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_HWP	5000
40 #define INTEL_CPUFREQ_TRANSITION_DELAY		500
41 
42 #ifdef CONFIG_ACPI
43 #include <acpi/processor.h>
44 #include <acpi/cppc_acpi.h>
45 #endif
46 
47 #define FRAC_BITS 8
48 #define int_tofp(X) ((int64_t)(X) << FRAC_BITS)
49 #define fp_toint(X) ((X) >> FRAC_BITS)
50 
51 #define ONE_EIGHTH_FP ((int64_t)1 << (FRAC_BITS - 3))
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:	Whether 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  * @turbo_disabled_mf:	The @turbo_disabled value reflected by cpuinfo.max_freq.
181  * @min_perf_pct:	Minimum capacity limit in percent of the maximum turbo
182  *			P-state capacity.
183  * @max_perf_pct:	Maximum capacity limit in percent of the maximum turbo
184  *			P-state capacity.
185  */
186 struct global_params {
187 	bool no_turbo;
188 	bool turbo_disabled;
189 	bool turbo_disabled_mf;
190 	int max_perf_pct;
191 	int min_perf_pct;
192 };
193 
194 /**
195  * struct cpudata -	Per CPU instance data storage
196  * @cpu:		CPU number for this instance data
197  * @policy:		CPUFreq policy value
198  * @update_util:	CPUFreq utility callback information
199  * @update_util_set:	CPUFreq utility callback is set
200  * @iowait_boost:	iowait-related boost fraction
201  * @last_update:	Time of the last update.
202  * @pstate:		Stores P state limits for this CPU
203  * @vid:		Stores VID limits for this CPU
204  * @last_sample_time:	Last Sample time
205  * @aperf_mperf_shift:	APERF vs MPERF counting frequency difference
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  * @epp_cached		Cached HWP energy-performance preference value
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 	s16 epp_cached;
263 	u64 hwp_req_cached;
264 	u64 hwp_cap_cached;
265 	u64 last_io_update;
266 	unsigned int sched_flags;
267 	u32 hwp_boost_min;
268 };
269 
270 static struct cpudata **all_cpu_data;
271 
272 /**
273  * struct pstate_funcs - Per CPU model specific callbacks
274  * @get_max:		Callback to get maximum non turbo effective P state
275  * @get_max_physical:	Callback to get maximum non turbo physical P state
276  * @get_min:		Callback to get minimum P state
277  * @get_turbo:		Callback to get turbo P state
278  * @get_scaling:	Callback to get frequency scaling factor
279  * @get_aperf_mperf_shift: Callback to get the APERF vs MPERF frequency difference
280  * @get_val:		Callback to convert P state to actual MSR write value
281  * @get_vid:		Callback to get VID data for Atom platforms
282  *
283  * Core and Atom CPU models have different way to get P State limits. This
284  * structure is used to store those callbacks.
285  */
286 struct pstate_funcs {
287 	int (*get_max)(void);
288 	int (*get_max_physical)(void);
289 	int (*get_min)(void);
290 	int (*get_turbo)(void);
291 	int (*get_scaling)(void);
292 	int (*get_aperf_mperf_shift)(void);
293 	u64 (*get_val)(struct cpudata*, int pstate);
294 	void (*get_vid)(struct cpudata *);
295 };
296 
297 static struct pstate_funcs pstate_funcs __read_mostly;
298 
299 static int hwp_active __read_mostly;
300 static int hwp_mode_bdw __read_mostly;
301 static bool per_cpu_limits __read_mostly;
302 static bool hwp_boost __read_mostly;
303 
304 static struct cpufreq_driver *intel_pstate_driver __read_mostly;
305 
306 #ifdef CONFIG_ACPI
307 static bool acpi_ppc;
308 #endif
309 
310 static struct global_params global;
311 
312 static DEFINE_MUTEX(intel_pstate_driver_lock);
313 static DEFINE_MUTEX(intel_pstate_limits_lock);
314 
315 #ifdef CONFIG_ACPI
316 
317 static bool intel_pstate_acpi_pm_profile_server(void)
318 {
319 	if (acpi_gbl_FADT.preferred_profile == PM_ENTERPRISE_SERVER ||
320 	    acpi_gbl_FADT.preferred_profile == PM_PERFORMANCE_SERVER)
321 		return true;
322 
323 	return false;
324 }
325 
326 static bool intel_pstate_get_ppc_enable_status(void)
327 {
328 	if (intel_pstate_acpi_pm_profile_server())
329 		return true;
330 
331 	return acpi_ppc;
332 }
333 
334 #ifdef CONFIG_ACPI_CPPC_LIB
335 
336 /* The work item is needed to avoid CPU hotplug locking issues */
337 static void intel_pstste_sched_itmt_work_fn(struct work_struct *work)
338 {
339 	sched_set_itmt_support();
340 }
341 
342 static DECLARE_WORK(sched_itmt_work, intel_pstste_sched_itmt_work_fn);
343 
344 static void intel_pstate_set_itmt_prio(int cpu)
345 {
346 	struct cppc_perf_caps cppc_perf;
347 	static u32 max_highest_perf = 0, min_highest_perf = U32_MAX;
348 	int ret;
349 
350 	ret = cppc_get_perf_caps(cpu, &cppc_perf);
351 	if (ret)
352 		return;
353 
354 	/*
355 	 * The priorities can be set regardless of whether or not
356 	 * sched_set_itmt_support(true) has been called and it is valid to
357 	 * update them at any time after it has been called.
358 	 */
359 	sched_set_itmt_core_prio(cppc_perf.highest_perf, cpu);
360 
361 	if (max_highest_perf <= min_highest_perf) {
362 		if (cppc_perf.highest_perf > max_highest_perf)
363 			max_highest_perf = cppc_perf.highest_perf;
364 
365 		if (cppc_perf.highest_perf < min_highest_perf)
366 			min_highest_perf = cppc_perf.highest_perf;
367 
368 		if (max_highest_perf > min_highest_perf) {
369 			/*
370 			 * This code can be run during CPU online under the
371 			 * CPU hotplug locks, so sched_set_itmt_support()
372 			 * cannot be called from here.  Queue up a work item
373 			 * to invoke it.
374 			 */
375 			schedule_work(&sched_itmt_work);
376 		}
377 	}
378 }
379 
380 static int intel_pstate_get_cppc_guranteed(int cpu)
381 {
382 	struct cppc_perf_caps cppc_perf;
383 	int ret;
384 
385 	ret = cppc_get_perf_caps(cpu, &cppc_perf);
386 	if (ret)
387 		return ret;
388 
389 	if (cppc_perf.guaranteed_perf)
390 		return cppc_perf.guaranteed_perf;
391 
392 	return cppc_perf.nominal_perf;
393 }
394 
395 #else /* CONFIG_ACPI_CPPC_LIB */
396 static void intel_pstate_set_itmt_prio(int cpu)
397 {
398 }
399 #endif /* CONFIG_ACPI_CPPC_LIB */
400 
401 static void intel_pstate_init_acpi_perf_limits(struct cpufreq_policy *policy)
402 {
403 	struct cpudata *cpu;
404 	int ret;
405 	int i;
406 
407 	if (hwp_active) {
408 		intel_pstate_set_itmt_prio(policy->cpu);
409 		return;
410 	}
411 
412 	if (!intel_pstate_get_ppc_enable_status())
413 		return;
414 
415 	cpu = all_cpu_data[policy->cpu];
416 
417 	ret = acpi_processor_register_performance(&cpu->acpi_perf_data,
418 						  policy->cpu);
419 	if (ret)
420 		return;
421 
422 	/*
423 	 * Check if the control value in _PSS is for PERF_CTL MSR, which should
424 	 * guarantee that the states returned by it map to the states in our
425 	 * list directly.
426 	 */
427 	if (cpu->acpi_perf_data.control_register.space_id !=
428 						ACPI_ADR_SPACE_FIXED_HARDWARE)
429 		goto err;
430 
431 	/*
432 	 * If there is only one entry _PSS, simply ignore _PSS and continue as
433 	 * usual without taking _PSS into account
434 	 */
435 	if (cpu->acpi_perf_data.state_count < 2)
436 		goto err;
437 
438 	pr_debug("CPU%u - ACPI _PSS perf data\n", policy->cpu);
439 	for (i = 0; i < cpu->acpi_perf_data.state_count; i++) {
440 		pr_debug("     %cP%d: %u MHz, %u mW, 0x%x\n",
441 			 (i == cpu->acpi_perf_data.state ? '*' : ' '), i,
442 			 (u32) cpu->acpi_perf_data.states[i].core_frequency,
443 			 (u32) cpu->acpi_perf_data.states[i].power,
444 			 (u32) cpu->acpi_perf_data.states[i].control);
445 	}
446 
447 	/*
448 	 * The _PSS table doesn't contain whole turbo frequency range.
449 	 * This just contains +1 MHZ above the max non turbo frequency,
450 	 * with control value corresponding to max turbo ratio. But
451 	 * when cpufreq set policy is called, it will call with this
452 	 * max frequency, which will cause a reduced performance as
453 	 * this driver uses real max turbo frequency as the max
454 	 * frequency. So correct this frequency in _PSS table to
455 	 * correct max turbo frequency based on the turbo state.
456 	 * Also need to convert to MHz as _PSS freq is in MHz.
457 	 */
458 	if (!global.turbo_disabled)
459 		cpu->acpi_perf_data.states[0].core_frequency =
460 					policy->cpuinfo.max_freq / 1000;
461 	cpu->valid_pss_table = true;
462 	pr_debug("_PPC limits will be enforced\n");
463 
464 	return;
465 
466  err:
467 	cpu->valid_pss_table = false;
468 	acpi_processor_unregister_performance(policy->cpu);
469 }
470 
471 static void intel_pstate_exit_perf_limits(struct cpufreq_policy *policy)
472 {
473 	struct cpudata *cpu;
474 
475 	cpu = all_cpu_data[policy->cpu];
476 	if (!cpu->valid_pss_table)
477 		return;
478 
479 	acpi_processor_unregister_performance(policy->cpu);
480 }
481 #else /* CONFIG_ACPI */
482 static inline void intel_pstate_init_acpi_perf_limits(struct cpufreq_policy *policy)
483 {
484 }
485 
486 static inline void intel_pstate_exit_perf_limits(struct cpufreq_policy *policy)
487 {
488 }
489 
490 static inline bool intel_pstate_acpi_pm_profile_server(void)
491 {
492 	return false;
493 }
494 #endif /* CONFIG_ACPI */
495 
496 #ifndef CONFIG_ACPI_CPPC_LIB
497 static int intel_pstate_get_cppc_guranteed(int cpu)
498 {
499 	return -ENOTSUPP;
500 }
501 #endif /* CONFIG_ACPI_CPPC_LIB */
502 
503 static inline void update_turbo_state(void)
504 {
505 	u64 misc_en;
506 	struct cpudata *cpu;
507 
508 	cpu = all_cpu_data[0];
509 	rdmsrl(MSR_IA32_MISC_ENABLE, misc_en);
510 	global.turbo_disabled =
511 		(misc_en & MSR_IA32_MISC_ENABLE_TURBO_DISABLE ||
512 		 cpu->pstate.max_pstate == cpu->pstate.turbo_pstate);
513 }
514 
515 static int min_perf_pct_min(void)
516 {
517 	struct cpudata *cpu = all_cpu_data[0];
518 	int turbo_pstate = cpu->pstate.turbo_pstate;
519 
520 	return turbo_pstate ?
521 		(cpu->pstate.min_pstate * 100 / turbo_pstate) : 0;
522 }
523 
524 static s16 intel_pstate_get_epb(struct cpudata *cpu_data)
525 {
526 	u64 epb;
527 	int ret;
528 
529 	if (!boot_cpu_has(X86_FEATURE_EPB))
530 		return -ENXIO;
531 
532 	ret = rdmsrl_on_cpu(cpu_data->cpu, MSR_IA32_ENERGY_PERF_BIAS, &epb);
533 	if (ret)
534 		return (s16)ret;
535 
536 	return (s16)(epb & 0x0f);
537 }
538 
539 static s16 intel_pstate_get_epp(struct cpudata *cpu_data, u64 hwp_req_data)
540 {
541 	s16 epp;
542 
543 	if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
544 		/*
545 		 * When hwp_req_data is 0, means that caller didn't read
546 		 * MSR_HWP_REQUEST, so need to read and get EPP.
