xref: /linux/drivers/cpufreq/intel_pstate.c (revision b4553f0cfea5ab5a02967e482bcafe7db6407afd)
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/cpufreq.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 #ifdef CONFIG_ACPI
43 #include <acpi/processor.h>
44 #include <acpi/cppc_acpi.h>
45 #endif
46 
47 #define FRAC_BITS 8
48 #define int_tofp(X) ((int64_t)(X) << FRAC_BITS)
49 #define fp_toint(X) ((X) >> FRAC_BITS)
50 
51 #define EXT_BITS 6
52 #define EXT_FRAC_BITS (EXT_BITS + FRAC_BITS)
53 #define fp_ext_toint(X) ((X) >> EXT_FRAC_BITS)
54 #define int_ext_tofp(X) ((int64_t)(X) << EXT_FRAC_BITS)
55 
56 static inline int32_t mul_fp(int32_t x, int32_t y)
57 {
58 	return ((int64_t)x * (int64_t)y) >> FRAC_BITS;
59 }
60 
61 static inline int32_t div_fp(s64 x, s64 y)
62 {
63 	return div64_s64((int64_t)x << FRAC_BITS, y);
64 }
65 
66 static inline int ceiling_fp(int32_t x)
67 {
68 	int mask, ret;
69 
70 	ret = fp_toint(x);
71 	mask = (1 << FRAC_BITS) - 1;
72 	if (x & mask)
73 		ret += 1;
74 	return ret;
75 }
76 
77 static inline u64 mul_ext_fp(u64 x, u64 y)
78 {
79 	return (x * y) >> EXT_FRAC_BITS;
80 }
81 
82 static inline u64 div_ext_fp(u64 x, u64 y)
83 {
84 	return div64_u64(x << EXT_FRAC_BITS, y);
85 }
86 
87 /**
88  * struct sample -	Store performance sample
89  * @core_avg_perf:	Ratio of APERF/MPERF which is the actual average
90  *			performance during last sample period
91  * @busy_scaled:	Scaled busy value which is used to calculate next
92  *			P state. This can be different than core_avg_perf
93  *			to account for cpu idle period
94  * @aperf:		Difference of actual performance frequency clock count
95  *			read from APERF MSR between last and current sample
96  * @mperf:		Difference of maximum performance frequency clock count
97  *			read from MPERF MSR between last and current sample
98  * @tsc:		Difference of time stamp counter between last and
99  *			current sample
100  * @time:		Current time from scheduler
101  *
102  * This structure is used in the cpudata structure to store performance sample
103  * data for choosing next P State.
104  */
105 struct sample {
106 	int32_t core_avg_perf;
107 	int32_t busy_scaled;
108 	u64 aperf;
109 	u64 mperf;
110 	u64 tsc;
111 	u64 time;
112 };
113 
114 /**
115  * struct pstate_data - Store P state data
116  * @current_pstate:	Current requested P state
117  * @min_pstate:		Min P state possible for this platform
118  * @max_pstate:		Max P state possible for this platform
119  * @max_pstate_physical:This is physical Max P state for a processor
120  *			This can be higher than the max_pstate which can
121  *			be limited by platform thermal design power limits
122  * @scaling:		Scaling factor to  convert frequency to cpufreq
123  *			frequency units
124  * @turbo_pstate:	Max Turbo P state possible for this platform
125  * @max_freq:		@max_pstate frequency in cpufreq units
126  * @turbo_freq:		@turbo_pstate frequency in cpufreq units
127  *
128  * Stores the per cpu model P state limits and current P state.
129  */
130 struct pstate_data {
131 	int	current_pstate;
132 	int	min_pstate;
133 	int	max_pstate;
134 	int	max_pstate_physical;
135 	int	scaling;
136 	int	turbo_pstate;
137 	unsigned int max_freq;
138 	unsigned int turbo_freq;
139 };
140 
141 /**
142  * struct vid_data -	Stores voltage information data
143  * @min:		VID data for this platform corresponding to
144  *			the lowest P state
145  * @max:		VID data corresponding to the highest P State.
146  * @turbo:		VID data for turbo P state
147  * @ratio:		Ratio of (vid max - vid min) /
148  *			(max P state - Min P State)
149  *
150  * Stores the voltage data for DVFS (Dynamic Voltage and Frequency Scaling)
151  * This data is used in Atom platforms, where in addition to target P state,
152  * the voltage data needs to be specified to select next P State.
153  */
154 struct vid_data {
155 	int min;
156 	int max;
157 	int turbo;
158 	int32_t ratio;
159 };
160 
161 /**
162  * struct _pid -	Stores PID data
163  * @setpoint:		Target set point for busyness or performance
164  * @integral:		Storage for accumulated error values
165  * @p_gain:		PID proportional gain
166  * @i_gain:		PID integral gain
167  * @d_gain:		PID derivative gain
168  * @deadband:		PID deadband
169  * @last_err:		Last error storage for integral part of PID calculation
170  *
171  * Stores PID coefficients and last error for PID controller.
172  */
173 struct _pid {
174 	int setpoint;
175 	int32_t integral;
176 	int32_t p_gain;
177 	int32_t i_gain;
178 	int32_t d_gain;
179 	int deadband;
180 	int32_t last_err;
181 };
182 
183 /**
184  * struct perf_limits - Store user and policy limits
185  * @no_turbo:		User requested turbo state from intel_pstate sysfs
186  * @turbo_disabled:	Platform turbo status either from msr
187  *			MSR_IA32_MISC_ENABLE or when maximum available pstate
188  *			matches the maximum turbo pstate
189  * @max_perf_pct:	Effective maximum performance limit in percentage, this
190  *			is minimum of either limits enforced by cpufreq policy
191  *			or limits from user set limits via intel_pstate sysfs
192  * @min_perf_pct:	Effective minimum performance limit in percentage, this
193  *			is maximum of either limits enforced by cpufreq policy
194  *			or limits from user set limits via intel_pstate sysfs
195  * @max_perf:		This is a scaled value between 0 to 255 for max_perf_pct
196  *			This value is used to limit max pstate
197  * @min_perf:		This is a scaled value between 0 to 255 for min_perf_pct
198  *			This value is used to limit min pstate
199  * @max_policy_pct:	The maximum performance in percentage enforced by
200  *			cpufreq setpolicy interface
201  * @max_sysfs_pct:	The maximum performance in percentage enforced by
202  *			intel pstate sysfs interface, unused when per cpu
203  *			controls are enforced
204  * @min_policy_pct:	The minimum performance in percentage enforced by
205  *			cpufreq setpolicy interface
206  * @min_sysfs_pct:	The minimum performance in percentage enforced by
207  *			intel pstate sysfs interface, unused when per cpu
208  *			controls are enforced
209  *
210  * Storage for user and policy defined limits.
211  */
212 struct perf_limits {
213 	int no_turbo;
214 	int turbo_disabled;
215 	int max_perf_pct;
216 	int min_perf_pct;
217 	int32_t max_perf;
218 	int32_t min_perf;
219 	int max_policy_pct;
220 	int max_sysfs_pct;
221 	int min_policy_pct;
222 	int min_sysfs_pct;
223 };
224 
225 /**
226  * struct cpudata -	Per CPU instance data storage
227  * @cpu:		CPU number for this instance data
228  * @policy:		CPUFreq policy value
229  * @update_util:	CPUFreq utility callback information
230  * @update_util_set:	CPUFreq utility callback is set
231  * @iowait_boost:	iowait-related boost fraction
232  * @last_update:	Time of the last update.
233  * @pstate:		Stores P state limits for this CPU
234  * @vid:		Stores VID limits for this CPU
235  * @pid:		Stores PID parameters for this CPU
236  * @last_sample_time:	Last Sample time
237  * @prev_aperf:		Last APERF value read from APERF MSR
238  * @prev_mperf:		Last MPERF value read from MPERF MSR
239  * @prev_tsc:		Last timestamp counter (TSC) value
240  * @prev_cummulative_iowait: IO Wait time difference from last and
241  *			current sample
242  * @sample:		Storage for storing last Sample data
243  * @perf_limits:	Pointer to perf_limit unique to this CPU
244  *			Not all field in the structure are applicable
245  *			when per cpu controls are enforced
246  * @acpi_perf_data:	Stores ACPI perf information read from _PSS
247  * @valid_pss_table:	Set to true for valid ACPI _PSS entries found
248  * @epp_powersave:	Last saved HWP energy performance preference
249  *			(EPP) or energy performance bias (EPB),
250  *			when policy switched to performance
251  * @epp_policy:		Last saved policy used to set EPP/EPB
252  * @epp_default:	Power on default HWP energy performance
253  *			preference/bias
254  * @epp_saved:		Saved EPP/EPB during system suspend or CPU offline
255  *			operation
256  *
257  * This structure stores per CPU instance data for all CPUs.
258  */
259 struct cpudata {
260 	int cpu;
261 
262 	unsigned int policy;
263 	struct update_util_data update_util;
264 	bool   update_util_set;
265 
266 	struct pstate_data pstate;
267 	struct vid_data vid;
268 	struct _pid pid;
269 
270 	u64	last_update;
271 	u64	last_sample_time;
272 	u64	prev_aperf;
273 	u64	prev_mperf;
274 	u64	prev_tsc;
275 	u64	prev_cummulative_iowait;
276 	struct sample sample;
277 	struct perf_limits *perf_limits;
278 #ifdef CONFIG_ACPI
279 	struct acpi_processor_performance acpi_perf_data;
280 	bool valid_pss_table;
281 #endif
282 	unsigned int iowait_boost;
283 	s16 epp_powersave;
284 	s16 epp_policy;
285 	s16 epp_default;
286 	s16 epp_saved;
287 };
288 
289 static struct cpudata **all_cpu_data;
290 
291 /**
292  * struct pstate_adjust_policy - Stores static PID configuration data
293  * @sample_rate_ms:	PID calculation sample rate in ms
294  * @sample_rate_ns:	Sample rate calculation in ns
295  * @deadband:		PID deadband
296  * @setpoint:		PID Setpoint
297  * @p_gain_pct:		PID proportional gain
298  * @i_gain_pct:		PID integral gain
299  * @d_gain_pct:		PID derivative gain
300  *
301  * Stores per CPU model static PID configuration data.
302  */
303 struct pstate_adjust_policy {
304 	int sample_rate_ms;
305 	s64 sample_rate_ns;
306 	int deadband;
307 	int setpoint;
308 	int p_gain_pct;
309 	int d_gain_pct;
310 	int i_gain_pct;
311 };
312 
313 /**
314  * struct pstate_funcs - Per CPU model specific callbacks
315  * @get_max:		Callback to get maximum non turbo effective P state
316  * @get_max_physical:	Callback to get maximum non turbo physical P state
317  * @get_min:		Callback to get minimum P state
318  * @get_turbo:		Callback to get turbo P state
319  * @get_scaling:	Callback to get frequency scaling factor
320  * @get_val:		Callback to convert P state to actual MSR write value
321  * @get_vid:		Callback to get VID data for Atom platforms
322  * @get_target_pstate:	Callback to a function to calculate next P state to use
323  *
324  * Core and Atom CPU models have different way to get P State limits. This
325  * structure is used to store those callbacks.
326  */
327 struct pstate_funcs {
328 	int (*get_max)(void);
329 	int (*get_max_physical)(void);
330 	int (*get_min)(void);
331 	int (*get_turbo)(void);
332 	int (*get_scaling)(void);
333 	u64 (*get_val)(struct cpudata*, int pstate);
334 	void (*get_vid)(struct cpudata *);
335 	int32_t (*get_target_pstate)(struct cpudata *);
336 };
337 
338 /**
339  * struct cpu_defaults- Per CPU model default config data
340  * @pid_policy:	PID config data
341  * @funcs:		Callback function data
342  */
343 struct cpu_defaults {
344 	struct pstate_adjust_policy pid_policy;
345 	struct pstate_funcs funcs;
346 };
347 
348 static inline int32_t get_target_pstate_use_performance(struct cpudata *cpu);
349 static inline int32_t get_target_pstate_use_cpu_load(struct cpudata *cpu);
350 
351 static struct pstate_adjust_policy pid_params __read_mostly;
352 static struct pstate_funcs pstate_funcs __read_mostly;
353 static int hwp_active __read_mostly;
354 static bool per_cpu_limits __read_mostly;
355 
356 static bool driver_registered __read_mostly;
357 
358 #ifdef CONFIG_ACPI
359 static bool acpi_ppc;
360 #endif
361 
362 static struct perf_limits performance_limits;
363 static struct perf_limits powersave_limits;
364 static struct perf_limits *limits;
365 
366 static void intel_pstate_init_limits(struct perf_limits *limits)
367 {
368 	memset(limits, 0, sizeof(*limits));
369 	limits->max_perf_pct = 100;
370 	limits->max_perf = int_ext_tofp(1);
371 	limits->max_policy_pct = 100;
372 	limits->max_sysfs_pct = 100;
373 }
374 
375 static void intel_pstate_set_performance_limits(struct perf_limits *limits)
376 {
377 	intel_pstate_init_limits(limits);
378 	limits->min_perf_pct = 100;
379 	limits->min_perf = int_ext_tofp(1);
380 	limits->min_sysfs_pct = 100;
381 }
382 
383 static DEFINE_MUTEX(intel_pstate_driver_lock);
384 static DEFINE_MUTEX(intel_pstate_limits_lock);
385 
386 #ifdef CONFIG_ACPI
387 
388 static bool intel_pstate_get_ppc_enable_status(void)
389 {
390 	if (acpi_gbl_FADT.preferred_profile == PM_ENTERPRISE_SERVER ||
391 	    acpi_gbl_FADT.preferred_profile == PM_PERFORMANCE_SERVER)
392 		return true;
393 
394 	return acpi_ppc;
395 }
396 
397 #ifdef CONFIG_ACPI_CPPC_LIB
398 
399 /* The work item is needed to avoid CPU hotplug locking issues */
400 static void intel_pstste_sched_itmt_work_fn(struct work_struct *work)
401 {
402 	sched_set_itmt_support();
403 }
404 
405 static DECLARE_WORK(sched_itmt_work, intel_pstste_sched_itmt_work_fn);
406 
407 static void intel_pstate_set_itmt_prio(int cpu)
408 {
409 	struct cppc_perf_caps cppc_perf;
410 	static u32 max_highest_perf = 0, min_highest_perf = U32_MAX;
411 	int ret;
412 
413 	ret = cppc_get_perf_caps(cpu, &cppc_perf);
414 	if (ret)
415 		return;
416 
417 	/*
418 	 * The priorities can be set regardless of whether or not
419 	 * sched_set_itmt_support(true) has been called and it is valid to
420 	 * update them at any time after it has been called.
