1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * intel_pstate.c: Native P state management for Intel processors
4  *
5  * (C) Copyright 2012 Intel Corporation
6  * Author: Dirk Brandewie <dirk.j.brandewie@intel.com>
7  */
8 
9 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
10 
11 #include <linux/kernel.h>
12 #include <linux/kernel_stat.h>
13 #include <linux/module.h>
14 #include <linux/ktime.h>
15 #include <linux/hrtimer.h>
16 #include <linux/tick.h>
17 #include <linux/slab.h>
18 #include <linux/sched/cpufreq.h>
19 #include <linux/list.h>
20 #include <linux/cpu.h>
21 #include <linux/cpufreq.h>
22 #include <linux/sysfs.h>
23 #include <linux/types.h>
24 #include <linux/fs.h>
25 #include <linux/acpi.h>
26 #include <linux/vmalloc.h>
27 #include <linux/pm_qos.h>
28 #include <trace/events/power.h>
29 
30 #include <asm/div64.h>
31 #include <asm/msr.h>
32 #include <asm/cpu_device_id.h>
33 #include <asm/cpufeature.h>
34 #include <asm/intel-family.h>
35 
36 #define INTEL_PSTATE_SAMPLING_INTERVAL	(10 * NSEC_PER_MSEC)
37 
38 #define INTEL_CPUFREQ_TRANSITION_LATENCY	20000
39 #define INTEL_CPUFREQ_TRANSITION_DELAY		500
40 
41 #ifdef CONFIG_ACPI
42 #include <acpi/processor.h>
43 #include <acpi/cppc_acpi.h>
44 #endif
45 
46 #define FRAC_BITS 8
47 #define int_tofp(X) ((int64_t)(X) << FRAC_BITS)
48 #define fp_toint(X) ((X) >> FRAC_BITS)
49 
50 #define ONE_EIGHTH_FP ((int64_t)1 << (FRAC_BITS - 3))
51 
52 #define EXT_BITS 6
53 #define EXT_FRAC_BITS (EXT_BITS + FRAC_BITS)
54 #define fp_ext_toint(X) ((X) >> EXT_FRAC_BITS)
55 #define int_ext_tofp(X) ((int64_t)(X) << EXT_FRAC_BITS)
56 
57 static inline int32_t mul_fp(int32_t x, int32_t y)
58 {
59 	return ((int64_t)x * (int64_t)y) >> FRAC_BITS;
60 }
61 
62 static inline int32_t div_fp(s64 x, s64 y)
63 {
64 	return div64_s64((int64_t)x << FRAC_BITS, y);
65 }
66 
67 static inline int ceiling_fp(int32_t x)
68 {
69 	int mask, ret;
70 
71 	ret = fp_toint(x);
72 	mask = (1 << FRAC_BITS) - 1;
73 	if (x & mask)
74 		ret += 1;
75 	return ret;
76 }
77 
78 static inline int32_t percent_fp(int percent)
79 {
80 	return div_fp(percent, 100);
81 }
82 
83 static inline u64 mul_ext_fp(u64 x, u64 y)
84 {
85 	return (x * y) >> EXT_FRAC_BITS;
86 }
87 
88 static inline u64 div_ext_fp(u64 x, u64 y)
89 {
90 	return div64_u64(x << EXT_FRAC_BITS, y);
91 }
92 
93 static inline int32_t percent_ext_fp(int percent)
94 {
95 	return div_ext_fp(percent, 100);
96 }
97 
98 /**
99  * struct sample -	Store performance sample
100  * @core_avg_perf:	Ratio of APERF/MPERF which is the actual average
101  *			performance during last sample period
102  * @busy_scaled:	Scaled busy value which is used to calculate next
103  *			P state. This can be different than core_avg_perf
104  *			to account for cpu idle period
105  * @aperf:		Difference of actual performance frequency clock count
106  *			read from APERF MSR between last and current sample
107  * @mperf:		Difference of maximum performance frequency clock count
108  *			read from MPERF MSR between last and current sample
109  * @tsc:		Difference of time stamp counter between last and
110  *			current sample
111  * @time:		Current time from scheduler
112  *
113  * This structure is used in the cpudata structure to store performance sample
114  * data for choosing next P State.
115  */
116 struct sample {
117 	int32_t core_avg_perf;
118 	int32_t busy_scaled;
119 	u64 aperf;
120 	u64 mperf;
121 	u64 tsc;
122 	u64 time;
123 };
124 
125 /**
126  * struct pstate_data - Store P state data
127  * @current_pstate:	Current requested P state
128  * @min_pstate:		Min P state possible for this platform
129  * @max_pstate:		Max P state possible for this platform
130  * @max_pstate_physical:This is physical Max P state for a processor
131  *			This can be higher than the max_pstate which can
132  *			be limited by platform thermal design power limits
133  * @scaling:		Scaling factor to  convert frequency to cpufreq
134  *			frequency units
135  * @turbo_pstate:	Max Turbo P state possible for this platform
136  * @max_freq:		@max_pstate frequency in cpufreq units
137  * @turbo_freq:		@turbo_pstate frequency in cpufreq units
138  *
139  * Stores the per cpu model P state limits and current P state.
140  */
141 struct pstate_data {
142 	int	current_pstate;
143 	int	min_pstate;
144 	int	max_pstate;
145 	int	max_pstate_physical;
146 	int	scaling;
147 	int	turbo_pstate;
148 	unsigned int max_freq;
149 	unsigned int turbo_freq;
150 };
151 
152 /**
153  * struct vid_data -	Stores voltage information data
154  * @min:		VID data for this platform corresponding to
155  *			the lowest P state
156  * @max:		VID data corresponding to the highest P State.
157  * @turbo:		VID data for turbo P state
158  * @ratio:		Ratio of (vid max - vid min) /
159  *			(max P state - Min P State)
160  *
161  * Stores the voltage data for DVFS (Dynamic Voltage and Frequency Scaling)
162  * This data is used in Atom platforms, where in addition to target P state,
163  * the voltage data needs to be specified to select next P State.
164  */
165 struct vid_data {
166 	int min;
167 	int max;
168 	int turbo;
169 	int32_t ratio;
170 };
171 
172 /**
173  * struct global_params - Global parameters, mostly tunable via sysfs.
174  * @no_turbo:		Whether or not to use turbo P-states.
175  * @turbo_disabled:	Whether or not turbo P-states are available at all,
176  *			based on the MSR_IA32_MISC_ENABLE value and whether or
177  *			not the maximum reported turbo P-state is different from
178  *			the maximum reported non-turbo one.
179  * @turbo_disabled_mf:	The @turbo_disabled value reflected by cpuinfo.max_freq.
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 	bool turbo_disabled_mf;
189 	int max_perf_pct;
190 	int min_perf_pct;
191 };
192 
193 /**
194  * struct cpudata -	Per CPU instance data storage
195  * @cpu:		CPU number for this instance data
196  * @policy:		CPUFreq policy value
197  * @update_util:	CPUFreq utility callback information
198  * @update_util_set:	CPUFreq utility callback is set
199  * @iowait_boost:	iowait-related boost fraction
200  * @last_update:	Time of the last update.
201  * @pstate:		Stores P state limits for this CPU
202  * @vid:		Stores VID limits for this CPU
203  * @last_sample_time:	Last Sample time
204  * @aperf_mperf_shift:	Number of clock cycles after aperf, merf is incremented
205  *			This shift is a multiplier to mperf delta to
206  *			calculate CPU busy.
207  * @prev_aperf:		Last APERF value read from APERF MSR
208  * @prev_mperf:		Last MPERF value read from MPERF MSR
209  * @prev_tsc:		Last timestamp counter (TSC) value
210  * @prev_cummulative_iowait: IO Wait time difference from last and
211  *			current sample
212  * @sample:		Storage for storing last Sample data
213  * @min_perf_ratio:	Minimum capacity in terms of PERF or HWP ratios
214  * @max_perf_ratio:	Maximum capacity in terms of PERF or HWP ratios
215  * @acpi_perf_data:	Stores ACPI perf information read from _PSS
216  * @valid_pss_table:	Set to true for valid ACPI _PSS entries found
217  * @epp_powersave:	Last saved HWP energy performance preference
218  *			(EPP) or energy performance bias (EPB),
219  *			when policy switched to performance
220  * @epp_policy:		Last saved policy used to set EPP/EPB
221  * @epp_default:	Power on default HWP energy performance
222  *			preference/bias
223  * @epp_saved:		Saved EPP/EPB during system suspend or CPU offline
224  *			operation
225  * @hwp_req_cached:	Cached value of the last HWP Request MSR
226  * @hwp_cap_cached:	Cached value of the last HWP Capabilities MSR
227  * @last_io_update:	Last time when IO wake flag was set
228  * @sched_flags:	Store scheduler flags for possible cross CPU update
229  * @hwp_boost_min:	Last HWP boosted min performance
230  *
231  * This structure stores per CPU instance data for all CPUs.
232  */
233 struct cpudata {
234 	int cpu;
235 
236 	unsigned int policy;
237 	struct update_util_data update_util;
238 	bool   update_util_set;
239 
240 	struct pstate_data pstate;
241 	struct vid_data vid;
242 
243 	u64	last_update;
244 	u64	last_sample_time;
245 	u64	aperf_mperf_shift;
246 	u64	prev_aperf;
247 	u64	prev_mperf;
248 	u64	prev_tsc;
249 	u64	prev_cummulative_iowait;
250 	struct sample sample;
251 	int32_t	min_perf_ratio;
252 	int32_t	max_perf_ratio;
253 #ifdef CONFIG_ACPI
254 	struct acpi_processor_performance acpi_perf_data;
255 	bool valid_pss_table;
256 #endif
257 	unsigned int iowait_boost;
258 	s16 epp_powersave;
259 	s16 epp_policy;
260 	s16 epp_default;
261 	s16 epp_saved;
262 	u64 hwp_req_cached;
263 	u64 hwp_cap_cached;
264 	u64 last_io_update;
265 	unsigned int sched_flags;
266 	u32 hwp_boost_min;
267 };
268 
269 static struct cpudata **all_cpu_data;
270 
271 /**
272  * struct pstate_funcs - Per CPU model specific callbacks
273  * @get_max:		Callback to get maximum non turbo effective P state
274  * @get_max_physical:	Callback to get maximum non turbo physical P state
275  * @get_min:		Callback to get minimum P state
276  * @get_turbo:		Callback to get turbo P state
277  * @get_scaling:	Callback to get frequency scaling factor
278  * @get_val:		Callback to convert P state to actual MSR write value
279  * @get_vid:		Callback to get VID data for Atom platforms
280  *
281  * Core and Atom CPU models have different way to get P State limits. This
282  * structure is used to store those callbacks.
283  */
284 struct pstate_funcs {
285 	int (*get_max)(void);
286 	int (*get_max_physical)(void);
287 	int (*get_min)(void);
288 	int (*get_turbo)(void);
289 	int (*get_scaling)(void);
290 	int (*get_aperf_mperf_shift)(void);
291 	u64 (*get_val)(struct cpudata*, int pstate);
292 	void (*get_vid)(struct cpudata *);
293 };
294 
295 static struct pstate_funcs pstate_funcs __read_mostly;
296 
297 static int hwp_active __read_mostly;
298 static int hwp_mode_bdw __read_mostly;
299 static bool per_cpu_limits __read_mostly;
300 static bool hwp_boost __read_mostly;
301 
302 static struct cpufreq_driver *intel_pstate_driver __read_mostly;
303 
304 #ifdef CONFIG_ACPI
305 static bool acpi_ppc;
306 #endif
307 
308 static struct global_params global;
309 
310 static DEFINE_MUTEX(intel_pstate_driver_lock);
311 static DEFINE_MUTEX(intel_pstate_limits_lock);
312 
313 #ifdef CONFIG_ACPI
314 
315 static bool intel_pstate_acpi_pm_profile_server(void)
316 {
317 	if (acpi_gbl_FADT.preferred_profile == PM_ENTERPRISE_SERVER ||
318 	    acpi_gbl_FADT.preferred_profile == PM_PERFORMANCE_SERVER)
319 		return true;
320 
321 	return false;
322 }
323 
324 static bool intel_pstate_get_ppc_enable_status(void)
325 {
326 	if (intel_pstate_acpi_pm_profile_server())
327 		return true;
328 
329 	return acpi_ppc;
330 }
331 
332 #ifdef CONFIG_ACPI_CPPC_LIB
333 
334 /* The work item is needed to avoid CPU hotplug locking issues */
335 static void intel_pstste_sched_itmt_work_fn(struct work_struct *work)
336 {
337 	sched_set_itmt_support();
338 }
339 
340 static DECLARE_WORK(sched_itmt_work, intel_pstste_sched_itmt_work_fn);
341 
342 static void intel_pstate_set_itmt_prio(int cpu)
343 {
344 	struct cppc_perf_caps cppc_perf;
345 	static u32 max_highest_perf = 0, min_highest_perf = U32_MAX;
346 	int ret;
347 
348 	ret = cppc_get_perf_caps(cpu, &cppc_perf);
349 	if (ret)
350 		return;
351 
352 	/*
353 	 * The priorities can be set regardless of whether or not
354 	 * sched_set_itmt_support(true) has been called and it is valid to
355 	 * update them at any time after it has been called.
