1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
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>
46#include <linux/freezer.h>
47#include <linux/cpu.h>
48#include <linux/thermal.h>
49#include <linux/slab.h>
50#include <linux/tick.h>
51#include <linux/debugfs.h>
52#include <linux/seq_file.h>
53#include <linux/sched/rt.h>
54
55#include <asm/nmi.h>
56#include <asm/msr.h>
57#include <asm/mwait.h>
58#include <asm/cpu_device_id.h>
59#include <asm/idle.h>
60#include <asm/hardirq.h>
61
62#define MAX_TARGET_RATIO (50U)
63
64
65
66
67
68#define CONFIDENCE_OK (3)
69
70
71
72#define DEFAULT_DURATION_JIFFIES (6)
73
74static unsigned int target_mwait;
75static struct dentry *debug_dir;
76
77
78static unsigned int set_target_ratio;
79static unsigned int current_ratio;
80static bool should_skip;
81static bool reduce_irq;
82static atomic_t idle_wakeup_counter;
83static unsigned int control_cpu;
84
85
86
87static bool clamping;
88
89
90static struct task_struct * __percpu *powerclamp_thread;
91static struct thermal_cooling_device *cooling_dev;
92static unsigned long *cpu_clamping_mask;
93
94
95
96static unsigned int duration;
97static unsigned int pkg_cstate_ratio_cur;
98static unsigned int window_size;
99
100static int duration_set(const char *arg, const struct kernel_param *kp)
101{
102 int ret = 0;
103 unsigned long new_duration;
104
105 ret = kstrtoul(arg, 10, &new_duration);
106 if (ret)
107 goto exit;
108 if (new_duration > 25 || new_duration < 6) {
109 pr_err("Out of recommended range %lu, between 6-25ms\n",
110 new_duration);
111 ret = -EINVAL;
112 }
113
114 duration = clamp(new_duration, 6ul, 25ul);
115 smp_mb();
116
117exit:
118
119 return ret;
120}
121
122static const struct kernel_param_ops duration_ops = {
123 .set = duration_set,
124 .get = param_get_int,
125};
126
127
128module_param_cb(duration, &duration_ops, &duration, 0644);
129MODULE_PARM_DESC(duration, "forced idle time for each attempt in msec.");
130
131struct powerclamp_calibration_data {
132 unsigned long confidence;
133
134
135
136
137
138 unsigned long steady_comp;
139
140
141 unsigned long dynamic_comp;
142
143
144};
145
146static struct powerclamp_calibration_data cal_data[MAX_TARGET_RATIO];
147
148static int window_size_set(const char *arg, const struct kernel_param *kp)
149{
150 int ret = 0;
151 unsigned long new_window_size;
152
153 ret = kstrtoul(arg, 10, &new_window_size);
154 if (ret)
155 goto exit_win;
156 if (new_window_size > 10 || new_window_size < 2) {
157 pr_err("Out of recommended window size %lu, between 2-10\n",
158 new_window_size);
159 ret = -EINVAL;
160 }
161
162 window_size = clamp(new_window_size, 2ul, 10ul);
163 smp_mb();
164
165exit_win:
166
167 return ret;
168}
169
170static const struct kernel_param_ops window_size_ops = {
171 .set = window_size_set,
172 .get = param_get_int,
173};
174
175module_param_cb(window_size, &window_size_ops, &window_size, 0644);
176MODULE_PARM_DESC(window_size, "sliding window in number of clamping cycles\n"
177 "\tpowerclamp controls idle ratio within this window. larger\n"
178 "\twindow size results in slower response time but more smooth\n"
179 "\tclamping results. default to 2.");
180
181static void find_target_mwait(void)
182{
183 unsigned int eax, ebx, ecx, edx;
184 unsigned int highest_cstate = 0;
185 unsigned int highest_subcstate = 0;
186 int i;
187
188 if (boot_cpu_data.cpuid_level < CPUID_MWAIT_LEAF)
189 return;
190
191 cpuid(CPUID_MWAIT_LEAF, &eax, &ebx, &ecx, &edx);
192
193 if (!(ecx & CPUID5_ECX_EXTENSIONS_SUPPORTED) ||
194 !(ecx & CPUID5_ECX_INTERRUPT_BREAK))