547 		 */
548 		if (!hwp_req_data) {
549 			epp = rdmsrl_on_cpu(cpu_data->cpu, MSR_HWP_REQUEST,
550 					    &hwp_req_data);
551 			if (epp)
552 				return epp;
553 		}
554 		epp = (hwp_req_data >> 24) & 0xff;
555 	} else {
556 		/* When there is no EPP present, HWP uses EPB settings */
557 		epp = intel_pstate_get_epb(cpu_data);
558 	}
559 
560 	return epp;
561 }
562 
563 static int intel_pstate_set_epb(int cpu, s16 pref)
564 {
565 	u64 epb;
566 	int ret;
567 
568 	if (!boot_cpu_has(X86_FEATURE_EPB))
569 		return -ENXIO;
570 
571 	ret = rdmsrl_on_cpu(cpu, MSR_IA32_ENERGY_PERF_BIAS, &epb);
572 	if (ret)
573 		return ret;
574 
575 	epb = (epb & ~0x0f) | pref;
576 	wrmsrl_on_cpu(cpu, MSR_IA32_ENERGY_PERF_BIAS, epb);
577 
578 	return 0;
579 }
580 
581 /*
582  * EPP/EPB display strings corresponding to EPP index in the
583  * energy_perf_strings[]
584  *	index		String
585  *-------------------------------------
586  *	0		default
587  *	1		performance
588  *	2		balance_performance
589  *	3		balance_power
590  *	4		power
591  */
592 static const char * const energy_perf_strings[] = {
593 	"default",
594 	"performance",
595 	"balance_performance",
596 	"balance_power",
597 	"power",
598 	NULL
599 };
600 static const unsigned int epp_values[] = {
601 	HWP_EPP_PERFORMANCE,
602 	HWP_EPP_BALANCE_PERFORMANCE,
603 	HWP_EPP_BALANCE_POWERSAVE,
604 	HWP_EPP_POWERSAVE
605 };
606 
607 static int intel_pstate_get_energy_pref_index(struct cpudata *cpu_data, int *raw_epp)
608 {
609 	s16 epp;
610 	int index = -EINVAL;
611 
612 	*raw_epp = 0;
613 	epp = intel_pstate_get_epp(cpu_data, 0);
614 	if (epp < 0)
615 		return epp;
616 
617 	if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
618 		if (epp == HWP_EPP_PERFORMANCE)
619 			return 1;
620 		if (epp == HWP_EPP_BALANCE_PERFORMANCE)
621 			return 2;
622 		if (epp == HWP_EPP_BALANCE_POWERSAVE)
623 			return 3;
624 		if (epp == HWP_EPP_POWERSAVE)
625 			return 4;
626 		*raw_epp = epp;
627 		return 0;
628 	} else if (boot_cpu_has(X86_FEATURE_EPB)) {
629 		/*
630 		 * Range:
631 		 *	0x00-0x03	:	Performance
632 		 *	0x04-0x07	:	Balance performance
633 		 *	0x08-0x0B	:	Balance power
634 		 *	0x0C-0x0F	:	Power
635 		 * The EPB is a 4 bit value, but our ranges restrict the
636 		 * value which can be set. Here only using top two bits
637 		 * effectively.
638 		 */
639 		index = (epp >> 2) + 1;
640 	}
641 
642 	return index;
643 }
644 
645 static int intel_pstate_set_epp(struct cpudata *cpu, u32 epp)
646 {
647 	/*
648 	 * Use the cached HWP Request MSR value, because in the active mode the
649 	 * register itself may be updated by intel_pstate_hwp_boost_up() or
650 	 * intel_pstate_hwp_boost_down() at any time.
651 	 */
652 	u64 value = READ_ONCE(cpu->hwp_req_cached);
653 
654 	value &= ~GENMASK_ULL(31, 24);
655 	value |= (u64)epp << 24;
656 	/*
657 	 * The only other updater of hwp_req_cached in the active mode,
658 	 * intel_pstate_hwp_set(), is called under the same lock as this
659 	 * function, so it cannot run in parallel with the update below.
660 	 */
661 	WRITE_ONCE(cpu->hwp_req_cached, value);
662 	return wrmsrl_on_cpu(cpu->cpu, MSR_HWP_REQUEST, value);
663 }
664 
665 static int intel_pstate_set_energy_pref_index(struct cpudata *cpu_data,
666 					      int pref_index, bool use_raw,
667 					      u32 raw_epp)
668 {
669 	int epp = -EINVAL;
670 	int ret;
671 
672 	if (!pref_index)
673 		epp = cpu_data->epp_default;
674 
675 	if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
676 		if (use_raw)
677 			epp = raw_epp;
678 		else if (epp == -EINVAL)
679 			epp = epp_values[pref_index - 1];
680 
681 		ret = intel_pstate_set_epp(cpu_data, epp);
682 	} else {
683 		if (epp == -EINVAL)
684 			epp = (pref_index - 1) << 2;
685 		ret = intel_pstate_set_epb(cpu_data->cpu, epp);
686 	}
687 
688 	return ret;
689 }
690 
691 static ssize_t show_energy_performance_available_preferences(
692 				struct cpufreq_policy *policy, char *buf)
693 {
694 	int i = 0;
695 	int ret = 0;
696 
697 	while (energy_perf_strings[i] != NULL)
698 		ret += sprintf(&buf[ret], "%s ", energy_perf_strings[i++]);
699 
700 	ret += sprintf(&buf[ret], "\n");
701 
702 	return ret;
703 }
704 
705 cpufreq_freq_attr_ro(energy_performance_available_preferences);
706 
707 static struct cpufreq_driver intel_pstate;
708 
709 static ssize_t store_energy_performance_preference(
710 		struct cpufreq_policy *policy, const char *buf, size_t count)
711 {
712 	struct cpudata *cpu = all_cpu_data[policy->cpu];
713 	char str_preference[21];
714 	bool raw = false;
715 	ssize_t ret;
716 	u32 epp = 0;
717 
718 	ret = sscanf(buf, "%20s", str_preference);
719 	if (ret != 1)
720 		return -EINVAL;
721 
722 	ret = match_string(energy_perf_strings, -1, str_preference);
723 	if (ret < 0) {
724 		if (!boot_cpu_has(X86_FEATURE_HWP_EPP))
725 			return ret;
726 
727 		ret = kstrtouint(buf, 10, &epp);
728 		if (ret)
729 			return ret;
730 
731 		if (epp > 255)
732 			return -EINVAL;
733 
734 		raw = true;
735 	}
736 
737 	/*
738 	 * This function runs with the policy R/W semaphore held, which
739 	 * guarantees that the driver pointer will not change while it is
740 	 * running.
741 	 */
742 	if (!intel_pstate_driver)
743 		return -EAGAIN;
744 
745 	mutex_lock(&intel_pstate_limits_lock);
746 
747 	if (intel_pstate_driver == &intel_pstate) {
748 		ret = intel_pstate_set_energy_pref_index(cpu, ret, raw, epp);
749 	} else {
750 		/*
751 		 * In the passive mode the governor needs to be stopped on the
752 		 * target CPU before the EPP update and restarted after it,
753 		 * which is super-heavy-weight, so make sure it is worth doing
754 		 * upfront.
755 		 */
756 		if (!raw)
757 			epp = ret ? epp_values[ret - 1] : cpu->epp_default;
758 
759 		if (cpu->epp_cached != epp) {
760 			int err;
761 
762 			cpufreq_stop_governor(policy);
763 			ret = intel_pstate_set_epp(cpu, epp);
764 			err = cpufreq_start_governor(policy);
765 			if (!ret) {
766 				cpu->epp_cached = epp;
767 				ret = err;
768 			}
769 		}
770 	}
771 
772 	mutex_unlock(&intel_pstate_limits_lock);
773 
774 	return ret ?: count;
775 }
776 
777 static ssize_t show_energy_performance_preference(
778 				struct cpufreq_policy *policy, char *buf)
779 {
780 	struct cpudata *cpu_data = all_cpu_data[policy->cpu];
781 	int preference, raw_epp;
782 
783 	preference = intel_pstate_get_energy_pref_index(cpu_data, &raw_epp);
784 	if (preference < 0)
785 		return preference;
786 
787 	if (raw_epp)
788 		return  sprintf(buf, "%d\n", raw_epp);
789 	else
790 		return  sprintf(buf, "%s\n", energy_perf_strings[preference]);
791 }
792 
793 cpufreq_freq_attr_rw(energy_performance_preference);
794 
795 static ssize_t show_base_frequency(struct cpufreq_policy *policy, char *buf)
796 {
797 	struct cpudata *cpu;
798 	u64 cap;
799 	int ratio;
800 
801 	ratio = intel_pstate_get_cppc_guranteed(policy->cpu);
802 	if (ratio <= 0) {
803 		rdmsrl_on_cpu(policy->cpu, MSR_HWP_CAPABILITIES, &cap);
804 		ratio = HWP_GUARANTEED_PERF(cap);
805 	}
806 
807 	cpu = all_cpu_data[policy->cpu];
808 
809 	return sprintf(buf, "%d\n", ratio * cpu->pstate.scaling);
810 }
811 
812 cpufreq_freq_attr_ro(base_frequency);
813 
814 static struct freq_attr *hwp_cpufreq_attrs[] = {
815 	&energy_performance_preference,
816 	&energy_performance_available_preferences,
817 	&base_frequency,
818 	NULL,
819 };
820 
821 static void intel_pstate_get_hwp_max(unsigned int cpu, int *phy_max,
822 				     int *current_max)
823 {
824 	u64 cap;
825 
826 	rdmsrl_on_cpu(cpu, MSR_HWP_CAPABILITIES, &cap);
827 	WRITE_ONCE(all_cpu_data[cpu]->hwp_cap_cached, cap);
828 	if (global.no_turbo)
829 		*current_max = HWP_GUARANTEED_PERF(cap);
830 	else
831 		*current_max = HWP_HIGHEST_PERF(cap);
832 
833 	*phy_max = HWP_HIGHEST_PERF(cap);
834 }
835 
836 static void intel_pstate_hwp_set(unsigned int cpu)
837 {
838 	struct cpudata *cpu_data = all_cpu_data[cpu];
839 	int max, min;
840 	u64 value;
841 	s16 epp;
842 
843 	max = cpu_data->max_perf_ratio;
844 	min = cpu_data->min_perf_ratio;
845 
846 	if (cpu_data->policy == CPUFREQ_POLICY_PERFORMANCE)
847 		min = max;
848 
849 	rdmsrl_on_cpu(cpu, MSR_HWP_REQUEST, &value);
850 
851 	value &= ~HWP_MIN_PERF(~0L);
852 	value |= HWP_MIN_PERF(min);
853 
854 	value &= ~HWP_MAX_PERF(~0L);
855 	value |= HWP_MAX_PERF(max);
856 
857 	if (cpu_data->epp_policy == cpu_data->policy)
858 		goto skip_epp;
859 
860 	cpu_data->epp_policy = cpu_data->policy;
861 
862 	if (cpu_data->epp_saved >= 0) {
863 		epp = cpu_data->epp_saved;
864 		cpu_data->epp_saved = -EINVAL;
865 		goto update_epp;
866 	}
867 
868 	if (cpu_data->policy == CPUFREQ_POLICY_PERFORMANCE) {
869 		epp = intel_pstate_get_epp(cpu_data, value);
870 		cpu_data->epp_powersave = epp;
871 		/* If EPP read was failed, then don't try to write */
872 		if (epp < 0)
873 			goto skip_epp;
874 
875 		epp = 0;
876 	} else {
877 		/* skip setting EPP, when saved value is invalid */
878 		if (cpu_data->epp_powersave < 0)
879 			goto skip_epp;
880 
881 		/*
882 		 * No need to restore EPP when it is not zero. This
883 		 * means:
884 		 *  - Policy is not changed
885 		 *  - user has manually changed
886 		 *  - Error reading EPB
887 		 */
888 		epp = intel_pstate_get_epp(cpu_data, value);
889 		if (epp)
890 			goto skip_epp;
891 
892 		epp = cpu_data->epp_powersave;
893 	}
894 update_epp:
895 	if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
896 		value &= ~GENMASK_ULL(31, 24);
897 		value |= (u64)epp << 24;
898 	} else {
899 		intel_pstate_set_epb(cpu, epp);
900 	}
901 skip_epp:
902 	WRITE_ONCE(cpu_data->hwp_req_cached, value);
903 	wrmsrl_on_cpu(cpu, MSR_HWP_REQUEST, value);
904 }
905 
906 static void intel_pstate_hwp_force_min_perf(int cpu)
907 {
908 	u64 value;
909 	int min_perf;
910 
911 	value = all_cpu_data[cpu]->hwp_req_cached;
912 	value &= ~GENMASK_ULL(31, 0);
913 	min_perf = HWP_LOWEST_PERF(all_cpu_data[cpu]->hwp_cap_cached);
914 
915 	/* Set hwp_max = hwp_min */
916 	value |= HWP_MAX_PERF(min_perf);
917 	value |= HWP_MIN_PERF(min_perf);
918 
919 	/* Set EPP to min */
920 	if (boot_cpu_has(X86_FEATURE_HWP_EPP))
921 		value |= HWP_ENERGY_PERF_PREFERENCE(HWP_EPP_POWERSAVE);
922 
923 	wrmsrl_on_cpu(cpu, MSR_HWP_REQUEST, value);
924 }
925 
926 static int intel_pstate_hwp_save_state(struct cpufreq_policy *policy)
927 {
928 	struct cpudata *cpu_data = all_cpu_data[policy->cpu];
929 
930 	if (!hwp_active)
931 		return 0;
932 
933 	cpu_data->epp_saved = intel_pstate_get_epp(cpu_data, 0);
934 
935 	return 0;
936 }
937 
938 #define POWER_CTL_EE_ENABLE	1
939 #define POWER_CTL_EE_DISABLE	2
940 
941 static int power_ctl_ee_state;
942 
943 static void set_power_ctl_ee_state(bool input)
944 {
945 	u64 power_ctl;
946 
947 	mutex_lock(&intel_pstate_driver_lock);
948 	rdmsrl(MSR_IA32_POWER_CTL, power_ctl);
949 	if (input) {
950 		power_ctl &= ~BIT(MSR_IA32_POWER_CTL_BIT_EE);
951 		power_ctl_ee_state = POWER_CTL_EE_ENABLE;
952 	} else {
953 		power_ctl |= BIT(MSR_IA32_POWER_CTL_BIT_EE);
954 		power_ctl_ee_state = POWER_CTL_EE_DISABLE;
955 	}
956 	wrmsrl(MSR_IA32_POWER_CTL, power_ctl);
957 	mutex_unlock(&intel_pstate_driver_lock);
958 }
959 
960 static void intel_pstate_hwp_enable(struct cpudata *cpudata);
961 
962 static int intel_pstate_resume(struct cpufreq_policy *policy)
963 {
964 
965 	/* Only restore if the system default is changed */
966 	if (power_ctl_ee_state == POWER_CTL_EE_ENABLE)
967 		set_power_ctl_ee_state(true);
968 	else if (power_ctl_ee_state == POWER_CTL_EE_DISABLE)
969 		set_power_ctl_ee_state(false);
970 
971 	if (!hwp_active)
972 		return 0;
973 
974 	mutex_lock(&intel_pstate_limits_lock);
975 
976 	if (policy->cpu == 0)
977 		intel_pstate_hwp_enable(all_cpu_data[policy->cpu]);
978 
979 	all_cpu_data[policy->cpu]->epp_policy = 0;
980 	intel_pstate_hwp_set(policy->cpu);
981 
982 	mutex_unlock(&intel_pstate_limits_lock);
983 
984 	return 0;
985 }
986 
987 static void intel_pstate_update_policies(void)
988 {
989 	int cpu;
990 
991 	for_each_possible_cpu(cpu)
992 		cpufreq_update_policy(cpu);
993 }
994 
995 static void intel_pstate_update_max_freq(unsigned int cpu)
996 {
997 	struct cpufreq_policy *policy = cpufreq_cpu_acquire(cpu);
998 	struct cpudata *cpudata;
999 
1000 	if (!policy)
1001 		return;
1002 
1003 	cpudata = all_cpu_data[cpu];
1004 	policy->cpuinfo.max_freq = global.turbo_disabled_mf ?
