xref: /linux/drivers/cpufreq/intel_pstate.c (revision 906fd46a65383cd639e5eec72a047efc33045d86)
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  * @min_perf_pct:	Minimum capacity limit in percent of the maximum turbo
177  *			P-state capacity.
178  * @max_perf_pct:	Maximum capacity limit in percent of the maximum turbo
179  *			P-state capacity.
180  */
181 struct global_params {
182 	bool no_turbo;
183 	bool turbo_disabled;
184 	int max_perf_pct;
185 	int min_perf_pct;
186 };
187 
188 /**
189  * struct cpudata -	Per CPU instance data storage
190  * @cpu:		CPU number for this instance data
191  * @policy:		CPUFreq policy value
192  * @update_util:	CPUFreq utility callback information
193  * @update_util_set:	CPUFreq utility callback is set
194  * @iowait_boost:	iowait-related boost fraction
195  * @last_update:	Time of the last update.
196  * @pstate:		Stores P state limits for this CPU
197  * @vid:		Stores VID limits for this CPU
198  * @last_sample_time:	Last Sample time
199  * @aperf_mperf_shift:	APERF vs MPERF counting frequency difference
200  * @prev_aperf:		Last APERF value read from APERF MSR
201  * @prev_mperf:		Last MPERF value read from MPERF MSR
202  * @prev_tsc:		Last timestamp counter (TSC) value
203  * @sample:		Storage for storing last Sample data
204  * @min_perf_ratio:	Minimum capacity in terms of PERF or HWP ratios
205  * @max_perf_ratio:	Maximum capacity in terms of PERF or HWP ratios
206  * @acpi_perf_data:	Stores ACPI perf information read from _PSS
207  * @valid_pss_table:	Set to true for valid ACPI _PSS entries found
208  * @epp_powersave:	Last saved HWP energy performance preference
209  *			(EPP) or energy performance bias (EPB),
210  *			when policy switched to performance
211  * @epp_policy:		Last saved policy used to set EPP/EPB
212  * @epp_default:	Power on default HWP energy performance
213  *			preference/bias
214  * @epp_cached:		Cached HWP energy-performance preference value
215  * @hwp_req_cached:	Cached value of the last HWP Request MSR
216  * @hwp_cap_cached:	Cached value of the last HWP Capabilities MSR
217  * @last_io_update:	Last time when IO wake flag was set
218  * @sched_flags:	Store scheduler flags for possible cross CPU update
219  * @hwp_boost_min:	Last HWP boosted min performance
220  * @suspended:		Whether or not the driver has been suspended.
221  * @hwp_notify_work:	workqueue for HWP notifications.
222  *
223  * This structure stores per CPU instance data for all CPUs.
224  */
225 struct cpudata {
226 	int cpu;
227 
228 	unsigned int policy;
229 	struct update_util_data update_util;
230 	bool   update_util_set;
231 
232 	struct pstate_data pstate;
233 	struct vid_data vid;
234 
235 	u64	last_update;
236 	u64	last_sample_time;
237 	u64	aperf_mperf_shift;
238 	u64	prev_aperf;
239 	u64	prev_mperf;
240 	u64	prev_tsc;
241 	struct sample sample;
242 	int32_t	min_perf_ratio;
243 	int32_t	max_perf_ratio;
244 #ifdef CONFIG_ACPI
245 	struct acpi_processor_performance acpi_perf_data;
246 	bool valid_pss_table;
247 #endif
248 	unsigned int iowait_boost;
249 	s16 epp_powersave;
250 	s16 epp_policy;
251 	s16 epp_default;
252 	s16 epp_cached;
253 	u64 hwp_req_cached;
254 	u64 hwp_cap_cached;
255 	u64 last_io_update;
256 	unsigned int sched_flags;
257 	u32 hwp_boost_min;
258 	bool suspended;
259 	struct delayed_work hwp_notify_work;
260 };
261 
262 static struct cpudata **all_cpu_data;
263 
264 /**
265  * struct pstate_funcs - Per CPU model specific callbacks
266  * @get_max:		Callback to get maximum non turbo effective P state
267  * @get_max_physical:	Callback to get maximum non turbo physical P state
268  * @get_min:		Callback to get minimum P state
269  * @get_turbo:		Callback to get turbo P state
270  * @get_scaling:	Callback to get frequency scaling factor
271  * @get_cpu_scaling:	Get frequency scaling factor for a given cpu
272  * @get_aperf_mperf_shift: Callback to get the APERF vs MPERF frequency difference
273  * @get_val:		Callback to convert P state to actual MSR write value
274  * @get_vid:		Callback to get VID data for Atom platforms
275  *
276  * Core and Atom CPU models have different way to get P State limits. This
277  * structure is used to store those callbacks.
278  */
279 struct pstate_funcs {
280 	int (*get_max)(int cpu);
281 	int (*get_max_physical)(int cpu);
282 	int (*get_min)(int cpu);
283 	int (*get_turbo)(int cpu);
284 	int (*get_scaling)(void);
285 	int (*get_cpu_scaling)(int cpu);
286 	int (*get_aperf_mperf_shift)(void);
287 	u64 (*get_val)(struct cpudata*, int pstate);
288 	void (*get_vid)(struct cpudata *);
289 };
290 
291 static struct pstate_funcs pstate_funcs __read_mostly;
292 
293 static bool hwp_active __ro_after_init;
294 static int hwp_mode_bdw __ro_after_init;
295 static bool per_cpu_limits __ro_after_init;
296 static bool hwp_forced __ro_after_init;
297 static bool hwp_boost __read_mostly;
298 
299 static struct cpufreq_driver *intel_pstate_driver __read_mostly;
300 
301 #define HYBRID_SCALING_FACTOR		78741
302 #define HYBRID_SCALING_FACTOR_MTL	80000
303 
304 static int hybrid_scaling_factor = HYBRID_SCALING_FACTOR;
305 
306 static inline int core_get_scaling(void)
307 {
308 	return 100000;
309 }
310 
311 #ifdef CONFIG_ACPI
312 static bool acpi_ppc;
313 #endif
314 
315 static struct global_params global;
316 
317 static DEFINE_MUTEX(intel_pstate_driver_lock);
318 static DEFINE_MUTEX(intel_pstate_limits_lock);
319 
320 #ifdef CONFIG_ACPI
321 
322 static bool intel_pstate_acpi_pm_profile_server(void)
323 {
324 	if (acpi_gbl_FADT.preferred_profile == PM_ENTERPRISE_SERVER ||
325 	    acpi_gbl_FADT.preferred_profile == PM_PERFORMANCE_SERVER)
326 		return true;
327 
328 	return false;
329 }
330 
331 static bool intel_pstate_get_ppc_enable_status(void)
332 {
333 	if (intel_pstate_acpi_pm_profile_server())
334 		return true;
335 
336 	return acpi_ppc;
337 }
338 
339 #ifdef CONFIG_ACPI_CPPC_LIB
340 
341 /* The work item is needed to avoid CPU hotplug locking issues */
342 static void intel_pstste_sched_itmt_work_fn(struct work_struct *work)
343 {
344 	sched_set_itmt_support();
345 }
346 
347 static DECLARE_WORK(sched_itmt_work, intel_pstste_sched_itmt_work_fn);
348 
349 #define CPPC_MAX_PERF	U8_MAX
350 
351 static void intel_pstate_set_itmt_prio(int cpu)
352 {
353 	struct cppc_perf_caps cppc_perf;
354 	static u32 max_highest_perf = 0, min_highest_perf = U32_MAX;
355 	int ret;
356 
357 	ret = cppc_get_perf_caps(cpu, &cppc_perf);
358 	/*
359 	 * If CPPC is not available, fall back to MSR_HWP_CAPABILITIES bits [8:0].
360 	 *
361 	 * Also, on some systems with overclocking enabled, CPPC.highest_perf is
362 	 * hardcoded to 0xff, so CPPC.highest_perf cannot be used to enable ITMT.
363 	 * Fall back to MSR_HWP_CAPABILITIES then too.
364 	 */
365 	if (ret || cppc_perf.highest_perf == CPPC_MAX_PERF)
366 		cppc_perf.highest_perf = HWP_HIGHEST_PERF(READ_ONCE(all_cpu_data[cpu]->hwp_cap_cached));
367 
368 	/*
369 	 * The priorities can be set regardless of whether or not
370 	 * sched_set_itmt_support(true) has been called and it is valid to
371 	 * update them at any time after it has been called.
372 	 */
373 	sched_set_itmt_core_prio(cppc_perf.highest_perf, cpu);
374 
375 	if (max_highest_perf <= min_highest_perf) {
376 		if (cppc_perf.highest_perf > max_highest_perf)
377 			max_highest_perf = cppc_perf.highest_perf;
378 
379 		if (cppc_perf.highest_perf < min_highest_perf)
380 			min_highest_perf = cppc_perf.highest_perf;
381 
382 		if (max_highest_perf > min_highest_perf) {
383 			/*
384 			 * This code can be run during CPU online under the
385 			 * CPU hotplug locks, so sched_set_itmt_support()
386 			 * cannot be called from here.  Queue up a work item
387 			 * to invoke it.
388 			 */
389 			schedule_work(&sched_itmt_work);
390 		}
391 	}
392 }
393 
394 static int intel_pstate_get_cppc_guaranteed(int cpu)
395 {
396 	struct cppc_perf_caps cppc_perf;
397 	int ret;
398 
399 	ret = cppc_get_perf_caps(cpu, &cppc_perf);
400 	if (ret)
401 		return ret;
402 
403 	if (cppc_perf.guaranteed_perf)
404 		return cppc_perf.guaranteed_perf;
405 
406 	return cppc_perf.nominal_perf;
407 }
408 
409 static int intel_pstate_cppc_get_scaling(int cpu)
410 {
411 	struct cppc_perf_caps cppc_perf;
412 	int ret;
413 
414 	ret = cppc_get_perf_caps(cpu, &cppc_perf);
415 
416 	/*
417 	 * If the nominal frequency and the nominal performance are not
418 	 * zero and the ratio between them is not 100, return the hybrid
419 	 * scaling factor.
420 	 */
421 	if (!ret && cppc_perf.nominal_perf && cppc_perf.nominal_freq &&
422 	    cppc_perf.nominal_perf * 100 != cppc_perf.nominal_freq)
423 		return hybrid_scaling_factor;
424 
425 	return core_get_scaling();
426 }
427 
428 #else /* CONFIG_ACPI_CPPC_LIB */
429 static inline void intel_pstate_set_itmt_prio(int cpu)
430 {
431 }
432 #endif /* CONFIG_ACPI_CPPC_LIB */
433 
434 static void intel_pstate_init_acpi_perf_limits(struct cpufreq_policy *policy)
435 {
436 	struct cpudata *cpu;
437 	int ret;
438 	int i;
439 
440 	if (hwp_active) {
441 		intel_pstate_set_itmt_prio(policy->cpu);
442 		return;
443 	}
444 
445 	if (!intel_pstate_get_ppc_enable_status())
446 		return;
447 
448 	cpu = all_cpu_data[policy->cpu];
449 
450 	ret = acpi_processor_register_performance(&cpu->acpi_perf_data,
451 						  policy->cpu);
452 	if (ret)
453 		return;
454 
455 	/*
456 	 * Check if the control value in _PSS is for PERF_CTL MSR, which should
457 	 * guarantee that the states returned by it map to the states in our
458 	 * list directly.
459 	 */
460 	if (cpu->acpi_perf_data.control_register.space_id !=
461 						ACPI_ADR_SPACE_FIXED_HARDWARE)
462 		goto err;
463 
464 	/*
465 	 * If there is only one entry _PSS, simply ignore _PSS and continue as
466 	 * usual without taking _PSS into account
467 	 */
468 	if (cpu->acpi_perf_data.state_count < 2)
469 		goto err;
470 
471 	pr_debug("CPU%u - ACPI _PSS perf data\n", policy->cpu);
472 	for (i = 0; i < cpu->acpi_perf_data.state_count; i++) {
473 		pr_debug("     %cP%d: %u MHz, %u mW, 0x%x\n",
474 			 (i == cpu->acpi_perf_data.state ? '*' : ' '), i,
475 			 (u32) cpu->acpi_perf_data.states[i].core_frequency,
476 			 (u32) cpu->acpi_perf_data.states[i].power,
477 			 (u32) cpu->acpi_perf_data.states[i].control);
478 	}
479 
480 	cpu->valid_pss_table = true;
481 	pr_debug("_PPC limits will be enforced\n");
482 
483 	return;
484 
485  err:
486 	cpu->valid_pss_table = false;
487 	acpi_processor_unregister_performance(policy->cpu);
488 }
489 
490 static void intel_pstate_exit_perf_limits(struct cpufreq_policy *policy)
491 {
492 	struct cpudata *cpu;
493 
494 	cpu = all_cpu_data[policy->cpu];
495 	if (!cpu->valid_pss_table)
496 		return;
497 
498 	acpi_processor_unregister_performance(policy->cpu);
499 }
500 #else /* CONFIG_ACPI */
501 static inline void intel_pstate_init_acpi_perf_limits(struct cpufreq_policy *policy)
502 {
503 }
504 
505 static inline void intel_pstate_exit_perf_limits(struct cpufreq_policy *policy)
506 {
507 }
508 
509 static inline bool intel_pstate_acpi_pm_profile_server(void)
510 {
511 	return false;
512 }
513 #endif /* CONFIG_ACPI */
514 
515 #ifndef CONFIG_ACPI_CPPC_LIB
516 static inline int intel_pstate_get_cppc_guaranteed(int cpu)
517 {
518 	return -ENOTSUPP;
519 }
520 
521 static int intel_pstate_cppc_get_scaling(int cpu)
522 {
523 	return core_get_scaling();
524 }
525 #endif /* CONFIG_ACPI_CPPC_LIB */
526 
527 static int intel_pstate_freq_to_hwp_rel(struct cpudata *cpu, int freq,
528 					unsigned int relation)
529 {
530 	if (freq == cpu->pstate.turbo_freq)
531 		return cpu->pstate.turbo_pstate;
532 
533 	if (freq == cpu->pstate.max_freq)
534 		return cpu->pstate.max_pstate;
535 
536 	switch (relation) {
537 	case CPUFREQ_RELATION_H:
538 		return freq / cpu->pstate.scaling;
539 	case CPUFREQ_RELATION_C:
540 		return DIV_ROUND_CLOSEST(freq, cpu->pstate.scaling);
541 	}
542 
543 	return DIV_ROUND_UP(freq, cpu->pstate.scaling);
544 }
545 
546 static int intel_pstate_freq_to_hwp(struct cpudata *cpu, int freq)
547 {
548 	return intel_pstate_freq_to_hwp_rel(cpu, freq, CPUFREQ_RELATION_L);
549 }
550 
551 /**
552  * intel_pstate_hybrid_hwp_adjust - Calibrate HWP performance levels.
553  * @cpu: Target CPU.
554  *
555  * On hybrid processors, HWP may expose more performance levels than there are
556  * P-states accessible through the PERF_CTL interface.  If that happens, the
557  * scaling factor between HWP performance levels and CPU frequency will be less
558  * than the scaling factor between P-state values and CPU frequency.
559  *
560  * In that case, adjust the CPU parameters used in computations accordingly.
561  */
562 static void intel_pstate_hybrid_hwp_adjust(struct cpudata *cpu)
563 {
564 	int perf_ctl_max_phys = cpu->pstate.max_pstate_physical;
565 	int perf_ctl_scaling = cpu->pstate.perf_ctl_scaling;
566 	int perf_ctl_turbo = pstate_funcs.get_turbo(cpu->cpu);
567 	int scaling = cpu->pstate.scaling;
568 	int freq;
569 
570 	pr_debug("CPU%d: perf_ctl_max_phys = %d\n", cpu->cpu, perf_ctl_max_phys);
571 	pr_debug("CPU%d: perf_ctl_turbo = %d\n", cpu->cpu, perf_ctl_turbo);
572 	pr_debug("CPU%d: perf_ctl_scaling = %d\n", cpu->cpu, perf_ctl_scaling);
573 	pr_debug("CPU%d: HWP_CAP guaranteed = %d\n", cpu->cpu, cpu->pstate.max_pstate);
574 	pr_debug("CPU%d: HWP_CAP highest = %d\n", cpu->cpu, cpu->pstate.turbo_pstate);
575 	pr_debug("CPU%d: HWP-to-frequency scaling factor: %d\n", cpu->cpu, scaling);
576 
577 	cpu->pstate.turbo_freq = rounddown(cpu->pstate.turbo_pstate * scaling,
578 					   perf_ctl_scaling);
579 	cpu->pstate.max_freq = rounddown(cpu->pstate.max_pstate * scaling,
580 					 perf_ctl_scaling);
581 
582 	freq = perf_ctl_max_phys * perf_ctl_scaling;
583 	cpu->pstate.max_pstate_physical = intel_pstate_freq_to_hwp(cpu, freq);
584 
585 	freq = cpu->pstate.min_pstate * perf_ctl_scaling;
586 	cpu->pstate.min_freq = freq;
587 	/*
588 	 * Cast the min P-state value retrieved via pstate_funcs.get_min() to
589 	 * the effective range of HWP performance levels.
