xref: /linux/drivers/cpufreq/intel_pstate.c (revision 6f7e0d9d08c7fa3847b1380b16dff7e14523933c)
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 <linux/vmalloc.h>
30 #include <trace/events/power.h>
31 
32 #include <asm/div64.h>
33 #include <asm/msr.h>
34 #include <asm/cpu_device_id.h>
35 #include <asm/cpufeature.h>
36 
37 #define ATOM_RATIOS		0x66a
38 #define ATOM_VIDS		0x66b
39 #define ATOM_TURBO_RATIOS	0x66c
40 #define ATOM_TURBO_VIDS		0x66d
41 
42 #define FRAC_BITS 8
43 #define int_tofp(X) ((int64_t)(X) << FRAC_BITS)
44 #define fp_toint(X) ((X) >> FRAC_BITS)
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(s64 x, s64 y)
52 {
53 	return div64_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 /**
68  * struct sample -	Store performance sample
69  * @core_pct_busy:	Ratio of APERF/MPERF in percent, which is actual
70  *			performance during last sample period
71  * @busy_scaled:	Scaled busy value which is used to calculate next
72  *			P state. This can be different than core_pct_busy
73  *			to account for cpu idle period
74  * @aperf:		Difference of actual performance frequency clock count
75  *			read from APERF MSR between last and current sample
76  * @mperf:		Difference of maximum performance frequency clock count
77  *			read from MPERF MSR between last and current sample
78  * @tsc:		Difference of time stamp counter between last and
79  *			current sample
80  * @freq:		Effective frequency calculated from APERF/MPERF
81  * @time:		Current time from scheduler
82  *
83  * This structure is used in the cpudata structure to store performance sample
84  * data for choosing next P State.
85  */
86 struct sample {
87 	int32_t core_pct_busy;
88 	int32_t busy_scaled;
89 	u64 aperf;
90 	u64 mperf;
91 	u64 tsc;
92 	int freq;
93 	u64 time;
94 };
95 
96 /**
97  * struct pstate_data - Store P state data
98  * @current_pstate:	Current requested P state
99  * @min_pstate:		Min P state possible for this platform
100  * @max_pstate:		Max P state possible for this platform
101  * @max_pstate_physical:This is physical Max P state for a processor
102  *			This can be higher than the max_pstate which can
103  *			be limited by platform thermal design power limits
104  * @scaling:		Scaling factor to  convert frequency to cpufreq
105  *			frequency units
106  * @turbo_pstate:	Max Turbo P state possible for this platform
107  *
108  * Stores the per cpu model P state limits and current P state.
109  */
110 struct pstate_data {
111 	int	current_pstate;
112 	int	min_pstate;
113 	int	max_pstate;
114 	int	max_pstate_physical;
115 	int	scaling;
116 	int	turbo_pstate;
117 };
118 
119 /**
120  * struct vid_data -	Stores voltage information data
121  * @min:		VID data for this platform corresponding to
122  *			the lowest P state
123  * @max:		VID data corresponding to the highest P State.
124  * @turbo:		VID data for turbo P state
125  * @ratio:		Ratio of (vid max - vid min) /
126  *			(max P state - Min P State)
127  *
128  * Stores the voltage data for DVFS (Dynamic Voltage and Frequency Scaling)
129  * This data is used in Atom platforms, where in addition to target P state,
130  * the voltage data needs to be specified to select next P State.
131  */
132 struct vid_data {
133 	int min;
134 	int max;
135 	int turbo;
136 	int32_t ratio;
137 };
138 
139 /**
140  * struct _pid -	Stores PID data
141  * @setpoint:		Target set point for busyness or performance
142  * @integral:		Storage for accumulated error values
143  * @p_gain:		PID proportional gain
144  * @i_gain:		PID integral gain
145  * @d_gain:		PID derivative gain
146  * @deadband:		PID deadband
147  * @last_err:		Last error storage for integral part of PID calculation
148  *
149  * Stores PID coefficients and last error for PID controller.
150  */
151 struct _pid {
152 	int setpoint;
153 	int32_t integral;
154 	int32_t p_gain;
155 	int32_t i_gain;
156 	int32_t d_gain;
157 	int deadband;
158 	int32_t last_err;
159 };
160 
161 /**
162  * struct cpudata -	Per CPU instance data storage
163  * @cpu:		CPU number for this instance data
164  * @update_util:	CPUFreq utility callback information
165  * @pstate:		Stores P state limits for this CPU
166  * @vid:		Stores VID limits for this CPU
167  * @pid:		Stores PID parameters for this CPU
168  * @last_sample_time:	Last Sample time
169  * @prev_aperf:		Last APERF value read from APERF MSR
170  * @prev_mperf:		Last MPERF value read from MPERF MSR
171  * @prev_tsc:		Last timestamp counter (TSC) value
172  * @prev_cummulative_iowait: IO Wait time difference from last and
173  *			current sample
174  * @sample:		Storage for storing last Sample data
175  *
176  * This structure stores per CPU instance data for all CPUs.
177  */
178 struct cpudata {
179 	int cpu;
180 
181 	struct update_util_data update_util;
182 
183 	struct pstate_data pstate;
184 	struct vid_data vid;
185 	struct _pid pid;
186 
187 	u64	last_sample_time;
188 	u64	prev_aperf;
189 	u64	prev_mperf;
190 	u64	prev_tsc;
191 	u64	prev_cummulative_iowait;
192 	struct sample sample;
193 };
194 
195 static struct cpudata **all_cpu_data;
196 
197 /**
198  * struct pid_adjust_policy - Stores static PID configuration data
199  * @sample_rate_ms:	PID calculation sample rate in ms
200  * @sample_rate_ns:	Sample rate calculation in ns
201  * @deadband:		PID deadband
202  * @setpoint:		PID Setpoint
203  * @p_gain_pct:		PID proportional gain
204  * @i_gain_pct:		PID integral gain
205  * @d_gain_pct:		PID derivative gain
206  *
207  * Stores per CPU model static PID configuration data.
208  */
209 struct pstate_adjust_policy {
210 	int sample_rate_ms;
211 	s64 sample_rate_ns;
212 	int deadband;
213 	int setpoint;
214 	int p_gain_pct;
215 	int d_gain_pct;
216 	int i_gain_pct;
217 };
218 
219 /**
220  * struct pstate_funcs - Per CPU model specific callbacks
221  * @get_max:		Callback to get maximum non turbo effective P state
222  * @get_max_physical:	Callback to get maximum non turbo physical P state
223  * @get_min:		Callback to get minimum P state
224  * @get_turbo:		Callback to get turbo P state
225  * @get_scaling:	Callback to get frequency scaling factor
226  * @get_val:		Callback to convert P state to actual MSR write value
227  * @get_vid:		Callback to get VID data for Atom platforms
228  * @get_target_pstate:	Callback to a function to calculate next P state to use
229  *
230  * Core and Atom CPU models have different way to get P State limits. This
231  * structure is used to store those callbacks.
