1 /*
2  * intel_pstate.c: Native P state management for Intel processors
3  *
4  * (C) Copyright 2012 Intel Corporation
5  * Author: Dirk Brandewie <dirk.j.brandewie@intel.com>
6  *
7  * This program is free software; you can redistribute it and/or
8  * modify it under the terms of the GNU General Public License
9  * as published by the Free Software Foundation; version 2
10  * of the License.
11  */
12 
13 #include <linux/kernel.h>
14 #include <linux/kernel_stat.h>
15 #include <linux/module.h>
16 #include <linux/ktime.h>
17 #include <linux/hrtimer.h>
18 #include <linux/tick.h>
19 #include <linux/slab.h>
20 #include <linux/sched.h>
21 #include <linux/list.h>
22 #include <linux/cpu.h>
23 #include <linux/cpufreq.h>
24 #include <linux/sysfs.h>
25 #include <linux/types.h>
26 #include <linux/fs.h>
27 #include <linux/debugfs.h>
28 #include <linux/acpi.h>
29 #include <trace/events/power.h>
30 
31 #include <asm/div64.h>
32 #include <asm/msr.h>
33 #include <asm/cpu_device_id.h>
34 #include <asm/cpufeature.h>
35 
36 #define BYT_RATIOS		0x66a
37 #define BYT_VIDS		0x66b
38 #define BYT_TURBO_RATIOS	0x66c
39 #define BYT_TURBO_VIDS		0x66d
40 
41 #define FRAC_BITS 8
42 #define int_tofp(X) ((int64_t)(X) << FRAC_BITS)
43 #define fp_toint(X) ((X) >> FRAC_BITS)
44 
45 
46 static inline int32_t mul_fp(int32_t x, int32_t y)
47 {
48 	return ((int64_t)x * (int64_t)y) >> FRAC_BITS;
49 }
50 
51 static inline int32_t div_fp(int32_t x, int32_t y)
52 {
53 	return div_s64((int64_t)x << FRAC_BITS, y);
54 }
55 
56 static inline int ceiling_fp(int32_t x)
57 {
58 	int mask, ret;
59 
60 	ret = fp_toint(x);
61 	mask = (1 << FRAC_BITS) - 1;
62 	if (x & mask)
63 		ret += 1;
64 	return ret;
65 }
66 
67 struct sample {
68 	int32_t core_pct_busy;
69 	u64 aperf;
70 	u64 mperf;
71 	int freq;
72 	ktime_t time;
73 };
74 
75 struct pstate_data {
76 	int	current_pstate;
77 	int	min_pstate;
78 	int	max_pstate;
79 	int	scaling;
80 	int	turbo_pstate;
81 };
82 
83 struct vid_data {
84 	int min;
85 	int max;
86 	int turbo;
87 	int32_t ratio;
88 };
89 
90 struct _pid {
91 	int setpoint;
92 	int32_t integral;
93 	int32_t p_gain;
94 	int32_t i_gain;
95 	int32_t d_gain;
96 	int deadband;
97 	int32_t last_err;
98 };
99 
100 struct cpudata {
101 	int cpu;
102 
103 	struct timer_list timer;
104 
105 	struct pstate_data pstate;
106 	struct vid_data vid;
107 	struct _pid pid;
108 
109 	ktime_t last_sample_time;
110 	u64	prev_aperf;
111 	u64	prev_mperf;
112 	struct sample sample;
113 };
114 
115 static struct cpudata **all_cpu_data;
116 struct pstate_adjust_policy {
117 	int sample_rate_ms;
118 	int deadband;
119 	int setpoint;
120 	int p_gain_pct;
121 	int d_gain_pct;
122 	int i_gain_pct;
123 };
124 
125 struct pstate_funcs {
126 	int (*get_max)(void);
127 	int (*get_min)(void);
128 	int (*get_turbo)(void);
129 	int (*get_scaling)(void);
130 	void (*set)(struct cpudata*, int pstate);
131 	void (*get_vid)(struct cpudata *);
132 };
133 
134 struct cpu_defaults {
135 	struct pstate_adjust_policy pid_policy;
136 	struct pstate_funcs funcs;
137 };
138 
139 static struct pstate_adjust_policy pid_params;
140 static struct pstate_funcs pstate_funcs;
141 static int hwp_active;
142 
143 struct perf_limits {
144 	int no_turbo;
145 	int turbo_disabled;
146 	int max_perf_pct;
147 	int min_perf_pct;
148 	int32_t max_perf;
149 	int32_t min_perf;
150 	int max_policy_pct;
151 	int max_sysfs_pct;
152 	int min_policy_pct;
153 	int min_sysfs_pct;
154 };
155 
156 static struct perf_limits limits = {
157 	.no_turbo = 0,
158 	.turbo_disabled = 0,
159 	.max_perf_pct = 100,
160 	.max_perf = int_tofp(1),
161 	.min_perf_pct = 0,
162 	.min_perf = 0,
163 	.max_policy_pct = 100,
164 	.max_sysfs_pct = 100,
165 	.min_policy_pct = 0,
166 	.min_sysfs_pct = 0,
167 };
168 
169 static inline void pid_reset(struct _pid *pid, int setpoint, int busy,
170 			     int deadband, int integral) {
171 	pid->setpoint = setpoint;
172 	pid->deadband  = deadband;
173 	pid->integral  = int_tofp(integral);
174 	pid->last_err  = int_tofp(setpoint) - int_tofp(busy);
175 }
176 
177 static inline void pid_p_gain_set(struct _pid *pid, int percent)
178 {
179 	pid->p_gain = div_fp(int_tofp(percent), int_tofp(100));
180 }
181 
182 static inline void pid_i_gain_set(struct _pid *pid, int percent)
183 {
184 	pid->i_gain = div_fp(int_tofp(percent), int_tofp(100));
185 }
186 
187 static inline void pid_d_gain_set(struct _pid *pid, int percent)
188 {
189 	pid->d_gain = div_fp(int_tofp(percent), int_tofp(100));
190 }
191 
192 static signed int pid_calc(struct _pid *pid, int32_t busy)
193 {
194 	signed int result;
195 	int32_t pterm, dterm, fp_error;
196 	int32_t integral_limit;
197 
198 	fp_error = int_tofp(pid->setpoint) - busy;
199 
200 	if (abs(fp_error) <= int_tofp(pid->deadband))
201 		return 0;
202 
203 	pterm = mul_fp(pid->p_gain, fp_error);
204 
205 	pid->integral += fp_error;
206 
207 	/*
208 	 * We limit the integral here so that it will never
209 	 * get higher than 30.  This prevents it from becoming
210 	 * too large an input over long periods of time and allows
211 	 * it to get factored out sooner.
