xref: /linux/drivers/cpufreq/intel_pstate.c (revision 1553419c3c10cf386496e68b90b5d0ce966ac614)
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 	if (ret)
359 		return;
360 
361 	/*
362 	 * On some systems with overclocking enabled, CPPC.highest_perf is hardcoded to 0xff.
363 	 * In this case we can't use CPPC.highest_perf to enable ITMT.
364 	 * In this case we can look at MSR_HWP_CAPABILITIES bits [8:0] to decide.
365 	 */
366 	if (cppc_perf.highest_perf == CPPC_MAX_PERF)
367 		cppc_perf.highest_perf = HWP_HIGHEST_PERF(READ_ONCE(all_cpu_data[cpu]->hwp_cap_cached));
368 
369 	/*
370 	 * The priorities can be set regardless of whether or not
371 	 * sched_set_itmt_support(true) has been called and it is valid to
372 	 * update them at any time after it has been called.
373 	 */
374 	sched_set_itmt_core_prio(cppc_perf.highest_perf, cpu);
375 
376 	if (max_highest_perf <= min_highest_perf) {
377 		if (cppc_perf.highest_perf > max_highest_perf)
378 			max_highest_perf = cppc_perf.highest_perf;
379 
380 		if (cppc_perf.highest_perf < min_highest_perf)
381 			min_highest_perf = cppc_perf.highest_perf;
382 
383 		if (max_highest_perf > min_highest_perf) {
384 			/*
385 			 * This code can be run during CPU online under the
386 			 * CPU hotplug locks, so sched_set_itmt_support()
387 			 * cannot be called from here.  Queue up a work item
388 			 * to invoke it.
389 			 */
390 			schedule_work(&sched_itmt_work);
391 		}
392 	}
393 }
394 
395 static int intel_pstate_get_cppc_guaranteed(int cpu)
396 {
397 	struct cppc_perf_caps cppc_perf;
398 	int ret;
399 
400 	ret = cppc_get_perf_caps(cpu, &cppc_perf);
401 	if (ret)
402 		return ret;
403 
404 	if (cppc_perf.guaranteed_perf)
405 		return cppc_perf.guaranteed_perf;
406 
407 	return cppc_perf.nominal_perf;
408 }
409 
410 static int intel_pstate_cppc_get_scaling(int cpu)
411 {
412 	struct cppc_perf_caps cppc_perf;
413 	int ret;
414 
415 	ret = cppc_get_perf_caps(cpu, &cppc_perf);
416 
417 	/*
418 	 * If the nominal frequency and the nominal performance are not
419 	 * zero and the ratio between them is not 100, return the hybrid
420 	 * scaling factor.
421 	 */
422 	if (!ret && cppc_perf.nominal_perf && cppc_perf.nominal_freq &&
423 	    cppc_perf.nominal_perf * 100 != cppc_perf.nominal_freq)
424 		return hybrid_scaling_factor;
425 
426 	return core_get_scaling();
427 }
428 
429 #else /* CONFIG_ACPI_CPPC_LIB */
430 static inline void intel_pstate_set_itmt_prio(int cpu)
431 {
432 }
433 #endif /* CONFIG_ACPI_CPPC_LIB */
434 
435 static void intel_pstate_init_acpi_perf_limits(struct cpufreq_policy *policy)
436 {
437 	struct cpudata *cpu;
438 	int ret;
439 	int i;
440 
441 	if (hwp_active) {
442 		intel_pstate_set_itmt_prio(policy->cpu);
443 		return;
444 	}
445 
446 	if (!intel_pstate_get_ppc_enable_status())
447 		return;
448 
449 	cpu = all_cpu_data[policy->cpu];
450 
451 	ret = acpi_processor_register_performance(&cpu->acpi_perf_data,
452 						  policy->cpu);
453 	if (ret)
454 		return;
455 
456 	/*
457 	 * Check if the control value in _PSS is for PERF_CTL MSR, which should
458 	 * guarantee that the states returned by it map to the states in our
459 	 * list directly.
460 	 */
461 	if (cpu->acpi_perf_data.control_register.space_id !=
462 						ACPI_ADR_SPACE_FIXED_HARDWARE)
463 		goto err;
464 
465 	/*
466 	 * If there is only one entry _PSS, simply ignore _PSS and continue as
467 	 * usual without taking _PSS into account
468 	 */
469 	if (cpu->acpi_perf_data.state_count < 2)
470 		goto err;
471 
472 	pr_debug("CPU%u - ACPI _PSS perf data\n", policy->cpu);
473 	for (i = 0; i < cpu->acpi_perf_data.state_count; i++) {
474 		pr_debug("     %cP%d: %u MHz, %u mW, 0x%x\n",
475 			 (i == cpu->acpi_perf_data.state ? '*' : ' '), i,
476 			 (u32) cpu->acpi_perf_data.states[i].core_frequency,
477 			 (u32) cpu->acpi_perf_data.states[i].power,
478 			 (u32) cpu->acpi_perf_data.states[i].control);
479 	}
480 
481 	cpu->valid_pss_table = true;
482 	pr_debug("_PPC limits will be enforced\n");
483 
484 	return;
485 
486  err:
487 	cpu->valid_pss_table = false;
488 	acpi_processor_unregister_performance(policy->cpu);
489 }
490 
491 static void intel_pstate_exit_perf_limits(struct cpufreq_policy *policy)
492 {
493 	struct cpudata *cpu;
494 
495 	cpu = all_cpu_data[policy->cpu];
496 	if (!cpu->valid_pss_table)
497 		return;
498 
499 	acpi_processor_unregister_performance(policy->cpu);
500 }
501 #else /* CONFIG_ACPI */
502 static inline void intel_pstate_init_acpi_perf_limits(struct cpufreq_policy *policy)
503 {
504 }
505 
506 static inline void intel_pstate_exit_perf_limits(struct cpufreq_policy *policy)
507 {
508 }
509 
510 static inline bool intel_pstate_acpi_pm_profile_server(void)
511 {
512 	return false;
513 }
514 #endif /* CONFIG_ACPI */
515 
516 #ifndef CONFIG_ACPI_CPPC_LIB
517 static inline int intel_pstate_get_cppc_guaranteed(int cpu)
518 {
519 	return -ENOTSUPP;
520 }
521 
522 static int intel_pstate_cppc_get_scaling(int cpu)
523 {
524 	return core_get_scaling();
525 }
526 #endif /* CONFIG_ACPI_CPPC_LIB */
527 
528 static int intel_pstate_freq_to_hwp_rel(struct cpudata *cpu, int freq,
529 					unsigned int relation)
530 {
531 	if (freq == cpu->pstate.turbo_freq)
532 		return cpu->pstate.turbo_pstate;
533 
534 	if (freq == cpu->pstate.max_freq)
535 		return cpu->pstate.max_pstate;
536 
537 	switch (relation) {
538 	case CPUFREQ_RELATION_H:
539 		return freq / cpu->pstate.scaling;
540 	case CPUFREQ_RELATION_C:
541 		return DIV_ROUND_CLOSEST(freq, cpu->pstate.scaling);
542 	}
543 
544 	return DIV_ROUND_UP(freq, cpu->pstate.scaling);
545 }
546 
547 static int intel_pstate_freq_to_hwp(struct cpudata *cpu, int freq)
548 {
549 	return intel_pstate_freq_to_hwp_rel(cpu, freq, CPUFREQ_RELATION_L);
550 }
551 
552 /**
553  * intel_pstate_hybrid_hwp_adjust - Calibrate HWP performance levels.
554  * @cpu: Target CPU.
555  *
556  * On hybrid processors, HWP may expose more performance levels than there are
557  * P-states accessible through the PERF_CTL interface.  If that happens, the
558  * scaling factor between HWP performance levels and CPU frequency will be less
559  * than the scaling factor between P-state values and CPU frequency.
560  *
561  * In that case, adjust the CPU parameters used in computations accordingly.
562  */
563 static void intel_pstate_hybrid_hwp_adjust(struct cpudata *cpu)
564 {
565 	int perf_ctl_max_phys = cpu->pstate.max_pstate_physical;
566 	int perf_ctl_scaling = cpu->pstate.perf_ctl_scaling;
567 	int perf_ctl_turbo = pstate_funcs.get_turbo(cpu->cpu);
568 	int scaling = cpu->pstate.scaling;
569 	int freq;
570 
571 	pr_debug("CPU%d: perf_ctl_max_phys = %d\n", cpu->cpu, perf_ctl_max_phys);
572 	pr_debug("CPU%d: perf_ctl_turbo = %d\n", cpu->cpu, perf_ctl_turbo);
573 	pr_debug("CPU%d: perf_ctl_scaling = %d\n", cpu->cpu, perf_ctl_scaling);
574 	pr_debug("CPU%d: HWP_CAP guaranteed = %d\n", cpu->cpu, cpu->pstate.max_pstate);
575 	pr_debug("CPU%d: HWP_CAP highest = %d\n", cpu->cpu, cpu->pstate.turbo_pstate);
576 	pr_debug("CPU%d: HWP-to-frequency scaling factor: %d\n", cpu->cpu, scaling);
577 
578 	cpu->pstate.turbo_freq = rounddown(cpu->pstate.turbo_pstate * scaling,
579 					   perf_ctl_scaling);
580 	cpu->pstate.max_freq = rounddown(cpu->pstate.max_pstate * scaling,
581 					 perf_ctl_scaling);
582 
583 	freq = perf_ctl_max_phys * perf_ctl_scaling;
584 	cpu->pstate.max_pstate_physical = intel_pstate_freq_to_hwp(cpu, freq);
585 
586 	freq = cpu->pstate.min_pstate * perf_ctl_scaling;
587 	cpu->pstate.min_freq = freq;
588 	/*
589 	 * Cast the min P-state value retrieved via pstate_funcs.get_min() to
590 	 * the effective range of HWP performance levels.
591 	 */
592 	cpu->pstate.min_pstate = intel_pstate_freq_to_hwp(cpu, freq);
593 }
594 
595 static bool turbo_is_disabled(void)
596 {
597 	u64 misc_en;
598 
599 	rdmsrl(MSR_IA32_MISC_ENABLE, misc_en);
600 
601 	return !!(misc_en & MSR_IA32_MISC_ENABLE_TURBO_DISABLE);
602 }
603 
604 static int min_perf_pct_min(void)
605 {
606 	struct cpudata *cpu = all_cpu_data[0];
607 	int turbo_pstate = cpu->pstate.turbo_pstate;
608 
609 	return turbo_pstate ?
610 		(cpu->pstate.min_pstate * 100 / turbo_pstate) : 0;
611 }
612 
613 static s16 intel_pstate_get_epb(struct cpudata *cpu_data)
614 {
615 	u64 epb;
616 	int ret;
617 
618 	if (!boot_cpu_has(X86_FEATURE_EPB))
619 		return -ENXIO;
620 
621 	ret = rdmsrl_on_cpu(cpu_data->cpu, MSR_IA32_ENERGY_PERF_BIAS, &epb);
622 	if (ret)
623 		return (s16)ret;
624 
625 	return (s16)(epb & 0x0f);
626 }
627 
628 static s16 intel_pstate_get_epp(struct cpudata *cpu_data, u64 hwp_req_data)
629 {
630 	s16 epp;
631 
632 	if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
633 		/*
634 		 * When hwp_req_data is 0, means that caller didn't read
635 		 * MSR_HWP_REQUEST, so need to read and get EPP.
636 		 */
637 		if (!hwp_req_data) {
638 			epp = rdmsrl_on_cpu(cpu_data->cpu, MSR_HWP_REQUEST,
639 					    &hwp_req_data);
640 			if (epp)
641 				return epp;
642 		}
643 		epp = (hwp_req_data >> 24) & 0xff;
644 	} else {
645 		/* When there is no EPP present, HWP uses EPB settings */
646 		epp = intel_pstate_get_epb(cpu_data);
647 	}
648 
649 	return epp;
650 }
651 
652 static int intel_pstate_set_epb(int cpu, s16 pref)
653 {
654 	u64 epb;
655 	int ret;
656 
657 	if (!boot_cpu_has(X86_FEATURE_EPB))
658 		return -ENXIO;
659 
660 	ret = rdmsrl_on_cpu(cpu, MSR_IA32_ENERGY_PERF_BIAS, &epb);
661 	if (ret)
662 		return ret;
663 
664 	epb = (epb & ~0x0f) | pref;
665 	wrmsrl_on_cpu(cpu, MSR_IA32_ENERGY_PERF_BIAS, epb);
666 
667 	return 0;
668 }
669 
670 /*
671  * EPP/EPB display strings corresponding to EPP index in the
672  * energy_perf_strings[]
673  *	index		String
674  *-------------------------------------
675  *	0		default
676  *	1		performance
677  *	2		balance_performance
678  *	3		balance_power
679  *	4		power
680  */
681 
682 enum energy_perf_value_index {
683 	EPP_INDEX_DEFAULT = 0,
684 	EPP_INDEX_PERFORMANCE,
685 	EPP_INDEX_BALANCE_PERFORMANCE,
686 	EPP_INDEX_BALANCE_POWERSAVE,
687 	EPP_INDEX_POWERSAVE,
688 };
689 
690 static const char * const energy_perf_strings[] = {
691 	[EPP_INDEX_DEFAULT] = "default",
692 	[EPP_INDEX_PERFORMANCE] = "performance",
693 	[EPP_INDEX_BALANCE_PERFORMANCE] = "balance_performance",
694 	[EPP_INDEX_BALANCE_POWERSAVE] = "balance_power",
695 	[EPP_INDEX_POWERSAVE] = "power",
696 	NULL
697 };
698 static unsigned int epp_values[] = {
699 	[EPP_INDEX_DEFAULT] = 0, /* Unused index */
700 	[EPP_INDEX_PERFORMANCE] = HWP_EPP_PERFORMANCE,
701 	[EPP_INDEX_BALANCE_PERFORMANCE] = HWP_EPP_BALANCE_PERFORMANCE,
702 	[EPP_INDEX_BALANCE_POWERSAVE] = HWP_EPP_BALANCE_POWERSAVE,
703 	[EPP_INDEX_POWERSAVE] = HWP_EPP_POWERSAVE,
704 };
705 
706 static int intel_pstate_get_energy_pref_index(struct cpudata *cpu_data, int *raw_epp)
707 {
708 	s16 epp;
709 	int index = -EINVAL;
710 
711 	*raw_epp = 0;
712 	epp = intel_pstate_get_epp(cpu_data, 0);
713 	if (epp < 0)
714 		return epp;
715 
716 	if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
717 		if (epp == epp_values[EPP_INDEX_PERFORMANCE])
718 			return EPP_INDEX_PERFORMANCE;
719 		if (epp == epp_values[EPP_INDEX_BALANCE_PERFORMANCE])
720 			return EPP_INDEX_BALANCE_PERFORMANCE;
721 		if (epp == epp_values[EPP_INDEX_BALANCE_POWERSAVE])
722 			return EPP_INDEX_BALANCE_POWERSAVE;
723 		if (epp == epp_values[EPP_INDEX_POWERSAVE])
724 			return EPP_INDEX_POWERSAVE;
725 		*raw_epp = epp;
726 		return 0;
727 	} else if (boot_cpu_has(X86_FEATURE_EPB)) {
728 		/*
729 		 * Range:
730 		 *	0x00-0x03	:	Performance
731 		 *	0x04-0x07	:	Balance performance
732 		 *	0x08-0x0B	:	Balance power
733 		 *	0x0C-0x0F	:	Power
734 		 * The EPB is a 4 bit value, but our ranges restrict the
735 		 * value which can be set. Here only using top two bits
736 		 * effectively.
737 		 */
738 		index = (epp >> 2) + 1;
739 	}
740 
741 	return index;
742 }
743 
744 static int intel_pstate_set_epp(struct cpudata *cpu, u32 epp)
745 {
746 	int ret;
747 
748 	/*
749 	 * Use the cached HWP Request MSR value, because in the active mode the
750 	 * register itself may be updated by intel_pstate_hwp_boost_up() or
751 	 * intel_pstate_hwp_boost_down() at any time.
752 	 */
753 	u64 value = READ_ONCE(cpu->hwp_req_cached);
754 
755 	value &= ~GENMASK_ULL(31, 24);
756 	value |= (u64)epp << 24;
757 	/*
758 	 * The only other updater of hwp_req_cached in the active mode,
759 	 * intel_pstate_hwp_set(), is called under the same lock as this
760 	 * function, so it cannot run in parallel with the update below.
