xref: /linux/drivers/cpufreq/intel_pstate.c (revision 001821b0e79716c4e17c71d8e053a23599a7a508)
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 	intel_pstate_get_hwp_cap(cpudata);
1157 
1158 	policy->cpuinfo.max_freq = READ_ONCE(global.no_turbo) ?
1159 			cpudata->pstate.max_freq : cpudata->pstate.turbo_freq;
1160 
1161 	refresh_frequency_limits(policy);
1162 }
1163 
1164 static void intel_pstate_update_limits(unsigned int cpu)
1165 {
1166 	struct cpufreq_policy *policy = cpufreq_cpu_acquire(cpu);
1167 
1168 	if (!policy)
1169 		return;
1170 
1171 	__intel_pstate_update_max_freq(all_cpu_data[cpu], policy);
1172 
1173 	cpufreq_cpu_release(policy);
1174 }
1175 
1176 static void intel_pstate_update_limits_for_all(void)
1177 {
1178 	int cpu;
1179 
1180 	for_each_possible_cpu(cpu)
1181 		intel_pstate_update_limits(cpu);
1182 }
1183 
1184 /************************** sysfs begin ************************/
1185 #define show_one(file_name, object)					\
1186 	static ssize_t show_##file_name					\
1187 	(struct kobject *kobj, struct kobj_attribute *attr, char *buf)	\
1188 	{								\
1189 		return sprintf(buf, "%u\n", global.object);		\
1190 	}
1191 
1192 static ssize_t intel_pstate_show_status(char *buf);
1193 static int intel_pstate_update_status(const char *buf, size_t size);
1194 
1195 static ssize_t show_status(struct kobject *kobj,
1196 			   struct kobj_attribute *attr, char *buf)
1197 {
1198 	ssize_t ret;
1199 
1200 	mutex_lock(&intel_pstate_driver_lock);
1201 	ret = intel_pstate_show_status(buf);
1202 	mutex_unlock(&intel_pstate_driver_lock);
1203 
1204 	return ret;
1205 }
1206 
1207 static ssize_t store_status(struct kobject *a, struct kobj_attribute *b,
1208 			    const char *buf, size_t count)
1209 {
1210 	char *p = memchr(buf, '\n', count);
1211 	int ret;
1212 
1213 	mutex_lock(&intel_pstate_driver_lock);
1214 	ret = intel_pstate_update_status(buf, p ? p - buf : count);
1215 	mutex_unlock(&intel_pstate_driver_lock);
1216 
1217 	return ret < 0 ? ret : count;
1218 }
1219 
1220 static ssize_t show_turbo_pct(struct kobject *kobj,
1221 				struct kobj_attribute *attr, char *buf)
1222 {
1223 	struct cpudata *cpu;
1224 	int total, no_turbo, turbo_pct;
1225 	uint32_t turbo_fp;
1226 
1227 	mutex_lock(&intel_pstate_driver_lock);
1228 
1229 	if (!intel_pstate_driver) {
1230 		mutex_unlock(&intel_pstate_driver_lock);
1231 		return -EAGAIN;
1232 	}
1233 
1234 	cpu = all_cpu_data[0];
1235 
1236 	total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
1237 	no_turbo = cpu->pstate.max_pstate - cpu->pstate.min_pstate + 1;
1238 	turbo_fp = div_fp(no_turbo, total);
1239 	turbo_pct = 100 - fp_toint(mul_fp(turbo_fp, int_tofp(100)));
1240 
1241 	mutex_unlock(&intel_pstate_driver_lock);
1242 
1243 	return sprintf(buf, "%u\n", turbo_pct);
1244 }
1245 
1246 static ssize_t show_num_pstates(struct kobject *kobj,
1247 				struct kobj_attribute *attr, char *buf)
1248 {
1249 	struct cpudata *cpu;
1250 	int total;
1251 
1252 	mutex_lock(&intel_pstate_driver_lock);
1253 
1254 	if (!intel_pstate_driver) {
1255 		mutex_unlock(&intel_pstate_driver_lock);
1256 		return -EAGAIN;
1257 	}
1258 
1259 	cpu = all_cpu_data[0];
1260 	total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
1261 
1262 	mutex_unlock(&intel_pstate_driver_lock);
1263 
1264 	return sprintf(buf, "%u\n", total);
1265 }
1266 
1267 static ssize_t show_no_turbo(struct kobject *kobj,
1268 			     struct kobj_attribute *attr, char *buf)
1269 {
1270 	ssize_t ret;
1271 
1272 	mutex_lock(&intel_pstate_driver_lock);
1273 
1274 	if (!intel_pstate_driver) {
1275 		mutex_unlock(&intel_pstate_driver_lock);
1276 		return -EAGAIN;
1277 	}
1278 
1279 	ret = sprintf(buf, "%u\n", global.no_turbo);
1280 
1281 	mutex_unlock(&intel_pstate_driver_lock);
1282 
1283 	return ret;
1284 }
1285 
1286 static ssize_t store_no_turbo(struct kobject *a, struct kobj_attribute *b,
1287 			      const char *buf, size_t count)
1288 {
1289 	unsigned int input;
1290 	bool no_turbo;
1291 
1292 	if (sscanf(buf, "%u", &input) != 1)
1293 		return -EINVAL;
1294 
1295 	mutex_lock(&intel_pstate_driver_lock);
1296 
1297 	if (!intel_pstate_driver) {
1298 		count = -EAGAIN;
1299 		goto unlock_driver;
1300 	}
1301 
1302 	no_turbo = !!clamp_t(int, input, 0, 1);
1303 
1304 	if (no_turbo == global.no_turbo)
1305 		goto unlock_driver;
1306 
1307 	if (global.turbo_disabled) {
1308 		pr_notice_once("Turbo disabled by BIOS or unavailable on processor\n");
1309 		count = -EPERM;
1310 		goto unlock_driver;
1311 	}
1312 
1313 	WRITE_ONCE(global.no_turbo, no_turbo);
1314 
1315 	mutex_lock(&intel_pstate_limits_lock);
1316 
1317 	if (no_turbo) {
1318 		struct cpudata *cpu = all_cpu_data[0];
1319 		int pct = cpu->pstate.max_pstate * 100 / cpu->pstate.turbo_pstate;
1320 
1321 		/* Squash the global minimum into the permitted range. */
1322 		if (global.min_perf_pct > pct)
1323 			global.min_perf_pct = pct;
1324 	}
1325 
1326 	mutex_unlock(&intel_pstate_limits_lock);
1327 
1328 	intel_pstate_update_limits_for_all();
1329 	arch_set_max_freq_ratio(no_turbo);
1330 
1331 unlock_driver:
1332 	mutex_unlock(&intel_pstate_driver_lock);
1333 
1334 	return count;
1335 }
1336 
1337 static void update_qos_request(enum freq_qos_req_type type)