212 	 *
213 	 * The value of 30 was chosen through experimentation.
214 	 */
215 	integral_limit = int_tofp(30);
216 	if (pid->integral > integral_limit)
217 		pid->integral = integral_limit;
218 	if (pid->integral < -integral_limit)
219 		pid->integral = -integral_limit;
220 
221 	dterm = mul_fp(pid->d_gain, fp_error - pid->last_err);
222 	pid->last_err = fp_error;
223 
224 	result = pterm + mul_fp(pid->integral, pid->i_gain) + dterm;
225 	result = result + (1 << (FRAC_BITS-1));
226 	return (signed int)fp_toint(result);
227 }
228 
229 static inline void intel_pstate_busy_pid_reset(struct cpudata *cpu)
230 {
231 	pid_p_gain_set(&cpu->pid, pid_params.p_gain_pct);
232 	pid_d_gain_set(&cpu->pid, pid_params.d_gain_pct);
233 	pid_i_gain_set(&cpu->pid, pid_params.i_gain_pct);
234 
235 	pid_reset(&cpu->pid, pid_params.setpoint, 100, pid_params.deadband, 0);
236 }
237 
238 static inline void intel_pstate_reset_all_pid(void)
239 {
240 	unsigned int cpu;
241 
242 	for_each_online_cpu(cpu) {
243 		if (all_cpu_data[cpu])
244 			intel_pstate_busy_pid_reset(all_cpu_data[cpu]);
245 	}
246 }
247 
248 static inline void update_turbo_state(void)
249 {
250 	u64 misc_en;
251 	struct cpudata *cpu;
252 
253 	cpu = all_cpu_data[0];
254 	rdmsrl(MSR_IA32_MISC_ENABLE, misc_en);
255 	limits.turbo_disabled =
256 		(misc_en & MSR_IA32_MISC_ENABLE_TURBO_DISABLE ||
257 		 cpu->pstate.max_pstate == cpu->pstate.turbo_pstate);
258 }
259 
260 #define PCT_TO_HWP(x) (x * 255 / 100)
261 static void intel_pstate_hwp_set(void)
262 {
263 	int min, max, cpu;
264 	u64 value, freq;
265 
266 	get_online_cpus();
267 
268 	for_each_online_cpu(cpu) {
269 		rdmsrl_on_cpu(cpu, MSR_HWP_REQUEST, &value);
270 		min = PCT_TO_HWP(limits.min_perf_pct);
271 		value &= ~HWP_MIN_PERF(~0L);
272 		value |= HWP_MIN_PERF(min);
273 
274 		max = PCT_TO_HWP(limits.max_perf_pct);
275 		if (limits.no_turbo) {
276 			rdmsrl( MSR_HWP_CAPABILITIES, freq);
277 			max = HWP_GUARANTEED_PERF(freq);
278 		}
279 
280 		value &= ~HWP_MAX_PERF(~0L);
281 		value |= HWP_MAX_PERF(max);
282 		wrmsrl_on_cpu(cpu, MSR_HWP_REQUEST, value);
283 	}
284 
285 	put_online_cpus();
286 }
287 
288 /************************** debugfs begin ************************/
289 static int pid_param_set(void *data, u64 val)
290 {
291 	*(u32 *)data = val;
292 	intel_pstate_reset_all_pid();
293 	return 0;
294 }
295 
296 static int pid_param_get(void *data, u64 *val)
297 {
298 	*val = *(u32 *)data;
299 	return 0;
300 }
301 DEFINE_SIMPLE_ATTRIBUTE(fops_pid_param, pid_param_get, pid_param_set, "%llu\n");
302 
303 struct pid_param {
304 	char *name;
305 	void *value;
306 };
307 
308 static struct pid_param pid_files[] = {
309 	{"sample_rate_ms", &pid_params.sample_rate_ms},
310 	{"d_gain_pct", &pid_params.d_gain_pct},
311 	{"i_gain_pct", &pid_params.i_gain_pct},
312 	{"deadband", &pid_params.deadband},
313 	{"setpoint", &pid_params.setpoint},
314 	{"p_gain_pct", &pid_params.p_gain_pct},
315 	{NULL, NULL}
316 };
317 
318 static void __init intel_pstate_debug_expose_params(void)
319 {
320 	struct dentry *debugfs_parent;
321 	int i = 0;
322 
323 	if (hwp_active)
324 		return;
325 	debugfs_parent = debugfs_create_dir("pstate_snb", NULL);
326 	if (IS_ERR_OR_NULL(debugfs_parent))
327 		return;
328 	while (pid_files[i].name) {
329 		debugfs_create_file(pid_files[i].name, 0660,
330 				    debugfs_parent, pid_files[i].value,
331 				    &fops_pid_param);
332 		i++;
333 	}
334 }
335 
336 /************************** debugfs end ************************/
337 
338 /************************** sysfs begin ************************/
339 #define show_one(file_name, object)					\
340 	static ssize_t show_##file_name					\
341 	(struct kobject *kobj, struct attribute *attr, char *buf)	\
342 	{								\
343 		return sprintf(buf, "%u\n", limits.object);		\
344 	}
345 
346 static ssize_t show_turbo_pct(struct kobject *kobj,
347 				struct attribute *attr, char *buf)
348 {
349 	struct cpudata *cpu;
350 	int total, no_turbo, turbo_pct;
351 	uint32_t turbo_fp;
352 
353 	cpu = all_cpu_data[0];
354 
355 	total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
356 	no_turbo = cpu->pstate.max_pstate - cpu->pstate.min_pstate + 1;
357 	turbo_fp = div_fp(int_tofp(no_turbo), int_tofp(total));