761 	 */
762 	WRITE_ONCE(cpu->hwp_req_cached, value);
763 	ret = wrmsrl_on_cpu(cpu->cpu, MSR_HWP_REQUEST, value);
764 	if (!ret)
765 		cpu->epp_cached = epp;
766 
767 	return ret;
768 }
769 
770 static int intel_pstate_set_energy_pref_index(struct cpudata *cpu_data,
771 					      int pref_index, bool use_raw,
772 					      u32 raw_epp)
773 {
774 	int epp = -EINVAL;
775 	int ret;
776 
777 	if (!pref_index)
778 		epp = cpu_data->epp_default;
779 
780 	if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
781 		if (use_raw)
782 			epp = raw_epp;
783 		else if (epp == -EINVAL)
784 			epp = epp_values[pref_index];
785 
786 		/*
787 		 * To avoid confusion, refuse to set EPP to any values different
788 		 * from 0 (performance) if the current policy is "performance",
789 		 * because those values would be overridden.
790 		 */
791 		if (epp > 0 && cpu_data->policy == CPUFREQ_POLICY_PERFORMANCE)
792 			return -EBUSY;
793 
794 		ret = intel_pstate_set_epp(cpu_data, epp);
795 	} else {
796 		if (epp == -EINVAL)
797 			epp = (pref_index - 1) << 2;
798 		ret = intel_pstate_set_epb(cpu_data->cpu, epp);
799 	}
800 
801 	return ret;
802 }
803 
804 static ssize_t show_energy_performance_available_preferences(
805 				struct cpufreq_policy *policy, char *buf)
806 {
807 	int i = 0;
808 	int ret = 0;
809 
810 	while (energy_perf_strings[i] != NULL)
811 		ret += sprintf(&buf[ret], "%s ", energy_perf_strings[i++]);
812 
813 	ret += sprintf(&buf[ret], "\n");
814 
815 	return ret;
816 }
817 
818 cpufreq_freq_attr_ro(energy_performance_available_preferences);
819 
820 static struct cpufreq_driver intel_pstate;
821 
822 static ssize_t store_energy_performance_preference(
823 		struct cpufreq_policy *policy, const char *buf, size_t count)
824 {
825 	struct cpudata *cpu = all_cpu_data[policy->cpu];
826 	char str_preference[21];
827 	bool raw = false;
828 	ssize_t ret;
829 	u32 epp = 0;
830 
831 	ret = sscanf(buf, "%20s", str_preference);
832 	if (ret != 1)
833 		return -EINVAL;
834 
835 	ret = match_string(energy_perf_strings, -1, str_preference);
836 	if (ret < 0) {
837 		if (!boot_cpu_has(X86_FEATURE_HWP_EPP))
838 			return ret;
839 
840 		ret = kstrtouint(buf, 10, &epp);
841 		if (ret)
842 			return ret;
843 
844 		if (epp > 255)
845 			return -EINVAL;
846 
847 		raw = true;
848 	}
849 
850 	/*
851 	 * This function runs with the policy R/W semaphore held, which
852 	 * guarantees that the driver pointer will not change while it is
853 	 * running.
854 	 */
855 	if (!intel_pstate_driver)
856 		return -EAGAIN;
857 
858 	mutex_lock(&intel_pstate_limits_lock);
859 
860 	if (intel_pstate_driver == &intel_pstate) {
861 		ret = intel_pstate_set_energy_pref_index(cpu, ret, raw, epp);
862 	} else {
863 		/*
864 		 * In the passive mode the governor needs to be stopped on the
865 		 * target CPU before the EPP update and restarted after it,
866 		 * which is super-heavy-weight, so make sure it is worth doing
867 		 * upfront.
868 		 */
869 		if (!raw)
870 			epp = ret ? epp_values[ret] : cpu->epp_default;
871 
872 		if (cpu->epp_cached != epp) {
873 			int err;
874 
875 			cpufreq_stop_governor(policy);
876 			ret = intel_pstate_set_epp(cpu, epp);
877 			err = cpufreq_start_governor(policy);
878 			if (!ret)
879 				ret = err;
880 		} else {
881 			ret = 0;
882 		}
883 	}
884 
885 	mutex_unlock(&intel_pstate_limits_lock);
886 
887 	return ret ?: count;
888 }
889 
890 static ssize_t show_energy_performance_preference(
891 				struct cpufreq_policy *policy, char *buf)
892 {
893 	struct cpudata *cpu_data = all_cpu_data[policy->cpu];
894 	int preference, raw_epp;
895 
896 	preference = intel_pstate_get_energy_pref_index(cpu_data, &raw_epp);
897 	if (preference < 0)
898 		return preference;
899 
900 	if (raw_epp)
901 		return  sprintf(buf, "%d\n", raw_epp);
902 	else
903 		return  sprintf(buf, "%s\n", energy_perf_strings[preference]);
904 }
905 
906 cpufreq_freq_attr_rw(energy_performance_preference);
907 
908 static ssize_t show_base_frequency(struct cpufreq_policy *policy, char *buf)
909 {
910 	struct cpudata *cpu = all_cpu_data[policy->cpu];
911 	int ratio, freq;
912 
913 	ratio = intel_pstate_get_cppc_guaranteed(policy->cpu);
914 	if (ratio <= 0) {
915 		u64 cap;
916 
917 		rdmsrl_on_cpu(policy->cpu, MSR_HWP_CAPABILITIES, &cap);
918 		ratio = HWP_GUARANTEED_PERF(cap);
919 	}
920 
921 	freq = ratio * cpu->pstate.scaling;
922 	if (cpu->pstate.scaling != cpu->pstate.perf_ctl_scaling)
923 		freq = rounddown(freq, cpu->pstate.perf_ctl_scaling);
924 
925 	return sprintf(buf, "%d\n", freq);
926 }
927 
928 cpufreq_freq_attr_ro(base_frequency);
929 
930 static struct freq_attr *hwp_cpufreq_attrs[] = {
931 	&energy_performance_preference,
932 	&energy_performance_available_preferences,
933 	&base_frequency,
934 	NULL,
935 };
936 
937 static void __intel_pstate_get_hwp_cap(struct cpudata *cpu)
938 {
939 	u64 cap;
940 
941 	rdmsrl_on_cpu(cpu->cpu, MSR_HWP_CAPABILITIES, &cap);
942 	WRITE_ONCE(cpu->hwp_cap_cached, cap);
943 	cpu->pstate.max_pstate = HWP_GUARANTEED_PERF(cap);
944 	cpu->pstate.turbo_pstate = HWP_HIGHEST_PERF(cap);
945 }
946 
947 static void intel_pstate_get_hwp_cap(struct cpudata *cpu)
948 {
949 	int scaling = cpu->pstate.scaling;
950 
951 	__intel_pstate_get_hwp_cap(cpu);
952 
953 	cpu->pstate.max_freq = cpu->pstate.max_pstate * scaling;
954 	cpu->pstate.turbo_freq = cpu->pstate.turbo_pstate * scaling;
955 	if (scaling != cpu->pstate.perf_ctl_scaling) {
956 		int perf_ctl_scaling = cpu->pstate.perf_ctl_scaling;
957 
958 		cpu->pstate.max_freq = rounddown(cpu->pstate.max_freq,
959 						 perf_ctl_scaling);
960 		cpu->pstate.turbo_freq = rounddown(cpu->pstate.turbo_freq,
961 						   perf_ctl_scaling);
962 	}
963 }
964 
965 static void intel_pstate_hwp_set(unsigned int cpu)
966 {
967 	struct cpudata *cpu_data = all_cpu_data[cpu];
968 	int max, min;
969 	u64 value;
970 	s16 epp;
971 
972 	max = cpu_data->max_perf_ratio;
973 	min = cpu_data->min_perf_ratio;
974 
975 	if (cpu_data->policy == CPUFREQ_POLICY_PERFORMANCE)
976 		min = max;
977 
978 	rdmsrl_on_cpu(cpu, MSR_HWP_REQUEST, &value);
979 
980 	value &= ~HWP_MIN_PERF(~0L);
981 	value |= HWP_MIN_PERF(min);
982 
983 	value &= ~HWP_MAX_PERF(~0L);
984 	value |= HWP_MAX_PERF(max);
985 
986 	if (cpu_data->epp_policy == cpu_data->policy)
987 		goto skip_epp;
988 
989 	cpu_data->epp_policy = cpu_data->policy;
990 
991 	if (cpu_data->policy == CPUFREQ_POLICY_PERFORMANCE) {
992 		epp = intel_pstate_get_epp(cpu_data, value);
993 		cpu_data->epp_powersave = epp;
994 		/* If EPP read was failed, then don't try to write */
995 		if (epp < 0)
996 			goto skip_epp;
997 
998 		epp = 0;
999 	} else {
1000 		/* skip setting EPP, when saved value is invalid */
1001 		if (cpu_data->epp_powersave < 0)
1002 			goto skip_epp;
1003 
1004 		/*
1005 		 * No need to restore EPP when it is not zero. This
1006 		 * means:
1007 		 *  - Policy is not changed
1008 		 *  - user has manually changed
1009 		 *  - Error reading EPB
1010 		 */
1011 		epp = intel_pstate_get_epp(cpu_data, value);
1012 		if (epp)
1013 			goto skip_epp;
1014 
1015 		epp = cpu_data->epp_powersave;
1016 	}
1017 	if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
1018 		value &= ~GENMASK_ULL(31, 24);
1019 		value |= (u64)epp << 24;
1020 	} else {
1021 		intel_pstate_set_epb(cpu, epp);
1022 	}
1023 skip_epp:
1024 	WRITE_ONCE(cpu_data->hwp_req_cached, value);
1025 	wrmsrl_on_cpu(cpu, MSR_HWP_REQUEST, value);
1026 }
1027 
1028 static void intel_pstate_disable_hwp_interrupt(struct cpudata *cpudata);
1029 
1030 static void intel_pstate_hwp_offline(struct cpudata *cpu)
1031 {
1032 	u64 value = READ_ONCE(cpu->hwp_req_cached);
1033 	int min_perf;
1034 
1035 	intel_pstate_disable_hwp_interrupt(cpu);
1036 
1037 	if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
1038 		/*
1039 		 * In case the EPP has been set to "performance" by the
1040 		 * active mode "performance" scaling algorithm, replace that
1041 		 * temporary value with the cached EPP one.
1042 		 */
1043 		value &= ~GENMASK_ULL(31, 24);
1044 		value |= HWP_ENERGY_PERF_PREFERENCE(cpu->epp_cached);
1045 		/*
1046 		 * However, make sure that EPP will be set to "performance" when
1047 		 * the CPU is brought back online again and the "performance"
1048 		 * scaling algorithm is still in effect.
1049 		 */
1050 		cpu->epp_policy = CPUFREQ_POLICY_UNKNOWN;
1051 	}
1052 
1053 	/*
1054 	 * Clear the desired perf field in the cached HWP request value to
1055 	 * prevent nonzero desired values from being leaked into the active
1056 	 * mode.
1057 	 */
1058 	value &= ~HWP_DESIRED_PERF(~0L);
1059 	WRITE_ONCE(cpu->hwp_req_cached, value);
1060 
1061 	value &= ~GENMASK_ULL(31, 0);
1062 	min_perf = HWP_LOWEST_PERF(READ_ONCE(cpu->hwp_cap_cached));
1063 
1064 	/* Set hwp_max = hwp_min */
1065 	value |= HWP_MAX_PERF(min_perf);
1066 	value |= HWP_MIN_PERF(min_perf);
1067 
1068 	/* Set EPP to min */
1069 	if (boot_cpu_has(X86_FEATURE_HWP_EPP))
1070 		value |= HWP_ENERGY_PERF_PREFERENCE(HWP_EPP_POWERSAVE);
1071 
1072 	wrmsrl_on_cpu(cpu->cpu, MSR_HWP_REQUEST, value);
1073 }
1074 
1075 #define POWER_CTL_EE_ENABLE	1
1076 #define POWER_CTL_EE_DISABLE	2
1077 
1078 static int power_ctl_ee_state;
1079 
1080 static void set_power_ctl_ee_state(bool input)
1081 {
1082 	u64 power_ctl;
1083 
1084 	mutex_lock(&intel_pstate_driver_lock);
1085 	rdmsrl(MSR_IA32_POWER_CTL, power_ctl);
1086 	if (input) {
1087 		power_ctl &= ~BIT(MSR_IA32_POWER_CTL_BIT_EE);
1088 		power_ctl_ee_state = POWER_CTL_EE_ENABLE;
1089 	} else {
1090 		power_ctl |= BIT(MSR_IA32_POWER_CTL_BIT_EE);
1091 		power_ctl_ee_state = POWER_CTL_EE_DISABLE;
1092 	}
1093 	wrmsrl(MSR_IA32_POWER_CTL, power_ctl);
1094 	mutex_unlock(&intel_pstate_driver_lock);
1095 }
1096 
1097 static void intel_pstate_hwp_enable(struct cpudata *cpudata);
1098 
1099 static void intel_pstate_hwp_reenable(struct cpudata *cpu)
1100 {
1101 	intel_pstate_hwp_enable(cpu);
1102 	wrmsrl_on_cpu(cpu->cpu, MSR_HWP_REQUEST, READ_ONCE(cpu->hwp_req_cached));
1103 }
1104 
1105 static int intel_pstate_suspend(struct cpufreq_policy *policy)
1106 {
1107 	struct cpudata *cpu = all_cpu_data[policy->cpu];
1108 
1109 	pr_debug("CPU %d suspending\n", cpu->cpu);
1110 
1111 	cpu->suspended = true;
1112 
1113 	/* disable HWP interrupt and cancel any pending work */
1114 	intel_pstate_disable_hwp_interrupt(cpu);
1115 
1116 	return 0;
1117 }
1118 
1119 static int intel_pstate_resume(struct cpufreq_policy *policy)
1120 {
1121 	struct cpudata *cpu = all_cpu_data[policy->cpu];
1122 
1123 	pr_debug("CPU %d resuming\n", cpu->cpu);
1124 
1125 	/* Only restore if the system default is changed */
1126 	if (power_ctl_ee_state == POWER_CTL_EE_ENABLE)
1127 		set_power_ctl_ee_state(true);
1128 	else if (power_ctl_ee_state == POWER_CTL_EE_DISABLE)
1129 		set_power_ctl_ee_state(false);
1130 
1131 	if (cpu->suspended && hwp_active) {
1132 		mutex_lock(&intel_pstate_limits_lock);
1133 
1134 		/* Re-enable HWP, because "online" has not done that. */
1135 		intel_pstate_hwp_reenable(cpu);
1136 
1137 		mutex_unlock(&intel_pstate_limits_lock);
1138 	}
1139 
1140 	cpu->suspended = false;
1141 
1142 	return 0;
1143 }
1144 
1145 static void intel_pstate_update_policies(void)
1146 {
1147 	int cpu;
1148 
1149 	for_each_possible_cpu(cpu)
1150 		cpufreq_update_policy(cpu);
1151 }
1152 
1153 static void __intel_pstate_update_max_freq(struct cpudata *cpudata,
1154 					   struct cpufreq_policy *policy)
1155 {
1156 	if (hwp_active)
1157 		intel_pstate_get_hwp_cap(cpudata);
1158 
1159 	policy->cpuinfo.max_freq = READ_ONCE(global.no_turbo) ?
