Commit | Line | Data |
---|---|---|
d2912cb1 | 1 | // SPDX-License-Identifier: GPL-2.0-only |
2aacdfff | 2 | /* |
3 | * drivers/cpufreq/cpufreq_governor.c | |
4 | * | |
5 | * CPUFREQ governors common code | |
6 | * | |
4471a34f VK |
7 | * Copyright (C) 2001 Russell King |
8 | * (C) 2003 Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>. | |
9 | * (C) 2003 Jun Nakajima <jun.nakajima@intel.com> | |
10 | * (C) 2009 Alexander Clouter <alex@digriz.org.uk> | |
11 | * (c) 2012 Viresh Kumar <viresh.kumar@linaro.org> | |
2aacdfff | 12 | */ |
13 | ||
4471a34f VK |
14 | #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt |
15 | ||
2aacdfff | 16 | #include <linux/export.h> |
17 | #include <linux/kernel_stat.h> | |
4d5dcc42 | 18 | #include <linux/slab.h> |
4471a34f VK |
19 | |
20 | #include "cpufreq_governor.h" | |
21 | ||
56026645 RW |
22 | #define CPUFREQ_DBS_MIN_SAMPLING_INTERVAL (2 * TICK_NSEC / NSEC_PER_USEC) |
23 | ||
8c8f77fd RW |
24 | static DEFINE_PER_CPU(struct cpu_dbs_info, cpu_dbs); |
25 | ||
1112e9d8 | 26 | static DEFINE_MUTEX(gov_dbs_data_mutex); |
2bb8d94f | 27 | |
aded387b | 28 | /* Common sysfs tunables */ |
dd2e65f2 | 29 | /* |
85750bcd | 30 | * sampling_rate_store - update sampling rate effective immediately if needed. |
aded387b VK |
31 | * |
32 | * If new rate is smaller than the old, simply updating | |
33 | * dbs.sampling_rate might not be appropriate. For example, if the | |
34 | * original sampling_rate was 1 second and the requested new sampling rate is 10 | |
35 | * ms because the user needs immediate reaction from ondemand governor, but not | |
36 | * sure if higher frequency will be required or not, then, the governor may | |
37 | * change the sampling rate too late; up to 1 second later. Thus, if we are | |
38 | * reducing the sampling rate, we need to make the new value effective | |
39 | * immediately. | |
40 | * | |
aded387b VK |
41 | * This must be called with dbs_data->mutex held, otherwise traversing |
42 | * policy_dbs_list isn't safe. | |
43 | */ | |
85750bcd | 44 | ssize_t sampling_rate_store(struct gov_attr_set *attr_set, const char *buf, |
aded387b VK |
45 | size_t count) |
46 | { | |
0dd3c1d6 | 47 | struct dbs_data *dbs_data = to_dbs_data(attr_set); |
aded387b | 48 | struct policy_dbs_info *policy_dbs; |
56026645 | 49 | unsigned int sampling_interval; |
aded387b | 50 | int ret; |
56026645 RW |
51 | |
52 | ret = sscanf(buf, "%u", &sampling_interval); | |
53 | if (ret != 1 || sampling_interval < CPUFREQ_DBS_MIN_SAMPLING_INTERVAL) | |
aded387b VK |
54 | return -EINVAL; |
55 | ||
56026645 RW |
56 | dbs_data->sampling_rate = sampling_interval; |
57 | ||
aded387b VK |
58 | /* |
59 | * We are operating under dbs_data->mutex and so the list and its | |
60 | * entries can't be freed concurrently. | |
61 | */ | |
0dd3c1d6 | 62 | list_for_each_entry(policy_dbs, &attr_set->policy_list, list) { |
26f0dbc9 | 63 | mutex_lock(&policy_dbs->update_mutex); |
aded387b VK |
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 | |