358 	turbo_pct = 100 - fp_toint(mul_fp(turbo_fp, int_tofp(100)));
359 	return sprintf(buf, "%u\n", turbo_pct);
360 }
361 
362 static ssize_t show_num_pstates(struct kobject *kobj,
363 				struct attribute *attr, char *buf)
364 {
365 	struct cpudata *cpu;
366 	int total;
367 
368 	cpu = all_cpu_data[0];
369 	total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
370 	return sprintf(buf, "%u\n", total);
371 }
372 
373 static ssize_t show_no_turbo(struct kobject *kobj,
374 			     struct attribute *attr, char *buf)
375 {
376 	ssize_t ret;
377 
378 	update_turbo_state();
379 	if (limits.turbo_disabled)
380 		ret = sprintf(buf, "%u\n", limits.turbo_disabled);
381 	else
382 		ret = sprintf(buf, "%u\n", limits.no_turbo);
383 
384 	return ret;
385 }
386 
387 static ssize_t store_no_turbo(struct kobject *a, struct attribute *b,
388 			      const char *buf, size_t count)
389 {
390 	unsigned int input;
391 	int ret;
392 
393 	ret = sscanf(buf, "%u", &input);
394 	if (ret != 1)
395 		return -EINVAL;
396 
397 	update_turbo_state();
398 	if (limits.turbo_disabled) {
399 		pr_warn("Turbo disabled by BIOS or unavailable on processor\n");
400 		return -EPERM;
401 	}
402 
403 	limits.no_turbo = clamp_t(int, input, 0, 1);
404 
405 	if (hwp_active)
406 		intel_pstate_hwp_set();
407 
408 	return count;
409 }
410 
411 static ssize_t store_max_perf_pct(struct kobject *a, struct attribute *b,
412 				  const char *buf, size_t count)
413 {
414 	unsigned int input;
415 	int ret;
416 
417 	ret = sscanf(buf, "%u", &input);
418 	if (ret != 1)
419 		return -EINVAL;
420 
421 	limits.max_sysfs_pct = clamp_t(int, input, 0 , 100);
422 	limits.max_perf_pct = min(limits.max_policy_pct, limits.max_sysfs_pct);
423 	limits.max_perf = div_fp(int_tofp(limits.max_perf_pct), int_tofp(100));
424 
425 	if (hwp_active)
426 		intel_pstate_hwp_set();
427 	return count;
428 }
429 
430 static ssize_t store_min_perf_pct(struct kobject *a, struct attribute *b,
431 				  const char *buf, size_t count)
432 {
433 	unsigned int input;
434 	int ret;
435 
436 	ret = sscanf(buf, "%u", &input);
437 	if (ret != 1)
438 		return -EINVAL;
439 
440 	limits.min_sysfs_pct = clamp_t(int, input, 0 , 100);
441 	limits.min_perf_pct = max(limits.min_policy_pct, limits.min_sysfs_pct);
442 	limits.min_perf = div_fp(int_tofp(limits.min_perf_pct), int_tofp(100));
443 
444 	if (hwp_active)
445 		intel_pstate_hwp_set();
446 	return count;
447 }
448 
449 show_one(max_perf_pct, max_perf_pct);
450 show_one(min_perf_pct, min_perf_pct);
451 
452 define_one_global_rw(no_turbo);
453 define_one_global_rw(max_perf_pct);
454 define_one_global_rw(min_perf_pct);
455 define_one_global_ro(turbo_pct);
456 define_one_global_ro(num_pstates);
457 
458 static struct attribute *intel_pstate_attributes[] = {
459 	&no_turbo.attr,
460 	&max_perf_pct.attr,
461 	&min_perf_pct.attr,
462 	&turbo_pct.attr,
463 	&num_pstates.attr,
464 	NULL
465 };
466 
467 static struct attribute_group intel_pstate_attr_group = {
468 	.attrs = intel_pstate_attributes,
469 };
470 
471 static void __init intel_pstate_sysfs_expose_params(void)
472 {
473 	struct kobject *intel_pstate_kobject;
474 	int rc;
475 
476 	intel_pstate_kobject = kobject_create_and_add("intel_pstate",
477 						&cpu_subsys.dev_root->kobj);
478 	BUG_ON(!intel_pstate_kobject);
479 	rc = sysfs_create_group(intel_pstate_kobject, &intel_pstate_attr_group);
480 	BUG_ON(rc);
481 }
482 /************************** sysfs end ************************/
483 
484 static void intel_pstate_hwp_enable(void)
485 {
486 	hwp_active++;
487 	pr_info("intel_pstate HWP enabled\n");
488 
489 	wrmsrl( MSR_PM_ENABLE, 0x1);
490 }
491 
492 static int byt_get_min_pstate(void)
493 {
494 	u64 value;
495 
496 	rdmsrl(BYT_RATIOS, value);
497 	return (value >> 8) & 0x7F;
498 }
499 
500 static int byt_get_max_pstate(void)
501 {
502 	u64 value;
503 
504 	rdmsrl(BYT_RATIOS, value);
505 	return (value >> 16) & 0x7F;
506 }
507 
508 static int byt_get_turbo_pstate(void)
509 {
510 	u64 value;
511 
512 	rdmsrl(BYT_TURBO_RATIOS, value);
513 	return value & 0x7F;
514 }
515 
516 static void byt_set_pstate(struct cpudata *cpudata, int pstate)
517 {
518 	u64 val;
519 	int32_t vid_fp;
520 	u32 vid;
521 
522 	val = pstate << 8;
