2 * intel_powerclamp.c - package c-state idle injection
4 * Copyright (c) 2012, Intel Corporation.
7 * Arjan van de Ven <arjan@linux.intel.com>
8 * Jacob Pan <jacob.jun.pan@linux.intel.com>
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.
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
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.
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()
35 * 2. synchronization with other hw blocks
40 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
42 #include <linux/module.h>
43 #include <linux/kernel.h>
44 #include <linux/delay.h>
45 #include <linux/kthread.h>
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>
52 #include <linux/sched/rt.h>
56 #include <asm/mwait.h>
57 #include <asm/cpu_device_id.h>
59 #include <asm/hardirq.h>
61 #define MAX_TARGET_RATIO (50U)
62 /* For each undisturbed clamping period (no extra wake ups during idle time),
63 * we increment the confidence counter for the given target ratio.
64 * CONFIDENCE_OK defines the level where runtime calibration results are
67 #define CONFIDENCE_OK (3)
68 /* Default idle injection duration, driver adjust sleep time to meet target
69 * idle ratio. Similar to frequency modulation.
71 #define DEFAULT_DURATION_JIFFIES (6)
73 static unsigned int target_mwait;
74 static struct dentry *debug_dir;
76 /* user selected target */
77 static unsigned int set_target_ratio;
78 static unsigned int current_ratio;
79 static bool should_skip;
80 static bool reduce_irq;
81 static atomic_t idle_wakeup_counter;
82 static unsigned int control_cpu; /* The cpu assigned to collect stat and update
83 * control parameters. default to BSP but BSP
88 static const struct sched_param sparam = {
89 .sched_priority = MAX_USER_RT_PRIO / 2,
91 struct powerclamp_worker_data {
92 struct kthread_worker *worker;
93 struct kthread_work balancing_work;
94 struct kthread_delayed_work idle_injection_work;
98 unsigned int window_size_now;
99 unsigned int target_ratio;
100 unsigned int duration_jiffies;
104 static struct powerclamp_worker_data * __percpu worker_data;
105 static struct thermal_cooling_device *cooling_dev;
106 static unsigned long *cpu_clamping_mask; /* bit map for tracking per cpu
107 * clamping kthread worker
110 static unsigned int duration;
111 static unsigned int pkg_cstate_ratio_cur;
112 static unsigned int window_size;
114 static int duration_set(const char *arg, const struct kernel_param *kp)
117 unsigned long new_duration;
119 ret = kstrtoul(arg, 10, &new_duration);
122 if (new_duration > 25 || new_duration < 6) {
123 pr_err("Out of recommended range %lu, between 6-25ms\n",
128 duration = clamp(new_duration, 6ul, 25ul);
136 static const struct kernel_param_ops duration_ops = {
138 .get = param_get_int,
142 module_param_cb(duration, &duration_ops, &duration, 0644);
143 MODULE_PARM_DESC(duration, "forced idle time for each attempt in msec.");
145 struct powerclamp_calibration_data {
146 unsigned long confidence; /* used for calibration, basically a counter
147 * gets incremented each time a clamping
148 * period is completed without extra wakeups
149 * once that counter is reached given level,
150 * compensation is deemed usable.
152 unsigned long steady_comp; /* steady state compensation used when
153 * no extra wakeups occurred.
155 unsigned long dynamic_comp; /* compensate excessive wakeup from idle
156 * mostly from external interrupts.
