xref: /linux/drivers/cpufreq/intel_pstate.c (revision fb1fe1041c04ee5ba362cf239e83a7c559beb0f3)
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 static bool driver_registered __read_mostly;
362 
363 #ifdef CONFIG_ACPI
364 static bool acpi_ppc;
365 #endif
366 
367 static struct perf_limits performance_limits = {
368 	.no_turbo = 0,
369 	.turbo_disabled = 0,
370 	.max_perf_pct = 100,
371 	.max_perf = int_ext_tofp(1),
372 	.min_perf_pct = 100,
373 	.min_perf = int_ext_tofp(1),
374 	.max_policy_pct = 100,
375 	.max_sysfs_pct = 100,
376 	.min_policy_pct = 0,
377 	.min_sysfs_pct = 0,
378 };
379 
380 static struct perf_limits powersave_limits = {
381 	.no_turbo = 0,
382 	.turbo_disabled = 0,
383 	.max_perf_pct = 100,
384 	.max_perf = int_ext_tofp(1),
385 	.min_perf_pct = 0,
386 	.min_perf = 0,
387 	.max_policy_pct = 100,
388 	.max_sysfs_pct = 100,
389 	.min_policy_pct = 0,
390 	.min_sysfs_pct = 0,
391 };
392 
393 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_PERFORMANCE
394 static struct perf_limits *limits = &performance_limits;
395 #else
396 static struct perf_limits *limits = &powersave_limits;
397 #endif
398 
399 static DEFINE_MUTEX(intel_pstate_driver_lock);
400 static DEFINE_MUTEX(intel_pstate_limits_lock);
401 
402 #ifdef CONFIG_ACPI
403 
404 static bool intel_pstate_get_ppc_enable_status(void)
405 {
406 	if (acpi_gbl_FADT.preferred_profile == PM_ENTERPRISE_SERVER ||
407 	    acpi_gbl_FADT.preferred_profile == PM_PERFORMANCE_SERVER)
408 		return true;
409 
410 	return acpi_ppc;
411 }
412 
413 #ifdef CONFIG_ACPI_CPPC_LIB
414 
415 /* The work item is needed to avoid CPU hotplug locking issues */
416 static void intel_pstste_sched_itmt_work_fn(struct work_struct *work)
417 {
418 	sched_set_itmt_support();
419 }
420 
421 static DECLARE_WORK(sched_itmt_work, intel_pstste_sched_itmt_work_fn);
422 
423 static void intel_pstate_set_itmt_prio(int cpu)
424 {
425 	struct cppc_perf_caps cppc_perf;
426 	static u32 max_highest_perf = 0, min_highest_perf = U32_MAX;
427 	int ret;
428 
429 	ret = cppc_get_perf_caps(cpu, &cppc_perf);
430 	if (ret)
431 		return;
432 
433 	/*
434 	 * The priorities can be set regardless of whether or not
435 	 * sched_set_itmt_support(true) has been called and it is valid to
436 	 * update them at any time after it has been called.
437 	 */
438 	sched_set_itmt_core_prio(cppc_perf.highest_perf, cpu);
439 
440 	if (max_highest_perf <= min_highest_perf) {
441 		if (cppc_perf.highest_perf > max_highest_perf)
442 			max_highest_perf = cppc_perf.highest_perf;
443 
444 		if (cppc_perf.highest_perf < min_highest_perf)
445 			min_highest_perf = cppc_perf.highest_perf;
446 
447 		if (max_highest_perf > min_highest_perf) {
448 			/*
449 			 * This code can be run during CPU online under the
450 			 * CPU hotplug locks, so sched_set_itmt_support()
451 			 * cannot be called from here.  Queue up a work item
452 			 * to invoke it.
453 			 */
454 			schedule_work(&sched_itmt_work);
455 		}
456 	}
457 }
458 #else
459 static void intel_pstate_set_itmt_prio(int cpu)
460 {
461 }
462 #endif
463 
464 static void intel_pstate_init_acpi_perf_limits(struct cpufreq_policy *policy)
465 {
466 	struct cpudata *cpu;
467 	int ret;
468 	int i;
469 
470 	if (hwp_active) {
471 		intel_pstate_set_itmt_prio(policy->cpu);
472 		return;
473 	}
474 
475 	if (!intel_pstate_get_ppc_enable_status())
476 		return;
477 
478 	cpu = all_cpu_data[policy->cpu];
479 
480 	ret = acpi_processor_register_performance(&cpu->acpi_perf_data,
481 						  policy->cpu);
482 	if (ret)
483 		return;
484 
485 	/*
486 	 * Check if the control value in _PSS is for PERF_CTL MSR, which should
487 	 * guarantee that the states returned by it map to the states in our
488 	 * list directly.
489 	 */
490 	if (cpu->acpi_perf_data.control_register.space_id !=
491 						ACPI_ADR_SPACE_FIXED_HARDWARE)
492 		goto err;
493 
494 	/*
495 	 * If there is only one entry _PSS, simply ignore _PSS and continue as
496 	 * usual without taking _PSS into account
497 	 */
498 	if (cpu->acpi_perf_data.state_count < 2)
499 		goto err;
500 
501 	pr_debug("CPU%u - ACPI _PSS perf data\n", policy->cpu);
502 	for (i = 0; i < cpu->acpi_perf_data.state_count; i++) {
503 		pr_debug("     %cP%d: %u MHz, %u mW, 0x%x\n",
504 			 (i == cpu->acpi_perf_data.state ? '*' : ' '), i,
505 			 (u32) cpu->acpi_perf_data.states[i].core_frequency,
506 			 (u32) cpu->acpi_perf_data.states[i].power,
507 			 (u32) cpu->acpi_perf_data.states[i].control);
508 	}
509 
510 	/*
511 	 * The _PSS table doesn't contain whole turbo frequency range.
512 	 * This just contains +1 MHZ above the max non turbo frequency,
513 	 * with control value corresponding to max turbo ratio. But
514 	 * when cpufreq set policy is called, it will call with this
515 	 * max frequency, which will cause a reduced performance as
516 	 * this driver uses real max turbo frequency as the max
517 	 * frequency. So correct this frequency in _PSS table to
518 	 * correct max turbo frequency based on the turbo state.
519 	 * Also need to convert to MHz as _PSS freq is in MHz.
520 	 */
521 	if (!limits->turbo_disabled)
522 		cpu->acpi_perf_data.states[0].core_frequency =
523 					policy->cpuinfo.max_freq / 1000;
524 	cpu->valid_pss_table = true;
525 	pr_debug("_PPC limits will be enforced\n");
526 
527 	return;
528 
529  err:
530 	cpu->valid_pss_table = false;
531 	acpi_processor_unregister_performance(policy->cpu);
532 }
533 
534 static void intel_pstate_exit_perf_limits(struct cpufreq_policy *policy)
535 {
536 	struct cpudata *cpu;
537 
538 	cpu = all_cpu_data[policy->cpu];
539 	if (!cpu->valid_pss_table)
540 		return;
541 
542 	acpi_processor_unregister_performance(policy->cpu);
543 }
544 #else
545 static inline void intel_pstate_init_acpi_perf_limits(struct cpufreq_policy *policy)
546 {
547 }
548 
549 static inline void intel_pstate_exit_perf_limits(struct cpufreq_policy *policy)
550 {
551 }
552 #endif
553 
554 static inline void pid_reset(struct _pid *pid, int setpoint, int busy,
555 			     int deadband, int integral) {
556 	pid->setpoint = int_tofp(setpoint);
557 	pid->deadband  = int_tofp(deadband);
558 	pid->integral  = int_tofp(integral);
559 	pid->last_err  = int_tofp(setpoint) - int_tofp(busy);
560 }
561 
562 static inline void pid_p_gain_set(struct _pid *pid, int percent)
563 {
564 	pid->p_gain = div_fp(percent, 100);
565 }
566 
567 static inline void pid_i_gain_set(struct _pid *pid, int percent)
568 {
569 	pid->i_gain = div_fp(percent, 100);
570 }
571 
572 static inline void pid_d_gain_set(struct _pid *pid, int percent)
573 {
574 	pid->d_gain = div_fp(percent, 100);
575 }
576 
577 static signed int pid_calc(struct _pid *pid, int32_t busy)
578 {
579 	signed int result;
580 	int32_t pterm, dterm, fp_error;
581 	int32_t integral_limit;
582 
583 	fp_error = pid->setpoint - busy;
584 
585 	if (abs(fp_error) <= pid->deadband)
586 		return 0;
587 
588 	pterm = mul_fp(pid->p_gain, fp_error);
589 
590 	pid->integral += fp_error;
591 
592 	/*
593 	 * We limit the integral here so that it will never
594 	 * get higher than 30.  This prevents it from becoming
595 	 * too large an input over long periods of time and allows
596 	 * it to get factored out sooner.
597 	 *
598 	 * The value of 30 was chosen through experimentation.
599 	 */
600 	integral_limit = int_tofp(30);
601 	if (pid->integral > integral_limit)
602 		pid->integral = integral_limit;
603 	if (pid->integral < -integral_limit)
604 		pid->integral = -integral_limit;
605 
606 	dterm = mul_fp(pid->d_gain, fp_error - pid->last_err);
607 	pid->last_err = fp_error;
608 
609 	result = pterm + mul_fp(pid->integral, pid->i_gain) + dterm;
610 	result = result + (1 << (FRAC_BITS-1));
611 	return (signed int)fp_toint(result);
612 }
613 
614 static inline void intel_pstate_busy_pid_reset(struct cpudata *cpu)
615 {
616 	pid_p_gain_set(&cpu->pid, pid_params.p_gain_pct);
617 	pid_d_gain_set(&cpu->pid, pid_params.d_gain_pct);
618 	pid_i_gain_set(&cpu->pid, pid_params.i_gain_pct);
619 
620 	pid_reset(&cpu->pid, pid_params.setpoint, 100, pid_params.deadband, 0);
621 }
622 
623 static inline void intel_pstate_reset_all_pid(void)
624 {
625 	unsigned int cpu;
626 
627 	for_each_online_cpu(cpu) {
628 		if (all_cpu_data[cpu])
629 			intel_pstate_busy_pid_reset(all_cpu_data[cpu]);
630 	}
631 }
632 
633 static inline void update_turbo_state(void)
634 {
635 	u64 misc_en;
636 	struct cpudata *cpu;
637 
638 	cpu = all_cpu_data[0];
639 	rdmsrl(MSR_IA32_MISC_ENABLE, misc_en);
640 	limits->turbo_disabled =
641 		(misc_en & MSR_IA32_MISC_ENABLE_TURBO_DISABLE ||
642 		 cpu->pstate.max_pstate == cpu->pstate.turbo_pstate);
643 }
644 
645 static s16 intel_pstate_get_epb(struct cpudata *cpu_data)
646 {
647 	u64 epb;
648 	int ret;
649 
650 	if (!static_cpu_has(X86_FEATURE_EPB))
651 		return -ENXIO;
652 
653 	ret = rdmsrl_on_cpu(cpu_data->cpu, MSR_IA32_ENERGY_PERF_BIAS, &epb);
654 	if (ret)
655 		return (s16)ret;
656 
657 	return (s16)(epb & 0x0f);
658 }
659 
660 static s16 intel_pstate_get_epp(struct cpudata *cpu_data, u64 hwp_req_data)
661 {
662 	s16 epp;
663 
664 	if (static_cpu_has(X86_FEATURE_HWP_EPP)) {
665 		/*
666 		 * When hwp_req_data is 0, means that caller didn't read
667 		 * MSR_HWP_REQUEST, so need to read and get EPP.