421 	 */
422 	sched_set_itmt_core_prio(cppc_perf.highest_perf, cpu);
423 
424 	if (max_highest_perf <= min_highest_perf) {
425 		if (cppc_perf.highest_perf > max_highest_perf)
426 			max_highest_perf = cppc_perf.highest_perf;
427 
428 		if (cppc_perf.highest_perf < min_highest_perf)
429 			min_highest_perf = cppc_perf.highest_perf;
430 
431 		if (max_highest_perf > min_highest_perf) {
432 			/*
433 			 * This code can be run during CPU online under the
434 			 * CPU hotplug locks, so sched_set_itmt_support()
435 			 * cannot be called from here.  Queue up a work item
436 			 * to invoke it.
437 			 */
438 			schedule_work(&sched_itmt_work);
439 		}
440 	}
441 }
442 #else
443 static void intel_pstate_set_itmt_prio(int cpu)
444 {
445 }
446 #endif
447 
448 static void intel_pstate_init_acpi_perf_limits(struct cpufreq_policy *policy)
449 {
450 	struct cpudata *cpu;
451 	int ret;
452 	int i;
453 
454 	if (hwp_active) {
455 		intel_pstate_set_itmt_prio(policy->cpu);
456 		return;
457 	}
458 
459 	if (!intel_pstate_get_ppc_enable_status())
460 		return;
461 
462 	cpu = all_cpu_data[policy->cpu];
463 
464 	ret = acpi_processor_register_performance(&cpu->acpi_perf_data,
465 						  policy->cpu);
466 	if (ret)
467 		return;
468 
469 	/*
470 	 * Check if the control value in _PSS is for PERF_CTL MSR, which should
471 	 * guarantee that the states returned by it map to the states in our
472 	 * list directly.
473 	 */
474 	if (cpu->acpi_perf_data.control_register.space_id !=
475 						ACPI_ADR_SPACE_FIXED_HARDWARE)
476 		goto err;
477 
478 	/*
479 	 * If there is only one entry _PSS, simply ignore _PSS and continue as
480 	 * usual without taking _PSS into account
481 	 */
482 	if (cpu->acpi_perf_data.state_count < 2)
483 		goto err;
484 
485 	pr_debug("CPU%u - ACPI _PSS perf data\n", policy->cpu);
486 	for (i = 0; i < cpu->acpi_perf_data.state_count; i++) {
487 		pr_debug("     %cP%d: %u MHz, %u mW, 0x%x\n",
488 			 (i == cpu->acpi_perf_data.state ? '*' : ' '), i,
489 			 (u32) cpu->acpi_perf_data.states[i].core_frequency,
490 			 (u32) cpu->acpi_perf_data.states[i].power,
491 			 (u32) cpu->acpi_perf_data.states[i].control);
492 	}
493 
494 	/*
495 	 * The _PSS table doesn't contain whole turbo frequency range.
496 	 * This just contains +1 MHZ above the max non turbo frequency,
497 	 * with control value corresponding to max turbo ratio. But
498 	 * when cpufreq set policy is called, it will call with this
499 	 * max frequency, which will cause a reduced performance as
500 	 * this driver uses real max turbo frequency as the max
501 	 * frequency. So correct this frequency in _PSS table to
502 	 * correct max turbo frequency based on the turbo state.
503 	 * Also need to convert to MHz as _PSS freq is in MHz.
504 	 */
505 	if (!limits->turbo_disabled)
506 		cpu->acpi_perf_data.states[0].core_frequency =
507 					policy->cpuinfo.max_freq / 1000;
508 	cpu->valid_pss_table = true;
509 	pr_debug("_PPC limits will be enforced\n");
510 
511 	return;
512 
513  err:
514 	cpu->valid_pss_table = false;
515 	acpi_processor_unregister_performance(policy->cpu);
516 }
517 
518 static void intel_pstate_exit_perf_limits(struct cpufreq_policy *policy)
519 {
520 	struct cpudata *cpu;
521 
522 	cpu = all_cpu_data[policy->cpu];
523 	if (!cpu->valid_pss_table)
524 		return;
525 
526 	acpi_processor_unregister_performance(policy->cpu);
527 }
528 #else
529 static inline void intel_pstate_init_acpi_perf_limits(struct cpufreq_policy *policy)
530 {
531 }
532 
533 static inline void intel_pstate_exit_perf_limits(struct cpufreq_policy *policy)
534 {
535 }
536 #endif
537 
538 static inline void pid_reset(struct _pid *pid, int setpoint, int busy,
539 			     int deadband, int integral) {
540 	pid->setpoint = int_tofp(setpoint);
541 	pid->deadband  = int_tofp(deadband);
542 	pid->integral  = int_tofp(integral);
543 	pid->last_err  = int_tofp(setpoint) - int_tofp(busy);
544 }
545 
546 static inline void pid_p_gain_set(struct _pid *pid, int percent)
547 {
548 	pid->p_gain = div_fp(percent, 100);
549 }
550 
551 static inline void pid_i_gain_set(struct _pid *pid, int percent)
552 {
553 	pid->i_gain = div_fp(percent, 100);
554 }
555 
556 static inline void pid_d_gain_set(struct _pid *pid, int percent)
557 {
558 	pid->d_gain = div_fp(percent, 100);
559 }
560 
561 static signed int pid_calc(struct _pid *pid, int32_t busy)
562 {
563 	signed int result;
564 	int32_t pterm, dterm, fp_error;
565 	int32_t integral_limit;
566 
567 	fp_error = pid->setpoint - busy;
568 
569 	if (abs(fp_error) <= pid->deadband)
570 		return 0;
571 
572 	pterm = mul_fp(pid->p_gain, fp_error);
573 
574 	pid->integral += fp_error;
575 
576 	/*
577 	 * We limit the integral here so that it will never
578 	 * get higher than 30.  This prevents it from becoming
579 	 * too large an input over long periods of time and allows
580 	 * it to get factored out sooner.
581 	 *
582 	 * The value of 30 was chosen through experimentation.
583 	 */
584 	integral_limit = int_tofp(30);
585 	if (pid->integral > integral_limit)
586 		pid->integral = integral_limit;
587 	if (pid->integral < -integral_limit)
588 		pid->integral = -integral_limit;
589 
590 	dterm = mul_fp(pid->d_gain, fp_error - pid->last_err);
591 	pid->last_err = fp_error;
592 
593 	result = pterm + mul_fp(pid->integral, pid->i_gain) + dterm;
594 	result = result + (1 << (FRAC_BITS-1));
595 	return (signed int)fp_toint(result);
596 }
597 
598 static inline void intel_pstate_busy_pid_reset(struct cpudata *cpu)
599 {
600 	pid_p_gain_set(&cpu->pid, pid_params.p_gain_pct);
601 	pid_d_gain_set(&cpu->pid, pid_params.d_gain_pct);
602 	pid_i_gain_set(&cpu->pid, pid_params.i_gain_pct);
603 
604 	pid_reset(&cpu->pid, pid_params.setpoint, 100, pid_params.deadband, 0);
605 }
606 
607 static inline void intel_pstate_reset_all_pid(void)
608 {
609 	unsigned int cpu;
610 
611 	for_each_online_cpu(cpu) {
612 		if (all_cpu_data[cpu])
613 			intel_pstate_busy_pid_reset(all_cpu_data[cpu]);
614 	}
615 }
616 
617 static inline void update_turbo_state(void)
618 {
619 	u64 misc_en;
620 	struct cpudata *cpu;
621 
622 	cpu = all_cpu_data[0];
623 	rdmsrl(MSR_IA32_MISC_ENABLE, misc_en);
624 	limits->turbo_disabled =
625 		(misc_en & MSR_IA32_MISC_ENABLE_TURBO_DISABLE ||
626 		 cpu->pstate.max_pstate == cpu->pstate.turbo_pstate);
627 }
628 
629 static s16 intel_pstate_get_epb(struct cpudata *cpu_data)
630 {
631 	u64 epb;
632 	int ret;
633 
634 	if (!static_cpu_has(X86_FEATURE_EPB))
635 		return -ENXIO;
636 
637 	ret = rdmsrl_on_cpu(cpu_data->cpu, MSR_IA32_ENERGY_PERF_BIAS, &epb);
638 	if (ret)
639 		return (s16)ret;
640 
641 	return (s16)(epb & 0x0f);
642 }
643 
644 static s16 intel_pstate_get_epp(struct cpudata *cpu_data, u64 hwp_req_data)
645 {
646 	s16 epp;
647 
648 	if (static_cpu_has(X86_FEATURE_HWP_EPP)) {
649 		/*
650 		 * When hwp_req_data is 0, means that caller didn't read
651 		 * MSR_HWP_REQUEST, so need to read and get EPP.
652 		 */
653 		if (!hwp_req_data) {
654 			epp = rdmsrl_on_cpu(cpu_data->cpu, MSR_HWP_REQUEST,
655 					    &hwp_req_data);
656 			if (epp)
657 				return epp;
658 		}
659 		epp = (hwp_req_data >> 24) & 0xff;
660 	} else {
661 		/* When there is no EPP present, HWP uses EPB settings */
662 		epp = intel_pstate_get_epb(cpu_data);
663 	}
664 
665 	return epp;
666 }
667 
668 static int intel_pstate_set_epb(int cpu, s16 pref)
669 {
670 	u64 epb;
671 	int ret;
672 
673 	if (!static_cpu_has(X86_FEATURE_EPB))
674 		return -ENXIO;
675 
676 	ret = rdmsrl_on_cpu(cpu, MSR_IA32_ENERGY_PERF_BIAS, &epb);
677 	if (ret)
678 		return ret;
679 
680 	epb = (epb & ~0x0f) | pref;
681 	wrmsrl_on_cpu(cpu, MSR_IA32_ENERGY_PERF_BIAS, epb);
682 
683 	return 0;
684 }
685 
686 /*
687  * EPP/EPB display strings corresponding to EPP index in the
688  * energy_perf_strings[]
689  *	index		String
690  *-------------------------------------
691  *	0		default
692  *	1		performance
693  *	2		balance_performance
694  *	3		balance_power
695  *	4		power
696  */
697 static const char * const energy_perf_strings[] = {
698 	"default",
699 	"performance",
700 	"balance_performance",
701 	"balance_power",
702 	"power",
703 	NULL
704 };
705 
706 static int intel_pstate_get_energy_pref_index(struct cpudata *cpu_data)
707 {
708 	s16 epp;
709 	int index = -EINVAL;
710 
711 	epp = intel_pstate_get_epp(cpu_data, 0);
712 	if (epp < 0)
713 		return epp;
714 
715 	if (static_cpu_has(X86_FEATURE_HWP_EPP)) {
716 		/*
717 		 * Range:
718 		 *	0x00-0x3F	:	Performance
719 		 *	0x40-0x7F	:	Balance performance
720 		 *	0x80-0xBF	:	Balance power
721 		 *	0xC0-0xFF	:	Power
722 		 * The EPP is a 8 bit value, but our ranges restrict the
723 		 * value which can be set. Here only using top two bits
724 		 * effectively.
725 		 */
726 		index = (epp >> 6) + 1;
727 	} else if (static_cpu_has(X86_FEATURE_EPB)) {
728 		/*
729 		 * Range:
730 		 *	0x00-0x03	:	Performance
731 		 *	0x04-0x07	:	Balance performance
732 		 *	0x08-0x0B	:	Balance power
733 		 *	0x0C-0x0F	:	Power
734 		 * The EPB is a 4 bit value, but our ranges restrict the
735 		 * value which can be set. Here only using top two bits
736 		 * effectively.
737 		 */
738 		index = (epp >> 2) + 1;
739 	}
740 
741 	return index;
742 }
743 
744 static int intel_pstate_set_energy_pref_index(struct cpudata *cpu_data,
745 					      int pref_index)
746 {
747 	int epp = -EINVAL;
748 	int ret;
749 
750 	if (!pref_index)
751 		epp = cpu_data->epp_default;
752 
753 	mutex_lock(&intel_pstate_limits_lock);
754 
755 	if (static_cpu_has(X86_FEATURE_HWP_EPP)) {
756 		u64 value;
757 
758 		ret = rdmsrl_on_cpu(cpu_data->cpu, MSR_HWP_REQUEST, &value);
759 		if (ret)
760 			goto return_pref;
761 
762 		value &= ~GENMASK_ULL(31, 24);
763 
764 		/*
765 		 * If epp is not default, convert from index into
766 		 * energy_perf_strings to epp value, by shifting 6
767 		 * bits left to use only top two bits in epp.
