blob: 642dd0f183a8d90f8804b64fd39ffee95e6bdfa8 [file] [log] [blame]
viresh kumar2aacdff2012-10-23 01:28:05 +02001/*
2 * drivers/cpufreq/cpufreq_governor.c
3 *
4 * CPUFREQ governors common code
5 *
Viresh Kumar4471a342012-10-26 00:47:42 +02006 * Copyright (C) 2001 Russell King
7 * (C) 2003 Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>.
8 * (C) 2003 Jun Nakajima <jun.nakajima@intel.com>
9 * (C) 2009 Alexander Clouter <alex@digriz.org.uk>
10 * (c) 2012 Viresh Kumar <viresh.kumar@linaro.org>
11 *
viresh kumar2aacdff2012-10-23 01:28:05 +020012 * This program is free software; you can redistribute it and/or modify
13 * it under the terms of the GNU General Public License version 2 as
14 * published by the Free Software Foundation.
15 */
16
Viresh Kumar4471a342012-10-26 00:47:42 +020017#define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
18
viresh kumar2aacdff2012-10-23 01:28:05 +020019#include <linux/export.h>
20#include <linux/kernel_stat.h>
Rafael J. Wysockiadaf9fc2016-03-10 20:44:47 +010021#include <linux/sched.h>
Viresh Kumar4d5dcc42013-03-27 15:58:58 +000022#include <linux/slab.h>
Viresh Kumar4471a342012-10-26 00:47:42 +020023
24#include "cpufreq_governor.h"
25
Rafael J. Wysocki8c8f77f2016-02-21 00:51:27 +010026static DEFINE_PER_CPU(struct cpu_dbs_info, cpu_dbs);
27
Rafael J. Wysocki1112e9d2016-02-21 00:53:06 +010028static DEFINE_MUTEX(gov_dbs_data_mutex);
Rafael J. Wysocki2bb8d942016-02-07 16:01:31 +010029
Viresh Kumaraded3872016-02-11 17:31:15 +053030/* Common sysfs tunables */
31/**
32 * store_sampling_rate - update sampling rate effective immediately if needed.
33 *
34 * If new rate is smaller than the old, simply updating
35 * dbs.sampling_rate might not be appropriate. For example, if the
36 * original sampling_rate was 1 second and the requested new sampling rate is 10
37 * ms because the user needs immediate reaction from ondemand governor, but not
38 * sure if higher frequency will be required or not, then, the governor may
39 * change the sampling rate too late; up to 1 second later. Thus, if we are
40 * reducing the sampling rate, we need to make the new value effective
41 * immediately.
42 *
Viresh Kumaraded3872016-02-11 17:31:15 +053043 * This must be called with dbs_data->mutex held, otherwise traversing
44 * policy_dbs_list isn't safe.
45 */
Rafael J. Wysocki0dd3c1d2016-03-22 02:47:51 +010046ssize_t store_sampling_rate(struct gov_attr_set *attr_set, const char *buf,
Viresh Kumaraded3872016-02-11 17:31:15 +053047 size_t count)
48{
Rafael J. Wysocki0dd3c1d2016-03-22 02:47:51 +010049 struct dbs_data *dbs_data = to_dbs_data(attr_set);
Viresh Kumaraded3872016-02-11 17:31:15 +053050 struct policy_dbs_info *policy_dbs;
51 unsigned int rate;
52 int ret;
53 ret = sscanf(buf, "%u", &rate);
54 if (ret != 1)
55 return -EINVAL;
56
57 dbs_data->sampling_rate = max(rate, dbs_data->min_sampling_rate);
58
59 /*
60 * We are operating under dbs_data->mutex and so the list and its
61 * entries can't be freed concurrently.
62 */
Rafael J. Wysocki0dd3c1d2016-03-22 02:47:51 +010063 list_for_each_entry(policy_dbs, &attr_set->policy_list, list) {
Viresh Kumaraded3872016-02-11 17:31:15 +053064 mutex_lock(&policy_dbs->timer_mutex);
65 /*
66 * On 32-bit architectures this may race with the
67 * sample_delay_ns read in dbs_update_util_handler(), but that
68 * really doesn't matter. If the read returns a value that's
69 * too big, the sample will be skipped, but the next invocation
70 * of dbs_update_util_handler() (when the update has been
Rafael J. Wysocki78347cd2016-02-15 02:20:11 +010071 * completed) will take a sample.
Viresh Kumaraded3872016-02-11 17:31:15 +053072 *
73 * If this runs in parallel with dbs_work_handler(), we may end
74 * up overwriting the sample_delay_ns value that it has just
Rafael J. Wysocki78347cd2016-02-15 02:20:11 +010075 * written, but it will be corrected next time a sample is
76 * taken, so it shouldn't be significant.
Viresh Kumaraded3872016-02-11 17:31:15 +053077 */
Rafael J. Wysocki78347cd2016-02-15 02:20:11 +010078 gov_update_sample_delay(policy_dbs, 0);
Viresh Kumaraded3872016-02-11 17:31:15 +053079 mutex_unlock(&policy_dbs->timer_mutex);
80 }
81
82 return count;
83}
84EXPORT_SYMBOL_GPL(store_sampling_rate);
85
Rafael J. Wysockia33cce12016-02-18 02:26:55 +010086/**
87 * gov_update_cpu_data - Update CPU load data.