1160 			cpudata->pstate.max_freq : cpudata->pstate.turbo_freq;
1161 
1162 	refresh_frequency_limits(policy);
1163 }
1164 
1165 static void intel_pstate_update_limits(unsigned int cpu)
1166 {
1167 	struct cpufreq_policy *policy = cpufreq_cpu_acquire(cpu);
1168 
1169 	if (!policy)
1170 		return;
1171 
1172 	__intel_pstate_update_max_freq(all_cpu_data[cpu], policy);
1173 
1174 	cpufreq_cpu_release(policy);
1175 }
1176 
1177 static void intel_pstate_update_limits_for_all(void)
1178 {
1179 	int cpu;
1180 
1181 	for_each_possible_cpu(cpu)
1182 		intel_pstate_update_limits(cpu);
1183 }
1184 
1185 /************************** sysfs begin ************************/
1186 #define show_one(file_name, object)					\
1187 	static ssize_t show_##file_name					\
1188 	(struct kobject *kobj, struct kobj_attribute *attr, char *buf)	\
1189 	{								\
1190 		return sprintf(buf, "%u\n", global.object);		\
1191 	}
1192 
1193 static ssize_t intel_pstate_show_status(char *buf);
1194 static int intel_pstate_update_status(const char *buf, size_t size);
1195 
1196 static ssize_t show_status(struct kobject *kobj,
1197 			   struct kobj_attribute *attr, char *buf)
1198 {
1199 	ssize_t ret;
1200 
1201 	mutex_lock(&intel_pstate_driver_lock);
1202 	ret = intel_pstate_show_status(buf);
1203 	mutex_unlock(&intel_pstate_driver_lock);
1204 
1205 	return ret;
1206 }
1207 
1208 static ssize_t store_status(struct kobject *a, struct kobj_attribute *b,
1209 			    const char *buf, size_t count)
1210 {
1211 	char *p = memchr(buf, '\n', count);
1212 	int ret;
1213 
1214 	mutex_lock(&intel_pstate_driver_lock);
1215 	ret = intel_pstate_update_status(buf, p ? p - buf : count);
1216 	mutex_unlock(&intel_pstate_driver_lock);
1217 
1218 	return ret < 0 ? ret : count;
1219 }
1220 
1221 static ssize_t show_turbo_pct(struct kobject *kobj,
1222 				struct kobj_attribute *attr, char *buf)
1223 {
1224 	struct cpudata *cpu;
1225 	int total, no_turbo, turbo_pct;
1226 	uint32_t turbo_fp;
1227 
1228 	mutex_lock(&intel_pstate_driver_lock);
1229 
1230 	if (!intel_pstate_driver) {
1231 		mutex_unlock(&intel_pstate_driver_lock);
1232 		return -EAGAIN;
1233 	}
1234 
1235 	cpu = all_cpu_data[0];
1236 
1237 	total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
1238 	no_turbo = cpu->pstate.max_pstate - cpu->pstate.min_pstate + 1;
1239 	turbo_fp = div_fp(no_turbo, total);
1240 	turbo_pct = 100 - fp_toint(mul_fp(turbo_fp, int_tofp(100)));
1241 
1242 	mutex_unlock(&intel_pstate_driver_lock);
1243 
1244 	return sprintf(buf, "%u\n", turbo_pct);
1245 }
1246 
1247 static ssize_t show_num_pstates(struct kobject *kobj,
1248 				struct kobj_attribute *attr, char *buf)
1249 {
1250 	struct cpudata *cpu;
1251 	int total;
1252 
1253 	mutex_lock(&intel_pstate_driver_lock);
1254 
1255 	if (!intel_pstate_driver) {
1256 		mutex_unlock(&intel_pstate_driver_lock);
1257 		return -EAGAIN;
1258 	}
1259 
1260 	cpu = all_cpu_data[0];
1261 	total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
1262 
1263 	mutex_unlock(&intel_pstate_driver_lock);
1264 
1265 	return sprintf(buf, "%u\n", total);
1266 }
1267 
1268 static ssize_t show_no_turbo(struct kobject *kobj,
1269 			     struct kobj_attribute *attr, char *buf)
1270 {
1271 	ssize_t ret;
1272 
1273 	mutex_lock(&intel_pstate_driver_lock);
1274 
1275 	if (!intel_pstate_driver) {
1276 		mutex_unlock(&intel_pstate_driver_lock);
1277 		return -EAGAIN;
1278 	}
1279 
1280 	ret = sprintf(buf, "%u\n", global.no_turbo);
1281 
1282 	mutex_unlock(&intel_pstate_driver_lock);
1283 
1284 	return ret;
1285 }
1286 
1287 static ssize_t store_no_turbo(struct kobject *a, struct kobj_attribute *b,
1288 			      const char *buf, size_t count)
1289 {
1290 	unsigned int input;
1291 	bool no_turbo;
1292 
1293 	if (sscanf(buf, "%u", &input) != 1)
1294 		return -EINVAL;
1295 
1296 	mutex_lock(&intel_pstate_driver_lock);
1297 
1298 	if (!intel_pstate_driver) {
1299 		count = -EAGAIN;
1300 		goto unlock_driver;
1301 	}
1302 
1303 	no_turbo = !!clamp_t(int, input, 0, 1);
1304 
1305 	if (no_turbo == global.no_turbo)
1306 		goto unlock_driver;
1307 
1308 	if (global.turbo_disabled) {
1309 		pr_notice_once("Turbo disabled by BIOS or unavailable on processor\n");
1310 		count = -EPERM;
1311 		goto unlock_driver;
1312 	}
1313 
1314 	WRITE_ONCE(global.no_turbo, no_turbo);
1315 
1316 	mutex_lock(&intel_pstate_limits_lock);
1317 
1318 	if (no_turbo) {
1319 		struct cpudata *cpu = all_cpu_data[0];
1320 		int pct = cpu->pstate.max_pstate * 100 / cpu->pstate.turbo_pstate;
1321 
1322 		/* Squash the global minimum into the permitted range. */
1323 		if (global.min_perf_pct > pct)
1324 			global.min_perf_pct = pct;
1325 	}
1326 
1327 	mutex_unlock(&intel_pstate_limits_lock);
1328 
1329 	intel_pstate_update_limits_for_all();
1330 	arch_set_max_freq_ratio(no_turbo);
1331 
1332 unlock_driver:
1333 	mutex_unlock(&intel_pstate_driver_lock);
1334 
1335 	return count;
1336 }
1337 
1338 static void update_qos_request(enum freq_qos_req_type type)
1339 {
1340 	struct freq_qos_request *req;
1341 	struct cpufreq_policy *policy;
1342 	int i;
1343 
1344 	for_each_possible_cpu(i) {
1345 		struct cpudata *cpu = all_cpu_data[i];
1346 		unsigned int freq, perf_pct;
1347 
1348 		policy = cpufreq_cpu_get(i);
1349 		if (!policy)
1350 			continue;
1351 
1352 		req = policy->driver_data;
1353 		cpufreq_cpu_put(policy);
1354 
1355 		if (!req)
1356 			continue;
1357 
1358 		if (hwp_active)
1359 			intel_pstate_get_hwp_cap(cpu);
1360 
1361 		if (type == FREQ_QOS_MIN) {
1362 			perf_pct = global.min_perf_pct;
1363 		} else {
1364 			req++;
1365 			perf_pct = global.max_perf_pct;
1366 		}
1367 
1368 		freq = DIV_ROUND_UP(cpu->pstate.turbo_freq * perf_pct, 100);
1369 
1370 		if (freq_qos_update_request(req, freq) < 0)
1371 			pr_warn("Failed to update freq constraint: CPU%d\n", i);
1372 	}
1373 }
1374 
1375 static ssize_t store_max_perf_pct(struct kobject *a, struct kobj_attribute *b,
1376 				  const char *buf, size_t count)
1377 {
1378 	unsigned int input;
1379 	int ret;
1380 
1381 	ret = sscanf(buf, "%u", &input);
1382 	if (ret != 1)
1383 		return -EINVAL;
1384 
1385 	mutex_lock(&intel_pstate_driver_lock);
1386 
1387 	if (!intel_pstate_driver) {
1388 		mutex_unlock(&intel_pstate_driver_lock);
1389 		return -EAGAIN;
1390 	}
1391 
1392 	mutex_lock(&intel_pstate_limits_lock);
1393 
1394 	global.max_perf_pct = clamp_t(int, input, global.min_perf_pct, 100);
1395 
1396 	mutex_unlock(&intel_pstate_limits_lock);
1397 
1398 	if (intel_pstate_driver == &intel_pstate)
1399 		intel_pstate_update_policies();
1400 	else
1401 		update_qos_request(FREQ_QOS_MAX);
1402 
1403 	mutex_unlock(&intel_pstate_driver_lock);
1404 
1405 	return count;
1406 }
1407 
1408 static ssize_t store_min_perf_pct(struct kobject *a, struct kobj_attribute *b,
1409 				  const char *buf, size_t count)
1410 {
1411 	unsigned int input;
1412 	int ret;
1413 
1414 	ret = sscanf(buf, "%u", &input);
1415 	if (ret != 1)
1416 		return -EINVAL;
1417 
1418 	mutex_lock(&intel_pstate_driver_lock);
1419 
1420 	if (!intel_pstate_driver) {
1421 		mutex_unlock(&intel_pstate_driver_lock);
1422 		return -EAGAIN;
1423 	}
1424 
1425 	mutex_lock(&intel_pstate_limits_lock);
1426 
1427 	global.min_perf_pct = clamp_t(int, input,
1428 				      min_perf_pct_min(), global.max_perf_pct);
1429 
1430 	mutex_unlock(&intel_pstate_limits_lock);
1431 
1432 	if (intel_pstate_driver == &intel_pstate)
1433 		intel_pstate_update_policies();
1434 	else
1435 		update_qos_request(FREQ_QOS_MIN);
1436 
1437 	mutex_unlock(&intel_pstate_driver_lock);
1438 
1439 	return count;
1440 }
1441 
1442 static ssize_t show_hwp_dynamic_boost(struct kobject *kobj,
1443 				struct kobj_attribute *attr, char *buf)
1444 {
1445 	return sprintf(buf, "%u\n", hwp_boost);
1446 }
1447 
1448 static ssize_t store_hwp_dynamic_boost(struct kobject *a,
1449 				       struct kobj_attribute *b,
1450 				       const char *buf, size_t count)
1451 {
1452 	unsigned int input;
1453 	int ret;
1454 
1455 	ret = kstrtouint(buf, 10, &input);
1456 	if (ret)
1457 		return ret;
1458 
1459 	mutex_lock(&intel_pstate_driver_lock);
1460 	hwp_boost = !!input;
1461 	intel_pstate_update_policies();
1462 	mutex_unlock(&intel_pstate_driver_lock);
1463 
1464 	return count;
1465 }
1466 
1467 static ssize_t show_energy_efficiency(struct kobject *kobj, struct kobj_attribute *attr,
1468 				      char *buf)
1469 {
1470 	u64 power_ctl;
1471 	int enable;
1472 
1473 	rdmsrl(MSR_IA32_POWER_CTL, power_ctl);
1474 	enable = !!(power_ctl & BIT(MSR_IA32_POWER_CTL_BIT_EE));
1475 	return sprintf(buf, "%d\n", !enable);
1476 }
1477 
1478 static ssize_t store_energy_efficiency(struct kobject *a, struct kobj_attribute *b,
1479 				       const char *buf, size_t count)
1480 {
1481 	bool input;
1482 	int ret;
1483 
1484 	ret = kstrtobool(buf, &input);
1485 	if (ret)
1486 		return ret;
1487 
1488 	set_power_ctl_ee_state(input);
1489 
1490 	return count;
1491 }
1492 
1493 show_one(max_perf_pct, max_perf_pct);
1494 show_one(min_perf_pct, min_perf_pct);
1495 
1496 define_one_global_rw(status);
1497 define_one_global_rw(no_turbo);
1498 define_one_global_rw(max_perf_pct);
1499 define_one_global_rw(min_perf_pct);
1500 define_one_global_ro(turbo_pct);
1501 define_one_global_ro(num_pstates);
1502 define_one_global_rw(hwp_dynamic_boost);
1503 define_one_global_rw(energy_efficiency);
1504 
1505 static struct attribute *intel_pstate_attributes[] = {
1506 	&status.attr,
1507 	&no_turbo.attr,
1508 	NULL
1509 };
1510 
1511 static const struct attribute_group intel_pstate_attr_group = {
1512 	.attrs = intel_pstate_attributes,
1513 };
1514 
1515 static const struct x86_cpu_id intel_pstate_cpu_ee_disable_ids[];
1516 
1517 static struct kobject *intel_pstate_kobject;
1518 
1519 static void __init intel_pstate_sysfs_expose_params(void)
1520 {
1521 	struct device *dev_root = bus_get_dev_root(&cpu_subsys);
1522 	int rc;
1523 
1524 	if (dev_root) {
1525 		intel_pstate_kobject = kobject_create_and_add("intel_pstate", &dev_root->kobj);
1526 		put_device(dev_root);
1527 	}
1528 	if (WARN_ON(!intel_pstate_kobject))
1529 		return;
1530 
1531 	rc = sysfs_create_group(intel_pstate_kobject, &intel_pstate_attr_group);
1532 	if (WARN_ON(rc))
1533 		return;
1534 
1535 	if (!boot_cpu_has(X86_FEATURE_HYBRID_CPU)) {
1536 		rc = sysfs_create_file(intel_pstate_kobject, &turbo_pct.attr);
1537 		WARN_ON(rc);
1538 
1539 		rc = sysfs_create_file(intel_pstate_kobject, &num_pstates.attr);
1540 		WARN_ON(rc);
1541 	}
1542 
1543 	/*
1544 	 * If per cpu limits are enforced there are no global limits, so
1545 	 * return without creating max/min_perf_pct attributes
1546 	 */
1547 	if (per_cpu_limits)
1548 		return;
1549 
1550 	rc = sysfs_create_file(intel_pstate_kobject, &max_perf_pct.attr);
1551 	WARN_ON(rc);
1552 
1553 	rc = sysfs_create_file(intel_pstate_kobject, &min_perf_pct.attr);
1554 	WARN_ON(rc);
1555 
1556 	if (x86_match_cpu(intel_pstate_cpu_ee_disable_ids)) {
1557 		rc = sysfs_create_file(intel_pstate_kobject, &energy_efficiency.attr);
1558 		WARN_ON(rc);
1559 	}
1560 }
1561 
1562 static void __init intel_pstate_sysfs_remove(void)
1563 {
1564 	if (!intel_pstate_kobject)
1565 		return;
1566 
1567 	sysfs_remove_group(intel_pstate_kobject, &intel_pstate_attr_group);
1568 
1569 	if (!boot_cpu_has(X86_FEATURE_HYBRID_CPU)) {
1570 		sysfs_remove_file(intel_pstate_kobject, &num_pstates.attr);
1571 		sysfs_remove_file(intel_pstate_kobject, &turbo_pct.attr);
1572 	}
1573 
1574 	if (!per_cpu_limits) {
1575 		sysfs_remove_file(intel_pstate_kobject, &max_perf_pct.attr);
1576 		sysfs_remove_file(intel_pstate_kobject, &min_perf_pct.attr);
1577 
1578 		if (x86_match_cpu(intel_pstate_cpu_ee_disable_ids))
1579 			sysfs_remove_file(intel_pstate_kobject, &energy_efficiency.attr);
1580 	}
1581 
1582 	kobject_put(intel_pstate_kobject);
1583 }
1584 
1585 static void intel_pstate_sysfs_expose_hwp_dynamic_boost(void)
1586 {
1587 	int rc;
1588 
1589 	if (!hwp_active)
1590 		return;
1591 
1592 	rc = sysfs_create_file(intel_pstate_kobject, &hwp_dynamic_boost.attr);
1593 	WARN_ON_ONCE(rc);
1594 }
1595 
1596 static void intel_pstate_sysfs_hide_hwp_dynamic_boost(void)
1597 {
1598 	if (!hwp_active)
1599 		return;
1600 
1601 	sysfs_remove_file(intel_pstate_kobject, &hwp_dynamic_boost.attr);
1602 }
1603 
1604 /************************** sysfs end ************************/
1605 
1606 static void intel_pstate_notify_work(struct work_struct *work)
1607 {
1608 	struct cpudata *cpudata =
1609 		container_of(to_delayed_work(work), struct cpudata, hwp_notify_work);
1610 	struct cpufreq_policy *policy = cpufreq_cpu_acquire(cpudata->cpu);
1611 
1612 	if (policy) {
1613 		__intel_pstate_update_max_freq(cpudata, policy);
1614 
1615 		cpufreq_cpu_release(policy);
1616 	}
1617 
1618 	wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_STATUS, 0);
1619 }
1620 
1621 static DEFINE_SPINLOCK(hwp_notify_lock);
1622 static cpumask_t hwp_intr_enable_mask;
1623 
1624 void notify_hwp_interrupt(void)
1625 {
1626 	unsigned int this_cpu = smp_processor_id();
1627 	unsigned long flags;
1628 	u64 value;
1629 
1630 	if (!hwp_active || !boot_cpu_has(X86_FEATURE_HWP_NOTIFY))
1631 		return;
1632 
1633 	rdmsrl_safe(MSR_HWP_STATUS, &value);
1634 	if (!(value & 0x01))
1635 		return;
1636 
1637 	spin_lock_irqsave(&hwp_notify_lock, flags);
1638 
1639 	if (!cpumask_test_cpu(this_cpu, &hwp_intr_enable_mask))
1640 		goto ack_intr;
1641 
1642 	schedule_delayed_work(&all_cpu_data[this_cpu]->hwp_notify_work,
1643 			      msecs_to_jiffies(10));
1644 
1645 	spin_unlock_irqrestore(&hwp_notify_lock, flags);
1646 
1647 	return;
1648 
1649 ack_intr:
1650 	wrmsrl_safe(MSR_HWP_STATUS, 0);
1651 	spin_unlock_irqrestore(&hwp_notify_lock, flags);
1652 }
1653 
1654 static void intel_pstate_disable_hwp_interrupt(struct cpudata *cpudata)
1655 {
1656 	bool cancel_work;
1657 
1658 	if (!boot_cpu_has(X86_FEATURE_HWP_NOTIFY))
1659 		return;
1660 
1661 	/* wrmsrl_on_cpu has to be outside spinlock as this can result in IPC */
1662 	wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_INTERRUPT, 0x00);
1663 
1664 	spin_lock_irq(&hwp_notify_lock);
1665 	cancel_work = cpumask_test_and_clear_cpu(cpudata->cpu, &hwp_intr_enable_mask);
1666 	spin_unlock_irq(&hwp_notify_lock);
1667 
1668 	if (cancel_work)
1669 		cancel_delayed_work_sync(&cpudata->hwp_notify_work);
1670 }
1671 
1672 static void intel_pstate_enable_hwp_interrupt(struct cpudata *cpudata)
1673 {
1674 	/* Enable HWP notification interrupt for guaranteed performance change */
1675 	if (boot_cpu_has(X86_FEATURE_HWP_NOTIFY)) {
1676 		spin_lock_irq(&hwp_notify_lock);
1677 		INIT_DELAYED_WORK(&cpudata->hwp_notify_work, intel_pstate_notify_work);
1678 		cpumask_set_cpu(cpudata->cpu, &hwp_intr_enable_mask);
1679 		spin_unlock_irq(&hwp_notify_lock);
1680 
1681 		/* wrmsrl_on_cpu has to be outside spinlock as this can result in IPC */
1682 		wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_INTERRUPT, 0x01);
1683 		wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_STATUS, 0);
1684 	}
1685 }
1686 
1687 static void intel_pstate_update_epp_defaults(struct cpudata *cpudata)
1688 {
1689 	cpudata->epp_default = intel_pstate_get_epp(cpudata, 0);
1690 
1691 	/*
1692 	 * If the EPP is set by firmware, which means that firmware enabled HWP
1693 	 * - Is equal or less than 0x80 (default balance_perf EPP)
1694 	 * - But less performance oriented than performance EPP
1695 	 *   then use this as new balance_perf EPP.