1005 			cpudata->pstate.max_freq : cpudata->pstate.turbo_freq;
1006 
1007 	refresh_frequency_limits(policy);
1008 
1009 	cpufreq_cpu_release(policy);
1010 }
1011 
1012 static void intel_pstate_update_limits(unsigned int cpu)
1013 {
1014 	mutex_lock(&intel_pstate_driver_lock);
1015 
1016 	update_turbo_state();
1017 	/*
1018 	 * If turbo has been turned on or off globally, policy limits for
1019 	 * all CPUs need to be updated to reflect that.
1020 	 */
1021 	if (global.turbo_disabled_mf != global.turbo_disabled) {
1022 		global.turbo_disabled_mf = global.turbo_disabled;
1023 		arch_set_max_freq_ratio(global.turbo_disabled);
1024 		for_each_possible_cpu(cpu)
1025 			intel_pstate_update_max_freq(cpu);
1026 	} else {
1027 		cpufreq_update_policy(cpu);
1028 	}
1029 
1030 	mutex_unlock(&intel_pstate_driver_lock);
1031 }
1032 
1033 /************************** sysfs begin ************************/
1034 #define show_one(file_name, object)					\
1035 	static ssize_t show_##file_name					\
1036 	(struct kobject *kobj, struct kobj_attribute *attr, char *buf)	\
1037 	{								\
1038 		return sprintf(buf, "%u\n", global.object);		\
1039 	}
1040 
1041 static ssize_t intel_pstate_show_status(char *buf);
1042 static int intel_pstate_update_status(const char *buf, size_t size);
1043 
1044 static ssize_t show_status(struct kobject *kobj,
1045 			   struct kobj_attribute *attr, char *buf)
1046 {
1047 	ssize_t ret;
1048 
1049 	mutex_lock(&intel_pstate_driver_lock);
1050 	ret = intel_pstate_show_status(buf);
1051 	mutex_unlock(&intel_pstate_driver_lock);
1052 
1053 	return ret;
1054 }
1055 
1056 static ssize_t store_status(struct kobject *a, struct kobj_attribute *b,
1057 			    const char *buf, size_t count)
1058 {
1059 	char *p = memchr(buf, '\n', count);
1060 	int ret;
1061 
1062 	mutex_lock(&intel_pstate_driver_lock);
1063 	ret = intel_pstate_update_status(buf, p ? p - buf : count);
1064 	mutex_unlock(&intel_pstate_driver_lock);
1065 
1066 	return ret < 0 ? ret : count;
1067 }
1068 
1069 static ssize_t show_turbo_pct(struct kobject *kobj,
1070 				struct kobj_attribute *attr, char *buf)
1071 {
1072 	struct cpudata *cpu;
1073 	int total, no_turbo, turbo_pct;
1074 	uint32_t turbo_fp;
1075 
1076 	mutex_lock(&intel_pstate_driver_lock);
1077 
1078 	if (!intel_pstate_driver) {
1079 		mutex_unlock(&intel_pstate_driver_lock);
1080 		return -EAGAIN;
1081 	}
1082 
1083 	cpu = all_cpu_data[0];
1084 
1085 	total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
1086 	no_turbo = cpu->pstate.max_pstate - cpu->pstate.min_pstate + 1;
1087 	turbo_fp = div_fp(no_turbo, total);
1088 	turbo_pct = 100 - fp_toint(mul_fp(turbo_fp, int_tofp(100)));
1089 
1090 	mutex_unlock(&intel_pstate_driver_lock);
1091 
1092 	return sprintf(buf, "%u\n", turbo_pct);
1093 }
1094 
1095 static ssize_t show_num_pstates(struct kobject *kobj,
1096 				struct kobj_attribute *attr, char *buf)
1097 {
1098 	struct cpudata *cpu;
1099 	int total;
1100 
1101 	mutex_lock(&intel_pstate_driver_lock);
1102 
1103 	if (!intel_pstate_driver) {
1104 		mutex_unlock(&intel_pstate_driver_lock);
1105 		return -EAGAIN;
1106 	}
1107 
1108 	cpu = all_cpu_data[0];
1109 	total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
1110 
1111 	mutex_unlock(&intel_pstate_driver_lock);
1112 
1113 	return sprintf(buf, "%u\n", total);
1114 }
1115 
1116 static ssize_t show_no_turbo(struct kobject *kobj,
1117 			     struct kobj_attribute *attr, char *buf)
1118 {
1119 	ssize_t ret;
1120 
1121 	mutex_lock(&intel_pstate_driver_lock);
1122 
1123 	if (!intel_pstate_driver) {
1124 		mutex_unlock(&intel_pstate_driver_lock);
1125 		return -EAGAIN;
1126 	}
1127 
1128 	update_turbo_state();
1129 	if (global.turbo_disabled)
1130 		ret = sprintf(buf, "%u\n", global.turbo_disabled);
1131 	else
1132 		ret = sprintf(buf, "%u\n", global.no_turbo);
1133 
1134 	mutex_unlock(&intel_pstate_driver_lock);
1135 
1136 	return ret;
1137 }
1138 
1139 static ssize_t store_no_turbo(struct kobject *a, struct kobj_attribute *b,
1140 			      const char *buf, size_t count)
1141 {
1142 	unsigned int input;
1143 	int ret;
1144 
1145 	ret = sscanf(buf, "%u", &input);
1146 	if (ret != 1)
1147 		return -EINVAL;
1148 
1149 	mutex_lock(&intel_pstate_driver_lock);
1150 
1151 	if (!intel_pstate_driver) {
1152 		mutex_unlock(&intel_pstate_driver_lock);
1153 		return -EAGAIN;
1154 	}
1155 
1156 	mutex_lock(&intel_pstate_limits_lock);
1157 
1158 	update_turbo_state();
1159 	if (global.turbo_disabled) {
1160 		pr_notice_once("Turbo disabled by BIOS or unavailable on processor\n");
1161 		mutex_unlock(&intel_pstate_limits_lock);
1162 		mutex_unlock(&intel_pstate_driver_lock);
1163 		return -EPERM;
1164 	}
1165 
1166 	global.no_turbo = clamp_t(int, input, 0, 1);
1167 
1168 	if (global.no_turbo) {
1169 		struct cpudata *cpu = all_cpu_data[0];
1170 		int pct = cpu->pstate.max_pstate * 100 / cpu->pstate.turbo_pstate;
1171 
1172 		/* Squash the global minimum into the permitted range. */
1173 		if (global.min_perf_pct > pct)
1174 			global.min_perf_pct = pct;
1175 	}
1176 
1177 	mutex_unlock(&intel_pstate_limits_lock);
1178 
1179 	intel_pstate_update_policies();
1180 
1181 	mutex_unlock(&intel_pstate_driver_lock);
1182 
1183 	return count;
1184 }
1185 
1186 static void update_qos_request(enum freq_qos_req_type type)
1187 {
1188 	int max_state, turbo_max, freq, i, perf_pct;
1189 	struct freq_qos_request *req;
1190 	struct cpufreq_policy *policy;
1191 
1192 	for_each_possible_cpu(i) {
1193 		struct cpudata *cpu = all_cpu_data[i];
1194 
1195 		policy = cpufreq_cpu_get(i);
1196 		if (!policy)
1197 			continue;
1198 
1199 		req = policy->driver_data;
1200 		cpufreq_cpu_put(policy);
1201 
1202 		if (!req)
1203 			continue;
1204 
1205 		if (hwp_active)
1206 			intel_pstate_get_hwp_max(i, &turbo_max, &max_state);
1207 		else
1208 			turbo_max = cpu->pstate.turbo_pstate;
1209 
1210 		if (type == FREQ_QOS_MIN) {
1211 			perf_pct = global.min_perf_pct;
1212 		} else {
1213 			req++;
1214 			perf_pct = global.max_perf_pct;
1215 		}
1216 
1217 		freq = DIV_ROUND_UP(turbo_max * perf_pct, 100);
1218 		freq *= cpu->pstate.scaling;
1219 
1220 		if (freq_qos_update_request(req, freq) < 0)
1221 			pr_warn("Failed to update freq constraint: CPU%d\n", i);
1222 	}
1223 }
1224 
1225 static ssize_t store_max_perf_pct(struct kobject *a, struct kobj_attribute *b,
1226 				  const char *buf, size_t count)
1227 {
1228 	unsigned int input;
1229 	int ret;
1230 
1231 	ret = sscanf(buf, "%u", &input);
1232 	if (ret != 1)
1233 		return -EINVAL;
1234 
1235 	mutex_lock(&intel_pstate_driver_lock);
1236 
1237 	if (!intel_pstate_driver) {
1238 		mutex_unlock(&intel_pstate_driver_lock);
1239 		return -EAGAIN;
1240 	}
1241 
1242 	mutex_lock(&intel_pstate_limits_lock);
1243 
1244 	global.max_perf_pct = clamp_t(int, input, global.min_perf_pct, 100);
1245 
1246 	mutex_unlock(&intel_pstate_limits_lock);
1247 
1248 	if (intel_pstate_driver == &intel_pstate)
1249 		intel_pstate_update_policies();
1250 	else
1251 		update_qos_request(FREQ_QOS_MAX);
1252 
1253 	mutex_unlock(&intel_pstate_driver_lock);
1254 
1255 	return count;
1256 }
1257 
1258 static ssize_t store_min_perf_pct(struct kobject *a, struct kobj_attribute *b,
1259 				  const char *buf, size_t count)
1260 {
1261 	unsigned int input;
1262 	int ret;
1263 
1264 	ret = sscanf(buf, "%u", &input);
1265 	if (ret != 1)
1266 		return -EINVAL;
1267 
1268 	mutex_lock(&intel_pstate_driver_lock);
1269 
1270 	if (!intel_pstate_driver) {
1271 		mutex_unlock(&intel_pstate_driver_lock);
1272 		return -EAGAIN;
1273 	}
1274 
1275 	mutex_lock(&intel_pstate_limits_lock);
1276 
1277 	global.min_perf_pct = clamp_t(int, input,
1278 				      min_perf_pct_min(), global.max_perf_pct);
1279 
1280 	mutex_unlock(&intel_pstate_limits_lock);
1281 
1282 	if (intel_pstate_driver == &intel_pstate)
1283 		intel_pstate_update_policies();
1284 	else
1285 		update_qos_request(FREQ_QOS_MIN);
1286 
1287 	mutex_unlock(&intel_pstate_driver_lock);
1288 
1289 	return count;
1290 }
1291 
1292 static ssize_t show_hwp_dynamic_boost(struct kobject *kobj,
1293 				struct kobj_attribute *attr, char *buf)
1294 {
1295 	return sprintf(buf, "%u\n", hwp_boost);
1296 }
1297 
1298 static ssize_t store_hwp_dynamic_boost(struct kobject *a,
1299 				       struct kobj_attribute *b,
1300 				       const char *buf, size_t count)
1301 {
1302 	unsigned int input;
1303 	int ret;
1304 
1305 	ret = kstrtouint(buf, 10, &input);
1306 	if (ret)
1307 		return ret;
1308 
1309 	mutex_lock(&intel_pstate_driver_lock);
1310 	hwp_boost = !!input;
1311 	intel_pstate_update_policies();
1312 	mutex_unlock(&intel_pstate_driver_lock);
1313 
1314 	return count;
1315 }
1316 
1317 static ssize_t show_energy_efficiency(struct kobject *kobj, struct kobj_attribute *attr,
1318 				      char *buf)
1319 {
1320 	u64 power_ctl;
1321 	int enable;
1322 
1323 	rdmsrl(MSR_IA32_POWER_CTL, power_ctl);
1324 	enable = !!(power_ctl & BIT(MSR_IA32_POWER_CTL_BIT_EE));
1325 	return sprintf(buf, "%d\n", !enable);
1326 }
1327 
1328 static ssize_t store_energy_efficiency(struct kobject *a, struct kobj_attribute *b,
1329 				       const char *buf, size_t count)
1330 {
1331 	bool input;
1332 	int ret;
1333 
1334 	ret = kstrtobool(buf, &input);
1335 	if (ret)
1336 		return ret;
1337 
1338 	set_power_ctl_ee_state(input);
1339 
1340 	return count;
1341 }
1342 
1343 show_one(max_perf_pct, max_perf_pct);
1344 show_one(min_perf_pct, min_perf_pct);
1345 
1346 define_one_global_rw(status);
1347 define_one_global_rw(no_turbo);
1348 define_one_global_rw(max_perf_pct);
1349 define_one_global_rw(min_perf_pct);
1350 define_one_global_ro(turbo_pct);
1351 define_one_global_ro(num_pstates);
1352 define_one_global_rw(hwp_dynamic_boost);
1353 define_one_global_rw(energy_efficiency);
1354 
1355 static struct attribute *intel_pstate_attributes[] = {
1356 	&status.attr,
1357 	&no_turbo.attr,
1358 	&turbo_pct.attr,
1359 	&num_pstates.attr,