Rafael J. Wysockia33cce12016-02-18 02:26:55 +010088 * @dbs_data: Top-level governor data pointer.
89 *
90 * Update CPU load data for all CPUs in the domain governed by @dbs_data
91 * (that may be a single policy or a bunch of them if governor tunables are
92 * system-wide).
93 *
94 * Call under the @dbs_data mutex.
95 */
Rafael J. Wysocki8c8f77f2016-02-21 00:51:27 +010096void gov_update_cpu_data(struct dbs_data *dbs_data)
Rafael J. Wysockia33cce12016-02-18 02:26:55 +010097{
98 struct policy_dbs_info *policy_dbs;
99
Rafael J. Wysocki0dd3c1d2016-03-22 02:47:51 +0100100 list_for_each_entry(policy_dbs, &dbs_data->attr_set.policy_list, list) {
Rafael J. Wysockia33cce12016-02-18 02:26:55 +0100101 unsigned int j;
102
103 for_each_cpu(j, policy_dbs->policy->cpus) {
Rafael J. Wysocki8c8f77f2016-02-21 00:51:27 +0100104 struct cpu_dbs_info *j_cdbs = &per_cpu(cpu_dbs, j);
Rafael J. Wysockia33cce12016-02-18 02:26:55 +0100105
Rafael J. Wysockib4f4b4b2016-04-28 01:19:03 +0200106 j_cdbs->prev_cpu_idle = get_cpu_idle_time(j, &j_cdbs->prev_update_time,
Rafael J. Wysockia33cce12016-02-18 02:26:55 +0100107 dbs_data->io_is_busy);
108 if (dbs_data->ignore_nice_load)
109 j_cdbs->prev_cpu_nice = kcpustat_cpu(j).cpustat[CPUTIME_NICE];
110 }
111 }
112}
113EXPORT_SYMBOL_GPL(gov_update_cpu_data);
114
Rafael J. Wysocki4cccf752016-02-15 02:19:31 +0100115unsigned int dbs_update(struct cpufreq_policy *policy)
Viresh Kumar4471a342012-10-26 00:47:42 +0200116{
Rafael J. Wysockibc505472016-02-07 16:24:26 +0100117 struct policy_dbs_info *policy_dbs = policy->governor_data;
118 struct dbs_data *dbs_data = policy_dbs->dbs_data;
Viresh Kumarff4b1782016-02-09 09:01:32 +0530119 unsigned int ignore_nice = dbs_data->ignore_nice_load;
Viresh Kumar4471a342012-10-26 00:47:42 +0200120 unsigned int max_load = 0;
Rafael J. Wysocki8847e032016-02-18 02:20:13 +0100121 unsigned int sampling_rate, io_busy, j;
Viresh Kumar4471a342012-10-26 00:47:42 +0200122
Rafael J. Wysocki57dc3bc2016-02-15 02:20:51 +0100123 /*
124 * Sometimes governors may use an additional multiplier to increase
125 * sample delays temporarily. Apply that multiplier to sampling_rate
126 * so as to keep the wake-up-from-idle detection logic a bit
127 * conservative.
128 */
129 sampling_rate = dbs_data->sampling_rate * policy_dbs->rate_mult;
Rafael J. Wysocki8847e032016-02-18 02:20:13 +0100130 /*
131 * For the purpose of ondemand, waiting for disk IO is an indication
132 * that you're performance critical, and not that the system is actually
133 * idle, so do not add the iowait time to the CPU idle time then.
134 */
135 io_busy = dbs_data->io_is_busy;
Viresh Kumar4471a342012-10-26 00:47:42 +0200136
Stratos Karafotisdfa5bb62013-06-05 19:01:25 +0300137 /* Get Absolute Load */
Viresh Kumar4471a342012-10-26 00:47:42 +0200138 for_each_cpu(j, policy->cpus) {
Rafael J. Wysocki8c8f77f2016-02-21 00:51:27 +0100139 struct cpu_dbs_info *j_cdbs = &per_cpu(cpu_dbs, j);
Rafael J. Wysockib4f4b4b2016-04-28 01:19:03 +0200140 u64 update_time, cur_idle_time;
141 unsigned int idle_time, time_elapsed;
Viresh Kumar4471a342012-10-26 00:47:42 +0200142 unsigned int load;
143
Rafael J. Wysockib4f4b4b2016-04-28 01:19:03 +0200144 cur_idle_time = get_cpu_idle_time(j, &update_time, io_busy);
Viresh Kumar4471a342012-10-26 00:47:42 +0200145
Rafael J. Wysockib4f4b4b2016-04-28 01:19:03 +0200146 time_elapsed = update_time - j_cdbs->prev_update_time;
147 j_cdbs->prev_update_time = update_time;
Viresh Kumar4471a342012-10-26 00:47:42 +0200148
Rafael J. Wysocki94862a62016-04-21 20:57:47 +0200149 idle_time = cur_idle_time - j_cdbs->prev_cpu_idle;
150 j_cdbs->prev_cpu_idle = cur_idle_time;
Viresh Kumar4471a342012-10-26 00:47:42 +0200151
152 if (ignore_nice) {
Rafael J. Wysocki679b8fe2016-02-15 02:15:50 +0100153 u64 cur_nice = kcpustat_cpu(j).cpustat[CPUTIME_NICE];
Viresh Kumar4471a342012-10-26 00:47:42 +0200154
Rafael J. Wysocki679b8fe2016-02-15 02:15:50 +0100155 idle_time += cputime_to_usecs(cur_nice - j_cdbs->prev_cpu_nice);
156 j_cdbs->prev_cpu_nice = cur_nice;
Viresh Kumar4471a342012-10-26 00:47:42 +0200157 }
158
Rafael J. Wysocki9485e4c2016-05-06 01:30:37 +0200159 if (unlikely(!time_elapsed)) {
Viresh Kumarc8ae4812014-06-09 14:21:24 +0530160 /*
Rafael J. Wysocki9485e4c2016-05-06 01:30:37 +0200161 * That can only happen when this function is called
162 * twice in a row with a very short interval between the
163 * calls, so the previous load value can be used then.