590 	 */
591 	cpu->pstate.min_pstate = intel_pstate_freq_to_hwp(cpu, freq);
592 }
593 
594 static bool turbo_is_disabled(void)
595 {
596 	u64 misc_en;
597 
598 	rdmsrl(MSR_IA32_MISC_ENABLE, misc_en);
599 
600 	return !!(misc_en & MSR_IA32_MISC_ENABLE_TURBO_DISABLE);
601 }
602 
603 static int min_perf_pct_min(void)
604 {
605 	struct cpudata *cpu = all_cpu_data[0];
606 	int turbo_pstate = cpu->pstate.turbo_pstate;
607 
608 	return turbo_pstate ?
609 		(cpu->pstate.min_pstate * 100 / turbo_pstate) : 0;
610 }
611 
612 static s16 intel_pstate_get_epb(struct cpudata *cpu_data)
613 {
614 	u64 epb;
615 	int ret;
616 
617 	if (!boot_cpu_has(X86_FEATURE_EPB))
618 		return -ENXIO;
619 
620 	ret = rdmsrl_on_cpu(cpu_data->cpu, MSR_IA32_ENERGY_PERF_BIAS, &epb);
621 	if (ret)
622 		return (s16)ret;
623 
624 	return (s16)(epb & 0x0f);
625 }
626 
627 static s16 intel_pstate_get_epp(struct cpudata *cpu_data, u64 hwp_req_data)
628 {
629 	s16 epp;
630 
631 	if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
632 		/*
633 		 * When hwp_req_data is 0, means that caller didn't read
634 		 * MSR_HWP_REQUEST, so need to read and get EPP.
635 		 */
636 		if (!hwp_req_data) {
637 			epp = rdmsrl_on_cpu(cpu_data->cpu, MSR_HWP_REQUEST,
638 					    &hwp_req_data);
639 			if (epp)
640 				return epp;
641 		}
642 		epp = (hwp_req_data >> 24) & 0xff;
643 	} else {
644 		/* When there is no EPP present, HWP uses EPB settings */
645 		epp = intel_pstate_get_epb(cpu_data);
646 	}
647 
648 	return epp;
649 }
650 
651 static int intel_pstate_set_epb(int cpu, s16 pref)
652 {
653 	u64 epb;
654 	int ret;
655 
656 	if (!boot_cpu_has(X86_FEATURE_EPB))
657 		return -ENXIO;
658 
659 	ret = rdmsrl_on_cpu(cpu, MSR_IA32_ENERGY_PERF_BIAS, &epb);
660 	if (ret)
661 		return ret;
662 
663 	epb = (epb & ~0x0f) | pref;
664 	wrmsrl_on_cpu(cpu, MSR_IA32_ENERGY_PERF_BIAS, epb);
665 
666 	return 0;
667 }
668 
669 /*
670  * EPP/EPB display strings corresponding to EPP index in the
671  * energy_perf_strings[]
672  *	index		String
673  *-------------------------------------
674  *	0		default
675  *	1		performance
676  *	2		balance_performance
677  *	3		balance_power
678  *	4		power
679  */
680 
681 enum energy_perf_value_index {
682 	EPP_INDEX_DEFAULT = 0,
683 	EPP_INDEX_PERFORMANCE,
684 	EPP_INDEX_BALANCE_PERFORMANCE,
685 	EPP_INDEX_BALANCE_POWERSAVE,
686 	EPP_INDEX_POWERSAVE,
687 };
688 
689 static const char * const energy_perf_strings[] = {
690 	[EPP_INDEX_DEFAULT] = "default",
691 	[EPP_INDEX_PERFORMANCE] = "performance",
692 	[EPP_INDEX_BALANCE_PERFORMANCE] = "balance_performance",
693 	[EPP_INDEX_BALANCE_POWERSAVE] = "balance_power",
694 	[EPP_INDEX_POWERSAVE] = "power",
695 	NULL
696 };
697 static unsigned int epp_values[] = {
698 	[EPP_INDEX_DEFAULT] = 0, /* Unused index */
699 	[EPP_INDEX_PERFORMANCE] = HWP_EPP_PERFORMANCE,
700 	[EPP_INDEX_BALANCE_PERFORMANCE] = HWP_EPP_BALANCE_PERFORMANCE,
701 	[EPP_INDEX_BALANCE_POWERSAVE] = HWP_EPP_BALANCE_POWERSAVE,
702 	[EPP_INDEX_POWERSAVE] = HWP_EPP_POWERSAVE,
703 };
704 
705 static int intel_pstate_get_energy_pref_index(struct cpudata *cpu_data, int *raw_epp)
706 {
707 	s16 epp;
708 	int index = -EINVAL;
709 
710 	*raw_epp = 0;
711 	epp = intel_pstate_get_epp(cpu_data, 0);
712 	if (epp < 0)
713 		return epp;
714 
715 	if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
716 		if (epp == epp_values[EPP_INDEX_PERFORMANCE])
717 			return EPP_INDEX_PERFORMANCE;
718 		if (epp == epp_values[EPP_INDEX_BALANCE_PERFORMANCE])
719 			return EPP_INDEX_BALANCE_PERFORMANCE;
720 		if (epp == epp_values[EPP_INDEX_BALANCE_POWERSAVE])
721 			return EPP_INDEX_BALANCE_POWERSAVE;
722 		if (epp == epp_values[EPP_INDEX_POWERSAVE])
723 			return EPP_INDEX_POWERSAVE;
724 		*raw_epp = epp;
725 		return 0;
726 	} else if (boot_cpu_has(X86_FEATURE_EPB)) {
727 		/*
728 		 * Range:
729 		 *	0x00-0x03	:	Performance
730 		 *	0x04-0x07	:	Balance performance
731 		 *	0x08-0x0B	:	Balance power
732 		 *	0x0C-0x0F	:	Power
733 		 * The EPB is a 4 bit value, but our ranges restrict the
734 		 * value which can be set. Here only using top two bits
735 		 * effectively.
736 		 */
737 		index = (epp >> 2) + 1;
738 	}
739 
740 	return index;
741 }
742 
743 static int intel_pstate_set_epp(struct cpudata *cpu, u32 epp)
744 {
745 	int ret;
746 
747 	/*
748 	 * Use the cached HWP Request MSR value, because in the active mode the
749 	 * register itself may be updated by intel_pstate_hwp_boost_up() or
750 	 * intel_pstate_hwp_boost_down() at any time.
751 	 */
752 	u64 value = READ_ONCE(cpu->hwp_req_cached);
753 
754 	value &= ~GENMASK_ULL(31, 24);
755 	value |= (u64)epp << 24;
756 	/*
757 	 * The only other updater of hwp_req_cached in the active mode,
758 	 * intel_pstate_hwp_set(), is called under the same lock as this
759 	 * function, so it cannot run in parallel with the update below.
760 	 */
761 	WRITE_ONCE(cpu->hwp_req_cached, value);
762 	ret = wrmsrl_on_cpu(cpu->cpu, MSR_HWP_REQUEST, value);
763 	if (!ret)
764 		cpu->epp_cached = epp;
765 
766 	return ret;
767 }
768 
769 static int intel_pstate_set_energy_pref_index(struct cpudata *cpu_data,
770 					      int pref_index, bool use_raw,
771 					      u32 raw_epp)
772 {
773 	int epp = -EINVAL;
774 	int ret;
775 
776 	if (!pref_index)
777 		epp = cpu_data->epp_default;
778 
779 	if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
780 		if (use_raw)
781 			epp = raw_epp;
782 		else if (epp == -EINVAL)
783 			epp = epp_values[pref_index];
784 
785 		/*
786 		 * To avoid confusion, refuse to set EPP to any values different
787 		 * from 0 (performance) if the current policy is "performance",
788 		 * because those values would be overridden.
789 		 */
790 		if (epp > 0 && cpu_data->policy == CPUFREQ_POLICY_PERFORMANCE)
791 			return -EBUSY;
792 
793 		ret = intel_pstate_set_epp(cpu_data, epp);
794 	} else {
795 		if (epp == -EINVAL)
796 			epp = (pref_index - 1) << 2;
797 		ret = intel_pstate_set_epb(cpu_data->cpu, epp);
798 	}
799 
800 	return ret;
801 }
802 
803 static ssize_t show_energy_performance_available_preferences(
804 				struct cpufreq_policy *policy, char *buf)
805 {
806 	int i = 0;
807 	int ret = 0;
808 
809 	while (energy_perf_strings[i] != NULL)
810 		ret += sprintf(&buf[ret], "%s ", energy_perf_strings[i++]);
811 
812 	ret += sprintf(&buf[ret], "\n");
813 
814 	return ret;
815 }
816 
817 cpufreq_freq_attr_ro(energy_performance_available_preferences);
818 
819 static struct cpufreq_driver intel_pstate;
820 
821 static ssize_t store_energy_performance_preference(
822 		struct cpufreq_policy *policy, const char *buf, size_t count)
823 {
824 	struct cpudata *cpu = all_cpu_data[policy->cpu];
825 	char str_preference[21];
826 	bool raw = false;
827 	ssize_t ret;
828 	u32 epp = 0;
829 
830 	ret = sscanf(buf, "%20s", str_preference);
831 	if (ret != 1)
832 		return -EINVAL;
833 
834 	ret = match_string(energy_perf_strings, -1, str_preference);
835 	if (ret < 0) {
836 		if (!boot_cpu_has(X86_FEATURE_HWP_EPP))
837 			return ret;
838 
839 		ret = kstrtouint(buf, 10, &epp);
840 		if (ret)
841 			return ret;
842 
843 		if (epp > 255)
844 			return -EINVAL;
845 
846 		raw = true;
847 	}
848 
849 	/*
850 	 * This function runs with the policy R/W semaphore held, which
851 	 * guarantees that the driver pointer will not change while it is
852 	 * running.
853 	 */
854 	if (!intel_pstate_driver)
855 		return -EAGAIN;
856 
857 	mutex_lock(&intel_pstate_limits_lock);
858 
859 	if (intel_pstate_driver == &intel_pstate) {
860 		ret = intel_pstate_set_energy_pref_index(cpu, ret, raw, epp);
861 	} else {
862 		/*
863 		 * In the passive mode the governor needs to be stopped on the
864 		 * target CPU before the EPP update and restarted after it,
865 		 * which is super-heavy-weight, so make sure it is worth doing
866 		 * upfront.
867 		 */
868 		if (!raw)
869 			epp = ret ? epp_values[ret] : cpu->epp_default;
870 
871 		if (cpu->epp_cached != epp) {
872 			int err;
873 
874 			cpufreq_stop_governor(policy);
875 			ret = intel_pstate_set_epp(cpu, epp);
876 			err = cpufreq_start_governor(policy);
877 			if (!ret)
878 				ret = err;
879 		} else {
880 			ret = 0;
881 		}
882 	}
883 
884 	mutex_unlock(&intel_pstate_limits_lock);
885 
886 	return ret ?: count;
887 }
888 
889 static ssize_t show_energy_performance_preference(
890 				struct cpufreq_policy *policy, char *buf)
891 {
892 	struct cpudata *cpu_data = all_cpu_data[policy->cpu];
893 	int preference, raw_epp;
894 
895 	preference = intel_pstate_get_energy_pref_index(cpu_data, &raw_epp);
896 	if (preference < 0)
897 		return preference;
898 
899 	if (raw_epp)
900 		return  sprintf(buf, "%d\n", raw_epp);
901 	else
902 		return  sprintf(buf, "%s\n", energy_perf_strings[preference]);
903 }
904 
905 cpufreq_freq_attr_rw(energy_performance_preference);
906 
907 static ssize_t show_base_frequency(struct cpufreq_policy *policy, char *buf)
908 {
909 	struct cpudata *cpu = all_cpu_data[policy->cpu];
910 	int ratio, freq;
911 
912 	ratio = intel_pstate_get_cppc_guaranteed(policy->cpu);
913 	if (ratio <= 0) {
914 		u64 cap;
915 
916 		rdmsrl_on_cpu(policy->cpu, MSR_HWP_CAPABILITIES, &cap);
917 		ratio = HWP_GUARANTEED_PERF(cap);
918 	}
919 
920 	freq = ratio * cpu->pstate.scaling;
921 	if (cpu->pstate.scaling != cpu->pstate.perf_ctl_scaling)
922 		freq = rounddown(freq, cpu->pstate.perf_ctl_scaling);
923 
924 	return sprintf(buf, "%d\n", freq);
925 }
926 
927 cpufreq_freq_attr_ro(base_frequency);
928 
929 static struct freq_attr *hwp_cpufreq_attrs[] = {
930 	&energy_performance_preference,
931 	&energy_performance_available_preferences,
932 	&base_frequency,
933 	NULL,
934 };
935 
936 static void __intel_pstate_get_hwp_cap(struct cpudata *cpu)
937 {
938 	u64 cap;
939 
940 	rdmsrl_on_cpu(cpu->cpu, MSR_HWP_CAPABILITIES, &cap);
941 	WRITE_ONCE(cpu->hwp_cap_cached, cap);
942 	cpu->pstate.max_pstate = HWP_GUARANTEED_PERF(cap);
943 	cpu->pstate.turbo_pstate = HWP_HIGHEST_PERF(cap);
944 }
945 
946 static void intel_pstate_get_hwp_cap(struct cpudata *cpu)
947 {
948 	int scaling = cpu->pstate.scaling;
949 
950 	__intel_pstate_get_hwp_cap(cpu);
951 
952 	cpu->pstate.max_freq = cpu->pstate.max_pstate * scaling;
953 	cpu->pstate.turbo_freq = cpu->pstate.turbo_pstate * scaling;
954 	if (scaling != cpu->pstate.perf_ctl_scaling) {
955 		int perf_ctl_scaling = cpu->pstate.perf_ctl_scaling;
956 
957 		cpu->pstate.max_freq = rounddown(cpu->pstate.max_freq,
958 						 perf_ctl_scaling);
959 		cpu->pstate.turbo_freq = rounddown(cpu->pstate.turbo_freq,
960 						   perf_ctl_scaling);
961 	}
962 }
963 
964 static void intel_pstate_hwp_set(unsigned int cpu)
965 {
966 	struct cpudata *cpu_data = all_cpu_data[cpu];
967 	int max, min;
968 	u64 value;
969 	s16 epp;
970 
971 	max = cpu_data->max_perf_ratio;
972 	min = cpu_data->min_perf_ratio;
973 
974 	if (cpu_data->policy == CPUFREQ_POLICY_PERFORMANCE)
975 		min = max;
976 
977 	rdmsrl_on_cpu(cpu, MSR_HWP_REQUEST, &value);
978 
979 	value &= ~HWP_MIN_PERF(~0L);
980 	value |= HWP_MIN_PERF(min);
981 
982 	value &= ~HWP_MAX_PERF(~0L);
983 	value |= HWP_MAX_PERF(max);
984 
985 	if (cpu_data->epp_policy == cpu_data->policy)
986 		goto skip_epp;
987 
988 	cpu_data->epp_policy = cpu_data->policy;
989 
990 	if (cpu_data->policy == CPUFREQ_POLICY_PERFORMANCE) {
991 		epp = intel_pstate_get_epp(cpu_data, value);
992 		cpu_data->epp_powersave = epp;
993 		/* If EPP read was failed, then don't try to write */
994 		if (epp < 0)
995 			goto skip_epp;
996 
997 		epp = 0;
998 	} else {
999 		/* skip setting EPP, when saved value is invalid */
1000 		if (cpu_data->epp_powersave < 0)
1001 			goto skip_epp;
1002 
1003 		/*
1004 		 * No need to restore EPP when it is not zero. This
1005 		 * means:
1006 		 *  - Policy is not changed
1007 		 *  - user has manually changed
1008 		 *  - Error reading EPB
1009 		 */
1010 		epp = intel_pstate_get_epp(cpu_data, value);
1011 		if (epp)
1012 			goto skip_epp;
1013 
1014 		epp = cpu_data->epp_powersave;
1015 	}
1016 	if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
1017 		value &= ~GENMASK_ULL(31, 24);
1018 		value |= (u64)epp << 24;
1019 	} else {
1020 		intel_pstate_set_epb(cpu, epp);
1021 	}
1022 skip_epp:
1023 	WRITE_ONCE(cpu_data->hwp_req_cached, value);
1024 	wrmsrl_on_cpu(cpu, MSR_HWP_REQUEST, value);
1025 }
1026 
1027 static void intel_pstate_disable_hwp_interrupt(struct cpudata *cpudata);
1028 
1029 static void intel_pstate_hwp_offline(struct cpudata *cpu)
1030 {
1031 	u64 value = READ_ONCE(cpu->hwp_req_cached);
1032 	int min_perf;
1033 
1034 	intel_pstate_disable_hwp_interrupt(cpu);
1035 
1036 	if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
1037 		/*
1038 		 * In case the EPP has been set to "performance" by the
1039 		 * active mode "performance" scaling algorithm, replace that
1040 		 * temporary value with the cached EPP one.
1041 		 */
1042 		value &= ~GENMASK_ULL(31, 24);
1043 		value |= HWP_ENERGY_PERF_PREFERENCE(cpu->epp_cached);
1044 		/*
1045 		 * However, make sure that EPP will be set to "performance" when
1046 		 * the CPU is brought back online again and the "performance"
1047 		 * scaling algorithm is still in effect.