1338 {
1339 	struct freq_qos_request *req;
1340 	struct cpufreq_policy *policy;
1341 	int i;
1342 
1343 	for_each_possible_cpu(i) {
1344 		struct cpudata *cpu = all_cpu_data[i];
1345 		unsigned int freq, perf_pct;
1346 
1347 		policy = cpufreq_cpu_get(i);
1348 		if (!policy)
1349 			continue;
1350 
1351 		req = policy->driver_data;
1352 		cpufreq_cpu_put(policy);
1353 
1354 		if (!req)
1355 			continue;
1356 
1357 		if (hwp_active)
1358 			intel_pstate_get_hwp_cap(cpu);
1359 
1360 		if (type == FREQ_QOS_MIN) {
1361 			perf_pct = global.min_perf_pct;
1362 		} else {
1363 			req++;
1364 			perf_pct = global.max_perf_pct;
1365 		}
1366 
1367 		freq = DIV_ROUND_UP(cpu->pstate.turbo_freq * perf_pct, 100);
1368 
1369 		if (freq_qos_update_request(req, freq) < 0)
1370 			pr_warn("Failed to update freq constraint: CPU%d\n", i);
1371 	}
1372 }
1373 
1374 static ssize_t store_max_perf_pct(struct kobject *a, struct kobj_attribute *b,
1375 				  const char *buf, size_t count)
1376 {
1377 	unsigned int input;
1378 	int ret;
1379 
1380 	ret = sscanf(buf, "%u", &input);
1381 	if (ret != 1)
1382 		return -EINVAL;
1383 
1384 	mutex_lock(&intel_pstate_driver_lock);
1385 
1386 	if (!intel_pstate_driver) {
1387 		mutex_unlock(&intel_pstate_driver_lock);
1388 		return -EAGAIN;
1389 	}
1390 
1391 	mutex_lock(&intel_pstate_limits_lock);
1392 
1393 	global.max_perf_pct = clamp_t(int, input, global.min_perf_pct, 100);
1394 
1395 	mutex_unlock(&intel_pstate_limits_lock);
1396 
1397 	if (intel_pstate_driver == &intel_pstate)
1398 		intel_pstate_update_policies();
1399 	else
1400 		update_qos_request(FREQ_QOS_MAX);
1401 
1402 	mutex_unlock(&intel_pstate_driver_lock);
1403 
1404 	return count;
1405 }
1406 
1407 static ssize_t store_min_perf_pct(struct kobject *a, struct kobj_attribute *b,
1408 				  const char *buf, size_t count)
1409 {
1410 	unsigned int input;
1411 	int ret;
1412 
1413 	ret = sscanf(buf, "%u", &input);
1414 	if (ret != 1)
1415 		return -EINVAL;
1416 
1417 	mutex_lock(&intel_pstate_driver_lock);
1418 
1419 	if (!intel_pstate_driver) {
1420 		mutex_unlock(&intel_pstate_driver_lock);
1421 		return -EAGAIN;
1422 	}
1423 
1424 	mutex_lock(&intel_pstate_limits_lock);
1425 
1426 	global.min_perf_pct = clamp_t(int, input,
1427 				      min_perf_pct_min(), global.max_perf_pct);
1428 
1429 	mutex_unlock(&intel_pstate_limits_lock);
1430 
1431 	if (intel_pstate_driver == &intel_pstate)
1432 		intel_pstate_update_policies();
1433 	else
1434 		update_qos_request(FREQ_QOS_MIN);
1435 
1436 	mutex_unlock(&intel_pstate_driver_lock);
1437 
1438 	return count;
1439 }
1440 
1441 static ssize_t show_hwp_dynamic_boost(struct kobject *kobj,
1442 				struct kobj_attribute *attr, char *buf)
1443 {
1444 	return sprintf(buf, "%u\n", hwp_boost);
1445 }
1446 
1447 static ssize_t store_hwp_dynamic_boost(struct kobject *a,
1448 				       struct kobj_attribute *b,
1449 				       const char *buf, size_t count)
1450 {
1451 	unsigned int input;
1452 	int ret;
1453 
1454 	ret = kstrtouint(buf, 10, &input);
1455 	if (ret)
1456 		return ret;
1457 
1458 	mutex_lock(&intel_pstate_driver_lock);
1459 	hwp_boost = !!input;
1460 	intel_pstate_update_policies();
1461 	mutex_unlock(&intel_pstate_driver_lock);
1462 
1463 	return count;
1464 }
1465 
1466 static ssize_t show_energy_efficiency(struct kobject *kobj, struct kobj_attribute *attr,
1467 				      char *buf)
1468 {
1469 	u64 power_ctl;
1470 	int enable;
1471 
1472 	rdmsrl(MSR_IA32_POWER_CTL, power_ctl);
1473 	enable = !!(power_ctl & BIT(MSR_IA32_POWER_CTL_BIT_EE));
1474 	return sprintf(buf, "%d\n", !enable);
1475 }
1476 
1477 static ssize_t store_energy_efficiency(struct kobject *a, struct kobj_attribute *b,
1478 				       const char *buf, size_t count)
1479 {
1480 	bool input;
1481 	int ret;
1482 
1483 	ret = kstrtobool(buf, &input);
1484 	if (ret)
1485 		return ret;
1486 
1487 	set_power_ctl_ee_state(input);
1488 
1489 	return count;
1490 }
1491 
1492 show_one(max_perf_pct, max_perf_pct);
1493 show_one(min_perf_pct, min_perf_pct);
1494 
1495 define_one_global_rw(status);
1496 define_one_global_rw(no_turbo);
1497 define_one_global_rw(max_perf_pct);
1498 define_one_global_rw(min_perf_pct);
1499 define_one_global_ro(turbo_pct);
1500 define_one_global_ro(num_pstates);
1501 define_one_global_rw(hwp_dynamic_boost);
1502 define_one_global_rw(energy_efficiency);
1503 
1504 static struct attribute *intel_pstate_attributes[] = {
1505 	&status.attr,
1506 	&no_turbo.attr,
1507 	NULL
1508 };
1509 
1510 static const struct attribute_group intel_pstate_attr_group = {
1511 	.attrs = intel_pstate_attributes,
1512 };
1513 
1514 static const struct x86_cpu_id intel_pstate_cpu_ee_disable_ids[];
1515 
1516 static struct kobject *intel_pstate_kobject;
1517 
1518 static void __init intel_pstate_sysfs_expose_params(void)
1519 {
1520 	struct device *dev_root = bus_get_dev_root(&cpu_subsys);
1521 	int rc;
1522 
1523 	if (dev_root) {
1524 		intel_pstate_kobject = kobject_create_and_add("intel_pstate", &dev_root->kobj);
1525 		put_device(dev_root);
1526 	}
1527 	if (WARN_ON(!intel_pstate_kobject))
1528 		return;
1529 
1530 	rc = sysfs_create_group(intel_pstate_kobject, &intel_pstate_attr_group);
1531 	if (WARN_ON(rc))
1532 		return;