1696 	 */
1697 	if (hwp_forced && cpudata->epp_default <= HWP_EPP_BALANCE_PERFORMANCE &&
1698 	    cpudata->epp_default > HWP_EPP_PERFORMANCE) {
1699 		epp_values[EPP_INDEX_BALANCE_PERFORMANCE] = cpudata->epp_default;
1700 		return;
1701 	}
1702 
1703 	/*
1704 	 * If this CPU gen doesn't call for change in balance_perf
1705 	 * EPP return.
1706 	 */
1707 	if (epp_values[EPP_INDEX_BALANCE_PERFORMANCE] == HWP_EPP_BALANCE_PERFORMANCE)
1708 		return;
1709 
1710 	/*
1711 	 * Use hard coded value per gen to update the balance_perf
1712 	 * and default EPP.
1713 	 */
1714 	cpudata->epp_default = epp_values[EPP_INDEX_BALANCE_PERFORMANCE];
1715 	intel_pstate_set_epp(cpudata, cpudata->epp_default);
1716 }
1717 
1718 static void intel_pstate_hwp_enable(struct cpudata *cpudata)
1719 {
1720 	/* First disable HWP notification interrupt till we activate again */
1721 	if (boot_cpu_has(X86_FEATURE_HWP_NOTIFY))
1722 		wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_INTERRUPT, 0x00);
1723 
1724 	wrmsrl_on_cpu(cpudata->cpu, MSR_PM_ENABLE, 0x1);
1725 
1726 	intel_pstate_enable_hwp_interrupt(cpudata);
1727 
1728 	if (cpudata->epp_default >= 0)
1729 		return;
1730 
1731 	intel_pstate_update_epp_defaults(cpudata);
1732 }
1733 
1734 static int atom_get_min_pstate(int not_used)
1735 {
1736 	u64 value;
1737 
1738 	rdmsrl(MSR_ATOM_CORE_RATIOS, value);
1739 	return (value >> 8) & 0x7F;
1740 }
1741 
1742 static int atom_get_max_pstate(int not_used)
1743 {
1744 	u64 value;
1745 
1746 	rdmsrl(MSR_ATOM_CORE_RATIOS, value);
1747 	return (value >> 16) & 0x7F;
1748 }
1749 
1750 static int atom_get_turbo_pstate(int not_used)
1751 {
1752 	u64 value;
1753 
1754 	rdmsrl(MSR_ATOM_CORE_TURBO_RATIOS, value);
1755 	return value & 0x7F;
1756 }
1757 
1758 static u64 atom_get_val(struct cpudata *cpudata, int pstate)
1759 {
1760 	u64 val;
1761 	int32_t vid_fp;
1762 	u32 vid;
1763 
1764 	val = (u64)pstate << 8;
1765 	if (READ_ONCE(global.no_turbo) && !global.turbo_disabled)
1766 		val |= (u64)1 << 32;
1767 
1768 	vid_fp = cpudata->vid.min + mul_fp(
1769 		int_tofp(pstate - cpudata->pstate.min_pstate),
1770 		cpudata->vid.ratio);
1771 
1772 	vid_fp = clamp_t(int32_t, vid_fp, cpudata->vid.min, cpudata->vid.max);
1773 	vid = ceiling_fp(vid_fp);
1774 
1775 	if (pstate > cpudata->pstate.max_pstate)
1776 		vid = cpudata->vid.turbo;
1777 
1778 	return val | vid;
1779 }
1780 
1781 static int silvermont_get_scaling(void)
1782 {
1783 	u64 value;
1784 	int i;
1785 	/* Defined in Table 35-6 from SDM (Sept 2015) */
1786 	static int silvermont_freq_table[] = {
1787 		83300, 100000, 133300, 116700, 80000};
1788 
1789 	rdmsrl(MSR_FSB_FREQ, value);
1790 	i = value & 0x7;
1791 	WARN_ON(i > 4);
1792 
1793 	return silvermont_freq_table[i];
1794 }
1795 
1796 static int airmont_get_scaling(void)
1797 {
1798 	u64 value;
1799 	int i;
1800 	/* Defined in Table 35-10 from SDM (Sept 2015) */
1801 	static int airmont_freq_table[] = {
1802 		83300, 100000, 133300, 116700, 80000,
1803 		93300, 90000, 88900, 87500};
1804 
1805 	rdmsrl(MSR_FSB_FREQ, value);
1806 	i = value & 0xF;
1807 	WARN_ON(i > 8);
1808 
1809 	return airmont_freq_table[i];
1810 }
1811 
1812 static void atom_get_vid(struct cpudata *cpudata)
1813 {
1814 	u64 value;
1815 
1816 	rdmsrl(MSR_ATOM_CORE_VIDS, value);
1817 	cpudata->vid.min = int_tofp((value >> 8) & 0x7f);
1818 	cpudata->vid.max = int_tofp((value >> 16) & 0x7f);
1819 	cpudata->vid.ratio = div_fp(
1820 		cpudata->vid.max - cpudata->vid.min,
1821 		int_tofp(cpudata->pstate.max_pstate -
1822 			cpudata->pstate.min_pstate));
1823 
1824 	rdmsrl(MSR_ATOM_CORE_TURBO_VIDS, value);
1825 	cpudata->vid.turbo = value & 0x7f;
1826 }
1827 
1828 static int core_get_min_pstate(int cpu)
1829 {
1830 	u64 value;
1831 
1832 	rdmsrl_on_cpu(cpu, MSR_PLATFORM_INFO, &value);
1833 	return (value >> 40) & 0xFF;
1834 }
1835 
1836 static int core_get_max_pstate_physical(int cpu)
1837 {
1838 	u64 value;
1839 
1840 	rdmsrl_on_cpu(cpu, MSR_PLATFORM_INFO, &value);
1841 	return (value >> 8) & 0xFF;
1842 }
1843 
1844 static int core_get_tdp_ratio(int cpu, u64 plat_info)
1845 {
1846 	/* Check how many TDP levels present */
1847 	if (plat_info & 0x600000000) {
1848 		u64 tdp_ctrl;
1849 		u64 tdp_ratio;
1850 		int tdp_msr;
1851 		int err;
1852 
1853 		/* Get the TDP level (0, 1, 2) to get ratios */
1854 		err = rdmsrl_safe_on_cpu(cpu, MSR_CONFIG_TDP_CONTROL, &tdp_ctrl);
1855 		if (err)
1856 			return err;
1857 
1858 		/* TDP MSR are continuous starting at 0x648 */
1859 		tdp_msr = MSR_CONFIG_TDP_NOMINAL + (tdp_ctrl & 0x03);
1860 		err = rdmsrl_safe_on_cpu(cpu, tdp_msr, &tdp_ratio);
1861 		if (err)
1862 			return err;
1863 
1864 		/* For level 1 and 2, bits[23:16] contain the ratio */
1865 		if (tdp_ctrl & 0x03)
1866 			tdp_ratio >>= 16;
1867 
1868 		tdp_ratio &= 0xff; /* ratios are only 8 bits long */
1869 		pr_debug("tdp_ratio %x\n", (int)tdp_ratio);
1870 
1871 		return (int)tdp_ratio;
1872 	}
1873 
1874 	return -ENXIO;
1875 }
1876 
1877 static int core_get_max_pstate(int cpu)
1878 {
1879 	u64 tar;
1880 	u64 plat_info;
1881 	int max_pstate;
1882 	int tdp_ratio;
1883 	int err;
1884 
1885 	rdmsrl_on_cpu(cpu, MSR_PLATFORM_INFO, &plat_info);
1886 	max_pstate = (plat_info >> 8) & 0xFF;
1887 
1888 	tdp_ratio = core_get_tdp_ratio(cpu, plat_info);
1889 	if (tdp_ratio <= 0)
1890 		return max_pstate;
1891 
1892 	if (hwp_active) {
1893 		/* Turbo activation ratio is not used on HWP platforms */
1894 		return tdp_ratio;
1895 	}
1896 
1897 	err = rdmsrl_safe_on_cpu(cpu, MSR_TURBO_ACTIVATION_RATIO, &tar);
1898 	if (!err) {
1899 		int tar_levels;
1900 
1901 		/* Do some sanity checking for safety */
1902 		tar_levels = tar & 0xff;
1903 		if (tdp_ratio - 1 == tar_levels) {
1904 			max_pstate = tar_levels;
1905 			pr_debug("max_pstate=TAC %x\n", max_pstate);
1906 		}
1907 	}
1908 
1909 	return max_pstate;
1910 }
1911 
1912 static int core_get_turbo_pstate(int cpu)
1913 {
1914 	u64 value;
1915 	int nont, ret;
1916 
1917 	rdmsrl_on_cpu(cpu, MSR_TURBO_RATIO_LIMIT, &value);
1918 	nont = core_get_max_pstate(cpu);
1919 	ret = (value) & 255;
1920 	if (ret <= nont)
1921 		ret = nont;
1922 	return ret;
1923 }
1924 
1925 static u64 core_get_val(struct cpudata *cpudata, int pstate)
1926 {
1927 	u64 val;
1928 
1929 	val = (u64)pstate << 8;
1930 	if (READ_ONCE(global.no_turbo) && !global.turbo_disabled)
1931 		val |= (u64)1 << 32;
1932 
1933 	return val;
1934 }
1935 
1936 static int knl_get_aperf_mperf_shift(void)
1937 {
1938 	return 10;
1939 }
1940 
1941 static int knl_get_turbo_pstate(int cpu)
1942 {
1943 	u64 value;
1944 	int nont, ret;
1945 
1946 	rdmsrl_on_cpu(cpu, MSR_TURBO_RATIO_LIMIT, &value);
1947 	nont = core_get_max_pstate(cpu);
1948 	ret = (((value) >> 8) & 0xFF);
1949 	if (ret <= nont)
1950 		ret = nont;
1951 	return ret;
1952 }
1953 
1954 static void hybrid_get_type(void *data)
1955 {
1956 	u8 *cpu_type = data;
1957 
1958 	*cpu_type = get_this_hybrid_cpu_type();
1959 }
1960 
1961 static int hwp_get_cpu_scaling(int cpu)
1962 {
1963 	u8 cpu_type = 0;
1964 
1965 	smp_call_function_single(cpu, hybrid_get_type, &cpu_type, 1);
1966 	/* P-cores have a smaller perf level-to-freqency scaling factor. */
1967 	if (cpu_type == 0x40)
1968 		return hybrid_scaling_factor;
1969 
1970 	/* Use default core scaling for E-cores */
1971 	if (cpu_type == 0x20)
1972 		return core_get_scaling();
1973 
1974 	/*
1975 	 * If reached here, this system is either non-hybrid (like Tiger
1976 	 * Lake) or hybrid-capable (like Alder Lake or Raptor Lake) with
1977 	 * no E cores (in which case CPUID for hybrid support is 0).
1978 	 *
1979 	 * The CPPC nominal_frequency field is 0 for non-hybrid systems,
1980 	 * so the default core scaling will be used for them.
1981 	 */
1982 	return intel_pstate_cppc_get_scaling(cpu);
1983 }
1984 
1985 static void intel_pstate_set_pstate(struct cpudata *cpu, int pstate)
1986 {
1987 	trace_cpu_frequency(pstate * cpu->pstate.scaling, cpu->cpu);
1988 	cpu->pstate.current_pstate = pstate;
1989 	/*
1990 	 * Generally, there is no guarantee that this code will always run on
1991 	 * the CPU being updated, so force the register update to run on the
1992 	 * right CPU.
1993 	 */
1994 	wrmsrl_on_cpu(cpu->cpu, MSR_IA32_PERF_CTL,
1995 		      pstate_funcs.get_val(cpu, pstate));
1996 }
1997 
1998 static void intel_pstate_set_min_pstate(struct cpudata *cpu)
1999 {
2000 	intel_pstate_set_pstate(cpu, cpu->pstate.min_pstate);
2001 }
2002 
2003 static void intel_pstate_get_cpu_pstates(struct cpudata *cpu)
2004 {
2005 	int perf_ctl_max_phys = pstate_funcs.get_max_physical(cpu->cpu);
2006 	int perf_ctl_scaling = pstate_funcs.get_scaling();
2007 
2008 	cpu->pstate.min_pstate = pstate_funcs.get_min(cpu->cpu);
2009 	cpu->pstate.max_pstate_physical = perf_ctl_max_phys;
2010 	cpu->pstate.perf_ctl_scaling = perf_ctl_scaling;
2011 
2012 	if (hwp_active && !hwp_mode_bdw) {
2013 		__intel_pstate_get_hwp_cap(cpu);
2014 
2015 		if (pstate_funcs.get_cpu_scaling) {
2016 			cpu->pstate.scaling = pstate_funcs.get_cpu_scaling(cpu->cpu);
2017 			if (cpu->pstate.scaling != perf_ctl_scaling)
2018 				intel_pstate_hybrid_hwp_adjust(cpu);
2019 		} else {
2020 			cpu->pstate.scaling = perf_ctl_scaling;
2021 		}
2022 	} else {
2023 		cpu->pstate.scaling = perf_ctl_scaling;
2024 		cpu->pstate.max_pstate = pstate_funcs.get_max(cpu->cpu);
2025 		cpu->pstate.turbo_pstate = pstate_funcs.get_turbo(cpu->cpu);
2026 	}
2027 
2028 	if (cpu->pstate.scaling == perf_ctl_scaling) {
2029 		cpu->pstate.min_freq = cpu->pstate.min_pstate * perf_ctl_scaling;
2030 		cpu->pstate.max_freq = cpu->pstate.max_pstate * perf_ctl_scaling;
2031 		cpu->pstate.turbo_freq = cpu->pstate.turbo_pstate * perf_ctl_scaling;
2032 	}
2033 
2034 	if (pstate_funcs.get_aperf_mperf_shift)
2035 		cpu->aperf_mperf_shift = pstate_funcs.get_aperf_mperf_shift();
2036 
2037 	if (pstate_funcs.get_vid)
2038 		pstate_funcs.get_vid(cpu);
2039 
2040 	intel_pstate_set_min_pstate(cpu);
2041 }
2042 
2043 /*
2044  * Long hold time will keep high perf limits for long time,
2045  * which negatively impacts perf/watt for some workloads,
2046  * like specpower. 3ms is based on experiements on some
2047  * workoads.