1048 		 */
1049 		cpu->epp_policy = CPUFREQ_POLICY_UNKNOWN;
1050 	}
1051 
1052 	/*
1053 	 * Clear the desired perf field in the cached HWP request value to
1054 	 * prevent nonzero desired values from being leaked into the active
1055 	 * mode.
1056 	 */
1057 	value &= ~HWP_DESIRED_PERF(~0L);
1058 	WRITE_ONCE(cpu->hwp_req_cached, value);
1059 
1060 	value &= ~GENMASK_ULL(31, 0);
1061 	min_perf = HWP_LOWEST_PERF(READ_ONCE(cpu->hwp_cap_cached));
1062 
1063 	/* Set hwp_max = hwp_min */
1064 	value |= HWP_MAX_PERF(min_perf);
1065 	value |= HWP_MIN_PERF(min_perf);
1066 
1067 	/* Set EPP to min */
1068 	if (boot_cpu_has(X86_FEATURE_HWP_EPP))
1069 		value |= HWP_ENERGY_PERF_PREFERENCE(HWP_EPP_POWERSAVE);
1070 
1071 	wrmsrl_on_cpu(cpu->cpu, MSR_HWP_REQUEST, value);
1072 }
1073 
1074 #define POWER_CTL_EE_ENABLE	1
1075 #define POWER_CTL_EE_DISABLE	2
1076 
1077 static int power_ctl_ee_state;
1078 
1079 static void set_power_ctl_ee_state(bool input)
1080 {
1081 	u64 power_ctl;
1082 
1083 	mutex_lock(&intel_pstate_driver_lock);
1084 	rdmsrl(MSR_IA32_POWER_CTL, power_ctl);
1085 	if (input) {
1086 		power_ctl &= ~BIT(MSR_IA32_POWER_CTL_BIT_EE);
1087 		power_ctl_ee_state = POWER_CTL_EE_ENABLE;
1088 	} else {
1089 		power_ctl |= BIT(MSR_IA32_POWER_CTL_BIT_EE);
1090 		power_ctl_ee_state = POWER_CTL_EE_DISABLE;
1091 	}
1092 	wrmsrl(MSR_IA32_POWER_CTL, power_ctl);
1093 	mutex_unlock(&intel_pstate_driver_lock);
1094 }
1095 
1096 static void intel_pstate_hwp_enable(struct cpudata *cpudata);
1097 
1098 static void intel_pstate_hwp_reenable(struct cpudata *cpu)
1099 {
1100 	intel_pstate_hwp_enable(cpu);
1101 	wrmsrl_on_cpu(cpu->cpu, MSR_HWP_REQUEST, READ_ONCE(cpu->hwp_req_cached));
1102 }
1103 
1104 static int intel_pstate_suspend(struct cpufreq_policy *policy)
1105 {
1106 	struct cpudata *cpu = all_cpu_data[policy->cpu];
1107 
1108 	pr_debug("CPU %d suspending\n", cpu->cpu);
1109 
1110 	cpu->suspended = true;
1111 
1112 	/* disable HWP interrupt and cancel any pending work */
1113 	intel_pstate_disable_hwp_interrupt(cpu);
1114 
1115 	return 0;
1116 }
1117 
1118 static int intel_pstate_resume(struct cpufreq_policy *policy)
1119 {
1120 	struct cpudata *cpu = all_cpu_data[policy->cpu];
1121 
1122 	pr_debug("CPU %d resuming\n", cpu->cpu);
1123 
1124 	/* Only restore if the system default is changed */
1125 	if (power_ctl_ee_state == POWER_CTL_EE_ENABLE)
1126 		set_power_ctl_ee_state(true);
1127 	else if (power_ctl_ee_state == POWER_CTL_EE_DISABLE)
1128 		set_power_ctl_ee_state(false);
1129 
1130 	if (cpu->suspended && hwp_active) {
1131 		mutex_lock(&intel_pstate_limits_lock);
1132 
1133 		/* Re-enable HWP, because "online" has not done that. */
1134 		intel_pstate_hwp_reenable(cpu);
1135 
1136 		mutex_unlock(&intel_pstate_limits_lock);
1137 	}
1138 
1139 	cpu->suspended = false;
1140 
1141 	return 0;
1142 }
1143 
1144 static void intel_pstate_update_policies(void)
1145 {
1146 	int cpu;
1147 
1148 	for_each_possible_cpu(cpu)
1149 		cpufreq_update_policy(cpu);
1150 }
1151 
1152 static void __intel_pstate_update_max_freq(struct cpudata *cpudata,
1153 					   struct cpufreq_policy *policy)
1154 {
1155 	if (hwp_active)
1156 		intel_pstate_get_hwp_cap(cpudata);
1157 
1158 	policy->cpuinfo.max_freq = READ_ONCE(global.no_turbo) ?
1159 			cpudata->pstate.max_freq : cpudata->pstate.turbo_freq;
1160 
1161 	refresh_frequency_limits(policy);
1162 }
1163 
1164 static void intel_pstate_update_limits(unsigned int cpu)
1165 {
1166 	struct cpufreq_policy *policy = cpufreq_cpu_acquire(cpu);
1167 
1168 	if (!policy)
1169 		return;
1170 
1171 	__intel_pstate_update_max_freq(all_cpu_data[cpu], policy);
1172 
1173 	cpufreq_cpu_release(policy);
1174 }
1175 
1176 static void intel_pstate_update_limits_for_all(void)
1177 {
1178 	int cpu;
1179 
1180 	for_each_possible_cpu(cpu)
1181 		intel_pstate_update_limits(cpu);
1182 }
1183 
1184 /************************** sysfs begin ************************/
1185 #define show_one(file_name, object)					\
1186 	static ssize_t show_##file_name					\
1187 	(struct kobject *kobj, struct kobj_attribute *attr, char *buf)	\
1188 	{								\
1189 		return sprintf(buf, "%u\n", global.object);		\
1190 	}
1191 
1192 static ssize_t intel_pstate_show_status(char *buf);
1193 static int intel_pstate_update_status(const char *buf, size_t size);
1194 
1195 static ssize_t show_status(struct kobject *kobj,
1196 			   struct kobj_attribute *attr, char *buf)
1197 {
1198 	ssize_t ret;
1199 
1200 	mutex_lock(&intel_pstate_driver_lock);
1201 	ret = intel_pstate_show_status(buf);
1202 	mutex_unlock(&intel_pstate_driver_lock);
1203 
1204 	return ret;
1205 }
1206 
1207 static ssize_t store_status(struct kobject *a, struct kobj_attribute *b,
1208 			    const char *buf, size_t count)
1209 {
1210 	char *p = memchr(buf, '\n', count);
1211 	int ret;
1212 
1213 	mutex_lock(&intel_pstate_driver_lock);
1214 	ret = intel_pstate_update_status(buf, p ? p - buf : count);
1215 	mutex_unlock(&intel_pstate_driver_lock);
1216 
1217 	return ret < 0 ? ret : count;
1218 }
1219 
1220 static ssize_t show_turbo_pct(struct kobject *kobj,
1221 				struct kobj_attribute *attr, char *buf)
1222 {
1223 	struct cpudata *cpu;
1224 	int total, no_turbo, turbo_pct;
1225 	uint32_t turbo_fp;
1226 
1227 	mutex_lock(&intel_pstate_driver_lock);
1228 
1229 	if (!intel_pstate_driver) {
1230 		mutex_unlock(&intel_pstate_driver_lock);
1231 		return -EAGAIN;
1232 	}
1233 
1234 	cpu = all_cpu_data[0];
1235 
1236 	total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
1237 	no_turbo = cpu->pstate.max_pstate - cpu->pstate.min_pstate + 1;
1238 	turbo_fp = div_fp(no_turbo, total);
1239 	turbo_pct = 100 - fp_toint(mul_fp(turbo_fp, int_tofp(100)));
1240 
1241 	mutex_unlock(&intel_pstate_driver_lock);
1242 
1243 	return sprintf(buf, "%u\n", turbo_pct);
1244 }
1245 
1246 static ssize_t show_num_pstates(struct kobject *kobj,
1247 				struct kobj_attribute *attr, char *buf)
1248 {
1249 	struct cpudata *cpu;
1250 	int total;
1251 
1252 	mutex_lock(&intel_pstate_driver_lock);
1253 
1254 	if (!intel_pstate_driver) {
1255 		mutex_unlock(&intel_pstate_driver_lock);
1256 		return -EAGAIN;
1257 	}
1258 
1259 	cpu = all_cpu_data[0];
1260 	total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
1261 
1262 	mutex_unlock(&intel_pstate_driver_lock);
1263 
1264 	return sprintf(buf, "%u\n", total);
1265 }
1266 
1267 static ssize_t show_no_turbo(struct kobject *kobj,
1268 			     struct kobj_attribute *attr, char *buf)
1269 {
1270 	ssize_t ret;
1271 
1272 	mutex_lock(&intel_pstate_driver_lock);
1273 
1274 	if (!intel_pstate_driver) {
1275 		mutex_unlock(&intel_pstate_driver_lock);
1276 		return -EAGAIN;
1277 	}
1278 
1279 	ret = sprintf(buf, "%u\n", global.no_turbo);
1280 
1281 	mutex_unlock(&intel_pstate_driver_lock);
1282 
1283 	return ret;
1284 }
1285 
1286 static ssize_t store_no_turbo(struct kobject *a, struct kobj_attribute *b,
1287 			      const char *buf, size_t count)
1288 {
1289 	unsigned int input;
1290 	bool no_turbo;
1291 
1292 	if (sscanf(buf, "%u", &input) != 1)
1293 		return -EINVAL;
1294 
1295 	mutex_lock(&intel_pstate_driver_lock);
1296 
1297 	if (!intel_pstate_driver) {
1298 		count = -EAGAIN;
1299 		goto unlock_driver;
1300 	}
1301 
1302 	no_turbo = !!clamp_t(int, input, 0, 1);
1303 
1304 	WRITE_ONCE(global.turbo_disabled, turbo_is_disabled());
1305 	if (global.turbo_disabled && !no_turbo) {
1306 		pr_notice("Turbo disabled by BIOS or unavailable on processor\n");
1307 		count = -EPERM;
1308 		if (global.no_turbo)
1309 			goto unlock_driver;
1310 		else
1311 			no_turbo = 1;
1312 	}
1313 
1314 	if (no_turbo == global.no_turbo) {
1315 		goto unlock_driver;
1316 	}
1317 
1318 	WRITE_ONCE(global.no_turbo, no_turbo);
1319 
1320 	mutex_lock(&intel_pstate_limits_lock);
1321 
1322 	if (no_turbo) {
1323 		struct cpudata *cpu = all_cpu_data[0];
1324 		int pct = cpu->pstate.max_pstate * 100 / cpu->pstate.turbo_pstate;
1325 
1326 		/* Squash the global minimum into the permitted range. */
1327 		if (global.min_perf_pct > pct)
1328 			global.min_perf_pct = pct;
1329 	}
1330 
1331 	mutex_unlock(&intel_pstate_limits_lock);
1332 
1333 	intel_pstate_update_limits_for_all();
1334 	arch_set_max_freq_ratio(no_turbo);
1335 
1336 unlock_driver:
1337 	mutex_unlock(&intel_pstate_driver_lock);
1338 
1339 	return count;
1340 }
1341 
1342 static void update_qos_request(enum freq_qos_req_type type)
1343 {
1344 	struct freq_qos_request *req;
1345 	struct cpufreq_policy *policy;
1346 	int i;
1347 
1348 	for_each_possible_cpu(i) {
1349 		struct cpudata *cpu = all_cpu_data[i];
1350 		unsigned int freq, perf_pct;
1351 
1352 		policy = cpufreq_cpu_get(i);
1353 		if (!policy)
1354 			continue;
1355 
1356 		req = policy->driver_data;
1357 		cpufreq_cpu_put(policy);
1358 
1359 		if (!req)
1360 			continue;
1361 
1362 		if (hwp_active)
1363 			intel_pstate_get_hwp_cap(cpu);
1364 
1365 		if (type == FREQ_QOS_MIN) {
1366 			perf_pct = global.min_perf_pct;
1367 		} else {
1368 			req++;
1369 			perf_pct = global.max_perf_pct;
1370 		}
1371 
1372 		freq = DIV_ROUND_UP(cpu->pstate.turbo_freq * perf_pct, 100);
1373 
1374 		if (freq_qos_update_request(req, freq) < 0)
1375 			pr_warn("Failed to update freq constraint: CPU%d\n", i);
1376 	}
1377 }
1378 
1379 static ssize_t store_max_perf_pct(struct kobject *a, struct kobj_attribute *b,
1380 				  const char *buf, size_t count)
1381 {
1382 	unsigned int input;
1383 	int ret;
1384 
1385 	ret = sscanf(buf, "%u", &input);
1386 	if (ret != 1)
1387 		return -EINVAL;
1388 
1389 	mutex_lock(&intel_pstate_driver_lock);
1390 
1391 	if (!intel_pstate_driver) {
1392 		mutex_unlock(&intel_pstate_driver_lock);
1393 		return -EAGAIN;
1394 	}
1395 
1396 	mutex_lock(&intel_pstate_limits_lock);
1397 
1398 	global.max_perf_pct = clamp_t(int, input, global.min_perf_pct, 100);
1399 
1400 	mutex_unlock(&intel_pstate_limits_lock);
1401 
1402 	if (intel_pstate_driver == &intel_pstate)
1403 		intel_pstate_update_policies();
1404 	else
1405 		update_qos_request(FREQ_QOS_MAX);
1406 
1407 	mutex_unlock(&intel_pstate_driver_lock);
1408 
1409 	return count;
1410 }
1411 
1412 static ssize_t store_min_perf_pct(struct kobject *a, struct kobj_attribute *b,
1413 				  const char *buf, size_t count)
1414 {
1415 	unsigned int input;
1416 	int ret;
1417 
1418 	ret = sscanf(buf, "%u", &input);
1419 	if (ret != 1)
1420 		return -EINVAL;
1421 
1422 	mutex_lock(&intel_pstate_driver_lock);
1423 
1424 	if (!intel_pstate_driver) {
1425 		mutex_unlock(&intel_pstate_driver_lock);
1426 		return -EAGAIN;
1427 	}
1428 
1429 	mutex_lock(&intel_pstate_limits_lock);
1430 
1431 	global.min_perf_pct = clamp_t(int, input,
1432 				      min_perf_pct_min(), global.max_perf_pct);
1433 
1434 	mutex_unlock(&intel_pstate_limits_lock);
1435 
1436 	if (intel_pstate_driver == &intel_pstate)
1437 		intel_pstate_update_policies();
1438 	else
1439 		update_qos_request(FREQ_QOS_MIN);
1440 
1441 	mutex_unlock(&intel_pstate_driver_lock);
1442 
1443 	return count;
1444 }
1445 
1446 static ssize_t show_hwp_dynamic_boost(struct kobject *kobj,
1447 				struct kobj_attribute *attr, char *buf)
1448 {
1449 	return sprintf(buf, "%u\n", hwp_boost);
1450 }
1451 
1452 static ssize_t store_hwp_dynamic_boost(struct kobject *a,
1453 				       struct kobj_attribute *b,
1454 				       const char *buf, size_t count)
1455 {
1456 	unsigned int input;
1457 	int ret;
1458 
1459 	ret = kstrtouint(buf, 10, &input);
1460 	if (ret)
1461 		return ret;
1462 
1463 	mutex_lock(&intel_pstate_driver_lock);
1464 	hwp_boost = !!input;
1465 	intel_pstate_update_policies();
1466 	mutex_unlock(&intel_pstate_driver_lock);
1467 
1468 	return count;
1469 }
1470 
1471 static ssize_t show_energy_efficiency(struct kobject *kobj, struct kobj_attribute *attr,
1472 				      char *buf)
1473 {
1474 	u64 power_ctl;
1475 	int enable;
1476 
1477 	rdmsrl(MSR_IA32_POWER_CTL, power_ctl);
1478 	enable = !!(power_ctl & BIT(MSR_IA32_POWER_CTL_BIT_EE));
1479 	return sprintf(buf, "%d\n", !enable);
1480 }
1481 
1482 static ssize_t store_energy_efficiency(struct kobject *a, struct kobj_attribute *b,
1483 				       const char *buf, size_t count)
1484 {
1485 	bool input;
1486 	int ret;
1487 
1488 	ret = kstrtobool(buf, &input);
1489 	if (ret)
1490 		return ret;
1491 
1492 	set_power_ctl_ee_state(input);
1493 
1494 	return count;
1495 }
1496 
1497 show_one(max_perf_pct, max_perf_pct);
1498 show_one(min_perf_pct, min_perf_pct);
1499 
1500 define_one_global_rw(status);
1501 define_one_global_rw(no_turbo);
1502 define_one_global_rw(max_perf_pct);
1503 define_one_global_rw(min_perf_pct);
1504 define_one_global_ro(turbo_pct);
1505 define_one_global_ro(num_pstates);
1506 define_one_global_rw(hwp_dynamic_boost);
1507 define_one_global_rw(energy_efficiency);
1508 
1509 static struct attribute *intel_pstate_attributes[] = {
1510 	&status.attr,
1511 	&no_turbo.attr,
1512 	NULL
1513 };
1514 
1515 static const struct attribute_group intel_pstate_attr_group = {
1516 	.attrs = intel_pstate_attributes,
1517 };
1518 
1519 static const struct x86_cpu_id intel_pstate_cpu_ee_disable_ids[];
1520 
1521 static struct kobject *intel_pstate_kobject;
1522 
1523 static void __init intel_pstate_sysfs_expose_params(void)
1524 {