232  */
233 struct pstate_funcs {
234 	int (*get_max)(void);
235 	int (*get_max_physical)(void);
236 	int (*get_min)(void);
237 	int (*get_turbo)(void);
238 	int (*get_scaling)(void);
239 	u64 (*get_val)(struct cpudata*, int pstate);
240 	void (*get_vid)(struct cpudata *);
241 	int32_t (*get_target_pstate)(struct cpudata *);
242 };
243 
244 /**
245  * struct cpu_defaults- Per CPU model default config data
246  * @pid_policy:	PID config data
247  * @funcs:		Callback function data
248  */
249 struct cpu_defaults {
250 	struct pstate_adjust_policy pid_policy;
251 	struct pstate_funcs funcs;
252 };
253 
254 static inline int32_t get_target_pstate_use_performance(struct cpudata *cpu);
255 static inline int32_t get_target_pstate_use_cpu_load(struct cpudata *cpu);
256 
257 static struct pstate_adjust_policy pid_params;
258 static struct pstate_funcs pstate_funcs;
259 static int hwp_active;
260 
261 
262 /**
263  * struct perf_limits - Store user and policy limits
264  * @no_turbo:		User requested turbo state from intel_pstate sysfs
265  * @turbo_disabled:	Platform turbo status either from msr
266  *			MSR_IA32_MISC_ENABLE or when maximum available pstate
267  *			matches the maximum turbo pstate
268  * @max_perf_pct:	Effective maximum performance limit in percentage, this
269  *			is minimum of either limits enforced by cpufreq policy
270  *			or limits from user set limits via intel_pstate sysfs
271  * @min_perf_pct:	Effective minimum performance limit in percentage, this
272  *			is maximum of either limits enforced by cpufreq policy
273  *			or limits from user set limits via intel_pstate sysfs
274  * @max_perf:		This is a scaled value between 0 to 255 for max_perf_pct
275  *			This value is used to limit max pstate
276  * @min_perf:		This is a scaled value between 0 to 255 for min_perf_pct
277  *			This value is used to limit min pstate
278  * @max_policy_pct:	The maximum performance in percentage enforced by
279  *			cpufreq setpolicy interface
280  * @max_sysfs_pct:	The maximum performance in percentage enforced by
281  *			intel pstate sysfs interface
282  * @min_policy_pct:	The minimum performance in percentage enforced by
283  *			cpufreq setpolicy interface
284  * @min_sysfs_pct:	The minimum performance in percentage enforced by
285  *			intel pstate sysfs interface
286  *
287  * Storage for user and policy defined limits.
288  */
289 struct perf_limits {
290 	int no_turbo;
291 	int turbo_disabled;
292 	int max_perf_pct;
293 	int min_perf_pct;
294 	int32_t max_perf;
295 	int32_t min_perf;
296 	int max_policy_pct;
297 	int max_sysfs_pct;
298 	int min_policy_pct;
299 	int min_sysfs_pct;
300 };
301 
302 static struct perf_limits performance_limits = {
303 	.no_turbo = 0,
304 	.turbo_disabled = 0,
305 	.max_perf_pct = 100,
306 	.max_perf = int_tofp(1),
307 	.min_perf_pct = 100,
308 	.min_perf = int_tofp(1),
309 	.max_policy_pct = 100,
310 	.max_sysfs_pct = 100,
311 	.min_policy_pct = 0,
312 	.min_sysfs_pct = 0,
313 };
314 
315 static struct perf_limits powersave_limits = {
316 	.no_turbo = 0,
317 	.turbo_disabled = 0,
318 	.max_perf_pct = 100,
319 	.max_perf = int_tofp(1),
320 	.min_perf_pct = 0,
321 	.min_perf = 0,
322 	.max_policy_pct = 100,
323 	.max_sysfs_pct = 100,
324 	.min_policy_pct = 0,
325 	.min_sysfs_pct = 0,
326 };
327 
328 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_PERFORMANCE
329 static struct perf_limits *limits = &performance_limits;
330 #else
331 static struct perf_limits *limits = &powersave_limits;
332 #endif
333 
334 static inline void pid_reset(struct _pid *pid, int setpoint, int busy,
335 			     int deadband, int integral) {
336 	pid->setpoint = int_tofp(setpoint);
337 	pid->deadband  = int_tofp(deadband);
338 	pid->integral  = int_tofp(integral);
339 	pid->last_err  = int_tofp(setpoint) - int_tofp(busy);
340 }
341 
342 static inline void pid_p_gain_set(struct _pid *pid, int percent)
343 {
344 	pid->p_gain = div_fp(int_tofp(percent), int_tofp(100));
345 }
346 
347 static inline void pid_i_gain_set(struct _pid *pid, int percent)
348 {
349 	pid->i_gain = div_fp(int_tofp(percent), int_tofp(100));
350 }
351 
352 static inline void pid_d_gain_set(struct _pid *pid, int percent)
353 {
354 	pid->d_gain = div_fp(int_tofp(percent), int_tofp(100));
355 }
356 
357 static signed int pid_calc(struct _pid *pid, int32_t busy)
358 {
359 	signed int result;
360 	int32_t pterm, dterm, fp_error;
361 	int32_t integral_limit;
362 
363 	fp_error = pid->setpoint - busy;
364 
365 	if (abs(fp_error) <= pid->deadband)
366 		return 0;
367 
368 	pterm = mul_fp(pid->p_gain, fp_error);
369 
370 	pid->integral += fp_error;
371 
372 	/*
373 	 * We limit the integral here so that it will never
374 	 * get higher than 30.  This prevents it from becoming
375 	 * too large an input over long periods of time and allows
376 	 * it to get factored out sooner.
377 	 *
378 	 * The value of 30 was chosen through experimentation.
379 	 */
380 	integral_limit = int_tofp(30);
381 	if (pid->integral > integral_limit)
382 		pid->integral = integral_limit;
383 	if (pid->integral < -integral_limit)
384 		pid->integral = -integral_limit;
385 
386 	dterm = mul_fp(pid->d_gain, fp_error - pid->last_err);
387 	pid->last_err = fp_error;
388 
389 	result = pterm + mul_fp(pid->integral, pid->i_gain) + dterm;
390 	result = result + (1 << (FRAC_BITS-1));
391 	return (signed int)fp_toint(result);
392 }
393 
394 static inline void intel_pstate_busy_pid_reset(struct cpudata *cpu)
395 {
396 	pid_p_gain_set(&cpu->pid, pid_params.p_gain_pct);
397 	pid_d_gain_set(&cpu->pid, pid_params.d_gain_pct);
398 	pid_i_gain_set(&cpu->pid, pid_params.i_gain_pct);
399 
400 	pid_reset(&cpu->pid, pid_params.setpoint, 100, pid_params.deadband, 0);
401 }
402 
403 static inline void intel_pstate_reset_all_pid(void)
404 {
405 	unsigned int cpu;
406 
407 	for_each_online_cpu(cpu) {
408 		if (all_cpu_data[cpu])
409 			intel_pstate_busy_pid_reset(all_cpu_data[cpu]);
410 	}
411 }
412 
413 static inline void update_turbo_state(void)
414 {
415 	u64 misc_en;
416 	struct cpudata *cpu;
417 
418 	cpu = all_cpu_data[0];
419 	rdmsrl(MSR_IA32_MISC_ENABLE, misc_en);
420 	limits->turbo_disabled =
421 		(misc_en & MSR_IA32_MISC_ENABLE_TURBO_DISABLE ||
422 		 cpu->pstate.max_pstate == cpu->pstate.turbo_pstate);
423 }
424 
425 static void intel_pstate_hwp_set(const struct cpumask *cpumask)
426 {
427 	int min, hw_min, max, hw_max, cpu, range, adj_range;
428 	u64 value, cap;
429 
430 	rdmsrl(MSR_HWP_CAPABILITIES, cap);
431 	hw_min = HWP_LOWEST_PERF(cap);
432 	hw_max = HWP_HIGHEST_PERF(cap);
433 	range = hw_max - hw_min;
434 
435 	for_each_cpu(cpu, cpumask) {
436 		rdmsrl_on_cpu(cpu, MSR_HWP_REQUEST, &value);
437 		adj_range = limits->min_perf_pct * range / 100;
438 		min = hw_min + adj_range;
439 		value &= ~HWP_MIN_PERF(~0L);
440 		value |= HWP_MIN_PERF(min);
441 
442 		adj_range = limits->max_perf_pct * range / 100;
443 		max = hw_min + adj_range;
444 		if (limits->no_turbo) {
445 			hw_max = HWP_GUARANTEED_PERF(cap);
