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/acpi.h>
30 #include <linux/vmalloc.h>
31 #include <trace/events/power.h>
32 
33 #include <asm/div64.h>
34 #include <asm/msr.h>
35 #include <asm/cpu_device_id.h>
36 #include <asm/cpufeature.h>
37 #include <asm/intel-family.h>
38 
39 #define INTEL_PSTATE_SAMPLING_INTERVAL	(10 * NSEC_PER_MSEC)
40 
41 #define INTEL_CPUFREQ_TRANSITION_LATENCY	20000
42 #define INTEL_CPUFREQ_TRANSITION_DELAY		500
43 
44 #ifdef CONFIG_ACPI
45 #include <acpi/processor.h>
46 #include <acpi/cppc_acpi.h>
47 #endif
48 
49 #define FRAC_BITS 8
50 #define int_tofp(X) ((int64_t)(X) << FRAC_BITS)
51 #define fp_toint(X) ((X) >> FRAC_BITS)
52 
53 #define ONE_EIGHTH_FP ((int64_t)1 << (FRAC_BITS - 3))
54 
55 #define EXT_BITS 6
56 #define EXT_FRAC_BITS (EXT_BITS + FRAC_BITS)
57 #define fp_ext_toint(X) ((X) >> EXT_FRAC_BITS)
58 #define int_ext_tofp(X) ((int64_t)(X) << EXT_FRAC_BITS)
59 
60 static inline int32_t mul_fp(int32_t x, int32_t y)
61 {
62 	return ((int64_t)x * (int64_t)y) >> FRAC_BITS;
63 }
64 
65 static inline int32_t div_fp(s64 x, s64 y)
66 {
67 	return div64_s64((int64_t)x << FRAC_BITS, y);
68 }
69 
70 static inline int ceiling_fp(int32_t x)
71 {
72 	int mask, ret;
73 
74 	ret = fp_toint(x);
75 	mask = (1 << FRAC_BITS) - 1;
76 	if (x & mask)
77 		ret += 1;
78 	return ret;
79 }
80 
81 static inline int32_t percent_fp(int percent)
82 {
83 	return div_fp(percent, 100);
84 }
85 
86 static inline u64 mul_ext_fp(u64 x, u64 y)
87 {
88 	return (x * y) >> EXT_FRAC_BITS;
89 }
90 
91 static inline u64 div_ext_fp(u64 x, u64 y)
92 {
93 	return div64_u64(x << EXT_FRAC_BITS, y);
94 }
95 
96 static inline int32_t percent_ext_fp(int percent)
97 {
98 	return div_ext_fp(percent, 100);
99 }
100 
101 /**
102  * struct sample -	Store performance sample
103  * @core_avg_perf:	Ratio of APERF/MPERF which is the actual average
104  *			performance during last sample period
105  * @busy_scaled:	Scaled busy value which is used to calculate next
106  *			P state. This can be different than core_avg_perf
107  *			to account for cpu idle period
108  * @aperf:		Difference of actual performance frequency clock count
109  *			read from APERF MSR between last and current sample
110  * @mperf:		Difference of maximum performance frequency clock count
111  *			read from MPERF MSR between last and current sample
112  * @tsc:		Difference of time stamp counter between last and
113  *			current sample
114  * @time:		Current time from scheduler
115  *
116  * This structure is used in the cpudata structure to store performance sample
117  * data for choosing next P State.
118  */
119 struct sample {
120 	int32_t core_avg_perf;
121 	int32_t busy_scaled;
122 	u64 aperf;
123 	u64 mperf;
124 	u64 tsc;
125 	u64 time;
126 };
127 
128 /**
129  * struct pstate_data - Store P state data
130  * @current_pstate:	Current requested P state
131  * @min_pstate:		Min P state possible for this platform
132  * @max_pstate:		Max P state possible for this platform
133  * @max_pstate_physical:This is physical Max P state for a processor
134  *			This can be higher than the max_pstate which can
135  *			be limited by platform thermal design power limits
136  * @scaling:		Scaling factor to  convert frequency to cpufreq
137  *			frequency units
138  * @turbo_pstate:	Max Turbo P state possible for this platform
139  * @max_freq:		@max_pstate frequency in cpufreq units
140  * @turbo_freq:		@turbo_pstate frequency in cpufreq units
141  *
142  * Stores the per cpu model P state limits and current P state.
143  */
144 struct pstate_data {
145 	int	current_pstate;
146 	int	min_pstate;
147 	int	max_pstate;
148 	int	max_pstate_physical;
149 	int	scaling;
150 	int	turbo_pstate;
151 	unsigned int max_freq;
152 	unsigned int turbo_freq;
153 };
154 
155 /**
156  * struct vid_data -	Stores voltage information data
157  * @min:		VID data for this platform corresponding to
158  *			the lowest P state
159  * @max:		VID data corresponding to the highest P State.
160  * @turbo:		VID data for turbo P state
161  * @ratio:		Ratio of (vid max - vid min) /
162  *			(max P state - Min P State)
163  *
164  * Stores the voltage data for DVFS (Dynamic Voltage and Frequency Scaling)
165  * This data is used in Atom platforms, where in addition to target P state,
166  * the voltage data needs to be specified to select next P State.
167  */
168 struct vid_data {
169 	int min;
170 	int max;
171 	int turbo;
172 	int32_t ratio;
173 };
174 
175 /**
176  * struct global_params - Global parameters, mostly tunable via sysfs.
177  * @no_turbo:		Whether or not to use turbo P-states.
178  * @turbo_disabled:	Whethet or not turbo P-states are available at all,
179  *			based on the MSR_IA32_MISC_ENABLE value and whether or
180  *			not the maximum reported turbo P-state is different from
181  *			the maximum reported non-turbo one.
182  * @min_perf_pct:	Minimum capacity limit in percent of the maximum turbo
183  *			P-state capacity.
184  * @max_perf_pct:	Maximum capacity limit in percent of the maximum turbo
185  *			P-state capacity.
186  */
187 struct global_params {
188 	bool no_turbo;
189 	bool turbo_disabled;
190 	int max_perf_pct;
191 	int min_perf_pct;
192 };
193 
194 /**
195  * struct cpudata -	Per CPU instance data storage
196  * @cpu:		CPU number for this instance data
197  * @policy:		CPUFreq policy value
198  * @update_util:	CPUFreq utility callback information
199  * @update_util_set:	CPUFreq utility callback is set
200  * @iowait_boost:	iowait-related boost fraction
201  * @last_update:	Time of the last update.
202  * @pstate:		Stores P state limits for this CPU
203  * @vid:		Stores VID limits for this CPU
204  * @last_sample_time:	Last Sample time
205  * @aperf_mperf_shift:	Number of clock cycles after aperf, merf is incremented
206  *			This shift is a multiplier to mperf delta to
207  *			calculate CPU busy.
208  * @prev_aperf:		Last APERF value read from APERF MSR
209  * @prev_mperf:		Last MPERF value read from MPERF MSR
210  * @prev_tsc:		Last timestamp counter (TSC) value
211  * @prev_cummulative_iowait: IO Wait time difference from last and
212  *			current sample
213  * @sample:		Storage for storing last Sample data
214  * @min_perf_ratio:	Minimum capacity in terms of PERF or HWP ratios
215  * @max_perf_ratio:	Maximum capacity in terms of PERF or HWP ratios
216  * @acpi_perf_data:	Stores ACPI perf information read from _PSS
217  * @valid_pss_table:	Set to true for valid ACPI _PSS entries found
218  * @epp_powersave:	Last saved HWP energy performance preference
219  *			(EPP) or energy performance bias (EPB),
220  *			when policy switched to performance
221  * @epp_policy:		Last saved policy used to set EPP/EPB
222  * @epp_default:	Power on default HWP energy performance
223  *			preference/bias
224  * @epp_saved:		Saved EPP/EPB during system suspend or CPU offline
225  *			operation
226  * @hwp_req_cached:	Cached value of the last HWP Request MSR
227  * @hwp_cap_cached:	Cached value of the last HWP Capabilities MSR
228  * @last_io_update:	Last time when IO wake flag was set
229  * @sched_flags:	Store scheduler flags for possible cross CPU update
230  * @hwp_boost_min:	Last HWP boosted min performance
231  *
232  * This structure stores per CPU instance data for all CPUs.
233  */
234 struct cpudata {
235 	int cpu;
236 
237 	unsigned int policy;
238 	struct update_util_data update_util;
239 	bool   update_util_set;
240 
241 	struct pstate_data pstate;
242 	struct vid_data vid;
243 
244 	u64	last_update;
245 	u64	last_sample_time;
246 	u64	aperf_mperf_shift;
247 	u64	prev_aperf;
248 	u64	prev_mperf;
249 	u64	prev_tsc;
250 	u64	prev_cummulative_iowait;
251 	struct sample sample;
252 	int32_t	min_perf_ratio;
253 	int32_t	max_perf_ratio;
254 #ifdef CONFIG_ACPI
255 	struct acpi_processor_performance acpi_perf_data;
256 	bool valid_pss_table;
257 #endif
258 	unsigned int iowait_boost;
259 	s16 epp_powersave;
260 	s16 epp_policy;
261 	s16 epp_default;
262 	s16 epp_saved;
263 	u64 hwp_req_cached;
264 	u64 hwp_cap_cached;
265 	u64 last_io_update;
266 	unsigned int sched_flags;
267 	u32 hwp_boost_min;
268 };
269 
270 static struct cpudata **all_cpu_data;
271 
272 /**
273  * struct pstate_funcs - Per CPU model specific callbacks
274  * @get_max:		Callback to get maximum non turbo effective P state
275  * @get_max_physical:	Callback to get maximum non turbo physical P state
276  * @get_min:		Callback to get minimum P state
277  * @get_turbo:		Callback to get turbo P state
278  * @get_scaling:	Callback to get frequency scaling factor
279  * @get_val:		Callback to convert P state to actual MSR write value
280  * @get_vid:		Callback to get VID data for Atom platforms
281  *
282  * Core and Atom CPU models have different way to get P State limits. This
283  * structure is used to store those callbacks.
284  */
285 struct pstate_funcs {
286 	int (*get_max)(void);
287 	int (*get_max_physical)(void);
288 	int (*get_min)(void);
289 	int (*get_turbo)(void);
290 	int (*get_scaling)(void);
291 	int (*get_aperf_mperf_shift)(void);
292 	u64 (*get_val)(struct cpudata*, int pstate);
293 	void (*get_vid)(struct cpudata *);
294 };
295 
296 static struct pstate_funcs pstate_funcs __read_mostly;
297 
298 static int hwp_active __read_mostly;
299 static int hwp_mode_bdw __read_mostly;
300 static bool per_cpu_limits __read_mostly;
301 static bool hwp_boost __read_mostly;
302 
303 static struct cpufreq_driver *intel_pstate_driver __read_mostly;
304 
305 #ifdef CONFIG_ACPI
306 static bool acpi_ppc;
307 #endif
308 
309 static struct global_params global;
310 
311 static DEFINE_MUTEX(intel_pstate_driver_lock);
312 static DEFINE_MUTEX(intel_pstate_limits_lock);
313 
314 #ifdef CONFIG_ACPI
315 
316 static bool intel_pstate_acpi_pm_profile_server(void)
317 {
318 	if (acpi_gbl_FADT.preferred_profile == PM_ENTERPRISE_SERVER ||
319 	    acpi_gbl_FADT.preferred_profile == PM_PERFORMANCE_SERVER)
320 		return true;
321 
322 	return false;
323 }
324 
325 static bool intel_pstate_get_ppc_enable_status(void)
326 {
327 	if (intel_pstate_acpi_pm_profile_server())
328 		return true;
329 
330 	return acpi_ppc;
331 }
332 
333 #ifdef CONFIG_ACPI_CPPC_LIB
334 
335 /* The work item is needed to avoid CPU hotplug locking issues */
336 static void intel_pstste_sched_itmt_work_fn(struct work_struct *work)
337 {
338 	sched_set_itmt_support();
339 }
340 
341 static DECLARE_WORK(sched_itmt_work, intel_pstste_sched_itmt_work_fn);
342 
343 static void intel_pstate_set_itmt_prio(int cpu)
344 {
345 	struct cppc_perf_caps cppc_perf;
346 	static u32 max_highest_perf = 0, min_highest_perf = U32_MAX;
347 	int ret;
348 
349 	ret = cppc_get_perf_caps(cpu, &cppc_perf);
350 	if (ret)
351 		return;
352 
353 	/*
354 	 * The priorities can be set regardless of whether or not
355 	 * sched_set_itmt_support(true) has been called and it is valid to
356 	 * update them at any time after it has been called.
357 	 */
358 	sched_set_itmt_core_prio(cppc_perf.highest_perf, cpu);
359 
360 	if (max_highest_perf <= min_highest_perf) {
361 		if (cppc_perf.highest_perf > max_highest_perf)
362 			max_highest_perf = cppc_perf.highest_perf;
363 
364 		if (cppc_perf.highest_perf < min_highest_perf)
365 			min_highest_perf = cppc_perf.highest_perf;
366 
367 		if (max_highest_perf > min_highest_perf) {
368 			/*
369 			 * This code can be run during CPU online under the
370 			 * CPU hotplug locks, so sched_set_itmt_support()
371 			 * cannot be called from here.  Queue up a work item
372 			 * to invoke it.