78347cdb | 70 | * completed) will take a sample. |
aded387b VK |
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 | |
78347cdb RW |
74 | * written, but it will be corrected next time a sample is |
75 | * taken, so it shouldn't be significant. | |
aded387b | 76 | */ |
78347cdb | 77 | gov_update_sample_delay(policy_dbs, 0); |
26f0dbc9 | 78 | mutex_unlock(&policy_dbs->update_mutex); |
aded387b VK |
79 | } |
80 | ||
81 | return count; | |
82 | } | |
85750bcd | 83 | EXPORT_SYMBOL_GPL(sampling_rate_store); |
aded387b | 84 | |
a33cce1c RW |
85 | /** |
86 | * gov_update_cpu_data - Update CPU load data. | |
a33cce1c RW |
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 | */ | |
8c8f77fd | 95 | void gov_update_cpu_data(struct dbs_data *dbs_data) |
a33cce1c RW |
96 | { |
97 | struct policy_dbs_info *policy_dbs; | |
98 | ||
0dd3c1d6 | 99 | list_for_each_entry(policy_dbs, &dbs_data->attr_set.policy_list, list) { |
a33cce1c RW |
100 | unsigned int j; |
101 | ||
102 | for_each_cpu(j, policy_dbs->policy->cpus) { | |
8c8f77fd | 103 | struct cpu_dbs_info *j_cdbs = &per_cpu(cpu_dbs, j); |
a33cce1c | 104 | |
b4f4b4b3 | 105 | j_cdbs->prev_cpu_idle = get_cpu_idle_time(j, &j_cdbs->prev_update_time, |
a33cce1c RW |
106 | dbs_data->io_is_busy); |
107 | if (dbs_data->ignore_nice_load) | |
5720821b | 108 | j_cdbs->prev_cpu_nice = kcpustat_field(&kcpustat_cpu(j), CPUTIME_NICE, j); |
a33cce1c RW |
109 | } |
110 | } | |
111 | } | |
112 | EXPORT_SYMBOL_GPL(gov_update_cpu_data); | |
113 | ||
4cccf755 | 114 | unsigned int dbs_update(struct cpufreq_policy *policy) |
4471a34f | 115 | { |
bc505475 RW |
116 | struct policy_dbs_info *policy_dbs = policy->governor_data; |
117 | struct dbs_data *dbs_data = policy_dbs->dbs_data; | |
ff4b1789 | 118 | unsigned int ignore_nice = dbs_data->ignore_nice_load; |
00bfe058 | 119 | unsigned int max_load = 0, idle_periods = UINT_MAX; |
8847e038 | 120 | unsigned int sampling_rate, io_busy, j; |
4471a34f | 121 | |
57dc3bcd RW |
122 | /* |
123 | * Sometimes governors may use an additional multiplier to increase | |
124 | * sample delays temporarily. Apply that multiplier to sampling_rate | |
125 | * so as to keep the wake-up-from-idle detection logic a bit | |
126 | * conservative. | |
127 | */ | |
128 | sampling_rate = dbs_data->sampling_rate * policy_dbs->rate_mult; | |
8847e038 RW |
129 | /* |
130 | * For the purpose of ondemand, waiting for disk IO is an indication | |
131 | * that you're performance critical, and not that the system is actually | |
132 | * idle, so do not add the iowait time to the CPU idle time then. | |
133 | */ | |
134 | io_busy = dbs_data->io_is_busy; | |
4471a34f | 135 | |
dfa5bb62 | 136 | /* Get Absolute Load */ |
4471a34f | 137 | for_each_cpu(j, policy->cpus) { |
8c8f77fd | 138 | struct cpu_dbs_info *j_cdbs = &per_cpu(cpu_dbs, j); |
b4f4b4b3 RW |
139 | u64 update_time, cur_idle_time; |
140 | unsigned int idle_time, time_elapsed; | |
4471a34f VK |
141 | unsigned int load; |
142 | ||