446 			if (hw_max < max)
447 				max = hw_max;
448 		}
449 
450 		value &= ~HWP_MAX_PERF(~0L);
451 		value |= HWP_MAX_PERF(max);
452 		wrmsrl_on_cpu(cpu, MSR_HWP_REQUEST, value);
453 	}
454 }
455 
456 static void intel_pstate_hwp_set_online_cpus(void)
457 {
458 	get_online_cpus();
459 	intel_pstate_hwp_set(cpu_online_mask);
460 	put_online_cpus();
461 }
462 
463 /************************** debugfs begin ************************/
464 static int pid_param_set(void *data, u64 val)
465 {
466 	*(u32 *)data = val;
467 	intel_pstate_reset_all_pid();
468 	return 0;
469 }
470 
471 static int pid_param_get(void *data, u64 *val)
472 {
473 	*val = *(u32 *)data;
474 	return 0;
475 }
476 DEFINE_SIMPLE_ATTRIBUTE(fops_pid_param, pid_param_get, pid_param_set, "%llu\n");
477 
478 struct pid_param {
479 	char *name;
480 	void *value;
481 };
482 
483 static struct pid_param pid_files[] = {
484 	{"sample_rate_ms", &pid_params.sample_rate_ms},
485 	{"d_gain_pct", &pid_params.d_gain_pct},
486 	{"i_gain_pct", &pid_params.i_gain_pct},
487 	{"deadband", &pid_params.deadband},
488 	{"setpoint", &pid_params.setpoint},
489 	{"p_gain_pct", &pid_params.p_gain_pct},
490 	{NULL, NULL}
491 };
492 
493 static void __init intel_pstate_debug_expose_params(void)
494 {
495 	struct dentry *debugfs_parent;
496 	int i = 0;
497 
498 	if (hwp_active)
499 		return;
500 	debugfs_parent = debugfs_create_dir("pstate_snb", NULL);
501 	if (IS_ERR_OR_NULL(debugfs_parent))
502 		return;
503 	while (pid_files[i].name) {
504 		debugfs_create_file(pid_files[i].name, 0660,
505 				    debugfs_parent, pid_files[i].value,
506 				    &fops_pid_param);
507 		i++;
508 	}
509 }
510 
511 /************************** debugfs end ************************/
512 
513 /************************** sysfs begin ************************/
514 #define show_one(file_name, object)					\
515 	static ssize_t show_##file_name					\
516 	(struct kobject *kobj, struct attribute *attr, char *buf)	\
517 	{								\
518 		return sprintf(buf, "%u\n", limits->object);		\
519 	}
520 
521 static ssize_t show_turbo_pct(struct kobject *kobj,
522 				struct attribute *attr, char *buf)
523 {
524 	struct cpudata *cpu;
525 	int total, no_turbo, turbo_pct;
526 	uint32_t turbo_fp;
527 
528 	cpu = all_cpu_data[0];
529 
530 	total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
531 	no_turbo = cpu->pstate.max_pstate - cpu->pstate.min_pstate + 1;
532 	turbo_fp = div_fp(int_tofp(no_turbo), int_tofp(total));
533 	turbo_pct = 100 - fp_toint(mul_fp(turbo_fp, int_tofp(100)));
534 	return sprintf(buf, "%u\n", turbo_pct);
535 }
536 
537 static ssize_t show_num_pstates(struct kobject *kobj,
538 				struct attribute *attr, char *buf)
539 {
540 	struct cpudata *cpu;
541 	int total;
542 
543 	cpu = all_cpu_data[0];
544 	total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
545 	return sprintf(buf, "%u\n", total);
546 }
547 
548 static ssize_t show_no_turbo(struct kobject *kobj,
549 			     struct attribute *attr, char *buf)
550 {
551 	ssize_t ret;
552 
553 	update_turbo_state();
554 	if (limits->turbo_disabled)
555 		ret = sprintf(buf, "%u\n", limits->turbo_disabled);
556 	else
557 		ret = sprintf(buf, "%u\n", limits->no_turbo);
558 
559 	return ret;
560 }
561 
562 static ssize_t store_no_turbo(struct kobject *a, struct attribute *b,
563 			      const char *buf, size_t count)
564 {
565 	unsigned int input;
566 	int ret;
567 
568 	ret = sscanf(buf, "%u", &input);
569 	if (ret != 1)
570 		return -EINVAL;
571 
572 	update_turbo_state();
573 	if (limits->turbo_disabled) {
574 		pr_warn("intel_pstate: Turbo disabled by BIOS or unavailable on processor\n");
575 		return -EPERM;
576 	}
577 
578 	limits->no_turbo = clamp_t(int, input, 0, 1);
579 
580 	if (hwp_active)
581 		intel_pstate_hwp_set_online_cpus();
582 
583 	return count;
584 }
585 
586 static ssize_t store_max_perf_pct(struct kobject *a, struct attribute *b,
587 				  const char *buf, size_t count)
588 {
589 	unsigned int input;
590 	int ret;
591 
592 	ret = sscanf(buf, "%u", &input);
593 	if (ret != 1)
594 		return -EINVAL;
595 
596 	limits->max_sysfs_pct = clamp_t(int, input, 0 , 100);
597 	limits->max_perf_pct = min(limits->max_policy_pct,
598 				   limits->max_sysfs_pct);
599 	limits->max_perf_pct = max(limits->min_policy_pct,
600 				   limits->max_perf_pct);
601 	limits->max_perf_pct = max(limits->min_perf_pct,
602 				   limits->max_perf_pct);
603 	limits->max_perf = div_fp(int_tofp(limits->max_perf_pct),
604 				  int_tofp(100));
605 
606 	if (hwp_active)
607 		intel_pstate_hwp_set_online_cpus();
608 	return count;
609 }
610 
611 static ssize_t store_min_perf_pct(struct kobject *a, struct attribute *b,
612 				  const char *buf, size_t count)
613 {
614 	unsigned int input;
615 	int ret;
616 
617 	ret = sscanf(buf, "%u", &input);
618 	if (ret != 1)
619 		return -EINVAL;
620 
621 	limits->min_sysfs_pct = clamp_t(int, input, 0 , 100);
622 	limits->min_perf_pct = max(limits->min_policy_pct,
623 				   limits->min_sysfs_pct);
624 	limits->min_perf_pct = min(limits->max_policy_pct,
625 				   limits->min_perf_pct);
626 	limits->min_perf_pct = min(limits->max_perf_pct,
627 				   limits->min_perf_pct);
628 	limits->min_perf = div_fp(int_tofp(limits->min_perf_pct),
629 				  int_tofp(100));
630 
631 	if (hwp_active)
632 		intel_pstate_hwp_set_online_cpus();
633 	return count;
634 }
635 
636 show_one(max_perf_pct, max_perf_pct);
637 show_one(min_perf_pct, min_perf_pct);
638 
639 define_one_global_rw(no_turbo);
640 define_one_global_rw(max_perf_pct);
641 define_one_global_rw(min_perf_pct);
642 define_one_global_ro(turbo_pct);
643 define_one_global_ro(num_pstates);
644 
645 static struct attribute *intel_pstate_attributes[] = {
646 	&no_turbo.attr,
647 	&max_perf_pct.attr,
648 	&min_perf_pct.attr,
649 	&turbo_pct.attr,
650 	&num_pstates.attr,
651 	NULL
652 };
653 
654 static struct attribute_group intel_pstate_attr_group = {
655 	.attrs = intel_pstate_attributes,
656 };
657 
658 static void __init intel_pstate_sysfs_expose_params(void)
659 {
660 	struct kobject *intel_pstate_kobject;
661 	int rc;
662 
663 	intel_pstate_kobject = kobject_create_and_add("intel_pstate",
664 						&cpu_subsys.dev_root->kobj);
665 	BUG_ON(!intel_pstate_kobject);
666 	rc = sysfs_create_group(intel_pstate_kobject, &intel_pstate_attr_group);
667 	BUG_ON(rc);
668 }
669 /************************** sysfs end ************************/
670 
671 static void intel_pstate_hwp_enable(struct cpudata *cpudata)
672 {
673 	/* First disable HWP notification interrupt as we don't process them */
674 	wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_INTERRUPT, 0x00);
675 
676 	wrmsrl_on_cpu(cpudata->cpu, MSR_PM_ENABLE, 0x1);
677 }
678 
679 static int atom_get_min_pstate(void)
680 {
681 	u64 value;
682 
683 	rdmsrl(ATOM_RATIOS, value);
684 	return (value >> 8) & 0x7F;
685 }
686 
687 static int atom_get_max_pstate(void)
688 {
689 	u64 value;
690 
691 	rdmsrl(ATOM_RATIOS, value);
692 	return (value >> 16) & 0x7F;
