xref: /linux/drivers/cpufreq/intel_pstate.c (revision 32786fdc9506aeba98278c1844d4bfb766863832)
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 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
14 
15 #include <linux/kernel.h>
16 #include <linux/kernel_stat.h>
17 #include <linux/module.h>
18 #include <linux/ktime.h>
19 #include <linux/hrtimer.h>
20 #include <linux/tick.h>
21 #include <linux/slab.h>
22 #include <linux/sched.h>
23 #include <linux/list.h>
24 #include <linux/cpu.h>
25 #include <linux/cpufreq.h>
26 #include <linux/sysfs.h>
27 #include <linux/types.h>
28 #include <linux/fs.h>
29 #include <linux/debugfs.h>
30 #include <linux/acpi.h>
31 #include <linux/vmalloc.h>
32 #include <trace/events/power.h>
33 
34 #include <asm/div64.h>
35 #include <asm/msr.h>
36 #include <asm/cpu_device_id.h>
37 #include <asm/cpufeature.h>
38 #include <asm/intel-family.h>
39 
40 #define INTEL_CPUFREQ_TRANSITION_LATENCY	20000
41 
42 #define ATOM_RATIOS		0x66a
43 #define ATOM_VIDS		0x66b
44 #define ATOM_TURBO_RATIOS	0x66c
45 #define ATOM_TURBO_VIDS		0x66d
46 
47 #ifdef CONFIG_ACPI
48 #include <acpi/processor.h>
49 #include <acpi/cppc_acpi.h>
50 #endif
51 
52 #define FRAC_BITS 8
53 #define int_tofp(X) ((int64_t)(X) << FRAC_BITS)
54 #define fp_toint(X) ((X) >> FRAC_BITS)
55 
56 #define EXT_BITS 6
57 #define EXT_FRAC_BITS (EXT_BITS + FRAC_BITS)
58 #define fp_ext_toint(X) ((X) >> EXT_FRAC_BITS)
59 #define int_ext_tofp(X) ((int64_t)(X) << EXT_FRAC_BITS)
60 
61 static inline int32_t mul_fp(int32_t x, int32_t y)
62 {
63 	return ((int64_t)x * (int64_t)y) >> FRAC_BITS;
64 }
65 
66 static inline int32_t div_fp(s64 x, s64 y)
67 {
68 	return div64_s64((int64_t)x << FRAC_BITS, y);
69 }
70 
71 static inline int ceiling_fp(int32_t x)
72 {
73 	int mask, ret;
74 
75 	ret = fp_toint(x);
76 	mask = (1 << FRAC_BITS) - 1;
77 	if (x & mask)
78 		ret += 1;
79 	return ret;
80 }
81 
82 static inline u64 mul_ext_fp(u64 x, u64 y)
83 {
84 	return (x * y) >> EXT_FRAC_BITS;
85 }
86 
87 static inline u64 div_ext_fp(u64 x, u64 y)
88 {
89 	return div64_u64(x << EXT_FRAC_BITS, y);
90 }
91 
92 /**
93  * struct sample -	Store performance sample
94  * @core_avg_perf:	Ratio of APERF/MPERF which is the actual average
95  *			performance during last sample period
96  * @busy_scaled:	Scaled busy value which is used to calculate next
97  *			P state. This can be different than core_avg_perf
98  *			to account for cpu idle period
99  * @aperf:		Difference of actual performance frequency clock count
100  *			read from APERF MSR between last and current sample
101  * @mperf:		Difference of maximum performance frequency clock count
102  *			read from MPERF MSR between last and current sample
103  * @tsc:		Difference of time stamp counter between last and
104  *			current sample
105  * @time:		Current time from scheduler
106  *
107  * This structure is used in the cpudata structure to store performance sample
108  * data for choosing next P State.
109  */
110 struct sample {
111 	int32_t core_avg_perf;
112 	int32_t busy_scaled;
113 	u64 aperf;
114 	u64 mperf;
115 	u64 tsc;
116 	u64 time;
117 };
118 
119 /**
120  * struct pstate_data - Store P state data
121  * @current_pstate:	Current requested P state
122  * @min_pstate:		Min P state possible for this platform
123  * @max_pstate:		Max P state possible for this platform
124  * @max_pstate_physical:This is physical Max P state for a processor
125  *			This can be higher than the max_pstate which can
126  *			be limited by platform thermal design power limits
127  * @scaling:		Scaling factor to  convert frequency to cpufreq
128  *			frequency units
129  * @turbo_pstate:	Max Turbo P state possible for this platform
130  * @max_freq:		@max_pstate frequency in cpufreq units
131  * @turbo_freq:		@turbo_pstate frequency in cpufreq units
132  *
133  * Stores the per cpu model P state limits and current P state.
134  */
135 struct pstate_data {
136 	int	current_pstate;
137 	int	min_pstate;
138 	int	max_pstate;
139 	int	max_pstate_physical;
140 	int	scaling;
141 	int	turbo_pstate;
142 	unsigned int max_freq;
143 	unsigned int turbo_freq;
144 };
145 
146 /**
147  * struct vid_data -	Stores voltage information data
148  * @min:		VID data for this platform corresponding to
149  *			the lowest P state
150  * @max:		VID data corresponding to the highest P State.
151  * @turbo:		VID data for turbo P state
152  * @ratio:		Ratio of (vid max - vid min) /
153  *			(max P state - Min P State)
154  *
155  * Stores the voltage data for DVFS (Dynamic Voltage and Frequency Scaling)
156  * This data is used in Atom platforms, where in addition to target P state,
157  * the voltage data needs to be specified to select next P State.
158  */
159 struct vid_data {
160 	int min;
161 	int max;
162 	int turbo;
163 	int32_t ratio;
164 };
165 
166 /**
167  * struct _pid -	Stores PID data
168  * @setpoint:		Target set point for busyness or performance
169  * @integral:		Storage for accumulated error values
170  * @p_gain:		PID proportional gain
171  * @i_gain:		PID integral gain
172  * @d_gain:		PID derivative gain
173  * @deadband:		PID deadband
174  * @last_err:		Last error storage for integral part of PID calculation
175  *
176  * Stores PID coefficients and last error for PID controller.
177  */
178 struct _pid {
179 	int setpoint;
180 	int32_t integral;
181 	int32_t p_gain;
182 	int32_t i_gain;
183 	int32_t d_gain;
184 	int deadband;
185 	int32_t last_err;
186 };
187 
188 /**
189  * struct perf_limits - Store user and policy limits
190  * @no_turbo:		User requested turbo state from intel_pstate sysfs
191  * @turbo_disabled:	Platform turbo status either from msr
192  *			MSR_IA32_MISC_ENABLE or when maximum available pstate
193  *			matches the maximum turbo pstate
194  * @max_perf_pct:	Effective maximum performance limit in percentage, this
195  *			is minimum of either limits enforced by cpufreq policy
196  *			or limits from user set limits via intel_pstate sysfs
197  * @min_perf_pct:	Effective minimum performance limit in percentage, this
198  *			is maximum of either limits enforced by cpufreq policy
199  *			or limits from user set limits via intel_pstate sysfs
200  * @max_perf:		This is a scaled value between 0 to 255 for max_perf_pct
201  *			This value is used to limit max pstate
202  * @min_perf:		This is a scaled value between 0 to 255 for min_perf_pct
203  *			This value is used to limit min pstate
204  * @max_policy_pct:	The maximum performance in percentage enforced by
205  *			cpufreq setpolicy interface
206  * @max_sysfs_pct:	The maximum performance in percentage enforced by
207  *			intel pstate sysfs interface, unused when per cpu
208  *			controls are enforced
209  * @min_policy_pct:	The minimum performance in percentage enforced by
210  *			cpufreq setpolicy interface
211  * @min_sysfs_pct:	The minimum performance in percentage enforced by
212  *			intel pstate sysfs interface, unused when per cpu
213  *			controls are enforced
214  *
215  * Storage for user and policy defined limits.
216  */
217 struct perf_limits {
218 	int no_turbo;
219 	int turbo_disabled;
220 	int max_perf_pct;
221 	int min_perf_pct;
222 	int32_t max_perf;
223 	int32_t min_perf;
224 	int max_policy_pct;
225 	int max_sysfs_pct;
226 	int min_policy_pct;
227 	int min_sysfs_pct;
228 };
229 
230 /**
231  * struct cpudata -	Per CPU instance data storage
232  * @cpu:		CPU number for this instance data
233  * @policy:		CPUFreq policy value
234  * @update_util:	CPUFreq utility callback information
235  * @update_util_set:	CPUFreq utility callback is set
236  * @iowait_boost:	iowait-related boost fraction
237  * @last_update:	Time of the last update.
238  * @pstate:		Stores P state limits for this CPU
239  * @vid:		Stores VID limits for this CPU
240  * @pid:		Stores PID parameters for this CPU
241  * @last_sample_time:	Last Sample time
242  * @prev_aperf:		Last APERF value read from APERF MSR
243  * @prev_mperf:		Last MPERF value read from MPERF MSR
244  * @prev_tsc:		Last timestamp counter (TSC) value
245  * @prev_cummulative_iowait: IO Wait time difference from last and
246  *			current sample
247  * @sample:		Storage for storing last Sample data
248  * @perf_limits:	Pointer to perf_limit unique to this CPU
249  *			Not all field in the structure are applicable
250  *			when per cpu controls are enforced
251  * @acpi_perf_data:	Stores ACPI perf information read from _PSS
252  * @valid_pss_table:	Set to true for valid ACPI _PSS entries found
253  * @epp_powersave:	Last saved HWP energy performance preference
254  *			(EPP) or energy performance bias (EPB),
255  *			when policy switched to performance
256  * @epp_policy:		Last saved policy used to set EPP/EPB
257  * @epp_default:	Power on default HWP energy performance
258  *			preference/bias
259  * @epp_saved:		Saved EPP/EPB during system suspend or CPU offline
260  *			operation
261  *
262  * This structure stores per CPU instance data for all CPUs.
263  */
264 struct cpudata {
265 	int cpu;
266 
267 	unsigned int policy;
268 	struct update_util_data update_util;
269 	bool   update_util_set;
270 
271 	struct pstate_data pstate;
272 	struct vid_data vid;
273 	struct _pid pid;
274 
275 	u64	last_update;
276 	u64	last_sample_time;
277 	u64	prev_aperf;
278 	u64	prev_mperf;
279 	u64	prev_tsc;
280 	u64	prev_cummulative_iowait;
281 	struct sample sample;
282 	struct perf_limits *perf_limits;
283 #ifdef CONFIG_ACPI
284 	struct acpi_processor_performance acpi_perf_data;
285 	bool valid_pss_table;
286 #endif
287 	unsigned int iowait_boost;
288 	s16 epp_powersave;
289 	s16 epp_policy;
290 	s16 epp_default;
291 	s16 epp_saved;
292 };
293 
294 static struct cpudata **all_cpu_data;
295 
296 /**
297  * struct pstate_adjust_policy - Stores static PID configuration data
298  * @sample_rate_ms:	PID calculation sample rate in ms
299  * @sample_rate_ns:	Sample rate calculation in ns
300  * @deadband:		PID deadband
301  * @setpoint:		PID Setpoint
302  * @p_gain_pct:		PID proportional gain
303  * @i_gain_pct:		PID integral gain
304  * @d_gain_pct:		PID derivative gain
305  *
306  * Stores per CPU model static PID configuration data.
307  */
308 struct pstate_adjust_policy {
309 	int sample_rate_ms;
310 	s64 sample_rate_ns;
311 	int deadband;
312 	int setpoint;
313 	int p_gain_pct;
314 	int d_gain_pct;
315 	int i_gain_pct;
316 };
317 
318 /**
319  * struct pstate_funcs - Per CPU model specific callbacks
320  * @get_max:		Callback to get maximum non turbo effective P state
321  * @get_max_physical:	Callback to get maximum non turbo physical P state
322  * @get_min:		Callback to get minimum P state
323  * @get_turbo:		Callback to get turbo P state
324  * @get_scaling:	Callback to get frequency scaling factor
325  * @get_val:		Callback to convert P state to actual MSR write value
326  * @get_vid:		Callback to get VID data for Atom platforms
327  * @get_target_pstate:	Callback to a function to calculate next P state to use
328  *
329  * Core and Atom CPU models have different way to get P State limits. This
330  * structure is used to store those callbacks.
