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 BYT_RATIOS		0x66a
36 #define BYT_VIDS		0x66b
37 #define BYT_TURBO_RATIOS	0x66c
38 #define BYT_TURBO_VIDS		0x66d
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 
45 static inline int32_t mul_fp(int32_t x, int32_t y)
46 {
47 	return ((int64_t)x * (int64_t)y) >> FRAC_BITS;
48 }
49 
50 static inline int32_t div_fp(int32_t x, int32_t y)
51 {
52 	return div_s64((int64_t)x << FRAC_BITS, y);
53 }
54 
55 static inline int ceiling_fp(int32_t x)
56 {
57 	int mask, ret;
58 
59 	ret = fp_toint(x);
60 	mask = (1 << FRAC_BITS) - 1;
61 	if (x & mask)
62 		ret += 1;
63 	return ret;
64 }
65 
66 struct sample {
67 	int32_t core_pct_busy;
68 	u64 aperf;
69 	u64 mperf;
70 	int freq;
71 	ktime_t time;
72 };
73 
74 struct pstate_data {
75 	int	current_pstate;
76 	int	min_pstate;
77 	int	max_pstate;
78 	int	scaling;
79 	int	turbo_pstate;
80 };
81 
82 struct vid_data {
83 	int min;
84 	int max;
85 	int turbo;
86 	int32_t ratio;
87 };
88 
89 struct _pid {
90 	int setpoint;
91 	int32_t integral;
92 	int32_t p_gain;
93 	int32_t i_gain;
94 	int32_t d_gain;
95 	int deadband;
96 	int32_t last_err;
97 };
98 
99 struct cpudata {
100 	int cpu;
101 
102 	struct timer_list timer;
103 
104 	struct pstate_data pstate;
105 	struct vid_data vid;
106 	struct _pid pid;
107 
108 	ktime_t last_sample_time;
109 	u64	prev_aperf;
110 	u64	prev_mperf;
111 	struct sample sample;
112 };
113 
114 static struct cpudata **all_cpu_data;
115 struct pstate_adjust_policy {
116 	int sample_rate_ms;
117 	int deadband;
118 	int setpoint;
119 	int p_gain_pct;
120 	int d_gain_pct;
121 	int i_gain_pct;
122 };
123 
124 struct pstate_funcs {
125 	int (*get_max)(void);
126 	int (*get_min)(void);
127 	int (*get_turbo)(void);
128 	int (*get_scaling)(void);
129 	void (*set)(struct cpudata*, int pstate);
130 	void (*get_vid)(struct cpudata *);
131 };
132 
133 struct cpu_defaults {
134 	struct pstate_adjust_policy pid_policy;
135 	struct pstate_funcs funcs;
136 };
137 
138 static struct pstate_adjust_policy pid_params;
139 static struct pstate_funcs pstate_funcs;
140 static int hwp_active;
141 
142 struct perf_limits {
143 	int no_turbo;
144 	int turbo_disabled;
145 	int max_perf_pct;
146 	int min_perf_pct;
147 	int32_t max_perf;
148 	int32_t min_perf;
149 	int max_policy_pct;
150 	int max_sysfs_pct;
151 	int min_policy_pct;
152 	int min_sysfs_pct;
153 };
154 
155 static struct perf_limits limits = {
156 	.no_turbo = 0,
157 	.turbo_disabled = 0,
158 	.max_perf_pct = 100,
159 	.max_perf = int_tofp(1),
160 	.min_perf_pct = 0,
161 	.min_perf = 0,
162 	.max_policy_pct = 100,
163 	.max_sysfs_pct = 100,
164 	.min_policy_pct = 0,
165 	.min_sysfs_pct = 0,
166 };
167 
168 static inline void pid_reset(struct _pid *pid, int setpoint, int busy,
169 			     int deadband, int integral) {
170 	pid->setpoint = setpoint;
171 	pid->deadband  = deadband;
172 	pid->integral  = int_tofp(integral);
173 	pid->last_err  = int_tofp(setpoint) - int_tofp(busy);
174 }
175 
176 static inline void pid_p_gain_set(struct _pid *pid, int percent)
177 {
178 	pid->p_gain = div_fp(int_tofp(percent), int_tofp(100));
179 }
180 
181 static inline void pid_i_gain_set(struct _pid *pid, int percent)
182 {
183 	pid->i_gain = div_fp(int_tofp(percent), int_tofp(100));
184 }
185 
186 static inline void pid_d_gain_set(struct _pid *pid, int percent)
187 {
188 	pid->d_gain = div_fp(int_tofp(percent), int_tofp(100));
189 }
190 
191 static signed int pid_calc(struct _pid *pid, int32_t busy)
192 {
193 	signed int result;
194 	int32_t pterm, dterm, fp_error;
195 	int32_t integral_limit;
196 
197 	fp_error = int_tofp(pid->setpoint) - busy;
198 
199 	if (abs(fp_error) <= int_tofp(pid->deadband))
200 		return 0;
201 
202 	pterm = mul_fp(pid->p_gain, fp_error);
203 
204 	pid->integral += fp_error;
205 
206 	/*
207 	 * We limit the integral here so that it will never
208 	 * get higher than 30.  This prevents it from becoming
209 	 * too large an input over long periods of time and allows
210 	 * it to get factored out sooner.
211 	 *
212 	 * The value of 30 was chosen through experimentation.
