]>
Commit | Line | Data |
---|---|---|
d6d71ee4 JP |
1 | /* |
2 | * intel_powerclamp.c - package c-state idle injection | |
3 | * | |
4 | * Copyright (c) 2012, Intel Corporation. | |
5 | * | |
6 | * Authors: | |
7 | * Arjan van de Ven <arjan@linux.intel.com> | |
8 | * Jacob Pan <jacob.jun.pan@linux.intel.com> | |
9 | * | |
10 | * This program is free software; you can redistribute it and/or modify it | |
11 | * under the terms and conditions of the GNU General Public License, | |
12 | * version 2, as published by the Free Software Foundation. | |
13 | * | |
14 | * This program is distributed in the hope it will be useful, but WITHOUT | |
15 | * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
16 | * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License for | |
17 | * more details. | |
18 | * | |
19 | * You should have received a copy of the GNU General Public License along with | |
20 | * this program; if not, write to the Free Software Foundation, Inc., | |
21 | * 51 Franklin St - Fifth Floor, Boston, MA 02110-1301 USA. | |
22 | * | |
23 | * | |
24 | * TODO: | |
25 | * 1. better handle wakeup from external interrupts, currently a fixed | |
26 | * compensation is added to clamping duration when excessive amount | |
27 | * of wakeups are observed during idle time. the reason is that in | |
28 | * case of external interrupts without need for ack, clamping down | |
29 | * cpu in non-irq context does not reduce irq. for majority of the | |
30 | * cases, clamping down cpu does help reduce irq as well, we should | |
31 | * be able to differenciate the two cases and give a quantitative | |
32 | * solution for the irqs that we can control. perhaps based on | |
33 | * get_cpu_iowait_time_us() | |
34 | * | |
35 | * 2. synchronization with other hw blocks | |
36 | * | |
37 | * | |
38 | */ | |
39 | ||
40 | #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt | |
41 | ||
42 | #include <linux/module.h> | |
43 | #include <linux/kernel.h> | |
44 | #include <linux/delay.h> | |
45 | #include <linux/kthread.h> | |
d6d71ee4 JP |
46 | #include <linux/cpu.h> |
47 | #include <linux/thermal.h> | |
48 | #include <linux/slab.h> | |
49 | #include <linux/tick.h> | |
50 | #include <linux/debugfs.h> | |
51 | #include <linux/seq_file.h> | |
19cc90f5 | 52 | #include <linux/sched/rt.h> |
d6d71ee4 JP |
53 | |
54 | #include <asm/nmi.h> | |
55 | #include <asm/msr.h> | |
56 | #include <asm/mwait.h> | |
57 | #include <asm/cpu_device_id.h> | |
d6d71ee4 JP |
58 | #include <asm/hardirq.h> |
59 | ||
60 | #define MAX_TARGET_RATIO (50U) | |
61 | /* For each undisturbed clamping period (no extra wake ups during idle time), | |
62 | * we increment the confidence counter for the given target ratio. | |
63 | * CONFIDENCE_OK defines the level where runtime calibration results are | |
64 | * valid. | |
65 | */ | |
66 | #define CONFIDENCE_OK (3) | |
67 | /* Default idle injection duration, driver adjust sleep time to meet target | |
68 | * idle ratio. Similar to frequency modulation. | |
69 | */ | |
70 | #define DEFAULT_DURATION_JIFFIES (6) | |
71 | ||
72 | static unsigned int target_mwait; | |
73 | static struct dentry *debug_dir; | |
74 | ||
75 | /* user selected target */ | |
76 | static unsigned int set_target_ratio; | |
77 | static unsigned int current_ratio; | |
78 | static bool should_skip; | |
79 | static bool reduce_irq; | |
80 | static atomic_t idle_wakeup_counter; | |
81 | static unsigned int control_cpu; /* The cpu assigned to collect stat and update | |
82 | * control parameters. default to BSP but BSP | |
83 | * can be offlined. | |
84 | */ | |
85 | static bool clamping; | |
86 | ||
8d962ac7 PM |
87 | static const struct sched_param sparam = { |