160 static struct powerclamp_calibration_data cal_data[MAX_TARGET_RATIO];
162 static int window_size_set(const char *arg, const struct kernel_param *kp)
165 unsigned long new_window_size;
167 ret = kstrtoul(arg, 10, &new_window_size);
170 if (new_window_size > 10 || new_window_size < 2) {
171 pr_err("Out of recommended window size %lu, between 2-10\n",
176 window_size = clamp(new_window_size, 2ul, 10ul);
184 static const struct kernel_param_ops window_size_ops = {
185 .set = window_size_set,
186 .get = param_get_int,
189 module_param_cb(window_size, &window_size_ops, &window_size, 0644);
190 MODULE_PARM_DESC(window_size, "sliding window in number of clamping cycles\n"
191 "\tpowerclamp controls idle ratio within this window. larger\n"
192 "\twindow size results in slower response time but more smooth\n"
193 "\tclamping results. default to 2.");
195 static void find_target_mwait(void)
197 unsigned int eax, ebx, ecx, edx;
198 unsigned int highest_cstate = 0;
199 unsigned int highest_subcstate = 0;
202 if (boot_cpu_data.cpuid_level < CPUID_MWAIT_LEAF)
205 cpuid(CPUID_MWAIT_LEAF, &eax, &ebx, &ecx, &edx);
207 if (!(ecx & CPUID5_ECX_EXTENSIONS_SUPPORTED) ||
208 !(ecx & CPUID5_ECX_INTERRUPT_BREAK))
211 edx >>= MWAIT_SUBSTATE_SIZE;
212 for (i = 0; i < 7 && edx; i++, edx >>= MWAIT_SUBSTATE_SIZE) {
213 if (edx & MWAIT_SUBSTATE_MASK) {
215 highest_subcstate = edx & MWAIT_SUBSTATE_MASK;
218 target_mwait = (highest_cstate << MWAIT_SUBSTATE_SIZE) |
219 (highest_subcstate - 1);
223 struct pkg_cstate_info {
229 #define PKG_CSTATE_INIT(id) { \
230 .msr_index = MSR_PKG_C##id##_RESIDENCY, \
234 static struct pkg_cstate_info pkg_cstates[] = {
245 static bool has_pkg_state_counter(void)
248 struct pkg_cstate_info *info = pkg_cstates;
250 /* check if any one of the counter msrs exists */
251 while (info->msr_index) {
252 if (!rdmsrl_safe(info->msr_index, &val))
260 static u64 pkg_state_counter(void)
264 struct pkg_cstate_info *info = pkg_cstates;
266 while (info->msr_index) {
268 if (!rdmsrl_safe(info->msr_index, &val))
279 static unsigned int get_compensation(int ratio)
281 unsigned int comp = 0;
283 /* we only use compensation if all adjacent ones are good */
285 cal_data[ratio].confidence >= CONFIDENCE_OK &&
286 cal_data[ratio + 1].confidence >= CONFIDENCE_OK &&
287 cal_data[ratio + 2].confidence >= CONFIDENCE_OK) {
288 comp = (cal_data[ratio].steady_comp +
289 cal_data[ratio + 1].steady_comp +
290 cal_data[ratio + 2].steady_comp) / 3;
291 } else if (ratio == MAX_TARGET_RATIO - 1 &&
292 cal_data[ratio].confidence >= CONFIDENCE_OK &&
293 cal_data[ratio - 1].confidence >= CONFIDENCE_OK &&
294 cal_data[ratio - 2].confidence >= CONFIDENCE_OK) {
295 comp = (cal_data[ratio].steady_comp +
296 cal_data[ratio - 1].steady_comp +
297 cal_data[ratio - 2].steady_comp) / 3;
298 } else if (cal_data[ratio].confidence >= CONFIDENCE_OK &&
299 cal_data[ratio - 1].confidence >= CONFIDENCE_OK &&
300 cal_data[ratio + 1].confidence >= CONFIDENCE_OK) {
301 comp = (cal_data[ratio].steady_comp +
302 cal_data[ratio - 1].steady_comp +
303 cal_data[ratio + 1].steady_comp) / 3;
306 /* REVISIT: simple penalty of double idle injection */
309 /* do not exceed limit */
310 if (comp + ratio >= MAX_TARGET_RATIO)
311 comp = MAX_TARGET_RATIO - ratio - 1;
316 static void adjust_compensation(int target_ratio, unsigned int win)
319 struct powerclamp_calibration_data *d = &cal_data[target_ratio];
322 * adjust compensations if confidence level has not been reached or
323 * there are too many wakeups during the last idle injection period, we
324 * cannot trust the data for compensation.