768 		 * The resultant epp need to shifted by 24 bits to
769 		 * epp position in MSR_HWP_REQUEST.
770 		 */
771 		if (epp == -EINVAL)
772 			epp = (pref_index - 1) << 6;
773 
774 		value |= (u64)epp << 24;
775 		ret = wrmsrl_on_cpu(cpu_data->cpu, MSR_HWP_REQUEST, value);
776 	} else {
777 		if (epp == -EINVAL)
778 			epp = (pref_index - 1) << 2;
779 		ret = intel_pstate_set_epb(cpu_data->cpu, epp);
780 	}
781 return_pref:
782 	mutex_unlock(&intel_pstate_limits_lock);
783 
784 	return ret;
785 }
786 
787 static ssize_t show_energy_performance_available_preferences(
788 				struct cpufreq_policy *policy, char *buf)
789 {
790 	int i = 0;
791 	int ret = 0;
792 
793 	while (energy_perf_strings[i] != NULL)
794 		ret += sprintf(&buf[ret], "%s ", energy_perf_strings[i++]);
795 
796 	ret += sprintf(&buf[ret], "\n");
797 
798 	return ret;
799 }
800 
801 cpufreq_freq_attr_ro(energy_performance_available_preferences);
802 
803 static ssize_t store_energy_performance_preference(
804 		struct cpufreq_policy *policy, const char *buf, size_t count)
805 {
806 	struct cpudata *cpu_data = all_cpu_data[policy->cpu];
807 	char str_preference[21];
808 	int ret, i = 0;
809 
810 	ret = sscanf(buf, "%20s", str_preference);
811 	if (ret != 1)
812 		return -EINVAL;
813 
814 	while (energy_perf_strings[i] != NULL) {
815 		if (!strcmp(str_preference, energy_perf_strings[i])) {
816 			intel_pstate_set_energy_pref_index(cpu_data, i);
817 			return count;
818 		}
819 		++i;
820 	}
821 
822 	return -EINVAL;
823 }
824 
825 static ssize_t show_energy_performance_preference(
826 				struct cpufreq_policy *policy, char *buf)
827 {
828 	struct cpudata *cpu_data = all_cpu_data[policy->cpu];
829 	int preference;
830 
831 	preference = intel_pstate_get_energy_pref_index(cpu_data);
832 	if (preference < 0)
833 		return preference;
834 
835 	return  sprintf(buf, "%s\n", energy_perf_strings[preference]);
836 }
837 
838 cpufreq_freq_attr_rw(energy_performance_preference);
839 
840 static struct freq_attr *hwp_cpufreq_attrs[] = {
841 	&energy_performance_preference,
842 	&energy_performance_available_preferences,
843 	NULL,
844 };
845 
846 static void intel_pstate_hwp_set(struct cpufreq_policy *policy)
847 {
848 	int min, hw_min, max, hw_max, cpu, range, adj_range;
849 	struct perf_limits *perf_limits = limits;
850 	u64 value, cap;
851 
852 	for_each_cpu(cpu, policy->cpus) {
853 		int max_perf_pct, min_perf_pct;
854 		struct cpudata *cpu_data = all_cpu_data[cpu];
855 		s16 epp;
856 
857 		if (per_cpu_limits)
858 			perf_limits = all_cpu_data[cpu]->perf_limits;
859 
860 		rdmsrl_on_cpu(cpu, MSR_HWP_CAPABILITIES, &cap);
861 		hw_min = HWP_LOWEST_PERF(cap);
862 		if (limits->no_turbo)
863 			hw_max = HWP_GUARANTEED_PERF(cap);
864 		else
865 			hw_max = HWP_HIGHEST_PERF(cap);
866 		range = hw_max - hw_min;
867 
868 		max_perf_pct = perf_limits->max_perf_pct;
869 		min_perf_pct = perf_limits->min_perf_pct;
870 
871 		rdmsrl_on_cpu(cpu, MSR_HWP_REQUEST, &value);
872 		adj_range = min_perf_pct * range / 100;
873 		min = hw_min + adj_range;
874 		value &= ~HWP_MIN_PERF(~0L);
875 		value |= HWP_MIN_PERF(min);
876 
877 		adj_range = max_perf_pct * range / 100;
878 		max = hw_min + adj_range;
879 
880 		value &= ~HWP_MAX_PERF(~0L);
881 		value |= HWP_MAX_PERF(max);
882 
883 		if (cpu_data->epp_policy == cpu_data->policy)
884 			goto skip_epp;
885 
886 		cpu_data->epp_policy = cpu_data->policy;
887 
888 		if (cpu_data->epp_saved >= 0) {
889 			epp = cpu_data->epp_saved;
890 			cpu_data->epp_saved = -EINVAL;
891 			goto update_epp;
892 		}
893 
894 		if (cpu_data->policy == CPUFREQ_POLICY_PERFORMANCE) {
895 			epp = intel_pstate_get_epp(cpu_data, value);
896 			cpu_data->epp_powersave = epp;
897 			/* If EPP read was failed, then don't try to write */
898 			if (epp < 0)
899 				goto skip_epp;
900 
901 
902 			epp = 0;
903 		} else {
904 			/* skip setting EPP, when saved value is invalid */
905 			if (cpu_data->epp_powersave < 0)
906 				goto skip_epp;
907 
908 			/*
909 			 * No need to restore EPP when it is not zero. This
910 			 * means:
911 			 *  - Policy is not changed
912 			 *  - user has manually changed
913 			 *  - Error reading EPB
914 			 */
915 			epp = intel_pstate_get_epp(cpu_data, value);
916 			if (epp)
917 				goto skip_epp;
918 
919 			epp = cpu_data->epp_powersave;
920 		}
921 update_epp:
922 		if (static_cpu_has(X86_FEATURE_HWP_EPP)) {
923 			value &= ~GENMASK_ULL(31, 24);
924 			value |= (u64)epp << 24;
925 		} else {
926 			intel_pstate_set_epb(cpu, epp);
927 		}
928 skip_epp:
929 		wrmsrl_on_cpu(cpu, MSR_HWP_REQUEST, value);
930 	}
931 }
932 
933 static int intel_pstate_hwp_set_policy(struct cpufreq_policy *policy)
934 {
935 	if (hwp_active)
936 		intel_pstate_hwp_set(policy);
937 
938 	return 0;
939 }
940 
941 static int intel_pstate_hwp_save_state(struct cpufreq_policy *policy)
942 {
943 	struct cpudata *cpu_data = all_cpu_data[policy->cpu];
944 
945 	if (!hwp_active)
946 		return 0;
947 
948 	cpu_data->epp_saved = intel_pstate_get_epp(cpu_data, 0);
949 
950 	return 0;
951 }
952 
953 static int intel_pstate_resume(struct cpufreq_policy *policy)
954 {
955 	int ret;
956 
957 	if (!hwp_active)
958 		return 0;
959 
960 	mutex_lock(&intel_pstate_limits_lock);
961 
962 	all_cpu_data[policy->cpu]->epp_policy = 0;
963 
964 	ret = intel_pstate_hwp_set_policy(policy);
965 
966 	mutex_unlock(&intel_pstate_limits_lock);
967 
968 	return ret;
969 }
970 
971 static void intel_pstate_update_policies(void)
972 	__releases(&intel_pstate_limits_lock)
973 	__acquires(&intel_pstate_limits_lock)
974 {
975 	struct perf_limits *saved_limits = limits;
976 	int cpu;
977 
978 	mutex_unlock(&intel_pstate_limits_lock);
979 
980 	for_each_possible_cpu(cpu)
981 		cpufreq_update_policy(cpu);
982 
983 	mutex_lock(&intel_pstate_limits_lock);
984 
985 	limits = saved_limits;
986 }
987 
988 /************************** debugfs begin ************************/
989 static int pid_param_set(void *data, u64 val)
990 {
991 	*(u32 *)data = val;
992 	intel_pstate_reset_all_pid();
993 	return 0;
994 }
995 
996 static int pid_param_get(void *data, u64 *val)
997 {
998 	*val = *(u32 *)data;
999 	return 0;
1000 }
1001 DEFINE_SIMPLE_ATTRIBUTE(fops_pid_param, pid_param_get, pid_param_set, "%llu\n");
1002 
1003 static struct dentry *debugfs_parent;
1004 
1005 struct pid_param {
1006 	char *name;
1007 	void *value;
1008 	struct dentry *dentry;
1009 };
1010 
1011 static struct pid_param pid_files[] = {
1012 	{"sample_rate_ms", &pid_params.sample_rate_ms, },
1013 	{"d_gain_pct", &pid_params.d_gain_pct, },
1014 	{"i_gain_pct", &pid_params.i_gain_pct, },
1015 	{"deadband", &pid_params.deadband, },
1016 	{"setpoint", &pid_params.setpoint, },
1017 	{"p_gain_pct", &pid_params.p_gain_pct, },
1018 	{NULL, NULL, }
1019 };
1020 
1021 static void intel_pstate_debug_expose_params(void)
1022 {
1023 	int i;
1024 
1025 	debugfs_parent = debugfs_create_dir("pstate_snb", NULL);
1026 	if (IS_ERR_OR_NULL(debugfs_parent))
1027 		return;
1028 
1029 	for (i = 0; pid_files[i].name; i++) {
1030 		struct dentry *dentry;
1031 
1032 		dentry = debugfs_create_file(pid_files[i].name, 0660,
1033 					     debugfs_parent, pid_files[i].value,
1034 					     &fops_pid_param);
1035 		if (!IS_ERR(dentry))
1036 			pid_files[i].dentry = dentry;
1037 	}
1038 }
1039 
1040 static void intel_pstate_debug_hide_params(void)
1041 {
1042 	int i;
1043 
1044 	if (IS_ERR_OR_NULL(debugfs_parent))
1045 		return;
1046 
1047 	for (i = 0; pid_files[i].name; i++) {
1048 		debugfs_remove(pid_files[i].dentry);
1049 		pid_files[i].dentry = NULL;
1050 	}
1051 
1052 	debugfs_remove(debugfs_parent);
1053 	debugfs_parent = NULL;
1054 }
1055 
1056 /************************** debugfs end ************************/
1057 
1058 /************************** sysfs begin ************************/
1059 #define show_one(file_name, object)					\
1060 	static ssize_t show_##file_name					\
1061 	(struct kobject *kobj, struct attribute *attr, char *buf)	\
1062 	{								\
1063 		return sprintf(buf, "%u\n", limits->object);		\
1064 	}
1065 
1066 static ssize_t intel_pstate_show_status(char *buf);
1067 static int intel_pstate_update_status(const char *buf, size_t size);
1068 
1069 static ssize_t show_status(struct kobject *kobj,
1070 			   struct attribute *attr, char *buf)
1071 {
1072 	ssize_t ret;
1073 
1074 	mutex_lock(&intel_pstate_driver_lock);
1075 	ret = intel_pstate_show_status(buf);
1076 	mutex_unlock(&intel_pstate_driver_lock);
1077 
1078 	return ret;
1079 }
1080 
1081 static ssize_t store_status(struct kobject *a, struct attribute *b,
1082 			    const char *buf, size_t count)
1083 {
1084 	char *p = memchr(buf, '\n', count);
1085 	int ret;
1086 
1087 	mutex_lock(&intel_pstate_driver_lock);
1088 	ret = intel_pstate_update_status(buf, p ? p - buf : count);
1089 	mutex_unlock(&intel_pstate_driver_lock);
1090 
1091 	return ret < 0 ? ret : count;
1092 }
1093 
1094 static ssize_t show_turbo_pct(struct kobject *kobj,
1095 				struct attribute *attr, char *buf)
1096 {
1097 	struct cpudata *cpu;
1098 	int total, no_turbo, turbo_pct;
1099 	uint32_t turbo_fp;
1100 
1101 	mutex_lock(&intel_pstate_driver_lock);
1102 
1103 	if (!driver_registered) {
1104 		mutex_unlock(&intel_pstate_driver_lock);
1105 		return -EAGAIN;
1106 	}
1107 
1108 	cpu = all_cpu_data[0];
1109 
1110 	total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
1111 	no_turbo = cpu->pstate.max_pstate - cpu->pstate.min_pstate + 1;
1112 	turbo_fp = div_fp(no_turbo, total);
1113 	turbo_pct = 100 - fp_toint(mul_fp(turbo_fp, int_tofp(100)));
1114 
1115 	mutex_unlock(&intel_pstate_driver_lock);
1116 
1117 	return sprintf(buf, "%u\n", turbo_pct);
1118 }
1119 
1120 static ssize_t show_num_pstates(struct kobject *kobj,
1121 				struct attribute *attr, char *buf)
1122 {
1123 	struct cpudata *cpu;
1124 	int total;
1125 
1126 	mutex_lock(&intel_pstate_driver_lock);
1127 
1128 	if (!driver_registered) {
1129 		mutex_unlock(&intel_pstate_driver_lock);
1130 		return -EAGAIN;
1131 	}
1132 
1133 	cpu = all_cpu_data[0];
1134 	total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