213 	 */
214 	integral_limit = int_tofp(30);
215 	if (pid->integral > integral_limit)
216 		pid->integral = integral_limit;
217 	if (pid->integral < -integral_limit)
218 		pid->integral = -integral_limit;
219 
220 	dterm = mul_fp(pid->d_gain, fp_error - pid->last_err);
221 	pid->last_err = fp_error;
222 
223 	result = pterm + mul_fp(pid->integral, pid->i_gain) + dterm;
224 	result = result + (1 << (FRAC_BITS-1));
225 	return (signed int)fp_toint(result);
226 }
227 
228 static inline void intel_pstate_busy_pid_reset(struct cpudata *cpu)
229 {
230 	pid_p_gain_set(&cpu->pid, pid_params.p_gain_pct);
231 	pid_d_gain_set(&cpu->pid, pid_params.d_gain_pct);
232 	pid_i_gain_set(&cpu->pid, pid_params.i_gain_pct);
233 
234 	pid_reset(&cpu->pid, pid_params.setpoint, 100, pid_params.deadband, 0);
235 }
236 
237 static inline void intel_pstate_reset_all_pid(void)
238 {
239 	unsigned int cpu;
240 
241 	for_each_online_cpu(cpu) {
242 		if (all_cpu_data[cpu])
243 			intel_pstate_busy_pid_reset(all_cpu_data[cpu]);
244 	}
245 }
246 
247 static inline void update_turbo_state(void)
248 {
249 	u64 misc_en;
250 	struct cpudata *cpu;
251 
252 	cpu = all_cpu_data[0];
253 	rdmsrl(MSR_IA32_MISC_ENABLE, misc_en);
254 	limits.turbo_disabled =
255 		(misc_en & MSR_IA32_MISC_ENABLE_TURBO_DISABLE ||
256 		 cpu->pstate.max_pstate == cpu->pstate.turbo_pstate);
257 }
258 
259 #define PCT_TO_HWP(x) (x * 255 / 100)
260 static void intel_pstate_hwp_set(void)
261 {
262 	int min, max, cpu;
263 	u64 value, freq;
264 
265 	get_online_cpus();
266 
267 	for_each_online_cpu(cpu) {
268 		rdmsrl_on_cpu(cpu, MSR_HWP_REQUEST, &value);
269 		min = PCT_TO_HWP(limits.min_perf_pct);
270 		value &= ~HWP_MIN_PERF(~0L);
271 		value |= HWP_MIN_PERF(min);
272 
273 		max = PCT_TO_HWP(limits.max_perf_pct);
274 		if (limits.no_turbo) {
275 			rdmsrl( MSR_HWP_CAPABILITIES, freq);
276 			max = HWP_GUARANTEED_PERF(freq);
277 		}
278 
279 		value &= ~HWP_MAX_PERF(~0L);
280 		value |= HWP_MAX_PERF(max);
281 		wrmsrl_on_cpu(cpu, MSR_HWP_REQUEST, value);
282 	}
283 
284 	put_online_cpus();
285 }
286 
287 /************************** debugfs begin ************************/
288 static int pid_param_set(void *data, u64 val)
289 {
290 	*(u32 *)data = val;
291 	intel_pstate_reset_all_pid();
292 	return 0;
293 }
294 
295 static int pid_param_get(void *data, u64 *val)
296 {
297 	*val = *(u32 *)data;
298 	return 0;
299 }
300 DEFINE_SIMPLE_ATTRIBUTE(fops_pid_param, pid_param_get, pid_param_set, "%llu\n");
301 
302 struct pid_param {
303 	char *name;
304 	void *value;
305 };
306 
307 static struct pid_param pid_files[] = {
308 	{"sample_rate_ms", &pid_params.sample_rate_ms},
309 	{"d_gain_pct", &pid_params.d_gain_pct},
310 	{"i_gain_pct", &pid_params.i_gain_pct},
311 	{"deadband", &pid_params.deadband},
312 	{"setpoint", &pid_params.setpoint},
313 	{"p_gain_pct", &pid_params.p_gain_pct},
314 	{NULL, NULL}
315 };
316 
317 static void __init intel_pstate_debug_expose_params(void)
318 {
319 	struct dentry *debugfs_parent;
320 	int i = 0;
321 
322 	if (hwp_active)
323 		return;
324 	debugfs_parent = debugfs_create_dir("pstate_snb", NULL);
325 	if (IS_ERR_OR_NULL(debugfs_parent))
326 		return;
327 	while (pid_files[i].name) {
328 		debugfs_create_file(pid_files[i].name, 0660,
329 				    debugfs_parent, pid_files[i].value,
330 				    &fops_pid_param);
331 		i++;
332 	}
333 }
334 
335 /************************** debugfs end ************************/
336 
337 /************************** sysfs begin ************************/
338 #define show_one(file_name, object)					\
339 	static ssize_t show_##file_name					\
340 	(struct kobject *kobj, struct attribute *attr, char *buf)	\
341 	{								\
342 		return sprintf(buf, "%u\n", limits.object);		\
343 	}
344 
345 static ssize_t show_turbo_pct(struct kobject *kobj,
346 				struct attribute *attr, char *buf)
347 {
348 	struct cpudata *cpu;
349 	int total, no_turbo, turbo_pct;
350 	uint32_t turbo_fp;
351 
352 	cpu = all_cpu_data[0];
353 