356 	 */
357 	sched_set_itmt_core_prio(cppc_perf.highest_perf, cpu);
358 
359 	if (max_highest_perf <= min_highest_perf) {
360 		if (cppc_perf.highest_perf > max_highest_perf)
361 			max_highest_perf = cppc_perf.highest_perf;
362 
363 		if (cppc_perf.highest_perf < min_highest_perf)
364 			min_highest_perf = cppc_perf.highest_perf;
365 
366 		if (max_highest_perf > min_highest_perf) {
367 			/*
368 			 * This code can be run during CPU online under the
369 			 * CPU hotplug locks, so sched_set_itmt_support()
370 			 * cannot be called from here.  Queue up a work item
371 			 * to invoke it.
372 			 */
373 			schedule_work(&sched_itmt_work);
374 		}
375 	}
376 }
377 
378 static int intel_pstate_get_cppc_guranteed(int cpu)
379 {
380 	struct cppc_perf_caps cppc_perf;
381 	int ret;
382 
383 	ret = cppc_get_perf_caps(cpu, &cppc_perf);
384 	if (ret)
385 		return ret;
386 
387 	if (cppc_perf.guaranteed_perf)
388 		return cppc_perf.guaranteed_perf;
389 
390 	return cppc_perf.nominal_perf;
391 }
392 
393 #else /* CONFIG_ACPI_CPPC_LIB */
394 static void intel_pstate_set_itmt_prio(int cpu)
395 {
396 }
397 #endif /* CONFIG_ACPI_CPPC_LIB */
398 
399 static void intel_pstate_init_acpi_perf_limits(struct cpufreq_policy *policy)
400 {
401 	struct cpudata *cpu;
402 	int ret;
403 	int i;
404 
405 	if (hwp_active) {
406 		intel_pstate_set_itmt_prio(policy->cpu);
407 		return;
408 	}
409 
410 	if (!intel_pstate_get_ppc_enable_status())
411 		return;
412 
413 	cpu = all_cpu_data[policy->cpu];
414 
415 	ret = acpi_processor_register_performance(&cpu->acpi_perf_data,
416 						  policy->cpu);
417 	if (ret)
418 		return;
419 
420 	/*
421 	 * Check if the control value in _PSS is for PERF_CTL MSR, which should
422 	 * guarantee that the states returned by it map to the states in our
423 	 * list directly.
424 	 */
425 	if (cpu->acpi_perf_data.control_register.space_id !=
426 						ACPI_ADR_SPACE_FIXED_HARDWARE)
427 		goto err;
428 
429 	/*
430 	 * If there is only one entry _PSS, simply ignore _PSS and continue as
431 	 * usual without taking _PSS into account
432 	 */
433 	if (cpu->acpi_perf_data.state_count < 2)
434 		goto err;
435 
436 	pr_debug("CPU%u - ACPI _PSS perf data\n", policy->cpu);
437 	for (i = 0; i < cpu->acpi_perf_data.state_count; i++) {
438 		pr_debug("     %cP%d: %u MHz, %u mW, 0x%x\n",
439 			 (i == cpu->acpi_perf_data.state ? '*' : ' '), i,
440 			 (u32) cpu->acpi_perf_data.states[i].core_frequency,
441 			 (u32) cpu->acpi_perf_data.states[i].power,
442 			 (u32) cpu->acpi_perf_data.states[i].control);
443 	}
444 
445 	/*
446 	 * The _PSS table doesn't contain whole turbo frequency range.
447 	 * This just contains +1 MHZ above the max non turbo frequency,
448 	 * with control value corresponding to max turbo ratio. But
449 	 * when cpufreq set policy is called, it will call with this
450 	 * max frequency, which will cause a reduced performance as
451 	 * this driver uses real max turbo frequency as the max
452 	 * frequency. So correct this frequency in _PSS table to
453 	 * correct max turbo frequency based on the turbo state.
454 	 * Also need to convert to MHz as _PSS freq is in MHz.
455 	 */
456 	if (!global.turbo_disabled)
457 		cpu->acpi_perf_data.states[0].core_frequency =
458 					policy->cpuinfo.max_freq / 1000;
459 	cpu->valid_pss_table = true;
460 	pr_debug("_PPC limits will be enforced\n");
461 
462 	return;
463 
464  err:
465 	cpu->valid_pss_table = false;
466 	acpi_processor_unregister_performance(policy->cpu);
467 }
468 
469 static void intel_pstate_exit_perf_limits(struct cpufreq_policy *policy)
470 {
471 	struct cpudata *cpu;
472 
473 	cpu = all_cpu_data[policy->cpu];
474 	if (!cpu->valid_pss_table)
475 		return;
476 
477 	acpi_processor_unregister_performance(policy->cpu);
478 }
479 #else /* CONFIG_ACPI */
480 static inline void intel_pstate_init_acpi_perf_limits(struct cpufreq_policy *policy)
481 {
482 }
483 
484 static inline void intel_pstate_exit_perf_limits(struct cpufreq_policy *policy)
485 {
486 }
487 
488 static inline bool intel_pstate_acpi_pm_profile_server(void)
489 {
490 	return false;
491 }
492 #endif /* CONFIG_ACPI */
493 
494 #ifndef CONFIG_ACPI_CPPC_LIB
495 static int intel_pstate_get_cppc_guranteed(int cpu)
496 {
497 	return -ENOTSUPP;
498 }
499 #endif /* CONFIG_ACPI_CPPC_LIB */
500 
501 static inline void update_turbo_state(void)
502 {
503 	u64 misc_en;
504 	struct cpudata *cpu;
505 
506 	cpu = all_cpu_data[0];
507 	rdmsrl(MSR_IA32_MISC_ENABLE, misc_en);
508 	global.turbo_disabled =
509 		(misc_en & MSR_IA32_MISC_ENABLE_TURBO_DISABLE ||
510 		 cpu->pstate.max_pstate == cpu->pstate.turbo_pstate);
511 }
512 
513 static int min_perf_pct_min(void)
514 {
515 	struct cpudata *cpu = all_cpu_data[0];
516 	int turbo_pstate = cpu->pstate.turbo_pstate;
517 
518 	return turbo_pstate ?
519 		(cpu->pstate.min_pstate * 100 / turbo_pstate) : 0;
520 }
521 
522 static s16 intel_pstate_get_epb(struct cpudata *cpu_data)
523 {
524 	u64 epb;
525 	int ret;
526 
527 	if (!boot_cpu_has(X86_FEATURE_EPB))
528 		return -ENXIO;
529 
530 	ret = rdmsrl_on_cpu(cpu_data->cpu, MSR_IA32_ENERGY_PERF_BIAS, &epb);
531 	if (ret)
532 		return (s16)ret;
533 
534 	return (s16)(epb & 0x0f);
535 }
536 
537 static s16 intel_pstate_get_epp(struct cpudata *cpu_data, u64 hwp_req_data)
538 {
539 	s16 epp;
540 
541 	if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
542 		/*
543 		 * When hwp_req_data is 0, means that caller didn't read
544 		 * MSR_HWP_REQUEST, so need to read and get EPP.
545 		 */
546 		if (!hwp_req_data) {
547 			epp = rdmsrl_on_cpu(cpu_data->cpu, MSR_HWP_REQUEST,
548 					    &hwp_req_data);
549 			if (epp)
550 				return epp;
551 		}
552 		epp = (hwp_req_data >> 24) & 0xff;
553 	} else {
554 		/* When there is no EPP present, HWP uses EPB settings */
555 		epp = intel_pstate_get_epb(cpu_data);
556 	}
557 
558 	return epp;
559 }
560 
561 static int intel_pstate_set_epb(int cpu, s16 pref)
562 {
563 	u64 epb;
564 	int ret;
565 
566 	if (!boot_cpu_has(X86_FEATURE_EPB))
567 		return -ENXIO;
568 
569 	ret = rdmsrl_on_cpu(cpu, MSR_IA32_ENERGY_PERF_BIAS, &epb);
570 	if (ret)
571 		return ret;
572 
573 	epb = (epb & ~0x0f) | pref;
574 	wrmsrl_on_cpu(cpu, MSR_IA32_ENERGY_PERF_BIAS, epb);
575 
576 	return 0;
577 }
578 
579 /*
580  * EPP/EPB display strings corresponding to EPP index in the
581  * energy_perf_strings[]
582  *	index		String
583  *-------------------------------------
584  *	0		default
585  *	1		performance
586  *	2		balance_performance
587  *	3		balance_power
588  *	4		power
589  */
590 static const char * const energy_perf_strings[] = {
591 	"default",
592 	"performance",
593 	"balance_performance",
594 	"balance_power",
595 	"power",
596 	NULL
597 };
598 static const unsigned int epp_values[] = {
599 	HWP_EPP_PERFORMANCE,
600 	HWP_EPP_BALANCE_PERFORMANCE,
601 	HWP_EPP_BALANCE_POWERSAVE,
602 	HWP_EPP_POWERSAVE
603 };
604 
605 static int intel_pstate_get_energy_pref_index(struct cpudata *cpu_data)
606 {
607 	s16 epp;
608 	int index = -EINVAL;
609 
610 	epp = intel_pstate_get_epp(cpu_data, 0);
611 	if (epp < 0)
612 		return epp;
613 
614 	if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
615 		if (epp == HWP_EPP_PERFORMANCE)
616 			return 1;
617 		if (epp <= HWP_EPP_BALANCE_PERFORMANCE)
618 			return 2;
619 		if (epp <= HWP_EPP_BALANCE_POWERSAVE)
620 			return 3;
621 		else
622 			return 4;
623 	} else if (boot_cpu_has(X86_FEATURE_EPB)) {
624 		/*
625 		 * Range:
626 		 *	0x00-0x03	:	Performance
627 		 *	0x04-0x07	:	Balance performance
628 		 *	0x08-0x0B	:	Balance power
629 		 *	0x0C-0x0F	:	Power
630 		 * The EPB is a 4 bit value, but our ranges restrict the
631 		 * value which can be set. Here only using top two bits
632 		 * effectively.