1360 	NULL
1361 };
1362 
1363 static const struct attribute_group intel_pstate_attr_group = {
1364 	.attrs = intel_pstate_attributes,
1365 };
1366 
1367 static const struct x86_cpu_id intel_pstate_cpu_ee_disable_ids[];
1368 
1369 static struct kobject *intel_pstate_kobject;
1370 
1371 static void __init intel_pstate_sysfs_expose_params(void)
1372 {
1373 	int rc;
1374 
1375 	intel_pstate_kobject = kobject_create_and_add("intel_pstate",
1376 						&cpu_subsys.dev_root->kobj);
1377 	if (WARN_ON(!intel_pstate_kobject))
1378 		return;
1379 
1380 	rc = sysfs_create_group(intel_pstate_kobject, &intel_pstate_attr_group);
1381 	if (WARN_ON(rc))
1382 		return;
1383 
1384 	/*
1385 	 * If per cpu limits are enforced there are no global limits, so
1386 	 * return without creating max/min_perf_pct attributes
1387 	 */
1388 	if (per_cpu_limits)
1389 		return;
1390 
1391 	rc = sysfs_create_file(intel_pstate_kobject, &max_perf_pct.attr);
1392 	WARN_ON(rc);
1393 
1394 	rc = sysfs_create_file(intel_pstate_kobject, &min_perf_pct.attr);
1395 	WARN_ON(rc);
1396 
1397 	if (x86_match_cpu(intel_pstate_cpu_ee_disable_ids)) {
1398 		rc = sysfs_create_file(intel_pstate_kobject, &energy_efficiency.attr);
1399 		WARN_ON(rc);
1400 	}
1401 }
1402 
1403 static void intel_pstate_sysfs_expose_hwp_dynamic_boost(void)
1404 {
1405 	int rc;
1406 
1407 	if (!hwp_active)
1408 		return;
1409 
1410 	rc = sysfs_create_file(intel_pstate_kobject, &hwp_dynamic_boost.attr);
1411 	WARN_ON_ONCE(rc);
1412 }
1413 
1414 static void intel_pstate_sysfs_hide_hwp_dynamic_boost(void)
1415 {
1416 	if (!hwp_active)
1417 		return;
1418 
1419 	sysfs_remove_file(intel_pstate_kobject, &hwp_dynamic_boost.attr);
1420 }
1421 
1422 /************************** sysfs end ************************/
1423 
1424 static void intel_pstate_hwp_enable(struct cpudata *cpudata)
1425 {
1426 	/* First disable HWP notification interrupt as we don't process them */
1427 	if (boot_cpu_has(X86_FEATURE_HWP_NOTIFY))
1428 		wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_INTERRUPT, 0x00);
1429 
1430 	wrmsrl_on_cpu(cpudata->cpu, MSR_PM_ENABLE, 0x1);
1431 	cpudata->epp_policy = 0;
1432 	if (cpudata->epp_default == -EINVAL)
1433 		cpudata->epp_default = intel_pstate_get_epp(cpudata, 0);
1434 }
1435 
1436 static int atom_get_min_pstate(void)
1437 {
1438 	u64 value;
1439 
1440 	rdmsrl(MSR_ATOM_CORE_RATIOS, value);
1441 	return (value >> 8) & 0x7F;
1442 }
1443 
1444 static int atom_get_max_pstate(void)
1445 {
1446 	u64 value;
1447 
1448 	rdmsrl(MSR_ATOM_CORE_RATIOS, value);
1449 	return (value >> 16) & 0x7F;
1450 }
1451 
1452 static int atom_get_turbo_pstate(void)
1453 {
1454 	u64 value;
1455 
1456 	rdmsrl(MSR_ATOM_CORE_TURBO_RATIOS, value);
1457 	return value & 0x7F;
1458 }
1459 
1460 static u64 atom_get_val(struct cpudata *cpudata, int pstate)
1461 {
1462 	u64 val;
1463 	int32_t vid_fp;
1464 	u32 vid;
1465 
1466 	val = (u64)pstate << 8;
1467 	if (global.no_turbo && !global.turbo_disabled)
1468 		val |= (u64)1 << 32;
1469 
1470 	vid_fp = cpudata->vid.min + mul_fp(
1471 		int_tofp(pstate - cpudata->pstate.min_pstate),
1472 		cpudata->vid.ratio);
1473 
1474 	vid_fp = clamp_t(int32_t, vid_fp, cpudata->vid.min, cpudata->vid.max);
1475 	vid = ceiling_fp(vid_fp);
1476 
1477 	if (pstate > cpudata->pstate.max_pstate)
1478 		vid = cpudata->vid.turbo;
1479 
1480 	return val | vid;
1481 }
1482 
1483 static int silvermont_get_scaling(void)
1484 {
1485 	u64 value;
1486 	int i;
1487 	/* Defined in Table 35-6 from SDM (Sept 2015) */
1488 	static int silvermont_freq_table[] = {
1489 		83300, 100000, 133300, 116700, 80000};
1490 
1491 	rdmsrl(MSR_FSB_FREQ, value);
1492 	i = value & 0x7;
1493 	WARN_ON(i > 4);
1494 
1495 	return silvermont_freq_table[i];
1496 }
1497 
1498 static int airmont_get_scaling(void)
1499 {
1500 	u64 value;
1501 	int i;
1502 	/* Defined in Table 35-10 from SDM (Sept 2015) */
1503 	static int airmont_freq_table[] = {
1504 		83300, 100000, 133300, 116700, 80000,
1505 		93300, 90000, 88900, 87500};
1506 
1507 	rdmsrl(MSR_FSB_FREQ, value);
1508 	i = value & 0xF;
1509 	WARN_ON(i > 8);
1510 
1511 	return airmont_freq_table[i];
1512 }
1513 
1514 static void atom_get_vid(struct cpudata *cpudata)
1515 {
1516 	u64 value;
1517 
1518 	rdmsrl(MSR_ATOM_CORE_VIDS, value);
1519 	cpudata->vid.min = int_tofp((value >> 8) & 0x7f);
1520 	cpudata->vid.max = int_tofp((value >> 16) & 0x7f);
1521 	cpudata->vid.ratio = div_fp(
1522 		cpudata->vid.max - cpudata->vid.min,
1523 		int_tofp(cpudata->pstate.max_pstate -
1524 			cpudata->pstate.min_pstate));
1525 
1526 	rdmsrl(MSR_ATOM_CORE_TURBO_VIDS, value);
1527 	cpudata->vid.turbo = value & 0x7f;
1528 }
1529 
1530 static int core_get_min_pstate(void)
1531 {
1532 	u64 value;
1533 
1534 	rdmsrl(MSR_PLATFORM_INFO, value);
1535 	return (value >> 40) & 0xFF;
1536 }
1537 
1538 static int core_get_max_pstate_physical(void)
1539 {
1540 	u64 value;
1541 
1542 	rdmsrl(MSR_PLATFORM_INFO, value);
1543 	return (value >> 8) & 0xFF;
1544 }
1545 
1546 static int core_get_tdp_ratio(u64 plat_info)
1547 {
1548 	/* Check how many TDP levels present */
1549 	if (plat_info & 0x600000000) {
1550 		u64 tdp_ctrl;
1551 		u64 tdp_ratio;
1552 		int tdp_msr;
1553 		int err;
1554 
1555 		/* Get the TDP level (0, 1, 2) to get ratios */
1556 		err = rdmsrl_safe(MSR_CONFIG_TDP_CONTROL, &tdp_ctrl);
1557 		if (err)
1558 			return err;
1559 
1560 		/* TDP MSR are continuous starting at 0x648 */
1561 		tdp_msr = MSR_CONFIG_TDP_NOMINAL + (tdp_ctrl & 0x03);
1562 		err = rdmsrl_safe(tdp_msr, &tdp_ratio);
1563 		if (err)
1564 			return err;
1565 
1566 		/* For level 1 and 2, bits[23:16] contain the ratio */
1567 		if (tdp_ctrl & 0x03)
1568 			tdp_ratio >>= 16;
1569 
1570 		tdp_ratio &= 0xff; /* ratios are only 8 bits long */
1571 		pr_debug("tdp_ratio %x\n", (int)tdp_ratio);
1572 
1573 		return (int)tdp_ratio;
1574 	}
1575 
1576 	return -ENXIO;
1577 }
1578 
1579 static int core_get_max_pstate(void)
1580 {
1581 	u64 tar;
1582 	u64 plat_info;
1583 	int max_pstate;
1584 	int tdp_ratio;
1585 	int err;
1586 
1587 	rdmsrl(MSR_PLATFORM_INFO, plat_info);
1588 	max_pstate = (plat_info >> 8) & 0xFF;
1589 
1590 	tdp_ratio = core_get_tdp_ratio(plat_info);
1591 	if (tdp_ratio <= 0)
1592 		return max_pstate;
1593 
1594 	if (hwp_active) {
1595 		/* Turbo activation ratio is not used on HWP platforms */
1596 		return tdp_ratio;
1597 	}
1598 
1599 	err = rdmsrl_safe(MSR_TURBO_ACTIVATION_RATIO, &tar);
1600 	if (!err) {
1601 		int tar_levels;
1602 
1603 		/* Do some sanity checking for safety */
1604 		tar_levels = tar & 0xff;
1605 		if (tdp_ratio - 1 == tar_levels) {
1606 			max_pstate = tar_levels;
1607 			pr_debug("max_pstate=TAC %x\n", max_pstate);
1608 		}
1609 	}
1610 
1611 	return max_pstate;
1612 }
1613 
1614 static int core_get_turbo_pstate(void)
1615 {
1616 	u64 value;
1617 	int nont, ret;
1618 
1619 	rdmsrl(MSR_TURBO_RATIO_LIMIT, value);
1620 	nont = core_get_max_pstate();
1621 	ret = (value) & 255;
1622 	if (ret <= nont)
1623 		ret = nont;
1624 	return ret;
1625 }
1626 
1627 static inline int core_get_scaling(void)
1628 {
1629 	return 100000;
1630 }
1631 
1632 static u64 core_get_val(struct cpudata *cpudata, int pstate)
1633 {
1634 	u64 val;
1635 
1636 	val = (u64)pstate << 8;
1637 	if (global.no_turbo && !global.turbo_disabled)
1638 		val |= (u64)1 << 32;
1639 
1640 	return val;
1641 }
1642 
1643 static int knl_get_aperf_mperf_shift(void)
1644 {
1645 	return 10;
1646 }
1647 
1648 static int knl_get_turbo_pstate(void)
1649 {
1650 	u64 value;
1651 	int nont, ret;
1652 
1653 	rdmsrl(MSR_TURBO_RATIO_LIMIT, value);
1654 	nont = core_get_max_pstate();
1655 	ret = (((value) >> 8) & 0xFF);
1656 	if (ret <= nont)
1657 		ret = nont;
1658 	return ret;
1659 }
1660 
1661 static void intel_pstate_set_pstate(struct cpudata *cpu, int pstate)
1662 {
1663 	trace_cpu_frequency(pstate * cpu->pstate.scaling, cpu->cpu);
1664 	cpu->pstate.current_pstate = pstate;
1665 	/*
1666 	 * Generally, there is no guarantee that this code will always run on
1667 	 * the CPU being updated, so force the register update to run on the
1668 	 * right CPU.
1669 	 */
1670 	wrmsrl_on_cpu(cpu->cpu, MSR_IA32_PERF_CTL,
1671 		      pstate_funcs.get_val(cpu, pstate));
1672 }
1673 
1674 static void intel_pstate_set_min_pstate(struct cpudata *cpu)
1675 {
1676 	intel_pstate_set_pstate(cpu, cpu->pstate.min_pstate);
1677 }
1678 
1679 static void intel_pstate_max_within_limits(struct cpudata *cpu)
1680 {
1681 	int pstate = max(cpu->pstate.min_pstate, cpu->max_perf_ratio);
1682 
1683 	update_turbo_state();
1684 	intel_pstate_set_pstate(cpu, pstate);
1685 }
1686 
1687 static void intel_pstate_get_cpu_pstates(struct cpudata *cpu)
1688 {
1689 	cpu->pstate.min_pstate = pstate_funcs.get_min();
1690 	cpu->pstate.max_pstate = pstate_funcs.get_max();
1691 	cpu->pstate.max_pstate_physical = pstate_funcs.get_max_physical();
1692 	cpu->pstate.turbo_pstate = pstate_funcs.get_turbo();
1693 	cpu->pstate.scaling = pstate_funcs.get_scaling();
1694 	cpu->pstate.max_freq = cpu->pstate.max_pstate * cpu->pstate.scaling;
1695 
1696 	if (hwp_active && !hwp_mode_bdw) {
1697 		unsigned int phy_max, current_max;
1698 
1699 		intel_pstate_get_hwp_max(cpu->cpu, &phy_max, &current_max);
1700 		cpu->pstate.turbo_freq = phy_max * cpu->pstate.scaling;
1701 		cpu->pstate.turbo_pstate = phy_max;
1702 	} else {
1703 		cpu->pstate.turbo_freq = cpu->pstate.turbo_pstate * cpu->pstate.scaling;
1704 	}
1705 
1706 	if (pstate_funcs.get_aperf_mperf_shift)
1707 		cpu->aperf_mperf_shift = pstate_funcs.get_aperf_mperf_shift();
1708 
1709 	if (pstate_funcs.get_vid)
1710 		pstate_funcs.get_vid(cpu);
1711 
1712 	intel_pstate_set_min_pstate(cpu);
1713 }
1714 
1715 /*
1716  * Long hold time will keep high perf limits for long time,
1717  * which negatively impacts perf/watt for some workloads,
1718  * like specpower. 3ms is based on experiements on some
1719  * workoads.