b4f4b4b3 | 143 | cur_idle_time = get_cpu_idle_time(j, &update_time, io_busy); |
4471a34f | 144 | |
b4f4b4b3 RW |
145 | time_elapsed = update_time - j_cdbs->prev_update_time; |
146 | j_cdbs->prev_update_time = update_time; | |
4471a34f | 147 | |
94862a62 RW |
148 | idle_time = cur_idle_time - j_cdbs->prev_cpu_idle; |
149 | j_cdbs->prev_cpu_idle = cur_idle_time; | |
4471a34f VK |
150 | |
151 | if (ignore_nice) { | |
5720821b | 152 | u64 cur_nice = kcpustat_field(&kcpustat_cpu(j), CPUTIME_NICE, j); |
679b8fe4 | 153 | |
7fb1327e | 154 | idle_time += div_u64(cur_nice - j_cdbs->prev_cpu_nice, NSEC_PER_USEC); |
679b8fe4 | 155 | j_cdbs->prev_cpu_nice = cur_nice; |
4471a34f VK |
156 | } |
157 | ||
9485e4ca RW |
158 | if (unlikely(!time_elapsed)) { |
159 | /* | |
160 | * That can only happen when this function is called | |
161 | * twice in a row with a very short interval between the | |
162 | * calls, so the previous load value can be used then. | |
163 | */ | |
18b46abd | 164 | load = j_cdbs->prev_load; |
75920196 | 165 | } else if (unlikely((int)idle_time > 2 * sampling_rate && |
9485e4ca | 166 | j_cdbs->prev_load)) { |
c8ae481b | 167 | /* |
9485e4ca RW |
168 | * If the CPU had gone completely idle and a task has |
169 | * just woken up on this CPU now, it would be unfair to | |
170 | * calculate 'load' the usual way for this elapsed | |
171 | * time-window, because it would show near-zero load, | |
172 | * irrespective of how CPU intensive that task actually | |
173 | * was. This is undesirable for latency-sensitive bursty | |
174 | * workloads. | |
175 | * | |
176 | * To avoid this, reuse the 'load' from the previous | |
177 | * time-window and give this task a chance to start with | |
178 | * a reasonably high CPU frequency. However, that | |
179 | * shouldn't be over-done, lest we get stuck at a high | |
180 | * load (high frequency) for too long, even when the | |
181 | * current system load has actually dropped down, so | |
182 | * clear prev_load to guarantee that the load will be | |
183 | * computed again next time. | |
184 | * | |
75920196 CY |
185 | * Detecting this situation is easy: an unusually large |
186 | * 'idle_time' (as compared to the sampling rate) | |
9485e4ca | 187 | * indicates this scenario. |
c8ae481b | 188 | */ |
9485e4ca | 189 | load = j_cdbs->prev_load; |
c8ae481b | 190 | j_cdbs->prev_load = 0; |
18b46abd | 191 | } else { |
9485e4ca RW |
192 | if (time_elapsed >= idle_time) { |
193 | load = 100 * (time_elapsed - idle_time) / time_elapsed; | |
194 | } else { | |
195 | /* | |
196 | * That can happen if idle_time is returned by | |
197 | * get_cpu_idle_time_jiffy(). In that case | |
198 | * idle_time is roughly equal to the difference | |
199 | * between time_elapsed and "busy time" obtained | |
200 | * from CPU statistics. Then, the "busy time" | |
201 | * can end up being greater than time_elapsed | |
202 | * (for example, if jiffies_64 and the CPU | |
203 | * statistics are updated by different CPUs), | |
