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 
35 #define SAMPLE_COUNT		3
36 
37 #define BYT_RATIOS	0x66a
38 #define BYT_VIDS        0x66b
39 
40 #define FRAC_BITS 8
41 #define int_tofp(X) ((int64_t)(X) << FRAC_BITS)
42 #define fp_toint(X) ((X) >> FRAC_BITS)
43 
44 static inline int32_t mul_fp(int32_t x, int32_t y)
45 {
46 	return ((int64_t)x * (int64_t)y) >> FRAC_BITS;
47 }
48 
49 static inline int32_t div_fp(int32_t x, int32_t y)
50 {
51 	return div_s64((int64_t)x << FRAC_BITS, (int64_t)y);
52 }
53 
54 static u64 energy_divisor;
55 
56 struct sample {
57 	int32_t core_pct_busy;
58 	u64 aperf;
59 	u64 mperf;
60 	int freq;
61 };
62 
63 struct pstate_data {
64 	int	current_pstate;
65 	int	min_pstate;
66 	int	max_pstate;
67 	int	turbo_pstate;
68 };
69 
70 struct vid_data {
71 	int32_t min;
72 	int32_t max;
73 	int32_t ratio;
74 };
75 
76 struct _pid {
77 	int setpoint;
78 	int32_t integral;
79 	int32_t p_gain;
80 	int32_t i_gain;
81 	int32_t d_gain;
82 	int deadband;
83 	int32_t last_err;
84 };
85 
86 struct cpudata {
87 	int cpu;
88 
89 	char name[64];
90 
91 	struct timer_list timer;
92 
93 	struct pstate_data pstate;
94 	struct vid_data vid;
95 	struct _pid pid;
96 
97 	u64	prev_aperf;
98 	u64	prev_mperf;
99 	int	sample_ptr;
100 	struct sample samples[SAMPLE_COUNT];
101 };
102 
103 static struct cpudata **all_cpu_data;
104 struct pstate_adjust_policy {
105 	int sample_rate_ms;
106 	int deadband;
107 	int setpoint;
108 	int p_gain_pct;
109 	int d_gain_pct;
110 	int i_gain_pct;
111 };
112 
113 struct pstate_funcs {
114 	int (*get_max)(void);
115 	int (*get_min)(void);
116 	int (*get_turbo)(void);
117 	void (*set)(struct cpudata*, int pstate);
118 	void (*get_vid)(struct cpudata *);
119 };
120 
121 struct cpu_defaults {
122 	struct pstate_adjust_policy pid_policy;
123 	struct pstate_funcs funcs;
124 };
125 
126 static struct pstate_adjust_policy pid_params;
127 static struct pstate_funcs pstate_funcs;
128 
129 struct perf_limits {
130 	int no_turbo;
131 	int max_perf_pct;
132 	int min_perf_pct;
133 	int32_t max_perf;
134 	int32_t min_perf;
135 	int max_policy_pct;
136 	int max_sysfs_pct;
137 };
138 
139 static struct perf_limits limits = {
140 	.no_turbo = 0,
141 	.max_perf_pct = 100,
142 	.max_perf = int_tofp(1),
143 	.min_perf_pct = 0,
144 	.min_perf = 0,
145 	.max_policy_pct = 100,
146 	.max_sysfs_pct = 100,
147 };
148 
149 static inline void pid_reset(struct _pid *pid, int setpoint, int busy,
150 			int deadband, int integral) {
151 	pid->setpoint = setpoint;
152 	pid->deadband  = deadband;
153 	pid->integral  = int_tofp(integral);
154 	pid->last_err  = setpoint - busy;
155 }
156 
157 static inline void pid_p_gain_set(struct _pid *pid, int percent)
158 {
159 	pid->p_gain = div_fp(int_tofp(percent), int_tofp(100));
160 }
161 
162 static inline void pid_i_gain_set(struct _pid *pid, int percent)
163 {
164 	pid->i_gain = div_fp(int_tofp(percent), int_tofp(100));
165 }
166 
167 static inline void pid_d_gain_set(struct _pid *pid, int percent)
168 {
169 
170 	pid->d_gain = div_fp(int_tofp(percent), int_tofp(100));
171 }
172 
173 static signed int pid_calc(struct _pid *pid, int32_t busy)
174 {
175 	signed int result;