633 		 */
634 		index = (epp >> 2) + 1;
635 	}
636 
637 	return index;
638 }
639 
640 static int intel_pstate_set_energy_pref_index(struct cpudata *cpu_data,
641 					      int pref_index)
642 {
643 	int epp = -EINVAL;
644 	int ret;
645 
646 	if (!pref_index)
647 		epp = cpu_data->epp_default;
648 
649 	mutex_lock(&intel_pstate_limits_lock);
650 
651 	if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
652 		u64 value;
653 
654 		ret = rdmsrl_on_cpu(cpu_data->cpu, MSR_HWP_REQUEST, &value);
655 		if (ret)
656 			goto return_pref;
657 
658 		value &= ~GENMASK_ULL(31, 24);
659 
660 		if (epp == -EINVAL)
661 			epp = epp_values[pref_index - 1];
662 
663 		value |= (u64)epp << 24;
664 		ret = wrmsrl_on_cpu(cpu_data->cpu, MSR_HWP_REQUEST, value);
665 	} else {
666 		if (epp == -EINVAL)
667 			epp = (pref_index - 1) << 2;
668 		ret = intel_pstate_set_epb(cpu_data->cpu, epp);
669 	}
670 return_pref:
671 	mutex_unlock(&intel_pstate_limits_lock);
672 
673 	return ret;
674 }
675 
676 static ssize_t show_energy_performance_available_preferences(
677 				struct cpufreq_policy *policy, char *buf)
678 {
679 	int i = 0;
680 	int ret = 0;
681 
682 	while (energy_perf_strings[i] != NULL)
683 		ret += sprintf(&buf[ret], "%s ", energy_perf_strings[i++]);
684 
685 	ret += sprintf(&buf[ret], "\n");
686 
687 	return ret;
688 }
689 
690 cpufreq_freq_attr_ro(energy_performance_available_preferences);
691 
692 static ssize_t store_energy_performance_preference(
693 		struct cpufreq_policy *policy, const char *buf, size_t count)
694 {
695 	struct cpudata *cpu_data = all_cpu_data[policy->cpu];
696 	char str_preference[21];
697 	int ret;
698 
699 	ret = sscanf(buf, "%20s", str_preference);
700 	if (ret != 1)
701 		return -EINVAL;
702 
703 	ret = match_string(energy_perf_strings, -1, str_preference);
704 	if (ret < 0)
705 		return ret;
706 
707 	intel_pstate_set_energy_pref_index(cpu_data, ret);
708 	return count;
709 }
710 
711 static ssize_t show_energy_performance_preference(
712 				struct cpufreq_policy *policy, char *buf)
713 {
714 	struct cpudata *cpu_data = all_cpu_data[policy->cpu];
715 	int preference;
716 
717 	preference = intel_pstate_get_energy_pref_index(cpu_data);
718 	if (preference < 0)
719 		return preference;
720 
721 	return  sprintf(buf, "%s\n", energy_perf_strings[preference]);
722 }
723 
724 cpufreq_freq_attr_rw(energy_performance_preference);
725 
726 static ssize_t show_base_frequency(struct cpufreq_policy *policy, char *buf)
727 {
728 	struct cpudata *cpu;
729 	u64 cap;
730 	int ratio;
731 
732 	ratio = intel_pstate_get_cppc_guranteed(policy->cpu);
733 	if (ratio <= 0) {
734 		rdmsrl_on_cpu(policy->cpu, MSR_HWP_CAPABILITIES, &cap);
735 		ratio = HWP_GUARANTEED_PERF(cap);
736 	}
737 
738 	cpu = all_cpu_data[policy->cpu];
739 
740 	return sprintf(buf, "%d\n", ratio * cpu->pstate.scaling);
741 }
742 
743 cpufreq_freq_attr_ro(base_frequency);
744 
745 static struct freq_attr *hwp_cpufreq_attrs[] = {
746 	&energy_performance_preference,
747 	&energy_performance_available_preferences,
748 	&base_frequency,
749 	NULL,
750 };
751 
752 static void intel_pstate_get_hwp_max(unsigned int cpu, int *phy_max,
753 				     int *current_max)
754 {
755 	u64 cap;
756 
757 	rdmsrl_on_cpu(cpu, MSR_HWP_CAPABILITIES, &cap);
758 	WRITE_ONCE(all_cpu_data[cpu]->hwp_cap_cached, cap);
759 	if (global.no_turbo)
760 		*current_max = HWP_GUARANTEED_PERF(cap);
761 	else
762 		*current_max = HWP_HIGHEST_PERF(cap);
763 
764 	*phy_max = HWP_HIGHEST_PERF(cap);
765 }
766 
767 static void intel_pstate_hwp_set(unsigned int cpu)
768 {
769 	struct cpudata *cpu_data = all_cpu_data[cpu];
770 	int max, min;
771 	u64 value;
772 	s16 epp;
773 
774 	max = cpu_data->max_perf_ratio;
775 	min = cpu_data->min_perf_ratio;
776 
777 	if (cpu_data->policy == CPUFREQ_POLICY_PERFORMANCE)
778 		min = max;
779 
780 	rdmsrl_on_cpu(cpu, MSR_HWP_REQUEST, &value);
781 
782 	value &= ~HWP_MIN_PERF(~0L);
783 	value |= HWP_MIN_PERF(min);
784 
785 	value &= ~HWP_MAX_PERF(~0L);
786 	value |= HWP_MAX_PERF(max);
787 
788 	if (cpu_data->epp_policy == cpu_data->policy)
789 		goto skip_epp;
790 
791 	cpu_data->epp_policy = cpu_data->policy;
792 
793 	if (cpu_data->epp_saved >= 0) {
794 		epp = cpu_data->epp_saved;
795 		cpu_data->epp_saved = -EINVAL;
796 		goto update_epp;
797 	}
798 
799 	if (cpu_data->policy == CPUFREQ_POLICY_PERFORMANCE) {
800 		epp = intel_pstate_get_epp(cpu_data, value);
801 		cpu_data->epp_powersave = epp;
802 		/* If EPP read was failed, then don't try to write */
803 		if (epp < 0)
804 			goto skip_epp;
805 
806 		epp = 0;
807 	} else {
808 		/* skip setting EPP, when saved value is invalid */
809 		if (cpu_data->epp_powersave < 0)
810 			goto skip_epp;
811 
812 		/*
813 		 * No need to restore EPP when it is not zero. This
814 		 * means:
815 		 *  - Policy is not changed
816 		 *  - user has manually changed
817 		 *  - Error reading EPB
818 		 */
819 		epp = intel_pstate_get_epp(cpu_data, value);
820 		if (epp)
821 			goto skip_epp;
822 
823 		epp = cpu_data->epp_powersave;
824 	}
825 update_epp:
826 	if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
827 		value &= ~GENMASK_ULL(31, 24);
828 		value |= (u64)epp << 24;
829 	} else {
830 		intel_pstate_set_epb(cpu, epp);
831 	}
832 skip_epp:
833 	WRITE_ONCE(cpu_data->hwp_req_cached, value);
834 	wrmsrl_on_cpu(cpu, MSR_HWP_REQUEST, value);
835 }
836 
837 static void intel_pstate_hwp_force_min_perf(int cpu)
838 {
839 	u64 value;
840 	int min_perf;
841 
842 	value = all_cpu_data[cpu]->hwp_req_cached;
843 	value &= ~GENMASK_ULL(31, 0);
844 	min_perf = HWP_LOWEST_PERF(all_cpu_data[cpu]->hwp_cap_cached);
845 
846 	/* Set hwp_max = hwp_min */
847 	value |= HWP_MAX_PERF(min_perf);
848 	value |= HWP_MIN_PERF(min_perf);
849 
850 	/* Set EPP to min */
851 	if (boot_cpu_has(X86_FEATURE_HWP_EPP))
852 		value |= HWP_ENERGY_PERF_PREFERENCE(HWP_EPP_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 static void intel_pstate_update_max_freq(unsigned int cpu)
898 {
899 	struct cpufreq_policy *policy = cpufreq_cpu_acquire(cpu);
900 	struct cpudata *cpudata;
901 
902 	if (!policy)
903 		return;
904 
905 	cpudata = all_cpu_data[cpu];
906 	policy->cpuinfo.max_freq = global.turbo_disabled_mf ?
907 			cpudata->pstate.max_freq : cpudata->pstate.turbo_freq;
908 
909 	refresh_frequency_limits(policy);
910 
911 	cpufreq_cpu_release(policy);
912 }
913 
914 static void intel_pstate_update_limits(unsigned int cpu)
915 {
916 	mutex_lock(&intel_pstate_driver_lock);
917 
918 	update_turbo_state();
919 	/*
920 	 * If turbo has been turned on or off globally, policy limits for
921 	 * all CPUs need to be updated to reflect that.
922 	 */
923 	if (global.turbo_disabled_mf != global.turbo_disabled) {
924 		global.turbo_disabled_mf = global.turbo_disabled;
925 		arch_set_max_freq_ratio(global.turbo_disabled);
926 		for_each_possible_cpu(cpu)
927 			intel_pstate_update_max_freq(cpu);
928 	} else {
929 		cpufreq_update_policy(cpu);
930 	}
931 
932 	mutex_unlock(&intel_pstate_driver_lock);
933 }
934 
935 /************************** sysfs begin ************************/
936 #define show_one(file_name, object)					\
937 	static ssize_t show_##file_name					\
938 	(struct kobject *kobj, struct kobj_attribute *attr, char *buf)	\
939 	{								\
940 		return sprintf(buf, "%u\n", global.object);		\
941 	}
942 
943 static ssize_t intel_pstate_show_status(char *buf);
944 static int intel_pstate_update_status(const char *buf, size_t size);
945 
946 static ssize_t show_status(struct kobject *kobj,
947 			   struct kobj_attribute *attr, char *buf)
948 {
949 	ssize_t ret;
950 
951 	mutex_lock(&intel_pstate_driver_lock);
952 	ret = intel_pstate_show_status(buf);
953 	mutex_unlock(&intel_pstate_driver_lock);
954 
955 	return ret;
956 }
957 
958 static ssize_t store_status(struct kobject *a, struct kobj_attribute *b,
959 			    const char *buf, size_t count)
960 {
961 	char *p = memchr(buf, '\n', count);
962 	int ret;
963 
964 	mutex_lock(&intel_pstate_driver_lock);
965 	ret = intel_pstate_update_status(buf, p ? p - buf : count);
966 	mutex_unlock(&intel_pstate_driver_lock);
967 
968 	return ret < 0 ? ret : count;
969 }
970 
971 static ssize_t show_turbo_pct(struct kobject *kobj,
972 				struct kobj_attribute *attr, char *buf)
973 {
974 	struct cpudata *cpu;
975 	int total, no_turbo, turbo_pct;
976 	uint32_t turbo_fp;
977 
978 	mutex_lock(&intel_pstate_driver_lock);
979 
980 	if (!intel_pstate_driver) {
981 		mutex_unlock(&intel_pstate_driver_lock);
982 		return -EAGAIN;
983 	}
984 
985 	cpu = all_cpu_data[0];
986 
987 	total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
988 	no_turbo = cpu->pstate.max_pstate - cpu->pstate.min_pstate + 1;
989 	turbo_fp = div_fp(no_turbo, total);
990 	turbo_pct = 100 - fp_toint(mul_fp(turbo_fp, int_tofp(100)));
991 
992 	mutex_unlock(&intel_pstate_driver_lock);
993 
994 	return sprintf(buf, "%u\n", turbo_pct);
995 }
996 
997 static ssize_t show_num_pstates(struct kobject *kobj,
998 				struct kobj_attribute *attr, char *buf)
999 {
1000 	struct cpudata *cpu;
1001 	int total;
1002 
1003 	mutex_lock(&intel_pstate_driver_lock);
1004 
1005 	if (!intel_pstate_driver) {
1006 		mutex_unlock(&intel_pstate_driver_lock);
1007 		return -EAGAIN;
1008 	}
1009 
1010 	cpu = all_cpu_data[0];
1011 	total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
1012 
1013 	mutex_unlock(&intel_pstate_driver_lock);
1014 
1015 	return sprintf(buf, "%u\n", total);
1016 }
1017 
1018 static ssize_t show_no_turbo(struct kobject *kobj,
1019 			     struct kobj_attribute *attr, char *buf)
1020 {
1021 	ssize_t ret;
1022 
1023 	mutex_lock(&intel_pstate_driver_lock);
1024 
1025 	if (!intel_pstate_driver) {
1026 		mutex_unlock(&intel_pstate_driver_lock);
1027 		return -EAGAIN;
1028 	}
1029 
1030 	update_turbo_state();
1031 	if (global.turbo_disabled)
1032 		ret = sprintf(buf, "%u\n", global.turbo_disabled);
1033 	else
1034 		ret = sprintf(buf, "%u\n", global.no_turbo);
1035 
1036 	mutex_unlock(&intel_pstate_driver_lock);
1037 
1038 	return ret;
1039 }
1040 
1041 static ssize_t store_no_turbo(struct kobject *a, struct kobj_attribute *b,
1042 			      const char *buf, size_t count)
1043 {
1044 	unsigned int input;
1045 	int ret;
1046 
1047 	ret = sscanf(buf, "%u", &input);
1048 	if (ret != 1)
1049 		return -EINVAL;
1050 
1051 	mutex_lock(&intel_pstate_driver_lock);
1052 
1053 	if (!intel_pstate_driver) {
1054 		mutex_unlock(&intel_pstate_driver_lock);
1055 		return -EAGAIN;
1056 	}
1057 
1058 	mutex_lock(&intel_pstate_limits_lock);
1059 
1060 	update_turbo_state();
1061 	if (global.turbo_disabled) {
1062 		pr_warn("Turbo disabled by BIOS or unavailable on processor\n");
1063 		mutex_unlock(&intel_pstate_limits_lock);
1064 		mutex_unlock(&intel_pstate_driver_lock);
1065 		return -EPERM;
1066 	}
1067 
1068 	global.no_turbo = clamp_t(int, input, 0, 1);
1069 
1070 	if (global.no_turbo) {
1071 		struct cpudata *cpu = all_cpu_data[0];
1072 		int pct = cpu->pstate.max_pstate * 100 / cpu->pstate.turbo_pstate;
1073 
1074 		/* Squash the global minimum into the permitted range. */
1075 		if (global.min_perf_pct > pct)
1076 			global.min_perf_pct = pct;
1077 	}
1078 
1079 	mutex_unlock(&intel_pstate_limits_lock);
1080 
1081 	intel_pstate_update_policies();
1082 
1083 	mutex_unlock(&intel_pstate_driver_lock);
1084 
1085 	return count;
1086 }
1087 
1088 static struct cpufreq_driver intel_pstate;
1089 
1090 static void update_qos_request(enum freq_qos_req_type type)
1091 {
1092 	int max_state, turbo_max, freq, i, perf_pct;
1093 	struct freq_qos_request *req;
1094 	struct cpufreq_policy *policy;
1095 
1096 	for_each_possible_cpu(i) {
1097 		struct cpudata *cpu = all_cpu_data[i];
1098 
1099 		policy = cpufreq_cpu_get(i);