1525 	struct device *dev_root = bus_get_dev_root(&cpu_subsys);
1526 	int rc;
1527 
1528 	if (dev_root) {
1529 		intel_pstate_kobject = kobject_create_and_add("intel_pstate", &dev_root->kobj);
1530 		put_device(dev_root);
1531 	}
1532 	if (WARN_ON(!intel_pstate_kobject))
1533 		return;
1534 
1535 	rc = sysfs_create_group(intel_pstate_kobject, &intel_pstate_attr_group);
1536 	if (WARN_ON(rc))
1537 		return;
1538 
1539 	if (!boot_cpu_has(X86_FEATURE_HYBRID_CPU)) {
1540 		rc = sysfs_create_file(intel_pstate_kobject, &turbo_pct.attr);
1541 		WARN_ON(rc);
1542 
1543 		rc = sysfs_create_file(intel_pstate_kobject, &num_pstates.attr);
1544 		WARN_ON(rc);
1545 	}
1546 
1547 	/*
1548 	 * If per cpu limits are enforced there are no global limits, so
1549 	 * return without creating max/min_perf_pct attributes
1550 	 */
1551 	if (per_cpu_limits)
1552 		return;
1553 
1554 	rc = sysfs_create_file(intel_pstate_kobject, &max_perf_pct.attr);
1555 	WARN_ON(rc);
1556 
1557 	rc = sysfs_create_file(intel_pstate_kobject, &min_perf_pct.attr);
1558 	WARN_ON(rc);
1559 
1560 	if (x86_match_cpu(intel_pstate_cpu_ee_disable_ids)) {
1561 		rc = sysfs_create_file(intel_pstate_kobject, &energy_efficiency.attr);
1562 		WARN_ON(rc);
1563 	}
1564 }
1565 
1566 static void __init intel_pstate_sysfs_remove(void)
1567 {
1568 	if (!intel_pstate_kobject)
1569 		return;
1570 
1571 	sysfs_remove_group(intel_pstate_kobject, &intel_pstate_attr_group);
1572 
1573 	if (!boot_cpu_has(X86_FEATURE_HYBRID_CPU)) {
1574 		sysfs_remove_file(intel_pstate_kobject, &num_pstates.attr);
1575 		sysfs_remove_file(intel_pstate_kobject, &turbo_pct.attr);
1576 	}
1577 
1578 	if (!per_cpu_limits) {
1579 		sysfs_remove_file(intel_pstate_kobject, &max_perf_pct.attr);
1580 		sysfs_remove_file(intel_pstate_kobject, &min_perf_pct.attr);
1581 
1582 		if (x86_match_cpu(intel_pstate_cpu_ee_disable_ids))
1583 			sysfs_remove_file(intel_pstate_kobject, &energy_efficiency.attr);
1584 	}
1585 
1586 	kobject_put(intel_pstate_kobject);
1587 }
1588 
1589 static void intel_pstate_sysfs_expose_hwp_dynamic_boost(void)
1590 {
1591 	int rc;
1592 
1593 	if (!hwp_active)
1594 		return;
1595 
1596 	rc = sysfs_create_file(intel_pstate_kobject, &hwp_dynamic_boost.attr);
1597 	WARN_ON_ONCE(rc);
1598 }
1599 
1600 static void intel_pstate_sysfs_hide_hwp_dynamic_boost(void)
1601 {
1602 	if (!hwp_active)
1603 		return;
1604 
1605 	sysfs_remove_file(intel_pstate_kobject, &hwp_dynamic_boost.attr);
1606 }
1607 
1608 /************************** sysfs end ************************/
1609 
1610 static void intel_pstate_notify_work(struct work_struct *work)
1611 {
1612 	struct cpudata *cpudata =
1613 		container_of(to_delayed_work(work), struct cpudata, hwp_notify_work);
1614 	struct cpufreq_policy *policy = cpufreq_cpu_acquire(cpudata->cpu);
1615 
1616 	if (policy) {
1617 		__intel_pstate_update_max_freq(cpudata, policy);
1618 
1619 		cpufreq_cpu_release(policy);
1620 	}
1621 
1622 	wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_STATUS, 0);
1623 }
1624 
1625 static DEFINE_SPINLOCK(hwp_notify_lock);
1626 static cpumask_t hwp_intr_enable_mask;
1627 
1628 void notify_hwp_interrupt(void)
1629 {
1630 	unsigned int this_cpu = smp_processor_id();
1631 	unsigned long flags;
1632 	u64 value;
1633 
1634 	if (!hwp_active || !boot_cpu_has(X86_FEATURE_HWP_NOTIFY))
1635 		return;
1636 
1637 	rdmsrl_safe(MSR_HWP_STATUS, &value);
1638 	if (!(value & 0x01))
1639 		return;
1640 
1641 	spin_lock_irqsave(&hwp_notify_lock, flags);
1642 
1643 	if (!cpumask_test_cpu(this_cpu, &hwp_intr_enable_mask))
1644 		goto ack_intr;
1645 
1646 	schedule_delayed_work(&all_cpu_data[this_cpu]->hwp_notify_work,
1647 			      msecs_to_jiffies(10));
1648 
1649 	spin_unlock_irqrestore(&hwp_notify_lock, flags);
1650 
1651 	return;
1652 
1653 ack_intr:
1654 	wrmsrl_safe(MSR_HWP_STATUS, 0);
1655 	spin_unlock_irqrestore(&hwp_notify_lock, flags);
1656 }
1657 
1658 static void intel_pstate_disable_hwp_interrupt(struct cpudata *cpudata)
1659 {
1660 	bool cancel_work;
1661 
1662 	if (!boot_cpu_has(X86_FEATURE_HWP_NOTIFY))
1663 		return;
1664 
1665 	/* wrmsrl_on_cpu has to be outside spinlock as this can result in IPC */
1666 	wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_INTERRUPT, 0x00);
1667 
1668 	spin_lock_irq(&hwp_notify_lock);
1669 	cancel_work = cpumask_test_and_clear_cpu(cpudata->cpu, &hwp_intr_enable_mask);
1670 	spin_unlock_irq(&hwp_notify_lock);
1671 
1672 	if (cancel_work)
1673 		cancel_delayed_work_sync(&cpudata->hwp_notify_work);
1674 }
1675 
1676 static void intel_pstate_enable_hwp_interrupt(struct cpudata *cpudata)
1677 {
1678 	/* Enable HWP notification interrupt for guaranteed performance change */
1679 	if (boot_cpu_has(X86_FEATURE_HWP_NOTIFY)) {
1680 		spin_lock_irq(&hwp_notify_lock);
1681 		INIT_DELAYED_WORK(&cpudata->hwp_notify_work, intel_pstate_notify_work);
1682 		cpumask_set_cpu(cpudata->cpu, &hwp_intr_enable_mask);
1683 		spin_unlock_irq(&hwp_notify_lock);
1684 
1685 		/* wrmsrl_on_cpu has to be outside spinlock as this can result in IPC */
1686 		wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_INTERRUPT, 0x01);
1687 		wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_STATUS, 0);
1688 	}
1689 }
1690 
1691 static void intel_pstate_update_epp_defaults(struct cpudata *cpudata)
1692 {
1693 	cpudata->epp_default = intel_pstate_get_epp(cpudata, 0);
1694 
1695 	/*
1696 	 * If the EPP is set by firmware, which means that firmware enabled HWP
1697 	 * - Is equal or less than 0x80 (default balance_perf EPP)
1698 	 * - But less performance oriented than performance EPP
1699 	 *   then use this as new balance_perf EPP.
1700 	 */
1701 	if (hwp_forced && cpudata->epp_default <= HWP_EPP_BALANCE_PERFORMANCE &&
1702 	    cpudata->epp_default > HWP_EPP_PERFORMANCE) {
1703 		epp_values[EPP_INDEX_BALANCE_PERFORMANCE] = cpudata->epp_default;
1704 		return;
1705 	}
1706 
1707 	/*
1708 	 * If this CPU gen doesn't call for change in balance_perf
1709 	 * EPP return.
1710 	 */
1711 	if (epp_values[EPP_INDEX_BALANCE_PERFORMANCE] == HWP_EPP_BALANCE_PERFORMANCE)
1712 		return;
1713 
1714 	/*
1715 	 * Use hard coded value per gen to update the balance_perf
1716 	 * and default EPP.
1717 	 */
1718 	cpudata->epp_default = epp_values[EPP_INDEX_BALANCE_PERFORMANCE];
1719 	intel_pstate_set_epp(cpudata, cpudata->epp_default);
1720 }
1721 
1722 static void intel_pstate_hwp_enable(struct cpudata *cpudata)
1723 {
1724 	/* First disable HWP notification interrupt till we activate again */
1725 	if (boot_cpu_has(X86_FEATURE_HWP_NOTIFY))
1726 		wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_INTERRUPT, 0x00);
1727 
1728 	wrmsrl_on_cpu(cpudata->cpu, MSR_PM_ENABLE, 0x1);
1729 
1730 	intel_pstate_enable_hwp_interrupt(cpudata);
1731 
1732 	if (cpudata->epp_default >= 0)
1733 		return;
1734 
1735 	intel_pstate_update_epp_defaults(cpudata);
1736 }
1737 
1738 static int atom_get_min_pstate(int not_used)
1739 {
1740 	u64 value;
1741 
1742 	rdmsrl(MSR_ATOM_CORE_RATIOS, value);
1743 	return (value >> 8) & 0x7F;
1744 }
1745 
1746 static int atom_get_max_pstate(int not_used)
1747 {
1748 	u64 value;
1749 
1750 	rdmsrl(MSR_ATOM_CORE_RATIOS, value);
1751 	return (value >> 16) & 0x7F;
1752 }
1753 
1754 static int atom_get_turbo_pstate(int not_used)
1755 {
1756 	u64 value;
1757 
1758 	rdmsrl(MSR_ATOM_CORE_TURBO_RATIOS, value);
1759 	return value & 0x7F;
1760 }
1761 
1762 static u64 atom_get_val(struct cpudata *cpudata, int pstate)
1763 {
1764 	u64 val;
1765 	int32_t vid_fp;
1766 	u32 vid;
1767 
1768 	val = (u64)pstate << 8;
1769 	if (READ_ONCE(global.no_turbo) && !READ_ONCE(global.turbo_disabled))
1770 		val |= (u64)1 << 32;
1771 
1772 	vid_fp = cpudata->vid.min + mul_fp(
1773 		int_tofp(pstate - cpudata->pstate.min_pstate),
1774 		cpudata->vid.ratio);
1775 
1776 	vid_fp = clamp_t(int32_t, vid_fp, cpudata->vid.min, cpudata->vid.max);
1777 	vid = ceiling_fp(vid_fp);
1778 
1779 	if (pstate > cpudata->pstate.max_pstate)
1780 		vid = cpudata->vid.turbo;
1781 
1782 	return val | vid;
1783 }
1784 
1785 static int silvermont_get_scaling(void)
1786 {
1787 	u64 value;
1788 	int i;
1789 	/* Defined in Table 35-6 from SDM (Sept 2015) */
1790 	static int silvermont_freq_table[] = {
1791 		83300, 100000, 133300, 116700, 80000};
1792 
1793 	rdmsrl(MSR_FSB_FREQ, value);
1794 	i = value & 0x7;
1795 	WARN_ON(i > 4);
1796 
1797 	return silvermont_freq_table[i];
1798 }
1799 
1800 static int airmont_get_scaling(void)
1801 {
1802 	u64 value;
1803 	int i;
1804 	/* Defined in Table 35-10 from SDM (Sept 2015) */
1805 	static int airmont_freq_table[] = {
1806 		83300, 100000, 133300, 116700, 80000,
1807 		93300, 90000, 88900, 87500};
1808 
1809 	rdmsrl(MSR_FSB_FREQ, value);
1810 	i = value & 0xF;
1811 	WARN_ON(i > 8);
1812 
1813 	return airmont_freq_table[i];
1814 }
1815 
1816 static void atom_get_vid(struct cpudata *cpudata)
1817 {
1818 	u64 value;
1819 
1820 	rdmsrl(MSR_ATOM_CORE_VIDS, value);
1821 	cpudata->vid.min = int_tofp((value >> 8) & 0x7f);
1822 	cpudata->vid.max = int_tofp((value >> 16) & 0x7f);
1823 	cpudata->vid.ratio = div_fp(
1824 		cpudata->vid.max - cpudata->vid.min,
1825 		int_tofp(cpudata->pstate.max_pstate -
1826 			cpudata->pstate.min_pstate));
1827 
1828 	rdmsrl(MSR_ATOM_CORE_TURBO_VIDS, value);
1829 	cpudata->vid.turbo = value & 0x7f;
1830 }
1831 
1832 static int core_get_min_pstate(int cpu)
1833 {
1834 	u64 value;
1835 
1836 	rdmsrl_on_cpu(cpu, MSR_PLATFORM_INFO, &value);
1837 	return (value >> 40) & 0xFF;
1838 }
1839 
1840 static int core_get_max_pstate_physical(int cpu)
1841 {
1842 	u64 value;
1843 
1844 	rdmsrl_on_cpu(cpu, MSR_PLATFORM_INFO, &value);
1845 	return (value >> 8) & 0xFF;
1846 }
1847 
1848 static int core_get_tdp_ratio(int cpu, u64 plat_info)
1849 {
1850 	/* Check how many TDP levels present */
1851 	if (plat_info & 0x600000000) {
1852 		u64 tdp_ctrl;
1853 		u64 tdp_ratio;
1854 		int tdp_msr;
1855 		int err;
1856 
1857 		/* Get the TDP level (0, 1, 2) to get ratios */
1858 		err = rdmsrl_safe_on_cpu(cpu, MSR_CONFIG_TDP_CONTROL, &tdp_ctrl);
1859 		if (err)
1860 			return err;
1861 
1862 		/* TDP MSR are continuous starting at 0x648 */
1863 		tdp_msr = MSR_CONFIG_TDP_NOMINAL + (tdp_ctrl & 0x03);
1864 		err = rdmsrl_safe_on_cpu(cpu, tdp_msr, &tdp_ratio);
1865 		if (err)
1866 			return err;
1867 
1868 		/* For level 1 and 2, bits[23:16] contain the ratio */
1869 		if (tdp_ctrl & 0x03)
1870 			tdp_ratio >>= 16;
1871 
1872 		tdp_ratio &= 0xff; /* ratios are only 8 bits long */
1873 		pr_debug("tdp_ratio %x\n", (int)tdp_ratio);
1874 
1875 		return (int)tdp_ratio;
1876 	}
1877 
1878 	return -ENXIO;
1879 }
1880 
1881 static int core_get_max_pstate(int cpu)
1882 {
1883 	u64 tar;
1884 	u64 plat_info;
1885 	int max_pstate;
1886 	int tdp_ratio;
1887 	int err;
1888 
1889 	rdmsrl_on_cpu(cpu, MSR_PLATFORM_INFO, &plat_info);
1890 	max_pstate = (plat_info >> 8) & 0xFF;
1891 
1892 	tdp_ratio = core_get_tdp_ratio(cpu, plat_info);
1893 	if (tdp_ratio <= 0)
1894 		return max_pstate;
1895 
1896 	if (hwp_active) {
1897 		/* Turbo activation ratio is not used on HWP platforms */
1898 		return tdp_ratio;
1899 	}
1900 
1901 	err = rdmsrl_safe_on_cpu(cpu, MSR_TURBO_ACTIVATION_RATIO, &tar);
1902 	if (!err) {
1903 		int tar_levels;
1904 
1905 		/* Do some sanity checking for safety */
1906 		tar_levels = tar & 0xff;
1907 		if (tdp_ratio - 1 == tar_levels) {
1908 			max_pstate = tar_levels;
1909 			pr_debug("max_pstate=TAC %x\n", max_pstate);
1910 		}
1911 	}
1912 
1913 	return max_pstate;
1914 }
1915 
1916 static int core_get_turbo_pstate(int cpu)
1917 {
1918 	u64 value;
1919 	int nont, ret;
1920 
1921 	rdmsrl_on_cpu(cpu, MSR_TURBO_RATIO_LIMIT, &value);
1922 	nont = core_get_max_pstate(cpu);
1923 	ret = (value) & 255;
1924 	if (ret <= nont)
1925 		ret = nont;
1926 	return ret;
1927 }
1928 
1929 static u64 core_get_val(struct cpudata *cpudata, int pstate)
1930 {
1931 	u64 val;
1932 
1933 	val = (u64)pstate << 8;
1934 	if (READ_ONCE(global.no_turbo) && !READ_ONCE(global.turbo_disabled))
1935 		val |= (u64)1 << 32;
1936 
1937 	return val;
1938 }
1939 
1940 static int knl_get_aperf_mperf_shift(void)
1941 {
1942 	return 10;
1943 }
1944 
1945 static int knl_get_turbo_pstate(int cpu)
1946 {
1947 	u64 value;
1948 	int nont, ret;
1949 
1950 	rdmsrl_on_cpu(cpu, MSR_TURBO_RATIO_LIMIT, &value);
1951 	nont = core_get_max_pstate(cpu);
1952 	ret = (((value) >> 8) & 0xFF);
1953 	if (ret <= nont)
1954 		ret = nont;
1955 	return ret;
1956 }
1957 
1958 static void hybrid_get_type(void *data)
1959 {
1960 	u8 *cpu_type = data;
1961 
1962 	*cpu_type = get_this_hybrid_cpu_type();
1963 }
1964 
1965 static int hwp_get_cpu_scaling(int cpu)
1966 {
1967 	u8 cpu_type = 0;
1968 
1969 	smp_call_function_single(cpu, hybrid_get_type, &cpu_type, 1);
1970 	/* P-cores have a smaller perf level-to-freqency scaling factor. */
1971 	if (cpu_type == 0x40)
1972 		return hybrid_scaling_factor;
1973 
1974 	/* Use default core scaling for E-cores */
1975 	if (cpu_type == 0x20)
1976 		return core_get_scaling();
1977 
1978 	/*
1979 	 * If reached here, this system is either non-hybrid (like Tiger
1980 	 * Lake) or hybrid-capable (like Alder Lake or Raptor Lake) with
1981 	 * no E cores (in which case CPUID for hybrid support is 0).