373 			 */
374 			schedule_work(&sched_itmt_work);
375 		}
376 	}
377 }
378 
379 static int intel_pstate_get_cppc_guranteed(int cpu)
380 {
381 	struct cppc_perf_caps cppc_perf;
382 	int ret;
383 
384 	ret = cppc_get_perf_caps(cpu, &cppc_perf);
385 	if (ret)
386 		return ret;
387 
388 	return cppc_perf.guaranteed_perf;
389 }
390 
391 #else /* CONFIG_ACPI_CPPC_LIB */
392 static void intel_pstate_set_itmt_prio(int cpu)
393 {
394 }
395 #endif /* CONFIG_ACPI_CPPC_LIB */
396 
397 static void intel_pstate_init_acpi_perf_limits(struct cpufreq_policy *policy)
398 {
399 	struct cpudata *cpu;
400 	int ret;
401 	int i;
402 
403 	if (hwp_active) {
404 		intel_pstate_set_itmt_prio(policy->cpu);
405 		return;
406 	}
407 
408 	if (!intel_pstate_get_ppc_enable_status())
409 		return;
410 
411 	cpu = all_cpu_data[policy->cpu];
412 
413 	ret = acpi_processor_register_performance(&cpu->acpi_perf_data,
414 						  policy->cpu);
415 	if (ret)
416 		return;
417 
418 	/*
419 	 * Check if the control value in _PSS is for PERF_CTL MSR, which should
420 	 * guarantee that the states returned by it map to the states in our
421 	 * list directly.
422 	 */
423 	if (cpu->acpi_perf_data.control_register.space_id !=
424 						ACPI_ADR_SPACE_FIXED_HARDWARE)
425 		goto err;
426 
427 	/*
428 	 * If there is only one entry _PSS, simply ignore _PSS and continue as
429 	 * usual without taking _PSS into account
430 	 */
431 	if (cpu->acpi_perf_data.state_count < 2)
432 		goto err;
433 
434 	pr_debug("CPU%u - ACPI _PSS perf data\n", policy->cpu);
435 	for (i = 0; i < cpu->acpi_perf_data.state_count; i++) {
436 		pr_debug("     %cP%d: %u MHz, %u mW, 0x%x\n",
437 			 (i == cpu->acpi_perf_data.state ? '*' : ' '), i,
438 			 (u32) cpu->acpi_perf_data.states[i].core_frequency,
439 			 (u32) cpu->acpi_perf_data.states[i].power,
440 			 (u32) cpu->acpi_perf_data.states[i].control);
441 	}
442 
443 	/*
444 	 * The _PSS table doesn't contain whole turbo frequency range.
445 	 * This just contains +1 MHZ above the max non turbo frequency,
446 	 * with control value corresponding to max turbo ratio. But
447 	 * when cpufreq set policy is called, it will call with this
448 	 * max frequency, which will cause a reduced performance as
449 	 * this driver uses real max turbo frequency as the max
450 	 * frequency. So correct this frequency in _PSS table to
451 	 * correct max turbo frequency based on the turbo state.
452 	 * Also need to convert to MHz as _PSS freq is in MHz.
453 	 */
454 	if (!global.turbo_disabled)
455 		cpu->acpi_perf_data.states[0].core_frequency =
456 					policy->cpuinfo.max_freq / 1000;
457 	cpu->valid_pss_table = true;
458 	pr_debug("_PPC limits will be enforced\n");
459 
460 	return;
461 
462  err:
463 	cpu->valid_pss_table = false;
464 	acpi_processor_unregister_performance(policy->cpu);
465 }
466 
467 static void intel_pstate_exit_perf_limits(struct cpufreq_policy *policy)
468 {
469 	struct cpudata *cpu;
470 
471 	cpu = all_cpu_data[policy->cpu];
472 	if (!cpu->valid_pss_table)
473 		return;
474 
475 	acpi_processor_unregister_performance(policy->cpu);
476 }
477 #else /* CONFIG_ACPI */
478 static inline void intel_pstate_init_acpi_perf_limits(struct cpufreq_policy *policy)
479 {
480 }
481 
482 static inline void intel_pstate_exit_perf_limits(struct cpufreq_policy *policy)
483 {
484 }
485 
486 static inline bool intel_pstate_acpi_pm_profile_server(void)
487 {
488 	return false;
489 }
490 #endif /* CONFIG_ACPI */
491 
492 #ifndef CONFIG_ACPI_CPPC_LIB
493 static int intel_pstate_get_cppc_guranteed(int cpu)
494 {
495 	return -ENOTSUPP;
496 }
497 #endif /* CONFIG_ACPI_CPPC_LIB */
498 
499 static inline void update_turbo_state(void)
500 {
501 	u64 misc_en;
502 	struct cpudata *cpu;
503 
504 	cpu = all_cpu_data[0];
505 	rdmsrl(MSR_IA32_MISC_ENABLE, misc_en);
506 	global.turbo_disabled =
507 		(misc_en & MSR_IA32_MISC_ENABLE_TURBO_DISABLE ||
508 		 cpu->pstate.max_pstate == cpu->pstate.turbo_pstate);
509 }
510 
511 static int min_perf_pct_min(void)
512 {
513 	struct cpudata *cpu = all_cpu_data[0];
514 	int turbo_pstate = cpu->pstate.turbo_pstate;
515 
516 	return turbo_pstate ?
517 		(cpu->pstate.min_pstate * 100 / turbo_pstate) : 0;
518 }
519 
520 static s16 intel_pstate_get_epb(struct cpudata *cpu_data)
521 {
522 	u64 epb;
523 	int ret;
524 
525 	if (!static_cpu_has(X86_FEATURE_EPB))
526 		return -ENXIO;
527 
528 	ret = rdmsrl_on_cpu(cpu_data->cpu, MSR_IA32_ENERGY_PERF_BIAS, &epb);
529 	if (ret)
530 		return (s16)ret;
531 
532 	return (s16)(epb & 0x0f);
533 }
534 
535 static s16 intel_pstate_get_epp(struct cpudata *cpu_data, u64 hwp_req_data)
536 {
537 	s16 epp;
538 
539 	if (static_cpu_has(X86_FEATURE_HWP_EPP)) {
540 		/*
541 		 * When hwp_req_data is 0, means that caller didn't read
542 		 * MSR_HWP_REQUEST, so need to read and get EPP.
543 		 */
544 		if (!hwp_req_data) {
545 			epp = rdmsrl_on_cpu(cpu_data->cpu, MSR_HWP_REQUEST,
546 					    &hwp_req_data);
547 			if (epp)
548 				return epp;
549 		}
550 		epp = (hwp_req_data >> 24) & 0xff;
551 	} else {
552 		/* When there is no EPP present, HWP uses EPB settings */
553 		epp = intel_pstate_get_epb(cpu_data);
554 	}
555 
556 	return epp;
557 }
558 
559 static int intel_pstate_set_epb(int cpu, s16 pref)
560 {
561 	u64 epb;
562 	int ret;
563 
564 	if (!static_cpu_has(X86_FEATURE_EPB))
565 		return -ENXIO;
566 
567 	ret = rdmsrl_on_cpu(cpu, MSR_IA32_ENERGY_PERF_BIAS, &epb);
568 	if (ret)
569 		return ret;
570 
571 	epb = (epb & ~0x0f) | pref;
572 	wrmsrl_on_cpu(cpu, MSR_IA32_ENERGY_PERF_BIAS, epb);
573 
574 	return 0;
575 }
576 
577 /*
578  * EPP/EPB display strings corresponding to EPP index in the
579  * energy_perf_strings[]
580  *	index		String
581  *-------------------------------------
582  *	0		default
583  *	1		performance
584  *	2		balance_performance
585  *	3		balance_power
586  *	4		power
587  */
588 static const char * const energy_perf_strings[] = {
589 	"default",
590 	"performance",
591 	"balance_performance",
592 	"balance_power",
593 	"power",
594 	NULL
595 };
596 static const unsigned int epp_values[] = {
597 	HWP_EPP_PERFORMANCE,
598 	HWP_EPP_BALANCE_PERFORMANCE,
599 	HWP_EPP_BALANCE_POWERSAVE,
600 	HWP_EPP_POWERSAVE
601 };
602 
603 static int intel_pstate_get_energy_pref_index(struct cpudata *cpu_data)
604 {
605 	s16 epp;
606 	int index = -EINVAL;
607 
608 	epp = intel_pstate_get_epp(cpu_data, 0);
609 	if (epp < 0)
610 		return epp;
611 
612 	if (static_cpu_has(X86_FEATURE_HWP_EPP)) {
613 		if (epp == HWP_EPP_PERFORMANCE)
614 			return 1;
615 		if (epp <= HWP_EPP_BALANCE_PERFORMANCE)
616 			return 2;
617 		if (epp <= HWP_EPP_BALANCE_POWERSAVE)
618 			return 3;
619 		else
620 			return 4;
621 	} else if (static_cpu_has(X86_FEATURE_EPB)) {
622 		/*
623 		 * Range:
624 		 *	0x00-0x03	:	Performance
625 		 *	0x04-0x07	:	Balance performance
626 		 *	0x08-0x0B	:	Balance power
627 		 *	0x0C-0x0F	:	Power
628 		 * The EPB is a 4 bit value, but our ranges restrict the
629 		 * value which can be set. Here only using top two bits
630 		 * effectively.