Viresh Kumarc8ae4812014-06-09 14:21:24 +0530164 */
Rafael J. Wysocki9485e4c2016-05-06 01:30:37 +0200165 load = j_cdbs->prev_load;
166 } else if (unlikely(time_elapsed > 2 * sampling_rate &&
167 j_cdbs->prev_load)) {
168 /*
169 * If the CPU had gone completely idle and a task has
170 * just woken up on this CPU now, it would be unfair to
171 * calculate 'load' the usual way for this elapsed
172 * time-window, because it would show near-zero load,
173 * irrespective of how CPU intensive that task actually
174 * was. This is undesirable for latency-sensitive bursty
175 * workloads.
176 *
177 * To avoid this, reuse the 'load' from the previous
178 * time-window and give this task a chance to start with
179 * a reasonably high CPU frequency. However, that
180 * shouldn't be over-done, lest we get stuck at a high
181 * load (high frequency) for too long, even when the
182 * current system load has actually dropped down, so
183 * clear prev_load to guarantee that the load will be
184 * computed again next time.
185 *
186 * Detecting this situation is easy: the governor's
187 * utilization update handler would not have run during
188 * CPU-idle periods. Hence, an unusually large
189 * 'time_elapsed' (as compared to the sampling rate)
190 * indicates this scenario.
191 */
192 load = j_cdbs->prev_load;
Viresh Kumarc8ae4812014-06-09 14:21:24 +0530193 j_cdbs->prev_load = 0;
Srivatsa S. Bhat18b46ab2014-06-08 02:11:43 +0530194 } else {
Rafael J. Wysocki9485e4c2016-05-06 01:30:37 +0200195 if (time_elapsed >= idle_time) {
196 load = 100 * (time_elapsed - idle_time) / time_elapsed;
197 } else {
198 /*
199 * That can happen if idle_time is returned by
200 * get_cpu_idle_time_jiffy(). In that case
201 * idle_time is roughly equal to the difference
202 * between time_elapsed and "busy time" obtained
203 * from CPU statistics. Then, the "busy time"
204 * can end up being greater than time_elapsed
205 * (for example, if jiffies_64 and the CPU
206 * statistics are updated by different CPUs),
207 * so idle_time may in fact be negative. That
208 * means, though, that the CPU was busy all
209 * the time (on the rough average) during the
210 * last sampling interval and 100 can be
211 * returned as the load.
212 */
213 load = (int)idle_time < 0 ? 100 : 0;
214 }
Srivatsa S. Bhat18b46ab2014-06-08 02:11:43 +0530215 j_cdbs->prev_load = load;
Srivatsa S. Bhat18b46ab2014-06-08 02:11:43 +0530216 }
Viresh Kumar4471a342012-10-26 00:47:42 +0200217
Viresh Kumar4471a342012-10-26 00:47:42 +0200218 if (load > max_load)
219 max_load = load;
220 }
Rafael J. Wysocki4cccf752016-02-15 02:19:31 +0100221 return max_load;
Viresh Kumar4471a342012-10-26 00:47:42 +0200222}
Rafael J. Wysocki4cccf752016-02-15 02:19:31 +0100223EXPORT_SYMBOL_GPL(dbs_update);
Viresh Kumar4471a342012-10-26 00:47:42 +0200224
Viresh Kumar70f43e52015-12-09 07:34:42 +0530225static void dbs_work_handler(struct work_struct *work)
Viresh Kumar43e0ee32015-07-18 11:31:00 +0530226{
Rafael J. Wysockie40e7b22016-02-10 17:07:44 +0100227 struct policy_dbs_info *policy_dbs;
Viresh Kumar3a91b0692015-10-29 08:08:38 +0530228 struct cpufreq_policy *policy;
Rafael J. Wysockiea59ee0d2016-02-07 16:09:51 +0100229 struct dbs_governor *gov;
Viresh Kumar43e0ee32015-07-18 11:31:00 +0530230
Rafael J. Wysockie40e7b22016-02-10 17:07:44 +0100231 policy_dbs = container_of(work, struct policy_dbs_info, work);
232 policy = policy_dbs->policy;
Rafael J. Wysockiea59ee0d2016-02-07 16:09:51 +0100233 gov = dbs_governor_of(policy);
Viresh Kumar3a91b0692015-10-29 08:08:38 +0530234
Viresh Kumar70f43e52015-12-09 07:34:42 +0530235 /*
Rafael J. Wysocki9be4fd22016-02-10 16:53:50 +0100236 * Make sure cpufreq_governor_limits() isn't evaluating load or the
237 * ondemand governor isn't updating the sampling rate in parallel.