1982 	 *
1983 	 * The CPPC nominal_frequency field is 0 for non-hybrid systems,
1984 	 * so the default core scaling will be used for them.
1985 	 */
1986 	return intel_pstate_cppc_get_scaling(cpu);
1987 }
1988 
1989 static void intel_pstate_set_pstate(struct cpudata *cpu, int pstate)
1990 {
1991 	trace_cpu_frequency(pstate * cpu->pstate.scaling, cpu->cpu);
1992 	cpu->pstate.current_pstate = pstate;
1993 	/*
1994 	 * Generally, there is no guarantee that this code will always run on
1995 	 * the CPU being updated, so force the register update to run on the
1996 	 * right CPU.
1997 	 */
1998 	wrmsrl_on_cpu(cpu->cpu, MSR_IA32_PERF_CTL,
1999 		      pstate_funcs.get_val(cpu, pstate));
2000 }
2001 
2002 static void intel_pstate_set_min_pstate(struct cpudata *cpu)
2003 {
2004 	intel_pstate_set_pstate(cpu, cpu->pstate.min_pstate);
2005 }
2006 
2007 static void intel_pstate_get_cpu_pstates(struct cpudata *cpu)
2008 {
2009 	int perf_ctl_max_phys = pstate_funcs.get_max_physical(cpu->cpu);
2010 	int perf_ctl_scaling = pstate_funcs.get_scaling();
2011 
2012 	cpu->pstate.min_pstate = pstate_funcs.get_min(cpu->cpu);
2013 	cpu->pstate.max_pstate_physical = perf_ctl_max_phys;
2014 	cpu->pstate.perf_ctl_scaling = perf_ctl_scaling;
2015 
2016 	if (hwp_active && !hwp_mode_bdw) {
2017 		__intel_pstate_get_hwp_cap(cpu);
2018 
2019 		if (pstate_funcs.get_cpu_scaling) {
2020 			cpu->pstate.scaling = pstate_funcs.get_cpu_scaling(cpu->cpu);
2021 			if (cpu->pstate.scaling != perf_ctl_scaling)
2022 				intel_pstate_hybrid_hwp_adjust(cpu);
2023 		} else {
2024 			cpu->pstate.scaling = perf_ctl_scaling;
2025 		}
2026 	} else {
2027 		cpu->pstate.scaling = perf_ctl_scaling;
2028 		cpu->pstate.max_pstate = pstate_funcs.get_max(cpu->cpu);
2029 		cpu->pstate.turbo_pstate = pstate_funcs.get_turbo(cpu->cpu);
2030 	}
2031 
2032 	if (cpu->pstate.scaling == perf_ctl_scaling) {
2033 		cpu->pstate.min_freq = cpu->pstate.min_pstate * perf_ctl_scaling;
2034 		cpu->pstate.max_freq = cpu->pstate.max_pstate * perf_ctl_scaling;
2035 		cpu->pstate.turbo_freq = cpu->pstate.turbo_pstate * perf_ctl_scaling;
2036 	}
2037 
2038 	if (pstate_funcs.get_aperf_mperf_shift)
2039 		cpu->aperf_mperf_shift = pstate_funcs.get_aperf_mperf_shift();
2040 
2041 	if (pstate_funcs.get_vid)
2042 		pstate_funcs.get_vid(cpu);
2043 
2044 	intel_pstate_set_min_pstate(cpu);
2045 }
2046 
2047 /*
2048  * Long hold time will keep high perf limits for long time,
2049  * which negatively impacts perf/watt for some workloads,
2050  * like specpower. 3ms is based on experiements on some
2051  * workoads.
2052  */
2053 static int hwp_boost_hold_time_ns = 3 * NSEC_PER_MSEC;
2054 
2055 static inline void intel_pstate_hwp_boost_up(struct cpudata *cpu)
2056 {
2057 	u64 hwp_req = READ_ONCE(cpu->hwp_req_cached);
2058 	u64 hwp_cap = READ_ONCE(cpu->hwp_cap_cached);
2059 	u32 max_limit = (hwp_req & 0xff00) >> 8;
2060 	u32 min_limit = (hwp_req & 0xff);
2061 	u32 boost_level1;
2062 
2063 	/*
2064 	 * Cases to consider (User changes via sysfs or boot time):
2065 	 * If, P0 (Turbo max) = P1 (Guaranteed max) = min:
2066 	 *	No boost, return.
2067 	 * If, P0 (Turbo max) > P1 (Guaranteed max) = min:
2068 	 *     Should result in one level boost only for P0.
2069 	 * If, P0 (Turbo max) = P1 (Guaranteed max) > min:
2070 	 *     Should result in two level boost:
2071 	 *         (min + p1)/2 and P1.
2072 	 * If, P0 (Turbo max) > P1 (Guaranteed max) > min:
2073 	 *     Should result in three level boost:
2074 	 *        (min + p1)/2, P1 and P0.
2075 	 */
2076 
2077 	/* If max and min are equal or already at max, nothing to boost */
2078 	if (max_limit == min_limit || cpu->hwp_boost_min >= max_limit)
2079 		return;
2080 
2081 	if (!cpu->hwp_boost_min)
2082 		cpu->hwp_boost_min = min_limit;
2083 
2084 	/* level at half way mark between min and guranteed */
2085 	boost_level1 = (HWP_GUARANTEED_PERF(hwp_cap) + min_limit) >> 1;
2086 
2087 	if (cpu->hwp_boost_min < boost_level1)
2088 		cpu->hwp_boost_min = boost_level1;
2089 	else if (cpu->hwp_boost_min < HWP_GUARANTEED_PERF(hwp_cap))
2090 		cpu->hwp_boost_min = HWP_GUARANTEED_PERF(hwp_cap);
2091 	else if (cpu->hwp_boost_min == HWP_GUARANTEED_PERF(hwp_cap) &&
2092 		 max_limit != HWP_GUARANTEED_PERF(hwp_cap))
2093 		cpu->hwp_boost_min = max_limit;
2094 	else
2095 		return;
2096 
2097 	hwp_req = (hwp_req & ~GENMASK_ULL(7, 0)) | cpu->hwp_boost_min;
2098 	wrmsrl(MSR_HWP_REQUEST, hwp_req);
2099 	cpu->last_update = cpu->sample.time;
2100 }
2101 
2102 static inline void intel_pstate_hwp_boost_down(struct cpudata *cpu)
2103 {
2104 	if (cpu->hwp_boost_min) {
2105 		bool expired;
2106 
2107 		/* Check if we are idle for hold time to boost down */
2108 		expired = time_after64(cpu->sample.time, cpu->last_update +
2109 				       hwp_boost_hold_time_ns);
2110 		if (expired) {
2111 			wrmsrl(MSR_HWP_REQUEST, cpu->hwp_req_cached);
2112 			cpu->hwp_boost_min = 0;
2113 		}
2114 	}
2115 	cpu->last_update = cpu->sample.time;
2116 }
2117 
2118 static inline void intel_pstate_update_util_hwp_local(struct cpudata *cpu,
2119 						      u64 time)
2120 {
2121 	cpu->sample.time = time;
2122 
2123 	if (cpu->sched_flags & SCHED_CPUFREQ_IOWAIT) {
2124 		bool do_io = false;
2125 
2126 		cpu->sched_flags = 0;
2127 		/*
2128 		 * Set iowait_boost flag and update time. Since IO WAIT flag
2129 		 * is set all the time, we can't just conclude that there is
2130 		 * some IO bound activity is scheduled on this CPU with just
2131 		 * one occurrence. If we receive at least two in two
2132 		 * consecutive ticks, then we treat as boost candidate.
2133 		 */
2134 		if (time_before64(time, cpu->last_io_update + 2 * TICK_NSEC))
2135 			do_io = true;
2136 
2137 		cpu->last_io_update = time;
2138 
2139 		if (do_io)
2140 			intel_pstate_hwp_boost_up(cpu);
2141 
2142 	} else {
2143 		intel_pstate_hwp_boost_down(cpu);
2144 	}
2145 }
2146 
2147 static inline void intel_pstate_update_util_hwp(struct update_util_data *data,
2148 						u64 time, unsigned int flags)
2149 {
2150 	struct cpudata *cpu = container_of(data, struct cpudata, update_util);
2151 
2152 	cpu->sched_flags |= flags;
2153 
2154 	if (smp_processor_id() == cpu->cpu)
2155 		intel_pstate_update_util_hwp_local(cpu, time);
2156 }
2157 
2158 static inline void intel_pstate_calc_avg_perf(struct cpudata *cpu)
2159 {
2160 	struct sample *sample = &cpu->sample;
2161 
2162 	sample->core_avg_perf = div_ext_fp(sample->aperf, sample->mperf);
2163 }
2164 
2165 static inline bool intel_pstate_sample(struct cpudata *cpu, u64 time)
2166 {
2167 	u64 aperf, mperf;
2168 	unsigned long flags;
2169 	u64 tsc;
2170 
2171 	local_irq_save(flags);
2172 	rdmsrl(MSR_IA32_APERF, aperf);
2173 	rdmsrl(MSR_IA32_MPERF, mperf);
2174 	tsc = rdtsc();
2175 	if (cpu->prev_mperf == mperf || cpu->prev_tsc == tsc) {
2176 		local_irq_restore(flags);
2177 		return false;
2178 	}
2179 	local_irq_restore(flags);
2180 
2181 	cpu->last_sample_time = cpu->sample.time;
2182 	cpu->sample.time = time;
2183 	cpu->sample.aperf = aperf;
2184 	cpu->sample.mperf = mperf;
2185 	cpu->sample.tsc =  tsc;
2186 	cpu->sample.aperf -= cpu->prev_aperf;
2187 	cpu->sample.mperf -= cpu->prev_mperf;
2188 	cpu->sample.tsc -= cpu->prev_tsc;
2189 
2190 	cpu->prev_aperf = aperf;
2191 	cpu->prev_mperf = mperf;
2192 	cpu->prev_tsc = tsc;
2193 	/*
2194 	 * First time this function is invoked in a given cycle, all of the
2195 	 * previous sample data fields are equal to zero or stale and they must
2196 	 * be populated with meaningful numbers for things to work, so assume
2197 	 * that sample.time will always be reset before setting the utilization
2198 	 * update hook and make the caller skip the sample then.
2199 	 */
2200 	if (cpu->last_sample_time) {
2201 		intel_pstate_calc_avg_perf(cpu);
2202 		return true;
2203 	}
2204 	return false;
2205 }
2206 
2207 static inline int32_t get_avg_frequency(struct cpudata *cpu)
2208 {
2209 	return mul_ext_fp(cpu->sample.core_avg_perf, cpu_khz);
2210 }
2211 
2212 static inline int32_t get_avg_pstate(struct cpudata *cpu)
2213 {
2214 	return mul_ext_fp(cpu->pstate.max_pstate_physical,
2215 			  cpu->sample.core_avg_perf);
2216 }
2217 
2218 static inline int32_t get_target_pstate(struct cpudata *cpu)
2219 {
2220 	struct sample *sample = &cpu->sample;
2221 	int32_t busy_frac;
2222 	int target, avg_pstate;
2223 
2224 	busy_frac = div_fp(sample->mperf << cpu->aperf_mperf_shift,
2225 			   sample->tsc);
2226 
2227 	if (busy_frac < cpu->iowait_boost)
2228 		busy_frac = cpu->iowait_boost;
2229 
2230 	sample->busy_scaled = busy_frac * 100;
2231 
2232 	target = READ_ONCE(global.no_turbo) ?
2233 			cpu->pstate.max_pstate : cpu->pstate.turbo_pstate;
2234 	target += target >> 2;
2235 	target = mul_fp(target, busy_frac);
2236 	if (target < cpu->pstate.min_pstate)
2237 		target = cpu->pstate.min_pstate;
2238 
2239 	/*
2240 	 * If the average P-state during the previous cycle was higher than the
2241 	 * current target, add 50% of the difference to the target to reduce
2242 	 * possible performance oscillations and offset possible performance
2243 	 * loss related to moving the workload from one CPU to another within
2244 	 * a package/module.