176 	int32_t pterm, dterm, fp_error;
177 	int32_t integral_limit;
178 
179 	fp_error = int_tofp(pid->setpoint) - busy;
180 
181 	if (abs(fp_error) <= int_tofp(pid->deadband))
182 		return 0;
183 
184 	pterm = mul_fp(pid->p_gain, fp_error);
185 
186 	pid->integral += fp_error;
187 
188 	/* limit the integral term */
189 	integral_limit = int_tofp(30);
190 	if (pid->integral > integral_limit)
191 		pid->integral = integral_limit;
192 	if (pid->integral < -integral_limit)
193 		pid->integral = -integral_limit;
194 
195 	dterm = mul_fp(pid->d_gain, fp_error - pid->last_err);
196 	pid->last_err = fp_error;
197 
198 	result = pterm + mul_fp(pid->integral, pid->i_gain) + dterm;
199 
200 	return (signed int)fp_toint(result);
201 }
202 
203 static inline void intel_pstate_busy_pid_reset(struct cpudata *cpu)
204 {
205 	pid_p_gain_set(&cpu->pid, pid_params.p_gain_pct);
206 	pid_d_gain_set(&cpu->pid, pid_params.d_gain_pct);
207 	pid_i_gain_set(&cpu->pid, pid_params.i_gain_pct);
208 
209 	pid_reset(&cpu->pid,
210 		pid_params.setpoint,
211 		100,
212 		pid_params.deadband,
213 		0);
214 }
215 
216 static inline void intel_pstate_reset_all_pid(void)
217 {
218 	unsigned int cpu;
219 	for_each_online_cpu(cpu) {
220 		if (all_cpu_data[cpu])
221 			intel_pstate_busy_pid_reset(all_cpu_data[cpu]);
222 	}
223 }
224 
225 /************************** debugfs begin ************************/
226 static int pid_param_set(void *data, u64 val)
227 {
228 	*(u32 *)data = val;
229 	intel_pstate_reset_all_pid();
230 	return 0;
231 }
232 static int pid_param_get(void *data, u64 *val)
233 {
234 	*val = *(u32 *)data;
235 	return 0;
236 }
237 DEFINE_SIMPLE_ATTRIBUTE(fops_pid_param, pid_param_get,
238 			pid_param_set, "%llu\n");
239 
240 struct pid_param {
241 	char *name;
242 	void *value;
243 };
244 
245 static struct pid_param pid_files[] = {
246 	{"sample_rate_ms", &pid_params.sample_rate_ms},
247 	{"d_gain_pct", &pid_params.d_gain_pct},
248 	{"i_gain_pct", &pid_params.i_gain_pct},
249 	{"deadband", &pid_params.deadband},
250 	{"setpoint", &pid_params.setpoint},
251 	{"p_gain_pct", &pid_params.p_gain_pct},
252 	{NULL, NULL}
253 };
254 
255 static struct dentry *debugfs_parent;
256 static void intel_pstate_debug_expose_params(void)
257 {
258 	int i = 0;
259 
260 	debugfs_parent = debugfs_create_dir("pstate_snb", NULL);
261 	if (IS_ERR_OR_NULL(debugfs_parent))
262 		return;
263 	while (pid_files[i].name) {
264 		debugfs_create_file(pid_files[i].name, 0660,
265 				debugfs_parent, pid_files[i].value,
266 				&fops_pid_param);
267 		i++;
268 	}
269 }
270 
271 /************************** debugfs end ************************/
272 
273 /************************** sysfs begin ************************/
274 #define show_one(file_name, object)					\
275 	static ssize_t show_##file_name					\
276 	(struct kobject *kobj, struct attribute *attr, char *buf)	\
277 	{								\
278 		return sprintf(buf, "%u\n", limits.object);		\
279 	}
280 
281 static ssize_t store_no_turbo(struct kobject *a, struct attribute *b,
282 				const char *buf, size_t count)
283 {
284 	unsigned int input;
285 	int ret;
286 	ret = sscanf(buf, "%u", &input);
287 	if (ret != 1)
288 		return -EINVAL;