354 	total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
355 	no_turbo = cpu->pstate.max_pstate - cpu->pstate.min_pstate + 1;
356 	turbo_fp = div_fp(int_tofp(no_turbo), int_tofp(total));
357 	turbo_pct = 100 - fp_toint(mul_fp(turbo_fp, int_tofp(100)));
358 	return sprintf(buf, "%u\n", turbo_pct);
359 }
360 
361 static ssize_t show_num_pstates(struct kobject *kobj,
362 				struct attribute *attr, char *buf)
363 {
364 	struct cpudata *cpu;
365 	int total;
366 
367 	cpu = all_cpu_data[0];
368 	total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
369 	return sprintf(buf, "%u\n", total);
370 }
371 
372 static ssize_t show_no_turbo(struct kobject *kobj,
373 			     struct attribute *attr, char *buf)
374 {
375 	ssize_t ret;
376 
377 	update_turbo_state();
378 	if (limits.turbo_disabled)
379 		ret = sprintf(buf, "%u\n", limits.turbo_disabled);
380 	else
381 		ret = sprintf(buf, "%u\n", limits.no_turbo);
382 
383 	return ret;
384 }
385 
386 static ssize_t store_no_turbo(struct kobject *a, struct attribute *b,
387 			      const char *buf, size_t count)
388 {
389 	unsigned int input;
390 	int ret;
391 
392 	ret = sscanf(buf, "%u", &input);
393 	if (ret != 1)
394 		return -EINVAL;
395 
396 	update_turbo_state();
397 	if (limits.turbo_disabled) {
398 		pr_warn("Turbo disabled by BIOS or unavailable on processor\n");
399 		return -EPERM;
400 	}
401 
402 	limits.no_turbo = clamp_t(int, input, 0, 1);
403 
404 	if (hwp_active)
405 		intel_pstate_hwp_set();
406 
407 	return count;
408 }
409 
410 static ssize_t store_max_perf_pct(struct kobject *a, struct attribute *b,
411 				  const char *buf, size_t count)
412 {
413 	unsigned int input;
414 	int ret;
415 
416 	ret = sscanf(buf, "%u", &input);
417 	if (ret != 1)
418 		return -EINVAL;
419 
420 	limits.max_sysfs_pct = clamp_t(int, input, 0 , 100);
421 	limits.max_perf_pct = min(limits.max_policy_pct, limits.max_sysfs_pct);
422 	limits.max_perf = div_fp(int_tofp(limits.max_perf_pct), int_tofp(100));
423 
424 	if (hwp_active)
425 		intel_pstate_hwp_set();
426 	return count;
427 }
428 
429 static ssize_t store_min_perf_pct(struct kobject *a, struct attribute *b,
430 				  const char *buf, size_t count)
431 {
432 	unsigned int input;
433 	int ret;
434 
435 	ret = sscanf(buf, "%u", &input);
436 	if (ret != 1)
437 		return -EINVAL;
438 
439 	limits.min_sysfs_pct = clamp_t(int, input, 0 , 100);
440 	limits.min_perf_pct = max(limits.min_policy_pct, limits.min_sysfs_pct);
441 	limits.min_perf = div_fp(int_tofp(limits.min_perf_pct), int_tofp(100));
442 
443 	if (hwp_active)
444 		intel_pstate_hwp_set();
445 	return count;
446 }
447 
448 show_one(max_perf_pct, max_perf_pct);
449 show_one(min_perf_pct, min_perf_pct);
450 
451 define_one_global_rw(no_turbo);
452 define_one_global_rw(max_perf_pct);
453 define_one_global_rw(min_perf_pct);
454 define_one_global_ro(turbo_pct);
455 define_one_global_ro(num_pstates);
456 
457 static struct attribute *intel_pstate_attributes[] = {
458 	&no_turbo.attr,
459 	&max_perf_pct.attr,
460 	&min_perf_pct.attr,
461 	&turbo_pct.attr,
462 	&num_pstates.attr,
463 	NULL
464 };
465 
466 static struct attribute_group intel_pstate_attr_group = {
467 	.attrs = intel_pstate_attributes,
468 };
469 
470 static void __init intel_pstate_sysfs_expose_params(void)
471 {
472 	struct kobject *intel_pstate_kobject;
473 	int rc;
474 
475 	intel_pstate_kobject = kobject_create_and_add("intel_pstate",
476 						&cpu_subsys.dev_root->kobj);
477 	BUG_ON(!intel_pstate_kobject);
478 	rc = sysfs_create_group(intel_pstate_kobject, &intel_pstate_attr_group);
479 	BUG_ON(rc);
480 }
481 /************************** sysfs end ************************/
482 
483 static void intel_pstate_hwp_enable(void)
484 {
485 	hwp_active++;
486 	pr_info("intel_pstate HWP enabled\n");
487 
488 	wrmsrl( MSR_PM_ENABLE, 0x1);
489 }
490 
491 static int byt_get_min_pstate(void)
492 {
493 	u64 value;
494 
495 	rdmsrl(BYT_RATIOS, value);
496 	return (value >> 8) & 0x7F;
497 }
498 