631 		 */
632 		index = (epp >> 2) + 1;
633 	}
634 
635 	return index;
636 }
637 
638 static int intel_pstate_set_energy_pref_index(struct cpudata *cpu_data,
639 					      int pref_index)
640 {
641 	int epp = -EINVAL;
642 	int ret;
643 
644 	if (!pref_index)
645 		epp = cpu_data->epp_default;
646 
647 	mutex_lock(&intel_pstate_limits_lock);
648 
649 	if (static_cpu_has(X86_FEATURE_HWP_EPP)) {
650 		u64 value;
651 
652 		ret = rdmsrl_on_cpu(cpu_data->cpu, MSR_HWP_REQUEST, &value);
653 		if (ret)
654 			goto return_pref;
655 
656 		value &= ~GENMASK_ULL(31, 24);
657 
658 		if (epp == -EINVAL)
659 			epp = epp_values[pref_index - 1];
660 
661 		value |= (u64)epp << 24;
662 		ret = wrmsrl_on_cpu(cpu_data->cpu, MSR_HWP_REQUEST, value);
663 	} else {
664 		if (epp == -EINVAL)
665 			epp = (pref_index - 1) << 2;
666 		ret = intel_pstate_set_epb(cpu_data->cpu, epp);
667 	}
668 return_pref:
669 	mutex_unlock(&intel_pstate_limits_lock);
670 
671 	return ret;
672 }
673 
674 static ssize_t show_energy_performance_available_preferences(
675 				struct cpufreq_policy *policy, char *buf)
676 {
677 	int i = 0;
678 	int ret = 0;
679 
680 	while (energy_perf_strings[i] != NULL)
681 		ret += sprintf(&buf[ret], "%s ", energy_perf_strings[i++]);
682 
683 	ret += sprintf(&buf[ret], "\n");
684 
685 	return ret;
686 }
687 
688 cpufreq_freq_attr_ro(energy_performance_available_preferences);
689 
690 static ssize_t store_energy_performance_preference(
691 		struct cpufreq_policy *policy, const char *buf, size_t count)
692 {
693 	struct cpudata *cpu_data = all_cpu_data[policy->cpu];
694 	char str_preference[21];
695 	int ret;
696 
697 	ret = sscanf(buf, "%20s", str_preference);
698 	if (ret != 1)
699 		return -EINVAL;
700 
701 	ret = match_string(energy_perf_strings, -1, str_preference);
702 	if (ret < 0)
703 		return ret;
704 
705 	intel_pstate_set_energy_pref_index(cpu_data, ret);
706 	return count;
707 }
708 
709 static ssize_t show_energy_performance_preference(
710 				struct cpufreq_policy *policy, char *buf)
711 {
712 	struct cpudata *cpu_data = all_cpu_data[policy->cpu];
713 	int preference;
714 
715 	preference = intel_pstate_get_energy_pref_index(cpu_data);
716 	if (preference < 0)
717 		return preference;
718 
719 	return  sprintf(buf, "%s\n", energy_perf_strings[preference]);
720 }
721 
722 cpufreq_freq_attr_rw(energy_performance_preference);
723 
724 static ssize_t show_base_frequency(struct cpufreq_policy *policy, char *buf)
725 {
726 	struct cpudata *cpu;
727 	u64 cap;
728 	int ratio;
729 
730 	ratio = intel_pstate_get_cppc_guranteed(policy->cpu);
731 	if (ratio <= 0) {
732 		rdmsrl_on_cpu(policy->cpu, MSR_HWP_CAPABILITIES, &cap);
733 		ratio = HWP_GUARANTEED_PERF(cap);
734 	}
735 
736 	cpu = all_cpu_data[policy->cpu];
737 
738 	return sprintf(buf, "%d\n", ratio * cpu->pstate.scaling);
739 }
740 
741 cpufreq_freq_attr_ro(base_frequency);
742 
743 static struct freq_attr *hwp_cpufreq_attrs[] = {
744 	&energy_performance_preference,
745 	&energy_performance_available_preferences,
746 	&base_frequency,
747 	NULL,
748 };
749 
750 static void intel_pstate_get_hwp_max(unsigned int cpu, int *phy_max,
751 				     int *current_max)
752 {
753 	u64 cap;
754 
755 	rdmsrl_on_cpu(cpu, MSR_HWP_CAPABILITIES, &cap);
756 	WRITE_ONCE(all_cpu_data[cpu]->hwp_cap_cached, cap);
757 	if (global.no_turbo)
758 		*current_max = HWP_GUARANTEED_PERF(cap);
759 	else
760 		*current_max = HWP_HIGHEST_PERF(cap);
761 
762 	*phy_max = HWP_HIGHEST_PERF(cap);
763 }
764 
765 static void intel_pstate_hwp_set(unsigned int cpu)
766 {
767 	struct cpudata *cpu_data = all_cpu_data[cpu];
768 	int max, min;
769 	u64 value;
770 	s16 epp;
771 
772 	max = cpu_data->max_perf_ratio;
773 	min = cpu_data->min_perf_ratio;
774 
775 	if (cpu_data->policy == CPUFREQ_POLICY_PERFORMANCE)
776 		min = max;
777 
778 	rdmsrl_on_cpu(cpu, MSR_HWP_REQUEST, &value);
779 
780 	value &= ~HWP_MIN_PERF(~0L);
781 	value |= HWP_MIN_PERF(min);
782 
783 	value &= ~HWP_MAX_PERF(~0L);
784 	value |= HWP_MAX_PERF(max);
785 
786 	if (cpu_data->epp_policy == cpu_data->policy)
787 		goto skip_epp;
788 
789 	cpu_data->epp_policy = cpu_data->policy;
790 
791 	if (cpu_data->epp_saved >= 0) {
792 		epp = cpu_data->epp_saved;
793 		cpu_data->epp_saved = -EINVAL;
794 		goto update_epp;
795 	}
796 
797 	if (cpu_data->policy == CPUFREQ_POLICY_PERFORMANCE) {
798 		epp = intel_pstate_get_epp(cpu_data, value);
799 		cpu_data->epp_powersave = epp;
800 		/* If EPP read was failed, then don't try to write */
801 		if (epp < 0)
802 			goto skip_epp;
803 
804 		epp = 0;
805 	} else {
806 		/* skip setting EPP, when saved value is invalid */
807 		if (cpu_data->epp_powersave < 0)
808 			goto skip_epp;
809 
810 		/*
811 		 * No need to restore EPP when it is not zero. This
812 		 * means:
813 		 *  - Policy is not changed
814 		 *  - user has manually changed
815 		 *  - Error reading EPB
816 		 */
817 		epp = intel_pstate_get_epp(cpu_data, value);
818 		if (epp)
819 			goto skip_epp;
820 
821 		epp = cpu_data->epp_powersave;
822 	}
823 update_epp:
824 	if (static_cpu_has(X86_FEATURE_HWP_EPP)) {
825 		value &= ~GENMASK_ULL(31, 24);
826 		value |= (u64)epp << 24;
827 	} else {
828 		intel_pstate_set_epb(cpu, epp);
829 	}
830 skip_epp:
831 	WRITE_ONCE(cpu_data->hwp_req_cached, value);
832 	wrmsrl_on_cpu(cpu, MSR_HWP_REQUEST, value);
833 }
834 
835 static void intel_pstate_hwp_force_min_perf(int cpu)
836 {
837 	u64 value;
838 	int min_perf;
839 
840 	value = all_cpu_data[cpu]->hwp_req_cached;
841 	value &= ~GENMASK_ULL(31, 0);
842 	min_perf = HWP_LOWEST_PERF(all_cpu_data[cpu]->hwp_cap_cached);
843 
844 	/* Set hwp_max = hwp_min */
845 	value |= HWP_MAX_PERF(min_perf);
846 	value |= HWP_MIN_PERF(min_perf);
847 
848 	/* Set EPP/EPB to min */
849 	if (static_cpu_has(X86_FEATURE_HWP_EPP))
850 		value |= HWP_ENERGY_PERF_PREFERENCE(HWP_EPP_POWERSAVE);
851 	else
852 		intel_pstate_set_epb(cpu, HWP_EPP_BALANCE_POWERSAVE);
853 
854 	wrmsrl_on_cpu(cpu, MSR_HWP_REQUEST, value);
855 }
856 
857 static int intel_pstate_hwp_save_state(struct cpufreq_policy *policy)
858 {
859 	struct cpudata *cpu_data = all_cpu_data[policy->cpu];
860 
861 	if (!hwp_active)
862 		return 0;
863 
864 	cpu_data->epp_saved = intel_pstate_get_epp(cpu_data, 0);
865 
866 	return 0;
867 }
868 
869 static void intel_pstate_hwp_enable(struct cpudata *cpudata);
870 
871 static int intel_pstate_resume(struct cpufreq_policy *policy)
872 {
873 	if (!hwp_active)
874 		return 0;
875 
876 	mutex_lock(&intel_pstate_limits_lock);
877 
878 	if (policy->cpu == 0)
879 		intel_pstate_hwp_enable(all_cpu_data[policy->cpu]);
880 
881 	all_cpu_data[policy->cpu]->epp_policy = 0;
882 	intel_pstate_hwp_set(policy->cpu);
883 
884 	mutex_unlock(&intel_pstate_limits_lock);
885 
886 	return 0;
887 }
888 
889 static void intel_pstate_update_policies(void)
890 {
891 	int cpu;
892 
893 	for_each_possible_cpu(cpu)
894 		cpufreq_update_policy(cpu);
895 }
896 
897 /************************** sysfs begin ************************/
898 #define show_one(file_name, object)					\
899 	static ssize_t show_##file_name					\
900 	(struct kobject *kobj, struct kobj_attribute *attr, char *buf)	\
901 	{								\
902 		return sprintf(buf, "%u\n", global.object);		\
903 	}
904 
905 static ssize_t intel_pstate_show_status(char *buf);
906 static int intel_pstate_update_status(const char *buf, size_t size);
907 
908 static ssize_t show_status(struct kobject *kobj,
909 			   struct kobj_attribute *attr, char *buf)
910 {
911 	ssize_t ret;
912 
913 	mutex_lock(&intel_pstate_driver_lock);
914 	ret = intel_pstate_show_status(buf);
915 	mutex_unlock(&intel_pstate_driver_lock);
916 
917 	return ret;
918 }
919 
920 static ssize_t store_status(struct kobject *a, struct kobj_attribute *b,
921 			    const char *buf, size_t count)
922 {
923 	char *p = memchr(buf, '\n', count);
924 	int ret;
925 
926 	mutex_lock(&intel_pstate_driver_lock);
927 	ret = intel_pstate_update_status(buf, p ? p - buf : count);
928 	mutex_unlock(&intel_pstate_driver_lock);
929 
930 	return ret < 0 ? ret : count;
931 }
932 
933 static ssize_t show_turbo_pct(struct kobject *kobj,
934 				struct kobj_attribute *attr, char *buf)
935 {
936 	struct cpudata *cpu;
937 	int total, no_turbo, turbo_pct;
938 	uint32_t turbo_fp;
939 
940 	mutex_lock(&intel_pstate_driver_lock);
941 
942 	if (!intel_pstate_driver) {
943 		mutex_unlock(&intel_pstate_driver_lock);
944 		return -EAGAIN;
945 	}
946 
947 	cpu = all_cpu_data[0];
948 
949 	total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
950 	no_turbo = cpu->pstate.max_pstate - cpu->pstate.min_pstate + 1;
951 	turbo_fp = div_fp(no_turbo, total);
952 	turbo_pct = 100 - fp_toint(mul_fp(turbo_fp, int_tofp(100)));
953 
954 	mutex_unlock(&intel_pstate_driver_lock);
955 
956 	return sprintf(buf, "%u\n", turbo_pct);
957 }
958 
959 static ssize_t show_num_pstates(struct kobject *kobj,
960 				struct kobj_attribute *attr, char *buf)
961 {
962 	struct cpudata *cpu;
963 	int total;
964 
965 	mutex_lock(&intel_pstate_driver_lock);
966 
967 	if (!intel_pstate_driver) {
968 		mutex_unlock(&intel_pstate_driver_lock);
969 		return -EAGAIN;
970 	}
971 
972 	cpu = all_cpu_data[0];
973 	total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
974 
975 	mutex_unlock(&intel_pstate_driver_lock);
976 
977 	return sprintf(buf, "%u\n", total);
978 }
979 
980 static ssize_t show_no_turbo(struct kobject *kobj,
981 			     struct kobj_attribute *attr, char *buf)
982 {
983 	ssize_t ret;
984 
985 	mutex_lock(&intel_pstate_driver_lock);
986 
987 	if (!intel_pstate_driver) {
988 		mutex_unlock(&intel_pstate_driver_lock);
989 		return -EAGAIN;
990 	}
991 
992 	update_turbo_state();
993 	if (global.turbo_disabled)
994 		ret = sprintf(buf, "%u\n", global.turbo_disabled);
995 	else
996 		ret = sprintf(buf, "%u\n", global.no_turbo);
997 
998 	mutex_unlock(&intel_pstate_driver_lock);
999 
1000 	return ret;
1001 }
1002 
1003 static ssize_t store_no_turbo(struct kobject *a, struct kobj_attribute *b,
1004 			      const char *buf, size_t count)
1005 {
1006 	unsigned int input;
1007 	int ret;
1008 
1009 	ret = sscanf(buf, "%u", &input);
1010 	if (ret != 1)
1011 		return -EINVAL;
1012 
1013 	mutex_lock(&intel_pstate_driver_lock);
1014 
1015 	if (!intel_pstate_driver) {
1016 		mutex_unlock(&intel_pstate_driver_lock);
1017 		return -EAGAIN;
1018 	}
1019 
1020 	mutex_lock(&intel_pstate_limits_lock);
1021 
1022 	update_turbo_state();
1023 	if (global.turbo_disabled) {
1024 		pr_warn("Turbo disabled by BIOS or unavailable on processor\n");
1025 		mutex_unlock(&intel_pstate_limits_lock);
1026 		mutex_unlock(&intel_pstate_driver_lock);
1027 		return -EPERM;
1028 	}
1029 
1030 	global.no_turbo = clamp_t(int, input, 0, 1);
1031 
1032 	if (global.no_turbo) {
1033 		struct cpudata *cpu = all_cpu_data[0];
1034 		int pct = cpu->pstate.max_pstate * 100 / cpu->pstate.turbo_pstate;
1035 
1036 		/* Squash the global minimum into the permitted range. */
1037 		if (global.min_perf_pct > pct)
1038 			global.min_perf_pct = pct;