88 | .sched_priority = MAX_USER_RT_PRIO / 2, | |
89 | }; | |
90 | struct powerclamp_worker_data { | |
91 | struct kthread_worker *worker; | |
92 | struct kthread_work balancing_work; | |
93 | struct kthread_delayed_work idle_injection_work; | |
8d962ac7 PM |
94 | unsigned int cpu; |
95 | unsigned int count; | |
96 | unsigned int guard; | |
97 | unsigned int window_size_now; | |
98 | unsigned int target_ratio; | |
99 | unsigned int duration_jiffies; | |
100 | bool clamping; | |
101 | }; | |
d6d71ee4 | 102 | |
8d962ac7 | 103 | static struct powerclamp_worker_data * __percpu worker_data; |
d6d71ee4 JP |
104 | static struct thermal_cooling_device *cooling_dev; |
105 | static unsigned long *cpu_clamping_mask; /* bit map for tracking per cpu | |
8d962ac7 | 106 | * clamping kthread worker |
d6d71ee4 JP |
107 | */ |
108 | ||
109 | static unsigned int duration; | |
110 | static unsigned int pkg_cstate_ratio_cur; | |
111 | static unsigned int window_size; | |
112 | ||
113 | static int duration_set(const char *arg, const struct kernel_param *kp) | |
114 | { | |
115 | int ret = 0; | |
116 | unsigned long new_duration; | |
117 | ||
118 | ret = kstrtoul(arg, 10, &new_duration); | |
119 | if (ret) | |
120 | goto exit; | |
121 | if (new_duration > 25 || new_duration < 6) { | |
122 | pr_err("Out of recommended range %lu, between 6-25ms\n", | |
123 | new_duration); | |
124 | ret = -EINVAL; | |
125 | } | |
126 | ||
127 | duration = clamp(new_duration, 6ul, 25ul); | |
128 | smp_mb(); | |
129 | ||
130 | exit: | |
131 | ||
132 | return ret; | |
133 | } | |
134 | ||
9c27847d | 135 | static const struct kernel_param_ops duration_ops = { |
d6d71ee4 JP |
136 | .set = duration_set, |
137 | .get = param_get_int, | |
138 | }; | |
139 | ||
140 | ||
141 | module_param_cb(duration, &duration_ops, &duration, 0644); | |
142 | MODULE_PARM_DESC(duration, "forced idle time for each attempt in msec."); | |
143 | ||
144 | struct powerclamp_calibration_data { | |
145 | unsigned long confidence; /* used for calibration, basically a counter | |
146 | * gets incremented each time a clamping | |
147 | * period is completed without extra wakeups | |
148 | * once that counter is reached given level, | |
149 | * compensation is deemed usable. | |
150 | */ | |
151 | unsigned long steady_comp; /* steady state compensation used when | |
152 | * no extra wakeups occurred. | |
153 | */ | |
154 | unsigned long dynamic_comp; /* compensate excessive wakeup from idle | |
155 | * mostly from external interrupts. | |
156 | */ | |
157 | }; | |
158 | ||
159 | static struct powerclamp_calibration_data cal_data[MAX_TARGET_RATIO]; | |
160 | ||
161 | static int window_size_set(const char *arg, const struct kernel_param *kp) | |
162 | { | |
163 | int ret = 0; | |
164 | unsigned long new_window_size; | |
165 | ||
166 | ret = kstrtoul(arg, 10, &new_window_size); | |
167 | if (ret) | |
168 | goto exit_win; | |
169 | if (new_window_size > 10 || new_window_size < 2) { | |
170 | pr_err("Out of recommended window size %lu, between 2-10\n", | |
171 | new_window_size); | |
172 | ret = -EINVAL; | |
173 | } | |
174 | ||
175 | window_size = clamp(new_window_size, 2ul, 10ul); | |
176 | smp_mb(); | |
177 | ||
178 | exit_win: | |
179 | ||
180 | return ret; | |
181 | } | |
182 | ||
9c27847d | 183 | static const struct kernel_param_ops window_size_ops = { |
d6d71ee4 JP |
184 | .set = window_size_set, |
185 | .get = param_get_int, | |
186 | }; | |
187 | ||
188 | module_param_cb(window_size, &window_size_ops, &window_size, 0644); | |
189 | MODULE_PARM_DESC(window_size, "sliding window in number of clamping cycles\n" | |
190 | "\tpowerclamp controls idle ratio within this window. larger\n" | |
