Mike Chan | ef96969 | 2010-06-22 11:26:45 -0700 | [diff] [blame] | 1 | /* |
| 2 | * drivers/cpufreq/cpufreq_interactive.c |
| 3 | * |
| 4 | * Copyright (C) 2010 Google, Inc. |
| 5 | * |
| 6 | * This software is licensed under the terms of the GNU General Public |
| 7 | * License version 2, as published by the Free Software Foundation, and |
| 8 | * may be copied, distributed, and modified under those terms. |
| 9 | * |
| 10 | * This program is distributed in the hope that it will be useful, |
| 11 | * but WITHOUT ANY WARRANTY; without even the implied warranty of |
| 12 | * MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the |
| 13 | * GNU General Public License for more details. |
| 14 | * |
| 15 | * Author: Mike Chan (mike@android.com) |
| 16 | * |
| 17 | */ |
| 18 | |
| 19 | #include <linux/cpu.h> |
| 20 | #include <linux/cpumask.h> |
| 21 | #include <linux/cpufreq.h> |
| 22 | #include <linux/module.h> |
| 23 | #include <linux/mutex.h> |
| 24 | #include <linux/sched.h> |
| 25 | #include <linux/sched/rt.h> |
| 26 | #include <linux/tick.h> |
| 27 | #include <linux/time.h> |
| 28 | #include <linux/timer.h> |
| 29 | #include <linux/workqueue.h> |
| 30 | #include <linux/kthread.h> |
| 31 | #include <linux/mutex.h> |
| 32 | |
Todd Poynor | ae01047 | 2012-02-16 16:27:59 -0800 | [diff] [blame] | 33 | #define CREATE_TRACE_POINTS |
| 34 | #include <trace/events/cpufreq_interactive.h> |
| 35 | |
Mike Chan | ef96969 | 2010-06-22 11:26:45 -0700 | [diff] [blame] | 36 | #include <asm/cputime.h> |
| 37 | |
| 38 | static atomic_t active_count = ATOMIC_INIT(0); |
| 39 | |
| 40 | struct cpufreq_interactive_cpuinfo { |
| 41 | struct timer_list cpu_timer; |
| 42 | int timer_idlecancel; |
| 43 | u64 time_in_idle; |
| 44 | u64 idle_exit_time; |
| 45 | u64 timer_run_time; |
| 46 | int idling; |
Todd Poynor | 1f408dc | 2012-04-06 19:59:36 -0700 | [diff] [blame] | 47 | u64 target_set_time; |
| 48 | u64 target_set_time_in_idle; |
Mike Chan | ef96969 | 2010-06-22 11:26:45 -0700 | [diff] [blame] | 49 | struct cpufreq_policy *policy; |
| 50 | struct cpufreq_frequency_table *freq_table; |
| 51 | unsigned int target_freq; |
| 52 | int governor_enabled; |
| 53 | }; |
| 54 | |
| 55 | static DEFINE_PER_CPU(struct cpufreq_interactive_cpuinfo, cpuinfo); |
| 56 | |
| 57 | /* Workqueues handle frequency scaling */ |
| 58 | static struct task_struct *up_task; |
| 59 | static struct workqueue_struct *down_wq; |
| 60 | static struct work_struct freq_scale_down_work; |
| 61 | static cpumask_t up_cpumask; |
| 62 | static spinlock_t up_cpumask_lock; |
| 63 | static cpumask_t down_cpumask; |
| 64 | static spinlock_t down_cpumask_lock; |
| 65 | static struct mutex set_speed_lock; |
| 66 | |
| 67 | /* Hi speed to bump to from lo speed when load burst (default max) */ |
| 68 | static u64 hispeed_freq; |
| 69 | |
| 70 | /* Go to hi speed when CPU load at or above this value. */ |
Todd Poynor | a380aa8 | 2012-04-17 17:39:34 -0700 | [diff] [blame^] | 71 | #define DEFAULT_GO_HISPEED_LOAD 85 |
Mike Chan | ef96969 | 2010-06-22 11:26:45 -0700 | [diff] [blame] | 72 | static unsigned long go_hispeed_load; |
| 73 | |
| 74 | /* |
| 75 | * The minimum amount of time to spend at a frequency before we can ramp down. |
| 76 | */ |
Todd Poynor | a380aa8 | 2012-04-17 17:39:34 -0700 | [diff] [blame^] | 77 | #define DEFAULT_MIN_SAMPLE_TIME (80 * USEC_PER_MSEC) |
Mike Chan | ef96969 | 2010-06-22 11:26:45 -0700 | [diff] [blame] | 78 | static unsigned long min_sample_time; |
| 79 | |
| 80 | /* |
| 81 | * The sample rate of the timer used to increase frequency |
| 82 | */ |