195 return;
196
197 edx >>= MWAIT_SUBSTATE_SIZE;
198 for (i = 0; i < 7 && edx; i++, edx >>= MWAIT_SUBSTATE_SIZE) {
199 if (edx & MWAIT_SUBSTATE_MASK) {
200 highest_cstate = i;
201 highest_subcstate = edx & MWAIT_SUBSTATE_MASK;
202 }
203 }
204 target_mwait = (highest_cstate << MWAIT_SUBSTATE_SIZE) |
205 (highest_subcstate - 1);
206
207}
208
209struct pkg_cstate_info {
210 bool skip;
211 int msr_index;
212 int cstate_id;
213};
214
215#define PKG_CSTATE_INIT(id) { \
216 .msr_index = MSR_PKG_C##id##_RESIDENCY, \
217 .cstate_id = id \
218 }
219
220static struct pkg_cstate_info pkg_cstates[] = {
221 PKG_CSTATE_INIT(2),
222 PKG_CSTATE_INIT(3),
223 PKG_CSTATE_INIT(6),
224 PKG_CSTATE_INIT(7),
225 PKG_CSTATE_INIT(8),
226 PKG_CSTATE_INIT(9),
227 PKG_CSTATE_INIT(10),
228 {NULL},
229};
230
231static bool has_pkg_state_counter(void)
232{
233 u64 val;
234 struct pkg_cstate_info *info = pkg_cstates;
235
236
237 while (info->msr_index) {
238 if (!rdmsrl_safe(info->msr_index, &val))
239 return true;
240 info++;
241 }
242
243 return false;
244}
245
246static u64 pkg_state_counter(void)
247{
248 u64 val;
249 u64 count = 0;
250 struct pkg_cstate_info *info = pkg_cstates;
251
252 while (info->msr_index) {
253 if (!info->skip) {
254 if (!rdmsrl_safe(info->msr_index, &val))
255 count += val;
256 else
257 info->skip = true;
258 }
259 info++;
260 }
261
262 return count;
263}
264
265static void noop_timer(unsigned long foo)
266{
267
268}
269
270static unsigned int get_compensation(int ratio)
271{
272 unsigned int comp = 0;
273
274
275 if (ratio == 1 &&
276 cal_data[ratio].confidence >= CONFIDENCE_OK &&
277 cal_data[ratio + 1].confidence >= CONFIDENCE_OK &&
278 cal_data[ratio + 2].confidence >= CONFIDENCE_OK) {
279 comp = (cal_data[ratio].steady_comp +
280 cal_data[ratio + 1].steady_comp +
281 cal_data[ratio + 2].steady_comp) / 3;
282 } else if (ratio == MAX_TARGET_RATIO - 1 &&
283 cal_data[ratio].confidence >= CONFIDENCE_OK &&
284 cal_data[ratio - 1].confidence >= CONFIDENCE_OK &&
285 cal_data[ratio - 2].confidence >= CONFIDENCE_OK) {
286 comp = (cal_data[ratio].steady_comp +
287 cal_data[ratio - 1].steady_comp +
288 cal_data[ratio - 2].steady_comp) / 3;
289 } else if (cal_data[ratio].confidence >= CONFIDENCE_OK &&
290 cal_data[ratio - 1].confidence >= CONFIDENCE_OK &&
291 cal_data[ratio + 1].confidence >= CONFIDENCE_OK) {
292 comp = (cal_data[ratio].steady_comp +
293 cal_data[ratio - 1].steady_comp +
294 cal_data[ratio + 1].steady_comp) / 3;
295 }
296
297
298 if (reduce_irq)
299 comp = ratio;
300
301 if (comp + ratio >= MAX_TARGET_RATIO)
302 comp = MAX_TARGET_RATIO - ratio - 1;
303
304 return comp;
305}
306
307static void adjust_compensation(int target_ratio, unsigned int win)
308{
309 int delta;
310 struct powerclamp_calibration_data *d = &cal_data[target_ratio];
311
312
313
314
315
316
317 if (d->confidence >= CONFIDENCE_OK ||
318 atomic_read(&idle_wakeup_counter) >
319 win * num_online_cpus())
320 return;
321
322 delta = set_target_ratio - current_ratio;
323
324 if (delta >= 0 && delta <= (1+target_ratio/10)) {
325 if (d->steady_comp)
326 d->steady_comp =
327 roundup(delta+d->steady_comp, 2)/2;
328 else
329 d->steady_comp = delta;
330 d->confidence++;
331 }
332}
333
334static bool powerclamp_adjust_controls(unsigned int target_ratio,
335 unsigned int guard, unsigned int win)
336{
337 static u64 msr_last, tsc_last;
338 u64 msr_now, tsc_now;
339 u64 val64;
340
341
342 msr_now = pkg_state_counter();
343 tsc_now = rdtsc();
344
345
346 if (!msr_last || !tsc_last)
347 current_ratio = 1;
348 else if (tsc_now-tsc_last) {