1100 		if (!policy)
1101 			continue;
1102 
1103 		req = policy->driver_data;
1104 		cpufreq_cpu_put(policy);
1105 
1106 		if (!req)
1107 			continue;
1108 
1109 		if (hwp_active)
1110 			intel_pstate_get_hwp_max(i, &turbo_max, &max_state);
1111 		else
1112 			turbo_max = cpu->pstate.turbo_pstate;
1113 
1114 		if (type == FREQ_QOS_MIN) {
1115 			perf_pct = global.min_perf_pct;
1116 		} else {
1117 			req++;
1118 			perf_pct = global.max_perf_pct;
1119 		}
1120 
1121 		freq = DIV_ROUND_UP(turbo_max * perf_pct, 100);
1122 		freq *= cpu->pstate.scaling;
1123 
1124 		if (freq_qos_update_request(req, freq) < 0)
1125 			pr_warn("Failed to update freq constraint: CPU%d\n", i);
1126 	}
1127 }
1128 
1129 static ssize_t store_max_perf_pct(struct kobject *a, struct kobj_attribute *b,
1130 				  const char *buf, size_t count)
1131 {
1132 	unsigned int input;
1133 	int ret;
1134 
1135 	ret = sscanf(buf, "%u", &input);
1136 	if (ret != 1)
1137 		return -EINVAL;
1138 
1139 	mutex_lock(&intel_pstate_driver_lock);
1140 
1141 	if (!intel_pstate_driver) {
1142 		mutex_unlock(&intel_pstate_driver_lock);
1143 		return -EAGAIN;
1144 	}
1145 
1146 	mutex_lock(&intel_pstate_limits_lock);
1147 
1148 	global.max_perf_pct = clamp_t(int, input, global.min_perf_pct, 100);
1149 
1150 	mutex_unlock(&intel_pstate_limits_lock);
1151 
1152 	if (intel_pstate_driver == &intel_pstate)
1153 		intel_pstate_update_policies();
1154 	else
1155 		update_qos_request(FREQ_QOS_MAX);
1156 
1157 	mutex_unlock(&intel_pstate_driver_lock);
1158 
1159 	return count;
1160 }
1161 
1162 static ssize_t store_min_perf_pct(struct kobject *a, struct kobj_attribute *b,
1163 				  const char *buf, size_t count)
1164 {
1165 	unsigned int input;
1166 	int ret;
1167 
1168 	ret = sscanf(buf, "%u", &input);
1169 	if (ret != 1)
1170 		return -EINVAL;
1171 
1172 	mutex_lock(&intel_pstate_driver_lock);
1173 
1174 	if (!intel_pstate_driver) {
1175 		mutex_unlock(&intel_pstate_driver_lock);
1176 		return -EAGAIN;
1177 	}
1178 
1179 	mutex_lock(&intel_pstate_limits_lock);
1180 
1181 	global.min_perf_pct = clamp_t(int, input,
1182 				      min_perf_pct_min(), global.max_perf_pct);
1183 
1184 	mutex_unlock(&intel_pstate_limits_lock);
1185 
1186 	if (intel_pstate_driver == &intel_pstate)
1187 		intel_pstate_update_policies();
1188 	else
1189 		update_qos_request(FREQ_QOS_MIN);
1190 
1191 	mutex_unlock(&intel_pstate_driver_lock);
1192 
1193 	return count;
1194 }
1195 
1196 static ssize_t show_hwp_dynamic_boost(struct kobject *kobj,
1197 				struct kobj_attribute *attr, char *buf)
1198 {
1199 	return sprintf(buf, "%u\n", hwp_boost);
1200 }
1201 
1202 static ssize_t store_hwp_dynamic_boost(struct kobject *a,
1203 				       struct kobj_attribute *b,
1204 				       const char *buf, size_t count)
1205 {
1206 	unsigned int input;
1207 	int ret;
1208 
1209 	ret = kstrtouint(buf, 10, &input);
1210 	if (ret)
1211 		return ret;
1212 
1213 	mutex_lock(&intel_pstate_driver_lock);
1214 	hwp_boost = !!input;
1215 	intel_pstate_update_policies();
1216 	mutex_unlock(&intel_pstate_driver_lock);
1217 
1218 	return count;
1219 }
1220 
1221 show_one(max_perf_pct, max_perf_pct);
1222 show_one(min_perf_pct, min_perf_pct);
1223 
1224 define_one_global_rw(status);
1225 define_one_global_rw(no_turbo);
1226 define_one_global_rw(max_perf_pct);
1227 define_one_global_rw(min_perf_pct);
1228 define_one_global_ro(turbo_pct);
1229 define_one_global_ro(num_pstates);
1230 define_one_global_rw(hwp_dynamic_boost);
1231 
1232 static struct attribute *intel_pstate_attributes[] = {
1233 	&status.attr,
1234 	&no_turbo.attr,
1235 	&turbo_pct.attr,
1236 	&num_pstates.attr,
1237 	NULL
1238 };
1239 
1240 static const struct attribute_group intel_pstate_attr_group = {
1241 	.attrs = intel_pstate_attributes,
1242 };
1243 
1244 static void __init intel_pstate_sysfs_expose_params(void)
1245 {
1246 	struct kobject *intel_pstate_kobject;
1247 	int rc;
1248 
1249 	intel_pstate_kobject = kobject_create_and_add("intel_pstate",
1250 						&cpu_subsys.dev_root->kobj);
1251 	if (WARN_ON(!intel_pstate_kobject))
1252 		return;
1253 
1254 	rc = sysfs_create_group(intel_pstate_kobject, &intel_pstate_attr_group);
1255 	if (WARN_ON(rc))
1256 		return;
1257 
1258 	/*
1259 	 * If per cpu limits are enforced there are no global limits, so
1260 	 * return without creating max/min_perf_pct attributes
1261 	 */
1262 	if (per_cpu_limits)
1263 		return;
1264 
1265 	rc = sysfs_create_file(intel_pstate_kobject, &max_perf_pct.attr);
1266 	WARN_ON(rc);
1267 
1268 	rc = sysfs_create_file(intel_pstate_kobject, &min_perf_pct.attr);
1269 	WARN_ON(rc);
1270 
1271 	if (hwp_active) {
1272 		rc = sysfs_create_file(intel_pstate_kobject,
1273 				       &hwp_dynamic_boost.attr);
1274 		WARN_ON(rc);
1275 	}
1276 }
1277 /************************** sysfs end ************************/
1278 
1279 static void intel_pstate_hwp_enable(struct cpudata *cpudata)
1280 {
1281 	/* First disable HWP notification interrupt as we don't process them */
1282 	if (boot_cpu_has(X86_FEATURE_HWP_NOTIFY))
1283 		wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_INTERRUPT, 0x00);
1284 
1285 	wrmsrl_on_cpu(cpudata->cpu, MSR_PM_ENABLE, 0x1);
1286 	cpudata->epp_policy = 0;
1287 	if (cpudata->epp_default == -EINVAL)
1288 		cpudata->epp_default = intel_pstate_get_epp(cpudata, 0);
1289 }
1290 
1291 #define MSR_IA32_POWER_CTL_BIT_EE	19
1292 
1293 /* Disable energy efficiency optimization */
1294 static void intel_pstate_disable_ee(int cpu)
1295 {
1296 	u64 power_ctl;
1297 	int ret;
1298 
1299 	ret = rdmsrl_on_cpu(cpu, MSR_IA32_POWER_CTL, &power_ctl);
1300 	if (ret)
1301 		return;
1302 
1303 	if (!(power_ctl & BIT(MSR_IA32_POWER_CTL_BIT_EE))) {
1304 		pr_info("Disabling energy efficiency optimization\n");
1305 		power_ctl |= BIT(MSR_IA32_POWER_CTL_BIT_EE);
1306 		wrmsrl_on_cpu(cpu, MSR_IA32_POWER_CTL, power_ctl);
1307 	}
1308 }
1309 
1310 static int atom_get_min_pstate(void)
1311 {
1312 	u64 value;
1313 
1314 	rdmsrl(MSR_ATOM_CORE_RATIOS, value);
1315 	return (value >> 8) & 0x7F;
1316 }
1317 
1318 static int atom_get_max_pstate(void)
1319 {
1320 	u64 value;
1321 
1322 	rdmsrl(MSR_ATOM_CORE_RATIOS, value);
1323 	return (value >> 16) & 0x7F;
1324 }
1325 
1326 static int atom_get_turbo_pstate(void)
1327 {
1328 	u64 value;
1329 
1330 	rdmsrl(MSR_ATOM_CORE_TURBO_RATIOS, value);
1331 	return value & 0x7F;
1332 }
1333 
1334 static u64 atom_get_val(struct cpudata *cpudata, int pstate)
1335 {
1336 	u64 val;
1337 	int32_t vid_fp;
1338 	u32 vid;
1339 
1340 	val = (u64)pstate << 8;
1341 	if (global.no_turbo && !global.turbo_disabled)
1342 		val |= (u64)1 << 32;
1343 
1344 	vid_fp = cpudata->vid.min + mul_fp(
1345 		int_tofp(pstate - cpudata->pstate.min_pstate),
1346 		cpudata->vid.ratio);
1347 
1348 	vid_fp = clamp_t(int32_t, vid_fp, cpudata->vid.min, cpudata->vid.max);
1349 	vid = ceiling_fp(vid_fp);
1350 
1351 	if (pstate > cpudata->pstate.max_pstate)
1352 		vid = cpudata->vid.turbo;
1353 
1354 	return val | vid;
1355 }
1356 
1357 static int silvermont_get_scaling(void)
1358 {
1359 	u64 value;
1360 	int i;
1361 	/* Defined in Table 35-6 from SDM (Sept 2015) */
1362 	static int silvermont_freq_table[] = {
1363 		83300, 100000, 133300, 116700, 80000};
1364 
1365 	rdmsrl(MSR_FSB_FREQ, value);
1366 	i = value & 0x7;
1367 	WARN_ON(i > 4);
1368 
1369 	return silvermont_freq_table[i];
1370 }
1371 
1372 static int airmont_get_scaling(void)
1373 {
1374 	u64 value;
1375 	int i;
1376 	/* Defined in Table 35-10 from SDM (Sept 2015) */
1377 	static int airmont_freq_table[] = {
1378 		83300, 100000, 133300, 116700, 80000,
1379 		93300, 90000, 88900, 87500};
1380 
1381 	rdmsrl(MSR_FSB_FREQ, value);
1382 	i = value & 0xF;
1383 	WARN_ON(i > 8);
1384 
1385 	return airmont_freq_table[i];
1386 }
1387 
1388 static void atom_get_vid(struct cpudata *cpudata)
1389 {
1390 	u64 value;
1391 
1392 	rdmsrl(MSR_ATOM_CORE_VIDS, value);
1393 	cpudata->vid.min = int_tofp((value >> 8) & 0x7f);
1394 	cpudata->vid.max = int_tofp((value >> 16) & 0x7f);
1395 	cpudata->vid.ratio = div_fp(
1396 		cpudata->vid.max - cpudata->vid.min,
1397 		int_tofp(cpudata->pstate.max_pstate -
1398 			cpudata->pstate.min_pstate));
1399 
1400 	rdmsrl(MSR_ATOM_CORE_TURBO_VIDS, value);
1401 	cpudata->vid.turbo = value & 0x7f;
1402 }
1403 
1404 static int core_get_min_pstate(void)
1405 {
1406 	u64 value;
1407 
1408 	rdmsrl(MSR_PLATFORM_INFO, value);
1409 	return (value >> 40) & 0xFF;
1410 }
1411 
1412 static int core_get_max_pstate_physical(void)
1413 {
1414 	u64 value;
1415 
1416 	rdmsrl(MSR_PLATFORM_INFO, value);
1417 	return (value >> 8) & 0xFF;
1418 }
1419 
1420 static int core_get_tdp_ratio(u64 plat_info)
1421 {
1422 	/* Check how many TDP levels present */
1423 	if (plat_info & 0x600000000) {
1424 		u64 tdp_ctrl;
1425 		u64 tdp_ratio;
1426 		int tdp_msr;
1427 		int err;
1428 
1429 		/* Get the TDP level (0, 1, 2) to get ratios */
1430 		err = rdmsrl_safe(MSR_CONFIG_TDP_CONTROL, &tdp_ctrl);
1431 		if (err)
1432 			return err;
1433 
1434 		/* TDP MSR are continuous starting at 0x648 */
1435 		tdp_msr = MSR_CONFIG_TDP_NOMINAL + (tdp_ctrl & 0x03);
1436 		err = rdmsrl_safe(tdp_msr, &tdp_ratio);
1437 		if (err)
1438 			return err;
1439 
1440 		/* For level 1 and 2, bits[23:16] contain the ratio */
1441 		if (tdp_ctrl & 0x03)
1442 			tdp_ratio >>= 16;
1443 
1444 		tdp_ratio &= 0xff; /* ratios are only 8 bits long */
1445 		pr_debug("tdp_ratio %x\n", (int)tdp_ratio);
1446 
1447 		return (int)tdp_ratio;
1448 	}
1449 
1450 	return -ENXIO;
1451 }
1452 
1453 static int core_get_max_pstate(void)
1454 {
1455 	u64 tar;
1456 	u64 plat_info;
1457 	int max_pstate;
1458 	int tdp_ratio;
1459 	int err;
1460 
1461 	rdmsrl(MSR_PLATFORM_INFO, plat_info);
1462 	max_pstate = (plat_info >> 8) & 0xFF;
1463 
1464 	tdp_ratio = core_get_tdp_ratio(plat_info);
1465 	if (tdp_ratio <= 0)
1466 		return max_pstate;
1467 
1468 	if (hwp_active) {
1469 		/* Turbo activation ratio is not used on HWP platforms */
1470 		return tdp_ratio;
1471 	}
1472 
1473 	err = rdmsrl_safe(MSR_TURBO_ACTIVATION_RATIO, &tar);
1474 	if (!err) {
1475 		int tar_levels;
1476 
1477 		/* Do some sanity checking for safety */
1478 		tar_levels = tar & 0xff;
1479 		if (tdp_ratio - 1 == tar_levels) {
1480 			max_pstate = tar_levels;
1481 			pr_debug("max_pstate=TAC %x\n", max_pstate);
1482 		}
1483 	}
1484 
1485 	return max_pstate;
1486 }
1487 
1488 static int core_get_turbo_pstate(void)
1489 {
1490 	u64 value;
1491 	int nont, ret;
1492 
1493 	rdmsrl(MSR_TURBO_RATIO_LIMIT, value);
1494 	nont = core_get_max_pstate();
1495 	ret = (value) & 255;
1496 	if (ret <= nont)
1497 		ret = nont;
1498 	return ret;
1499 }
1500 
1501 static inline int core_get_scaling(void)
1502 {
1503 	return 100000;
1504 }
1505 
1506 static u64 core_get_val(struct cpudata *cpudata, int pstate)
1507 {
1508 	u64 val;
1509 
1510 	val = (u64)pstate << 8;
1511 	if (global.no_turbo && !global.turbo_disabled)
1512 		val |= (u64)1 << 32;
1513 
1514 	return val;
1515 }
1516 
1517 static int knl_get_aperf_mperf_shift(void)
1518 {
1519 	return 10;
1520 }
1521 
1522 static int knl_get_turbo_pstate(void)
1523 {
1524 	u64 value;
1525 	int nont, ret;
1526 
1527 	rdmsrl(MSR_TURBO_RATIO_LIMIT, value);
1528 	nont = core_get_max_pstate();
1529 	ret = (((value) >> 8) & 0xFF);
1530 	if (ret <= nont)
1531 		ret = nont;
1532 	return ret;
1533 }
1534 
1535 static void intel_pstate_set_pstate(struct cpudata *cpu, int pstate)
1536 {
1537 	trace_cpu_frequency(pstate * cpu->pstate.scaling, cpu->cpu);
1538 	cpu->pstate.current_pstate = pstate;
1539 	/*
1540 	 * Generally, there is no guarantee that this code will always run on
1541 	 * the CPU being updated, so force the register update to run on the
1542 	 * right CPU.