Todd Poynor | a380aa8 | 2012-04-17 17:39:34 -0700 | [diff] [blame^] | 83 | #define DEFAULT_TIMER_RATE (20 * USEC_PER_MSEC) |
Mike Chan | ef96969 | 2010-06-22 11:26:45 -0700 | [diff] [blame] | 84 | static unsigned long timer_rate; |
| 85 | |
Todd Poynor | cbbe17d | 2012-04-13 20:18:02 -0700 | [diff] [blame] | 86 | /* |
| 87 | * Wait this long before raising speed above hispeed, by default a single |
| 88 | * timer interval. |
| 89 | */ |
| 90 | #define DEFAULT_ABOVE_HISPEED_DELAY DEFAULT_TIMER_RATE |
| 91 | static unsigned long above_hispeed_delay_val; |
| 92 | |
Mike Chan | ef96969 | 2010-06-22 11:26:45 -0700 | [diff] [blame] | 93 | static int cpufreq_governor_interactive(struct cpufreq_policy *policy, |
| 94 | unsigned int event); |
| 95 | |
| 96 | #ifndef CONFIG_CPU_FREQ_DEFAULT_GOV_INTERACTIVE |
| 97 | static |
| 98 | #endif |
| 99 | struct cpufreq_governor cpufreq_gov_interactive = { |
| 100 | .name = "interactive", |
| 101 | .governor = cpufreq_governor_interactive, |
| 102 | .max_transition_latency = 10000000, |
| 103 | .owner = THIS_MODULE, |
| 104 | }; |
| 105 | |
| 106 | static void cpufreq_interactive_timer(unsigned long data) |
| 107 | { |
| 108 | unsigned int delta_idle; |
| 109 | unsigned int delta_time; |
| 110 | int cpu_load; |
| 111 | int load_since_change; |
| 112 | u64 time_in_idle; |
| 113 | u64 idle_exit_time; |
| 114 | struct cpufreq_interactive_cpuinfo *pcpu = |
| 115 | &per_cpu(cpuinfo, data); |
| 116 | u64 now_idle; |
| 117 | unsigned int new_freq; |
| 118 | unsigned int index; |
| 119 | unsigned long flags; |
| 120 | |
| 121 | smp_rmb(); |
| 122 | |
| 123 | if (!pcpu->governor_enabled) |
| 124 | goto exit; |
| 125 | |
| 126 | /* |
| 127 | * Once pcpu->timer_run_time is updated to >= pcpu->idle_exit_time, |
| 128 | * this lets idle exit know the current idle time sample has |
| 129 | * been processed, and idle exit can generate a new sample and |
| 130 | * re-arm the timer. This prevents a concurrent idle |
| 131 | * exit on that CPU from writing a new set of info at the same time |
| 132 | * the timer function runs (the timer function can't use that info |
| 133 | * until more time passes). |
| 134 | */ |
| 135 | time_in_idle = pcpu->time_in_idle; |
| 136 | idle_exit_time = pcpu->idle_exit_time; |
| 137 | now_idle = get_cpu_idle_time_us(data, &pcpu->timer_run_time); |
| 138 | smp_wmb(); |
| 139 | |
| 140 | /* If we raced with cancelling a timer, skip. */ |
| 141 | if (!idle_exit_time) |
| 142 | goto exit; |
| 143 | |
| 144 | delta_idle = (unsigned int)(now_idle - time_in_idle); |
| 145 | delta_time = (unsigned int)(pcpu->timer_run_time - idle_exit_time); |
| 146 | |
| 147 | /* |
| 148 | * If timer ran less than 1ms after short-term sample started, retry. |
| 149 | */ |
| 150 | if (delta_time < 1000) |
| 151 | goto rearm; |
| 152 | |
| 153 | if (delta_idle > delta_time) |
| 154 | cpu_load = 0; |
| 155 | else |
| 156 | cpu_load = 100 * (delta_time - delta_idle) / delta_time; |
| 157 | |
Todd Poynor | 1f408dc | 2012-04-06 19:59:36 -0700 | [diff] [blame] | 158 | delta_idle = (unsigned int)(now_idle - pcpu->target_set_time_in_idle); |
| 159 | delta_time = (unsigned int)(pcpu->timer_run_time - |
| 160 | pcpu->target_set_time); |
Mike Chan | ef96969 | 2010-06-22 11:26:45 -0700 | [diff] [blame] | 161 | |
| 162 | if ((delta_time == 0) || (delta_idle > delta_time)) |
| 163 | load_since_change = 0; |
| 164 | else |
| 165 | load_since_change = |
| 166 | 100 * (delta_time - delta_idle) / delta_time; |
| 167 | |
| 168 | /* |
| 169 | * Choose greater of short-term load (since last idle timer |
| 170 | * started or timer function re-armed itself) or long-term load |