204 | * so idle_time may in fact be negative. That | |
205 | * means, though, that the CPU was busy all | |
206 | * the time (on the rough average) during the | |
207 | * last sampling interval and 100 can be | |
208 | * returned as the load. | |
209 | */ | |
210 | load = (int)idle_time < 0 ? 100 : 0; | |
211 | } | |
18b46abd | 212 | j_cdbs->prev_load = load; |
18b46abd | 213 | } |
4471a34f | 214 | |
75920196 CY |
215 | if (unlikely((int)idle_time > 2 * sampling_rate)) { |
216 | unsigned int periods = idle_time / sampling_rate; | |
00bfe058 SK |
217 | |
218 | if (periods < idle_periods) | |
219 | idle_periods = periods; | |
220 | } | |
221 | ||
4471a34f VK |
222 | if (load > max_load) |
223 | max_load = load; | |
224 | } | |
00bfe058 SK |
225 | |
226 | policy_dbs->idle_periods = idle_periods; | |
227 | ||
4cccf755 | 228 | return max_load; |
4471a34f | 229 | } |
4cccf755 | 230 | EXPORT_SYMBOL_GPL(dbs_update); |
4471a34f | 231 | |
70f43e5e | 232 | static void dbs_work_handler(struct work_struct *work) |
43e0ee36 | 233 | { |
e40e7b25 | 234 | struct policy_dbs_info *policy_dbs; |
3a91b069 | 235 | struct cpufreq_policy *policy; |
ea59ee0d | 236 | struct dbs_governor *gov; |
43e0ee36 | 237 | |
e40e7b25 RW |
238 | policy_dbs = container_of(work, struct policy_dbs_info, work); |
239 | policy = policy_dbs->policy; | |
ea59ee0d | 240 | gov = dbs_governor_of(policy); |
3a91b069 | 241 | |
70f43e5e | 242 | /* |
9be4fd2c RW |
243 | * Make sure cpufreq_governor_limits() isn't evaluating load or the |
244 | * ondemand governor isn't updating the sampling rate in parallel. | |
70f43e5e | 245 | */ |
26f0dbc9 VK |
246 | mutex_lock(&policy_dbs->update_mutex); |
247 | gov_update_sample_delay(policy_dbs, gov->gov_dbs_update(policy)); | |
248 | mutex_unlock(&policy_dbs->update_mutex); | |
70f43e5e | 249 | |
e4db2813 RW |
250 | /* Allow the utilization update handler to queue up more work. */ |
251 | atomic_set(&policy_dbs->work_count, 0); | |
9be4fd2c | 252 | /* |
e4db2813 RW |
253 | * If the update below is reordered with respect to the sample delay |
254 | * modification, the utilization update handler may end up using a stale | |
255 | * sample delay value. | |
9be4fd2c | 256 | */ |
e4db2813 RW |
257 | smp_wmb(); |
258 | policy_dbs->work_in_progress = false; | |
9be4fd2c RW |
259 | } |
260 | ||
261 | static void dbs_irq_work(struct irq_work *irq_work) | |
262 | { | |
e40e7b25 | 263 | struct policy_dbs_info *policy_dbs; |
70f43e5e | 264 | |
e40e7b25 | 265 | policy_dbs = container_of(irq_work, struct policy_dbs_info, irq_work); |
539a4c42 | 266 | schedule_work_on(smp_processor_id(), &policy_dbs->work); |
70f43e5e VK |
267 | } |
268 | ||
9be4fd2c | 269 | static void dbs_update_util_handler(struct update_util_data *data, u64 time, |
58919e83 | 270 | unsigned int flags) |
9be4fd2c RW |
271 | { |
272 | struct cpu_dbs_info *cdbs = container_of(data, struct cpu_dbs_info, update_util); | |
e40e7b25 | 273 | struct policy_dbs_info *policy_dbs = cdbs->policy_dbs; |
27de3482 | 274 | u64 delta_ns, lst; |