191 | "\twindow size results in slower response time but more smooth\n" | |
192 | "\tclamping results. default to 2."); | |
193 | ||
194 | static void find_target_mwait(void) | |
195 | { | |
196 | unsigned int eax, ebx, ecx, edx; | |
197 | unsigned int highest_cstate = 0; | |
198 | unsigned int highest_subcstate = 0; | |
199 | int i; | |
200 | ||
201 | if (boot_cpu_data.cpuid_level < CPUID_MWAIT_LEAF) | |
202 | return; | |
203 | ||
204 | cpuid(CPUID_MWAIT_LEAF, &eax, &ebx, &ecx, &edx); | |
205 | ||
206 | if (!(ecx & CPUID5_ECX_EXTENSIONS_SUPPORTED) || | |
207 | !(ecx & CPUID5_ECX_INTERRUPT_BREAK)) | |
208 | return; | |
209 | ||
210 | edx >>= MWAIT_SUBSTATE_SIZE; | |
211 | for (i = 0; i < 7 && edx; i++, edx >>= MWAIT_SUBSTATE_SIZE) { | |
212 | if (edx & MWAIT_SUBSTATE_MASK) { | |
213 | highest_cstate = i; | |
214 | highest_subcstate = edx & MWAIT_SUBSTATE_MASK; | |
215 | } | |
216 | } | |
217 | target_mwait = (highest_cstate << MWAIT_SUBSTATE_SIZE) | | |
218 | (highest_subcstate - 1); | |
219 | ||
220 | } | |
221 | ||
d8186113 JP |
222 | struct pkg_cstate_info { |
223 | bool skip; | |
224 | int msr_index; | |
225 | int cstate_id; | |
226 | }; | |
227 | ||
228 | #define PKG_CSTATE_INIT(id) { \ | |
229 | .msr_index = MSR_PKG_C##id##_RESIDENCY, \ | |
230 | .cstate_id = id \ | |
231 | } | |
232 | ||
233 | static struct pkg_cstate_info pkg_cstates[] = { | |
234 | PKG_CSTATE_INIT(2), | |
235 | PKG_CSTATE_INIT(3), | |
236 | PKG_CSTATE_INIT(6), | |
237 | PKG_CSTATE_INIT(7), | |
238 | PKG_CSTATE_INIT(8), | |
239 | PKG_CSTATE_INIT(9), | |
240 | PKG_CSTATE_INIT(10), | |
241 | {NULL}, | |
242 | }; | |
243 | ||
7734e3ac YS |
244 | static bool has_pkg_state_counter(void) |
245 | { | |
d8186113 JP |
246 | u64 val; |
247 | struct pkg_cstate_info *info = pkg_cstates; | |
248 | ||
249 | /* check if any one of the counter msrs exists */ | |
250 | while (info->msr_index) { | |
251 | if (!rdmsrl_safe(info->msr_index, &val)) | |
252 | return true; | |
253 | info++; | |
254 | } | |
255 | ||
256 | return false; | |
7734e3ac YS |
257 | } |
258 | ||
d6d71ee4 JP |
259 | static u64 pkg_state_counter(void) |
260 | { | |
261 | u64 val; | |
262 | u64 count = 0; | |
d8186113 JP |
263 | struct pkg_cstate_info *info = pkg_cstates; |
264 | ||
265 | while (info->msr_index) { | |
266 | if (!info->skip) { | |
267 | if (!rdmsrl_safe(info->msr_index, &val)) | |
268 | count += val; | |
269 | else | |
270 | info->skip = true; | |
271 | } | |
272 | info++; | |
d6d71ee4 JP |
273 | } |
274 | ||
275 | return count; | |
276 | } | |
277 | ||
d6d71ee4 JP |
278 | static unsigned int get_compensation(int ratio) |
279 | { | |
280 | unsigned int comp = 0; | |
281 | ||
282 | /* we only use compensation if all adjacent ones are good */ | |
283 | if (ratio == 1 && | |
284 | cal_data[ratio].confidence >= CONFIDENCE_OK && | |
285 | cal_data[ratio + 1].confidence >= CONFIDENCE_OK && | |
286 | cal_data[ratio + 2].confidence >= CONFIDENCE_OK) { | |
287 | comp = (cal_data[ratio].steady_comp + | |
288 | cal_data[ratio + 1].steady_comp + | |
289 | cal_data[ratio + 2].steady_comp) / 3; | |
290 | } else if (ratio == MAX_TARGET_RATIO - 1 && | |
291 | cal_data[ratio].confidence >= CONFIDENCE_OK && | |
292 | cal_data[ratio - 1].confidence >= CONFIDENCE_OK && | |
293 | cal_data[ratio - 2].confidence >= CONFIDENCE_OK) { | |
294 | comp = (cal_data[ratio].steady_comp + | |
295 | cal_data[ratio - 1].steady_comp + | |
296 | cal_data[ratio - 2].steady_comp) / 3; | |
297 | } else if (cal_data[ratio].confidence >= CONFIDENCE_OK && | |
298 | cal_data[ratio - 1].confidence >= CONFIDENCE_OK && | |
299 | cal_data[ratio + 1].confidence >= CONFIDENCE_OK) { | |
300 | comp = (cal_data[ratio].steady_comp + | |