1720  */
1721 static int hwp_boost_hold_time_ns = 3 * NSEC_PER_MSEC;
1722 
1723 static inline void intel_pstate_hwp_boost_up(struct cpudata *cpu)
1724 {
1725 	u64 hwp_req = READ_ONCE(cpu->hwp_req_cached);
1726 	u32 max_limit = (hwp_req & 0xff00) >> 8;
1727 	u32 min_limit = (hwp_req & 0xff);
1728 	u32 boost_level1;
1729 
1730 	/*
1731 	 * Cases to consider (User changes via sysfs or boot time):
1732 	 * If, P0 (Turbo max) = P1 (Guaranteed max) = min:
1733 	 *	No boost, return.
1734 	 * If, P0 (Turbo max) > P1 (Guaranteed max) = min:
1735 	 *     Should result in one level boost only for P0.
1736 	 * If, P0 (Turbo max) = P1 (Guaranteed max) > min:
1737 	 *     Should result in two level boost:
1738 	 *         (min + p1)/2 and P1.
1739 	 * If, P0 (Turbo max) > P1 (Guaranteed max) > min:
1740 	 *     Should result in three level boost:
1741 	 *        (min + p1)/2, P1 and P0.
1742 	 */
1743 
1744 	/* If max and min are equal or already at max, nothing to boost */
1745 	if (max_limit == min_limit || cpu->hwp_boost_min >= max_limit)
1746 		return;
1747 
1748 	if (!cpu->hwp_boost_min)
1749 		cpu->hwp_boost_min = min_limit;
1750 
1751 	/* level at half way mark between min and guranteed */
1752 	boost_level1 = (HWP_GUARANTEED_PERF(cpu->hwp_cap_cached) + min_limit) >> 1;
1753 
1754 	if (cpu->hwp_boost_min < boost_level1)
1755 		cpu->hwp_boost_min = boost_level1;
1756 	else if (cpu->hwp_boost_min < HWP_GUARANTEED_PERF(cpu->hwp_cap_cached))
1757 		cpu->hwp_boost_min = HWP_GUARANTEED_PERF(cpu->hwp_cap_cached);
1758 	else if (cpu->hwp_boost_min == HWP_GUARANTEED_PERF(cpu->hwp_cap_cached) &&
1759 		 max_limit != HWP_GUARANTEED_PERF(cpu->hwp_cap_cached))
1760 		cpu->hwp_boost_min = max_limit;
1761 	else
1762 		return;
1763 
1764 	hwp_req = (hwp_req & ~GENMASK_ULL(7, 0)) | cpu->hwp_boost_min;
1765 	wrmsrl(MSR_HWP_REQUEST, hwp_req);
1766 	cpu->last_update = cpu->sample.time;
1767 }
1768 
1769 static inline void intel_pstate_hwp_boost_down(struct cpudata *cpu)
1770 {
1771 	if (cpu->hwp_boost_min) {
1772 		bool expired;
1773 
1774 		/* Check if we are idle for hold time to boost down */
1775 		expired = time_after64(cpu->sample.time, cpu->last_update +
1776 				       hwp_boost_hold_time_ns);
1777 		if (expired) {
1778 			wrmsrl(MSR_HWP_REQUEST, cpu->hwp_req_cached);
1779 			cpu->hwp_boost_min = 0;
1780 		}
1781 	}
1782 	cpu->last_update = cpu->sample.time;
1783 }
1784 
1785 static inline void intel_pstate_update_util_hwp_local(struct cpudata *cpu,
1786 						      u64 time)
1787 {
1788 	cpu->sample.time = time;
1789 
1790 	if (cpu->sched_flags & SCHED_CPUFREQ_IOWAIT) {
1791 		bool do_io = false;
1792 
1793 		cpu->sched_flags = 0;
1794 		/*
1795 		 * Set iowait_boost flag and update time. Since IO WAIT flag
1796 		 * is set all the time, we can't just conclude that there is
1797 		 * some IO bound activity is scheduled on this CPU with just
1798 		 * one occurrence. If we receive at least two in two
1799 		 * consecutive ticks, then we treat as boost candidate.
1800 		 */
1801 		if (time_before64(time, cpu->last_io_update + 2 * TICK_NSEC))
1802 			do_io = true;
1803 
1804 		cpu->last_io_update = time;
1805 
1806 		if (do_io)
1807 			intel_pstate_hwp_boost_up(cpu);
1808 
1809 	} else {
1810 		intel_pstate_hwp_boost_down(cpu);
1811 	}
1812 }
1813 
1814 static inline void intel_pstate_update_util_hwp(struct update_util_data *data,
1815 						u64 time, unsigned int flags)
1816 {
1817 	struct cpudata *cpu = container_of(data, struct cpudata, update_util);
1818 
1819 	cpu->sched_flags |= flags;
1820 
1821 	if (smp_processor_id() == cpu->cpu)
1822 		intel_pstate_update_util_hwp_local(cpu, time);
1823 }
1824 
1825 static inline void intel_pstate_calc_avg_perf(struct cpudata *cpu)
1826 {
1827 	struct sample *sample = &cpu->sample;
1828 
1829 	sample->core_avg_perf = div_ext_fp(sample->aperf, sample->mperf);
1830 }
1831 
1832 static inline bool intel_pstate_sample(struct cpudata *cpu, u64 time)
1833 {
1834 	u64 aperf, mperf;
1835 	unsigned long flags;
1836 	u64 tsc;
1837 
1838 	local_irq_save(flags);
1839 	rdmsrl(MSR_IA32_APERF, aperf);
1840 	rdmsrl(MSR_IA32_MPERF, mperf);
1841 	tsc = rdtsc();
1842 	if (cpu->prev_mperf == mperf || cpu->prev_tsc == tsc) {
1843 		local_irq_restore(flags);
1844 		return false;
1845 	}
1846 	local_irq_restore(flags);
1847 
1848 	cpu->last_sample_time = cpu->sample.time;
1849 	cpu->sample.time = time;
1850 	cpu->sample.aperf = aperf;
1851 	cpu->sample.mperf = mperf;
1852 	cpu->sample.tsc =  tsc;
1853 	cpu->sample.aperf -= cpu->prev_aperf;
1854 	cpu->sample.mperf -= cpu->prev_mperf;
1855 	cpu->sample.tsc -= cpu->prev_tsc;
1856 
1857 	cpu->prev_aperf = aperf;
1858 	cpu->prev_mperf = mperf;
1859 	cpu->prev_tsc = tsc;
1860 	/*
1861 	 * First time this function is invoked in a given cycle, all of the
1862 	 * previous sample data fields are equal to zero or stale and they must
1863 	 * be populated with meaningful numbers for things to work, so assume
1864 	 * that sample.time will always be reset before setting the utilization
1865 	 * update hook and make the caller skip the sample then.
1866 	 */
1867 	if (cpu->last_sample_time) {
1868 		intel_pstate_calc_avg_perf(cpu);
1869 		return true;
1870 	}
1871 	return false;
1872 }
1873 
1874 static inline int32_t get_avg_frequency(struct cpudata *cpu)
1875 {
1876 	return mul_ext_fp(cpu->sample.core_avg_perf, cpu_khz);
1877 }
1878 
1879 static inline int32_t get_avg_pstate(struct cpudata *cpu)
1880 {
1881 	return mul_ext_fp(cpu->pstate.max_pstate_physical,
1882 			  cpu->sample.core_avg_perf);
1883 }
1884 
1885 static inline int32_t get_target_pstate(struct cpudata *cpu)
1886 {
1887 	struct sample *sample = &cpu->sample;
1888 	int32_t busy_frac;
1889 	int target, avg_pstate;
1890 
1891 	busy_frac = div_fp(sample->mperf << cpu->aperf_mperf_shift,
1892 			   sample->tsc);
1893 
1894 	if (busy_frac < cpu->iowait_boost)
1895 		busy_frac = cpu->iowait_boost;
1896 
1897 	sample->busy_scaled = busy_frac * 100;
1898 
1899 	target = global.no_turbo || global.turbo_disabled ?
1900 			cpu->pstate.max_pstate : cpu->pstate.turbo_pstate;
1901 	target += target >> 2;
1902 	target = mul_fp(target, busy_frac);
1903 	if (target < cpu->pstate.min_pstate)
1904 		target = cpu->pstate.min_pstate;
1905 
1906 	/*
1907 	 * If the average P-state during the previous cycle was higher than the
1908 	 * current target, add 50% of the difference to the target to reduce
1909 	 * possible performance oscillations and offset possible performance
1910 	 * loss related to moving the workload from one CPU to another within
1911 	 * a package/module.
1912 	 */
1913 	avg_pstate = get_avg_pstate(cpu);
1914 	if (avg_pstate > target)
1915 		target += (avg_pstate - target) >> 1;
1916 
1917 	return target;
1918 }
1919 
1920 static int intel_pstate_prepare_request(struct cpudata *cpu, int pstate)
1921 {
1922 	int min_pstate = max(cpu->pstate.min_pstate, cpu->min_perf_ratio);
1923 	int max_pstate = max(min_pstate, cpu->max_perf_ratio);
1924 
1925 	return clamp_t(int, pstate, min_pstate, max_pstate);
1926 }
1927 
1928 static void intel_pstate_update_pstate(struct cpudata *cpu, int pstate)
1929 {
1930 	if (pstate == cpu->pstate.current_pstate)
1931 		return;
1932 
1933 	cpu->pstate.current_pstate = pstate;
1934 	wrmsrl(MSR_IA32_PERF_CTL, pstate_funcs.get_val(cpu, pstate));
1935 }
1936 
1937 static void intel_pstate_adjust_pstate(struct cpudata *cpu)
1938 {
1939 	int from = cpu->pstate.current_pstate;
1940 	struct sample *sample;
1941 	int target_pstate;
1942 
1943 	update_turbo_state();
1944 
1945 	target_pstate = get_target_pstate(cpu);
1946 	target_pstate = intel_pstate_prepare_request(cpu, target_pstate);
1947 	trace_cpu_frequency(target_pstate * cpu->pstate.scaling, cpu->cpu);
1948 	intel_pstate_update_pstate(cpu, target_pstate);
1949 
1950 	sample = &cpu->sample;
1951 	trace_pstate_sample(mul_ext_fp(100, sample->core_avg_perf),
1952 		fp_toint(sample->busy_scaled),
1953 		from,
1954 		cpu->pstate.current_pstate,
1955 		sample->mperf,
1956 		sample->aperf,
1957 		sample->tsc,
1958 		get_avg_frequency(cpu),
1959 		fp_toint(cpu->iowait_boost * 100));
1960 }
1961 
1962 static void intel_pstate_update_util(struct update_util_data *data, u64 time,
1963 				     unsigned int flags)
1964 {
1965 	struct cpudata *cpu = container_of(data, struct cpudata, update_util);
1966 	u64 delta_ns;
1967 
1968 	/* Don't allow remote callbacks */
1969 	if (smp_processor_id() != cpu->cpu)
1970 		return;
1971 
1972 	delta_ns = time - cpu->last_update;
1973 	if (flags & SCHED_CPUFREQ_IOWAIT) {
1974 		/* Start over if the CPU may have been idle. */
1975 		if (delta_ns > TICK_NSEC) {
1976 			cpu->iowait_boost = ONE_EIGHTH_FP;
1977 		} else if (cpu->iowait_boost >= ONE_EIGHTH_FP) {
1978 			cpu->iowait_boost <<= 1;
1979 			if (cpu->iowait_boost > int_tofp(1))
1980 				cpu->iowait_boost = int_tofp(1);
1981 		} else {
1982 			cpu->iowait_boost = ONE_EIGHTH_FP;
1983 		}
1984 	} else if (cpu->iowait_boost) {
1985 		/* Clear iowait_boost if the CPU may have been idle. */
1986 		if (delta_ns > TICK_NSEC)
1987 			cpu->iowait_boost = 0;
1988 		else
1989 			cpu->iowait_boost >>= 1;
1990 	}
1991 	cpu->last_update = time;
1992 	delta_ns = time - cpu->sample.time;
1993 	if ((s64)delta_ns < INTEL_PSTATE_SAMPLING_INTERVAL)
1994 		return;
1995 
1996 	if (intel_pstate_sample(cpu, time))
1997 		intel_pstate_adjust_pstate(cpu);
1998 }
1999 
2000 static struct pstate_funcs core_funcs = {
2001 	.get_max = core_get_max_pstate,
2002 	.get_max_physical = core_get_max_pstate_physical,
2003 	.get_min = core_get_min_pstate,
2004 	.get_turbo = core_get_turbo_pstate,
2005 	.get_scaling = core_get_scaling,
2006 	.get_val = core_get_val,
2007 };
2008 
2009 static const struct pstate_funcs silvermont_funcs = {
2010 	.get_max = atom_get_max_pstate,
2011 	.get_max_physical = atom_get_max_pstate,
2012 	.get_min = atom_get_min_pstate,
2013 	.get_turbo = atom_get_turbo_pstate,
2014 	.get_val = atom_get_val,
2015 	.get_scaling = silvermont_get_scaling,
2016 	.get_vid = atom_get_vid,
2017 };
2018 
2019 static const struct pstate_funcs airmont_funcs = {
2020 	.get_max = atom_get_max_pstate,
2021 	.get_max_physical = atom_get_max_pstate,
2022 	.get_min = atom_get_min_pstate,
2023 	.get_turbo = atom_get_turbo_pstate,
2024 	.get_val = atom_get_val,
2025 	.get_scaling = airmont_get_scaling,