1039 	}
1040 
1041 	mutex_unlock(&intel_pstate_limits_lock);
1042 
1043 	intel_pstate_update_policies();
1044 
1045 	mutex_unlock(&intel_pstate_driver_lock);
1046 
1047 	return count;
1048 }
1049 
1050 static ssize_t store_max_perf_pct(struct kobject *a, struct kobj_attribute *b,
1051 				  const char *buf, size_t count)
1052 {
1053 	unsigned int input;
1054 	int ret;
1055 
1056 	ret = sscanf(buf, "%u", &input);
1057 	if (ret != 1)
1058 		return -EINVAL;
1059 
1060 	mutex_lock(&intel_pstate_driver_lock);
1061 
1062 	if (!intel_pstate_driver) {
1063 		mutex_unlock(&intel_pstate_driver_lock);
1064 		return -EAGAIN;
1065 	}
1066 
1067 	mutex_lock(&intel_pstate_limits_lock);
1068 
1069 	global.max_perf_pct = clamp_t(int, input, global.min_perf_pct, 100);
1070 
1071 	mutex_unlock(&intel_pstate_limits_lock);
1072 
1073 	intel_pstate_update_policies();
1074 
1075 	mutex_unlock(&intel_pstate_driver_lock);
1076 
1077 	return count;
1078 }
1079 
1080 static ssize_t store_min_perf_pct(struct kobject *a, struct kobj_attribute *b,
1081 				  const char *buf, size_t count)
1082 {
1083 	unsigned int input;
1084 	int ret;
1085 
1086 	ret = sscanf(buf, "%u", &input);
1087 	if (ret != 1)
1088 		return -EINVAL;
1089 
1090 	mutex_lock(&intel_pstate_driver_lock);
1091 
1092 	if (!intel_pstate_driver) {
1093 		mutex_unlock(&intel_pstate_driver_lock);
1094 		return -EAGAIN;
1095 	}
1096 
1097 	mutex_lock(&intel_pstate_limits_lock);
1098 
1099 	global.min_perf_pct = clamp_t(int, input,
1100 				      min_perf_pct_min(), global.max_perf_pct);
1101 
1102 	mutex_unlock(&intel_pstate_limits_lock);
1103 
1104 	intel_pstate_update_policies();
1105 
1106 	mutex_unlock(&intel_pstate_driver_lock);
1107 
1108 	return count;
1109 }
1110 
1111 static ssize_t show_hwp_dynamic_boost(struct kobject *kobj,
1112 				struct kobj_attribute *attr, char *buf)
1113 {
1114 	return sprintf(buf, "%u\n", hwp_boost);
1115 }
1116 
1117 static ssize_t store_hwp_dynamic_boost(struct kobject *a,
1118 				       struct kobj_attribute *b,
1119 				       const char *buf, size_t count)
1120 {
1121 	unsigned int input;
1122 	int ret;
1123 
1124 	ret = kstrtouint(buf, 10, &input);
1125 	if (ret)
1126 		return ret;
1127 
1128 	mutex_lock(&intel_pstate_driver_lock);
1129 	hwp_boost = !!input;
1130 	intel_pstate_update_policies();
1131 	mutex_unlock(&intel_pstate_driver_lock);
1132 
1133 	return count;
1134 }
1135 
1136 show_one(max_perf_pct, max_perf_pct);
1137 show_one(min_perf_pct, min_perf_pct);
1138 
1139 define_one_global_rw(status);
1140 define_one_global_rw(no_turbo);
1141 define_one_global_rw(max_perf_pct);
1142 define_one_global_rw(min_perf_pct);
1143 define_one_global_ro(turbo_pct);
1144 define_one_global_ro(num_pstates);
1145 define_one_global_rw(hwp_dynamic_boost);
1146 
1147 static struct attribute *intel_pstate_attributes[] = {
1148 	&status.attr,
1149 	&no_turbo.attr,
1150 	&turbo_pct.attr,
1151 	&num_pstates.attr,
1152 	NULL
1153 };
1154 
1155 static const struct attribute_group intel_pstate_attr_group = {
1156 	.attrs = intel_pstate_attributes,
1157 };
1158 
1159 static void __init intel_pstate_sysfs_expose_params(void)
1160 {
1161 	struct kobject *intel_pstate_kobject;
1162 	int rc;
1163 
1164 	intel_pstate_kobject = kobject_create_and_add("intel_pstate",
1165 						&cpu_subsys.dev_root->kobj);
1166 	if (WARN_ON(!intel_pstate_kobject))
1167 		return;
1168 
1169 	rc = sysfs_create_group(intel_pstate_kobject, &intel_pstate_attr_group);
1170 	if (WARN_ON(rc))
1171 		return;
1172 
1173 	/*
1174 	 * If per cpu limits are enforced there are no global limits, so
1175 	 * return without creating max/min_perf_pct attributes
1176 	 */
1177 	if (per_cpu_limits)
1178 		return;
1179 
1180 	rc = sysfs_create_file(intel_pstate_kobject, &max_perf_pct.attr);
1181 	WARN_ON(rc);
1182 
1183 	rc = sysfs_create_file(intel_pstate_kobject, &min_perf_pct.attr);
1184 	WARN_ON(rc);
1185 
1186 	if (hwp_active) {
1187 		rc = sysfs_create_file(intel_pstate_kobject,
1188 				       &hwp_dynamic_boost.attr);
1189 		WARN_ON(rc);
1190 	}
1191 }
1192 /************************** sysfs end ************************/
1193 
1194 static void intel_pstate_hwp_enable(struct cpudata *cpudata)
1195 {
1196 	/* First disable HWP notification interrupt as we don't process them */
1197 	if (static_cpu_has(X86_FEATURE_HWP_NOTIFY))
1198 		wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_INTERRUPT, 0x00);
1199 
1200 	wrmsrl_on_cpu(cpudata->cpu, MSR_PM_ENABLE, 0x1);
1201 	cpudata->epp_policy = 0;
1202 	if (cpudata->epp_default == -EINVAL)
1203 		cpudata->epp_default = intel_pstate_get_epp(cpudata, 0);
1204 }
1205 
1206 #define MSR_IA32_POWER_CTL_BIT_EE	19
1207 
1208 /* Disable energy efficiency optimization */
1209 static void intel_pstate_disable_ee(int cpu)
1210 {
1211 	u64 power_ctl;
1212 	int ret;
1213 
1214 	ret = rdmsrl_on_cpu(cpu, MSR_IA32_POWER_CTL, &power_ctl);
1215 	if (ret)
1216 		return;
1217 
1218 	if (!(power_ctl & BIT(MSR_IA32_POWER_CTL_BIT_EE))) {
1219 		pr_info("Disabling energy efficiency optimization\n");
1220 		power_ctl |= BIT(MSR_IA32_POWER_CTL_BIT_EE);
1221 		wrmsrl_on_cpu(cpu, MSR_IA32_POWER_CTL, power_ctl);
1222 	}
1223 }
1224 
1225 static int atom_get_min_pstate(void)
1226 {
1227 	u64 value;
1228 
1229 	rdmsrl(MSR_ATOM_CORE_RATIOS, value);
1230 	return (value >> 8) & 0x7F;
1231 }
1232 
1233 static int atom_get_max_pstate(void)
1234 {
1235 	u64 value;
1236 
1237 	rdmsrl(MSR_ATOM_CORE_RATIOS, value);
1238 	return (value >> 16) & 0x7F;
1239 }
1240 
1241 static int atom_get_turbo_pstate(void)
1242 {
1243 	u64 value;
1244 
1245 	rdmsrl(MSR_ATOM_CORE_TURBO_RATIOS, value);
1246 	return value & 0x7F;
1247 }
1248 
1249 static u64 atom_get_val(struct cpudata *cpudata, int pstate)
1250 {
1251 	u64 val;
1252 	int32_t vid_fp;
1253 	u32 vid;
1254 
1255 	val = (u64)pstate << 8;
1256 	if (global.no_turbo && !global.turbo_disabled)
1257 		val |= (u64)1 << 32;
1258 
1259 	vid_fp = cpudata->vid.min + mul_fp(
1260 		int_tofp(pstate - cpudata->pstate.min_pstate),
1261 		cpudata->vid.ratio);
1262 
1263 	vid_fp = clamp_t(int32_t, vid_fp, cpudata->vid.min, cpudata->vid.max);
1264 	vid = ceiling_fp(vid_fp);
1265 
1266 	if (pstate > cpudata->pstate.max_pstate)
1267 		vid = cpudata->vid.turbo;
1268 
1269 	return val | vid;
1270 }
1271 
1272 static int silvermont_get_scaling(void)
1273 {
1274 	u64 value;
1275 	int i;
1276 	/* Defined in Table 35-6 from SDM (Sept 2015) */
1277 	static int silvermont_freq_table[] = {
1278 		83300, 100000, 133300, 116700, 80000};
1279 
1280 	rdmsrl(MSR_FSB_FREQ, value);
1281 	i = value & 0x7;
1282 	WARN_ON(i > 4);
1283 
1284 	return silvermont_freq_table[i];
1285 }
1286 
1287 static int airmont_get_scaling(void)
1288 {
1289 	u64 value;
1290 	int i;
1291 	/* Defined in Table 35-10 from SDM (Sept 2015) */
1292 	static int airmont_freq_table[] = {
1293 		83300, 100000, 133300, 116700, 80000,
1294 		93300, 90000, 88900, 87500};
1295 
1296 	rdmsrl(MSR_FSB_FREQ, value);
1297 	i = value & 0xF;
1298 	WARN_ON(i > 8);
1299 
1300 	return airmont_freq_table[i];
1301 }
1302 
1303 static void atom_get_vid(struct cpudata *cpudata)
1304 {
1305 	u64 value;
1306 
1307 	rdmsrl(MSR_ATOM_CORE_VIDS, value);
1308 	cpudata->vid.min = int_tofp((value >> 8) & 0x7f);
1309 	cpudata->vid.max = int_tofp((value >> 16) & 0x7f);
1310 	cpudata->vid.ratio = div_fp(
1311 		cpudata->vid.max - cpudata->vid.min,
1312 		int_tofp(cpudata->pstate.max_pstate -
1313 			cpudata->pstate.min_pstate));
1314 
1315 	rdmsrl(MSR_ATOM_CORE_TURBO_VIDS, value);
1316 	cpudata->vid.turbo = value & 0x7f;
1317 }
1318 
1319 static int core_get_min_pstate(void)
1320 {
1321 	u64 value;
1322 
1323 	rdmsrl(MSR_PLATFORM_INFO, value);
1324 	return (value >> 40) & 0xFF;
1325 }
1326 
1327 static int core_get_max_pstate_physical(void)
1328 {
1329 	u64 value;
1330 
1331 	rdmsrl(MSR_PLATFORM_INFO, value);
1332 	return (value >> 8) & 0xFF;
1333 }
1334 
1335 static int core_get_tdp_ratio(u64 plat_info)
1336 {
1337 	/* Check how many TDP levels present */
1338 	if (plat_info & 0x600000000) {
1339 		u64 tdp_ctrl;
1340 		u64 tdp_ratio;
1341 		int tdp_msr;
1342 		int err;
1343 
1344 		/* Get the TDP level (0, 1, 2) to get ratios */
1345 		err = rdmsrl_safe(MSR_CONFIG_TDP_CONTROL, &tdp_ctrl);
1346 		if (err)
1347 			return err;
1348 
1349 		/* TDP MSR are continuous starting at 0x648 */
1350 		tdp_msr = MSR_CONFIG_TDP_NOMINAL + (tdp_ctrl & 0x03);
1351 		err = rdmsrl_safe(tdp_msr, &tdp_ratio);
1352 		if (err)
1353 			return err;
1354 
1355 		/* For level 1 and 2, bits[23:16] contain the ratio */
1356 		if (tdp_ctrl & 0x03)
1357 			tdp_ratio >>= 16;
1358 
1359 		tdp_ratio &= 0xff; /* ratios are only 8 bits long */
1360 		pr_debug("tdp_ratio %x\n", (int)tdp_ratio);
1361 
1362 		return (int)tdp_ratio;
1363 	}
1364 
1365 	return -ENXIO;
1366 }
1367 
1368 static int core_get_max_pstate(void)
1369 {
1370 	u64 tar;
1371 	u64 plat_info;
1372 	int max_pstate;
1373 	int tdp_ratio;
1374 	int err;
1375 
1376 	rdmsrl(MSR_PLATFORM_INFO, plat_info);
1377 	max_pstate = (plat_info >> 8) & 0xFF;
1378 
1379 	tdp_ratio = core_get_tdp_ratio(plat_info);
1380 	if (tdp_ratio <= 0)
1381 		return max_pstate;
1382 
1383 	if (hwp_active) {
1384 		/* Turbo activation ratio is not used on HWP platforms */
1385 		return tdp_ratio;
1386 	}
1387 
1388 	err = rdmsrl_safe(MSR_TURBO_ACTIVATION_RATIO, &tar);
1389 	if (!err) {
1390 		int tar_levels;
1391 
1392 		/* Do some sanity checking for safety */
1393 		tar_levels = tar & 0xff;
1394 		if (tdp_ratio - 1 == tar_levels) {
1395 			max_pstate = tar_levels;
1396 			pr_debug("max_pstate=TAC %x\n", max_pstate);
1397 		}
1398 	}
1399 
1400 	return max_pstate;
1401 }
1402 
1403 static int core_get_turbo_pstate(void)
1404 {
1405 	u64 value;
1406 	int nont, ret;
1407 
1408 	rdmsrl(MSR_TURBO_RATIO_LIMIT, value);
1409 	nont = core_get_max_pstate();
1410 	ret = (value) & 255;
1411 	if (ret <= nont)
1412 		ret = nont;
1413 	return ret;
1414 }
1415 
1416 static inline int core_get_scaling(void)
1417 {
1418 	return 100000;
1419 }
1420 
1421 static u64 core_get_val(struct cpudata *cpudata, int pstate)
1422 {
1423 	u64 val;
1424 
1425 	val = (u64)pstate << 8;
1426 	if (global.no_turbo && !global.turbo_disabled)
1427 		val |= (u64)1 << 32;
1428 
1429 	return val;
1430 }
1431 
1432 static int knl_get_aperf_mperf_shift(void)
1433 {
1434 	return 10;
1435 }
1436 
1437 static int knl_get_turbo_pstate(void)
1438 {
1439 	u64 value;
1440 	int nont, ret;
1441 
1442 	rdmsrl(MSR_TURBO_RATIO_LIMIT, value);
1443 	nont = core_get_max_pstate();
1444 	ret = (((value) >> 8) & 0xFF);
1445 	if (ret <= nont)
1446 		ret = nont;
1447 	return ret;
1448 }
1449 
1450 static void intel_pstate_set_pstate(struct cpudata *cpu, int pstate)
1451 {
1452 	trace_cpu_frequency(pstate * cpu->pstate.scaling, cpu->cpu);
1453 	cpu->pstate.current_pstate = pstate;
1454 	/*
1455 	 * Generally, there is no guarantee that this code will always run on
1456 	 * the CPU being updated, so force the register update to run on the
1457 	 * right CPU.
