xref: /linux/drivers/cpufreq/intel_pstate.c (revision 78c3925c048c752334873f56c3a3d1c9d53e0416)
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 <linux/bitfield.h>
29 #include <trace/events/power.h>
30 
31 #include <asm/cpu.h>
32 #include <asm/div64.h>
33 #include <asm/msr.h>
34 #include <asm/cpu_device_id.h>
35 #include <asm/cpufeature.h>
36 #include <asm/intel-family.h>
37 #include "../drivers/thermal/intel/thermal_interrupt.h"
38 
39 #define INTEL_PSTATE_SAMPLING_INTERVAL	(10 * NSEC_PER_MSEC)
40 
41 #define INTEL_CPUFREQ_TRANSITION_LATENCY	20000
42 #define INTEL_CPUFREQ_TRANSITION_DELAY_HWP	5000
43 #define INTEL_CPUFREQ_TRANSITION_DELAY		500
44 
45 #ifdef CONFIG_ACPI
46 #include <acpi/processor.h>
47 #include <acpi/cppc_acpi.h>
48 #endif
49 
50 #define FRAC_BITS 8
51 #define int_tofp(X) ((int64_t)(X) << FRAC_BITS)
52 #define fp_toint(X) ((X) >> FRAC_BITS)
53 
54 #define ONE_EIGHTH_FP ((int64_t)1 << (FRAC_BITS - 3))
55 
56 #define EXT_BITS 6
57 #define EXT_FRAC_BITS (EXT_BITS + FRAC_BITS)
58 #define fp_ext_toint(X) ((X) >> EXT_FRAC_BITS)
59 #define int_ext_tofp(X) ((int64_t)(X) << EXT_FRAC_BITS)
60 
61 static inline int32_t mul_fp(int32_t x, int32_t y)
62 {
63 	return ((int64_t)x * (int64_t)y) >> FRAC_BITS;
64 }
65 
66 static inline int32_t div_fp(s64 x, s64 y)
67 {
68 	return div64_s64((int64_t)x << FRAC_BITS, y);
69 }
70 
71 static inline int ceiling_fp(int32_t x)
72 {
73 	int mask, ret;
74 
75 	ret = fp_toint(x);
76 	mask = (1 << FRAC_BITS) - 1;
77 	if (x & mask)
78 		ret += 1;
79 	return ret;
80 }
81 
82 static inline u64 mul_ext_fp(u64 x, u64 y)
83 {
84 	return (x * y) >> EXT_FRAC_BITS;
85 }
86 
87 static inline u64 div_ext_fp(u64 x, u64 y)
88 {
89 	return div64_u64(x << EXT_FRAC_BITS, y);
90 }
91 
92 /**
93  * struct sample -	Store performance sample
94  * @core_avg_perf:	Ratio of APERF/MPERF which is the actual average
95  *			performance during last sample period
96  * @busy_scaled:	Scaled busy value which is used to calculate next
97  *			P state. This can be different than core_avg_perf
98  *			to account for cpu idle period
99  * @aperf:		Difference of actual performance frequency clock count
100  *			read from APERF MSR between last and current sample
101  * @mperf:		Difference of maximum performance frequency clock count
102  *			read from MPERF MSR between last and current sample
103  * @tsc:		Difference of time stamp counter between last and
104  *			current sample
105  * @time:		Current time from scheduler
106  *
107  * This structure is used in the cpudata structure to store performance sample
108  * data for choosing next P State.
109  */
110 struct sample {
111 	int32_t core_avg_perf;
112 	int32_t busy_scaled;
113 	u64 aperf;
114 	u64 mperf;
115 	u64 tsc;
116 	u64 time;
117 };
118 
119 /**
120  * struct pstate_data - Store P state data
121  * @current_pstate:	Current requested P state
122  * @min_pstate:		Min P state possible for this platform
123  * @max_pstate:		Max P state possible for this platform
124  * @max_pstate_physical:This is physical Max P state for a processor
125  *			This can be higher than the max_pstate which can
126  *			be limited by platform thermal design power limits
127  * @perf_ctl_scaling:	PERF_CTL P-state to frequency scaling factor
128  * @scaling:		Scaling factor between performance and frequency
129  * @turbo_pstate:	Max Turbo P state possible for this platform
130  * @min_freq:		@min_pstate frequency in cpufreq units
131  * @max_freq:		@max_pstate frequency in cpufreq units
132  * @turbo_freq:		@turbo_pstate frequency in cpufreq units
133  *
134  * Stores the per cpu model P state limits and current P state.
135  */
136 struct pstate_data {
137 	int	current_pstate;
138 	int	min_pstate;
139 	int	max_pstate;
140 	int	max_pstate_physical;
141 	int	perf_ctl_scaling;
142 	int	scaling;
143 	int	turbo_pstate;
144 	unsigned int min_freq;
145 	unsigned int max_freq;
146 	unsigned int turbo_freq;
147 };
148 
149 /**
150  * struct vid_data -	Stores voltage information data
151  * @min:		VID data for this platform corresponding to
152  *			the lowest P state
153  * @max:		VID data corresponding to the highest P State.
154  * @turbo:		VID data for turbo P state
155  * @ratio:		Ratio of (vid max - vid min) /
156  *			(max P state - Min P State)
157  *
158  * Stores the voltage data for DVFS (Dynamic Voltage and Frequency Scaling)
159  * This data is used in Atom platforms, where in addition to target P state,
160  * the voltage data needs to be specified to select next P State.
161  */
162 struct vid_data {
163 	int min;
164 	int max;
165 	int turbo;
166 	int32_t ratio;
167 };
168 
169 /**
170  * struct global_params - Global parameters, mostly tunable via sysfs.
171  * @no_turbo:		Whether or not to use turbo P-states.
172  * @turbo_disabled:	Whether or not turbo P-states are available at all,
173  *			based on the MSR_IA32_MISC_ENABLE value and whether or
174  *			not the maximum reported turbo P-state is different from
175  *			the maximum reported non-turbo one.
176  * @turbo_disabled_mf:	The @turbo_disabled value reflected by cpuinfo.max_freq.
177  * @min_perf_pct:	Minimum capacity limit in percent of the maximum turbo
178  *			P-state capacity.
179  * @max_perf_pct:	Maximum capacity limit in percent of the maximum turbo
180  *			P-state capacity.
181  */
182 struct global_params {
183 	bool no_turbo;
184 	bool turbo_disabled;
185 	bool turbo_disabled_mf;
186 	int max_perf_pct;
187 	int min_perf_pct;
188 };
189 
190 /**
191  * struct cpudata -	Per CPU instance data storage
192  * @cpu:		CPU number for this instance data
193  * @policy:		CPUFreq policy value
194  * @update_util:	CPUFreq utility callback information
195  * @update_util_set:	CPUFreq utility callback is set
196  * @iowait_boost:	iowait-related boost fraction
197  * @last_update:	Time of the last update.
198  * @pstate:		Stores P state limits for this CPU
199  * @vid:		Stores VID limits for this CPU
200  * @last_sample_time:	Last Sample time
201  * @aperf_mperf_shift:	APERF vs MPERF counting frequency difference
202  * @prev_aperf:		Last APERF value read from APERF MSR
203  * @prev_mperf:		Last MPERF value read from MPERF MSR
204  * @prev_tsc:		Last timestamp counter (TSC) value
205  * @sample:		Storage for storing last Sample data
206  * @min_perf_ratio:	Minimum capacity in terms of PERF or HWP ratios
207  * @max_perf_ratio:	Maximum capacity in terms of PERF or HWP ratios
208  * @acpi_perf_data:	Stores ACPI perf information read from _PSS
209  * @valid_pss_table:	Set to true for valid ACPI _PSS entries found
210  * @epp_powersave:	Last saved HWP energy performance preference
211  *			(EPP) or energy performance bias (EPB),
212  *			when policy switched to performance
213  * @epp_policy:		Last saved policy used to set EPP/EPB
214  * @epp_default:	Power on default HWP energy performance
215  *			preference/bias
216  * @epp_cached		Cached HWP energy-performance preference value
217  * @hwp_req_cached:	Cached value of the last HWP Request MSR
218  * @hwp_cap_cached:	Cached value of the last HWP Capabilities MSR
219  * @last_io_update:	Last time when IO wake flag was set
220  * @sched_flags:	Store scheduler flags for possible cross CPU update
221  * @hwp_boost_min:	Last HWP boosted min performance
222  * @suspended:		Whether or not the driver has been suspended.
223  * @hwp_notify_work:	workqueue for HWP notifications.
224  *
225  * This structure stores per CPU instance data for all CPUs.
226  */
227 struct cpudata {
228 	int cpu;
229 
230 	unsigned int policy;
231 	struct update_util_data update_util;
232 	bool   update_util_set;
233 
234 	struct pstate_data pstate;
235 	struct vid_data vid;
236 
237 	u64	last_update;
238 	u64	last_sample_time;
239 	u64	aperf_mperf_shift;
240 	u64	prev_aperf;
241 	u64	prev_mperf;
242 	u64	prev_tsc;
243 	struct sample sample;
244 	int32_t	min_perf_ratio;
245 	int32_t	max_perf_ratio;
246 #ifdef CONFIG_ACPI
247 	struct acpi_processor_performance acpi_perf_data;
248 	bool valid_pss_table;
249 #endif
250 	unsigned int iowait_boost;
251 	s16 epp_powersave;
252 	s16 epp_policy;
253 	s16 epp_default;
254 	s16 epp_cached;
255 	u64 hwp_req_cached;
256 	u64 hwp_cap_cached;
257 	u64 last_io_update;
258 	unsigned int sched_flags;
259 	u32 hwp_boost_min;
260 	bool suspended;
261 	struct delayed_work hwp_notify_work;
262 };
263 
264 static struct cpudata **all_cpu_data;
265 
266 /**
267  * struct pstate_funcs - Per CPU model specific callbacks
268  * @get_max:		Callback to get maximum non turbo effective P state
269  * @get_max_physical:	Callback to get maximum non turbo physical P state
270  * @get_min:		Callback to get minimum P state
271  * @get_turbo:		Callback to get turbo P state
272  * @get_scaling:	Callback to get frequency scaling factor
273  * @get_cpu_scaling:	Get frequency scaling factor for a given cpu
274  * @get_aperf_mperf_shift: Callback to get the APERF vs MPERF frequency difference
275  * @get_val:		Callback to convert P state to actual MSR write value
276  * @get_vid:		Callback to get VID data for Atom platforms
277  *
278  * Core and Atom CPU models have different way to get P State limits. This
279  * structure is used to store those callbacks.
280  */
281 struct pstate_funcs {
282 	int (*get_max)(int cpu);
283 	int (*get_max_physical)(int cpu);
284 	int (*get_min)(int cpu);
285 	int (*get_turbo)(int cpu);
286 	int (*get_scaling)(void);
287 	int (*get_cpu_scaling)(int cpu);
288 	int (*get_aperf_mperf_shift)(void);
289 	u64 (*get_val)(struct cpudata*, int pstate);
290 	void (*get_vid)(struct cpudata *);
291 };
292 
293 static struct pstate_funcs pstate_funcs __read_mostly;
294 
295 static int hwp_active __read_mostly;
296 static int hwp_mode_bdw __read_mostly;
297 static bool per_cpu_limits __read_mostly;
298 static bool hwp_boost __read_mostly;
299 static bool hwp_forced __read_mostly;
300 
301 static struct cpufreq_driver *intel_pstate_driver __read_mostly;
302 
303 #define HYBRID_SCALING_FACTOR		78741
304 #define HYBRID_SCALING_FACTOR_MTL	80000
305 
306 static int hybrid_scaling_factor = HYBRID_SCALING_FACTOR;
307 
308 static inline int core_get_scaling(void)
309 {
310 	return 100000;
311 }
312 
313 #ifdef CONFIG_ACPI
314 static bool acpi_ppc;
315 #endif
316 
317 static struct global_params global;
318 
319 static DEFINE_MUTEX(intel_pstate_driver_lock);
320 static DEFINE_MUTEX(intel_pstate_limits_lock);
321 
322 #ifdef CONFIG_ACPI
323 
324 static bool intel_pstate_acpi_pm_profile_server(void)
325 {
326 	if (acpi_gbl_FADT.preferred_profile == PM_ENTERPRISE_SERVER ||
327 	    acpi_gbl_FADT.preferred_profile == PM_PERFORMANCE_SERVER)
328 		return true;
329 
330 	return false;
331 }
332 
333 static bool intel_pstate_get_ppc_enable_status(void)
334 {
335 	if (intel_pstate_acpi_pm_profile_server())
336 		return true;
337 
338 	return acpi_ppc;
339 }
340 
341 #ifdef CONFIG_ACPI_CPPC_LIB
342 
343 /* The work item is needed to avoid CPU hotplug locking issues */
344 static void intel_pstste_sched_itmt_work_fn(struct work_struct *work)
345 {
346 	sched_set_itmt_support();
347 }
348 
349 static DECLARE_WORK(sched_itmt_work, intel_pstste_sched_itmt_work_fn);
350 
351 #define CPPC_MAX_PERF	U8_MAX
352 
353 static void intel_pstate_set_itmt_prio(int cpu)
354 {
355 	struct cppc_perf_caps cppc_perf;
356 	static u32 max_highest_perf = 0, min_highest_perf = U32_MAX;
357 	int ret;
358 
359 	ret = cppc_get_perf_caps(cpu, &cppc_perf);
360 	if (ret)
361 		return;
362 
363 	/*
364 	 * On some systems with overclocking enabled, CPPC.highest_perf is hardcoded to 0xff.
365 	 * In this case we can't use CPPC.highest_perf to enable ITMT.
366 	 * In this case we can look at MSR_HWP_CAPABILITIES bits [8:0] to decide.
367 	 */
368 	if (cppc_perf.highest_perf == CPPC_MAX_PERF)
369 		cppc_perf.highest_perf = HWP_HIGHEST_PERF(READ_ONCE(all_cpu_data[cpu]->hwp_cap_cached));
370 
371 	/*
372 	 * The priorities can be set regardless of whether or not
373 	 * sched_set_itmt_support(true) has been called and it is valid to
374 	 * update them at any time after it has been called.
375 	 */
376 	sched_set_itmt_core_prio(cppc_perf.highest_perf, cpu);
377 
378 	if (max_highest_perf <= min_highest_perf) {
379 		if (cppc_perf.highest_perf > max_highest_perf)
380 			max_highest_perf = cppc_perf.highest_perf;
381 
382 		if (cppc_perf.highest_perf < min_highest_perf)
383 			min_highest_perf = cppc_perf.highest_perf;
384 
385 		if (max_highest_perf > min_highest_perf) {
386 			/*
387 			 * This code can be run during CPU online under the
388 			 * CPU hotplug locks, so sched_set_itmt_support()
389 			 * cannot be called from here.  Queue up a work item
390 			 * to invoke it.
391 			 */
392 			schedule_work(&sched_itmt_work);
393 		}
394 	}
395 }
396 
397 static int intel_pstate_get_cppc_guaranteed(int cpu)
398 {
399 	struct cppc_perf_caps cppc_perf;
400 	int ret;
401 
402 	ret = cppc_get_perf_caps(cpu, &cppc_perf);
403 	if (ret)
404 		return ret;
405 
406 	if (cppc_perf.guaranteed_perf)
407 		return cppc_perf.guaranteed_perf;
408 
409 	return cppc_perf.nominal_perf;
410 }
411 
412 static int intel_pstate_cppc_get_scaling(int cpu)
413 {
414 	struct cppc_perf_caps cppc_perf;
415 	int ret;
416 
417 	ret = cppc_get_perf_caps(cpu, &cppc_perf);
418 
419 	/*
420 	 * If the nominal frequency and the nominal performance are not
421 	 * zero and the ratio between them is not 100, return the hybrid
422 	 * scaling factor.
423 	 */
424 	if (!ret && cppc_perf.nominal_perf && cppc_perf.nominal_freq &&
425 	    cppc_perf.nominal_perf * 100 != cppc_perf.nominal_freq)
426 		return hybrid_scaling_factor;
427 
428 	return core_get_scaling();
429 }
430 
431 #else /* CONFIG_ACPI_CPPC_LIB */
432 static inline void intel_pstate_set_itmt_prio(int cpu)
433 {
434 }
435 #endif /* CONFIG_ACPI_CPPC_LIB */
436 
437 static void intel_pstate_init_acpi_perf_limits(struct cpufreq_policy *policy)
438 {
439 	struct cpudata *cpu;
440 	int ret;
441 	int i;
442 
443 	if (hwp_active) {
444 		intel_pstate_set_itmt_prio(policy->cpu);
445 		return;
446 	}
447 
448 	if (!intel_pstate_get_ppc_enable_status())
449 		return;
450 
451 	cpu = all_cpu_data[policy->cpu];
452 
453 	ret = acpi_processor_register_performance(&cpu->acpi_perf_data,
454 						  policy->cpu);
455 	if (ret)
456 		return;
457 
458 	/*
459 	 * Check if the control value in _PSS is for PERF_CTL MSR, which should
460 	 * guarantee that the states returned by it map to the states in our
461 	 * list directly.