1543 	 */
1544 	wrmsrl_on_cpu(cpu->cpu, MSR_IA32_PERF_CTL,
1545 		      pstate_funcs.get_val(cpu, pstate));
1546 }
1547 
1548 static void intel_pstate_set_min_pstate(struct cpudata *cpu)
1549 {
1550 	intel_pstate_set_pstate(cpu, cpu->pstate.min_pstate);
1551 }
1552 
1553 static void intel_pstate_max_within_limits(struct cpudata *cpu)
1554 {
1555 	int pstate = max(cpu->pstate.min_pstate, cpu->max_perf_ratio);
1556 
1557 	update_turbo_state();
1558 	intel_pstate_set_pstate(cpu, pstate);
1559 }
1560 
1561 static void intel_pstate_get_cpu_pstates(struct cpudata *cpu)
1562 {
1563 	cpu->pstate.min_pstate = pstate_funcs.get_min();
1564 	cpu->pstate.max_pstate = pstate_funcs.get_max();
1565 	cpu->pstate.max_pstate_physical = pstate_funcs.get_max_physical();
1566 	cpu->pstate.turbo_pstate = pstate_funcs.get_turbo();
1567 	cpu->pstate.scaling = pstate_funcs.get_scaling();
1568 	cpu->pstate.max_freq = cpu->pstate.max_pstate * cpu->pstate.scaling;
1569 
1570 	if (hwp_active && !hwp_mode_bdw) {
1571 		unsigned int phy_max, current_max;
1572 
1573 		intel_pstate_get_hwp_max(cpu->cpu, &phy_max, &current_max);
1574 		cpu->pstate.turbo_freq = phy_max * cpu->pstate.scaling;
1575 	} else {
1576 		cpu->pstate.turbo_freq = cpu->pstate.turbo_pstate * cpu->pstate.scaling;
1577 	}
1578 
1579 	if (pstate_funcs.get_aperf_mperf_shift)
1580 		cpu->aperf_mperf_shift = pstate_funcs.get_aperf_mperf_shift();
1581 
1582 	if (pstate_funcs.get_vid)
1583 		pstate_funcs.get_vid(cpu);
1584 
1585 	intel_pstate_set_min_pstate(cpu);
1586 }
1587 
1588 /*
1589  * Long hold time will keep high perf limits for long time,
1590  * which negatively impacts perf/watt for some workloads,
1591  * like specpower. 3ms is based on experiements on some
1592  * workoads.
1593  */
1594 static int hwp_boost_hold_time_ns = 3 * NSEC_PER_MSEC;
1595 
1596 static inline void intel_pstate_hwp_boost_up(struct cpudata *cpu)
1597 {
1598 	u64 hwp_req = READ_ONCE(cpu->hwp_req_cached);
1599 	u32 max_limit = (hwp_req & 0xff00) >> 8;
1600 	u32 min_limit = (hwp_req & 0xff);
1601 	u32 boost_level1;
1602 
1603 	/*
1604 	 * Cases to consider (User changes via sysfs or boot time):
1605 	 * If, P0 (Turbo max) = P1 (Guaranteed max) = min:
1606 	 *	No boost, return.
1607 	 * If, P0 (Turbo max) > P1 (Guaranteed max) = min:
1608 	 *     Should result in one level boost only for P0.
1609 	 * If, P0 (Turbo max) = P1 (Guaranteed max) > min:
1610 	 *     Should result in two level boost:
1611 	 *         (min + p1)/2 and P1.
1612 	 * If, P0 (Turbo max) > P1 (Guaranteed max) > min:
1613 	 *     Should result in three level boost:
1614 	 *        (min + p1)/2, P1 and P0.
1615 	 */
1616 
1617 	/* If max and min are equal or already at max, nothing to boost */
1618 	if (max_limit == min_limit || cpu->hwp_boost_min >= max_limit)
1619 		return;
1620 
1621 	if (!cpu->hwp_boost_min)
1622 		cpu->hwp_boost_min = min_limit;
1623 
1624 	/* level at half way mark between min and guranteed */
1625 	boost_level1 = (HWP_GUARANTEED_PERF(cpu->hwp_cap_cached) + min_limit) >> 1;
1626 
1627 	if (cpu->hwp_boost_min < boost_level1)
1628 		cpu->hwp_boost_min = boost_level1;
1629 	else if (cpu->hwp_boost_min < HWP_GUARANTEED_PERF(cpu->hwp_cap_cached))
1630 		cpu->hwp_boost_min = HWP_GUARANTEED_PERF(cpu->hwp_cap_cached);
1631 	else if (cpu->hwp_boost_min == HWP_GUARANTEED_PERF(cpu->hwp_cap_cached) &&
1632 		 max_limit != HWP_GUARANTEED_PERF(cpu->hwp_cap_cached))
1633 		cpu->hwp_boost_min = max_limit;
1634 	else
1635 		return;
1636 
1637 	hwp_req = (hwp_req & ~GENMASK_ULL(7, 0)) | cpu->hwp_boost_min;
1638 	wrmsrl(MSR_HWP_REQUEST, hwp_req);
1639 	cpu->last_update = cpu->sample.time;
1640 }
1641 
1642 static inline void intel_pstate_hwp_boost_down(struct cpudata *cpu)
1643 {
1644 	if (cpu->hwp_boost_min) {
1645 		bool expired;
1646 
1647 		/* Check if we are idle for hold time to boost down */
1648 		expired = time_after64(cpu->sample.time, cpu->last_update +
1649 				       hwp_boost_hold_time_ns);
1650 		if (expired) {
1651 			wrmsrl(MSR_HWP_REQUEST, cpu->hwp_req_cached);
1652 			cpu->hwp_boost_min = 0;
1653 		}
1654 	}
1655 	cpu->last_update = cpu->sample.time;
1656 }
1657 
1658 static inline void intel_pstate_update_util_hwp_local(struct cpudata *cpu,
1659 						      u64 time)
1660 {
1661 	cpu->sample.time = time;
1662 
1663 	if (cpu->sched_flags & SCHED_CPUFREQ_IOWAIT) {
1664 		bool do_io = false;
1665 
1666 		cpu->sched_flags = 0;
1667 		/*
1668 		 * Set iowait_boost flag and update time. Since IO WAIT flag
1669 		 * is set all the time, we can't just conclude that there is
1670 		 * some IO bound activity is scheduled on this CPU with just
1671 		 * one occurrence. If we receive at least two in two
1672 		 * consecutive ticks, then we treat as boost candidate.
1673 		 */
1674 		if (time_before64(time, cpu->last_io_update + 2 * TICK_NSEC))
1675 			do_io = true;
1676 
1677 		cpu->last_io_update = time;
1678 
1679 		if (do_io)
1680 			intel_pstate_hwp_boost_up(cpu);
1681 
1682 	} else {
1683 		intel_pstate_hwp_boost_down(cpu);
1684 	}
1685 }
1686 
1687 static inline void intel_pstate_update_util_hwp(struct update_util_data *data,
1688 						u64 time, unsigned int flags)
1689 {
1690 	struct cpudata *cpu = container_of(data, struct cpudata, update_util);
1691 
1692 	cpu->sched_flags |= flags;
1693 
1694 	if (smp_processor_id() == cpu->cpu)
1695 		intel_pstate_update_util_hwp_local(cpu, time);
1696 }
1697 
1698 static inline void intel_pstate_calc_avg_perf(struct cpudata *cpu)
1699 {
1700 	struct sample *sample = &cpu->sample;
1701 
1702 	sample->core_avg_perf = div_ext_fp(sample->aperf, sample->mperf);
1703 }
1704 
1705 static inline bool intel_pstate_sample(struct cpudata *cpu, u64 time)
1706 {
1707 	u64 aperf, mperf;
1708 	unsigned long flags;
1709 	u64 tsc;
1710 
1711 	local_irq_save(flags);
1712 	rdmsrl(MSR_IA32_APERF, aperf);
1713 	rdmsrl(MSR_IA32_MPERF, mperf);
1714 	tsc = rdtsc();
1715 	if (cpu->prev_mperf == mperf || cpu->prev_tsc == tsc) {
1716 		local_irq_restore(flags);
1717 		return false;
1718 	}
1719 	local_irq_restore(flags);
1720 
1721 	cpu->last_sample_time = cpu->sample.time;
1722 	cpu->sample.time = time;
1723 	cpu->sample.aperf = aperf;
1724 	cpu->sample.mperf = mperf;
1725 	cpu->sample.tsc =  tsc;
1726 	cpu->sample.aperf -= cpu->prev_aperf;
1727 	cpu->sample.mperf -= cpu->prev_mperf;
1728 	cpu->sample.tsc -= cpu->prev_tsc;
1729 
1730 	cpu->prev_aperf = aperf;
1731 	cpu->prev_mperf = mperf;
1732 	cpu->prev_tsc = tsc;
1733 	/*
1734 	 * First time this function is invoked in a given cycle, all of the
1735 	 * previous sample data fields are equal to zero or stale and they must
1736 	 * be populated with meaningful numbers for things to work, so assume
1737 	 * that sample.time will always be reset before setting the utilization
1738 	 * update hook and make the caller skip the sample then.
1739 	 */
1740 	if (cpu->last_sample_time) {
1741 		intel_pstate_calc_avg_perf(cpu);
1742 		return true;
1743 	}
1744 	return false;
1745 }
1746 
1747 static inline int32_t get_avg_frequency(struct cpudata *cpu)
1748 {
1749 	return mul_ext_fp(cpu->sample.core_avg_perf, cpu_khz);
1750 }
1751 
1752 static inline int32_t get_avg_pstate(struct cpudata *cpu)
1753 {
1754 	return mul_ext_fp(cpu->pstate.max_pstate_physical,
1755 			  cpu->sample.core_avg_perf);
1756 }
1757 
1758 static inline int32_t get_target_pstate(struct cpudata *cpu)
1759 {
1760 	struct sample *sample = &cpu->sample;
1761 	int32_t busy_frac;
1762 	int target, avg_pstate;
1763 
1764 	busy_frac = div_fp(sample->mperf << cpu->aperf_mperf_shift,
1765 			   sample->tsc);
1766 
1767 	if (busy_frac < cpu->iowait_boost)
1768 		busy_frac = cpu->iowait_boost;
1769 
1770 	sample->busy_scaled = busy_frac * 100;
1771 
1772 	target = global.no_turbo || global.turbo_disabled ?
1773 			cpu->pstate.max_pstate : cpu->pstate.turbo_pstate;
1774 	target += target >> 2;
1775 	target = mul_fp(target, busy_frac);
1776 	if (target < cpu->pstate.min_pstate)
1777 		target = cpu->pstate.min_pstate;
1778 
1779 	/*
1780 	 * If the average P-state during the previous cycle was higher than the
1781 	 * current target, add 50% of the difference to the target to reduce
1782 	 * possible performance oscillations and offset possible performance
1783 	 * loss related to moving the workload from one CPU to another within
1784 	 * a package/module.