499 static int byt_get_max_pstate(void)
500 {
501 	u64 value;
502 
503 	rdmsrl(BYT_RATIOS, value);
504 	return (value >> 16) & 0x7F;
505 }
506 
507 static int byt_get_turbo_pstate(void)
508 {
509 	u64 value;
510 
511 	rdmsrl(BYT_TURBO_RATIOS, value);
512 	return value & 0x7F;
513 }
514 
515 static void byt_set_pstate(struct cpudata *cpudata, int pstate)
516 {
517 	u64 val;
518 	int32_t vid_fp;
519 	u32 vid;
520 
521 	val = pstate << 8;
522 	if (limits.no_turbo && !limits.turbo_disabled)
523 		val |= (u64)1 << 32;
524 
525 	vid_fp = cpudata->vid.min + mul_fp(
526 		int_tofp(pstate - cpudata->pstate.min_pstate),
527 		cpudata->vid.ratio);
528 
529 	vid_fp = clamp_t(int32_t, vid_fp, cpudata->vid.min, cpudata->vid.max);
530 	vid = ceiling_fp(vid_fp);
531 
532 	if (pstate > cpudata->pstate.max_pstate)
533 		vid = cpudata->vid.turbo;
534 
535 	val |= vid;
536 
537 	wrmsrl(MSR_IA32_PERF_CTL, val);
538 }
539 
540 #define BYT_BCLK_FREQS 5
541 static int byt_freq_table[BYT_BCLK_FREQS] = { 833, 1000, 1333, 1167, 800};
542 
543 static int byt_get_scaling(void)
544 {
545 	u64 value;
546 	int i;
547 
548 	rdmsrl(MSR_FSB_FREQ, value);
549 	i = value & 0x3;
550 
551 	BUG_ON(i > BYT_BCLK_FREQS);
552 
553 	return byt_freq_table[i] * 100;
554 }
555 
556 static void byt_get_vid(struct cpudata *cpudata)
557 {
558 	u64 value;
559 
560 	rdmsrl(BYT_VIDS, value);
561 	cpudata->vid.min = int_tofp((value >> 8) & 0x7f);
562 	cpudata->vid.max = int_tofp((value >> 16) & 0x7f);
563 	cpudata->vid.ratio = div_fp(
564 		cpudata->vid.max - cpudata->vid.min,
565 		int_tofp(cpudata->pstate.max_pstate -
566 			cpudata->pstate.min_pstate));
567 
568 	rdmsrl(BYT_TURBO_VIDS, value);
569 	cpudata->vid.turbo = value & 0x7f;
570 }
571 
572 static int core_get_min_pstate(void)
573 {
574 	u64 value;
575 
576 	rdmsrl(MSR_PLATFORM_INFO, value);
577 	return (value >> 40) & 0xFF;
578 }
579 
580 static int core_get_max_pstate(void)
581 {
582 	u64 value;
583 
584 	rdmsrl(MSR_PLATFORM_INFO, value);
585 	return (value >> 8) & 0xFF;
586 }
587 
588 static int core_get_turbo_pstate(void)
589 {
590 	u64 value;
591 	int nont, ret;
592 
593 	rdmsrl(MSR_NHM_TURBO_RATIO_LIMIT, value);
594 	nont = core_get_max_pstate();
595 	ret = (value) & 255;
596 	if (ret <= nont)
597 		ret = nont;
598 	return ret;
599 }
600 
601 static inline int core_get_scaling(void)
602 {
603 	return 100000;
604 }
605 
606 static void core_set_pstate(struct cpudata *cpudata, int pstate)
607 {
608 	u64 val;
609 
610 	val = pstate << 8;
611 	if (limits.no_turbo && !limits.turbo_disabled)
612 		val |= (u64)1 << 32;
613 
614 	wrmsrl_on_cpu(cpudata->cpu, MSR_IA32_PERF_CTL, val);
615 }
616 
617 static struct cpu_defaults core_params = {
618 	.pid_policy = {
619 		.sample_rate_ms = 10,
620 		.deadband = 0,
621 		.setpoint = 97,
622 		.p_gain_pct = 20,
623 		.d_gain_pct = 0,
624 		.i_gain_pct = 0,
625 	},
626 	.funcs = {
627 		.get_max = core_get_max_pstate,
628 		.get_min = core_get_min_pstate,
629 		.get_turbo = core_get_turbo_pstate,
630 		.get_scaling = core_get_scaling,
631 		.set = core_set_pstate,
632 	},
633 };
634 
635 static struct cpu_defaults byt_params = {
636 	.pid_policy = {
637 		.sample_rate_ms = 10,
638 		.deadband = 0,
639 		.setpoint = 97,
640 		.p_gain_pct = 14,
641 		.d_gain_pct = 0,
642 		.i_gain_pct = 4,
643 	},
644 	.funcs = {
645 		.get_max = byt_get_max_pstate,
646 		.get_min = byt_get_min_pstate,
647 		.get_turbo = byt_get_turbo_pstate,
648 		.set = byt_set_pstate,
649 		.get_scaling = byt_get_scaling,
650 		.get_vid = byt_get_vid,
651 	},
652 };
653 
654 static void intel_pstate_get_min_max(struct cpudata *cpu, int *min, int *max)
655 {
656 	int max_perf = cpu->pstate.turbo_pstate;
657 	int max_perf_adj;
658 	int min_perf;
659 
660 	if (limits.no_turbo || limits.turbo_disabled)
661 		max_perf = cpu->pstate.max_pstate;
662 
663 	/*
664 	 * performance can be limited by user through sysfs, by cpufreq
665 	 * policy, or by cpu specific default values determined through
666 	 * experimentation.