462 	 */
463 	if (cpu->acpi_perf_data.control_register.space_id !=
464 						ACPI_ADR_SPACE_FIXED_HARDWARE)
465 		goto err;
466 
467 	/*
468 	 * If there is only one entry _PSS, simply ignore _PSS and continue as
469 	 * usual without taking _PSS into account
470 	 */
471 	if (cpu->acpi_perf_data.state_count < 2)
472 		goto err;
473 
474 	pr_debug("CPU%u - ACPI _PSS perf data\n", policy->cpu);
475 	for (i = 0; i < cpu->acpi_perf_data.state_count; i++) {
476 		pr_debug("     %cP%d: %u MHz, %u mW, 0x%x\n",
477 			 (i == cpu->acpi_perf_data.state ? '*' : ' '), i,
478 			 (u32) cpu->acpi_perf_data.states[i].core_frequency,
479 			 (u32) cpu->acpi_perf_data.states[i].power,
480 			 (u32) cpu->acpi_perf_data.states[i].control);
481 	}
482 
483 	cpu->valid_pss_table = true;
484 	pr_debug("_PPC limits will be enforced\n");
485 
486 	return;
487 
488  err:
489 	cpu->valid_pss_table = false;
490 	acpi_processor_unregister_performance(policy->cpu);
491 }
492 
493 static void intel_pstate_exit_perf_limits(struct cpufreq_policy *policy)
494 {
495 	struct cpudata *cpu;
496 
497 	cpu = all_cpu_data[policy->cpu];
498 	if (!cpu->valid_pss_table)
499 		return;
500 
501 	acpi_processor_unregister_performance(policy->cpu);
502 }
503 #else /* CONFIG_ACPI */
504 static inline void intel_pstate_init_acpi_perf_limits(struct cpufreq_policy *policy)
505 {
506 }
507 
508 static inline void intel_pstate_exit_perf_limits(struct cpufreq_policy *policy)
509 {
510 }
511 
512 static inline bool intel_pstate_acpi_pm_profile_server(void)
513 {
514 	return false;
515 }
516 #endif /* CONFIG_ACPI */
517 
518 #ifndef CONFIG_ACPI_CPPC_LIB
519 static inline int intel_pstate_get_cppc_guaranteed(int cpu)
520 {
521 	return -ENOTSUPP;
522 }
523 
524 static int intel_pstate_cppc_get_scaling(int cpu)
525 {
526 	return core_get_scaling();
527 }
528 #endif /* CONFIG_ACPI_CPPC_LIB */
529 
530 static int intel_pstate_freq_to_hwp_rel(struct cpudata *cpu, int freq,
531 					unsigned int relation)
532 {
533 	if (freq == cpu->pstate.turbo_freq)
534 		return cpu->pstate.turbo_pstate;
535 
536 	if (freq == cpu->pstate.max_freq)
537 		return cpu->pstate.max_pstate;
538 
539 	switch (relation) {
540 	case CPUFREQ_RELATION_H:
541 		return freq / cpu->pstate.scaling;
542 	case CPUFREQ_RELATION_C:
543 		return DIV_ROUND_CLOSEST(freq, cpu->pstate.scaling);
544 	}
545 
546 	return DIV_ROUND_UP(freq, cpu->pstate.scaling);
547 }
548 
549 static int intel_pstate_freq_to_hwp(struct cpudata *cpu, int freq)
550 {
551 	return intel_pstate_freq_to_hwp_rel(cpu, freq, CPUFREQ_RELATION_L);
552 }
553 
554 /**
555  * intel_pstate_hybrid_hwp_adjust - Calibrate HWP performance levels.
556  * @cpu: Target CPU.
557  *
558  * On hybrid processors, HWP may expose more performance levels than there are
559  * P-states accessible through the PERF_CTL interface.  If that happens, the
560  * scaling factor between HWP performance levels and CPU frequency will be less
561  * than the scaling factor between P-state values and CPU frequency.
562  *
563  * In that case, adjust the CPU parameters used in computations accordingly.
564  */
565 static void intel_pstate_hybrid_hwp_adjust(struct cpudata *cpu)
566 {
567 	int perf_ctl_max_phys = cpu->pstate.max_pstate_physical;
568 	int perf_ctl_scaling = cpu->pstate.perf_ctl_scaling;
569 	int perf_ctl_turbo = pstate_funcs.get_turbo(cpu->cpu);
570 	int scaling = cpu->pstate.scaling;
571 	int freq;
572 
573 	pr_debug("CPU%d: perf_ctl_max_phys = %d\n", cpu->cpu, perf_ctl_max_phys);
574 	pr_debug("CPU%d: perf_ctl_turbo = %d\n", cpu->cpu, perf_ctl_turbo);
575 	pr_debug("CPU%d: perf_ctl_scaling = %d\n", cpu->cpu, perf_ctl_scaling);
576 	pr_debug("CPU%d: HWP_CAP guaranteed = %d\n", cpu->cpu, cpu->pstate.max_pstate);
577 	pr_debug("CPU%d: HWP_CAP highest = %d\n", cpu->cpu, cpu->pstate.turbo_pstate);
578 	pr_debug("CPU%d: HWP-to-frequency scaling factor: %d\n", cpu->cpu, scaling);
579 
580 	cpu->pstate.turbo_freq = rounddown(cpu->pstate.turbo_pstate * scaling,
581 					   perf_ctl_scaling);
582 	cpu->pstate.max_freq = rounddown(cpu->pstate.max_pstate * scaling,
583 					 perf_ctl_scaling);
584 
585 	freq = perf_ctl_max_phys * perf_ctl_scaling;
586 	cpu->pstate.max_pstate_physical = intel_pstate_freq_to_hwp(cpu, freq);
587 
588 	freq = cpu->pstate.min_pstate * perf_ctl_scaling;
589 	cpu->pstate.min_freq = freq;
590 	/*
591 	 * Cast the min P-state value retrieved via pstate_funcs.get_min() to
592 	 * the effective range of HWP performance levels.
593 	 */
594 	cpu->pstate.min_pstate = intel_pstate_freq_to_hwp(cpu, freq);
595 }
596 
597 static inline void update_turbo_state(void)
598 {
599 	u64 misc_en;
600 
601 	rdmsrl(MSR_IA32_MISC_ENABLE, misc_en);
602 	global.turbo_disabled = misc_en & MSR_IA32_MISC_ENABLE_TURBO_DISABLE;
603 }
604 
605 static int min_perf_pct_min(void)
606 {
607 	struct cpudata *cpu = all_cpu_data[0];
608 	int turbo_pstate = cpu->pstate.turbo_pstate;
609 
610 	return turbo_pstate ?
611 		(cpu->pstate.min_pstate * 100 / turbo_pstate) : 0;
612 }
613 
614 static s16 intel_pstate_get_epb(struct cpudata *cpu_data)
615 {
616 	u64 epb;
617 	int ret;
618 
619 	if (!boot_cpu_has(X86_FEATURE_EPB))
620 		return -ENXIO;
621 
622 	ret = rdmsrl_on_cpu(cpu_data->cpu, MSR_IA32_ENERGY_PERF_BIAS, &epb);
623 	if (ret)
624 		return (s16)ret;
625 
626 	return (s16)(epb & 0x0f);
627 }
628 
629 static s16 intel_pstate_get_epp(struct cpudata *cpu_data, u64 hwp_req_data)
630 {
631 	s16 epp;
632 
633 	if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
634 		/*
635 		 * When hwp_req_data is 0, means that caller didn't read
636 		 * MSR_HWP_REQUEST, so need to read and get EPP.
637 		 */
638 		if (!hwp_req_data) {
639 			epp = rdmsrl_on_cpu(cpu_data->cpu, MSR_HWP_REQUEST,
640 					    &hwp_req_data);
641 			if (epp)
642 				return epp;
643 		}
644 		epp = (hwp_req_data >> 24) & 0xff;
645 	} else {
646 		/* When there is no EPP present, HWP uses EPB settings */
647 		epp = intel_pstate_get_epb(cpu_data);
648 	}
649 
650 	return epp;
651 }
652 
653 static int intel_pstate_set_epb(int cpu, s16 pref)
654 {
655 	u64 epb;
656 	int ret;
657 
658 	if (!boot_cpu_has(X86_FEATURE_EPB))
659 		return -ENXIO;
660 
661 	ret = rdmsrl_on_cpu(cpu, MSR_IA32_ENERGY_PERF_BIAS, &epb);
662 	if (ret)
663 		return ret;
664 
665 	epb = (epb & ~0x0f) | pref;
666 	wrmsrl_on_cpu(cpu, MSR_IA32_ENERGY_PERF_BIAS, epb);
667 
668 	return 0;
669 }
670 
671 /*
672  * EPP/EPB display strings corresponding to EPP index in the
673  * energy_perf_strings[]
674  *	index		String
675  *-------------------------------------
676  *	0		default
677  *	1		performance
678  *	2		balance_performance
679  *	3		balance_power
680  *	4		power
681  */
682 
683 enum energy_perf_value_index {
684 	EPP_INDEX_DEFAULT = 0,
685 	EPP_INDEX_PERFORMANCE,
686 	EPP_INDEX_BALANCE_PERFORMANCE,
687 	EPP_INDEX_BALANCE_POWERSAVE,
688 	EPP_INDEX_POWERSAVE,
689 };
690 
691 static const char * const energy_perf_strings[] = {
692 	[EPP_INDEX_DEFAULT] = "default",
693 	[EPP_INDEX_PERFORMANCE] = "performance",
694 	[EPP_INDEX_BALANCE_PERFORMANCE] = "balance_performance",
695 	[EPP_INDEX_BALANCE_POWERSAVE] = "balance_power",
696 	[EPP_INDEX_POWERSAVE] = "power",
697 	NULL
698 };
699 static unsigned int epp_values[] = {
700 	[EPP_INDEX_DEFAULT] = 0, /* Unused index */
701 	[EPP_INDEX_PERFORMANCE] = HWP_EPP_PERFORMANCE,
702 	[EPP_INDEX_BALANCE_PERFORMANCE] = HWP_EPP_BALANCE_PERFORMANCE,
703 	[EPP_INDEX_BALANCE_POWERSAVE] = HWP_EPP_BALANCE_POWERSAVE,
704 	[EPP_INDEX_POWERSAVE] = HWP_EPP_POWERSAVE,
705 };
706 
707 static int intel_pstate_get_energy_pref_index(struct cpudata *cpu_data, int *raw_epp)
708 {
709 	s16 epp;
710 	int index = -EINVAL;
711 
712 	*raw_epp = 0;
713 	epp = intel_pstate_get_epp(cpu_data, 0);
714 	if (epp < 0)
715 		return epp;
716 
717 	if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
718 		if (epp == epp_values[EPP_INDEX_PERFORMANCE])
719 			return EPP_INDEX_PERFORMANCE;
720 		if (epp == epp_values[EPP_INDEX_BALANCE_PERFORMANCE])
721 			return EPP_INDEX_BALANCE_PERFORMANCE;
722 		if (epp == epp_values[EPP_INDEX_BALANCE_POWERSAVE])
723 			return EPP_INDEX_BALANCE_POWERSAVE;
724 		if (epp == epp_values[EPP_INDEX_POWERSAVE])
725 			return EPP_INDEX_POWERSAVE;
726 		*raw_epp = epp;
727 		return 0;
728 	} else if (boot_cpu_has(X86_FEATURE_EPB)) {
729 		/*
730 		 * Range:
731 		 *	0x00-0x03	:	Performance
732 		 *	0x04-0x07	:	Balance performance
733 		 *	0x08-0x0B	:	Balance power
734 		 *	0x0C-0x0F	:	Power
735 		 * The EPB is a 4 bit value, but our ranges restrict the
736 		 * value which can be set. Here only using top two bits
737 		 * effectively.
738 		 */
739 		index = (epp >> 2) + 1;
740 	}
741 
742 	return index;
743 }
744 
745 static int intel_pstate_set_epp(struct cpudata *cpu, u32 epp)
746 {
747 	int ret;
748 
749 	/*
750 	 * Use the cached HWP Request MSR value, because in the active mode the
751 	 * register itself may be updated by intel_pstate_hwp_boost_up() or
752 	 * intel_pstate_hwp_boost_down() at any time.
753 	 */
754 	u64 value = READ_ONCE(cpu->hwp_req_cached);
755 
756 	value &= ~GENMASK_ULL(31, 24);
757 	value |= (u64)epp << 24;
758 	/*
759 	 * The only other updater of hwp_req_cached in the active mode,
760 	 * intel_pstate_hwp_set(), is called under the same lock as this
761 	 * function, so it cannot run in parallel with the update below.
762 	 */
763 	WRITE_ONCE(cpu->hwp_req_cached, value);
764 	ret = wrmsrl_on_cpu(cpu->cpu, MSR_HWP_REQUEST, value);
765 	if (!ret)
766 		cpu->epp_cached = epp;
767 
768 	return ret;
769 }
770 
771 static int intel_pstate_set_energy_pref_index(struct cpudata *cpu_data,
772 					      int pref_index, bool use_raw,
773 					      u32 raw_epp)
774 {
775 	int epp = -EINVAL;
776 	int ret;
777 
778 	if (!pref_index)
779 		epp = cpu_data->epp_default;
780 
781 	if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
782 		if (use_raw)
783 			epp = raw_epp;
784 		else if (epp == -EINVAL)
785 			epp = epp_values[pref_index];
786 
787 		/*
788 		 * To avoid confusion, refuse to set EPP to any values different
789 		 * from 0 (performance) if the current policy is "performance",
790 		 * because those values would be overridden.
791 		 */
792 		if (epp > 0 && cpu_data->policy == CPUFREQ_POLICY_PERFORMANCE)
793 			return -EBUSY;
794 
795 		ret = intel_pstate_set_epp(cpu_data, epp);
796 	} else {
797 		if (epp == -EINVAL)
798 			epp = (pref_index - 1) << 2;
799 		ret = intel_pstate_set_epb(cpu_data->cpu, epp);
800 	}
801 
802 	return ret;
803 }
804 
805 static ssize_t show_energy_performance_available_preferences(
806 				struct cpufreq_policy *policy, char *buf)
807 {
808 	int i = 0;
809 	int ret = 0;
810 
811 	while (energy_perf_strings[i] != NULL)
812 		ret += sprintf(&buf[ret], "%s ", energy_perf_strings[i++]);
813 
814 	ret += sprintf(&buf[ret], "\n");
815 
816 	return ret;
817 }
818 
819 cpufreq_freq_attr_ro(energy_performance_available_preferences);
820 
821 static struct cpufreq_driver intel_pstate;
822 
823 static ssize_t store_energy_performance_preference(
824 		struct cpufreq_policy *policy, const char *buf, size_t count)
825 {
826 	struct cpudata *cpu = all_cpu_data[policy->cpu];
827 	char str_preference[21];
828 	bool raw = false;
829 	ssize_t ret;
830 	u32 epp = 0;
831 
832 	ret = sscanf(buf, "%20s", str_preference);
833 	if (ret != 1)
834 		return -EINVAL;
835 
836 	ret = match_string(energy_perf_strings, -1, str_preference);
837 	if (ret < 0) {
838 		if (!boot_cpu_has(X86_FEATURE_HWP_EPP))
839 			return ret;
840 
841 		ret = kstrtouint(buf, 10, &epp);
842 		if (ret)
843 			return ret;
844 
845 		if (epp > 255)
846 			return -EINVAL;
847 
848 		raw = true;
849 	}
850 
851 	/*
852 	 * This function runs with the policy R/W semaphore held, which
853 	 * guarantees that the driver pointer will not change while it is
854 	 * running.
855 	 */
856 	if (!intel_pstate_driver)
857 		return -EAGAIN;
858 
859 	mutex_lock(&intel_pstate_limits_lock);
860 
861 	if (intel_pstate_driver == &intel_pstate) {
862 		ret = intel_pstate_set_energy_pref_index(cpu, ret, raw, epp);
863 	} else {
864 		/*
865 		 * In the passive mode the governor needs to be stopped on the
866 		 * target CPU before the EPP update and restarted after it,
867 		 * which is super-heavy-weight, so make sure it is worth doing
868 		 * upfront.