1135 
1136 	mutex_unlock(&intel_pstate_driver_lock);
1137 
1138 	return sprintf(buf, "%u\n", total);
1139 }
1140 
1141 static ssize_t show_no_turbo(struct kobject *kobj,
1142 			     struct attribute *attr, char *buf)
1143 {
1144 	ssize_t ret;
1145 
1146 	mutex_lock(&intel_pstate_driver_lock);
1147 
1148 	if (!driver_registered) {
1149 		mutex_unlock(&intel_pstate_driver_lock);
1150 		return -EAGAIN;
1151 	}
1152 
1153 	update_turbo_state();
1154 	if (limits->turbo_disabled)
1155 		ret = sprintf(buf, "%u\n", limits->turbo_disabled);
1156 	else
1157 		ret = sprintf(buf, "%u\n", limits->no_turbo);
1158 
1159 	mutex_unlock(&intel_pstate_driver_lock);
1160 
1161 	return ret;
1162 }
1163 
1164 static ssize_t store_no_turbo(struct kobject *a, struct attribute *b,
1165 			      const char *buf, size_t count)
1166 {
1167 	unsigned int input;
1168 	int ret;
1169 
1170 	ret = sscanf(buf, "%u", &input);
1171 	if (ret != 1)
1172 		return -EINVAL;
1173 
1174 	mutex_lock(&intel_pstate_driver_lock);
1175 
1176 	if (!driver_registered) {
1177 		mutex_unlock(&intel_pstate_driver_lock);
1178 		return -EAGAIN;
1179 	}
1180 
1181 	mutex_lock(&intel_pstate_limits_lock);
1182 
1183 	update_turbo_state();
1184 	if (limits->turbo_disabled) {
1185 		pr_warn("Turbo disabled by BIOS or unavailable on processor\n");
1186 		mutex_unlock(&intel_pstate_limits_lock);
1187 		mutex_unlock(&intel_pstate_driver_lock);
1188 		return -EPERM;
1189 	}
1190 
1191 	limits->no_turbo = clamp_t(int, input, 0, 1);
1192 
1193 	intel_pstate_update_policies();
1194 
1195 	mutex_unlock(&intel_pstate_limits_lock);
1196 
1197 	mutex_unlock(&intel_pstate_driver_lock);
1198 
1199 	return count;
1200 }
1201 
1202 static ssize_t store_max_perf_pct(struct kobject *a, struct attribute *b,
1203 				  const char *buf, size_t count)
1204 {
1205 	unsigned int input;
1206 	int ret;
1207 
1208 	ret = sscanf(buf, "%u", &input);
1209 	if (ret != 1)
1210 		return -EINVAL;
1211 
1212 	mutex_lock(&intel_pstate_driver_lock);
1213 
1214 	if (!driver_registered) {
1215 		mutex_unlock(&intel_pstate_driver_lock);
1216 		return -EAGAIN;
1217 	}
1218 
1219 	mutex_lock(&intel_pstate_limits_lock);
1220 
1221 	limits->max_sysfs_pct = clamp_t(int, input, 0 , 100);
1222 	limits->max_perf_pct = min(limits->max_policy_pct,
1223 				   limits->max_sysfs_pct);
1224 	limits->max_perf_pct = max(limits->min_policy_pct,
1225 				   limits->max_perf_pct);
1226 	limits->max_perf_pct = max(limits->min_perf_pct,
1227 				   limits->max_perf_pct);
1228 	limits->max_perf = div_ext_fp(limits->max_perf_pct, 100);
1229 
1230 	intel_pstate_update_policies();
1231 
1232 	mutex_unlock(&intel_pstate_limits_lock);
1233 
1234 	mutex_unlock(&intel_pstate_driver_lock);
1235 
1236 	return count;
1237 }
1238 
1239 static ssize_t store_min_perf_pct(struct kobject *a, struct attribute *b,
1240 				  const char *buf, size_t count)
1241 {
1242 	unsigned int input;
1243 	int ret;
1244 
1245 	ret = sscanf(buf, "%u", &input);
1246 	if (ret != 1)
1247 		return -EINVAL;
1248 
1249 	mutex_lock(&intel_pstate_driver_lock);
1250 
1251 	if (!driver_registered) {
1252 		mutex_unlock(&intel_pstate_driver_lock);
1253 		return -EAGAIN;
1254 	}
1255 
1256 	mutex_lock(&intel_pstate_limits_lock);
1257 
1258 	limits->min_sysfs_pct = clamp_t(int, input, 0 , 100);
1259 	limits->min_perf_pct = max(limits->min_policy_pct,
1260 				   limits->min_sysfs_pct);
1261 	limits->min_perf_pct = min(limits->max_policy_pct,
1262 				   limits->min_perf_pct);
1263 	limits->min_perf_pct = min(limits->max_perf_pct,
1264 				   limits->min_perf_pct);
1265 	limits->min_perf = div_ext_fp(limits->min_perf_pct, 100);
1266 
1267 	intel_pstate_update_policies();
1268 
1269 	mutex_unlock(&intel_pstate_limits_lock);
1270 
1271 	mutex_unlock(&intel_pstate_driver_lock);
1272 
1273 	return count;
1274 }
1275 
1276 show_one(max_perf_pct, max_perf_pct);
1277 show_one(min_perf_pct, min_perf_pct);
1278 
1279 define_one_global_rw(status);
1280 define_one_global_rw(no_turbo);
1281 define_one_global_rw(max_perf_pct);
1282 define_one_global_rw(min_perf_pct);
1283 define_one_global_ro(turbo_pct);
1284 define_one_global_ro(num_pstates);
1285 
1286 static struct attribute *intel_pstate_attributes[] = {
1287 	&status.attr,
1288 	&no_turbo.attr,
1289 	&turbo_pct.attr,
1290 	&num_pstates.attr,
1291 	NULL
1292 };
1293 
1294 static struct attribute_group intel_pstate_attr_group = {
1295 	.attrs = intel_pstate_attributes,
1296 };
1297 
1298 static void __init intel_pstate_sysfs_expose_params(void)
1299 {
1300 	struct kobject *intel_pstate_kobject;
1301 	int rc;
1302 
1303 	intel_pstate_kobject = kobject_create_and_add("intel_pstate",
1304 						&cpu_subsys.dev_root->kobj);
1305 	if (WARN_ON(!intel_pstate_kobject))
1306 		return;
1307 
1308 	rc = sysfs_create_group(intel_pstate_kobject, &intel_pstate_attr_group);
1309 	if (WARN_ON(rc))
1310 		return;
1311 
1312 	/*
1313 	 * If per cpu limits are enforced there are no global limits, so
1314 	 * return without creating max/min_perf_pct attributes
1315 	 */
1316 	if (per_cpu_limits)
1317 		return;
1318 
1319 	rc = sysfs_create_file(intel_pstate_kobject, &max_perf_pct.attr);
1320 	WARN_ON(rc);
1321 
1322 	rc = sysfs_create_file(intel_pstate_kobject, &min_perf_pct.attr);
1323 	WARN_ON(rc);
1324 
1325 }
1326 /************************** sysfs end ************************/
1327 
1328 static void intel_pstate_hwp_enable(struct cpudata *cpudata)
1329 {
1330 	/* First disable HWP notification interrupt as we don't process them */
1331 	if (static_cpu_has(X86_FEATURE_HWP_NOTIFY))
1332 		wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_INTERRUPT, 0x00);
1333 
1334 	wrmsrl_on_cpu(cpudata->cpu, MSR_PM_ENABLE, 0x1);
1335 	cpudata->epp_policy = 0;
1336 	if (cpudata->epp_default == -EINVAL)
1337 		cpudata->epp_default = intel_pstate_get_epp(cpudata, 0);
1338 }
1339 
1340 #define MSR_IA32_POWER_CTL_BIT_EE	19
1341 
1342 /* Disable energy efficiency optimization */
1343 static void intel_pstate_disable_ee(int cpu)
1344 {
1345 	u64 power_ctl;
1346 	int ret;
1347 
1348 	ret = rdmsrl_on_cpu(cpu, MSR_IA32_POWER_CTL, &power_ctl);
1349 	if (ret)
1350 		return;
1351 
1352 	if (!(power_ctl & BIT(MSR_IA32_POWER_CTL_BIT_EE))) {
1353 		pr_info("Disabling energy efficiency optimization\n");
1354 		power_ctl |= BIT(MSR_IA32_POWER_CTL_BIT_EE);
1355 		wrmsrl_on_cpu(cpu, MSR_IA32_POWER_CTL, power_ctl);
1356 	}
1357 }
1358 
1359 static int atom_get_min_pstate(void)
1360 {
1361 	u64 value;
1362 
1363 	rdmsrl(MSR_ATOM_CORE_RATIOS, value);
1364 	return (value >> 8) & 0x7F;
1365 }
1366 
1367 static int atom_get_max_pstate(void)
1368 {
1369 	u64 value;
1370 
1371 	rdmsrl(MSR_ATOM_CORE_RATIOS, value);
1372 	return (value >> 16) & 0x7F;
1373 }
1374 
1375 static int atom_get_turbo_pstate(void)
1376 {
1377 	u64 value;
1378 
1379 	rdmsrl(MSR_ATOM_CORE_TURBO_RATIOS, value);
1380 	return value & 0x7F;
1381 }
1382 
1383 static u64 atom_get_val(struct cpudata *cpudata, int pstate)
1384 {
1385 	u64 val;
1386 	int32_t vid_fp;
1387 	u32 vid;
1388 
1389 	val = (u64)pstate << 8;
1390 	if (limits->no_turbo && !limits->turbo_disabled)
1391 		val |= (u64)1 << 32;
1392 
1393 	vid_fp = cpudata->vid.min + mul_fp(
1394 		int_tofp(pstate - cpudata->pstate.min_pstate),
1395 		cpudata->vid.ratio);
1396 
1397 	vid_fp = clamp_t(int32_t, vid_fp, cpudata->vid.min, cpudata->vid.max);
1398 	vid = ceiling_fp(vid_fp);
1399 
1400 	if (pstate > cpudata->pstate.max_pstate)
1401 		vid = cpudata->vid.turbo;
1402 
1403 	return val | vid;
1404 }
1405 
1406 static int silvermont_get_scaling(void)
1407 {
1408 	u64 value;
1409 	int i;
1410 	/* Defined in Table 35-6 from SDM (Sept 2015) */
1411 	static int silvermont_freq_table[] = {
1412 		83300, 100000, 133300, 116700, 80000};
1413 
1414 	rdmsrl(MSR_FSB_FREQ, value);
1415 	i = value & 0x7;
1416 	WARN_ON(i > 4);
1417 
1418 	return silvermont_freq_table[i];
1419 }
1420 
1421 static int airmont_get_scaling(void)
1422 {
1423 	u64 value;
1424 	int i;
1425 	/* Defined in Table 35-10 from SDM (Sept 2015) */
1426 	static int airmont_freq_table[] = {
1427 		83300, 100000, 133300, 116700, 80000,
1428 		93300, 90000, 88900, 87500};
1429 
1430 	rdmsrl(MSR_FSB_FREQ, value);
1431 	i = value & 0xF;
1432 	WARN_ON(i > 8);
1433 
1434 	return airmont_freq_table[i];
1435 }
1436 
1437 static void atom_get_vid(struct cpudata *cpudata)
1438 {
1439 	u64 value;
1440 
1441 	rdmsrl(MSR_ATOM_CORE_VIDS, value);
1442 	cpudata->vid.min = int_tofp((value >> 8) & 0x7f);
1443 	cpudata->vid.max = int_tofp((value >> 16) & 0x7f);
1444 	cpudata->vid.ratio = div_fp(
1445 		cpudata->vid.max - cpudata->vid.min,
1446 		int_tofp(cpudata->pstate.max_pstate -
1447 			cpudata->pstate.min_pstate));
1448 
1449 	rdmsrl(MSR_ATOM_CORE_TURBO_VIDS, value);
1450 	cpudata->vid.turbo = value & 0x7f;
1451 }
1452 
1453 static int core_get_min_pstate(void)
1454 {
1455 	u64 value;
1456 
1457 	rdmsrl(MSR_PLATFORM_INFO, value);
1458 	return (value >> 40) & 0xFF;
1459 }
1460 
1461 static int core_get_max_pstate_physical(void)
1462 {
1463 	u64 value;
1464 
1465 	rdmsrl(MSR_PLATFORM_INFO, value);
1466 	return (value >> 8) & 0xFF;
1467 }
1468 
1469 static int core_get_tdp_ratio(u64 plat_info)
1470 {
1471 	/* Check how many TDP levels present */
1472 	if (plat_info & 0x600000000) {
1473 		u64 tdp_ctrl;
1474 		u64 tdp_ratio;
1475 		int tdp_msr;
1476 		int err;
1477 
1478 		/* Get the TDP level (0, 1, 2) to get ratios */
1479 		err = rdmsrl_safe(MSR_CONFIG_TDP_CONTROL, &tdp_ctrl);
1480 		if (err)
1481 			return err;
1482 
1483 		/* TDP MSR are continuous starting at 0x648 */
1484 		tdp_msr = MSR_CONFIG_TDP_NOMINAL + (tdp_ctrl & 0x03);
1485 		err = rdmsrl_safe(tdp_msr, &tdp_ratio);
1486 		if (err)
1487 			return err;
1488 
1489 		/* For level 1 and 2, bits[23:16] contain the ratio */
1490 		if (tdp_ctrl & 0x03)
1491 			tdp_ratio >>= 16;
1492 
1493 		tdp_ratio &= 0xff; /* ratios are only 8 bits long */
1494 		pr_debug("tdp_ratio %x\n", (int)tdp_ratio);
1495 
1496 		return (int)tdp_ratio;
1497 	}
1498 
1499 	return -ENXIO;
1500 }
1501 
1502 static int core_get_max_pstate(void)
1503 {
1504 	u64 tar;