667 	 */
668 	max_perf_adj = fp_toint(mul_fp(int_tofp(max_perf), limits.max_perf));
669 	*max = clamp_t(int, max_perf_adj,
670 			cpu->pstate.min_pstate, cpu->pstate.turbo_pstate);
671 
672 	min_perf = fp_toint(mul_fp(int_tofp(max_perf), limits.min_perf));
673 	*min = clamp_t(int, min_perf, cpu->pstate.min_pstate, max_perf);
674 }
675 
676 static void intel_pstate_set_pstate(struct cpudata *cpu, int pstate)
677 {
678 	int max_perf, min_perf;
679 
680 	update_turbo_state();
681 
682 	intel_pstate_get_min_max(cpu, &min_perf, &max_perf);
683 
684 	pstate = clamp_t(int, pstate, min_perf, max_perf);
685 
686 	if (pstate == cpu->pstate.current_pstate)
687 		return;
688 
689 	trace_cpu_frequency(pstate * cpu->pstate.scaling, cpu->cpu);
690 
691 	cpu->pstate.current_pstate = pstate;
692 
693 	pstate_funcs.set(cpu, pstate);
694 }
695 
696 static void intel_pstate_get_cpu_pstates(struct cpudata *cpu)
697 {
698 	cpu->pstate.min_pstate = pstate_funcs.get_min();
699 	cpu->pstate.max_pstate = pstate_funcs.get_max();
700 	cpu->pstate.turbo_pstate = pstate_funcs.get_turbo();
701 	cpu->pstate.scaling = pstate_funcs.get_scaling();
702 
703 	if (pstate_funcs.get_vid)
704 		pstate_funcs.get_vid(cpu);
705 	intel_pstate_set_pstate(cpu, cpu->pstate.min_pstate);
706 }
707 
708 static inline void intel_pstate_calc_busy(struct cpudata *cpu)
709 {
710 	struct sample *sample = &cpu->sample;
711 	int64_t core_pct;
712 
713 	core_pct = int_tofp(sample->aperf) * int_tofp(100);
714 	core_pct = div64_u64(core_pct, int_tofp(sample->mperf));
715 
716 	sample->freq = fp_toint(
717 		mul_fp(int_tofp(
718 			cpu->pstate.max_pstate * cpu->pstate.scaling / 100),
719 			core_pct));
720 
721 	sample->core_pct_busy = (int32_t)core_pct;
722 }
723 
724 static inline void intel_pstate_sample(struct cpudata *cpu)
725 {
726 	u64 aperf, mperf;
727 	unsigned long flags;
728 
729 	local_irq_save(flags);
730 	rdmsrl(MSR_IA32_APERF, aperf);
731 	rdmsrl(MSR_IA32_MPERF, mperf);
732 	local_irq_restore(flags);
733 
734 	cpu->last_sample_time = cpu->sample.time;
735 	cpu->sample.time = ktime_get();
736 	cpu->sample.aperf = aperf;
737 	cpu->sample.mperf = mperf;
738 	cpu->sample.aperf -= cpu->prev_aperf;
739 	cpu->sample.mperf -= cpu->prev_mperf;
740 
741 	intel_pstate_calc_busy(cpu);
742 
743 	cpu->prev_aperf = aperf;
744 	cpu->prev_mperf = mperf;
745 }
746 
747 static inline void intel_hwp_set_sample_time(struct cpudata *cpu)
748 {
749 	int delay;
750 
751 	delay = msecs_to_jiffies(50);
752 	mod_timer_pinned(&cpu->timer, jiffies + delay);
753 }
754 
755 static inline void intel_pstate_set_sample_time(struct cpudata *cpu)
756 {
757 	int delay;
758 
759 	delay = msecs_to_jiffies(pid_params.sample_rate_ms);
760 	mod_timer_pinned(&cpu->timer, jiffies + delay);
761 }
762 
763 static inline int32_t intel_pstate_get_scaled_busy(struct cpudata *cpu)
764 {
765 	int32_t core_busy, max_pstate, current_pstate, sample_ratio;
766 	u32 duration_us;
767 	u32 sample_time;
768 
769 	/*
770 	 * core_busy is the ratio of actual performance to max
771 	 * max_pstate is the max non turbo pstate available
772 	 * current_pstate was the pstate that was requested during
773 	 * 	the last sample period.
774 	 *
775 	 * We normalize core_busy, which was our actual percent
776 	 * performance to what we requested during the last sample
777 	 * period. The result will be a percentage of busy at a
778 	 * specified pstate.
779 	 */
780 	core_busy = cpu->sample.core_pct_busy;
781 	max_pstate = int_tofp(cpu->pstate.max_pstate);
782 	current_pstate = int_tofp(cpu->pstate.current_pstate);
783 	core_busy = mul_fp(core_busy, div_fp(max_pstate, current_pstate));
784 
785 	/*
786 	 * Since we have a deferred timer, it will not fire unless
787 	 * we are in C0.  So, determine if the actual elapsed time
788 	 * is significantly greater (3x) than our sample interval.  If it
789 	 * is, then we were idle for a long enough period of time
790 	 * to adjust our busyness.
791 	 */
792 	sample_time = pid_params.sample_rate_ms  * USEC_PER_MSEC;
793 	duration_us = (u32) ktime_us_delta(cpu->sample.time,
794 					   cpu->last_sample_time);
795 	if (duration_us > sample_time * 3) {
796 		sample_ratio = div_fp(int_tofp(sample_time),
797 				      int_tofp(duration_us));
798 		core_busy = mul_fp(core_busy, sample_ratio);
799 	}
800 
801 	return core_busy;
802 }
803 
