10a5cfeae8c5e34317805402c63d0c11e3aaac5e
[cascardo/linux.git] / drivers / cpufreq / cpufreq_governor.c
1 /*
2  * drivers/cpufreq/cpufreq_governor.c
3  *
4  * CPUFREQ governors common code
5  *
6  * 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  *
12  * 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
17 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
18
19 #include <linux/export.h>
20 #include <linux/kernel_stat.h>
21 #include <linux/sched.h>
22 #include <linux/slab.h>
23
24 #include "cpufreq_governor.h"
25
26 static DEFINE_PER_CPU(struct cpu_dbs_info, cpu_dbs);
27
28 static DEFINE_MUTEX(gov_dbs_data_mutex);
29
30 /* 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  *
43  * This must be called with dbs_data->mutex held, otherwise traversing
44  * policy_dbs_list isn't safe.
45  */
46 ssize_t store_sampling_rate(struct dbs_data *dbs_data, const char *buf,
47                             size_t count)
48 {
49         struct policy_dbs_info *policy_dbs;
50         unsigned int rate;
51         int ret;
52         ret = sscanf(buf, "%u", &rate);
53         if (ret != 1)
54                 return -EINVAL;
55
56         dbs_data->sampling_rate = max(rate, dbs_data->min_sampling_rate);
57
58         /*
59          * We are operating under dbs_data->mutex and so the list and its
60          * entries can't be freed concurrently.
61          */
62         list_for_each_entry(policy_dbs, &dbs_data->policy_dbs_list, list) {
63                 mutex_lock(&policy_dbs->timer_mutex);
64                 /*
65                  * On 32-bit architectures this may race with the
66                  * sample_delay_ns read in dbs_update_util_handler(), but that
67                  * really doesn't matter.  If the read returns a value that's
68                  * too big, the sample will be skipped, but the next invocation
69                  * of dbs_update_util_handler() (when the update has been
70                  * completed) will take a sample.
71                  *
72                  * If this runs in parallel with dbs_work_handler(), we may end
73                  * up overwriting the sample_delay_ns value that it has just
74                  * written, but it will be corrected next time a sample is
75                  * taken, so it shouldn't be significant.
76                  */
77                 gov_update_sample_delay(policy_dbs, 0);
78                 mutex_unlock(&policy_dbs->timer_mutex);
79         }
80
81         return count;
82 }
83 EXPORT_SYMBOL_GPL(store_sampling_rate);
84
85 /**
86  * gov_update_cpu_data - Update CPU load data.
87  * @dbs_data: Top-level governor data pointer.
88  *
89  * Update CPU load data for all CPUs in the domain governed by @dbs_data
90  * (that may be a single policy or a bunch of them if governor tunables are
91  * system-wide).
92  *
93  * Call under the @dbs_data mutex.
94  */
95 void gov_update_cpu_data(struct dbs_data *dbs_data)
96 {
97         struct policy_dbs_info *policy_dbs;
98
99         list_for_each_entry(policy_dbs, &dbs_data->policy_dbs_list, list) {
100                 unsigned int j;
101
102                 for_each_cpu(j, policy_dbs->policy->cpus) {
103                         struct cpu_dbs_info *j_cdbs = &per_cpu(cpu_dbs, j);
104
105                         j_cdbs->prev_cpu_idle = get_cpu_idle_time(j, &j_cdbs->prev_cpu_wall,
106                                                                   dbs_data->io_is_busy);
107                         if (dbs_data->ignore_nice_load)
108                                 j_cdbs->prev_cpu_nice = kcpustat_cpu(j).cpustat[CPUTIME_NICE];
109                 }
110         }
111 }
112 EXPORT_SYMBOL_GPL(gov_update_cpu_data);
113
114 static inline struct dbs_data *to_dbs_data(struct kobject *kobj)
115 {
116         return container_of(kobj, struct dbs_data, kobj);
117 }
118
119 static inline struct governor_attr *to_gov_attr(struct attribute *attr)
120 {
121         return container_of(attr, struct governor_attr, attr);
122 }
123
124 static ssize_t governor_show(struct kobject *kobj, struct attribute *attr,
125                              char *buf)
126 {
127         struct dbs_data *dbs_data = to_dbs_data(kobj);
128         struct governor_attr *gattr = to_gov_attr(attr);
129
130         return gattr->show(dbs_data, buf);
131 }
132
133 static ssize_t governor_store(struct kobject *kobj, struct attribute *attr,
134                               const char *buf, size_t count)
135 {
136         struct dbs_data *dbs_data = to_dbs_data(kobj);
137         struct governor_attr *gattr = to_gov_attr(attr);
138         int ret = -EBUSY;
139
140         mutex_lock(&dbs_data->mutex);
141
142         if (dbs_data->usage_count)
143                 ret = gattr->store(dbs_data, buf, count);
144
145         mutex_unlock(&dbs_data->mutex);
146
147         return ret;
148 }
149
150 /*
151  * Sysfs Ops for accessing governor attributes.