301 | cal_data[ratio - 1].steady_comp + | |
302 | cal_data[ratio + 1].steady_comp) / 3; | |
303 | } | |
304 | ||
305 | /* REVISIT: simple penalty of double idle injection */ | |
306 | if (reduce_irq) | |
307 | comp = ratio; | |
308 | /* do not exceed limit */ | |
309 | if (comp + ratio >= MAX_TARGET_RATIO) | |
310 | comp = MAX_TARGET_RATIO - ratio - 1; | |
311 | ||
312 | return comp; | |
313 | } | |
314 | ||
315 | static void adjust_compensation(int target_ratio, unsigned int win) | |
316 | { | |
317 | int delta; | |
318 | struct powerclamp_calibration_data *d = &cal_data[target_ratio]; | |
319 | ||
320 | /* | |
321 | * adjust compensations if confidence level has not been reached or | |
322 | * there are too many wakeups during the last idle injection period, we | |
323 | * cannot trust the data for compensation. | |
324 | */ | |
325 | if (d->confidence >= CONFIDENCE_OK || | |
326 | atomic_read(&idle_wakeup_counter) > | |
327 | win * num_online_cpus()) | |
328 | return; | |
329 | ||
330 | delta = set_target_ratio - current_ratio; | |
331 | /* filter out bad data */ | |
332 | if (delta >= 0 && delta <= (1+target_ratio/10)) { | |
333 | if (d->steady_comp) | |
334 | d->steady_comp = | |
335 | roundup(delta+d->steady_comp, 2)/2; | |
336 | else | |
337 | d->steady_comp = delta; | |
338 | d->confidence++; | |
339 | } | |
340 | } | |
341 | ||
342 | static bool powerclamp_adjust_controls(unsigned int target_ratio, | |
343 | unsigned int guard, unsigned int win) | |
344 | { | |
345 | static u64 msr_last, tsc_last; | |
346 | u64 msr_now, tsc_now; | |
347 | u64 val64; | |
348 | ||
349 | /* check result for the last window */ | |
350 | msr_now = pkg_state_counter(); | |
4ea1636b | 351 | tsc_now = rdtsc(); |
d6d71ee4 JP |
352 | |
353 | /* calculate pkg cstate vs tsc ratio */ | |
354 | if (!msr_last || !tsc_last) | |
355 | current_ratio = 1; | |
356 | else if (tsc_now-tsc_last) { | |
357 | val64 = 100*(msr_now-msr_last); | |
358 | do_div(val64, (tsc_now-tsc_last)); | |
359 | current_ratio = val64; | |
360 | } | |
361 | ||
362 | /* update record */ | |
363 | msr_last = msr_now; | |
364 | tsc_last = tsc_now; | |
365 | ||
366 | adjust_compensation(target_ratio, win); | |
367 | /* | |
368 | * too many external interrupts, set flag such | |
369 | * that we can take measure later. | |
370 | */ | |
371 | reduce_irq = atomic_read(&idle_wakeup_counter) >= | |
372 | 2 * win * num_online_cpus(); | |
373 | ||
374 | atomic_set(&idle_wakeup_counter, 0); | |
375 | /* if we are above target+guard, skip */ | |
376 | return set_target_ratio + guard <= current_ratio; | |
377 | } | |
378 | ||
8d962ac7 | 379 | static void clamp_balancing_func(struct kthread_work *work) |
d6d71ee4 | 380 | { |
8d962ac7 PM |
381 | struct powerclamp_worker_data *w_data; |
382 | int sleeptime; | |
383 | unsigned long target_jiffies; | |
384 | unsigned int compensated_ratio; | |
385 | int interval; /* jiffies to sleep for each attempt */ | |
d6d71ee4 | 386 | |
8d962ac7 PM |
387 | w_data = container_of(work, struct powerclamp_worker_data, |
388 | balancing_work); | |
d6d71ee4 | 389 | |
8d962ac7 PM |
390 | /* |
391 | * make sure user selected ratio does not take effect until | |
392 | * the next round. adjust target_ratio if user has changed | |
393 | * target such that we can converge quickly. | |
394 | */ | |
395 | w_data->target_ratio = READ_ONCE(set_target_ratio); | |
396 | w_data->guard = 1 + w_data->target_ratio / 20; | |
397 | w_data->window_size_now = window_size; | |
398 | w_data->duration_jiffies = msecs_to_jiffies(duration); | |
399 | w_data->count++; | |
400 | ||
401 | /* | |
402 | * systems may have different ability to enter package level | |
