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/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>
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
89 static struct task_struct * __percpu *powerclamp_thread;
90 static struct thermal_cooling_device *cooling_dev;
91 static unsigned long *cpu_clamping_mask; /* bit map for tracking per cpu
95 static unsigned int duration;
96 static unsigned int pkg_cstate_ratio_cur;
97 static unsigned int window_size;
99 static int duration_set(const char *arg, const struct kernel_param *kp)
102 unsigned long new_duration;
104 ret = kstrtoul(arg, 10, &new_duration);
107 if (new_duration > 25 || new_duration < 6) {
108 pr_err("Out of recommended range %lu, between 6-25ms\n",
113 duration = clamp(new_duration, 6ul, 25ul);
121 static struct kernel_param_ops duration_ops = {
123 .get = param_get_int,
127 module_param_cb(duration, &duration_ops, &duration, 0644);
128 MODULE_PARM_DESC(duration, "forced idle time for each attempt in msec.");
130 struct powerclamp_calibration_data {
131 unsigned long confidence; /* used for calibration, basically a counter
132 * gets incremented each time a clamping
133 * period is completed without extra wakeups
134 * once that counter is reached given level,
135 * compensation is deemed usable.
137 unsigned long steady_comp; /* steady state compensation used when
138 * no extra wakeups occurred.
140 unsigned long dynamic_comp; /* compensate excessive wakeup from idle
141 * mostly from external interrupts.
145 static struct powerclamp_calibration_data cal_data[MAX_TARGET_RATIO];
147 static int window_size_set(const char *arg, const struct kernel_param *kp)
150 unsigned long new_window_size;
152 ret = kstrtoul(arg, 10, &new_window_size);
155 if (new_window_size > 10 || new_window_size < 2) {
156 pr_err("Out of recommended window size %lu, between 2-10\n",
161 window_size = clamp(new_window_size, 2ul, 10ul);
169 static struct kernel_param_ops window_size_ops = {
170 .set = window_size_set,
171 .get = param_get_int,
174 module_param_cb(window_size, &window_size_ops, &window_size, 0644);
175 MODULE_PARM_DESC(window_size, "sliding window in number of clamping cycles\n"
176 "\tpowerclamp controls idle ratio within this window. larger\n"
177 "\twindow size results in slower response time but more smooth\n"
178 "\tclamping results. default to 2.");
180 static void find_target_mwait(void)
182 unsigned int eax, ebx, ecx, edx;
183 unsigned int highest_cstate = 0;
184 unsigned int highest_subcstate = 0;
187 if (boot_cpu_data.cpuid_level < CPUID_MWAIT_LEAF)
190 cpuid(CPUID_MWAIT_LEAF, &eax, &ebx, &ecx, &edx);
192 if (!(ecx & CPUID5_ECX_EXTENSIONS_SUPPORTED) ||
193 !(ecx & CPUID5_ECX_INTERRUPT_BREAK))
196 edx >>= MWAIT_SUBSTATE_SIZE;
197 for (i = 0; i < 7 && edx; i++, edx >>= MWAIT_SUBSTATE_SIZE) {
198 if (edx & MWAIT_SUBSTATE_MASK) {
200 highest_subcstate = edx & MWAIT_SUBSTATE_MASK;
203 target_mwait = (highest_cstate << MWAIT_SUBSTATE_SIZE) |
204 (highest_subcstate - 1);
208 static u64 pkg_state_counter(void)
219 if (!rdmsrl_safe(MSR_PKG_C2_RESIDENCY, &val))
226 if (!rdmsrl_safe(MSR_PKG_C3_RESIDENCY, &val))
233 if (!rdmsrl_safe(MSR_PKG_C6_RESIDENCY, &val))
240 if (!rdmsrl_safe(MSR_PKG_C7_RESIDENCY, &val))
249 static void noop_timer(unsigned long foo)
251 /* empty... just the fact that we get the interrupt wakes us up */
254 static unsigned int get_compensation(int ratio)
256 unsigned int comp = 0;
258 /* we only use compensation if all adjacent ones are good */
260 cal_data[ratio].confidence >= CONFIDENCE_OK &&
261 cal_data[ratio + 1].confidence >= CONFIDENCE_OK &&
262 cal_data[ratio + 2].confidence >= CONFIDENCE_OK) {
263 comp = (cal_data[ratio].steady_comp +
264 cal_data[ratio + 1].steady_comp +
265 cal_data[ratio + 2].steady_comp) / 3;
266 } else if (ratio == MAX_TARGET_RATIO - 1 &&
267 cal_data[ratio].confidence >= CONFIDENCE_OK &&
268 cal_data[ratio - 1].confidence >= CONFIDENCE_OK &&
269 cal_data[ratio - 2].confidence >= CONFIDENCE_OK) {
270 comp = (cal_data[ratio].steady_comp +
271 cal_data[ratio - 1].steady_comp +
272 cal_data[ratio - 2].steady_comp) / 3;
273 } else if (cal_data[ratio].confidence >= CONFIDENCE_OK &&
274 cal_data[ratio - 1].confidence >= CONFIDENCE_OK &&
275 cal_data[ratio + 1].confidence >= CONFIDENCE_OK) {
276 comp = (cal_data[ratio].steady_comp +
277 cal_data[ratio - 1].steady_comp +
278 cal_data[ratio + 1].steady_comp) / 3;
281 /* REVISIT: simple penalty of double idle injection */
284 /* do not exceed limit */
285 if (comp + ratio >= MAX_TARGET_RATIO)
286 comp = MAX_TARGET_RATIO - ratio - 1;
291 static void adjust_compensation(int target_ratio, unsigned int win)
294 struct powerclamp_calibration_data *d = &cal_data[target_ratio];
297 * adjust compensations if confidence level has not been reached or
298 * there are too many wakeups during the last idle injection period, we
299 * cannot trust the data for compensation.