668 		 */
669 		if (!hwp_req_data) {
670 			epp = rdmsrl_on_cpu(cpu_data->cpu, MSR_HWP_REQUEST,
671 					    &hwp_req_data);
672 			if (epp)
673 				return epp;
674 		}
675 		epp = (hwp_req_data >> 24) & 0xff;
676 	} else {
677 		/* When there is no EPP present, HWP uses EPB settings */
678 		epp = intel_pstate_get_epb(cpu_data);
679 	}
680 
681 	return epp;
682 }
683 
684 static int intel_pstate_set_epb(int cpu, s16 pref)
685 {
686 	u64 epb;
687 	int ret;
688 
689 	if (!static_cpu_has(X86_FEATURE_EPB))
690 		return -ENXIO;
691 
692 	ret = rdmsrl_on_cpu(cpu, MSR_IA32_ENERGY_PERF_BIAS, &epb);
693 	if (ret)
694 		return ret;
695 
696 	epb = (epb & ~0x0f) | pref;
697 	wrmsrl_on_cpu(cpu, MSR_IA32_ENERGY_PERF_BIAS, epb);
698 
699 	return 0;
700 }
701 
702 /*
703  * EPP/EPB display strings corresponding to EPP index in the
704  * energy_perf_strings[]
705  *	index		String
706  *-------------------------------------
707  *	0		default
708  *	1		performance
709  *	2		balance_performance
710  *	3		balance_power
711  *	4		power
712  */
713 static const char * const energy_perf_strings[] = {
714 	"default",
715 	"performance",
716 	"balance_performance",
717 	"balance_power",
718 	"power",
719 	NULL
720 };
721 
722 static int intel_pstate_get_energy_pref_index(struct cpudata *cpu_data)
723 {
724 	s16 epp;
725 	int index = -EINVAL;
726 
727 	epp = intel_pstate_get_epp(cpu_data, 0);
728 	if (epp < 0)
729 		return epp;
730 
731 	if (static_cpu_has(X86_FEATURE_HWP_EPP)) {
732 		/*
733 		 * Range:
734 		 *	0x00-0x3F	:	Performance
735 		 *	0x40-0x7F	:	Balance performance
736 		 *	0x80-0xBF	:	Balance power
737 		 *	0xC0-0xFF	:	Power
738 		 * The EPP is a 8 bit value, but our ranges restrict the
739 		 * value which can be set. Here only using top two bits
740 		 * effectively.
741 		 */
742 		index = (epp >> 6) + 1;
743 	} else if (static_cpu_has(X86_FEATURE_EPB)) {
744 		/*
745 		 * Range:
746 		 *	0x00-0x03	:	Performance
747 		 *	0x04-0x07	:	Balance performance
748 		 *	0x08-0x0B	:	Balance power
749 		 *	0x0C-0x0F	:	Power
750 		 * The EPB is a 4 bit value, but our ranges restrict the
751 		 * value which can be set. Here only using top two bits
752 		 * effectively.
753 		 */
754 		index = (epp >> 2) + 1;
755 	}
756 
757 	return index;
758 }
759 
760 static int intel_pstate_set_energy_pref_index(struct cpudata *cpu_data,
761 					      int pref_index)
762 {
763 	int epp = -EINVAL;
764 	int ret;
765 
766 	if (!pref_index)
767 		epp = cpu_data->epp_default;
768 
769 	mutex_lock(&intel_pstate_limits_lock);
770 
771 	if (static_cpu_has(X86_FEATURE_HWP_EPP)) {
772 		u64 value;
773 
774 		ret = rdmsrl_on_cpu(cpu_data->cpu, MSR_HWP_REQUEST, &value);
775 		if (ret)
776 			goto return_pref;
777 
778 		value &= ~GENMASK_ULL(31, 24);
779 
780 		/*
781 		 * If epp is not default, convert from index into
782 		 * energy_perf_strings to epp value, by shifting 6
783 		 * bits left to use only top two bits in epp.
784 		 * The resultant epp need to shifted by 24 bits to
785 		 * epp position in MSR_HWP_REQUEST.
786 		 */
787 		if (epp == -EINVAL)
788 			epp = (pref_index - 1) << 6;
789 
790 		value |= (u64)epp << 24;
791 		ret = wrmsrl_on_cpu(cpu_data->cpu, MSR_HWP_REQUEST, value);
792 	} else {
793 		if (epp == -EINVAL)
794 			epp = (pref_index - 1) << 2;
795 		ret = intel_pstate_set_epb(cpu_data->cpu, epp);
796 	}
797 return_pref:
798 	mutex_unlock(&intel_pstate_limits_lock);
799 
800 	return ret;
801 }
802 
803 static ssize_t show_energy_performance_available_preferences(
804 				struct cpufreq_policy *policy, char *buf)
805 {
806 	int i = 0;
807 	int ret = 0;
808 
809 	while (energy_perf_strings[i] != NULL)
810 		ret += sprintf(&buf[ret], "%s ", energy_perf_strings[i++]);
811 
812 	ret += sprintf(&buf[ret], "\n");
813 
814 	return ret;
815 }
816 
817 cpufreq_freq_attr_ro(energy_performance_available_preferences);
818 
819 static ssize_t store_energy_performance_preference(
820 		struct cpufreq_policy *policy, const char *buf, size_t count)
821 {
822 	struct cpudata *cpu_data = all_cpu_data[policy->cpu];
823 	char str_preference[21];
824 	int ret, i = 0;
825 
826 	ret = sscanf(buf, "%20s", str_preference);
827 	if (ret != 1)
828 		return -EINVAL;
829 
830 	while (energy_perf_strings[i] != NULL) {
831 		if (!strcmp(str_preference, energy_perf_strings[i])) {
832 			intel_pstate_set_energy_pref_index(cpu_data, i);
833 			return count;
834 		}
835 		++i;
836 	}
837 
838 	return -EINVAL;
839 }
840 
841 static ssize_t show_energy_performance_preference(
842 				struct cpufreq_policy *policy, char *buf)
843 {
844 	struct cpudata *cpu_data = all_cpu_data[policy->cpu];
845 	int preference;
846 
847 	preference = intel_pstate_get_energy_pref_index(cpu_data);
848 	if (preference < 0)
849 		return preference;
850 
851 	return  sprintf(buf, "%s\n", energy_perf_strings[preference]);
852 }
853 
854 cpufreq_freq_attr_rw(energy_performance_preference);
855 
856 static struct freq_attr *hwp_cpufreq_attrs[] = {
857 	&energy_performance_preference,
858 	&energy_performance_available_preferences,
859 	NULL,
860 };
861 
862 static void intel_pstate_hwp_set(struct cpufreq_policy *policy)
863 {
864 	int min, hw_min, max, hw_max, cpu, range, adj_range;
865 	struct perf_limits *perf_limits = limits;
866 	u64 value, cap;
867 
868 	for_each_cpu(cpu, policy->cpus) {
869 		int max_perf_pct, min_perf_pct;
870 		struct cpudata *cpu_data = all_cpu_data[cpu];
871 		s16 epp;
872 
873 		if (per_cpu_limits)
874 			perf_limits = all_cpu_data[cpu]->perf_limits;
875 
876 		rdmsrl_on_cpu(cpu, MSR_HWP_CAPABILITIES, &cap);
877 		hw_min = HWP_LOWEST_PERF(cap);
878 		hw_max = HWP_HIGHEST_PERF(cap);
879 		range = hw_max - hw_min;
880 
881 		max_perf_pct = perf_limits->max_perf_pct;
882 		min_perf_pct = perf_limits->min_perf_pct;
883 
884 		rdmsrl_on_cpu(cpu, MSR_HWP_REQUEST, &value);
885 		adj_range = min_perf_pct * range / 100;
886 		min = hw_min + adj_range;
887 		value &= ~HWP_MIN_PERF(~0L);
888 		value |= HWP_MIN_PERF(min);
889 
890 		adj_range = max_perf_pct * range / 100;
891 		max = hw_min + adj_range;
892 		if (limits->no_turbo) {
893 			hw_max = HWP_GUARANTEED_PERF(cap);
894 			if (hw_max < max)
895 				max = hw_max;
896 		}
897 
898 		value &= ~HWP_MAX_PERF(~0L);
899 		value |= HWP_MAX_PERF(max);
900 
901 		if (cpu_data->epp_policy == cpu_data->policy)
902 			goto skip_epp;
903 
904 		cpu_data->epp_policy = cpu_data->policy;
905 
906 		if (cpu_data->epp_saved >= 0) {
907 			epp = cpu_data->epp_saved;
908 			cpu_data->epp_saved = -EINVAL;
909 			goto update_epp;
910 		}
911 
912 		if (cpu_data->policy == CPUFREQ_POLICY_PERFORMANCE) {
913 			epp = intel_pstate_get_epp(cpu_data, value);
914 			cpu_data->epp_powersave = epp;
915 			/* If EPP read was failed, then don't try to write */
916 			if (epp < 0)
917 				goto skip_epp;
918 
919 
920 			epp = 0;
921 		} else {
922 			/* skip setting EPP, when saved value is invalid */
923 			if (cpu_data->epp_powersave < 0)
924 				goto skip_epp;
925 
926 			/*
927 			 * No need to restore EPP when it is not zero. This
928 			 * means:
929 			 *  - Policy is not changed
930 			 *  - user has manually changed
931 			 *  - Error reading EPB
932 			 */
933 			epp = intel_pstate_get_epp(cpu_data, value);
934 			if (epp)
935 				goto skip_epp;
936 
937 			epp = cpu_data->epp_powersave;
938 		}
939 update_epp:
940 		if (static_cpu_has(X86_FEATURE_HWP_EPP)) {
941 			value &= ~GENMASK_ULL(31, 24);
942 			value |= (u64)epp << 24;
943 		} else {
944 			intel_pstate_set_epb(cpu, epp);
945 		}
946 skip_epp:
947 		wrmsrl_on_cpu(cpu, MSR_HWP_REQUEST, value);
948 	}
949 }
950 
951 static int intel_pstate_hwp_set_policy(struct cpufreq_policy *policy)
952 {
953 	if (hwp_active)
954 		intel_pstate_hwp_set(policy);
955 
956 	return 0;
957 }
958 
959 static int intel_pstate_hwp_save_state(struct cpufreq_policy *policy)
960 {
961 	struct cpudata *cpu_data = all_cpu_data[policy->cpu];
962 
963 	if (!hwp_active)
964 		return 0;
965 
966 	cpu_data->epp_saved = intel_pstate_get_epp(cpu_data, 0);
967 
968 	return 0;
969 }
970 
971 static int intel_pstate_resume(struct cpufreq_policy *policy)
972 {
973 	int ret;
974 
975 	if (!hwp_active)
976 		return 0;
977 
978 	mutex_lock(&intel_pstate_limits_lock);
979 
980 	all_cpu_data[policy->cpu]->epp_policy = 0;
981 
982 	ret = intel_pstate_hwp_set_policy(policy);
983 
984 	mutex_unlock(&intel_pstate_limits_lock);
985 
986 	return ret;
987 }
988 
989 static void intel_pstate_update_policies(void)
990 {
991 	int cpu;
992 
993 	for_each_possible_cpu(cpu)
994 		cpufreq_update_policy(cpu);
995 }
996 
997 /************************** debugfs begin ************************/
998 static int pid_param_set(void *data, u64 val)
999 {
1000 	*(u32 *)data = val;
1001 	intel_pstate_reset_all_pid();
1002 	return 0;
1003 }
1004 
1005 static int pid_param_get(void *data, u64 *val)
1006 {
1007 	*val = *(u32 *)data;
1008 	return 0;
1009 }
1010 DEFINE_SIMPLE_ATTRIBUTE(fops_pid_param, pid_param_get, pid_param_set, "%llu\n");