1505 	u64 plat_info;
1506 	int max_pstate;
1507 	int tdp_ratio;
1508 	int err;
1509 
1510 	rdmsrl(MSR_PLATFORM_INFO, plat_info);
1511 	max_pstate = (plat_info >> 8) & 0xFF;
1512 
1513 	tdp_ratio = core_get_tdp_ratio(plat_info);
1514 	if (tdp_ratio <= 0)
1515 		return max_pstate;
1516 
1517 	if (hwp_active) {
1518 		/* Turbo activation ratio is not used on HWP platforms */
1519 		return tdp_ratio;
1520 	}
1521 
1522 	err = rdmsrl_safe(MSR_TURBO_ACTIVATION_RATIO, &tar);
1523 	if (!err) {
1524 		int tar_levels;
1525 
1526 		/* Do some sanity checking for safety */
1527 		tar_levels = tar & 0xff;
1528 		if (tdp_ratio - 1 == tar_levels) {
1529 			max_pstate = tar_levels;
1530 			pr_debug("max_pstate=TAC %x\n", max_pstate);
1531 		}
1532 	}
1533 
1534 	return max_pstate;
1535 }
1536 
1537 static int core_get_turbo_pstate(void)
1538 {
1539 	u64 value;
1540 	int nont, ret;
1541 
1542 	rdmsrl(MSR_TURBO_RATIO_LIMIT, value);
1543 	nont = core_get_max_pstate();
1544 	ret = (value) & 255;
1545 	if (ret <= nont)
1546 		ret = nont;
1547 	return ret;
1548 }
1549 
1550 static inline int core_get_scaling(void)
1551 {
1552 	return 100000;
1553 }
1554 
1555 static u64 core_get_val(struct cpudata *cpudata, int pstate)
1556 {
1557 	u64 val;
1558 
1559 	val = (u64)pstate << 8;
1560 	if (limits->no_turbo && !limits->turbo_disabled)
1561 		val |= (u64)1 << 32;
1562 
1563 	return val;
1564 }
1565 
1566 static int knl_get_turbo_pstate(void)
1567 {
1568 	u64 value;
1569 	int nont, ret;
1570 
1571 	rdmsrl(MSR_TURBO_RATIO_LIMIT, value);
1572 	nont = core_get_max_pstate();
1573 	ret = (((value) >> 8) & 0xFF);
1574 	if (ret <= nont)
1575 		ret = nont;
1576 	return ret;
1577 }
1578 
1579 static struct cpu_defaults core_params = {
1580 	.pid_policy = {
1581 		.sample_rate_ms = 10,
1582 		.deadband = 0,
1583 		.setpoint = 97,
1584 		.p_gain_pct = 20,
1585 		.d_gain_pct = 0,
1586 		.i_gain_pct = 0,
1587 	},
1588 	.funcs = {
1589 		.get_max = core_get_max_pstate,
1590 		.get_max_physical = core_get_max_pstate_physical,
1591 		.get_min = core_get_min_pstate,
1592 		.get_turbo = core_get_turbo_pstate,
1593 		.get_scaling = core_get_scaling,
1594 		.get_val = core_get_val,
1595 		.get_target_pstate = get_target_pstate_use_performance,
1596 	},
1597 };
1598 
1599 static const struct cpu_defaults silvermont_params = {
1600 	.pid_policy = {
1601 		.sample_rate_ms = 10,
1602 		.deadband = 0,
1603 		.setpoint = 60,
1604 		.p_gain_pct = 14,
1605 		.d_gain_pct = 0,
1606 		.i_gain_pct = 4,
1607 	},
1608 	.funcs = {
1609 		.get_max = atom_get_max_pstate,
1610 		.get_max_physical = atom_get_max_pstate,
1611 		.get_min = atom_get_min_pstate,
1612 		.get_turbo = atom_get_turbo_pstate,
1613 		.get_val = atom_get_val,
1614 		.get_scaling = silvermont_get_scaling,
1615 		.get_vid = atom_get_vid,
1616 		.get_target_pstate = get_target_pstate_use_cpu_load,
1617 	},
1618 };
1619 
1620 static const struct cpu_defaults airmont_params = {
1621 	.pid_policy = {
1622 		.sample_rate_ms = 10,
1623 		.deadband = 0,
1624 		.setpoint = 60,
1625 		.p_gain_pct = 14,
1626 		.d_gain_pct = 0,
1627 		.i_gain_pct = 4,
1628 	},
1629 	.funcs = {
1630 		.get_max = atom_get_max_pstate,
1631 		.get_max_physical = atom_get_max_pstate,
1632 		.get_min = atom_get_min_pstate,
1633 		.get_turbo = atom_get_turbo_pstate,
1634 		.get_val = atom_get_val,
1635 		.get_scaling = airmont_get_scaling,
1636 		.get_vid = atom_get_vid,
1637 		.get_target_pstate = get_target_pstate_use_cpu_load,
1638 	},
1639 };
1640 
1641 static const struct cpu_defaults knl_params = {
1642 	.pid_policy = {
1643 		.sample_rate_ms = 10,
1644 		.deadband = 0,
1645 		.setpoint = 97,
1646 		.p_gain_pct = 20,
1647 		.d_gain_pct = 0,
1648 		.i_gain_pct = 0,
1649 	},
1650 	.funcs = {
1651 		.get_max = core_get_max_pstate,
1652 		.get_max_physical = core_get_max_pstate_physical,
1653 		.get_min = core_get_min_pstate,
1654 		.get_turbo = knl_get_turbo_pstate,
1655 		.get_scaling = core_get_scaling,
1656 		.get_val = core_get_val,
1657 		.get_target_pstate = get_target_pstate_use_performance,
1658 	},
1659 };
1660 
1661 static const struct cpu_defaults bxt_params = {
1662 	.pid_policy = {
1663 		.sample_rate_ms = 10,
1664 		.deadband = 0,
1665 		.setpoint = 60,
1666 		.p_gain_pct = 14,
1667 		.d_gain_pct = 0,
1668 		.i_gain_pct = 4,
1669 	},
1670 	.funcs = {
1671 		.get_max = core_get_max_pstate,
1672 		.get_max_physical = core_get_max_pstate_physical,
1673 		.get_min = core_get_min_pstate,
1674 		.get_turbo = core_get_turbo_pstate,
1675 		.get_scaling = core_get_scaling,
1676 		.get_val = core_get_val,
1677 		.get_target_pstate = get_target_pstate_use_cpu_load,
1678 	},
1679 };
1680 
1681 static void intel_pstate_get_min_max(struct cpudata *cpu, int *min, int *max)
1682 {
1683 	int max_perf = cpu->pstate.turbo_pstate;
1684 	int max_perf_adj;
1685 	int min_perf;
1686 	struct perf_limits *perf_limits = limits;
1687 
1688 	if (limits->no_turbo || limits->turbo_disabled)
1689 		max_perf = cpu->pstate.max_pstate;
1690 
1691 	if (per_cpu_limits)
1692 		perf_limits = cpu->perf_limits;
1693 
1694 	/*
1695 	 * performance can be limited by user through sysfs, by cpufreq
1696 	 * policy, or by cpu specific default values determined through
1697 	 * experimentation.
1698 	 */
1699 	max_perf_adj = fp_ext_toint(max_perf * perf_limits->max_perf);
1700 	*max = clamp_t(int, max_perf_adj,
1701 			cpu->pstate.min_pstate, cpu->pstate.turbo_pstate);
1702 
1703 	min_perf = fp_ext_toint(max_perf * perf_limits->min_perf);
1704 	*min = clamp_t(int, min_perf, cpu->pstate.min_pstate, max_perf);
1705 }
1706 
1707 static void intel_pstate_set_pstate(struct cpudata *cpu, int pstate)
1708 {
1709 	trace_cpu_frequency(pstate * cpu->pstate.scaling, cpu->cpu);
1710 	cpu->pstate.current_pstate = pstate;
1711 	/*
1712 	 * Generally, there is no guarantee that this code will always run on
1713 	 * the CPU being updated, so force the register update to run on the
1714 	 * right CPU.
1715 	 */
1716 	wrmsrl_on_cpu(cpu->cpu, MSR_IA32_PERF_CTL,
1717 		      pstate_funcs.get_val(cpu, pstate));
1718 }
1719 
1720 static void intel_pstate_set_min_pstate(struct cpudata *cpu)
1721 {
1722 	intel_pstate_set_pstate(cpu, cpu->pstate.min_pstate);
1723 }
1724 
1725 static void intel_pstate_max_within_limits(struct cpudata *cpu)
1726 {
1727 	int min_pstate, max_pstate;
1728 
1729 	update_turbo_state();
1730 	intel_pstate_get_min_max(cpu, &min_pstate, &max_pstate);
1731 	intel_pstate_set_pstate(cpu, max_pstate);
1732 }
1733 
1734 static void intel_pstate_get_cpu_pstates(struct cpudata *cpu)
1735 {
1736 	cpu->pstate.min_pstate = pstate_funcs.get_min();
1737 	cpu->pstate.max_pstate = pstate_funcs.get_max();
1738 	cpu->pstate.max_pstate_physical = pstate_funcs.get_max_physical();
1739 	cpu->pstate.turbo_pstate = pstate_funcs.get_turbo();
1740 	cpu->pstate.scaling = pstate_funcs.get_scaling();
1741 	cpu->pstate.max_freq = cpu->pstate.max_pstate * cpu->pstate.scaling;
1742 	cpu->pstate.turbo_freq = cpu->pstate.turbo_pstate * cpu->pstate.scaling;
1743 
1744 	if (pstate_funcs.get_vid)
1745 		pstate_funcs.get_vid(cpu);
1746 
1747 	intel_pstate_set_min_pstate(cpu);
1748 }
1749 
1750 static inline void intel_pstate_calc_avg_perf(struct cpudata *cpu)
1751 {
1752 	struct sample *sample = &cpu->sample;
1753 
1754 	sample->core_avg_perf = div_ext_fp(sample->aperf, sample->mperf);
1755 }
1756 
1757 static inline bool intel_pstate_sample(struct cpudata *cpu, u64 time)
1758 {
1759 	u64 aperf, mperf;
1760 	unsigned long flags;
1761 	u64 tsc;
1762 
1763 	local_irq_save(flags);
1764 	rdmsrl(MSR_IA32_APERF, aperf);
1765 	rdmsrl(MSR_IA32_MPERF, mperf);
1766 	tsc = rdtsc();
1767 	if (cpu->prev_mperf == mperf || cpu->prev_tsc == tsc) {
1768 		local_irq_restore(flags);
1769 		return false;
1770 	}
1771 	local_irq_restore(flags);
1772 
1773 	cpu->last_sample_time = cpu->sample.time;
1774 	cpu->sample.time = time;
1775 	cpu->sample.aperf = aperf;
1776 	cpu->sample.mperf = mperf;
1777 	cpu->sample.tsc =  tsc;
1778 	cpu->sample.aperf -= cpu->prev_aperf;
1779 	cpu->sample.mperf -= cpu->prev_mperf;
1780 	cpu->sample.tsc -= cpu->prev_tsc;
1781 
1782 	cpu->prev_aperf = aperf;
1783 	cpu->prev_mperf = mperf;
1784 	cpu->prev_tsc = tsc;
1785 	/*
1786 	 * First time this function is invoked in a given cycle, all of the
1787 	 * previous sample data fields are equal to zero or stale and they must
1788 	 * be populated with meaningful numbers for things to work, so assume
1789 	 * that sample.time will always be reset before setting the utilization
1790 	 * update hook and make the caller skip the sample then.
1791 	 */
1792 	return !!cpu->last_sample_time;
1793 }
1794 
1795 static inline int32_t get_avg_frequency(struct cpudata *cpu)
1796 {
1797 	return mul_ext_fp(cpu->sample.core_avg_perf,
1798 			  cpu->pstate.max_pstate_physical * cpu->pstate.scaling);
1799 }
1800 
1801 static inline int32_t get_avg_pstate(struct cpudata *cpu)
1802 {
1803 	return mul_ext_fp(cpu->pstate.max_pstate_physical,
1804 			  cpu->sample.core_avg_perf);
1805 }
1806 
1807 static inline int32_t get_target_pstate_use_cpu_load(struct cpudata *cpu)
1808 {
1809 	struct sample *sample = &cpu->sample;
1810 	int32_t busy_frac, boost;
1811 	int target, avg_pstate;
1812 
1813 	busy_frac = div_fp(sample->mperf, sample->tsc);
1814 
1815 	boost = cpu->iowait_boost;
1816 	cpu->iowait_boost >>= 1;
1817 
1818 	if (busy_frac < boost)
1819 		busy_frac = boost;
1820 
1821 	sample->busy_scaled = busy_frac * 100;
1822 
1823 	target = limits->no_turbo || limits->turbo_disabled ?
1824 			cpu->pstate.max_pstate : cpu->pstate.turbo_pstate;
1825 	target += target >> 2;
1826 	target = mul_fp(target, busy_frac);
1827 	if (target < cpu->pstate.min_pstate)
1828 		target = cpu->pstate.min_pstate;
1829 
1830 	/*
1831 	 * If the average P-state during the previous cycle was higher than the
1832 	 * current target, add 50% of the difference to the target to reduce
1833 	 * possible performance oscillations and offset possible performance
1834 	 * loss related to moving the workload from one CPU to another within
1835 	 * a package/module.