152  *
153  * All show/store invocations for governor specific sysfs attributes, will first
154  * call the below show/store callbacks and the attribute specific callback will
155  * be called from within it.
156  */
157 static const struct sysfs_ops governor_sysfs_ops = {
158         .show   = governor_show,
159         .store  = governor_store,
160 };
161
162 unsigned int dbs_update(struct cpufreq_policy *policy)
163 {
164         struct policy_dbs_info *policy_dbs = policy->governor_data;
165         struct dbs_data *dbs_data = policy_dbs->dbs_data;
166         unsigned int ignore_nice = dbs_data->ignore_nice_load;
167         unsigned int max_load = 0;
168         unsigned int sampling_rate, io_busy, j;
169
170         /*
171          * Sometimes governors may use an additional multiplier to increase
172          * sample delays temporarily.  Apply that multiplier to sampling_rate
173          * so as to keep the wake-up-from-idle detection logic a bit
174          * conservative.
175          */
176         sampling_rate = dbs_data->sampling_rate * policy_dbs->rate_mult;
177         /*
178          * For the purpose of ondemand, waiting for disk IO is an indication
179          * that you're performance critical, and not that the system is actually
180          * idle, so do not add the iowait time to the CPU idle time then.
181          */
182         io_busy = dbs_data->io_is_busy;
183
184         /* Get Absolute Load */
185         for_each_cpu(j, policy->cpus) {
186                 struct cpu_dbs_info *j_cdbs = &per_cpu(cpu_dbs, j);
187                 u64 cur_wall_time, cur_idle_time;
188                 unsigned int idle_time, wall_time;
189                 unsigned int load;
190
191                 cur_idle_time = get_cpu_idle_time(j, &cur_wall_time, io_busy);
192
193                 wall_time = cur_wall_time - j_cdbs->prev_cpu_wall;
194                 j_cdbs->prev_cpu_wall = cur_wall_time;
195
196                 if (cur_idle_time <= j_cdbs->prev_cpu_idle) {
197                         idle_time = 0;
198                 } else {
199                         idle_time = cur_idle_time - j_cdbs->prev_cpu_idle;
200                         j_cdbs->prev_cpu_idle = cur_idle_time;
201                 }
202
203                 if (ignore_nice) {
204                         u64 cur_nice = kcpustat_cpu(j).cpustat[CPUTIME_NICE];
205
206                         idle_time += cputime_to_usecs(cur_nice - j_cdbs->prev_cpu_nice);
207                         j_cdbs->prev_cpu_nice = cur_nice;
208                 }
209
210                 if (unlikely(!wall_time || wall_time < idle_time))
211                         continue;
212
213                 /*
214                  * If the CPU had gone completely idle, and a task just woke up
215                  * on this CPU now, it would be unfair to calculate 'load' the
216                  * usual way for this elapsed time-window, because it will show
217                  * near-zero load, irrespective of how CPU intensive that task
218                  * actually is. This is undesirable for latency-sensitive bursty
219                  * workloads.