1533 
1534 	if (!boot_cpu_has(X86_FEATURE_HYBRID_CPU)) {
1535 		rc = sysfs_create_file(intel_pstate_kobject, &turbo_pct.attr);
1536 		WARN_ON(rc);
1537 
1538 		rc = sysfs_create_file(intel_pstate_kobject, &num_pstates.attr);
1539 		WARN_ON(rc);
1540 	}
1541 
1542 	/*
1543 	 * If per cpu limits are enforced there are no global limits, so
1544 	 * return without creating max/min_perf_pct attributes
1545 	 */
1546 	if (per_cpu_limits)
1547 		return;
1548 
1549 	rc = sysfs_create_file(intel_pstate_kobject, &max_perf_pct.attr);
1550 	WARN_ON(rc);
1551 
1552 	rc = sysfs_create_file(intel_pstate_kobject, &min_perf_pct.attr);
1553 	WARN_ON(rc);
1554 
1555 	if (x86_match_cpu(intel_pstate_cpu_ee_disable_ids)) {
1556 		rc = sysfs_create_file(intel_pstate_kobject, &energy_efficiency.attr);
1557 		WARN_ON(rc);
1558 	}
1559 }
1560 
1561 static void __init intel_pstate_sysfs_remove(void)
1562 {
1563 	if (!intel_pstate_kobject)
1564 		return;
1565 
1566 	sysfs_remove_group(intel_pstate_kobject, &intel_pstate_attr_group);
1567 
1568 	if (!boot_cpu_has(X86_FEATURE_HYBRID_CPU)) {
1569 		sysfs_remove_file(intel_pstate_kobject, &num_pstates.attr);
1570 		sysfs_remove_file(intel_pstate_kobject, &turbo_pct.attr);
1571 	}
1572 
1573 	if (!per_cpu_limits) {
1574 		sysfs_remove_file(intel_pstate_kobject, &max_perf_pct.attr);
1575 		sysfs_remove_file(intel_pstate_kobject, &min_perf_pct.attr);
1576 
1577 		if (x86_match_cpu(intel_pstate_cpu_ee_disable_ids))
1578 			sysfs_remove_file(intel_pstate_kobject, &energy_efficiency.attr);
1579 	}
1580 
1581 	kobject_put(intel_pstate_kobject);
1582 }
1583 
1584 static void intel_pstate_sysfs_expose_hwp_dynamic_boost(void)
1585 {
1586 	int rc;
1587 
1588 	if (!hwp_active)
1589 		return;
1590 
1591 	rc = sysfs_create_file(intel_pstate_kobject, &hwp_dynamic_boost.attr);
1592 	WARN_ON_ONCE(rc);
1593 }
1594 
1595 static void intel_pstate_sysfs_hide_hwp_dynamic_boost(void)
1596 {
1597 	if (!hwp_active)
1598 		return;
1599 
1600 	sysfs_remove_file(intel_pstate_kobject, &hwp_dynamic_boost.attr);
1601 }
1602 
1603 /************************** sysfs end ************************/
1604 
1605 static void intel_pstate_notify_work(struct work_struct *work)
1606 {
1607 	struct cpudata *cpudata =
1608 		container_of(to_delayed_work(work), struct cpudata, hwp_notify_work);
1609 	struct cpufreq_policy *policy = cpufreq_cpu_acquire(cpudata->cpu);
1610 
1611 	if (policy) {
1612 		__intel_pstate_update_max_freq(cpudata, policy);
1613 
1614 		cpufreq_cpu_release(policy);
1615 	}
1616 
1617 	wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_STATUS, 0);
1618 }
1619 
1620 static DEFINE_SPINLOCK(hwp_notify_lock);
1621 static cpumask_t hwp_intr_enable_mask;
1622 
1623 void notify_hwp_interrupt(void)
1624 {
1625 	unsigned int this_cpu = smp_processor_id();
1626 	unsigned long flags;
1627 	u64 value;
1628 
1629 	if (!hwp_active || !boot_cpu_has(X86_FEATURE_HWP_NOTIFY))
1630 		return;
1631 
1632 	rdmsrl_safe(MSR_HWP_STATUS, &value);
1633 	if (!(value & 0x01))
1634 		return;
1635 
1636 	spin_lock_irqsave(&hwp_notify_lock, flags);
1637 
1638 	if (!cpumask_test_cpu(this_cpu, &hwp_intr_enable_mask))
1639 		goto ack_intr;
1640 
1641 	schedule_delayed_work(&all_cpu_data[this_cpu]->hwp_notify_work,
1642 			      msecs_to_jiffies(10));
1643 
1644 	spin_unlock_irqrestore(&hwp_notify_lock, flags);
1645 
1646 	return;
1647 
1648 ack_intr:
1649 	wrmsrl_safe(MSR_HWP_STATUS, 0);
1650 	spin_unlock_irqrestore(&hwp_notify_lock, flags);
1651 }
1652 
1653 static void intel_pstate_disable_hwp_interrupt(struct cpudata *cpudata)
1654 {
1655 	bool cancel_work;
1656 
1657 	if (!boot_cpu_has(X86_FEATURE_HWP_NOTIFY))
1658 		return;
1659 
1660 	/* wrmsrl_on_cpu has to be outside spinlock as this can result in IPC */
1661 	wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_INTERRUPT, 0x00);
1662 
1663 	spin_lock_irq(&hwp_notify_lock);
1664 	cancel_work = cpumask_test_and_clear_cpu(cpudata->cpu, &hwp_intr_enable_mask);
1665 	spin_unlock_irq(&hwp_notify_lock);
1666 
1667 	if (cancel_work)
1668 		cancel_delayed_work_sync(&cpudata->hwp_notify_work);
1669 }
1670 
1671 static void intel_pstate_enable_hwp_interrupt(struct cpudata *cpudata)
1672 {
1673 	/* Enable HWP notification interrupt for guaranteed performance change */
1674 	if (boot_cpu_has(X86_FEATURE_HWP_NOTIFY)) {
1675 		spin_lock_irq(&hwp_notify_lock);
1676 		INIT_DELAYED_WORK(&cpudata->hwp_notify_work, intel_pstate_notify_work);
1677 		cpumask_set_cpu(cpudata->cpu, &hwp_intr_enable_mask);
1678 		spin_unlock_irq(&hwp_notify_lock);
1679 
1680 		/* wrmsrl_on_cpu has to be outside spinlock as this can result in IPC */
1681 		wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_INTERRUPT, 0x01);
1682 		wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_STATUS, 0);
1683 	}
1684 }
1685 
1686 static void intel_pstate_update_epp_defaults(struct cpudata *cpudata)
1687 {
1688 	cpudata->epp_default = intel_pstate_get_epp(cpudata, 0);
1689 
1690 	/*
1691 	 * If the EPP is set by firmware, which means that firmware enabled HWP
1692 	 * - Is equal or less than 0x80 (default balance_perf EPP)
1693 	 * - But less performance oriented than performance EPP
1694 	 *   then use this as new balance_perf EPP.