70f43e5e | 275 | |
03639978 | 276 | if (!cpufreq_this_cpu_can_update(policy_dbs->policy)) |
674e7541 VK |
277 | return; |
278 | ||
70f43e5e | 279 | /* |
9be4fd2c RW |
280 | * The work may not be allowed to be queued up right now. |
281 | * Possible reasons: | |
282 | * - Work has already been queued up or is in progress. | |
9be4fd2c | 283 | * - It is too early (too little time from the previous sample). |
70f43e5e | 284 | */ |
e4db2813 RW |
285 | if (policy_dbs->work_in_progress) |
286 | return; | |
287 | ||
288 | /* | |
289 | * If the reads below are reordered before the check above, the value | |
290 | * of sample_delay_ns used in the computation may be stale. | |
291 | */ | |
292 | smp_rmb(); | |
27de3482 RW |
293 | lst = READ_ONCE(policy_dbs->last_sample_time); |
294 | delta_ns = time - lst; | |
e4db2813 RW |
295 | if ((s64)delta_ns < policy_dbs->sample_delay_ns) |
296 | return; | |
297 | ||
298 | /* | |
299 | * If the policy is not shared, the irq_work may be queued up right away | |
300 | * at this point. Otherwise, we need to ensure that only one of the | |
301 | * CPUs sharing the policy will do that. | |
302 | */ | |
27de3482 RW |
303 | if (policy_dbs->is_shared) { |
304 | if (!atomic_add_unless(&policy_dbs->work_count, 1, 1)) | |
305 | return; | |
306 | ||
307 | /* | |
308 | * If another CPU updated last_sample_time in the meantime, we | |
309 | * shouldn't be here, so clear the work counter and bail out. | |
310 | */ | |
311 | if (unlikely(lst != READ_ONCE(policy_dbs->last_sample_time))) { | |
312 | atomic_set(&policy_dbs->work_count, 0); | |
313 | return; | |
314 | } | |
315 | } | |
e4db2813 RW |
316 | |
317 | policy_dbs->last_sample_time = time; | |
318 | policy_dbs->work_in_progress = true; | |
319 | irq_work_queue(&policy_dbs->irq_work); | |
43e0ee36 | 320 | } |
4447266b | 321 | |
0bed612b RW |
322 | static void gov_set_update_util(struct policy_dbs_info *policy_dbs, |
323 | unsigned int delay_us) | |
324 | { | |
325 | struct cpufreq_policy *policy = policy_dbs->policy; | |
326 | int cpu; | |
327 | ||
328 | gov_update_sample_delay(policy_dbs, delay_us); | |
329 | policy_dbs->last_sample_time = 0; | |
330 | ||
331 | for_each_cpu(cpu, policy->cpus) { | |
332 | struct cpu_dbs_info *cdbs = &per_cpu(cpu_dbs, cpu); | |
333 | ||
334 | cpufreq_add_update_util_hook(cpu, &cdbs->update_util, | |
335 | dbs_update_util_handler); | |
336 | } | |
337 | } | |
338 | ||
339 | static inline void gov_clear_update_util(struct cpufreq_policy *policy) | |
340 | { | |
341 | int i; | |
342 | ||
343 | for_each_cpu(i, policy->cpus) | |
344 | cpufreq_remove_update_util_hook(i); | |
345 | ||
cc69b389 | 346 | synchronize_rcu(); |
0bed612b RW |
347 | } |
348 | ||
bc505475 RW |
349 | static struct policy_dbs_info *alloc_policy_dbs_info(struct cpufreq_policy *policy, |
350 | struct dbs_governor *gov) | |
44152cb8 | 351 | { |
e40e7b25 | 352 | struct policy_dbs_info *policy_dbs; |
44152cb8 VK |
353 | int j; |
354 | ||
7d5a9956 RW |