403 | * c-states, thus we need to compensate the injected idle ratio | |
404 | * to achieve the actual target reported by the HW. | |
405 | */ | |
406 | compensated_ratio = w_data->target_ratio + | |
407 | get_compensation(w_data->target_ratio); | |
408 | if (compensated_ratio <= 0) | |
409 | compensated_ratio = 1; | |
410 | interval = w_data->duration_jiffies * 100 / compensated_ratio; | |
411 | ||
412 | /* align idle time */ | |
413 | target_jiffies = roundup(jiffies, interval); | |
414 | sleeptime = target_jiffies - jiffies; | |
415 | if (sleeptime <= 0) | |
416 | sleeptime = 1; | |
417 | ||
418 | if (clamping && w_data->clamping && cpu_online(w_data->cpu)) | |
419 | kthread_queue_delayed_work(w_data->worker, | |
420 | &w_data->idle_injection_work, | |
421 | sleeptime); | |
422 | } | |
423 | ||
424 | static void clamp_idle_injection_func(struct kthread_work *work) | |
425 | { | |
426 | struct powerclamp_worker_data *w_data; | |
8d962ac7 PM |
427 | |
428 | w_data = container_of(work, struct powerclamp_worker_data, | |
429 | idle_injection_work.work); | |
430 | ||
431 | /* | |
432 | * only elected controlling cpu can collect stats and update | |
433 | * control parameters. | |
434 | */ | |
435 | if (w_data->cpu == control_cpu && | |
436 | !(w_data->count % w_data->window_size_now)) { | |
437 | should_skip = | |
438 | powerclamp_adjust_controls(w_data->target_ratio, | |
439 | w_data->guard, | |
440 | w_data->window_size_now); | |
441 | smp_mb(); | |
d6d71ee4 | 442 | } |
d6d71ee4 | 443 | |
8d962ac7 PM |
444 | if (should_skip) |
445 | goto balance; | |
446 | ||
feb6cd6a | 447 | play_idle(jiffies_to_msecs(w_data->duration_jiffies)); |
d6d71ee4 | 448 | |
8d962ac7 PM |
449 | balance: |
450 | if (clamping && w_data->clamping && cpu_online(w_data->cpu)) | |
451 | kthread_queue_work(w_data->worker, &w_data->balancing_work); | |
d6d71ee4 JP |
452 | } |
453 | ||
454 | /* | |
455 | * 1 HZ polling while clamping is active, useful for userspace | |
456 | * to monitor actual idle ratio. | |
457 | */ | |
458 | static void poll_pkg_cstate(struct work_struct *dummy); | |
459 | static DECLARE_DELAYED_WORK(poll_pkg_cstate_work, poll_pkg_cstate); | |
460 | static void poll_pkg_cstate(struct work_struct *dummy) | |
461 | { | |
462 | static u64 msr_last; | |
463 | static u64 tsc_last; | |
464 | static unsigned long jiffies_last; | |
465 | ||
466 | u64 msr_now; | |
467 | unsigned long jiffies_now; | |
468 | u64 tsc_now; | |
469 | u64 val64; | |
470 | ||
471 | msr_now = pkg_state_counter(); | |
4ea1636b | 472 | tsc_now = rdtsc(); |
d6d71ee4 JP |
473 | jiffies_now = jiffies; |
474 | ||
475 | /* calculate pkg cstate vs tsc ratio */ | |
476 | if (!msr_last || !tsc_last) | |
477 | pkg_cstate_ratio_cur = 1; | |
478 | else { | |
479 | if (tsc_now - tsc_last) { | |
480 | val64 = 100 * (msr_now - msr_last); | |
481 | do_div(val64, (tsc_now - tsc_last)); | |
482 | pkg_cstate_ratio_cur = val64; | |
483 | } | |
484 | } | |
485 | ||
486 | /* update record */ | |
487 | msr_last = msr_now; | |
488 | jiffies_last = jiffies_now; | |
489 | tsc_last = tsc_now; | |
490 | ||
491 | if (true == clamping) | |
492 | schedule_delayed_work(&poll_pkg_cstate_work, HZ); | |
493 | } | |
494 | ||
8d962ac7 | 495 | static void start_power_clamp_worker(unsigned long cpu) |
14f3f7d8 | 496 | { |
8d962ac7 PM |
497 | struct powerclamp_worker_data *w_data = per_cpu_ptr(worker_data, cpu); |
498 | struct kthread_worker *worker; | |
499 | ||
cb91fef1 | 500 | worker = kthread_create_worker_on_cpu(cpu, 0, "kidle_inject/%ld", cpu); |
8d962ac7 | 501 | if (IS_ERR(worker)) |
14f3f7d8 PM |
502 | return; |
503 | ||
8d962ac7 PM |
504 | w_data->worker = worker; |
505 | w_data->count = 0; | |
506 | w_data->cpu = cpu; | |