1695 	 */
1696 	if (hwp_forced && cpudata->epp_default <= HWP_EPP_BALANCE_PERFORMANCE &&
1697 	    cpudata->epp_default > HWP_EPP_PERFORMANCE) {
1698 		epp_values[EPP_INDEX_BALANCE_PERFORMANCE] = cpudata->epp_default;
1699 		return;
1700 	}
1701 
1702 	/*
1703 	 * If this CPU gen doesn't call for change in balance_perf
1704 	 * EPP return.
1705 	 */
1706 	if (epp_values[EPP_INDEX_BALANCE_PERFORMANCE] == HWP_EPP_BALANCE_PERFORMANCE)
1707 		return;
1708 
1709 	/*
1710 	 * Use hard coded value per gen to update the balance_perf
1711 	 * and default EPP.
1712 	 */
1713 	cpudata->epp_default = epp_values[EPP_INDEX_BALANCE_PERFORMANCE];
1714 	intel_pstate_set_epp(cpudata, cpudata->epp_default);
1715 }
1716 
1717 static void intel_pstate_hwp_enable(struct cpudata *cpudata)
1718 {
1719 	/* First disable HWP notification interrupt till we activate again */
1720 	if (boot_cpu_has(X86_FEATURE_HWP_NOTIFY))
1721 		wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_INTERRUPT, 0x00);
1722 
1723 	wrmsrl_on_cpu(cpudata->cpu, MSR_PM_ENABLE, 0x1);
1724 
1725 	intel_pstate_enable_hwp_interrupt(cpudata);
1726 
1727 	if (cpudata->epp_default >= 0)
1728 		return;
1729 
1730 	intel_pstate_update_epp_defaults(cpudata);
1731 }
1732 
1733 static int atom_get_min_pstate(int not_used)
1734 {
1735 	u64 value;
1736 
1737 	rdmsrl(MSR_ATOM_CORE_RATIOS, value);
1738 	return (value >> 8) & 0x7F;
1739 }
1740 
1741 static int atom_get_max_pstate(int not_used)
1742 {
1743 	u64 value;
1744 
1745 	rdmsrl(MSR_ATOM_CORE_RATIOS, value);
1746 	return (value >> 16) & 0x7F;
1747 }
1748 
1749 static int atom_get_turbo_pstate(int not_used)
1750 {
1751 	u64 value;
1752 
1753 	rdmsrl(MSR_ATOM_CORE_TURBO_RATIOS, value);
1754 	return value & 0x7F;
1755 }
1756 
1757 static u64 atom_get_val(struct cpudata *cpudata, int pstate)
1758 {
1759 	u64 val;
1760 	int32_t vid_fp;
1761 	u32 vid;
1762 
1763 	val = (u64)pstate << 8;
1764 	if (READ_ONCE(global.no_turbo) && !global.turbo_disabled)
1765 		val |= (u64)1 << 32;
1766 
1767 	vid_fp = cpudata->vid.min + mul_fp(
1768 		int_tofp(pstate - cpudata->pstate.min_pstate),
1769 		cpudata->vid.ratio);
1770 
1771 	vid_fp = clamp_t(int32_t, vid_fp, cpudata->vid.min, cpudata->vid.max);
1772 	vid = ceiling_fp(vid_fp);
1773 
1774 	if (pstate > cpudata->pstate.max_pstate)
1775 		vid = cpudata->vid.turbo;
1776 
1777 	return val | vid;
1778 }
1779 
1780 static int silvermont_get_scaling(void)
1781 {
1782 	u64 value;
1783 	int i;
1784 	/* Defined in Table 35-6 from SDM (Sept 2015) */
1785 	static int silvermont_freq_table[] = {
1786 		83300, 100000, 133300, 116700, 80000};
1787 
1788 	rdmsrl(MSR_FSB_FREQ, value);
1789 	i = value & 0x7;
1790 	WARN_ON(i > 4);
1791 
1792 	return silvermont_freq_table[i];
1793 }
1794 
1795 static int airmont_get_scaling(void)
1796 {
1797 	u64 value;
1798 	int i;
1799 	/* Defined in Table 35-10 from SDM (Sept 2015) */
1800 	static int airmont_freq_table[] = {
1801 		83300, 100000, 133300, 116700, 80000,
1802 		93300, 90000, 88900, 87500};
1803 
1804 	rdmsrl(MSR_FSB_FREQ, value);
1805 	i = value & 0xF;
1806 	WARN_ON(i > 8);
1807 
1808 	return airmont_freq_table[i];
1809 }
1810 
1811 static void atom_get_vid(struct cpudata *cpudata)
1812 {
1813 	u64 value;
1814 
1815 	rdmsrl(MSR_ATOM_CORE_VIDS, value);
1816 	cpudata->vid.min = int_tofp((value >> 8) & 0x7f);
1817 	cpudata->vid.max = int_tofp((value >> 16) & 0x7f);
1818 	cpudata->vid.ratio = div_fp(
1819 		cpudata->vid.max - cpudata->vid.min,
1820 		int_tofp(cpudata->pstate.max_pstate -
1821 			cpudata->pstate.min_pstate));
1822 
1823 	rdmsrl(MSR_ATOM_CORE_TURBO_VIDS, value);
1824 	cpudata->vid.turbo = value & 0x7f;
1825 }
1826 
1827 static int core_get_min_pstate(int cpu)
1828 {
1829 	u64 value;
1830 
1831 	rdmsrl_on_cpu(cpu, MSR_PLATFORM_INFO, &value);
1832 	return (value >> 40) & 0xFF;
1833 }
1834 
1835 static int core_get_max_pstate_physical(int cpu)
1836 {
1837 	u64 value;
1838 
1839 	rdmsrl_on_cpu(cpu, MSR_PLATFORM_INFO, &value);
1840 	return (value >> 8) & 0xFF;
1841 }
1842 
1843 static int core_get_tdp_ratio(int cpu, u64 plat_info)
1844 {
1845 	/* Check how many TDP levels present */
1846 	if (plat_info & 0x600000000) {
1847 		u64 tdp_ctrl;
1848 		u64 tdp_ratio;
1849 		int tdp_msr;
1850 		int err;
1851 
1852 		/* Get the TDP level (0, 1, 2) to get ratios */
1853 		err = rdmsrl_safe_on_cpu(cpu, MSR_CONFIG_TDP_CONTROL, &tdp_ctrl);
1854 		if (err)
1855 			return err;
1856 
1857 		/* TDP MSR are continuous starting at 0x648 */
1858 		tdp_msr = MSR_CONFIG_TDP_NOMINAL + (tdp_ctrl & 0x03);
1859 		err = rdmsrl_safe_on_cpu(cpu, tdp_msr, &tdp_ratio);
1860 		if (err)
1861 			return err;
1862 
1863 		/* For level 1 and 2, bits[23:16] contain the ratio */
1864 		if (tdp_ctrl & 0x03)
1865 			tdp_ratio >>= 16;
1866 
1867 		tdp_ratio &= 0xff; /* ratios are only 8 bits long */
1868 		pr_debug("tdp_ratio %x\n", (int)tdp_ratio);
1869 
1870 		return (int)tdp_ratio;
1871 	}
1872 
1873 	return -ENXIO;
1874 }
1875 
1876 static int core_get_max_pstate(int cpu)
1877 {
1878 	u64 tar;
1879 	u64 plat_info;
1880 	int max_pstate;