869 		 */
870 		if (!raw)
871 			epp = ret ? epp_values[ret] : cpu->epp_default;
872 
873 		if (cpu->epp_cached != epp) {
874 			int err;
875 
876 			cpufreq_stop_governor(policy);
877 			ret = intel_pstate_set_epp(cpu, epp);
878 			err = cpufreq_start_governor(policy);
879 			if (!ret)
880 				ret = err;
881 		} else {
882 			ret = 0;
883 		}
884 	}
885 
886 	mutex_unlock(&intel_pstate_limits_lock);
887 
888 	return ret ?: count;
889 }
890 
891 static ssize_t show_energy_performance_preference(
892 				struct cpufreq_policy *policy, char *buf)
893 {
894 	struct cpudata *cpu_data = all_cpu_data[policy->cpu];
895 	int preference, raw_epp;
896 
897 	preference = intel_pstate_get_energy_pref_index(cpu_data, &raw_epp);
898 	if (preference < 0)
899 		return preference;
900 
901 	if (raw_epp)
902 		return  sprintf(buf, "%d\n", raw_epp);
903 	else
904 		return  sprintf(buf, "%s\n", energy_perf_strings[preference]);
905 }
906 
907 cpufreq_freq_attr_rw(energy_performance_preference);
908 
909 static ssize_t show_base_frequency(struct cpufreq_policy *policy, char *buf)
910 {
911 	struct cpudata *cpu = all_cpu_data[policy->cpu];
912 	int ratio, freq;
913 
914 	ratio = intel_pstate_get_cppc_guaranteed(policy->cpu);
915 	if (ratio <= 0) {
916 		u64 cap;
917 
918 		rdmsrl_on_cpu(policy->cpu, MSR_HWP_CAPABILITIES, &cap);
919 		ratio = HWP_GUARANTEED_PERF(cap);
920 	}
921 
922 	freq = ratio * cpu->pstate.scaling;
923 	if (cpu->pstate.scaling != cpu->pstate.perf_ctl_scaling)
924 		freq = rounddown(freq, cpu->pstate.perf_ctl_scaling);
925 
926 	return sprintf(buf, "%d\n", freq);
927 }
928 
929 cpufreq_freq_attr_ro(base_frequency);
930 
931 static struct freq_attr *hwp_cpufreq_attrs[] = {
932 	&energy_performance_preference,
933 	&energy_performance_available_preferences,
934 	&base_frequency,
935 	NULL,
936 };
937 
938 static void __intel_pstate_get_hwp_cap(struct cpudata *cpu)
939 {
940 	u64 cap;
941 
942 	rdmsrl_on_cpu(cpu->cpu, MSR_HWP_CAPABILITIES, &cap);
943 	WRITE_ONCE(cpu->hwp_cap_cached, cap);
944 	cpu->pstate.max_pstate = HWP_GUARANTEED_PERF(cap);
945 	cpu->pstate.turbo_pstate = HWP_HIGHEST_PERF(cap);
946 }
947 
948 static void intel_pstate_get_hwp_cap(struct cpudata *cpu)
949 {
950 	int scaling = cpu->pstate.scaling;
951 
952 	__intel_pstate_get_hwp_cap(cpu);
953 
954 	cpu->pstate.max_freq = cpu->pstate.max_pstate * scaling;
955 	cpu->pstate.turbo_freq = cpu->pstate.turbo_pstate * scaling;
956 	if (scaling != cpu->pstate.perf_ctl_scaling) {
957 		int perf_ctl_scaling = cpu->pstate.perf_ctl_scaling;
958 
959 		cpu->pstate.max_freq = rounddown(cpu->pstate.max_freq,
960 						 perf_ctl_scaling);
961 		cpu->pstate.turbo_freq = rounddown(cpu->pstate.turbo_freq,
962 						   perf_ctl_scaling);
963 	}
964 }
965 
966 static void intel_pstate_hwp_set(unsigned int cpu)
967 {
968 	struct cpudata *cpu_data = all_cpu_data[cpu];
969 	int max, min;
970 	u64 value;
971 	s16 epp;
972 
973 	max = cpu_data->max_perf_ratio;
974 	min = cpu_data->min_perf_ratio;
975 
976 	if (cpu_data->policy == CPUFREQ_POLICY_PERFORMANCE)
977 		min = max;
978 
979 	rdmsrl_on_cpu(cpu, MSR_HWP_REQUEST, &value);
980 
981 	value &= ~HWP_MIN_PERF(~0L);
982 	value |= HWP_MIN_PERF(min);
983 
984 	value &= ~HWP_MAX_PERF(~0L);
985 	value |= HWP_MAX_PERF(max);
986 
987 	if (cpu_data->epp_policy == cpu_data->policy)
988 		goto skip_epp;
989 
990 	cpu_data->epp_policy = cpu_data->policy;
991 
992 	if (cpu_data->policy == CPUFREQ_POLICY_PERFORMANCE) {
993 		epp = intel_pstate_get_epp(cpu_data, value);
994 		cpu_data->epp_powersave = epp;
995 		/* If EPP read was failed, then don't try to write */
996 		if (epp < 0)
997 			goto skip_epp;
998 
999 		epp = 0;
1000 	} else {
1001 		/* skip setting EPP, when saved value is invalid */
1002 		if (cpu_data->epp_powersave < 0)
1003 			goto skip_epp;
1004 
1005 		/*
1006 		 * No need to restore EPP when it is not zero. This
1007 		 * means:
1008 		 *  - Policy is not changed
1009 		 *  - user has manually changed
1010 		 *  - Error reading EPB
1011 		 */
1012 		epp = intel_pstate_get_epp(cpu_data, value);
1013 		if (epp)
1014 			goto skip_epp;
1015 
1016 		epp = cpu_data->epp_powersave;
1017 	}
1018 	if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
1019 		value &= ~GENMASK_ULL(31, 24);
1020 		value |= (u64)epp << 24;
1021 	} else {
1022 		intel_pstate_set_epb(cpu, epp);
1023 	}
1024 skip_epp:
1025 	WRITE_ONCE(cpu_data->hwp_req_cached, value);
1026 	wrmsrl_on_cpu(cpu, MSR_HWP_REQUEST, value);
1027 }
1028 
1029 static void intel_pstate_disable_hwp_interrupt(struct cpudata *cpudata);
1030 
1031 static void intel_pstate_hwp_offline(struct cpudata *cpu)
1032 {
1033 	u64 value = READ_ONCE(cpu->hwp_req_cached);
1034 	int min_perf;
1035 
1036 	intel_pstate_disable_hwp_interrupt(cpu);
1037 
1038 	if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
1039 		/*
1040 		 * In case the EPP has been set to "performance" by the
1041 		 * active mode "performance" scaling algorithm, replace that
1042 		 * temporary value with the cached EPP one.
1043 		 */
1044 		value &= ~GENMASK_ULL(31, 24);
1045 		value |= HWP_ENERGY_PERF_PREFERENCE(cpu->epp_cached);
1046 		/*
1047 		 * However, make sure that EPP will be set to "performance" when
1048 		 * the CPU is brought back online again and the "performance"
1049 		 * scaling algorithm is still in effect.
1050 		 */
1051 		cpu->epp_policy = CPUFREQ_POLICY_UNKNOWN;
1052 	}
1053 
1054 	/*
1055 	 * Clear the desired perf field in the cached HWP request value to
1056 	 * prevent nonzero desired values from being leaked into the active
1057 	 * mode.
1058 	 */
1059 	value &= ~HWP_DESIRED_PERF(~0L);
1060 	WRITE_ONCE(cpu->hwp_req_cached, value);
1061 
1062 	value &= ~GENMASK_ULL(31, 0);
1063 	min_perf = HWP_LOWEST_PERF(READ_ONCE(cpu->hwp_cap_cached));
1064 
1065 	/* Set hwp_max = hwp_min */
1066 	value |= HWP_MAX_PERF(min_perf);
1067 	value |= HWP_MIN_PERF(min_perf);
1068 
1069 	/* Set EPP to min */
1070 	if (boot_cpu_has(X86_FEATURE_HWP_EPP))
1071 		value |= HWP_ENERGY_PERF_PREFERENCE(HWP_EPP_POWERSAVE);
1072 
1073 	wrmsrl_on_cpu(cpu->cpu, MSR_HWP_REQUEST, value);
1074 }
1075 
1076 #define POWER_CTL_EE_ENABLE	1
1077 #define POWER_CTL_EE_DISABLE	2
1078 
1079 static int power_ctl_ee_state;
1080 
1081 static void set_power_ctl_ee_state(bool input)
1082 {
1083 	u64 power_ctl;
1084 
1085 	mutex_lock(&intel_pstate_driver_lock);
1086 	rdmsrl(MSR_IA32_POWER_CTL, power_ctl);
1087 	if (input) {
1088 		power_ctl &= ~BIT(MSR_IA32_POWER_CTL_BIT_EE);
1089 		power_ctl_ee_state = POWER_CTL_EE_ENABLE;
1090 	} else {
1091 		power_ctl |= BIT(MSR_IA32_POWER_CTL_BIT_EE);
1092 		power_ctl_ee_state = POWER_CTL_EE_DISABLE;
1093 	}
1094 	wrmsrl(MSR_IA32_POWER_CTL, power_ctl);
1095 	mutex_unlock(&intel_pstate_driver_lock);
1096 }
1097 
1098 static void intel_pstate_hwp_enable(struct cpudata *cpudata);
1099 
1100 static void intel_pstate_hwp_reenable(struct cpudata *cpu)
1101 {
1102 	intel_pstate_hwp_enable(cpu);
1103 	wrmsrl_on_cpu(cpu->cpu, MSR_HWP_REQUEST, READ_ONCE(cpu->hwp_req_cached));
1104 }
1105 
1106 static int intel_pstate_suspend(struct cpufreq_policy *policy)
1107 {
1108 	struct cpudata *cpu = all_cpu_data[policy->cpu];
1109 
1110 	pr_debug("CPU %d suspending\n", cpu->cpu);
1111 
1112 	cpu->suspended = true;
1113 
1114 	/* disable HWP interrupt and cancel any pending work */
1115 	intel_pstate_disable_hwp_interrupt(cpu);
1116 
1117 	return 0;
1118 }
1119 
1120 static int intel_pstate_resume(struct cpufreq_policy *policy)
1121 {
1122 	struct cpudata *cpu = all_cpu_data[policy->cpu];
1123 
1124 	pr_debug("CPU %d resuming\n", cpu->cpu);
1125 
1126 	/* Only restore if the system default is changed */
1127 	if (power_ctl_ee_state == POWER_CTL_EE_ENABLE)
1128 		set_power_ctl_ee_state(true);
1129 	else if (power_ctl_ee_state == POWER_CTL_EE_DISABLE)
1130 		set_power_ctl_ee_state(false);
1131 
1132 	if (cpu->suspended && hwp_active) {
1133 		mutex_lock(&intel_pstate_limits_lock);
1134 
1135 		/* Re-enable HWP, because "online" has not done that. */
1136 		intel_pstate_hwp_reenable(cpu);
1137 
1138 		mutex_unlock(&intel_pstate_limits_lock);
1139 	}
1140 
1141 	cpu->suspended = false;
1142 
1143 	return 0;
1144 }
1145 
1146 static void intel_pstate_update_policies(void)
1147 {
1148 	int cpu;
1149 
1150 	for_each_possible_cpu(cpu)
1151 		cpufreq_update_policy(cpu);
1152 }
1153 
1154 static void __intel_pstate_update_max_freq(struct cpudata *cpudata,
1155 					   struct cpufreq_policy *policy)
1156 {
1157 	policy->cpuinfo.max_freq = global.turbo_disabled_mf ?
1158 			cpudata->pstate.max_freq : cpudata->pstate.turbo_freq;
1159 	refresh_frequency_limits(policy);
1160 }
1161 
1162 static void intel_pstate_update_max_freq(unsigned int cpu)
1163 {
1164 	struct cpufreq_policy *policy = cpufreq_cpu_acquire(cpu);
1165 
1166 	if (!policy)
1167 		return;
1168 
1169 	__intel_pstate_update_max_freq(all_cpu_data[cpu], policy);
1170 
1171 	cpufreq_cpu_release(policy);
1172 }
1173 
1174 static void intel_pstate_update_limits(unsigned int cpu)
1175 {
1176 	mutex_lock(&intel_pstate_driver_lock);
1177 
1178 	update_turbo_state();
1179 	/*
1180 	 * If turbo has been turned on or off globally, policy limits for
1181 	 * all CPUs need to be updated to reflect that.
1182 	 */
1183 	if (global.turbo_disabled_mf != global.turbo_disabled) {
1184 		global.turbo_disabled_mf = global.turbo_disabled;
1185 		arch_set_max_freq_ratio(global.turbo_disabled);
1186 		for_each_possible_cpu(cpu)
1187 			intel_pstate_update_max_freq(cpu);
1188 	} else {
1189 		cpufreq_update_policy(cpu);
1190 	}
1191 
1192 	mutex_unlock(&intel_pstate_driver_lock);
1193 }
1194 
1195 /************************** sysfs begin ************************/
1196 #define show_one(file_name, object)					\
1197 	static ssize_t show_##file_name					\
1198 	(struct kobject *kobj, struct kobj_attribute *attr, char *buf)	\
1199 	{								\
1200 		return sprintf(buf, "%u\n", global.object);		\
1201 	}
1202 
1203 static ssize_t intel_pstate_show_status(char *buf);
1204 static int intel_pstate_update_status(const char *buf, size_t size);
1205 
1206 static ssize_t show_status(struct kobject *kobj,
1207 			   struct kobj_attribute *attr, char *buf)
1208 {
1209 	ssize_t ret;
1210 
1211 	mutex_lock(&intel_pstate_driver_lock);
1212 	ret = intel_pstate_show_status(buf);
1213 	mutex_unlock(&intel_pstate_driver_lock);
1214 
1215 	return ret;
1216 }
1217 
1218 static ssize_t store_status(struct kobject *a, struct kobj_attribute *b,
1219 			    const char *buf, size_t count)
1220 {
1221 	char *p = memchr(buf, '\n', count);
1222 	int ret;
1223 
1224 	mutex_lock(&intel_pstate_driver_lock);
1225 	ret = intel_pstate_update_status(buf, p ? p - buf : count);
1226 	mutex_unlock(&intel_pstate_driver_lock);
1227 
1228 	return ret < 0 ? ret : count;
1229 }
1230 
1231 static ssize_t show_turbo_pct(struct kobject *kobj,
1232 				struct kobj_attribute *attr, char *buf)
1233 {
1234 	struct cpudata *cpu;
1235 	int total, no_turbo, turbo_pct;
1236 	uint32_t turbo_fp;
1237 
1238 	mutex_lock(&intel_pstate_driver_lock);
1239 
1240 	if (!intel_pstate_driver) {
1241 		mutex_unlock(&intel_pstate_driver_lock);
1242 		return -EAGAIN;
1243 	}
1244 
1245 	cpu = all_cpu_data[0];
1246 
1247 	total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
1248 	no_turbo = cpu->pstate.max_pstate - cpu->pstate.min_pstate + 1;
1249 	turbo_fp = div_fp(no_turbo, total);
1250 	turbo_pct = 100 - fp_toint(mul_fp(turbo_fp, int_tofp(100)));
1251 
1252 	mutex_unlock(&intel_pstate_driver_lock);
1253 
1254 	return sprintf(buf, "%u\n", turbo_pct);
1255 }
1256 
1257 static ssize_t show_num_pstates(struct kobject *kobj,
1258 				struct kobj_attribute *attr, char *buf)
1259 {
1260 	struct cpudata *cpu;
1261 	int total;
1262 
1263 	mutex_lock(&intel_pstate_driver_lock);
1264 
1265 	if (!intel_pstate_driver) {
1266 		mutex_unlock(&intel_pstate_driver_lock);
1267 		return -EAGAIN;
1268 	}
1269 
1270 	cpu = all_cpu_data[0];
1271 	total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
1272 
1273 	mutex_unlock(&intel_pstate_driver_lock);
1274 
1275 	return sprintf(buf, "%u\n", total);
1276 }
1277 
1278 static ssize_t show_no_turbo(struct kobject *kobj,
1279 			     struct kobj_attribute *attr, char *buf)
1280 {
1281 	ssize_t ret;
1282 
1283 	mutex_lock(&intel_pstate_driver_lock);
1284 
1285 	if (!intel_pstate_driver) {
1286 		mutex_unlock(&intel_pstate_driver_lock);
1287 		return -EAGAIN;
1288 	}
1289 
1290 	update_turbo_state();
1291 	if (global.turbo_disabled)
1292 		ret = sprintf(buf, "%u\n", global.turbo_disabled);
1293 	else
1294 		ret = sprintf(buf, "%u\n", global.no_turbo);
1295 
1296 	mutex_unlock(&intel_pstate_driver_lock);
1297 
1298 	return ret;
1299 }
1300 
1301 static ssize_t store_no_turbo(struct kobject *a, struct kobj_attribute *b,
1302 			      const char *buf, size_t count)
1303 {
1304 	unsigned int input;
1305 	int ret;
1306 
1307 	ret = sscanf(buf, "%u", &input);
1308 	if (ret != 1)
1309 		return -EINVAL;
1310 
1311 	mutex_lock(&intel_pstate_driver_lock);
1312 
1313 	if (!intel_pstate_driver) {
1314 		mutex_unlock(&intel_pstate_driver_lock);
1315 		return -EAGAIN;
1316 	}
1317 
1318 	mutex_lock(&intel_pstate_limits_lock);
1319 
1320 	update_turbo_state();
1321 	if (global.turbo_disabled) {
1322 		pr_notice_once("Turbo disabled by BIOS or unavailable on processor\n");
1323 		mutex_unlock(&intel_pstate_limits_lock);
1324 		mutex_unlock(&intel_pstate_driver_lock);
1325 		return -EPERM;
1326 	}
1327 
1328 	global.no_turbo = clamp_t(int, input, 0, 1);
1329 
1330 	if (global.no_turbo) {
1331 		struct cpudata *cpu = all_cpu_data[0];
1332 		int pct = cpu->pstate.max_pstate * 100 / cpu->pstate.turbo_pstate;
1333 
1334 		/* Squash the global minimum into the permitted range. */
1335 		if (global.min_perf_pct > pct)
1336 			global.min_perf_pct = pct;
1337 	}
1338 
1339 	mutex_unlock(&intel_pstate_limits_lock);
1340 
1341 	intel_pstate_update_policies();
1342 	arch_set_max_freq_ratio(global.no_turbo);
1343 
1344 	mutex_unlock(&intel_pstate_driver_lock);
1345 
1346 	return count;
1347 }
1348 
1349 static void update_qos_request(enum freq_qos_req_type type)
1350 {
1351 	struct freq_qos_request *req;
1352 	struct cpufreq_policy *policy;
1353 	int i;
1354 
1355 	for_each_possible_cpu(i) {
1356 		struct cpudata *cpu = all_cpu_data[i];
1357 		unsigned int freq, perf_pct;
1358 
1359 		policy = cpufreq_cpu_get(i);
1360 		if (!policy)
1361 			continue;
1362 
1363 		req = policy->driver_data;
1364 		cpufreq_cpu_put(policy);
1365 
1366 		if (!req)
1367 			continue;
1368 
1369 		if (hwp_active)
1370 			intel_pstate_get_hwp_cap(cpu);
1371 
1372 		if (type == FREQ_QOS_MIN) {
1373 			perf_pct = global.min_perf_pct;
1374 		} else {
1375 			req++;
1376 			perf_pct = global.max_perf_pct;
1377 		}
1378 
1379 		freq = DIV_ROUND_UP(cpu->pstate.turbo_freq * perf_pct, 100);
1380 
1381 		if (freq_qos_update_request(req, freq) < 0)
1382 			pr_warn("Failed to update freq constraint: CPU%d\n", i);
1383 	}
1384 }
1385 
1386 static ssize_t store_max_perf_pct(struct kobject *a, struct kobj_attribute *b,
1387 				  const char *buf, size_t count)
1388 {
1389 	unsigned int input;
1390 	int ret;
1391 
1392 	ret = sscanf(buf, "%u", &input);
1393 	if (ret != 1)
1394 		return -EINVAL;
1395 
1396 	mutex_lock(&intel_pstate_driver_lock);
1397 
1398 	if (!intel_pstate_driver) {
1399 		mutex_unlock(&intel_pstate_driver_lock);
1400 		return -EAGAIN;
1401 	}
1402 
1403 	mutex_lock(&intel_pstate_limits_lock);
1404 
1405 	global.max_perf_pct = clamp_t(int, input, global.min_perf_pct, 100);
1406 
1407 	mutex_unlock(&intel_pstate_limits_lock);
1408 
1409 	if (intel_pstate_driver == &intel_pstate)
1410 		intel_pstate_update_policies();
1411 	else
1412 		update_qos_request(FREQ_QOS_MAX);
1413 
1414 	mutex_unlock(&intel_pstate_driver_lock);
1415 
1416 	return count;
1417 }
1418 
1419 static ssize_t store_min_perf_pct(struct kobject *a, struct kobj_attribute *b,
1420 				  const char *buf, size_t count)
1421 {
1422 	unsigned int input;
1423 	int ret;
1424 
1425 	ret = sscanf(buf, "%u", &input);
1426 	if (ret != 1)
1427 		return -EINVAL;
1428 
1429 	mutex_lock(&intel_pstate_driver_lock);
1430 
1431 	if (!intel_pstate_driver) {
1432 		mutex_unlock(&intel_pstate_driver_lock);
1433 		return -EAGAIN;
1434 	}
1435 
1436 	mutex_lock(&intel_pstate_limits_lock);
1437 
1438 	global.min_perf_pct = clamp_t(int, input,