220                  *
221                  * To avoid this, we reuse the 'load' from the previous
222                  * time-window and give this task a chance to start with a
223                  * reasonably high CPU frequency. (However, we shouldn't over-do
224                  * this copy, lest we get stuck at a high load (high frequency)
225                  * for too long, even when the current system load has actually
226                  * dropped down. So we perform the copy only once, upon the
227                  * first wake-up from idle.)
228                  *
229                  * Detecting this situation is easy: the governor's utilization
230                  * update handler would not have run during CPU-idle periods.
231                  * Hence, an unusually large 'wall_time' (as compared to the
232                  * sampling rate) indicates this scenario.
233                  *
234                  * prev_load can be zero in two cases and we must recalculate it
235                  * for both cases:
236                  * - during long idle intervals
237                  * - explicitly set to zero
238                  */
239                 if (unlikely(wall_time > (2 * sampling_rate) &&
240                              j_cdbs->prev_load)) {
241                         load = j_cdbs->prev_load;
242
243                         /*
244                          * Perform a destructive copy, to ensure that we copy
245                          * the previous load only once, upon the first wake-up
246                          * from idle.
247                          */
248                         j_cdbs->prev_load = 0;
249                 } else {
250                         load = 100 * (wall_time - idle_time) / wall_time;
251                         j_cdbs->prev_load = load;
252                 }
253
254                 if (load > max_load)
255                         max_load = load;
256         }
257         return max_load;
258 }
259 EXPORT_SYMBOL_GPL(dbs_update);
260
261 static void gov_set_update_util(struct policy_dbs_info *policy_dbs,
262                                 unsigned int delay_us)
263 {
264         struct cpufreq_policy *policy = policy_dbs->policy;
265         int cpu;
266
267         gov_update_sample_delay(policy_dbs, delay_us);
268         policy_dbs->last_sample_time = 0;
269
270         for_each_cpu(cpu, policy->cpus) {
271                 struct cpu_dbs_info *cdbs = &per_cpu(cpu_dbs, cpu);
272
273                 cpufreq_set_update_util_data(cpu, &cdbs->update_util);
274         }
275 }
276
277 static inline void gov_clear_update_util(struct cpufreq_policy *policy)
278 {
279         int i;
280
281         for_each_cpu(i, policy->cpus)
282                 cpufreq_set_update_util_data(i, NULL);
283
284         synchronize_sched();
285 }
286
287 static void gov_cancel_work(struct cpufreq_policy *policy)
288 {
289         struct policy_dbs_info *policy_dbs = policy->governor_data;
290
291         gov_clear_update_util(policy_dbs->policy);
292         irq_work_sync(&policy_dbs->irq_work);
293         cancel_work_sync(&policy_dbs->work);
294         atomic_set(&policy_dbs->work_count, 0);
295         policy_dbs->work_in_progress = false;
296 }
297
298 static void dbs_work_handler(struct work_struct *work)
299 {
300         struct policy_dbs_info *policy_dbs;
301         struct cpufreq_policy *policy;
302         struct dbs_governor *gov;
303
304         policy_dbs = container_of(work, struct policy_dbs_info, work);
305         policy = policy_dbs->policy;
306         gov = dbs_governor_of(policy);
307
308         /*
309          * Make sure cpufreq_governor_limits() isn't evaluating load or the
310          * ondemand governor isn't updating the sampling rate in parallel.
311          */
312         mutex_lock(&policy_dbs->timer_mutex);
313         gov_update_sample_delay(policy_dbs, gov->gov_dbs_timer(policy));
314         mutex_unlock(&policy_dbs->timer_mutex);
315
316         /* Allow the utilization update handler to queue up more work. */
317         atomic_set(&policy_dbs->work_count, 0);
318         /*
319          * If the update below is reordered with respect to the sample delay
320          * modification, the utilization update handler may end up using a stale
321          * sample delay value.