507 | w_data->clamping = true; | |
508 | set_bit(cpu, cpu_clamping_mask); | |
8d962ac7 PM |
509 | sched_setscheduler(worker->task, SCHED_FIFO, &sparam); |
510 | kthread_init_work(&w_data->balancing_work, clamp_balancing_func); | |
511 | kthread_init_delayed_work(&w_data->idle_injection_work, | |
512 | clamp_idle_injection_func); | |
513 | kthread_queue_work(w_data->worker, &w_data->balancing_work); | |
514 | } | |
515 | ||
516 | static void stop_power_clamp_worker(unsigned long cpu) | |
517 | { | |
518 | struct powerclamp_worker_data *w_data = per_cpu_ptr(worker_data, cpu); | |
519 | ||
520 | if (!w_data->worker) | |
521 | return; | |
522 | ||
523 | w_data->clamping = false; | |
524 | /* | |
525 | * Make sure that all works that get queued after this point see | |
526 | * the clamping disabled. The counter part is not needed because | |
527 | * there is an implicit memory barrier when the queued work | |
528 | * is proceed. | |
529 | */ | |
530 | smp_wmb(); | |
531 | kthread_cancel_work_sync(&w_data->balancing_work); | |
532 | kthread_cancel_delayed_work_sync(&w_data->idle_injection_work); | |
533 | /* | |
534 | * The balancing work still might be queued here because | |
535 | * the handling of the "clapming" variable, cancel, and queue | |
536 | * operations are not synchronized via a lock. But it is not | |
537 | * a big deal. The balancing work is fast and destroy kthread | |
538 | * will wait for it. | |
539 | */ | |
8d962ac7 PM |
540 | clear_bit(w_data->cpu, cpu_clamping_mask); |
541 | kthread_destroy_worker(w_data->worker); | |
542 | ||
543 | w_data->worker = NULL; | |
14f3f7d8 PM |
544 | } |
545 | ||
d6d71ee4 JP |
546 | static int start_power_clamp(void) |
547 | { | |
548 | unsigned long cpu; | |
d6d71ee4 | 549 | |
c8165dc0 | 550 | set_target_ratio = clamp(set_target_ratio, 0U, MAX_TARGET_RATIO - 1); |
d6d71ee4 JP |
551 | /* prevent cpu hotplug */ |
552 | get_online_cpus(); | |
553 | ||
554 | /* prefer BSP */ | |
555 | control_cpu = 0; | |
556 | if (!cpu_online(control_cpu)) | |
557 | control_cpu = smp_processor_id(); | |
558 | ||
559 | clamping = true; | |
560 | schedule_delayed_work(&poll_pkg_cstate_work, 0); | |
561 | ||
8d962ac7 | 562 | /* start one kthread worker per online cpu */ |
d6d71ee4 | 563 | for_each_online_cpu(cpu) { |
8d962ac7 | 564 | start_power_clamp_worker(cpu); |
d6d71ee4 JP |
565 | } |
566 | put_online_cpus(); | |
567 | ||
568 | return 0; | |
569 | } | |
570 | ||
571 | static void end_power_clamp(void) | |
572 | { | |
573 | int i; | |
d6d71ee4 | 574 | |
d6d71ee4 | 575 | /* |
8d962ac7 PM |
576 | * Block requeuing in all the kthread workers. They will flush and |
577 | * stop faster. | |
d6d71ee4 | 578 | */ |
8d962ac7 | 579 | clamping = false; |
d6d71ee4 JP |
580 | if (bitmap_weight(cpu_clamping_mask, num_possible_cpus())) { |
581 | for_each_set_bit(i, cpu_clamping_mask, num_possible_cpus()) { | |
8d962ac7 PM |
582 | pr_debug("clamping worker for cpu %d alive, destroy\n", |
583 | i); | |
584 | stop_power_clamp_worker(i); | |
d6d71ee4 JP |
585 | } |
586 | } | |
587 | } | |
588 | ||
cb91fef1 | 589 | static int powerclamp_cpu_online(unsigned int cpu) |
d6d71ee4 | 590 | { |
cb91fef1 SAS |
591 | if (clamping == false) |
592 | return 0; | |
593 | start_power_clamp_worker(cpu); | |
594 | /* prefer BSP as controlling CPU */ | |
595 | if (cpu == 0) { | |
596 | control_cpu = 0; | |
597 | smp_mb(); | |
d6d71ee4 | 598 | } |
cb91fef1 | 599 | return 0; |
d6d71ee4 JP |
600 | } |
601 | ||
cb91fef1 SAS |
602 | static int powerclamp_cpu_predown(unsigned int cpu) |
603 | { | |
604 | if (clamping == false) | |
605 | return 0; | |
d6d71ee4 | 606 | |
cb91fef1 SAS |
607 | stop_power_clamp_worker(cpu); |
608 | if (cpu != control_cpu) | |