1458 	 */
1459 	wrmsrl_on_cpu(cpu->cpu, MSR_IA32_PERF_CTL,
1460 		      pstate_funcs.get_val(cpu, pstate));
1461 }
1462 
1463 static void intel_pstate_set_min_pstate(struct cpudata *cpu)
1464 {
1465 	intel_pstate_set_pstate(cpu, cpu->pstate.min_pstate);
1466 }
1467 
1468 static void intel_pstate_max_within_limits(struct cpudata *cpu)
1469 {
1470 	int pstate = max(cpu->pstate.min_pstate, cpu->max_perf_ratio);
1471 
1472 	update_turbo_state();
1473 	intel_pstate_set_pstate(cpu, pstate);
1474 }
1475 
1476 static void intel_pstate_get_cpu_pstates(struct cpudata *cpu)
1477 {
1478 	cpu->pstate.min_pstate = pstate_funcs.get_min();
1479 	cpu->pstate.max_pstate = pstate_funcs.get_max();
1480 	cpu->pstate.max_pstate_physical = pstate_funcs.get_max_physical();
1481 	cpu->pstate.turbo_pstate = pstate_funcs.get_turbo();
1482 	cpu->pstate.scaling = pstate_funcs.get_scaling();
1483 	cpu->pstate.max_freq = cpu->pstate.max_pstate * cpu->pstate.scaling;
1484 
1485 	if (hwp_active && !hwp_mode_bdw) {
1486 		unsigned int phy_max, current_max;
1487 
1488 		intel_pstate_get_hwp_max(cpu->cpu, &phy_max, &current_max);
1489 		cpu->pstate.turbo_freq = phy_max * cpu->pstate.scaling;
1490 	} else {
1491 		cpu->pstate.turbo_freq = cpu->pstate.turbo_pstate * cpu->pstate.scaling;
1492 	}
1493 
1494 	if (pstate_funcs.get_aperf_mperf_shift)
1495 		cpu->aperf_mperf_shift = pstate_funcs.get_aperf_mperf_shift();
1496 
1497 	if (pstate_funcs.get_vid)
1498 		pstate_funcs.get_vid(cpu);
1499 
1500 	intel_pstate_set_min_pstate(cpu);
1501 }
1502 
1503 /*
1504  * Long hold time will keep high perf limits for long time,
1505  * which negatively impacts perf/watt for some workloads,
1506  * like specpower. 3ms is based on experiements on some
1507  * workoads.
1508  */
1509 static int hwp_boost_hold_time_ns = 3 * NSEC_PER_MSEC;
1510 
1511 static inline void intel_pstate_hwp_boost_up(struct cpudata *cpu)
1512 {
1513 	u64 hwp_req = READ_ONCE(cpu->hwp_req_cached);
1514 	u32 max_limit = (hwp_req & 0xff00) >> 8;
1515 	u32 min_limit = (hwp_req & 0xff);
1516 	u32 boost_level1;
1517 
1518 	/*
1519 	 * Cases to consider (User changes via sysfs or boot time):
1520 	 * If, P0 (Turbo max) = P1 (Guaranteed max) = min:
1521 	 *	No boost, return.
1522 	 * If, P0 (Turbo max) > P1 (Guaranteed max) = min:
1523 	 *     Should result in one level boost only for P0.
1524 	 * If, P0 (Turbo max) = P1 (Guaranteed max) > min:
1525 	 *     Should result in two level boost:
1526 	 *         (min + p1)/2 and P1.
1527 	 * If, P0 (Turbo max) > P1 (Guaranteed max) > min:
1528 	 *     Should result in three level boost:
1529 	 *        (min + p1)/2, P1 and P0.
1530 	 */
1531 
1532 	/* If max and min are equal or already at max, nothing to boost */
1533 	if (max_limit == min_limit || cpu->hwp_boost_min >= max_limit)
1534 		return;
1535 
1536 	if (!cpu->hwp_boost_min)
1537 		cpu->hwp_boost_min = min_limit;
1538 
1539 	/* level at half way mark between min and guranteed */
1540 	boost_level1 = (HWP_GUARANTEED_PERF(cpu->hwp_cap_cached) + min_limit) >> 1;
1541 
1542 	if (cpu->hwp_boost_min < boost_level1)
1543 		cpu->hwp_boost_min = boost_level1;
1544 	else if (cpu->hwp_boost_min < HWP_GUARANTEED_PERF(cpu->hwp_cap_cached))
1545 		cpu->hwp_boost_min = HWP_GUARANTEED_PERF(cpu->hwp_cap_cached);
1546 	else if (cpu->hwp_boost_min == HWP_GUARANTEED_PERF(cpu->hwp_cap_cached) &&
1547 		 max_limit != HWP_GUARANTEED_PERF(cpu->hwp_cap_cached))
1548 		cpu->hwp_boost_min = max_limit;
1549 	else
1550 		return;
1551 
1552 	hwp_req = (hwp_req & ~GENMASK_ULL(7, 0)) | cpu->hwp_boost_min;
1553 	wrmsrl(MSR_HWP_REQUEST, hwp_req);
1554 	cpu->last_update = cpu->sample.time;
1555 }
1556 
1557 static inline void intel_pstate_hwp_boost_down(struct cpudata *cpu)
1558 {
1559 	if (cpu->hwp_boost_min) {
1560 		bool expired;
1561 
1562 		/* Check if we are idle for hold time to boost down */
1563 		expired = time_after64(cpu->sample.time, cpu->last_update +
1564 				       hwp_boost_hold_time_ns);
1565 		if (expired) {
1566 			wrmsrl(MSR_HWP_REQUEST, cpu->hwp_req_cached);
1567 			cpu->hwp_boost_min = 0;
1568 		}
1569 	}
1570 	cpu->last_update = cpu->sample.time;
1571 }
1572 
1573 static inline void intel_pstate_update_util_hwp_local(struct cpudata *cpu,
1574 						      u64 time)
1575 {
1576 	cpu->sample.time = time;
1577 
1578 	if (cpu->sched_flags & SCHED_CPUFREQ_IOWAIT) {
1579 		bool do_io = false;
1580 
1581 		cpu->sched_flags = 0;
1582 		/*
1583 		 * Set iowait_boost flag and update time. Since IO WAIT flag
1584 		 * is set all the time, we can't just conclude that there is
1585 		 * some IO bound activity is scheduled on this CPU with just
1586 		 * one occurrence. If we receive at least two in two
1587 		 * consecutive ticks, then we treat as boost candidate.
1588 		 */
1589 		if (time_before64(time, cpu->last_io_update + 2 * TICK_NSEC))
1590 			do_io = true;
1591 
1592 		cpu->last_io_update = time;
1593 
1594 		if (do_io)
1595 			intel_pstate_hwp_boost_up(cpu);
1596 
1597 	} else {
1598 		intel_pstate_hwp_boost_down(cpu);
1599 	}
1600 }
1601 
1602 static inline void intel_pstate_update_util_hwp(struct update_util_data *data,
1603 						u64 time, unsigned int flags)
1604 {
1605 	struct cpudata *cpu = container_of(data, struct cpudata, update_util);
1606 
1607 	cpu->sched_flags |= flags;
1608 
1609 	if (smp_processor_id() == cpu->cpu)
1610 		intel_pstate_update_util_hwp_local(cpu, time);
1611 }
1612 
1613 static inline void intel_pstate_calc_avg_perf(struct cpudata *cpu)
1614 {
1615 	struct sample *sample = &cpu->sample;
1616 
1617 	sample->core_avg_perf = div_ext_fp(sample->aperf, sample->mperf);
1618 }
1619 
1620 static inline bool intel_pstate_sample(struct cpudata *cpu, u64 time)
1621 {
1622 	u64 aperf, mperf;
1623 	unsigned long flags;
1624 	u64 tsc;
1625 
1626 	local_irq_save(flags);
1627 	rdmsrl(MSR_IA32_APERF, aperf);
1628 	rdmsrl(MSR_IA32_MPERF, mperf);
1629 	tsc = rdtsc();
1630 	if (cpu->prev_mperf == mperf || cpu->prev_tsc == tsc) {
1631 		local_irq_restore(flags);
1632 		return false;
1633 	}
1634 	local_irq_restore(flags);
1635 
1636 	cpu->last_sample_time = cpu->sample.time;
1637 	cpu->sample.time = time;
1638 	cpu->sample.aperf = aperf;
1639 	cpu->sample.mperf = mperf;
1640 	cpu->sample.tsc =  tsc;
1641 	cpu->sample.aperf -= cpu->prev_aperf;
1642 	cpu->sample.mperf -= cpu->prev_mperf;
1643 	cpu->sample.tsc -= cpu->prev_tsc;
1644 
1645 	cpu->prev_aperf = aperf;
1646 	cpu->prev_mperf = mperf;
1647 	cpu->prev_tsc = tsc;
1648 	/*
1649 	 * First time this function is invoked in a given cycle, all of the
1650 	 * previous sample data fields are equal to zero or stale and they must
1651 	 * be populated with meaningful numbers for things to work, so assume
1652 	 * that sample.time will always be reset before setting the utilization
1653 	 * update hook and make the caller skip the sample then.
1654 	 */
1655 	if (cpu->last_sample_time) {
1656 		intel_pstate_calc_avg_perf(cpu);
1657 		return true;
1658 	}
1659 	return false;
1660 }
1661 
1662 static inline int32_t get_avg_frequency(struct cpudata *cpu)
1663 {
1664 	return mul_ext_fp(cpu->sample.core_avg_perf, cpu_khz);
1665 }
1666 
1667 static inline int32_t get_avg_pstate(struct cpudata *cpu)
1668 {
1669 	return mul_ext_fp(cpu->pstate.max_pstate_physical,
1670 			  cpu->sample.core_avg_perf);
1671 }
1672 
1673 static inline int32_t get_target_pstate(struct cpudata *cpu)
1674 {
1675 	struct sample *sample = &cpu->sample;
1676 	int32_t busy_frac;
1677 	int target, avg_pstate;
1678 
1679 	busy_frac = div_fp(sample->mperf << cpu->aperf_mperf_shift,
1680 			   sample->tsc);
1681 
1682 	if (busy_frac < cpu->iowait_boost)
1683 		busy_frac = cpu->iowait_boost;
1684 
1685 	sample->busy_scaled = busy_frac * 100;
1686 
1687 	target = global.no_turbo || global.turbo_disabled ?
1688 			cpu->pstate.max_pstate : cpu->pstate.turbo_pstate;
1689 	target += target >> 2;
1690 	target = mul_fp(target, busy_frac);
1691 	if (target < cpu->pstate.min_pstate)
1692 		target = cpu->pstate.min_pstate;
1693 
1694 	/*
1695 	 * If the average P-state during the previous cycle was higher than the
1696 	 * current target, add 50% of the difference to the target to reduce
1697 	 * possible performance oscillations and offset possible performance
1698 	 * loss related to moving the workload from one CPU to another within
1699 	 * a package/module.