1785 	 */
1786 	avg_pstate = get_avg_pstate(cpu);
1787 	if (avg_pstate > target)
1788 		target += (avg_pstate - target) >> 1;
1789 
1790 	return target;
1791 }
1792 
1793 static int intel_pstate_prepare_request(struct cpudata *cpu, int pstate)
1794 {
1795 	int min_pstate = max(cpu->pstate.min_pstate, cpu->min_perf_ratio);
1796 	int max_pstate = max(min_pstate, cpu->max_perf_ratio);
1797 
1798 	return clamp_t(int, pstate, min_pstate, max_pstate);
1799 }
1800 
1801 static void intel_pstate_update_pstate(struct cpudata *cpu, int pstate)
1802 {
1803 	if (pstate == cpu->pstate.current_pstate)
1804 		return;
1805 
1806 	cpu->pstate.current_pstate = pstate;
1807 	wrmsrl(MSR_IA32_PERF_CTL, pstate_funcs.get_val(cpu, pstate));
1808 }
1809 
1810 static void intel_pstate_adjust_pstate(struct cpudata *cpu)
1811 {
1812 	int from = cpu->pstate.current_pstate;
1813 	struct sample *sample;
1814 	int target_pstate;
1815 
1816 	update_turbo_state();
1817 
1818 	target_pstate = get_target_pstate(cpu);
1819 	target_pstate = intel_pstate_prepare_request(cpu, target_pstate);
1820 	trace_cpu_frequency(target_pstate * cpu->pstate.scaling, cpu->cpu);
1821 	intel_pstate_update_pstate(cpu, target_pstate);
1822 
1823 	sample = &cpu->sample;
1824 	trace_pstate_sample(mul_ext_fp(100, sample->core_avg_perf),
1825 		fp_toint(sample->busy_scaled),
1826 		from,
1827 		cpu->pstate.current_pstate,
1828 		sample->mperf,
1829 		sample->aperf,
1830 		sample->tsc,
1831 		get_avg_frequency(cpu),
1832 		fp_toint(cpu->iowait_boost * 100));
1833 }
1834 
1835 static void intel_pstate_update_util(struct update_util_data *data, u64 time,
1836 				     unsigned int flags)
1837 {
1838 	struct cpudata *cpu = container_of(data, struct cpudata, update_util);
1839 	u64 delta_ns;
1840 
1841 	/* Don't allow remote callbacks */
1842 	if (smp_processor_id() != cpu->cpu)
1843 		return;
1844 
1845 	delta_ns = time - cpu->last_update;
1846 	if (flags & SCHED_CPUFREQ_IOWAIT) {
1847 		/* Start over if the CPU may have been idle. */
1848 		if (delta_ns > TICK_NSEC) {
1849 			cpu->iowait_boost = ONE_EIGHTH_FP;
1850 		} else if (cpu->iowait_boost >= ONE_EIGHTH_FP) {
1851 			cpu->iowait_boost <<= 1;
1852 			if (cpu->iowait_boost > int_tofp(1))
1853 				cpu->iowait_boost = int_tofp(1);
1854 		} else {
1855 			cpu->iowait_boost = ONE_EIGHTH_FP;
1856 		}
1857 	} else if (cpu->iowait_boost) {
1858 		/* Clear iowait_boost if the CPU may have been idle. */
1859 		if (delta_ns > TICK_NSEC)
1860 			cpu->iowait_boost = 0;
1861 		else
1862 			cpu->iowait_boost >>= 1;
1863 	}
1864 	cpu->last_update = time;
1865 	delta_ns = time - cpu->sample.time;
1866 	if ((s64)delta_ns < INTEL_PSTATE_SAMPLING_INTERVAL)
1867 		return;
1868 
1869 	if (intel_pstate_sample(cpu, time))
1870 		intel_pstate_adjust_pstate(cpu);
1871 }
1872 
1873 static struct pstate_funcs core_funcs = {
1874 	.get_max = core_get_max_pstate,
1875 	.get_max_physical = core_get_max_pstate_physical,
1876 	.get_min = core_get_min_pstate,
1877 	.get_turbo = core_get_turbo_pstate,
1878 	.get_scaling = core_get_scaling,
1879 	.get_val = core_get_val,
1880 };
1881 
1882 static const struct pstate_funcs silvermont_funcs = {
1883 	.get_max = atom_get_max_pstate,
1884 	.get_max_physical = atom_get_max_pstate,
1885 	.get_min = atom_get_min_pstate,
1886 	.get_turbo = atom_get_turbo_pstate,
1887 	.get_val = atom_get_val,
1888 	.get_scaling = silvermont_get_scaling,
1889 	.get_vid = atom_get_vid,
1890 };
1891 
1892 static const struct pstate_funcs airmont_funcs = {
1893 	.get_max = atom_get_max_pstate,
1894 	.get_max_physical = atom_get_max_pstate,
1895 	.get_min = atom_get_min_pstate,
1896 	.get_turbo = atom_get_turbo_pstate,
1897 	.get_val = atom_get_val,
1898 	.get_scaling = airmont_get_scaling,
1899 	.get_vid = atom_get_vid,
1900 };
1901 
1902 static const struct pstate_funcs knl_funcs = {
1903 	.get_max = core_get_max_pstate,
1904 	.get_max_physical = core_get_max_pstate_physical,
1905 	.get_min = core_get_min_pstate,
1906 	.get_turbo = knl_get_turbo_pstate,
1907 	.get_aperf_mperf_shift = knl_get_aperf_mperf_shift,
1908 	.get_scaling = core_get_scaling,
1909 	.get_val = core_get_val,
1910 };
1911 
1912 #define X86_MATCH(model, policy)					 \
1913 	X86_MATCH_VENDOR_FAM_MODEL_FEATURE(INTEL, 6, INTEL_FAM6_##model, \
1914 					   X86_FEATURE_APERFMPERF, &policy)
1915 
1916 static const struct x86_cpu_id intel_pstate_cpu_ids[] = {
1917 	X86_MATCH(SANDYBRIDGE,		core_funcs),
1918 	X86_MATCH(SANDYBRIDGE_X,	core_funcs),
1919 	X86_MATCH(ATOM_SILVERMONT,	silvermont_funcs),
1920 	X86_MATCH(IVYBRIDGE,		core_funcs),
1921 	X86_MATCH(HASWELL,		core_funcs),
1922 	X86_MATCH(BROADWELL,		core_funcs),
1923 	X86_MATCH(IVYBRIDGE_X,		core_funcs),
1924 	X86_MATCH(HASWELL_X,		core_funcs),
1925 	X86_MATCH(HASWELL_L,		core_funcs),
1926 	X86_MATCH(HASWELL_G,		core_funcs),
1927 	X86_MATCH(BROADWELL_G,		core_funcs),
1928 	X86_MATCH(ATOM_AIRMONT,		airmont_funcs),
1929 	X86_MATCH(SKYLAKE_L,		core_funcs),
1930 	X86_MATCH(BROADWELL_X,		core_funcs),
1931 	X86_MATCH(SKYLAKE,		core_funcs),
1932 	X86_MATCH(BROADWELL_D,		core_funcs),
1933 	X86_MATCH(XEON_PHI_KNL,		knl_funcs),
1934 	X86_MATCH(XEON_PHI_KNM,		knl_funcs),
1935 	X86_MATCH(ATOM_GOLDMONT,	core_funcs),
1936 	X86_MATCH(ATOM_GOLDMONT_PLUS,	core_funcs),
1937 	X86_MATCH(SKYLAKE_X,		core_funcs),
1938 	{}
1939 };
1940 MODULE_DEVICE_TABLE(x86cpu, intel_pstate_cpu_ids);
1941 
1942 static const struct x86_cpu_id intel_pstate_cpu_oob_ids[] __initconst = {
1943 	X86_MATCH(BROADWELL_D,		core_funcs),
1944 	X86_MATCH(BROADWELL_X,		core_funcs),
1945 	X86_MATCH(SKYLAKE_X,		core_funcs),
1946 	{}
1947 };
1948 
1949 static const struct x86_cpu_id intel_pstate_cpu_ee_disable_ids[] = {
1950 	X86_MATCH(KABYLAKE,		core_funcs),
1951 	{}
1952 };
1953 
1954 static const struct x86_cpu_id intel_pstate_hwp_boost_ids[] = {
1955 	X86_MATCH(SKYLAKE_X,		core_funcs),
1956 	X86_MATCH(SKYLAKE,		core_funcs),
1957 	{}
1958 };
1959 
1960 static int intel_pstate_init_cpu(unsigned int cpunum)
1961 {
1962 	struct cpudata *cpu;
1963 
1964 	cpu = all_cpu_data[cpunum];
1965 
1966 	if (!cpu) {
1967 		cpu = kzalloc(sizeof(*cpu), GFP_KERNEL);
1968 		if (!cpu)
1969 			return -ENOMEM;
1970 
1971 		all_cpu_data[cpunum] = cpu;
1972 
1973 		cpu->epp_default = -EINVAL;
1974 		cpu->epp_powersave = -EINVAL;
1975 		cpu->epp_saved = -EINVAL;
1976 	}
1977 
1978 	cpu = all_cpu_data[cpunum];
1979 
1980 	cpu->cpu = cpunum;
1981 
1982 	if (hwp_active) {
1983 		const struct x86_cpu_id *id;
1984 
1985 		id = x86_match_cpu(intel_pstate_cpu_ee_disable_ids);
1986 		if (id)
1987 			intel_pstate_disable_ee(cpunum);
1988 
1989 		intel_pstate_hwp_enable(cpu);
1990 
1991 		id = x86_match_cpu(intel_pstate_hwp_boost_ids);
1992 		if (id && intel_pstate_acpi_pm_profile_server())
1993 			hwp_boost = true;
1994 	}
1995 
1996 	intel_pstate_get_cpu_pstates(cpu);
1997 
1998 	pr_debug("controlling: cpu %d\n", cpunum);
1999 
2000 	return 0;
2001 }
2002 
2003 static void intel_pstate_set_update_util_hook(unsigned int cpu_num)
2004 {
2005 	struct cpudata *cpu = all_cpu_data[cpu_num];
2006 
2007 	if (hwp_active && !hwp_boost)
2008 		return;
2009 
2010 	if (cpu->update_util_set)
2011 		return;
2012 
2013 	/* Prevent intel_pstate_update_util() from using stale data. */
2014 	cpu->sample.time = 0;
2015 	cpufreq_add_update_util_hook(cpu_num, &cpu->update_util,
2016 				     (hwp_active ?
2017 				      intel_pstate_update_util_hwp :
2018 				      intel_pstate_update_util));
2019 	cpu->update_util_set = true;
2020 }
2021 
2022 static void intel_pstate_clear_update_util_hook(unsigned int cpu)
2023 {
2024 	struct cpudata *cpu_data = all_cpu_data[cpu];
2025 
2026 	if (!cpu_data->update_util_set)
2027 		return;
2028 
2029 	cpufreq_remove_update_util_hook(cpu);
2030 	cpu_data->update_util_set = false;
2031 	synchronize_rcu();
2032 }
2033 
2034 static int intel_pstate_get_max_freq(struct cpudata *cpu)
2035 {
2036 	return global.turbo_disabled || global.no_turbo ?
2037 			cpu->pstate.max_freq : cpu->pstate.turbo_freq;
2038 }
2039 
2040 static void intel_pstate_update_perf_limits(struct cpudata *cpu,
2041 					    unsigned int policy_min,
2042 					    unsigned int policy_max)
2043 {
2044 	int max_freq = intel_pstate_get_max_freq(cpu);
2045 	int32_t max_policy_perf, min_policy_perf;
2046 	int max_state, turbo_max;
2047 
2048 	/*
2049 	 * HWP needs some special consideration, because on BDX the
2050 	 * HWP_REQUEST uses abstract value to represent performance
2051 	 * rather than pure ratios.
2052 	 */
2053 	if (hwp_active) {
2054 		intel_pstate_get_hwp_max(cpu->cpu, &turbo_max, &max_state);
2055 	} else {
2056 		max_state = global.no_turbo || global.turbo_disabled ?
2057 			cpu->pstate.max_pstate : cpu->pstate.turbo_pstate;
2058 		turbo_max = cpu->pstate.turbo_pstate;
2059 	}
2060 
2061 	max_policy_perf = max_state * policy_max / max_freq;
2062 	if (policy_max == policy_min) {
2063 		min_policy_perf = max_policy_perf;
2064 	} else {
2065 		min_policy_perf = max_state * policy_min / max_freq;
2066 		min_policy_perf = clamp_t(int32_t, min_policy_perf,
2067 					  0, max_policy_perf);
2068 	}
2069 
2070 	pr_debug("cpu:%d max_state %d min_policy_perf:%d max_policy_perf:%d\n",
2071 		 cpu->cpu, max_state, min_policy_perf, max_policy_perf);
2072 
2073 	/* Normalize user input to [min_perf, max_perf] */
2074 	if (per_cpu_limits) {
2075 		cpu->min_perf_ratio = min_policy_perf;
2076 		cpu->max_perf_ratio = max_policy_perf;
2077 	} else {
2078 		int32_t global_min, global_max;
2079 
2080 		/* Global limits are in percent of the maximum turbo P-state. */
2081 		global_max = DIV_ROUND_UP(turbo_max * global.max_perf_pct, 100);
2082 		global_min = DIV_ROUND_UP(turbo_max * global.min_perf_pct, 100);
2083 		global_min = clamp_t(int32_t, global_min, 0, global_max);
2084 
2085 		pr_debug("cpu:%d global_min:%d global_max:%d\n", cpu->cpu,
2086 			 global_min, global_max);
2087 
2088 		cpu->min_perf_ratio = max(min_policy_perf, global_min);
2089 		cpu->min_perf_ratio = min(cpu->min_perf_ratio, max_policy_perf);
2090 		cpu->max_perf_ratio = min(max_policy_perf, global_max);
2091 		cpu->max_perf_ratio = max(min_policy_perf, cpu->max_perf_ratio);
2092 
2093 		/* Make sure min_perf <= max_perf */
2094 		cpu->min_perf_ratio = min(cpu->min_perf_ratio,
2095 					  cpu->max_perf_ratio);
2096 
2097 	}
2098 	pr_debug("cpu:%d max_perf_ratio:%d min_perf_ratio:%d\n", cpu->cpu,
2099 		 cpu->max_perf_ratio,
2100 		 cpu->min_perf_ratio);
2101 }
2102 
2103 static int intel_pstate_set_policy(struct cpufreq_policy *policy)
2104 {
2105 	struct cpudata *cpu;
2106 
2107 	if (!policy->cpuinfo.max_freq)
2108 		return -ENODEV;
2109 
2110 	pr_debug("set_policy cpuinfo.max %u policy->max %u\n",
2111 		 policy->cpuinfo.max_freq, policy->max);
2112 
2113 	cpu = all_cpu_data[policy->cpu];
2114 	cpu->policy = policy->policy;
2115 
2116 	mutex_lock(&intel_pstate_limits_lock);
2117 
2118 	intel_pstate_update_perf_limits(cpu, policy->min, policy->max);
2119 
2120 	if (cpu->policy == CPUFREQ_POLICY_PERFORMANCE) {
2121 		/*
2122 		 * NOHZ_FULL CPUs need this as the governor callback may not
2123 		 * be invoked on them.
2124 		 */
2125 		intel_pstate_clear_update_util_hook(policy->cpu);
2126 		intel_pstate_max_within_limits(cpu);
2127 	} else {
2128 		intel_pstate_set_update_util_hook(policy->cpu);
2129 	}
2130 
2131 	if (hwp_active) {
2132 		/*
2133 		 * When hwp_boost was active before and dynamically it
2134 		 * was turned off, in that case we need to clear the
2135 		 * update util hook.