289 	limits.no_turbo = clamp_t(int, input, 0 , 1);
290 
291 	return count;
292 }
293 
294 static ssize_t store_max_perf_pct(struct kobject *a, struct attribute *b,
295 				const char *buf, size_t count)
296 {
297 	unsigned int input;
298 	int ret;
299 	ret = sscanf(buf, "%u", &input);
300 	if (ret != 1)
301 		return -EINVAL;
302 
303 	limits.max_sysfs_pct = clamp_t(int, input, 0 , 100);
304 	limits.max_perf_pct = min(limits.max_policy_pct, limits.max_sysfs_pct);
305 	limits.max_perf = div_fp(int_tofp(limits.max_perf_pct), int_tofp(100));
306 	return count;
307 }
308 
309 static ssize_t store_min_perf_pct(struct kobject *a, struct attribute *b,
310 				const char *buf, size_t count)
311 {
312 	unsigned int input;
313 	int ret;
314 	ret = sscanf(buf, "%u", &input);
315 	if (ret != 1)
316 		return -EINVAL;
317 	limits.min_perf_pct = clamp_t(int, input, 0 , 100);
318 	limits.min_perf = div_fp(int_tofp(limits.min_perf_pct), int_tofp(100));
319 
320 	return count;
321 }
322 
323 show_one(no_turbo, no_turbo);
324 show_one(max_perf_pct, max_perf_pct);
325 show_one(min_perf_pct, min_perf_pct);
326 
327 define_one_global_rw(no_turbo);
328 define_one_global_rw(max_perf_pct);
329 define_one_global_rw(min_perf_pct);
330 
331 static struct attribute *intel_pstate_attributes[] = {
332 	&no_turbo.attr,
333 	&max_perf_pct.attr,
334 	&min_perf_pct.attr,
335 	NULL
336 };
337 
338 static struct attribute_group intel_pstate_attr_group = {
339 	.attrs = intel_pstate_attributes,
340 };
341 static struct kobject *intel_pstate_kobject;
342 
343 static void intel_pstate_sysfs_expose_params(void)
344 {
345 	int rc;
346 
347 	intel_pstate_kobject = kobject_create_and_add("intel_pstate",
348 						&cpu_subsys.dev_root->kobj);
349 	BUG_ON(!intel_pstate_kobject);
350 	rc = sysfs_create_group(intel_pstate_kobject,
351 				&intel_pstate_attr_group);
352 	BUG_ON(rc);
353 }
354 
355 /************************** sysfs end ************************/
356 static int byt_get_min_pstate(void)
357 {
358 	u64 value;
359 	rdmsrl(BYT_RATIOS, value);
360 	return value & 0xFF;
361 }
362 
363 static int byt_get_max_pstate(void)
364 {
365 	u64 value;
366 	rdmsrl(BYT_RATIOS, value);
367 	return (value >> 16) & 0xFF;
368 }
369 
370 static void byt_set_pstate(struct cpudata *cpudata, int pstate)
371 {
372 	u64 val;
373 	int32_t vid_fp;
374 	u32 vid;
375 
376 	val = pstate << 8;
377 	if (limits.no_turbo)
378 		val |= (u64)1 << 32;
379 
380 	vid_fp = cpudata->vid.min + mul_fp(
381 		int_tofp(pstate - cpudata->pstate.min_pstate),
382 		cpudata->vid.ratio);
383 
384 	vid_fp = clamp_t(int32_t, vid_fp, cpudata->vid.min, cpudata->vid.max);
385 	vid = fp_toint(vid_fp);
386 
387 	val |= vid;
388 
389 	wrmsrl(MSR_IA32_PERF_CTL, val);
390 }
391 
392 static void byt_get_vid(struct cpudata *cpudata)
393 {
394 	u64 value;
395 
396 	rdmsrl(BYT_VIDS, value);
397 	cpudata->vid.min = int_tofp((value >> 8) & 0x7f);
398 	cpudata->vid.max = int_tofp((value >> 16) & 0x7f);
399 	cpudata->vid.ratio = div_fp(
400 		cpudata->vid.max - cpudata->vid.min,
401 		int_tofp(cpudata->pstate.max_pstate -
402 			cpudata->pstate.min_pstate));
403 }
404 
405 
406 static int core_get_min_pstate(void)
407 {
408 	u64 value;