1011 
1012 static struct dentry *debugfs_parent;
1013 
1014 struct pid_param {
1015 	char *name;
1016 	void *value;
1017 	struct dentry *dentry;
1018 };
1019 
1020 static struct pid_param pid_files[] = {
1021 	{"sample_rate_ms", &pid_params.sample_rate_ms, },
1022 	{"d_gain_pct", &pid_params.d_gain_pct, },
1023 	{"i_gain_pct", &pid_params.i_gain_pct, },
1024 	{"deadband", &pid_params.deadband, },
1025 	{"setpoint", &pid_params.setpoint, },
1026 	{"p_gain_pct", &pid_params.p_gain_pct, },
1027 	{NULL, NULL, }
1028 };
1029 
1030 static void intel_pstate_debug_expose_params(void)
1031 {
1032 	int i;
1033 
1034 	debugfs_parent = debugfs_create_dir("pstate_snb", NULL);
1035 	if (IS_ERR_OR_NULL(debugfs_parent))
1036 		return;
1037 
1038 	for (i = 0; pid_files[i].name; i++) {
1039 		struct dentry *dentry;
1040 
1041 		dentry = debugfs_create_file(pid_files[i].name, 0660,
1042 					     debugfs_parent, pid_files[i].value,
1043 					     &fops_pid_param);
1044 		if (!IS_ERR(dentry))
1045 			pid_files[i].dentry = dentry;
1046 	}
1047 }
1048 
1049 static void intel_pstate_debug_hide_params(void)
1050 {
1051 	int i;
1052 
1053 	if (IS_ERR_OR_NULL(debugfs_parent))
1054 		return;
1055 
1056 	for (i = 0; pid_files[i].name; i++) {
1057 		debugfs_remove(pid_files[i].dentry);
1058 		pid_files[i].dentry = NULL;
1059 	}
1060 
1061 	debugfs_remove(debugfs_parent);
1062 	debugfs_parent = NULL;
1063 }
1064 
1065 /************************** debugfs end ************************/
1066 
1067 /************************** sysfs begin ************************/
1068 #define show_one(file_name, object)					\
1069 	static ssize_t show_##file_name					\
1070 	(struct kobject *kobj, struct attribute *attr, char *buf)	\
1071 	{								\
1072 		return sprintf(buf, "%u\n", limits->object);		\
1073 	}
1074 
1075 static ssize_t intel_pstate_show_status(char *buf);
1076 static int intel_pstate_update_status(const char *buf, size_t size);
1077 
1078 static ssize_t show_status(struct kobject *kobj,
1079 			   struct attribute *attr, char *buf)
1080 {
1081 	ssize_t ret;
1082 
1083 	mutex_lock(&intel_pstate_driver_lock);
1084 	ret = intel_pstate_show_status(buf);
1085 	mutex_unlock(&intel_pstate_driver_lock);
1086 
1087 	return ret;
1088 }
1089 
1090 static ssize_t store_status(struct kobject *a, struct attribute *b,
1091 			    const char *buf, size_t count)
1092 {
1093 	char *p = memchr(buf, '\n', count);
1094 	int ret;
1095 
1096 	mutex_lock(&intel_pstate_driver_lock);
1097 	ret = intel_pstate_update_status(buf, p ? p - buf : count);
1098 	mutex_unlock(&intel_pstate_driver_lock);
1099 
1100 	return ret < 0 ? ret : count;
1101 }
1102 
1103 static ssize_t show_turbo_pct(struct kobject *kobj,
1104 				struct attribute *attr, char *buf)
1105 {
1106 	struct cpudata *cpu;
1107 	int total, no_turbo, turbo_pct;
1108 	uint32_t turbo_fp;
1109 
1110 	mutex_lock(&intel_pstate_driver_lock);
1111 
1112 	if (!driver_registered) {
1113 		mutex_unlock(&intel_pstate_driver_lock);
1114 		return -EAGAIN;
1115 	}
1116 
1117 	cpu = all_cpu_data[0];
1118 
1119 	total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
1120 	no_turbo = cpu->pstate.max_pstate - cpu->pstate.min_pstate + 1;
1121 	turbo_fp = div_fp(no_turbo, total);
1122 	turbo_pct = 100 - fp_toint(mul_fp(turbo_fp, int_tofp(100)));
1123 
1124 	mutex_unlock(&intel_pstate_driver_lock);
1125 
1126 	return sprintf(buf, "%u\n", turbo_pct);
1127 }
1128 
1129 static ssize_t show_num_pstates(struct kobject *kobj,
1130 				struct attribute *attr, char *buf)
1131 {
1132 	struct cpudata *cpu;
1133 	int total;
1134 
1135 	mutex_lock(&intel_pstate_driver_lock);
1136 
1137 	if (!driver_registered) {
1138 		mutex_unlock(&intel_pstate_driver_lock);
1139 		return -EAGAIN;
1140 	}
1141 
1142 	cpu = all_cpu_data[0];
1143 	total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
1144 
1145 	mutex_unlock(&intel_pstate_driver_lock);
1146 
1147 	return sprintf(buf, "%u\n", total);
1148 }
1149 
1150 static ssize_t show_no_turbo(struct kobject *kobj,
1151 			     struct attribute *attr, char *buf)
1152 {
1153 	ssize_t ret;
1154 
1155 	mutex_lock(&intel_pstate_driver_lock);
1156 
1157 	if (!driver_registered) {
1158 		mutex_unlock(&intel_pstate_driver_lock);
1159 		return -EAGAIN;
1160 	}
1161 
1162 	update_turbo_state();
1163 	if (limits->turbo_disabled)
1164 		ret = sprintf(buf, "%u\n", limits->turbo_disabled);
1165 	else
1166 		ret = sprintf(buf, "%u\n", limits->no_turbo);
1167 
1168 	mutex_unlock(&intel_pstate_driver_lock);
1169 
1170 	return ret;
1171 }
1172 
1173 static ssize_t store_no_turbo(struct kobject *a, struct attribute *b,
1174 			      const char *buf, size_t count)
1175 {
1176 	unsigned int input;
1177 	int ret;
1178 
1179 	ret = sscanf(buf, "%u", &input);
1180 	if (ret != 1)
1181 		return -EINVAL;
1182 
1183 	mutex_lock(&intel_pstate_driver_lock);
1184 
1185 	if (!driver_registered) {
1186 		mutex_unlock(&intel_pstate_driver_lock);
1187 		return -EAGAIN;
1188 	}
1189 
1190 	mutex_lock(&intel_pstate_limits_lock);
1191 
1192 	update_turbo_state();
1193 	if (limits->turbo_disabled) {
1194 		pr_warn("Turbo disabled by BIOS or unavailable on processor\n");
1195 		mutex_unlock(&intel_pstate_limits_lock);
1196 		mutex_unlock(&intel_pstate_driver_lock);
1197 		return -EPERM;
1198 	}
1199 
1200 	limits->no_turbo = clamp_t(int, input, 0, 1);
1201 
1202 	mutex_unlock(&intel_pstate_limits_lock);
1203 
1204 	intel_pstate_update_policies();
1205 
1206 	mutex_unlock(&intel_pstate_driver_lock);
1207 
1208 	return count;
1209 }
1210 
1211 static ssize_t store_max_perf_pct(struct kobject *a, struct attribute *b,
1212 				  const char *buf, size_t count)
1213 {
1214 	unsigned int input;
1215 	int ret;
1216 
1217 	ret = sscanf(buf, "%u", &input);
1218 	if (ret != 1)
1219 		return -EINVAL;
1220 
1221 	mutex_lock(&intel_pstate_driver_lock);
1222 
1223 	if (!driver_registered) {
1224 		mutex_unlock(&intel_pstate_driver_lock);
1225 		return -EAGAIN;
1226 	}
1227 
1228 	mutex_lock(&intel_pstate_limits_lock);
1229 
1230 	limits->max_sysfs_pct = clamp_t(int, input, 0 , 100);
1231 	limits->max_perf_pct = min(limits->max_policy_pct,
1232 				   limits->max_sysfs_pct);
1233 	limits->max_perf_pct = max(limits->min_policy_pct,
1234 				   limits->max_perf_pct);
1235 	limits->max_perf_pct = max(limits->min_perf_pct,
1236 				   limits->max_perf_pct);
1237 	limits->max_perf = div_ext_fp(limits->max_perf_pct, 100);
1238 
1239 	mutex_unlock(&intel_pstate_limits_lock);
1240 
1241 	intel_pstate_update_policies();
1242 
1243 	mutex_unlock(&intel_pstate_driver_lock);
1244 
1245 	return count;
1246 }
1247 
1248 static ssize_t store_min_perf_pct(struct kobject *a, struct attribute *b,
1249 				  const char *buf, size_t count)
1250 {
1251 	unsigned int input;
1252 	int ret;
1253 
1254 	ret = sscanf(buf, "%u", &input);
1255 	if (ret != 1)
1256 		return -EINVAL;
1257 
1258 	mutex_lock(&intel_pstate_driver_lock);
1259 
1260 	if (!driver_registered) {
1261 		mutex_unlock(&intel_pstate_driver_lock);
1262 		return -EAGAIN;
1263 	}
1264 
1265 	mutex_lock(&intel_pstate_limits_lock);
1266 
1267 	limits->min_sysfs_pct = clamp_t(int, input, 0 , 100);
1268 	limits->min_perf_pct = max(limits->min_policy_pct,
1269 				   limits->min_sysfs_pct);
1270 	limits->min_perf_pct = min(limits->max_policy_pct,
1271 				   limits->min_perf_pct);
1272 	limits->min_perf_pct = min(limits->max_perf_pct,
1273 				   limits->min_perf_pct);
1274 	limits->min_perf = div_ext_fp(limits->min_perf_pct, 100);
1275 
1276 	mutex_unlock(&intel_pstate_limits_lock);
1277 
1278 	intel_pstate_update_policies();
1279 
1280 	mutex_unlock(&intel_pstate_driver_lock);
1281 
1282 	return count;
1283 }
1284 
1285 show_one(max_perf_pct, max_perf_pct);
1286 show_one(min_perf_pct, min_perf_pct);
1287 
1288 define_one_global_rw(status);
1289 define_one_global_rw(no_turbo);
1290 define_one_global_rw(max_perf_pct);
1291 define_one_global_rw(min_perf_pct);
1292 define_one_global_ro(turbo_pct);
1293 define_one_global_ro(num_pstates);
1294 
1295 static struct attribute *intel_pstate_attributes[] = {
1296 	&status.attr,
1297 	&no_turbo.attr,
1298 	&turbo_pct.attr,
1299 	&num_pstates.attr,
1300 	NULL
1301 };
1302 
1303 static struct attribute_group intel_pstate_attr_group = {
1304 	.attrs = intel_pstate_attributes,
1305 };
1306 
1307 static void __init intel_pstate_sysfs_expose_params(void)
1308 {
1309 	struct kobject *intel_pstate_kobject;
1310 	int rc;
1311 
1312 	intel_pstate_kobject = kobject_create_and_add("intel_pstate",
1313 						&cpu_subsys.dev_root->kobj);
1314 	if (WARN_ON(!intel_pstate_kobject))
1315 		return;
1316 
1317 	rc = sysfs_create_group(intel_pstate_kobject, &intel_pstate_attr_group);
1318 	if (WARN_ON(rc))
1319 		return;
1320 
1321 	/*
1322 	 * If per cpu limits are enforced there are no global limits, so
1323 	 * return without creating max/min_perf_pct attributes
1324 	 */
1325 	if (per_cpu_limits)
1326 		return;
1327 
1328 	rc = sysfs_create_file(intel_pstate_kobject, &max_perf_pct.attr);