355 | /* Allocate memory for per-policy governor data. */ |
356 | policy_dbs = gov->alloc(); | |
e40e7b25 | 357 | if (!policy_dbs) |
bc505475 | 358 | return NULL; |
44152cb8 | 359 | |
581c214b | 360 | policy_dbs->policy = policy; |
26f0dbc9 | 361 | mutex_init(&policy_dbs->update_mutex); |
686cc637 | 362 | atomic_set(&policy_dbs->work_count, 0); |
e40e7b25 RW |
363 | init_irq_work(&policy_dbs->irq_work, dbs_irq_work); |
364 | INIT_WORK(&policy_dbs->work, dbs_work_handler); | |
cea6a9e7 RW |
365 | |
366 | /* Set policy_dbs for all CPUs, online+offline */ | |
367 | for_each_cpu(j, policy->related_cpus) { | |
8c8f77fd | 368 | struct cpu_dbs_info *j_cdbs = &per_cpu(cpu_dbs, j); |
cea6a9e7 RW |
369 | |
370 | j_cdbs->policy_dbs = policy_dbs; | |
cea6a9e7 | 371 | } |
bc505475 | 372 | return policy_dbs; |
44152cb8 VK |
373 | } |
374 | ||
8c8f77fd | 375 | static void free_policy_dbs_info(struct policy_dbs_info *policy_dbs, |
7bdad34d | 376 | struct dbs_governor *gov) |
44152cb8 | 377 | { |
44152cb8 VK |
378 | int j; |
379 | ||
26f0dbc9 | 380 | mutex_destroy(&policy_dbs->update_mutex); |
5e4500d8 | 381 | |
8c8f77fd RW |
382 | for_each_cpu(j, policy_dbs->policy->related_cpus) { |
383 | struct cpu_dbs_info *j_cdbs = &per_cpu(cpu_dbs, j); | |
44152cb8 | 384 | |
cea6a9e7 RW |
385 | j_cdbs->policy_dbs = NULL; |
386 | j_cdbs->update_util.func = NULL; | |
387 | } | |
7d5a9956 | 388 | gov->free(policy_dbs); |
44152cb8 VK |
389 | } |
390 | ||
e788892b | 391 | int cpufreq_dbs_governor_init(struct cpufreq_policy *policy) |
4471a34f | 392 | { |
ea59ee0d | 393 | struct dbs_governor *gov = dbs_governor_of(policy); |
1112e9d8 | 394 | struct dbs_data *dbs_data; |
bc505475 | 395 | struct policy_dbs_info *policy_dbs; |
1112e9d8 | 396 | int ret = 0; |
4471a34f | 397 | |
a72c4959 VK |
398 | /* State should be equivalent to EXIT */ |
399 | if (policy->governor_data) | |
400 | return -EBUSY; | |
401 | ||
bc505475 RW |
402 | policy_dbs = alloc_policy_dbs_info(policy, gov); |
403 | if (!policy_dbs) | |
404 | return -ENOMEM; | |
44152cb8 | 405 | |
1112e9d8 RW |
406 | /* Protect gov->gdbs_data against concurrent updates. */ |
407 | mutex_lock(&gov_dbs_data_mutex); | |
408 | ||
409 | dbs_data = gov->gdbs_data; | |
bc505475 RW |
410 | if (dbs_data) { |
411 | if (WARN_ON(have_governor_per_policy())) { | |
412 | ret = -EINVAL; | |
413 | goto free_policy_dbs_info; | |
414 | } | |
bc505475 RW |
415 | policy_dbs->dbs_data = dbs_data; |
416 | policy->governor_data = policy_dbs; | |
c54df071 | 417 | |
0dd3c1d6 | 418 | gov_attr_set_get(&dbs_data->attr_set, &policy_dbs->list); |
1112e9d8 | 419 | goto out; |
714a2d9c | 420 | } |
4d5dcc42 | 421 | |
714a2d9c | 422 | dbs_data = kzalloc(sizeof(*dbs_data), GFP_KERNEL); |
bc505475 RW |
423 | if (!dbs_data) { |
424 | ret = -ENOMEM; | |
425 | goto free_policy_dbs_info; | |
426 | } | |
44152cb8 | 427 | |
0dd3c1d6 | 428 | gov_attr_set_init(&dbs_data->attr_set, &policy_dbs->list); |
4d5dcc42 | 429 | |
9a15fb2c | 430 | ret = gov->init(dbs_data); |
714a2d9c | 431 | if (ret) |