409 	rdmsrl(MSR_PLATFORM_INFO, value);
410 	return (value >> 40) & 0xFF;
411 }
412 
413 static int core_get_max_pstate(void)
414 {
415 	u64 value;
416 	rdmsrl(MSR_PLATFORM_INFO, value);
417 	return (value >> 8) & 0xFF;
418 }
419 
420 static int core_get_turbo_pstate(void)
421 {
422 	u64 value;
423 	int nont, ret;
424 	rdmsrl(MSR_NHM_TURBO_RATIO_LIMIT, value);
425 	nont = core_get_max_pstate();
426 	ret = ((value) & 255);
427 	if (ret <= nont)
428 		ret = nont;
429 	return ret;
430 }
431 
432 static void core_set_pstate(struct cpudata *cpudata, int pstate)
433 {
434 	u64 val;
435 
436 	val = pstate << 8;
437 	if (limits.no_turbo)
438 		val |= (u64)1 << 32;
439 
440 	wrmsrl(MSR_IA32_PERF_CTL, val);
441 }
442 
443 static struct cpu_defaults core_params = {
444 	.pid_policy = {
445 		.sample_rate_ms = 10,
446 		.deadband = 0,
447 		.setpoint = 97,
448 		.p_gain_pct = 20,
449 		.d_gain_pct = 0,
450 		.i_gain_pct = 0,
451 	},
452 	.funcs = {
453 		.get_max = core_get_max_pstate,
454 		.get_min = core_get_min_pstate,
455 		.get_turbo = core_get_turbo_pstate,
456 		.set = core_set_pstate,
457 	},
458 };
459 
460 static struct cpu_defaults byt_params = {
461 	.pid_policy = {
462 		.sample_rate_ms = 10,
463 		.deadband = 0,
464 		.setpoint = 97,
465 		.p_gain_pct = 14,
466 		.d_gain_pct = 0,
467 		.i_gain_pct = 4,
468 	},
469 	.funcs = {
470 		.get_max = byt_get_max_pstate,
471 		.get_min = byt_get_min_pstate,
472 		.get_turbo = byt_get_max_pstate,
473 		.set = byt_set_pstate,
474 		.get_vid = byt_get_vid,
475 	},
476 };
477 
478 
479 static void intel_pstate_get_min_max(struct cpudata *cpu, int *min, int *max)
480 {
481 	int max_perf = cpu->pstate.turbo_pstate;
482 	int max_perf_adj;
483 	int min_perf;
484 	if (limits.no_turbo)
485 		max_perf = cpu->pstate.max_pstate;
486 
487 	max_perf_adj = fp_toint(mul_fp(int_tofp(max_perf), limits.max_perf));
488 	*max = clamp_t(int, max_perf_adj,
489 			cpu->pstate.min_pstate, cpu->pstate.turbo_pstate);
490 
491 	min_perf = fp_toint(mul_fp(int_tofp(max_perf), limits.min_perf));
492 	*min = clamp_t(int, min_perf,
493 			cpu->pstate.min_pstate, max_perf);
494 }
495 
496 static void intel_pstate_set_pstate(struct cpudata *cpu, int pstate)
497 {
498 	int max_perf, min_perf;
499 
500 	intel_pstate_get_min_max(cpu, &min_perf, &max_perf);
501 
502 	pstate = clamp_t(int, pstate, min_perf, max_perf);
503 
504 	if (pstate == cpu->pstate.current_pstate)
505 		return;
506 
507 	trace_cpu_frequency(pstate * 100000, cpu->cpu);
508 
509 	cpu->pstate.current_pstate = pstate;
510 
511 	pstate_funcs.set(cpu, pstate);
512 }
513 
514 static inline void intel_pstate_pstate_increase(struct cpudata *cpu, int steps)
515 {
516 	int target;
517 	target = cpu->pstate.current_pstate + steps;
518 
519 	intel_pstate_set_pstate(cpu, target);
520 }
521 
522 static inline void intel_pstate_pstate_decrease(struct cpudata *cpu, int steps)
523 {
524 	int target;
525 	target = cpu->pstate.current_pstate - steps;
526 	intel_pstate_set_pstate(cpu, target);
527 }
528 
529 static void intel_pstate_get_cpu_pstates(struct cpudata *cpu)
530 {
531 	sprintf(cpu->name, "Intel 2nd generation core");
532 
533 	cpu->pstate.min_pstate = pstate_funcs.get_min();