331  */
332 struct pstate_funcs {
333 	int (*get_max)(void);
334 	int (*get_max_physical)(void);
335 	int (*get_min)(void);
336 	int (*get_turbo)(void);
337 	int (*get_scaling)(void);
338 	u64 (*get_val)(struct cpudata*, int pstate);
339 	void (*get_vid)(struct cpudata *);
340 	int32_t (*get_target_pstate)(struct cpudata *);
341 };
342 
343 /**
344  * struct cpu_defaults- Per CPU model default config data
345  * @pid_policy:	PID config data
346  * @funcs:		Callback function data
347  */
348 struct cpu_defaults {
349 	struct pstate_adjust_policy pid_policy;
350 	struct pstate_funcs funcs;
351 };
352 
353 static inline int32_t get_target_pstate_use_performance(struct cpudata *cpu);
354 static inline int32_t get_target_pstate_use_cpu_load(struct cpudata *cpu);
355 
356 static struct pstate_adjust_policy pid_params __read_mostly;
357 static struct pstate_funcs pstate_funcs __read_mostly;
358 static int hwp_active __read_mostly;
359 static bool per_cpu_limits __read_mostly;
360 
361 #ifdef CONFIG_ACPI
362 static bool acpi_ppc;
363 #endif
364 
365 static struct perf_limits performance_limits = {
366 	.no_turbo = 0,
367 	.turbo_disabled = 0,
368 	.max_perf_pct = 100,
369 	.max_perf = int_ext_tofp(1),
370 	.min_perf_pct = 100,
371 	.min_perf = int_ext_tofp(1),
372 	.max_policy_pct = 100,
373 	.max_sysfs_pct = 100,
374 	.min_policy_pct = 0,
375 	.min_sysfs_pct = 0,
376 };
377 
378 static struct perf_limits powersave_limits = {
379 	.no_turbo = 0,
380 	.turbo_disabled = 0,
381 	.max_perf_pct = 100,
382 	.max_perf = int_ext_tofp(1),
383 	.min_perf_pct = 0,
384 	.min_perf = 0,
385 	.max_policy_pct = 100,
386 	.max_sysfs_pct = 100,
387 	.min_policy_pct = 0,
388 	.min_sysfs_pct = 0,
389 };
390 
391 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_PERFORMANCE
392 static struct perf_limits *limits = &performance_limits;
393 #else
394 static struct perf_limits *limits = &powersave_limits;
395 #endif
396 
397 static DEFINE_MUTEX(intel_pstate_limits_lock);
398 
399 #ifdef CONFIG_ACPI
400 
401 static bool intel_pstate_get_ppc_enable_status(void)
402 {
403 	if (acpi_gbl_FADT.preferred_profile == PM_ENTERPRISE_SERVER ||
404 	    acpi_gbl_FADT.preferred_profile == PM_PERFORMANCE_SERVER)
405 		return true;
406 
407 	return acpi_ppc;
408 }
409 
410 #ifdef CONFIG_ACPI_CPPC_LIB
411 
412 /* The work item is needed to avoid CPU hotplug locking issues */
413 static void intel_pstste_sched_itmt_work_fn(struct work_struct *work)
414 {
415 	sched_set_itmt_support();
416 }
417 
418 static DECLARE_WORK(sched_itmt_work, intel_pstste_sched_itmt_work_fn);
419 
420 static void intel_pstate_set_itmt_prio(int cpu)
421 {
422 	struct cppc_perf_caps cppc_perf;
423 	static u32 max_highest_perf = 0, min_highest_perf = U32_MAX;
424 	int ret;
425 
426 	ret = cppc_get_perf_caps(cpu, &cppc_perf);
427 	if (ret)
428 		return;
429 
430 	/*
431 	 * The priorities can be set regardless of whether or not
432 	 * sched_set_itmt_support(true) has been called and it is valid to
433 	 * update them at any time after it has been called.
434 	 */
435 	sched_set_itmt_core_prio(cppc_perf.highest_perf, cpu);
436 
437 	if (max_highest_perf <= min_highest_perf) {
438 		if (cppc_perf.highest_perf > max_highest_perf)
439 			max_highest_perf = cppc_perf.highest_perf;
440 
441 		if (cppc_perf.highest_perf < min_highest_perf)
442 			min_highest_perf = cppc_perf.highest_perf;
443 
444 		if (max_highest_perf > min_highest_perf) {
445 			/*
446 			 * This code can be run during CPU online under the
447 			 * CPU hotplug locks, so sched_set_itmt_support()
448 			 * cannot be called from here.  Queue up a work item
449 			 * to invoke it.
450 			 */
451 			schedule_work(&sched_itmt_work);
452 		}
453 	}
454 }
455 #else
456 static void intel_pstate_set_itmt_prio(int cpu)
457 {
458 }
459 #endif
460 
461 static void intel_pstate_init_acpi_perf_limits(struct cpufreq_policy *policy)
462 {
463 	struct cpudata *cpu;
464 	int ret;
465 	int i;
466 
467 	if (hwp_active) {
468 		intel_pstate_set_itmt_prio(policy->cpu);
469 		return;
470 	}
471 
472 	if (!intel_pstate_get_ppc_enable_status())
473 		return;
474 
475 	cpu = all_cpu_data[policy->cpu];
476 
477 	ret = acpi_processor_register_performance(&cpu->acpi_perf_data,
478 						  policy->cpu);
479 	if (ret)
480 		return;
481 
482 	/*
483 	 * Check if the control value in _PSS is for PERF_CTL MSR, which should
484 	 * guarantee that the states returned by it map to the states in our
485 	 * list directly.
486 	 */
487 	if (cpu->acpi_perf_data.control_register.space_id !=
488 						ACPI_ADR_SPACE_FIXED_HARDWARE)
489 		goto err;
490 
491 	/*
492 	 * If there is only one entry _PSS, simply ignore _PSS and continue as
493 	 * usual without taking _PSS into account
494 	 */
495 	if (cpu->acpi_perf_data.state_count < 2)
496 		goto err;
497 
498 	pr_debug("CPU%u - ACPI _PSS perf data\n", policy->cpu);
499 	for (i = 0; i < cpu->acpi_perf_data.state_count; i++) {
500 		pr_debug("     %cP%d: %u MHz, %u mW, 0x%x\n",
501 			 (i == cpu->acpi_perf_data.state ? '*' : ' '), i,
502 			 (u32) cpu->acpi_perf_data.states[i].core_frequency,
503 			 (u32) cpu->acpi_perf_data.states[i].power,
504 			 (u32) cpu->acpi_perf_data.states[i].control);
505 	}
506 
507 	/*
508 	 * The _PSS table doesn't contain whole turbo frequency range.
509 	 * This just contains +1 MHZ above the max non turbo frequency,
510 	 * with control value corresponding to max turbo ratio. But
511 	 * when cpufreq set policy is called, it will call with this
512 	 * max frequency, which will cause a reduced performance as
513 	 * this driver uses real max turbo frequency as the max
514 	 * frequency. So correct this frequency in _PSS table to
515 	 * correct max turbo frequency based on the turbo state.
516 	 * Also need to convert to MHz as _PSS freq is in MHz.
517 	 */
518 	if (!limits->turbo_disabled)
519 		cpu->acpi_perf_data.states[0].core_frequency =
520 					policy->cpuinfo.max_freq / 1000;
521 	cpu->valid_pss_table = true;
522 	pr_debug("_PPC limits will be enforced\n");
523 
524 	return;
525 
526  err:
527 	cpu->valid_pss_table = false;
528 	acpi_processor_unregister_performance(policy->cpu);
529 }
530 
531 static void intel_pstate_exit_perf_limits(struct cpufreq_policy *policy)
532 {
533 	struct cpudata *cpu;
534 
535 	cpu = all_cpu_data[policy->cpu];
536 	if (!cpu->valid_pss_table)
537 		return;
538 
539 	acpi_processor_unregister_performance(policy->cpu);
540 }
541 
542 #else
543 static inline void intel_pstate_init_acpi_perf_limits(struct cpufreq_policy *policy)
544 {
545 }
546 
547 static inline void intel_pstate_exit_perf_limits(struct cpufreq_policy *policy)
548 {
549 }
550 #endif
551 
552 static inline void pid_reset(struct _pid *pid, int setpoint, int busy,
553 			     int deadband, int integral) {
554 	pid->setpoint = int_tofp(setpoint);
555 	pid->deadband  = int_tofp(deadband);
556 	pid->integral  = int_tofp(integral);
557 	pid->last_err  = int_tofp(setpoint) - int_tofp(busy);
558 }
559 
560 static inline void pid_p_gain_set(struct _pid *pid, int percent)
561 {
562 	pid->p_gain = div_fp(percent, 100);
563 }
564 
565 static inline void pid_i_gain_set(struct _pid *pid, int percent)
566 {
567 	pid->i_gain = div_fp(percent, 100);
568 }
569 
570 static inline void pid_d_gain_set(struct _pid *pid, int percent)
571 {
572 	pid->d_gain = div_fp(percent, 100);
573 }
574 
575 static signed int pid_calc(struct _pid *pid, int32_t busy)
576 {
577 	signed int result;
578 	int32_t pterm, dterm, fp_error;
579 	int32_t integral_limit;
580 
581 	fp_error = pid->setpoint - busy;
582 
583 	if (abs(fp_error) <= pid->deadband)
584 		return 0;
585 
586 	pterm = mul_fp(pid->p_gain, fp_error);
587 
588 	pid->integral += fp_error;
589 
590 	/*
591 	 * We limit the integral here so that it will never
592 	 * get higher than 30.  This prevents it from becoming
593 	 * too large an input over long periods of time and allows
594 	 * it to get factored out sooner.
595 	 *
596 	 * The value of 30 was chosen through experimentation.
597 	 */
598 	integral_limit = int_tofp(30);
599 	if (pid->integral > integral_limit)
600 		pid->integral = integral_limit;
601 	if (pid->integral < -integral_limit)
602 		pid->integral = -integral_limit;
603 
604 	dterm = mul_fp(pid->d_gain, fp_error - pid->last_err);
605 	pid->last_err = fp_error;
606 
607 	result = pterm + mul_fp(pid->integral, pid->i_gain) + dterm;
608 	result = result + (1 << (FRAC_BITS-1));
609 	return (signed int)fp_toint(result);
610 }
611 
612 static inline void intel_pstate_busy_pid_reset(struct cpudata *cpu)
613 {
614 	pid_p_gain_set(&cpu->pid, pid_params.p_gain_pct);
615 	pid_d_gain_set(&cpu->pid, pid_params.d_gain_pct);
616 	pid_i_gain_set(&cpu->pid, pid_params.i_gain_pct);
617 
618 	pid_reset(&cpu->pid, pid_params.setpoint, 100, pid_params.deadband, 0);
619 }
620 
621 static inline void intel_pstate_reset_all_pid(void)
622 {
623 	unsigned int cpu;
624 
625 	for_each_online_cpu(cpu) {
626 		if (all_cpu_data[cpu])
627 			intel_pstate_busy_pid_reset(all_cpu_data[cpu]);
628 	}
629 }
630 
631 static inline void update_turbo_state(void)
632 {
633 	u64 misc_en;
634 	struct cpudata *cpu;
635 
636 	cpu = all_cpu_data[0];
637 	rdmsrl(MSR_IA32_MISC_ENABLE, misc_en);
638 	limits->turbo_disabled =
639 		(misc_en & MSR_IA32_MISC_ENABLE_TURBO_DISABLE ||
640 		 cpu->pstate.max_pstate == cpu->pstate.turbo_pstate);
641 }
642 
643 static s16 intel_pstate_get_epb(struct cpudata *cpu_data)
644 {
645 	u64 epb;
646 	int ret;
647 
648 	if (!static_cpu_has(X86_FEATURE_EPB))
649 		return -ENXIO;
650 
651 	ret = rdmsrl_on_cpu(cpu_data->cpu, MSR_IA32_ENERGY_PERF_BIAS, &epb);
652 	if (ret)
653 		return (s16)ret;
654 
655 	return (s16)(epb & 0x0f);
656 }
657 
658 static s16 intel_pstate_get_epp(struct cpudata *cpu_data, u64 hwp_req_data)
659 {
660 	s16 epp;
661 
662 	if (static_cpu_has(X86_FEATURE_HWP_EPP)) {
663 		/*
664 		 * When hwp_req_data is 0, means that caller didn't read
665 		 * MSR_HWP_REQUEST, so need to read and get EPP.
