpu_idle = get_cpu_idle_time(j, |
272 | j_dbs_info->prev_cpu_idle_down = j_dbs_info->prev_cpu_idle_up; | 328 | &dbs_info->prev_cpu_wall); |
329 | if (dbs_tuners_ins.ignore_nice) | ||
330 | dbs_info->prev_cpu_nice = kstat_cpu(j).cpustat.nice; | ||
273 | } | 331 | } |
274 | mutex_unlock(&dbs_mutex); | 332 | mutex_unlock(&dbs_mutex); |
275 | 333 | ||
@@ -281,7 +339,6 @@ static ssize_t store_freq_step(struct cpufreq_policy *policy, | |||
281 | { | 339 | { |
282 | unsigned int input; | 340 | unsigned int input; |
283 | int ret; | 341 | int ret; |
284 | |||
285 | ret = sscanf(buf, "%u", &input); | 342 | ret = sscanf(buf, "%u", &input); |
286 | 343 | ||
287 | if (ret != 1) | 344 | if (ret != 1) |
@@ -310,7 +367,7 @@ define_one_rw(down_threshold); | |||
310 | define_one_rw(ignore_nice_load); | 367 | define_one_rw(ignore_nice_load); |
311 | define_one_rw(freq_step); | 368 | define_one_rw(freq_step); |
312 | 369 | ||
313 | static struct attribute * dbs_attributes[] = { | 370 | static struct attribute *dbs_attributes[] = { |
314 | &sampling_rate_max.attr, | 371 | &sampling_rate_max.attr, |
315 | &sampling_rate_min.attr, | 372 | &sampling_rate_min.attr, |
316 | &sampling_rate.attr, | 373 | &sampling_rate.attr, |
@@ -329,55 +386,78 @@ static struct attribute_group dbs_attr_group = { | |||
329 | 386 | ||
330 | /************************** sysfs end ************************/ | 387 | /************************** sysfs end ************************/ |
331 | 388 | ||
332 | static void dbs_check_cpu(int cpu) | 389 | static void dbs_check_cpu(struct cpu_dbs_info_s *this_dbs_info) |
333 | { | 390 | { |
334 | unsigned int idle_ticks, up_idle_ticks, down_idle_ticks; | 391 | unsigned int load = 0; |
335 | unsigned int tmp_idle_ticks, total_idle_ticks; | ||
336 | unsigned int freq_target; | 392 | unsigned int freq_target; |
337 | unsigned int freq_down_sampling_rate; | ||
338 | struct cpu_dbs_info_s *this_dbs_info = &per_cpu(cpu_dbs_info, cpu); | ||
339 | struct cpufreq_policy *policy; | ||
340 | 393 | ||
341 | if (!this_dbs_info->enable) | 394 | struct cpufreq_policy *policy; |
342 | return; | 395 | unsigned int j; |
343 | 396 | ||
344 | policy = this_dbs_info->cur_policy; | 397 | policy = this_dbs_info->cur_policy; |
345 | 398 | ||
346 | /* | 399 | /* |
347 | * The default safe range is 20% to 80% | 400 | * Every sampling_rate, we check, if current idle time is less |
348 | * Every sampling_rate, we check | 401 | * than 20% (default), then we try to increase frequency |
349 | * - If current idle time is less than 20%, then we try to | 402 | * Every sampling_rate*sampling_down_factor, we check, if current |
350 | * increase frequency | 403 | * idle time is more than 80%, then we try to decrease frequency |
351 | * Every sampling_rate*sampling_down_factor, we check | ||
352 | * - If current idle time is more than 80%, then we try to | ||
353 | * decrease frequency | ||
354 | * | 404 | * |
355 | * Any frequency increase takes it to the maximum frequency. | 405 | * Any frequency increase takes it to the maximum frequency. |
356 | * Frequency reduction happens at minimum steps of | 406 | * Frequency reduction happens at minimum steps of |
357 | * 5% (default) of max_frequency | 407 | * 5% (default) of maximum frequency |
358 | */ | 408 | */ |
359 | 409 | ||
360 | /* Check for frequency increase */ | 410 | /* Get Absolute Load */ |
361 | idle_ticks = UINT_MAX; | 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; | ||
362 | 415 | ||
363 | /* Check for frequency increase */ | 416 | j_dbs_info = &per_cpu(cpu_dbs_info, j); |
364 | total_idle_ticks = get_cpu_idle_time(cpu); | 417 | |
365 | tmp_idle_ticks = total_idle_ticks - | 418 | cur_idle_time = get_cpu_idle_time(j, &cur_wall_time); |
366 | this_dbs_info->prev_cpu_idle_up; | 419 | |