534 	cpu->pstate.max_pstate = pstate_funcs.get_max();
535 	cpu->pstate.turbo_pstate = pstate_funcs.get_turbo();
536 
537 	if (pstate_funcs.get_vid)
538 		pstate_funcs.get_vid(cpu);
539 
540 	/*
541 	 * goto max pstate so we don't slow up boot if we are built-in if we are
542 	 * a module we will take care of it during normal operation
543 	 */
544 	intel_pstate_set_pstate(cpu, cpu->pstate.max_pstate);
545 }
546 
547 static inline void intel_pstate_calc_busy(struct cpudata *cpu,
548 					struct sample *sample)
549 {
550 	u64 core_pct;
551 	core_pct = div64_u64(int_tofp(sample->aperf * 100),
552 			     sample->mperf);
553 	sample->freq = fp_toint(cpu->pstate.max_pstate * core_pct * 1000);
554 
555 	sample->core_pct_busy = core_pct;
556 }
557 
558 static inline void intel_pstate_sample(struct cpudata *cpu)
559 {
560 	u64 aperf, mperf;
561 
562 	rdmsrl(MSR_IA32_APERF, aperf);
563 	rdmsrl(MSR_IA32_MPERF, mperf);
564 
565 	cpu->sample_ptr = (cpu->sample_ptr + 1) % SAMPLE_COUNT;
566 	cpu->samples[cpu->sample_ptr].aperf = aperf;
567 	cpu->samples[cpu->sample_ptr].mperf = mperf;
568 	cpu->samples[cpu->sample_ptr].aperf -= cpu->prev_aperf;
569 	cpu->samples[cpu->sample_ptr].mperf -= cpu->prev_mperf;
570 
571 	intel_pstate_calc_busy(cpu, &cpu->samples[cpu->sample_ptr]);
572 
573 	cpu->prev_aperf = aperf;
574 	cpu->prev_mperf = mperf;
575 }
576 
577 static inline void intel_pstate_set_sample_time(struct cpudata *cpu)
578 {
579 	int sample_time, delay;
580 
581 	sample_time = pid_params.sample_rate_ms;
582 	delay = msecs_to_jiffies(sample_time);
583 	mod_timer_pinned(&cpu->timer, jiffies + delay);
584 }
585 
586 static inline int32_t intel_pstate_get_scaled_busy(struct cpudata *cpu)
587 {
588 	int32_t core_busy, max_pstate, current_pstate;
589 
590 	core_busy = cpu->samples[cpu->sample_ptr].core_pct_busy;
591 	max_pstate = int_tofp(cpu->pstate.max_pstate);
592 	current_pstate = int_tofp(cpu->pstate.current_pstate);
593 	return mul_fp(core_busy, div_fp(max_pstate, current_pstate));
594 }
595 
596 static inline void intel_pstate_adjust_busy_pstate(struct cpudata *cpu)
597 {
598 	int32_t busy_scaled;
599 	struct _pid *pid;
600 	signed int ctl = 0;
601 	int steps;
602 
603 	pid = &cpu->pid;
604 	busy_scaled = intel_pstate_get_scaled_busy(cpu);
605 
606 	ctl = pid_calc(pid, busy_scaled);
607 
608 	steps = abs(ctl);
609 
610 	if (ctl < 0)
611 		intel_pstate_pstate_increase(cpu, steps);
612 	else
613 		intel_pstate_pstate_decrease(cpu, steps);
614 }
615 
616 static void intel_pstate_timer_func(unsigned long __data)
617 {
618 	struct cpudata *cpu = (struct cpudata *) __data;
619 	struct sample *sample;
620 	u64 energy;
621 
622 	intel_pstate_sample(cpu);
623 
624 	sample = &cpu->samples[cpu->sample_ptr];
625 	rdmsrl(MSR_PKG_ENERGY_STATUS, energy);
626 
627 	intel_pstate_adjust_busy_pstate(cpu);
628 
629 	trace_pstate_sample(fp_toint(sample->core_pct_busy),
630 			fp_toint(intel_pstate_get_scaled_busy(cpu)),
631 			cpu->pstate.current_pstate,
632 			sample->mperf,
633 			sample->aperf,
634 			div64_u64(energy, energy_divisor),
635 			sample->freq);
636 
637 	intel_pstate_set_sample_time(cpu);
638 }
639 
640 #define ICPU(model, policy) \