1836 	 */
1837 	avg_pstate = get_avg_pstate(cpu);
1838 	if (avg_pstate > target)
1839 		target += (avg_pstate - target) >> 1;
1840 
1841 	return target;
1842 }
1843 
1844 static inline int32_t get_target_pstate_use_performance(struct cpudata *cpu)
1845 {
1846 	int32_t perf_scaled, max_pstate, current_pstate, sample_ratio;
1847 	u64 duration_ns;
1848 
1849 	/*
1850 	 * perf_scaled is the ratio of the average P-state during the last
1851 	 * sampling period to the P-state requested last time (in percent).
1852 	 *
1853 	 * That measures the system's response to the previous P-state
1854 	 * selection.
1855 	 */
1856 	max_pstate = cpu->pstate.max_pstate_physical;
1857 	current_pstate = cpu->pstate.current_pstate;
1858 	perf_scaled = mul_ext_fp(cpu->sample.core_avg_perf,
1859 			       div_fp(100 * max_pstate, current_pstate));
1860 
1861 	/*
1862 	 * Since our utilization update callback will not run unless we are
1863 	 * in C0, check if the actual elapsed time is significantly greater (3x)
1864 	 * than our sample interval.  If it is, then we were idle for a long
1865 	 * enough period of time to adjust our performance metric.
1866 	 */
1867 	duration_ns = cpu->sample.time - cpu->last_sample_time;
1868 	if ((s64)duration_ns > pid_params.sample_rate_ns * 3) {
1869 		sample_ratio = div_fp(pid_params.sample_rate_ns, duration_ns);
1870 		perf_scaled = mul_fp(perf_scaled, sample_ratio);
1871 	} else {
1872 		sample_ratio = div_fp(100 * cpu->sample.mperf, cpu->sample.tsc);
1873 		if (sample_ratio < int_tofp(1))
1874 			perf_scaled = 0;
1875 	}
1876 
1877 	cpu->sample.busy_scaled = perf_scaled;
1878 	return cpu->pstate.current_pstate - pid_calc(&cpu->pid, perf_scaled);
1879 }
1880 
1881 static int intel_pstate_prepare_request(struct cpudata *cpu, int pstate)
1882 {
1883 	int max_perf, min_perf;
1884 
1885 	intel_pstate_get_min_max(cpu, &min_perf, &max_perf);
1886 	pstate = clamp_t(int, pstate, min_perf, max_perf);
1887 	return pstate;
1888 }
1889 
1890 static void intel_pstate_update_pstate(struct cpudata *cpu, int pstate)
1891 {
1892 	if (pstate == cpu->pstate.current_pstate)
1893 		return;
1894 
1895 	cpu->pstate.current_pstate = pstate;
1896 	wrmsrl(MSR_IA32_PERF_CTL, pstate_funcs.get_val(cpu, pstate));
1897 }
1898 
1899 static inline void intel_pstate_adjust_busy_pstate(struct cpudata *cpu)
1900 {
1901 	int from, target_pstate;
1902 	struct sample *sample;
1903 
1904 	from = cpu->pstate.current_pstate;
1905 
1906 	target_pstate = cpu->policy == CPUFREQ_POLICY_PERFORMANCE ?
1907 		cpu->pstate.turbo_pstate : pstate_funcs.get_target_pstate(cpu);
1908 
1909 	update_turbo_state();
1910 
1911 	target_pstate = intel_pstate_prepare_request(cpu, target_pstate);
1912 	trace_cpu_frequency(target_pstate * cpu->pstate.scaling, cpu->cpu);
1913 	intel_pstate_update_pstate(cpu, target_pstate);
1914 
1915 	sample = &cpu->sample;
1916 	trace_pstate_sample(mul_ext_fp(100, sample->core_avg_perf),
1917 		fp_toint(sample->busy_scaled),
1918 		from,
1919 		cpu->pstate.current_pstate,
1920 		sample->mperf,
1921 		sample->aperf,
1922 		sample->tsc,
1923 		get_avg_frequency(cpu),
1924 		fp_toint(cpu->iowait_boost * 100));
1925 }
1926 
1927 static void intel_pstate_update_util(struct update_util_data *data, u64 time,
1928 				     unsigned int flags)
1929 {
1930 	struct cpudata *cpu = container_of(data, struct cpudata, update_util);
1931 	u64 delta_ns;
1932 
1933 	if (pstate_funcs.get_target_pstate == get_target_pstate_use_cpu_load) {
1934 		if (flags & SCHED_CPUFREQ_IOWAIT) {
1935 			cpu->iowait_boost = int_tofp(1);
1936 		} else if (cpu->iowait_boost) {
1937 			/* Clear iowait_boost if the CPU may have been idle. */
1938 			delta_ns = time - cpu->last_update;
1939 			if (delta_ns > TICK_NSEC)
1940 				cpu->iowait_boost = 0;
1941 		}
1942 		cpu->last_update = time;
1943 	}
1944 
1945 	delta_ns = time - cpu->sample.time;
1946 	if ((s64)delta_ns >= pid_params.sample_rate_ns) {
1947 		bool sample_taken = intel_pstate_sample(cpu, time);
1948 
1949 		if (sample_taken) {
1950 			intel_pstate_calc_avg_perf(cpu);
1951 			if (!hwp_active)
1952 				intel_pstate_adjust_busy_pstate(cpu);
1953 		}
1954 	}
1955 }
1956 
1957 #define ICPU(model, policy) \
1958 	{ X86_VENDOR_INTEL, 6, model, X86_FEATURE_APERFMPERF,\
1959 			(unsigned long)&policy }
1960 
1961 static const struct x86_cpu_id intel_pstate_cpu_ids[] = {
1962 	ICPU(INTEL_FAM6_SANDYBRIDGE, 		core_params),
1963 	ICPU(INTEL_FAM6_SANDYBRIDGE_X,		core_params),
1964 	ICPU(INTEL_FAM6_ATOM_SILVERMONT1,	silvermont_params),
1965 	ICPU(INTEL_FAM6_IVYBRIDGE,		core_params),
1966 	ICPU(INTEL_FAM6_HASWELL_CORE,		core_params),
1967 	ICPU(INTEL_FAM6_BROADWELL_CORE,		core_params),
1968 	ICPU(INTEL_FAM6_IVYBRIDGE_X,		core_params),
1969 	ICPU(INTEL_FAM6_HASWELL_X,		core_params),
1970 	ICPU(INTEL_FAM6_HASWELL_ULT,		core_params),
1971 	ICPU(INTEL_FAM6_HASWELL_GT3E,		core_params),
1972 	ICPU(INTEL_FAM6_BROADWELL_GT3E,		core_params),
1973 	ICPU(INTEL_FAM6_ATOM_AIRMONT,		airmont_params),
1974 	ICPU(INTEL_FAM6_SKYLAKE_MOBILE,		core_params),
1975 	ICPU(INTEL_FAM6_BROADWELL_X,		core_params),
1976 	ICPU(INTEL_FAM6_SKYLAKE_DESKTOP,	core_params),
1977 	ICPU(INTEL_FAM6_BROADWELL_XEON_D,	core_params),
1978 	ICPU(INTEL_FAM6_XEON_PHI_KNL,		knl_params),
1979 	ICPU(INTEL_FAM6_XEON_PHI_KNM,		knl_params),
1980 	ICPU(INTEL_FAM6_ATOM_GOLDMONT,		bxt_params),
1981 	{}
1982 };
1983 MODULE_DEVICE_TABLE(x86cpu, intel_pstate_cpu_ids);
1984 
1985 static const struct x86_cpu_id intel_pstate_cpu_oob_ids[] __initconst = {
1986 	ICPU(INTEL_FAM6_BROADWELL_XEON_D, core_params),
1987 	ICPU(INTEL_FAM6_BROADWELL_X, core_params),
1988 	ICPU(INTEL_FAM6_SKYLAKE_X, core_params),
1989 	{}
1990 };
1991 
1992 static const struct x86_cpu_id intel_pstate_cpu_ee_disable_ids[] = {
1993 	ICPU(INTEL_FAM6_KABYLAKE_DESKTOP, core_params),
1994 	{}
1995 };
1996 
1997 static int intel_pstate_init_cpu(unsigned int cpunum)
1998 {
1999 	struct cpudata *cpu;
2000 
2001 	cpu = all_cpu_data[cpunum];
2002 
2003 	if (!cpu) {
2004 		unsigned int size = sizeof(struct cpudata);
2005 
2006 		if (per_cpu_limits)
2007 			size += sizeof(struct perf_limits);
2008 
2009 		cpu = kzalloc(size, GFP_KERNEL);
2010 		if (!cpu)
2011 			return -ENOMEM;
2012 
2013 		all_cpu_data[cpunum] = cpu;
2014 		if (per_cpu_limits)
2015 			cpu->perf_limits = (struct perf_limits *)(cpu + 1);
2016 
2017 		cpu->epp_default = -EINVAL;
2018 		cpu->epp_powersave = -EINVAL;
2019 		cpu->epp_saved = -EINVAL;
2020 	}
2021 
2022 	cpu = all_cpu_data[cpunum];
2023 
2024 	cpu->cpu = cpunum;
2025 
2026 	if (hwp_active) {
2027 		const struct x86_cpu_id *id;
2028 
2029 		id = x86_match_cpu(intel_pstate_cpu_ee_disable_ids);
2030 		if (id)
2031 			intel_pstate_disable_ee(cpunum);
2032 
2033 		intel_pstate_hwp_enable(cpu);
2034 		pid_params.sample_rate_ms = 50;
2035 		pid_params.sample_rate_ns = 50 * NSEC_PER_MSEC;
2036 	}
2037 
2038 	intel_pstate_get_cpu_pstates(cpu);
2039 
2040 	intel_pstate_busy_pid_reset(cpu);
2041 
2042 	pr_debug("controlling: cpu %d\n", cpunum);
2043 
2044 	return 0;
2045 }
2046 
2047 static unsigned int intel_pstate_get(unsigned int cpu_num)
2048 {
2049 	struct cpudata *cpu = all_cpu_data[cpu_num];
2050 
2051 	return cpu ? get_avg_frequency(cpu) : 0;
2052 }
2053 
2054 static void intel_pstate_set_update_util_hook(unsigned int cpu_num)
2055 {
2056 	struct cpudata *cpu = all_cpu_data[cpu_num];
2057 
2058 	if (cpu->update_util_set)
2059 		return;
2060 
2061 	/* Prevent intel_pstate_update_util() from using stale data. */
2062 	cpu->sample.time = 0;
2063 	cpufreq_add_update_util_hook(cpu_num, &cpu->update_util,
2064 				     intel_pstate_update_util);
2065 	cpu->update_util_set = true;
2066 }
2067 
2068 static void intel_pstate_clear_update_util_hook(unsigned int cpu)
2069 {
2070 	struct cpudata *cpu_data = all_cpu_data[cpu];
2071 
2072 	if (!cpu_data->update_util_set)
2073 		return;
2074 
2075 	cpufreq_remove_update_util_hook(cpu);
2076 	cpu_data->update_util_set = false;
2077 	synchronize_sched();
2078 }
2079 
2080 static void intel_pstate_update_perf_limits(struct cpufreq_policy *policy,
2081 					    struct perf_limits *limits)
2082 {
2083 
2084 	limits->max_policy_pct = DIV_ROUND_UP(policy->max * 100,
2085 					      policy->cpuinfo.max_freq);
2086 	limits->max_policy_pct = clamp_t(int, limits->max_policy_pct, 0, 100);
2087 	if (policy->max == policy->min) {
2088 		limits->min_policy_pct = limits->max_policy_pct;
2089 	} else {
2090 		limits->min_policy_pct = DIV_ROUND_UP(policy->min * 100,
2091 						      policy->cpuinfo.max_freq);
2092 		limits->min_policy_pct = clamp_t(int, limits->min_policy_pct,
2093 						 0, 100);
2094 	}
2095 
2096 	/* Normalize user input to [min_policy_pct, max_policy_pct] */
2097 	limits->min_perf_pct = max(limits->min_policy_pct,
2098 				   limits->min_sysfs_pct);
2099 	limits->min_perf_pct = min(limits->max_policy_pct,
2100 				   limits->min_perf_pct);
2101 	limits->max_perf_pct = min(limits->max_policy_pct,
2102 				   limits->max_sysfs_pct);
2103 	limits->max_perf_pct = max(limits->min_policy_pct,
2104 				   limits->max_perf_pct);
2105 
2106 	/* Make sure min_perf_pct <= max_perf_pct */
2107 	limits->min_perf_pct = min(limits->max_perf_pct, limits->min_perf_pct);
2108 
2109 	limits->min_perf = div_ext_fp(limits->min_perf_pct, 100);
2110 	limits->max_perf = div_ext_fp(limits->max_perf_pct, 100);
2111 	limits->max_perf = round_up(limits->max_perf, EXT_FRAC_BITS);
2112 	limits->min_perf = round_up(limits->min_perf, EXT_FRAC_BITS);
2113 
2114 	pr_debug("cpu:%d max_perf_pct:%d min_perf_pct:%d\n", policy->cpu,
2115 		 limits->max_perf_pct, limits->min_perf_pct);
2116 }
2117 
2118 static int intel_pstate_set_policy(struct cpufreq_policy *policy)
2119 {
2120 	struct cpudata *cpu;
2121 	struct perf_limits *perf_limits = NULL;
2122 
2123 	if (!policy->cpuinfo.max_freq)
2124 		return -ENODEV;
2125 
2126 	pr_debug("set_policy cpuinfo.max %u policy->max %u\n",
2127 		 policy->cpuinfo.max_freq, policy->max);
2128 
2129 	cpu = all_cpu_data[policy->cpu];
2130 	cpu->policy = policy->policy;
2131 
2132 	if (cpu->pstate.max_pstate_physical > cpu->pstate.max_pstate &&
2133 	    policy->max < policy->cpuinfo.max_freq &&
2134 	    policy->max > cpu->pstate.max_pstate * cpu->pstate.scaling) {
2135 		pr_debug("policy->max > max non turbo frequency\n");
2136 		policy->max = policy->cpuinfo.max_freq;
2137 	}
2138 
2139 	if (per_cpu_limits)
2140 		perf_limits = cpu->perf_limits;
2141 
2142 	mutex_lock(&intel_pstate_limits_lock);
2143 
2144 	if (policy->policy == CPUFREQ_POLICY_PERFORMANCE) {
2145 		pr_debug("set performance\n");
2146 		if (!perf_limits) {
2147 			limits = &performance_limits;
2148 			perf_limits = limits;
2149 		}
2150 	} else {
2151 		pr_debug("set powersave\n");
2152 		if (!perf_limits) {
2153 			limits = &powersave_limits;
2154 			perf_limits = limits;
2155 		}
2156 
2157 	}
2158 
2159 	intel_pstate_update_perf_limits(policy, perf_limits);
2160 
2161 	if (cpu->policy == CPUFREQ_POLICY_PERFORMANCE) {
2162 		/*
2163 		 * NOHZ_FULL CPUs need this as the governor callback may not
2164 		 * be invoked on them.