1700 	 */
1701 	avg_pstate = get_avg_pstate(cpu);
1702 	if (avg_pstate > target)
1703 		target += (avg_pstate - target) >> 1;
1704 
1705 	return target;
1706 }
1707 
1708 static int intel_pstate_prepare_request(struct cpudata *cpu, int pstate)
1709 {
1710 	int min_pstate = max(cpu->pstate.min_pstate, cpu->min_perf_ratio);
1711 	int max_pstate = max(min_pstate, cpu->max_perf_ratio);
1712 
1713 	return clamp_t(int, pstate, min_pstate, max_pstate);
1714 }
1715 
1716 static void intel_pstate_update_pstate(struct cpudata *cpu, int pstate)
1717 {
1718 	if (pstate == cpu->pstate.current_pstate)
1719 		return;
1720 
1721 	cpu->pstate.current_pstate = pstate;
1722 	wrmsrl(MSR_IA32_PERF_CTL, pstate_funcs.get_val(cpu, pstate));
1723 }
1724 
1725 static void intel_pstate_adjust_pstate(struct cpudata *cpu)
1726 {
1727 	int from = cpu->pstate.current_pstate;
1728 	struct sample *sample;
1729 	int target_pstate;
1730 
1731 	update_turbo_state();
1732 
1733 	target_pstate = get_target_pstate(cpu);
1734 	target_pstate = intel_pstate_prepare_request(cpu, target_pstate);
1735 	trace_cpu_frequency(target_pstate * cpu->pstate.scaling, cpu->cpu);
1736 	intel_pstate_update_pstate(cpu, target_pstate);
1737 
1738 	sample = &cpu->sample;
1739 	trace_pstate_sample(mul_ext_fp(100, sample->core_avg_perf),
1740 		fp_toint(sample->busy_scaled),
1741 		from,
1742 		cpu->pstate.current_pstate,
1743 		sample->mperf,
1744 		sample->aperf,
1745 		sample->tsc,
1746 		get_avg_frequency(cpu),
1747 		fp_toint(cpu->iowait_boost * 100));
1748 }
1749 
1750 static void intel_pstate_update_util(struct update_util_data *data, u64 time,
1751 				     unsigned int flags)
1752 {
1753 	struct cpudata *cpu = container_of(data, struct cpudata, update_util);
1754 	u64 delta_ns;
1755 
1756 	/* Don't allow remote callbacks */
1757 	if (smp_processor_id() != cpu->cpu)
1758 		return;
1759 
1760 	delta_ns = time - cpu->last_update;
1761 	if (flags & SCHED_CPUFREQ_IOWAIT) {
1762 		/* Start over if the CPU may have been idle. */
1763 		if (delta_ns > TICK_NSEC) {
1764 			cpu->iowait_boost = ONE_EIGHTH_FP;
1765 		} else if (cpu->iowait_boost) {
1766 			cpu->iowait_boost <<= 1;
1767 			if (cpu->iowait_boost > int_tofp(1))
1768 				cpu->iowait_boost = int_tofp(1);
1769 		} else {
1770 			cpu->iowait_boost = ONE_EIGHTH_FP;
1771 		}
1772 	} else if (cpu->iowait_boost) {
1773 		/* Clear iowait_boost if the CPU may have been idle. */
1774 		if (delta_ns > TICK_NSEC)
1775 			cpu->iowait_boost = 0;
1776 		else
1777 			cpu->iowait_boost >>= 1;
1778 	}
1779 	cpu->last_update = time;
1780 	delta_ns = time - cpu->sample.time;
1781 	if ((s64)delta_ns < INTEL_PSTATE_SAMPLING_INTERVAL)
1782 		return;
1783 
1784 	if (intel_pstate_sample(cpu, time))
1785 		intel_pstate_adjust_pstate(cpu);
1786 }
1787 
1788 static struct pstate_funcs core_funcs = {
1789 	.get_max = core_get_max_pstate,
1790 	.get_max_physical = core_get_max_pstate_physical,
1791 	.get_min = core_get_min_pstate,
1792 	.get_turbo = core_get_turbo_pstate,
1793 	.get_scaling = core_get_scaling,
1794 	.get_val = core_get_val,
1795 };
1796 
1797 static const struct pstate_funcs silvermont_funcs = {
1798 	.get_max = atom_get_max_pstate,
1799 	.get_max_physical = atom_get_max_pstate,
1800 	.get_min = atom_get_min_pstate,
1801 	.get_turbo = atom_get_turbo_pstate,
1802 	.get_val = atom_get_val,
1803 	.get_scaling = silvermont_get_scaling,
1804 	.get_vid = atom_get_vid,
1805 };
1806 
1807 static const struct pstate_funcs airmont_funcs = {
1808 	.get_max = atom_get_max_pstate,
1809 	.get_max_physical = atom_get_max_pstate,
1810 	.get_min = atom_get_min_pstate,
1811 	.get_turbo = atom_get_turbo_pstate,
1812 	.get_val = atom_get_val,
1813 	.get_scaling = airmont_get_scaling,
1814 	.get_vid = atom_get_vid,
1815 };
1816 
1817 static const struct pstate_funcs knl_funcs = {
1818 	.get_max = core_get_max_pstate,
1819 	.get_max_physical = core_get_max_pstate_physical,
1820 	.get_min = core_get_min_pstate,
1821 	.get_turbo = knl_get_turbo_pstate,
1822 	.get_aperf_mperf_shift = knl_get_aperf_mperf_shift,
1823 	.get_scaling = core_get_scaling,
1824 	.get_val = core_get_val,
1825 };
1826 
1827 #define ICPU(model, policy) \
1828 	{ X86_VENDOR_INTEL, 6, model, X86_FEATURE_APERFMPERF,\
1829 			(unsigned long)&policy }
1830 
1831 static const struct x86_cpu_id intel_pstate_cpu_ids[] = {
1832 	ICPU(INTEL_FAM6_SANDYBRIDGE, 		core_funcs),
1833 	ICPU(INTEL_FAM6_SANDYBRIDGE_X,		core_funcs),
1834 	ICPU(INTEL_FAM6_ATOM_SILVERMONT,	silvermont_funcs),
1835 	ICPU(INTEL_FAM6_IVYBRIDGE,		core_funcs),
1836 	ICPU(INTEL_FAM6_HASWELL_CORE,		core_funcs),
1837 	ICPU(INTEL_FAM6_BROADWELL_CORE,		core_funcs),
1838 	ICPU(INTEL_FAM6_IVYBRIDGE_X,		core_funcs),
1839 	ICPU(INTEL_FAM6_HASWELL_X,		core_funcs),
1840 	ICPU(INTEL_FAM6_HASWELL_ULT,		core_funcs),
1841 	ICPU(INTEL_FAM6_HASWELL_GT3E,		core_funcs),
1842 	ICPU(INTEL_FAM6_BROADWELL_GT3E,		core_funcs),
1843 	ICPU(INTEL_FAM6_ATOM_AIRMONT,		airmont_funcs),
1844 	ICPU(INTEL_FAM6_SKYLAKE_MOBILE,		core_funcs),
1845 	ICPU(INTEL_FAM6_BROADWELL_X,		core_funcs),
1846 	ICPU(INTEL_FAM6_SKYLAKE_DESKTOP,	core_funcs),
1847 	ICPU(INTEL_FAM6_BROADWELL_XEON_D,	core_funcs),
1848 	ICPU(INTEL_FAM6_XEON_PHI_KNL,		knl_funcs),
1849 	ICPU(INTEL_FAM6_XEON_PHI_KNM,		knl_funcs),
1850 	ICPU(INTEL_FAM6_ATOM_GOLDMONT,		core_funcs),
1851 	ICPU(INTEL_FAM6_ATOM_GOLDMONT_PLUS,     core_funcs),
1852 	ICPU(INTEL_FAM6_SKYLAKE_X,		core_funcs),
1853 	{}
1854 };
1855 MODULE_DEVICE_TABLE(x86cpu, intel_pstate_cpu_ids);
1856 
1857 static const struct x86_cpu_id intel_pstate_cpu_oob_ids[] __initconst = {
1858 	ICPU(INTEL_FAM6_BROADWELL_XEON_D, core_funcs),
1859 	ICPU(INTEL_FAM6_BROADWELL_X, core_funcs),
1860 	ICPU(INTEL_FAM6_SKYLAKE_X, core_funcs),
1861 	{}
1862 };
1863 
1864 static const struct x86_cpu_id intel_pstate_cpu_ee_disable_ids[] = {
1865 	ICPU(INTEL_FAM6_KABYLAKE_DESKTOP, core_funcs),
1866 	{}
1867 };
1868 
1869 static const struct x86_cpu_id intel_pstate_hwp_boost_ids[] = {
1870 	ICPU(INTEL_FAM6_SKYLAKE_X, core_funcs),
1871 	ICPU(INTEL_FAM6_SKYLAKE_DESKTOP, core_funcs),
1872 	{}
1873 };
1874 
1875 static int intel_pstate_init_cpu(unsigned int cpunum)
1876 {
1877 	struct cpudata *cpu;
1878 
1879 	cpu = all_cpu_data[cpunum];
1880 
1881 	if (!cpu) {
1882 		cpu = kzalloc(sizeof(*cpu), GFP_KERNEL);
1883 		if (!cpu)
1884 			return -ENOMEM;
1885 
1886 		all_cpu_data[cpunum] = cpu;
1887 
1888 		cpu->epp_default = -EINVAL;
1889 		cpu->epp_powersave = -EINVAL;
1890 		cpu->epp_saved = -EINVAL;
1891 	}
1892 
1893 	cpu = all_cpu_data[cpunum];
1894 
1895 	cpu->cpu = cpunum;
1896 
1897 	if (hwp_active) {
1898 		const struct x86_cpu_id *id;
1899 
1900 		id = x86_match_cpu(intel_pstate_cpu_ee_disable_ids);
1901 		if (id)
1902 			intel_pstate_disable_ee(cpunum);
1903 
1904 		intel_pstate_hwp_enable(cpu);
1905 
1906 		id = x86_match_cpu(intel_pstate_hwp_boost_ids);
1907 		if (id && intel_pstate_acpi_pm_profile_server())
1908 			hwp_boost = true;
1909 	}
1910 
1911 	intel_pstate_get_cpu_pstates(cpu);
1912 
1913 	pr_debug("controlling: cpu %d\n", cpunum);
1914 
1915 	return 0;
1916 }
1917 
1918 static void intel_pstate_set_update_util_hook(unsigned int cpu_num)
1919 {
1920 	struct cpudata *cpu = all_cpu_data[cpu_num];
1921 
1922 	if (hwp_active && !hwp_boost)
1923 		return;
1924 
1925 	if (cpu->update_util_set)
1926 		return;
1927 
1928 	/* Prevent intel_pstate_update_util() from using stale data. */
1929 	cpu->sample.time = 0;
1930 	cpufreq_add_update_util_hook(cpu_num, &cpu->update_util,
1931 				     (hwp_active ?
1932 				      intel_pstate_update_util_hwp :
1933 				      intel_pstate_update_util));
1934 	cpu->update_util_set = true;
1935 }
1936 
1937 static void intel_pstate_clear_update_util_hook(unsigned int cpu)
1938 {
1939 	struct cpudata *cpu_data = all_cpu_data[cpu];
1940 
1941 	if (!cpu_data->update_util_set)
1942 		return;
1943 
1944 	cpufreq_remove_update_util_hook(cpu);
1945 	cpu_data->update_util_set = false;
1946 	synchronize_rcu();
1947 }
1948 
1949 static int intel_pstate_get_max_freq(struct cpudata *cpu)
1950 {
1951 	return global.turbo_disabled || global.no_turbo ?
1952 			cpu->pstate.max_freq : cpu->pstate.turbo_freq;
1953 }
1954 
1955 static void intel_pstate_update_perf_limits(struct cpufreq_policy *policy,
1956 					    struct cpudata *cpu)
1957 {
1958 	int max_freq = intel_pstate_get_max_freq(cpu);
1959 	int32_t max_policy_perf, min_policy_perf;
1960 	int max_state, turbo_max;
1961 
1962 	/*
1963 	 * HWP needs some special consideration, because on BDX the
1964 	 * HWP_REQUEST uses abstract value to represent performance
1965 	 * rather than pure ratios.
1966 	 */
1967 	if (hwp_active) {
1968 		intel_pstate_get_hwp_max(cpu->cpu, &turbo_max, &max_state);
1969 	} else {
1970 		max_state = global.no_turbo || global.turbo_disabled ?
1971 			cpu->pstate.max_pstate : cpu->pstate.turbo_pstate;
1972 		turbo_max = cpu->pstate.turbo_pstate;
1973 	}
1974 
1975 	max_policy_perf = max_state * policy->max / max_freq;
1976 	if (policy->max == policy->min) {
1977 		min_policy_perf = max_policy_perf;
1978 	} else {
1979 		min_policy_perf = max_state * policy->min / max_freq;
1980 		min_policy_perf = clamp_t(int32_t, min_policy_perf,
1981 					  0, max_policy_perf);
1982 	}
1983 
1984 	pr_debug("cpu:%d max_state %d min_policy_perf:%d max_policy_perf:%d\n",
1985 		 policy->cpu, max_state,
1986 		 min_policy_perf, max_policy_perf);
1987 
1988 	/* Normalize user input to [min_perf, max_perf] */
1989 	if (per_cpu_limits) {
1990 		cpu->min_perf_ratio = min_policy_perf;
1991 		cpu->max_perf_ratio = max_policy_perf;
1992 	} else {
1993 		int32_t global_min, global_max;
1994 
1995 		/* Global limits are in percent of the maximum turbo P-state. */
1996 		global_max = DIV_ROUND_UP(turbo_max * global.max_perf_pct, 100);
1997 		global_min = DIV_ROUND_UP(turbo_max * global.min_perf_pct, 100);
1998 		global_min = clamp_t(int32_t, global_min, 0, global_max);
1999 
2000 		pr_debug("cpu:%d global_min:%d global_max:%d\n", policy->cpu,
2001 			 global_min, global_max);
2002 
2003 		cpu->min_perf_ratio = max(min_policy_perf, global_min);
2004 		cpu->min_perf_ratio = min(cpu->min_perf_ratio, max_policy_perf);
2005 		cpu->max_perf_ratio = min(max_policy_perf, global_max);
2006 		cpu->max_perf_ratio = max(min_policy_perf, cpu->max_perf_ratio);
2007 
2008 		/* Make sure min_perf <= max_perf */
2009 		cpu->min_perf_ratio = min(cpu->min_perf_ratio,
2010 					  cpu->max_perf_ratio);
2011 
2012 	}
2013 	pr_debug("cpu:%d max_perf_ratio:%d min_perf_ratio:%d\n", policy->cpu,
2014 		 cpu->max_perf_ratio,
2015 		 cpu->min_perf_ratio);
2016 }
2017 
2018 static int intel_pstate_set_policy(struct cpufreq_policy *policy)
2019 {
2020 	struct cpudata *cpu;
2021 
2022 	if (!policy->cpuinfo.max_freq)
2023 		return -ENODEV;
2024 
2025 	pr_debug("set_policy cpuinfo.max %u policy->max %u\n",
2026 		 policy->cpuinfo.max_freq, policy->max);
2027 
2028 	cpu = all_cpu_data[policy->cpu];
2029 	cpu->policy = policy->policy;
2030 
2031 	mutex_lock(&intel_pstate_limits_lock);
2032 
2033 	intel_pstate_update_perf_limits(policy, cpu);
2034 
2035 	if (cpu->policy == CPUFREQ_POLICY_PERFORMANCE) {
2036 		/*
2037 		 * NOHZ_FULL CPUs need this as the governor callback may not
2038 		 * be invoked on them.