804 static inline void intel_pstate_adjust_busy_pstate(struct cpudata *cpu)
805 {
806 	int32_t busy_scaled;
807 	struct _pid *pid;
808 	signed int ctl;
809 
810 	pid = &cpu->pid;
811 	busy_scaled = intel_pstate_get_scaled_busy(cpu);
812 
813 	ctl = pid_calc(pid, busy_scaled);
814 
815 	/* Negative values of ctl increase the pstate and vice versa */
816 	intel_pstate_set_pstate(cpu, cpu->pstate.current_pstate - ctl);
817 }
818 
819 static void intel_hwp_timer_func(unsigned long __data)
820 {
821 	struct cpudata *cpu = (struct cpudata *) __data;
822 
823 	intel_pstate_sample(cpu);
824 	intel_hwp_set_sample_time(cpu);
825 }
826 
827 static void intel_pstate_timer_func(unsigned long __data)
828 {
829 	struct cpudata *cpu = (struct cpudata *) __data;
830 	struct sample *sample;
831 
832 	intel_pstate_sample(cpu);
833 
834 	sample = &cpu->sample;
835 
836 	intel_pstate_adjust_busy_pstate(cpu);
837 
838 	trace_pstate_sample(fp_toint(sample->core_pct_busy),
839 			fp_toint(intel_pstate_get_scaled_busy(cpu)),
840 			cpu->pstate.current_pstate,
841 			sample->mperf,
842 			sample->aperf,
843 			sample->freq);
844 
845 	intel_pstate_set_sample_time(cpu);
846 }
847 
848 #define ICPU(model, policy) \
849 	{ X86_VENDOR_INTEL, 6, model, X86_FEATURE_APERFMPERF,\
850 			(unsigned long)&policy }
851 
852 static const struct x86_cpu_id intel_pstate_cpu_ids[] = {
853 	ICPU(0x2a, core_params),
854 	ICPU(0x2d, core_params),
855 	ICPU(0x37, byt_params),
856 	ICPU(0x3a, core_params),
857 	ICPU(0x3c, core_params),
858 	ICPU(0x3d, core_params),
859 	ICPU(0x3e, core_params),
860 	ICPU(0x3f, core_params),
861 	ICPU(0x45, core_params),
862 	ICPU(0x46, core_params),
863 	ICPU(0x47, core_params),
864 	ICPU(0x4c, byt_params),
865 	ICPU(0x4e, core_params),
866 	ICPU(0x4f, core_params),
867 	ICPU(0x56, core_params),
868 	{}
869 };
870 MODULE_DEVICE_TABLE(x86cpu, intel_pstate_cpu_ids);
871 
872 static const struct x86_cpu_id intel_pstate_cpu_oob_ids[] = {
873 	ICPU(0x56, core_params),
874 	{}
875 };
876 
877 static int intel_pstate_init_cpu(unsigned int cpunum)
878 {
879 	struct cpudata *cpu;
880 
881 	if (!all_cpu_data[cpunum])
882 		all_cpu_data[cpunum] = kzalloc(sizeof(struct cpudata),
883 					       GFP_KERNEL);
884 	if (!all_cpu_data[cpunum])
885 		return -ENOMEM;
886 
887 	cpu = all_cpu_data[cpunum];
888 
889 	cpu->cpu = cpunum;
890 	intel_pstate_get_cpu_pstates(cpu);
891 
892 	init_timer_deferrable(&cpu->timer);
893 	cpu->timer.data = (unsigned long)cpu;
894 	cpu->timer.expires = jiffies + HZ/100;
895 
896 	if (!hwp_active)
897 		cpu->timer.function = intel_pstate_timer_func;
898 	else
899 		cpu->timer.function = intel_hwp_timer_func;
900 
901 	intel_pstate_busy_pid_reset(cpu);
902 	intel_pstate_sample(cpu);
903 
904 	add_timer_on(&cpu->timer, cpunum);
905 
906 	pr_debug("Intel pstate controlling: cpu %d\n", cpunum);
907 
908 	return 0;
909 }
910 
911 static unsigned int intel_pstate_get(unsigned int cpu_num)
912 {
913 	struct sample *sample;
914 	struct cpudata *cpu;
915 
916 	cpu = all_cpu_data[cpu_num];
917 	if (!cpu)
918 		return 0;
919 	sample = &cpu->sample;
920 	return sample->freq;
921 }
922 
923 static int intel_pstate_set_policy(struct cpufreq_policy *policy)
924 {
925 	if (!policy->cpuinfo.max_freq)
926 		return -ENODEV;
927 
928 	if (policy->policy == CPUFREQ_POLICY_PERFORMANCE &&
929 	    policy->max >= policy->cpuinfo.max_freq) {
930 		limits.min_policy_pct = 100;
931 		limits.min_perf_pct = 100;
932 		limits.min_perf = int_tofp(1);
933 		limits.max_policy_pct = 100;
934 		limits.max_perf_pct = 100;
935 		limits.max_perf = int_tofp(1);
936 		limits.no_turbo = 0;
937 		return 0;
938 	}
939 
940 	limits.min_policy_pct = (policy->min * 100) / policy->cpuinfo.max_freq;
941 	limits.min_policy_pct = clamp_t(int, limits.min_policy_pct, 0 , 100);
942 	limits.min_perf_pct = max(limits.min_policy_pct, limits.min_sysfs_pct);
943 	limits.min_perf = div_fp(int_tofp(limits.min_perf_pct), int_tofp(100));
944 
945 	limits.max_policy_pct = (policy->max * 100) / policy->cpuinfo.max_freq;
946 	limits.max_policy_pct = clamp_t(int, limits.max_policy_pct, 0 , 100);
947 	limits.max_perf_pct = min(limits.max_policy_pct, limits.max_sysfs_pct);
948 	limits.max_perf = div_fp(int_tofp(limits.max_perf_pct), int_tofp(100));
949 
950 	if (hwp_active)
951 		intel_pstate_hwp_set();
952 
953 	return 0;
954 }
955 