523 	if (limits.no_turbo && !limits.turbo_disabled)
524 		val |= (u64)1 << 32;
525 
526 	vid_fp = cpudata->vid.min + mul_fp(
527 		int_tofp(pstate - cpudata->pstate.min_pstate),
528 		cpudata->vid.ratio);
529 
530 	vid_fp = clamp_t(int32_t, vid_fp, cpudata->vid.min, cpudata->vid.max);
531 	vid = ceiling_fp(vid_fp);
532 
533 	if (pstate > cpudata->pstate.max_pstate)
534 		vid = cpudata->vid.turbo;
535 
536 	val |= vid;
537 
538 	wrmsrl(MSR_IA32_PERF_CTL, val);
539 }
540 
541 #define BYT_BCLK_FREQS 5
542 static int byt_freq_table[BYT_BCLK_FREQS] = { 833, 1000, 1333, 1167, 800};
543 
544 static int byt_get_scaling(void)
545 {
546 	u64 value;
547 	int i;
548 
549 	rdmsrl(MSR_FSB_FREQ, value);
550 	i = value & 0x3;
551 
552 	BUG_ON(i > BYT_BCLK_FREQS);
553 
554 	return byt_freq_table[i] * 100;
555 }
556 
557 static void byt_get_vid(struct cpudata *cpudata)
558 {
559 	u64 value;
560 
561 	rdmsrl(BYT_VIDS, value);
562 	cpudata->vid.min = int_tofp((value >> 8) & 0x7f);
563 	cpudata->vid.max = int_tofp((value >> 16) & 0x7f);
564 	cpudata->vid.ratio = div_fp(
565 		cpudata->vid.max - cpudata->vid.min,
566 		int_tofp(cpudata->pstate.max_pstate -
567 			cpudata->pstate.min_pstate));
568 
569 	rdmsrl(BYT_TURBO_VIDS, value);
570 	cpudata->vid.turbo = value & 0x7f;
571 }
572 
573 static int core_get_min_pstate(void)
574 {
575 	u64 value;
576 
577 	rdmsrl(MSR_PLATFORM_INFO, value);
578 	return (value >> 40) & 0xFF;
579 }
580 
581 static int core_get_max_pstate(void)
582 {
583 	u64 value;
584 
585 	rdmsrl(MSR_PLATFORM_INFO, value);
586 	return (value >> 8) & 0xFF;
587 }
588 
589 static int core_get_turbo_pstate(void)
590 {
591 	u64 value;
592 	int nont, ret;
593 
594 	rdmsrl(MSR_NHM_TURBO_RATIO_LIMIT, value);
595 	nont = core_get_max_pstate();
596 	ret = (value) & 255;
597 	if (ret <= nont)
598 		ret = nont;
599 	return ret;
600 }
601 
602 static inline int core_get_scaling(void)
603 {
604 	return 100000;
605 }
606 
607 static void core_set_pstate(struct cpudata *cpudata, int pstate)
608 {
609 	u64 val;
610 
611 	val = pstate << 8;
612 	if (limits.no_turbo && !limits.turbo_disabled)
613 		val |= (u64)1 << 32;
614 
615 	wrmsrl_on_cpu(cpudata->cpu, MSR_IA32_PERF_CTL, val);
616 }
617 
618 static int knl_get_turbo_pstate(void)
619 {
620 	u64 value;
621 	int nont, ret;
622 
623 	rdmsrl(MSR_NHM_TURBO_RATIO_LIMIT, value);
624 	nont = core_get_max_pstate();
625 	ret = (((value) >> 8) & 0xFF);
626 	if (ret <= nont)
627 		ret = nont;
628 	return ret;
629 }
630 
631 static struct cpu_defaults core_params = {
632 	.pid_policy = {
633 		.sample_rate_ms = 10,
634 		.deadband = 0,
635 		.setpoint = 97,
636 		.p_gain_pct = 20,
637 		.d_gain_pct = 0,
638 		.i_gain_pct = 0,
639 	},
640 	.funcs = {
641 		.get_max = core_get_max_pstate,
642 		.get_min = core_get_min_pstate,
643 		.get_turbo = core_get_turbo_pstate,
644 		.get_scaling = core_get_scaling,
645 		.set = core_set_pstate,
646 	},
647 };
648 
649 static struct cpu_defaults byt_params = {
650 	.pid_policy = {
651 		.sample_rate_ms = 10,
652 		.deadband = 0,
653 		.setpoint = 60,
654 		.p_gain_pct = 14,
655 		.d_gain_pct = 0,
656 		.i_gain_pct = 4,
657 	},
658 	.funcs = {
659 		.get_max = byt_get_max_pstate,
660 		.get_min = byt_get_min_pstate,
661 		.get_turbo = byt_get_turbo_pstate,
662 		.set = byt_set_pstate,
663 		.get_scaling = byt_get_scaling,
664 		.get_vid = byt_get_vid,
665 	},
666 };
667 
668 static struct cpu_defaults knl_params = {
669 	.pid_policy = {
670 		.sample_rate_ms = 10,
671 		.deadband = 0,
672 		.setpoint = 97,
673 		.p_gain_pct = 20,
674 		.d_gain_pct = 0,
675 		.i_gain_pct = 0,
676 	},
677 	.funcs = {
678 		.get_max = core_get_max_pstate,
679 		.get_min = core_get_min_pstate,
680 		.get_turbo = knl_get_turbo_pstate,
681 		.set = core_set_pstate,
682 	},
683 };
684 
685 static void intel_pstate_get_min_max(struct cpudata *cpu, int *min, int *max)
686 {
687 	int max_perf = cpu->pstate.turbo_pstate;
688 	int max_perf_adj;
689 	int min_perf;
690 
691 	if (limits.no_turbo || limits.turbo_disabled)
692 		max_perf = cpu->pstate.max_pstate;
693 
694 	/*
695 	 * performance can be limited by user through sysfs, by cpufreq
696 	 * policy, or by cpu specific default values determined through
697 	 * experimentation.