349 val64 = 100*(msr_now-msr_last);
350 do_div(val64, (tsc_now-tsc_last));
351 current_ratio = val64;
352 }
353
354
355 msr_last = msr_now;
356 tsc_last = tsc_now;
357
358 adjust_compensation(target_ratio, win);
359
360
361
362
363 reduce_irq = atomic_read(&idle_wakeup_counter) >=
364 2 * win * num_online_cpus();
365
366 atomic_set(&idle_wakeup_counter, 0);
367
368 return set_target_ratio + guard <= current_ratio;
369}
370
371static int clamp_thread(void *arg)
372{
373 int cpunr = (unsigned long)arg;
374 DEFINE_TIMER(wakeup_timer, noop_timer, 0, 0);
375 static const struct sched_param param = {
376 .sched_priority = MAX_USER_RT_PRIO/2,
377 };
378 unsigned int count = 0;
379 unsigned int target_ratio;
380
381 set_bit(cpunr, cpu_clamping_mask);
382 set_freezable();
383 init_timer_on_stack(&wakeup_timer);
384 sched_setscheduler(current, SCHED_FIFO, ¶m);
385
386 while (true == clamping && !kthread_should_stop() &&
387 cpu_online(cpunr)) {
388 int sleeptime;
389 unsigned long target_jiffies;
390 unsigned int guard;
391 unsigned int compensated_ratio;
392 int interval;
393 unsigned int duration_jiffies = msecs_to_jiffies(duration);
394 unsigned int window_size_now;
395
396 try_to_freeze();
397
398
399
400
401
402 target_ratio = set_target_ratio;
403 guard = 1 + target_ratio/20;
404 window_size_now = window_size;
405 count++;
406
407
408
409
410
411
412 compensated_ratio = target_ratio +
413 get_compensation(target_ratio);
414 if (compensated_ratio <= 0)
415 compensated_ratio = 1;
416 interval = duration_jiffies * 100 / compensated_ratio;
417
418
419 target_jiffies = roundup(jiffies, interval);
420 sleeptime = target_jiffies - jiffies;
421 if (sleeptime <= 0)
422 sleeptime = 1;
423 schedule_timeout_interruptible(sleeptime);
424
425
426
427
428 if (cpunr == control_cpu && !(count%window_size_now)) {
429 should_skip =
430 powerclamp_adjust_controls(target_ratio,
431 guard, window_size_now);
432 smp_mb();
433 }
434
435 if (should_skip)
436 continue;
437
438 target_jiffies = jiffies + duration_jiffies;
439 mod_timer(&wakeup_timer, target_jiffies);
440 if (unlikely(local_softirq_pending()))
441 continue;
442
443
444
445
446 preempt_disable();
447
448 while (time_before(jiffies, target_jiffies)) {
449 unsigned long ecx = 1;
450 unsigned long eax = target_mwait;
451
452
453
454
455
456 local_touch_nmi();
457 stop_critical_timings();
458 mwait_idle_with_hints(eax, ecx);
459 start_critical_timings();
460 atomic_inc(&idle_wakeup_counter);
461 }
462 preempt_enable();
463 }
464 del_timer_sync(&wakeup_timer);
465 clear_bit(cpunr, cpu_clamping_mask);
466
467 return 0;
468}
469
470
471
472
473
474static void poll_pkg_cstate(struct work_struct *dummy);
475static DECLARE_DELAYED_WORK(poll_pkg_cstate_work, poll_pkg_cstate);
476static void poll_pkg_cstate(struct work_struct *dummy)
477{
478 static u64 msr_last;
479 static u64 tsc_last;
480 static unsigned long jiffies_last;
481
482 u64 msr_now;
483 unsigned long jiffies_now;
484 u64 tsc_now;
485 u64 val64;
486
487 msr_now = pkg_state_counter();
488 tsc_now = rdtsc();
489 jiffies_now = jiffies;
490
491
492 if (!msr_last || !tsc_last)
493 pkg_cstate_ratio_cur = 1;
494 else {
495 if (tsc_now - tsc_last) {
496 val64 = 100 * (msr_now - msr_last);
497 do_div(val64, (tsc_now - tsc_last));
498 pkg_cstate_ratio_cur = val64;
499 }
500 }
501
502
503 msr_last = msr_now;
504 jiffies_last = jiffies_now;
505 tsc_last = tsc_now;
506
507 if (true == clamping)
508 schedule_delayed_work(&poll_pkg_cstate_work, HZ);
509}
510
511static int start_power_clamp(void)
512{
513 unsigned long cpu;
514 struct task_struct *thread;