2048  */
2049 static int hwp_boost_hold_time_ns = 3 * NSEC_PER_MSEC;
2050 
2051 static inline void intel_pstate_hwp_boost_up(struct cpudata *cpu)
2052 {
2053 	u64 hwp_req = READ_ONCE(cpu->hwp_req_cached);
2054 	u64 hwp_cap = READ_ONCE(cpu->hwp_cap_cached);
2055 	u32 max_limit = (hwp_req & 0xff00) >> 8;
2056 	u32 min_limit = (hwp_req & 0xff);
2057 	u32 boost_level1;
2058 
2059 	/*
2060 	 * Cases to consider (User changes via sysfs or boot time):
2061 	 * If, P0 (Turbo max) = P1 (Guaranteed max) = min:
2062 	 *	No boost, return.
2063 	 * If, P0 (Turbo max) > P1 (Guaranteed max) = min:
2064 	 *     Should result in one level boost only for P0.
2065 	 * If, P0 (Turbo max) = P1 (Guaranteed max) > min:
2066 	 *     Should result in two level boost:
2067 	 *         (min + p1)/2 and P1.
2068 	 * If, P0 (Turbo max) > P1 (Guaranteed max) > min:
2069 	 *     Should result in three level boost:
2070 	 *        (min + p1)/2, P1 and P0.
2071 	 */
2072 
2073 	/* If max and min are equal or already at max, nothing to boost */
2074 	if (max_limit == min_limit || cpu->hwp_boost_min >= max_limit)
2075 		return;
2076 
2077 	if (!cpu->hwp_boost_min)
2078 		cpu->hwp_boost_min = min_limit;
2079 
2080 	/* level at half way mark between min and guranteed */
2081 	boost_level1 = (HWP_GUARANTEED_PERF(hwp_cap) + min_limit) >> 1;
2082 
2083 	if (cpu->hwp_boost_min < boost_level1)
2084 		cpu->hwp_boost_min = boost_level1;
2085 	else if (cpu->hwp_boost_min < HWP_GUARANTEED_PERF(hwp_cap))
2086 		cpu->hwp_boost_min = HWP_GUARANTEED_PERF(hwp_cap);
2087 	else if (cpu->hwp_boost_min == HWP_GUARANTEED_PERF(hwp_cap) &&
2088 		 max_limit != HWP_GUARANTEED_PERF(hwp_cap))
2089 		cpu->hwp_boost_min = max_limit;
2090 	else
2091 		return;
2092 
2093 	hwp_req = (hwp_req & ~GENMASK_ULL(7, 0)) | cpu->hwp_boost_min;
2094 	wrmsrl(MSR_HWP_REQUEST, hwp_req);
2095 	cpu->last_update = cpu->sample.time;
2096 }
2097 
2098 static inline void intel_pstate_hwp_boost_down(struct cpudata *cpu)
2099 {
2100 	if (cpu->hwp_boost_min) {
2101 		bool expired;
2102 
2103 		/* Check if we are idle for hold time to boost down */
2104 		expired = time_after64(cpu->sample.time, cpu->last_update +
2105 				       hwp_boost_hold_time_ns);
2106 		if (expired) {
2107 			wrmsrl(MSR_HWP_REQUEST, cpu->hwp_req_cached);
2108 			cpu->hwp_boost_min = 0;
2109 		}
2110 	}
2111 	cpu->last_update = cpu->sample.time;
2112 }
2113 
2114 static inline void intel_pstate_update_util_hwp_local(struct cpudata *cpu,
2115 						      u64 time)
2116 {
2117 	cpu->sample.time = time;
2118 
2119 	if (cpu->sched_flags & SCHED_CPUFREQ_IOWAIT) {
2120 		bool do_io = false;
2121 
2122 		cpu->sched_flags = 0;
2123 		/*
2124 		 * Set iowait_boost flag and update time. Since IO WAIT flag
2125 		 * is set all the time, we can't just conclude that there is
2126 		 * some IO bound activity is scheduled on this CPU with just
2127 		 * one occurrence. If we receive at least two in two
2128 		 * consecutive ticks, then we treat as boost candidate.
2129 		 */
2130 		if (time_before64(time, cpu->last_io_update + 2 * TICK_NSEC))
2131 			do_io = true;
2132 
2133 		cpu->last_io_update = time;
2134 
2135 		if (do_io)
2136 			intel_pstate_hwp_boost_up(cpu);
2137 
2138 	} else {
2139 		intel_pstate_hwp_boost_down(cpu);
2140 	}
2141 }
2142 
2143 static inline void intel_pstate_update_util_hwp(struct update_util_data *data,
2144 						u64 time, unsigned int flags)
2145 {
2146 	struct cpudata *cpu = container_of(data, struct cpudata, update_util);
2147 
2148 	cpu->sched_flags |= flags;
2149 
2150 	if (smp_processor_id() == cpu->cpu)
2151 		intel_pstate_update_util_hwp_local(cpu, time);
2152 }
2153 
2154 static inline void intel_pstate_calc_avg_perf(struct cpudata *cpu)
2155 {
2156 	struct sample *sample = &cpu->sample;
2157 
2158 	sample->core_avg_perf = div_ext_fp(sample->aperf, sample->mperf);
2159 }
2160 
2161 static inline bool intel_pstate_sample(struct cpudata *cpu, u64 time)
2162 {
2163 	u64 aperf, mperf;
2164 	unsigned long flags;
2165 	u64 tsc;
2166 
2167 	local_irq_save(flags);
2168 	rdmsrl(MSR_IA32_APERF, aperf);
2169 	rdmsrl(MSR_IA32_MPERF, mperf);
2170 	tsc = rdtsc();
2171 	if (cpu->prev_mperf == mperf || cpu->prev_tsc == tsc) {
2172 		local_irq_restore(flags);
2173 		return false;
2174 	}
2175 	local_irq_restore(flags);
2176 
2177 	cpu->last_sample_time = cpu->sample.time;
2178 	cpu->sample.time = time;
2179 	cpu->sample.aperf = aperf;
2180 	cpu->sample.mperf = mperf;
2181 	cpu->sample.tsc =  tsc;
2182 	cpu->sample.aperf -= cpu->prev_aperf;
2183 	cpu->sample.mperf -= cpu->prev_mperf;
2184 	cpu->sample.tsc -= cpu->prev_tsc;
2185 
2186 	cpu->prev_aperf = aperf;
2187 	cpu->prev_mperf = mperf;
2188 	cpu->prev_tsc = tsc;
2189 	/*
2190 	 * First time this function is invoked in a given cycle, all of the
2191 	 * previous sample data fields are equal to zero or stale and they must
2192 	 * be populated with meaningful numbers for things to work, so assume
2193 	 * that sample.time will always be reset before setting the utilization
2194 	 * update hook and make the caller skip the sample then.
2195 	 */
2196 	if (cpu->last_sample_time) {
2197 		intel_pstate_calc_avg_perf(cpu);
2198 		return true;
2199 	}
2200 	return false;
2201 }
2202 
2203 static inline int32_t get_avg_frequency(struct cpudata *cpu)
2204 {
2205 	return mul_ext_fp(cpu->sample.core_avg_perf, cpu_khz);
2206 }
2207 
2208 static inline int32_t get_avg_pstate(struct cpudata *cpu)
2209 {
2210 	return mul_ext_fp(cpu->pstate.max_pstate_physical,
2211 			  cpu->sample.core_avg_perf);
2212 }
2213 
2214 static inline int32_t get_target_pstate(struct cpudata *cpu)
2215 {
2216 	struct sample *sample = &cpu->sample;
2217 	int32_t busy_frac;
2218 	int target, avg_pstate;
2219 
2220 	busy_frac = div_fp(sample->mperf << cpu->aperf_mperf_shift,
2221 			   sample->tsc);
2222 
2223 	if (busy_frac < cpu->iowait_boost)
2224 		busy_frac = cpu->iowait_boost;
2225 
2226 	sample->busy_scaled = busy_frac * 100;
2227 
2228 	target = READ_ONCE(global.no_turbo) ?
2229 			cpu->pstate.max_pstate : cpu->pstate.turbo_pstate;
2230 	target += target >> 2;
2231 	target = mul_fp(target, busy_frac);
2232 	if (target < cpu->pstate.min_pstate)
2233 		target = cpu->pstate.min_pstate;
2234 
2235 	/*
2236 	 * If the average P-state during the previous cycle was higher than the
2237 	 * current target, add 50% of the difference to the target to reduce
2238 	 * possible performance oscillations and offset possible performance
2239 	 * loss related to moving the workload from one CPU to another within
2240 	 * a package/module.
2241 	 */
2242 	avg_pstate = get_avg_pstate(cpu);
2243 	if (avg_pstate > target)
2244 		target += (avg_pstate - target) >> 1;
2245 
2246 	return target;
2247 }
2248 
2249 static int intel_pstate_prepare_request(struct cpudata *cpu, int pstate)
2250 {
2251 	int min_pstate = max(cpu->pstate.min_pstate, cpu->min_perf_ratio);
2252 	int max_pstate = max(min_pstate, cpu->max_perf_ratio);
2253 
2254 	return clamp_t(int, pstate, min_pstate, max_pstate);
2255 }
2256 
2257 static void intel_pstate_update_pstate(struct cpudata *cpu, int pstate)
2258 {
2259 	if (pstate == cpu->pstate.current_pstate)
2260 		return;
2261 
2262 	cpu->pstate.current_pstate = pstate;
2263 	wrmsrl(MSR_IA32_PERF_CTL, pstate_funcs.get_val(cpu, pstate));
2264 }
2265 
2266 static void intel_pstate_adjust_pstate(struct cpudata *cpu)
2267 {
2268 	int from = cpu->pstate.current_pstate;
2269 	struct sample *sample;
2270 	int target_pstate;
2271 
2272 	target_pstate = get_target_pstate(cpu);
2273 	target_pstate = intel_pstate_prepare_request(cpu, target_pstate);
2274 	trace_cpu_frequency(target_pstate * cpu->pstate.scaling, cpu->cpu);
2275 	intel_pstate_update_pstate(cpu, target_pstate);
2276 
2277 	sample = &cpu->sample;
2278 	trace_pstate_sample(mul_ext_fp(100, sample->core_avg_perf),
2279 		fp_toint(sample->busy_scaled),
2280 		from,
2281 		cpu->pstate.current_pstate,
2282 		sample->mperf,
2283 		sample->aperf,
2284 		sample->tsc,
2285 		get_avg_frequency(cpu),
2286 		fp_toint(cpu->iowait_boost * 100));
2287 }
2288 
2289 static void intel_pstate_update_util(struct update_util_data *data, u64 time,
2290 				     unsigned int flags)
2291 {
2292 	struct cpudata *cpu = container_of(data, struct cpudata, update_util);
2293 	u64 delta_ns;
2294 
2295 	/* Don't allow remote callbacks */
2296 	if (smp_processor_id() != cpu->cpu)
2297 		return;
2298 
2299 	delta_ns = time - cpu->last_update;
2300 	if (flags & SCHED_CPUFREQ_IOWAIT) {
2301 		/* Start over if the CPU may have been idle. */
2302 		if (delta_ns > TICK_NSEC) {
2303 			cpu->iowait_boost = ONE_EIGHTH_FP;
2304 		} else if (cpu->iowait_boost >= ONE_EIGHTH_FP) {
2305 			cpu->iowait_boost <<= 1;
2306 			if (cpu->iowait_boost > int_tofp(1))
2307 				cpu->iowait_boost = int_tofp(1);
2308 		} else {
2309 			cpu->iowait_boost = ONE_EIGHTH_FP;
2310 		}
2311 	} else if (cpu->iowait_boost) {
2312 		/* Clear iowait_boost if the CPU may have been idle. */
2313 		if (delta_ns > TICK_NSEC)
2314 			cpu->iowait_boost = 0;
2315 		else
2316 			cpu->iowait_boost >>= 1;
2317 	}
2318 	cpu->last_update = time;
2319 	delta_ns = time - cpu->sample.time;
2320 	if ((s64)delta_ns < INTEL_PSTATE_SAMPLING_INTERVAL)
2321 		return;
2322 
2323 	if (intel_pstate_sample(cpu, time))
2324 		intel_pstate_adjust_pstate(cpu);
2325 }
2326 
2327 static struct pstate_funcs core_funcs = {
2328 	.get_max = core_get_max_pstate,
2329 	.get_max_physical = core_get_max_pstate_physical,
2330 	.get_min = core_get_min_pstate,
2331 	.get_turbo = core_get_turbo_pstate,
2332 	.get_scaling = core_get_scaling,
2333 	.get_val = core_get_val,
2334 };
2335 
2336 static const struct pstate_funcs silvermont_funcs = {
2337 	.get_max = atom_get_max_pstate,
2338 	.get_max_physical = atom_get_max_pstate,
2339 	.get_min = atom_get_min_pstate,
2340 	.get_turbo = atom_get_turbo_pstate,
2341 	.get_val = atom_get_val,
2342 	.get_scaling = silvermont_get_scaling,
2343 	.get_vid = atom_get_vid,
2344 };
2345 
2346 static const struct pstate_funcs airmont_funcs = {
2347 	.get_max = atom_get_max_pstate,
2348 	.get_max_physical = atom_get_max_pstate,
2349 	.get_min = atom_get_min_pstate,
2350 	.get_turbo = atom_get_turbo_pstate,
2351 	.get_val = atom_get_val,
2352 	.get_scaling = airmont_get_scaling,
2353 	.get_vid = atom_get_vid,
2354 };
2355 
2356 static const struct pstate_funcs knl_funcs = {
2357 	.get_max = core_get_max_pstate,
2358 	.get_max_physical = core_get_max_pstate_physical,
2359 	.get_min = core_get_min_pstate,
2360 	.get_turbo = knl_get_turbo_pstate,
2361 	.get_aperf_mperf_shift = knl_get_aperf_mperf_shift,
2362 	.get_scaling = core_get_scaling,
2363 	.get_val = core_get_val,
2364 };
2365 
2366 #define X86_MATCH(model, policy)					 \
2367 	X86_MATCH_VENDOR_FAM_MODEL_FEATURE(INTEL, 6, INTEL_FAM6_##model, \
2368 					   X86_FEATURE_APERFMPERF, &policy)
2369 
2370 static const struct x86_cpu_id intel_pstate_cpu_ids[] = {
2371 	X86_MATCH(SANDYBRIDGE,		core_funcs),
2372 	X86_MATCH(SANDYBRIDGE_X,	core_funcs),
2373 	X86_MATCH(ATOM_SILVERMONT,	silvermont_funcs),
2374 	X86_MATCH(IVYBRIDGE,		core_funcs),
2375 	X86_MATCH(HASWELL,		core_funcs),
2376 	X86_MATCH(BROADWELL,		core_funcs),
2377 	X86_MATCH(IVYBRIDGE_X,		core_funcs),
2378 	X86_MATCH(HASWELL_X,		core_funcs),
2379 	X86_MATCH(HASWELL_L,		core_funcs),
2380 	X86_MATCH(HASWELL_G,		core_funcs),
2381 	X86_MATCH(BROADWELL_G,		core_funcs),
2382 	X86_MATCH(ATOM_AIRMONT,		airmont_funcs),
2383 	X86_MATCH(SKYLAKE_L,		core_funcs),
2384 	X86_MATCH(BROADWELL_X,		core_funcs),
2385 	X86_MATCH(SKYLAKE,		core_funcs),
2386 	X86_MATCH(BROADWELL_D,		core_funcs),
2387 	X86_MATCH(XEON_PHI_KNL,		knl_funcs),
2388 	X86_MATCH(XEON_PHI_KNM,		knl_funcs),
2389 	X86_MATCH(ATOM_GOLDMONT,	core_funcs),
2390 	X86_MATCH(ATOM_GOLDMONT_PLUS,	core_funcs),
2391 	X86_MATCH(SKYLAKE_X,		core_funcs),
2392 	X86_MATCH(COMETLAKE,		core_funcs),
2393 	X86_MATCH(ICELAKE_X,		core_funcs),
2394 	X86_MATCH(TIGERLAKE,		core_funcs),
2395 	X86_MATCH(SAPPHIRERAPIDS_X,	core_funcs),
2396 	X86_MATCH(EMERALDRAPIDS_X,      core_funcs),
2397 	{}
2398 };
2399 MODULE_DEVICE_TABLE(x86cpu, intel_pstate_cpu_ids);
2400 
2401 #ifdef CONFIG_ACPI
2402 static const struct x86_cpu_id intel_pstate_cpu_oob_ids[] __initconst = {
2403 	X86_MATCH(BROADWELL_D,		core_funcs),
2404 	X86_MATCH(BROADWELL_X,		core_funcs),
2405 	X86_MATCH(SKYLAKE_X,		core_funcs),
2406 	X86_MATCH(ICELAKE_X,		core_funcs),
2407 	X86_MATCH(SAPPHIRERAPIDS_X,	core_funcs),
2408 	{}
2409 };
2410 #endif
2411 
2412 static const struct x86_cpu_id intel_pstate_cpu_ee_disable_ids[] = {
2413 	X86_MATCH(KABYLAKE,		core_funcs),
2414 	{}
2415 };
2416 
2417 static int intel_pstate_init_cpu(unsigned int cpunum)
2418 {
2419 	struct cpudata *cpu;
2420 
2421 	cpu = all_cpu_data[cpunum];
2422 
2423 	if (!cpu) {
2424 		cpu = kzalloc(sizeof(*cpu), GFP_KERNEL);
2425 		if (!cpu)
2426 			return -ENOMEM;
2427 
2428 		WRITE_ONCE(all_cpu_data[cpunum], cpu);
2429 
2430 		cpu->cpu = cpunum;
2431 
2432 		cpu->epp_default = -EINVAL;
2433 
2434 		if (hwp_active) {
2435 			intel_pstate_hwp_enable(cpu);
2436 
2437 			if (intel_pstate_acpi_pm_profile_server())
2438 				hwp_boost = true;
2439 		}
2440 	} else if (hwp_active) {
2441 		/*
2442 		 * Re-enable HWP in case this happens after a resume from ACPI
2443 		 * S3 if the CPU was offline during the whole system/resume
2444 		 * cycle.