666 		 */
667 		if (!hwp_req_data) {
668 			epp = rdmsrl_on_cpu(cpu_data->cpu, MSR_HWP_REQUEST,
669 					    &hwp_req_data);
670 			if (epp)
671 				return epp;
672 		}
673 		epp = (hwp_req_data >> 24) & 0xff;
674 	} else {
675 		/* When there is no EPP present, HWP uses EPB settings */
676 		epp = intel_pstate_get_epb(cpu_data);
677 	}
678 
679 	return epp;
680 }
681 
682 static int intel_pstate_set_epb(int cpu, s16 pref)
683 {
684 	u64 epb;
685 	int ret;
686 
687 	if (!static_cpu_has(X86_FEATURE_EPB))
688 		return -ENXIO;
689 
690 	ret = rdmsrl_on_cpu(cpu, MSR_IA32_ENERGY_PERF_BIAS, &epb);
691 	if (ret)
692 		return ret;
693 
694 	epb = (epb & ~0x0f) | pref;
695 	wrmsrl_on_cpu(cpu, MSR_IA32_ENERGY_PERF_BIAS, epb);
696 
697 	return 0;
698 }
699 
700 /*
701  * EPP/EPB display strings corresponding to EPP index in the
702  * energy_perf_strings[]
703  *	index		String
704  *-------------------------------------
705  *	0		default
706  *	1		performance
707  *	2		balance_performance
708  *	3		balance_power
709  *	4		power
710  */
711 static const char * const energy_perf_strings[] = {
712 	"default",
713 	"performance",
714 	"balance_performance",
715 	"balance_power",
716 	"power",
717 	NULL
718 };
719 
720 static int intel_pstate_get_energy_pref_index(struct cpudata *cpu_data)
721 {
722 	s16 epp;
723 	int index = -EINVAL;
724 
725 	epp = intel_pstate_get_epp(cpu_data, 0);
726 	if (epp < 0)
727 		return epp;
728 
729 	if (static_cpu_has(X86_FEATURE_HWP_EPP)) {
730 		/*
731 		 * Range:
732 		 *	0x00-0x3F	:	Performance
733 		 *	0x40-0x7F	:	Balance performance
734 		 *	0x80-0xBF	:	Balance power
735 		 *	0xC0-0xFF	:	Power
736 		 * The EPP is a 8 bit value, but our ranges restrict the
737 		 * value which can be set. Here only using top two bits
738 		 * effectively.
739 		 */
740 		index = (epp >> 6) + 1;
741 	} else if (static_cpu_has(X86_FEATURE_EPB)) {
742 		/*
743 		 * Range:
744 		 *	0x00-0x03	:	Performance
745 		 *	0x04-0x07	:	Balance performance
746 		 *	0x08-0x0B	:	Balance power
747 		 *	0x0C-0x0F	:	Power
748 		 * The EPB is a 4 bit value, but our ranges restrict the
749 		 * value which can be set. Here only using top two bits
750 		 * effectively.
751 		 */
752 		index = (epp >> 2) + 1;
753 	}
754 
755 	return index;
756 }
757 
758 static int intel_pstate_set_energy_pref_index(struct cpudata *cpu_data,
759 					      int pref_index)
760 {
761 	int epp = -EINVAL;
762 	int ret;
763 
764 	if (!pref_index)
765 		epp = cpu_data->epp_default;
766 
767 	mutex_lock(&intel_pstate_limits_lock);
768 
769 	if (static_cpu_has(X86_FEATURE_HWP_EPP)) {
770 		u64 value;
771 
772 		ret = rdmsrl_on_cpu(cpu_data->cpu, MSR_HWP_REQUEST, &value);
773 		if (ret)
774 			goto return_pref;
775 
776 		value &= ~GENMASK_ULL(31, 24);
777 
778 		/*
779 		 * If epp is not default, convert from index into
780 		 * energy_perf_strings to epp value, by shifting 6
781 		 * bits left to use only top two bits in epp.
782 		 * The resultant epp need to shifted by 24 bits to
783 		 * epp position in MSR_HWP_REQUEST.
784 		 */
785 		if (epp == -EINVAL)
786 			epp = (pref_index - 1) << 6;
787 
788 		value |= (u64)epp << 24;
789 		ret = wrmsrl_on_cpu(cpu_data->cpu, MSR_HWP_REQUEST, value);
790 	} else {
791 		if (epp == -EINVAL)
792 			epp = (pref_index - 1) << 2;
793 		ret = intel_pstate_set_epb(cpu_data->cpu, epp);
794 	}
795 return_pref:
796 	mutex_unlock(&intel_pstate_limits_lock);
797 
798 	return ret;
799 }
800 
801 static ssize_t show_energy_performance_available_preferences(
802 				struct cpufreq_policy *policy, char *buf)
803 {
804 	int i = 0;
805 	int ret = 0;
806 
807 	while (energy_perf_strings[i] != NULL)
808 		ret += sprintf(&buf[ret], "%s ", energy_perf_strings[i++]);
809 
810 	ret += sprintf(&buf[ret], "\n");
811 
812 	return ret;
813 }
814 
815 cpufreq_freq_attr_ro(energy_performance_available_preferences);
816 
817 static ssize_t store_energy_performance_preference(
818 		struct cpufreq_policy *policy, const char *buf, size_t count)
819 {
820 	struct cpudata *cpu_data = all_cpu_data[policy->cpu];
821 	char str_preference[21];
822 	int ret, i = 0;
823 
824 	ret = sscanf(buf, "%20s", str_preference);
825 	if (ret != 1)
826 		return -EINVAL;
827 
828 	while (energy_perf_strings[i] != NULL) {
829 		if (!strcmp(str_preference, energy_perf_strings[i])) {
830 			intel_pstate_set_energy_pref_index(cpu_data, i);
831 			return count;
832 		}
833 		++i;
834 	}
835 
836 	return -EINVAL;
837 }
838 
839 static ssize_t show_energy_performance_preference(
840 				struct cpufreq_policy *policy, char *buf)
841 {
842 	struct cpudata *cpu_data = all_cpu_data[policy->cpu];
843 	int preference;
844 
845 	preference = intel_pstate_get_energy_pref_index(cpu_data);
846 	if (preference < 0)
847 		return preference;
848 
849 	return  sprintf(buf, "%s\n", energy_perf_strings[preference]);
850 }
851 
852 cpufreq_freq_attr_rw(energy_performance_preference);
853 
854 static struct freq_attr *hwp_cpufreq_attrs[] = {
855 	&energy_performance_preference,
856 	&energy_performance_available_preferences,
857 	NULL,
858 };
859 
860 static void intel_pstate_hwp_set(const struct cpumask *cpumask)
861 {
862 	int min, hw_min, max, hw_max, cpu, range, adj_range;
863 	struct perf_limits *perf_limits = limits;
864 	u64 value, cap;
865 
866 	for_each_cpu(cpu, cpumask) {
867 		int max_perf_pct, min_perf_pct;
868 		struct cpudata *cpu_data = all_cpu_data[cpu];
869 		s16 epp;
870 
871 		if (per_cpu_limits)
872 			perf_limits = all_cpu_data[cpu]->perf_limits;
873 
874 		rdmsrl_on_cpu(cpu, MSR_HWP_CAPABILITIES, &cap);
875 		hw_min = HWP_LOWEST_PERF(cap);
876 		hw_max = HWP_HIGHEST_PERF(cap);
877 		range = hw_max - hw_min;
878 
879 		max_perf_pct = perf_limits->max_perf_pct;
880 		min_perf_pct = perf_limits->min_perf_pct;
881 
882 		rdmsrl_on_cpu(cpu, MSR_HWP_REQUEST, &value);
883 		adj_range = min_perf_pct * range / 100;
884 		min = hw_min + adj_range;
885 		value &= ~HWP_MIN_PERF(~0L);
886 		value |= HWP_MIN_PERF(min);
887 
888 		adj_range = max_perf_pct * range / 100;
889 		max = hw_min + adj_range;
890 		if (limits->no_turbo) {
891 			hw_max = HWP_GUARANTEED_PERF(cap);
892 			if (hw_max < max)
893 				max = hw_max;
894 		}
895 
896 		value &= ~HWP_MAX_PERF(~0L);
897 		value |= HWP_MAX_PERF(max);
898 
899 		if (cpu_data->epp_policy == cpu_data->policy)
900 			goto skip_epp;
901 
902 		cpu_data->epp_policy = cpu_data->policy;
903 
904 		if (cpu_data->epp_saved >= 0) {
905 			epp = cpu_data->epp_saved;
906 			cpu_data->epp_saved = -EINVAL;
907 			goto update_epp;
908 		}
909 
910 		if (cpu_data->policy == CPUFREQ_POLICY_PERFORMANCE) {
911 			epp = intel_pstate_get_epp(cpu_data, value);
912 			cpu_data->epp_powersave = epp;
913 			/* If EPP read was failed, then don't try to write */
914 			if (epp < 0)
915 				goto skip_epp;
916 
917 
918 			epp = 0;
919 		} else {
920 			/* skip setting EPP, when saved value is invalid */
921 			if (cpu_data->epp_powersave < 0)
922 				goto skip_epp;
923 
924 			/*
925 			 * No need to restore EPP when it is not zero. This
926 			 * means:
927 			 *  - Policy is not changed
928 			 *  - user has manually changed
929 			 *  - Error reading EPB
930 			 */
931 			epp = intel_pstate_get_epp(cpu_data, value);
932 			if (epp)
933 				goto skip_epp;
934 
935 			epp = cpu_data->epp_powersave;
936 		}
937 update_epp:
938 		if (static_cpu_has(X86_FEATURE_HWP_EPP)) {
939 			value &= ~GENMASK_ULL(31, 24);
940 			value |= (u64)epp << 24;
941 		} else {
942 			intel_pstate_set_epb(cpu, epp);
943 		}
944 skip_epp:
945 		wrmsrl_on_cpu(cpu, MSR_HWP_REQUEST, value);
946 	}
947 }
948 
949 static int intel_pstate_hwp_set_policy(struct cpufreq_policy *policy)
950 {
951 	if (hwp_active)
952 		intel_pstate_hwp_set(policy->cpus);
953 
954 	return 0;
955 }
956 
957 static int intel_pstate_hwp_save_state(struct cpufreq_policy *policy)
958 {
959 	struct cpudata *cpu_data = all_cpu_data[policy->cpu];
960 
961 	if (!hwp_active)
962 		return 0;
963 
964 	cpu_data->epp_saved = intel_pstate_get_epp(cpu_data, 0);
965 
966 	return 0;
967 }
968 
969 static int intel_pstate_resume(struct cpufreq_policy *policy)
970 {
971 	if (!hwp_active)
972 		return 0;
973 
974 	all_cpu_data[policy->cpu]->epp_policy = 0;
975 
976 	return intel_pstate_hwp_set_policy(policy);
977 }
978 
979 static void intel_pstate_hwp_set_online_cpus(void)
980 {
981 	get_online_cpus();
982 	intel_pstate_hwp_set(cpu_online_mask);
983 	put_online_cpus();
984 }
985 
986 /************************** debugfs begin ************************/
987 static int pid_param_set(void *data, u64 val)
988 {
989 	*(u32 *)data = val;
990 	intel_pstate_reset_all_pid();
991 	return 0;
992 }
993 
994 static int pid_param_get(void *data, u64 *val)
995 {
996 	*val = *(u32 *)data;
997 	return 0;
998 }
999 DEFINE_SIMPLE_ATTRIBUTE(fops_pid_param, pid_param_get, pid_param_set, "%llu\n");
1000 
1001 struct pid_param {
1002 	char *name;
1003 	void *value;
1004 };
1005 
1006 static struct pid_param pid_files[] = {
1007 	{"sample_rate_ms", &pid_params.sample_rate_ms},
1008 	{"d_gain_pct", &pid_params.d_gain_pct},
1009 	{"i_gain_pct", &pid_params.i_gain_pct},
1010 	{"deadband", &pid_params.deadband},
1011 	{"setpoint", &pid_params.setpoint},
1012 	{"p_gain_pct", &pid_params.p_gain_pct},
1013 	{NULL, NULL}
1014 };
1015 
1016 static void __init intel_pstate_debug_expose_params(void)
1017 {
1018 	struct dentry *debugfs_parent;
1019 	int i = 0;
1020 
1021 	if (hwp_active ||
1022 	    pstate_funcs.get_target_pstate == get_target_pstate_use_cpu_load)