301 if (d->confidence >= CONFIDENCE_OK ||
302 atomic_read(&idle_wakeup_counter) >
303 win * num_online_cpus())
306 delta = set_target_ratio - current_ratio;
307 /* filter out bad data */
308 if (delta >= 0 && delta <= (1+target_ratio/10)) {
311 roundup(delta+d->steady_comp, 2)/2;
313 d->steady_comp = delta;
318 static bool powerclamp_adjust_controls(unsigned int target_ratio,
319 unsigned int guard, unsigned int win)
321 static u64 msr_last, tsc_last;
322 u64 msr_now, tsc_now;
325 /* check result for the last window */
326 msr_now = pkg_state_counter();
329 /* calculate pkg cstate vs tsc ratio */
330 if (!msr_last || !tsc_last)
332 else if (tsc_now-tsc_last) {
333 val64 = 100*(msr_now-msr_last);
334 do_div(val64, (tsc_now-tsc_last));
335 current_ratio = val64;
342 adjust_compensation(target_ratio, win);
344 * too many external interrupts, set flag such
345 * that we can take measure later.
347 reduce_irq = atomic_read(&idle_wakeup_counter) >=
348 2 * win * num_online_cpus();
350 atomic_set(&idle_wakeup_counter, 0);
351 /* if we are above target+guard, skip */
352 return set_target_ratio + guard <= current_ratio;
355 static int clamp_thread(void *arg)
357 int cpunr = (unsigned long)arg;
358 DEFINE_TIMER(wakeup_timer, noop_timer, 0, 0);
359 static const struct sched_param param = {
360 .sched_priority = MAX_USER_RT_PRIO/2,
362 unsigned int count = 0;
363 unsigned int target_ratio;
365 set_bit(cpunr, cpu_clamping_mask);
367 init_timer_on_stack(&wakeup_timer);
368 sched_setscheduler(current, SCHED_FIFO, ¶m);
370 while (true == clamping && !kthread_should_stop() &&
373 unsigned long target_jiffies;
375 unsigned int compensation = 0;
376 int interval; /* jiffies to sleep for each attempt */
377 unsigned int duration_jiffies = msecs_to_jiffies(duration);
378 unsigned int window_size_now;
382 * make sure user selected ratio does not take effect until
383 * the next round. adjust target_ratio if user has changed
384 * target such that we can converge quickly.
386 target_ratio = set_target_ratio;
387 guard = 1 + target_ratio/20;
388 window_size_now = window_size;
392 * systems may have different ability to enter package level
393 * c-states, thus we need to compensate the injected idle ratio
394 * to achieve the actual target reported by the HW.
396 compensation = get_compensation(target_ratio);
397 interval = duration_jiffies*100/(target_ratio+compensation);
399 /* align idle time */
400 target_jiffies = roundup(jiffies, interval);
401 sleeptime = target_jiffies - jiffies;
404 schedule_timeout_interruptible(sleeptime);
406 * only elected controlling cpu can collect stats and update
407 * control parameters.
409 if (cpunr == control_cpu && !(count%window_size_now)) {
411 powerclamp_adjust_controls(target_ratio,
412 guard, window_size_now);
419 target_jiffies = jiffies + duration_jiffies;
420 mod_timer(&wakeup_timer, target_jiffies);
421 if (unlikely(local_softirq_pending()))
424 * stop tick sched during idle time, interrupts are still
425 * allowed. thus jiffies are updated properly.
