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