Viresh Kumar70f43e52015-12-09 07:34:42 +0530238 */
Rafael J. Wysockie40e7b22016-02-10 17:07:44 +0100239 mutex_lock(&policy_dbs->timer_mutex);
Rafael J. Wysocki07aa4402016-02-15 02:22:13 +0100240 gov_update_sample_delay(policy_dbs, gov->gov_dbs_timer(policy));
Rafael J. Wysockie40e7b22016-02-10 17:07:44 +0100241 mutex_unlock(&policy_dbs->timer_mutex);
Viresh Kumar70f43e52015-12-09 07:34:42 +0530242
Rafael J. Wysockie4db2812016-02-15 02:13:42 +0100243 /* Allow the utilization update handler to queue up more work. */
244 atomic_set(&policy_dbs->work_count, 0);
Rafael J. Wysocki9be4fd22016-02-10 16:53:50 +0100245 /*
Rafael J. Wysockie4db2812016-02-15 02:13:42 +0100246 * If the update below is reordered with respect to the sample delay
247 * modification, the utilization update handler may end up using a stale
248 * sample delay value.
Rafael J. Wysocki9be4fd22016-02-10 16:53:50 +0100249 */
Rafael J. Wysockie4db2812016-02-15 02:13:42 +0100250 smp_wmb();
251 policy_dbs->work_in_progress = false;
Viresh Kumar70f43e52015-12-09 07:34:42 +0530252}
253
Rafael J. Wysocki9be4fd22016-02-10 16:53:50 +0100254static void dbs_irq_work(struct irq_work *irq_work)
Viresh Kumar70f43e52015-12-09 07:34:42 +0530255{
Rafael J. Wysockie40e7b22016-02-10 17:07:44 +0100256 struct policy_dbs_info *policy_dbs;
Rafael J. Wysocki9be4fd22016-02-10 16:53:50 +0100257
Rafael J. Wysockie40e7b22016-02-10 17:07:44 +0100258 policy_dbs = container_of(irq_work, struct policy_dbs_info, irq_work);
Rafael J. Wysocki539a4c42016-03-22 01:17:43 +0100259 schedule_work_on(smp_processor_id(), &policy_dbs->work);
Rafael J. Wysocki9be4fd22016-02-10 16:53:50 +0100260}
261
Rafael J. Wysocki9be4fd22016-02-10 16:53:50 +0100262static void dbs_update_util_handler(struct update_util_data *data, u64 time,
Rafael J. Wysocki58919e82016-08-16 22:14:55 +0200263 unsigned int flags)
Rafael J. Wysocki9be4fd22016-02-10 16:53:50 +0100264{
265 struct cpu_dbs_info *cdbs = container_of(data, struct cpu_dbs_info, update_util);
Rafael J. Wysockie40e7b22016-02-10 17:07:44 +0100266 struct policy_dbs_info *policy_dbs = cdbs->policy_dbs;
Rafael J. Wysocki27de3482016-02-22 14:14:34 +0100267 u64 delta_ns, lst;
Viresh Kumar70f43e52015-12-09 07:34:42 +0530268
269 /*
Rafael J. Wysocki9be4fd22016-02-10 16:53:50 +0100270 * The work may not be allowed to be queued up right now.
271 * Possible reasons:
272 * - Work has already been queued up or is in progress.
Rafael J. Wysocki9be4fd22016-02-10 16:53:50 +0100273 * - It is too early (too little time from the previous sample).
Viresh Kumar70f43e52015-12-09 07:34:42 +0530274 */
Rafael J. Wysockie4db2812016-02-15 02:13:42 +0100275 if (policy_dbs->work_in_progress)
276 return;
Rafael J. Wysocki9be4fd22016-02-10 16:53:50 +0100277
Rafael J. Wysockie4db2812016-02-15 02:13:42 +0100278 /*
279 * If the reads below are reordered before the check above, the value
280 * of sample_delay_ns used in the computation may be stale.
281 */
282 smp_rmb();
Rafael J. Wysocki27de3482016-02-22 14:14:34 +0100283 lst = READ_ONCE(policy_dbs->last_sample_time);
284 delta_ns = time - lst;
Rafael J. Wysockie4db2812016-02-15 02:13:42 +0100285 if ((s64)delta_ns < policy_dbs->sample_delay_ns)
286 return;
287
288 /*
289 * If the policy is not shared, the irq_work may be queued up right away
290 * at this point. Otherwise, we need to ensure that only one of the
291 * CPUs sharing the policy will do that.
292 */
Rafael J. Wysocki27de3482016-02-22 14:14:34 +0100293 if (policy_dbs->is_shared) {
294 if (!atomic_add_unless(&policy_dbs->work_count, 1, 1))
295 return;
296
297 /*
298 * If another CPU updated last_sample_time in the meantime, we
299 * shouldn't be here, so clear the work counter and bail out.