2445 		 */
2446 		intel_pstate_hwp_reenable(cpu);
2447 	}
2448 
2449 	cpu->epp_powersave = -EINVAL;
2450 	cpu->epp_policy = 0;
2451 
2452 	intel_pstate_get_cpu_pstates(cpu);
2453 
2454 	pr_debug("controlling: cpu %d\n", cpunum);
2455 
2456 	return 0;
2457 }
2458 
2459 static void intel_pstate_set_update_util_hook(unsigned int cpu_num)
2460 {
2461 	struct cpudata *cpu = all_cpu_data[cpu_num];
2462 
2463 	if (hwp_active && !hwp_boost)
2464 		return;
2465 
2466 	if (cpu->update_util_set)
2467 		return;
2468 
2469 	/* Prevent intel_pstate_update_util() from using stale data. */
2470 	cpu->sample.time = 0;
2471 	cpufreq_add_update_util_hook(cpu_num, &cpu->update_util,
2472 				     (hwp_active ?
2473 				      intel_pstate_update_util_hwp :
2474 				      intel_pstate_update_util));
2475 	cpu->update_util_set = true;
2476 }
2477 
2478 static void intel_pstate_clear_update_util_hook(unsigned int cpu)
2479 {
2480 	struct cpudata *cpu_data = all_cpu_data[cpu];
2481 
2482 	if (!cpu_data->update_util_set)
2483 		return;
2484 
2485 	cpufreq_remove_update_util_hook(cpu);
2486 	cpu_data->update_util_set = false;
2487 	synchronize_rcu();
2488 }
2489 
2490 static int intel_pstate_get_max_freq(struct cpudata *cpu)
2491 {
2492 	return READ_ONCE(global.no_turbo) ?
2493 			cpu->pstate.max_freq : cpu->pstate.turbo_freq;
2494 }
2495 
2496 static void intel_pstate_update_perf_limits(struct cpudata *cpu,
2497 					    unsigned int policy_min,
2498 					    unsigned int policy_max)
2499 {
2500 	int perf_ctl_scaling = cpu->pstate.perf_ctl_scaling;
2501 	int32_t max_policy_perf, min_policy_perf;
2502 
2503 	max_policy_perf = policy_max / perf_ctl_scaling;
2504 	if (policy_max == policy_min) {
2505 		min_policy_perf = max_policy_perf;
2506 	} else {
2507 		min_policy_perf = policy_min / perf_ctl_scaling;
2508 		min_policy_perf = clamp_t(int32_t, min_policy_perf,
2509 					  0, max_policy_perf);
2510 	}
2511 
2512 	/*
2513 	 * HWP needs some special consideration, because HWP_REQUEST uses
2514 	 * abstract values to represent performance rather than pure ratios.
2515 	 */
2516 	if (hwp_active && cpu->pstate.scaling != perf_ctl_scaling) {
2517 		int freq;
2518 
2519 		freq = max_policy_perf * perf_ctl_scaling;
2520 		max_policy_perf = intel_pstate_freq_to_hwp(cpu, freq);
2521 		freq = min_policy_perf * perf_ctl_scaling;
2522 		min_policy_perf = intel_pstate_freq_to_hwp(cpu, freq);
2523 	}
2524 
2525 	pr_debug("cpu:%d min_policy_perf:%d max_policy_perf:%d\n",
2526 		 cpu->cpu, min_policy_perf, max_policy_perf);
2527 
2528 	/* Normalize user input to [min_perf, max_perf] */
2529 	if (per_cpu_limits) {
2530 		cpu->min_perf_ratio = min_policy_perf;
2531 		cpu->max_perf_ratio = max_policy_perf;
2532 	} else {
2533 		int turbo_max = cpu->pstate.turbo_pstate;
2534 		int32_t global_min, global_max;
2535 
2536 		/* Global limits are in percent of the maximum turbo P-state. */
2537 		global_max = DIV_ROUND_UP(turbo_max * global.max_perf_pct, 100);
2538 		global_min = DIV_ROUND_UP(turbo_max * global.min_perf_pct, 100);
2539 		global_min = clamp_t(int32_t, global_min, 0, global_max);
2540 
2541 		pr_debug("cpu:%d global_min:%d global_max:%d\n", cpu->cpu,
2542 			 global_min, global_max);
2543 
2544 		cpu->min_perf_ratio = max(min_policy_perf, global_min);
2545 		cpu->min_perf_ratio = min(cpu->min_perf_ratio, max_policy_perf);
2546 		cpu->max_perf_ratio = min(max_policy_perf, global_max);
2547 		cpu->max_perf_ratio = max(min_policy_perf, cpu->max_perf_ratio);
2548 
2549 		/* Make sure min_perf <= max_perf */
2550 		cpu->min_perf_ratio = min(cpu->min_perf_ratio,
2551 					  cpu->max_perf_ratio);
2552 
2553 	}
2554 	pr_debug("cpu:%d max_perf_ratio:%d min_perf_ratio:%d\n", cpu->cpu,
2555 		 cpu->max_perf_ratio,
2556 		 cpu->min_perf_ratio);
2557 }
2558 
2559 static int intel_pstate_set_policy(struct cpufreq_policy *policy)
2560 {
2561 	struct cpudata *cpu;
2562 
2563 	if (!policy->cpuinfo.max_freq)
2564 		return -ENODEV;
2565 
2566 	pr_debug("set_policy cpuinfo.max %u policy->max %u\n",
2567 		 policy->cpuinfo.max_freq, policy->max);
2568 
2569 	cpu = all_cpu_data[policy->cpu];
2570 	cpu->policy = policy->policy;
2571 
2572 	mutex_lock(&intel_pstate_limits_lock);
2573 
2574 	intel_pstate_update_perf_limits(cpu, policy->min, policy->max);
2575 
2576 	if (cpu->policy == CPUFREQ_POLICY_PERFORMANCE) {
2577 		int pstate = max(cpu->pstate.min_pstate, cpu->max_perf_ratio);
2578 
2579 		/*
2580 		 * NOHZ_FULL CPUs need this as the governor callback may not
2581 		 * be invoked on them.
2582 		 */
2583 		intel_pstate_clear_update_util_hook(policy->cpu);
2584 		intel_pstate_set_pstate(cpu, pstate);
2585 	} else {
2586 		intel_pstate_set_update_util_hook(policy->cpu);
2587 	}
2588 
2589 	if (hwp_active) {
2590 		/*
2591 		 * When hwp_boost was active before and dynamically it
2592 		 * was turned off, in that case we need to clear the
2593 		 * update util hook.
2594 		 */
2595 		if (!hwp_boost)
2596 			intel_pstate_clear_update_util_hook(policy->cpu);
2597 		intel_pstate_hwp_set(policy->cpu);
2598 	}
2599 	/*
2600 	 * policy->cur is never updated with the intel_pstate driver, but it
2601 	 * is used as a stale frequency value. So, keep it within limits.
2602 	 */
2603 	policy->cur = policy->min;
2604 
2605 	mutex_unlock(&intel_pstate_limits_lock);
2606 
2607 	return 0;
2608 }
2609 
2610 static void intel_pstate_adjust_policy_max(struct cpudata *cpu,
2611 					   struct cpufreq_policy_data *policy)
2612 {
2613 	if (!hwp_active &&
2614 	    cpu->pstate.max_pstate_physical > cpu->pstate.max_pstate &&
2615 	    policy->max < policy->cpuinfo.max_freq &&
2616 	    policy->max > cpu->pstate.max_freq) {
2617 		pr_debug("policy->max > max non turbo frequency\n");
2618 		policy->max = policy->cpuinfo.max_freq;
2619 	}
2620 }
2621 
2622 static void intel_pstate_verify_cpu_policy(struct cpudata *cpu,
2623 					   struct cpufreq_policy_data *policy)
2624 {
2625 	int max_freq;
2626 
2627 	if (hwp_active) {
2628 		intel_pstate_get_hwp_cap(cpu);
2629 		max_freq = READ_ONCE(global.no_turbo) ?
2630 				cpu->pstate.max_freq : cpu->pstate.turbo_freq;
2631 	} else {
2632 		max_freq = intel_pstate_get_max_freq(cpu);
2633 	}
2634 	cpufreq_verify_within_limits(policy, policy->cpuinfo.min_freq, max_freq);
2635 
2636 	intel_pstate_adjust_policy_max(cpu, policy);
2637 }
2638 
2639 static int intel_pstate_verify_policy(struct cpufreq_policy_data *policy)
2640 {
2641 	intel_pstate_verify_cpu_policy(all_cpu_data[policy->cpu], policy);
2642 
2643 	return 0;
2644 }
2645 
2646 static int intel_cpufreq_cpu_offline(struct cpufreq_policy *policy)
2647 {
2648 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2649 
2650 	pr_debug("CPU %d going offline\n", cpu->cpu);
2651 
2652 	if (cpu->suspended)
2653 		return 0;
2654 
2655 	/*
2656 	 * If the CPU is an SMT thread and it goes offline with the performance
2657 	 * settings different from the minimum, it will prevent its sibling
2658 	 * from getting to lower performance levels, so force the minimum
2659 	 * performance on CPU offline to prevent that from happening.
2660 	 */
2661 	if (hwp_active)
2662 		intel_pstate_hwp_offline(cpu);
2663 	else
2664 		intel_pstate_set_min_pstate(cpu);
2665 
2666 	intel_pstate_exit_perf_limits(policy);
2667 
2668 	return 0;
2669 }
2670 
2671 static int intel_pstate_cpu_online(struct cpufreq_policy *policy)
2672 {
2673 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2674 
2675 	pr_debug("CPU %d going online\n", cpu->cpu);
2676 
2677 	intel_pstate_init_acpi_perf_limits(policy);
2678 
2679 	if (hwp_active) {
2680 		/*
2681 		 * Re-enable HWP and clear the "suspended" flag to let "resume"
2682 		 * know that it need not do that.
2683 		 */
2684 		intel_pstate_hwp_reenable(cpu);
2685 		cpu->suspended = false;
2686 	}
2687 
2688 	return 0;
2689 }
2690 
2691 static int intel_pstate_cpu_offline(struct cpufreq_policy *policy)
2692 {
2693 	intel_pstate_clear_update_util_hook(policy->cpu);
2694 
2695 	return intel_cpufreq_cpu_offline(policy);
2696 }
2697 
2698 static int intel_pstate_cpu_exit(struct cpufreq_policy *policy)
2699 {
2700 	pr_debug("CPU %d exiting\n", policy->cpu);
2701 
2702 	policy->fast_switch_possible = false;
2703 
2704 	return 0;
2705 }
2706 
2707 static int __intel_pstate_cpu_init(struct cpufreq_policy *policy)
2708 {
2709 	struct cpudata *cpu;
2710 	int rc;
2711 
2712 	rc = intel_pstate_init_cpu(policy->cpu);
2713 	if (rc)
2714 		return rc;
2715 
2716 	cpu = all_cpu_data[policy->cpu];
2717 
2718 	cpu->max_perf_ratio = 0xFF;
2719 	cpu->min_perf_ratio = 0;
2720 
2721 	/* cpuinfo and default policy values */
2722 	policy->cpuinfo.min_freq = cpu->pstate.min_freq;
2723 	policy->cpuinfo.max_freq = READ_ONCE(global.no_turbo) ?
2724 			cpu->pstate.max_freq : cpu->pstate.turbo_freq;
2725 
2726 	policy->min = policy->cpuinfo.min_freq;
2727 	policy->max = policy->cpuinfo.max_freq;
2728 
2729 	intel_pstate_init_acpi_perf_limits(policy);
2730 
2731 	policy->fast_switch_possible = true;
2732 
2733 	return 0;
2734 }
2735 
2736 static int intel_pstate_cpu_init(struct cpufreq_policy *policy)
2737 {
2738 	int ret = __intel_pstate_cpu_init(policy);
2739 
2740 	if (ret)
2741 		return ret;
2742 
2743 	/*
2744 	 * Set the policy to powersave to provide a valid fallback value in case
2745 	 * the default cpufreq governor is neither powersave nor performance.
2746 	 */
2747 	policy->policy = CPUFREQ_POLICY_POWERSAVE;
2748 
2749 	if (hwp_active) {
2750 		struct cpudata *cpu = all_cpu_data[policy->cpu];
2751 
2752 		cpu->epp_cached = intel_pstate_get_epp(cpu, 0);
2753 	}
2754 
2755 	return 0;
2756 }
2757 
2758 static struct cpufreq_driver intel_pstate = {
2759 	.flags		= CPUFREQ_CONST_LOOPS,
2760 	.verify		= intel_pstate_verify_policy,
2761 	.setpolicy	= intel_pstate_set_policy,
2762 	.suspend	= intel_pstate_suspend,
2763 	.resume		= intel_pstate_resume,
2764 	.init		= intel_pstate_cpu_init,
2765 	.exit		= intel_pstate_cpu_exit,
2766 	.offline	= intel_pstate_cpu_offline,
2767 	.online		= intel_pstate_cpu_online,
2768 	.update_limits	= intel_pstate_update_limits,
2769 	.name		= "intel_pstate",
2770 };
2771 
2772 static int intel_cpufreq_verify_policy(struct cpufreq_policy_data *policy)
2773 {
2774 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2775 
2776 	intel_pstate_verify_cpu_policy(cpu, policy);
2777 	intel_pstate_update_perf_limits(cpu, policy->min, policy->max);
2778 
2779 	return 0;
2780 }
2781 
2782 /* Use of trace in passive mode:
2783  *
2784  * In passive mode the trace core_busy field (also known as the
2785  * performance field, and lablelled as such on the graphs; also known as
2786  * core_avg_perf) is not needed and so is re-assigned to indicate if the
2787  * driver call was via the normal or fast switch path. Various graphs
2788  * output from the intel_pstate_tracer.py utility that include core_busy
2789  * (or performance or core_avg_perf) have a fixed y-axis from 0 to 100%,
2790  * so we use 10 to indicate the normal path through the driver, and
2791  * 90 to indicate the fast switch path through the driver.