641 	{ X86_VENDOR_INTEL, 6, model, X86_FEATURE_APERFMPERF,\
642 			(unsigned long)&policy }
643 
644 static const struct x86_cpu_id intel_pstate_cpu_ids[] = {
645 	ICPU(0x2a, core_params),
646 	ICPU(0x2d, core_params),
647 	ICPU(0x37, byt_params),
648 	ICPU(0x3a, core_params),
649 	ICPU(0x3c, core_params),
650 	ICPU(0x3e, core_params),
651 	ICPU(0x3f, core_params),
652 	ICPU(0x45, core_params),
653 	ICPU(0x46, core_params),
654 	{}
655 };
656 MODULE_DEVICE_TABLE(x86cpu, intel_pstate_cpu_ids);
657 
658 static int intel_pstate_init_cpu(unsigned int cpunum)
659 {
660 
661 	const struct x86_cpu_id *id;
662 	struct cpudata *cpu;
663 
664 	id = x86_match_cpu(intel_pstate_cpu_ids);
665 	if (!id)
666 		return -ENODEV;
667 
668 	all_cpu_data[cpunum] = kzalloc(sizeof(struct cpudata), GFP_KERNEL);
669 	if (!all_cpu_data[cpunum])
670 		return -ENOMEM;
671 
672 	cpu = all_cpu_data[cpunum];
673 
674 	intel_pstate_get_cpu_pstates(cpu);
675 	if (!cpu->pstate.current_pstate) {
676 		all_cpu_data[cpunum] = NULL;
677 		kfree(cpu);
678 		return -ENODATA;
679 	}
680 
681 	cpu->cpu = cpunum;
682 
683 	init_timer_deferrable(&cpu->timer);
684 	cpu->timer.function = intel_pstate_timer_func;
685 	cpu->timer.data =
686 		(unsigned long)cpu;
687 	cpu->timer.expires = jiffies + HZ/100;
688 	intel_pstate_busy_pid_reset(cpu);
689 	intel_pstate_sample(cpu);
690 	intel_pstate_set_pstate(cpu, cpu->pstate.max_pstate);
691 
692 	add_timer_on(&cpu->timer, cpunum);
693 
694 	pr_info("Intel pstate controlling: cpu %d\n", cpunum);
695 
696 	return 0;
697 }
698 
699 static unsigned int intel_pstate_get(unsigned int cpu_num)
700 {
701 	struct sample *sample;
702 	struct cpudata *cpu;
703 
704 	cpu = all_cpu_data[cpu_num];
705 	if (!cpu)
706 		return 0;
707 	sample = &cpu->samples[cpu->sample_ptr];
708 	return sample->freq;
709 }
710 
711 static int intel_pstate_set_policy(struct cpufreq_policy *policy)
712 {
713 	struct cpudata *cpu;
714 
715 	cpu = all_cpu_data[policy->cpu];
716 
717 	if (!policy->cpuinfo.max_freq)
718 		return -ENODEV;
719 
720 	if (policy->policy == CPUFREQ_POLICY_PERFORMANCE) {
721 		limits.min_perf_pct = 100;
722 		limits.min_perf = int_tofp(1);
723 		limits.max_perf_pct = 100;
724 		limits.max_perf = int_tofp(1);
725 		limits.no_turbo = 0;
726 		return 0;
727 	}
728 	limits.min_perf_pct = (policy->min * 100) / policy->cpuinfo.max_freq;
729 	limits.min_perf_pct = clamp_t(int, limits.min_perf_pct, 0 , 100);
730 	limits.min_perf = div_fp(int_tofp(limits.min_perf_pct), int_tofp(100));
731 
732 	limits.max_policy_pct = policy->max * 100 / policy->cpuinfo.max_freq;
733 	limits.max_policy_pct = clamp_t(int, limits.max_policy_pct, 0 , 100);
734 	limits.max_perf_pct = min(limits.max_policy_pct, limits.max_sysfs_pct);
735 	limits.max_perf = div_fp(int_tofp(limits.max_perf_pct), int_tofp(100));
736 
737 	return 0;
738 }
739 
740 static int intel_pstate_verify_policy(struct cpufreq_policy *policy)
741 {
742 	cpufreq_verify_within_cpu_limits(policy);
743 
744 	if ((policy->policy != CPUFREQ_POLICY_POWERSAVE) &&
745 		(policy->policy != CPUFREQ_POLICY_PERFORMANCE))
746 		return -EINVAL;
747 
748 	return 0;
749 }
750 
751 static int intel_pstate_cpu_exit(struct cpufreq_policy *policy)