1439 				      min_perf_pct_min(), global.max_perf_pct);
1440 
1441 	mutex_unlock(&intel_pstate_limits_lock);
1442 
1443 	if (intel_pstate_driver == &intel_pstate)
1444 		intel_pstate_update_policies();
1445 	else
1446 		update_qos_request(FREQ_QOS_MIN);
1447 
1448 	mutex_unlock(&intel_pstate_driver_lock);
1449 
1450 	return count;
1451 }
1452 
1453 static ssize_t show_hwp_dynamic_boost(struct kobject *kobj,
1454 				struct kobj_attribute *attr, char *buf)
1455 {
1456 	return sprintf(buf, "%u\n", hwp_boost);
1457 }
1458 
1459 static ssize_t store_hwp_dynamic_boost(struct kobject *a,
1460 				       struct kobj_attribute *b,
1461 				       const char *buf, size_t count)
1462 {
1463 	unsigned int input;
1464 	int ret;
1465 
1466 	ret = kstrtouint(buf, 10, &input);
1467 	if (ret)
1468 		return ret;
1469 
1470 	mutex_lock(&intel_pstate_driver_lock);
1471 	hwp_boost = !!input;
1472 	intel_pstate_update_policies();
1473 	mutex_unlock(&intel_pstate_driver_lock);
1474 
1475 	return count;
1476 }
1477 
1478 static ssize_t show_energy_efficiency(struct kobject *kobj, struct kobj_attribute *attr,
1479 				      char *buf)
1480 {
1481 	u64 power_ctl;
1482 	int enable;
1483 
1484 	rdmsrl(MSR_IA32_POWER_CTL, power_ctl);
1485 	enable = !!(power_ctl & BIT(MSR_IA32_POWER_CTL_BIT_EE));
1486 	return sprintf(buf, "%d\n", !enable);
1487 }
1488 
1489 static ssize_t store_energy_efficiency(struct kobject *a, struct kobj_attribute *b,
1490 				       const char *buf, size_t count)
1491 {
1492 	bool input;
1493 	int ret;
1494 
1495 	ret = kstrtobool(buf, &input);
1496 	if (ret)
1497 		return ret;
1498 
1499 	set_power_ctl_ee_state(input);
1500 
1501 	return count;
1502 }
1503 
1504 show_one(max_perf_pct, max_perf_pct);
1505 show_one(min_perf_pct, min_perf_pct);
1506 
1507 define_one_global_rw(status);
1508 define_one_global_rw(no_turbo);
1509 define_one_global_rw(max_perf_pct);
1510 define_one_global_rw(min_perf_pct);
1511 define_one_global_ro(turbo_pct);
1512 define_one_global_ro(num_pstates);
1513 define_one_global_rw(hwp_dynamic_boost);
1514 define_one_global_rw(energy_efficiency);
1515 
1516 static struct attribute *intel_pstate_attributes[] = {
1517 	&status.attr,
1518 	&no_turbo.attr,
1519 	NULL
1520 };
1521 
1522 static const struct attribute_group intel_pstate_attr_group = {
1523 	.attrs = intel_pstate_attributes,
1524 };
1525 
1526 static const struct x86_cpu_id intel_pstate_cpu_ee_disable_ids[];
1527 
1528 static struct kobject *intel_pstate_kobject;
1529 
1530 static void __init intel_pstate_sysfs_expose_params(void)
1531 {
1532 	struct device *dev_root = bus_get_dev_root(&cpu_subsys);
1533 	int rc;
1534 
1535 	if (dev_root) {
1536 		intel_pstate_kobject = kobject_create_and_add("intel_pstate", &dev_root->kobj);
1537 		put_device(dev_root);
1538 	}
1539 	if (WARN_ON(!intel_pstate_kobject))
1540 		return;
1541 
1542 	rc = sysfs_create_group(intel_pstate_kobject, &intel_pstate_attr_group);
1543 	if (WARN_ON(rc))
1544 		return;
1545 
1546 	if (!boot_cpu_has(X86_FEATURE_HYBRID_CPU)) {
1547 		rc = sysfs_create_file(intel_pstate_kobject, &turbo_pct.attr);
1548 		WARN_ON(rc);
1549 
1550 		rc = sysfs_create_file(intel_pstate_kobject, &num_pstates.attr);
1551 		WARN_ON(rc);
1552 	}
1553 
1554 	/*
1555 	 * If per cpu limits are enforced there are no global limits, so
1556 	 * return without creating max/min_perf_pct attributes
1557 	 */
1558 	if (per_cpu_limits)
1559 		return;
1560 
1561 	rc = sysfs_create_file(intel_pstate_kobject, &max_perf_pct.attr);
1562 	WARN_ON(rc);
1563 
1564 	rc = sysfs_create_file(intel_pstate_kobject, &min_perf_pct.attr);
1565 	WARN_ON(rc);
1566 
1567 	if (x86_match_cpu(intel_pstate_cpu_ee_disable_ids)) {
1568 		rc = sysfs_create_file(intel_pstate_kobject, &energy_efficiency.attr);
1569 		WARN_ON(rc);
1570 	}
1571 }
1572 
1573 static void __init intel_pstate_sysfs_remove(void)
1574 {
1575 	if (!intel_pstate_kobject)
1576 		return;
1577 
1578 	sysfs_remove_group(intel_pstate_kobject, &intel_pstate_attr_group);
1579 
1580 	if (!boot_cpu_has(X86_FEATURE_HYBRID_CPU)) {
1581 		sysfs_remove_file(intel_pstate_kobject, &num_pstates.attr);
1582 		sysfs_remove_file(intel_pstate_kobject, &turbo_pct.attr);
1583 	}
1584 
1585 	if (!per_cpu_limits) {
1586 		sysfs_remove_file(intel_pstate_kobject, &max_perf_pct.attr);
1587 		sysfs_remove_file(intel_pstate_kobject, &min_perf_pct.attr);
1588 
1589 		if (x86_match_cpu(intel_pstate_cpu_ee_disable_ids))
1590 			sysfs_remove_file(intel_pstate_kobject, &energy_efficiency.attr);
1591 	}
1592 
1593 	kobject_put(intel_pstate_kobject);
1594 }
1595 
1596 static void intel_pstate_sysfs_expose_hwp_dynamic_boost(void)
1597 {
1598 	int rc;
1599 
1600 	if (!hwp_active)
1601 		return;
1602 
1603 	rc = sysfs_create_file(intel_pstate_kobject, &hwp_dynamic_boost.attr);
1604 	WARN_ON_ONCE(rc);
1605 }
1606 
1607 static void intel_pstate_sysfs_hide_hwp_dynamic_boost(void)
1608 {
1609 	if (!hwp_active)
1610 		return;
1611 
1612 	sysfs_remove_file(intel_pstate_kobject, &hwp_dynamic_boost.attr);
1613 }
1614 
1615 /************************** sysfs end ************************/
1616 
1617 static void intel_pstate_notify_work(struct work_struct *work)
1618 {
1619 	struct cpudata *cpudata =
1620 		container_of(to_delayed_work(work), struct cpudata, hwp_notify_work);
1621 	struct cpufreq_policy *policy = cpufreq_cpu_acquire(cpudata->cpu);
1622 
1623 	if (policy) {
1624 		intel_pstate_get_hwp_cap(cpudata);
1625 		__intel_pstate_update_max_freq(cpudata, policy);
1626 
1627 		cpufreq_cpu_release(policy);
1628 	}
1629 
1630 	wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_STATUS, 0);
1631 }
1632 
1633 static DEFINE_SPINLOCK(hwp_notify_lock);
1634 static cpumask_t hwp_intr_enable_mask;
1635 
1636 void notify_hwp_interrupt(void)
1637 {
1638 	unsigned int this_cpu = smp_processor_id();
1639 	struct cpudata *cpudata;
1640 	unsigned long flags;
1641 	u64 value;
1642 
1643 	if (!READ_ONCE(hwp_active) || !boot_cpu_has(X86_FEATURE_HWP_NOTIFY))
1644 		return;
1645 
1646 	rdmsrl_safe(MSR_HWP_STATUS, &value);
1647 	if (!(value & 0x01))
1648 		return;
1649 
1650 	spin_lock_irqsave(&hwp_notify_lock, flags);
1651 
1652 	if (!cpumask_test_cpu(this_cpu, &hwp_intr_enable_mask))
1653 		goto ack_intr;
1654 
1655 	/*
1656 	 * Currently we never free all_cpu_data. And we can't reach here
1657 	 * without this allocated. But for safety for future changes, added
1658 	 * check.
1659 	 */
1660 	if (unlikely(!READ_ONCE(all_cpu_data)))
1661 		goto ack_intr;
1662 
1663 	/*
1664 	 * The free is done during cleanup, when cpufreq registry is failed.
1665 	 * We wouldn't be here if it fails on init or switch status. But for
1666 	 * future changes, added check.
1667 	 */
1668 	cpudata = READ_ONCE(all_cpu_data[this_cpu]);
1669 	if (unlikely(!cpudata))
1670 		goto ack_intr;
1671 
1672 	schedule_delayed_work(&cpudata->hwp_notify_work, msecs_to_jiffies(10));
1673 
1674 	spin_unlock_irqrestore(&hwp_notify_lock, flags);
1675 
1676 	return;
1677 
1678 ack_intr:
1679 	wrmsrl_safe(MSR_HWP_STATUS, 0);
1680 	spin_unlock_irqrestore(&hwp_notify_lock, flags);
1681 }
1682 
1683 static void intel_pstate_disable_hwp_interrupt(struct cpudata *cpudata)
1684 {
1685 	unsigned long flags;
1686 
1687 	if (!boot_cpu_has(X86_FEATURE_HWP_NOTIFY))
1688 		return;
1689 
1690 	/* wrmsrl_on_cpu has to be outside spinlock as this can result in IPC */
1691 	wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_INTERRUPT, 0x00);
1692 
1693 	spin_lock_irqsave(&hwp_notify_lock, flags);
1694 	if (cpumask_test_and_clear_cpu(cpudata->cpu, &hwp_intr_enable_mask))
1695 		cancel_delayed_work(&cpudata->hwp_notify_work);
1696 	spin_unlock_irqrestore(&hwp_notify_lock, flags);
1697 }
1698 
1699 static void intel_pstate_enable_hwp_interrupt(struct cpudata *cpudata)
1700 {
1701 	/* Enable HWP notification interrupt for guaranteed performance change */
1702 	if (boot_cpu_has(X86_FEATURE_HWP_NOTIFY)) {
1703 		unsigned long flags;
1704 
1705 		spin_lock_irqsave(&hwp_notify_lock, flags);
1706 		INIT_DELAYED_WORK(&cpudata->hwp_notify_work, intel_pstate_notify_work);
1707 		cpumask_set_cpu(cpudata->cpu, &hwp_intr_enable_mask);
1708 		spin_unlock_irqrestore(&hwp_notify_lock, flags);
1709 
1710 		/* wrmsrl_on_cpu has to be outside spinlock as this can result in IPC */
1711 		wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_INTERRUPT, 0x01);
1712 		wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_STATUS, 0);
1713 	}
1714 }
1715 
1716 static void intel_pstate_update_epp_defaults(struct cpudata *cpudata)
1717 {
1718 	cpudata->epp_default = intel_pstate_get_epp(cpudata, 0);
1719 
1720 	/*
1721 	 * If the EPP is set by firmware, which means that firmware enabled HWP
1722 	 * - Is equal or less than 0x80 (default balance_perf EPP)
1723 	 * - But less performance oriented than performance EPP
1724 	 *   then use this as new balance_perf EPP.
1725 	 */
1726 	if (hwp_forced && cpudata->epp_default <= HWP_EPP_BALANCE_PERFORMANCE &&
1727 	    cpudata->epp_default > HWP_EPP_PERFORMANCE) {
1728 		epp_values[EPP_INDEX_BALANCE_PERFORMANCE] = cpudata->epp_default;
1729 		return;
1730 	}
1731 
1732 	/*
1733 	 * If this CPU gen doesn't call for change in balance_perf
1734 	 * EPP return.
1735 	 */
1736 	if (epp_values[EPP_INDEX_BALANCE_PERFORMANCE] == HWP_EPP_BALANCE_PERFORMANCE)
1737 		return;
1738 
1739 	/*
1740 	 * Use hard coded value per gen to update the balance_perf
1741 	 * and default EPP.
1742 	 */
1743 	cpudata->epp_default = epp_values[EPP_INDEX_BALANCE_PERFORMANCE];
1744 	intel_pstate_set_epp(cpudata, cpudata->epp_default);
1745 }
1746 
1747 static void intel_pstate_hwp_enable(struct cpudata *cpudata)
1748 {
1749 	/* First disable HWP notification interrupt till we activate again */
1750 	if (boot_cpu_has(X86_FEATURE_HWP_NOTIFY))
1751 		wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_INTERRUPT, 0x00);
1752 
1753 	wrmsrl_on_cpu(cpudata->cpu, MSR_PM_ENABLE, 0x1);
1754 
1755 	intel_pstate_enable_hwp_interrupt(cpudata);
1756 
1757 	if (cpudata->epp_default >= 0)
1758 		return;
1759 
1760 	intel_pstate_update_epp_defaults(cpudata);
1761 }
1762 
1763 static int atom_get_min_pstate(int not_used)
1764 {
1765 	u64 value;
1766 
1767 	rdmsrl(MSR_ATOM_CORE_RATIOS, value);
1768 	return (value >> 8) & 0x7F;
1769 }
1770 
1771 static int atom_get_max_pstate(int not_used)
1772 {
1773 	u64 value;
1774 
1775 	rdmsrl(MSR_ATOM_CORE_RATIOS, value);
1776 	return (value >> 16) & 0x7F;
1777 }
1778 
1779 static int atom_get_turbo_pstate(int not_used)
1780 {
1781 	u64 value;
1782 
1783 	rdmsrl(MSR_ATOM_CORE_TURBO_RATIOS, value);
1784 	return value & 0x7F;
1785 }
1786 
1787 static u64 atom_get_val(struct cpudata *cpudata, int pstate)
1788 {
1789 	u64 val;
1790 	int32_t vid_fp;
1791 	u32 vid;
1792 
1793 	val = (u64)pstate << 8;
1794 	if (global.no_turbo && !global.turbo_disabled)
1795 		val |= (u64)1 << 32;
1796 
1797 	vid_fp = cpudata->vid.min + mul_fp(
1798 		int_tofp(pstate - cpudata->pstate.min_pstate),
1799 		cpudata->vid.ratio);
1800 
1801 	vid_fp = clamp_t(int32_t, vid_fp, cpudata->vid.min, cpudata->vid.max);
1802 	vid = ceiling_fp(vid_fp);
1803 
1804 	if (pstate > cpudata->pstate.max_pstate)
1805 		vid = cpudata->vid.turbo;
1806 
1807 	return val | vid;
1808 }
1809 
1810 static int silvermont_get_scaling(void)
1811 {
1812 	u64 value;
1813 	int i;
1814 	/* Defined in Table 35-6 from SDM (Sept 2015) */
1815 	static int silvermont_freq_table[] = {
1816 		83300, 100000, 133300, 116700, 80000};
1817 
1818 	rdmsrl(MSR_FSB_FREQ, value);
1819 	i = value & 0x7;
1820 	WARN_ON(i > 4);
1821 
1822 	return silvermont_freq_table[i];
1823 }
1824 
1825 static int airmont_get_scaling(void)
1826 {
1827 	u64 value;
1828 	int i;
1829 	/* Defined in Table 35-10 from SDM (Sept 2015) */
1830 	static int airmont_freq_table[] = {
1831 		83300, 100000, 133300, 116700, 80000,
1832 		93300, 90000, 88900, 87500};
1833 
1834 	rdmsrl(MSR_FSB_FREQ, value);
1835 	i = value & 0xF;
1836 	WARN_ON(i > 8);
1837 
1838 	return airmont_freq_table[i];
1839 }
1840 
1841 static void atom_get_vid(struct cpudata *cpudata)
1842 {
1843 	u64 value;
1844 
1845 	rdmsrl(MSR_ATOM_CORE_VIDS, value);
1846 	cpudata->vid.min = int_tofp((value >> 8) & 0x7f);
1847 	cpudata->vid.max = int_tofp((value >> 16) & 0x7f);
1848 	cpudata->vid.ratio = div_fp(
1849 		cpudata->vid.max - cpudata->vid.min,
1850 		int_tofp(cpudata->pstate.max_pstate -
1851 			cpudata->pstate.min_pstate));
1852 
1853 	rdmsrl(MSR_ATOM_CORE_TURBO_VIDS, value);
1854 	cpudata->vid.turbo = value & 0x7f;
1855 }
1856 
1857 static int core_get_min_pstate(int cpu)
1858 {
1859 	u64 value;
1860 
1861 	rdmsrl_on_cpu(cpu, MSR_PLATFORM_INFO, &value);
1862 	return (value >> 40) & 0xFF;
1863 }
1864 
1865 static int core_get_max_pstate_physical(int cpu)
1866 {
1867 	u64 value;
1868 
1869 	rdmsrl_on_cpu(cpu, MSR_PLATFORM_INFO, &value);
1870 	return (value >> 8) & 0xFF;
1871 }
1872 
1873 static int core_get_tdp_ratio(int cpu, u64 plat_info)
1874 {
1875 	/* Check how many TDP levels present */
1876 	if (plat_info & 0x600000000) {
1877 		u64 tdp_ctrl;
1878 		u64 tdp_ratio;
1879 		int tdp_msr;
1880 		int err;
1881 
1882 		/* Get the TDP level (0, 1, 2) to get ratios */
1883 		err = rdmsrl_safe_on_cpu(cpu, MSR_CONFIG_TDP_CONTROL, &tdp_ctrl);
1884 		if (err)
1885 			return err;
1886 
1887 		/* TDP MSR are continuous starting at 0x648 */
1888 		tdp_msr = MSR_CONFIG_TDP_NOMINAL + (tdp_ctrl & 0x03);
1889 		err = rdmsrl_safe_on_cpu(cpu, tdp_msr, &tdp_ratio);
1890 		if (err)
1891 			return err;
1892 
1893 		/* For level 1 and 2, bits[23:16] contain the ratio */
1894 		if (tdp_ctrl & 0x03)
1895 			tdp_ratio >>= 16;
1896 
1897 		tdp_ratio &= 0xff; /* ratios are only 8 bits long */
1898 		pr_debug("tdp_ratio %x\n", (int)tdp_ratio);
1899 
1900 		return (int)tdp_ratio;
1901 	}
1902 
1903 	return -ENXIO;
1904 }
1905 
1906 static int core_get_max_pstate(int cpu)
1907 {
1908 	u64 tar;
1909 	u64 plat_info;
1910 	int max_pstate;
1911 	int tdp_ratio;
1912 	int err;
1913 
1914 	rdmsrl_on_cpu(cpu, MSR_PLATFORM_INFO, &plat_info);
1915 	max_pstate = (plat_info >> 8) & 0xFF;
1916 
1917 	tdp_ratio = core_get_tdp_ratio(cpu, plat_info);
1918 	if (tdp_ratio <= 0)
1919 		return max_pstate;
1920 
1921 	if (hwp_active) {
1922 		/* Turbo activation ratio is not used on HWP platforms */
1923 		return tdp_ratio;
1924 	}
1925 
1926 	err = rdmsrl_safe_on_cpu(cpu, MSR_TURBO_ACTIVATION_RATIO, &tar);
1927 	if (!err) {
1928 		int tar_levels;
1929 
1930 		/* Do some sanity checking for safety */
1931 		tar_levels = tar & 0xff;
1932 		if (tdp_ratio - 1 == tar_levels) {
1933 			max_pstate = tar_levels;
1934 			pr_debug("max_pstate=TAC %x\n", max_pstate);
1935 		}
1936 	}
1937 
1938 	return max_pstate;
1939 }
1940 
1941 static int core_get_turbo_pstate(int cpu)
1942 {
1943 	u64 value;
1944 	int nont, ret;
1945 
1946 	rdmsrl_on_cpu(cpu, MSR_TURBO_RATIO_LIMIT, &value);
1947 	nont = core_get_max_pstate(cpu);
1948 	ret = (value) & 255;
1949 	if (ret <= nont)
1950 		ret = nont;
1951 	return ret;
1952 }
1953 
1954 static u64 core_get_val(struct cpudata *cpudata, int pstate)
1955 {
1956 	u64 val;
1957 
1958 	val = (u64)pstate << 8;
1959 	if (global.no_turbo && !global.turbo_disabled)
1960 		val |= (u64)1 << 32;
1961 
1962 	return val;
1963 }
1964 
1965 static int knl_get_aperf_mperf_shift(void)
1966 {
1967 	return 10;
1968 }
1969 
1970 static int knl_get_turbo_pstate(int cpu)
1971 {
1972 	u64 value;
1973 	int nont, ret;
1974 
1975 	rdmsrl_on_cpu(cpu, MSR_TURBO_RATIO_LIMIT, &value);
1976 	nont = core_get_max_pstate(cpu);
1977 	ret = (((value) >> 8) & 0xFF);
1978 	if (ret <= nont)
1979 		ret = nont;
1980 	return ret;
1981 }
1982 
1983 static void hybrid_get_type(void *data)
1984 {
1985 	u8 *cpu_type = data;
1986 
1987 	*cpu_type = get_this_hybrid_cpu_type();
1988 }
1989 
1990 static int hwp_get_cpu_scaling(int cpu)
1991 {
1992 	u8 cpu_type = 0;
1993 
1994 	smp_call_function_single(cpu, hybrid_get_type, &cpu_type, 1);
1995 	/* P-cores have a smaller perf level-to-freqency scaling factor. */
1996 	if (cpu_type == 0x40)
1997 		return hybrid_scaling_factor;
1998 
1999 	/* Use default core scaling for E-cores */
2000 	if (cpu_type == 0x20)
2001 		return core_get_scaling();
2002 
2003 	/*
2004 	 * If reached here, this system is either non-hybrid (like Tiger
2005 	 * Lake) or hybrid-capable (like Alder Lake or Raptor Lake) with
2006 	 * no E cores (in which case CPUID for hybrid support is 0).