693 }
694 
695 static int atom_get_turbo_pstate(void)
696 {
697 	u64 value;
698 
699 	rdmsrl(ATOM_TURBO_RATIOS, value);
700 	return value & 0x7F;
701 }
702 
703 static u64 atom_get_val(struct cpudata *cpudata, int pstate)
704 {
705 	u64 val;
706 	int32_t vid_fp;
707 	u32 vid;
708 
709 	val = (u64)pstate << 8;
710 	if (limits->no_turbo && !limits->turbo_disabled)
711 		val |= (u64)1 << 32;
712 
713 	vid_fp = cpudata->vid.min + mul_fp(
714 		int_tofp(pstate - cpudata->pstate.min_pstate),
715 		cpudata->vid.ratio);
716 
717 	vid_fp = clamp_t(int32_t, vid_fp, cpudata->vid.min, cpudata->vid.max);
718 	vid = ceiling_fp(vid_fp);
719 
720 	if (pstate > cpudata->pstate.max_pstate)
721 		vid = cpudata->vid.turbo;
722 
723 	return val | vid;
724 }
725 
726 static int silvermont_get_scaling(void)
727 {
728 	u64 value;
729 	int i;
730 	/* Defined in Table 35-6 from SDM (Sept 2015) */
731 	static int silvermont_freq_table[] = {
732 		83300, 100000, 133300, 116700, 80000};
733 
734 	rdmsrl(MSR_FSB_FREQ, value);
735 	i = value & 0x7;
736 	WARN_ON(i > 4);
737 
738 	return silvermont_freq_table[i];
739 }
740 
741 static int airmont_get_scaling(void)
742 {
743 	u64 value;
744 	int i;
745 	/* Defined in Table 35-10 from SDM (Sept 2015) */
746 	static int airmont_freq_table[] = {
747 		83300, 100000, 133300, 116700, 80000,
748 		93300, 90000, 88900, 87500};
749 
750 	rdmsrl(MSR_FSB_FREQ, value);
751 	i = value & 0xF;
752 	WARN_ON(i > 8);
753 
754 	return airmont_freq_table[i];
755 }
756 
757 static void atom_get_vid(struct cpudata *cpudata)
758 {
759 	u64 value;
760 
761 	rdmsrl(ATOM_VIDS, value);
762 	cpudata->vid.min = int_tofp((value >> 8) & 0x7f);
763 	cpudata->vid.max = int_tofp((value >> 16) & 0x7f);
764 	cpudata->vid.ratio = div_fp(
765 		cpudata->vid.max - cpudata->vid.min,
766 		int_tofp(cpudata->pstate.max_pstate -
767 			cpudata->pstate.min_pstate));
768 
769 	rdmsrl(ATOM_TURBO_VIDS, value);
770 	cpudata->vid.turbo = value & 0x7f;
771 }
772 
773 static int core_get_min_pstate(void)
774 {
775 	u64 value;
776 
777 	rdmsrl(MSR_PLATFORM_INFO, value);
778 	return (value >> 40) & 0xFF;
779 }
780 
781 static int core_get_max_pstate_physical(void)
782 {
783 	u64 value;
784 
785 	rdmsrl(MSR_PLATFORM_INFO, value);
786 	return (value >> 8) & 0xFF;
787 }
788 
789 static int core_get_max_pstate(void)
790 {
791 	u64 tar;
792 	u64 plat_info;
793 	int max_pstate;
794 	int err;
795 
796 	rdmsrl(MSR_PLATFORM_INFO, plat_info);
797 	max_pstate = (plat_info >> 8) & 0xFF;
798 
799 	err = rdmsrl_safe(MSR_TURBO_ACTIVATION_RATIO, &tar);
800 	if (!err) {
801 		/* Do some sanity checking for safety */
802 		if (plat_info & 0x600000000) {
803 			u64 tdp_ctrl;
804 			u64 tdp_ratio;
805 			int tdp_msr;
806 
807 			err = rdmsrl_safe(MSR_CONFIG_TDP_CONTROL, &tdp_ctrl);
808 			if (err)
809 				goto skip_tar;
810 
811 			tdp_msr = MSR_CONFIG_TDP_NOMINAL + tdp_ctrl;
812 			err = rdmsrl_safe(tdp_msr, &tdp_ratio);
813 			if (err)
814 				goto skip_tar;
815 
816 			/* For level 1 and 2, bits[23:16] contain the ratio */
817 			if (tdp_ctrl)
818 				tdp_ratio >>= 16;
819 
820 			tdp_ratio &= 0xff; /* ratios are only 8 bits long */
821 			if (tdp_ratio - 1 == tar) {
822 				max_pstate = tar;
823 				pr_debug("max_pstate=TAC %x\n", max_pstate);
824 			} else {
825 				goto skip_tar;
826 			}
827 		}
828 	}
829 
830 skip_tar:
831 	return max_pstate;
832 }
833 
834 static int core_get_turbo_pstate(void)
835 {
836 	u64 value;
837 	int nont, ret;
838 
839 	rdmsrl(MSR_NHM_TURBO_RATIO_LIMIT, value);
840 	nont = core_get_max_pstate();
841 	ret = (value) & 255;
842 	if (ret <= nont)
843 		ret = nont;
844 	return ret;
845 }
846 
847 static inline int core_get_scaling(void)
848 {
849 	return 100000;
850 }
851 
852 static u64 core_get_val(struct cpudata *cpudata, int pstate)
853 {
854 	u64 val;
855 
856 	val = (u64)pstate << 8;
857 	if (limits->no_turbo && !limits->turbo_disabled)
858 		val |= (u64)1 << 32;
859 
860 	return val;
861 }
862 
863 static int knl_get_turbo_pstate(void)
864 {
865 	u64 value;
866 	int nont, ret;
867 
868 	rdmsrl(MSR_NHM_TURBO_RATIO_LIMIT, value);
869 	nont = core_get_max_pstate();
870 	ret = (((value) >> 8) & 0xFF);
871 	if (ret <= nont)
872 		ret = nont;
873 	return ret;
874 }
875 
876 static struct cpu_defaults core_params = {
877 	.pid_policy = {
878 		.sample_rate_ms = 10,
879 		.deadband = 0,
880 		.setpoint = 97,
881 		.p_gain_pct = 20,
882 		.d_gain_pct = 0,
883 		.i_gain_pct = 0,
884 	},
885 	.funcs = {
886 		.get_max = core_get_max_pstate,
887 		.get_max_physical = core_get_max_pstate_physical,
888 		.get_min = core_get_min_pstate,
889 		.get_turbo = core_get_turbo_pstate,
890 		.get_scaling = core_get_scaling,
891 		.get_val = core_get_val,
892 		.get_target_pstate = get_target_pstate_use_performance,
893 	},
894 };
895 
896 static struct cpu_defaults silvermont_params = {
897 	.pid_policy = {
898 		.sample_rate_ms = 10,
899 		.deadband = 0,
900 		.setpoint = 60,
901 		.p_gain_pct = 14,
902 		.d_gain_pct = 0,
903 		.i_gain_pct = 4,
904 	},
905 	.funcs = {
906 		.get_max = atom_get_max_pstate,
907 		.get_max_physical = atom_get_max_pstate,
908 		.get_min = atom_get_min_pstate,
909 		.get_turbo = atom_get_turbo_pstate,
910 		.get_val = atom_get_val,
911 		.get_scaling = silvermont_get_scaling,
912 		.get_vid = atom_get_vid,
913 		.get_target_pstate = get_target_pstate_use_cpu_load,
914 	},
915 };
916 
917 static struct cpu_defaults airmont_params = {
918 	.pid_policy = {
919 		.sample_rate_ms = 10,
920 		.deadband = 0,
921 		.setpoint = 60,
922 		.p_gain_pct = 14,
923 		.d_gain_pct = 0,
924 		.i_gain_pct = 4,
925 	},
926 	.funcs = {
927 		.get_max = atom_get_max_pstate,
928 		.get_max_physical = atom_get_max_pstate,
929 		.get_min = atom_get_min_pstate,
930 		.get_turbo = atom_get_turbo_pstate,
931 		.get_val = atom_get_val,
932 		.get_scaling = airmont_get_scaling,
933 		.get_vid = atom_get_vid,
934 		.get_target_pstate = get_target_pstate_use_cpu_load,
935 	},
936 };
937 
938 static struct cpu_defaults knl_params = {
939 	.pid_policy = {
940 		.sample_rate_ms = 10,
941 		.deadband = 0,
942 		.setpoint = 97,
943 		.p_gain_pct = 20,
944 		.d_gain_pct = 0,
945 		.i_gain_pct = 0,
946 	},
947 	.funcs = {
948 		.get_max = core_get_max_pstate,
949 		.get_max_physical = core_get_max_pstate_physical,
950 		.get_min = core_get_min_pstate,
951 		.get_turbo = knl_get_turbo_pstate,
952 		.get_scaling = core_get_scaling,
953 		.get_val = core_get_val,
954 		.get_target_pstate = get_target_pstate_use_performance,
955 	},
956 };
957 
958 static void intel_pstate_get_min_max(struct cpudata *cpu, int *min, int *max)
959 {
960 	int max_perf = cpu->pstate.turbo_pstate;
961 	int max_perf_adj;
962 	int min_perf;
963 
964 	if (limits->no_turbo || limits->turbo_disabled)
965 		max_perf = cpu->pstate.max_pstate;
966 
967 	/*
968 	 * performance can be limited by user through sysfs, by cpufreq
969 	 * policy, or by cpu specific default values determined through
970 	 * experimentation.