2136 		 */
2137 		if (!hwp_boost)
2138 			intel_pstate_clear_update_util_hook(policy->cpu);
2139 		intel_pstate_hwp_set(policy->cpu);
2140 	}
2141 
2142 	mutex_unlock(&intel_pstate_limits_lock);
2143 
2144 	return 0;
2145 }
2146 
2147 static void intel_pstate_adjust_policy_max(struct cpudata *cpu,
2148 					   struct cpufreq_policy_data *policy)
2149 {
2150 	if (!hwp_active &&
2151 	    cpu->pstate.max_pstate_physical > cpu->pstate.max_pstate &&
2152 	    policy->max < policy->cpuinfo.max_freq &&
2153 	    policy->max > cpu->pstate.max_freq) {
2154 		pr_debug("policy->max > max non turbo frequency\n");
2155 		policy->max = policy->cpuinfo.max_freq;
2156 	}
2157 }
2158 
2159 static void intel_pstate_verify_cpu_policy(struct cpudata *cpu,
2160 					   struct cpufreq_policy_data *policy)
2161 {
2162 	update_turbo_state();
2163 	cpufreq_verify_within_limits(policy, policy->cpuinfo.min_freq,
2164 				     intel_pstate_get_max_freq(cpu));
2165 
2166 	intel_pstate_adjust_policy_max(cpu, policy);
2167 }
2168 
2169 static int intel_pstate_verify_policy(struct cpufreq_policy_data *policy)
2170 {
2171 	intel_pstate_verify_cpu_policy(all_cpu_data[policy->cpu], policy);
2172 
2173 	return 0;
2174 }
2175 
2176 static void intel_cpufreq_stop_cpu(struct cpufreq_policy *policy)
2177 {
2178 	intel_pstate_set_min_pstate(all_cpu_data[policy->cpu]);
2179 }
2180 
2181 static void intel_pstate_stop_cpu(struct cpufreq_policy *policy)
2182 {
2183 	pr_debug("CPU %d exiting\n", policy->cpu);
2184 
2185 	intel_pstate_clear_update_util_hook(policy->cpu);
2186 	if (hwp_active) {
2187 		intel_pstate_hwp_save_state(policy);
2188 		intel_pstate_hwp_force_min_perf(policy->cpu);
2189 	} else {
2190 		intel_cpufreq_stop_cpu(policy);
2191 	}
2192 }
2193 
2194 static int intel_pstate_cpu_exit(struct cpufreq_policy *policy)
2195 {
2196 	intel_pstate_exit_perf_limits(policy);
2197 
2198 	policy->fast_switch_possible = false;
2199 
2200 	return 0;
2201 }
2202 
2203 static int __intel_pstate_cpu_init(struct cpufreq_policy *policy)
2204 {
2205 	struct cpudata *cpu;
2206 	int rc;
2207 
2208 	rc = intel_pstate_init_cpu(policy->cpu);
2209 	if (rc)
2210 		return rc;
2211 
2212 	cpu = all_cpu_data[policy->cpu];
2213 
2214 	cpu->max_perf_ratio = 0xFF;
2215 	cpu->min_perf_ratio = 0;
2216 
2217 	policy->min = cpu->pstate.min_pstate * cpu->pstate.scaling;
2218 	policy->max = cpu->pstate.turbo_pstate * cpu->pstate.scaling;
2219 
2220 	/* cpuinfo and default policy values */
2221 	policy->cpuinfo.min_freq = cpu->pstate.min_pstate * cpu->pstate.scaling;
2222 	update_turbo_state();
2223 	global.turbo_disabled_mf = global.turbo_disabled;
2224 	policy->cpuinfo.max_freq = global.turbo_disabled ?
2225 			cpu->pstate.max_pstate : cpu->pstate.turbo_pstate;
2226 	policy->cpuinfo.max_freq *= cpu->pstate.scaling;
2227 
2228 	if (hwp_active) {
2229 		unsigned int max_freq;
2230 
2231 		max_freq = global.turbo_disabled ?
2232 			cpu->pstate.max_freq : cpu->pstate.turbo_freq;
2233 		if (max_freq < policy->cpuinfo.max_freq)
2234 			policy->cpuinfo.max_freq = max_freq;
2235 	}
2236 
2237 	intel_pstate_init_acpi_perf_limits(policy);
2238 
2239 	policy->fast_switch_possible = true;
2240 
2241 	return 0;
2242 }
2243 
2244 static int intel_pstate_cpu_init(struct cpufreq_policy *policy)
2245 {
2246 	int ret = __intel_pstate_cpu_init(policy);
2247 
2248 	if (ret)
2249 		return ret;
2250 
2251 	/*
2252 	 * Set the policy to powersave to provide a valid fallback value in case
2253 	 * the default cpufreq governor is neither powersave nor performance.
2254 	 */
2255 	policy->policy = CPUFREQ_POLICY_POWERSAVE;
2256 
2257 	return 0;
2258 }
2259 
2260 static struct cpufreq_driver intel_pstate = {
2261 	.flags		= CPUFREQ_CONST_LOOPS,
2262 	.verify		= intel_pstate_verify_policy,
2263 	.setpolicy	= intel_pstate_set_policy,
2264 	.suspend	= intel_pstate_hwp_save_state,
2265 	.resume		= intel_pstate_resume,
2266 	.init		= intel_pstate_cpu_init,
2267 	.exit		= intel_pstate_cpu_exit,
2268 	.stop_cpu	= intel_pstate_stop_cpu,
2269 	.update_limits	= intel_pstate_update_limits,
2270 	.name		= "intel_pstate",
2271 };
2272 
2273 static int intel_cpufreq_verify_policy(struct cpufreq_policy_data *policy)
2274 {
2275 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2276 
2277 	intel_pstate_verify_cpu_policy(cpu, policy);
2278 	intel_pstate_update_perf_limits(cpu, policy->min, policy->max);
2279 
2280 	return 0;
2281 }
2282 
2283 /* Use of trace in passive mode:
2284  *
2285  * In passive mode the trace core_busy field (also known as the
2286  * performance field, and lablelled as such on the graphs; also known as
2287  * core_avg_perf) is not needed and so is re-assigned to indicate if the
2288  * driver call was via the normal or fast switch path. Various graphs
2289  * output from the intel_pstate_tracer.py utility that include core_busy
2290  * (or performance or core_avg_perf) have a fixed y-axis from 0 to 100%,
2291  * so we use 10 to indicate the the normal path through the driver, and
2292  * 90 to indicate the fast switch path through the driver.
2293  * The scaled_busy field is not used, and is set to 0.
2294  */
2295 
2296 #define	INTEL_PSTATE_TRACE_TARGET 10
2297 #define	INTEL_PSTATE_TRACE_FAST_SWITCH 90
2298 
2299 static void intel_cpufreq_trace(struct cpudata *cpu, unsigned int trace_type, int old_pstate)
2300 {
2301 	struct sample *sample;
2302 
2303 	if (!trace_pstate_sample_enabled())
2304 		return;
2305 
2306 	if (!intel_pstate_sample(cpu, ktime_get()))
2307 		return;
2308 
2309 	sample = &cpu->sample;
2310 	trace_pstate_sample(trace_type,
2311 		0,
2312 		old_pstate,
2313 		cpu->pstate.current_pstate,
2314 		sample->mperf,
2315 		sample->aperf,
2316 		sample->tsc,
2317 		get_avg_frequency(cpu),
2318 		fp_toint(cpu->iowait_boost * 100));
2319 }
2320 
2321 static int intel_cpufreq_target(struct cpufreq_policy *policy,
2322 				unsigned int target_freq,
2323 				unsigned int relation)
2324 {
2325 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2326 	struct cpufreq_freqs freqs;
2327 	int target_pstate, old_pstate;
2328 
2329 	update_turbo_state();
2330 
2331 	freqs.old = policy->cur;
2332 	freqs.new = target_freq;
2333 
2334 	cpufreq_freq_transition_begin(policy, &freqs);
2335 	switch (relation) {
2336 	case CPUFREQ_RELATION_L:
2337 		target_pstate = DIV_ROUND_UP(freqs.new, cpu->pstate.scaling);
2338 		break;
2339 	case CPUFREQ_RELATION_H:
2340 		target_pstate = freqs.new / cpu->pstate.scaling;
2341 		break;
2342 	default:
2343 		target_pstate = DIV_ROUND_CLOSEST(freqs.new, cpu->pstate.scaling);
2344 		break;
2345 	}
2346 	target_pstate = intel_pstate_prepare_request(cpu, target_pstate);
2347 	old_pstate = cpu->pstate.current_pstate;
2348 	if (target_pstate != cpu->pstate.current_pstate) {
2349 		cpu->pstate.current_pstate = target_pstate;
2350 		wrmsrl_on_cpu(policy->cpu, MSR_IA32_PERF_CTL,
2351 			      pstate_funcs.get_val(cpu, target_pstate));
2352 	}
2353 	freqs.new = target_pstate * cpu->pstate.scaling;
2354 	intel_cpufreq_trace(cpu, INTEL_PSTATE_TRACE_TARGET, old_pstate);
2355 	cpufreq_freq_transition_end(policy, &freqs, false);
2356 
2357 	return 0;
2358 }
2359 
2360 static unsigned int intel_cpufreq_fast_switch(struct cpufreq_policy *policy,
2361 					      unsigned int target_freq)
2362 {
2363 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2364 	int target_pstate, old_pstate;
2365 
2366 	update_turbo_state();
2367 
2368 	target_pstate = DIV_ROUND_UP(target_freq, cpu->pstate.scaling);
2369 	target_pstate = intel_pstate_prepare_request(cpu, target_pstate);
2370 	old_pstate = cpu->pstate.current_pstate;
2371 	intel_pstate_update_pstate(cpu, target_pstate);
2372 	intel_cpufreq_trace(cpu, INTEL_PSTATE_TRACE_FAST_SWITCH, old_pstate);
2373 	return target_pstate * cpu->pstate.scaling;
2374 }
2375 
2376 static int intel_cpufreq_cpu_init(struct cpufreq_policy *policy)
2377 {
2378 	int max_state, turbo_max, min_freq, max_freq, ret;
2379 	struct freq_qos_request *req;
2380 	struct cpudata *cpu;
2381 	struct device *dev;
2382 
2383 	dev = get_cpu_device(policy->cpu);
2384 	if (!dev)
2385 		return -ENODEV;
2386 
2387 	ret = __intel_pstate_cpu_init(policy);
2388 	if (ret)
2389 		return ret;
2390 
2391 	policy->cpuinfo.transition_latency = INTEL_CPUFREQ_TRANSITION_LATENCY;
2392 	policy->transition_delay_us = INTEL_CPUFREQ_TRANSITION_DELAY;
2393 	/* This reflects the intel_pstate_get_cpu_pstates() setting. */
2394 	policy->cur = policy->cpuinfo.min_freq;
2395 
2396 	req = kcalloc(2, sizeof(*req), GFP_KERNEL);
2397 	if (!req) {
2398 		ret = -ENOMEM;
2399 		goto pstate_exit;
2400 	}
2401 
2402 	cpu = all_cpu_data[policy->cpu];
2403 
2404 	if (hwp_active)
2405 		intel_pstate_get_hwp_max(policy->cpu, &turbo_max, &max_state);
2406 	else
2407 		turbo_max = cpu->pstate.turbo_pstate;
2408 
2409 	min_freq = DIV_ROUND_UP(turbo_max * global.min_perf_pct, 100);
2410 	min_freq *= cpu->pstate.scaling;
2411 	max_freq = DIV_ROUND_UP(turbo_max * global.max_perf_pct, 100);
2412 	max_freq *= cpu->pstate.scaling;
2413 
2414 	ret = freq_qos_add_request(&policy->constraints, req, FREQ_QOS_MIN,
2415 				   min_freq);
2416 	if (ret < 0) {
2417 		dev_err(dev, "Failed to add min-freq constraint (%d)\n", ret);
2418 		goto free_req;
2419 	}
2420 
2421 	ret = freq_qos_add_request(&policy->constraints, req + 1, FREQ_QOS_MAX,
2422 				   max_freq);
2423 	if (ret < 0) {
2424 		dev_err(dev, "Failed to add max-freq constraint (%d)\n", ret);
2425 		goto remove_min_req;
2426 	}
2427 
2428 	policy->driver_data = req;
2429 
2430 	return 0;
2431 
2432 remove_min_req:
2433 	freq_qos_remove_request(req);
2434 free_req:
2435 	kfree(req);
2436 pstate_exit:
2437 	intel_pstate_exit_perf_limits(policy);
2438 
2439 	return ret;
2440 }
2441 
2442 static int intel_cpufreq_cpu_exit(struct cpufreq_policy *policy)
2443 {
2444 	struct freq_qos_request *req;
2445 
2446 	req = policy->driver_data;
2447 
2448 	freq_qos_remove_request(req + 1);
2449 	freq_qos_remove_request(req);
2450 	kfree(req);
2451 
2452 	return intel_pstate_cpu_exit(policy);
2453 }
2454 
2455 static struct cpufreq_driver intel_cpufreq = {
2456 	.flags		= CPUFREQ_CONST_LOOPS,
2457 	.verify		= intel_cpufreq_verify_policy,
2458 	.target		= intel_cpufreq_target,
2459 	.fast_switch	= intel_cpufreq_fast_switch,
2460 	.init		= intel_cpufreq_cpu_init,
2461 	.exit		= intel_cpufreq_cpu_exit,
2462 	.stop_cpu	= intel_cpufreq_stop_cpu,
2463 	.update_limits	= intel_pstate_update_limits,
2464 	.name		= "intel_cpufreq",
2465 };
2466 
2467 static struct cpufreq_driver *default_driver = &intel_pstate;
2468 
2469 static void intel_pstate_driver_cleanup(void)
2470 {
2471 	unsigned int cpu;
2472 
2473 	get_online_cpus();
2474 	for_each_online_cpu(cpu) {
2475 		if (all_cpu_data[cpu]) {
2476 			if (intel_pstate_driver == &intel_pstate)
2477 				intel_pstate_clear_update_util_hook(cpu);
2478 
2479 			kfree(all_cpu_data[cpu]);
2480 			all_cpu_data[cpu] = NULL;
2481 		}
2482 	}
2483 	put_online_cpus();
2484 	intel_pstate_driver = NULL;
2485 }
2486 
2487 static int intel_pstate_register_driver(struct cpufreq_driver *driver)
2488 {
2489 	int ret;
2490 
2491 	memset(&global, 0, sizeof(global));
2492 	global.max_perf_pct = 100;
2493 
2494 	intel_pstate_driver = driver;
2495 	ret = cpufreq_register_driver(intel_pstate_driver);
2496 	if (ret) {
2497 		intel_pstate_driver_cleanup();
2498 		return ret;
2499 	}
2500 
2501 	global.min_perf_pct = min_perf_pct_min();
2502 
2503 	return 0;
2504 }
2505 
2506 static int intel_pstate_unregister_driver(void)
2507 {
2508 	if (hwp_active)
2509 		return -EBUSY;
2510 
2511 	cpufreq_unregister_driver(intel_pstate_driver);
2512 	intel_pstate_driver_cleanup();
2513 
2514 	return 0;
2515 }
2516 
2517 static ssize_t intel_pstate_show_status(char *buf)
2518 {
2519 	if (!intel_pstate_driver)
2520 		return sprintf(buf, "off\n");
2521 
2522 	return sprintf(buf, "%s\n", intel_pstate_driver == &intel_pstate ?