1329 	WARN_ON(rc);
1330 
1331 	rc = sysfs_create_file(intel_pstate_kobject, &min_perf_pct.attr);
1332 	WARN_ON(rc);
1333 
1334 }
1335 /************************** sysfs end ************************/
1336 
1337 static void intel_pstate_hwp_enable(struct cpudata *cpudata)
1338 {
1339 	/* First disable HWP notification interrupt as we don't process them */
1340 	if (static_cpu_has(X86_FEATURE_HWP_NOTIFY))
1341 		wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_INTERRUPT, 0x00);
1342 
1343 	wrmsrl_on_cpu(cpudata->cpu, MSR_PM_ENABLE, 0x1);
1344 	cpudata->epp_policy = 0;
1345 	if (cpudata->epp_default == -EINVAL)
1346 		cpudata->epp_default = intel_pstate_get_epp(cpudata, 0);
1347 }
1348 
1349 static int atom_get_min_pstate(void)
1350 {
1351 	u64 value;
1352 
1353 	rdmsrl(ATOM_RATIOS, value);
1354 	return (value >> 8) & 0x7F;
1355 }
1356 
1357 static int atom_get_max_pstate(void)
1358 {
1359 	u64 value;
1360 
1361 	rdmsrl(ATOM_RATIOS, value);
1362 	return (value >> 16) & 0x7F;
1363 }
1364 
1365 static int atom_get_turbo_pstate(void)
1366 {
1367 	u64 value;
1368 
1369 	rdmsrl(ATOM_TURBO_RATIOS, value);
1370 	return value & 0x7F;
1371 }
1372 
1373 static u64 atom_get_val(struct cpudata *cpudata, int pstate)
1374 {
1375 	u64 val;
1376 	int32_t vid_fp;
1377 	u32 vid;
1378 
1379 	val = (u64)pstate << 8;
1380 	if (limits->no_turbo && !limits->turbo_disabled)
1381 		val |= (u64)1 << 32;
1382 
1383 	vid_fp = cpudata->vid.min + mul_fp(
1384 		int_tofp(pstate - cpudata->pstate.min_pstate),
1385 		cpudata->vid.ratio);
1386 
1387 	vid_fp = clamp_t(int32_t, vid_fp, cpudata->vid.min, cpudata->vid.max);
1388 	vid = ceiling_fp(vid_fp);
1389 
1390 	if (pstate > cpudata->pstate.max_pstate)
1391 		vid = cpudata->vid.turbo;
1392 
1393 	return val | vid;
1394 }
1395 
1396 static int silvermont_get_scaling(void)
1397 {
1398 	u64 value;
1399 	int i;
1400 	/* Defined in Table 35-6 from SDM (Sept 2015) */
1401 	static int silvermont_freq_table[] = {
1402 		83300, 100000, 133300, 116700, 80000};
1403 
1404 	rdmsrl(MSR_FSB_FREQ, value);
1405 	i = value & 0x7;
1406 	WARN_ON(i > 4);
1407 
1408 	return silvermont_freq_table[i];
1409 }
1410 
1411 static int airmont_get_scaling(void)
1412 {
1413 	u64 value;
1414 	int i;
1415 	/* Defined in Table 35-10 from SDM (Sept 2015) */
1416 	static int airmont_freq_table[] = {
1417 		83300, 100000, 133300, 116700, 80000,
1418 		93300, 90000, 88900, 87500};
1419 
1420 	rdmsrl(MSR_FSB_FREQ, value);
1421 	i = value & 0xF;
1422 	WARN_ON(i > 8);
1423 
1424 	return airmont_freq_table[i];
1425 }
1426 
1427 static void atom_get_vid(struct cpudata *cpudata)
1428 {
1429 	u64 value;
1430 
1431 	rdmsrl(ATOM_VIDS, value);
1432 	cpudata->vid.min = int_tofp((value >> 8) & 0x7f);
1433 	cpudata->vid.max = int_tofp((value >> 16) & 0x7f);
1434 	cpudata->vid.ratio = div_fp(
1435 		cpudata->vid.max - cpudata->vid.min,
1436 		int_tofp(cpudata->pstate.max_pstate -
1437 			cpudata->pstate.min_pstate));
1438 
1439 	rdmsrl(ATOM_TURBO_VIDS, value);
1440 	cpudata->vid.turbo = value & 0x7f;
1441 }
1442 
1443 static int core_get_min_pstate(void)
1444 {
1445 	u64 value;
1446 
1447 	rdmsrl(MSR_PLATFORM_INFO, value);
1448 	return (value >> 40) & 0xFF;
1449 }
1450 
1451 static int core_get_max_pstate_physical(void)
1452 {
1453 	u64 value;
1454 
1455 	rdmsrl(MSR_PLATFORM_INFO, value);
1456 	return (value >> 8) & 0xFF;
1457 }
1458 
1459 static int core_get_max_pstate(void)
1460 {
1461 	u64 tar;
1462 	u64 plat_info;
1463 	int max_pstate;
1464 	int err;
1465 
1466 	rdmsrl(MSR_PLATFORM_INFO, plat_info);
1467 	max_pstate = (plat_info >> 8) & 0xFF;
1468 
1469 	err = rdmsrl_safe(MSR_TURBO_ACTIVATION_RATIO, &tar);
1470 	if (!err) {
1471 		/* Do some sanity checking for safety */
1472 		if (plat_info & 0x600000000) {
1473 			u64 tdp_ctrl;
1474 			u64 tdp_ratio;
1475 			int tdp_msr;
1476 
1477 			err = rdmsrl_safe(MSR_CONFIG_TDP_CONTROL, &tdp_ctrl);
1478 			if (err)
1479 				goto skip_tar;
1480 
1481 			tdp_msr = MSR_CONFIG_TDP_NOMINAL + (tdp_ctrl & 0x3);
1482 			err = rdmsrl_safe(tdp_msr, &tdp_ratio);
1483 			if (err)
1484 				goto skip_tar;
1485 
1486 			/* For level 1 and 2, bits[23:16] contain the ratio */
1487 			if (tdp_ctrl)
1488 				tdp_ratio >>= 16;
1489 
1490 			tdp_ratio &= 0xff; /* ratios are only 8 bits long */
1491 			if (tdp_ratio - 1 == tar) {
1492 				max_pstate = tar;
1493 				pr_debug("max_pstate=TAC %x\n", max_pstate);
1494 			} else {
1495 				goto skip_tar;
1496 			}
1497 		}
1498 	}
1499 
1500 skip_tar:
1501 	return max_pstate;
1502 }
1503 
1504 static int core_get_turbo_pstate(void)
1505 {
1506 	u64 value;
1507 	int nont, ret;
1508 
1509 	rdmsrl(MSR_TURBO_RATIO_LIMIT, value);
1510 	nont = core_get_max_pstate();
1511 	ret = (value) & 255;
1512 	if (ret <= nont)
1513 		ret = nont;
1514 	return ret;
1515 }
1516 
1517 static inline int core_get_scaling(void)
1518 {
1519 	return 100000;
1520 }
1521 
1522 static u64 core_get_val(struct cpudata *cpudata, int pstate)
1523 {
1524 	u64 val;
1525 
1526 	val = (u64)pstate << 8;
1527 	if (limits->no_turbo && !limits->turbo_disabled)
1528 		val |= (u64)1 << 32;
1529 
1530 	return val;
1531 }
1532 
1533 static int knl_get_turbo_pstate(void)
1534 {
1535 	u64 value;
1536 	int nont, ret;
1537 
1538 	rdmsrl(MSR_TURBO_RATIO_LIMIT, value);
1539 	nont = core_get_max_pstate();
1540 	ret = (((value) >> 8) & 0xFF);
1541 	if (ret <= nont)
1542 		ret = nont;
1543 	return ret;
1544 }
1545 
1546 static struct cpu_defaults core_params = {
1547 	.pid_policy = {
1548 		.sample_rate_ms = 10,
1549 		.deadband = 0,
1550 		.setpoint = 97,
1551 		.p_gain_pct = 20,
1552 		.d_gain_pct = 0,
1553 		.i_gain_pct = 0,
1554 	},
1555 	.funcs = {
1556 		.get_max = core_get_max_pstate,
1557 		.get_max_physical = core_get_max_pstate_physical,
1558 		.get_min = core_get_min_pstate,
1559 		.get_turbo = core_get_turbo_pstate,
1560 		.get_scaling = core_get_scaling,
1561 		.get_val = core_get_val,
1562 		.get_target_pstate = get_target_pstate_use_performance,
1563 	},
1564 };
1565 
1566 static const struct cpu_defaults silvermont_params = {
1567 	.pid_policy = {
1568 		.sample_rate_ms = 10,
1569 		.deadband = 0,
1570 		.setpoint = 60,
1571 		.p_gain_pct = 14,
1572 		.d_gain_pct = 0,
1573 		.i_gain_pct = 4,
1574 	},
1575 	.funcs = {
1576 		.get_max = atom_get_max_pstate,
1577 		.get_max_physical = atom_get_max_pstate,
1578 		.get_min = atom_get_min_pstate,
1579 		.get_turbo = atom_get_turbo_pstate,
1580 		.get_val = atom_get_val,
1581 		.get_scaling = silvermont_get_scaling,
1582 		.get_vid = atom_get_vid,
1583 		.get_target_pstate = get_target_pstate_use_cpu_load,
1584 	},
1585 };
1586 
1587 static const struct cpu_defaults airmont_params = {
1588 	.pid_policy = {
1589 		.sample_rate_ms = 10,
1590 		.deadband = 0,
1591 		.setpoint = 60,
1592 		.p_gain_pct = 14,
1593 		.d_gain_pct = 0,
1594 		.i_gain_pct = 4,
1595 	},
1596 	.funcs = {
1597 		.get_max = atom_get_max_pstate,
1598 		.get_max_physical = atom_get_max_pstate,
1599 		.get_min = atom_get_min_pstate,
1600 		.get_turbo = atom_get_turbo_pstate,
1601 		.get_val = atom_get_val,
1602 		.get_scaling = airmont_get_scaling,
1603 		.get_vid = atom_get_vid,
1604 		.get_target_pstate = get_target_pstate_use_cpu_load,
1605 	},
1606 };
1607 
1608 static const struct cpu_defaults knl_params = {
1609 	.pid_policy = {
1610 		.sample_rate_ms = 10,
1611 		.deadband = 0,
1612 		.setpoint = 97,
1613 		.p_gain_pct = 20,
1614 		.d_gain_pct = 0,
1615 		.i_gain_pct = 0,
1616 	},
1617 	.funcs = {
1618 		.get_max = core_get_max_pstate,
1619 		.get_max_physical = core_get_max_pstate_physical,
1620 		.get_min = core_get_min_pstate,
1621 		.get_turbo = knl_get_turbo_pstate,
1622 		.get_scaling = core_get_scaling,
1623 		.get_val = core_get_val,
1624 		.get_target_pstate = get_target_pstate_use_performance,
1625 	},
1626 };
1627 
1628 static const struct cpu_defaults bxt_params = {
1629 	.pid_policy = {
1630 		.sample_rate_ms = 10,
1631 		.deadband = 0,
1632 		.setpoint = 60,
1633 		.p_gain_pct = 14,
1634 		.d_gain_pct = 0,
1635 		.i_gain_pct = 4,
1636 	},
1637 	.funcs = {
1638 		.get_max = core_get_max_pstate,
1639 		.get_max_physical = core_get_max_pstate_physical,
1640 		.get_min = core_get_min_pstate,
1641 		.get_turbo = core_get_turbo_pstate,
1642 		.get_scaling = core_get_scaling,
1643 		.get_val = core_get_val,
1644 		.get_target_pstate = get_target_pstate_use_cpu_load,
1645 	},
1646 };
1647 
1648 static void intel_pstate_get_min_max(struct cpudata *cpu, int *min, int *max)
1649 {
1650 	int max_perf = cpu->pstate.turbo_pstate;
1651 	int max_perf_adj;
1652 	int min_perf;
1653 	struct perf_limits *perf_limits = limits;
1654 
1655 	if (limits->no_turbo || limits->turbo_disabled)
1656 		max_perf = cpu->pstate.max_pstate;
1657 
1658 	if (per_cpu_limits)
1659 		perf_limits = cpu->perf_limits;
1660 
1661 	/*
1662 	 * performance can be limited by user through sysfs, by cpufreq
1663 	 * policy, or by cpu specific default values determined through
1664 	 * experimentation.