971 	 */
972 	max_perf_adj = fp_toint(max_perf * limits->max_perf);
973 	*max = clamp_t(int, max_perf_adj,
974 			cpu->pstate.min_pstate, cpu->pstate.turbo_pstate);
975 
976 	min_perf = fp_toint(max_perf * limits->min_perf);
977 	*min = clamp_t(int, min_perf, cpu->pstate.min_pstate, max_perf);
978 }
979 
980 static inline void intel_pstate_record_pstate(struct cpudata *cpu, int pstate)
981 {
982 	trace_cpu_frequency(pstate * cpu->pstate.scaling, cpu->cpu);
983 	cpu->pstate.current_pstate = pstate;
984 }
985 
986 static void intel_pstate_set_min_pstate(struct cpudata *cpu)
987 {
988 	int pstate = cpu->pstate.min_pstate;
989 
990 	intel_pstate_record_pstate(cpu, pstate);
991 	/*
992 	 * Generally, there is no guarantee that this code will always run on
993 	 * the CPU being updated, so force the register update to run on the
994 	 * right CPU.
995 	 */
996 	wrmsrl_on_cpu(cpu->cpu, MSR_IA32_PERF_CTL,
997 		      pstate_funcs.get_val(cpu, pstate));
998 }
999 
1000 static void intel_pstate_get_cpu_pstates(struct cpudata *cpu)
1001 {
1002 	cpu->pstate.min_pstate = pstate_funcs.get_min();
1003 	cpu->pstate.max_pstate = pstate_funcs.get_max();
1004 	cpu->pstate.max_pstate_physical = pstate_funcs.get_max_physical();
1005 	cpu->pstate.turbo_pstate = pstate_funcs.get_turbo();
1006 	cpu->pstate.scaling = pstate_funcs.get_scaling();
1007 
1008 	if (pstate_funcs.get_vid)
1009 		pstate_funcs.get_vid(cpu);
1010 
1011 	intel_pstate_set_min_pstate(cpu);
1012 }
1013 
1014 static inline void intel_pstate_calc_busy(struct cpudata *cpu)
1015 {
1016 	struct sample *sample = &cpu->sample;
1017 	int64_t core_pct;
1018 
1019 	core_pct = int_tofp(sample->aperf) * int_tofp(100);
1020 	core_pct = div64_u64(core_pct, int_tofp(sample->mperf));
1021 
1022 	sample->core_pct_busy = (int32_t)core_pct;
1023 }
1024 
1025 static inline bool intel_pstate_sample(struct cpudata *cpu, u64 time)
1026 {
1027 	u64 aperf, mperf;
1028 	unsigned long flags;
1029 	u64 tsc;
1030 
1031 	local_irq_save(flags);
1032 	rdmsrl(MSR_IA32_APERF, aperf);
1033 	rdmsrl(MSR_IA32_MPERF, mperf);
1034 	tsc = rdtsc();
1035 	if (cpu->prev_mperf == mperf || cpu->prev_tsc == tsc) {
1036 		local_irq_restore(flags);
1037 		return false;
1038 	}
1039 	local_irq_restore(flags);
1040 
1041 	cpu->last_sample_time = cpu->sample.time;
1042 	cpu->sample.time = time;
1043 	cpu->sample.aperf = aperf;
1044 	cpu->sample.mperf = mperf;
1045 	cpu->sample.tsc =  tsc;
1046 	cpu->sample.aperf -= cpu->prev_aperf;
1047 	cpu->sample.mperf -= cpu->prev_mperf;
1048 	cpu->sample.tsc -= cpu->prev_tsc;
1049 
1050 	cpu->prev_aperf = aperf;
1051 	cpu->prev_mperf = mperf;
1052 	cpu->prev_tsc = tsc;
1053 	/*
1054 	 * First time this function is invoked in a given cycle, all of the
1055 	 * previous sample data fields are equal to zero or stale and they must
1056 	 * be populated with meaningful numbers for things to work, so assume
1057 	 * that sample.time will always be reset before setting the utilization
1058 	 * update hook and make the caller skip the sample then.
1059 	 */
1060 	return !!cpu->last_sample_time;
1061 }
1062 
1063 static inline int32_t get_avg_frequency(struct cpudata *cpu)
1064 {
1065 	return div64_u64(cpu->pstate.max_pstate_physical * cpu->sample.aperf *
1066 		cpu->pstate.scaling, cpu->sample.mperf);
1067 }
1068 
1069 static inline int32_t get_target_pstate_use_cpu_load(struct cpudata *cpu)
1070 {
1071 	struct sample *sample = &cpu->sample;
1072 	u64 cummulative_iowait, delta_iowait_us;
1073 	u64 delta_iowait_mperf;
1074 	u64 mperf, now;
1075 	int32_t cpu_load;
1076 
1077 	cummulative_iowait = get_cpu_iowait_time_us(cpu->cpu, &now);
1078 
1079 	/*
1080 	 * Convert iowait time into number of IO cycles spent at max_freq.
1081 	 * IO is considered as busy only for the cpu_load algorithm. For
1082 	 * performance this is not needed since we always try to reach the
1083 	 * maximum P-State, so we are already boosting the IOs.
1084 	 */
1085 	delta_iowait_us = cummulative_iowait - cpu->prev_cummulative_iowait;
1086 	delta_iowait_mperf = div64_u64(delta_iowait_us * cpu->pstate.scaling *
1087 		cpu->pstate.max_pstate, MSEC_PER_SEC);
1088 
1089 	mperf = cpu->sample.mperf + delta_iowait_mperf;
1090 	cpu->prev_cummulative_iowait = cummulative_iowait;
1091 
1092 	/*
1093 	 * The load can be estimated as the ratio of the mperf counter
1094 	 * running at a constant frequency during active periods
1095 	 * (C0) and the time stamp counter running at the same frequency
1096 	 * also during C-states.