300 */
301 if (unlikely(lst != READ_ONCE(policy_dbs->last_sample_time))) {
302 atomic_set(&policy_dbs->work_count, 0);
303 return;
304 }
305 }
Rafael J. Wysockie4db2812016-02-15 02:13:42 +0100306
307 policy_dbs->last_sample_time = time;
308 policy_dbs->work_in_progress = true;
309 irq_work_queue(&policy_dbs->irq_work);
Viresh Kumar43e0ee32015-07-18 11:31:00 +0530310}
Viresh Kumar44472662013-01-31 17:28:02 +0000311
Rafael J. Wysocki0bed6122016-04-02 01:08:43 +0200312static void gov_set_update_util(struct policy_dbs_info *policy_dbs,
313 unsigned int delay_us)
314{
315 struct cpufreq_policy *policy = policy_dbs->policy;
316 int cpu;
317
318 gov_update_sample_delay(policy_dbs, delay_us);
319 policy_dbs->last_sample_time = 0;
320
321 for_each_cpu(cpu, policy->cpus) {
322 struct cpu_dbs_info *cdbs = &per_cpu(cpu_dbs, cpu);
323
324 cpufreq_add_update_util_hook(cpu, &cdbs->update_util,
325 dbs_update_util_handler);
326 }
327}
328
329static inline void gov_clear_update_util(struct cpufreq_policy *policy)
330{
331 int i;
332
333 for_each_cpu(i, policy->cpus)
334 cpufreq_remove_update_util_hook(i);
335
336 synchronize_sched();
337}
338
Rafael J. Wysockibc505472016-02-07 16:24:26 +0100339static struct policy_dbs_info *alloc_policy_dbs_info(struct cpufreq_policy *policy,
340 struct dbs_governor *gov)
Viresh Kumar44152cb2015-07-18 11:30:59 +0530341{
Rafael J. Wysockie40e7b22016-02-10 17:07:44 +0100342 struct policy_dbs_info *policy_dbs;
Viresh Kumar44152cb2015-07-18 11:30:59 +0530343 int j;
344
Rafael J. Wysocki7d5a9952016-02-18 18:40:14 +0100345 /* Allocate memory for per-policy governor data. */
346 policy_dbs = gov->alloc();
Rafael J. Wysockie40e7b22016-02-10 17:07:44 +0100347 if (!policy_dbs)
Rafael J. Wysockibc505472016-02-07 16:24:26 +0100348 return NULL;
Viresh Kumar44152cb2015-07-18 11:30:59 +0530349
Viresh Kumar581c2142016-02-11 17:31:14 +0530350 policy_dbs->policy = policy;
Rafael J. Wysockie40e7b22016-02-10 17:07:44 +0100351 mutex_init(&policy_dbs->timer_mutex);
Rafael J. Wysocki686cc632016-02-08 23:41:10 +0100352 atomic_set(&policy_dbs->work_count, 0);
Rafael J. Wysockie40e7b22016-02-10 17:07:44 +0100353 init_irq_work(&policy_dbs->irq_work, dbs_irq_work);
354 INIT_WORK(&policy_dbs->work, dbs_work_handler);
Rafael J. Wysockicea6a9e2016-02-07 16:25:02 +0100355
356 /* Set policy_dbs for all CPUs, online+offline */
357 for_each_cpu(j, policy->related_cpus) {
Rafael J. Wysocki8c8f77f2016-02-21 00:51:27 +0100358 struct cpu_dbs_info *j_cdbs = &per_cpu(cpu_dbs, j);
Rafael J. Wysockicea6a9e2016-02-07 16:25:02 +0100359
360 j_cdbs->policy_dbs = policy_dbs;
Rafael J. Wysockicea6a9e2016-02-07 16:25:02 +0100361 }
Rafael J. Wysockibc505472016-02-07 16:24:26 +0100362 return policy_dbs;
Viresh Kumar44152cb2015-07-18 11:30:59 +0530363}
364
Rafael J. Wysocki8c8f77f2016-02-21 00:51:27 +0100365static void free_policy_dbs_info(struct policy_dbs_info *policy_dbs,
Rafael J. Wysocki7bdad342016-02-07 16:05:07 +0100366 struct dbs_governor *gov)
Viresh Kumar44152cb2015-07-18 11:30:59 +0530367{
Viresh Kumar44152cb2015-07-18 11:30:59 +0530368 int j;
369
Rafael J. Wysockie40e7b22016-02-10 17:07:44 +0100370 mutex_destroy(&policy_dbs->timer_mutex);
Viresh Kumar5e4500d2015-12-03 09:37:52 +0530371
Rafael J. Wysocki8c8f77f2016-02-21 00:51:27 +0100372 for_each_cpu(j, policy_dbs->policy->related_cpus) {
373 struct cpu_dbs_info *j_cdbs = &per_cpu(cpu_dbs, j);
Viresh Kumar44152cb2015-07-18 11:30:59 +0530374
Rafael J. Wysockicea6a9e2016-02-07 16:25:02 +0100375 j_cdbs->policy_dbs = NULL;
376 j_cdbs->update_util.func = NULL;