428 tick_nohz_idle_enter();
429 /* mwait until target jiffies is reached */
430 while (time_before(jiffies, target_jiffies)) {
431 unsigned long ecx = 1;
432 unsigned long eax = target_mwait;
435 * REVISIT: may call enter_idle() to notify drivers who
436 * can save power during cpu idle. same for exit_idle()
439 stop_critical_timings();
440 __monitor((void *)¤t_thread_info()->flags, 0, 0);
441 cpu_relax(); /* allow HT sibling to run */
443 start_critical_timings();
444 atomic_inc(&idle_wakeup_counter);
446 tick_nohz_idle_exit();
447 preempt_enable_no_resched();
449 del_timer_sync(&wakeup_timer);
450 clear_bit(cpunr, cpu_clamping_mask);
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 int start_power_clamp(void)
499 struct task_struct *thread;
501 /* check if pkg cstate counter is completely 0, abort in this case */
502 if (!pkg_state_counter()) {
503 pr_err("pkg cstate counter not functional, abort\n");
507 set_target_ratio = clamp(set_target_ratio, 0U, MAX_TARGET_RATIO - 1);
508 /* prevent cpu hotplug */
513 if (!cpu_online(control_cpu))
514 control_cpu = smp_processor_id();
517 schedule_delayed_work(&poll_pkg_cstate_work, 0);
519 /* start one thread per online cpu */
520 for_each_online_cpu(cpu) {
521 struct task_struct **p =
522 per_cpu_ptr(powerclamp_thread, cpu);
524 thread = kthread_create_on_node(clamp_thread,
527 "kidle_inject/%ld", cpu);
528 /* bind to cpu here */
529 if (likely(!IS_ERR(thread))) {
530 kthread_bind(thread, cpu);
531 wake_up_process(thread);
541 static void end_power_clamp(void)
544 struct task_struct *thread;
548 * make clamping visible to other cpus and give per cpu clamping threads
549 * sometime to exit, or gets killed later.
553 if (bitmap_weight(cpu_clamping_mask, num_possible_cpus())) {
554 for_each_set_bit(i, cpu_clamping_mask, num_possible_cpus()) {
555 pr_debug("clamping thread for cpu %d alive, kill\n", i);
556 thread = *per_cpu_ptr(powerclamp_thread, i);
557 kthread_stop(thread);
562 static int powerclamp_cpu_callback(struct notifier_block *nfb,
563 unsigned long action, void *hcpu)
565 unsigned long cpu = (unsigned long)hcpu;
566 struct task_struct *thread;
567 struct task_struct **percpu_thread =
568 per_cpu_ptr(powerclamp_thread, cpu);
570 if (false == clamping)
575 thread = kthread_create_on_node(clamp_thread,
578 "kidle_inject/%lu", cpu);
579 if (likely(!IS_ERR(thread))) {
580 kthread_bind(thread, cpu);
581 wake_up_process(thread);
582 *percpu_thread = thread;
584 /* prefer BSP as controlling CPU */
591 if (test_bit(cpu, cpu_clamping_mask)) {
592 pr_err("cpu %lu dead but powerclamping thread is not\n",
594 kthread_stop(*percpu_thread);
596 if (cpu == control_cpu) {
597 control_cpu = smp_processor_id();
606 static struct notifier_block powerclamp_cpu_notifier = {
607 .notifier_call = powerclamp_cpu_callback,
610 static int powerclamp_get_max_state(struct thermal_cooling_device *cdev,
611 unsigned long *state)
613 *state = MAX_TARGET_RATIO;
618 static int powerclamp_get_cur_state(struct thermal_cooling_device *cdev,
619 unsigned long *state)
621 if (true == clamping)
622 *state = pkg_cstate_ratio_cur;
624 /* to save power, do not poll idle ratio while not clamping */
625 *state = -1; /* indicates invalid state */
630 static int powerclamp_set_cur_state(struct thermal_cooling_device *cdev,
631 unsigned long new_target_ratio)
635 new_target_ratio = clamp(new_target_ratio, 0UL,
636 (unsigned long) (MAX_TARGET_RATIO-1));
637 if (set_target_ratio == 0 && new_target_ratio > 0) {
638 pr_info("Start idle injection to reduce power\n");
639 set_target_ratio = new_target_ratio;