2245 	 */
2246 	avg_pstate = get_avg_pstate(cpu);
2247 	if (avg_pstate > target)
2248 		target += (avg_pstate - target) >> 1;
2249 
2250 	return target;
2251 }
2252 
2253 static int intel_pstate_prepare_request(struct cpudata *cpu, int pstate)
2254 {
2255 	int min_pstate = max(cpu->pstate.min_pstate, cpu->min_perf_ratio);
2256 	int max_pstate = max(min_pstate, cpu->max_perf_ratio);
2257 
2258 	return clamp_t(int, pstate, min_pstate, max_pstate);
2259 }
2260 
2261 static void intel_pstate_update_pstate(struct cpudata *cpu, int pstate)
2262 {
2263 	if (pstate == cpu->pstate.current_pstate)
2264 		return;
2265 
2266 	cpu->pstate.current_pstate = pstate;
2267 	wrmsrl(MSR_IA32_PERF_CTL, pstate_funcs.get_val(cpu, pstate));
2268 }
2269 
2270 static void intel_pstate_adjust_pstate(struct cpudata *cpu)
2271 {
2272 	int from = cpu->pstate.current_pstate;
2273 	struct sample *sample;
2274 	int target_pstate;
2275 
2276 	target_pstate = get_target_pstate(cpu);
2277 	target_pstate = intel_pstate_prepare_request(cpu, target_pstate);
2278 	trace_cpu_frequency(target_pstate * cpu->pstate.scaling, cpu->cpu);
2279 	intel_pstate_update_pstate(cpu, target_pstate);
2280 
2281 	sample = &cpu->sample;
2282 	trace_pstate_sample(mul_ext_fp(100, sample->core_avg_perf),
2283 		fp_toint(sample->busy_scaled),
2284 		from,
2285 		cpu->pstate.current_pstate,
2286 		sample->mperf,
2287 		sample->aperf,
2288 		sample->tsc,
2289 		get_avg_frequency(cpu),
2290 		fp_toint(cpu->iowait_boost * 100));
2291 }
2292 
2293 static void intel_pstate_update_util(struct update_util_data *data, u64 time,
2294 				     unsigned int flags)
2295 {
2296 	struct cpudata *cpu = container_of(data, struct cpudata, update_util);
2297 	u64 delta_ns;
2298 
2299 	/* Don't allow remote callbacks */
2300 	if (smp_processor_id() != cpu->cpu)
2301 		return;
2302 
2303 	delta_ns = time - cpu->last_update;
2304 	if (flags & SCHED_CPUFREQ_IOWAIT) {
2305 		/* Start over if the CPU may have been idle. */
2306 		if (delta_ns > TICK_NSEC) {
2307 			cpu->iowait_boost = ONE_EIGHTH_FP;
2308 		} else if (cpu->iowait_boost >= ONE_EIGHTH_FP) {
2309 			cpu->iowait_boost <<= 1;
2310 			if (cpu->iowait_boost > int_tofp(1))
2311 				cpu->iowait_boost = int_tofp(1);
2312 		} else {
2313 			cpu->iowait_boost = ONE_EIGHTH_FP;
2314 		}
2315 	} else if (cpu->iowait_boost) {
2316 		/* Clear iowait_boost if the CPU may have been idle. */
2317 		if (delta_ns > TICK_NSEC)
2318 			cpu->iowait_boost = 0;
2319 		else
2320 			cpu->iowait_boost >>= 1;
2321 	}
2322 	cpu->last_update = time;
2323 	delta_ns = time - cpu->sample.time;
2324 	if ((s64)delta_ns < INTEL_PSTATE_SAMPLING_INTERVAL)
2325 		return;
2326 
2327 	if (intel_pstate_sample(cpu, time))
2328 		intel_pstate_adjust_pstate(cpu);
2329 }
2330 
2331 static struct pstate_funcs core_funcs = {
2332 	.get_max = core_get_max_pstate,
2333 	.get_max_physical = core_get_max_pstate_physical,
2334 	.get_min = core_get_min_pstate,
2335 	.get_turbo = core_get_turbo_pstate,
2336 	.get_scaling = core_get_scaling,
2337 	.get_val = core_get_val,
2338 };
2339 
2340 static const struct pstate_funcs silvermont_funcs = {
2341 	.get_max = atom_get_max_pstate,
2342 	.get_max_physical = atom_get_max_pstate,
2343 	.get_min = atom_get_min_pstate,
2344 	.get_turbo = atom_get_turbo_pstate,
2345 	.get_val = atom_get_val,
2346 	.get_scaling = silvermont_get_scaling,
2347 	.get_vid = atom_get_vid,
2348 };
2349 
2350 static const struct pstate_funcs airmont_funcs = {
2351 	.get_max = atom_get_max_pstate,
2352 	.get_max_physical = atom_get_max_pstate,
2353 	.get_min = atom_get_min_pstate,
2354 	.get_turbo = atom_get_turbo_pstate,
2355 	.get_val = atom_get_val,
2356 	.get_scaling = airmont_get_scaling,
2357 	.get_vid = atom_get_vid,
2358 };
2359 
2360 static const struct pstate_funcs knl_funcs = {
2361 	.get_max = core_get_max_pstate,
2362 	.get_max_physical = core_get_max_pstate_physical,
2363 	.get_min = core_get_min_pstate,
2364 	.get_turbo = knl_get_turbo_pstate,
2365 	.get_aperf_mperf_shift = knl_get_aperf_mperf_shift,
2366 	.get_scaling = core_get_scaling,
2367 	.get_val = core_get_val,
2368 };
2369 
2370 #define X86_MATCH(model, policy)					 \
2371 	X86_MATCH_VENDOR_FAM_MODEL_FEATURE(INTEL, 6, INTEL_FAM6_##model, \
2372 					   X86_FEATURE_APERFMPERF, &policy)
2373 
2374 static const struct x86_cpu_id intel_pstate_cpu_ids[] = {
2375 	X86_MATCH(SANDYBRIDGE,		core_funcs),
2376 	X86_MATCH(SANDYBRIDGE_X,	core_funcs),
2377 	X86_MATCH(ATOM_SILVERMONT,	silvermont_funcs),
2378 	X86_MATCH(IVYBRIDGE,		core_funcs),
2379 	X86_MATCH(HASWELL,		core_funcs),
2380 	X86_MATCH(BROADWELL,		core_funcs),
2381 	X86_MATCH(IVYBRIDGE_X,		core_funcs),
2382 	X86_MATCH(HASWELL_X,		core_funcs),
2383 	X86_MATCH(HASWELL_L,		core_funcs),
2384 	X86_MATCH(HASWELL_G,		core_funcs),
2385 	X86_MATCH(BROADWELL_G,		core_funcs),
2386 	X86_MATCH(ATOM_AIRMONT,		airmont_funcs),
2387 	X86_MATCH(SKYLAKE_L,		core_funcs),
2388 	X86_MATCH(BROADWELL_X,		core_funcs),
2389 	X86_MATCH(SKYLAKE,		core_funcs),
2390 	X86_MATCH(BROADWELL_D,		core_funcs),
2391 	X86_MATCH(XEON_PHI_KNL,		knl_funcs),
2392 	X86_MATCH(XEON_PHI_KNM,		knl_funcs),
2393 	X86_MATCH(ATOM_GOLDMONT,	core_funcs),
2394 	X86_MATCH(ATOM_GOLDMONT_PLUS,	core_funcs),
2395 	X86_MATCH(SKYLAKE_X,		core_funcs),
2396 	X86_MATCH(COMETLAKE,		core_funcs),
2397 	X86_MATCH(ICELAKE_X,		core_funcs),
2398 	X86_MATCH(TIGERLAKE,		core_funcs),
2399 	X86_MATCH(SAPPHIRERAPIDS_X,	core_funcs),
2400 	X86_MATCH(EMERALDRAPIDS_X,      core_funcs),
2401 	{}
2402 };
2403 MODULE_DEVICE_TABLE(x86cpu, intel_pstate_cpu_ids);
2404 
2405 #ifdef CONFIG_ACPI
2406 static const struct x86_cpu_id intel_pstate_cpu_oob_ids[] __initconst = {
2407 	X86_MATCH(BROADWELL_D,		core_funcs),
2408 	X86_MATCH(BROADWELL_X,		core_funcs),
2409 	X86_MATCH(SKYLAKE_X,		core_funcs),
2410 	X86_MATCH(ICELAKE_X,		core_funcs),
2411 	X86_MATCH(SAPPHIRERAPIDS_X,	core_funcs),
2412 	{}
2413 };
2414 #endif
2415 
2416 static const struct x86_cpu_id intel_pstate_cpu_ee_disable_ids[] = {
2417 	X86_MATCH(KABYLAKE,		core_funcs),
2418 	{}
2419 };
2420 
2421 static int intel_pstate_init_cpu(unsigned int cpunum)
2422 {
2423 	struct cpudata *cpu;
2424 
2425 	cpu = all_cpu_data[cpunum];
2426 
2427 	if (!cpu) {
2428 		cpu = kzalloc(sizeof(*cpu), GFP_KERNEL);
2429 		if (!cpu)
2430 			return -ENOMEM;
2431 
2432 		WRITE_ONCE(all_cpu_data[cpunum], cpu);
2433 
2434 		cpu->cpu = cpunum;
2435 
2436 		cpu->epp_default = -EINVAL;
2437 
2438 		if (hwp_active) {
2439 			intel_pstate_hwp_enable(cpu);
2440 
2441 			if (intel_pstate_acpi_pm_profile_server())
2442 				hwp_boost = true;
2443 		}
2444 	} else if (hwp_active) {
2445 		/*
2446 		 * Re-enable HWP in case this happens after a resume from ACPI
2447 		 * S3 if the CPU was offline during the whole system/resume
2448 		 * cycle.
2449 		 */
2450 		intel_pstate_hwp_reenable(cpu);
2451 	}
2452 
2453 	cpu->epp_powersave = -EINVAL;
2454 	cpu->epp_policy = 0;
2455 
2456 	intel_pstate_get_cpu_pstates(cpu);
2457 
2458 	pr_debug("controlling: cpu %d\n", cpunum);
2459 
2460 	return 0;
2461 }
2462 
2463 static void intel_pstate_set_update_util_hook(unsigned int cpu_num)
2464 {
2465 	struct cpudata *cpu = all_cpu_data[cpu_num];
2466 
2467 	if (hwp_active && !hwp_boost)
2468 		return;
2469 
2470 	if (cpu->update_util_set)
2471 		return;
2472 
2473 	/* Prevent intel_pstate_update_util() from using stale data. */
2474 	cpu->sample.time = 0;
2475 	cpufreq_add_update_util_hook(cpu_num, &cpu->update_util,
2476 				     (hwp_active ?
2477 				      intel_pstate_update_util_hwp :
2478 				      intel_pstate_update_util));
2479 	cpu->update_util_set = true;
2480 }
2481 
2482 static void intel_pstate_clear_update_util_hook(unsigned int cpu)
2483 {
2484 	struct cpudata *cpu_data = all_cpu_data[cpu];
2485 
2486 	if (!cpu_data->update_util_set)
2487 		return;
2488 
2489 	cpufreq_remove_update_util_hook(cpu);
2490 	cpu_data->update_util_set = false;
2491 	synchronize_rcu();
2492 }
2493 
2494 static int intel_pstate_get_max_freq(struct cpudata *cpu)
2495 {
2496 	return READ_ONCE(global.no_turbo) ?
2497 			cpu->pstate.max_freq : cpu->pstate.turbo_freq;
2498 }
2499 
2500 static void intel_pstate_update_perf_limits(struct cpudata *cpu,
2501 					    unsigned int policy_min,
2502 					    unsigned int policy_max)
2503 {
2504 	int perf_ctl_scaling = cpu->pstate.perf_ctl_scaling;
2505 	int32_t max_policy_perf, min_policy_perf;
2506 
2507 	max_policy_perf = policy_max / perf_ctl_scaling;
2508 	if (policy_max == policy_min) {
2509 		min_policy_perf = max_policy_perf;
2510 	} else {
2511 		min_policy_perf = policy_min / perf_ctl_scaling;
2512 		min_policy_perf = clamp_t(int32_t, min_policy_perf,
2513 					  0, max_policy_perf);
2514 	}
2515 
2516 	/*
2517 	 * HWP needs some special consideration, because HWP_REQUEST uses
2518 	 * abstract values to represent performance rather than pure ratios.
2519 	 */
2520 	if (hwp_active && cpu->pstate.scaling != perf_ctl_scaling) {
2521 		int freq;
2522 
2523 		freq = max_policy_perf * perf_ctl_scaling;
2524 		max_policy_perf = intel_pstate_freq_to_hwp(cpu, freq);
2525 		freq = min_policy_perf * perf_ctl_scaling;
2526 		min_policy_perf = intel_pstate_freq_to_hwp(cpu, freq);
2527 	}
2528 
2529 	pr_debug("cpu:%d min_policy_perf:%d max_policy_perf:%d\n",
2530 		 cpu->cpu, min_policy_perf, max_policy_perf);
2531 
2532 	/* Normalize user input to [min_perf, max_perf] */
2533 	if (per_cpu_limits) {
2534 		cpu->min_perf_ratio = min_policy_perf;
2535 		cpu->max_perf_ratio = max_policy_perf;
2536 	} else {
2537 		int turbo_max = cpu->pstate.turbo_pstate;
2538 		int32_t global_min, global_max;
2539 
2540 		/* Global limits are in percent of the maximum turbo P-state. */
2541 		global_max = DIV_ROUND_UP(turbo_max * global.max_perf_pct, 100);
2542 		global_min = DIV_ROUND_UP(turbo_max * global.min_perf_pct, 100);
2543 		global_min = clamp_t(int32_t, global_min, 0, global_max);
2544 
2545 		pr_debug("cpu:%d global_min:%d global_max:%d\n", cpu->cpu,
2546 			 global_min, global_max);
2547 
2548 		cpu->min_perf_ratio = max(min_policy_perf, global_min);
2549 		cpu->min_perf_ratio = min(cpu->min_perf_ratio, max_policy_perf);
2550 		cpu->max_perf_ratio = min(max_policy_perf, global_max);
2551 		cpu->max_perf_ratio = max(min_policy_perf, cpu->max_perf_ratio);
2552 
2553 		/* Make sure min_perf <= max_perf */
2554 		cpu->min_perf_ratio = min(cpu->min_perf_ratio,
2555 					  cpu->max_perf_ratio);
2556 
2557 	}
2558 	pr_debug("cpu:%d max_perf_ratio:%d min_perf_ratio:%d\n", cpu->cpu,
2559 		 cpu->max_perf_ratio,
2560 		 cpu->min_perf_ratio);
2561 }
2562 
2563 static int intel_pstate_set_policy(struct cpufreq_policy *policy)
2564 {
2565 	struct cpudata *cpu;
2566 
2567 	if (!policy->cpuinfo.max_freq)
2568 		return -ENODEV;
2569 
2570 	pr_debug("set_policy cpuinfo.max %u policy->max %u\n",
2571 		 policy->cpuinfo.max_freq, policy->max);
2572 
2573 	cpu = all_cpu_data[policy->cpu];
2574 	cpu->policy = policy->policy;
2575 
2576 	mutex_lock(&intel_pstate_limits_lock);
2577 
2578 	intel_pstate_update_perf_limits(cpu, policy->min, policy->max);
2579 
2580 	if (cpu->policy == CPUFREQ_POLICY_PERFORMANCE) {
2581 		int pstate = max(cpu->pstate.min_pstate, cpu->max_perf_ratio);
2582 
2583 		/*
2584 		 * NOHZ_FULL CPUs need this as the governor callback may not
2585 		 * be invoked on them.
2586 		 */
2587 		intel_pstate_clear_update_util_hook(policy->cpu);
2588 		intel_pstate_set_pstate(cpu, pstate);
2589 	} else {
2590 		intel_pstate_set_update_util_hook(policy->cpu);
2591 	}
2592 
2593 	if (hwp_active) {
2594 		/*
2595 		 * When hwp_boost was active before and dynamically it
2596 		 * was turned off, in that case we need to clear the
2597 		 * update util hook.
2598 		 */
2599 		if (!hwp_boost)
2600 			intel_pstate_clear_update_util_hook(policy->cpu);
2601 		intel_pstate_hwp_set(policy->cpu);
2602 	}
2603 	/*
2604 	 * policy->cur is never updated with the intel_pstate driver, but it
2605 	 * is used as a stale frequency value. So, keep it within limits.
2606 	 */
2607 	policy->cur = policy->min;
2608 
2609 	mutex_unlock(&intel_pstate_limits_lock);
2610 
2611 	return 0;
2612 }
2613 
2614 static void intel_pstate_adjust_policy_max(struct cpudata *cpu,
2615 					   struct cpufreq_policy_data *policy)
2616 {
2617 	if (!hwp_active &&
2618 	    cpu->pstate.max_pstate_physical > cpu->pstate.max_pstate &&
2619 	    policy->max < policy->cpuinfo.max_freq &&
2620 	    policy->max > cpu->pstate.max_freq) {
2621 		pr_debug("policy->max > max non turbo frequency\n");
2622 		policy->max = policy->cpuinfo.max_freq;
2623 	}
2624 }
2625 
2626 static void intel_pstate_verify_cpu_policy(struct cpudata *cpu,
2627 					   struct cpufreq_policy_data *policy)
2628 {
2629 	int max_freq;
2630 
2631 	if (hwp_active) {
2632 		intel_pstate_get_hwp_cap(cpu);
2633 		max_freq = READ_ONCE(global.no_turbo) ?
2634 				cpu->pstate.max_freq : cpu->pstate.turbo_freq;
2635 	} else {
2636 		max_freq = intel_pstate_get_max_freq(cpu);
2637 	}
2638 	cpufreq_verify_within_limits(policy, policy->cpuinfo.min_freq, max_freq);
2639 
2640 	intel_pstate_adjust_policy_max(cpu, policy);
2641 }
2642 
2643 static int intel_pstate_verify_policy(struct cpufreq_policy_data *policy)
2644 {
2645 	intel_pstate_verify_cpu_policy(all_cpu_data[policy->cpu], policy);
2646 
2647 	return 0;
2648 }
2649 
2650 static int intel_cpufreq_cpu_offline(struct cpufreq_policy *policy)
2651 {
2652 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2653 
2654 	pr_debug("CPU %d going offline\n", cpu->cpu);
2655 
2656 	if (cpu->suspended)
2657 		return 0;
2658 
2659 	/*
2660 	 * If the CPU is an SMT thread and it goes offline with the performance
2661 	 * settings different from the minimum, it will prevent its sibling
2662 	 * from getting to lower performance levels, so force the minimum
2663 	 * performance on CPU offline to prevent that from happening.
2664 	 */
2665 	if (hwp_active)
2666 		intel_pstate_hwp_offline(cpu);
2667 	else
2668 		intel_pstate_set_min_pstate(cpu);
2669 
2670 	intel_pstate_exit_perf_limits(policy);
2671 
2672 	return 0;
2673 }
2674 
2675 static int intel_pstate_cpu_online(struct cpufreq_policy *policy)
2676 {
2677 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2678 
2679 	pr_debug("CPU %d going online\n", cpu->cpu);
2680 
2681 	intel_pstate_init_acpi_perf_limits(policy);
2682 
2683 	if (hwp_active) {
2684 		/*
2685 		 * Re-enable HWP and clear the "suspended" flag to let "resume"
2686 		 * know that it need not do that.
2687 		 */
2688 		intel_pstate_hwp_reenable(cpu);
2689 		cpu->suspended = false;
2690 	}
2691 
2692 	return 0;
2693 }
2694 
2695 static int intel_pstate_cpu_offline(struct cpufreq_policy *policy)
2696 {
2697 	intel_pstate_clear_update_util_hook(policy->cpu);
2698 
2699 	return intel_cpufreq_cpu_offline(policy);
2700 }
2701 
2702 static int intel_pstate_cpu_exit(struct cpufreq_policy *policy)
2703 {
2704 	pr_debug("CPU %d exiting\n", policy->cpu);
2705 
2706 	policy->fast_switch_possible = false;
2707 
2708 	return 0;
2709 }
2710 
2711 static int __intel_pstate_cpu_init(struct cpufreq_policy *policy)
2712 {
2713 	struct cpudata *cpu;
2714 	int rc;
2715 
2716 	rc = intel_pstate_init_cpu(policy->cpu);
2717 	if (rc)
2718 		return rc;
2719 
2720 	cpu = all_cpu_data[policy->cpu];
2721 
2722 	cpu->max_perf_ratio = 0xFF;
2723 	cpu->min_perf_ratio = 0;
2724 
2725 	/* cpuinfo and default policy values */
2726 	policy->cpuinfo.min_freq = cpu->pstate.min_freq;
2727 	policy->cpuinfo.max_freq = READ_ONCE(global.no_turbo) ?
2728 			cpu->pstate.max_freq : cpu->pstate.turbo_freq;
2729 
2730 	policy->min = policy->cpuinfo.min_freq;
2731 	policy->max = policy->cpuinfo.max_freq;
2732 
2733 	intel_pstate_init_acpi_perf_limits(policy);
2734 
2735 	policy->fast_switch_possible = true;
2736 
2737 	return 0;
2738 }
2739 
2740 static int intel_pstate_cpu_init(struct cpufreq_policy *policy)
2741 {
2742 	int ret = __intel_pstate_cpu_init(policy);
2743 
2744 	if (ret)
2745 		return ret;
2746 
2747 	/*
2748 	 * Set the policy to powersave to provide a valid fallback value in case
2749 	 * the default cpufreq governor is neither powersave nor performance.