956 static int intel_pstate_verify_policy(struct cpufreq_policy *policy)
957 {
958 	cpufreq_verify_within_cpu_limits(policy);
959 
960 	if (policy->policy != CPUFREQ_POLICY_POWERSAVE &&
961 	    policy->policy != CPUFREQ_POLICY_PERFORMANCE)
962 		return -EINVAL;
963 
964 	return 0;
965 }
966 
967 static void intel_pstate_stop_cpu(struct cpufreq_policy *policy)
968 {
969 	int cpu_num = policy->cpu;
970 	struct cpudata *cpu = all_cpu_data[cpu_num];
971 
972 	pr_info("intel_pstate CPU %d exiting\n", cpu_num);
973 
974 	del_timer_sync(&all_cpu_data[cpu_num]->timer);
975 	if (hwp_active)
976 		return;
977 
978 	intel_pstate_set_pstate(cpu, cpu->pstate.min_pstate);
979 }
980 
981 static int intel_pstate_cpu_init(struct cpufreq_policy *policy)
982 {
983 	struct cpudata *cpu;
984 	int rc;
985 
986 	rc = intel_pstate_init_cpu(policy->cpu);
987 	if (rc)
988 		return rc;
989 
990 	cpu = all_cpu_data[policy->cpu];
991 
992 	if (limits.min_perf_pct == 100 && limits.max_perf_pct == 100)
993 		policy->policy = CPUFREQ_POLICY_PERFORMANCE;
994 	else
995 		policy->policy = CPUFREQ_POLICY_POWERSAVE;
996 
997 	policy->min = cpu->pstate.min_pstate * cpu->pstate.scaling;
998 	policy->max = cpu->pstate.turbo_pstate * cpu->pstate.scaling;
999 
1000 	/* cpuinfo and default policy values */
1001 	policy->cpuinfo.min_freq = cpu->pstate.min_pstate * cpu->pstate.scaling;
1002 	policy->cpuinfo.max_freq =
1003 		cpu->pstate.turbo_pstate * cpu->pstate.scaling;
1004 	policy->cpuinfo.transition_latency = CPUFREQ_ETERNAL;
1005 	cpumask_set_cpu(policy->cpu, policy->cpus);
1006 
1007 	return 0;
1008 }
1009 
1010 static struct cpufreq_driver intel_pstate_driver = {
1011 	.flags		= CPUFREQ_CONST_LOOPS,
1012 	.verify		= intel_pstate_verify_policy,
1013 	.setpolicy	= intel_pstate_set_policy,
1014 	.get		= intel_pstate_get,
1015 	.init		= intel_pstate_cpu_init,
1016 	.stop_cpu	= intel_pstate_stop_cpu,
1017 	.name		= "intel_pstate",
1018 };
1019 
1020 static int __initdata no_load;
1021 static int __initdata no_hwp;
1022 static int __initdata hwp_only;
1023 static unsigned int force_load;
1024 
1025 static int intel_pstate_msrs_not_valid(void)
1026 {
1027 	/* Check that all the msr's we are using are valid. */
1028 	u64 aperf, mperf, tmp;
1029 
1030 	rdmsrl(MSR_IA32_APERF, aperf);
1031 	rdmsrl(MSR_IA32_MPERF, mperf);
1032 
1033 	if (!pstate_funcs.get_max() ||
1034 	    !pstate_funcs.get_min() ||
1035 	    !pstate_funcs.get_turbo())
1036 		return -ENODEV;
1037 
1038 	rdmsrl(MSR_IA32_APERF, tmp);
1039 	if (!(tmp - aperf))
1040 		return -ENODEV;
1041 
1042 	rdmsrl(MSR_IA32_MPERF, tmp);
1043 	if (!(tmp - mperf))
1044 		return -ENODEV;
1045 
1046 	return 0;
1047 }
1048 
1049 static void copy_pid_params(struct pstate_adjust_policy *policy)
1050 {
1051 	pid_params.sample_rate_ms = policy->sample_rate_ms;
1052 	pid_params.p_gain_pct = policy->p_gain_pct;
1053 	pid_params.i_gain_pct = policy->i_gain_pct;
1054 	pid_params.d_gain_pct = policy->d_gain_pct;
1055 	pid_params.deadband = policy->deadband;
1056 	pid_params.setpoint = policy->setpoint;
1057 }
1058 
1059 static void copy_cpu_funcs(struct pstate_funcs *funcs)
1060 {
1061 	pstate_funcs.get_max   = funcs->get_max;
1062 	pstate_funcs.get_min   = funcs->get_min;
1063 	pstate_funcs.get_turbo = funcs->get_turbo;
1064 	pstate_funcs.get_scaling = funcs->get_scaling;
1065 	pstate_funcs.set       = funcs->set;
1066 	pstate_funcs.get_vid   = funcs->get_vid;
1067 }
1068 
1069 #if IS_ENABLED(CONFIG_ACPI)
1070 #include <acpi/processor.h>
1071 
1072 static bool intel_pstate_no_acpi_pss(void)
1073 {
1074 	int i;
1075 
1076 	for_each_possible_cpu(i) {
1077 		acpi_status status;
1078 		union acpi_object *pss;
1079 		struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
1080 		struct acpi_processor *pr = per_cpu(processors, i);
1081 
1082 		if (!pr)
1083 			continue;
1084 
1085 		status = acpi_evaluate_object(pr->handle, "_PSS", NULL, &buffer);
1086 		if (ACPI_FAILURE(status))
1087 			continue;
1088 
1089 		pss = buffer.pointer;
1090 		if (pss && pss->type == ACPI_TYPE_PACKAGE) {
1091 			kfree(pss);
1092 			return false;
1093 		}
1094 
1095 		kfree(pss);
1096 	}
1097 
1098 	return true;
1099 }
1100 
1101 static bool intel_pstate_has_acpi_ppc(void)
1102 {
1103 	int i;
1104 
1105 	for_each_possible_cpu(i) {
1106 		struct acpi_processor *pr = per_cpu(processors, i);
1107 
1108 		if (!pr)
1109 			continue;