1023 		return;
1024 
1025 	debugfs_parent = debugfs_create_dir("pstate_snb", NULL);
1026 	if (IS_ERR_OR_NULL(debugfs_parent))
1027 		return;
1028 	while (pid_files[i].name) {
1029 		debugfs_create_file(pid_files[i].name, 0660,
1030 				    debugfs_parent, pid_files[i].value,
1031 				    &fops_pid_param);
1032 		i++;
1033 	}
1034 }
1035 
1036 /************************** debugfs end ************************/
1037 
1038 /************************** sysfs begin ************************/
1039 #define show_one(file_name, object)					\
1040 	static ssize_t show_##file_name					\
1041 	(struct kobject *kobj, struct attribute *attr, char *buf)	\
1042 	{								\
1043 		return sprintf(buf, "%u\n", limits->object);		\
1044 	}
1045 
1046 static ssize_t show_turbo_pct(struct kobject *kobj,
1047 				struct attribute *attr, char *buf)
1048 {
1049 	struct cpudata *cpu;
1050 	int total, no_turbo, turbo_pct;
1051 	uint32_t turbo_fp;
1052 
1053 	cpu = all_cpu_data[0];
1054 
1055 	total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
1056 	no_turbo = cpu->pstate.max_pstate - cpu->pstate.min_pstate + 1;
1057 	turbo_fp = div_fp(no_turbo, total);
1058 	turbo_pct = 100 - fp_toint(mul_fp(turbo_fp, int_tofp(100)));
1059 	return sprintf(buf, "%u\n", turbo_pct);
1060 }
1061 
1062 static ssize_t show_num_pstates(struct kobject *kobj,
1063 				struct attribute *attr, char *buf)
1064 {
1065 	struct cpudata *cpu;
1066 	int total;
1067 
1068 	cpu = all_cpu_data[0];
1069 	total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
1070 	return sprintf(buf, "%u\n", total);
1071 }
1072 
1073 static ssize_t show_no_turbo(struct kobject *kobj,
1074 			     struct attribute *attr, char *buf)
1075 {
1076 	ssize_t ret;
1077 
1078 	update_turbo_state();
1079 	if (limits->turbo_disabled)
1080 		ret = sprintf(buf, "%u\n", limits->turbo_disabled);
1081 	else
1082 		ret = sprintf(buf, "%u\n", limits->no_turbo);
1083 
1084 	return ret;
1085 }
1086 
1087 static ssize_t store_no_turbo(struct kobject *a, struct attribute *b,
1088 			      const char *buf, size_t count)
1089 {
1090 	unsigned int input;
1091 	int ret;
1092 
1093 	ret = sscanf(buf, "%u", &input);
1094 	if (ret != 1)
1095 		return -EINVAL;
1096 
1097 	mutex_lock(&intel_pstate_limits_lock);
1098 
1099 	update_turbo_state();
1100 	if (limits->turbo_disabled) {
1101 		pr_warn("Turbo disabled by BIOS or unavailable on processor\n");
1102 		mutex_unlock(&intel_pstate_limits_lock);
1103 		return -EPERM;
1104 	}
1105 
1106 	limits->no_turbo = clamp_t(int, input, 0, 1);
1107 
1108 	if (hwp_active)
1109 		intel_pstate_hwp_set_online_cpus();
1110 
1111 	mutex_unlock(&intel_pstate_limits_lock);
1112 
1113 	return count;
1114 }
1115 
1116 static ssize_t store_max_perf_pct(struct kobject *a, struct attribute *b,
1117 				  const char *buf, size_t count)
1118 {
1119 	unsigned int input;
1120 	int ret;
1121 
1122 	ret = sscanf(buf, "%u", &input);
1123 	if (ret != 1)
1124 		return -EINVAL;
1125 
1126 	mutex_lock(&intel_pstate_limits_lock);
1127 
1128 	limits->max_sysfs_pct = clamp_t(int, input, 0 , 100);
1129 	limits->max_perf_pct = min(limits->max_policy_pct,
1130 				   limits->max_sysfs_pct);
1131 	limits->max_perf_pct = max(limits->min_policy_pct,
1132 				   limits->max_perf_pct);
1133 	limits->max_perf_pct = max(limits->min_perf_pct,
1134 				   limits->max_perf_pct);
1135 	limits->max_perf = div_ext_fp(limits->max_perf_pct, 100);
1136 
1137 	if (hwp_active)
1138 		intel_pstate_hwp_set_online_cpus();
1139 
1140 	mutex_unlock(&intel_pstate_limits_lock);
1141 
1142 	return count;
1143 }
1144 
1145 static ssize_t store_min_perf_pct(struct kobject *a, struct attribute *b,
1146 				  const char *buf, size_t count)
1147 {
1148 	unsigned int input;
1149 	int ret;
1150 
1151 	ret = sscanf(buf, "%u", &input);
1152 	if (ret != 1)
1153 		return -EINVAL;
1154 
1155 	mutex_lock(&intel_pstate_limits_lock);
1156 
1157 	limits->min_sysfs_pct = clamp_t(int, input, 0 , 100);
1158 	limits->min_perf_pct = max(limits->min_policy_pct,
1159 				   limits->min_sysfs_pct);
1160 	limits->min_perf_pct = min(limits->max_policy_pct,
1161 				   limits->min_perf_pct);
1162 	limits->min_perf_pct = min(limits->max_perf_pct,
1163 				   limits->min_perf_pct);
1164 	limits->min_perf = div_ext_fp(limits->min_perf_pct, 100);
1165 
1166 	if (hwp_active)
1167 		intel_pstate_hwp_set_online_cpus();
1168 
1169 	mutex_unlock(&intel_pstate_limits_lock);
1170 
1171 	return count;
1172 }
1173 
1174 show_one(max_perf_pct, max_perf_pct);
1175 show_one(min_perf_pct, min_perf_pct);
1176 
1177 define_one_global_rw(no_turbo);
1178 define_one_global_rw(max_perf_pct);
1179 define_one_global_rw(min_perf_pct);
1180 define_one_global_ro(turbo_pct);
1181 define_one_global_ro(num_pstates);
1182 
1183 static struct attribute *intel_pstate_attributes[] = {
1184 	&no_turbo.attr,
1185 	&turbo_pct.attr,
1186 	&num_pstates.attr,
1187 	NULL
1188 };
1189 
1190 static struct attribute_group intel_pstate_attr_group = {
1191 	.attrs = intel_pstate_attributes,
1192 };
1193 
1194 static void __init intel_pstate_sysfs_expose_params(void)
1195 {
1196 	struct kobject *intel_pstate_kobject;
1197 	int rc;
1198 
1199 	intel_pstate_kobject = kobject_create_and_add("intel_pstate",
1200 						&cpu_subsys.dev_root->kobj);
1201 	if (WARN_ON(!intel_pstate_kobject))
1202 		return;
1203 
1204 	rc = sysfs_create_group(intel_pstate_kobject, &intel_pstate_attr_group);
1205 	if (WARN_ON(rc))
1206 		return;
1207 
1208 	/*
1209 	 * If per cpu limits are enforced there are no global limits, so
1210 	 * return without creating max/min_perf_pct attributes
1211 	 */
1212 	if (per_cpu_limits)
1213 		return;
1214 
1215 	rc = sysfs_create_file(intel_pstate_kobject, &max_perf_pct.attr);
1216 	WARN_ON(rc);
1217 
1218 	rc = sysfs_create_file(intel_pstate_kobject, &min_perf_pct.attr);
1219 	WARN_ON(rc);
1220 
1221 }
1222 /************************** sysfs end ************************/
1223 
1224 static void intel_pstate_hwp_enable(struct cpudata *cpudata)
1225 {
1226 	/* First disable HWP notification interrupt as we don't process them */
1227 	if (static_cpu_has(X86_FEATURE_HWP_NOTIFY))
1228 		wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_INTERRUPT, 0x00);
1229 
1230 	wrmsrl_on_cpu(cpudata->cpu, MSR_PM_ENABLE, 0x1);
1231 	cpudata->epp_policy = 0;
1232 	if (cpudata->epp_default == -EINVAL)
1233 		cpudata->epp_default = intel_pstate_get_epp(cpudata, 0);
1234 }
1235 
1236 static int atom_get_min_pstate(void)
1237 {
1238 	u64 value;
1239 
1240 	rdmsrl(ATOM_RATIOS, value);
1241 	return (value >> 8) & 0x7F;
1242 }
1243 
1244 static int atom_get_max_pstate(void)
1245 {
1246 	u64 value;
1247 
1248 	rdmsrl(ATOM_RATIOS, value);
1249 	return (value >> 16) & 0x7F;
1250 }
1251 
1252 static int atom_get_turbo_pstate(void)
1253 {
1254 	u64 value;
1255 
1256 	rdmsrl(ATOM_TURBO_RATIOS, value);
1257 	return value & 0x7F;
1258 }
1259 
1260 static u64 atom_get_val(struct cpudata *cpudata, int pstate)
1261 {
1262 	u64 val;
1263 	int32_t vid_fp;
1264 	u32 vid;
1265 
1266 	val = (u64)pstate << 8;
1267 	if (limits->no_turbo && !limits->turbo_disabled)
1268 		val |= (u64)1 << 32;
1269 
1270 	vid_fp = cpudata->vid.min + mul_fp(
1271 		int_tofp(pstate - cpudata->pstate.min_pstate),
1272 		cpudata->vid.ratio);
1273 
1274 	vid_fp = clamp_t(int32_t, vid_fp, cpudata->vid.min, cpudata->vid.max);
1275 	vid = ceiling_fp(vid_fp);
1276 
1277 	if (pstate > cpudata->pstate.max_pstate)
1278 		vid = cpudata->vid.turbo;
1279 
1280 	return val | vid;
1281 }
1282 
1283 static int silvermont_get_scaling(void)
1284 {
1285 	u64 value;
1286 	int i;
1287 	/* Defined in Table 35-6 from SDM (Sept 2015) */
1288 	static int silvermont_freq_table[] = {
1289 		83300, 100000, 133300, 116700, 80000};
1290 
1291 	rdmsrl(MSR_FSB_FREQ, value);
1292 	i = value & 0x7;
1293 	WARN_ON(i > 4);
1294 
1295 	return silvermont_freq_table[i];
1296 }
1297 
1298 static int airmont_get_scaling(void)
1299 {
1300 	u64 value;
1301 	int i;
1302 	/* Defined in Table 35-10 from SDM (Sept 2015) */
1303 	static int airmont_freq_table[] = {
1304 		83300, 100000, 133300, 116700, 80000,
1305 		93300, 90000, 88900, 87500};
1306 
1307 	rdmsrl(MSR_FSB_FREQ, value);
1308 	i = value & 0xF;
1309 	WARN_ON(i > 8);
1310 
1311 	return airmont_freq_table[i];
1312 }
1313 
1314 static void atom_get_vid(struct cpudata *cpudata)
1315 {
1316 	u64 value;
1317 
1318 	rdmsrl(ATOM_VIDS, value);
1319 	cpudata->vid.min = int_tofp((value >> 8) & 0x7f);
1320 	cpudata->vid.max = int_tofp((value >> 16) & 0x7f);
1321 	cpudata->vid.ratio = div_fp(
1322 		cpudata->vid.max - cpudata->vid.min,
1323 		int_tofp(cpudata->pstate.max_pstate -
1324 			cpudata->pstate.min_pstate));
1325 
1326 	rdmsrl(ATOM_TURBO_VIDS, value);
1327 	cpudata->vid.turbo = value & 0x7f;
1328 }
1329 
1330 static int core_get_min_pstate(void)
1331 {
1332 	u64 value;
1333 
1334 	rdmsrl(MSR_PLATFORM_INFO, value);
1335 	return (value >> 40) & 0xFF;
1336 }
1337 
1338 static int core_get_max_pstate_physical(void)
1339 {
1340 	u64 value;
1341 
1342 	rdmsrl(MSR_PLATFORM_INFO, value);
1343 	return (value >> 8) & 0xFF;
1344 }
1345 
1346 static int core_get_max_pstate(void)
1347 {
1348 	u64 tar;
1349 	u64 plat_info;
1350 	int max_pstate;
1351 	int err;
1352 
1353 	rdmsrl(MSR_PLATFORM_INFO, plat_info);
1354 	max_pstate = (plat_info >> 8) & 0xFF;
1355 
1356 	err = rdmsrl_safe(MSR_TURBO_ACTIVATION_RATIO, &tar);
1357 	if (!err) {
1358 		/* Do some sanity checking for safety */
1359 		if (plat_info & 0x600000000) {