2039 		 */
2040 		intel_pstate_clear_update_util_hook(policy->cpu);
2041 		intel_pstate_max_within_limits(cpu);
2042 	} else {
2043 		intel_pstate_set_update_util_hook(policy->cpu);
2044 	}
2045 
2046 	if (hwp_active) {
2047 		/*
2048 		 * When hwp_boost was active before and dynamically it
2049 		 * was turned off, in that case we need to clear the
2050 		 * update util hook.
2051 		 */
2052 		if (!hwp_boost)
2053 			intel_pstate_clear_update_util_hook(policy->cpu);
2054 		intel_pstate_hwp_set(policy->cpu);
2055 	}
2056 
2057 	mutex_unlock(&intel_pstate_limits_lock);
2058 
2059 	return 0;
2060 }
2061 
2062 static void intel_pstate_adjust_policy_max(struct cpufreq_policy *policy,
2063 					 struct cpudata *cpu)
2064 {
2065 	if (!hwp_active &&
2066 	    cpu->pstate.max_pstate_physical > cpu->pstate.max_pstate &&
2067 	    policy->max < policy->cpuinfo.max_freq &&
2068 	    policy->max > cpu->pstate.max_freq) {
2069 		pr_debug("policy->max > max non turbo frequency\n");
2070 		policy->max = policy->cpuinfo.max_freq;
2071 	}
2072 }
2073 
2074 static int intel_pstate_verify_policy(struct cpufreq_policy *policy)
2075 {
2076 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2077 
2078 	update_turbo_state();
2079 	cpufreq_verify_within_limits(policy, policy->cpuinfo.min_freq,
2080 				     intel_pstate_get_max_freq(cpu));
2081 
2082 	if (policy->policy != CPUFREQ_POLICY_POWERSAVE &&
2083 	    policy->policy != CPUFREQ_POLICY_PERFORMANCE)
2084 		return -EINVAL;
2085 
2086 	intel_pstate_adjust_policy_max(policy, cpu);
2087 
2088 	return 0;
2089 }
2090 
2091 static void intel_cpufreq_stop_cpu(struct cpufreq_policy *policy)
2092 {
2093 	intel_pstate_set_min_pstate(all_cpu_data[policy->cpu]);
2094 }
2095 
2096 static void intel_pstate_stop_cpu(struct cpufreq_policy *policy)
2097 {
2098 	pr_debug("CPU %d exiting\n", policy->cpu);
2099 
2100 	intel_pstate_clear_update_util_hook(policy->cpu);
2101 	if (hwp_active) {
2102 		intel_pstate_hwp_save_state(policy);
2103 		intel_pstate_hwp_force_min_perf(policy->cpu);
2104 	} else {
2105 		intel_cpufreq_stop_cpu(policy);
2106 	}
2107 }
2108 
2109 static int intel_pstate_cpu_exit(struct cpufreq_policy *policy)
2110 {
2111 	intel_pstate_exit_perf_limits(policy);
2112 
2113 	policy->fast_switch_possible = false;
2114 
2115 	return 0;
2116 }
2117 
2118 static int __intel_pstate_cpu_init(struct cpufreq_policy *policy)
2119 {
2120 	struct cpudata *cpu;
2121 	int rc;
2122 
2123 	rc = intel_pstate_init_cpu(policy->cpu);
2124 	if (rc)
2125 		return rc;
2126 
2127 	cpu = all_cpu_data[policy->cpu];
2128 
2129 	cpu->max_perf_ratio = 0xFF;
2130 	cpu->min_perf_ratio = 0;
2131 
2132 	policy->min = cpu->pstate.min_pstate * cpu->pstate.scaling;
2133 	policy->max = cpu->pstate.turbo_pstate * cpu->pstate.scaling;
2134 
2135 	/* cpuinfo and default policy values */
2136 	policy->cpuinfo.min_freq = cpu->pstate.min_pstate * cpu->pstate.scaling;
2137 	update_turbo_state();
2138 	policy->cpuinfo.max_freq = global.turbo_disabled ?
2139 			cpu->pstate.max_pstate : cpu->pstate.turbo_pstate;
2140 	policy->cpuinfo.max_freq *= cpu->pstate.scaling;
2141 
2142 	if (hwp_active) {
2143 		unsigned int max_freq;
2144 
2145 		max_freq = global.turbo_disabled ?
2146 			cpu->pstate.max_freq : cpu->pstate.turbo_freq;
2147 		if (max_freq < policy->cpuinfo.max_freq)
2148 			policy->cpuinfo.max_freq = max_freq;
2149 	}
2150 
2151 	intel_pstate_init_acpi_perf_limits(policy);
2152 
2153 	policy->fast_switch_possible = true;
2154 
2155 	return 0;
2156 }
2157 
2158 static int intel_pstate_cpu_init(struct cpufreq_policy *policy)
2159 {
2160 	int ret = __intel_pstate_cpu_init(policy);
2161 
2162 	if (ret)
2163 		return ret;
2164 
2165 	if (IS_ENABLED(CONFIG_CPU_FREQ_DEFAULT_GOV_PERFORMANCE))
2166 		policy->policy = CPUFREQ_POLICY_PERFORMANCE;
2167 	else
2168 		policy->policy = CPUFREQ_POLICY_POWERSAVE;
2169 
2170 	return 0;
2171 }
2172 
2173 static struct cpufreq_driver intel_pstate = {
2174 	.flags		= CPUFREQ_CONST_LOOPS,
2175 	.verify		= intel_pstate_verify_policy,
2176 	.setpolicy	= intel_pstate_set_policy,
2177 	.suspend	= intel_pstate_hwp_save_state,
2178 	.resume		= intel_pstate_resume,
2179 	.init		= intel_pstate_cpu_init,
2180 	.exit		= intel_pstate_cpu_exit,
2181 	.stop_cpu	= intel_pstate_stop_cpu,
2182 	.name		= "intel_pstate",
2183 };
2184 
2185 static int intel_cpufreq_verify_policy(struct cpufreq_policy *policy)
2186 {
2187 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2188 
2189 	update_turbo_state();
2190 	cpufreq_verify_within_limits(policy, policy->cpuinfo.min_freq,
2191 				     intel_pstate_get_max_freq(cpu));
2192 
2193 	intel_pstate_adjust_policy_max(policy, cpu);
2194 
2195 	intel_pstate_update_perf_limits(policy, cpu);
2196 
2197 	return 0;
2198 }
2199 
2200 /* Use of trace in passive mode:
2201  *
2202  * In passive mode the trace core_busy field (also known as the
2203  * performance field, and lablelled as such on the graphs; also known as
2204  * core_avg_perf) is not needed and so is re-assigned to indicate if the
2205  * driver call was via the normal or fast switch path. Various graphs
2206  * output from the intel_pstate_tracer.py utility that include core_busy
2207  * (or performance or core_avg_perf) have a fixed y-axis from 0 to 100%,
2208  * so we use 10 to indicate the the normal path through the driver, and
2209  * 90 to indicate the fast switch path through the driver.
2210  * The scaled_busy field is not used, and is set to 0.
2211  */
2212 
2213 #define	INTEL_PSTATE_TRACE_TARGET 10
2214 #define	INTEL_PSTATE_TRACE_FAST_SWITCH 90
2215 
2216 static void intel_cpufreq_trace(struct cpudata *cpu, unsigned int trace_type, int old_pstate)
2217 {
2218 	struct sample *sample;
2219 
2220 	if (!trace_pstate_sample_enabled())
2221 		return;
2222 
2223 	if (!intel_pstate_sample(cpu, ktime_get()))
2224 		return;
2225 
2226 	sample = &cpu->sample;
2227 	trace_pstate_sample(trace_type,
2228 		0,
2229 		old_pstate,
2230 		cpu->pstate.current_pstate,
2231 		sample->mperf,
2232 		sample->aperf,
2233 		sample->tsc,
2234 		get_avg_frequency(cpu),
2235 		fp_toint(cpu->iowait_boost * 100));
2236 }
2237 
2238 static int intel_cpufreq_target(struct cpufreq_policy *policy,
2239 				unsigned int target_freq,
2240 				unsigned int relation)
2241 {
2242 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2243 	struct cpufreq_freqs freqs;
2244 	int target_pstate, old_pstate;
2245 
2246 	update_turbo_state();
2247 
2248 	freqs.old = policy->cur;
2249 	freqs.new = target_freq;
2250 
2251 	cpufreq_freq_transition_begin(policy, &freqs);
2252 	switch (relation) {
2253 	case CPUFREQ_RELATION_L:
2254 		target_pstate = DIV_ROUND_UP(freqs.new, cpu->pstate.scaling);
2255 		break;
2256 	case CPUFREQ_RELATION_H:
2257 		target_pstate = freqs.new / cpu->pstate.scaling;
2258 		break;
2259 	default:
2260 		target_pstate = DIV_ROUND_CLOSEST(freqs.new, cpu->pstate.scaling);
2261 		break;
2262 	}
2263 	target_pstate = intel_pstate_prepare_request(cpu, target_pstate);
2264 	old_pstate = cpu->pstate.current_pstate;
2265 	if (target_pstate != cpu->pstate.current_pstate) {
2266 		cpu->pstate.current_pstate = target_pstate;
2267 		wrmsrl_on_cpu(policy->cpu, MSR_IA32_PERF_CTL,
2268 			      pstate_funcs.get_val(cpu, target_pstate));
2269 	}
2270 	freqs.new = target_pstate * cpu->pstate.scaling;
2271 	intel_cpufreq_trace(cpu, INTEL_PSTATE_TRACE_TARGET, old_pstate);
2272 	cpufreq_freq_transition_end(policy, &freqs, false);
2273 
2274 	return 0;
2275 }
2276 
2277 static unsigned int intel_cpufreq_fast_switch(struct cpufreq_policy *policy,
2278 					      unsigned int target_freq)
2279 {
2280 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2281 	int target_pstate, old_pstate;
2282 
2283 	update_turbo_state();
2284 
2285 	target_pstate = DIV_ROUND_UP(target_freq, cpu->pstate.scaling);
2286 	target_pstate = intel_pstate_prepare_request(cpu, target_pstate);
2287 	old_pstate = cpu->pstate.current_pstate;
2288 	intel_pstate_update_pstate(cpu, target_pstate);
2289 	intel_cpufreq_trace(cpu, INTEL_PSTATE_TRACE_FAST_SWITCH, old_pstate);
2290 	return target_pstate * cpu->pstate.scaling;
2291 }
2292 
2293 static int intel_cpufreq_cpu_init(struct cpufreq_policy *policy)
2294 {
2295 	int ret = __intel_pstate_cpu_init(policy);
2296 
2297 	if (ret)
2298 		return ret;
2299 
2300 	policy->cpuinfo.transition_latency = INTEL_CPUFREQ_TRANSITION_LATENCY;
2301 	policy->transition_delay_us = INTEL_CPUFREQ_TRANSITION_DELAY;
2302 	/* This reflects the intel_pstate_get_cpu_pstates() setting. */
2303 	policy->cur = policy->cpuinfo.min_freq;
2304 
2305 	return 0;
2306 }
2307 
2308 static struct cpufreq_driver intel_cpufreq = {
2309 	.flags		= CPUFREQ_CONST_LOOPS,
2310 	.verify		= intel_cpufreq_verify_policy,
2311 	.target		= intel_cpufreq_target,
2312 	.fast_switch	= intel_cpufreq_fast_switch,
2313 	.init		= intel_cpufreq_cpu_init,
2314 	.exit		= intel_pstate_cpu_exit,
2315 	.stop_cpu	= intel_cpufreq_stop_cpu,
2316 	.name		= "intel_cpufreq",
2317 };
2318 
2319 static struct cpufreq_driver *default_driver = &intel_pstate;
2320 
2321 static void intel_pstate_driver_cleanup(void)
2322 {
2323 	unsigned int cpu;
2324 
2325 	get_online_cpus();
2326 	for_each_online_cpu(cpu) {
2327 		if (all_cpu_data[cpu]) {
2328 			if (intel_pstate_driver == &intel_pstate)
2329 				intel_pstate_clear_update_util_hook(cpu);
2330 
2331 			kfree(all_cpu_data[cpu]);
2332 			all_cpu_data[cpu] = NULL;
2333 		}
2334 	}
2335 	put_online_cpus();
2336 	intel_pstate_driver = NULL;
2337 }
2338 
2339 static int intel_pstate_register_driver(struct cpufreq_driver *driver)
2340 {
2341 	int ret;
2342 
2343 	memset(&global, 0, sizeof(global));
2344 	global.max_perf_pct = 100;
2345 
2346 	intel_pstate_driver = driver;
2347 	ret = cpufreq_register_driver(intel_pstate_driver);
2348 	if (ret) {
2349 		intel_pstate_driver_cleanup();
2350 		return ret;
2351 	}
2352 
2353 	global.min_perf_pct = min_perf_pct_min();
2354 
2355 	return 0;
2356 }
2357 
2358 static int intel_pstate_unregister_driver(void)
2359 {
2360 	if (hwp_active)
2361 		return -EBUSY;
2362 
2363 	cpufreq_unregister_driver(intel_pstate_driver);
2364 	intel_pstate_driver_cleanup();
2365 
2366 	return 0;
2367 }
2368 
2369 static ssize_t intel_pstate_show_status(char *buf)
2370 {
2371 	if (!intel_pstate_driver)
2372 		return sprintf(buf, "off\n");
2373 
2374 	return sprintf(buf, "%s\n", intel_pstate_driver == &intel_pstate ?