2750 	 */
2751 	policy->policy = CPUFREQ_POLICY_POWERSAVE;
2752 
2753 	if (hwp_active) {
2754 		struct cpudata *cpu = all_cpu_data[policy->cpu];
2755 
2756 		cpu->epp_cached = intel_pstate_get_epp(cpu, 0);
2757 	}
2758 
2759 	return 0;
2760 }
2761 
2762 static struct cpufreq_driver intel_pstate = {
2763 	.flags		= CPUFREQ_CONST_LOOPS,
2764 	.verify		= intel_pstate_verify_policy,
2765 	.setpolicy	= intel_pstate_set_policy,
2766 	.suspend	= intel_pstate_suspend,
2767 	.resume		= intel_pstate_resume,
2768 	.init		= intel_pstate_cpu_init,
2769 	.exit		= intel_pstate_cpu_exit,
2770 	.offline	= intel_pstate_cpu_offline,
2771 	.online		= intel_pstate_cpu_online,
2772 	.update_limits	= intel_pstate_update_limits,
2773 	.name		= "intel_pstate",
2774 };
2775 
2776 static int intel_cpufreq_verify_policy(struct cpufreq_policy_data *policy)
2777 {
2778 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2779 
2780 	intel_pstate_verify_cpu_policy(cpu, policy);
2781 	intel_pstate_update_perf_limits(cpu, policy->min, policy->max);
2782 
2783 	return 0;
2784 }
2785 
2786 /* Use of trace in passive mode:
2787  *
2788  * In passive mode the trace core_busy field (also known as the
2789  * performance field, and lablelled as such on the graphs; also known as
2790  * core_avg_perf) is not needed and so is re-assigned to indicate if the
2791  * driver call was via the normal or fast switch path. Various graphs
2792  * output from the intel_pstate_tracer.py utility that include core_busy
2793  * (or performance or core_avg_perf) have a fixed y-axis from 0 to 100%,
2794  * so we use 10 to indicate the normal path through the driver, and
2795  * 90 to indicate the fast switch path through the driver.
2796  * The scaled_busy field is not used, and is set to 0.
2797  */
2798 
2799 #define	INTEL_PSTATE_TRACE_TARGET 10
2800 #define	INTEL_PSTATE_TRACE_FAST_SWITCH 90
2801 
2802 static void intel_cpufreq_trace(struct cpudata *cpu, unsigned int trace_type, int old_pstate)
2803 {
2804 	struct sample *sample;
2805 
2806 	if (!trace_pstate_sample_enabled())
2807 		return;
2808 
2809 	if (!intel_pstate_sample(cpu, ktime_get()))
2810 		return;
2811 
2812 	sample = &cpu->sample;
2813 	trace_pstate_sample(trace_type,
2814 		0,
2815 		old_pstate,
2816 		cpu->pstate.current_pstate,
2817 		sample->mperf,
2818 		sample->aperf,
2819 		sample->tsc,
2820 		get_avg_frequency(cpu),
2821 		fp_toint(cpu->iowait_boost * 100));
2822 }
2823 
2824 static void intel_cpufreq_hwp_update(struct cpudata *cpu, u32 min, u32 max,
2825 				     u32 desired, bool fast_switch)
2826 {
2827 	u64 prev = READ_ONCE(cpu->hwp_req_cached), value = prev;
2828 
2829 	value &= ~HWP_MIN_PERF(~0L);
2830 	value |= HWP_MIN_PERF(min);
2831 
2832 	value &= ~HWP_MAX_PERF(~0L);
2833 	value |= HWP_MAX_PERF(max);
2834 
2835 	value &= ~HWP_DESIRED_PERF(~0L);
2836 	value |= HWP_DESIRED_PERF(desired);
2837 
2838 	if (value == prev)
2839 		return;
2840 
2841 	WRITE_ONCE(cpu->hwp_req_cached, value);
2842 	if (fast_switch)
2843 		wrmsrl(MSR_HWP_REQUEST, value);
2844 	else
2845 		wrmsrl_on_cpu(cpu->cpu, MSR_HWP_REQUEST, value);
2846 }
2847 
2848 static void intel_cpufreq_perf_ctl_update(struct cpudata *cpu,
2849 					  u32 target_pstate, bool fast_switch)
2850 {
2851 	if (fast_switch)
2852 		wrmsrl(MSR_IA32_PERF_CTL,
2853 		       pstate_funcs.get_val(cpu, target_pstate));
2854 	else
2855 		wrmsrl_on_cpu(cpu->cpu, MSR_IA32_PERF_CTL,
2856 			      pstate_funcs.get_val(cpu, target_pstate));
2857 }
2858 
2859 static int intel_cpufreq_update_pstate(struct cpufreq_policy *policy,
2860 				       int target_pstate, bool fast_switch)
2861 {
2862 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2863 	int old_pstate = cpu->pstate.current_pstate;
2864 
2865 	target_pstate = intel_pstate_prepare_request(cpu, target_pstate);
2866 	if (hwp_active) {
2867 		int max_pstate = policy->strict_target ?
2868 					target_pstate : cpu->max_perf_ratio;
2869 
2870 		intel_cpufreq_hwp_update(cpu, target_pstate, max_pstate, 0,
2871 					 fast_switch);
2872 	} else if (target_pstate != old_pstate) {
2873 		intel_cpufreq_perf_ctl_update(cpu, target_pstate, fast_switch);
2874 	}
2875 
2876 	cpu->pstate.current_pstate = target_pstate;
2877 
2878 	intel_cpufreq_trace(cpu, fast_switch ? INTEL_PSTATE_TRACE_FAST_SWITCH :
2879 			    INTEL_PSTATE_TRACE_TARGET, old_pstate);
2880 
2881 	return target_pstate;
2882 }
2883 
2884 static int intel_cpufreq_target(struct cpufreq_policy *policy,
2885 				unsigned int target_freq,
2886 				unsigned int relation)
2887 {
2888 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2889 	struct cpufreq_freqs freqs;
2890 	int target_pstate;
2891 
2892 	freqs.old = policy->cur;
2893 	freqs.new = target_freq;
2894 
2895 	cpufreq_freq_transition_begin(policy, &freqs);
2896 
2897 	target_pstate = intel_pstate_freq_to_hwp_rel(cpu, freqs.new, relation);
2898 	target_pstate = intel_cpufreq_update_pstate(policy, target_pstate, false);
2899 
2900 	freqs.new = target_pstate * cpu->pstate.scaling;
2901 
2902 	cpufreq_freq_transition_end(policy, &freqs, false);
2903 
2904 	return 0;
2905 }
2906 
2907 static unsigned int intel_cpufreq_fast_switch(struct cpufreq_policy *policy,
2908 					      unsigned int target_freq)
2909 {
2910 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2911 	int target_pstate;
2912 
2913 	target_pstate = intel_pstate_freq_to_hwp(cpu, target_freq);
2914 
2915 	target_pstate = intel_cpufreq_update_pstate(policy, target_pstate, true);
2916 
2917 	return target_pstate * cpu->pstate.scaling;
2918 }
2919 
2920 static void intel_cpufreq_adjust_perf(unsigned int cpunum,
2921 				      unsigned long min_perf,
2922 				      unsigned long target_perf,
2923 				      unsigned long capacity)
2924 {
2925 	struct cpudata *cpu = all_cpu_data[cpunum];
2926 	u64 hwp_cap = READ_ONCE(cpu->hwp_cap_cached);
2927 	int old_pstate = cpu->pstate.current_pstate;
2928 	int cap_pstate, min_pstate, max_pstate, target_pstate;
2929 
2930 	cap_pstate = READ_ONCE(global.no_turbo) ?
2931 					HWP_GUARANTEED_PERF(hwp_cap) :
2932 					HWP_HIGHEST_PERF(hwp_cap);
2933 
2934 	/* Optimization: Avoid unnecessary divisions. */
2935 
2936 	target_pstate = cap_pstate;
2937 	if (target_perf < capacity)
2938 		target_pstate = DIV_ROUND_UP(cap_pstate * target_perf, capacity);
2939 
2940 	min_pstate = cap_pstate;
2941 	if (min_perf < capacity)
2942 		min_pstate = DIV_ROUND_UP(cap_pstate * min_perf, capacity);
2943 
2944 	if (min_pstate < cpu->pstate.min_pstate)
2945 		min_pstate = cpu->pstate.min_pstate;
2946 
2947 	if (min_pstate < cpu->min_perf_ratio)
2948 		min_pstate = cpu->min_perf_ratio;
2949 
2950 	if (min_pstate > cpu->max_perf_ratio)
2951 		min_pstate = cpu->max_perf_ratio;
2952 
2953 	max_pstate = min(cap_pstate, cpu->max_perf_ratio);
2954 	if (max_pstate < min_pstate)
2955 		max_pstate = min_pstate;
2956 
2957 	target_pstate = clamp_t(int, target_pstate, min_pstate, max_pstate);
2958 
2959 	intel_cpufreq_hwp_update(cpu, min_pstate, max_pstate, target_pstate, true);
2960 
2961 	cpu->pstate.current_pstate = target_pstate;
2962 	intel_cpufreq_trace(cpu, INTEL_PSTATE_TRACE_FAST_SWITCH, old_pstate);
2963 }
2964 
2965 static int intel_cpufreq_cpu_init(struct cpufreq_policy *policy)
2966 {
2967 	struct freq_qos_request *req;
2968 	struct cpudata *cpu;
2969 	struct device *dev;
2970 	int ret, freq;
2971 
2972 	dev = get_cpu_device(policy->cpu);
2973 	if (!dev)
2974 		return -ENODEV;
2975 
2976 	ret = __intel_pstate_cpu_init(policy);
2977 	if (ret)
2978 		return ret;
2979 
2980 	policy->cpuinfo.transition_latency = INTEL_CPUFREQ_TRANSITION_LATENCY;
2981 	/* This reflects the intel_pstate_get_cpu_pstates() setting. */
2982 	policy->cur = policy->cpuinfo.min_freq;
2983 
2984 	req = kcalloc(2, sizeof(*req), GFP_KERNEL);
2985 	if (!req) {
2986 		ret = -ENOMEM;
2987 		goto pstate_exit;
2988 	}
2989 
2990 	cpu = all_cpu_data[policy->cpu];
2991 
2992 	if (hwp_active) {
2993 		u64 value;
2994 
2995 		policy->transition_delay_us = INTEL_CPUFREQ_TRANSITION_DELAY_HWP;
2996 
2997 		intel_pstate_get_hwp_cap(cpu);
2998 
2999 		rdmsrl_on_cpu(cpu->cpu, MSR_HWP_REQUEST, &value);
3000 		WRITE_ONCE(cpu->hwp_req_cached, value);
3001 
3002 		cpu->epp_cached = intel_pstate_get_epp(cpu, value);
3003 	} else {
3004 		policy->transition_delay_us = INTEL_CPUFREQ_TRANSITION_DELAY;
3005 	}
3006 
3007 	freq = DIV_ROUND_UP(cpu->pstate.turbo_freq * global.min_perf_pct, 100);
3008 
3009 	ret = freq_qos_add_request(&policy->constraints, req, FREQ_QOS_MIN,
3010 				   freq);
3011 	if (ret < 0) {
3012 		dev_err(dev, "Failed to add min-freq constraint (%d)\n", ret);
3013 		goto free_req;
3014 	}
3015 
3016 	freq = DIV_ROUND_UP(cpu->pstate.turbo_freq * global.max_perf_pct, 100);
3017 
3018 	ret = freq_qos_add_request(&policy->constraints, req + 1, FREQ_QOS_MAX,
3019 				   freq);
3020 	if (ret < 0) {
3021 		dev_err(dev, "Failed to add max-freq constraint (%d)\n", ret);
3022 		goto remove_min_req;
3023 	}
3024 
3025 	policy->driver_data = req;
3026 
3027 	return 0;
3028 
3029 remove_min_req:
3030 	freq_qos_remove_request(req);
3031 free_req:
3032 	kfree(req);
3033 pstate_exit:
3034 	intel_pstate_exit_perf_limits(policy);
3035 
3036 	return ret;
3037 }
3038 
3039 static int intel_cpufreq_cpu_exit(struct cpufreq_policy *policy)
3040 {
3041 	struct freq_qos_request *req;
3042 
3043 	req = policy->driver_data;
3044 
3045 	freq_qos_remove_request(req + 1);
3046 	freq_qos_remove_request(req);
3047 	kfree(req);
3048 
3049 	return intel_pstate_cpu_exit(policy);
3050 }
3051 
3052 static int intel_cpufreq_suspend(struct cpufreq_policy *policy)
3053 {
3054 	intel_pstate_suspend(policy);
3055 
3056 	if (hwp_active) {
3057 		struct cpudata *cpu = all_cpu_data[policy->cpu];
3058 		u64 value = READ_ONCE(cpu->hwp_req_cached);
3059 
3060 		/*
3061 		 * Clear the desired perf field in MSR_HWP_REQUEST in case
3062 		 * intel_cpufreq_adjust_perf() is in use and the last value
3063 		 * written by it may not be suitable.
3064 		 */
3065 		value &= ~HWP_DESIRED_PERF(~0L);
3066 		wrmsrl_on_cpu(cpu->cpu, MSR_HWP_REQUEST, value);
3067 		WRITE_ONCE(cpu->hwp_req_cached, value);
3068 	}
3069 
3070 	return 0;
3071 }
3072 
3073 static struct cpufreq_driver intel_cpufreq = {
3074 	.flags		= CPUFREQ_CONST_LOOPS,
3075 	.verify		= intel_cpufreq_verify_policy,
3076 	.target		= intel_cpufreq_target,
3077 	.fast_switch	= intel_cpufreq_fast_switch,
3078 	.init		= intel_cpufreq_cpu_init,
3079 	.exit		= intel_cpufreq_cpu_exit,
3080 	.offline	= intel_cpufreq_cpu_offline,
3081 	.online		= intel_pstate_cpu_online,
3082 	.suspend	= intel_cpufreq_suspend,
3083 	.resume		= intel_pstate_resume,
3084 	.update_limits	= intel_pstate_update_limits,
3085 	.name		= "intel_cpufreq",
3086 };
3087 
3088 static struct cpufreq_driver *default_driver;
3089 
3090 static void intel_pstate_driver_cleanup(void)
3091 {
3092 	unsigned int cpu;
3093 
3094 	cpus_read_lock();
3095 	for_each_online_cpu(cpu) {
3096 		if (all_cpu_data[cpu]) {
3097 			if (intel_pstate_driver == &intel_pstate)
3098 				intel_pstate_clear_update_util_hook(cpu);
3099 
3100 			kfree(all_cpu_data[cpu]);
3101 			WRITE_ONCE(all_cpu_data[cpu], NULL);
3102 		}
3103 	}
3104 	cpus_read_unlock();
3105 
3106 	intel_pstate_driver = NULL;
3107 }
3108 
3109 static int intel_pstate_register_driver(struct cpufreq_driver *driver)
3110 {
3111 	int ret;
3112 
3113 	if (driver == &intel_pstate)
3114 		intel_pstate_sysfs_expose_hwp_dynamic_boost();
3115 
3116 	memset(&global, 0, sizeof(global));
3117 	global.max_perf_pct = 100;
3118 	global.turbo_disabled = turbo_is_disabled();
3119 	global.no_turbo = global.turbo_disabled;
3120 
3121 	arch_set_max_freq_ratio(global.turbo_disabled);
3122 
3123 	intel_pstate_driver = driver;
3124 	ret = cpufreq_register_driver(intel_pstate_driver);
3125 	if (ret) {
3126 		intel_pstate_driver_cleanup();
3127 		return ret;
3128 	}
3129 
3130 	global.min_perf_pct = min_perf_pct_min();
3131 
3132 	return 0;
3133 }
3134 
3135 static ssize_t intel_pstate_show_status(char *buf)
3136 {
3137 	if (!intel_pstate_driver)
3138 		return sprintf(buf, "off\n");
3139 
3140 	return sprintf(buf, "%s\n", intel_pstate_driver == &intel_pstate ?