322          */
323         smp_wmb();
324         policy_dbs->work_in_progress = false;
325 }
326
327 static void dbs_irq_work(struct irq_work *irq_work)
328 {
329         struct policy_dbs_info *policy_dbs;
330
331         policy_dbs = container_of(irq_work, struct policy_dbs_info, irq_work);
332         schedule_work_on(smp_processor_id(), &policy_dbs->work);
333 }
334
335 static void dbs_update_util_handler(struct update_util_data *data, u64 time,
336                                     unsigned long util, unsigned long max)
337 {
338         struct cpu_dbs_info *cdbs = container_of(data, struct cpu_dbs_info, update_util);
339         struct policy_dbs_info *policy_dbs = cdbs->policy_dbs;
340         u64 delta_ns, lst;
341
342         /*
343          * The work may not be allowed to be queued up right now.
344          * Possible reasons:
345          * - Work has already been queued up or is in progress.
346          * - It is too early (too little time from the previous sample).
347          */
348         if (policy_dbs->work_in_progress)
349                 return;
350
351         /*
352          * If the reads below are reordered before the check above, the value
353          * of sample_delay_ns used in the computation may be stale.
354          */
355         smp_rmb();
356         lst = READ_ONCE(policy_dbs->last_sample_time);
357         delta_ns = time - lst;
358         if ((s64)delta_ns < policy_dbs->sample_delay_ns)
359                 return;
360
361         /*
362          * If the policy is not shared, the irq_work may be queued up right away
363          * at this point.  Otherwise, we need to ensure that only one of the
364          * CPUs sharing the policy will do that.
365          */
366         if (policy_dbs->is_shared) {
367                 if (!atomic_add_unless(&policy_dbs->work_count, 1, 1))
368                         return;
369
370                 /*
371                  * If another CPU updated last_sample_time in the meantime, we
372                  * shouldn't be here, so clear the work counter and bail out.
373                  */
374                 if (unlikely(lst != READ_ONCE(policy_dbs->last_sample_time))) {
375                         atomic_set(&policy_dbs->work_count, 0);
376                         return;
377                 }
378         }
379
380         policy_dbs->last_sample_time = time;
381         policy_dbs->work_in_progress = true;
382         irq_work_queue(&policy_dbs->irq_work);
383 }
384
385 static struct policy_dbs_info *alloc_policy_dbs_info(struct cpufreq_policy *policy,
386                                                      struct dbs_governor *gov)
387 {
388         struct policy_dbs_info *policy_dbs;
389         int j;
390
391         /* Allocate memory for per-policy governor data. */
392         policy_dbs = gov->alloc();
393         if (!policy_dbs)
394                 return NULL;
395
396         policy_dbs->policy = policy;
397         mutex_init(&policy_dbs->timer_mutex);
398         atomic_set(&policy_dbs->work_count, 0);
399         init_irq_work(&policy_dbs->irq_work, dbs_irq_work);
400         INIT_WORK(&policy_dbs->work, dbs_work_handler);
401
402         /* Set policy_dbs for all CPUs, online+offline */
403         for_each_cpu(j, policy->related_cpus) {
404                 struct cpu_dbs_info *j_cdbs = &per_cpu(cpu_dbs, j);
405
406                 j_cdbs->policy_dbs = policy_dbs;
407                 j_cdbs->update_util.func = dbs_update_util_handler;
408         }
409         return policy_dbs;
410 }
411
412 static void free_policy_dbs_info(struct policy_dbs_info *policy_dbs,
413                                  struct dbs_governor *gov)
414 {
415         int j;
416
417         mutex_destroy(&policy_dbs->timer_mutex);
418
419         for_each_cpu(j, policy_dbs->policy->related_cpus) {
420                 struct cpu_dbs_info *j_cdbs = &per_cpu(cpu_dbs, j);
421