| 171 | * (since last frequency change). |
| 172 | */ |
| 173 | if (load_since_change > cpu_load) |
| 174 | cpu_load = load_since_change; |
| 175 | |
| 176 | if (cpu_load >= go_hispeed_load) { |
Todd Poynor | f87b9d5 | 2012-04-06 19:50:12 -0700 | [diff] [blame] | 177 | if (pcpu->policy->cur == pcpu->policy->min) { |
Mike Chan | ef96969 | 2010-06-22 11:26:45 -0700 | [diff] [blame] | 178 | new_freq = hispeed_freq; |
Todd Poynor | f87b9d5 | 2012-04-06 19:50:12 -0700 | [diff] [blame] | 179 | } else { |
Mike Chan | ef96969 | 2010-06-22 11:26:45 -0700 | [diff] [blame] | 180 | new_freq = pcpu->policy->max * cpu_load / 100; |
Todd Poynor | f87b9d5 | 2012-04-06 19:50:12 -0700 | [diff] [blame] | 181 | |
| 182 | if (new_freq < hispeed_freq) |
| 183 | new_freq = hispeed_freq; |
Todd Poynor | cbbe17d | 2012-04-13 20:18:02 -0700 | [diff] [blame] | 184 | |
| 185 | if (pcpu->target_freq == hispeed_freq && |
| 186 | new_freq > hispeed_freq && |
| 187 | pcpu->timer_run_time - pcpu->target_set_time |
| 188 | < above_hispeed_delay_val) { |
| 189 | trace_cpufreq_interactive_notyet(data, cpu_load, |
| 190 | pcpu->target_freq, |
| 191 | new_freq); |
| 192 | goto rearm; |
| 193 | } |
Todd Poynor | f87b9d5 | 2012-04-06 19:50:12 -0700 | [diff] [blame] | 194 | } |
Mike Chan | ef96969 | 2010-06-22 11:26:45 -0700 | [diff] [blame] | 195 | } else { |
Todd Poynor | c1128fc | 2012-04-06 01:13:09 -0700 | [diff] [blame] | 196 | new_freq = pcpu->policy->max * cpu_load / 100; |
Mike Chan | ef96969 | 2010-06-22 11:26:45 -0700 | [diff] [blame] | 197 | } |
| 198 | |
| 199 | if (cpufreq_frequency_table_target(pcpu->policy, pcpu->freq_table, |
| 200 | new_freq, CPUFREQ_RELATION_H, |
| 201 | &index)) { |
| 202 | pr_warn_once("timer %d: cpufreq_frequency_table_target error\n", |
| 203 | (int) data); |
| 204 | goto rearm; |
| 205 | } |
| 206 | |
| 207 | new_freq = pcpu->freq_table[index].frequency; |
| 208 | |
Mike Chan | ef96969 | 2010-06-22 11:26:45 -0700 | [diff] [blame] | 209 | /* |
| 210 | * Do not scale down unless we have been at this frequency for the |
| 211 | * minimum sample time. |
| 212 | */ |
| 213 | if (new_freq < pcpu->target_freq) { |
Todd Poynor | 1f408dc | 2012-04-06 19:59:36 -0700 | [diff] [blame] | 214 | if (pcpu->timer_run_time - pcpu->target_set_time |
Todd Poynor | ae01047 | 2012-02-16 16:27:59 -0800 | [diff] [blame] | 215 | < min_sample_time) { |
| 216 | trace_cpufreq_interactive_notyet(data, cpu_load, |
| 217 | pcpu->target_freq, new_freq); |
Mike Chan | ef96969 | 2010-06-22 11:26:45 -0700 | [diff] [blame] | 218 | goto rearm; |
Todd Poynor | ae01047 | 2012-02-16 16:27:59 -0800 | [diff] [blame] | 219 | } |
Mike Chan | ef96969 | 2010-06-22 11:26:45 -0700 | [diff] [blame] | 220 | } |
| 221 | |
Todd Poynor | 1f408dc | 2012-04-06 19:59:36 -0700 | [diff] [blame] | 222 | pcpu->target_set_time_in_idle = now_idle; |
| 223 | pcpu->target_set_time = pcpu->timer_run_time; |
| 224 | |
| 225 | if (pcpu->target_freq == new_freq) { |
| 226 | trace_cpufreq_interactive_already(data, cpu_load, |
| 227 | pcpu->target_freq, new_freq); |
| 228 | goto rearm_if_notmax; |
| 229 | } |
| 230 | |
Todd Poynor | ae01047 | 2012-02-16 16:27:59 -0800 | [diff] [blame] | 231 | trace_cpufreq_interactive_target(data, cpu_load, pcpu->target_freq, |
| 232 | new_freq); |
| 233 | |
Mike Chan | ef96969 | 2010-06-22 11:26:45 -0700 | [diff] [blame] | 234 | if (new_freq < pcpu->target_freq) { |
| 235 | pcpu->target_freq = new_freq; |
| 236 | spin_lock_irqsave(&down_cpumask_lock, flags); |
| 237 | cpumask_set_cpu(data, &down_cpumask); |
| 238 | spin_unlock_irqrestore(&down_cpumask_lock, flags); |
| 239 | queue_work(down_wq, &freq_scale_down_work); |