e40e7b25 | 432 | goto free_policy_dbs_info; |
4d5dcc42 | 433 | |
56026645 RW |
434 | /* |
435 | * The sampling interval should not be less than the transition latency | |
436 | * of the CPU and it also cannot be too small for dbs_update() to work | |
437 | * correctly. | |
438 | */ | |
439 | dbs_data->sampling_rate = max_t(unsigned int, | |
440 | CPUFREQ_DBS_MIN_SAMPLING_INTERVAL, | |
441 | cpufreq_policy_transition_delay_us(policy)); | |
2361be23 | 442 | |
8eec1020 | 443 | if (!have_governor_per_policy()) |
7bdad34d | 444 | gov->gdbs_data = dbs_data; |
4d5dcc42 | 445 | |
c54df071 | 446 | policy_dbs->dbs_data = dbs_data; |
0dd3c1d6 | 447 | policy->governor_data = policy_dbs; |
c54df071 | 448 | |
c4435630 | 449 | gov->kobj_type.sysfs_ops = &governor_sysfs_ops; |
0dd3c1d6 | 450 | ret = kobject_init_and_add(&dbs_data->attr_set.kobj, &gov->kobj_type, |
c4435630 VK |
451 | get_governor_parent_kobj(policy), |
452 | "%s", gov->gov.name); | |
fafd5e8a | 453 | if (!ret) |
1112e9d8 | 454 | goto out; |
4d5dcc42 | 455 | |
fafd5e8a | 456 | /* Failure, so roll back. */ |
666f4ccc | 457 | pr_err("initialization failed (dbs_data kobject init error %d)\n", ret); |
4d5dcc42 | 458 | |
4ebe36c9 VK |
459 | kobject_put(&dbs_data->attr_set.kobj); |
460 | ||
e4b133cc VK |
461 | policy->governor_data = NULL; |
462 | ||
8eec1020 | 463 | if (!have_governor_per_policy()) |
7bdad34d | 464 | gov->gdbs_data = NULL; |
9a15fb2c | 465 | gov->exit(dbs_data); |
bc505475 RW |
466 | kfree(dbs_data); |
467 | ||
e40e7b25 | 468 | free_policy_dbs_info: |
8c8f77fd | 469 | free_policy_dbs_info(policy_dbs, gov); |
1112e9d8 RW |
470 | |
471 | out: | |
472 | mutex_unlock(&gov_dbs_data_mutex); | |
714a2d9c VK |
473 | return ret; |
474 | } | |
e788892b | 475 | EXPORT_SYMBOL_GPL(cpufreq_dbs_governor_init); |
4d5dcc42 | 476 | |
e788892b | 477 | void cpufreq_dbs_governor_exit(struct cpufreq_policy *policy) |
714a2d9c | 478 | { |
ea59ee0d | 479 | struct dbs_governor *gov = dbs_governor_of(policy); |
bc505475 RW |
480 | struct policy_dbs_info *policy_dbs = policy->governor_data; |
481 | struct dbs_data *dbs_data = policy_dbs->dbs_data; | |
0dd3c1d6 | 482 | unsigned int count; |
a72c4959 | 483 | |
1112e9d8 RW |
484 | /* Protect gov->gdbs_data against concurrent updates. */ |
485 | mutex_lock(&gov_dbs_data_mutex); | |
486 | ||
0dd3c1d6 | 487 | count = gov_attr_set_put(&dbs_data->attr_set, &policy_dbs->list); |
2361be23 | 488 | |
0dd3c1d6 | 489 | policy->governor_data = NULL; |
e4b133cc | 490 | |
0dd3c1d6 | 491 | if (!count) { |
8eec1020 | 492 | if (!have_governor_per_policy()) |
7bdad34d | 493 | gov->gdbs_data = NULL; |
4471a34f | 494 | |
9a15fb2c | 495 | gov->exit(dbs_data); |
714a2d9c | 496 | kfree(dbs_data); |
4d5dcc42 | 497 | } |
44152cb8 | 498 | |
8c8f77fd | 499 | free_policy_dbs_info(policy_dbs, gov); |
1112e9d8 RW |
500 | |
501 | mutex_unlock(&gov_dbs_data_mutex); | |
714a2d9c | 502 | } |
e788892b | 503 | EXPORT_SYMBOL_GPL(cpufreq_dbs_governor_exit); |
4d5dcc42 | 504 | |