752 {
753 	int cpu = policy->cpu;
754 
755 	del_timer(&all_cpu_data[cpu]->timer);
756 	kfree(all_cpu_data[cpu]);
757 	all_cpu_data[cpu] = NULL;
758 	return 0;
759 }
760 
761 static int intel_pstate_cpu_init(struct cpufreq_policy *policy)
762 {
763 	struct cpudata *cpu;
764 	int rc;
765 
766 	rc = intel_pstate_init_cpu(policy->cpu);
767 	if (rc)
768 		return rc;
769 
770 	cpu = all_cpu_data[policy->cpu];
771 
772 	if (!limits.no_turbo &&
773 		limits.min_perf_pct == 100 && limits.max_perf_pct == 100)
774 		policy->policy = CPUFREQ_POLICY_PERFORMANCE;
775 	else
776 		policy->policy = CPUFREQ_POLICY_POWERSAVE;
777 
778 	policy->min = cpu->pstate.min_pstate * 100000;
779 	policy->max = cpu->pstate.turbo_pstate * 100000;
780 
781 	/* cpuinfo and default policy values */
782 	policy->cpuinfo.min_freq = cpu->pstate.min_pstate * 100000;
783 	policy->cpuinfo.max_freq = cpu->pstate.turbo_pstate * 100000;
784 	policy->cpuinfo.transition_latency = CPUFREQ_ETERNAL;
785 	cpumask_set_cpu(policy->cpu, policy->cpus);
786 
787 	return 0;
788 }
789 
790 static struct cpufreq_driver intel_pstate_driver = {
791 	.flags		= CPUFREQ_CONST_LOOPS,
792 	.verify		= intel_pstate_verify_policy,
793 	.setpolicy	= intel_pstate_set_policy,
794 	.get		= intel_pstate_get,
795 	.init		= intel_pstate_cpu_init,
796 	.exit		= intel_pstate_cpu_exit,
797 	.name		= "intel_pstate",
798 };
799 
800 static int __initdata no_load;
801 
802 static int intel_pstate_msrs_not_valid(void)
803 {
804 	/* Check that all the msr's we are using are valid. */
805 	u64 aperf, mperf, tmp;
806 
807 	rdmsrl(MSR_IA32_APERF, aperf);
808 	rdmsrl(MSR_IA32_MPERF, mperf);
809 
810 	if (!pstate_funcs.get_max() ||
811 		!pstate_funcs.get_min() ||
812 		!pstate_funcs.get_turbo())
813 		return -ENODEV;
814 
815 	rdmsrl(MSR_IA32_APERF, tmp);
816 	if (!(tmp - aperf))
817 		return -ENODEV;
818 
819 	rdmsrl(MSR_IA32_MPERF, tmp);
820 	if (!(tmp - mperf))
821 		return -ENODEV;
822 
823 	return 0;
824 }
825 
826 static void copy_pid_params(struct pstate_adjust_policy *policy)
827 {
828 	pid_params.sample_rate_ms = policy->sample_rate_ms;
829 	pid_params.p_gain_pct = policy->p_gain_pct;
830 	pid_params.i_gain_pct = policy->i_gain_pct;
831 	pid_params.d_gain_pct = policy->d_gain_pct;
832 	pid_params.deadband = policy->deadband;
833 	pid_params.setpoint = policy->setpoint;
834 }
835 
836 static void copy_cpu_funcs(struct pstate_funcs *funcs)
837 {
838 	pstate_funcs.get_max   = funcs->get_max;
839 	pstate_funcs.get_min   = funcs->get_min;
840 	pstate_funcs.get_turbo = funcs->get_turbo;
841 	pstate_funcs.set       = funcs->set;
842 	pstate_funcs.get_vid   = funcs->get_vid;
843 }
844 
845 #if IS_ENABLED(CONFIG_ACPI)
846 #include <acpi/processor.h>
847 
848 static bool intel_pstate_no_acpi_pss(void)
849 {
850 	int i;
851 
852 	for_each_possible_cpu(i) {
853 		acpi_status status;
854 		union acpi_object *pss;
855 		struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
856 		struct acpi_processor *pr = per_cpu(processors, i);
857 
858 		if (!pr)
859 			continue;
860 
861 		status = acpi_evaluate_object(pr->handle, "_PSS", NULL, &buffer);
862 		if (ACPI_FAILURE(status))
863 			continue;
864 