377 }
Rafael J. Wysocki7d5a9952016-02-18 18:40:14 +0100378 gov->free(policy_dbs);
Viresh Kumar44152cb2015-07-18 11:30:59 +0530379}
380
Rafael J. Wysockie7888922016-06-02 23:24:15 +0200381int cpufreq_dbs_governor_init(struct cpufreq_policy *policy)
Viresh Kumar714a2d92015-06-04 16:43:27 +0530382{
Rafael J. Wysockiea59ee0d2016-02-07 16:09:51 +0100383 struct dbs_governor *gov = dbs_governor_of(policy);
Rafael J. Wysocki1112e9d2016-02-21 00:53:06 +0100384 struct dbs_data *dbs_data;
Rafael J. Wysockibc505472016-02-07 16:24:26 +0100385 struct policy_dbs_info *policy_dbs;
Viresh Kumar714a2d92015-06-04 16:43:27 +0530386 unsigned int latency;
Rafael J. Wysocki1112e9d2016-02-21 00:53:06 +0100387 int ret = 0;
Viresh Kumar714a2d92015-06-04 16:43:27 +0530388
Viresh Kumara72c4952015-07-18 11:31:01 +0530389 /* State should be equivalent to EXIT */
390 if (policy->governor_data)
391 return -EBUSY;
392
Rafael J. Wysockibc505472016-02-07 16:24:26 +0100393 policy_dbs = alloc_policy_dbs_info(policy, gov);
394 if (!policy_dbs)
395 return -ENOMEM;
396
Rafael J. Wysocki1112e9d2016-02-21 00:53:06 +0100397 /* Protect gov->gdbs_data against concurrent updates. */
398 mutex_lock(&gov_dbs_data_mutex);
399
400 dbs_data = gov->gdbs_data;
Viresh Kumar714a2d92015-06-04 16:43:27 +0530401 if (dbs_data) {
Rafael J. Wysockibc505472016-02-07 16:24:26 +0100402 if (WARN_ON(have_governor_per_policy())) {
403 ret = -EINVAL;
404 goto free_policy_dbs_info;
405 }
Rafael J. Wysockibc505472016-02-07 16:24:26 +0100406 policy_dbs->dbs_data = dbs_data;
407 policy->governor_data = policy_dbs;
Viresh Kumarc54df072016-02-10 11:00:25 +0530408
Rafael J. Wysocki0dd3c1d2016-03-22 02:47:51 +0100409 gov_attr_set_get(&dbs_data->attr_set, &policy_dbs->list);
Rafael J. Wysocki1112e9d2016-02-21 00:53:06 +0100410 goto out;
Viresh Kumar714a2d92015-06-04 16:43:27 +0530411 }
412
413 dbs_data = kzalloc(sizeof(*dbs_data), GFP_KERNEL);
Rafael J. Wysockibc505472016-02-07 16:24:26 +0100414 if (!dbs_data) {
415 ret = -ENOMEM;
416 goto free_policy_dbs_info;
417 }
Viresh Kumar44152cb2015-07-18 11:30:59 +0530418
Rafael J. Wysocki0dd3c1d2016-03-22 02:47:51 +0100419 gov_attr_set_init(&dbs_data->attr_set, &policy_dbs->list);
Viresh Kumar714a2d92015-06-04 16:43:27 +0530420
Rafael J. Wysocki9a15fb22016-05-18 22:59:49 +0200421 ret = gov->init(dbs_data);
Viresh Kumar714a2d92015-06-04 16:43:27 +0530422 if (ret)
Rafael J. Wysockie40e7b22016-02-10 17:07:44 +0100423 goto free_policy_dbs_info;
Viresh Kumar714a2d92015-06-04 16:43:27 +0530424
425 /* policy latency is in ns. Convert it to us first */
426 latency = policy->cpuinfo.transition_latency / 1000;
427 if (latency == 0)
428 latency = 1;
429
430 /* Bring kernel and HW constraints together */
431 dbs_data->min_sampling_rate = max(dbs_data->min_sampling_rate,
432 MIN_LATENCY_MULTIPLIER * latency);
Viresh Kumarff4b1782016-02-09 09:01:32 +0530433 dbs_data->sampling_rate = max(dbs_data->min_sampling_rate,
434 LATENCY_MULTIPLIER * latency);
Viresh Kumar714a2d92015-06-04 16:43:27 +0530435
Viresh Kumar8eec1022015-10-15 21:35:22 +0530436 if (!have_governor_per_policy())
Rafael J. Wysocki7bdad342016-02-07 16:05:07 +0100437 gov->gdbs_data = dbs_data;
Viresh Kumar714a2d92015-06-04 16:43:27 +0530438
Rafael J. Wysocki0dd3c1d2016-03-22 02:47:51 +0100439 policy_dbs->dbs_data = dbs_data;
Rafael J. Wysockibc505472016-02-07 16:24:26 +0100440 policy->governor_data = policy_dbs;
Viresh Kumare4b133c2016-01-25 22:33:46 +0530441
Viresh Kumarc4435632016-02-09 09:01:33 +0530442 gov->kobj_type.sysfs_ops = &governor_sysfs_ops;