1360 			u64 tdp_ctrl;
1361 			u64 tdp_ratio;
1362 			int tdp_msr;
1363 
1364 			err = rdmsrl_safe(MSR_CONFIG_TDP_CONTROL, &tdp_ctrl);
1365 			if (err)
1366 				goto skip_tar;
1367 
1368 			tdp_msr = MSR_CONFIG_TDP_NOMINAL + (tdp_ctrl & 0x3);
1369 			err = rdmsrl_safe(tdp_msr, &tdp_ratio);
1370 			if (err)
1371 				goto skip_tar;
1372 
1373 			/* For level 1 and 2, bits[23:16] contain the ratio */
1374 			if (tdp_ctrl)
1375 				tdp_ratio >>= 16;
1376 
1377 			tdp_ratio &= 0xff; /* ratios are only 8 bits long */
1378 			if (tdp_ratio - 1 == tar) {
1379 				max_pstate = tar;
1380 				pr_debug("max_pstate=TAC %x\n", max_pstate);
1381 			} else {
1382 				goto skip_tar;
1383 			}
1384 		}
1385 	}
1386 
1387 skip_tar:
1388 	return max_pstate;
1389 }
1390 
1391 static int core_get_turbo_pstate(void)
1392 {
1393 	u64 value;
1394 	int nont, ret;
1395 
1396 	rdmsrl(MSR_TURBO_RATIO_LIMIT, value);
1397 	nont = core_get_max_pstate();
1398 	ret = (value) & 255;
1399 	if (ret <= nont)
1400 		ret = nont;
1401 	return ret;
1402 }
1403 
1404 static inline int core_get_scaling(void)
1405 {
1406 	return 100000;
1407 }
1408 
1409 static u64 core_get_val(struct cpudata *cpudata, int pstate)
1410 {
1411 	u64 val;
1412 
1413 	val = (u64)pstate << 8;
1414 	if (limits->no_turbo && !limits->turbo_disabled)
1415 		val |= (u64)1 << 32;
1416 
1417 	return val;
1418 }
1419 
1420 static int knl_get_turbo_pstate(void)
1421 {
1422 	u64 value;
1423 	int nont, ret;
1424 
1425 	rdmsrl(MSR_TURBO_RATIO_LIMIT, value);
1426 	nont = core_get_max_pstate();
1427 	ret = (((value) >> 8) & 0xFF);
1428 	if (ret <= nont)
1429 		ret = nont;
1430 	return ret;
1431 }
1432 
1433 static struct cpu_defaults core_params = {
1434 	.pid_policy = {
1435 		.sample_rate_ms = 10,
1436 		.deadband = 0,
1437 		.setpoint = 97,
1438 		.p_gain_pct = 20,
1439 		.d_gain_pct = 0,
1440 		.i_gain_pct = 0,
1441 	},
1442 	.funcs = {
1443 		.get_max = core_get_max_pstate,
1444 		.get_max_physical = core_get_max_pstate_physical,
1445 		.get_min = core_get_min_pstate,
1446 		.get_turbo = core_get_turbo_pstate,
1447 		.get_scaling = core_get_scaling,
1448 		.get_val = core_get_val,
1449 		.get_target_pstate = get_target_pstate_use_performance,
1450 	},
1451 };
1452 
1453 static const struct cpu_defaults silvermont_params = {
1454 	.pid_policy = {
1455 		.sample_rate_ms = 10,
1456 		.deadband = 0,
1457 		.setpoint = 60,
1458 		.p_gain_pct = 14,
1459 		.d_gain_pct = 0,
1460 		.i_gain_pct = 4,
1461 	},
1462 	.funcs = {
1463 		.get_max = atom_get_max_pstate,
1464 		.get_max_physical = atom_get_max_pstate,
1465 		.get_min = atom_get_min_pstate,
1466 		.get_turbo = atom_get_turbo_pstate,
1467 		.get_val = atom_get_val,
1468 		.get_scaling = silvermont_get_scaling,
1469 		.get_vid = atom_get_vid,
1470 		.get_target_pstate = get_target_pstate_use_cpu_load,
1471 	},
1472 };
1473 
1474 static const struct cpu_defaults airmont_params = {
1475 	.pid_policy = {
1476 		.sample_rate_ms = 10,
1477 		.deadband = 0,
1478 		.setpoint = 60,
1479 		.p_gain_pct = 14,
1480 		.d_gain_pct = 0,
1481 		.i_gain_pct = 4,
1482 	},
1483 	.funcs = {
1484 		.get_max = atom_get_max_pstate,
1485 		.get_max_physical = atom_get_max_pstate,
1486 		.get_min = atom_get_min_pstate,
1487 		.get_turbo = atom_get_turbo_pstate,
1488 		.get_val = atom_get_val,
1489 		.get_scaling = airmont_get_scaling,
1490 		.get_vid = atom_get_vid,
1491 		.get_target_pstate = get_target_pstate_use_cpu_load,
1492 	},
1493 };
1494 
1495 static const struct cpu_defaults knl_params = {
1496 	.pid_policy = {
1497 		.sample_rate_ms = 10,
1498 		.deadband = 0,
1499 		.setpoint = 97,
1500 		.p_gain_pct = 20,
1501 		.d_gain_pct = 0,
1502 		.i_gain_pct = 0,
1503 	},
1504 	.funcs = {
1505 		.get_max = core_get_max_pstate,
1506 		.get_max_physical = core_get_max_pstate_physical,
1507 		.get_min = core_get_min_pstate,
1508 		.get_turbo = knl_get_turbo_pstate,
1509 		.get_scaling = core_get_scaling,
1510 		.get_val = core_get_val,
1511 		.get_target_pstate = get_target_pstate_use_performance,
1512 	},
1513 };
1514 
1515 static const struct cpu_defaults bxt_params = {
1516 	.pid_policy = {
1517 		.sample_rate_ms = 10,
1518 		.deadband = 0,
1519 		.setpoint = 60,
1520 		.p_gain_pct = 14,
1521 		.d_gain_pct = 0,
1522 		.i_gain_pct = 4,
1523 	},
1524 	.funcs = {
1525 		.get_max = core_get_max_pstate,
1526 		.get_max_physical = core_get_max_pstate_physical,
1527 		.get_min = core_get_min_pstate,
1528 		.get_turbo = core_get_turbo_pstate,
1529 		.get_scaling = core_get_scaling,
1530 		.get_val = core_get_val,
1531 		.get_target_pstate = get_target_pstate_use_cpu_load,
1532 	},
1533 };
1534 
1535 static void intel_pstate_get_min_max(struct cpudata *cpu, int *min, int *max)
1536 {
1537 	int max_perf = cpu->pstate.turbo_pstate;
1538 	int max_perf_adj;
1539 	int min_perf;
1540 	struct perf_limits *perf_limits = limits;
1541 
1542 	if (limits->no_turbo || limits->turbo_disabled)
1543 		max_perf = cpu->pstate.max_pstate;
1544 
1545 	if (per_cpu_limits)
1546 		perf_limits = cpu->perf_limits;
1547 
1548 	/*
1549 	 * performance can be limited by user through sysfs, by cpufreq
1550 	 * policy, or by cpu specific default values determined through
1551 	 * experimentation.
1552 	 */
1553 	max_perf_adj = fp_ext_toint(max_perf * perf_limits->max_perf);
1554 	*max = clamp_t(int, max_perf_adj,
1555 			cpu->pstate.min_pstate, cpu->pstate.turbo_pstate);
1556 
1557 	min_perf = fp_ext_toint(max_perf * perf_limits->min_perf);
1558 	*min = clamp_t(int, min_perf, cpu->pstate.min_pstate, max_perf);
1559 }
1560 
1561 static void intel_pstate_set_pstate(struct cpudata *cpu, int pstate)
1562 {
1563 	trace_cpu_frequency(pstate * cpu->pstate.scaling, cpu->cpu);
1564 	cpu->pstate.current_pstate = pstate;
1565 	/*
1566 	 * Generally, there is no guarantee that this code will always run on
1567 	 * the CPU being updated, so force the register update to run on the
1568 	 * right CPU.
1569 	 */
1570 	wrmsrl_on_cpu(cpu->cpu, MSR_IA32_PERF_CTL,
1571 		      pstate_funcs.get_val(cpu, pstate));
1572 }
1573 
1574 static void intel_pstate_set_min_pstate(struct cpudata *cpu)
1575 {
1576 	intel_pstate_set_pstate(cpu, cpu->pstate.min_pstate);
1577 }
1578 
1579 static void intel_pstate_max_within_limits(struct cpudata *cpu)
1580 {
1581 	int min_pstate, max_pstate;
1582 
1583 	update_turbo_state();
1584 	intel_pstate_get_min_max(cpu, &min_pstate, &max_pstate);
1585 	intel_pstate_set_pstate(cpu, max_pstate);
1586 }
1587 
1588 static void intel_pstate_get_cpu_pstates(struct cpudata *cpu)
1589 {
1590 	cpu->pstate.min_pstate = pstate_funcs.get_min();
1591 	cpu->pstate.max_pstate = pstate_funcs.get_max();
1592 	cpu->pstate.max_pstate_physical = pstate_funcs.get_max_physical();
1593 	cpu->pstate.turbo_pstate = pstate_funcs.get_turbo();
1594 	cpu->pstate.scaling = pstate_funcs.get_scaling();
1595 	cpu->pstate.max_freq = cpu->pstate.max_pstate * cpu->pstate.scaling;
1596 	cpu->pstate.turbo_freq = cpu->pstate.turbo_pstate * cpu->pstate.scaling;
1597 
1598 	if (pstate_funcs.get_vid)
1599 		pstate_funcs.get_vid(cpu);
1600 
1601 	intel_pstate_set_min_pstate(cpu);
1602 }
1603 
1604 static inline void intel_pstate_calc_avg_perf(struct cpudata *cpu)
1605 {
1606 	struct sample *sample = &cpu->sample;
1607 
1608 	sample->core_avg_perf = div_ext_fp(sample->aperf, sample->mperf);
1609 }
1610 
1611 static inline bool intel_pstate_sample(struct cpudata *cpu, u64 time)
1612 {
1613 	u64 aperf, mperf;
1614 	unsigned long flags;
1615 	u64 tsc;
1616 
1617 	local_irq_save(flags);
1618 	rdmsrl(MSR_IA32_APERF, aperf);
1619 	rdmsrl(MSR_IA32_MPERF, mperf);
1620 	tsc = rdtsc();
1621 	if (cpu->prev_mperf == mperf || cpu->prev_tsc == tsc) {
1622 		local_irq_restore(flags);
1623 		return false;
1624 	}
1625 	local_irq_restore(flags);
1626 
1627 	cpu->last_sample_time = cpu->sample.time;
1628 	cpu->sample.time = time;
1629 	cpu->sample.aperf = aperf;
1630 	cpu->sample.mperf = mperf;
1631 	cpu->sample.tsc =  tsc;
1632 	cpu->sample.aperf -= cpu->prev_aperf;
1633 	cpu->sample.mperf -= cpu->prev_mperf;
1634 	cpu->sample.tsc -= cpu->prev_tsc;
1635 
1636 	cpu->prev_aperf = aperf;
1637 	cpu->prev_mperf = mperf;
1638 	cpu->prev_tsc = tsc;
1639 	/*
1640 	 * First time this function is invoked in a given cycle, all of the
1641 	 * previous sample data fields are equal to zero or stale and they must
1642 	 * be populated with meaningful numbers for things to work, so assume
1643 	 * that sample.time will always be reset before setting the utilization
1644 	 * update hook and make the caller skip the sample then.
1645 	 */
1646 	return !!cpu->last_sample_time;
1647 }
1648 
1649 static inline int32_t get_avg_frequency(struct cpudata *cpu)
1650 {
1651 	return mul_ext_fp(cpu->sample.core_avg_perf,
1652 			  cpu->pstate.max_pstate_physical * cpu->pstate.scaling);
1653 }
1654 
1655 static inline int32_t get_avg_pstate(struct cpudata *cpu)
1656 {
1657 	return mul_ext_fp(cpu->pstate.max_pstate_physical,
1658 			  cpu->sample.core_avg_perf);
1659 }
1660 
1661 static inline int32_t get_target_pstate_use_cpu_load(struct cpudata *cpu)
1662 {
1663 	struct sample *sample = &cpu->sample;
1664 	int32_t busy_frac, boost;
1665 	int target, avg_pstate;
1666 
1667 	busy_frac = div_fp(sample->mperf, sample->tsc);
1668 
1669 	boost = cpu->iowait_boost;
1670 	cpu->iowait_boost >>= 1;
1671 
1672 	if (busy_frac < boost)
1673 		busy_frac = boost;
1674 
1675 	sample->busy_scaled = busy_frac * 100;
1676 
1677 	target = limits->no_turbo || limits->turbo_disabled ?