865 		pss = buffer.pointer;
866 		if (pss && pss->type == ACPI_TYPE_PACKAGE) {
867 			kfree(pss);
868 			return false;
869 		}
870 
871 		kfree(pss);
872 	}
873 
874 	return true;
875 }
876 
877 struct hw_vendor_info {
878 	u16  valid;
879 	char oem_id[ACPI_OEM_ID_SIZE];
880 	char oem_table_id[ACPI_OEM_TABLE_ID_SIZE];
881 };
882 
883 /* Hardware vendor-specific info that has its own power management modes */
884 static struct hw_vendor_info vendor_info[] = {
885 	{1, "HP    ", "ProLiant"},
886 	{0, "", ""},
887 };
888 
889 static bool intel_pstate_platform_pwr_mgmt_exists(void)
890 {
891 	struct acpi_table_header hdr;
892 	struct hw_vendor_info *v_info;
893 
894 	if (acpi_disabled
895 	    || ACPI_FAILURE(acpi_get_table_header(ACPI_SIG_FADT, 0, &hdr)))
896 		return false;
897 
898 	for (v_info = vendor_info; v_info->valid; v_info++) {
899 		if (!strncmp(hdr.oem_id, v_info->oem_id, ACPI_OEM_ID_SIZE)
900 		    && !strncmp(hdr.oem_table_id, v_info->oem_table_id, ACPI_OEM_TABLE_ID_SIZE)
901 		    && intel_pstate_no_acpi_pss())
902 			return true;
903 	}
904 
905 	return false;
906 }
907 #else /* CONFIG_ACPI not enabled */
908 static inline bool intel_pstate_platform_pwr_mgmt_exists(void) { return false; }
909 #endif /* CONFIG_ACPI */
910 
911 static int __init intel_pstate_init(void)
912 {
913 	int cpu, rc = 0;
914 	const struct x86_cpu_id *id;
915 	struct cpu_defaults *cpu_info;
916 	u64 units;
917 
918 	if (no_load)
919 		return -ENODEV;
920 
921 	id = x86_match_cpu(intel_pstate_cpu_ids);
922 	if (!id)
923 		return -ENODEV;
924 
925 	/*
926 	 * The Intel pstate driver will be ignored if the platform
927 	 * firmware has its own power management modes.
928 	 */
929 	if (intel_pstate_platform_pwr_mgmt_exists())
930 		return -ENODEV;
931 
932 	cpu_info = (struct cpu_defaults *)id->driver_data;
933 
934 	copy_pid_params(&cpu_info->pid_policy);
935 	copy_cpu_funcs(&cpu_info->funcs);
936 
937 	if (intel_pstate_msrs_not_valid())
938 		return -ENODEV;
939 
940 	pr_info("Intel P-state driver initializing.\n");
941 
942 	all_cpu_data = vzalloc(sizeof(void *) * num_possible_cpus());
943 	if (!all_cpu_data)
944 		return -ENOMEM;
945 
946 	rc = cpufreq_register_driver(&intel_pstate_driver);
947 	if (rc)
948 		goto out;
949 
950 	rdmsrl(MSR_RAPL_POWER_UNIT, units);
951 	energy_divisor = 1 << ((units >> 8) & 0x1f); /* bits{12:8} */
952 
953 	intel_pstate_debug_expose_params();
954 	intel_pstate_sysfs_expose_params();
955 
956 	return rc;
957 out:
958 	get_online_cpus();
959 	for_each_online_cpu(cpu) {
960 		if (all_cpu_data[cpu]) {
961 			del_timer_sync(&all_cpu_data[cpu]->timer);
962 			kfree(all_cpu_data[cpu]);
963 		}
964 	}
965 
966 	put_online_cpus();
967 	vfree(all_cpu_data);
968 	return -ENODEV;
969 }
970 device_initcall(intel_pstate_init);
971 
972 static int __init intel_pstate_setup(char *str)
973 {
974 	if (!str)
975 		return -EINVAL;
976 
977 	if (!strcmp(str, "disable"))
978 		no_load = 1;
979 	return 0;
980 }
981 early_param("intel_pstate", intel_pstate_setup);
982 
983 MODULE_AUTHOR("Dirk Brandewie <dirk.j.brandewie@intel.com>");
984 MODULE_DESCRIPTION("'intel_pstate' - P state driver Intel Core processors");
985 MODULE_LICENSE("GPL");
986