| 240 | } else { |
| 241 | pcpu->target_freq = new_freq; |
| 242 | spin_lock_irqsave(&up_cpumask_lock, flags); |
| 243 | cpumask_set_cpu(data, &up_cpumask); |
| 244 | spin_unlock_irqrestore(&up_cpumask_lock, flags); |
| 245 | wake_up_process(up_task); |
| 246 | } |
| 247 | |
| 248 | rearm_if_notmax: |
| 249 | /* |
| 250 | * Already set max speed and don't see a need to change that, |
| 251 | * wait until next idle to re-evaluate, don't need timer. |
| 252 | */ |
| 253 | if (pcpu->target_freq == pcpu->policy->max) |
| 254 | goto exit; |
| 255 | |
| 256 | rearm: |
| 257 | if (!timer_pending(&pcpu->cpu_timer)) { |
| 258 | /* |
| 259 | * If already at min: if that CPU is idle, don't set timer. |
| 260 | * Else cancel the timer if that CPU goes idle. We don't |
| 261 | * need to re-evaluate speed until the next idle exit. |
| 262 | */ |
| 263 | if (pcpu->target_freq == pcpu->policy->min) { |
| 264 | smp_rmb(); |
| 265 | |
| 266 | if (pcpu->idling) |
| 267 | goto exit; |
| 268 | |
| 269 | pcpu->timer_idlecancel = 1; |
| 270 | } |
| 271 | |
| 272 | pcpu->time_in_idle = get_cpu_idle_time_us( |
| 273 | data, &pcpu->idle_exit_time); |
| 274 | mod_timer(&pcpu->cpu_timer, |
| 275 | jiffies + usecs_to_jiffies(timer_rate)); |
| 276 | } |
| 277 | |
| 278 | exit: |
| 279 | return; |
| 280 | } |
| 281 | |
| 282 | static void cpufreq_interactive_idle_start(void) |
| 283 | { |
| 284 | struct cpufreq_interactive_cpuinfo *pcpu = |
| 285 | &per_cpu(cpuinfo, smp_processor_id()); |
| 286 | int pending; |
| 287 | |
| 288 | if (!pcpu->governor_enabled) |
| 289 | return; |
| 290 | |
| 291 | pcpu->idling = 1; |
| 292 | smp_wmb(); |
| 293 | pending = timer_pending(&pcpu->cpu_timer); |
| 294 | |
| 295 | if (pcpu->target_freq != pcpu->policy->min) { |
| 296 | #ifdef CONFIG_SMP |
| 297 | /* |
| 298 | * Entering idle while not at lowest speed. On some |
| 299 | * platforms this can hold the other CPU(s) at that speed |
| 300 | * even though the CPU is idle. Set a timer to re-evaluate |
| 301 | * speed so this idle CPU doesn't hold the other CPUs above |
| 302 | * min indefinitely. This should probably be a quirk of |
| 303 | * the CPUFreq driver. |
| 304 | */ |
| 305 | if (!pending) { |
| 306 | pcpu->time_in_idle = get_cpu_idle_time_us( |
| 307 | smp_processor_id(), &pcpu->idle_exit_time); |
| 308 | pcpu->timer_idlecancel = 0; |
| 309 | mod_timer(&pcpu->cpu_timer, |
| 310 | jiffies + usecs_to_jiffies(timer_rate)); |
| 311 | } |
| 312 | #endif |
| 313 | } else { |
| 314 | /* |
| 315 | * If at min speed and entering idle after load has |
| 316 | * already been evaluated, and a timer has been set just in |
| 317 | * case the CPU suddenly goes busy, cancel that timer. The |
| 318 | * CPU didn't go busy; we'll recheck things upon idle exit. |
| 319 | */ |
| 320 | if (pending && pcpu->timer_idlecancel) { |
| 321 | del_timer(&pcpu->cpu_timer); |
| 322 | /* |
| 323 | * Ensure last timer run time is after current idle |
| 324 | * sample start time, so next idle exit will always |
| 325 | * start a new idle sampling period. |
| 326 | */ |
| 327 | pcpu->idle_exit_time = 0; |
| 328 | pcpu->timer_idlecancel = 0; |
| 329 | } |
| 330 | } |
| 331 | |
| 332 | } |
| 333 | |
| 334 | static void cpufreq_interactive_idle_end(void) |
| 335 | { |
| 336 | struct cpufreq_interactive_cpuinfo *pcpu = |
| 337 | &per_cpu(cpuinfo, smp_processor_id()); |
| 338 | |
| 339 | pcpu->idling = 0; |
| 340 | smp_wmb(); |
| 341 | |
| 342 | /* |
| 343 | * Arm the timer for 1-2 ticks later if not already, and if the timer |
| 344 | * function has already processed the previous load sampling |