2007 	 *
2008 	 * The CPPC nominal_frequency field is 0 for non-hybrid systems,
2009 	 * so the default core scaling will be used for them.
2010 	 */
2011 	return intel_pstate_cppc_get_scaling(cpu);
2012 }
2013 
2014 static void intel_pstate_set_pstate(struct cpudata *cpu, int pstate)
2015 {
2016 	trace_cpu_frequency(pstate * cpu->pstate.scaling, cpu->cpu);
2017 	cpu->pstate.current_pstate = pstate;
2018 	/*
2019 	 * Generally, there is no guarantee that this code will always run on
2020 	 * the CPU being updated, so force the register update to run on the
2021 	 * right CPU.
2022 	 */
2023 	wrmsrl_on_cpu(cpu->cpu, MSR_IA32_PERF_CTL,
2024 		      pstate_funcs.get_val(cpu, pstate));
2025 }
2026 
2027 static void intel_pstate_set_min_pstate(struct cpudata *cpu)
2028 {
2029 	intel_pstate_set_pstate(cpu, cpu->pstate.min_pstate);
2030 }
2031 
2032 static void intel_pstate_max_within_limits(struct cpudata *cpu)
2033 {
2034 	int pstate = max(cpu->pstate.min_pstate, cpu->max_perf_ratio);
2035 
2036 	update_turbo_state();
2037 	intel_pstate_set_pstate(cpu, pstate);
2038 }
2039 
2040 static void intel_pstate_get_cpu_pstates(struct cpudata *cpu)
2041 {
2042 	int perf_ctl_max_phys = pstate_funcs.get_max_physical(cpu->cpu);
2043 	int perf_ctl_scaling = pstate_funcs.get_scaling();
2044 
2045 	cpu->pstate.min_pstate = pstate_funcs.get_min(cpu->cpu);
2046 	cpu->pstate.max_pstate_physical = perf_ctl_max_phys;
2047 	cpu->pstate.perf_ctl_scaling = perf_ctl_scaling;
2048 
2049 	if (hwp_active && !hwp_mode_bdw) {
2050 		__intel_pstate_get_hwp_cap(cpu);
2051 
2052 		if (pstate_funcs.get_cpu_scaling) {
2053 			cpu->pstate.scaling = pstate_funcs.get_cpu_scaling(cpu->cpu);
2054 			if (cpu->pstate.scaling != perf_ctl_scaling)
2055 				intel_pstate_hybrid_hwp_adjust(cpu);
2056 		} else {
2057 			cpu->pstate.scaling = perf_ctl_scaling;
2058 		}
2059 	} else {
2060 		cpu->pstate.scaling = perf_ctl_scaling;
2061 		cpu->pstate.max_pstate = pstate_funcs.get_max(cpu->cpu);
2062 		cpu->pstate.turbo_pstate = pstate_funcs.get_turbo(cpu->cpu);
2063 	}
2064 
2065 	if (cpu->pstate.scaling == perf_ctl_scaling) {
2066 		cpu->pstate.min_freq = cpu->pstate.min_pstate * perf_ctl_scaling;
2067 		cpu->pstate.max_freq = cpu->pstate.max_pstate * perf_ctl_scaling;
2068 		cpu->pstate.turbo_freq = cpu->pstate.turbo_pstate * perf_ctl_scaling;
2069 	}
2070 
2071 	if (pstate_funcs.get_aperf_mperf_shift)
2072 		cpu->aperf_mperf_shift = pstate_funcs.get_aperf_mperf_shift();
2073 
2074 	if (pstate_funcs.get_vid)
2075 		pstate_funcs.get_vid(cpu);
2076 
2077 	intel_pstate_set_min_pstate(cpu);
2078 }
2079 
2080 /*
2081  * Long hold time will keep high perf limits for long time,
2082  * which negatively impacts perf/watt for some workloads,
2083  * like specpower. 3ms is based on experiements on some
2084  * workoads.
2085  */
2086 static int hwp_boost_hold_time_ns = 3 * NSEC_PER_MSEC;
2087 
2088 static inline void intel_pstate_hwp_boost_up(struct cpudata *cpu)
2089 {
2090 	u64 hwp_req = READ_ONCE(cpu->hwp_req_cached);
2091 	u64 hwp_cap = READ_ONCE(cpu->hwp_cap_cached);
2092 	u32 max_limit = (hwp_req & 0xff00) >> 8;
2093 	u32 min_limit = (hwp_req & 0xff);
2094 	u32 boost_level1;
2095 
2096 	/*
2097 	 * Cases to consider (User changes via sysfs or boot time):
2098 	 * If, P0 (Turbo max) = P1 (Guaranteed max) = min:
2099 	 *	No boost, return.
2100 	 * If, P0 (Turbo max) > P1 (Guaranteed max) = min:
2101 	 *     Should result in one level boost only for P0.
2102 	 * If, P0 (Turbo max) = P1 (Guaranteed max) > min:
2103 	 *     Should result in two level boost:
2104 	 *         (min + p1)/2 and P1.
2105 	 * If, P0 (Turbo max) > P1 (Guaranteed max) > min:
2106 	 *     Should result in three level boost:
2107 	 *        (min + p1)/2, P1 and P0.
2108 	 */
2109 
2110 	/* If max and min are equal or already at max, nothing to boost */
2111 	if (max_limit == min_limit || cpu->hwp_boost_min >= max_limit)
2112 		return;
2113 
2114 	if (!cpu->hwp_boost_min)
2115 		cpu->hwp_boost_min = min_limit;
2116 
2117 	/* level at half way mark between min and guranteed */
2118 	boost_level1 = (HWP_GUARANTEED_PERF(hwp_cap) + min_limit) >> 1;
2119 
2120 	if (cpu->hwp_boost_min < boost_level1)
2121 		cpu->hwp_boost_min = boost_level1;
2122 	else if (cpu->hwp_boost_min < HWP_GUARANTEED_PERF(hwp_cap))
2123 		cpu->hwp_boost_min = HWP_GUARANTEED_PERF(hwp_cap);
2124 	else if (cpu->hwp_boost_min == HWP_GUARANTEED_PERF(hwp_cap) &&
2125 		 max_limit != HWP_GUARANTEED_PERF(hwp_cap))
2126 		cpu->hwp_boost_min = max_limit;
2127 	else
2128 		return;
2129 
2130 	hwp_req = (hwp_req & ~GENMASK_ULL(7, 0)) | cpu->hwp_boost_min;
2131 	wrmsrl(MSR_HWP_REQUEST, hwp_req);
2132 	cpu->last_update = cpu->sample.time;
2133 }
2134 
2135 static inline void intel_pstate_hwp_boost_down(struct cpudata *cpu)
2136 {
2137 	if (cpu->hwp_boost_min) {
2138 		bool expired;
2139 
2140 		/* Check if we are idle for hold time to boost down */
2141 		expired = time_after64(cpu->sample.time, cpu->last_update +
2142 				       hwp_boost_hold_time_ns);
2143 		if (expired) {
2144 			wrmsrl(MSR_HWP_REQUEST, cpu->hwp_req_cached);
2145 			cpu->hwp_boost_min = 0;
2146 		}
2147 	}
2148 	cpu->last_update = cpu->sample.time;
2149 }
2150 
2151 static inline void intel_pstate_update_util_hwp_local(struct cpudata *cpu,
2152 						      u64 time)
2153 {
2154 	cpu->sample.time = time;
2155 
2156 	if (cpu->sched_flags & SCHED_CPUFREQ_IOWAIT) {
2157 		bool do_io = false;
2158 
2159 		cpu->sched_flags = 0;
2160 		/*
2161 		 * Set iowait_boost flag and update time. Since IO WAIT flag
2162 		 * is set all the time, we can't just conclude that there is
2163 		 * some IO bound activity is scheduled on this CPU with just
2164 		 * one occurrence. If we receive at least two in two
2165 		 * consecutive ticks, then we treat as boost candidate.
2166 		 */
2167 		if (time_before64(time, cpu->last_io_update + 2 * TICK_NSEC))
2168 			do_io = true;
2169 
2170 		cpu->last_io_update = time;
2171 
2172 		if (do_io)
2173 			intel_pstate_hwp_boost_up(cpu);
2174 
2175 	} else {
2176 		intel_pstate_hwp_boost_down(cpu);
2177 	}
2178 }
2179 
2180 static inline void intel_pstate_update_util_hwp(struct update_util_data *data,
2181 						u64 time, unsigned int flags)
2182 {
2183 	struct cpudata *cpu = container_of(data, struct cpudata, update_util);
2184 
2185 	cpu->sched_flags |= flags;
2186 
2187 	if (smp_processor_id() == cpu->cpu)
2188 		intel_pstate_update_util_hwp_local(cpu, time);
2189 }
2190 
2191 static inline void intel_pstate_calc_avg_perf(struct cpudata *cpu)
2192 {
2193 	struct sample *sample = &cpu->sample;
2194 
2195 	sample->core_avg_perf = div_ext_fp(sample->aperf, sample->mperf);
2196 }
2197 
2198 static inline bool intel_pstate_sample(struct cpudata *cpu, u64 time)
2199 {
2200 	u64 aperf, mperf;
2201 	unsigned long flags;
2202 	u64 tsc;
2203 
2204 	local_irq_save(flags);
2205 	rdmsrl(MSR_IA32_APERF, aperf);
2206 	rdmsrl(MSR_IA32_MPERF, mperf);
2207 	tsc = rdtsc();
2208 	if (cpu->prev_mperf == mperf || cpu->prev_tsc == tsc) {
2209 		local_irq_restore(flags);
2210 		return false;
2211 	}
2212 	local_irq_restore(flags);
2213 
2214 	cpu->last_sample_time = cpu->sample.time;
2215 	cpu->sample.time = time;
2216 	cpu->sample.aperf = aperf;
2217 	cpu->sample.mperf = mperf;
2218 	cpu->sample.tsc =  tsc;
2219 	cpu->sample.aperf -= cpu->prev_aperf;
2220 	cpu->sample.mperf -= cpu->prev_mperf;
2221 	cpu->sample.tsc -= cpu->prev_tsc;
2222 
2223 	cpu->prev_aperf = aperf;
2224 	cpu->prev_mperf = mperf;
2225 	cpu->prev_tsc = tsc;
2226 	/*
2227 	 * First time this function is invoked in a given cycle, all of the
2228 	 * previous sample data fields are equal to zero or stale and they must
2229 	 * be populated with meaningful numbers for things to work, so assume
2230 	 * that sample.time will always be reset before setting the utilization
2231 	 * update hook and make the caller skip the sample then.
2232 	 */
2233 	if (cpu->last_sample_time) {
2234 		intel_pstate_calc_avg_perf(cpu);
2235 		return true;
2236 	}
2237 	return false;
2238 }
2239 
2240 static inline int32_t get_avg_frequency(struct cpudata *cpu)
2241 {
2242 	return mul_ext_fp(cpu->sample.core_avg_perf, cpu_khz);
2243 }
2244 
2245 static inline int32_t get_avg_pstate(struct cpudata *cpu)
2246 {
2247 	return mul_ext_fp(cpu->pstate.max_pstate_physical,
2248 			  cpu->sample.core_avg_perf);
2249 }
2250 
2251 static inline int32_t get_target_pstate(struct cpudata *cpu)
2252 {
2253 	struct sample *sample = &cpu->sample;
2254 	int32_t busy_frac;
2255 	int target, avg_pstate;
2256 
2257 	busy_frac = div_fp(sample->mperf << cpu->aperf_mperf_shift,
2258 			   sample->tsc);
2259 
2260 	if (busy_frac < cpu->iowait_boost)
2261 		busy_frac = cpu->iowait_boost;
2262 
2263 	sample->busy_scaled = busy_frac * 100;
2264 
2265 	target = global.no_turbo || global.turbo_disabled ?
2266 			cpu->pstate.max_pstate : cpu->pstate.turbo_pstate;
2267 	target += target >> 2;
2268 	target = mul_fp(target, busy_frac);
2269 	if (target < cpu->pstate.min_pstate)
2270 		target = cpu->pstate.min_pstate;
2271 
2272 	/*
2273 	 * If the average P-state during the previous cycle was higher than the
2274 	 * current target, add 50% of the difference to the target to reduce
2275 	 * possible performance oscillations and offset possible performance
2276 	 * loss related to moving the workload from one CPU to another within
2277 	 * a package/module.