515
516 set_target_ratio = clamp(set_target_ratio, 0U, MAX_TARGET_RATIO - 1);
517
518 get_online_cpus();
519
520
521 control_cpu = 0;
522 if (!cpu_online(control_cpu))
523 control_cpu = smp_processor_id();
524
525 clamping = true;
526 schedule_delayed_work(&poll_pkg_cstate_work, 0);
527
528
529 for_each_online_cpu(cpu) {
530 struct task_struct **p =
531 per_cpu_ptr(powerclamp_thread, cpu);
532
533 thread = kthread_create_on_node(clamp_thread,
534 (void *) cpu,
535 cpu_to_node(cpu),
536 "kidle_inject/%ld", cpu);
537
538 if (likely(!IS_ERR(thread))) {
539 kthread_bind(thread, cpu);
540 wake_up_process(thread);
541 *p = thread;
542 }
543
544 }
545 put_online_cpus();
546
547 return 0;
548}
549
550static void end_power_clamp(void)
551{
552 int i;
553 struct task_struct *thread;
554
555 clamping = false;
556
557
558
559
560 smp_mb();
561 msleep(20);
562 if (bitmap_weight(cpu_clamping_mask, num_possible_cpus())) {
563 for_each_set_bit(i, cpu_clamping_mask, num_possible_cpus()) {
564 pr_debug("clamping thread for cpu %d alive, kill\n", i);
565 thread = *per_cpu_ptr(powerclamp_thread, i);
566 kthread_stop(thread);
567 }
568 }
569}
570
571static int powerclamp_cpu_callback(struct notifier_block *nfb,
572 unsigned long action, void *hcpu)
573{
574 unsigned long cpu = (unsigned long)hcpu;
575 struct task_struct *thread;
576 struct task_struct **percpu_thread =
577 per_cpu_ptr(powerclamp_thread, cpu);
578
579 if (false == clamping)
580 goto exit_ok;
581
582 switch (action) {
583 case CPU_ONLINE:
584 thread = kthread_create_on_node(clamp_thread,
585 (void *) cpu,
586 cpu_to_node(cpu),
587 "kidle_inject/%lu", cpu);
588 if (likely(!IS_ERR(thread))) {
589 kthread_bind(thread, cpu);
590 wake_up_process(thread);
591 *percpu_thread = thread;
592 }
593
594 if (cpu == 0) {
595 control_cpu = 0;
596 smp_mb();
597 }
598 break;
599 case CPU_DEAD:
600 if (test_bit(cpu, cpu_clamping_mask)) {
601 pr_err("cpu %lu dead but powerclamping thread is not\n",
602 cpu);
603 kthread_stop(*percpu_thread);
604 }
605 if (cpu == control_cpu) {
606 control_cpu = smp_processor_id();
607 smp_mb();
608 }
609 }
610
611exit_ok:
612 return NOTIFY_OK;
613}
614
615static struct notifier_block powerclamp_cpu_notifier = {
616 .notifier_call = powerclamp_cpu_callback,
617};
618
619static int powerclamp_get_max_state(struct thermal_cooling_device *cdev,
620 unsigned long *state)
621{
622 *state = MAX_TARGET_RATIO;
623
624 return 0;
625}
626
627static int powerclamp_get_cur_state(struct thermal_cooling_device *cdev,
628 unsigned long *state)
629{
630 if (true == clamping)
631 *state = pkg_cstate_ratio_cur;
632 else
633
634 *state = -1;
635
636 return 0;
637}
638
639static int powerclamp_set_cur_state(struct thermal_cooling_device *cdev,
640 unsigned long new_target_ratio)
641{
642 int ret = 0;
643
644 new_target_ratio = clamp(new_target_ratio, 0UL,
645 (unsigned long) (MAX_TARGET_RATIO-1));
646 if (set_target_ratio == 0 && new_target_ratio > 0) {
647 pr_info("Start idle injection to reduce power\n");
648 set_target_ratio = new_target_ratio;
649 ret = start_power_clamp();
650 goto exit_set;
651 } else if (set_target_ratio > 0 && new_target_ratio == 0) {
652 pr_info("Stop forced idle injection\n");
653 end_power_clamp();
654 set_target_ratio = 0;
655 } else {
656 set_target_ratio = new_target_ratio;
657
658 smp_mb();
659 }
660
661exit_set:
662 return ret;
663}
664
665
666static struct thermal_cooling_device_ops powerclamp_cooling_ops = {
667 .get_max_state = powerclamp_get_max_state,
668 .get_cur_state = powerclamp_get_cur_state,