2165 		 */
2166 		intel_pstate_clear_update_util_hook(policy->cpu);
2167 		intel_pstate_max_within_limits(cpu);
2168 	}
2169 
2170 	intel_pstate_set_update_util_hook(policy->cpu);
2171 
2172 	intel_pstate_hwp_set_policy(policy);
2173 
2174 	mutex_unlock(&intel_pstate_limits_lock);
2175 
2176 	return 0;
2177 }
2178 
2179 static int intel_pstate_verify_policy(struct cpufreq_policy *policy)
2180 {
2181 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2182 	struct perf_limits *perf_limits;
2183 
2184 	if (policy->policy == CPUFREQ_POLICY_PERFORMANCE)
2185 		perf_limits = &performance_limits;
2186 	else
2187 		perf_limits = &powersave_limits;
2188 
2189 	update_turbo_state();
2190 	policy->cpuinfo.max_freq = perf_limits->turbo_disabled ||
2191 					perf_limits->no_turbo ?
2192 					cpu->pstate.max_freq :
2193 					cpu->pstate.turbo_freq;
2194 
2195 	cpufreq_verify_within_cpu_limits(policy);
2196 
2197 	if (policy->policy != CPUFREQ_POLICY_POWERSAVE &&
2198 	    policy->policy != CPUFREQ_POLICY_PERFORMANCE)
2199 		return -EINVAL;
2200 
2201 	/* When per-CPU limits are used, sysfs limits are not used */
2202 	if (!per_cpu_limits) {
2203 		unsigned int max_freq, min_freq;
2204 
2205 		max_freq = policy->cpuinfo.max_freq *
2206 					perf_limits->max_sysfs_pct / 100;
2207 		min_freq = policy->cpuinfo.max_freq *
2208 					perf_limits->min_sysfs_pct / 100;
2209 		cpufreq_verify_within_limits(policy, min_freq, max_freq);
2210 	}
2211 
2212 	return 0;
2213 }
2214 
2215 static void intel_cpufreq_stop_cpu(struct cpufreq_policy *policy)
2216 {
2217 	intel_pstate_set_min_pstate(all_cpu_data[policy->cpu]);
2218 }
2219 
2220 static void intel_pstate_stop_cpu(struct cpufreq_policy *policy)
2221 {
2222 	pr_debug("CPU %d exiting\n", policy->cpu);
2223 
2224 	intel_pstate_clear_update_util_hook(policy->cpu);
2225 	if (hwp_active)
2226 		intel_pstate_hwp_save_state(policy);
2227 	else
2228 		intel_cpufreq_stop_cpu(policy);
2229 }
2230 
2231 static int intel_pstate_cpu_exit(struct cpufreq_policy *policy)
2232 {
2233 	intel_pstate_exit_perf_limits(policy);
2234 
2235 	policy->fast_switch_possible = false;
2236 
2237 	return 0;
2238 }
2239 
2240 static int __intel_pstate_cpu_init(struct cpufreq_policy *policy)
2241 {
2242 	struct cpudata *cpu;
2243 	int rc;
2244 
2245 	rc = intel_pstate_init_cpu(policy->cpu);
2246 	if (rc)
2247 		return rc;
2248 
2249 	cpu = all_cpu_data[policy->cpu];
2250 
2251 	if (per_cpu_limits)
2252 		intel_pstate_init_limits(cpu->perf_limits);
2253 
2254 	policy->min = cpu->pstate.min_pstate * cpu->pstate.scaling;
2255 	policy->max = cpu->pstate.turbo_pstate * cpu->pstate.scaling;
2256 
2257 	/* cpuinfo and default policy values */
2258 	policy->cpuinfo.min_freq = cpu->pstate.min_pstate * cpu->pstate.scaling;
2259 	update_turbo_state();
2260 	policy->cpuinfo.max_freq = limits->turbo_disabled ?
2261 			cpu->pstate.max_pstate : cpu->pstate.turbo_pstate;
2262 	policy->cpuinfo.max_freq *= cpu->pstate.scaling;
2263 
2264 	intel_pstate_init_acpi_perf_limits(policy);
2265 	cpumask_set_cpu(policy->cpu, policy->cpus);
2266 
2267 	policy->fast_switch_possible = true;
2268 
2269 	return 0;
2270 }
2271 
2272 static int intel_pstate_cpu_init(struct cpufreq_policy *policy)
2273 {
2274 	int ret = __intel_pstate_cpu_init(policy);
2275 
2276 	if (ret)
2277 		return ret;
2278 
2279 	policy->cpuinfo.transition_latency = CPUFREQ_ETERNAL;
2280 	if (limits->min_perf_pct == 100 && limits->max_perf_pct == 100)
2281 		policy->policy = CPUFREQ_POLICY_PERFORMANCE;
2282 	else
2283 		policy->policy = CPUFREQ_POLICY_POWERSAVE;
2284 
2285 	return 0;
2286 }
2287 
2288 static struct cpufreq_driver intel_pstate = {
2289 	.flags		= CPUFREQ_CONST_LOOPS,
2290 	.verify		= intel_pstate_verify_policy,
2291 	.setpolicy	= intel_pstate_set_policy,
2292 	.suspend	= intel_pstate_hwp_save_state,
2293 	.resume		= intel_pstate_resume,
2294 	.get		= intel_pstate_get,
2295 	.init		= intel_pstate_cpu_init,
2296 	.exit		= intel_pstate_cpu_exit,
2297 	.stop_cpu	= intel_pstate_stop_cpu,
2298 	.name		= "intel_pstate",
2299 };
2300 
2301 static int intel_cpufreq_verify_policy(struct cpufreq_policy *policy)
2302 {
2303 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2304 
2305 	update_turbo_state();
2306 	policy->cpuinfo.max_freq = limits->turbo_disabled ?
2307 			cpu->pstate.max_freq : cpu->pstate.turbo_freq;
2308 
2309 	cpufreq_verify_within_cpu_limits(policy);
2310 
2311 	return 0;
2312 }
2313 
2314 static unsigned int intel_cpufreq_turbo_update(struct cpudata *cpu,
2315 					       struct cpufreq_policy *policy,
2316 					       unsigned int target_freq)
2317 {
2318 	unsigned int max_freq;
2319 
2320 	update_turbo_state();
2321 
2322 	max_freq = limits->no_turbo || limits->turbo_disabled ?
2323 			cpu->pstate.max_freq : cpu->pstate.turbo_freq;
2324 	policy->cpuinfo.max_freq = max_freq;
2325 	if (policy->max > max_freq)
2326 		policy->max = max_freq;
2327 
2328 	if (target_freq > max_freq)
2329 		target_freq = max_freq;
2330 
2331 	return target_freq;
2332 }
2333 
2334 static int intel_cpufreq_target(struct cpufreq_policy *policy,
2335 				unsigned int target_freq,
2336 				unsigned int relation)
2337 {
2338 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2339 	struct cpufreq_freqs freqs;
2340 	int target_pstate;
2341 
2342 	freqs.old = policy->cur;
2343 	freqs.new = intel_cpufreq_turbo_update(cpu, policy, target_freq);
2344 
2345 	cpufreq_freq_transition_begin(policy, &freqs);
2346 	switch (relation) {
2347 	case CPUFREQ_RELATION_L:
2348 		target_pstate = DIV_ROUND_UP(freqs.new, cpu->pstate.scaling);
2349 		break;
2350 	case CPUFREQ_RELATION_H:
2351 		target_pstate = freqs.new / cpu->pstate.scaling;
2352 		break;
2353 	default:
2354 		target_pstate = DIV_ROUND_CLOSEST(freqs.new, cpu->pstate.scaling);
2355 		break;
2356 	}
2357 	target_pstate = intel_pstate_prepare_request(cpu, target_pstate);
2358 	if (target_pstate != cpu->pstate.current_pstate) {
2359 		cpu->pstate.current_pstate = target_pstate;
2360 		wrmsrl_on_cpu(policy->cpu, MSR_IA32_PERF_CTL,
2361 			      pstate_funcs.get_val(cpu, target_pstate));
2362 	}
2363 	freqs.new = target_pstate * cpu->pstate.scaling;
2364 	cpufreq_freq_transition_end(policy, &freqs, false);
2365 
2366 	return 0;
2367 }
2368 
2369 static unsigned int intel_cpufreq_fast_switch(struct cpufreq_policy *policy,
2370 					      unsigned int target_freq)
2371 {
2372 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2373 	int target_pstate;
2374 
2375 	target_freq = intel_cpufreq_turbo_update(cpu, policy, target_freq);
2376 	target_pstate = DIV_ROUND_UP(target_freq, cpu->pstate.scaling);
2377 	target_pstate = intel_pstate_prepare_request(cpu, target_pstate);
2378 	intel_pstate_update_pstate(cpu, target_pstate);
2379 	return target_pstate * cpu->pstate.scaling;
2380 }
2381 
2382 static int intel_cpufreq_cpu_init(struct cpufreq_policy *policy)
2383 {
2384 	int ret = __intel_pstate_cpu_init(policy);
2385 
2386 	if (ret)
2387 		return ret;
2388 
2389 	policy->cpuinfo.transition_latency = INTEL_CPUFREQ_TRANSITION_LATENCY;
2390 	/* This reflects the intel_pstate_get_cpu_pstates() setting. */
2391 	policy->cur = policy->cpuinfo.min_freq;
2392 
2393 	return 0;
2394 }
2395 
2396 static struct cpufreq_driver intel_cpufreq = {
2397 	.flags		= CPUFREQ_CONST_LOOPS,
2398 	.verify		= intel_cpufreq_verify_policy,
2399 	.target		= intel_cpufreq_target,
2400 	.fast_switch	= intel_cpufreq_fast_switch,
2401 	.init		= intel_cpufreq_cpu_init,
2402 	.exit		= intel_pstate_cpu_exit,
2403 	.stop_cpu	= intel_cpufreq_stop_cpu,
2404 	.name		= "intel_cpufreq",
2405 };
2406 
2407 static struct cpufreq_driver *intel_pstate_driver = &intel_pstate;
2408 
2409 static void intel_pstate_driver_cleanup(void)
2410 {
2411 	unsigned int cpu;
2412 
2413 	get_online_cpus();
2414 	for_each_online_cpu(cpu) {
2415 		if (all_cpu_data[cpu]) {
2416 			if (intel_pstate_driver == &intel_pstate)
2417 				intel_pstate_clear_update_util_hook(cpu);
2418 
2419 			kfree(all_cpu_data[cpu]);
2420 			all_cpu_data[cpu] = NULL;
2421 		}
2422 	}
2423 	put_online_cpus();
2424 }
2425 
2426 static int intel_pstate_register_driver(void)
2427 {
2428 	int ret;
2429 
2430 	intel_pstate_init_limits(&powersave_limits);
2431 	intel_pstate_set_performance_limits(&performance_limits);
2432 	if (IS_ENABLED(CONFIG_CPU_FREQ_DEFAULT_GOV_PERFORMANCE) &&
2433 	    intel_pstate_driver == &intel_pstate)
2434 		limits = &performance_limits;
2435 	else
2436 		limits = &powersave_limits;
2437 
2438 	ret = cpufreq_register_driver(intel_pstate_driver);
2439 	if (ret) {
2440 		intel_pstate_driver_cleanup();
2441 		return ret;
2442 	}
2443 
2444 	mutex_lock(&intel_pstate_limits_lock);
2445 	driver_registered = true;
2446 	mutex_unlock(&intel_pstate_limits_lock);
2447 
2448 	if (intel_pstate_driver == &intel_pstate && !hwp_active &&
2449 	    pstate_funcs.get_target_pstate != get_target_pstate_use_cpu_load)
2450 		intel_pstate_debug_expose_params();
2451 
2452 	return 0;
2453 }
2454 
2455 static int intel_pstate_unregister_driver(void)
2456 {
2457 	if (hwp_active)
2458 		return -EBUSY;
2459 
2460 	if (intel_pstate_driver == &intel_pstate && !hwp_active &&
2461 	    pstate_funcs.get_target_pstate != get_target_pstate_use_cpu_load)
2462 		intel_pstate_debug_hide_params();
2463 
2464 	mutex_lock(&intel_pstate_limits_lock);
2465 	driver_registered = false;
2466 	mutex_unlock(&intel_pstate_limits_lock);
2467 
2468 	cpufreq_unregister_driver(intel_pstate_driver);
2469 	intel_pstate_driver_cleanup();
2470 
2471 	return 0;
2472 }
2473 
2474 static ssize_t intel_pstate_show_status(char *buf)
2475 {
2476 	if (!driver_registered)
2477 		return sprintf(buf, "off\n");
2478 
2479 	return sprintf(buf, "%s\n", intel_pstate_driver == &intel_pstate ?