326 if (d->confidence >= CONFIDENCE_OK ||
327 atomic_read(&idle_wakeup_counter) >
328 win * num_online_cpus())
331 delta = set_target_ratio - current_ratio;
332 /* filter out bad data */
333 if (delta >= 0 && delta <= (1+target_ratio/10)) {
336 roundup(delta+d->steady_comp, 2)/2;
338 d->steady_comp = delta;
343 static bool powerclamp_adjust_controls(unsigned int target_ratio,
344 unsigned int guard, unsigned int win)
346 static u64 msr_last, tsc_last;
347 u64 msr_now, tsc_now;
350 /* check result for the last window */
351 msr_now = pkg_state_counter();
354 /* calculate pkg cstate vs tsc ratio */
355 if (!msr_last || !tsc_last)
357 else if (tsc_now-tsc_last) {
358 val64 = 100*(msr_now-msr_last);
359 do_div(val64, (tsc_now-tsc_last));
360 current_ratio = val64;
367 adjust_compensation(target_ratio, win);
369 * too many external interrupts, set flag such
370 * that we can take measure later.
372 reduce_irq = atomic_read(&idle_wakeup_counter) >=
373 2 * win * num_online_cpus();
375 atomic_set(&idle_wakeup_counter, 0);
376 /* if we are above target+guard, skip */
377 return set_target_ratio + guard <= current_ratio;
380 static void clamp_balancing_func(struct kthread_work *work)
382 struct powerclamp_worker_data *w_data;
384 unsigned long target_jiffies;
385 unsigned int compensated_ratio;
386 int interval; /* jiffies to sleep for each attempt */
388 w_data = container_of(work, struct powerclamp_worker_data,
392 * make sure user selected ratio does not take effect until
393 * the next round. adjust target_ratio if user has changed
394 * target such that we can converge quickly.
396 w_data->target_ratio = READ_ONCE(set_target_ratio);
397 w_data->guard = 1 + w_data->target_ratio / 20;
398 w_data->window_size_now = window_size;
399 w_data->duration_jiffies = msecs_to_jiffies(duration);
403 * systems may have different ability to enter package level
404 * c-states, thus we need to compensate the injected idle ratio
405 * to achieve the actual target reported by the HW.
407 compensated_ratio = w_data->target_ratio +
408 get_compensation(w_data->target_ratio);
409 if (compensated_ratio <= 0)
410 compensated_ratio = 1;
411 interval = w_data->duration_jiffies * 100 / compensated_ratio;
413 /* align idle time */
414 target_jiffies = roundup(jiffies, interval);
415 sleeptime = target_jiffies - jiffies;
419 if (clamping && w_data->clamping && cpu_online(w_data->cpu))
420 kthread_queue_delayed_work(w_data->worker,
421 &w_data->idle_injection_work,
425 static void clamp_idle_injection_func(struct kthread_work *work)
427 struct powerclamp_worker_data *w_data;
429 w_data = container_of(work, struct powerclamp_worker_data,
430 idle_injection_work.work);
433 * only elected controlling cpu can collect stats and update
434 * control parameters.
436 if (w_data->cpu == control_cpu &&
437 !(w_data->count % w_data->window_size_now)) {
439 powerclamp_adjust_controls(w_data->target_ratio,
441 w_data->window_size_now);
448 play_idle(jiffies_to_msecs(w_data->duration_jiffies));
451 if (clamping && w_data->clamping && cpu_online(w_data->cpu))
452 kthread_queue_work(w_data->worker, &w_data->balancing_work);
456 * 1 HZ polling while clamping is active, useful for userspace
457 * to monitor actual idle ratio.