1881 	int tdp_ratio;
1882 	int err;
1883 
1884 	rdmsrl_on_cpu(cpu, MSR_PLATFORM_INFO, &plat_info);
1885 	max_pstate = (plat_info >> 8) & 0xFF;
1886 
1887 	tdp_ratio = core_get_tdp_ratio(cpu, plat_info);
1888 	if (tdp_ratio <= 0)
1889 		return max_pstate;
1890 
1891 	if (hwp_active) {
1892 		/* Turbo activation ratio is not used on HWP platforms */
1893 		return tdp_ratio;
1894 	}
1895 
1896 	err = rdmsrl_safe_on_cpu(cpu, MSR_TURBO_ACTIVATION_RATIO, &tar);
1897 	if (!err) {
1898 		int tar_levels;
1899 
1900 		/* Do some sanity checking for safety */
1901 		tar_levels = tar & 0xff;
1902 		if (tdp_ratio - 1 == tar_levels) {
1903 			max_pstate = tar_levels;
1904 			pr_debug("max_pstate=TAC %x\n", max_pstate);
1905 		}
1906 	}
1907 
1908 	return max_pstate;
1909 }
1910 
1911 static int core_get_turbo_pstate(int cpu)
1912 {
1913 	u64 value;
1914 	int nont, ret;
1915 
1916 	rdmsrl_on_cpu(cpu, MSR_TURBO_RATIO_LIMIT, &value);
1917 	nont = core_get_max_pstate(cpu);
1918 	ret = (value) & 255;
1919 	if (ret <= nont)
1920 		ret = nont;
1921 	return ret;
1922 }
1923 
1924 static u64 core_get_val(struct cpudata *cpudata, int pstate)
1925 {
1926 	u64 val;
1927 
1928 	val = (u64)pstate << 8;
1929 	if (READ_ONCE(global.no_turbo) && !global.turbo_disabled)
1930 		val |= (u64)1 << 32;
1931 
1932 	return val;
1933 }
1934 
1935 static int knl_get_aperf_mperf_shift(void)
1936 {
1937 	return 10;
1938 }
1939 
1940 static int knl_get_turbo_pstate(int cpu)
1941 {
1942 	u64 value;
1943 	int nont, ret;
1944 
1945 	rdmsrl_on_cpu(cpu, MSR_TURBO_RATIO_LIMIT, &value);
1946 	nont = core_get_max_pstate(cpu);
1947 	ret = (((value) >> 8) & 0xFF);
1948 	if (ret <= nont)
1949 		ret = nont;
1950 	return ret;
1951 }
1952 
1953 static void hybrid_get_type(void *data)
1954 {
1955 	u8 *cpu_type = data;
1956 
1957 	*cpu_type = get_this_hybrid_cpu_type();
1958 }
1959 
1960 static int hwp_get_cpu_scaling(int cpu)
1961 {
1962 	u8 cpu_type = 0;
1963 
1964 	smp_call_function_single(cpu, hybrid_get_type, &cpu_type, 1);
1965 	/* P-cores have a smaller perf level-to-freqency scaling factor. */
1966 	if (cpu_type == 0x40)
1967 		return hybrid_scaling_factor;
1968 
1969 	/* Use default core scaling for E-cores */
1970 	if (cpu_type == 0x20)
1971 		return core_get_scaling();
1972 
1973 	/*
1974 	 * If reached here, this system is either non-hybrid (like Tiger
1975 	 * Lake) or hybrid-capable (like Alder Lake or Raptor Lake) with
1976 	 * no E cores (in which case CPUID for hybrid support is 0).
1977 	 *
1978 	 * The CPPC nominal_frequency field is 0 for non-hybrid systems,
1979 	 * so the default core scaling will be used for them.
1980 	 */
1981 	return intel_pstate_cppc_get_scaling(cpu);
1982 }
1983 
1984 static void intel_pstate_set_pstate(struct cpudata *cpu, int pstate)
1985 {
1986 	trace_cpu_frequency(pstate * cpu->pstate.scaling, cpu->cpu);
1987 	cpu->pstate.current_pstate = pstate;
1988 	/*
1989 	 * Generally, there is no guarantee that this code will always run on
1990 	 * the CPU being updated, so force the register update to run on the
1991 	 * right CPU.
1992 	 */
1993 	wrmsrl_on_cpu(cpu->cpu, MSR_IA32_PERF_CTL,
1994 		      pstate_funcs.get_val(cpu, pstate));
1995 }
1996 
1997 static void intel_pstate_set_min_pstate(struct cpudata *cpu)
1998 {
1999 	intel_pstate_set_pstate(cpu, cpu->pstate.min_pstate);
2000 }
2001 
2002 static void intel_pstate_get_cpu_pstates(struct cpudata *cpu)
2003 {
2004 	int perf_ctl_max_phys = pstate_funcs.get_max_physical(cpu->cpu);
2005 	int perf_ctl_scaling = pstate_funcs.get_scaling();
2006 
2007 	cpu->pstate.min_pstate = pstate_funcs.get_min(cpu->cpu);
2008 	cpu->pstate.max_pstate_physical = perf_ctl_max_phys;
2009 	cpu->pstate.perf_ctl_scaling = perf_ctl_scaling;
2010 
2011 	if (hwp_active && !hwp_mode_bdw) {
2012 		__intel_pstate_get_hwp_cap(cpu);
2013 
2014 		if (pstate_funcs.get_cpu_scaling) {
2015 			cpu->pstate.scaling = pstate_funcs.get_cpu_scaling(cpu->cpu);
2016 			if (cpu->pstate.scaling != perf_ctl_scaling)
2017 				intel_pstate_hybrid_hwp_adjust(cpu);
2018 		} else {
2019 			cpu->pstate.scaling = perf_ctl_scaling;
2020 		}
2021 	} else {
2022 		cpu->pstate.scaling = perf_ctl_scaling;
2023 		cpu->pstate.max_pstate = pstate_funcs.get_max(cpu->cpu);
2024 		cpu->pstate.turbo_pstate = pstate_funcs.get_turbo(cpu->cpu);
2025 	}
2026 
2027 	if (cpu->pstate.scaling == perf_ctl_scaling) {
2028 		cpu->pstate.min_freq = cpu->pstate.min_pstate * perf_ctl_scaling;
2029 		cpu->pstate.max_freq = cpu->pstate.max_pstate * perf_ctl_scaling;
2030 		cpu->pstate.turbo_freq = cpu->pstate.turbo_pstate * perf_ctl_scaling;
2031 	}
2032 
2033 	if (pstate_funcs.get_aperf_mperf_shift)
2034 		cpu->aperf_mperf_shift = pstate_funcs.get_aperf_mperf_shift();
2035 
2036 	if (pstate_funcs.get_vid)
2037 		pstate_funcs.get_vid(cpu);
2038 
2039 	intel_pstate_set_min_pstate(cpu);
2040 }
2041 
2042 /*
2043  * Long hold time will keep high perf limits for long time,
2044  * which negatively impacts perf/watt for some workloads,
2045  * like specpower. 3ms is based on experiements on some
2046  * workoads.