2278 	 */
2279 	avg_pstate = get_avg_pstate(cpu);
2280 	if (avg_pstate > target)
2281 		target += (avg_pstate - target) >> 1;
2282 
2283 	return target;
2284 }
2285 
2286 static int intel_pstate_prepare_request(struct cpudata *cpu, int pstate)
2287 {
2288 	int min_pstate = max(cpu->pstate.min_pstate, cpu->min_perf_ratio);
2289 	int max_pstate = max(min_pstate, cpu->max_perf_ratio);
2290 
2291 	return clamp_t(int, pstate, min_pstate, max_pstate);
2292 }
2293 
2294 static void intel_pstate_update_pstate(struct cpudata *cpu, int pstate)
2295 {
2296 	if (pstate == cpu->pstate.current_pstate)
2297 		return;
2298 
2299 	cpu->pstate.current_pstate = pstate;
2300 	wrmsrl(MSR_IA32_PERF_CTL, pstate_funcs.get_val(cpu, pstate));
2301 }
2302 
2303 static void intel_pstate_adjust_pstate(struct cpudata *cpu)
2304 {
2305 	int from = cpu->pstate.current_pstate;
2306 	struct sample *sample;
2307 	int target_pstate;
2308 
2309 	update_turbo_state();
2310 
2311 	target_pstate = get_target_pstate(cpu);
2312 	target_pstate = intel_pstate_prepare_request(cpu, target_pstate);
2313 	trace_cpu_frequency(target_pstate * cpu->pstate.scaling, cpu->cpu);
2314 	intel_pstate_update_pstate(cpu, target_pstate);
2315 
2316 	sample = &cpu->sample;
2317 	trace_pstate_sample(mul_ext_fp(100, sample->core_avg_perf),
2318 		fp_toint(sample->busy_scaled),
2319 		from,
2320 		cpu->pstate.current_pstate,
2321 		sample->mperf,
2322 		sample->aperf,
2323 		sample->tsc,
2324 		get_avg_frequency(cpu),
2325 		fp_toint(cpu->iowait_boost * 100));
2326 }
2327 
2328 static void intel_pstate_update_util(struct update_util_data *data, u64 time,
2329 				     unsigned int flags)
2330 {
2331 	struct cpudata *cpu = container_of(data, struct cpudata, update_util);
2332 	u64 delta_ns;
2333 
2334 	/* Don't allow remote callbacks */
2335 	if (smp_processor_id() != cpu->cpu)
2336 		return;
2337 
2338 	delta_ns = time - cpu->last_update;
2339 	if (flags & SCHED_CPUFREQ_IOWAIT) {
2340 		/* Start over if the CPU may have been idle. */
2341 		if (delta_ns > TICK_NSEC) {
2342 			cpu->iowait_boost = ONE_EIGHTH_FP;
2343 		} else if (cpu->iowait_boost >= ONE_EIGHTH_FP) {
2344 			cpu->iowait_boost <<= 1;
2345 			if (cpu->iowait_boost > int_tofp(1))
2346 				cpu->iowait_boost = int_tofp(1);
2347 		} else {
2348 			cpu->iowait_boost = ONE_EIGHTH_FP;
2349 		}
2350 	} else if (cpu->iowait_boost) {
2351 		/* Clear iowait_boost if the CPU may have been idle. */
2352 		if (delta_ns > TICK_NSEC)
2353 			cpu->iowait_boost = 0;
2354 		else
2355 			cpu->iowait_boost >>= 1;
2356 	}
2357 	cpu->last_update = time;
2358 	delta_ns = time - cpu->sample.time;
2359 	if ((s64)delta_ns < INTEL_PSTATE_SAMPLING_INTERVAL)
2360 		return;
2361 
2362 	if (intel_pstate_sample(cpu, time))
2363 		intel_pstate_adjust_pstate(cpu);
2364 }
2365 
2366 static struct pstate_funcs core_funcs = {
2367 	.get_max = core_get_max_pstate,
2368 	.get_max_physical = core_get_max_pstate_physical,
2369 	.get_min = core_get_min_pstate,
2370 	.get_turbo = core_get_turbo_pstate,
2371 	.get_scaling = core_get_scaling,
2372 	.get_val = core_get_val,
2373 };
2374 
2375 static const struct pstate_funcs silvermont_funcs = {
2376 	.get_max = atom_get_max_pstate,
2377 	.get_max_physical = atom_get_max_pstate,
2378 	.get_min = atom_get_min_pstate,
2379 	.get_turbo = atom_get_turbo_pstate,
2380 	.get_val = atom_get_val,
2381 	.get_scaling = silvermont_get_scaling,
2382 	.get_vid = atom_get_vid,
2383 };
2384 
2385 static const struct pstate_funcs airmont_funcs = {
2386 	.get_max = atom_get_max_pstate,
2387 	.get_max_physical = atom_get_max_pstate,
2388 	.get_min = atom_get_min_pstate,
2389 	.get_turbo = atom_get_turbo_pstate,
2390 	.get_val = atom_get_val,
2391 	.get_scaling = airmont_get_scaling,
2392 	.get_vid = atom_get_vid,
2393 };
2394 
2395 static const struct pstate_funcs knl_funcs = {
2396 	.get_max = core_get_max_pstate,
2397 	.get_max_physical = core_get_max_pstate_physical,
2398 	.get_min = core_get_min_pstate,
2399 	.get_turbo = knl_get_turbo_pstate,
2400 	.get_aperf_mperf_shift = knl_get_aperf_mperf_shift,
2401 	.get_scaling = core_get_scaling,
2402 	.get_val = core_get_val,
2403 };
2404 
2405 #define X86_MATCH(model, policy)					 \
2406 	X86_MATCH_VENDOR_FAM_MODEL_FEATURE(INTEL, 6, INTEL_FAM6_##model, \
2407 					   X86_FEATURE_APERFMPERF, &policy)
2408 
2409 static const struct x86_cpu_id intel_pstate_cpu_ids[] = {
2410 	X86_MATCH(SANDYBRIDGE,		core_funcs),
2411 	X86_MATCH(SANDYBRIDGE_X,	core_funcs),
2412 	X86_MATCH(ATOM_SILVERMONT,	silvermont_funcs),
2413 	X86_MATCH(IVYBRIDGE,		core_funcs),
2414 	X86_MATCH(HASWELL,		core_funcs),
2415 	X86_MATCH(BROADWELL,		core_funcs),
2416 	X86_MATCH(IVYBRIDGE_X,		core_funcs),
2417 	X86_MATCH(HASWELL_X,		core_funcs),
2418 	X86_MATCH(HASWELL_L,		core_funcs),
2419 	X86_MATCH(HASWELL_G,		core_funcs),
2420 	X86_MATCH(BROADWELL_G,		core_funcs),
2421 	X86_MATCH(ATOM_AIRMONT,		airmont_funcs),
2422 	X86_MATCH(SKYLAKE_L,		core_funcs),
2423 	X86_MATCH(BROADWELL_X,		core_funcs),
2424 	X86_MATCH(SKYLAKE,		core_funcs),
2425 	X86_MATCH(BROADWELL_D,		core_funcs),
2426 	X86_MATCH(XEON_PHI_KNL,		knl_funcs),
2427 	X86_MATCH(XEON_PHI_KNM,		knl_funcs),
2428 	X86_MATCH(ATOM_GOLDMONT,	core_funcs),
2429 	X86_MATCH(ATOM_GOLDMONT_PLUS,	core_funcs),
2430 	X86_MATCH(SKYLAKE_X,		core_funcs),
2431 	X86_MATCH(COMETLAKE,		core_funcs),
2432 	X86_MATCH(ICELAKE_X,		core_funcs),
2433 	X86_MATCH(TIGERLAKE,		core_funcs),
2434 	X86_MATCH(SAPPHIRERAPIDS_X,	core_funcs),
2435 	X86_MATCH(EMERALDRAPIDS_X,      core_funcs),
2436 	{}
2437 };
2438 MODULE_DEVICE_TABLE(x86cpu, intel_pstate_cpu_ids);
2439 
2440 static const struct x86_cpu_id intel_pstate_cpu_oob_ids[] __initconst = {
2441 	X86_MATCH(BROADWELL_D,		core_funcs),
2442 	X86_MATCH(BROADWELL_X,		core_funcs),
2443 	X86_MATCH(SKYLAKE_X,		core_funcs),
2444 	X86_MATCH(ICELAKE_X,		core_funcs),
2445 	X86_MATCH(SAPPHIRERAPIDS_X,	core_funcs),
2446 	{}
2447 };
2448 
2449 static const struct x86_cpu_id intel_pstate_cpu_ee_disable_ids[] = {
2450 	X86_MATCH(KABYLAKE,		core_funcs),
2451 	{}
2452 };
2453 
2454 static int intel_pstate_init_cpu(unsigned int cpunum)
2455 {
2456 	struct cpudata *cpu;
2457 
2458 	cpu = all_cpu_data[cpunum];
2459 
2460 	if (!cpu) {
2461 		cpu = kzalloc(sizeof(*cpu), GFP_KERNEL);
2462 		if (!cpu)
2463 			return -ENOMEM;
2464 
2465 		WRITE_ONCE(all_cpu_data[cpunum], cpu);
2466 
2467 		cpu->cpu = cpunum;
2468 
2469 		cpu->epp_default = -EINVAL;
2470 
2471 		if (hwp_active) {
2472 			intel_pstate_hwp_enable(cpu);
2473 
2474 			if (intel_pstate_acpi_pm_profile_server())
2475 				hwp_boost = true;
2476 		}
2477 	} else if (hwp_active) {
2478 		/*
2479 		 * Re-enable HWP in case this happens after a resume from ACPI
2480 		 * S3 if the CPU was offline during the whole system/resume
2481 		 * cycle.
2482 		 */
2483 		intel_pstate_hwp_reenable(cpu);
2484 	}
2485 
2486 	cpu->epp_powersave = -EINVAL;
2487 	cpu->epp_policy = 0;
2488 
2489 	intel_pstate_get_cpu_pstates(cpu);
2490 
2491 	pr_debug("controlling: cpu %d\n", cpunum);
2492 
2493 	return 0;
2494 }
2495 
2496 static void intel_pstate_set_update_util_hook(unsigned int cpu_num)
2497 {
2498 	struct cpudata *cpu = all_cpu_data[cpu_num];
2499 
2500 	if (hwp_active && !hwp_boost)
2501 		return;
2502 
2503 	if (cpu->update_util_set)
2504 		return;
2505 
2506 	/* Prevent intel_pstate_update_util() from using stale data. */
2507 	cpu->sample.time = 0;
2508 	cpufreq_add_update_util_hook(cpu_num, &cpu->update_util,
2509 				     (hwp_active ?
2510 				      intel_pstate_update_util_hwp :
2511 				      intel_pstate_update_util));
2512 	cpu->update_util_set = true;
2513 }
2514 
2515 static void intel_pstate_clear_update_util_hook(unsigned int cpu)
2516 {
2517 	struct cpudata *cpu_data = all_cpu_data[cpu];
2518 
2519 	if (!cpu_data->update_util_set)
2520 		return;
2521 
2522 	cpufreq_remove_update_util_hook(cpu);
2523 	cpu_data->update_util_set = false;
2524 	synchronize_rcu();
2525 }
2526 
2527 static int intel_pstate_get_max_freq(struct cpudata *cpu)
2528 {
2529 	return global.turbo_disabled || global.no_turbo ?
2530 			cpu->pstate.max_freq : cpu->pstate.turbo_freq;
2531 }
2532 
2533 static void intel_pstate_update_perf_limits(struct cpudata *cpu,
2534 					    unsigned int policy_min,
2535 					    unsigned int policy_max)
2536 {
2537 	int perf_ctl_scaling = cpu->pstate.perf_ctl_scaling;
2538 	int32_t max_policy_perf, min_policy_perf;
2539 
2540 	max_policy_perf = policy_max / perf_ctl_scaling;
2541 	if (policy_max == policy_min) {
2542 		min_policy_perf = max_policy_perf;
2543 	} else {
2544 		min_policy_perf = policy_min / perf_ctl_scaling;
2545 		min_policy_perf = clamp_t(int32_t, min_policy_perf,
2546 					  0, max_policy_perf);
2547 	}
2548 
2549 	/*
2550 	 * HWP needs some special consideration, because HWP_REQUEST uses
2551 	 * abstract values to represent performance rather than pure ratios.
2552 	 */
2553 	if (hwp_active && cpu->pstate.scaling != perf_ctl_scaling) {
2554 		int freq;
2555 
2556 		freq = max_policy_perf * perf_ctl_scaling;
2557 		max_policy_perf = intel_pstate_freq_to_hwp(cpu, freq);
2558 		freq = min_policy_perf * perf_ctl_scaling;
2559 		min_policy_perf = intel_pstate_freq_to_hwp(cpu, freq);
2560 	}
2561 
2562 	pr_debug("cpu:%d min_policy_perf:%d max_policy_perf:%d\n",
2563 		 cpu->cpu, min_policy_perf, max_policy_perf);
2564 
2565 	/* Normalize user input to [min_perf, max_perf] */
2566 	if (per_cpu_limits) {
2567 		cpu->min_perf_ratio = min_policy_perf;
2568 		cpu->max_perf_ratio = max_policy_perf;
2569 	} else {
2570 		int turbo_max = cpu->pstate.turbo_pstate;
2571 		int32_t global_min, global_max;
2572 
2573 		/* Global limits are in percent of the maximum turbo P-state. */
2574 		global_max = DIV_ROUND_UP(turbo_max * global.max_perf_pct, 100);
2575 		global_min = DIV_ROUND_UP(turbo_max * global.min_perf_pct, 100);
2576 		global_min = clamp_t(int32_t, global_min, 0, global_max);
2577 
2578 		pr_debug("cpu:%d global_min:%d global_max:%d\n", cpu->cpu,
2579 			 global_min, global_max);
2580 
2581 		cpu->min_perf_ratio = max(min_policy_perf, global_min);
2582 		cpu->min_perf_ratio = min(cpu->min_perf_ratio, max_policy_perf);
2583 		cpu->max_perf_ratio = min(max_policy_perf, global_max);
2584 		cpu->max_perf_ratio = max(min_policy_perf, cpu->max_perf_ratio);
2585 
2586 		/* Make sure min_perf <= max_perf */
2587 		cpu->min_perf_ratio = min(cpu->min_perf_ratio,
2588 					  cpu->max_perf_ratio);
2589 
2590 	}
2591 	pr_debug("cpu:%d max_perf_ratio:%d min_perf_ratio:%d\n", cpu->cpu,
2592 		 cpu->max_perf_ratio,
2593 		 cpu->min_perf_ratio);
2594 }
2595 
2596 static int intel_pstate_set_policy(struct cpufreq_policy *policy)
2597 {
2598 	struct cpudata *cpu;
2599 
2600 	if (!policy->cpuinfo.max_freq)
2601 		return -ENODEV;
2602 
2603 	pr_debug("set_policy cpuinfo.max %u policy->max %u\n",
2604 		 policy->cpuinfo.max_freq, policy->max);
2605 
2606 	cpu = all_cpu_data[policy->cpu];
2607 	cpu->policy = policy->policy;
2608 
2609 	mutex_lock(&intel_pstate_limits_lock);
2610 
2611 	intel_pstate_update_perf_limits(cpu, policy->min, policy->max);
2612 
2613 	if (cpu->policy == CPUFREQ_POLICY_PERFORMANCE) {
2614 		/*
2615 		 * NOHZ_FULL CPUs need this as the governor callback may not
2616 		 * be invoked on them.
2617 		 */
2618 		intel_pstate_clear_update_util_hook(policy->cpu);
2619 		intel_pstate_max_within_limits(cpu);
2620 	} else {
2621 		intel_pstate_set_update_util_hook(policy->cpu);
2622 	}
2623 
2624 	if (hwp_active) {
2625 		/*
2626 		 * When hwp_boost was active before and dynamically it
2627 		 * was turned off, in that case we need to clear the
2628 		 * update util hook.
2629 		 */
2630 		if (!hwp_boost)
2631 			intel_pstate_clear_update_util_hook(policy->cpu);
2632 		intel_pstate_hwp_set(policy->cpu);
2633 	}
2634 	/*
2635 	 * policy->cur is never updated with the intel_pstate driver, but it
2636 	 * is used as a stale frequency value. So, keep it within limits.
2637 	 */
2638 	policy->cur = policy->min;
2639 
2640 	mutex_unlock(&intel_pstate_limits_lock);
2641 
2642 	return 0;
2643 }
2644 
2645 static void intel_pstate_adjust_policy_max(struct cpudata *cpu,
2646 					   struct cpufreq_policy_data *policy)
2647 {
2648 	if (!hwp_active &&
2649 	    cpu->pstate.max_pstate_physical > cpu->pstate.max_pstate &&
2650 	    policy->max < policy->cpuinfo.max_freq &&
2651 	    policy->max > cpu->pstate.max_freq) {
2652 		pr_debug("policy->max > max non turbo frequency\n");
2653 		policy->max = policy->cpuinfo.max_freq;
2654 	}
2655 }
2656 
2657 static void intel_pstate_verify_cpu_policy(struct cpudata *cpu,
2658 					   struct cpufreq_policy_data *policy)
2659 {
2660 	int max_freq;
2661 
2662 	update_turbo_state();
2663 	if (hwp_active) {
2664 		intel_pstate_get_hwp_cap(cpu);
2665 		max_freq = global.no_turbo || global.turbo_disabled ?
2666 				cpu->pstate.max_freq : cpu->pstate.turbo_freq;
2667 	} else {
2668 		max_freq = intel_pstate_get_max_freq(cpu);
2669 	}
2670 	cpufreq_verify_within_limits(policy, policy->cpuinfo.min_freq, max_freq);
2671 
2672 	intel_pstate_adjust_policy_max(cpu, policy);
2673 }
2674 
2675 static int intel_pstate_verify_policy(struct cpufreq_policy_data *policy)
2676 {
2677 	intel_pstate_verify_cpu_policy(all_cpu_data[policy->cpu], policy);
2678 
2679 	return 0;
2680 }
2681 
2682 static int intel_cpufreq_cpu_offline(struct cpufreq_policy *policy)
2683 {
2684 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2685 
2686 	pr_debug("CPU %d going offline\n", cpu->cpu);
2687 
2688 	if (cpu->suspended)
2689 		return 0;
2690 
2691 	/*
2692 	 * If the CPU is an SMT thread and it goes offline with the performance
2693 	 * settings different from the minimum, it will prevent its sibling
2694 	 * from getting to lower performance levels, so force the minimum
2695 	 * performance on CPU offline to prevent that from happening.
2696 	 */
2697 	if (hwp_active)
2698 		intel_pstate_hwp_offline(cpu);
2699 	else
2700 		intel_pstate_set_min_pstate(cpu);
2701 
2702 	intel_pstate_exit_perf_limits(policy);
2703 
2704 	return 0;
2705 }
2706 
2707 static int intel_pstate_cpu_online(struct cpufreq_policy *policy)
2708 {
2709 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2710 
2711 	pr_debug("CPU %d going online\n", cpu->cpu);
2712 
2713 	intel_pstate_init_acpi_perf_limits(policy);
2714 
2715 	if (hwp_active) {
2716 		/*
2717 		 * Re-enable HWP and clear the "suspended" flag to let "resume"
2718 		 * know that it need not do that.