2375 					"active" : "passive");
2376 }
2377 
2378 static int intel_pstate_update_status(const char *buf, size_t size)
2379 {
2380 	int ret;
2381 
2382 	if (size == 3 && !strncmp(buf, "off", size))
2383 		return intel_pstate_driver ?
2384 			intel_pstate_unregister_driver() : -EINVAL;
2385 
2386 	if (size == 6 && !strncmp(buf, "active", size)) {
2387 		if (intel_pstate_driver) {
2388 			if (intel_pstate_driver == &intel_pstate)
2389 				return 0;
2390 
2391 			ret = intel_pstate_unregister_driver();
2392 			if (ret)
2393 				return ret;
2394 		}
2395 
2396 		return intel_pstate_register_driver(&intel_pstate);
2397 	}
2398 
2399 	if (size == 7 && !strncmp(buf, "passive", size)) {
2400 		if (intel_pstate_driver) {
2401 			if (intel_pstate_driver == &intel_cpufreq)
2402 				return 0;
2403 
2404 			ret = intel_pstate_unregister_driver();
2405 			if (ret)
2406 				return ret;
2407 		}
2408 
2409 		return intel_pstate_register_driver(&intel_cpufreq);
2410 	}
2411 
2412 	return -EINVAL;
2413 }
2414 
2415 static int no_load __initdata;
2416 static int no_hwp __initdata;
2417 static int hwp_only __initdata;
2418 static unsigned int force_load __initdata;
2419 
2420 static int __init intel_pstate_msrs_not_valid(void)
2421 {
2422 	if (!pstate_funcs.get_max() ||
2423 	    !pstate_funcs.get_min() ||
2424 	    !pstate_funcs.get_turbo())
2425 		return -ENODEV;
2426 
2427 	return 0;
2428 }
2429 
2430 static void __init copy_cpu_funcs(struct pstate_funcs *funcs)
2431 {
2432 	pstate_funcs.get_max   = funcs->get_max;
2433 	pstate_funcs.get_max_physical = funcs->get_max_physical;
2434 	pstate_funcs.get_min   = funcs->get_min;
2435 	pstate_funcs.get_turbo = funcs->get_turbo;
2436 	pstate_funcs.get_scaling = funcs->get_scaling;
2437 	pstate_funcs.get_val   = funcs->get_val;
2438 	pstate_funcs.get_vid   = funcs->get_vid;
2439 	pstate_funcs.get_aperf_mperf_shift = funcs->get_aperf_mperf_shift;
2440 }
2441 
2442 #ifdef CONFIG_ACPI
2443 
2444 static bool __init intel_pstate_no_acpi_pss(void)
2445 {
2446 	int i;
2447 
2448 	for_each_possible_cpu(i) {
2449 		acpi_status status;
2450 		union acpi_object *pss;
2451 		struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
2452 		struct acpi_processor *pr = per_cpu(processors, i);
2453 
2454 		if (!pr)
2455 			continue;
2456 
2457 		status = acpi_evaluate_object(pr->handle, "_PSS", NULL, &buffer);
2458 		if (ACPI_FAILURE(status))
2459 			continue;
2460 
2461 		pss = buffer.pointer;
2462 		if (pss && pss->type == ACPI_TYPE_PACKAGE) {
2463 			kfree(pss);
2464 			return false;
2465 		}
2466 
2467 		kfree(pss);
2468 	}
2469 
2470 	pr_debug("ACPI _PSS not found\n");
2471 	return true;
2472 }
2473 
2474 static bool __init intel_pstate_no_acpi_pcch(void)
2475 {
2476 	acpi_status status;
2477 	acpi_handle handle;
2478 
2479 	status = acpi_get_handle(NULL, "\\_SB", &handle);
2480 	if (ACPI_FAILURE(status))
2481 		goto not_found;
2482 
2483 	if (acpi_has_method(handle, "PCCH"))
2484 		return false;
2485 
2486 not_found:
2487 	pr_debug("ACPI PCCH not found\n");
2488 	return true;
2489 }
2490 
2491 static bool __init intel_pstate_has_acpi_ppc(void)
2492 {
2493 	int i;
2494 
2495 	for_each_possible_cpu(i) {
2496 		struct acpi_processor *pr = per_cpu(processors, i);
2497 
2498 		if (!pr)
2499 			continue;
2500 		if (acpi_has_method(pr->handle, "_PPC"))
2501 			return true;
2502 	}
2503 	pr_debug("ACPI _PPC not found\n");
2504 	return false;
2505 }
2506 
2507 enum {
2508 	PSS,
2509 	PPC,
2510 };
2511 
2512 /* Hardware vendor-specific info that has its own power management modes */
2513 static struct acpi_platform_list plat_info[] __initdata = {
2514 	{"HP    ", "ProLiant", 0, ACPI_SIG_FADT, all_versions, 0, PSS},
2515 	{"ORACLE", "X4-2    ", 0, ACPI_SIG_FADT, all_versions, 0, PPC},
2516 	{"ORACLE", "X4-2L   ", 0, ACPI_SIG_FADT, all_versions, 0, PPC},
2517 	{"ORACLE", "X4-2B   ", 0, ACPI_SIG_FADT, all_versions, 0, PPC},
2518 	{"ORACLE", "X3-2    ", 0, ACPI_SIG_FADT, all_versions, 0, PPC},
2519 	{"ORACLE", "X3-2L   ", 0, ACPI_SIG_FADT, all_versions, 0, PPC},
2520 	{"ORACLE", "X3-2B   ", 0, ACPI_SIG_FADT, all_versions, 0, PPC},
2521 	{"ORACLE", "X4470M2 ", 0, ACPI_SIG_FADT, all_versions, 0, PPC},
2522 	{"ORACLE", "X4270M3 ", 0, ACPI_SIG_FADT, all_versions, 0, PPC},
2523 	{"ORACLE", "X4270M2 ", 0, ACPI_SIG_FADT, all_versions, 0, PPC},
2524 	{"ORACLE", "X4170M2 ", 0, ACPI_SIG_FADT, all_versions, 0, PPC},
2525 	{"ORACLE", "X4170 M3", 0, ACPI_SIG_FADT, all_versions, 0, PPC},
2526 	{"ORACLE", "X4275 M3", 0, ACPI_SIG_FADT, all_versions, 0, PPC},
2527 	{"ORACLE", "X6-2    ", 0, ACPI_SIG_FADT, all_versions, 0, PPC},
2528 	{"ORACLE", "Sudbury ", 0, ACPI_SIG_FADT, all_versions, 0, PPC},
2529 	{ } /* End */
2530 };
2531 
2532 static bool __init intel_pstate_platform_pwr_mgmt_exists(void)
2533 {
2534 	const struct x86_cpu_id *id;
2535 	u64 misc_pwr;
2536 	int idx;
2537 
2538 	id = x86_match_cpu(intel_pstate_cpu_oob_ids);
2539 	if (id) {
2540 		rdmsrl(MSR_MISC_PWR_MGMT, misc_pwr);
2541 		if (misc_pwr & (1 << 8)) {
2542 			pr_debug("Bit 8 in the MISC_PWR_MGMT MSR set\n");
2543 			return true;
2544 		}
2545 	}
2546 
2547 	idx = acpi_match_platform_list(plat_info);
2548 	if (idx < 0)
2549 		return false;
2550 
2551 	switch (plat_info[idx].data) {
2552 	case PSS:
2553 		if (!intel_pstate_no_acpi_pss())
2554 			return false;
2555 
2556 		return intel_pstate_no_acpi_pcch();
2557 	case PPC:
2558 		return intel_pstate_has_acpi_ppc() && !force_load;
2559 	}
2560 
2561 	return false;
2562 }
2563 
2564 static void intel_pstate_request_control_from_smm(void)
2565 {
2566 	/*
2567 	 * It may be unsafe to request P-states control from SMM if _PPC support
2568 	 * has not been enabled.
2569 	 */
2570 	if (acpi_ppc)
2571 		acpi_processor_pstate_control();
2572 }
2573 #else /* CONFIG_ACPI not enabled */
2574 static inline bool intel_pstate_platform_pwr_mgmt_exists(void) { return false; }
2575 static inline bool intel_pstate_has_acpi_ppc(void) { return false; }
2576 static inline void intel_pstate_request_control_from_smm(void) {}
2577 #endif /* CONFIG_ACPI */
2578 
2579 #define INTEL_PSTATE_HWP_BROADWELL	0x01
2580 
2581 #define ICPU_HWP(model, hwp_mode) \
2582 	{ X86_VENDOR_INTEL, 6, model, X86_FEATURE_HWP, hwp_mode }
2583 
2584 static const struct x86_cpu_id hwp_support_ids[] __initconst = {
2585 	ICPU_HWP(INTEL_FAM6_BROADWELL_X, INTEL_PSTATE_HWP_BROADWELL),
2586 	ICPU_HWP(INTEL_FAM6_BROADWELL_XEON_D, INTEL_PSTATE_HWP_BROADWELL),
2587 	ICPU_HWP(X86_MODEL_ANY, 0),
2588 	{}
2589 };
2590 
2591 static int __init intel_pstate_init(void)
2592 {
2593 	const struct x86_cpu_id *id;
2594 	int rc;
2595 
2596 	if (no_load)
2597 		return -ENODEV;
2598 
2599 	id = x86_match_cpu(hwp_support_ids);
2600 	if (id) {
2601 		copy_cpu_funcs(&core_funcs);
2602 		if (!no_hwp) {
2603 			hwp_active++;
2604 			hwp_mode_bdw = id->driver_data;
2605 			intel_pstate.attr = hwp_cpufreq_attrs;
2606 			goto hwp_cpu_matched;
2607 		}
2608 	} else {
2609 		id = x86_match_cpu(intel_pstate_cpu_ids);
2610 		if (!id) {
2611 			pr_info("CPU ID not supported\n");
2612 			return -ENODEV;
2613 		}
2614 
2615 		copy_cpu_funcs((struct pstate_funcs *)id->driver_data);
2616 	}
2617 
2618 	if (intel_pstate_msrs_not_valid()) {
2619 		pr_info("Invalid MSRs\n");
2620 		return -ENODEV;
2621 	}
2622 
2623 hwp_cpu_matched:
2624 	/*
2625 	 * The Intel pstate driver will be ignored if the platform
2626 	 * firmware has its own power management modes.
2627 	 */
2628 	if (intel_pstate_platform_pwr_mgmt_exists()) {
2629 		pr_info("P-states controlled by the platform\n");
2630 		return -ENODEV;
2631 	}
2632 
2633 	if (!hwp_active && hwp_only)
2634 		return -ENOTSUPP;
2635 
2636 	pr_info("Intel P-state driver initializing\n");
2637 
2638 	all_cpu_data = vzalloc(array_size(sizeof(void *), num_possible_cpus()));
2639 	if (!all_cpu_data)
2640 		return -ENOMEM;
2641 
2642 	intel_pstate_request_control_from_smm();
2643 
2644 	intel_pstate_sysfs_expose_params();
2645 
2646 	mutex_lock(&intel_pstate_driver_lock);
2647 	rc = intel_pstate_register_driver(default_driver);
2648 	mutex_unlock(&intel_pstate_driver_lock);
2649 	if (rc)
2650 		return rc;
2651 
2652 	if (hwp_active)
2653 		pr_info("HWP enabled\n");
2654 
2655 	return 0;
2656 }
2657 device_initcall(intel_pstate_init);
2658 
2659 static int __init intel_pstate_setup(char *str)
2660 {
2661 	if (!str)
2662 		return -EINVAL;
2663 
2664 	if (!strcmp(str, "disable")) {
2665 		no_load = 1;
2666 	} else if (!strcmp(str, "passive")) {
2667 		pr_info("Passive mode enabled\n");
2668 		default_driver = &intel_cpufreq;
2669 		no_hwp = 1;
2670 	}
2671 	if (!strcmp(str, "no_hwp")) {
2672 		pr_info("HWP disabled\n");
2673 		no_hwp = 1;
2674 	}
2675 	if (!strcmp(str, "force"))
2676 		force_load = 1;
2677 	if (!strcmp(str, "hwp_only"))
2678 		hwp_only = 1;
2679 	if (!strcmp(str, "per_cpu_perf_limits"))
2680 		per_cpu_limits = true;
2681 
2682 #ifdef CONFIG_ACPI
2683 	if (!strcmp(str, "support_acpi_ppc"))
2684 		acpi_ppc = true;
2685 #endif
2686 
2687 	return 0;
2688 }
2689 early_param("intel_pstate", intel_pstate_setup);
2690 
2691 MODULE_AUTHOR("Dirk Brandewie <dirk.j.brandewie@intel.com>");
2692 MODULE_DESCRIPTION("'intel_pstate' - P state driver Intel Core processors");
2693 MODULE_LICENSE("GPL");
2694