698 	 */
699 	max_perf_adj = fp_toint(mul_fp(int_tofp(max_perf), limits.max_perf));
700 	*max = clamp_t(int, max_perf_adj,
701 			cpu->pstate.min_pstate, cpu->pstate.turbo_pstate);
702 
703 	min_perf = fp_toint(mul_fp(int_tofp(max_perf), limits.min_perf));
704 	*min = clamp_t(int, min_perf, cpu->pstate.min_pstate, max_perf);
705 }
706 
707 static void intel_pstate_set_pstate(struct cpudata *cpu, int pstate)
708 {
709 	int max_perf, min_perf;
710 
711 	update_turbo_state();
712 
713 	intel_pstate_get_min_max(cpu, &min_perf, &max_perf);
714 
715 	pstate = clamp_t(int, pstate, min_perf, max_perf);
716 
717 	if (pstate == cpu->pstate.current_pstate)
718 		return;
719 
720 	trace_cpu_frequency(pstate * cpu->pstate.scaling, cpu->cpu);
721 
722 	cpu->pstate.current_pstate = pstate;
723 
724 	pstate_funcs.set(cpu, pstate);
725 }
726 
727 static void intel_pstate_get_cpu_pstates(struct cpudata *cpu)
728 {
729 	cpu->pstate.min_pstate = pstate_funcs.get_min();
730 	cpu->pstate.max_pstate = pstate_funcs.get_max();
731 	cpu->pstate.turbo_pstate = pstate_funcs.get_turbo();
732 	cpu->pstate.scaling = pstate_funcs.get_scaling();
733 
734 	if (pstate_funcs.get_vid)
735 		pstate_funcs.get_vid(cpu);
736 	intel_pstate_set_pstate(cpu, cpu->pstate.min_pstate);
737 }
738 
739 static inline void intel_pstate_calc_busy(struct cpudata *cpu)
740 {
741 	struct sample *sample = &cpu->sample;
742 	int64_t core_pct;
743 
744 	core_pct = int_tofp(sample->aperf) * int_tofp(100);
745 	core_pct = div64_u64(core_pct, int_tofp(sample->mperf));
746 
747 	sample->freq = fp_toint(
748 		mul_fp(int_tofp(
749 			cpu->pstate.max_pstate * cpu->pstate.scaling / 100),
750 			core_pct));
751 
752 	sample->core_pct_busy = (int32_t)core_pct;
753 }
754 
755 static inline void intel_pstate_sample(struct cpudata *cpu)
756 {
757 	u64 aperf, mperf;
758 	unsigned long flags;
759 
760 	local_irq_save(flags);
761 	rdmsrl(MSR_IA32_APERF, aperf);
762 	rdmsrl(MSR_IA32_MPERF, mperf);
763 	local_irq_restore(flags);
764 
765 	cpu->last_sample_time = cpu->sample.time;
766 	cpu->sample.time = ktime_get();
767 	cpu->sample.aperf = aperf;
768 	cpu->sample.mperf = mperf;
769 	cpu->sample.aperf -= cpu->prev_aperf;
770 	cpu->sample.mperf -= cpu->prev_mperf;
771 
772 	intel_pstate_calc_busy(cpu);
773 
774 	cpu->prev_aperf = aperf;
775 	cpu->prev_mperf = mperf;
776 }
777 
778 static inline void intel_hwp_set_sample_time(struct cpudata *cpu)
779 {
780 	int delay;
781 
782 	delay = msecs_to_jiffies(50);
783 	mod_timer_pinned(&cpu->timer, jiffies + delay);
784 }
785 
786 static inline void intel_pstate_set_sample_time(struct cpudata *cpu)
787 {
788 	int delay;
789 
790 	delay = msecs_to_jiffies(pid_params.sample_rate_ms);
791 	mod_timer_pinned(&cpu->timer, jiffies + delay);
792 }
793 
794 static inline int32_t intel_pstate_get_scaled_busy(struct cpudata *cpu)
795 {
796 	int32_t core_busy, max_pstate, current_pstate, sample_ratio;
797 	u32 duration_us;
798 	u32 sample_time;
799 
800 	/*
801 	 * core_busy is the ratio of actual performance to max
802 	 * max_pstate is the max non turbo pstate available
803 	 * current_pstate was the pstate that was requested during
804 	 * 	the last sample period.
805 	 *
806 	 * We normalize core_busy, which was our actual percent
807 	 * performance to what we requested during the last sample
808 	 * period. The result will be a percentage of busy at a
809 	 * specified pstate.
810 	 */
811 	core_busy = cpu->sample.core_pct_busy;
812 	max_pstate = int_tofp(cpu->pstate.max_pstate);
813 	current_pstate = int_tofp(cpu->pstate.current_pstate);
814 	core_busy = mul_fp(core_busy, div_fp(max_pstate, current_pstate));
815 
816 	/*
817 	 * Since we have a deferred timer, it will not fire unless
818 	 * we are in C0.  So, determine if the actual elapsed time
819 	 * is significantly greater (3x) than our sample interval.  If it
820 	 * is, then we were idle for a long enough period of time
821 	 * to adjust our busyness.