Rafael J. Wysocki0dd3c1d2016-03-22 02:47:51 +0100443 ret = kobject_init_and_add(&dbs_data->attr_set.kobj, &gov->kobj_type,
Viresh Kumarc4435632016-02-09 09:01:33 +0530444 get_governor_parent_kobj(policy),
445 "%s", gov->gov.name);
Rafael J. Wysockifafd5e82016-02-08 23:57:22 +0100446 if (!ret)
Rafael J. Wysocki1112e9d2016-02-21 00:53:06 +0100447 goto out;
Viresh Kumar714a2d92015-06-04 16:43:27 +0530448
Rafael J. Wysockifafd5e82016-02-08 23:57:22 +0100449 /* Failure, so roll back. */
Viresh Kumar666f4cc2016-05-18 17:55:27 +0530450 pr_err("initialization failed (dbs_data kobject init error %d)\n", ret);
Viresh Kumar714a2d92015-06-04 16:43:27 +0530451
Viresh Kumare4b133c2016-01-25 22:33:46 +0530452 policy->governor_data = NULL;
453
Viresh Kumar8eec1022015-10-15 21:35:22 +0530454 if (!have_governor_per_policy())
Rafael J. Wysocki7bdad342016-02-07 16:05:07 +0100455 gov->gdbs_data = NULL;
Rafael J. Wysocki9a15fb22016-05-18 22:59:49 +0200456 gov->exit(dbs_data);
Rafael J. Wysockibc505472016-02-07 16:24:26 +0100457 kfree(dbs_data);
458
Rafael J. Wysockie40e7b22016-02-10 17:07:44 +0100459free_policy_dbs_info:
Rafael J. Wysocki8c8f77f2016-02-21 00:51:27 +0100460 free_policy_dbs_info(policy_dbs, gov);
Rafael J. Wysocki1112e9d2016-02-21 00:53:06 +0100461
462out:
463 mutex_unlock(&gov_dbs_data_mutex);
Viresh Kumar714a2d92015-06-04 16:43:27 +0530464 return ret;
465}
Rafael J. Wysockie7888922016-06-02 23:24:15 +0200466EXPORT_SYMBOL_GPL(cpufreq_dbs_governor_init);
Viresh Kumar714a2d92015-06-04 16:43:27 +0530467
Rafael J. Wysockie7888922016-06-02 23:24:15 +0200468void cpufreq_dbs_governor_exit(struct cpufreq_policy *policy)
Viresh Kumar714a2d92015-06-04 16:43:27 +0530469{
Rafael J. Wysockiea59ee0d2016-02-07 16:09:51 +0100470 struct dbs_governor *gov = dbs_governor_of(policy);
Rafael J. Wysockibc505472016-02-07 16:24:26 +0100471 struct policy_dbs_info *policy_dbs = policy->governor_data;
472 struct dbs_data *dbs_data = policy_dbs->dbs_data;
Rafael J. Wysocki0dd3c1d2016-03-22 02:47:51 +0100473 unsigned int count;
Viresh Kumara72c4952015-07-18 11:31:01 +0530474
Rafael J. Wysocki1112e9d2016-02-21 00:53:06 +0100475 /* Protect gov->gdbs_data against concurrent updates. */
476 mutex_lock(&gov_dbs_data_mutex);
477
Rafael J. Wysocki0dd3c1d2016-03-22 02:47:51 +0100478 count = gov_attr_set_put(&dbs_data->attr_set, &policy_dbs->list);
479
480 policy->governor_data = NULL;
Viresh Kumarc54df072016-02-10 11:00:25 +0530481
482 if (!count) {
Viresh Kumar8eec1022015-10-15 21:35:22 +0530483 if (!have_governor_per_policy())
Rafael J. Wysocki7bdad342016-02-07 16:05:07 +0100484 gov->gdbs_data = NULL;
Viresh Kumar714a2d92015-06-04 16:43:27 +0530485
Rafael J. Wysocki9a15fb22016-05-18 22:59:49 +0200486 gov->exit(dbs_data);
Viresh Kumar714a2d92015-06-04 16:43:27 +0530487 kfree(dbs_data);
488 }
Viresh Kumar44152cb2015-07-18 11:30:59 +0530489
Rafael J. Wysocki8c8f77f2016-02-21 00:51:27 +0100490 free_policy_dbs_info(policy_dbs, gov);
Rafael J. Wysocki1112e9d2016-02-21 00:53:06 +0100491
492 mutex_unlock(&gov_dbs_data_mutex);
Viresh Kumar714a2d92015-06-04 16:43:27 +0530493}
Rafael J. Wysockie7888922016-06-02 23:24:15 +0200494EXPORT_SYMBOL_GPL(cpufreq_dbs_governor_exit);
Viresh Kumar714a2d92015-06-04 16:43:27 +0530495
Rafael J. Wysockie7888922016-06-02 23:24:15 +0200496int cpufreq_dbs_governor_start(struct cpufreq_policy *policy)
Viresh Kumar714a2d92015-06-04 16:43:27 +0530497{
Rafael J. Wysockiea59ee0d2016-02-07 16:09:51 +0100498 struct dbs_governor *gov = dbs_governor_of(policy);
Rafael J. Wysockibc505472016-02-07 16:24:26 +0100499 struct policy_dbs_info *policy_dbs = policy->governor_data;