367 | this_dbs_info->prev_cpu_idle_up = total_idle_ticks; | 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; | ||
423 | |||
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; | ||
427 | |||
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); | ||
368 | 440 | ||
369 | if (tmp_idle_ticks < idle_ticks) | 441 | j_dbs_info->prev_cpu_nice = kstat_cpu(j).cpustat.nice; |
370 | idle_ticks = tmp_idle_ticks; | 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 | } | ||
371 | 450 | ||
372 | /* Scale idle ticks by 100 and compare with up and down ticks */ | 451 | /* |
373 | idle_ticks *= 100; | 452 | * break out if we 'cannot' reduce the speed as the user might |
374 | up_idle_ticks = (100 - dbs_tuners_ins.up_threshold) * | 453 | * want freq_step to be zero |
375 | usecs_to_jiffies(dbs_tuners_ins.sampling_rate); | 454 | */ |
455 | if (dbs_tuners_ins.freq_step == 0) | ||
456 | return; | ||
376 | 457 | ||
377 | if (idle_ticks < up_idle_ticks) { | 458 | /* Check for frequency increase */ |
459 | if (load > dbs_tuners_ins.up_threshold) { | ||
378 | this_dbs_info->down_skip = 0; | 460 | this_dbs_info->down_skip = 0; |
379 | this_dbs_info->prev_cpu_idle_down = | ||
380 | this_dbs_info->prev_cpu_idle_up; | ||
381 | 461 | ||
382 | /* if we are already at full speed then break out early */ | 462 | /* if we are already at full speed then break out early */ |
383 | if (this_dbs_info->requested_freq == policy->max) | 463 | if (this_dbs_info->requested_freq == policy->max) |
@@ -398,49 +478,24 @@ static void dbs_check_cpu(int cpu) | |||
398 | return; | 478 | return; |
399 | } | 479 | } |
400 | 480 | ||
401 | /* Check for frequency decrease */ | 481 | /* |
402 | this_dbs_info->down_skip++; | 482 | * The optimal frequency is the frequency that is the lowest that |
403 | if (this_dbs_info->down_skip < dbs_tuners_ins.sampling_down_factor) | 483 | * can support the current CPU usage without triggering the up |
404 | return; | 484 | * policy. To be safe, we focus 10 points under the threshold. |
405 | 485 | */ | |
406 | /* Check for frequency decrease */ | 486 | if (load < (dbs_tuners_ins.down_threshold - 10)) { |
407 | total_idle_ticks = this_dbs_info->prev_cpu_idle_up; | ||
408 | tmp_idle_ticks = total_idle_ticks - | ||
409 | this_dbs_info->prev_cpu_idle_down; | ||
410 | this_dbs_info->prev_cpu_idle_down = total_idle_ticks; | ||
411 | |||
412 | if (tmp_idle_ticks < idle_ticks) | ||
413 | idle_ticks = tmp_idle_ticks; | ||
414 | |||
415 | /* Scale idle ticks by 100 and compare with up and down ticks */ | ||
416 | idle_ticks *= 100; | ||
417 | this_dbs_info->down_skip = 0; | ||
418 | |||
419 | freq_down_sampling_rate = dbs_tuners_ins.sampling_rate * | ||
420 | dbs_tuners_ins.sampling_down_factor; | ||
421 | down_idle_ticks = (100 - dbs_tuners_ins.down_threshold) * | ||
422 | usecs_to_jiffies(freq_down_sampling_rate); | ||
423 | |||
424 | if (idle_ticks > down_idle_ticks) { | ||
425 | /* | ||
426 | * if we are already at the lowest speed then break out early | ||
427 | * or if we 'cannot' reduce the speed as the user might want | ||
428 | * freq_target to be zero | ||
429 | */ | ||
430 | if (this_dbs_info->requested_freq == policy->min | ||
431 | || dbs_tuners_ins.freq_step == 0) | ||
432 | return; | ||
433 | |||
434 | freq_target = (dbs_tuners_ins.freq_step * policy->max) / 100; | 487 | freq_target = (dbs_tuners_ins.freq_step * policy->max) / 100; |
435 | 488 | ||
436 | /* max freq cannot be less than 100. But who knows.... */ | ||
437 | if (unlikely(freq_target == 0)) | ||
438 | freq_target = 5; | ||
439 | |||
440 | this_dbs_info->requested_freq -= freq_target; | 489 | this_dbs_info->requested_freq -= freq_target; |
441 | if (this_dbs_info->requested_freq < policy->min) | 490 | if (this_dbs_info->requested_freq < policy->min) |