2026 	.get_vid = atom_get_vid,
2027 };
2028 
2029 static const struct pstate_funcs knl_funcs = {
2030 	.get_max = core_get_max_pstate,
2031 	.get_max_physical = core_get_max_pstate_physical,
2032 	.get_min = core_get_min_pstate,
2033 	.get_turbo = knl_get_turbo_pstate,
2034 	.get_aperf_mperf_shift = knl_get_aperf_mperf_shift,
2035 	.get_scaling = core_get_scaling,
2036 	.get_val = core_get_val,
2037 };
2038 
2039 #define X86_MATCH(model, policy)					 \
2040 	X86_MATCH_VENDOR_FAM_MODEL_FEATURE(INTEL, 6, INTEL_FAM6_##model, \
2041 					   X86_FEATURE_APERFMPERF, &policy)
2042 
2043 static const struct x86_cpu_id intel_pstate_cpu_ids[] = {
2044 	X86_MATCH(SANDYBRIDGE,		core_funcs),
2045 	X86_MATCH(SANDYBRIDGE_X,	core_funcs),
2046 	X86_MATCH(ATOM_SILVERMONT,	silvermont_funcs),
2047 	X86_MATCH(IVYBRIDGE,		core_funcs),
2048 	X86_MATCH(HASWELL,		core_funcs),
2049 	X86_MATCH(BROADWELL,		core_funcs),
2050 	X86_MATCH(IVYBRIDGE_X,		core_funcs),
2051 	X86_MATCH(HASWELL_X,		core_funcs),
2052 	X86_MATCH(HASWELL_L,		core_funcs),
2053 	X86_MATCH(HASWELL_G,		core_funcs),
2054 	X86_MATCH(BROADWELL_G,		core_funcs),
2055 	X86_MATCH(ATOM_AIRMONT,		airmont_funcs),
2056 	X86_MATCH(SKYLAKE_L,		core_funcs),
2057 	X86_MATCH(BROADWELL_X,		core_funcs),
2058 	X86_MATCH(SKYLAKE,		core_funcs),
2059 	X86_MATCH(BROADWELL_D,		core_funcs),
2060 	X86_MATCH(XEON_PHI_KNL,		knl_funcs),
2061 	X86_MATCH(XEON_PHI_KNM,		knl_funcs),
2062 	X86_MATCH(ATOM_GOLDMONT,	core_funcs),
2063 	X86_MATCH(ATOM_GOLDMONT_PLUS,	core_funcs),
2064 	X86_MATCH(SKYLAKE_X,		core_funcs),
2065 	{}
2066 };
2067 MODULE_DEVICE_TABLE(x86cpu, intel_pstate_cpu_ids);
2068 
2069 static const struct x86_cpu_id intel_pstate_cpu_oob_ids[] __initconst = {
2070 	X86_MATCH(BROADWELL_D,		core_funcs),
2071 	X86_MATCH(BROADWELL_X,		core_funcs),
2072 	X86_MATCH(SKYLAKE_X,		core_funcs),
2073 	{}
2074 };
2075 
2076 static const struct x86_cpu_id intel_pstate_cpu_ee_disable_ids[] = {
2077 	X86_MATCH(KABYLAKE,		core_funcs),
2078 	{}
2079 };
2080 
2081 static const struct x86_cpu_id intel_pstate_hwp_boost_ids[] = {
2082 	X86_MATCH(SKYLAKE_X,		core_funcs),
2083 	X86_MATCH(SKYLAKE,		core_funcs),
2084 	{}
2085 };
2086 
2087 static int intel_pstate_init_cpu(unsigned int cpunum)
2088 {
2089 	struct cpudata *cpu;
2090 
2091 	cpu = all_cpu_data[cpunum];
2092 
2093 	if (!cpu) {
2094 		cpu = kzalloc(sizeof(*cpu), GFP_KERNEL);
2095 		if (!cpu)
2096 			return -ENOMEM;
2097 
2098 		all_cpu_data[cpunum] = cpu;
2099 
2100 		cpu->epp_default = -EINVAL;
2101 		cpu->epp_powersave = -EINVAL;
2102 		cpu->epp_saved = -EINVAL;
2103 	}
2104 
2105 	cpu = all_cpu_data[cpunum];
2106 
2107 	cpu->cpu = cpunum;
2108 
2109 	if (hwp_active) {
2110 		const struct x86_cpu_id *id;
2111 
2112 		intel_pstate_hwp_enable(cpu);
2113 
2114 		id = x86_match_cpu(intel_pstate_hwp_boost_ids);
2115 		if (id && intel_pstate_acpi_pm_profile_server())
2116 			hwp_boost = true;
2117 	}
2118 
2119 	intel_pstate_get_cpu_pstates(cpu);
2120 
2121 	pr_debug("controlling: cpu %d\n", cpunum);
2122 
2123 	return 0;
2124 }
2125 
2126 static void intel_pstate_set_update_util_hook(unsigned int cpu_num)
2127 {
2128 	struct cpudata *cpu = all_cpu_data[cpu_num];
2129 
2130 	if (hwp_active && !hwp_boost)
2131 		return;
2132 
2133 	if (cpu->update_util_set)
2134 		return;
2135 
2136 	/* Prevent intel_pstate_update_util() from using stale data. */
2137 	cpu->sample.time = 0;
2138 	cpufreq_add_update_util_hook(cpu_num, &cpu->update_util,
2139 				     (hwp_active ?
2140 				      intel_pstate_update_util_hwp :
2141 				      intel_pstate_update_util));
2142 	cpu->update_util_set = true;
2143 }
2144 
2145 static void intel_pstate_clear_update_util_hook(unsigned int cpu)
2146 {
2147 	struct cpudata *cpu_data = all_cpu_data[cpu];
2148 
2149 	if (!cpu_data->update_util_set)
2150 		return;
2151 
2152 	cpufreq_remove_update_util_hook(cpu);
2153 	cpu_data->update_util_set = false;
2154 	synchronize_rcu();
2155 }
2156 
2157 static int intel_pstate_get_max_freq(struct cpudata *cpu)
2158 {
2159 	return global.turbo_disabled || global.no_turbo ?
2160 			cpu->pstate.max_freq : cpu->pstate.turbo_freq;
2161 }
2162 
2163 static void intel_pstate_update_perf_limits(struct cpudata *cpu,
2164 					    unsigned int policy_min,
2165 					    unsigned int policy_max)
2166 {
2167 	int max_freq = intel_pstate_get_max_freq(cpu);
2168 	int32_t max_policy_perf, min_policy_perf;
2169 	int max_state, turbo_max;
2170 
2171 	/*
2172 	 * HWP needs some special consideration, because on BDX the
2173 	 * HWP_REQUEST uses abstract value to represent performance
2174 	 * rather than pure ratios.
2175 	 */
2176 	if (hwp_active) {
2177 		intel_pstate_get_hwp_max(cpu->cpu, &turbo_max, &max_state);
2178 	} else {
2179 		max_state = global.no_turbo || global.turbo_disabled ?
2180 			cpu->pstate.max_pstate : cpu->pstate.turbo_pstate;
2181 		turbo_max = cpu->pstate.turbo_pstate;
2182 	}
2183 
2184 	max_policy_perf = max_state * policy_max / max_freq;
2185 	if (policy_max == policy_min) {
2186 		min_policy_perf = max_policy_perf;
2187 	} else {
2188 		min_policy_perf = max_state * policy_min / max_freq;
2189 		min_policy_perf = clamp_t(int32_t, min_policy_perf,
2190 					  0, max_policy_perf);
2191 	}
2192 
2193 	pr_debug("cpu:%d max_state %d min_policy_perf:%d max_policy_perf:%d\n",
2194 		 cpu->cpu, max_state, min_policy_perf, max_policy_perf);
2195 
2196 	/* Normalize user input to [min_perf, max_perf] */
2197 	if (per_cpu_limits) {
2198 		cpu->min_perf_ratio = min_policy_perf;
2199 		cpu->max_perf_ratio = max_policy_perf;
2200 	} else {
2201 		int32_t global_min, global_max;
2202 
2203 		/* Global limits are in percent of the maximum turbo P-state. */
2204 		global_max = DIV_ROUND_UP(turbo_max * global.max_perf_pct, 100);
2205 		global_min = DIV_ROUND_UP(turbo_max * global.min_perf_pct, 100);
2206 		global_min = clamp_t(int32_t, global_min, 0, global_max);
2207 
2208 		pr_debug("cpu:%d global_min:%d global_max:%d\n", cpu->cpu,
2209 			 global_min, global_max);
2210 
2211 		cpu->min_perf_ratio = max(min_policy_perf, global_min);
2212 		cpu->min_perf_ratio = min(cpu->min_perf_ratio, max_policy_perf);
2213 		cpu->max_perf_ratio = min(max_policy_perf, global_max);
2214 		cpu->max_perf_ratio = max(min_policy_perf, cpu->max_perf_ratio);
2215 
2216 		/* Make sure min_perf <= max_perf */
2217 		cpu->min_perf_ratio = min(cpu->min_perf_ratio,
2218 					  cpu->max_perf_ratio);
2219 
2220 	}
2221 	pr_debug("cpu:%d max_perf_ratio:%d min_perf_ratio:%d\n", cpu->cpu,
2222 		 cpu->max_perf_ratio,
2223 		 cpu->min_perf_ratio);
2224 }
2225 
2226 static int intel_pstate_set_policy(struct cpufreq_policy *policy)
2227 {
2228 	struct cpudata *cpu;
2229 
2230 	if (!policy->cpuinfo.max_freq)
2231 		return -ENODEV;
2232 
2233 	pr_debug("set_policy cpuinfo.max %u policy->max %u\n",
2234 		 policy->cpuinfo.max_freq, policy->max);
2235 
2236 	cpu = all_cpu_data[policy->cpu];
2237 	cpu->policy = policy->policy;
2238 
2239 	mutex_lock(&intel_pstate_limits_lock);
2240 
2241 	intel_pstate_update_perf_limits(cpu, policy->min, policy->max);
2242 
2243 	if (cpu->policy == CPUFREQ_POLICY_PERFORMANCE) {
2244 		/*
2245 		 * NOHZ_FULL CPUs need this as the governor callback may not
2246 		 * be invoked on them.
2247 		 */
2248 		intel_pstate_clear_update_util_hook(policy->cpu);
2249 		intel_pstate_max_within_limits(cpu);
2250 	} else {
2251 		intel_pstate_set_update_util_hook(policy->cpu);
2252 	}
2253 
2254 	if (hwp_active) {
2255 		/*
2256 		 * When hwp_boost was active before and dynamically it
2257 		 * was turned off, in that case we need to clear the
2258 		 * update util hook.
2259 		 */
2260 		if (!hwp_boost)
2261 			intel_pstate_clear_update_util_hook(policy->cpu);
2262 		intel_pstate_hwp_set(policy->cpu);
2263 	}
2264 
2265 	mutex_unlock(&intel_pstate_limits_lock);
2266 
2267 	return 0;
2268 }
2269 
2270 static void intel_pstate_adjust_policy_max(struct cpudata *cpu,
2271 					   struct cpufreq_policy_data *policy)
2272 {
2273 	if (!hwp_active &&
2274 	    cpu->pstate.max_pstate_physical > cpu->pstate.max_pstate &&
2275 	    policy->max < policy->cpuinfo.max_freq &&
2276 	    policy->max > cpu->pstate.max_freq) {
2277 		pr_debug("policy->max > max non turbo frequency\n");
2278 		policy->max = policy->cpuinfo.max_freq;
2279 	}
2280 }
2281 
2282 static void intel_pstate_verify_cpu_policy(struct cpudata *cpu,
2283 					   struct cpufreq_policy_data *policy)
2284 {
2285 	update_turbo_state();
2286 	cpufreq_verify_within_limits(policy, policy->cpuinfo.min_freq,
2287 				     intel_pstate_get_max_freq(cpu));
2288 
2289 	intel_pstate_adjust_policy_max(cpu, policy);
2290 }
2291 
2292 static int intel_pstate_verify_policy(struct cpufreq_policy_data *policy)
2293 {
2294 	intel_pstate_verify_cpu_policy(all_cpu_data[policy->cpu], policy);
2295 
2296 	return 0;
2297 }
2298 
2299 static void intel_cpufreq_stop_cpu(struct cpufreq_policy *policy)
2300 {
2301 	if (hwp_active)
2302 		intel_pstate_hwp_force_min_perf(policy->cpu);
2303 	else
2304 		intel_pstate_set_min_pstate(all_cpu_data[policy->cpu]);
2305 }
2306 
2307 static void intel_pstate_stop_cpu(struct cpufreq_policy *policy)
2308 {
2309 	pr_debug("CPU %d exiting\n", policy->cpu);
2310 
2311 	intel_pstate_clear_update_util_hook(policy->cpu);
2312 	if (hwp_active)
2313 		intel_pstate_hwp_save_state(policy);
2314 
2315 	intel_cpufreq_stop_cpu(policy);
2316 }
2317 
2318 static int intel_pstate_cpu_exit(struct cpufreq_policy *policy)
2319 {
2320 	intel_pstate_exit_perf_limits(policy);
2321 
2322 	policy->fast_switch_possible = false;
2323 
2324 	return 0;
2325 }
2326 
2327 static int __intel_pstate_cpu_init(struct cpufreq_policy *policy)
2328 {
2329 	struct cpudata *cpu;
2330 	int rc;
2331 
2332 	rc = intel_pstate_init_cpu(policy->cpu);
2333 	if (rc)
2334 		return rc;
2335 
2336 	cpu = all_cpu_data[policy->cpu];
2337 
2338 	cpu->max_perf_ratio = 0xFF;
2339 	cpu->min_perf_ratio = 0;
2340 
2341 	policy->min = cpu->pstate.min_pstate * cpu->pstate.scaling;
2342 	policy->max = cpu->pstate.turbo_pstate * cpu->pstate.scaling;
2343 
2344 	/* cpuinfo and default policy values */
2345 	policy->cpuinfo.min_freq = cpu->pstate.min_pstate * cpu->pstate.scaling;
2346 	update_turbo_state();
2347 	global.turbo_disabled_mf = global.turbo_disabled;
2348 	policy->cpuinfo.max_freq = global.turbo_disabled ?