| 345 | * interval. (If the timer is not pending but has not processed |
| 346 | * the previous interval, it is probably racing with us on another |
| 347 | * CPU. Let it compute load based on the previous sample and then |
| 348 | * re-arm the timer for another interval when it's done, rather |
| 349 | * than updating the interval start time to be "now", which doesn't |
| 350 | * give the timer function enough time to make a decision on this |
| 351 | * run.) |
| 352 | */ |
| 353 | if (timer_pending(&pcpu->cpu_timer) == 0 && |
| 354 | pcpu->timer_run_time >= pcpu->idle_exit_time && |
| 355 | pcpu->governor_enabled) { |
| 356 | pcpu->time_in_idle = |
| 357 | get_cpu_idle_time_us(smp_processor_id(), |
| 358 | &pcpu->idle_exit_time); |
| 359 | pcpu->timer_idlecancel = 0; |
| 360 | mod_timer(&pcpu->cpu_timer, |
| 361 | jiffies + usecs_to_jiffies(timer_rate)); |
| 362 | } |
| 363 | |
| 364 | } |
| 365 | |
| 366 | static int cpufreq_interactive_up_task(void *data) |
| 367 | { |
| 368 | unsigned int cpu; |
| 369 | cpumask_t tmp_mask; |
| 370 | unsigned long flags; |
| 371 | struct cpufreq_interactive_cpuinfo *pcpu; |
| 372 | |
| 373 | while (1) { |
| 374 | set_current_state(TASK_INTERRUPTIBLE); |
| 375 | spin_lock_irqsave(&up_cpumask_lock, flags); |
| 376 | |
| 377 | if (cpumask_empty(&up_cpumask)) { |
| 378 | spin_unlock_irqrestore(&up_cpumask_lock, flags); |
| 379 | schedule(); |
| 380 | |
| 381 | if (kthread_should_stop()) |
| 382 | break; |
| 383 | |
| 384 | spin_lock_irqsave(&up_cpumask_lock, flags); |
| 385 | } |
| 386 | |
| 387 | set_current_state(TASK_RUNNING); |
| 388 | tmp_mask = up_cpumask; |
| 389 | cpumask_clear(&up_cpumask); |
| 390 | spin_unlock_irqrestore(&up_cpumask_lock, flags); |
| 391 | |
| 392 | for_each_cpu(cpu, &tmp_mask) { |
| 393 | unsigned int j; |
| 394 | unsigned int max_freq = 0; |
| 395 | |
| 396 | pcpu = &per_cpu(cpuinfo, cpu); |
| 397 | smp_rmb(); |
| 398 | |
| 399 | if (!pcpu->governor_enabled) |
| 400 | continue; |
| 401 | |
| 402 | mutex_lock(&set_speed_lock); |
| 403 | |
| 404 | for_each_cpu(j, pcpu->policy->cpus) { |
| 405 | struct cpufreq_interactive_cpuinfo *pjcpu = |
| 406 | &per_cpu(cpuinfo, j); |
| 407 | |
| 408 | if (pjcpu->target_freq > max_freq) |
| 409 | max_freq = pjcpu->target_freq; |
| 410 | } |
| 411 | |
| 412 | if (max_freq != pcpu->policy->cur) |
| 413 | __cpufreq_driver_target(pcpu->policy, |
| 414 | max_freq, |
| 415 | CPUFREQ_RELATION_H); |
| 416 | mutex_unlock(&set_speed_lock); |
Todd Poynor | ae01047 | 2012-02-16 16:27:59 -0800 | [diff] [blame] | 417 | trace_cpufreq_interactive_up(cpu, pcpu->target_freq, |
| 418 | pcpu->policy->cur); |
Mike Chan | ef96969 | 2010-06-22 11:26:45 -0700 | [diff] [blame] | 419 | } |
| 420 | } |
| 421 | |
| 422 | return 0; |
| 423 | } |
| 424 | |
| 425 | static void cpufreq_interactive_freq_down(struct work_struct *work) |
| 426 | { |
| 427 | unsigned int cpu; |
| 428 | cpumask_t tmp_mask; |
| 429 | unsigned long flags; |
| 430 | struct cpufreq_interactive_cpuinfo *pcpu; |
| 431 | |
| 432 | spin_lock_irqsave(&down_cpumask_lock, flags); |
| 433 | tmp_mask = down_cpumask; |
| 434 | cpumask_clear(&down_cpumask); |
| 435 | spin_unlock_irqrestore(&down_cpumask_lock, flags); |
| 436 | |
| 437 | for_each_cpu(cpu, &tmp_mask) { |
| 438 | unsigned int j; |
| 439 | unsigned int max_freq = 0; |
| 440 | |
| 441 | pcpu = &per_cpu(cpuinfo, cpu); |
| 442 | smp_rmb(); |
| 443 | |
| 444 | if (!pcpu->governor_enabled) |
| 445 | continue; |
| 446 | |
| 447 | mutex_lock(&set_speed_lock); |
| 448 | |
| 449 | for_each_cpu(j, pcpu->policy->cpus) { |
| 450 | struct cpufreq_interactive_cpuinfo *pjcpu = |