442 | this_dbs_info->requested_freq = policy->min; | 491 | this_dbs_info->requested_freq = policy->min; |
443 | 492 | ||
493 | /* | ||
494 | * if we cannot reduce the frequency anymore, break out early | ||
495 | */ | ||
496 | if (policy->cur == policy->min) | ||
497 | return; | ||
498 | |||
444 | __cpufreq_driver_target(policy, this_dbs_info->requested_freq, | 499 | __cpufreq_driver_target(policy, this_dbs_info->requested_freq, |
445 | CPUFREQ_RELATION_H); | 500 | CPUFREQ_RELATION_H); |
446 | return; | 501 | return; |
@@ -449,27 +504,45 @@ static void dbs_check_cpu(int cpu) | |||
449 | 504 | ||
450 | static void do_dbs_timer(struct work_struct *work) | 505 | static void do_dbs_timer(struct work_struct *work) |
451 | { | 506 | { |
452 | int i; | 507 | struct cpu_dbs_info_s *dbs_info = |
453 | mutex_lock(&dbs_mutex); | 508 | container_of(work, struct cpu_dbs_info_s, work.work); |
454 | for_each_online_cpu(i) | 509 | unsigned int cpu = dbs_info->cpu; |
455 | dbs_check_cpu(i); | 510 | |
456 | schedule_delayed_work(&dbs_work, | 511 | /* We want all CPUs to do sampling nearly on same jiffy */ |
457 | usecs_to_jiffies(dbs_tuners_ins.sampling_rate)); | 512 | int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate); |
458 | mutex_unlock(&dbs_mutex); | 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); | ||
459 | } | 528 | } |
460 | 529 | ||
461 | static inline void dbs_timer_init(void) | 530 | static inline void dbs_timer_init(struct cpu_dbs_info_s *dbs_info) |
462 | { | 531 | { |
463 | init_timer_deferrable(&dbs_work.timer); | 532 | /* We want all CPUs to do sampling nearly on same jiffy */ |
464 | schedule_delayed_work(&dbs_work, | 533 | int delay = usecs_to_jiffies(dbs_tuners_ins.sampling_rate); |
465 | usecs_to_jiffies(dbs_tuners_ins.sampling_rate)); | 534 | delay -= jiffies % delay; |
466 | return; | 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); | ||
467 | } | 540 | } |
468 | 541 | ||
469 | static inline void dbs_timer_exit(void) | 542 | static inline void dbs_timer_exit(struct cpu_dbs_info_s *dbs_info) |
470 | { | 543 | { |
471 | cancel_delayed_work(&dbs_work); | 544 | dbs_info->enable = 0; |
472 | return; | 545 | cancel_delayed_work(&dbs_info->work); |
473 | } | 546 | } |
474 | 547 | ||
475 | static int cpufreq_governor_dbs(struct cpufreq_policy *policy, | 548 | static int cpufreq_governor_dbs(struct cpufreq_policy *policy, |
@@ -503,11 +576,13 @@ static int cpufreq_governor_dbs(struct cpufreq_policy *policy, | |||
503 | j_dbs_info = &per_cpu(cpu_dbs_info, j); | 576 | j_dbs_info = &per_cpu(cpu_dbs_info, j); |
504 | j_dbs_info->cur_policy = policy; | 577 | j_dbs_info->cur_policy = policy; |
505 | 578 | ||
506 | j_dbs_info->prev_cpu_idle_up = get_cpu_idle_time(cpu); | 579 | j_dbs_info->prev_cpu_idle = get_cpu_idle_time(j, |
507 | j_dbs_info->prev_cpu_idle_down | 580 | &j_dbs_info->prev_cpu_wall); |
508 | = j_dbs_info->prev_cpu_idle_up; | 581 | if (dbs_tuners_ins.ignore_nice) { |
582 | j_dbs_info->prev_cpu_nice = | ||
583 | kstat_cpu(j).cpustat.nice; | ||
584 | } | ||
509 | } | 585 | } |
510 | this_dbs_info->enable = 1; | ||
511 | this_dbs_info->down_skip = 0; | 586 | this_dbs_info->down_skip = 0; |
512 | this_dbs_info->requested_freq = policy->cur; | 587 | this_dbs_info->requested_freq = policy->cur; |
513 | 588 | ||
@@ -523,38 +598,36 @@ static int cpufreq_governor_dbs(struct cpufreq_policy *policy, | |||
523 | if (latency == 0) | 598 | if (latency == 0) |
524 | latency = 1; | 599 | latency = 1; |
525 | 600 | ||
526 | def_sampling_rate = 10 * latency * | 601 | def_sampling_rate = |
527 | DEF_SAMPLING_RATE_LATENCY_MULTIPLIER; | 602 | max(latency * LATENCY_MULTIPLIER, |
528 | 603 | MIN_STAT_SAMPLING_RATE); | |
529 | if (def_sampling_rate < MIN_STAT_SAMPLING_RATE) | ||