1097 	 */
1098 	cpu_load = div64_u64(int_tofp(100) * mperf, sample->tsc);
1099 	cpu->sample.busy_scaled = cpu_load;
1100 
1101 	return cpu->pstate.current_pstate - pid_calc(&cpu->pid, cpu_load);
1102 }
1103 
1104 static inline int32_t get_target_pstate_use_performance(struct cpudata *cpu)
1105 {
1106 	int32_t core_busy, max_pstate, current_pstate, sample_ratio;
1107 	u64 duration_ns;
1108 
1109 	intel_pstate_calc_busy(cpu);
1110 
1111 	/*
1112 	 * core_busy is the ratio of actual performance to max
1113 	 * max_pstate is the max non turbo pstate available
1114 	 * current_pstate was the pstate that was requested during
1115 	 * 	the last sample period.
1116 	 *
1117 	 * We normalize core_busy, which was our actual percent
1118 	 * performance to what we requested during the last sample
1119 	 * period. The result will be a percentage of busy at a
1120 	 * specified pstate.
1121 	 */
1122 	core_busy = cpu->sample.core_pct_busy;
1123 	max_pstate = int_tofp(cpu->pstate.max_pstate_physical);
1124 	current_pstate = int_tofp(cpu->pstate.current_pstate);
1125 	core_busy = mul_fp(core_busy, div_fp(max_pstate, current_pstate));
1126 
1127 	/*
1128 	 * Since our utilization update callback will not run unless we are
1129 	 * in C0, check if the actual elapsed time is significantly greater (3x)
1130 	 * than our sample interval.  If it is, then we were idle for a long
1131 	 * enough period of time to adjust our busyness.
1132 	 */
1133 	duration_ns = cpu->sample.time - cpu->last_sample_time;
1134 	if ((s64)duration_ns > pid_params.sample_rate_ns * 3) {
1135 		sample_ratio = div_fp(int_tofp(pid_params.sample_rate_ns),
1136 				      int_tofp(duration_ns));
1137 		core_busy = mul_fp(core_busy, sample_ratio);
1138 	} else {
1139 		sample_ratio = div_fp(100 * cpu->sample.mperf, cpu->sample.tsc);
1140 		if (sample_ratio < int_tofp(1))
1141 			core_busy = 0;
1142 	}
1143 
1144 	cpu->sample.busy_scaled = core_busy;
1145 	return cpu->pstate.current_pstate - pid_calc(&cpu->pid, core_busy);
1146 }
1147 
1148 static inline void intel_pstate_update_pstate(struct cpudata *cpu, int pstate)
1149 {
1150 	int max_perf, min_perf;
1151 
1152 	update_turbo_state();
1153 
1154 	intel_pstate_get_min_max(cpu, &min_perf, &max_perf);
1155 	pstate = clamp_t(int, pstate, min_perf, max_perf);
1156 	if (pstate == cpu->pstate.current_pstate)
1157 		return;
1158 
1159 	intel_pstate_record_pstate(cpu, pstate);
1160 	wrmsrl(MSR_IA32_PERF_CTL, pstate_funcs.get_val(cpu, pstate));
1161 }
1162 
1163 static inline void intel_pstate_adjust_busy_pstate(struct cpudata *cpu)
1164 {
1165 	int from, target_pstate;
1166 	struct sample *sample;
1167 
1168 	from = cpu->pstate.current_pstate;
1169 
1170 	target_pstate = pstate_funcs.get_target_pstate(cpu);
1171 
1172 	intel_pstate_update_pstate(cpu, target_pstate);
1173 
1174 	sample = &cpu->sample;
1175 	trace_pstate_sample(fp_toint(sample->core_pct_busy),
1176 		fp_toint(sample->busy_scaled),
1177 		from,
1178 		cpu->pstate.current_pstate,
1179 		sample->mperf,
1180 		sample->aperf,
1181 		sample->tsc,
1182 		get_avg_frequency(cpu));
1183 }
1184 
1185 static void intel_pstate_update_util(struct update_util_data *data, u64 time,
1186 				     unsigned long util, unsigned long max)
1187 {
1188 	struct cpudata *cpu = container_of(data, struct cpudata, update_util);
1189 	u64 delta_ns = time - cpu->sample.time;
1190 
1191 	if ((s64)delta_ns >= pid_params.sample_rate_ns) {
1192 		bool sample_taken = intel_pstate_sample(cpu, time);
1193 
1194 		if (sample_taken && !hwp_active)
1195 			intel_pstate_adjust_busy_pstate(cpu);
1196 	}
1197 }
1198 
1199 #define ICPU(model, policy) \
1200 	{ X86_VENDOR_INTEL, 6, model, X86_FEATURE_APERFMPERF,\
1201 			(unsigned long)&policy }
1202 
1203 static const struct x86_cpu_id intel_pstate_cpu_ids[] = {
1204 	ICPU(0x2a, core_params),
1205 	ICPU(0x2d, core_params),
1206 	ICPU(0x37, silvermont_params),
1207 	ICPU(0x3a, core_params),
1208 	ICPU(0x3c, core_params),
1209 	ICPU(0x3d, core_params),
1210 	ICPU(0x3e, core_params),
1211 	ICPU(0x3f, core_params),
1212 	ICPU(0x45, core_params),
1213 	ICPU(0x46, core_params),
1214 	ICPU(0x47, core_params),
1215 	ICPU(0x4c, airmont_params),
1216 	ICPU(0x4e, core_params),
1217 	ICPU(0x4f, core_params),
1218 	ICPU(0x5e, core_params),
1219 	ICPU(0x56, core_params),
1220 	ICPU(0x57, knl_params),
1221 	{}
1222 };
1223 MODULE_DEVICE_TABLE(x86cpu, intel_pstate_cpu_ids);
1224 
1225 static const struct x86_cpu_id intel_pstate_cpu_oob_ids[] = {
1226 	ICPU(0x56, core_params),
1227 	{}
1228 };
1229 
1230 static int intel_pstate_init_cpu(unsigned int cpunum)
1231 {
1232 	struct cpudata *cpu;
1233 
1234 	if (!all_cpu_data[cpunum])
1235 		all_cpu_data[cpunum] = kzalloc(sizeof(struct cpudata),
1236 					       GFP_KERNEL);
1237 	if (!all_cpu_data[cpunum])
1238 		return -ENOMEM;
1239 
1240 	cpu = all_cpu_data[cpunum];
1241 
1242 	cpu->cpu = cpunum;
1243 
1244 	if (hwp_active) {
1245 		intel_pstate_hwp_enable(cpu);
1246 		pid_params.sample_rate_ms = 50;
1247 		pid_params.sample_rate_ns = 50 * NSEC_PER_MSEC;
1248 	}
1249 
1250 	intel_pstate_get_cpu_pstates(cpu);
1251 
1252 	intel_pstate_busy_pid_reset(cpu);
1253 
1254 	cpu->update_util.func = intel_pstate_update_util;
1255 
1256 	pr_debug("intel_pstate: controlling: cpu %d\n", cpunum);
1257 
1258 	return 0;
1259 }
1260 
1261 static unsigned int intel_pstate_get(unsigned int cpu_num)
1262 {
1263 	struct sample *sample;
1264 	struct cpudata *cpu;
1265 
1266 	cpu = all_cpu_data[cpu_num];
1267 	if (!cpu)
1268 		return 0;
1269 	sample = &cpu->sample;
1270 	return get_avg_frequency(cpu);
1271 }
1272 
1273 static void intel_pstate_set_update_util_hook(unsigned int cpu_num)
1274 {
1275 	struct cpudata *cpu = all_cpu_data[cpu_num];
1276 
1277 	/* Prevent intel_pstate_update_util() from using stale data. */
1278 	cpu->sample.time = 0;
1279 	cpufreq_set_update_util_data(cpu_num, &cpu->update_util);