422                 j_cdbs->policy_dbs = NULL;
423                 j_cdbs->update_util.func = NULL;
424         }
425         gov->free(policy_dbs);
426 }
427
428 static int cpufreq_governor_init(struct cpufreq_policy *policy)
429 {
430         struct dbs_governor *gov = dbs_governor_of(policy);
431         struct dbs_data *dbs_data;
432         struct policy_dbs_info *policy_dbs;
433         unsigned int latency;
434         int ret = 0;
435
436         /* State should be equivalent to EXIT */
437         if (policy->governor_data)
438                 return -EBUSY;
439
440         policy_dbs = alloc_policy_dbs_info(policy, gov);
441         if (!policy_dbs)
442                 return -ENOMEM;
443
444         /* Protect gov->gdbs_data against concurrent updates. */
445         mutex_lock(&gov_dbs_data_mutex);
446
447         dbs_data = gov->gdbs_data;
448         if (dbs_data) {
449                 if (WARN_ON(have_governor_per_policy())) {
450                         ret = -EINVAL;
451                         goto free_policy_dbs_info;
452                 }
453                 policy_dbs->dbs_data = dbs_data;
454                 policy->governor_data = policy_dbs;
455
456                 mutex_lock(&dbs_data->mutex);
457                 dbs_data->usage_count++;
458                 list_add(&policy_dbs->list, &dbs_data->policy_dbs_list);
459                 mutex_unlock(&dbs_data->mutex);
460                 goto out;
461         }
462
463         dbs_data = kzalloc(sizeof(*dbs_data), GFP_KERNEL);
464         if (!dbs_data) {
465                 ret = -ENOMEM;
466                 goto free_policy_dbs_info;
467         }
468
469         INIT_LIST_HEAD(&dbs_data->policy_dbs_list);
470         mutex_init(&dbs_data->mutex);
471
472         ret = gov->init(dbs_data, !policy->governor->initialized);
473         if (ret)
474                 goto free_policy_dbs_info;
475
476         /* policy latency is in ns. Convert it to us first */
477         latency = policy->cpuinfo.transition_latency / 1000;
478         if (latency == 0)
479                 latency = 1;
480
481         /* Bring kernel and HW constraints together */
482         dbs_data->min_sampling_rate = max(dbs_data->min_sampling_rate,
483                                           MIN_LATENCY_MULTIPLIER * latency);
484         dbs_data->sampling_rate = max(dbs_data->min_sampling_rate,
485                                       LATENCY_MULTIPLIER * latency);
486
487         if (!have_governor_per_policy())
488                 gov->gdbs_data = dbs_data;
489
490         policy->governor_data = policy_dbs;
491
492         policy_dbs->dbs_data = dbs_data;
493         dbs_data->usage_count = 1;
494         list_add(&policy_dbs->list, &dbs_data->policy_dbs_list);
495
496         gov->kobj_type.sysfs_ops = &governor_sysfs_ops;
497         ret = kobject_init_and_add(&dbs_data->kobj, &gov->kobj_type,
498                                    get_governor_parent_kobj(policy),
499                                    "%s", gov->gov.name);
500         if (!ret)
501                 goto out;
502
503         /* Failure, so roll back. */
504         pr_err("cpufreq: Governor initialization failed (dbs_data kobject init error %d)\n", ret);
505
506         policy->governor_data = NULL;
507
508         if (!have_governor_per_policy())
509                 gov->gdbs_data = NULL;
510         gov->exit(dbs_data, !policy->governor->initialized);
511         kfree(dbs_data);
512
513 free_policy_dbs_info:
514         free_policy_dbs_info(policy_dbs, gov);
515
516 out:
517         mutex_unlock(&gov_dbs_data_mutex);
518         return ret;
519 }
520
521 static int cpufreq_governor_exit(struct cpufreq_policy *policy)
522 {
523         struct dbs_governor *gov = dbs_governor_of(policy);
524         struct policy_dbs_info *policy_dbs = policy->governor_data;
525         struct dbs_data *dbs_data = policy_dbs->dbs_data;
526         int count;
527