1665 	 */
1666 	max_perf_adj = fp_ext_toint(max_perf * perf_limits->max_perf);
1667 	*max = clamp_t(int, max_perf_adj,
1668 			cpu->pstate.min_pstate, cpu->pstate.turbo_pstate);
1669 
1670 	min_perf = fp_ext_toint(max_perf * perf_limits->min_perf);
1671 	*min = clamp_t(int, min_perf, cpu->pstate.min_pstate, max_perf);
1672 }
1673 
1674 static void intel_pstate_set_pstate(struct cpudata *cpu, int pstate)
1675 {
1676 	trace_cpu_frequency(pstate * cpu->pstate.scaling, cpu->cpu);
1677 	cpu->pstate.current_pstate = pstate;
1678 	/*
1679 	 * Generally, there is no guarantee that this code will always run on
1680 	 * the CPU being updated, so force the register update to run on the
1681 	 * right CPU.
1682 	 */
1683 	wrmsrl_on_cpu(cpu->cpu, MSR_IA32_PERF_CTL,
1684 		      pstate_funcs.get_val(cpu, pstate));
1685 }
1686 
1687 static void intel_pstate_set_min_pstate(struct cpudata *cpu)
1688 {
1689 	intel_pstate_set_pstate(cpu, cpu->pstate.min_pstate);
1690 }
1691 
1692 static void intel_pstate_max_within_limits(struct cpudata *cpu)
1693 {
1694 	int min_pstate, max_pstate;
1695 
1696 	update_turbo_state();
1697 	intel_pstate_get_min_max(cpu, &min_pstate, &max_pstate);
1698 	intel_pstate_set_pstate(cpu, max_pstate);
1699 }
1700 
1701 static void intel_pstate_get_cpu_pstates(struct cpudata *cpu)
1702 {
1703 	cpu->pstate.min_pstate = pstate_funcs.get_min();
1704 	cpu->pstate.max_pstate = pstate_funcs.get_max();
1705 	cpu->pstate.max_pstate_physical = pstate_funcs.get_max_physical();
1706 	cpu->pstate.turbo_pstate = pstate_funcs.get_turbo();
1707 	cpu->pstate.scaling = pstate_funcs.get_scaling();
1708 	cpu->pstate.max_freq = cpu->pstate.max_pstate * cpu->pstate.scaling;
1709 	cpu->pstate.turbo_freq = cpu->pstate.turbo_pstate * cpu->pstate.scaling;
1710 
1711 	if (pstate_funcs.get_vid)
1712 		pstate_funcs.get_vid(cpu);
1713 
1714 	intel_pstate_set_min_pstate(cpu);
1715 }
1716 
1717 static inline void intel_pstate_calc_avg_perf(struct cpudata *cpu)
1718 {
1719 	struct sample *sample = &cpu->sample;
1720 
1721 	sample->core_avg_perf = div_ext_fp(sample->aperf, sample->mperf);
1722 }
1723 
1724 static inline bool intel_pstate_sample(struct cpudata *cpu, u64 time)
1725 {
1726 	u64 aperf, mperf;
1727 	unsigned long flags;
1728 	u64 tsc;
1729 
1730 	local_irq_save(flags);
1731 	rdmsrl(MSR_IA32_APERF, aperf);
1732 	rdmsrl(MSR_IA32_MPERF, mperf);
1733 	tsc = rdtsc();
1734 	if (cpu->prev_mperf == mperf || cpu->prev_tsc == tsc) {
1735 		local_irq_restore(flags);
1736 		return false;
1737 	}
1738 	local_irq_restore(flags);
1739 
1740 	cpu->last_sample_time = cpu->sample.time;
1741 	cpu->sample.time = time;
1742 	cpu->sample.aperf = aperf;
1743 	cpu->sample.mperf = mperf;
1744 	cpu->sample.tsc =  tsc;
1745 	cpu->sample.aperf -= cpu->prev_aperf;
1746 	cpu->sample.mperf -= cpu->prev_mperf;
1747 	cpu->sample.tsc -= cpu->prev_tsc;
1748 
1749 	cpu->prev_aperf = aperf;
1750 	cpu->prev_mperf = mperf;
1751 	cpu->prev_tsc = tsc;
1752 	/*
1753 	 * First time this function is invoked in a given cycle, all of the
1754 	 * previous sample data fields are equal to zero or stale and they must
1755 	 * be populated with meaningful numbers for things to work, so assume
1756 	 * that sample.time will always be reset before setting the utilization
1757 	 * update hook and make the caller skip the sample then.
1758 	 */
1759 	return !!cpu->last_sample_time;
1760 }
1761 
1762 static inline int32_t get_avg_frequency(struct cpudata *cpu)
1763 {
1764 	return mul_ext_fp(cpu->sample.core_avg_perf,
1765 			  cpu->pstate.max_pstate_physical * cpu->pstate.scaling);
1766 }
1767 
1768 static inline int32_t get_avg_pstate(struct cpudata *cpu)
1769 {
1770 	return mul_ext_fp(cpu->pstate.max_pstate_physical,
1771 			  cpu->sample.core_avg_perf);
1772 }
1773 
1774 static inline int32_t get_target_pstate_use_cpu_load(struct cpudata *cpu)
1775 {
1776 	struct sample *sample = &cpu->sample;
1777 	int32_t busy_frac, boost;
1778 	int target, avg_pstate;
1779 
1780 	busy_frac = div_fp(sample->mperf, sample->tsc);
1781 
1782 	boost = cpu->iowait_boost;
1783 	cpu->iowait_boost >>= 1;
1784 
1785 	if (busy_frac < boost)
1786 		busy_frac = boost;
1787 
1788 	sample->busy_scaled = busy_frac * 100;
1789 
1790 	target = limits->no_turbo || limits->turbo_disabled ?
1791 			cpu->pstate.max_pstate : cpu->pstate.turbo_pstate;
1792 	target += target >> 2;
1793 	target = mul_fp(target, busy_frac);
1794 	if (target < cpu->pstate.min_pstate)
1795 		target = cpu->pstate.min_pstate;
1796 
1797 	/*
1798 	 * If the average P-state during the previous cycle was higher than the
1799 	 * current target, add 50% of the difference to the target to reduce
1800 	 * possible performance oscillations and offset possible performance
1801 	 * loss related to moving the workload from one CPU to another within
1802 	 * a package/module.
1803 	 */
1804 	avg_pstate = get_avg_pstate(cpu);
1805 	if (avg_pstate > target)
1806 		target += (avg_pstate - target) >> 1;
1807 
1808 	return target;
1809 }
1810 
1811 static inline int32_t get_target_pstate_use_performance(struct cpudata *cpu)
1812 {
1813 	int32_t perf_scaled, max_pstate, current_pstate, sample_ratio;
1814 	u64 duration_ns;
1815 
1816 	/*
1817 	 * perf_scaled is the ratio of the average P-state during the last
1818 	 * sampling period to the P-state requested last time (in percent).
1819 	 *
1820 	 * That measures the system's response to the previous P-state
1821 	 * selection.
1822 	 */
1823 	max_pstate = cpu->pstate.max_pstate_physical;
1824 	current_pstate = cpu->pstate.current_pstate;
1825 	perf_scaled = mul_ext_fp(cpu->sample.core_avg_perf,
1826 			       div_fp(100 * max_pstate, current_pstate));
1827 
1828 	/*
1829 	 * Since our utilization update callback will not run unless we are
1830 	 * in C0, check if the actual elapsed time is significantly greater (3x)
1831 	 * than our sample interval.  If it is, then we were idle for a long
1832 	 * enough period of time to adjust our performance metric.
1833 	 */
1834 	duration_ns = cpu->sample.time - cpu->last_sample_time;
1835 	if ((s64)duration_ns > pid_params.sample_rate_ns * 3) {
1836 		sample_ratio = div_fp(pid_params.sample_rate_ns, duration_ns);
1837 		perf_scaled = mul_fp(perf_scaled, sample_ratio);
1838 	} else {
1839 		sample_ratio = div_fp(100 * cpu->sample.mperf, cpu->sample.tsc);
1840 		if (sample_ratio < int_tofp(1))
1841 			perf_scaled = 0;
1842 	}
1843 
1844 	cpu->sample.busy_scaled = perf_scaled;
1845 	return cpu->pstate.current_pstate - pid_calc(&cpu->pid, perf_scaled);
1846 }
1847 
1848 static int intel_pstate_prepare_request(struct cpudata *cpu, int pstate)
1849 {
1850 	int max_perf, min_perf;
1851 
1852 	intel_pstate_get_min_max(cpu, &min_perf, &max_perf);
1853 	pstate = clamp_t(int, pstate, min_perf, max_perf);
1854 	trace_cpu_frequency(pstate * cpu->pstate.scaling, cpu->cpu);
1855 	return pstate;
1856 }
1857 
1858 static void intel_pstate_update_pstate(struct cpudata *cpu, int pstate)
1859 {
1860 	pstate = intel_pstate_prepare_request(cpu, pstate);
1861 	if (pstate == cpu->pstate.current_pstate)
1862 		return;
1863 
1864 	cpu->pstate.current_pstate = pstate;
1865 	wrmsrl(MSR_IA32_PERF_CTL, pstate_funcs.get_val(cpu, pstate));
1866 }
1867 
1868 static inline void intel_pstate_adjust_busy_pstate(struct cpudata *cpu)
1869 {
1870 	int from, target_pstate;
1871 	struct sample *sample;
1872 
1873 	from = cpu->pstate.current_pstate;
1874 
1875 	target_pstate = cpu->policy == CPUFREQ_POLICY_PERFORMANCE ?