1678 			cpu->pstate.max_pstate : cpu->pstate.turbo_pstate;
1679 	target += target >> 2;
1680 	target = mul_fp(target, busy_frac);
1681 	if (target < cpu->pstate.min_pstate)
1682 		target = cpu->pstate.min_pstate;
1683 
1684 	/*
1685 	 * If the average P-state during the previous cycle was higher than the
1686 	 * current target, add 50% of the difference to the target to reduce
1687 	 * possible performance oscillations and offset possible performance
1688 	 * loss related to moving the workload from one CPU to another within
1689 	 * a package/module.
1690 	 */
1691 	avg_pstate = get_avg_pstate(cpu);
1692 	if (avg_pstate > target)
1693 		target += (avg_pstate - target) >> 1;
1694 
1695 	return target;
1696 }
1697 
1698 static inline int32_t get_target_pstate_use_performance(struct cpudata *cpu)
1699 {
1700 	int32_t perf_scaled, max_pstate, current_pstate, sample_ratio;
1701 	u64 duration_ns;
1702 
1703 	/*
1704 	 * perf_scaled is the ratio of the average P-state during the last
1705 	 * sampling period to the P-state requested last time (in percent).
1706 	 *
1707 	 * That measures the system's response to the previous P-state
1708 	 * selection.
1709 	 */
1710 	max_pstate = cpu->pstate.max_pstate_physical;
1711 	current_pstate = cpu->pstate.current_pstate;
1712 	perf_scaled = mul_ext_fp(cpu->sample.core_avg_perf,
1713 			       div_fp(100 * max_pstate, current_pstate));
1714 
1715 	/*
1716 	 * Since our utilization update callback will not run unless we are
1717 	 * in C0, check if the actual elapsed time is significantly greater (3x)
1718 	 * than our sample interval.  If it is, then we were idle for a long
1719 	 * enough period of time to adjust our performance metric.
1720 	 */
1721 	duration_ns = cpu->sample.time - cpu->last_sample_time;
1722 	if ((s64)duration_ns > pid_params.sample_rate_ns * 3) {
1723 		sample_ratio = div_fp(pid_params.sample_rate_ns, duration_ns);
1724 		perf_scaled = mul_fp(perf_scaled, sample_ratio);
1725 	} else {
1726 		sample_ratio = div_fp(100 * cpu->sample.mperf, cpu->sample.tsc);
1727 		if (sample_ratio < int_tofp(1))
1728 			perf_scaled = 0;
1729 	}
1730 
1731 	cpu->sample.busy_scaled = perf_scaled;
1732 	return cpu->pstate.current_pstate - pid_calc(&cpu->pid, perf_scaled);
1733 }
1734 
1735 static int intel_pstate_prepare_request(struct cpudata *cpu, int pstate)
1736 {
1737 	int max_perf, min_perf;
1738 
1739 	intel_pstate_get_min_max(cpu, &min_perf, &max_perf);
1740 	pstate = clamp_t(int, pstate, min_perf, max_perf);
1741 	trace_cpu_frequency(pstate * cpu->pstate.scaling, cpu->cpu);
1742 	return pstate;
1743 }
1744 
1745 static void intel_pstate_update_pstate(struct cpudata *cpu, int pstate)
1746 {
1747 	pstate = intel_pstate_prepare_request(cpu, pstate);
1748 	if (pstate == cpu->pstate.current_pstate)
1749 		return;
1750 
1751 	cpu->pstate.current_pstate = pstate;
1752 	wrmsrl(MSR_IA32_PERF_CTL, pstate_funcs.get_val(cpu, pstate));
1753 }
1754 
1755 static inline void intel_pstate_adjust_busy_pstate(struct cpudata *cpu)
1756 {
1757 	int from, target_pstate;
1758 	struct sample *sample;
1759 
1760 	from = cpu->pstate.current_pstate;
1761 
1762 	target_pstate = cpu->policy == CPUFREQ_POLICY_PERFORMANCE ?
1763 		cpu->pstate.turbo_pstate : pstate_funcs.get_target_pstate(cpu);
1764 
1765 	update_turbo_state();
1766 
1767 	intel_pstate_update_pstate(cpu, target_pstate);
1768 
1769 	sample = &cpu->sample;
1770 	trace_pstate_sample(mul_ext_fp(100, sample->core_avg_perf),
1771 		fp_toint(sample->busy_scaled),
1772 		from,
1773 		cpu->pstate.current_pstate,
1774 		sample->mperf,
1775 		sample->aperf,
1776 		sample->tsc,
1777 		get_avg_frequency(cpu),
1778 		fp_toint(cpu->iowait_boost * 100));
1779 }
1780 
1781 static void intel_pstate_update_util(struct update_util_data *data, u64 time,
1782 				     unsigned int flags)
1783 {
1784 	struct cpudata *cpu = container_of(data, struct cpudata, update_util);
1785 	u64 delta_ns;
1786 
1787 	if (pstate_funcs.get_target_pstate == get_target_pstate_use_cpu_load) {
1788 		if (flags & SCHED_CPUFREQ_IOWAIT) {
1789 			cpu->iowait_boost = int_tofp(1);
1790 		} else if (cpu->iowait_boost) {
1791 			/* Clear iowait_boost if the CPU may have been idle. */
1792 			delta_ns = time - cpu->last_update;
1793 			if (delta_ns > TICK_NSEC)
1794 				cpu->iowait_boost = 0;
1795 		}
1796 		cpu->last_update = time;
1797 	}
1798 
1799 	delta_ns = time - cpu->sample.time;
1800 	if ((s64)delta_ns >= pid_params.sample_rate_ns) {
1801 		bool sample_taken = intel_pstate_sample(cpu, time);
1802 
1803 		if (sample_taken) {
1804 			intel_pstate_calc_avg_perf(cpu);
1805 			if (!hwp_active)
1806 				intel_pstate_adjust_busy_pstate(cpu);
1807 		}
1808 	}
1809 }
1810 
1811 #define ICPU(model, policy) \
1812 	{ X86_VENDOR_INTEL, 6, model, X86_FEATURE_APERFMPERF,\
1813 			(unsigned long)&policy }
1814 
1815 static const struct x86_cpu_id intel_pstate_cpu_ids[] = {
1816 	ICPU(INTEL_FAM6_SANDYBRIDGE, 		core_params),
1817 	ICPU(INTEL_FAM6_SANDYBRIDGE_X,		core_params),
1818 	ICPU(INTEL_FAM6_ATOM_SILVERMONT1,	silvermont_params),
1819 	ICPU(INTEL_FAM6_IVYBRIDGE,		core_params),
1820 	ICPU(INTEL_FAM6_HASWELL_CORE,		core_params),
1821 	ICPU(INTEL_FAM6_BROADWELL_CORE,		core_params),
1822 	ICPU(INTEL_FAM6_IVYBRIDGE_X,		core_params),
1823 	ICPU(INTEL_FAM6_HASWELL_X,		core_params),
1824 	ICPU(INTEL_FAM6_HASWELL_ULT,		core_params),
1825 	ICPU(INTEL_FAM6_HASWELL_GT3E,		core_params),
1826 	ICPU(INTEL_FAM6_BROADWELL_GT3E,		core_params),
1827 	ICPU(INTEL_FAM6_ATOM_AIRMONT,		airmont_params),
1828 	ICPU(INTEL_FAM6_SKYLAKE_MOBILE,		core_params),
1829 	ICPU(INTEL_FAM6_BROADWELL_X,		core_params),
1830 	ICPU(INTEL_FAM6_SKYLAKE_DESKTOP,	core_params),
1831 	ICPU(INTEL_FAM6_BROADWELL_XEON_D,	core_params),
1832 	ICPU(INTEL_FAM6_XEON_PHI_KNL,		knl_params),
1833 	ICPU(INTEL_FAM6_XEON_PHI_KNM,		knl_params),
1834 	ICPU(INTEL_FAM6_ATOM_GOLDMONT,		bxt_params),
1835 	{}
1836 };
1837 MODULE_DEVICE_TABLE(x86cpu, intel_pstate_cpu_ids);
1838 
1839 static const struct x86_cpu_id intel_pstate_cpu_oob_ids[] __initconst = {
1840 	ICPU(INTEL_FAM6_BROADWELL_XEON_D, core_params),
1841 	ICPU(INTEL_FAM6_BROADWELL_X, core_params),
1842 	ICPU(INTEL_FAM6_SKYLAKE_X, core_params),
1843 	{}
1844 };
1845 
1846 static int intel_pstate_init_cpu(unsigned int cpunum)
1847 {
1848 	struct cpudata *cpu;
1849 
1850 	cpu = all_cpu_data[cpunum];
1851 
1852 	if (!cpu) {
1853 		unsigned int size = sizeof(struct cpudata);
1854 
1855 		if (per_cpu_limits)
1856 			size += sizeof(struct perf_limits);
1857 
1858 		cpu = kzalloc(size, GFP_KERNEL);
1859 		if (!cpu)
1860 			return -ENOMEM;
1861 
1862 		all_cpu_data[cpunum] = cpu;
1863 		if (per_cpu_limits)
1864 			cpu->perf_limits = (struct perf_limits *)(cpu + 1);
1865 
1866 		cpu->epp_default = -EINVAL;
1867 		cpu->epp_powersave = -EINVAL;
1868 		cpu->epp_saved = -EINVAL;
1869 	}
1870 
1871 	cpu = all_cpu_data[cpunum];
1872 
1873 	cpu->cpu = cpunum;
1874 
1875 	if (hwp_active) {
1876 		intel_pstate_hwp_enable(cpu);
1877 		pid_params.sample_rate_ms = 50;
1878 		pid_params.sample_rate_ns = 50 * NSEC_PER_MSEC;
1879 	}
1880 
1881 	intel_pstate_get_cpu_pstates(cpu);
1882 
1883 	intel_pstate_busy_pid_reset(cpu);
1884 
1885 	pr_debug("controlling: cpu %d\n", cpunum);
1886 
1887 	return 0;
1888 }
1889 
1890 static unsigned int intel_pstate_get(unsigned int cpu_num)
1891 {
1892 	struct cpudata *cpu = all_cpu_data[cpu_num];
1893 
1894 	return cpu ? get_avg_frequency(cpu) : 0;
1895 }
1896 
1897 static void intel_pstate_set_update_util_hook(unsigned int cpu_num)
1898 {
1899 	struct cpudata *cpu = all_cpu_data[cpu_num];
1900 
1901 	if (cpu->update_util_set)
1902 		return;
1903 
1904 	/* Prevent intel_pstate_update_util() from using stale data. */
1905 	cpu->sample.time = 0;
1906 	cpufreq_add_update_util_hook(cpu_num, &cpu->update_util,
1907 				     intel_pstate_update_util);
1908 	cpu->update_util_set = true;
1909 }
1910 
1911 static void intel_pstate_clear_update_util_hook(unsigned int cpu)
1912 {
1913 	struct cpudata *cpu_data = all_cpu_data[cpu];
1914 
1915 	if (!cpu_data->update_util_set)
1916 		return;
1917 