528         /* Protect gov->gdbs_data against concurrent updates. */
529         mutex_lock(&gov_dbs_data_mutex);
530
531         mutex_lock(&dbs_data->mutex);
532         list_del(&policy_dbs->list);
533         count = --dbs_data->usage_count;
534         mutex_unlock(&dbs_data->mutex);
535
536         if (!count) {
537                 kobject_put(&dbs_data->kobj);
538
539                 policy->governor_data = NULL;
540
541                 if (!have_governor_per_policy())
542                         gov->gdbs_data = NULL;
543
544                 gov->exit(dbs_data, policy->governor->initialized == 1);
545                 mutex_destroy(&dbs_data->mutex);
546                 kfree(dbs_data);
547         } else {
548                 policy->governor_data = NULL;
549         }
550
551         free_policy_dbs_info(policy_dbs, gov);
552
553         mutex_unlock(&gov_dbs_data_mutex);
554         return 0;
555 }
556
557 static int cpufreq_governor_start(struct cpufreq_policy *policy)
558 {
559         struct dbs_governor *gov = dbs_governor_of(policy);
560         struct policy_dbs_info *policy_dbs = policy->governor_data;
561         struct dbs_data *dbs_data = policy_dbs->dbs_data;
562         unsigned int sampling_rate, ignore_nice, j;
563         unsigned int io_busy;
564
565         if (!policy->cur)
566                 return -EINVAL;
567
568         policy_dbs->is_shared = policy_is_shared(policy);
569         policy_dbs->rate_mult = 1;
570
571         sampling_rate = dbs_data->sampling_rate;
572         ignore_nice = dbs_data->ignore_nice_load;
573         io_busy = dbs_data->io_is_busy;
574
575         for_each_cpu(j, policy->cpus) {
576                 struct cpu_dbs_info *j_cdbs = &per_cpu(cpu_dbs, j);
577                 unsigned int prev_load;
578
579                 j_cdbs->prev_cpu_idle = get_cpu_idle_time(j, &j_cdbs->prev_cpu_wall, io_busy);
580
581                 prev_load = j_cdbs->prev_cpu_wall - j_cdbs->prev_cpu_idle;
582                 j_cdbs->prev_load = 100 * prev_load / (unsigned int)j_cdbs->prev_cpu_wall;
583
584                 if (ignore_nice)
585                         j_cdbs->prev_cpu_nice = kcpustat_cpu(j).cpustat[CPUTIME_NICE];
586         }
587
588         gov->start(policy);
589
590         gov_set_update_util(policy_dbs, sampling_rate);
591         return 0;
592 }
593
594 static int cpufreq_governor_stop(struct cpufreq_policy *policy)
595 {
596         gov_cancel_work(policy);
597         return 0;
598 }
599
600 static int cpufreq_governor_limits(struct cpufreq_policy *policy)
601 {
602         struct policy_dbs_info *policy_dbs = policy->governor_data;
603
604         mutex_lock(&policy_dbs->timer_mutex);
605
606         if (policy->max < policy->cur)
607                 __cpufreq_driver_target(policy, policy->max, CPUFREQ_RELATION_H);
608         else if (policy->min > policy->cur)
609                 __cpufreq_driver_target(policy, policy->min, CPUFREQ_RELATION_L);
610
611         gov_update_sample_delay(policy_dbs, 0);
612
613         mutex_unlock(&policy_dbs->timer_mutex);
614
615         return 0;
616 }
617
618 int cpufreq_governor_dbs(struct cpufreq_policy *policy, unsigned int event)
619 {
620         if (event == CPUFREQ_GOV_POLICY_INIT) {
621                 return cpufreq_governor_init(policy);
622         } else if (policy->governor_data) {
623                 switch (event) {
624                 case CPUFREQ_GOV_POLICY_EXIT:
625                         return cpufreq_governor_exit(policy);
626                 case CPUFREQ_GOV_START:
627                         return cpufreq_governor_start(policy);
628                 case CPUFREQ_GOV_STOP:
629                         return cpufreq_governor_stop(policy);
630                 case CPUFREQ_GOV_LIMITS:
631                         return cpufreq_governor_limits(policy);
632                 }
633         }
634         return -EINVAL;
635 }
636 EXPORT_SYMBOL_GPL(cpufreq_governor_dbs);