459 static void poll_pkg_cstate(struct work_struct *dummy);
460 static DECLARE_DELAYED_WORK(poll_pkg_cstate_work, poll_pkg_cstate);
461 static void poll_pkg_cstate(struct work_struct *dummy)
465 static unsigned long jiffies_last;
468 unsigned long jiffies_now;
472 msr_now = pkg_state_counter();
474 jiffies_now = jiffies;
476 /* calculate pkg cstate vs tsc ratio */
477 if (!msr_last || !tsc_last)
478 pkg_cstate_ratio_cur = 1;
480 if (tsc_now - tsc_last) {
481 val64 = 100 * (msr_now - msr_last);
482 do_div(val64, (tsc_now - tsc_last));
483 pkg_cstate_ratio_cur = val64;
489 jiffies_last = jiffies_now;
492 if (true == clamping)
493 schedule_delayed_work(&poll_pkg_cstate_work, HZ);
496 static void start_power_clamp_worker(unsigned long cpu)
498 struct powerclamp_worker_data *w_data = per_cpu_ptr(worker_data, cpu);
499 struct kthread_worker *worker;
501 worker = kthread_create_worker_on_cpu(cpu, 0, "kidle_inject/%ld", cpu);
505 w_data->worker = worker;
508 w_data->clamping = true;
509 set_bit(cpu, cpu_clamping_mask);
510 sched_setscheduler(worker->task, SCHED_FIFO, &sparam);
511 kthread_init_work(&w_data->balancing_work, clamp_balancing_func);
512 kthread_init_delayed_work(&w_data->idle_injection_work,
513 clamp_idle_injection_func);
514 kthread_queue_work(w_data->worker, &w_data->balancing_work);
517 static void stop_power_clamp_worker(unsigned long cpu)
519 struct powerclamp_worker_data *w_data = per_cpu_ptr(worker_data, cpu);
524 w_data->clamping = false;
526 * Make sure that all works that get queued after this point see
527 * the clamping disabled. The counter part is not needed because
528 * there is an implicit memory barrier when the queued work
532 kthread_cancel_work_sync(&w_data->balancing_work);
533 kthread_cancel_delayed_work_sync(&w_data->idle_injection_work);
535 * The balancing work still might be queued here because
536 * the handling of the "clapming" variable, cancel, and queue
537 * operations are not synchronized via a lock. But it is not
538 * a big deal. The balancing work is fast and destroy kthread
541 clear_bit(w_data->cpu, cpu_clamping_mask);
542 kthread_destroy_worker(w_data->worker);
544 w_data->worker = NULL;
547 static int start_power_clamp(void)
551 set_target_ratio = clamp(set_target_ratio, 0U, MAX_TARGET_RATIO - 1);
552 /* prevent cpu hotplug */
557 if (!cpu_online(control_cpu))
558 control_cpu = smp_processor_id();
561 schedule_delayed_work(&poll_pkg_cstate_work, 0);
563 /* start one kthread worker per online cpu */
564 for_each_online_cpu(cpu) {
565 start_power_clamp_worker(cpu);
572 static void end_power_clamp(void)
577 * Block requeuing in all the kthread workers. They will flush and
581 if (bitmap_weight(cpu_clamping_mask, num_possible_cpus())) {
582 for_each_set_bit(i, cpu_clamping_mask, num_possible_cpus()) {
583 pr_debug("clamping worker for cpu %d alive, destroy\n",
585 stop_power_clamp_worker(i);
590 static int powerclamp_cpu_online(unsigned int cpu)
592 if (clamping == false)
594 start_power_clamp_worker(cpu);
595 /* prefer BSP as controlling CPU */
603 static int powerclamp_cpu_predown(unsigned int cpu)
605 if (clamping == false)
608 stop_power_clamp_worker(cpu);
609 if (cpu != control_cpu)
612 control_cpu = cpumask_first(cpu_online_mask);
613 if (control_cpu == cpu)
614 control_cpu = cpumask_next(cpu, cpu_online_mask);
619 static int powerclamp_get_max_state(struct thermal_cooling_device *cdev,
620 unsigned long *state)
622 *state = MAX_TARGET_RATIO;
627 static int powerclamp_get_cur_state(struct thermal_cooling_device *cdev,
628 unsigned long *state)
630 if (true == clamping)