2349 			cpu->pstate.max_pstate : cpu->pstate.turbo_pstate;
2350 	policy->cpuinfo.max_freq *= cpu->pstate.scaling;
2351 
2352 	if (hwp_active) {
2353 		unsigned int max_freq;
2354 
2355 		max_freq = global.turbo_disabled ?
2356 			cpu->pstate.max_freq : cpu->pstate.turbo_freq;
2357 		if (max_freq < policy->cpuinfo.max_freq)
2358 			policy->cpuinfo.max_freq = max_freq;
2359 	}
2360 
2361 	intel_pstate_init_acpi_perf_limits(policy);
2362 
2363 	policy->fast_switch_possible = true;
2364 
2365 	return 0;
2366 }
2367 
2368 static int intel_pstate_cpu_init(struct cpufreq_policy *policy)
2369 {
2370 	int ret = __intel_pstate_cpu_init(policy);
2371 
2372 	if (ret)
2373 		return ret;
2374 
2375 	/*
2376 	 * Set the policy to powersave to provide a valid fallback value in case
2377 	 * the default cpufreq governor is neither powersave nor performance.
2378 	 */
2379 	policy->policy = CPUFREQ_POLICY_POWERSAVE;
2380 
2381 	return 0;
2382 }
2383 
2384 static struct cpufreq_driver intel_pstate = {
2385 	.flags		= CPUFREQ_CONST_LOOPS,
2386 	.verify		= intel_pstate_verify_policy,
2387 	.setpolicy	= intel_pstate_set_policy,
2388 	.suspend	= intel_pstate_hwp_save_state,
2389 	.resume		= intel_pstate_resume,
2390 	.init		= intel_pstate_cpu_init,
2391 	.exit		= intel_pstate_cpu_exit,
2392 	.stop_cpu	= intel_pstate_stop_cpu,
2393 	.update_limits	= intel_pstate_update_limits,
2394 	.name		= "intel_pstate",
2395 };
2396 
2397 static int intel_cpufreq_verify_policy(struct cpufreq_policy_data *policy)
2398 {
2399 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2400 
2401 	intel_pstate_verify_cpu_policy(cpu, policy);
2402 	intel_pstate_update_perf_limits(cpu, policy->min, policy->max);
2403 
2404 	return 0;
2405 }
2406 
2407 /* Use of trace in passive mode:
2408  *
2409  * In passive mode the trace core_busy field (also known as the
2410  * performance field, and lablelled as such on the graphs; also known as
2411  * core_avg_perf) is not needed and so is re-assigned to indicate if the
2412  * driver call was via the normal or fast switch path. Various graphs
2413  * output from the intel_pstate_tracer.py utility that include core_busy
2414  * (or performance or core_avg_perf) have a fixed y-axis from 0 to 100%,
2415  * so we use 10 to indicate the the normal path through the driver, and
2416  * 90 to indicate the fast switch path through the driver.
2417  * The scaled_busy field is not used, and is set to 0.
2418  */
2419 
2420 #define	INTEL_PSTATE_TRACE_TARGET 10
2421 #define	INTEL_PSTATE_TRACE_FAST_SWITCH 90
2422 
2423 static void intel_cpufreq_trace(struct cpudata *cpu, unsigned int trace_type, int old_pstate)
2424 {
2425 	struct sample *sample;
2426 
2427 	if (!trace_pstate_sample_enabled())
2428 		return;
2429 
2430 	if (!intel_pstate_sample(cpu, ktime_get()))
2431 		return;
2432 
2433 	sample = &cpu->sample;
2434 	trace_pstate_sample(trace_type,
2435 		0,
2436 		old_pstate,
2437 		cpu->pstate.current_pstate,
2438 		sample->mperf,
2439 		sample->aperf,
2440 		sample->tsc,
2441 		get_avg_frequency(cpu),
2442 		fp_toint(cpu->iowait_boost * 100));
2443 }
2444 
2445 static void intel_cpufreq_adjust_hwp(struct cpudata *cpu, u32 target_pstate,
2446 				     bool fast_switch)
2447 {
2448 	u64 prev = READ_ONCE(cpu->hwp_req_cached), value = prev;
2449 
2450 	value &= ~HWP_MIN_PERF(~0L);
2451 	value |= HWP_MIN_PERF(target_pstate);
2452 
2453 	/*
2454 	 * The entire MSR needs to be updated in order to update the HWP min
2455 	 * field in it, so opportunistically update the max too if needed.
2456 	 */
2457 	value &= ~HWP_MAX_PERF(~0L);
2458 	value |= HWP_MAX_PERF(cpu->max_perf_ratio);
2459 
2460 	if (value == prev)
2461 		return;
2462 
2463 	WRITE_ONCE(cpu->hwp_req_cached, value);
2464 	if (fast_switch)
2465 		wrmsrl(MSR_HWP_REQUEST, value);
2466 	else
2467 		wrmsrl_on_cpu(cpu->cpu, MSR_HWP_REQUEST, value);
2468 }
2469 
2470 static void intel_cpufreq_adjust_perf_ctl(struct cpudata *cpu,
2471 					  u32 target_pstate, bool fast_switch)
2472 {
2473 	if (fast_switch)
2474 		wrmsrl(MSR_IA32_PERF_CTL,
2475 		       pstate_funcs.get_val(cpu, target_pstate));
2476 	else
2477 		wrmsrl_on_cpu(cpu->cpu, MSR_IA32_PERF_CTL,
2478 			      pstate_funcs.get_val(cpu, target_pstate));
2479 }
2480 
2481 static int intel_cpufreq_update_pstate(struct cpudata *cpu, int target_pstate,
2482 				       bool fast_switch)
2483 {
2484 	int old_pstate = cpu->pstate.current_pstate;
2485 
2486 	target_pstate = intel_pstate_prepare_request(cpu, target_pstate);
2487 	if (target_pstate != old_pstate) {
2488 		cpu->pstate.current_pstate = target_pstate;
2489 		if (hwp_active)
2490 			intel_cpufreq_adjust_hwp(cpu, target_pstate,
2491 						 fast_switch);
2492 		else
2493 			intel_cpufreq_adjust_perf_ctl(cpu, target_pstate,
2494 						      fast_switch);
2495 	}
2496 
2497 	intel_cpufreq_trace(cpu, fast_switch ? INTEL_PSTATE_TRACE_FAST_SWITCH :
2498 			    INTEL_PSTATE_TRACE_TARGET, old_pstate);
2499 
2500 	return target_pstate;
2501 }
2502 
2503 static int intel_cpufreq_target(struct cpufreq_policy *policy,
2504 				unsigned int target_freq,
2505 				unsigned int relation)
2506 {
2507 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2508 	struct cpufreq_freqs freqs;
2509 	int target_pstate;
2510 
2511 	update_turbo_state();
2512 
2513 	freqs.old = policy->cur;
2514 	freqs.new = target_freq;
2515 
2516 	cpufreq_freq_transition_begin(policy, &freqs);
2517 
2518 	switch (relation) {
2519 	case CPUFREQ_RELATION_L:
2520 		target_pstate = DIV_ROUND_UP(freqs.new, cpu->pstate.scaling);
2521 		break;
2522 	case CPUFREQ_RELATION_H:
2523 		target_pstate = freqs.new / cpu->pstate.scaling;
2524 		break;
2525 	default:
2526 		target_pstate = DIV_ROUND_CLOSEST(freqs.new, cpu->pstate.scaling);
2527 		break;
2528 	}
2529 
2530 	target_pstate = intel_cpufreq_update_pstate(cpu, target_pstate, false);
2531 
2532 	freqs.new = target_pstate * cpu->pstate.scaling;
2533 
2534 	cpufreq_freq_transition_end(policy, &freqs, false);
2535 
2536 	return 0;
2537 }
2538 
2539 static unsigned int intel_cpufreq_fast_switch(struct cpufreq_policy *policy,
2540 					      unsigned int target_freq)
2541 {
2542 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2543 	int target_pstate;
2544 
2545 	update_turbo_state();
2546 
2547 	target_pstate = DIV_ROUND_UP(target_freq, cpu->pstate.scaling);
2548 
2549 	target_pstate = intel_cpufreq_update_pstate(cpu, target_pstate, true);
2550 
2551 	return target_pstate * cpu->pstate.scaling;
2552 }
2553 
2554 static int intel_cpufreq_cpu_init(struct cpufreq_policy *policy)
2555 {
2556 	int max_state, turbo_max, min_freq, max_freq, ret;
2557 	struct freq_qos_request *req;
2558 	struct cpudata *cpu;
2559 	struct device *dev;
2560 
2561 	dev = get_cpu_device(policy->cpu);
2562 	if (!dev)
2563 		return -ENODEV;
2564 
2565 	ret = __intel_pstate_cpu_init(policy);
2566 	if (ret)
2567 		return ret;
2568 
2569 	policy->cpuinfo.transition_latency = INTEL_CPUFREQ_TRANSITION_LATENCY;
2570 	/* This reflects the intel_pstate_get_cpu_pstates() setting. */
2571 	policy->cur = policy->cpuinfo.min_freq;
2572 
2573 	req = kcalloc(2, sizeof(*req), GFP_KERNEL);
2574 	if (!req) {
2575 		ret = -ENOMEM;
2576 		goto pstate_exit;
2577 	}
2578 
2579 	cpu = all_cpu_data[policy->cpu];
2580 
2581 	if (hwp_active) {
2582 		u64 value;
2583 
2584 		intel_pstate_get_hwp_max(policy->cpu, &turbo_max, &max_state);
2585 		policy->transition_delay_us = INTEL_CPUFREQ_TRANSITION_DELAY_HWP;
2586 		rdmsrl_on_cpu(cpu->cpu, MSR_HWP_REQUEST, &value);
2587 		WRITE_ONCE(cpu->hwp_req_cached, value);
2588 		cpu->epp_cached = (value & GENMASK_ULL(31, 24)) >> 24;
2589 	} else {
2590 		turbo_max = cpu->pstate.turbo_pstate;
2591 		policy->transition_delay_us = INTEL_CPUFREQ_TRANSITION_DELAY;
2592 	}
2593 
2594 	min_freq = DIV_ROUND_UP(turbo_max * global.min_perf_pct, 100);
2595 	min_freq *= cpu->pstate.scaling;
2596 	max_freq = DIV_ROUND_UP(turbo_max * global.max_perf_pct, 100);
2597 	max_freq *= cpu->pstate.scaling;
2598 
2599 	ret = freq_qos_add_request(&policy->constraints, req, FREQ_QOS_MIN,
2600 				   min_freq);
2601 	if (ret < 0) {
2602 		dev_err(dev, "Failed to add min-freq constraint (%d)\n", ret);
2603 		goto free_req;
2604 	}
2605 
2606 	ret = freq_qos_add_request(&policy->constraints, req + 1, FREQ_QOS_MAX,
2607 				   max_freq);
2608 	if (ret < 0) {
2609 		dev_err(dev, "Failed to add max-freq constraint (%d)\n", ret);
2610 		goto remove_min_req;
2611 	}
2612 
2613 	policy->driver_data = req;
2614 
2615 	return 0;
2616 
2617 remove_min_req:
2618 	freq_qos_remove_request(req);
2619 free_req:
2620 	kfree(req);
2621 pstate_exit:
2622 	intel_pstate_exit_perf_limits(policy);
2623 
2624 	return ret;
2625 }
2626 
2627 static int intel_cpufreq_cpu_exit(struct cpufreq_policy *policy)
2628 {
2629 	struct freq_qos_request *req;
2630 
2631 	req = policy->driver_data;
2632 
2633 	freq_qos_remove_request(req + 1);
2634 	freq_qos_remove_request(req);
2635 	kfree(req);
2636 
2637 	return intel_pstate_cpu_exit(policy);
2638 }
2639 
2640 static struct cpufreq_driver intel_cpufreq = {
2641 	.flags		= CPUFREQ_CONST_LOOPS,
2642 	.verify		= intel_cpufreq_verify_policy,
2643 	.target		= intel_cpufreq_target,
2644 	.fast_switch	= intel_cpufreq_fast_switch,
2645 	.init		= intel_cpufreq_cpu_init,
2646 	.exit		= intel_cpufreq_cpu_exit,
2647 	.stop_cpu	= intel_cpufreq_stop_cpu,
2648 	.update_limits	= intel_pstate_update_limits,
2649 	.name		= "intel_cpufreq",
2650 };
2651 
2652 static struct cpufreq_driver *default_driver;
2653 
2654 static void intel_pstate_driver_cleanup(void)
2655 {
2656 	unsigned int cpu;
2657 
2658 	get_online_cpus();
2659 	for_each_online_cpu(cpu) {
2660 		if (all_cpu_data[cpu]) {
2661 			if (intel_pstate_driver == &intel_pstate)
2662 				intel_pstate_clear_update_util_hook(cpu);
2663 
2664 			kfree(all_cpu_data[cpu]);
2665 			all_cpu_data[cpu] = NULL;
2666 		}
2667 	}
2668 	put_online_cpus();
2669 
2670 	if (intel_pstate_driver == &intel_pstate)
2671 		intel_pstate_sysfs_hide_hwp_dynamic_boost();
2672 
2673 	intel_pstate_driver = NULL;
2674 }
2675 
2676 static int intel_pstate_register_driver(struct cpufreq_driver *driver)
2677 {
2678 	int ret;
2679 
2680 	if (driver == &intel_pstate)
2681 		intel_pstate_sysfs_expose_hwp_dynamic_boost();
2682 
2683 	memset(&global, 0, sizeof(global));
2684 	global.max_perf_pct = 100;
2685 
2686 	intel_pstate_driver = driver;
2687 	ret = cpufreq_register_driver(intel_pstate_driver);
2688 	if (ret) {
2689 		intel_pstate_driver_cleanup();
2690 		return ret;
2691 	}
2692 
2693 	global.min_perf_pct = min_perf_pct_min();
2694 
2695 	return 0;
2696 }
2697 
2698 static int intel_pstate_unregister_driver(void)
2699 {
2700 	cpufreq_unregister_driver(intel_pstate_driver);
2701 	intel_pstate_driver_cleanup();
2702 
2703 	return 0;
2704 }
2705 
2706 static ssize_t intel_pstate_show_status(char *buf)
2707 {
2708 	if (!intel_pstate_driver)
2709 		return sprintf(buf, "off\n");
2710 
2711 	return sprintf(buf, "%s\n", intel_pstate_driver == &intel_pstate ?