669 .set_cur_state = powerclamp_set_cur_state,
670};
671
672static const struct x86_cpu_id __initconst intel_powerclamp_ids[] = {
673 { X86_VENDOR_INTEL, X86_FAMILY_ANY, X86_MODEL_ANY, X86_FEATURE_MWAIT },
674 {}
675};
676MODULE_DEVICE_TABLE(x86cpu, intel_powerclamp_ids);
677
678static int __init powerclamp_probe(void)
679{
680
681 if (!x86_match_cpu(intel_powerclamp_ids)) {
682 pr_err("CPU does not support MWAIT");
683 return -ENODEV;
684 }
685
686
687 if (!has_pkg_state_counter()) {
688 pr_info("No package C-state available");
689 return -ENODEV;
690 }
691
692
693 find_target_mwait();
694
695 return 0;
696}
697
698static int powerclamp_debug_show(struct seq_file *m, void *unused)
699{
700 int i = 0;
701
702 seq_printf(m, "controlling cpu: %d\n", control_cpu);
703 seq_printf(m, "pct confidence steady dynamic (compensation)\n");
704 for (i = 0; i < MAX_TARGET_RATIO; i++) {
705 seq_printf(m, "%d\t%lu\t%lu\t%lu\n",
706 i,
707 cal_data[i].confidence,
708 cal_data[i].steady_comp,
709 cal_data[i].dynamic_comp);
710 }
711
712 return 0;
713}
714
715static int powerclamp_debug_open(struct inode *inode,
716 struct file *file)
717{
718 return single_open(file, powerclamp_debug_show, inode->i_private);
719}
720
721static const struct file_operations powerclamp_debug_fops = {
722 .open = powerclamp_debug_open,
723 .read = seq_read,
724 .llseek = seq_lseek,
725 .release = single_release,
726 .owner = THIS_MODULE,
727};
728
729static inline void powerclamp_create_debug_files(void)
730{
731 debug_dir = debugfs_create_dir("intel_powerclamp", NULL);
732 if (!debug_dir)
733 return;
734
735 if (!debugfs_create_file("powerclamp_calib", S_IRUGO, debug_dir,
736 cal_data, &powerclamp_debug_fops))
737 goto file_error;
738
739 return;
740
741file_error:
742 debugfs_remove_recursive(debug_dir);
743}
744
745static int __init powerclamp_init(void)
746{
747 int retval;
748 int bitmap_size;
749
750 bitmap_size = BITS_TO_LONGS(num_possible_cpus()) * sizeof(long);
751 cpu_clamping_mask = kzalloc(bitmap_size, GFP_KERNEL);
752 if (!cpu_clamping_mask)
753 return -ENOMEM;
754
755
756 retval = powerclamp_probe();
757 if (retval)
758 goto exit_free;
759
760
761 window_size = 2;
762 register_hotcpu_notifier(&powerclamp_cpu_notifier);
763
764 powerclamp_thread = alloc_percpu(struct task_struct *);
765 if (!powerclamp_thread) {
766 retval = -ENOMEM;
767 goto exit_unregister;
768 }
769
770 cooling_dev = thermal_cooling_device_register("intel_powerclamp", NULL,
771 &powerclamp_cooling_ops);
772 if (IS_ERR(cooling_dev)) {
773 retval = -ENODEV;
774 goto exit_free_thread;
775 }
776
777 if (!duration)
778 duration = jiffies_to_msecs(DEFAULT_DURATION_JIFFIES);
779
780 powerclamp_create_debug_files();
781
782 return 0;
783
784exit_free_thread:
785 free_percpu(powerclamp_thread);
786exit_unregister:
787 unregister_hotcpu_notifier(&powerclamp_cpu_notifier);
788exit_free:
789 kfree(cpu_clamping_mask);
790 return retval;
791}
792module_init(powerclamp_init);
793
794static void __exit powerclamp_exit(void)
795{
796 unregister_hotcpu_notifier(&powerclamp_cpu_notifier);
797 end_power_clamp();
798 free_percpu(powerclamp_thread);
799 thermal_cooling_device_unregister(cooling_dev);
800 kfree(cpu_clamping_mask);
801
802 cancel_delayed_work_sync(&poll_pkg_cstate_work);
803 debugfs_remove_recursive(debug_dir);
804}
805module_exit(powerclamp_exit);
806
807MODULE_LICENSE("GPL");
808MODULE_AUTHOR("Arjan van de Ven <arjan@linux.intel.com>");
809MODULE_AUTHOR("Jacob Pan <jacob.jun.pan@linux.intel.com>");
810MODULE_DESCRIPTION("Package Level C-state Idle Injection for Intel CPUs");
811