640 ret = start_power_clamp();
642 } else if (set_target_ratio > 0 && new_target_ratio == 0) {
643 pr_info("Stop forced idle injection\n");
644 set_target_ratio = 0;
646 } else /* adjust currently running */ {
647 set_target_ratio = new_target_ratio;
648 /* make new set_target_ratio visible to other cpus */
656 /* bind to generic thermal layer as cooling device*/
657 static struct thermal_cooling_device_ops powerclamp_cooling_ops = {
658 .get_max_state = powerclamp_get_max_state,
659 .get_cur_state = powerclamp_get_cur_state,
660 .set_cur_state = powerclamp_set_cur_state,
663 /* runs on Nehalem and later */
664 static const struct x86_cpu_id intel_powerclamp_ids[] = {
665 { X86_VENDOR_INTEL, 6, 0x1a},
666 { X86_VENDOR_INTEL, 6, 0x1c},
667 { X86_VENDOR_INTEL, 6, 0x1e},
668 { X86_VENDOR_INTEL, 6, 0x1f},
669 { X86_VENDOR_INTEL, 6, 0x25},
670 { X86_VENDOR_INTEL, 6, 0x26},
671 { X86_VENDOR_INTEL, 6, 0x2a},
672 { X86_VENDOR_INTEL, 6, 0x2c},
673 { X86_VENDOR_INTEL, 6, 0x2d},
674 { X86_VENDOR_INTEL, 6, 0x2e},
675 { X86_VENDOR_INTEL, 6, 0x2f},
676 { X86_VENDOR_INTEL, 6, 0x3a},
679 MODULE_DEVICE_TABLE(x86cpu, intel_powerclamp_ids);
681 static int powerclamp_probe(void)
683 if (!x86_match_cpu(intel_powerclamp_ids)) {
684 pr_err("Intel powerclamp does not run on family %d model %d\n",
685 boot_cpu_data.x86, boot_cpu_data.x86_model);
688 if (!boot_cpu_has(X86_FEATURE_NONSTOP_TSC) ||
689 !boot_cpu_has(X86_FEATURE_CONSTANT_TSC) ||
690 !boot_cpu_has(X86_FEATURE_MWAIT) ||
691 !boot_cpu_has(X86_FEATURE_ARAT))
694 /* find the deepest mwait value */
700 static int powerclamp_debug_show(struct seq_file *m, void *unused)
704 seq_printf(m, "controlling cpu: %d\n", control_cpu);
705 seq_printf(m, "pct confidence steady dynamic (compensation)\n");
706 for (i = 0; i < MAX_TARGET_RATIO; i++) {
707 seq_printf(m, "%d\t%lu\t%lu\t%lu\n",
709 cal_data[i].confidence,
710 cal_data[i].steady_comp,
711 cal_data[i].dynamic_comp);
717 static int powerclamp_debug_open(struct inode *inode,
720 return single_open(file, powerclamp_debug_show, inode->i_private);
723 static const struct file_operations powerclamp_debug_fops = {
724 .open = powerclamp_debug_open,
727 .release = single_release,
728 .owner = THIS_MODULE,
731 static inline void powerclamp_create_debug_files(void)
733 debug_dir = debugfs_create_dir("intel_powerclamp", NULL);
737 if (!debugfs_create_file("powerclamp_calib", S_IRUGO, debug_dir,
738 cal_data, &powerclamp_debug_fops))
744 debugfs_remove_recursive(debug_dir);
747 static int 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();
761 /* set default limit, maybe adjusted during runtime based on feedback */
763 register_hotcpu_notifier(&powerclamp_cpu_notifier);
764 powerclamp_thread = alloc_percpu(struct task_struct *);
765 cooling_dev = thermal_cooling_device_register("intel_powerclamp", NULL,
766 &powerclamp_cooling_ops);
767 if (IS_ERR(cooling_dev))
771 duration = jiffies_to_msecs(DEFAULT_DURATION_JIFFIES);
772 powerclamp_create_debug_files();
776 module_init(powerclamp_init);
778 static void powerclamp_exit(void)
780 unregister_hotcpu_notifier(&powerclamp_cpu_notifier);
782 free_percpu(powerclamp_thread);
783 thermal_cooling_device_unregister(cooling_dev);
784 kfree(cpu_clamping_mask);
786 cancel_delayed_work_sync(&poll_pkg_cstate_work);
787 debugfs_remove_recursive(debug_dir);
789 module_exit(powerclamp_exit);
791 MODULE_LICENSE("GPL");
792 MODULE_AUTHOR("Arjan van de Ven <arjan@linux.intel.com>");
793 MODULE_AUTHOR("Jacob Pan <jacob.jun.pan@linux.intel.com>");
794 MODULE_DESCRIPTION("Package Level C-state Idle Injection for Intel CPUs");