2047  */
2048 static int hwp_boost_hold_time_ns = 3 * NSEC_PER_MSEC;
2049 
2050 static inline void intel_pstate_hwp_boost_up(struct cpudata *cpu)
2051 {
2052 	u64 hwp_req = READ_ONCE(cpu->hwp_req_cached);
2053 	u64 hwp_cap = READ_ONCE(cpu->hwp_cap_cached);
2054 	u32 max_limit = (hwp_req & 0xff00) >> 8;
2055 	u32 min_limit = (hwp_req & 0xff);
2056 	u32 boost_level1;
2057 
2058 	/*
2059 	 * Cases to consider (User changes via sysfs or boot time):
2060 	 * If, P0 (Turbo max) = P1 (Guaranteed max) = min:
2061 	 *	No boost, return.
2062 	 * If, P0 (Turbo max) > P1 (Guaranteed max) = min:
2063 	 *     Should result in one level boost only for P0.
2064 	 * If, P0 (Turbo max) = P1 (Guaranteed max) > min:
2065 	 *     Should result in two level boost:
2066 	 *         (min + p1)/2 and P1.
2067 	 * If, P0 (Turbo max) > P1 (Guaranteed max) > min:
2068 	 *     Should result in three level boost:
2069 	 *        (min + p1)/2, P1 and P0.
2070 	 */
2071 
2072 	/* If max and min are equal or already at max, nothing to boost */
2073 	if (max_limit == min_limit || cpu->hwp_boost_min >= max_limit)
2074 		return;
2075 
2076 	if (!cpu->hwp_boost_min)
2077 		cpu->hwp_boost_min = min_limit;
2078 
2079 	/* level at half way mark between min and guranteed */
2080 	boost_level1 = (HWP_GUARANTEED_PERF(hwp_cap) + min_limit) >> 1;
2081 
2082 	if (cpu->hwp_boost_min < boost_level1)
2083 		cpu->hwp_boost_min = boost_level1;
2084 	else if (cpu->hwp_boost_min < HWP_GUARANTEED_PERF(hwp_cap))
2085 		cpu->hwp_boost_min = HWP_GUARANTEED_PERF(hwp_cap);
2086 	else if (cpu->hwp_boost_min == HWP_GUARANTEED_PERF(hwp_cap) &&
2087 		 max_limit != HWP_GUARANTEED_PERF(hwp_cap))
2088 		cpu->hwp_boost_min = max_limit;
2089 	else
2090 		return;
2091 
2092 	hwp_req = (hwp_req & ~GENMASK_ULL(7, 0)) | cpu->hwp_boost_min;
2093 	wrmsrl(MSR_HWP_REQUEST, hwp_req);
2094 	cpu->last_update = cpu->sample.time;
2095 }
2096 
2097 static inline void intel_pstate_hwp_boost_down(struct cpudata *cpu)
2098 {
2099 	if (cpu->hwp_boost_min) {
2100 		bool expired;
2101 
2102 		/* Check if we are idle for hold time to boost down */
2103 		expired = time_after64(cpu->sample.time, cpu->last_update +
2104 				       hwp_boost_hold_time_ns);
2105 		if (expired) {
2106 			wrmsrl(MSR_HWP_REQUEST, cpu->hwp_req_cached);
2107 			cpu->hwp_boost_min = 0;
2108 		}
2109 	}
2110 	cpu->last_update = cpu->sample.time;
2111 }
2112 
2113 static inline void intel_pstate_update_util_hwp_local(struct cpudata *cpu,
2114 						      u64 time)
2115 {
2116 	cpu->sample.time = time;
2117 
2118 	if (cpu->sched_flags & SCHED_CPUFREQ_IOWAIT) {
2119 		bool do_io = false;
2120 
2121 		cpu->sched_flags = 0;
2122 		/*
2123 		 * Set iowait_boost flag and update time. Since IO WAIT flag
2124 		 * is set all the time, we can't just conclude that there is
2125 		 * some IO bound activity is scheduled on this CPU with just
2126 		 * one occurrence. If we receive at least two in two
2127 		 * consecutive ticks, then we treat as boost candidate.
2128 		 */
2129 		if (time_before64(time, cpu->last_io_update + 2 * TICK_NSEC))
2130 			do_io = true;
2131 
2132 		cpu->last_io_update = time;
2133 
2134 		if (do_io)
2135 			intel_pstate_hwp_boost_up(cpu);
2136 
2137 	} else {
2138 		intel_pstate_hwp_boost_down(cpu);
2139 	}
2140 }
2141 
2142 static inline void intel_pstate_update_util_hwp(struct update_util_data *data,
2143 						u64 time, unsigned int flags)
2144 {
2145 	struct cpudata *cpu = container_of(data, struct cpudata, update_util);
2146 
2147 	cpu->sched_flags |= flags;
2148 
2149 	if (smp_processor_id() == cpu->cpu)
2150 		intel_pstate_update_util_hwp_local(cpu, time);
2151 }
2152 
2153 static inline void intel_pstate_calc_avg_perf(struct cpudata *cpu)
2154 {
2155 	struct sample *sample = &cpu->sample;
2156 
2157 	sample->core_avg_perf = div_ext_fp(sample->aperf, sample->mperf);
2158 }
2159 
2160 static inline bool intel_pstate_sample(struct cpudata *cpu, u64 time)
2161 {
2162 	u64 aperf, mperf;
2163 	unsigned long flags;
2164 	u64 tsc;
2165 
2166 	local_irq_save(flags);
2167 	rdmsrl(MSR_IA32_APERF, aperf);
2168 	rdmsrl(MSR_IA32_MPERF, mperf);
2169 	tsc = rdtsc();
2170 	if (cpu->prev_mperf == mperf || cpu->prev_tsc == tsc) {
2171 		local_irq_restore(flags);
2172 		return false;
2173 	}
2174 	local_irq_restore(flags);
2175 
2176 	cpu->last_sample_time = cpu->sample.time;
2177 	cpu->sample.time = time;
2178 	cpu->sample.aperf = aperf;
2179 	cpu->sample.mperf = mperf;
2180 	cpu->sample.tsc =  tsc;
2181 	cpu->sample.aperf -= cpu->prev_aperf;
2182 	cpu->sample.mperf -= cpu->prev_mperf;
2183 	cpu->sample.tsc -= cpu->prev_tsc;
2184 
2185 	cpu->prev_aperf = aperf;
2186 	cpu->prev_mperf = mperf;
2187 	cpu->prev_tsc = tsc;
2188 	/*
2189 	 * First time this function is invoked in a given cycle, all of the
2190 	 * previous sample data fields are equal to zero or stale and they must
2191 	 * be populated with meaningful numbers for things to work, so assume
2192 	 * that sample.time will always be reset before setting the utilization
2193 	 * update hook and make the caller skip the sample then.