2719 		 */
2720 		intel_pstate_hwp_reenable(cpu);
2721 		cpu->suspended = false;
2722 	}
2723 
2724 	return 0;
2725 }
2726 
2727 static int intel_pstate_cpu_offline(struct cpufreq_policy *policy)
2728 {
2729 	intel_pstate_clear_update_util_hook(policy->cpu);
2730 
2731 	return intel_cpufreq_cpu_offline(policy);
2732 }
2733 
2734 static int intel_pstate_cpu_exit(struct cpufreq_policy *policy)
2735 {
2736 	pr_debug("CPU %d exiting\n", policy->cpu);
2737 
2738 	policy->fast_switch_possible = false;
2739 
2740 	return 0;
2741 }
2742 
2743 static int __intel_pstate_cpu_init(struct cpufreq_policy *policy)
2744 {
2745 	struct cpudata *cpu;
2746 	int rc;
2747 
2748 	rc = intel_pstate_init_cpu(policy->cpu);
2749 	if (rc)
2750 		return rc;
2751 
2752 	cpu = all_cpu_data[policy->cpu];
2753 
2754 	cpu->max_perf_ratio = 0xFF;
2755 	cpu->min_perf_ratio = 0;
2756 
2757 	/* cpuinfo and default policy values */
2758 	policy->cpuinfo.min_freq = cpu->pstate.min_freq;
2759 	update_turbo_state();
2760 	global.turbo_disabled_mf = global.turbo_disabled;
2761 	policy->cpuinfo.max_freq = global.turbo_disabled ?
2762 			cpu->pstate.max_freq : cpu->pstate.turbo_freq;
2763 
2764 	policy->min = policy->cpuinfo.min_freq;
2765 	policy->max = policy->cpuinfo.max_freq;
2766 
2767 	intel_pstate_init_acpi_perf_limits(policy);
2768 
2769 	policy->fast_switch_possible = true;
2770 
2771 	return 0;
2772 }
2773 
2774 static int intel_pstate_cpu_init(struct cpufreq_policy *policy)
2775 {
2776 	int ret = __intel_pstate_cpu_init(policy);
2777 
2778 	if (ret)
2779 		return ret;
2780 
2781 	/*
2782 	 * Set the policy to powersave to provide a valid fallback value in case
2783 	 * the default cpufreq governor is neither powersave nor performance.
2784 	 */
2785 	policy->policy = CPUFREQ_POLICY_POWERSAVE;
2786 
2787 	if (hwp_active) {
2788 		struct cpudata *cpu = all_cpu_data[policy->cpu];
2789 
2790 		cpu->epp_cached = intel_pstate_get_epp(cpu, 0);
2791 	}
2792 
2793 	return 0;
2794 }
2795 
2796 static struct cpufreq_driver intel_pstate = {
2797 	.flags		= CPUFREQ_CONST_LOOPS,
2798 	.verify		= intel_pstate_verify_policy,
2799 	.setpolicy	= intel_pstate_set_policy,
2800 	.suspend	= intel_pstate_suspend,
2801 	.resume		= intel_pstate_resume,
2802 	.init		= intel_pstate_cpu_init,
2803 	.exit		= intel_pstate_cpu_exit,
2804 	.offline	= intel_pstate_cpu_offline,
2805 	.online		= intel_pstate_cpu_online,
2806 	.update_limits	= intel_pstate_update_limits,
2807 	.name		= "intel_pstate",
2808 };
2809 
2810 static int intel_cpufreq_verify_policy(struct cpufreq_policy_data *policy)
2811 {
2812 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2813 
2814 	intel_pstate_verify_cpu_policy(cpu, policy);
2815 	intel_pstate_update_perf_limits(cpu, policy->min, policy->max);
2816 
2817 	return 0;
2818 }
2819 
2820 /* Use of trace in passive mode:
2821  *
2822  * In passive mode the trace core_busy field (also known as the
2823  * performance field, and lablelled as such on the graphs; also known as
2824  * core_avg_perf) is not needed and so is re-assigned to indicate if the
2825  * driver call was via the normal or fast switch path. Various graphs
2826  * output from the intel_pstate_tracer.py utility that include core_busy
2827  * (or performance or core_avg_perf) have a fixed y-axis from 0 to 100%,
2828  * so we use 10 to indicate the normal path through the driver, and
2829  * 90 to indicate the fast switch path through the driver.
2830  * The scaled_busy field is not used, and is set to 0.
2831  */
2832 
2833 #define	INTEL_PSTATE_TRACE_TARGET 10
2834 #define	INTEL_PSTATE_TRACE_FAST_SWITCH 90
2835 
2836 static void intel_cpufreq_trace(struct cpudata *cpu, unsigned int trace_type, int old_pstate)
2837 {
2838 	struct sample *sample;
2839 
2840 	if (!trace_pstate_sample_enabled())
2841 		return;
2842 
2843 	if (!intel_pstate_sample(cpu, ktime_get()))
2844 		return;
2845 
2846 	sample = &cpu->sample;
2847 	trace_pstate_sample(trace_type,
2848 		0,
2849 		old_pstate,
2850 		cpu->pstate.current_pstate,
2851 		sample->mperf,
2852 		sample->aperf,
2853 		sample->tsc,
2854 		get_avg_frequency(cpu),
2855 		fp_toint(cpu->iowait_boost * 100));
2856 }
2857 
2858 static void intel_cpufreq_hwp_update(struct cpudata *cpu, u32 min, u32 max,
2859 				     u32 desired, bool fast_switch)
2860 {
2861 	u64 prev = READ_ONCE(cpu->hwp_req_cached), value = prev;
2862 
2863 	value &= ~HWP_MIN_PERF(~0L);
2864 	value |= HWP_MIN_PERF(min);
2865 
2866 	value &= ~HWP_MAX_PERF(~0L);
2867 	value |= HWP_MAX_PERF(max);
2868 
2869 	value &= ~HWP_DESIRED_PERF(~0L);
2870 	value |= HWP_DESIRED_PERF(desired);
2871 
2872 	if (value == prev)
2873 		return;
2874 
2875 	WRITE_ONCE(cpu->hwp_req_cached, value);
2876 	if (fast_switch)
2877 		wrmsrl(MSR_HWP_REQUEST, value);
2878 	else
2879 		wrmsrl_on_cpu(cpu->cpu, MSR_HWP_REQUEST, value);
2880 }
2881 
2882 static void intel_cpufreq_perf_ctl_update(struct cpudata *cpu,
2883 					  u32 target_pstate, bool fast_switch)
2884 {
2885 	if (fast_switch)
2886 		wrmsrl(MSR_IA32_PERF_CTL,
2887 		       pstate_funcs.get_val(cpu, target_pstate));
2888 	else
2889 		wrmsrl_on_cpu(cpu->cpu, MSR_IA32_PERF_CTL,
2890 			      pstate_funcs.get_val(cpu, target_pstate));
2891 }
2892 
2893 static int intel_cpufreq_update_pstate(struct cpufreq_policy *policy,
2894 				       int target_pstate, bool fast_switch)
2895 {
2896 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2897 	int old_pstate = cpu->pstate.current_pstate;
2898 
2899 	target_pstate = intel_pstate_prepare_request(cpu, target_pstate);
2900 	if (hwp_active) {
2901 		int max_pstate = policy->strict_target ?
2902 					target_pstate : cpu->max_perf_ratio;
2903 
2904 		intel_cpufreq_hwp_update(cpu, target_pstate, max_pstate, 0,
2905 					 fast_switch);
2906 	} else if (target_pstate != old_pstate) {
2907 		intel_cpufreq_perf_ctl_update(cpu, target_pstate, fast_switch);
2908 	}
2909 
2910 	cpu->pstate.current_pstate = target_pstate;
2911 
2912 	intel_cpufreq_trace(cpu, fast_switch ? INTEL_PSTATE_TRACE_FAST_SWITCH :
2913 			    INTEL_PSTATE_TRACE_TARGET, old_pstate);
2914 
2915 	return target_pstate;
2916 }
2917 
2918 static int intel_cpufreq_target(struct cpufreq_policy *policy,
2919 				unsigned int target_freq,
2920 				unsigned int relation)
2921 {
2922 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2923 	struct cpufreq_freqs freqs;
2924 	int target_pstate;
2925 
2926 	update_turbo_state();
2927 
2928 	freqs.old = policy->cur;
2929 	freqs.new = target_freq;
2930 
2931 	cpufreq_freq_transition_begin(policy, &freqs);
2932 
2933 	target_pstate = intel_pstate_freq_to_hwp_rel(cpu, freqs.new, relation);
2934 	target_pstate = intel_cpufreq_update_pstate(policy, target_pstate, false);
2935 
2936 	freqs.new = target_pstate * cpu->pstate.scaling;
2937 
2938 	cpufreq_freq_transition_end(policy, &freqs, false);
2939 
2940 	return 0;
2941 }
2942 
2943 static unsigned int intel_cpufreq_fast_switch(struct cpufreq_policy *policy,
2944 					      unsigned int target_freq)
2945 {
2946 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2947 	int target_pstate;
2948 
2949 	update_turbo_state();
2950 
2951 	target_pstate = intel_pstate_freq_to_hwp(cpu, target_freq);
2952 
2953 	target_pstate = intel_cpufreq_update_pstate(policy, target_pstate, true);
2954 
2955 	return target_pstate * cpu->pstate.scaling;
2956 }
2957 
2958 static void intel_cpufreq_adjust_perf(unsigned int cpunum,
2959 				      unsigned long min_perf,
2960 				      unsigned long target_perf,
2961 				      unsigned long capacity)
2962 {
2963 	struct cpudata *cpu = all_cpu_data[cpunum];
2964 	u64 hwp_cap = READ_ONCE(cpu->hwp_cap_cached);
2965 	int old_pstate = cpu->pstate.current_pstate;
2966 	int cap_pstate, min_pstate, max_pstate, target_pstate;
2967 
2968 	update_turbo_state();
2969 	cap_pstate = global.turbo_disabled ? HWP_GUARANTEED_PERF(hwp_cap) :
2970 					     HWP_HIGHEST_PERF(hwp_cap);
2971 
2972 	/* Optimization: Avoid unnecessary divisions. */
2973 
2974 	target_pstate = cap_pstate;
2975 	if (target_perf < capacity)
2976 		target_pstate = DIV_ROUND_UP(cap_pstate * target_perf, capacity);
2977 
2978 	min_pstate = cap_pstate;
2979 	if (min_perf < capacity)
2980 		min_pstate = DIV_ROUND_UP(cap_pstate * min_perf, capacity);
2981 
2982 	if (min_pstate < cpu->pstate.min_pstate)
2983 		min_pstate = cpu->pstate.min_pstate;
2984 
2985 	if (min_pstate < cpu->min_perf_ratio)
2986 		min_pstate = cpu->min_perf_ratio;
2987 
2988 	if (min_pstate > cpu->max_perf_ratio)
2989 		min_pstate = cpu->max_perf_ratio;
2990 
2991 	max_pstate = min(cap_pstate, cpu->max_perf_ratio);
2992 	if (max_pstate < min_pstate)
2993 		max_pstate = min_pstate;
2994 
2995 	target_pstate = clamp_t(int, target_pstate, min_pstate, max_pstate);
2996 
2997 	intel_cpufreq_hwp_update(cpu, min_pstate, max_pstate, target_pstate, true);
2998 
2999 	cpu->pstate.current_pstate = target_pstate;
3000 	intel_cpufreq_trace(cpu, INTEL_PSTATE_TRACE_FAST_SWITCH, old_pstate);
3001 }
3002 
3003 static int intel_cpufreq_cpu_init(struct cpufreq_policy *policy)
3004 {
3005 	struct freq_qos_request *req;
3006 	struct cpudata *cpu;
3007 	struct device *dev;
3008 	int ret, freq;
3009 
3010 	dev = get_cpu_device(policy->cpu);
3011 	if (!dev)
3012 		return -ENODEV;
3013 
3014 	ret = __intel_pstate_cpu_init(policy);
3015 	if (ret)
3016 		return ret;
3017 
3018 	policy->cpuinfo.transition_latency = INTEL_CPUFREQ_TRANSITION_LATENCY;
3019 	/* This reflects the intel_pstate_get_cpu_pstates() setting. */
3020 	policy->cur = policy->cpuinfo.min_freq;
3021 
3022 	req = kcalloc(2, sizeof(*req), GFP_KERNEL);
3023 	if (!req) {
3024 		ret = -ENOMEM;
3025 		goto pstate_exit;
3026 	}
3027 
3028 	cpu = all_cpu_data[policy->cpu];
3029 
3030 	if (hwp_active) {
3031 		u64 value;
3032 
3033 		policy->transition_delay_us = INTEL_CPUFREQ_TRANSITION_DELAY_HWP;
3034 
3035 		intel_pstate_get_hwp_cap(cpu);
3036 
3037 		rdmsrl_on_cpu(cpu->cpu, MSR_HWP_REQUEST, &value);
3038 		WRITE_ONCE(cpu->hwp_req_cached, value);
3039 
3040 		cpu->epp_cached = intel_pstate_get_epp(cpu, value);
3041 	} else {
3042 		policy->transition_delay_us = INTEL_CPUFREQ_TRANSITION_DELAY;
3043 	}
3044 
3045 	freq = DIV_ROUND_UP(cpu->pstate.turbo_freq * global.min_perf_pct, 100);
3046 
3047 	ret = freq_qos_add_request(&policy->constraints, req, FREQ_QOS_MIN,
3048 				   freq);
3049 	if (ret < 0) {
3050 		dev_err(dev, "Failed to add min-freq constraint (%d)\n", ret);
3051 		goto free_req;
3052 	}
3053 
3054 	freq = DIV_ROUND_UP(cpu->pstate.turbo_freq * global.max_perf_pct, 100);
3055 
3056 	ret = freq_qos_add_request(&policy->constraints, req + 1, FREQ_QOS_MAX,
3057 				   freq);
3058 	if (ret < 0) {
3059 		dev_err(dev, "Failed to add max-freq constraint (%d)\n", ret);
3060 		goto remove_min_req;
3061 	}
3062 
3063 	policy->driver_data = req;
3064 
3065 	return 0;
3066 
3067 remove_min_req:
3068 	freq_qos_remove_request(req);
3069 free_req:
3070 	kfree(req);
3071 pstate_exit:
3072 	intel_pstate_exit_perf_limits(policy);
3073 
3074 	return ret;
3075 }
3076 
3077 static int intel_cpufreq_cpu_exit(struct cpufreq_policy *policy)
3078 {
3079 	struct freq_qos_request *req;
3080 
3081 	req = policy->driver_data;
3082 
3083 	freq_qos_remove_request(req + 1);
3084 	freq_qos_remove_request(req);
3085 	kfree(req);
3086 
3087 	return intel_pstate_cpu_exit(policy);
3088 }
3089 
3090 static int intel_cpufreq_suspend(struct cpufreq_policy *policy)
3091 {
3092 	intel_pstate_suspend(policy);
3093 
3094 	if (hwp_active) {
3095 		struct cpudata *cpu = all_cpu_data[policy->cpu];
3096 		u64 value = READ_ONCE(cpu->hwp_req_cached);
3097 
3098 		/*
3099 		 * Clear the desired perf field in MSR_HWP_REQUEST in case
3100 		 * intel_cpufreq_adjust_perf() is in use and the last value
3101 		 * written by it may not be suitable.
3102 		 */
3103 		value &= ~HWP_DESIRED_PERF(~0L);
3104 		wrmsrl_on_cpu(cpu->cpu, MSR_HWP_REQUEST, value);
3105 		WRITE_ONCE(cpu->hwp_req_cached, value);
3106 	}
3107 
3108 	return 0;
3109 }
3110 
3111 static struct cpufreq_driver intel_cpufreq = {
3112 	.flags		= CPUFREQ_CONST_LOOPS,
3113 	.verify		= intel_cpufreq_verify_policy,
3114 	.target		= intel_cpufreq_target,
3115 	.fast_switch	= intel_cpufreq_fast_switch,
3116 	.init		= intel_cpufreq_cpu_init,
3117 	.exit		= intel_cpufreq_cpu_exit,
3118 	.offline	= intel_cpufreq_cpu_offline,
3119 	.online		= intel_pstate_cpu_online,
3120 	.suspend	= intel_cpufreq_suspend,
3121 	.resume		= intel_pstate_resume,
3122 	.update_limits	= intel_pstate_update_limits,
3123 	.name		= "intel_cpufreq",
3124 };
3125 
3126 static struct cpufreq_driver *default_driver;
3127 
3128 static void intel_pstate_driver_cleanup(void)
3129 {
3130 	unsigned int cpu;
3131 
3132 	cpus_read_lock();
3133 	for_each_online_cpu(cpu) {
3134 		if (all_cpu_data[cpu]) {
3135 			if (intel_pstate_driver == &intel_pstate)
3136 				intel_pstate_clear_update_util_hook(cpu);
3137 
3138 			spin_lock(&hwp_notify_lock);
3139 			kfree(all_cpu_data[cpu]);
3140 			WRITE_ONCE(all_cpu_data[cpu], NULL);
3141 			spin_unlock(&hwp_notify_lock);
3142 		}
3143 	}
3144 	cpus_read_unlock();
3145 
3146 	intel_pstate_driver = NULL;
3147 }
3148 
3149 static int intel_pstate_register_driver(struct cpufreq_driver *driver)
3150 {
3151 	int ret;
3152 
3153 	if (driver == &intel_pstate)
3154 		intel_pstate_sysfs_expose_hwp_dynamic_boost();
3155 
3156 	memset(&global, 0, sizeof(global));
3157 	global.max_perf_pct = 100;
3158 
3159 	intel_pstate_driver = driver;
3160 	ret = cpufreq_register_driver(intel_pstate_driver);
3161 	if (ret) {
3162 		intel_pstate_driver_cleanup();
3163 		return ret;
3164 	}
3165 
3166 	global.min_perf_pct = min_perf_pct_min();
3167 
3168 	return 0;
3169 }
3170 
3171 static ssize_t intel_pstate_show_status(char *buf)
3172 {
3173 	if (!intel_pstate_driver)
3174 		return sprintf(buf, "off\n");
3175 
3176 	return sprintf(buf, "%s\n", intel_pstate_driver == &intel_pstate ?