530 | def_sampling_rate = MIN_STAT_SAMPLING_RATE; | ||
531 | 604 | ||
532 | dbs_tuners_ins.sampling_rate = def_sampling_rate; | 605 | dbs_tuners_ins.sampling_rate = def_sampling_rate; |
533 | 606 | ||
534 | dbs_timer_init(); | ||
535 | cpufreq_register_notifier( | 607 | cpufreq_register_notifier( |
536 | &dbs_cpufreq_notifier_block, | 608 | &dbs_cpufreq_notifier_block, |
537 | CPUFREQ_TRANSITION_NOTIFIER); | 609 | CPUFREQ_TRANSITION_NOTIFIER); |
538 | } | 610 | } |
611 | dbs_timer_init(this_dbs_info); | ||
539 | 612 | ||
540 | mutex_unlock(&dbs_mutex); | 613 | mutex_unlock(&dbs_mutex); |
614 | |||
541 | break; | 615 | break; |
542 | 616 | ||
543 | case CPUFREQ_GOV_STOP: | 617 | case CPUFREQ_GOV_STOP: |
544 | mutex_lock(&dbs_mutex); | 618 | mutex_lock(&dbs_mutex); |
545 | this_dbs_info->enable = 0; | 619 | dbs_timer_exit(this_dbs_info); |
546 | sysfs_remove_group(&policy->kobj, &dbs_attr_group); | 620 | sysfs_remove_group(&policy->kobj, &dbs_attr_group); |
547 | dbs_enable--; | 621 | dbs_enable--; |
622 | |||
548 | /* | 623 | /* |
549 | * Stop the timerschedule work, when this governor | 624 | * Stop the timerschedule work, when this governor |
550 | * is used for first time | 625 | * is used for first time |
551 | */ | 626 | */ |
552 | if (dbs_enable == 0) { | 627 | if (dbs_enable == 0) |
553 | dbs_timer_exit(); | ||
554 | cpufreq_unregister_notifier( | 628 | cpufreq_unregister_notifier( |
555 | &dbs_cpufreq_notifier_block, | 629 | &dbs_cpufreq_notifier_block, |
556 | CPUFREQ_TRANSITION_NOTIFIER); | 630 | CPUFREQ_TRANSITION_NOTIFIER); |
557 | } | ||
558 | 631 | ||
559 | mutex_unlock(&dbs_mutex); | 632 | mutex_unlock(&dbs_mutex); |
560 | 633 | ||
@@ -571,6 +644,7 @@ static int cpufreq_governor_dbs(struct cpufreq_policy *policy, | |||
571 | this_dbs_info->cur_policy, | 644 | this_dbs_info->cur_policy, |
572 | policy->min, CPUFREQ_RELATION_L); | 645 | policy->min, CPUFREQ_RELATION_L); |
573 | mutex_unlock(&dbs_mutex); | 646 | mutex_unlock(&dbs_mutex); |
647 | |||
574 | break; | 648 | break; |
575 | } | 649 | } |
576 | return 0; | 650 | return 0; |
@@ -588,23 +662,33 @@ struct cpufreq_governor cpufreq_gov_conservative = { | |||
588 | 662 | ||
589 | static int __init cpufreq_gov_dbs_init(void) | 663 | static int __init cpufreq_gov_dbs_init(void) |
590 | { | 664 | { |
591 | return cpufreq_register_governor(&cpufreq_gov_conservative); | 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; | ||
592 | } | 678 | } |
593 | 679 | ||
594 | static void __exit cpufreq_gov_dbs_exit(void) | 680 | static void __exit cpufreq_gov_dbs_exit(void) |
595 | { | 681 | { |
596 | /* Make sure that the scheduled work is indeed not running */ | ||
597 | flush_scheduled_work(); | ||
598 | |||
599 | cpufreq_unregister_governor(&cpufreq_gov_conservative); | 682 | cpufreq_unregister_governor(&cpufreq_gov_conservative); |
683 | destroy_workqueue(kconservative_wq); | ||
600 | } | 684 | } |
601 | 685 | ||
602 | 686 | ||
603 | MODULE_AUTHOR ("Alexander Clouter <alex-kernel@digriz.org.uk>"); | 687 | MODULE_AUTHOR("Alexander Clouter <alex@digriz.org.uk>"); |
604 | MODULE_DESCRIPTION ("'cpufreq_conservative' - A dynamic cpufreq governor for " | 688 | MODULE_DESCRIPTION("'cpufreq_conservative' - A dynamic cpufreq governor for " |
605 | "Low Latency Frequency Transition capable processors " | 689 | "Low Latency Frequency Transition capable processors " |
606 | "optimised for use in a battery environment"); | 690 | "optimised for use in a battery environment"); |
607 | MODULE_LICENSE ("GPL"); | 691 | MODULE_LICENSE("GPL"); |
608 | 692 | ||
609 | #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_CONSERVATIVE | 693 | #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_CONSERVATIVE |
610 | fs_initcall(cpufreq_gov_dbs_init); | 694 | fs_initcall(cpufreq_gov_dbs_init); |