e788892b | 505 | int cpufreq_dbs_governor_start(struct cpufreq_policy *policy) |
714a2d9c | 506 | { |
ea59ee0d | 507 | struct dbs_governor *gov = dbs_governor_of(policy); |
bc505475 RW |
508 | struct policy_dbs_info *policy_dbs = policy->governor_data; |
509 | struct dbs_data *dbs_data = policy_dbs->dbs_data; | |
702c9e54 | 510 | unsigned int sampling_rate, ignore_nice, j; |
8847e038 | 511 | unsigned int io_busy; |
714a2d9c VK |
512 | |
513 | if (!policy->cur) | |
514 | return -EINVAL; | |
515 | ||
e4db2813 | 516 | policy_dbs->is_shared = policy_is_shared(policy); |
57dc3bcd | 517 | policy_dbs->rate_mult = 1; |
e4db2813 | 518 | |
ff4b1789 VK |
519 | sampling_rate = dbs_data->sampling_rate; |
520 | ignore_nice = dbs_data->ignore_nice_load; | |
8847e038 | 521 | io_busy = dbs_data->io_is_busy; |
4471a34f | 522 | |
714a2d9c | 523 | for_each_cpu(j, policy->cpus) { |
8c8f77fd | 524 | struct cpu_dbs_info *j_cdbs = &per_cpu(cpu_dbs, j); |
4471a34f | 525 | |
b4f4b4b3 | 526 | j_cdbs->prev_cpu_idle = get_cpu_idle_time(j, &j_cdbs->prev_update_time, io_busy); |
ba1ca654 RW |
527 | /* |
528 | * Make the first invocation of dbs_update() compute the load. | |
529 | */ | |
530 | j_cdbs->prev_load = 0; | |
18b46abd | 531 | |
714a2d9c | 532 | if (ignore_nice) |
5720821b | 533 | j_cdbs->prev_cpu_nice = kcpustat_field(&kcpustat_cpu(j), CPUTIME_NICE, j); |
714a2d9c | 534 | } |
2abfa876 | 535 | |
702c9e54 | 536 | gov->start(policy); |
4471a34f | 537 | |
e40e7b25 | 538 | gov_set_update_util(policy_dbs, sampling_rate); |
714a2d9c VK |
539 | return 0; |
540 | } | |
e788892b | 541 | EXPORT_SYMBOL_GPL(cpufreq_dbs_governor_start); |
714a2d9c | 542 | |
e788892b | 543 | void cpufreq_dbs_governor_stop(struct cpufreq_policy *policy) |
714a2d9c | 544 | { |
f6709b8a RW |
545 | struct policy_dbs_info *policy_dbs = policy->governor_data; |
546 | ||
547 | gov_clear_update_util(policy_dbs->policy); | |
548 | irq_work_sync(&policy_dbs->irq_work); | |
549 | cancel_work_sync(&policy_dbs->work); | |
550 | atomic_set(&policy_dbs->work_count, 0); | |
551 | policy_dbs->work_in_progress = false; | |
714a2d9c | 552 | } |
e788892b | 553 | EXPORT_SYMBOL_GPL(cpufreq_dbs_governor_stop); |
4471a34f | 554 | |
e788892b | 555 | void cpufreq_dbs_governor_limits(struct cpufreq_policy *policy) |
714a2d9c | 556 | { |
2a3eb51e HW |
557 | struct policy_dbs_info *policy_dbs; |
558 | ||
559 | /* Protect gov->gdbs_data against cpufreq_dbs_governor_exit() */ | |
560 | mutex_lock(&gov_dbs_data_mutex); | |
561 | policy_dbs = policy->governor_data; | |
562 | if (!policy_dbs) | |
563 | goto out; | |
8eeed095 | 564 | |
26f0dbc9 | 565 | mutex_lock(&policy_dbs->update_mutex); |
bf2be2de | 566 | cpufreq_policy_apply_limits(policy); |
4cccf755 | 567 | gov_update_sample_delay(policy_dbs, 0); |
26f0dbc9 | 568 | mutex_unlock(&policy_dbs->update_mutex); |
2a3eb51e HW |
569 | |
570 | out: | |
571 | mutex_unlock(&gov_dbs_data_mutex); | |
4471a34f | 572 | } |
e788892b | 573 | EXPORT_SYMBOL_GPL(cpufreq_dbs_governor_limits); |