2792  * The scaled_busy field is not used, and is set to 0.
2793  */
2794 
2795 #define	INTEL_PSTATE_TRACE_TARGET 10
2796 #define	INTEL_PSTATE_TRACE_FAST_SWITCH 90
2797 
2798 static void intel_cpufreq_trace(struct cpudata *cpu, unsigned int trace_type, int old_pstate)
2799 {
2800 	struct sample *sample;
2801 
2802 	if (!trace_pstate_sample_enabled())
2803 		return;
2804 
2805 	if (!intel_pstate_sample(cpu, ktime_get()))
2806 		return;
2807 
2808 	sample = &cpu->sample;
2809 	trace_pstate_sample(trace_type,
2810 		0,
2811 		old_pstate,
2812 		cpu->pstate.current_pstate,
2813 		sample->mperf,
2814 		sample->aperf,
2815 		sample->tsc,
2816 		get_avg_frequency(cpu),
2817 		fp_toint(cpu->iowait_boost * 100));
2818 }
2819 
2820 static void intel_cpufreq_hwp_update(struct cpudata *cpu, u32 min, u32 max,
2821 				     u32 desired, bool fast_switch)
2822 {
2823 	u64 prev = READ_ONCE(cpu->hwp_req_cached), value = prev;
2824 
2825 	value &= ~HWP_MIN_PERF(~0L);
2826 	value |= HWP_MIN_PERF(min);
2827 
2828 	value &= ~HWP_MAX_PERF(~0L);
2829 	value |= HWP_MAX_PERF(max);
2830 
2831 	value &= ~HWP_DESIRED_PERF(~0L);
2832 	value |= HWP_DESIRED_PERF(desired);
2833 
2834 	if (value == prev)
2835 		return;
2836 
2837 	WRITE_ONCE(cpu->hwp_req_cached, value);
2838 	if (fast_switch)
2839 		wrmsrl(MSR_HWP_REQUEST, value);
2840 	else
2841 		wrmsrl_on_cpu(cpu->cpu, MSR_HWP_REQUEST, value);
2842 }
2843 
2844 static void intel_cpufreq_perf_ctl_update(struct cpudata *cpu,
2845 					  u32 target_pstate, bool fast_switch)
2846 {
2847 	if (fast_switch)
2848 		wrmsrl(MSR_IA32_PERF_CTL,
2849 		       pstate_funcs.get_val(cpu, target_pstate));
2850 	else
2851 		wrmsrl_on_cpu(cpu->cpu, MSR_IA32_PERF_CTL,
2852 			      pstate_funcs.get_val(cpu, target_pstate));
2853 }
2854 
2855 static int intel_cpufreq_update_pstate(struct cpufreq_policy *policy,
2856 				       int target_pstate, bool fast_switch)
2857 {
2858 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2859 	int old_pstate = cpu->pstate.current_pstate;
2860 
2861 	target_pstate = intel_pstate_prepare_request(cpu, target_pstate);
2862 	if (hwp_active) {
2863 		int max_pstate = policy->strict_target ?
2864 					target_pstate : cpu->max_perf_ratio;
2865 
2866 		intel_cpufreq_hwp_update(cpu, target_pstate, max_pstate, 0,
2867 					 fast_switch);
2868 	} else if (target_pstate != old_pstate) {
2869 		intel_cpufreq_perf_ctl_update(cpu, target_pstate, fast_switch);
2870 	}
2871 
2872 	cpu->pstate.current_pstate = target_pstate;
2873 
2874 	intel_cpufreq_trace(cpu, fast_switch ? INTEL_PSTATE_TRACE_FAST_SWITCH :
2875 			    INTEL_PSTATE_TRACE_TARGET, old_pstate);
2876 
2877 	return target_pstate;
2878 }
2879 
2880 static int intel_cpufreq_target(struct cpufreq_policy *policy,
2881 				unsigned int target_freq,
2882 				unsigned int relation)
2883 {
2884 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2885 	struct cpufreq_freqs freqs;
2886 	int target_pstate;
2887 
2888 	freqs.old = policy->cur;
2889 	freqs.new = target_freq;
2890 
2891 	cpufreq_freq_transition_begin(policy, &freqs);
2892 
2893 	target_pstate = intel_pstate_freq_to_hwp_rel(cpu, freqs.new, relation);
2894 	target_pstate = intel_cpufreq_update_pstate(policy, target_pstate, false);
2895 
2896 	freqs.new = target_pstate * cpu->pstate.scaling;
2897 
2898 	cpufreq_freq_transition_end(policy, &freqs, false);
2899 
2900 	return 0;
2901 }
2902 
2903 static unsigned int intel_cpufreq_fast_switch(struct cpufreq_policy *policy,
2904 					      unsigned int target_freq)
2905 {
2906 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2907 	int target_pstate;
2908 
2909 	target_pstate = intel_pstate_freq_to_hwp(cpu, target_freq);
2910 
2911 	target_pstate = intel_cpufreq_update_pstate(policy, target_pstate, true);
2912 
2913 	return target_pstate * cpu->pstate.scaling;
2914 }
2915 
2916 static void intel_cpufreq_adjust_perf(unsigned int cpunum,
2917 				      unsigned long min_perf,
2918 				      unsigned long target_perf,
2919 				      unsigned long capacity)
2920 {
2921 	struct cpudata *cpu = all_cpu_data[cpunum];
2922 	u64 hwp_cap = READ_ONCE(cpu->hwp_cap_cached);
2923 	int old_pstate = cpu->pstate.current_pstate;
2924 	int cap_pstate, min_pstate, max_pstate, target_pstate;
2925 
2926 	cap_pstate = READ_ONCE(global.no_turbo) ?
2927 					HWP_GUARANTEED_PERF(hwp_cap) :
2928 					HWP_HIGHEST_PERF(hwp_cap);
2929 
2930 	/* Optimization: Avoid unnecessary divisions. */
2931 
2932 	target_pstate = cap_pstate;
2933 	if (target_perf < capacity)
2934 		target_pstate = DIV_ROUND_UP(cap_pstate * target_perf, capacity);
2935 
2936 	min_pstate = cap_pstate;
2937 	if (min_perf < capacity)
2938 		min_pstate = DIV_ROUND_UP(cap_pstate * min_perf, capacity);
2939 
2940 	if (min_pstate < cpu->pstate.min_pstate)
2941 		min_pstate = cpu->pstate.min_pstate;
2942 
2943 	if (min_pstate < cpu->min_perf_ratio)
2944 		min_pstate = cpu->min_perf_ratio;
2945 
2946 	if (min_pstate > cpu->max_perf_ratio)
2947 		min_pstate = cpu->max_perf_ratio;
2948 
2949 	max_pstate = min(cap_pstate, cpu->max_perf_ratio);
2950 	if (max_pstate < min_pstate)
2951 		max_pstate = min_pstate;
2952 
2953 	target_pstate = clamp_t(int, target_pstate, min_pstate, max_pstate);
2954 
2955 	intel_cpufreq_hwp_update(cpu, min_pstate, max_pstate, target_pstate, true);
2956 
2957 	cpu->pstate.current_pstate = target_pstate;
2958 	intel_cpufreq_trace(cpu, INTEL_PSTATE_TRACE_FAST_SWITCH, old_pstate);
2959 }
2960 
2961 static int intel_cpufreq_cpu_init(struct cpufreq_policy *policy)
2962 {
2963 	struct freq_qos_request *req;
2964 	struct cpudata *cpu;
2965 	struct device *dev;
2966 	int ret, freq;
2967 
2968 	dev = get_cpu_device(policy->cpu);
2969 	if (!dev)
2970 		return -ENODEV;
2971 
2972 	ret = __intel_pstate_cpu_init(policy);
2973 	if (ret)
2974 		return ret;
2975 
2976 	policy->cpuinfo.transition_latency = INTEL_CPUFREQ_TRANSITION_LATENCY;
2977 	/* This reflects the intel_pstate_get_cpu_pstates() setting. */
2978 	policy->cur = policy->cpuinfo.min_freq;
2979 
2980 	req = kcalloc(2, sizeof(*req), GFP_KERNEL);
2981 	if (!req) {
2982 		ret = -ENOMEM;
2983 		goto pstate_exit;
2984 	}
2985 
2986 	cpu = all_cpu_data[policy->cpu];
2987 
2988 	if (hwp_active) {
2989 		u64 value;
2990 
2991 		policy->transition_delay_us = INTEL_CPUFREQ_TRANSITION_DELAY_HWP;
2992 
2993 		intel_pstate_get_hwp_cap(cpu);
2994 
2995 		rdmsrl_on_cpu(cpu->cpu, MSR_HWP_REQUEST, &value);
2996 		WRITE_ONCE(cpu->hwp_req_cached, value);
2997 
2998 		cpu->epp_cached = intel_pstate_get_epp(cpu, value);
2999 	} else {
3000 		policy->transition_delay_us = INTEL_CPUFREQ_TRANSITION_DELAY;
3001 	}
3002 
3003 	freq = DIV_ROUND_UP(cpu->pstate.turbo_freq * global.min_perf_pct, 100);
3004 
3005 	ret = freq_qos_add_request(&policy->constraints, req, FREQ_QOS_MIN,
3006 				   freq);
3007 	if (ret < 0) {
3008 		dev_err(dev, "Failed to add min-freq constraint (%d)\n", ret);
3009 		goto free_req;
3010 	}
3011 
3012 	freq = DIV_ROUND_UP(cpu->pstate.turbo_freq * global.max_perf_pct, 100);
3013 
3014 	ret = freq_qos_add_request(&policy->constraints, req + 1, FREQ_QOS_MAX,
3015 				   freq);
3016 	if (ret < 0) {
3017 		dev_err(dev, "Failed to add max-freq constraint (%d)\n", ret);
3018 		goto remove_min_req;
3019 	}
3020 
3021 	policy->driver_data = req;
3022 
3023 	return 0;
3024 
3025 remove_min_req:
3026 	freq_qos_remove_request(req);
3027 free_req:
3028 	kfree(req);
3029 pstate_exit:
3030 	intel_pstate_exit_perf_limits(policy);
3031 
3032 	return ret;
3033 }
3034 
3035 static int intel_cpufreq_cpu_exit(struct cpufreq_policy *policy)
3036 {
3037 	struct freq_qos_request *req;
3038 
3039 	req = policy->driver_data;
3040 
3041 	freq_qos_remove_request(req + 1);
3042 	freq_qos_remove_request(req);
3043 	kfree(req);
3044 
3045 	return intel_pstate_cpu_exit(policy);
3046 }
3047 
3048 static int intel_cpufreq_suspend(struct cpufreq_policy *policy)
3049 {
3050 	intel_pstate_suspend(policy);
3051 
3052 	if (hwp_active) {
3053 		struct cpudata *cpu = all_cpu_data[policy->cpu];
3054 		u64 value = READ_ONCE(cpu->hwp_req_cached);
3055 
3056 		/*
3057 		 * Clear the desired perf field in MSR_HWP_REQUEST in case
3058 		 * intel_cpufreq_adjust_perf() is in use and the last value
3059 		 * written by it may not be suitable.
3060 		 */
3061 		value &= ~HWP_DESIRED_PERF(~0L);
3062 		wrmsrl_on_cpu(cpu->cpu, MSR_HWP_REQUEST, value);
3063 		WRITE_ONCE(cpu->hwp_req_cached, value);
3064 	}
3065 
3066 	return 0;
3067 }
3068 
3069 static struct cpufreq_driver intel_cpufreq = {
3070 	.flags		= CPUFREQ_CONST_LOOPS,
3071 	.verify		= intel_cpufreq_verify_policy,
3072 	.target		= intel_cpufreq_target,
3073 	.fast_switch	= intel_cpufreq_fast_switch,
3074 	.init		= intel_cpufreq_cpu_init,
3075 	.exit		= intel_cpufreq_cpu_exit,
3076 	.offline	= intel_cpufreq_cpu_offline,
3077 	.online		= intel_pstate_cpu_online,
3078 	.suspend	= intel_cpufreq_suspend,
3079 	.resume		= intel_pstate_resume,
3080 	.update_limits	= intel_pstate_update_limits,
3081 	.name		= "intel_cpufreq",
3082 };
3083 
3084 static struct cpufreq_driver *default_driver;
3085 
3086 static void intel_pstate_driver_cleanup(void)
3087 {
3088 	unsigned int cpu;
3089 
3090 	cpus_read_lock();
3091 	for_each_online_cpu(cpu) {
3092 		if (all_cpu_data[cpu]) {
3093 			if (intel_pstate_driver == &intel_pstate)
3094 				intel_pstate_clear_update_util_hook(cpu);
3095 
3096 			kfree(all_cpu_data[cpu]);
3097 			WRITE_ONCE(all_cpu_data[cpu], NULL);
3098 		}
3099 	}
3100 	cpus_read_unlock();
3101 
3102 	intel_pstate_driver = NULL;
3103 }
3104 
3105 static int intel_pstate_register_driver(struct cpufreq_driver *driver)
3106 {
3107 	int ret;
3108 
3109 	if (driver == &intel_pstate)
3110 		intel_pstate_sysfs_expose_hwp_dynamic_boost();
3111 
3112 	memset(&global, 0, sizeof(global));
3113 	global.max_perf_pct = 100;
3114 	global.turbo_disabled = turbo_is_disabled();
3115 	global.no_turbo = global.turbo_disabled;
3116 
3117 	arch_set_max_freq_ratio(global.turbo_disabled);
3118 
3119 	intel_pstate_driver = driver;
3120 	ret = cpufreq_register_driver(intel_pstate_driver);
3121 	if (ret) {
3122 		intel_pstate_driver_cleanup();
3123 		return ret;
3124 	}
3125 
3126 	global.min_perf_pct = min_perf_pct_min();
3127 
3128 	return 0;
3129 }
3130 
3131 static ssize_t intel_pstate_show_status(char *buf)
3132 {
3133 	if (!intel_pstate_driver)
3134 		return sprintf(buf, "off\n");
3135 
3136 	return sprintf(buf, "%s\n", intel_pstate_driver == &intel_pstate ?