| 451 | &per_cpu(cpuinfo, j); |
| 452 | |
| 453 | if (pjcpu->target_freq > max_freq) |
| 454 | max_freq = pjcpu->target_freq; |
| 455 | } |
| 456 | |
| 457 | if (max_freq != pcpu->policy->cur) |
| 458 | __cpufreq_driver_target(pcpu->policy, max_freq, |
| 459 | CPUFREQ_RELATION_H); |
| 460 | |
| 461 | mutex_unlock(&set_speed_lock); |
Todd Poynor | ae01047 | 2012-02-16 16:27:59 -0800 | [diff] [blame] | 462 | trace_cpufreq_interactive_down(cpu, pcpu->target_freq, |
| 463 | pcpu->policy->cur); |
Mike Chan | ef96969 | 2010-06-22 11:26:45 -0700 | [diff] [blame] | 464 | } |
| 465 | } |
| 466 | |
| 467 | static ssize_t show_hispeed_freq(struct kobject *kobj, |
| 468 | struct attribute *attr, char *buf) |
| 469 | { |
| 470 | return sprintf(buf, "%llu\n", hispeed_freq); |
| 471 | } |
| 472 | |
| 473 | static ssize_t store_hispeed_freq(struct kobject *kobj, |
| 474 | struct attribute *attr, const char *buf, |
| 475 | size_t count) |
| 476 | { |
| 477 | int ret; |
| 478 | u64 val; |
| 479 | |
| 480 | ret = strict_strtoull(buf, 0, &val); |
| 481 | if (ret < 0) |
| 482 | return ret; |
| 483 | hispeed_freq = val; |
| 484 | return count; |
| 485 | } |
| 486 | |
| 487 | static struct global_attr hispeed_freq_attr = __ATTR(hispeed_freq, 0644, |
| 488 | show_hispeed_freq, store_hispeed_freq); |
| 489 | |
| 490 | |
| 491 | static ssize_t show_go_hispeed_load(struct kobject *kobj, |
| 492 | struct attribute *attr, char *buf) |
| 493 | { |
| 494 | return sprintf(buf, "%lu\n", go_hispeed_load); |
| 495 | } |
| 496 | |
| 497 | static ssize_t store_go_hispeed_load(struct kobject *kobj, |
| 498 | struct attribute *attr, const char *buf, size_t count) |
| 499 | { |
| 500 | int ret; |
| 501 | unsigned long val; |
| 502 | |
| 503 | ret = strict_strtoul(buf, 0, &val); |
| 504 | if (ret < 0) |
| 505 | return ret; |
| 506 | go_hispeed_load = val; |
| 507 | return count; |
| 508 | } |
| 509 | |
| 510 | static struct global_attr go_hispeed_load_attr = __ATTR(go_hispeed_load, 0644, |
| 511 | show_go_hispeed_load, store_go_hispeed_load); |
| 512 | |
| 513 | static ssize_t show_min_sample_time(struct kobject *kobj, |
| 514 | struct attribute *attr, char *buf) |
| 515 | { |
| 516 | return sprintf(buf, "%lu\n", min_sample_time); |
| 517 | } |
| 518 | |
| 519 | static ssize_t store_min_sample_time(struct kobject *kobj, |
| 520 | struct attribute *attr, const char *buf, size_t count) |
| 521 | { |
| 522 | int ret; |
| 523 | unsigned long val; |
| 524 | |
| 525 | ret = strict_strtoul(buf, 0, &val); |
| 526 | if (ret < 0) |
| 527 | return ret; |
| 528 | min_sample_time = val; |
| 529 | return count; |
| 530 | } |
| 531 | |
| 532 | static struct global_attr min_sample_time_attr = __ATTR(min_sample_time, 0644, |
| 533 | show_min_sample_time, store_min_sample_time); |
| 534 | |
Todd Poynor | cbbe17d | 2012-04-13 20:18:02 -0700 | [diff] [blame] | 535 | static ssize_t show_above_hispeed_delay(struct kobject *kobj, |
| 536 | struct attribute *attr, char *buf) |
| 537 | { |
| 538 | return sprintf(buf, "%lu\n", above_hispeed_delay_val); |
| 539 | } |
| 540 | |
| 541 | static ssize_t store_above_hispeed_delay(struct kobject *kobj, |
| 542 | struct attribute *attr, |
| 543 | const char *buf, size_t count) |
| 544 | { |
| 545 | int ret; |
| 546 | unsigned long val; |
| 547 | |
| 548 | ret = strict_strtoul(buf, 0, &val); |
| 549 | if (ret < 0) |
| 550 | return ret; |
| 551 | above_hispeed_delay_val = val; |
| 552 | return count; |
| 553 | } |
| 554 | |
| 555 | define_one_global_rw(above_hispeed_delay); |
| 556 | |
Mike Chan | ef96969 | 2010-06-22 11:26:45 -0700 | [diff] [blame] | 557 | static ssize_t show_timer_rate(struct kobject *kobj, |