631 *state = pkg_cstate_ratio_cur;
633 /* to save power, do not poll idle ratio while not clamping */
634 *state = -1; /* indicates invalid state */
639 static int powerclamp_set_cur_state(struct thermal_cooling_device *cdev,
640 unsigned long new_target_ratio)
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();
651 } else if (set_target_ratio > 0 && new_target_ratio == 0) {
652 pr_info("Stop forced idle injection\n");
654 set_target_ratio = 0;
655 } else /* adjust currently running */ {
656 set_target_ratio = new_target_ratio;
657 /* make new set_target_ratio visible to other cpus */
665 /* bind to generic thermal layer as cooling device*/
666 static 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,
672 static const struct x86_cpu_id __initconst intel_powerclamp_ids[] = {
673 { X86_VENDOR_INTEL, X86_FAMILY_ANY, X86_MODEL_ANY, X86_FEATURE_MWAIT },
676 MODULE_DEVICE_TABLE(x86cpu, intel_powerclamp_ids);
678 static int __init powerclamp_probe(void)
681 if (!x86_match_cpu(intel_powerclamp_ids)) {
682 pr_err("CPU does not support MWAIT");
686 /* The goal for idle time alignment is to achieve package cstate. */
687 if (!has_pkg_state_counter()) {
688 pr_info("No package C-state available");
692 /* find the deepest mwait value */
698 static int powerclamp_debug_show(struct seq_file *m, void *unused)
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",
707 cal_data[i].confidence,
708 cal_data[i].steady_comp,
709 cal_data[i].dynamic_comp);
715 static int powerclamp_debug_open(struct inode *inode,
718 return single_open(file, powerclamp_debug_show, inode->i_private);
721 static const struct file_operations powerclamp_debug_fops = {
722 .open = powerclamp_debug_open,
725 .release = single_release,
726 .owner = THIS_MODULE,
729 static inline void powerclamp_create_debug_files(void)
731 debug_dir = debugfs_create_dir("intel_powerclamp", NULL);
735 if (!debugfs_create_file("powerclamp_calib", S_IRUGO, debug_dir,
736 cal_data, &powerclamp_debug_fops))
742 debugfs_remove_recursive(debug_dir);
745 static enum cpuhp_state hp_state;
747 static int __init powerclamp_init(void)
752 bitmap_size = BITS_TO_LONGS(num_possible_cpus()) * sizeof(long);
753 cpu_clamping_mask = kzalloc(bitmap_size, GFP_KERNEL);
754 if (!cpu_clamping_mask)
757 /* probe cpu features and ids here */
758 retval = powerclamp_probe();
762 /* set default limit, maybe adjusted during runtime based on feedback */
764 retval = cpuhp_setup_state_nocalls(CPUHP_AP_ONLINE_DYN,
765 "thermal/intel_powerclamp:online",
766 powerclamp_cpu_online,
767 powerclamp_cpu_predown);
773 worker_data = alloc_percpu(struct powerclamp_worker_data);
776 goto exit_unregister;
779 cooling_dev = thermal_cooling_device_register("intel_powerclamp", NULL,
780 &powerclamp_cooling_ops);
781 if (IS_ERR(cooling_dev)) {
783 goto exit_free_thread;
787 duration = jiffies_to_msecs(DEFAULT_DURATION_JIFFIES);
789 powerclamp_create_debug_files();
794 free_percpu(worker_data);
796 cpuhp_remove_state_nocalls(hp_state);
798 kfree(cpu_clamping_mask);
801 module_init(powerclamp_init);
803 static void __exit powerclamp_exit(void)
806 cpuhp_remove_state_nocalls(hp_state);
807 free_percpu(worker_data);
808 thermal_cooling_device_unregister(cooling_dev);
809 kfree(cpu_clamping_mask);
811 cancel_delayed_work_sync(&poll_pkg_cstate_work);
812 debugfs_remove_recursive(debug_dir);
814 module_exit(powerclamp_exit);
816 MODULE_LICENSE("GPL");
817 MODULE_AUTHOR("Arjan van de Ven <arjan@linux.intel.com>");
818 MODULE_AUTHOR("Jacob Pan <jacob.jun.pan@linux.intel.com>");
819 MODULE_DESCRIPTION("Package Level C-state Idle Injection for Intel CPUs");