1280 }
1281 
1282 static void intel_pstate_clear_update_util_hook(unsigned int cpu)
1283 {
1284 	cpufreq_set_update_util_data(cpu, NULL);
1285 	synchronize_sched();
1286 }
1287 
1288 static void intel_pstate_set_performance_limits(struct perf_limits *limits)
1289 {
1290 	limits->no_turbo = 0;
1291 	limits->turbo_disabled = 0;
1292 	limits->max_perf_pct = 100;
1293 	limits->max_perf = int_tofp(1);
1294 	limits->min_perf_pct = 100;
1295 	limits->min_perf = int_tofp(1);
1296 	limits->max_policy_pct = 100;
1297 	limits->max_sysfs_pct = 100;
1298 	limits->min_policy_pct = 0;
1299 	limits->min_sysfs_pct = 0;
1300 }
1301 
1302 static int intel_pstate_set_policy(struct cpufreq_policy *policy)
1303 {
1304 	if (!policy->cpuinfo.max_freq)
1305 		return -ENODEV;
1306 
1307 	intel_pstate_clear_update_util_hook(policy->cpu);
1308 
1309 	if (policy->policy == CPUFREQ_POLICY_PERFORMANCE) {
1310 		limits = &performance_limits;
1311 		if (policy->max >= policy->cpuinfo.max_freq) {
1312 			pr_debug("intel_pstate: set performance\n");
1313 			intel_pstate_set_performance_limits(limits);
1314 			goto out;
1315 		}
1316 	} else {
1317 		pr_debug("intel_pstate: set powersave\n");
1318 		limits = &powersave_limits;
1319 	}
1320 
1321 	limits->min_policy_pct = (policy->min * 100) / policy->cpuinfo.max_freq;
1322 	limits->min_policy_pct = clamp_t(int, limits->min_policy_pct, 0 , 100);
1323 	limits->max_policy_pct = DIV_ROUND_UP(policy->max * 100,
1324 					      policy->cpuinfo.max_freq);
1325 	limits->max_policy_pct = clamp_t(int, limits->max_policy_pct, 0 , 100);
1326 
1327 	/* Normalize user input to [min_policy_pct, max_policy_pct] */
1328 	limits->min_perf_pct = max(limits->min_policy_pct,
1329 				   limits->min_sysfs_pct);
1330 	limits->min_perf_pct = min(limits->max_policy_pct,
1331 				   limits->min_perf_pct);
1332 	limits->max_perf_pct = min(limits->max_policy_pct,
1333 				   limits->max_sysfs_pct);
1334 	limits->max_perf_pct = max(limits->min_policy_pct,
1335 				   limits->max_perf_pct);
1336 	limits->max_perf = round_up(limits->max_perf, FRAC_BITS);
1337 
1338 	/* Make sure min_perf_pct <= max_perf_pct */
1339 	limits->min_perf_pct = min(limits->max_perf_pct, limits->min_perf_pct);
1340 
1341 	limits->min_perf = div_fp(int_tofp(limits->min_perf_pct),
1342 				  int_tofp(100));
1343 	limits->max_perf = div_fp(int_tofp(limits->max_perf_pct),
1344 				  int_tofp(100));
1345 
1346  out:
1347 	intel_pstate_set_update_util_hook(policy->cpu);
1348 
1349 	if (hwp_active)
1350 		intel_pstate_hwp_set(policy->cpus);
1351 
1352 	return 0;
1353 }
1354 
1355 static int intel_pstate_verify_policy(struct cpufreq_policy *policy)
1356 {
1357 	cpufreq_verify_within_cpu_limits(policy);
1358 
1359 	if (policy->policy != CPUFREQ_POLICY_POWERSAVE &&
1360 	    policy->policy != CPUFREQ_POLICY_PERFORMANCE)
1361 		return -EINVAL;
1362 
1363 	return 0;
1364 }
1365 
1366 static void intel_pstate_stop_cpu(struct cpufreq_policy *policy)
1367 {
1368 	int cpu_num = policy->cpu;
1369 	struct cpudata *cpu = all_cpu_data[cpu_num];
1370 
1371 	pr_debug("intel_pstate: CPU %d exiting\n", cpu_num);
1372 
1373 	intel_pstate_clear_update_util_hook(cpu_num);
1374 
1375 	if (hwp_active)
1376 		return;
1377 
1378 	intel_pstate_set_min_pstate(cpu);
1379 }
1380 
1381 static int intel_pstate_cpu_init(struct cpufreq_policy *policy)
1382 {
1383 	struct cpudata *cpu;
1384 	int rc;
1385 
1386 	rc = intel_pstate_init_cpu(policy->cpu);
1387 	if (rc)
1388 		return rc;
1389 
1390 	cpu = all_cpu_data[policy->cpu];
1391 
1392 	if (limits->min_perf_pct == 100 && limits->max_perf_pct == 100)
1393 		policy->policy = CPUFREQ_POLICY_PERFORMANCE;
1394 	else
1395 		policy->policy = CPUFREQ_POLICY_POWERSAVE;
1396 
1397 	policy->min = cpu->pstate.min_pstate * cpu->pstate.scaling;
1398 	policy->max = cpu->pstate.turbo_pstate * cpu->pstate.scaling;
1399 
1400 	/* cpuinfo and default policy values */
1401 	policy->cpuinfo.min_freq = cpu->pstate.min_pstate * cpu->pstate.scaling;
1402 	policy->cpuinfo.max_freq =
1403 		cpu->pstate.turbo_pstate * cpu->pstate.scaling;
1404 	policy->cpuinfo.transition_latency = CPUFREQ_ETERNAL;
1405 	cpumask_set_cpu(policy->cpu, policy->cpus);
1406 
1407 	return 0;
1408 }
1409 
1410 static struct cpufreq_driver intel_pstate_driver = {
1411 	.flags		= CPUFREQ_CONST_LOOPS,
1412 	.verify		= intel_pstate_verify_policy,
1413 	.setpolicy	= intel_pstate_set_policy,
1414 	.get		= intel_pstate_get,
1415 	.init		= intel_pstate_cpu_init,
1416 	.stop_cpu	= intel_pstate_stop_cpu,
1417 	.name		= "intel_pstate",
1418 };
1419 
1420 static int __initdata no_load;
1421 static int __initdata no_hwp;
1422 static int __initdata hwp_only;
1423 static unsigned int force_load;
1424 
1425 static int intel_pstate_msrs_not_valid(void)
1426 {
1427 	if (!pstate_funcs.get_max() ||
1428 	    !pstate_funcs.get_min() ||
1429 	    !pstate_funcs.get_turbo())
1430 		return -ENODEV;
1431 
1432 	return 0;
1433 }
1434 
1435 static void copy_pid_params(struct pstate_adjust_policy *policy)
1436 {
1437 	pid_params.sample_rate_ms = policy->sample_rate_ms;
1438 	pid_params.sample_rate_ns = pid_params.sample_rate_ms * NSEC_PER_MSEC;
1439 	pid_params.p_gain_pct = policy->p_gain_pct;
1440 	pid_params.i_gain_pct = policy->i_gain_pct;
1441 	pid_params.d_gain_pct = policy->d_gain_pct;
1442 	pid_params.deadband = policy->deadband;
1443 	pid_params.setpoint = policy->setpoint;
1444 }
1445 
1446 static void copy_cpu_funcs(struct pstate_funcs *funcs)
1447 {
1448 	pstate_funcs.get_max   = funcs->get_max;
1449 	pstate_funcs.get_max_physical = funcs->get_max_physical;
1450 	pstate_funcs.get_min   = funcs->get_min;
1451 	pstate_funcs.get_turbo = funcs->get_turbo;
1452 	pstate_funcs.get_scaling = funcs->get_scaling;
1453 	pstate_funcs.get_val   = funcs->get_val;
1454 	pstate_funcs.get_vid   = funcs->get_vid;