| 558 | struct attribute *attr, char *buf) |
| 559 | { |
| 560 | return sprintf(buf, "%lu\n", timer_rate); |
| 561 | } |
| 562 | |
| 563 | static ssize_t store_timer_rate(struct kobject *kobj, |
| 564 | struct attribute *attr, const char *buf, size_t count) |
| 565 | { |
| 566 | int ret; |
| 567 | unsigned long val; |
| 568 | |
| 569 | ret = strict_strtoul(buf, 0, &val); |
| 570 | if (ret < 0) |
| 571 | return ret; |
| 572 | timer_rate = val; |
| 573 | return count; |
| 574 | } |
| 575 | |
| 576 | static struct global_attr timer_rate_attr = __ATTR(timer_rate, 0644, |
| 577 | show_timer_rate, store_timer_rate); |
| 578 | |
| 579 | static struct attribute *interactive_attributes[] = { |
| 580 | &hispeed_freq_attr.attr, |
| 581 | &go_hispeed_load_attr.attr, |
Todd Poynor | cbbe17d | 2012-04-13 20:18:02 -0700 | [diff] [blame] | 582 | &above_hispeed_delay.attr, |
Mike Chan | ef96969 | 2010-06-22 11:26:45 -0700 | [diff] [blame] | 583 | &min_sample_time_attr.attr, |
| 584 | &timer_rate_attr.attr, |
| 585 | NULL, |
| 586 | }; |
| 587 | |
| 588 | static struct attribute_group interactive_attr_group = { |
| 589 | .attrs = interactive_attributes, |
| 590 | .name = "interactive", |
| 591 | }; |
| 592 | |
| 593 | static int cpufreq_governor_interactive(struct cpufreq_policy *policy, |
| 594 | unsigned int event) |
| 595 | { |
| 596 | int rc; |
| 597 | unsigned int j; |
| 598 | struct cpufreq_interactive_cpuinfo *pcpu; |
| 599 | struct cpufreq_frequency_table *freq_table; |
| 600 | |
| 601 | switch (event) { |
| 602 | case CPUFREQ_GOV_START: |
| 603 | if (!cpu_online(policy->cpu)) |
| 604 | return -EINVAL; |
| 605 | |
| 606 | freq_table = |
| 607 | cpufreq_frequency_get_table(policy->cpu); |
| 608 | |
| 609 | for_each_cpu(j, policy->cpus) { |
| 610 | pcpu = &per_cpu(cpuinfo, j); |
| 611 | pcpu->policy = policy; |
| 612 | pcpu->target_freq = policy->cur; |
| 613 | pcpu->freq_table = freq_table; |
Todd Poynor | 1f408dc | 2012-04-06 19:59:36 -0700 | [diff] [blame] | 614 | pcpu->target_set_time_in_idle = |
Mike Chan | ef96969 | 2010-06-22 11:26:45 -0700 | [diff] [blame] | 615 | get_cpu_idle_time_us(j, |
Todd Poynor | 1f408dc | 2012-04-06 19:59:36 -0700 | [diff] [blame] | 616 | &pcpu->target_set_time); |
Mike Chan | ef96969 | 2010-06-22 11:26:45 -0700 | [diff] [blame] | 617 | pcpu->governor_enabled = 1; |
| 618 | smp_wmb(); |
| 619 | } |
| 620 | |
| 621 | if (!hispeed_freq) |
| 622 | hispeed_freq = policy->max; |
| 623 | |
| 624 | /* |
| 625 | * Do not register the idle hook and create sysfs |
| 626 | * entries if we have already done so. |
| 627 | */ |
| 628 | if (atomic_inc_return(&active_count) > 1) |
| 629 | return 0; |
| 630 | |
| 631 | rc = sysfs_create_group(cpufreq_global_kobject, |
| 632 | &interactive_attr_group); |
| 633 | if (rc) |
| 634 | return rc; |
| 635 | |
| 636 | break; |
| 637 | |
| 638 | case CPUFREQ_GOV_STOP: |
| 639 | for_each_cpu(j, policy->cpus) { |
| 640 | pcpu = &per_cpu(cpuinfo, j); |
| 641 | pcpu->governor_enabled = 0; |
| 642 | smp_wmb(); |
| 643 | del_timer_sync(&pcpu->cpu_timer); |
| 644 | |
| 645 | /* |
| 646 | * Reset idle exit time since we may cancel the timer |
| 647 | * before it can run after the last idle exit time, |
| 648 | * to avoid tripping the check in idle exit for a timer |
| 649 | * that is trying to run. |
| 650 | */ |
| 651 | pcpu->idle_exit_time = 0; |
| 652 | } |
| 653 | |
| 654 | flush_work(&freq_scale_down_work); |
| 655 | if (atomic_dec_return(&active_count) > 0) |
| 656 | return 0; |
| 657 | |
| 658 | sysfs_remove_group(cpufreq_global_kobject, |
| 659 | &interactive_attr_group); |
| 660 | |