1918 	cpufreq_remove_update_util_hook(cpu);
1919 	cpu_data->update_util_set = false;
1920 	synchronize_sched();
1921 }
1922 
1923 static void intel_pstate_set_performance_limits(struct perf_limits *limits)
1924 {
1925 	limits->no_turbo = 0;
1926 	limits->turbo_disabled = 0;
1927 	limits->max_perf_pct = 100;
1928 	limits->max_perf = int_ext_tofp(1);
1929 	limits->min_perf_pct = 100;
1930 	limits->min_perf = int_ext_tofp(1);
1931 	limits->max_policy_pct = 100;
1932 	limits->max_sysfs_pct = 100;
1933 	limits->min_policy_pct = 0;
1934 	limits->min_sysfs_pct = 0;
1935 }
1936 
1937 static void intel_pstate_update_perf_limits(struct cpufreq_policy *policy,
1938 					    struct perf_limits *limits)
1939 {
1940 
1941 	limits->max_policy_pct = DIV_ROUND_UP(policy->max * 100,
1942 					      policy->cpuinfo.max_freq);
1943 	limits->max_policy_pct = clamp_t(int, limits->max_policy_pct, 0, 100);
1944 	if (policy->max == policy->min) {
1945 		limits->min_policy_pct = limits->max_policy_pct;
1946 	} else {
1947 		limits->min_policy_pct = DIV_ROUND_UP(policy->min * 100,
1948 						      policy->cpuinfo.max_freq);
1949 		limits->min_policy_pct = clamp_t(int, limits->min_policy_pct,
1950 						 0, 100);
1951 	}
1952 
1953 	/* Normalize user input to [min_policy_pct, max_policy_pct] */
1954 	limits->min_perf_pct = max(limits->min_policy_pct,
1955 				   limits->min_sysfs_pct);
1956 	limits->min_perf_pct = min(limits->max_policy_pct,
1957 				   limits->min_perf_pct);
1958 	limits->max_perf_pct = min(limits->max_policy_pct,
1959 				   limits->max_sysfs_pct);
1960 	limits->max_perf_pct = max(limits->min_policy_pct,
1961 				   limits->max_perf_pct);
1962 
1963 	/* Make sure min_perf_pct <= max_perf_pct */
1964 	limits->min_perf_pct = min(limits->max_perf_pct, limits->min_perf_pct);
1965 
1966 	limits->min_perf = div_ext_fp(limits->min_perf_pct, 100);
1967 	limits->max_perf = div_ext_fp(limits->max_perf_pct, 100);
1968 	limits->max_perf = round_up(limits->max_perf, EXT_FRAC_BITS);
1969 	limits->min_perf = round_up(limits->min_perf, EXT_FRAC_BITS);
1970 
1971 	pr_debug("cpu:%d max_perf_pct:%d min_perf_pct:%d\n", policy->cpu,
1972 		 limits->max_perf_pct, limits->min_perf_pct);
1973 }
1974 
1975 static int intel_pstate_set_policy(struct cpufreq_policy *policy)
1976 {
1977 	struct cpudata *cpu;
1978 	struct perf_limits *perf_limits = NULL;
1979 
1980 	if (!policy->cpuinfo.max_freq)
1981 		return -ENODEV;
1982 
1983 	pr_debug("set_policy cpuinfo.max %u policy->max %u\n",
1984 		 policy->cpuinfo.max_freq, policy->max);
1985 
1986 	cpu = all_cpu_data[policy->cpu];
1987 	cpu->policy = policy->policy;
1988 
1989 	if (cpu->pstate.max_pstate_physical > cpu->pstate.max_pstate &&
1990 	    policy->max < policy->cpuinfo.max_freq &&
1991 	    policy->max > cpu->pstate.max_pstate * cpu->pstate.scaling) {
1992 		pr_debug("policy->max > max non turbo frequency\n");
1993 		policy->max = policy->cpuinfo.max_freq;
1994 	}
1995 
1996 	if (per_cpu_limits)
1997 		perf_limits = cpu->perf_limits;
1998 
1999 	mutex_lock(&intel_pstate_limits_lock);
2000 
2001 	if (policy->policy == CPUFREQ_POLICY_PERFORMANCE) {
2002 		if (!perf_limits) {
2003 			limits = &performance_limits;
2004 			perf_limits = limits;
2005 		}
2006 		if (policy->max >= policy->cpuinfo.max_freq) {
2007 			pr_debug("set performance\n");
2008 			intel_pstate_set_performance_limits(perf_limits);
2009 			goto out;
2010 		}
2011 	} else {
2012 		pr_debug("set powersave\n");
2013 		if (!perf_limits) {
2014 			limits = &powersave_limits;
2015 			perf_limits = limits;
2016 		}
2017 
2018 	}
2019 
2020 	intel_pstate_update_perf_limits(policy, perf_limits);
2021  out:
2022 	if (cpu->policy == CPUFREQ_POLICY_PERFORMANCE) {
2023 		/*
2024 		 * NOHZ_FULL CPUs need this as the governor callback may not
2025 		 * be invoked on them.
2026 		 */
2027 		intel_pstate_clear_update_util_hook(policy->cpu);
2028 		intel_pstate_max_within_limits(cpu);
2029 	}
2030 
2031 	intel_pstate_set_update_util_hook(policy->cpu);
2032 
2033 	intel_pstate_hwp_set_policy(policy);
2034 
2035 	mutex_unlock(&intel_pstate_limits_lock);
2036 
2037 	return 0;
2038 }
2039 
2040 static int intel_pstate_verify_policy(struct cpufreq_policy *policy)
2041 {
2042 	cpufreq_verify_within_cpu_limits(policy);
2043 
2044 	if (policy->policy != CPUFREQ_POLICY_POWERSAVE &&
2045 	    policy->policy != CPUFREQ_POLICY_PERFORMANCE)
2046 		return -EINVAL;
2047 
2048 	return 0;
2049 }
2050 
2051 static void intel_cpufreq_stop_cpu(struct cpufreq_policy *policy)
2052 {
2053 	intel_pstate_set_min_pstate(all_cpu_data[policy->cpu]);
2054 }
2055 
2056 static void intel_pstate_stop_cpu(struct cpufreq_policy *policy)
2057 {
2058 	pr_debug("CPU %d exiting\n", policy->cpu);
2059 
2060 	intel_pstate_clear_update_util_hook(policy->cpu);
2061 	if (hwp_active)
2062 		intel_pstate_hwp_save_state(policy);
2063 	else
2064 		intel_cpufreq_stop_cpu(policy);
2065 }
2066 
2067 static int intel_pstate_cpu_exit(struct cpufreq_policy *policy)
2068 {
2069 	intel_pstate_exit_perf_limits(policy);
2070 
2071 	policy->fast_switch_possible = false;
2072 
2073 	return 0;
2074 }
2075 
2076 static int __intel_pstate_cpu_init(struct cpufreq_policy *policy)
2077 {
2078 	struct cpudata *cpu;
2079 	int rc;
2080 
2081 	rc = intel_pstate_init_cpu(policy->cpu);
2082 	if (rc)
2083 		return rc;
2084 
2085 	cpu = all_cpu_data[policy->cpu];
2086 
2087 	/*
2088 	 * We need sane value in the cpu->perf_limits, so inherit from global
2089 	 * perf_limits limits, which are seeded with values based on the
2090 	 * CONFIG_CPU_FREQ_DEFAULT_GOV_*, during boot up.
2091 	 */
2092 	if (per_cpu_limits)
2093 		memcpy(cpu->perf_limits, limits, sizeof(struct perf_limits));
2094 
2095 	policy->min = cpu->pstate.min_pstate * cpu->pstate.scaling;
2096 	policy->max = cpu->pstate.turbo_pstate * cpu->pstate.scaling;
2097 
2098 	/* cpuinfo and default policy values */
2099 	policy->cpuinfo.min_freq = cpu->pstate.min_pstate * cpu->pstate.scaling;
2100 	update_turbo_state();
2101 	policy->cpuinfo.max_freq = limits->turbo_disabled ?
2102 			cpu->pstate.max_pstate : cpu->pstate.turbo_pstate;
2103 	policy->cpuinfo.max_freq *= cpu->pstate.scaling;
2104 
2105 	intel_pstate_init_acpi_perf_limits(policy);
2106 	cpumask_set_cpu(policy->cpu, policy->cpus);
2107 
2108 	policy->fast_switch_possible = true;
2109 
2110 	return 0;
2111 }
2112 
2113 static int intel_pstate_cpu_init(struct cpufreq_policy *policy)
2114 {
2115 	int ret = __intel_pstate_cpu_init(policy);
2116 
2117 	if (ret)
2118 		return ret;
2119 
2120 	policy->cpuinfo.transition_latency = CPUFREQ_ETERNAL;
2121 	if (limits->min_perf_pct == 100 && limits->max_perf_pct == 100)
2122 		policy->policy = CPUFREQ_POLICY_PERFORMANCE;
2123 	else
2124 		policy->policy = CPUFREQ_POLICY_POWERSAVE;
2125 
2126 	return 0;
2127 }
2128 
2129 static struct cpufreq_driver intel_pstate = {
2130 	.flags		= CPUFREQ_CONST_LOOPS,
2131 	.verify		= intel_pstate_verify_policy,
2132 	.setpolicy	= intel_pstate_set_policy,
2133 	.suspend	= intel_pstate_hwp_save_state,
2134 	.resume		= intel_pstate_resume,
2135 	.get		= intel_pstate_get,
2136 	.init		= intel_pstate_cpu_init,
2137 	.exit		= intel_pstate_cpu_exit,
2138 	.stop_cpu	= intel_pstate_stop_cpu,
2139 	.name		= "intel_pstate",
2140 };
2141 
2142 static int intel_cpufreq_verify_policy(struct cpufreq_policy *policy)
2143 {
2144 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2145 	struct perf_limits *perf_limits = limits;
2146 
2147 	update_turbo_state();
2148 	policy->cpuinfo.max_freq = limits->turbo_disabled ?
2149 			cpu->pstate.max_freq : cpu->pstate.turbo_freq;
2150 
2151 	cpufreq_verify_within_cpu_limits(policy);
2152 
2153 	if (per_cpu_limits)
2154 		perf_limits = cpu->perf_limits;
2155 
2156 	intel_pstate_update_perf_limits(policy, perf_limits);
2157 
2158 	return 0;
2159 }
2160 
2161 static unsigned int intel_cpufreq_turbo_update(struct cpudata *cpu,
2162 					       struct cpufreq_policy *policy,
2163 					       unsigned int target_freq)
2164 {
2165 	unsigned int max_freq;
2166 
2167 	update_turbo_state();
2168 
2169 	max_freq = limits->no_turbo || limits->turbo_disabled ?