3177 					"active" : "passive");
3178 }
3179 
3180 static int intel_pstate_update_status(const char *buf, size_t size)
3181 {
3182 	if (size == 3 && !strncmp(buf, "off", size)) {
3183 		if (!intel_pstate_driver)
3184 			return -EINVAL;
3185 
3186 		if (hwp_active)
3187 			return -EBUSY;
3188 
3189 		cpufreq_unregister_driver(intel_pstate_driver);
3190 		intel_pstate_driver_cleanup();
3191 		return 0;
3192 	}
3193 
3194 	if (size == 6 && !strncmp(buf, "active", size)) {
3195 		if (intel_pstate_driver) {
3196 			if (intel_pstate_driver == &intel_pstate)
3197 				return 0;
3198 
3199 			cpufreq_unregister_driver(intel_pstate_driver);
3200 		}
3201 
3202 		return intel_pstate_register_driver(&intel_pstate);
3203 	}
3204 
3205 	if (size == 7 && !strncmp(buf, "passive", size)) {
3206 		if (intel_pstate_driver) {
3207 			if (intel_pstate_driver == &intel_cpufreq)
3208 				return 0;
3209 
3210 			cpufreq_unregister_driver(intel_pstate_driver);
3211 			intel_pstate_sysfs_hide_hwp_dynamic_boost();
3212 		}
3213 
3214 		return intel_pstate_register_driver(&intel_cpufreq);
3215 	}
3216 
3217 	return -EINVAL;
3218 }
3219 
3220 static int no_load __initdata;
3221 static int no_hwp __initdata;
3222 static int hwp_only __initdata;
3223 static unsigned int force_load __initdata;
3224 
3225 static int __init intel_pstate_msrs_not_valid(void)
3226 {
3227 	if (!pstate_funcs.get_max(0) ||
3228 	    !pstate_funcs.get_min(0) ||
3229 	    !pstate_funcs.get_turbo(0))
3230 		return -ENODEV;
3231 
3232 	return 0;
3233 }
3234 
3235 static void __init copy_cpu_funcs(struct pstate_funcs *funcs)
3236 {
3237 	pstate_funcs.get_max   = funcs->get_max;
3238 	pstate_funcs.get_max_physical = funcs->get_max_physical;
3239 	pstate_funcs.get_min   = funcs->get_min;
3240 	pstate_funcs.get_turbo = funcs->get_turbo;
3241 	pstate_funcs.get_scaling = funcs->get_scaling;
3242 	pstate_funcs.get_val   = funcs->get_val;
3243 	pstate_funcs.get_vid   = funcs->get_vid;
3244 	pstate_funcs.get_aperf_mperf_shift = funcs->get_aperf_mperf_shift;
3245 }
3246 
3247 #ifdef CONFIG_ACPI
3248 
3249 static bool __init intel_pstate_no_acpi_pss(void)
3250 {
3251 	int i;
3252 
3253 	for_each_possible_cpu(i) {
3254 		acpi_status status;
3255 		union acpi_object *pss;
3256 		struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
3257 		struct acpi_processor *pr = per_cpu(processors, i);
3258 
3259 		if (!pr)
3260 			continue;
3261 
3262 		status = acpi_evaluate_object(pr->handle, "_PSS", NULL, &buffer);
3263 		if (ACPI_FAILURE(status))
3264 			continue;
3265 
3266 		pss = buffer.pointer;
3267 		if (pss && pss->type == ACPI_TYPE_PACKAGE) {
3268 			kfree(pss);
3269 			return false;
3270 		}
3271 
3272 		kfree(pss);
3273 	}
3274 
3275 	pr_debug("ACPI _PSS not found\n");
3276 	return true;
3277 }
3278 
3279 static bool __init intel_pstate_no_acpi_pcch(void)
3280 {
3281 	acpi_status status;
3282 	acpi_handle handle;
3283 
3284 	status = acpi_get_handle(NULL, "\\_SB", &handle);
3285 	if (ACPI_FAILURE(status))
3286 		goto not_found;
3287 
3288 	if (acpi_has_method(handle, "PCCH"))
3289 		return false;
3290 
3291 not_found:
3292 	pr_debug("ACPI PCCH not found\n");
3293 	return true;
3294 }
3295 
3296 static bool __init intel_pstate_has_acpi_ppc(void)
3297 {
3298 	int i;
3299 
3300 	for_each_possible_cpu(i) {
3301 		struct acpi_processor *pr = per_cpu(processors, i);
3302 
3303 		if (!pr)
3304 			continue;
3305 		if (acpi_has_method(pr->handle, "_PPC"))
3306 			return true;
3307 	}
3308 	pr_debug("ACPI _PPC not found\n");
3309 	return false;
3310 }
3311 
3312 enum {
3313 	PSS,
3314 	PPC,
3315 };
3316 
3317 /* Hardware vendor-specific info that has its own power management modes */
3318 static struct acpi_platform_list plat_info[] __initdata = {
3319 	{"HP    ", "ProLiant", 0, ACPI_SIG_FADT, all_versions, NULL, PSS},
3320 	{"ORACLE", "X4-2    ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3321 	{"ORACLE", "X4-2L   ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3322 	{"ORACLE", "X4-2B   ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3323 	{"ORACLE", "X3-2    ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3324 	{"ORACLE", "X3-2L   ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3325 	{"ORACLE", "X3-2B   ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3326 	{"ORACLE", "X4470M2 ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3327 	{"ORACLE", "X4270M3 ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3328 	{"ORACLE", "X4270M2 ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3329 	{"ORACLE", "X4170M2 ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3330 	{"ORACLE", "X4170 M3", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3331 	{"ORACLE", "X4275 M3", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3332 	{"ORACLE", "X6-2    ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3333 	{"ORACLE", "Sudbury ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3334 	{ } /* End */
3335 };
3336 
3337 #define BITMASK_OOB	(BIT(8) | BIT(18))
3338 
3339 static bool __init intel_pstate_platform_pwr_mgmt_exists(void)
3340 {
3341 	const struct x86_cpu_id *id;
3342 	u64 misc_pwr;
3343 	int idx;
3344 
3345 	id = x86_match_cpu(intel_pstate_cpu_oob_ids);
3346 	if (id) {
3347 		rdmsrl(MSR_MISC_PWR_MGMT, misc_pwr);
3348 		if (misc_pwr & BITMASK_OOB) {
3349 			pr_debug("Bit 8 or 18 in the MISC_PWR_MGMT MSR set\n");
3350 			pr_debug("P states are controlled in Out of Band mode by the firmware/hardware\n");
3351 			return true;
3352 		}
3353 	}
3354 
3355 	idx = acpi_match_platform_list(plat_info);
3356 	if (idx < 0)
3357 		return false;
3358 
3359 	switch (plat_info[idx].data) {
3360 	case PSS:
3361 		if (!intel_pstate_no_acpi_pss())
3362 			return false;
3363 
3364 		return intel_pstate_no_acpi_pcch();
3365 	case PPC:
3366 		return intel_pstate_has_acpi_ppc() && !force_load;
3367 	}
3368 
3369 	return false;
3370 }
3371 
3372 static void intel_pstate_request_control_from_smm(void)
3373 {
3374 	/*
3375 	 * It may be unsafe to request P-states control from SMM if _PPC support
3376 	 * has not been enabled.
3377 	 */
3378 	if (acpi_ppc)
3379 		acpi_processor_pstate_control();
3380 }
3381 #else /* CONFIG_ACPI not enabled */
3382 static inline bool intel_pstate_platform_pwr_mgmt_exists(void) { return false; }
3383 static inline bool intel_pstate_has_acpi_ppc(void) { return false; }
3384 static inline void intel_pstate_request_control_from_smm(void) {}
3385 #endif /* CONFIG_ACPI */
3386 
3387 #define INTEL_PSTATE_HWP_BROADWELL	0x01
3388 
3389 #define X86_MATCH_HWP(model, hwp_mode)					\
3390 	X86_MATCH_VENDOR_FAM_MODEL_FEATURE(INTEL, 6, INTEL_FAM6_##model, \
3391 					   X86_FEATURE_HWP, hwp_mode)
3392 
3393 static const struct x86_cpu_id hwp_support_ids[] __initconst = {
3394 	X86_MATCH_HWP(BROADWELL_X,	INTEL_PSTATE_HWP_BROADWELL),
3395 	X86_MATCH_HWP(BROADWELL_D,	INTEL_PSTATE_HWP_BROADWELL),
3396 	X86_MATCH_HWP(ANY,		0),
3397 	{}
3398 };
3399 
3400 static bool intel_pstate_hwp_is_enabled(void)
3401 {
3402 	u64 value;
3403 
3404 	rdmsrl(MSR_PM_ENABLE, value);
3405 	return !!(value & 0x1);
3406 }
3407 
3408 #define POWERSAVE_MASK			GENMASK(7, 0)
3409 #define BALANCE_POWER_MASK		GENMASK(15, 8)
3410 #define BALANCE_PERFORMANCE_MASK	GENMASK(23, 16)
3411 #define PERFORMANCE_MASK		GENMASK(31, 24)
3412 
3413 #define HWP_SET_EPP_VALUES(powersave, balance_power, balance_perf, performance) \
3414 	(FIELD_PREP_CONST(POWERSAVE_MASK, powersave) |\
3415 	 FIELD_PREP_CONST(BALANCE_POWER_MASK, balance_power) |\
3416 	 FIELD_PREP_CONST(BALANCE_PERFORMANCE_MASK, balance_perf) |\
3417 	 FIELD_PREP_CONST(PERFORMANCE_MASK, performance))
3418 
3419 #define HWP_SET_DEF_BALANCE_PERF_EPP(balance_perf) \
3420 	(HWP_SET_EPP_VALUES(HWP_EPP_POWERSAVE, HWP_EPP_BALANCE_POWERSAVE,\
3421 	 balance_perf, HWP_EPP_PERFORMANCE))
3422 
3423 static const struct x86_cpu_id intel_epp_default[] = {
3424 	/*
3425 	 * Set EPP value as 102, this is the max suggested EPP
3426 	 * which can result in one core turbo frequency for
3427 	 * AlderLake Mobile CPUs.
3428 	 */
3429 	X86_MATCH_INTEL_FAM6_MODEL(ALDERLAKE_L, HWP_SET_DEF_BALANCE_PERF_EPP(102)),
3430 	X86_MATCH_INTEL_FAM6_MODEL(SAPPHIRERAPIDS_X, HWP_SET_DEF_BALANCE_PERF_EPP(32)),
3431 	X86_MATCH_INTEL_FAM6_MODEL(METEORLAKE_L, HWP_SET_EPP_VALUES(HWP_EPP_POWERSAVE,
3432 							HWP_EPP_BALANCE_POWERSAVE, 115, 16)),
3433 	{}
3434 };
3435 
3436 static const struct x86_cpu_id intel_hybrid_scaling_factor[] = {
3437 	X86_MATCH_INTEL_FAM6_MODEL(METEORLAKE_L, HYBRID_SCALING_FACTOR_MTL),
3438 	{}
3439 };
3440 
3441 static int __init intel_pstate_init(void)
3442 {
3443 	static struct cpudata **_all_cpu_data;
3444 	const struct x86_cpu_id *id;
3445 	int rc;
3446 
3447 	if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL)
3448 		return -ENODEV;
3449 
3450 	id = x86_match_cpu(hwp_support_ids);
3451 	if (id) {
3452 		hwp_forced = intel_pstate_hwp_is_enabled();
3453 
3454 		if (hwp_forced)
3455 			pr_info("HWP enabled by BIOS\n");
3456 		else if (no_load)
3457 			return -ENODEV;
3458 
3459 		copy_cpu_funcs(&core_funcs);
3460 		/*
3461 		 * Avoid enabling HWP for processors without EPP support,
3462 		 * because that means incomplete HWP implementation which is a
3463 		 * corner case and supporting it is generally problematic.
3464 		 *
3465 		 * If HWP is enabled already, though, there is no choice but to
3466 		 * deal with it.
3467 		 */
3468 		if ((!no_hwp && boot_cpu_has(X86_FEATURE_HWP_EPP)) || hwp_forced) {
3469 			WRITE_ONCE(hwp_active, 1);
3470 			hwp_mode_bdw = id->driver_data;
3471 			intel_pstate.attr = hwp_cpufreq_attrs;
3472 			intel_cpufreq.attr = hwp_cpufreq_attrs;
3473 			intel_cpufreq.flags |= CPUFREQ_NEED_UPDATE_LIMITS;
3474 			intel_cpufreq.adjust_perf = intel_cpufreq_adjust_perf;
3475 			if (!default_driver)
3476 				default_driver = &intel_pstate;
3477 
3478 			pstate_funcs.get_cpu_scaling = hwp_get_cpu_scaling;
3479 
3480 			goto hwp_cpu_matched;
3481 		}
3482 		pr_info("HWP not enabled\n");
3483 	} else {
3484 		if (no_load)
3485 			return -ENODEV;
3486 
3487 		id = x86_match_cpu(intel_pstate_cpu_ids);
3488 		if (!id) {
3489 			pr_info("CPU model not supported\n");
3490 			return -ENODEV;
3491 		}
3492 
3493 		copy_cpu_funcs((struct pstate_funcs *)id->driver_data);
3494 	}
3495 
3496 	if (intel_pstate_msrs_not_valid()) {
3497 		pr_info("Invalid MSRs\n");
3498 		return -ENODEV;
3499 	}
3500 	/* Without HWP start in the passive mode. */
3501 	if (!default_driver)
3502 		default_driver = &intel_cpufreq;
3503 
3504 hwp_cpu_matched:
3505 	/*
3506 	 * The Intel pstate driver will be ignored if the platform
3507 	 * firmware has its own power management modes.
3508 	 */
3509 	if (intel_pstate_platform_pwr_mgmt_exists()) {
3510 		pr_info("P-states controlled by the platform\n");
3511 		return -ENODEV;
3512 	}
3513 
3514 	if (!hwp_active && hwp_only)
3515 		return -ENOTSUPP;
3516 
3517 	pr_info("Intel P-state driver initializing\n");
3518 
3519 	_all_cpu_data = vzalloc(array_size(sizeof(void *), num_possible_cpus()));
3520 	if (!_all_cpu_data)
3521 		return -ENOMEM;
3522 
3523 	WRITE_ONCE(all_cpu_data, _all_cpu_data);
3524 
3525 	intel_pstate_request_control_from_smm();
3526 
3527 	intel_pstate_sysfs_expose_params();
3528 
3529 	if (hwp_active) {
3530 		const struct x86_cpu_id *id = x86_match_cpu(intel_epp_default);
3531 		const struct x86_cpu_id *hybrid_id = x86_match_cpu(intel_hybrid_scaling_factor);
3532 
3533 		if (id) {
3534 			epp_values[EPP_INDEX_POWERSAVE] =
3535 					FIELD_GET(POWERSAVE_MASK, id->driver_data);
3536 			epp_values[EPP_INDEX_BALANCE_POWERSAVE] =
3537 					FIELD_GET(BALANCE_POWER_MASK, id->driver_data);
3538 			epp_values[EPP_INDEX_BALANCE_PERFORMANCE] =
3539 					FIELD_GET(BALANCE_PERFORMANCE_MASK, id->driver_data);
3540 			epp_values[EPP_INDEX_PERFORMANCE] =
3541 					FIELD_GET(PERFORMANCE_MASK, id->driver_data);
3542 			pr_debug("Updated EPPs powersave:%x balanced power:%x balanced perf:%x performance:%x\n",
3543 				 epp_values[EPP_INDEX_POWERSAVE],
3544 				 epp_values[EPP_INDEX_BALANCE_POWERSAVE],
3545 				 epp_values[EPP_INDEX_BALANCE_PERFORMANCE],
3546 				 epp_values[EPP_INDEX_PERFORMANCE]);
3547 		}
3548 
3549 		if (hybrid_id) {
3550 			hybrid_scaling_factor = hybrid_id->driver_data;
3551 			pr_debug("hybrid scaling factor: %d\n", hybrid_scaling_factor);
3552 		}
3553 
3554 	}
3555 
3556 	mutex_lock(&intel_pstate_driver_lock);
3557 	rc = intel_pstate_register_driver(default_driver);
3558 	mutex_unlock(&intel_pstate_driver_lock);
3559 	if (rc) {
3560 		intel_pstate_sysfs_remove();
3561 		return rc;
3562 	}
3563 
3564 	if (hwp_active) {
3565 		const struct x86_cpu_id *id;
3566 
3567 		id = x86_match_cpu(intel_pstate_cpu_ee_disable_ids);
3568 		if (id) {
3569 			set_power_ctl_ee_state(false);
3570 			pr_info("Disabling energy efficiency optimization\n");
3571 		}
3572 
3573 		pr_info("HWP enabled\n");
3574 	} else if (boot_cpu_has(X86_FEATURE_HYBRID_CPU)) {
3575 		pr_warn("Problematic setup: Hybrid processor with disabled HWP\n");
3576 	}
3577 
3578 	return 0;
3579 }
3580 device_initcall(intel_pstate_init);
3581 
3582 static int __init intel_pstate_setup(char *str)
3583 {
3584 	if (!str)
3585 		return -EINVAL;
3586 
3587 	if (!strcmp(str, "disable"))
3588 		no_load = 1;
3589 	else if (!strcmp(str, "active"))
3590 		default_driver = &intel_pstate;
3591 	else if (!strcmp(str, "passive"))
3592 		default_driver = &intel_cpufreq;
3593 
3594 	if (!strcmp(str, "no_hwp"))
3595 		no_hwp = 1;
3596 
3597 	if (!strcmp(str, "force"))
3598 		force_load = 1;
3599 	if (!strcmp(str, "hwp_only"))
3600 		hwp_only = 1;
3601 	if (!strcmp(str, "per_cpu_perf_limits"))
3602 		per_cpu_limits = true;
3603 
3604 #ifdef CONFIG_ACPI
3605 	if (!strcmp(str, "support_acpi_ppc"))
3606 		acpi_ppc = true;
3607 #endif
3608 
3609 	return 0;
3610 }
3611 early_param("intel_pstate", intel_pstate_setup);
3612 
3613 MODULE_AUTHOR("Dirk Brandewie <dirk.j.brandewie@intel.com>");
3614 MODULE_DESCRIPTION("'intel_pstate' - P state driver Intel Core processors");
3615