1876 		cpu->pstate.turbo_pstate : pstate_funcs.get_target_pstate(cpu);
1877 
1878 	update_turbo_state();
1879 
1880 	intel_pstate_update_pstate(cpu, target_pstate);
1881 
1882 	sample = &cpu->sample;
1883 	trace_pstate_sample(mul_ext_fp(100, sample->core_avg_perf),
1884 		fp_toint(sample->busy_scaled),
1885 		from,
1886 		cpu->pstate.current_pstate,
1887 		sample->mperf,
1888 		sample->aperf,
1889 		sample->tsc,
1890 		get_avg_frequency(cpu),
1891 		fp_toint(cpu->iowait_boost * 100));
1892 }
1893 
1894 static void intel_pstate_update_util(struct update_util_data *data, u64 time,
1895 				     unsigned int flags)
1896 {
1897 	struct cpudata *cpu = container_of(data, struct cpudata, update_util);
1898 	u64 delta_ns;
1899 
1900 	if (pstate_funcs.get_target_pstate == get_target_pstate_use_cpu_load) {
1901 		if (flags & SCHED_CPUFREQ_IOWAIT) {
1902 			cpu->iowait_boost = int_tofp(1);
1903 		} else if (cpu->iowait_boost) {
1904 			/* Clear iowait_boost if the CPU may have been idle. */
1905 			delta_ns = time - cpu->last_update;
1906 			if (delta_ns > TICK_NSEC)
1907 				cpu->iowait_boost = 0;
1908 		}
1909 		cpu->last_update = time;
1910 	}
1911 
1912 	delta_ns = time - cpu->sample.time;
1913 	if ((s64)delta_ns >= pid_params.sample_rate_ns) {
1914 		bool sample_taken = intel_pstate_sample(cpu, time);
1915 
1916 		if (sample_taken) {
1917 			intel_pstate_calc_avg_perf(cpu);
1918 			if (!hwp_active)
1919 				intel_pstate_adjust_busy_pstate(cpu);
1920 		}
1921 	}
1922 }
1923 
1924 #define ICPU(model, policy) \
1925 	{ X86_VENDOR_INTEL, 6, model, X86_FEATURE_APERFMPERF,\
1926 			(unsigned long)&policy }
1927 
1928 static const struct x86_cpu_id intel_pstate_cpu_ids[] = {
1929 	ICPU(INTEL_FAM6_SANDYBRIDGE, 		core_params),
1930 	ICPU(INTEL_FAM6_SANDYBRIDGE_X,		core_params),
1931 	ICPU(INTEL_FAM6_ATOM_SILVERMONT1,	silvermont_params),
1932 	ICPU(INTEL_FAM6_IVYBRIDGE,		core_params),
1933 	ICPU(INTEL_FAM6_HASWELL_CORE,		core_params),
1934 	ICPU(INTEL_FAM6_BROADWELL_CORE,		core_params),
1935 	ICPU(INTEL_FAM6_IVYBRIDGE_X,		core_params),
1936 	ICPU(INTEL_FAM6_HASWELL_X,		core_params),
1937 	ICPU(INTEL_FAM6_HASWELL_ULT,		core_params),
1938 	ICPU(INTEL_FAM6_HASWELL_GT3E,		core_params),
1939 	ICPU(INTEL_FAM6_BROADWELL_GT3E,		core_params),
1940 	ICPU(INTEL_FAM6_ATOM_AIRMONT,		airmont_params),
1941 	ICPU(INTEL_FAM6_SKYLAKE_MOBILE,		core_params),
1942 	ICPU(INTEL_FAM6_BROADWELL_X,		core_params),
1943 	ICPU(INTEL_FAM6_SKYLAKE_DESKTOP,	core_params),
1944 	ICPU(INTEL_FAM6_BROADWELL_XEON_D,	core_params),
1945 	ICPU(INTEL_FAM6_XEON_PHI_KNL,		knl_params),
1946 	ICPU(INTEL_FAM6_XEON_PHI_KNM,		knl_params),
1947 	ICPU(INTEL_FAM6_ATOM_GOLDMONT,		bxt_params),
1948 	{}
1949 };
1950 MODULE_DEVICE_TABLE(x86cpu, intel_pstate_cpu_ids);
1951 
1952 static const struct x86_cpu_id intel_pstate_cpu_oob_ids[] __initconst = {
1953 	ICPU(INTEL_FAM6_BROADWELL_XEON_D, core_params),
1954 	ICPU(INTEL_FAM6_BROADWELL_X, core_params),
1955 	ICPU(INTEL_FAM6_SKYLAKE_X, core_params),
1956 	{}
1957 };
1958 
1959 static int intel_pstate_init_cpu(unsigned int cpunum)
1960 {
1961 	struct cpudata *cpu;
1962 
1963 	cpu = all_cpu_data[cpunum];
1964 
1965 	if (!cpu) {
1966 		unsigned int size = sizeof(struct cpudata);
1967 
1968 		if (per_cpu_limits)
1969 			size += sizeof(struct perf_limits);
1970 
1971 		cpu = kzalloc(size, GFP_KERNEL);
1972 		if (!cpu)
1973 			return -ENOMEM;
1974 
1975 		all_cpu_data[cpunum] = cpu;
1976 		if (per_cpu_limits)
1977 			cpu->perf_limits = (struct perf_limits *)(cpu + 1);
1978 
1979 		cpu->epp_default = -EINVAL;
1980 		cpu->epp_powersave = -EINVAL;
1981 		cpu->epp_saved = -EINVAL;
1982 	}
1983 
1984 	cpu = all_cpu_data[cpunum];
1985 
1986 	cpu->cpu = cpunum;
1987 
1988 	if (hwp_active) {
1989 		intel_pstate_hwp_enable(cpu);
1990 		pid_params.sample_rate_ms = 50;
1991 		pid_params.sample_rate_ns = 50 * NSEC_PER_MSEC;
1992 	}
1993 
1994 	intel_pstate_get_cpu_pstates(cpu);
1995 
1996 	intel_pstate_busy_pid_reset(cpu);
1997 
1998 	pr_debug("controlling: cpu %d\n", cpunum);
1999 
2000 	return 0;
2001 }
2002 
2003 static unsigned int intel_pstate_get(unsigned int cpu_num)
2004 {
2005 	struct cpudata *cpu = all_cpu_data[cpu_num];
2006 
2007 	return cpu ? get_avg_frequency(cpu) : 0;
2008 }
2009 
2010 static void intel_pstate_set_update_util_hook(unsigned int cpu_num)
2011 {
2012 	struct cpudata *cpu = all_cpu_data[cpu_num];
2013 
2014 	if (cpu->update_util_set)
2015 		return;
2016 
2017 	/* Prevent intel_pstate_update_util() from using stale data. */
2018 	cpu->sample.time = 0;
2019 	cpufreq_add_update_util_hook(cpu_num, &cpu->update_util,
2020 				     intel_pstate_update_util);
2021 	cpu->update_util_set = true;
2022 }
2023 
2024 static void intel_pstate_clear_update_util_hook(unsigned int cpu)
2025 {
2026 	struct cpudata *cpu_data = all_cpu_data[cpu];
2027 
2028 	if (!cpu_data->update_util_set)
2029 		return;
2030 
2031 	cpufreq_remove_update_util_hook(cpu);
2032 	cpu_data->update_util_set = false;
2033 	synchronize_sched();
2034 }
2035 
2036 static void intel_pstate_set_performance_limits(struct perf_limits *limits)
2037 {
2038 	limits->no_turbo = 0;
2039 	limits->turbo_disabled = 0;
2040 	limits->max_perf_pct = 100;
2041 	limits->max_perf = int_ext_tofp(1);
2042 	limits->min_perf_pct = 100;
2043 	limits->min_perf = int_ext_tofp(1);
2044 	limits->max_policy_pct = 100;
2045 	limits->max_sysfs_pct = 100;
2046 	limits->min_policy_pct = 0;
2047 	limits->min_sysfs_pct = 0;
2048 }
2049 
2050 static void intel_pstate_update_perf_limits(struct cpufreq_policy *policy,
2051 					    struct perf_limits *limits)
2052 {
2053 
2054 	limits->max_policy_pct = DIV_ROUND_UP(policy->max * 100,
2055 					      policy->cpuinfo.max_freq);
2056 	limits->max_policy_pct = clamp_t(int, limits->max_policy_pct, 0, 100);
2057 	if (policy->max == policy->min) {
2058 		limits->min_policy_pct = limits->max_policy_pct;
2059 	} else {
2060 		limits->min_policy_pct = DIV_ROUND_UP(policy->min * 100,
2061 						      policy->cpuinfo.max_freq);
2062 		limits->min_policy_pct = clamp_t(int, limits->min_policy_pct,
2063 						 0, 100);
2064 	}
2065 
2066 	/* Normalize user input to [min_policy_pct, max_policy_pct] */
2067 	limits->min_perf_pct = max(limits->min_policy_pct,
2068 				   limits->min_sysfs_pct);
2069 	limits->min_perf_pct = min(limits->max_policy_pct,
2070 				   limits->min_perf_pct);
2071 	limits->max_perf_pct = min(limits->max_policy_pct,
2072 				   limits->max_sysfs_pct);
2073 	limits->max_perf_pct = max(limits->min_policy_pct,
2074 				   limits->max_perf_pct);
2075 
2076 	/* Make sure min_perf_pct <= max_perf_pct */
2077 	limits->min_perf_pct = min(limits->max_perf_pct, limits->min_perf_pct);
2078 
2079 	limits->min_perf = div_ext_fp(limits->min_perf_pct, 100);
2080 	limits->max_perf = div_ext_fp(limits->max_perf_pct, 100);
2081 	limits->max_perf = round_up(limits->max_perf, EXT_FRAC_BITS);
2082 	limits->min_perf = round_up(limits->min_perf, EXT_FRAC_BITS);
2083 
2084 	pr_debug("cpu:%d max_perf_pct:%d min_perf_pct:%d\n", policy->cpu,
2085 		 limits->max_perf_pct, limits->min_perf_pct);
2086 }
2087 
2088 static int intel_pstate_set_policy(struct cpufreq_policy *policy)
2089 {
2090 	struct cpudata *cpu;
2091 	struct perf_limits *perf_limits = NULL;
2092 
2093 	if (!policy->cpuinfo.max_freq)
2094 		return -ENODEV;
2095 
2096 	pr_debug("set_policy cpuinfo.max %u policy->max %u\n",
2097 		 policy->cpuinfo.max_freq, policy->max);
2098 
2099 	cpu = all_cpu_data[policy->cpu];
2100 	cpu->policy = policy->policy;
2101 
2102 	if (cpu->pstate.max_pstate_physical > cpu->pstate.max_pstate &&
2103 	    policy->max < policy->cpuinfo.max_freq &&
2104 	    policy->max > cpu->pstate.max_pstate * cpu->pstate.scaling) {
2105 		pr_debug("policy->max > max non turbo frequency\n");
2106 		policy->max = policy->cpuinfo.max_freq;
2107 	}
2108 
2109 	if (per_cpu_limits)
2110 		perf_limits = cpu->perf_limits;
2111 
2112 	mutex_lock(&intel_pstate_limits_lock);
2113 
2114 	if (policy->policy == CPUFREQ_POLICY_PERFORMANCE) {
2115 		if (!perf_limits) {
2116 			limits = &performance_limits;
2117 			perf_limits = limits;
2118 		}
2119 		if (policy->max >= policy->cpuinfo.max_freq &&
2120 		    !limits->no_turbo) {
2121 			pr_debug("set performance\n");
2122 			intel_pstate_set_performance_limits(perf_limits);
2123 			goto out;
2124 		}
2125 	} else {
2126 		pr_debug("set powersave\n");
2127 		if (!perf_limits) {
2128 			limits = &powersave_limits;
2129 			perf_limits = limits;
2130 		}
2131 
2132 	}
2133 
2134 	intel_pstate_update_perf_limits(policy, perf_limits);
2135  out:
2136 	if (cpu->policy == CPUFREQ_POLICY_PERFORMANCE) {
2137 		/*
2138 		 * NOHZ_FULL CPUs need this as the governor callback may not
2139 		 * be invoked on them.