1110 		if (acpi_has_method(pr->handle, "_PPC"))
1111 			return true;
1112 	}
1113 	return false;
1114 }
1115 
1116 enum {
1117 	PSS,
1118 	PPC,
1119 };
1120 
1121 struct hw_vendor_info {
1122 	u16  valid;
1123 	char oem_id[ACPI_OEM_ID_SIZE];
1124 	char oem_table_id[ACPI_OEM_TABLE_ID_SIZE];
1125 	int  oem_pwr_table;
1126 };
1127 
1128 /* Hardware vendor-specific info that has its own power management modes */
1129 static struct hw_vendor_info vendor_info[] = {
1130 	{1, "HP    ", "ProLiant", PSS},
1131 	{1, "ORACLE", "X4-2    ", PPC},
1132 	{1, "ORACLE", "X4-2L   ", PPC},
1133 	{1, "ORACLE", "X4-2B   ", PPC},
1134 	{1, "ORACLE", "X3-2    ", PPC},
1135 	{1, "ORACLE", "X3-2L   ", PPC},
1136 	{1, "ORACLE", "X3-2B   ", PPC},
1137 	{1, "ORACLE", "X4470M2 ", PPC},
1138 	{1, "ORACLE", "X4270M3 ", PPC},
1139 	{1, "ORACLE", "X4270M2 ", PPC},
1140 	{1, "ORACLE", "X4170M2 ", PPC},
1141 	{0, "", ""},
1142 };
1143 
1144 static bool intel_pstate_platform_pwr_mgmt_exists(void)
1145 {
1146 	struct acpi_table_header hdr;
1147 	struct hw_vendor_info *v_info;
1148 	const struct x86_cpu_id *id;
1149 	u64 misc_pwr;
1150 
1151 	id = x86_match_cpu(intel_pstate_cpu_oob_ids);
1152 	if (id) {
1153 		rdmsrl(MSR_MISC_PWR_MGMT, misc_pwr);
1154 		if ( misc_pwr & (1 << 8))
1155 			return true;
1156 	}
1157 
1158 	if (acpi_disabled ||
1159 	    ACPI_FAILURE(acpi_get_table_header(ACPI_SIG_FADT, 0, &hdr)))
1160 		return false;
1161 
1162 	for (v_info = vendor_info; v_info->valid; v_info++) {
1163 		if (!strncmp(hdr.oem_id, v_info->oem_id, ACPI_OEM_ID_SIZE) &&
1164 			!strncmp(hdr.oem_table_id, v_info->oem_table_id,
1165 						ACPI_OEM_TABLE_ID_SIZE))
1166 			switch (v_info->oem_pwr_table) {
1167 			case PSS:
1168 				return intel_pstate_no_acpi_pss();
1169 			case PPC:
1170 				return intel_pstate_has_acpi_ppc() &&
1171 					(!force_load);
1172 			}
1173 	}
1174 
1175 	return false;
1176 }
1177 #else /* CONFIG_ACPI not enabled */
1178 static inline bool intel_pstate_platform_pwr_mgmt_exists(void) { return false; }
1179 static inline bool intel_pstate_has_acpi_ppc(void) { return false; }
1180 #endif /* CONFIG_ACPI */
1181 
1182 static int __init intel_pstate_init(void)
1183 {
1184 	int cpu, rc = 0;
1185 	const struct x86_cpu_id *id;
1186 	struct cpu_defaults *cpu_info;
1187 	struct cpuinfo_x86 *c = &boot_cpu_data;
1188 
1189 	if (no_load)
1190 		return -ENODEV;
1191 
1192 	id = x86_match_cpu(intel_pstate_cpu_ids);
1193 	if (!id)
1194 		return -ENODEV;
1195 
1196 	/*
1197 	 * The Intel pstate driver will be ignored if the platform
1198 	 * firmware has its own power management modes.
1199 	 */
1200 	if (intel_pstate_platform_pwr_mgmt_exists())
1201 		return -ENODEV;
1202 
1203 	cpu_info = (struct cpu_defaults *)id->driver_data;
1204 
1205 	copy_pid_params(&cpu_info->pid_policy);
1206 	copy_cpu_funcs(&cpu_info->funcs);
1207 
1208 	if (intel_pstate_msrs_not_valid())
1209 		return -ENODEV;
1210 
1211 	pr_info("Intel P-state driver initializing.\n");
1212 
1213 	all_cpu_data = vzalloc(sizeof(void *) * num_possible_cpus());
1214 	if (!all_cpu_data)
1215 		return -ENOMEM;
1216 
1217 	if (cpu_has(c,X86_FEATURE_HWP) && !no_hwp)
1218 		intel_pstate_hwp_enable();
1219 
1220 	if (!hwp_active && hwp_only)
1221 		goto out;
1222 
1223 	rc = cpufreq_register_driver(&intel_pstate_driver);
1224 	if (rc)
1225 		goto out;
1226 
1227 	intel_pstate_debug_expose_params();
1228 	intel_pstate_sysfs_expose_params();
1229 
1230 	return rc;
1231 out:
1232 	get_online_cpus();
1233 	for_each_online_cpu(cpu) {
1234 		if (all_cpu_data[cpu]) {
1235 			del_timer_sync(&all_cpu_data[cpu]->timer);
1236 			kfree(all_cpu_data[cpu]);
1237 		}
1238 	}
1239 
1240 	put_online_cpus();
1241 	vfree(all_cpu_data);
1242 	return -ENODEV;
1243 }
1244 device_initcall(intel_pstate_init);
1245 
1246 static int __init intel_pstate_setup(char *str)
1247 {
1248 	if (!str)
1249 		return -EINVAL;
1250 
1251 	if (!strcmp(str, "disable"))
1252 		no_load = 1;
1253 	if (!strcmp(str, "no_hwp"))
1254 		no_hwp = 1;
1255 	if (!strcmp(str, "force"))
1256 		force_load = 1;
1257 	if (!strcmp(str, "hwp_only"))
1258 		hwp_only = 1;
1259 	return 0;
1260 }
1261 early_param("intel_pstate", intel_pstate_setup);
1262 
1263 MODULE_AUTHOR("Dirk Brandewie <dirk.j.brandewie@intel.com>");
1264 MODULE_DESCRIPTION("'intel_pstate' - P state driver Intel Core processors");
1265 MODULE_LICENSE("GPL");
1266