3141 					"active" : "passive");
3142 }
3143 
3144 static int intel_pstate_update_status(const char *buf, size_t size)
3145 {
3146 	if (size == 3 && !strncmp(buf, "off", size)) {
3147 		if (!intel_pstate_driver)
3148 			return -EINVAL;
3149 
3150 		if (hwp_active)
3151 			return -EBUSY;
3152 
3153 		cpufreq_unregister_driver(intel_pstate_driver);
3154 		intel_pstate_driver_cleanup();
3155 		return 0;
3156 	}
3157 
3158 	if (size == 6 && !strncmp(buf, "active", size)) {
3159 		if (intel_pstate_driver) {
3160 			if (intel_pstate_driver == &intel_pstate)
3161 				return 0;
3162 
3163 			cpufreq_unregister_driver(intel_pstate_driver);
3164 		}
3165 
3166 		return intel_pstate_register_driver(&intel_pstate);
3167 	}
3168 
3169 	if (size == 7 && !strncmp(buf, "passive", size)) {
3170 		if (intel_pstate_driver) {
3171 			if (intel_pstate_driver == &intel_cpufreq)
3172 				return 0;
3173 
3174 			cpufreq_unregister_driver(intel_pstate_driver);
3175 			intel_pstate_sysfs_hide_hwp_dynamic_boost();
3176 		}
3177 
3178 		return intel_pstate_register_driver(&intel_cpufreq);
3179 	}
3180 
3181 	return -EINVAL;
3182 }
3183 
3184 static int no_load __initdata;
3185 static int no_hwp __initdata;
3186 static int hwp_only __initdata;
3187 static unsigned int force_load __initdata;
3188 
3189 static int __init intel_pstate_msrs_not_valid(void)
3190 {
3191 	if (!pstate_funcs.get_max(0) ||
3192 	    !pstate_funcs.get_min(0) ||
3193 	    !pstate_funcs.get_turbo(0))
3194 		return -ENODEV;
3195 
3196 	return 0;
3197 }
3198 
3199 static void __init copy_cpu_funcs(struct pstate_funcs *funcs)
3200 {
3201 	pstate_funcs.get_max   = funcs->get_max;
3202 	pstate_funcs.get_max_physical = funcs->get_max_physical;
3203 	pstate_funcs.get_min   = funcs->get_min;
3204 	pstate_funcs.get_turbo = funcs->get_turbo;
3205 	pstate_funcs.get_scaling = funcs->get_scaling;
3206 	pstate_funcs.get_val   = funcs->get_val;
3207 	pstate_funcs.get_vid   = funcs->get_vid;
3208 	pstate_funcs.get_aperf_mperf_shift = funcs->get_aperf_mperf_shift;
3209 }
3210 
3211 #ifdef CONFIG_ACPI
3212 
3213 static bool __init intel_pstate_no_acpi_pss(void)
3214 {
3215 	int i;
3216 
3217 	for_each_possible_cpu(i) {
3218 		acpi_status status;
3219 		union acpi_object *pss;
3220 		struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
3221 		struct acpi_processor *pr = per_cpu(processors, i);
3222 
3223 		if (!pr)
3224 			continue;
3225 
3226 		status = acpi_evaluate_object(pr->handle, "_PSS", NULL, &buffer);
3227 		if (ACPI_FAILURE(status))
3228 			continue;
3229 
3230 		pss = buffer.pointer;
3231 		if (pss && pss->type == ACPI_TYPE_PACKAGE) {
3232 			kfree(pss);
3233 			return false;
3234 		}
3235 
3236 		kfree(pss);
3237 	}
3238 
3239 	pr_debug("ACPI _PSS not found\n");
3240 	return true;
3241 }
3242 
3243 static bool __init intel_pstate_no_acpi_pcch(void)
3244 {
3245 	acpi_status status;
3246 	acpi_handle handle;
3247 
3248 	status = acpi_get_handle(NULL, "\\_SB", &handle);
3249 	if (ACPI_FAILURE(status))
3250 		goto not_found;
3251 
3252 	if (acpi_has_method(handle, "PCCH"))
3253 		return false;
3254 
3255 not_found:
3256 	pr_debug("ACPI PCCH not found\n");
3257 	return true;
3258 }
3259 
3260 static bool __init intel_pstate_has_acpi_ppc(void)
3261 {
3262 	int i;
3263 
3264 	for_each_possible_cpu(i) {
3265 		struct acpi_processor *pr = per_cpu(processors, i);
3266 
3267 		if (!pr)
3268 			continue;
3269 		if (acpi_has_method(pr->handle, "_PPC"))
3270 			return true;
3271 	}
3272 	pr_debug("ACPI _PPC not found\n");
3273 	return false;
3274 }
3275 
3276 enum {
3277 	PSS,
3278 	PPC,
3279 };
3280 
3281 /* Hardware vendor-specific info that has its own power management modes */
3282 static struct acpi_platform_list plat_info[] __initdata = {
3283 	{"HP    ", "ProLiant", 0, ACPI_SIG_FADT, all_versions, NULL, PSS},
3284 	{"ORACLE", "X4-2    ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3285 	{"ORACLE", "X4-2L   ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3286 	{"ORACLE", "X4-2B   ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3287 	{"ORACLE", "X3-2    ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3288 	{"ORACLE", "X3-2L   ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3289 	{"ORACLE", "X3-2B   ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3290 	{"ORACLE", "X4470M2 ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3291 	{"ORACLE", "X4270M3 ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3292 	{"ORACLE", "X4270M2 ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3293 	{"ORACLE", "X4170M2 ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3294 	{"ORACLE", "X4170 M3", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3295 	{"ORACLE", "X4275 M3", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3296 	{"ORACLE", "X6-2    ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3297 	{"ORACLE", "Sudbury ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3298 	{ } /* End */
3299 };
3300 
3301 #define BITMASK_OOB	(BIT(8) | BIT(18))
3302 
3303 static bool __init intel_pstate_platform_pwr_mgmt_exists(void)
3304 {
3305 	const struct x86_cpu_id *id;
3306 	u64 misc_pwr;
3307 	int idx;
3308 
3309 	id = x86_match_cpu(intel_pstate_cpu_oob_ids);
3310 	if (id) {
3311 		rdmsrl(MSR_MISC_PWR_MGMT, misc_pwr);
3312 		if (misc_pwr & BITMASK_OOB) {
3313 			pr_debug("Bit 8 or 18 in the MISC_PWR_MGMT MSR set\n");
3314 			pr_debug("P states are controlled in Out of Band mode by the firmware/hardware\n");
3315 			return true;
3316 		}
3317 	}
3318 
3319 	idx = acpi_match_platform_list(plat_info);
3320 	if (idx < 0)
3321 		return false;
3322 
3323 	switch (plat_info[idx].data) {
3324 	case PSS:
3325 		if (!intel_pstate_no_acpi_pss())
3326 			return false;
3327 
3328 		return intel_pstate_no_acpi_pcch();
3329 	case PPC:
3330 		return intel_pstate_has_acpi_ppc() && !force_load;
3331 	}
3332 
3333 	return false;
3334 }
3335 
3336 static void intel_pstate_request_control_from_smm(void)
3337 {
3338 	/*
3339 	 * It may be unsafe to request P-states control from SMM if _PPC support
3340 	 * has not been enabled.
3341 	 */
3342 	if (acpi_ppc)
3343 		acpi_processor_pstate_control();
3344 }
3345 #else /* CONFIG_ACPI not enabled */
3346 static inline bool intel_pstate_platform_pwr_mgmt_exists(void) { return false; }
3347 static inline bool intel_pstate_has_acpi_ppc(void) { return false; }
3348 static inline void intel_pstate_request_control_from_smm(void) {}
3349 #endif /* CONFIG_ACPI */
3350 
3351 #define INTEL_PSTATE_HWP_BROADWELL	0x01
3352 
3353 #define X86_MATCH_HWP(model, hwp_mode)					\
3354 	X86_MATCH_VENDOR_FAM_MODEL_FEATURE(INTEL, 6, INTEL_FAM6_##model, \
3355 					   X86_FEATURE_HWP, hwp_mode)
3356 
3357 static const struct x86_cpu_id hwp_support_ids[] __initconst = {
3358 	X86_MATCH_HWP(BROADWELL_X,	INTEL_PSTATE_HWP_BROADWELL),
3359 	X86_MATCH_HWP(BROADWELL_D,	INTEL_PSTATE_HWP_BROADWELL),
3360 	X86_MATCH_HWP(ANY,		0),
3361 	{}
3362 };
3363 
3364 static bool intel_pstate_hwp_is_enabled(void)
3365 {
3366 	u64 value;
3367 
3368 	rdmsrl(MSR_PM_ENABLE, value);
3369 	return !!(value & 0x1);
3370 }
3371 
3372 #define POWERSAVE_MASK			GENMASK(7, 0)
3373 #define BALANCE_POWER_MASK		GENMASK(15, 8)
3374 #define BALANCE_PERFORMANCE_MASK	GENMASK(23, 16)
3375 #define PERFORMANCE_MASK		GENMASK(31, 24)
3376 
3377 #define HWP_SET_EPP_VALUES(powersave, balance_power, balance_perf, performance) \
3378 	(FIELD_PREP_CONST(POWERSAVE_MASK, powersave) |\
3379 	 FIELD_PREP_CONST(BALANCE_POWER_MASK, balance_power) |\
3380 	 FIELD_PREP_CONST(BALANCE_PERFORMANCE_MASK, balance_perf) |\
3381 	 FIELD_PREP_CONST(PERFORMANCE_MASK, performance))
3382 
3383 #define HWP_SET_DEF_BALANCE_PERF_EPP(balance_perf) \
3384 	(HWP_SET_EPP_VALUES(HWP_EPP_POWERSAVE, HWP_EPP_BALANCE_POWERSAVE,\
3385 	 balance_perf, HWP_EPP_PERFORMANCE))
3386 
3387 static const struct x86_cpu_id intel_epp_default[] = {
3388 	/*
3389 	 * Set EPP value as 102, this is the max suggested EPP
3390 	 * which can result in one core turbo frequency for
3391 	 * AlderLake Mobile CPUs.
3392 	 */
3393 	X86_MATCH_INTEL_FAM6_MODEL(ALDERLAKE_L, HWP_SET_DEF_BALANCE_PERF_EPP(102)),
3394 	X86_MATCH_INTEL_FAM6_MODEL(SAPPHIRERAPIDS_X, HWP_SET_DEF_BALANCE_PERF_EPP(32)),
3395 	X86_MATCH_INTEL_FAM6_MODEL(METEORLAKE_L, HWP_SET_EPP_VALUES(HWP_EPP_POWERSAVE,
3396 							HWP_EPP_BALANCE_POWERSAVE, 115, 16)),
3397 	{}
3398 };
3399 
3400 static const struct x86_cpu_id intel_hybrid_scaling_factor[] = {
3401 	X86_MATCH_INTEL_FAM6_MODEL(METEORLAKE_L, HYBRID_SCALING_FACTOR_MTL),
3402 	{}
3403 };
3404 
3405 static int __init intel_pstate_init(void)
3406 {
3407 	static struct cpudata **_all_cpu_data;
3408 	const struct x86_cpu_id *id;
3409 	int rc;
3410 
3411 	if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL)
3412 		return -ENODEV;
3413 
3414 	id = x86_match_cpu(hwp_support_ids);
3415 	if (id) {
3416 		hwp_forced = intel_pstate_hwp_is_enabled();
3417 
3418 		if (hwp_forced)
3419 			pr_info("HWP enabled by BIOS\n");
3420 		else if (no_load)
3421 			return -ENODEV;
3422 
3423 		copy_cpu_funcs(&core_funcs);
3424 		/*
3425 		 * Avoid enabling HWP for processors without EPP support,
3426 		 * because that means incomplete HWP implementation which is a
3427 		 * corner case and supporting it is generally problematic.
3428 		 *
3429 		 * If HWP is enabled already, though, there is no choice but to
3430 		 * deal with it.
3431 		 */
3432 		if ((!no_hwp && boot_cpu_has(X86_FEATURE_HWP_EPP)) || hwp_forced) {
3433 			hwp_active = true;
3434 			hwp_mode_bdw = id->driver_data;
3435 			intel_pstate.attr = hwp_cpufreq_attrs;
3436 			intel_cpufreq.attr = hwp_cpufreq_attrs;
3437 			intel_cpufreq.flags |= CPUFREQ_NEED_UPDATE_LIMITS;
3438 			intel_cpufreq.adjust_perf = intel_cpufreq_adjust_perf;
3439 			if (!default_driver)
3440 				default_driver = &intel_pstate;
3441 
3442 			pstate_funcs.get_cpu_scaling = hwp_get_cpu_scaling;
3443 
3444 			goto hwp_cpu_matched;
3445 		}
3446 		pr_info("HWP not enabled\n");
3447 	} else {
3448 		if (no_load)
3449 			return -ENODEV;
3450 
3451 		id = x86_match_cpu(intel_pstate_cpu_ids);
3452 		if (!id) {
3453 			pr_info("CPU model not supported\n");
3454 			return -ENODEV;
3455 		}
3456 
3457 		copy_cpu_funcs((struct pstate_funcs *)id->driver_data);
3458 	}
3459 
3460 	if (intel_pstate_msrs_not_valid()) {
3461 		pr_info("Invalid MSRs\n");
3462 		return -ENODEV;
3463 	}
3464 	/* Without HWP start in the passive mode. */
3465 	if (!default_driver)
3466 		default_driver = &intel_cpufreq;
3467 
3468 hwp_cpu_matched:
3469 	/*
3470 	 * The Intel pstate driver will be ignored if the platform
3471 	 * firmware has its own power management modes.
3472 	 */
3473 	if (intel_pstate_platform_pwr_mgmt_exists()) {
3474 		pr_info("P-states controlled by the platform\n");
3475 		return -ENODEV;
3476 	}
3477 
3478 	if (!hwp_active && hwp_only)
3479 		return -ENOTSUPP;
3480 
3481 	pr_info("Intel P-state driver initializing\n");
3482 
3483 	_all_cpu_data = vzalloc(array_size(sizeof(void *), num_possible_cpus()));
3484 	if (!_all_cpu_data)
3485 		return -ENOMEM;
3486 
3487 	WRITE_ONCE(all_cpu_data, _all_cpu_data);
3488 
3489 	intel_pstate_request_control_from_smm();
3490 
3491 	intel_pstate_sysfs_expose_params();
3492 
3493 	if (hwp_active) {
3494 		const struct x86_cpu_id *id = x86_match_cpu(intel_epp_default);
3495 		const struct x86_cpu_id *hybrid_id = x86_match_cpu(intel_hybrid_scaling_factor);
3496 
3497 		if (id) {
3498 			epp_values[EPP_INDEX_POWERSAVE] =
3499 					FIELD_GET(POWERSAVE_MASK, id->driver_data);
3500 			epp_values[EPP_INDEX_BALANCE_POWERSAVE] =
3501 					FIELD_GET(BALANCE_POWER_MASK, id->driver_data);
3502 			epp_values[EPP_INDEX_BALANCE_PERFORMANCE] =
3503 					FIELD_GET(BALANCE_PERFORMANCE_MASK, id->driver_data);
3504 			epp_values[EPP_INDEX_PERFORMANCE] =
3505 					FIELD_GET(PERFORMANCE_MASK, id->driver_data);
3506 			pr_debug("Updated EPPs powersave:%x balanced power:%x balanced perf:%x performance:%x\n",
3507 				 epp_values[EPP_INDEX_POWERSAVE],
3508 				 epp_values[EPP_INDEX_BALANCE_POWERSAVE],
3509 				 epp_values[EPP_INDEX_BALANCE_PERFORMANCE],
3510 				 epp_values[EPP_INDEX_PERFORMANCE]);
3511 		}
3512 
3513 		if (hybrid_id) {
3514 			hybrid_scaling_factor = hybrid_id->driver_data;
3515 			pr_debug("hybrid scaling factor: %d\n", hybrid_scaling_factor);
3516 		}
3517 
3518 	}
3519 
3520 	mutex_lock(&intel_pstate_driver_lock);
3521 	rc = intel_pstate_register_driver(default_driver);
3522 	mutex_unlock(&intel_pstate_driver_lock);
3523 	if (rc) {
3524 		intel_pstate_sysfs_remove();
3525 		return rc;
3526 	}
3527 
3528 	if (hwp_active) {
3529 		const struct x86_cpu_id *id;
3530 
3531 		id = x86_match_cpu(intel_pstate_cpu_ee_disable_ids);
3532 		if (id) {
3533 			set_power_ctl_ee_state(false);
3534 			pr_info("Disabling energy efficiency optimization\n");
3535 		}
3536 
3537 		pr_info("HWP enabled\n");
3538 	} else if (boot_cpu_has(X86_FEATURE_HYBRID_CPU)) {
3539 		pr_warn("Problematic setup: Hybrid processor with disabled HWP\n");
3540 	}
3541 
3542 	return 0;
3543 }
3544 device_initcall(intel_pstate_init);
3545 
3546 static int __init intel_pstate_setup(char *str)
3547 {
3548 	if (!str)
3549 		return -EINVAL;
3550 
3551 	if (!strcmp(str, "disable"))
3552 		no_load = 1;
3553 	else if (!strcmp(str, "active"))
3554 		default_driver = &intel_pstate;
3555 	else if (!strcmp(str, "passive"))
3556 		default_driver = &intel_cpufreq;
3557 
3558 	if (!strcmp(str, "no_hwp"))
3559 		no_hwp = 1;
3560 
3561 	if (!strcmp(str, "force"))
3562 		force_load = 1;
3563 	if (!strcmp(str, "hwp_only"))
3564 		hwp_only = 1;
3565 	if (!strcmp(str, "per_cpu_perf_limits"))
3566 		per_cpu_limits = true;
3567 
3568 #ifdef CONFIG_ACPI
3569 	if (!strcmp(str, "support_acpi_ppc"))
3570 		acpi_ppc = true;
3571 #endif
3572 
3573 	return 0;
3574 }
3575 early_param("intel_pstate", intel_pstate_setup);
3576 
3577 MODULE_AUTHOR("Dirk Brandewie <dirk.j.brandewie@intel.com>");
3578 MODULE_DESCRIPTION("'intel_pstate' - P state driver Intel Core processors");
3579