609 | return 0; | |
d6d71ee4 | 610 | |
cb91fef1 SAS |
611 | control_cpu = cpumask_first(cpu_online_mask); |
612 | if (control_cpu == cpu) | |
613 | control_cpu = cpumask_next(cpu, cpu_online_mask); | |
614 | smp_mb(); | |
615 | return 0; | |
d6d71ee4 | 616 | } |
d6d71ee4 JP |
617 | |
618 | static int powerclamp_get_max_state(struct thermal_cooling_device *cdev, | |
619 | unsigned long *state) | |
620 | { | |
621 | *state = MAX_TARGET_RATIO; | |
622 | ||
623 | return 0; | |
624 | } | |
625 | ||
626 | static int powerclamp_get_cur_state(struct thermal_cooling_device *cdev, | |
627 | unsigned long *state) | |
628 | { | |
629 | if (true == clamping) | |
630 | *state = pkg_cstate_ratio_cur; | |
631 | else | |
632 | /* to save power, do not poll idle ratio while not clamping */ | |
633 | *state = -1; /* indicates invalid state */ | |
634 | ||
635 | return 0; | |
636 | } | |
637 | ||
638 | static int powerclamp_set_cur_state(struct thermal_cooling_device *cdev, | |
639 | unsigned long new_target_ratio) | |
640 | { | |
641 | int ret = 0; | |
642 | ||
643 | new_target_ratio = clamp(new_target_ratio, 0UL, | |
644 | (unsigned long) (MAX_TARGET_RATIO-1)); | |
645 | if (set_target_ratio == 0 && new_target_ratio > 0) { | |
646 | pr_info("Start idle injection to reduce power\n"); | |
647 | set_target_ratio = new_target_ratio; | |
648 | ret = start_power_clamp(); | |
649 | goto exit_set; | |
650 | } else if (set_target_ratio > 0 && new_target_ratio == 0) { | |
651 | pr_info("Stop forced idle injection\n"); | |
d6d71ee4 | 652 | end_power_clamp(); |
70c50ee7 | 653 | set_target_ratio = 0; |
d6d71ee4 JP |
654 | } else /* adjust currently running */ { |
655 | set_target_ratio = new_target_ratio; | |
656 | /* make new set_target_ratio visible to other cpus */ | |
657 | smp_mb(); | |
658 | } | |
659 | ||
660 | exit_set: | |
661 | return ret; | |
662 | } | |
663 | ||
664 | /* bind to generic thermal layer as cooling device*/ | |
665 | static struct thermal_cooling_device_ops powerclamp_cooling_ops = { | |
666 | .get_max_state = powerclamp_get_max_state, | |
667 | .get_cur_state = powerclamp_get_cur_state, | |
668 | .set_cur_state = powerclamp_set_cur_state, | |
669 | }; | |
670 | ||
ec638db8 JP |
671 | static const struct x86_cpu_id __initconst intel_powerclamp_ids[] = { |
672 | { X86_VENDOR_INTEL, X86_FAMILY_ANY, X86_MODEL_ANY, X86_FEATURE_MWAIT }, | |
673 | {} | |
674 | }; | |
675 | MODULE_DEVICE_TABLE(x86cpu, intel_powerclamp_ids); | |
676 | ||
4d2b6e4a | 677 | static int __init powerclamp_probe(void) |
d6d71ee4 | 678 | { |
ec638db8 JP |
679 | |
680 | if (!x86_match_cpu(intel_powerclamp_ids)) { | |
3105f234 | 681 | pr_err("CPU does not support MWAIT"); |
d6d71ee4 JP |
682 | return -ENODEV; |
683 | } | |
b721ca0d JP |
684 | |
685 | /* The goal for idle time alignment is to achieve package cstate. */ | |
686 | if (!has_pkg_state_counter()) { | |
687 | pr_info("No package C-state available"); | |
d6d71ee4 | 688 | return -ENODEV; |
b721ca0d | 689 | } |
d6d71ee4 JP |
690 | |
691 | /* find the deepest mwait value */ | |
692 | find_target_mwait(); | |
693 | ||
694 | return 0; | |
695 | } | |
696 | ||
697 | static int powerclamp_debug_show(struct seq_file *m, void *unused) | |
698 | { | |
699 | int i = 0; | |
700 | ||
701 | seq_printf(m, "controlling cpu: %d\n", control_cpu); | |
702 | seq_printf(m, "pct confidence steady dynamic (compensation)\n"); | |
703 | for (i = 0; i < MAX_TARGET_RATIO; i++) { | |
704 | seq_printf(m, "%d\t%lu\t%lu\t%lu\n", | |
705 | i, | |
706 | cal_data[i].confidence, | |
707 | cal_data[i].steady_comp, | |
708 | cal_data[i].dynamic_comp); | |
709 | } | |
710 | ||
711 | return 0; | |
712 | } | |
713 | ||