2194 	 */
2195 	if (cpu->last_sample_time) {
2196 		intel_pstate_calc_avg_perf(cpu);
2197 		return true;
2198 	}
2199 	return false;
2200 }
2201 
2202 static inline int32_t get_avg_frequency(struct cpudata *cpu)
2203 {
2204 	return mul_ext_fp(cpu->sample.core_avg_perf, cpu_khz);
2205 }
2206 
2207 static inline int32_t get_avg_pstate(struct cpudata *cpu)
2208 {
2209 	return mul_ext_fp(cpu->pstate.max_pstate_physical,
2210 			  cpu->sample.core_avg_perf);
2211 }
2212 
2213 static inline int32_t get_target_pstate(struct cpudata *cpu)
2214 {
2215 	struct sample *sample = &cpu->sample;
2216 	int32_t busy_frac;
2217 	int target, avg_pstate;
2218 
2219 	busy_frac = div_fp(sample->mperf << cpu->aperf_mperf_shift,
2220 			   sample->tsc);
2221 
2222 	if (busy_frac < cpu->iowait_boost)
2223 		busy_frac = cpu->iowait_boost;
2224 
2225 	sample->busy_scaled = busy_frac * 100;
2226 
2227 	target = READ_ONCE(global.no_turbo) ?
2228 			cpu->pstate.max_pstate : cpu->pstate.turbo_pstate;
2229 	target += target >> 2;
2230 	target = mul_fp(target, busy_frac);
2231 	if (target < cpu->pstate.min_pstate)
2232 		target = cpu->pstate.min_pstate;
2233 
2234 	/*
2235 	 * If the average P-state during the previous cycle was higher than the
2236 	 * current target, add 50% of the difference to the target to reduce
2237 	 * possible performance oscillations and offset possible performance
2238 	 * loss related to moving the workload from one CPU to another within
2239 	 * a package/module.
2240 	 */
2241 	avg_pstate = get_avg_pstate(cpu);
2242 	if (avg_pstate > target)
2243 		target += (avg_pstate - target) >> 1;
2244 
2245 	return target;
2246 }
2247 
2248 static int intel_pstate_prepare_request(struct cpudata *cpu, int pstate)
2249 {
2250 	int min_pstate = max(cpu->pstate.min_pstate, cpu->min_perf_ratio);
2251 	int max_pstate = max(min_pstate, cpu->max_perf_ratio);
2252 
2253 	return clamp_t(int, pstate, min_pstate, max_pstate);
2254 }
2255 
2256 static void intel_pstate_update_pstate(struct cpudata *cpu, int pstate)
2257 {
2258 	if (pstate == cpu->pstate.current_pstate)
2259 		return;
2260 
2261 	cpu->pstate.current_pstate = pstate;
2262 	wrmsrl(MSR_IA32_PERF_CTL, pstate_funcs.get_val(cpu, pstate));
2263 }
2264 
2265 static void intel_pstate_adjust_pstate(struct cpudata *cpu)
2266 {
2267 	int from = cpu->pstate.current_pstate;
2268 	struct sample *sample;
2269 	int target_pstate;
2270 
2271 	target_pstate = get_target_pstate(cpu);
2272 	target_pstate = intel_pstate_prepare_request(cpu, target_pstate);
2273 	trace_cpu_frequency(target_pstate * cpu->pstate.scaling, cpu->cpu);
2274 	intel_pstate_update_pstate(cpu, target_pstate);
2275 
2276 	sample = &cpu->sample;
2277 	trace_pstate_sample(mul_ext_fp(100, sample->core_avg_perf),
2278 		fp_toint(sample->busy_scaled),
2279 		from,
2280 		cpu->pstate.current_pstate,
2281 		sample->mperf,
2282 		sample->aperf,
2283 		sample->tsc,
2284 		get_avg_frequency(cpu),
2285 		fp_toint(cpu->iowait_boost * 100));
2286 }
2287 
2288 static void intel_pstate_update_util(struct update_util_data *data, u64 time,
2289 				     unsigned int flags)
2290 {
2291 	struct cpudata *cpu = container_of(data, struct cpudata, update_util);
2292 	u64 delta_ns;
2293 
2294 	/* Don't allow remote callbacks */
2295 	if (smp_processor_id() != cpu->cpu)
2296 		return;
2297 
2298 	delta_ns = time - cpu->last_update;
2299 	if (flags & SCHED_CPUFREQ_IOWAIT) {
2300 		/* Start over if the CPU may have been idle. */
2301 		if (delta_ns > TICK_NSEC) {
2302 			cpu->iowait_boost = ONE_EIGHTH_FP;
2303 		} else if (cpu->iowait_boost >= ONE_EIGHTH_FP) {
2304 			cpu->iowait_boost <<= 1;
2305 			if (cpu->iowait_boost > int_tofp(1))
2306 				cpu->iowait_boost = int_tofp(1);
2307 		} else {
2308 			cpu->iowait_boost = ONE_EIGHTH_FP;
2309 		}
2310 	} else if (cpu->iowait_boost) {
2311 		/* Clear iowait_boost if the CPU may have been idle. */
2312 		if (delta_ns > TICK_NSEC)
2313 			cpu->iowait_boost = 0;
2314 		else
2315 			cpu->iowait_boost >>= 1;
2316 	}
2317 	cpu->last_update = time;
2318 	delta_ns = time - cpu->sample.time;
2319 	if ((s64)delta_ns < INTEL_PSTATE_SAMPLING_INTERVAL)
2320 		return;
2321 
2322 	if (intel_pstate_sample(cpu, time))
2323 		intel_pstate_adjust_pstate(cpu);
2324 }
2325 
2326 static struct pstate_funcs core_funcs = {
2327 	.get_max = core_get_max_pstate,
2328 	.get_max_physical = core_get_max_pstate_physical,
2329 	.get_min = core_get_min_pstate,