822 	 */
823 	sample_time = pid_params.sample_rate_ms  * USEC_PER_MSEC;
824 	duration_us = (u32) ktime_us_delta(cpu->sample.time,
825 					   cpu->last_sample_time);
826 	if (duration_us > sample_time * 3) {
827 		sample_ratio = div_fp(int_tofp(sample_time),
828 				      int_tofp(duration_us));
829 		core_busy = mul_fp(core_busy, sample_ratio);
830 	}
831 
832 	return core_busy;
833 }
834 
835 static inline void intel_pstate_adjust_busy_pstate(struct cpudata *cpu)
836 {
837 	int32_t busy_scaled;
838 	struct _pid *pid;
839 	signed int ctl;
840 
841 	pid = &cpu->pid;
842 	busy_scaled = intel_pstate_get_scaled_busy(cpu);
843 
844 	ctl = pid_calc(pid, busy_scaled);
845 
846 	/* Negative values of ctl increase the pstate and vice versa */
847 	intel_pstate_set_pstate(cpu, cpu->pstate.current_pstate - ctl);
848 }
849 
850 static void intel_hwp_timer_func(unsigned long __data)
851 {
852 	struct cpudata *cpu = (struct cpudata *) __data;
853 
854 	intel_pstate_sample(cpu);
855 	intel_hwp_set_sample_time(cpu);
856 }
857 
858 static void intel_pstate_timer_func(unsigned long __data)
859 {
860 	struct cpudata *cpu = (struct cpudata *) __data;
861 	struct sample *sample;
862 
863 	intel_pstate_sample(cpu);
864 
865 	sample = &cpu->sample;
866 
867 	intel_pstate_adjust_busy_pstate(cpu);
868 
869 	trace_pstate_sample(fp_toint(sample->core_pct_busy),
870 			fp_toint(intel_pstate_get_scaled_busy(cpu)),
871 			cpu->pstate.current_pstate,
872 			sample->mperf,
873 			sample->aperf,
874 			sample->freq);
875 
876 	intel_pstate_set_sample_time(cpu);
877 }
878 
879 #define ICPU(model, policy) \
880 	{ X86_VENDOR_INTEL, 6, model, X86_FEATURE_APERFMPERF,\
881 			(unsigned long)&policy }
882 
883 static const struct x86_cpu_id intel_pstate_cpu_ids[] = {
884 	ICPU(0x2a, core_params),
885 	ICPU(0x2d, core_params),
886 	ICPU(0x37, byt_params),
887 	ICPU(0x3a, core_params),
888 	ICPU(0x3c, core_params),
889 	ICPU(0x3d, core_params),
890 	ICPU(0x3e, core_params),
891 	ICPU(0x3f, core_params),
892 	ICPU(0x45, core_params),
893 	ICPU(0x46, core_params),
894 	ICPU(0x47, core_params),
895 	ICPU(0x4c, byt_params),
896 	ICPU(0x4e, core_params),
897 	ICPU(0x4f, core_params),
898 	ICPU(0x56, core_params),
899 	ICPU(0x57, knl_params),
900 	{}
901 };
902 MODULE_DEVICE_TABLE(x86cpu, intel_pstate_cpu_ids);
903 
904 static const struct x86_cpu_id intel_pstate_cpu_oob_ids[] = {
905 	ICPU(0x56, core_params),
906 	{}
907 };
908 
909 static int intel_pstate_init_cpu(unsigned int cpunum)
910 {
911 	struct cpudata *cpu;
912 
913 	if (!all_cpu_data[cpunum])
914 		all_cpu_data[cpunum] = kzalloc(sizeof(struct cpudata),
915 					       GFP_KERNEL);
916 	if (!all_cpu_data[cpunum])
917 		return -ENOMEM;
918 
919 	cpu = all_cpu_data[cpunum];
920 
921 	cpu->cpu = cpunum;
922 	intel_pstate_get_cpu_pstates(cpu);
923 
924 	init_timer_deferrable(&cpu->timer);
925 	cpu->timer.data = (unsigned long)cpu;
926 	cpu->timer.expires = jiffies + HZ/100;
927 
928 	if (!hwp_active)
929 		cpu->timer.function = intel_pstate_timer_func;
930 	else
931 		cpu->timer.function = intel_hwp_timer_func;
932 
933 	intel_pstate_busy_pid_reset(cpu);
934 	intel_pstate_sample(cpu);
935 
936 	add_timer_on(&cpu->timer, cpunum);
937 
938 	pr_debug("Intel pstate controlling: cpu %d\n", cpunum);
939 
940 	return 0;
941 }
942 
943 static unsigned int intel_pstate_get(unsigned int cpu_num)
944 {
945 	struct sample *sample;
946 	struct cpudata *cpu;
947 
948 	cpu = all_cpu_data[cpu_num];
949 	if (!cpu)
950 		return 0;
951 	sample = &cpu->sample;
952 	return sample->freq;
953 }
954 
955 static int intel_pstate_set_policy(struct cpufreq_policy *policy)
956 {
957 	if (!policy->cpuinfo.max_freq)
958 		return -ENODEV;
959 
960 	if (policy->policy == CPUFREQ_POLICY_PERFORMANCE &&
961 	    policy->max >= policy->cpuinfo.max_freq) {
962 		limits.min_policy_pct = 100;
963 		limits.min_perf_pct = 100;
964 		limits.min_perf = int_tofp(1);
965 		limits.max_policy_pct = 100;
966 		limits.max_perf_pct = 100;
967 		limits.max_perf = int_tofp(1);
968 		limits.no_turbo = 0;
969 		return 0;
970 	}
971 
972 	limits.min_policy_pct = (policy->min * 100) / policy->cpuinfo.max_freq;
973 	limits.min_policy_pct = clamp_t(int, limits.min_policy_pct, 0 , 100);
974 	limits.min_perf_pct = max(limits.min_policy_pct, limits.min_sysfs_pct);
975 	limits.min_perf = div_fp(int_tofp(limits.min_perf_pct), int_tofp(100));
976 