2140 		 */
2141 		intel_pstate_clear_update_util_hook(policy->cpu);
2142 		intel_pstate_max_within_limits(cpu);
2143 	}
2144 
2145 	intel_pstate_set_update_util_hook(policy->cpu);
2146 
2147 	intel_pstate_hwp_set_policy(policy);
2148 
2149 	mutex_unlock(&intel_pstate_limits_lock);
2150 
2151 	return 0;
2152 }
2153 
2154 static int intel_pstate_verify_policy(struct cpufreq_policy *policy)
2155 {
2156 	cpufreq_verify_within_cpu_limits(policy);
2157 
2158 	if (policy->policy != CPUFREQ_POLICY_POWERSAVE &&
2159 	    policy->policy != CPUFREQ_POLICY_PERFORMANCE)
2160 		return -EINVAL;
2161 
2162 	/* When per-CPU limits are used, sysfs limits are not used */
2163 	if (!per_cpu_limits) {
2164 		unsigned int max_freq, min_freq;
2165 
2166 		max_freq = policy->cpuinfo.max_freq *
2167 						limits->max_sysfs_pct / 100;
2168 		min_freq = policy->cpuinfo.max_freq *
2169 						limits->min_sysfs_pct / 100;
2170 		cpufreq_verify_within_limits(policy, min_freq, max_freq);
2171 	}
2172 
2173 	return 0;
2174 }
2175 
2176 static void intel_cpufreq_stop_cpu(struct cpufreq_policy *policy)
2177 {
2178 	intel_pstate_set_min_pstate(all_cpu_data[policy->cpu]);
2179 }
2180 
2181 static void intel_pstate_stop_cpu(struct cpufreq_policy *policy)
2182 {
2183 	pr_debug("CPU %d exiting\n", policy->cpu);
2184 
2185 	intel_pstate_clear_update_util_hook(policy->cpu);
2186 	if (hwp_active)
2187 		intel_pstate_hwp_save_state(policy);
2188 	else
2189 		intel_cpufreq_stop_cpu(policy);
2190 }
2191 
2192 static int intel_pstate_cpu_exit(struct cpufreq_policy *policy)
2193 {
2194 	intel_pstate_exit_perf_limits(policy);
2195 
2196 	policy->fast_switch_possible = false;
2197 
2198 	return 0;
2199 }
2200 
2201 static int __intel_pstate_cpu_init(struct cpufreq_policy *policy)
2202 {
2203 	struct cpudata *cpu;
2204 	int rc;
2205 
2206 	rc = intel_pstate_init_cpu(policy->cpu);
2207 	if (rc)
2208 		return rc;
2209 
2210 	cpu = all_cpu_data[policy->cpu];
2211 
2212 	/*
2213 	 * We need sane value in the cpu->perf_limits, so inherit from global
2214 	 * perf_limits limits, which are seeded with values based on the
2215 	 * CONFIG_CPU_FREQ_DEFAULT_GOV_*, during boot up.
2216 	 */
2217 	if (per_cpu_limits)
2218 		memcpy(cpu->perf_limits, limits, sizeof(struct perf_limits));
2219 
2220 	policy->min = cpu->pstate.min_pstate * cpu->pstate.scaling;
2221 	policy->max = cpu->pstate.turbo_pstate * cpu->pstate.scaling;
2222 
2223 	/* cpuinfo and default policy values */
2224 	policy->cpuinfo.min_freq = cpu->pstate.min_pstate * cpu->pstate.scaling;
2225 	update_turbo_state();
2226 	policy->cpuinfo.max_freq = limits->turbo_disabled ?
2227 			cpu->pstate.max_pstate : cpu->pstate.turbo_pstate;
2228 	policy->cpuinfo.max_freq *= cpu->pstate.scaling;
2229 
2230 	intel_pstate_init_acpi_perf_limits(policy);
2231 	cpumask_set_cpu(policy->cpu, policy->cpus);
2232 
2233 	policy->fast_switch_possible = true;
2234 
2235 	return 0;
2236 }
2237 
2238 static int intel_pstate_cpu_init(struct cpufreq_policy *policy)
2239 {
2240 	int ret = __intel_pstate_cpu_init(policy);
2241 
2242 	if (ret)
2243 		return ret;
2244 
2245 	policy->cpuinfo.transition_latency = CPUFREQ_ETERNAL;
2246 	if (limits->min_perf_pct == 100 && limits->max_perf_pct == 100)
2247 		policy->policy = CPUFREQ_POLICY_PERFORMANCE;
2248 	else
2249 		policy->policy = CPUFREQ_POLICY_POWERSAVE;
2250 
2251 	return 0;
2252 }
2253 
2254 static struct cpufreq_driver intel_pstate = {
2255 	.flags		= CPUFREQ_CONST_LOOPS,
2256 	.verify		= intel_pstate_verify_policy,
2257 	.setpolicy	= intel_pstate_set_policy,
2258 	.suspend	= intel_pstate_hwp_save_state,
2259 	.resume		= intel_pstate_resume,
2260 	.get		= intel_pstate_get,
2261 	.init		= intel_pstate_cpu_init,
2262 	.exit		= intel_pstate_cpu_exit,
2263 	.stop_cpu	= intel_pstate_stop_cpu,
2264 	.name		= "intel_pstate",
2265 };
2266 
2267 static int intel_cpufreq_verify_policy(struct cpufreq_policy *policy)
2268 {
2269 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2270 	struct perf_limits *perf_limits = limits;
2271 
2272 	update_turbo_state();
2273 	policy->cpuinfo.max_freq = limits->turbo_disabled ?
2274 			cpu->pstate.max_freq : cpu->pstate.turbo_freq;
2275 
2276 	cpufreq_verify_within_cpu_limits(policy);
2277 
2278 	if (per_cpu_limits)
2279 		perf_limits = cpu->perf_limits;
2280 
2281 	mutex_lock(&intel_pstate_limits_lock);
2282 
2283 	intel_pstate_update_perf_limits(policy, perf_limits);
2284 
2285 	mutex_unlock(&intel_pstate_limits_lock);
2286 
2287 	return 0;
2288 }
2289 
2290 static unsigned int intel_cpufreq_turbo_update(struct cpudata *cpu,
2291 					       struct cpufreq_policy *policy,
2292 					       unsigned int target_freq)
2293 {
2294 	unsigned int max_freq;
2295 
2296 	update_turbo_state();
2297 
2298 	max_freq = limits->no_turbo || limits->turbo_disabled ?
2299 			cpu->pstate.max_freq : cpu->pstate.turbo_freq;
2300 	policy->cpuinfo.max_freq = max_freq;
2301 	if (policy->max > max_freq)
2302 		policy->max = max_freq;
2303 
2304 	if (target_freq > max_freq)
2305 		target_freq = max_freq;
2306 
2307 	return target_freq;
2308 }
2309 
2310 static int intel_cpufreq_target(struct cpufreq_policy *policy,
2311 				unsigned int target_freq,
2312 				unsigned int relation)
2313 {
2314 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2315 	struct cpufreq_freqs freqs;
2316 	int target_pstate;
2317 
2318 	freqs.old = policy->cur;
2319 	freqs.new = intel_cpufreq_turbo_update(cpu, policy, target_freq);
2320 
2321 	cpufreq_freq_transition_begin(policy, &freqs);
2322 	switch (relation) {
2323 	case CPUFREQ_RELATION_L:
2324 		target_pstate = DIV_ROUND_UP(freqs.new, cpu->pstate.scaling);
2325 		break;
2326 	case CPUFREQ_RELATION_H:
2327 		target_pstate = freqs.new / cpu->pstate.scaling;
2328 		break;
2329 	default:
2330 		target_pstate = DIV_ROUND_CLOSEST(freqs.new, cpu->pstate.scaling);
2331 		break;
2332 	}
2333 	target_pstate = intel_pstate_prepare_request(cpu, target_pstate);
2334 	if (target_pstate != cpu->pstate.current_pstate) {
2335 		cpu->pstate.current_pstate = target_pstate;
2336 		wrmsrl_on_cpu(policy->cpu, MSR_IA32_PERF_CTL,
2337 			      pstate_funcs.get_val(cpu, target_pstate));
2338 	}
2339 	cpufreq_freq_transition_end(policy, &freqs, false);
2340 
2341 	return 0;
2342 }
2343 
2344 static unsigned int intel_cpufreq_fast_switch(struct cpufreq_policy *policy,
2345 					      unsigned int target_freq)
2346 {
2347 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2348 	int target_pstate;
2349 
2350 	target_freq = intel_cpufreq_turbo_update(cpu, policy, target_freq);
2351 	target_pstate = DIV_ROUND_UP(target_freq, cpu->pstate.scaling);
2352 	intel_pstate_update_pstate(cpu, target_pstate);
2353 	return target_freq;
2354 }
2355 
2356 static int intel_cpufreq_cpu_init(struct cpufreq_policy *policy)
2357 {
2358 	int ret = __intel_pstate_cpu_init(policy);
2359 
2360 	if (ret)
2361 		return ret;
2362 
2363 	policy->cpuinfo.transition_latency = INTEL_CPUFREQ_TRANSITION_LATENCY;
2364 	/* This reflects the intel_pstate_get_cpu_pstates() setting. */
2365 	policy->cur = policy->cpuinfo.min_freq;
2366 
2367 	return 0;
2368 }
2369 
2370 static struct cpufreq_driver intel_cpufreq = {
2371 	.flags		= CPUFREQ_CONST_LOOPS,
2372 	.verify		= intel_cpufreq_verify_policy,
2373 	.target		= intel_cpufreq_target,
2374 	.fast_switch	= intel_cpufreq_fast_switch,
2375 	.init		= intel_cpufreq_cpu_init,
2376 	.exit		= intel_pstate_cpu_exit,
2377 	.stop_cpu	= intel_cpufreq_stop_cpu,
2378 	.name		= "intel_cpufreq",
2379 };
2380 
2381 static struct cpufreq_driver *intel_pstate_driver = &intel_pstate;
2382 
2383 static void intel_pstate_driver_cleanup(void)
2384 {
2385 	unsigned int cpu;
2386 
2387 	get_online_cpus();
2388 	for_each_online_cpu(cpu) {
2389 		if (all_cpu_data[cpu]) {
2390 			if (intel_pstate_driver == &intel_pstate)
2391 				intel_pstate_clear_update_util_hook(cpu);
2392 
2393 			kfree(all_cpu_data[cpu]);
2394 			all_cpu_data[cpu] = NULL;
2395 		}
2396 	}
2397 	put_online_cpus();
2398 }
2399 
2400 static int intel_pstate_register_driver(void)
2401 {
2402 	int ret;
2403 
2404 	ret = cpufreq_register_driver(intel_pstate_driver);
2405 	if (ret) {
2406 		intel_pstate_driver_cleanup();
2407 		return ret;
2408 	}
2409 
2410 	mutex_lock(&intel_pstate_limits_lock);
2411 	driver_registered = true;
2412 	mutex_unlock(&intel_pstate_limits_lock);
2413 
2414 	if (intel_pstate_driver == &intel_pstate && !hwp_active &&
2415 	    pstate_funcs.get_target_pstate != get_target_pstate_use_cpu_load)
2416 		intel_pstate_debug_expose_params();
2417 
2418 	return 0;
2419 }
2420 
2421 static int intel_pstate_unregister_driver(void)
2422 {
2423 	if (hwp_active)
2424 		return -EBUSY;
2425 
2426 	if (intel_pstate_driver == &intel_pstate && !hwp_active &&
2427 	    pstate_funcs.get_target_pstate != get_target_pstate_use_cpu_load)
2428 		intel_pstate_debug_hide_params();
2429 
2430 	mutex_lock(&intel_pstate_limits_lock);
2431 	driver_registered = false;
2432 	mutex_unlock(&intel_pstate_limits_lock);
2433 
2434 	cpufreq_unregister_driver(intel_pstate_driver);
2435 	intel_pstate_driver_cleanup();
2436 
2437 	return 0;
2438 }
2439 
2440 static ssize_t intel_pstate_show_status(char *buf)
2441 {
2442 	if (!driver_registered)
2443 		return sprintf(buf, "off\n");
2444 
2445 	return sprintf(buf, "%s\n", intel_pstate_driver == &intel_pstate ?
2446 					"active" : "passive");
2447 }
2448 
2449 static int intel_pstate_update_status(const char *buf, size_t size)
2450 {
2451 	int ret;
2452 
2453 	if (size == 3 && !strncmp(buf, "off", size))
2454 		return driver_registered ?