714 | static int powerclamp_debug_open(struct inode *inode, | |
715 | struct file *file) | |
716 | { | |
717 | return single_open(file, powerclamp_debug_show, inode->i_private); | |
718 | } | |
719 | ||
720 | static const struct file_operations powerclamp_debug_fops = { | |
721 | .open = powerclamp_debug_open, | |
722 | .read = seq_read, | |
723 | .llseek = seq_lseek, | |
724 | .release = single_release, | |
725 | .owner = THIS_MODULE, | |
726 | }; | |
727 | ||
728 | static inline void powerclamp_create_debug_files(void) | |
729 | { | |
730 | debug_dir = debugfs_create_dir("intel_powerclamp", NULL); | |
731 | if (!debug_dir) | |
732 | return; | |
733 | ||
734 | if (!debugfs_create_file("powerclamp_calib", S_IRUGO, debug_dir, | |
735 | cal_data, &powerclamp_debug_fops)) | |
736 | goto file_error; | |
737 | ||
738 | return; | |
739 | ||
740 | file_error: | |
741 | debugfs_remove_recursive(debug_dir); | |
742 | } | |
743 | ||
cb91fef1 SAS |
744 | static enum cpuhp_state hp_state; |
745 | ||
4d2b6e4a | 746 | static int __init powerclamp_init(void) |
d6d71ee4 JP |
747 | { |
748 | int retval; | |
749 | int bitmap_size; | |
750 | ||
751 | bitmap_size = BITS_TO_LONGS(num_possible_cpus()) * sizeof(long); | |
752 | cpu_clamping_mask = kzalloc(bitmap_size, GFP_KERNEL); | |
753 | if (!cpu_clamping_mask) | |
754 | return -ENOMEM; | |
755 | ||
756 | /* probe cpu features and ids here */ | |
757 | retval = powerclamp_probe(); | |
758 | if (retval) | |
c32a5087 | 759 | goto exit_free; |
760 | ||
d6d71ee4 JP |
761 | /* set default limit, maybe adjusted during runtime based on feedback */ |
762 | window_size = 2; | |
cb91fef1 SAS |
763 | retval = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN, |
764 | "thermal/intel_powerclamp:online", | |
765 | powerclamp_cpu_online, | |
766 | powerclamp_cpu_predown); | |
767 | if (retval < 0) | |
768 | goto exit_free; | |
769 | ||
770 | hp_state = retval; | |
c32a5087 | 771 | |
8d962ac7 PM |
772 | worker_data = alloc_percpu(struct powerclamp_worker_data); |
773 | if (!worker_data) { | |
c32a5087 | 774 | retval = -ENOMEM; |
775 | goto exit_unregister; | |
776 | } | |
777 | ||
d6d71ee4 JP |
778 | cooling_dev = thermal_cooling_device_register("intel_powerclamp", NULL, |
779 | &powerclamp_cooling_ops); | |
c32a5087 | 780 | if (IS_ERR(cooling_dev)) { |
781 | retval = -ENODEV; | |
782 | goto exit_free_thread; | |
783 | } | |
d6d71ee4 JP |
784 | |
785 | if (!duration) | |
786 | duration = jiffies_to_msecs(DEFAULT_DURATION_JIFFIES); | |
c32a5087 | 787 | |
d6d71ee4 JP |
788 | powerclamp_create_debug_files(); |
789 | ||
790 | return 0; | |
c32a5087 | 791 | |
792 | exit_free_thread: | |
8d962ac7 | 793 | free_percpu(worker_data); |
c32a5087 | 794 | exit_unregister: |
cb91fef1 | 795 | cpuhp_remove_state_nocalls(hp_state); |
c32a5087 | 796 | exit_free: |
797 | kfree(cpu_clamping_mask); | |
798 | return retval; | |
d6d71ee4 JP |
799 | } |
800 | module_init(powerclamp_init); | |
801 | ||
4d2b6e4a | 802 | static void __exit powerclamp_exit(void) |
d6d71ee4 | 803 | { |
d6d71ee4 | 804 | end_power_clamp(); |
cb91fef1 | 805 | cpuhp_remove_state_nocalls(hp_state); |
8d962ac7 | 806 | free_percpu(worker_data); |
d6d71ee4 JP |
807 | thermal_cooling_device_unregister(cooling_dev); |
808 | kfree(cpu_clamping_mask); | |
809 | ||
810 | cancel_delayed_work_sync(&poll_pkg_cstate_work); | |
811 | debugfs_remove_recursive(debug_dir); | |
812 | } | |
813 | module_exit(powerclamp_exit); | |
814 | ||
815 | MODULE_LICENSE("GPL"); | |
816 | MODULE_AUTHOR("Arjan van de Ven <arjan@linux.intel.com>"); | |
817 | MODULE_AUTHOR("Jacob Pan <jacob.jun.pan@linux.intel.com>"); | |
818 | MODULE_DESCRIPTION("Package Level C-state Idle Injection for Intel CPUs"); |