2523 					"active" : "passive");
2524 }
2525 
2526 static int intel_pstate_update_status(const char *buf, size_t size)
2527 {
2528 	int ret;
2529 
2530 	if (size == 3 && !strncmp(buf, "off", size))
2531 		return intel_pstate_driver ?
2532 			intel_pstate_unregister_driver() : -EINVAL;
2533 
2534 	if (size == 6 && !strncmp(buf, "active", size)) {
2535 		if (intel_pstate_driver) {
2536 			if (intel_pstate_driver == &intel_pstate)
2537 				return 0;
2538 
2539 			ret = intel_pstate_unregister_driver();
2540 			if (ret)
2541 				return ret;
2542 		}
2543 
2544 		return intel_pstate_register_driver(&intel_pstate);
2545 	}
2546 
2547 	if (size == 7 && !strncmp(buf, "passive", size)) {
2548 		if (intel_pstate_driver) {
2549 			if (intel_pstate_driver == &intel_cpufreq)
2550 				return 0;
2551 
2552 			ret = intel_pstate_unregister_driver();
2553 			if (ret)
2554 				return ret;
2555 		}
2556 
2557 		return intel_pstate_register_driver(&intel_cpufreq);
2558 	}
2559 
2560 	return -EINVAL;
2561 }
2562 
2563 static int no_load __initdata;
2564 static int no_hwp __initdata;
2565 static int hwp_only __initdata;
2566 static unsigned int force_load __initdata;
2567 
2568 static int __init intel_pstate_msrs_not_valid(void)
2569 {
2570 	if (!pstate_funcs.get_max() ||
2571 	    !pstate_funcs.get_min() ||
2572 	    !pstate_funcs.get_turbo())
2573 		return -ENODEV;
2574 
2575 	return 0;
2576 }
2577 
2578 static void __init copy_cpu_funcs(struct pstate_funcs *funcs)
2579 {
2580 	pstate_funcs.get_max   = funcs->get_max;
2581 	pstate_funcs.get_max_physical = funcs->get_max_physical;
2582 	pstate_funcs.get_min   = funcs->get_min;
2583 	pstate_funcs.get_turbo = funcs->get_turbo;
2584 	pstate_funcs.get_scaling = funcs->get_scaling;
2585 	pstate_funcs.get_val   = funcs->get_val;
2586 	pstate_funcs.get_vid   = funcs->get_vid;
2587 	pstate_funcs.get_aperf_mperf_shift = funcs->get_aperf_mperf_shift;
2588 }
2589 
2590 #ifdef CONFIG_ACPI
2591 
2592 static bool __init intel_pstate_no_acpi_pss(void)
2593 {
2594 	int i;
2595 
2596 	for_each_possible_cpu(i) {
2597 		acpi_status status;
2598 		union acpi_object *pss;
2599 		struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
2600 		struct acpi_processor *pr = per_cpu(processors, i);
2601 
2602 		if (!pr)
2603 			continue;
2604 
2605 		status = acpi_evaluate_object(pr->handle, "_PSS", NULL, &buffer);
2606 		if (ACPI_FAILURE(status))
2607 			continue;
2608 
2609 		pss = buffer.pointer;
2610 		if (pss && pss->type == ACPI_TYPE_PACKAGE) {
2611 			kfree(pss);
2612 			return false;
2613 		}
2614 
2615 		kfree(pss);
2616 	}
2617 
2618 	pr_debug("ACPI _PSS not found\n");
2619 	return true;
2620 }
2621 
2622 static bool __init intel_pstate_no_acpi_pcch(void)
2623 {
2624 	acpi_status status;
2625 	acpi_handle handle;
2626 
2627 	status = acpi_get_handle(NULL, "\\_SB", &handle);
2628 	if (ACPI_FAILURE(status))
2629 		goto not_found;
2630 
2631 	if (acpi_has_method(handle, "PCCH"))
2632 		return false;
2633 
2634 not_found:
2635 	pr_debug("ACPI PCCH not found\n");
2636 	return true;
2637 }
2638 
2639 static bool __init intel_pstate_has_acpi_ppc(void)
2640 {
2641 	int i;
2642 
2643 	for_each_possible_cpu(i) {
2644 		struct acpi_processor *pr = per_cpu(processors, i);
2645 
2646 		if (!pr)
2647 			continue;
2648 		if (acpi_has_method(pr->handle, "_PPC"))
2649 			return true;
2650 	}
2651 	pr_debug("ACPI _PPC not found\n");
2652 	return false;
2653 }
2654 
2655 enum {
2656 	PSS,
2657 	PPC,
2658 };
2659 
2660 /* Hardware vendor-specific info that has its own power management modes */
2661 static struct acpi_platform_list plat_info[] __initdata = {
2662 	{"HP    ", "ProLiant", 0, ACPI_SIG_FADT, all_versions, NULL, PSS},
2663 	{"ORACLE", "X4-2    ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
2664 	{"ORACLE", "X4-2L   ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
2665 	{"ORACLE", "X4-2B   ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
2666 	{"ORACLE", "X3-2    ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
2667 	{"ORACLE", "X3-2L   ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
2668 	{"ORACLE", "X3-2B   ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
2669 	{"ORACLE", "X4470M2 ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
2670 	{"ORACLE", "X4270M3 ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
2671 	{"ORACLE", "X4270M2 ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
2672 	{"ORACLE", "X4170M2 ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
2673 	{"ORACLE", "X4170 M3", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
2674 	{"ORACLE", "X4275 M3", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
2675 	{"ORACLE", "X6-2    ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
2676 	{"ORACLE", "Sudbury ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
2677 	{ } /* End */
2678 };
2679 
2680 static bool __init intel_pstate_platform_pwr_mgmt_exists(void)
2681 {
2682 	const struct x86_cpu_id *id;
2683 	u64 misc_pwr;
2684 	int idx;
2685 
2686 	id = x86_match_cpu(intel_pstate_cpu_oob_ids);
2687 	if (id) {
2688 		rdmsrl(MSR_MISC_PWR_MGMT, misc_pwr);
2689 		if (misc_pwr & (1 << 8)) {
2690 			pr_debug("Bit 8 in the MISC_PWR_MGMT MSR set\n");
2691 			return true;
2692 		}
2693 	}
2694 
2695 	idx = acpi_match_platform_list(plat_info);
2696 	if (idx < 0)
2697 		return false;
2698 
2699 	switch (plat_info[idx].data) {
2700 	case PSS:
2701 		if (!intel_pstate_no_acpi_pss())
2702 			return false;
2703 
2704 		return intel_pstate_no_acpi_pcch();
2705 	case PPC:
2706 		return intel_pstate_has_acpi_ppc() && !force_load;
2707 	}
2708 
2709 	return false;
2710 }
2711 
2712 static void intel_pstate_request_control_from_smm(void)
2713 {
2714 	/*
2715 	 * It may be unsafe to request P-states control from SMM if _PPC support
2716 	 * has not been enabled.
2717 	 */
2718 	if (acpi_ppc)
2719 		acpi_processor_pstate_control();
2720 }
2721 #else /* CONFIG_ACPI not enabled */
2722 static inline bool intel_pstate_platform_pwr_mgmt_exists(void) { return false; }
2723 static inline bool intel_pstate_has_acpi_ppc(void) { return false; }
2724 static inline void intel_pstate_request_control_from_smm(void) {}
2725 #endif /* CONFIG_ACPI */
2726 
2727 #define INTEL_PSTATE_HWP_BROADWELL	0x01
2728 
2729 #define X86_MATCH_HWP(model, hwp_mode)					\
2730 	X86_MATCH_VENDOR_FAM_MODEL_FEATURE(INTEL, 6, INTEL_FAM6_##model, \
2731 					   X86_FEATURE_HWP, hwp_mode)
2732 
2733 static const struct x86_cpu_id hwp_support_ids[] __initconst = {
2734 	X86_MATCH_HWP(BROADWELL_X,	INTEL_PSTATE_HWP_BROADWELL),
2735 	X86_MATCH_HWP(BROADWELL_D,	INTEL_PSTATE_HWP_BROADWELL),
2736 	X86_MATCH_HWP(ANY,		0),
2737 	{}
2738 };
2739 
2740 static int __init intel_pstate_init(void)
2741 {
2742 	const struct x86_cpu_id *id;
2743 	int rc;
2744 
2745 	if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL)
2746 		return -ENODEV;
2747 
2748 	if (no_load)
2749 		return -ENODEV;
2750 
2751 	id = x86_match_cpu(hwp_support_ids);
2752 	if (id) {
2753 		copy_cpu_funcs(&core_funcs);
2754 		if (!no_hwp) {
2755 			hwp_active++;
2756 			hwp_mode_bdw = id->driver_data;
2757 			intel_pstate.attr = hwp_cpufreq_attrs;
2758 			goto hwp_cpu_matched;
2759 		}
2760 	} else {
2761 		id = x86_match_cpu(intel_pstate_cpu_ids);
2762 		if (!id) {
2763 			pr_info("CPU model not supported\n");
2764 			return -ENODEV;
2765 		}
2766 
2767 		copy_cpu_funcs((struct pstate_funcs *)id->driver_data);
2768 	}
2769 
2770 	if (intel_pstate_msrs_not_valid()) {
2771 		pr_info("Invalid MSRs\n");
2772 		return -ENODEV;
2773 	}
2774 
2775 hwp_cpu_matched:
2776 	/*
2777 	 * The Intel pstate driver will be ignored if the platform
2778 	 * firmware has its own power management modes.
2779 	 */
2780 	if (intel_pstate_platform_pwr_mgmt_exists()) {
2781 		pr_info("P-states controlled by the platform\n");
2782 		return -ENODEV;
2783 	}
2784 
2785 	if (!hwp_active && hwp_only)
2786 		return -ENOTSUPP;
2787 
2788 	pr_info("Intel P-state driver initializing\n");
2789 
2790 	all_cpu_data = vzalloc(array_size(sizeof(void *), num_possible_cpus()));
2791 	if (!all_cpu_data)
2792 		return -ENOMEM;
2793 
2794 	intel_pstate_request_control_from_smm();
2795 
2796 	intel_pstate_sysfs_expose_params();
2797 
2798 	mutex_lock(&intel_pstate_driver_lock);
2799 	rc = intel_pstate_register_driver(default_driver);
2800 	mutex_unlock(&intel_pstate_driver_lock);
2801 	if (rc)
2802 		return rc;
2803 
2804 	if (hwp_active)
2805 		pr_info("HWP enabled\n");
2806 
2807 	return 0;
2808 }
2809 device_initcall(intel_pstate_init);
2810 
2811 static int __init intel_pstate_setup(char *str)
2812 {
2813 	if (!str)
2814 		return -EINVAL;
2815 
2816 	if (!strcmp(str, "disable")) {
2817 		no_load = 1;
2818 	} else if (!strcmp(str, "passive")) {
2819 		pr_info("Passive mode enabled\n");
2820 		default_driver = &intel_cpufreq;
2821 		no_hwp = 1;
2822 	}
2823 	if (!strcmp(str, "no_hwp")) {
2824 		pr_info("HWP disabled\n");
2825 		no_hwp = 1;
2826 	}
2827 	if (!strcmp(str, "force"))
2828 		force_load = 1;
2829 	if (!strcmp(str, "hwp_only"))
2830 		hwp_only = 1;
2831 	if (!strcmp(str, "per_cpu_perf_limits"))
2832 		per_cpu_limits = true;
2833 
2834 #ifdef CONFIG_ACPI
2835 	if (!strcmp(str, "support_acpi_ppc"))
2836 		acpi_ppc = true;
2837 #endif
2838 
2839 	return 0;
2840 }
2841 early_param("intel_pstate", intel_pstate_setup);
2842 
2843 MODULE_AUTHOR("Dirk Brandewie <dirk.j.brandewie@intel.com>");
2844 MODULE_DESCRIPTION("'intel_pstate' - P state driver Intel Core processors");
2845 MODULE_LICENSE("GPL");
2846