2712 					"active" : "passive");
2713 }
2714 
2715 static int intel_pstate_update_status(const char *buf, size_t size)
2716 {
2717 	int ret;
2718 
2719 	if (size == 3 && !strncmp(buf, "off", size))
2720 		return intel_pstate_driver ?
2721 			intel_pstate_unregister_driver() : -EINVAL;
2722 
2723 	if (size == 6 && !strncmp(buf, "active", size)) {
2724 		if (intel_pstate_driver) {
2725 			if (intel_pstate_driver == &intel_pstate)
2726 				return 0;
2727 
2728 			ret = intel_pstate_unregister_driver();
2729 			if (ret)
2730 				return ret;
2731 		}
2732 
2733 		return intel_pstate_register_driver(&intel_pstate);
2734 	}
2735 
2736 	if (size == 7 && !strncmp(buf, "passive", size)) {
2737 		if (intel_pstate_driver) {
2738 			if (intel_pstate_driver == &intel_cpufreq)
2739 				return 0;
2740 
2741 			ret = intel_pstate_unregister_driver();
2742 			if (ret)
2743 				return ret;
2744 		}
2745 
2746 		return intel_pstate_register_driver(&intel_cpufreq);
2747 	}
2748 
2749 	return -EINVAL;
2750 }
2751 
2752 static int no_load __initdata;
2753 static int no_hwp __initdata;
2754 static int hwp_only __initdata;
2755 static unsigned int force_load __initdata;
2756 
2757 static int __init intel_pstate_msrs_not_valid(void)
2758 {
2759 	if (!pstate_funcs.get_max() ||
2760 	    !pstate_funcs.get_min() ||
2761 	    !pstate_funcs.get_turbo())
2762 		return -ENODEV;
2763 
2764 	return 0;
2765 }
2766 
2767 static void __init copy_cpu_funcs(struct pstate_funcs *funcs)
2768 {
2769 	pstate_funcs.get_max   = funcs->get_max;
2770 	pstate_funcs.get_max_physical = funcs->get_max_physical;
2771 	pstate_funcs.get_min   = funcs->get_min;
2772 	pstate_funcs.get_turbo = funcs->get_turbo;
2773 	pstate_funcs.get_scaling = funcs->get_scaling;
2774 	pstate_funcs.get_val   = funcs->get_val;
2775 	pstate_funcs.get_vid   = funcs->get_vid;
2776 	pstate_funcs.get_aperf_mperf_shift = funcs->get_aperf_mperf_shift;
2777 }
2778 
2779 #ifdef CONFIG_ACPI
2780 
2781 static bool __init intel_pstate_no_acpi_pss(void)
2782 {
2783 	int i;
2784 
2785 	for_each_possible_cpu(i) {
2786 		acpi_status status;
2787 		union acpi_object *pss;
2788 		struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
2789 		struct acpi_processor *pr = per_cpu(processors, i);
2790 
2791 		if (!pr)
2792 			continue;
2793 
2794 		status = acpi_evaluate_object(pr->handle, "_PSS", NULL, &buffer);
2795 		if (ACPI_FAILURE(status))
2796 			continue;
2797 
2798 		pss = buffer.pointer;
2799 		if (pss && pss->type == ACPI_TYPE_PACKAGE) {
2800 			kfree(pss);
2801 			return false;
2802 		}
2803 
2804 		kfree(pss);
2805 	}
2806 
2807 	pr_debug("ACPI _PSS not found\n");
2808 	return true;
2809 }
2810 
2811 static bool __init intel_pstate_no_acpi_pcch(void)
2812 {
2813 	acpi_status status;
2814 	acpi_handle handle;
2815 
2816 	status = acpi_get_handle(NULL, "\\_SB", &handle);
2817 	if (ACPI_FAILURE(status))
2818 		goto not_found;
2819 
2820 	if (acpi_has_method(handle, "PCCH"))
2821 		return false;
2822 
2823 not_found:
2824 	pr_debug("ACPI PCCH not found\n");
2825 	return true;
2826 }
2827 
2828 static bool __init intel_pstate_has_acpi_ppc(void)
2829 {
2830 	int i;
2831 
2832 	for_each_possible_cpu(i) {
2833 		struct acpi_processor *pr = per_cpu(processors, i);
2834 
2835 		if (!pr)
2836 			continue;
2837 		if (acpi_has_method(pr->handle, "_PPC"))
2838 			return true;
2839 	}
2840 	pr_debug("ACPI _PPC not found\n");
2841 	return false;
2842 }
2843 
2844 enum {
2845 	PSS,
2846 	PPC,
2847 };
2848 
2849 /* Hardware vendor-specific info that has its own power management modes */
2850 static struct acpi_platform_list plat_info[] __initdata = {
2851 	{"HP    ", "ProLiant", 0, ACPI_SIG_FADT, all_versions, NULL, PSS},
2852 	{"ORACLE", "X4-2    ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
2853 	{"ORACLE", "X4-2L   ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
2854 	{"ORACLE", "X4-2B   ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
2855 	{"ORACLE", "X3-2    ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
2856 	{"ORACLE", "X3-2L   ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
2857 	{"ORACLE", "X3-2B   ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
2858 	{"ORACLE", "X4470M2 ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
2859 	{"ORACLE", "X4270M3 ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
2860 	{"ORACLE", "X4270M2 ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
2861 	{"ORACLE", "X4170M2 ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
2862 	{"ORACLE", "X4170 M3", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
2863 	{"ORACLE", "X4275 M3", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
2864 	{"ORACLE", "X6-2    ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
2865 	{"ORACLE", "Sudbury ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
2866 	{ } /* End */
2867 };
2868 
2869 #define BITMASK_OOB	(BIT(8) | BIT(18))
2870 
2871 static bool __init intel_pstate_platform_pwr_mgmt_exists(void)
2872 {
2873 	const struct x86_cpu_id *id;
2874 	u64 misc_pwr;
2875 	int idx;
2876 
2877 	id = x86_match_cpu(intel_pstate_cpu_oob_ids);
2878 	if (id) {
2879 		rdmsrl(MSR_MISC_PWR_MGMT, misc_pwr);
2880 		if (misc_pwr & BITMASK_OOB) {
2881 			pr_debug("Bit 8 or 18 in the MISC_PWR_MGMT MSR set\n");
2882 			pr_debug("P states are controlled in Out of Band mode by the firmware/hardware\n");
2883 			return true;
2884 		}
2885 	}
2886 
2887 	idx = acpi_match_platform_list(plat_info);
2888 	if (idx < 0)
2889 		return false;
2890 
2891 	switch (plat_info[idx].data) {
2892 	case PSS:
2893 		if (!intel_pstate_no_acpi_pss())
2894 			return false;
2895 
2896 		return intel_pstate_no_acpi_pcch();
2897 	case PPC:
2898 		return intel_pstate_has_acpi_ppc() && !force_load;
2899 	}
2900 
2901 	return false;
2902 }
2903 
2904 static void intel_pstate_request_control_from_smm(void)
2905 {
2906 	/*
2907 	 * It may be unsafe to request P-states control from SMM if _PPC support
2908 	 * has not been enabled.
2909 	 */
2910 	if (acpi_ppc)
2911 		acpi_processor_pstate_control();
2912 }
2913 #else /* CONFIG_ACPI not enabled */
2914 static inline bool intel_pstate_platform_pwr_mgmt_exists(void) { return false; }
2915 static inline bool intel_pstate_has_acpi_ppc(void) { return false; }
2916 static inline void intel_pstate_request_control_from_smm(void) {}
2917 #endif /* CONFIG_ACPI */
2918 
2919 #define INTEL_PSTATE_HWP_BROADWELL	0x01
2920 
2921 #define X86_MATCH_HWP(model, hwp_mode)					\
2922 	X86_MATCH_VENDOR_FAM_MODEL_FEATURE(INTEL, 6, INTEL_FAM6_##model, \
2923 					   X86_FEATURE_HWP, hwp_mode)
2924 
2925 static const struct x86_cpu_id hwp_support_ids[] __initconst = {
2926 	X86_MATCH_HWP(BROADWELL_X,	INTEL_PSTATE_HWP_BROADWELL),
2927 	X86_MATCH_HWP(BROADWELL_D,	INTEL_PSTATE_HWP_BROADWELL),
2928 	X86_MATCH_HWP(ANY,		0),
2929 	{}
2930 };
2931 
2932 static int __init intel_pstate_init(void)
2933 {
2934 	const struct x86_cpu_id *id;
2935 	int rc;
2936 
2937 	if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL)
2938 		return -ENODEV;
2939 
2940 	if (no_load)
2941 		return -ENODEV;
2942 
2943 	id = x86_match_cpu(hwp_support_ids);
2944 	if (id) {
2945 		copy_cpu_funcs(&core_funcs);
2946 		/*
2947 		 * Avoid enabling HWP for processors without EPP support,
2948 		 * because that means incomplete HWP implementation which is a
2949 		 * corner case and supporting it is generally problematic.
2950 		 */
2951 		if (!no_hwp && boot_cpu_has(X86_FEATURE_HWP_EPP)) {
2952 			hwp_active++;
2953 			hwp_mode_bdw = id->driver_data;
2954 			intel_pstate.attr = hwp_cpufreq_attrs;
2955 			intel_cpufreq.attr = hwp_cpufreq_attrs;
2956 			if (!default_driver)
2957 				default_driver = &intel_pstate;
2958 
2959 			goto hwp_cpu_matched;
2960 		}
2961 	} else {
2962 		id = x86_match_cpu(intel_pstate_cpu_ids);
2963 		if (!id) {
2964 			pr_info("CPU model not supported\n");
2965 			return -ENODEV;
2966 		}
2967 
2968 		copy_cpu_funcs((struct pstate_funcs *)id->driver_data);
2969 	}
2970 
2971 	if (intel_pstate_msrs_not_valid()) {
2972 		pr_info("Invalid MSRs\n");
2973 		return -ENODEV;
2974 	}
2975 	/* Without HWP start in the passive mode. */
2976 	if (!default_driver)
2977 		default_driver = &intel_cpufreq;
2978 
2979 hwp_cpu_matched:
2980 	/*
2981 	 * The Intel pstate driver will be ignored if the platform
2982 	 * firmware has its own power management modes.
2983 	 */
2984 	if (intel_pstate_platform_pwr_mgmt_exists()) {
2985 		pr_info("P-states controlled by the platform\n");
2986 		return -ENODEV;
2987 	}
2988 
2989 	if (!hwp_active && hwp_only)
2990 		return -ENOTSUPP;
2991 
2992 	pr_info("Intel P-state driver initializing\n");
2993 
2994 	all_cpu_data = vzalloc(array_size(sizeof(void *), num_possible_cpus()));
2995 	if (!all_cpu_data)
2996 		return -ENOMEM;
2997 
2998 	intel_pstate_request_control_from_smm();
2999 
3000 	intel_pstate_sysfs_expose_params();
3001 
3002 	mutex_lock(&intel_pstate_driver_lock);
3003 	rc = intel_pstate_register_driver(default_driver);
3004 	mutex_unlock(&intel_pstate_driver_lock);
3005 	if (rc)
3006 		return rc;
3007 
3008 	if (hwp_active) {
3009 		const struct x86_cpu_id *id;
3010 
3011 		id = x86_match_cpu(intel_pstate_cpu_ee_disable_ids);
3012 		if (id) {
3013 			set_power_ctl_ee_state(false);
3014 			pr_info("Disabling energy efficiency optimization\n");
3015 		}
3016 
3017 		pr_info("HWP enabled\n");
3018 	}
3019 
3020 	return 0;
3021 }
3022 device_initcall(intel_pstate_init);
3023 
3024 static int __init intel_pstate_setup(char *str)
3025 {
3026 	if (!str)
3027 		return -EINVAL;
3028 
3029 	if (!strcmp(str, "disable"))
3030 		no_load = 1;
3031 	else if (!strcmp(str, "active"))
3032 		default_driver = &intel_pstate;
3033 	else if (!strcmp(str, "passive"))
3034 		default_driver = &intel_cpufreq;
3035 
3036 	if (!strcmp(str, "no_hwp")) {
3037 		pr_info("HWP disabled\n");
3038 		no_hwp = 1;
3039 	}
3040 	if (!strcmp(str, "force"))
3041 		force_load = 1;
3042 	if (!strcmp(str, "hwp_only"))
3043 		hwp_only = 1;
3044 	if (!strcmp(str, "per_cpu_perf_limits"))
3045 		per_cpu_limits = true;
3046 
3047 #ifdef CONFIG_ACPI
3048 	if (!strcmp(str, "support_acpi_ppc"))
3049 		acpi_ppc = true;
3050 #endif
3051 
3052 	return 0;
3053 }
3054 early_param("intel_pstate", intel_pstate_setup);
3055 
3056 MODULE_AUTHOR("Dirk Brandewie <dirk.j.brandewie@intel.com>");
3057 MODULE_DESCRIPTION("'intel_pstate' - P state driver Intel Core processors");
3058 MODULE_LICENSE("GPL");
3059