3137 					"active" : "passive");
3138 }
3139 
3140 static int intel_pstate_update_status(const char *buf, size_t size)
3141 {
3142 	if (size == 3 && !strncmp(buf, "off", size)) {
3143 		if (!intel_pstate_driver)
3144 			return -EINVAL;
3145 
3146 		if (hwp_active)
3147 			return -EBUSY;
3148 
3149 		cpufreq_unregister_driver(intel_pstate_driver);
3150 		intel_pstate_driver_cleanup();
3151 		return 0;
3152 	}
3153 
3154 	if (size == 6 && !strncmp(buf, "active", size)) {
3155 		if (intel_pstate_driver) {
3156 			if (intel_pstate_driver == &intel_pstate)
3157 				return 0;
3158 
3159 			cpufreq_unregister_driver(intel_pstate_driver);
3160 		}
3161 
3162 		return intel_pstate_register_driver(&intel_pstate);
3163 	}
3164 
3165 	if (size == 7 && !strncmp(buf, "passive", size)) {
3166 		if (intel_pstate_driver) {
3167 			if (intel_pstate_driver == &intel_cpufreq)
3168 				return 0;
3169 
3170 			cpufreq_unregister_driver(intel_pstate_driver);
3171 			intel_pstate_sysfs_hide_hwp_dynamic_boost();
3172 		}
3173 
3174 		return intel_pstate_register_driver(&intel_cpufreq);
3175 	}
3176 
3177 	return -EINVAL;
3178 }
3179 
3180 static int no_load __initdata;
3181 static int no_hwp __initdata;
3182 static int hwp_only __initdata;
3183 static unsigned int force_load __initdata;
3184 
3185 static int __init intel_pstate_msrs_not_valid(void)
3186 {
3187 	if (!pstate_funcs.get_max(0) ||
3188 	    !pstate_funcs.get_min(0) ||
3189 	    !pstate_funcs.get_turbo(0))
3190 		return -ENODEV;
3191 
3192 	return 0;
3193 }
3194 
3195 static void __init copy_cpu_funcs(struct pstate_funcs *funcs)
3196 {
3197 	pstate_funcs.get_max   = funcs->get_max;
3198 	pstate_funcs.get_max_physical = funcs->get_max_physical;
3199 	pstate_funcs.get_min   = funcs->get_min;
3200 	pstate_funcs.get_turbo = funcs->get_turbo;
3201 	pstate_funcs.get_scaling = funcs->get_scaling;
3202 	pstate_funcs.get_val   = funcs->get_val;
3203 	pstate_funcs.get_vid   = funcs->get_vid;
3204 	pstate_funcs.get_aperf_mperf_shift = funcs->get_aperf_mperf_shift;
3205 }
3206 
3207 #ifdef CONFIG_ACPI
3208 
3209 static bool __init intel_pstate_no_acpi_pss(void)
3210 {
3211 	int i;
3212 
3213 	for_each_possible_cpu(i) {
3214 		acpi_status status;
3215 		union acpi_object *pss;
3216 		struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
3217 		struct acpi_processor *pr = per_cpu(processors, i);
3218 
3219 		if (!pr)
3220 			continue;
3221 
3222 		status = acpi_evaluate_object(pr->handle, "_PSS", NULL, &buffer);
3223 		if (ACPI_FAILURE(status))
3224 			continue;
3225 
3226 		pss = buffer.pointer;
3227 		if (pss && pss->type == ACPI_TYPE_PACKAGE) {
3228 			kfree(pss);
3229 			return false;
3230 		}
3231 
3232 		kfree(pss);
3233 	}
3234 
3235 	pr_debug("ACPI _PSS not found\n");
3236 	return true;
3237 }
3238 
3239 static bool __init intel_pstate_no_acpi_pcch(void)
3240 {
3241 	acpi_status status;
3242 	acpi_handle handle;
3243 
3244 	status = acpi_get_handle(NULL, "\\_SB", &handle);
3245 	if (ACPI_FAILURE(status))
3246 		goto not_found;
3247 
3248 	if (acpi_has_method(handle, "PCCH"))
3249 		return false;
3250 
3251 not_found:
3252 	pr_debug("ACPI PCCH not found\n");
3253 	return true;
3254 }
3255 
3256 static bool __init intel_pstate_has_acpi_ppc(void)
3257 {
3258 	int i;
3259 
3260 	for_each_possible_cpu(i) {
3261 		struct acpi_processor *pr = per_cpu(processors, i);
3262 
3263 		if (!pr)
3264 			continue;
3265 		if (acpi_has_method(pr->handle, "_PPC"))
3266 			return true;
3267 	}
3268 	pr_debug("ACPI _PPC not found\n");
3269 	return false;
3270 }
3271 
3272 enum {
3273 	PSS,
3274 	PPC,
3275 };
3276 
3277 /* Hardware vendor-specific info that has its own power management modes */
3278 static struct acpi_platform_list plat_info[] __initdata = {
3279 	{"HP    ", "ProLiant", 0, ACPI_SIG_FADT, all_versions, NULL, PSS},
3280 	{"ORACLE", "X4-2    ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3281 	{"ORACLE", "X4-2L   ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3282 	{"ORACLE", "X4-2B   ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3283 	{"ORACLE", "X3-2    ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3284 	{"ORACLE", "X3-2L   ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3285 	{"ORACLE", "X3-2B   ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3286 	{"ORACLE", "X4470M2 ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3287 	{"ORACLE", "X4270M3 ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3288 	{"ORACLE", "X4270M2 ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3289 	{"ORACLE", "X4170M2 ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3290 	{"ORACLE", "X4170 M3", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3291 	{"ORACLE", "X4275 M3", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3292 	{"ORACLE", "X6-2    ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3293 	{"ORACLE", "Sudbury ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3294 	{ } /* End */
3295 };
3296 
3297 #define BITMASK_OOB	(BIT(8) | BIT(18))
3298 
3299 static bool __init intel_pstate_platform_pwr_mgmt_exists(void)
3300 {
3301 	const struct x86_cpu_id *id;
3302 	u64 misc_pwr;
3303 	int idx;
3304 
3305 	id = x86_match_cpu(intel_pstate_cpu_oob_ids);
3306 	if (id) {
3307 		rdmsrl(MSR_MISC_PWR_MGMT, misc_pwr);
3308 		if (misc_pwr & BITMASK_OOB) {
3309 			pr_debug("Bit 8 or 18 in the MISC_PWR_MGMT MSR set\n");
3310 			pr_debug("P states are controlled in Out of Band mode by the firmware/hardware\n");
3311 			return true;
3312 		}
3313 	}
3314 
3315 	idx = acpi_match_platform_list(plat_info);
3316 	if (idx < 0)
3317 		return false;
3318 
3319 	switch (plat_info[idx].data) {
3320 	case PSS:
3321 		if (!intel_pstate_no_acpi_pss())
3322 			return false;
3323 
3324 		return intel_pstate_no_acpi_pcch();
3325 	case PPC:
3326 		return intel_pstate_has_acpi_ppc() && !force_load;
3327 	}
3328 
3329 	return false;
3330 }
3331 
3332 static void intel_pstate_request_control_from_smm(void)
3333 {
3334 	/*
3335 	 * It may be unsafe to request P-states control from SMM if _PPC support
3336 	 * has not been enabled.
3337 	 */
3338 	if (acpi_ppc)
3339 		acpi_processor_pstate_control();
3340 }
3341 #else /* CONFIG_ACPI not enabled */
3342 static inline bool intel_pstate_platform_pwr_mgmt_exists(void) { return false; }
3343 static inline bool intel_pstate_has_acpi_ppc(void) { return false; }
3344 static inline void intel_pstate_request_control_from_smm(void) {}
3345 #endif /* CONFIG_ACPI */
3346 
3347 #define INTEL_PSTATE_HWP_BROADWELL	0x01
3348 
3349 #define X86_MATCH_HWP(model, hwp_mode)					\
3350 	X86_MATCH_VENDOR_FAM_MODEL_FEATURE(INTEL, 6, INTEL_FAM6_##model, \
3351 					   X86_FEATURE_HWP, hwp_mode)
3352 
3353 static const struct x86_cpu_id hwp_support_ids[] __initconst = {
3354 	X86_MATCH_HWP(BROADWELL_X,	INTEL_PSTATE_HWP_BROADWELL),
3355 	X86_MATCH_HWP(BROADWELL_D,	INTEL_PSTATE_HWP_BROADWELL),
3356 	X86_MATCH_HWP(ANY,		0),
3357 	{}
3358 };
3359 
3360 static bool intel_pstate_hwp_is_enabled(void)
3361 {
3362 	u64 value;
3363 
3364 	rdmsrl(MSR_PM_ENABLE, value);
3365 	return !!(value & 0x1);
3366 }
3367 
3368 #define POWERSAVE_MASK			GENMASK(7, 0)
3369 #define BALANCE_POWER_MASK		GENMASK(15, 8)
3370 #define BALANCE_PERFORMANCE_MASK	GENMASK(23, 16)
3371 #define PERFORMANCE_MASK		GENMASK(31, 24)
3372 
3373 #define HWP_SET_EPP_VALUES(powersave, balance_power, balance_perf, performance) \
3374 	(FIELD_PREP_CONST(POWERSAVE_MASK, powersave) |\
3375 	 FIELD_PREP_CONST(BALANCE_POWER_MASK, balance_power) |\
3376 	 FIELD_PREP_CONST(BALANCE_PERFORMANCE_MASK, balance_perf) |\
3377 	 FIELD_PREP_CONST(PERFORMANCE_MASK, performance))
3378 
3379 #define HWP_SET_DEF_BALANCE_PERF_EPP(balance_perf) \
3380 	(HWP_SET_EPP_VALUES(HWP_EPP_POWERSAVE, HWP_EPP_BALANCE_POWERSAVE,\
3381 	 balance_perf, HWP_EPP_PERFORMANCE))
3382 
3383 static const struct x86_cpu_id intel_epp_default[] = {
3384 	/*
3385 	 * Set EPP value as 102, this is the max suggested EPP
3386 	 * which can result in one core turbo frequency for
3387 	 * AlderLake Mobile CPUs.
3388 	 */
3389 	X86_MATCH_INTEL_FAM6_MODEL(ALDERLAKE_L, HWP_SET_DEF_BALANCE_PERF_EPP(102)),
3390 	X86_MATCH_INTEL_FAM6_MODEL(SAPPHIRERAPIDS_X, HWP_SET_DEF_BALANCE_PERF_EPP(32)),
3391 	X86_MATCH_INTEL_FAM6_MODEL(METEORLAKE_L, HWP_SET_EPP_VALUES(HWP_EPP_POWERSAVE,
3392 							HWP_EPP_BALANCE_POWERSAVE, 115, 16)),
3393 	{}
3394 };
3395 
3396 static const struct x86_cpu_id intel_hybrid_scaling_factor[] = {
3397 	X86_MATCH_INTEL_FAM6_MODEL(METEORLAKE_L, HYBRID_SCALING_FACTOR_MTL),
3398 	{}
3399 };
3400 
3401 static int __init intel_pstate_init(void)
3402 {
3403 	static struct cpudata **_all_cpu_data;
3404 	const struct x86_cpu_id *id;
3405 	int rc;
3406 
3407 	if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL)
3408 		return -ENODEV;
3409 
3410 	id = x86_match_cpu(hwp_support_ids);
3411 	if (id) {
3412 		hwp_forced = intel_pstate_hwp_is_enabled();
3413 
3414 		if (hwp_forced)
3415 			pr_info("HWP enabled by BIOS\n");
3416 		else if (no_load)
3417 			return -ENODEV;
3418 
3419 		copy_cpu_funcs(&core_funcs);
3420 		/*
3421 		 * Avoid enabling HWP for processors without EPP support,
3422 		 * because that means incomplete HWP implementation which is a
3423 		 * corner case and supporting it is generally problematic.
3424 		 *
3425 		 * If HWP is enabled already, though, there is no choice but to
3426 		 * deal with it.
3427 		 */
3428 		if ((!no_hwp && boot_cpu_has(X86_FEATURE_HWP_EPP)) || hwp_forced) {
3429 			hwp_active = true;
3430 			hwp_mode_bdw = id->driver_data;
3431 			intel_pstate.attr = hwp_cpufreq_attrs;
3432 			intel_cpufreq.attr = hwp_cpufreq_attrs;
3433 			intel_cpufreq.flags |= CPUFREQ_NEED_UPDATE_LIMITS;
3434 			intel_cpufreq.adjust_perf = intel_cpufreq_adjust_perf;
3435 			if (!default_driver)
3436 				default_driver = &intel_pstate;
3437 
3438 			pstate_funcs.get_cpu_scaling = hwp_get_cpu_scaling;
3439 
3440 			goto hwp_cpu_matched;
3441 		}
3442 		pr_info("HWP not enabled\n");
3443 	} else {
3444 		if (no_load)
3445 			return -ENODEV;
3446 
3447 		id = x86_match_cpu(intel_pstate_cpu_ids);
3448 		if (!id) {
3449 			pr_info("CPU model not supported\n");
3450 			return -ENODEV;
3451 		}
3452 
3453 		copy_cpu_funcs((struct pstate_funcs *)id->driver_data);
3454 	}
3455 
3456 	if (intel_pstate_msrs_not_valid()) {
3457 		pr_info("Invalid MSRs\n");
3458 		return -ENODEV;
3459 	}
3460 	/* Without HWP start in the passive mode. */
3461 	if (!default_driver)
3462 		default_driver = &intel_cpufreq;
3463 
3464 hwp_cpu_matched:
3465 	/*
3466 	 * The Intel pstate driver will be ignored if the platform
3467 	 * firmware has its own power management modes.
3468 	 */
3469 	if (intel_pstate_platform_pwr_mgmt_exists()) {
3470 		pr_info("P-states controlled by the platform\n");
3471 		return -ENODEV;
3472 	}
3473 
3474 	if (!hwp_active && hwp_only)
3475 		return -ENOTSUPP;
3476 
3477 	pr_info("Intel P-state driver initializing\n");
3478 
3479 	_all_cpu_data = vzalloc(array_size(sizeof(void *), num_possible_cpus()));
3480 	if (!_all_cpu_data)
3481 		return -ENOMEM;
3482 
3483 	WRITE_ONCE(all_cpu_data, _all_cpu_data);
3484 
3485 	intel_pstate_request_control_from_smm();
3486 
3487 	intel_pstate_sysfs_expose_params();
3488 
3489 	if (hwp_active) {
3490 		const struct x86_cpu_id *id = x86_match_cpu(intel_epp_default);
3491 		const struct x86_cpu_id *hybrid_id = x86_match_cpu(intel_hybrid_scaling_factor);
3492 
3493 		if (id) {
3494 			epp_values[EPP_INDEX_POWERSAVE] =
3495 					FIELD_GET(POWERSAVE_MASK, id->driver_data);
3496 			epp_values[EPP_INDEX_BALANCE_POWERSAVE] =
3497 					FIELD_GET(BALANCE_POWER_MASK, id->driver_data);
3498 			epp_values[EPP_INDEX_BALANCE_PERFORMANCE] =
3499 					FIELD_GET(BALANCE_PERFORMANCE_MASK, id->driver_data);
3500 			epp_values[EPP_INDEX_PERFORMANCE] =
3501 					FIELD_GET(PERFORMANCE_MASK, id->driver_data);
3502 			pr_debug("Updated EPPs powersave:%x balanced power:%x balanced perf:%x performance:%x\n",
3503 				 epp_values[EPP_INDEX_POWERSAVE],
3504 				 epp_values[EPP_INDEX_BALANCE_POWERSAVE],
3505 				 epp_values[EPP_INDEX_BALANCE_PERFORMANCE],
3506 				 epp_values[EPP_INDEX_PERFORMANCE]);
3507 		}
3508 
3509 		if (hybrid_id) {
3510 			hybrid_scaling_factor = hybrid_id->driver_data;
3511 			pr_debug("hybrid scaling factor: %d\n", hybrid_scaling_factor);
3512 		}
3513 
3514 	}
3515 
3516 	mutex_lock(&intel_pstate_driver_lock);
3517 	rc = intel_pstate_register_driver(default_driver);
3518 	mutex_unlock(&intel_pstate_driver_lock);
3519 	if (rc) {
3520 		intel_pstate_sysfs_remove();
3521 		return rc;
3522 	}
3523 
3524 	if (hwp_active) {
3525 		const struct x86_cpu_id *id;
3526 
3527 		id = x86_match_cpu(intel_pstate_cpu_ee_disable_ids);
3528 		if (id) {
3529 			set_power_ctl_ee_state(false);
3530 			pr_info("Disabling energy efficiency optimization\n");
3531 		}
3532 
3533 		pr_info("HWP enabled\n");
3534 	} else if (boot_cpu_has(X86_FEATURE_HYBRID_CPU)) {
3535 		pr_warn("Problematic setup: Hybrid processor with disabled HWP\n");
3536 	}
3537 
3538 	return 0;
3539 }
3540 device_initcall(intel_pstate_init);
3541 
3542 static int __init intel_pstate_setup(char *str)
3543 {
3544 	if (!str)
3545 		return -EINVAL;
3546 
3547 	if (!strcmp(str, "disable"))
3548 		no_load = 1;
3549 	else if (!strcmp(str, "active"))
3550 		default_driver = &intel_pstate;
3551 	else if (!strcmp(str, "passive"))
3552 		default_driver = &intel_cpufreq;
3553 
3554 	if (!strcmp(str, "no_hwp"))
3555 		no_hwp = 1;
3556 
3557 	if (!strcmp(str, "force"))
3558 		force_load = 1;
3559 	if (!strcmp(str, "hwp_only"))
3560 		hwp_only = 1;
3561 	if (!strcmp(str, "per_cpu_perf_limits"))
3562 		per_cpu_limits = true;
3563 
3564 #ifdef CONFIG_ACPI
3565 	if (!strcmp(str, "support_acpi_ppc"))
3566 		acpi_ppc = true;
3567 #endif
3568 
3569 	return 0;
3570 }
3571 early_param("intel_pstate", intel_pstate_setup);
3572 
3573 MODULE_AUTHOR("Dirk Brandewie <dirk.j.brandewie@intel.com>");
3574 MODULE_DESCRIPTION("'intel_pstate' - P state driver Intel Core processors");
3575