2480 					"active" : "passive");
2481 }
2482 
2483 static int intel_pstate_update_status(const char *buf, size_t size)
2484 {
2485 	int ret;
2486 
2487 	if (size == 3 && !strncmp(buf, "off", size))
2488 		return driver_registered ?
2489 			intel_pstate_unregister_driver() : -EINVAL;
2490 
2491 	if (size == 6 && !strncmp(buf, "active", size)) {
2492 		if (driver_registered) {
2493 			if (intel_pstate_driver == &intel_pstate)
2494 				return 0;
2495 
2496 			ret = intel_pstate_unregister_driver();
2497 			if (ret)
2498 				return ret;
2499 		}
2500 
2501 		intel_pstate_driver = &intel_pstate;
2502 		return intel_pstate_register_driver();
2503 	}
2504 
2505 	if (size == 7 && !strncmp(buf, "passive", size)) {
2506 		if (driver_registered) {
2507 			if (intel_pstate_driver != &intel_pstate)
2508 				return 0;
2509 
2510 			ret = intel_pstate_unregister_driver();
2511 			if (ret)
2512 				return ret;
2513 		}
2514 
2515 		intel_pstate_driver = &intel_cpufreq;
2516 		return intel_pstate_register_driver();
2517 	}
2518 
2519 	return -EINVAL;
2520 }
2521 
2522 static int no_load __initdata;
2523 static int no_hwp __initdata;
2524 static int hwp_only __initdata;
2525 static unsigned int force_load __initdata;
2526 
2527 static int __init intel_pstate_msrs_not_valid(void)
2528 {
2529 	if (!pstate_funcs.get_max() ||
2530 	    !pstate_funcs.get_min() ||
2531 	    !pstate_funcs.get_turbo())
2532 		return -ENODEV;
2533 
2534 	return 0;
2535 }
2536 
2537 static void __init copy_pid_params(struct pstate_adjust_policy *policy)
2538 {
2539 	pid_params.sample_rate_ms = policy->sample_rate_ms;
2540 	pid_params.sample_rate_ns = pid_params.sample_rate_ms * NSEC_PER_MSEC;
2541 	pid_params.p_gain_pct = policy->p_gain_pct;
2542 	pid_params.i_gain_pct = policy->i_gain_pct;
2543 	pid_params.d_gain_pct = policy->d_gain_pct;
2544 	pid_params.deadband = policy->deadband;
2545 	pid_params.setpoint = policy->setpoint;
2546 }
2547 
2548 #ifdef CONFIG_ACPI
2549 static void intel_pstate_use_acpi_profile(void)
2550 {
2551 	if (acpi_gbl_FADT.preferred_profile == PM_MOBILE)
2552 		pstate_funcs.get_target_pstate =
2553 				get_target_pstate_use_cpu_load;
2554 }
2555 #else
2556 static void intel_pstate_use_acpi_profile(void)
2557 {
2558 }
2559 #endif
2560 
2561 static void __init copy_cpu_funcs(struct pstate_funcs *funcs)
2562 {
2563 	pstate_funcs.get_max   = funcs->get_max;
2564 	pstate_funcs.get_max_physical = funcs->get_max_physical;
2565 	pstate_funcs.get_min   = funcs->get_min;
2566 	pstate_funcs.get_turbo = funcs->get_turbo;
2567 	pstate_funcs.get_scaling = funcs->get_scaling;
2568 	pstate_funcs.get_val   = funcs->get_val;
2569 	pstate_funcs.get_vid   = funcs->get_vid;
2570 	pstate_funcs.get_target_pstate = funcs->get_target_pstate;
2571 
2572 	intel_pstate_use_acpi_profile();
2573 }
2574 
2575 #ifdef CONFIG_ACPI
2576 
2577 static bool __init intel_pstate_no_acpi_pss(void)
2578 {
2579 	int i;
2580 
2581 	for_each_possible_cpu(i) {
2582 		acpi_status status;
2583 		union acpi_object *pss;
2584 		struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
2585 		struct acpi_processor *pr = per_cpu(processors, i);
2586 
2587 		if (!pr)
2588 			continue;
2589 
2590 		status = acpi_evaluate_object(pr->handle, "_PSS", NULL, &buffer);
2591 		if (ACPI_FAILURE(status))
2592 			continue;
2593 
2594 		pss = buffer.pointer;
2595 		if (pss && pss->type == ACPI_TYPE_PACKAGE) {
2596 			kfree(pss);
2597 			return false;
2598 		}
2599 
2600 		kfree(pss);
2601 	}
2602 
2603 	return true;
2604 }
2605 
2606 static bool __init intel_pstate_has_acpi_ppc(void)
2607 {
2608 	int i;
2609 
2610 	for_each_possible_cpu(i) {
2611 		struct acpi_processor *pr = per_cpu(processors, i);
2612 
2613 		if (!pr)
2614 			continue;
2615 		if (acpi_has_method(pr->handle, "_PPC"))
2616 			return true;
2617 	}
2618 	return false;
2619 }
2620 
2621 enum {
2622 	PSS,
2623 	PPC,
2624 };
2625 
2626 struct hw_vendor_info {
2627 	u16  valid;
2628 	char oem_id[ACPI_OEM_ID_SIZE];
2629 	char oem_table_id[ACPI_OEM_TABLE_ID_SIZE];
2630 	int  oem_pwr_table;
2631 };
2632 
2633 /* Hardware vendor-specific info that has its own power management modes */
2634 static struct hw_vendor_info vendor_info[] __initdata = {
2635 	{1, "HP    ", "ProLiant", PSS},
2636 	{1, "ORACLE", "X4-2    ", PPC},
2637 	{1, "ORACLE", "X4-2L   ", PPC},
2638 	{1, "ORACLE", "X4-2B   ", PPC},
2639 	{1, "ORACLE", "X3-2    ", PPC},
2640 	{1, "ORACLE", "X3-2L   ", PPC},
2641 	{1, "ORACLE", "X3-2B   ", PPC},
2642 	{1, "ORACLE", "X4470M2 ", PPC},
2643 	{1, "ORACLE", "X4270M3 ", PPC},
2644 	{1, "ORACLE", "X4270M2 ", PPC},
2645 	{1, "ORACLE", "X4170M2 ", PPC},
2646 	{1, "ORACLE", "X4170 M3", PPC},
2647 	{1, "ORACLE", "X4275 M3", PPC},
2648 	{1, "ORACLE", "X6-2    ", PPC},
2649 	{1, "ORACLE", "Sudbury ", PPC},
2650 	{0, "", ""},
2651 };
2652 
2653 static bool __init intel_pstate_platform_pwr_mgmt_exists(void)
2654 {
2655 	struct acpi_table_header hdr;
2656 	struct hw_vendor_info *v_info;
2657 	const struct x86_cpu_id *id;
2658 	u64 misc_pwr;
2659 
2660 	id = x86_match_cpu(intel_pstate_cpu_oob_ids);
2661 	if (id) {
2662 		rdmsrl(MSR_MISC_PWR_MGMT, misc_pwr);
2663 		if ( misc_pwr & (1 << 8))
2664 			return true;
2665 	}
2666 
2667 	if (acpi_disabled ||
2668 	    ACPI_FAILURE(acpi_get_table_header(ACPI_SIG_FADT, 0, &hdr)))
2669 		return false;
2670 
2671 	for (v_info = vendor_info; v_info->valid; v_info++) {
2672 		if (!strncmp(hdr.oem_id, v_info->oem_id, ACPI_OEM_ID_SIZE) &&
2673 			!strncmp(hdr.oem_table_id, v_info->oem_table_id,
2674 						ACPI_OEM_TABLE_ID_SIZE))
2675 			switch (v_info->oem_pwr_table) {
2676 			case PSS:
2677 				return intel_pstate_no_acpi_pss();
2678 			case PPC:
2679 				return intel_pstate_has_acpi_ppc() &&
2680 					(!force_load);
2681 			}
2682 	}
2683 
2684 	return false;
2685 }
2686 
2687 static void intel_pstate_request_control_from_smm(void)
2688 {
2689 	/*
2690 	 * It may be unsafe to request P-states control from SMM if _PPC support
2691 	 * has not been enabled.
2692 	 */
2693 	if (acpi_ppc)
2694 		acpi_processor_pstate_control();
2695 }
2696 #else /* CONFIG_ACPI not enabled */
2697 static inline bool intel_pstate_platform_pwr_mgmt_exists(void) { return false; }
2698 static inline bool intel_pstate_has_acpi_ppc(void) { return false; }
2699 static inline void intel_pstate_request_control_from_smm(void) {}
2700 #endif /* CONFIG_ACPI */
2701 
2702 static const struct x86_cpu_id hwp_support_ids[] __initconst = {
2703 	{ X86_VENDOR_INTEL, 6, X86_MODEL_ANY, X86_FEATURE_HWP },
2704 	{}
2705 };
2706 
2707 static int __init intel_pstate_init(void)
2708 {
2709 	const struct x86_cpu_id *id;
2710 	struct cpu_defaults *cpu_def;
2711 	int rc = 0;
2712 
2713 	if (no_load)
2714 		return -ENODEV;
2715 
2716 	if (x86_match_cpu(hwp_support_ids) && !no_hwp) {
2717 		copy_cpu_funcs(&core_params.funcs);
2718 		hwp_active++;
2719 		intel_pstate.attr = hwp_cpufreq_attrs;
2720 		goto hwp_cpu_matched;
2721 	}
2722 
2723 	id = x86_match_cpu(intel_pstate_cpu_ids);
2724 	if (!id)
2725 		return -ENODEV;
2726 
2727 	cpu_def = (struct cpu_defaults *)id->driver_data;
2728 
2729 	copy_pid_params(&cpu_def->pid_policy);
2730 	copy_cpu_funcs(&cpu_def->funcs);
2731 
2732 	if (intel_pstate_msrs_not_valid())
2733 		return -ENODEV;
2734 
2735 hwp_cpu_matched:
2736 	/*
2737 	 * The Intel pstate driver will be ignored if the platform
2738 	 * firmware has its own power management modes.
2739 	 */
2740 	if (intel_pstate_platform_pwr_mgmt_exists())
2741 		return -ENODEV;
2742 
2743 	if (!hwp_active && hwp_only)
2744 		return -ENOTSUPP;
2745 
2746 	pr_info("Intel P-state driver initializing\n");
2747 
2748 	all_cpu_data = vzalloc(sizeof(void *) * num_possible_cpus());
2749 	if (!all_cpu_data)
2750 		return -ENOMEM;
2751 
2752 	intel_pstate_request_control_from_smm();
2753 
2754 	intel_pstate_sysfs_expose_params();
2755 
2756 	mutex_lock(&intel_pstate_driver_lock);
2757 	rc = intel_pstate_register_driver();
2758 	mutex_unlock(&intel_pstate_driver_lock);
2759 	if (rc)
2760 		return rc;
2761 
2762 	if (hwp_active)
2763 		pr_info("HWP enabled\n");
2764 
2765 	return 0;
2766 }
2767 device_initcall(intel_pstate_init);
2768 
2769 static int __init intel_pstate_setup(char *str)
2770 {
2771 	if (!str)
2772 		return -EINVAL;
2773 
2774 	if (!strcmp(str, "disable")) {
2775 		no_load = 1;
2776 	} else if (!strcmp(str, "passive")) {
2777 		pr_info("Passive mode enabled\n");
2778 		intel_pstate_driver = &intel_cpufreq;
2779 		no_hwp = 1;
2780 	}
2781 	if (!strcmp(str, "no_hwp")) {
2782 		pr_info("HWP disabled\n");
2783 		no_hwp = 1;
2784 	}
2785 	if (!strcmp(str, "force"))
2786 		force_load = 1;
2787 	if (!strcmp(str, "hwp_only"))
2788 		hwp_only = 1;
2789 	if (!strcmp(str, "per_cpu_perf_limits"))
2790 		per_cpu_limits = true;
2791 
2792 #ifdef CONFIG_ACPI
2793 	if (!strcmp(str, "support_acpi_ppc"))
2794 		acpi_ppc = true;
2795 #endif
2796 
2797 	return 0;
2798 }
2799 early_param("intel_pstate", intel_pstate_setup);
2800 
2801 MODULE_AUTHOR("Dirk Brandewie <dirk.j.brandewie@intel.com>");
2802 MODULE_DESCRIPTION("'intel_pstate' - P state driver Intel Core processors");
2803 MODULE_LICENSE("GPL");
2804