977 	limits.max_policy_pct = (policy->max * 100) / policy->cpuinfo.max_freq;
978 	limits.max_policy_pct = clamp_t(int, limits.max_policy_pct, 0 , 100);
979 	limits.max_perf_pct = min(limits.max_policy_pct, limits.max_sysfs_pct);
980 	limits.max_perf = div_fp(int_tofp(limits.max_perf_pct), int_tofp(100));
981 
982 	if (hwp_active)
983 		intel_pstate_hwp_set();
984 
985 	return 0;
986 }
987 
988 static int intel_pstate_verify_policy(struct cpufreq_policy *policy)
989 {
990 	cpufreq_verify_within_cpu_limits(policy);
991 
992 	if (policy->policy != CPUFREQ_POLICY_POWERSAVE &&
993 	    policy->policy != CPUFREQ_POLICY_PERFORMANCE)
994 		return -EINVAL;
995 
996 	return 0;
997 }
998 
999 static void intel_pstate_stop_cpu(struct cpufreq_policy *policy)
1000 {
1001 	int cpu_num = policy->cpu;
1002 	struct cpudata *cpu = all_cpu_data[cpu_num];
1003 
1004 	pr_info("intel_pstate CPU %d exiting\n", cpu_num);
1005 
1006 	del_timer_sync(&all_cpu_data[cpu_num]->timer);
1007 	if (hwp_active)
1008 		return;
1009 
1010 	intel_pstate_set_pstate(cpu, cpu->pstate.min_pstate);
1011 }
1012 
1013 static int intel_pstate_cpu_init(struct cpufreq_policy *policy)
1014 {
1015 	struct cpudata *cpu;
1016 	int rc;
1017 
1018 	rc = intel_pstate_init_cpu(policy->cpu);
1019 	if (rc)
1020 		return rc;
1021 
1022 	cpu = all_cpu_data[policy->cpu];
1023 
1024 	if (limits.min_perf_pct == 100 && limits.max_perf_pct == 100)
1025 		policy->policy = CPUFREQ_POLICY_PERFORMANCE;
1026 	else
1027 		policy->policy = CPUFREQ_POLICY_POWERSAVE;
1028 
1029 	policy->min = cpu->pstate.min_pstate * cpu->pstate.scaling;
1030 	policy->max = cpu->pstate.turbo_pstate * cpu->pstate.scaling;
1031 
1032 	/* cpuinfo and default policy values */
1033 	policy->cpuinfo.min_freq = cpu->pstate.min_pstate * cpu->pstate.scaling;
1034 	policy->cpuinfo.max_freq =
1035 		cpu->pstate.turbo_pstate * cpu->pstate.scaling;
1036 	policy->cpuinfo.transition_latency = CPUFREQ_ETERNAL;
1037 	cpumask_set_cpu(policy->cpu, policy->cpus);
1038 
1039 	return 0;
1040 }
1041 
1042 static struct cpufreq_driver intel_pstate_driver = {
1043 	.flags		= CPUFREQ_CONST_LOOPS,
1044 	.verify		= intel_pstate_verify_policy,
1045 	.setpolicy	= intel_pstate_set_policy,
1046 	.get		= intel_pstate_get,
1047 	.init		= intel_pstate_cpu_init,
1048 	.stop_cpu	= intel_pstate_stop_cpu,
1049 	.name		= "intel_pstate",
1050 };
1051 
1052 static int __initdata no_load;
1053 static int __initdata no_hwp;
1054 static int __initdata hwp_only;
1055 static unsigned int force_load;
1056 
1057 static int intel_pstate_msrs_not_valid(void)
1058 {
1059 	if (!pstate_funcs.get_max() ||
1060 	    !pstate_funcs.get_min() ||
1061 	    !pstate_funcs.get_turbo())
1062 		return -ENODEV;
1063 
1064 	return 0;
1065 }
1066 
1067 static void copy_pid_params(struct pstate_adjust_policy *policy)
1068 {
1069 	pid_params.sample_rate_ms = policy->sample_rate_ms;
1070 	pid_params.p_gain_pct = policy->p_gain_pct;
1071 	pid_params.i_gain_pct = policy->i_gain_pct;
1072 	pid_params.d_gain_pct = policy->d_gain_pct;
1073 	pid_params.deadband = policy->deadband;
1074 	pid_params.setpoint = policy->setpoint;
1075 }
1076 
1077 static void copy_cpu_funcs(struct pstate_funcs *funcs)
1078 {
1079 	pstate_funcs.get_max   = funcs->get_max;
1080 	pstate_funcs.get_min   = funcs->get_min;
1081 	pstate_funcs.get_turbo = funcs->get_turbo;
1082 	pstate_funcs.get_scaling = funcs->get_scaling;
1083 	pstate_funcs.set       = funcs->set;
1084 	pstate_funcs.get_vid   = funcs->get_vid;
1085 }
1086 
1087 #if IS_ENABLED(CONFIG_ACPI)
1088 #include <acpi/processor.h>
1089 
1090 static bool intel_pstate_no_acpi_pss(void)
1091 {
1092 	int i;
1093 
1094 	for_each_possible_cpu(i) {
1095 		acpi_status status;
1096 		union acpi_object *pss;
1097 		struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
1098 		struct acpi_processor *pr = per_cpu(processors, i);
1099 
1100 		if (!pr)
1101 			continue;
1102 
1103 		status = acpi_evaluate_object(pr->handle, "_PSS", NULL, &buffer);
1104 		if (ACPI_FAILURE(status))
1105 			continue;
1106 
1107 		pss = buffer.pointer;
1108 		if (pss && pss->type == ACPI_TYPE_PACKAGE) {
1109 			kfree(pss);
1110 			return false;
1111 		}
1112 
1113 		kfree(pss);
1114 	}
1115 
1116 	return true;
1117 }
1118 
1119 static bool intel_pstate_has_acpi_ppc(void)
1120 {
1121 	int i;
1122 
1123 	for_each_possible_cpu(i) {
1124 		struct acpi_processor *pr = per_cpu(processors, i);