2170 			cpu->pstate.max_freq : cpu->pstate.turbo_freq;
2171 	policy->cpuinfo.max_freq = max_freq;
2172 	if (policy->max > max_freq)
2173 		policy->max = max_freq;
2174 
2175 	if (target_freq > max_freq)
2176 		target_freq = max_freq;
2177 
2178 	return target_freq;
2179 }
2180 
2181 static int intel_cpufreq_target(struct cpufreq_policy *policy,
2182 				unsigned int target_freq,
2183 				unsigned int relation)
2184 {
2185 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2186 	struct cpufreq_freqs freqs;
2187 	int target_pstate;
2188 
2189 	freqs.old = policy->cur;
2190 	freqs.new = intel_cpufreq_turbo_update(cpu, policy, target_freq);
2191 
2192 	cpufreq_freq_transition_begin(policy, &freqs);
2193 	switch (relation) {
2194 	case CPUFREQ_RELATION_L:
2195 		target_pstate = DIV_ROUND_UP(freqs.new, cpu->pstate.scaling);
2196 		break;
2197 	case CPUFREQ_RELATION_H:
2198 		target_pstate = freqs.new / cpu->pstate.scaling;
2199 		break;
2200 	default:
2201 		target_pstate = DIV_ROUND_CLOSEST(freqs.new, cpu->pstate.scaling);
2202 		break;
2203 	}
2204 	target_pstate = intel_pstate_prepare_request(cpu, target_pstate);
2205 	if (target_pstate != cpu->pstate.current_pstate) {
2206 		cpu->pstate.current_pstate = target_pstate;
2207 		wrmsrl_on_cpu(policy->cpu, MSR_IA32_PERF_CTL,
2208 			      pstate_funcs.get_val(cpu, target_pstate));
2209 	}
2210 	cpufreq_freq_transition_end(policy, &freqs, false);
2211 
2212 	return 0;
2213 }
2214 
2215 static unsigned int intel_cpufreq_fast_switch(struct cpufreq_policy *policy,
2216 					      unsigned int target_freq)
2217 {
2218 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2219 	int target_pstate;
2220 
2221 	target_freq = intel_cpufreq_turbo_update(cpu, policy, target_freq);
2222 	target_pstate = DIV_ROUND_UP(target_freq, cpu->pstate.scaling);
2223 	intel_pstate_update_pstate(cpu, target_pstate);
2224 	return target_freq;
2225 }
2226 
2227 static int intel_cpufreq_cpu_init(struct cpufreq_policy *policy)
2228 {
2229 	int ret = __intel_pstate_cpu_init(policy);
2230 
2231 	if (ret)
2232 		return ret;
2233 
2234 	policy->cpuinfo.transition_latency = INTEL_CPUFREQ_TRANSITION_LATENCY;
2235 	/* This reflects the intel_pstate_get_cpu_pstates() setting. */
2236 	policy->cur = policy->cpuinfo.min_freq;
2237 
2238 	return 0;
2239 }
2240 
2241 static struct cpufreq_driver intel_cpufreq = {
2242 	.flags		= CPUFREQ_CONST_LOOPS,
2243 	.verify		= intel_cpufreq_verify_policy,
2244 	.target		= intel_cpufreq_target,
2245 	.fast_switch	= intel_cpufreq_fast_switch,
2246 	.init		= intel_cpufreq_cpu_init,
2247 	.exit		= intel_pstate_cpu_exit,
2248 	.stop_cpu	= intel_cpufreq_stop_cpu,
2249 	.name		= "intel_cpufreq",
2250 };
2251 
2252 static struct cpufreq_driver *intel_pstate_driver = &intel_pstate;
2253 
2254 static int no_load __initdata;
2255 static int no_hwp __initdata;
2256 static int hwp_only __initdata;
2257 static unsigned int force_load __initdata;
2258 
2259 static int __init intel_pstate_msrs_not_valid(void)
2260 {
2261 	if (!pstate_funcs.get_max() ||
2262 	    !pstate_funcs.get_min() ||
2263 	    !pstate_funcs.get_turbo())
2264 		return -ENODEV;
2265 
2266 	return 0;
2267 }
2268 
2269 static void __init copy_pid_params(struct pstate_adjust_policy *policy)
2270 {
2271 	pid_params.sample_rate_ms = policy->sample_rate_ms;
2272 	pid_params.sample_rate_ns = pid_params.sample_rate_ms * NSEC_PER_MSEC;
2273 	pid_params.p_gain_pct = policy->p_gain_pct;
2274 	pid_params.i_gain_pct = policy->i_gain_pct;
2275 	pid_params.d_gain_pct = policy->d_gain_pct;
2276 	pid_params.deadband = policy->deadband;
2277 	pid_params.setpoint = policy->setpoint;
2278 }
2279 
2280 #ifdef CONFIG_ACPI
2281 static void intel_pstate_use_acpi_profile(void)
2282 {
2283 	if (acpi_gbl_FADT.preferred_profile == PM_MOBILE)
2284 		pstate_funcs.get_target_pstate =
2285 				get_target_pstate_use_cpu_load;
2286 }
2287 #else
2288 static void intel_pstate_use_acpi_profile(void)
2289 {
2290 }
2291 #endif
2292 
2293 static void __init copy_cpu_funcs(struct pstate_funcs *funcs)
2294 {
2295 	pstate_funcs.get_max   = funcs->get_max;
2296 	pstate_funcs.get_max_physical = funcs->get_max_physical;
2297 	pstate_funcs.get_min   = funcs->get_min;
2298 	pstate_funcs.get_turbo = funcs->get_turbo;
2299 	pstate_funcs.get_scaling = funcs->get_scaling;
2300 	pstate_funcs.get_val   = funcs->get_val;
2301 	pstate_funcs.get_vid   = funcs->get_vid;
2302 	pstate_funcs.get_target_pstate = funcs->get_target_pstate;
2303 
2304 	intel_pstate_use_acpi_profile();
2305 }
2306 
2307 #ifdef CONFIG_ACPI
2308 
2309 static bool __init intel_pstate_no_acpi_pss(void)
2310 {
2311 	int i;
2312 
2313 	for_each_possible_cpu(i) {
2314 		acpi_status status;
2315 		union acpi_object *pss;
2316 		struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
2317 		struct acpi_processor *pr = per_cpu(processors, i);
2318 
2319 		if (!pr)
2320 			continue;
2321 
2322 		status = acpi_evaluate_object(pr->handle, "_PSS", NULL, &buffer);
2323 		if (ACPI_FAILURE(status))
2324 			continue;
2325 
2326 		pss = buffer.pointer;
2327 		if (pss && pss->type == ACPI_TYPE_PACKAGE) {
2328 			kfree(pss);
2329 			return false;
2330 		}
2331 
2332 		kfree(pss);
2333 	}
2334 
2335 	return true;
2336 }
2337 
2338 static bool __init intel_pstate_has_acpi_ppc(void)
2339 {
2340 	int i;
2341 
2342 	for_each_possible_cpu(i) {
2343 		struct acpi_processor *pr = per_cpu(processors, i);
2344 
2345 		if (!pr)
2346 			continue;
2347 		if (acpi_has_method(pr->handle, "_PPC"))
2348 			return true;
2349 	}
2350 	return false;
2351 }
2352 
2353 enum {
2354 	PSS,
2355 	PPC,
2356 };
2357 
2358 struct hw_vendor_info {
2359 	u16  valid;
2360 	char oem_id[ACPI_OEM_ID_SIZE];
2361 	char oem_table_id[ACPI_OEM_TABLE_ID_SIZE];
2362 	int  oem_pwr_table;
2363 };
2364 
2365 /* Hardware vendor-specific info that has its own power management modes */
2366 static struct hw_vendor_info vendor_info[] __initdata = {
2367 	{1, "HP    ", "ProLiant", PSS},
2368 	{1, "ORACLE", "X4-2    ", PPC},
2369 	{1, "ORACLE", "X4-2L   ", PPC},
2370 	{1, "ORACLE", "X4-2B   ", PPC},
2371 	{1, "ORACLE", "X3-2    ", PPC},
2372 	{1, "ORACLE", "X3-2L   ", PPC},
2373 	{1, "ORACLE", "X3-2B   ", PPC},
2374 	{1, "ORACLE", "X4470M2 ", PPC},
2375 	{1, "ORACLE", "X4270M3 ", PPC},
2376 	{1, "ORACLE", "X4270M2 ", PPC},
2377 	{1, "ORACLE", "X4170M2 ", PPC},
2378 	{1, "ORACLE", "X4170 M3", PPC},
2379 	{1, "ORACLE", "X4275 M3", PPC},
2380 	{1, "ORACLE", "X6-2    ", PPC},
2381 	{1, "ORACLE", "Sudbury ", PPC},
2382 	{0, "", ""},
2383 };
2384 
2385 static bool __init intel_pstate_platform_pwr_mgmt_exists(void)
2386 {
2387 	struct acpi_table_header hdr;
2388 	struct hw_vendor_info *v_info;
2389 	const struct x86_cpu_id *id;
2390 	u64 misc_pwr;
2391 
2392 	id = x86_match_cpu(intel_pstate_cpu_oob_ids);
2393 	if (id) {
2394 		rdmsrl(MSR_MISC_PWR_MGMT, misc_pwr);
2395 		if ( misc_pwr & (1 << 8))
2396 			return true;
2397 	}
2398 
2399 	if (acpi_disabled ||
2400 	    ACPI_FAILURE(acpi_get_table_header(ACPI_SIG_FADT, 0, &hdr)))
2401 		return false;
2402 
2403 	for (v_info = vendor_info; v_info->valid; v_info++) {
2404 		if (!strncmp(hdr.oem_id, v_info->oem_id, ACPI_OEM_ID_SIZE) &&
2405 			!strncmp(hdr.oem_table_id, v_info->oem_table_id,
2406 						ACPI_OEM_TABLE_ID_SIZE))
2407 			switch (v_info->oem_pwr_table) {
2408 			case PSS:
2409 				return intel_pstate_no_acpi_pss();
2410 			case PPC:
2411 				return intel_pstate_has_acpi_ppc() &&
2412 					(!force_load);
2413 			}
2414 	}
2415 
2416 	return false;
2417 }
2418 
2419 static void intel_pstate_request_control_from_smm(void)
2420 {
2421 	/*
2422 	 * It may be unsafe to request P-states control from SMM if _PPC support
2423 	 * has not been enabled.
2424 	 */
2425 	if (acpi_ppc)
2426 		acpi_processor_pstate_control();
2427 }
2428 #else /* CONFIG_ACPI not enabled */
2429 static inline bool intel_pstate_platform_pwr_mgmt_exists(void) { return false; }
2430 static inline bool intel_pstate_has_acpi_ppc(void) { return false; }
2431 static inline void intel_pstate_request_control_from_smm(void) {}
2432 #endif /* CONFIG_ACPI */
2433 
2434 static const struct x86_cpu_id hwp_support_ids[] __initconst = {
2435 	{ X86_VENDOR_INTEL, 6, X86_MODEL_ANY, X86_FEATURE_HWP },
2436 	{}
2437 };
2438 
2439 static int __init intel_pstate_init(void)
2440 {
2441 	int cpu, rc = 0;
2442 	const struct x86_cpu_id *id;
2443 	struct cpu_defaults *cpu_def;
2444 
2445 	if (no_load)
2446 		return -ENODEV;
2447 
2448 	if (x86_match_cpu(hwp_support_ids) && !no_hwp) {
2449 		copy_cpu_funcs(&core_params.funcs);
2450 		hwp_active++;
2451 		intel_pstate.attr = hwp_cpufreq_attrs;
2452 		goto hwp_cpu_matched;
2453 	}
2454 
2455 	id = x86_match_cpu(intel_pstate_cpu_ids);
2456 	if (!id)
2457 		return -ENODEV;
2458 
2459 	cpu_def = (struct cpu_defaults *)id->driver_data;
2460 
2461 	copy_pid_params(&cpu_def->pid_policy);
2462 	copy_cpu_funcs(&cpu_def->funcs);
2463 
2464 	if (intel_pstate_msrs_not_valid())
2465 		return -ENODEV;
2466 
2467 hwp_cpu_matched:
2468 	/*
2469 	 * The Intel pstate driver will be ignored if the platform
2470 	 * firmware has its own power management modes.
2471 	 */
2472 	if (intel_pstate_platform_pwr_mgmt_exists())
2473 		return -ENODEV;
2474 
2475 	pr_info("Intel P-state driver initializing\n");
2476 
2477 	all_cpu_data = vzalloc(sizeof(void *) * num_possible_cpus());
2478 	if (!all_cpu_data)
2479 		return -ENOMEM;
2480 
2481 	if (!hwp_active && hwp_only)
2482 		goto out;
2483 
2484 	intel_pstate_request_control_from_smm();
2485 
2486 	rc = cpufreq_register_driver(intel_pstate_driver);
2487 	if (rc)
2488 		goto out;
2489 
2490 	intel_pstate_debug_expose_params();
2491 	intel_pstate_sysfs_expose_params();
2492 
2493 	if (hwp_active)
2494 		pr_info("HWP enabled\n");
2495 
2496 	return rc;
2497 out:
2498 	get_online_cpus();
2499 	for_each_online_cpu(cpu) {
2500 		if (all_cpu_data[cpu]) {
2501 			if (intel_pstate_driver == &intel_pstate)
2502 				intel_pstate_clear_update_util_hook(cpu);
2503 
2504 			kfree(all_cpu_data[cpu]);
2505 		}
2506 	}
2507 
2508 	put_online_cpus();
2509 	vfree(all_cpu_data);
2510 	return -ENODEV;
2511 }
2512 device_initcall(intel_pstate_init);
2513 
2514 static int __init intel_pstate_setup(char *str)
2515 {
2516 	if (!str)
2517 		return -EINVAL;
2518 
2519 	if (!strcmp(str, "disable")) {
2520 		no_load = 1;
2521 	} else if (!strcmp(str, "passive")) {
2522 		pr_info("Passive mode enabled\n");
2523 		intel_pstate_driver = &intel_cpufreq;
2524 		no_hwp = 1;
2525 	}
2526 	if (!strcmp(str, "no_hwp")) {
2527 		pr_info("HWP disabled\n");
2528 		no_hwp = 1;
2529 	}
2530 	if (!strcmp(str, "force"))
2531 		force_load = 1;
2532 	if (!strcmp(str, "hwp_only"))
2533 		hwp_only = 1;
2534 	if (!strcmp(str, "per_cpu_perf_limits"))
2535 		per_cpu_limits = true;
2536 
2537 #ifdef CONFIG_ACPI
2538 	if (!strcmp(str, "support_acpi_ppc"))
2539 		acpi_ppc = true;
2540 #endif
2541 
2542 	return 0;
2543 }
2544 early_param("intel_pstate", intel_pstate_setup);
2545 
2546 MODULE_AUTHOR("Dirk Brandewie <dirk.j.brandewie@intel.com>");
2547 MODULE_DESCRIPTION("'intel_pstate' - P state driver Intel Core processors");
2548 MODULE_LICENSE("GPL");
2549