500 struct dbs_data *dbs_data = policy_dbs->dbs_data;
Rafael J. Wysocki702c9e52016-02-18 02:21:21 +0100501 unsigned int sampling_rate, ignore_nice, j;
Rafael J. Wysocki8847e032016-02-18 02:20:13 +0100502 unsigned int io_busy;
Viresh Kumar714a2d92015-06-04 16:43:27 +0530503
504 if (!policy->cur)
505 return -EINVAL;
506
Rafael J. Wysockie4db2812016-02-15 02:13:42 +0100507 policy_dbs->is_shared = policy_is_shared(policy);
Rafael J. Wysocki57dc3bc2016-02-15 02:20:51 +0100508 policy_dbs->rate_mult = 1;
Rafael J. Wysockie4db2812016-02-15 02:13:42 +0100509
Viresh Kumarff4b1782016-02-09 09:01:32 +0530510 sampling_rate = dbs_data->sampling_rate;
511 ignore_nice = dbs_data->ignore_nice_load;
Rafael J. Wysocki8847e032016-02-18 02:20:13 +0100512 io_busy = dbs_data->io_is_busy;
Viresh Kumar714a2d92015-06-04 16:43:27 +0530513
Viresh Kumar714a2d92015-06-04 16:43:27 +0530514 for_each_cpu(j, policy->cpus) {
Rafael J. Wysocki8c8f77f2016-02-21 00:51:27 +0100515 struct cpu_dbs_info *j_cdbs = &per_cpu(cpu_dbs, j);
Viresh Kumar714a2d92015-06-04 16:43:27 +0530516
Rafael J. Wysockib4f4b4b2016-04-28 01:19:03 +0200517 j_cdbs->prev_cpu_idle = get_cpu_idle_time(j, &j_cdbs->prev_update_time, io_busy);
Rafael J. Wysockiba1ca652016-04-25 16:21:34 +0200518 /*
519 * Make the first invocation of dbs_update() compute the load.
520 */
521 j_cdbs->prev_load = 0;
Viresh Kumar714a2d92015-06-04 16:43:27 +0530522
523 if (ignore_nice)
524 j_cdbs->prev_cpu_nice = kcpustat_cpu(j).cpustat[CPUTIME_NICE];
Viresh Kumar714a2d92015-06-04 16:43:27 +0530525 }
526
Rafael J. Wysocki702c9e52016-02-18 02:21:21 +0100527 gov->start(policy);
Viresh Kumar714a2d92015-06-04 16:43:27 +0530528
Rafael J. Wysockie40e7b22016-02-10 17:07:44 +0100529 gov_set_update_util(policy_dbs, sampling_rate);
Viresh Kumar714a2d92015-06-04 16:43:27 +0530530 return 0;
531}
Rafael J. Wysockie7888922016-06-02 23:24:15 +0200532EXPORT_SYMBOL_GPL(cpufreq_dbs_governor_start);
Viresh Kumar714a2d92015-06-04 16:43:27 +0530533
Rafael J. Wysockie7888922016-06-02 23:24:15 +0200534void cpufreq_dbs_governor_stop(struct cpufreq_policy *policy)
Viresh Kumar714a2d92015-06-04 16:43:27 +0530535{
Rafael J. Wysockif6709b82016-06-09 01:45:32 +0200536 struct policy_dbs_info *policy_dbs = policy->governor_data;
537
538 gov_clear_update_util(policy_dbs->policy);
539 irq_work_sync(&policy_dbs->irq_work);
540 cancel_work_sync(&policy_dbs->work);
541 atomic_set(&policy_dbs->work_count, 0);
542 policy_dbs->work_in_progress = false;
Viresh Kumar714a2d92015-06-04 16:43:27 +0530543}
Rafael J. Wysockie7888922016-06-02 23:24:15 +0200544EXPORT_SYMBOL_GPL(cpufreq_dbs_governor_stop);
Viresh Kumar714a2d92015-06-04 16:43:27 +0530545
Rafael J. Wysockie7888922016-06-02 23:24:15 +0200546void cpufreq_dbs_governor_limits(struct cpufreq_policy *policy)
Viresh Kumar714a2d92015-06-04 16:43:27 +0530547{
Rafael J. Wysockibc505472016-02-07 16:24:26 +0100548 struct policy_dbs_info *policy_dbs = policy->governor_data;
Viresh Kumar714a2d92015-06-04 16:43:27 +0530549
Rafael J. Wysockie9751892016-02-07 16:23:49 +0100550 mutex_lock(&policy_dbs->timer_mutex);
Viresh Kumarbf2be2d2016-05-18 17:55:31 +0530551 cpufreq_policy_apply_limits(policy);
Rafael J. Wysocki4cccf752016-02-15 02:19:31 +0100552 gov_update_sample_delay(policy_dbs, 0);
553
Rafael J. Wysockie9751892016-02-07 16:23:49 +0100554 mutex_unlock(&policy_dbs->timer_mutex);
Viresh Kumar714a2d92015-06-04 16:43:27 +0530555}
Rafael J. Wysockie7888922016-06-02 23:24:15 +0200556EXPORT_SYMBOL_GPL(cpufreq_dbs_governor_limits);