1125 
1126 		if (!pr)
1127 			continue;
1128 		if (acpi_has_method(pr->handle, "_PPC"))
1129 			return true;
1130 	}
1131 	return false;
1132 }
1133 
1134 enum {
1135 	PSS,
1136 	PPC,
1137 };
1138 
1139 struct hw_vendor_info {
1140 	u16  valid;
1141 	char oem_id[ACPI_OEM_ID_SIZE];
1142 	char oem_table_id[ACPI_OEM_TABLE_ID_SIZE];
1143 	int  oem_pwr_table;
1144 };
1145 
1146 /* Hardware vendor-specific info that has its own power management modes */
1147 static struct hw_vendor_info vendor_info[] = {
1148 	{1, "HP    ", "ProLiant", PSS},
1149 	{1, "ORACLE", "X4-2    ", PPC},
1150 	{1, "ORACLE", "X4-2L   ", PPC},
1151 	{1, "ORACLE", "X4-2B   ", PPC},
1152 	{1, "ORACLE", "X3-2    ", PPC},
1153 	{1, "ORACLE", "X3-2L   ", PPC},
1154 	{1, "ORACLE", "X3-2B   ", PPC},
1155 	{1, "ORACLE", "X4470M2 ", PPC},
1156 	{1, "ORACLE", "X4270M3 ", PPC},
1157 	{1, "ORACLE", "X4270M2 ", PPC},
1158 	{1, "ORACLE", "X4170M2 ", PPC},
1159 	{0, "", ""},
1160 };
1161 
1162 static bool intel_pstate_platform_pwr_mgmt_exists(void)
1163 {
1164 	struct acpi_table_header hdr;
1165 	struct hw_vendor_info *v_info;
1166 	const struct x86_cpu_id *id;
1167 	u64 misc_pwr;
1168 
1169 	id = x86_match_cpu(intel_pstate_cpu_oob_ids);
1170 	if (id) {
1171 		rdmsrl(MSR_MISC_PWR_MGMT, misc_pwr);
1172 		if ( misc_pwr & (1 << 8))
1173 			return true;
1174 	}
1175 
1176 	if (acpi_disabled ||
1177 	    ACPI_FAILURE(acpi_get_table_header(ACPI_SIG_FADT, 0, &hdr)))
1178 		return false;
1179 
1180 	for (v_info = vendor_info; v_info->valid; v_info++) {
1181 		if (!strncmp(hdr.oem_id, v_info->oem_id, ACPI_OEM_ID_SIZE) &&
1182 			!strncmp(hdr.oem_table_id, v_info->oem_table_id,
1183 						ACPI_OEM_TABLE_ID_SIZE))
1184 			switch (v_info->oem_pwr_table) {
1185 			case PSS:
1186 				return intel_pstate_no_acpi_pss();
1187 			case PPC:
1188 				return intel_pstate_has_acpi_ppc() &&
1189 					(!force_load);
1190 			}
1191 	}
1192 
1193 	return false;
1194 }
1195 #else /* CONFIG_ACPI not enabled */
1196 static inline bool intel_pstate_platform_pwr_mgmt_exists(void) { return false; }
1197 static inline bool intel_pstate_has_acpi_ppc(void) { return false; }
1198 #endif /* CONFIG_ACPI */
1199 
1200 static int __init intel_pstate_init(void)
1201 {
1202 	int cpu, rc = 0;
1203 	const struct x86_cpu_id *id;
1204 	struct cpu_defaults *cpu_def;
1205 
1206 	if (no_load)
1207 		return -ENODEV;
1208 
1209 	id = x86_match_cpu(intel_pstate_cpu_ids);
1210 	if (!id)
1211 		return -ENODEV;
1212 
1213 	/*
1214 	 * The Intel pstate driver will be ignored if the platform
1215 	 * firmware has its own power management modes.
1216 	 */
1217 	if (intel_pstate_platform_pwr_mgmt_exists())
1218 		return -ENODEV;
1219 
1220 	cpu_def = (struct cpu_defaults *)id->driver_data;
1221 
1222 	copy_pid_params(&cpu_def->pid_policy);
1223 	copy_cpu_funcs(&cpu_def->funcs);
1224 
1225 	if (intel_pstate_msrs_not_valid())
1226 		return -ENODEV;
1227 
1228 	pr_info("Intel P-state driver initializing.\n");
1229 
1230 	all_cpu_data = vzalloc(sizeof(void *) * num_possible_cpus());
1231 	if (!all_cpu_data)
1232 		return -ENOMEM;
1233 
1234 	if (static_cpu_has_safe(X86_FEATURE_HWP) && !no_hwp)
1235 		intel_pstate_hwp_enable();
1236 
1237 	if (!hwp_active && hwp_only)
1238 		goto out;
1239 
1240 	rc = cpufreq_register_driver(&intel_pstate_driver);
1241 	if (rc)
1242 		goto out;
1243 
1244 	intel_pstate_debug_expose_params();
1245 	intel_pstate_sysfs_expose_params();
1246 
1247 	return rc;
1248 out:
1249 	get_online_cpus();
1250 	for_each_online_cpu(cpu) {
1251 		if (all_cpu_data[cpu]) {
1252 			del_timer_sync(&all_cpu_data[cpu]->timer);
1253 			kfree(all_cpu_data[cpu]);
1254 		}
1255 	}
1256 
1257 	put_online_cpus();
1258 	vfree(all_cpu_data);
1259 	return -ENODEV;
1260 }
1261 device_initcall(intel_pstate_init);
1262 
1263 static int __init intel_pstate_setup(char *str)
1264 {
1265 	if (!str)
1266 		return -EINVAL;
1267 
1268 	if (!strcmp(str, "disable"))
1269 		no_load = 1;
1270 	if (!strcmp(str, "no_hwp"))
1271 		no_hwp = 1;
1272 	if (!strcmp(str, "force"))
1273 		force_load = 1;
1274 	if (!strcmp(str, "hwp_only"))
1275 		hwp_only = 1;
1276 	return 0;
1277 }
1278 early_param("intel_pstate", intel_pstate_setup);
1279 
1280 MODULE_AUTHOR("Dirk Brandewie <dirk.j.brandewie@intel.com>");
1281 MODULE_DESCRIPTION("'intel_pstate' - P state driver Intel Core processors");
1282 MODULE_LICENSE("GPL");
1283