2455 			intel_pstate_unregister_driver() : -EINVAL;
2456 
2457 	if (size == 6 && !strncmp(buf, "active", size)) {
2458 		if (driver_registered) {
2459 			if (intel_pstate_driver == &intel_pstate)
2460 				return 0;
2461 
2462 			ret = intel_pstate_unregister_driver();
2463 			if (ret)
2464 				return ret;
2465 		}
2466 
2467 		intel_pstate_driver = &intel_pstate;
2468 		return intel_pstate_register_driver();
2469 	}
2470 
2471 	if (size == 7 && !strncmp(buf, "passive", size)) {
2472 		if (driver_registered) {
2473 			if (intel_pstate_driver != &intel_pstate)
2474 				return 0;
2475 
2476 			ret = intel_pstate_unregister_driver();
2477 			if (ret)
2478 				return ret;
2479 		}
2480 
2481 		intel_pstate_driver = &intel_cpufreq;
2482 		return intel_pstate_register_driver();
2483 	}
2484 
2485 	return -EINVAL;
2486 }
2487 
2488 static int no_load __initdata;
2489 static int no_hwp __initdata;
2490 static int hwp_only __initdata;
2491 static unsigned int force_load __initdata;
2492 
2493 static int __init intel_pstate_msrs_not_valid(void)
2494 {
2495 	if (!pstate_funcs.get_max() ||
2496 	    !pstate_funcs.get_min() ||
2497 	    !pstate_funcs.get_turbo())
2498 		return -ENODEV;
2499 
2500 	return 0;
2501 }
2502 
2503 static void __init copy_pid_params(struct pstate_adjust_policy *policy)
2504 {
2505 	pid_params.sample_rate_ms = policy->sample_rate_ms;
2506 	pid_params.sample_rate_ns = pid_params.sample_rate_ms * NSEC_PER_MSEC;
2507 	pid_params.p_gain_pct = policy->p_gain_pct;
2508 	pid_params.i_gain_pct = policy->i_gain_pct;
2509 	pid_params.d_gain_pct = policy->d_gain_pct;
2510 	pid_params.deadband = policy->deadband;
2511 	pid_params.setpoint = policy->setpoint;
2512 }
2513 
2514 #ifdef CONFIG_ACPI
2515 static void intel_pstate_use_acpi_profile(void)
2516 {
2517 	if (acpi_gbl_FADT.preferred_profile == PM_MOBILE)
2518 		pstate_funcs.get_target_pstate =
2519 				get_target_pstate_use_cpu_load;
2520 }
2521 #else
2522 static void intel_pstate_use_acpi_profile(void)
2523 {
2524 }
2525 #endif
2526 
2527 static void __init copy_cpu_funcs(struct pstate_funcs *funcs)
2528 {
2529 	pstate_funcs.get_max   = funcs->get_max;
2530 	pstate_funcs.get_max_physical = funcs->get_max_physical;
2531 	pstate_funcs.get_min   = funcs->get_min;
2532 	pstate_funcs.get_turbo = funcs->get_turbo;
2533 	pstate_funcs.get_scaling = funcs->get_scaling;
2534 	pstate_funcs.get_val   = funcs->get_val;
2535 	pstate_funcs.get_vid   = funcs->get_vid;
2536 	pstate_funcs.get_target_pstate = funcs->get_target_pstate;
2537 
2538 	intel_pstate_use_acpi_profile();
2539 }
2540 
2541 #ifdef CONFIG_ACPI
2542 
2543 static bool __init intel_pstate_no_acpi_pss(void)
2544 {
2545 	int i;
2546 
2547 	for_each_possible_cpu(i) {
2548 		acpi_status status;
2549 		union acpi_object *pss;
2550 		struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
2551 		struct acpi_processor *pr = per_cpu(processors, i);
2552 
2553 		if (!pr)
2554 			continue;
2555 
2556 		status = acpi_evaluate_object(pr->handle, "_PSS", NULL, &buffer);
2557 		if (ACPI_FAILURE(status))
2558 			continue;
2559 
2560 		pss = buffer.pointer;
2561 		if (pss && pss->type == ACPI_TYPE_PACKAGE) {
2562 			kfree(pss);
2563 			return false;
2564 		}
2565 
2566 		kfree(pss);
2567 	}
2568 
2569 	return true;
2570 }
2571 
2572 static bool __init intel_pstate_has_acpi_ppc(void)
2573 {
2574 	int i;
2575 
2576 	for_each_possible_cpu(i) {
2577 		struct acpi_processor *pr = per_cpu(processors, i);
2578 
2579 		if (!pr)
2580 			continue;
2581 		if (acpi_has_method(pr->handle, "_PPC"))
2582 			return true;
2583 	}
2584 	return false;
2585 }
2586 
2587 enum {
2588 	PSS,
2589 	PPC,
2590 };
2591 
2592 struct hw_vendor_info {
2593 	u16  valid;
2594 	char oem_id[ACPI_OEM_ID_SIZE];
2595 	char oem_table_id[ACPI_OEM_TABLE_ID_SIZE];
2596 	int  oem_pwr_table;
2597 };
2598 
2599 /* Hardware vendor-specific info that has its own power management modes */
2600 static struct hw_vendor_info vendor_info[] __initdata = {
2601 	{1, "HP    ", "ProLiant", PSS},
2602 	{1, "ORACLE", "X4-2    ", PPC},
2603 	{1, "ORACLE", "X4-2L   ", PPC},
2604 	{1, "ORACLE", "X4-2B   ", PPC},
2605 	{1, "ORACLE", "X3-2    ", PPC},
2606 	{1, "ORACLE", "X3-2L   ", PPC},
2607 	{1, "ORACLE", "X3-2B   ", PPC},
2608 	{1, "ORACLE", "X4470M2 ", PPC},
2609 	{1, "ORACLE", "X4270M3 ", PPC},
2610 	{1, "ORACLE", "X4270M2 ", PPC},
2611 	{1, "ORACLE", "X4170M2 ", PPC},
2612 	{1, "ORACLE", "X4170 M3", PPC},
2613 	{1, "ORACLE", "X4275 M3", PPC},
2614 	{1, "ORACLE", "X6-2    ", PPC},
2615 	{1, "ORACLE", "Sudbury ", PPC},
2616 	{0, "", ""},
2617 };
2618 
2619 static bool __init intel_pstate_platform_pwr_mgmt_exists(void)
2620 {
2621 	struct acpi_table_header hdr;
2622 	struct hw_vendor_info *v_info;
2623 	const struct x86_cpu_id *id;
2624 	u64 misc_pwr;
2625 
2626 	id = x86_match_cpu(intel_pstate_cpu_oob_ids);
2627 	if (id) {
2628 		rdmsrl(MSR_MISC_PWR_MGMT, misc_pwr);
2629 		if ( misc_pwr & (1 << 8))
2630 			return true;
2631 	}
2632 
2633 	if (acpi_disabled ||
2634 	    ACPI_FAILURE(acpi_get_table_header(ACPI_SIG_FADT, 0, &hdr)))
2635 		return false;
2636 
2637 	for (v_info = vendor_info; v_info->valid; v_info++) {
2638 		if (!strncmp(hdr.oem_id, v_info->oem_id, ACPI_OEM_ID_SIZE) &&
2639 			!strncmp(hdr.oem_table_id, v_info->oem_table_id,
2640 						ACPI_OEM_TABLE_ID_SIZE))
2641 			switch (v_info->oem_pwr_table) {
2642 			case PSS:
2643 				return intel_pstate_no_acpi_pss();
2644 			case PPC:
2645 				return intel_pstate_has_acpi_ppc() &&
2646 					(!force_load);
2647 			}
2648 	}
2649 
2650 	return false;
2651 }
2652 
2653 static void intel_pstate_request_control_from_smm(void)
2654 {
2655 	/*
2656 	 * It may be unsafe to request P-states control from SMM if _PPC support
2657 	 * has not been enabled.
2658 	 */
2659 	if (acpi_ppc)
2660 		acpi_processor_pstate_control();
2661 }
2662 #else /* CONFIG_ACPI not enabled */
2663 static inline bool intel_pstate_platform_pwr_mgmt_exists(void) { return false; }
2664 static inline bool intel_pstate_has_acpi_ppc(void) { return false; }
2665 static inline void intel_pstate_request_control_from_smm(void) {}
2666 #endif /* CONFIG_ACPI */
2667 
2668 static const struct x86_cpu_id hwp_support_ids[] __initconst = {
2669 	{ X86_VENDOR_INTEL, 6, X86_MODEL_ANY, X86_FEATURE_HWP },
2670 	{}
2671 };
2672 
2673 static int __init intel_pstate_init(void)
2674 {
2675 	const struct x86_cpu_id *id;
2676 	struct cpu_defaults *cpu_def;
2677 	int rc = 0;
2678 
2679 	if (no_load)
2680 		return -ENODEV;
2681 
2682 	if (x86_match_cpu(hwp_support_ids) && !no_hwp) {
2683 		copy_cpu_funcs(&core_params.funcs);
2684 		hwp_active++;
2685 		intel_pstate.attr = hwp_cpufreq_attrs;
2686 		goto hwp_cpu_matched;
2687 	}
2688 
2689 	id = x86_match_cpu(intel_pstate_cpu_ids);
2690 	if (!id)
2691 		return -ENODEV;
2692 
2693 	cpu_def = (struct cpu_defaults *)id->driver_data;
2694 
2695 	copy_pid_params(&cpu_def->pid_policy);
2696 	copy_cpu_funcs(&cpu_def->funcs);
2697 
2698 	if (intel_pstate_msrs_not_valid())
2699 		return -ENODEV;
2700 
2701 hwp_cpu_matched:
2702 	/*
2703 	 * The Intel pstate driver will be ignored if the platform
2704 	 * firmware has its own power management modes.
2705 	 */
2706 	if (intel_pstate_platform_pwr_mgmt_exists())
2707 		return -ENODEV;
2708 
2709 	if (!hwp_active && hwp_only)
2710 		return -ENOTSUPP;
2711 
2712 	pr_info("Intel P-state driver initializing\n");
2713 
2714 	all_cpu_data = vzalloc(sizeof(void *) * num_possible_cpus());
2715 	if (!all_cpu_data)
2716 		return -ENOMEM;
2717 
2718 	intel_pstate_request_control_from_smm();
2719 
2720 	intel_pstate_sysfs_expose_params();
2721 
2722 	mutex_lock(&intel_pstate_driver_lock);
2723 	rc = intel_pstate_register_driver();
2724 	mutex_unlock(&intel_pstate_driver_lock);
2725 	if (rc)
2726 		return rc;
2727 
2728 	if (hwp_active)
2729 		pr_info("HWP enabled\n");
2730 
2731 	return 0;
2732 }
2733 device_initcall(intel_pstate_init);
2734 
2735 static int __init intel_pstate_setup(char *str)
2736 {
2737 	if (!str)
2738 		return -EINVAL;
2739 
2740 	if (!strcmp(str, "disable")) {
2741 		no_load = 1;
2742 	} else if (!strcmp(str, "passive")) {
2743 		pr_info("Passive mode enabled\n");
2744 		intel_pstate_driver = &intel_cpufreq;
2745 		no_hwp = 1;
2746 	}
2747 	if (!strcmp(str, "no_hwp")) {
2748 		pr_info("HWP disabled\n");
2749 		no_hwp = 1;
2750 	}
2751 	if (!strcmp(str, "force"))
2752 		force_load = 1;
2753 	if (!strcmp(str, "hwp_only"))
2754 		hwp_only = 1;
2755 	if (!strcmp(str, "per_cpu_perf_limits"))
2756 		per_cpu_limits = true;
2757 
2758 #ifdef CONFIG_ACPI
2759 	if (!strcmp(str, "support_acpi_ppc"))
2760 		acpi_ppc = true;
2761 #endif
2762 
2763 	return 0;
2764 }
2765 early_param("intel_pstate", intel_pstate_setup);
2766 
2767 MODULE_AUTHOR("Dirk Brandewie <dirk.j.brandewie@intel.com>");
2768 MODULE_DESCRIPTION("'intel_pstate' - P state driver Intel Core processors");
2769 MODULE_LICENSE("GPL");
2770