1455 	pstate_funcs.get_target_pstate = funcs->get_target_pstate;
1456 
1457 }
1458 
1459 #if IS_ENABLED(CONFIG_ACPI)
1460 #include <acpi/processor.h>
1461 
1462 static bool intel_pstate_no_acpi_pss(void)
1463 {
1464 	int i;
1465 
1466 	for_each_possible_cpu(i) {
1467 		acpi_status status;
1468 		union acpi_object *pss;
1469 		struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
1470 		struct acpi_processor *pr = per_cpu(processors, i);
1471 
1472 		if (!pr)
1473 			continue;
1474 
1475 		status = acpi_evaluate_object(pr->handle, "_PSS", NULL, &buffer);
1476 		if (ACPI_FAILURE(status))
1477 			continue;
1478 
1479 		pss = buffer.pointer;
1480 		if (pss && pss->type == ACPI_TYPE_PACKAGE) {
1481 			kfree(pss);
1482 			return false;
1483 		}
1484 
1485 		kfree(pss);
1486 	}
1487 
1488 	return true;
1489 }
1490 
1491 static bool intel_pstate_has_acpi_ppc(void)
1492 {
1493 	int i;
1494 
1495 	for_each_possible_cpu(i) {
1496 		struct acpi_processor *pr = per_cpu(processors, i);
1497 
1498 		if (!pr)
1499 			continue;
1500 		if (acpi_has_method(pr->handle, "_PPC"))
1501 			return true;
1502 	}
1503 	return false;
1504 }
1505 
1506 enum {
1507 	PSS,
1508 	PPC,
1509 };
1510 
1511 struct hw_vendor_info {
1512 	u16  valid;
1513 	char oem_id[ACPI_OEM_ID_SIZE];
1514 	char oem_table_id[ACPI_OEM_TABLE_ID_SIZE];
1515 	int  oem_pwr_table;
1516 };
1517 
1518 /* Hardware vendor-specific info that has its own power management modes */
1519 static struct hw_vendor_info vendor_info[] = {
1520 	{1, "HP    ", "ProLiant", PSS},
1521 	{1, "ORACLE", "X4-2    ", PPC},
1522 	{1, "ORACLE", "X4-2L   ", PPC},
1523 	{1, "ORACLE", "X4-2B   ", PPC},
1524 	{1, "ORACLE", "X3-2    ", PPC},
1525 	{1, "ORACLE", "X3-2L   ", PPC},
1526 	{1, "ORACLE", "X3-2B   ", PPC},
1527 	{1, "ORACLE", "X4470M2 ", PPC},
1528 	{1, "ORACLE", "X4270M3 ", PPC},
1529 	{1, "ORACLE", "X4270M2 ", PPC},
1530 	{1, "ORACLE", "X4170M2 ", PPC},
1531 	{1, "ORACLE", "X4170 M3", PPC},
1532 	{1, "ORACLE", "X4275 M3", PPC},
1533 	{1, "ORACLE", "X6-2    ", PPC},
1534 	{1, "ORACLE", "Sudbury ", PPC},
1535 	{0, "", ""},
1536 };
1537 
1538 static bool intel_pstate_platform_pwr_mgmt_exists(void)
1539 {
1540 	struct acpi_table_header hdr;
1541 	struct hw_vendor_info *v_info;
1542 	const struct x86_cpu_id *id;
1543 	u64 misc_pwr;
1544 
1545 	id = x86_match_cpu(intel_pstate_cpu_oob_ids);
1546 	if (id) {
1547 		rdmsrl(MSR_MISC_PWR_MGMT, misc_pwr);
1548 		if ( misc_pwr & (1 << 8))
1549 			return true;
1550 	}
1551 
1552 	if (acpi_disabled ||
1553 	    ACPI_FAILURE(acpi_get_table_header(ACPI_SIG_FADT, 0, &hdr)))
1554 		return false;
1555 
1556 	for (v_info = vendor_info; v_info->valid; v_info++) {
1557 		if (!strncmp(hdr.oem_id, v_info->oem_id, ACPI_OEM_ID_SIZE) &&
1558 			!strncmp(hdr.oem_table_id, v_info->oem_table_id,
1559 						ACPI_OEM_TABLE_ID_SIZE))
1560 			switch (v_info->oem_pwr_table) {
1561 			case PSS:
1562 				return intel_pstate_no_acpi_pss();
1563 			case PPC:
1564 				return intel_pstate_has_acpi_ppc() &&
1565 					(!force_load);
1566 			}
1567 	}
1568 
1569 	return false;
1570 }
1571 #else /* CONFIG_ACPI not enabled */
1572 static inline bool intel_pstate_platform_pwr_mgmt_exists(void) { return false; }
1573 static inline bool intel_pstate_has_acpi_ppc(void) { return false; }
1574 #endif /* CONFIG_ACPI */
1575 
1576 static const struct x86_cpu_id hwp_support_ids[] __initconst = {
1577 	{ X86_VENDOR_INTEL, 6, X86_MODEL_ANY, X86_FEATURE_HWP },
1578 	{}
1579 };
1580 
1581 static int __init intel_pstate_init(void)
1582 {
1583 	int cpu, rc = 0;
1584 	const struct x86_cpu_id *id;
1585 	struct cpu_defaults *cpu_def;
1586 
1587 	if (no_load)
1588 		return -ENODEV;
1589 
1590 	if (x86_match_cpu(hwp_support_ids) && !no_hwp) {
1591 		copy_cpu_funcs(&core_params.funcs);
1592 		hwp_active++;
1593 		goto hwp_cpu_matched;
1594 	}
1595 
1596 	id = x86_match_cpu(intel_pstate_cpu_ids);
1597 	if (!id)
1598 		return -ENODEV;
1599 
1600 	cpu_def = (struct cpu_defaults *)id->driver_data;
1601 
1602 	copy_pid_params(&cpu_def->pid_policy);
1603 	copy_cpu_funcs(&cpu_def->funcs);
1604 
1605 	if (intel_pstate_msrs_not_valid())
1606 		return -ENODEV;
1607 
1608 hwp_cpu_matched:
1609 	/*
1610 	 * The Intel pstate driver will be ignored if the platform
1611 	 * firmware has its own power management modes.
1612 	 */
1613 	if (intel_pstate_platform_pwr_mgmt_exists())
1614 		return -ENODEV;
1615 
1616 	pr_info("Intel P-state driver initializing.\n");
1617 
1618 	all_cpu_data = vzalloc(sizeof(void *) * num_possible_cpus());
1619 	if (!all_cpu_data)
1620 		return -ENOMEM;
1621 
1622 	if (!hwp_active && hwp_only)
1623 		goto out;
1624 
1625 	rc = cpufreq_register_driver(&intel_pstate_driver);
1626 	if (rc)
1627 		goto out;
1628 
1629 	intel_pstate_debug_expose_params();
1630 	intel_pstate_sysfs_expose_params();
1631 
1632 	if (hwp_active)
1633 		pr_info("intel_pstate: HWP enabled\n");
1634 
1635 	return rc;
1636 out:
1637 	get_online_cpus();
1638 	for_each_online_cpu(cpu) {
1639 		if (all_cpu_data[cpu]) {
1640 			intel_pstate_clear_update_util_hook(cpu);
1641 			kfree(all_cpu_data[cpu]);
1642 		}
1643 	}
1644 
1645 	put_online_cpus();
1646 	vfree(all_cpu_data);
1647 	return -ENODEV;
1648 }
1649 device_initcall(intel_pstate_init);
1650 
1651 static int __init intel_pstate_setup(char *str)
1652 {
1653 	if (!str)
1654 		return -EINVAL;
1655 
1656 	if (!strcmp(str, "disable"))
1657 		no_load = 1;
1658 	if (!strcmp(str, "no_hwp")) {
1659 		pr_info("intel_pstate: HWP disabled\n");
1660 		no_hwp = 1;
1661 	}
1662 	if (!strcmp(str, "force"))
1663 		force_load = 1;
1664 	if (!strcmp(str, "hwp_only"))
1665 		hwp_only = 1;
1666 	return 0;
1667 }
1668 early_param("intel_pstate", intel_pstate_setup);
1669 
1670 MODULE_AUTHOR("Dirk Brandewie <dirk.j.brandewie@intel.com>");
1671 MODULE_DESCRIPTION("'intel_pstate' - P state driver Intel Core processors");
1672 MODULE_LICENSE("GPL");
1673