| 661 | break; |
| 662 | |
| 663 | case CPUFREQ_GOV_LIMITS: |
| 664 | if (policy->max < policy->cur) |
| 665 | __cpufreq_driver_target(policy, |
| 666 | policy->max, CPUFREQ_RELATION_H); |
| 667 | else if (policy->min > policy->cur) |
| 668 | __cpufreq_driver_target(policy, |
| 669 | policy->min, CPUFREQ_RELATION_L); |
| 670 | break; |
| 671 | } |
| 672 | return 0; |
| 673 | } |
| 674 | |
| 675 | static int cpufreq_interactive_idle_notifier(struct notifier_block *nb, |
| 676 | unsigned long val, |
| 677 | void *data) |
| 678 | { |
| 679 | switch (val) { |
| 680 | case IDLE_START: |
| 681 | cpufreq_interactive_idle_start(); |
| 682 | break; |
| 683 | case IDLE_END: |
| 684 | cpufreq_interactive_idle_end(); |
| 685 | break; |
| 686 | } |
| 687 | |
| 688 | return 0; |
| 689 | } |
| 690 | |
| 691 | static struct notifier_block cpufreq_interactive_idle_nb = { |
| 692 | .notifier_call = cpufreq_interactive_idle_notifier, |
| 693 | }; |
| 694 | |
| 695 | static int __init cpufreq_interactive_init(void) |
| 696 | { |
| 697 | unsigned int i; |
| 698 | struct cpufreq_interactive_cpuinfo *pcpu; |
| 699 | struct sched_param param = { .sched_priority = MAX_RT_PRIO-1 }; |
| 700 | |
| 701 | go_hispeed_load = DEFAULT_GO_HISPEED_LOAD; |
| 702 | min_sample_time = DEFAULT_MIN_SAMPLE_TIME; |
Todd Poynor | cbbe17d | 2012-04-13 20:18:02 -0700 | [diff] [blame] | 703 | above_hispeed_delay_val = DEFAULT_ABOVE_HISPEED_DELAY; |
Mike Chan | ef96969 | 2010-06-22 11:26:45 -0700 | [diff] [blame] | 704 | timer_rate = DEFAULT_TIMER_RATE; |
| 705 | |
| 706 | /* Initalize per-cpu timers */ |
| 707 | for_each_possible_cpu(i) { |
| 708 | pcpu = &per_cpu(cpuinfo, i); |
| 709 | init_timer(&pcpu->cpu_timer); |
| 710 | pcpu->cpu_timer.function = cpufreq_interactive_timer; |
| 711 | pcpu->cpu_timer.data = i; |
| 712 | } |
| 713 | |
| 714 | up_task = kthread_create(cpufreq_interactive_up_task, NULL, |
| 715 | "kinteractiveup"); |
| 716 | if (IS_ERR(up_task)) |
| 717 | return PTR_ERR(up_task); |
| 718 | |
| 719 | sched_setscheduler_nocheck(up_task, SCHED_FIFO, ¶m); |
| 720 | get_task_struct(up_task); |
| 721 | |
| 722 | /* No rescuer thread, bind to CPU queuing the work for possibly |
| 723 | warm cache (probably doesn't matter much). */ |
| 724 | down_wq = alloc_workqueue("knteractive_down", 0, 1); |
| 725 | |
| 726 | if (!down_wq) |
| 727 | goto err_freeuptask; |
| 728 | |
| 729 | INIT_WORK(&freq_scale_down_work, |
| 730 | cpufreq_interactive_freq_down); |
| 731 | |
| 732 | spin_lock_init(&up_cpumask_lock); |
| 733 | spin_lock_init(&down_cpumask_lock); |
| 734 | mutex_init(&set_speed_lock); |
| 735 | |
| 736 | idle_notifier_register(&cpufreq_interactive_idle_nb); |
| 737 | |
| 738 | return cpufreq_register_governor(&cpufreq_gov_interactive); |
| 739 | |
| 740 | err_freeuptask: |
| 741 | put_task_struct(up_task); |
| 742 | return -ENOMEM; |
| 743 | } |
| 744 | |
| 745 | #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_INTERACTIVE |
| 746 | fs_initcall(cpufreq_interactive_init); |
| 747 | #else |
| 748 | module_init(cpufreq_interactive_init); |
| 749 | #endif |
| 750 | |
| 751 | static void __exit cpufreq_interactive_exit(void) |
| 752 | { |
| 753 | cpufreq_unregister_governor(&cpufreq_gov_interactive); |
| 754 | kthread_stop(up_task); |
| 755 | put_task_struct(up_task); |
| 756 | destroy_workqueue(down_wq); |
| 757 | } |
| 758 | |
| 759 | module_exit(cpufreq_interactive_exit); |
| 760 | |
| 761 | MODULE_AUTHOR("Mike Chan <mike@android.com>"); |
| 762 | MODULE_DESCRIPTION("'cpufreq_interactive' - A cpufreq governor for " |
| 763 | "Latency sensitive workloads"); |
| 764 | MODULE_LICENSE("GPL"); |