2330 	.get_turbo = core_get_turbo_pstate,
2331 	.get_scaling = core_get_scaling,
2332 	.get_val = core_get_val,
2333 };
2334 
2335 static const struct pstate_funcs silvermont_funcs = {
2336 	.get_max = atom_get_max_pstate,
2337 	.get_max_physical = atom_get_max_pstate,
2338 	.get_min = atom_get_min_pstate,
2339 	.get_turbo = atom_get_turbo_pstate,
2340 	.get_val = atom_get_val,
2341 	.get_scaling = silvermont_get_scaling,
2342 	.get_vid = atom_get_vid,
2343 };
2344 
2345 static const struct pstate_funcs airmont_funcs = {
2346 	.get_max = atom_get_max_pstate,
2347 	.get_max_physical = atom_get_max_pstate,
2348 	.get_min = atom_get_min_pstate,
2349 	.get_turbo = atom_get_turbo_pstate,
2350 	.get_val = atom_get_val,
2351 	.get_scaling = airmont_get_scaling,
2352 	.get_vid = atom_get_vid,
2353 };
2354 
2355 static const struct pstate_funcs knl_funcs = {
2356 	.get_max = core_get_max_pstate,
2357 	.get_max_physical = core_get_max_pstate_physical,
2358 	.get_min = core_get_min_pstate,
2359 	.get_turbo = knl_get_turbo_pstate,
2360 	.get_aperf_mperf_shift = knl_get_aperf_mperf_shift,
2361 	.get_scaling = core_get_scaling,
2362 	.get_val = core_get_val,
2363 };
2364 
2365 #define X86_MATCH(model, policy)					 \
2366 	X86_MATCH_VENDOR_FAM_MODEL_FEATURE(INTEL, 6, INTEL_FAM6_##model, \
2367 					   X86_FEATURE_APERFMPERF, &policy)
2368 
2369 static const struct x86_cpu_id intel_pstate_cpu_ids[] = {
2370 	X86_MATCH(SANDYBRIDGE,		core_funcs),
2371 	X86_MATCH(SANDYBRIDGE_X,	core_funcs),
2372 	X86_MATCH(ATOM_SILVERMONT,	silvermont_funcs),
2373 	X86_MATCH(IVYBRIDGE,		core_funcs),
2374 	X86_MATCH(HASWELL,		core_funcs),
2375 	X86_MATCH(BROADWELL,		core_funcs),
2376 	X86_MATCH(IVYBRIDGE_X,		core_funcs),
2377 	X86_MATCH(HASWELL_X,		core_funcs),
2378 	X86_MATCH(HASWELL_L,		core_funcs),
2379 	X86_MATCH(HASWELL_G,		core_funcs),
2380 	X86_MATCH(BROADWELL_G,		core_funcs),
2381 	X86_MATCH(ATOM_AIRMONT,		airmont_funcs),
2382 	X86_MATCH(SKYLAKE_L,		core_funcs),
2383 	X86_MATCH(BROADWELL_X,		core_funcs),
2384 	X86_MATCH(SKYLAKE,		core_funcs),
2385 	X86_MATCH(BROADWELL_D,		core_funcs),
2386 	X86_MATCH(XEON_PHI_KNL,		knl_funcs),
2387 	X86_MATCH(XEON_PHI_KNM,		knl_funcs),
2388 	X86_MATCH(ATOM_GOLDMONT,	core_funcs),
2389 	X86_MATCH(ATOM_GOLDMONT_PLUS,	core_funcs),
2390 	X86_MATCH(SKYLAKE_X,		core_funcs),
2391 	X86_MATCH(COMETLAKE,		core_funcs),
2392 	X86_MATCH(ICELAKE_X,		core_funcs),
2393 	X86_MATCH(TIGERLAKE,		core_funcs),
2394 	X86_MATCH(SAPPHIRERAPIDS_X,	core_funcs),
2395 	X86_MATCH(EMERALDRAPIDS_X,      core_funcs),
2396 	{}
2397 };
2398 MODULE_DEVICE_TABLE(x86cpu, intel_pstate_cpu_ids);
2399 
2400 #ifdef CONFIG_ACPI
2401 static const struct x86_cpu_id intel_pstate_cpu_oob_ids[] __initconst = {
2402 	X86_MATCH(BROADWELL_D,		core_funcs),
2403 	X86_MATCH(BROADWELL_X,		core_funcs),
2404 	X86_MATCH(SKYLAKE_X,		core_funcs),
2405 	X86_MATCH(ICELAKE_X,		core_funcs),
2406 	X86_MATCH(SAPPHIRERAPIDS_X,	core_funcs),
2407 	{}
2408 };
2409 #endif
2410 
2411 static const struct x86_cpu_id intel_pstate_cpu_ee_disable_ids[] = {
2412 	X86_MATCH(KABYLAKE,		core_funcs),
2413 	{}
2414 };
2415 
2416 static int intel_pstate_init_cpu(unsigned int cpunum)
2417 {
2418 	struct cpudata *cpu;
2419 
2420 	cpu = all_cpu_data[cpunum];
2421 
2422 	if (!cpu) {
2423 		cpu = kzalloc(sizeof(*cpu), GFP_KERNEL);
2424 		if (!cpu)
2425 			return -ENOMEM;
2426 
2427 		WRITE_ONCE(all_cpu_data[cpunum], cpu);
2428 
2429 		cpu->cpu = cpunum;
2430 
2431 		cpu->epp_default = -EINVAL;
2432 
2433 		if (hwp_active) {
2434 			intel_pstate_hwp_enable(cpu);
2435 
2436 			if (intel_pstate_acpi_pm_profile_server())
2437 				hwp_boost = true;
2438 		}
2439 	} else if (hwp_active) {
2440 		/*
2441 		 * Re-enable HWP in case this happens after a resume from ACPI
2442 		 * S3 if the CPU was offline during the whole system/resume
2443 		 * cycle.
2444 		 */
2445 		intel_pstate_hwp_reenable(cpu);
2446 	}
2447 
2448 	cpu->epp_powersave = -EINVAL;
2449 	cpu->epp_policy = 0;
2450 
2451 	intel_pstate_get_cpu_pstates(cpu);
2452 
2453 	pr_debug("controlling: cpu %d\n", cpunum);
2454 
2455 	return 0;
2456 }
2457 
2458 static void intel_pstate_set_update_util_hook(unsigned int cpu_num)
2459 {
2460 	struct cpudata *cpu = all_cpu_data[cpu_num];
2461 
2462 	if (hwp_active && !hwp_boost)
2463 		return;
2464 
2465 	if (cpu->update_util_set)
2466 		return;
2467 
2468 	/* Prevent intel_pstate_update_util() from using stale data. */
2469 	cpu->sample.time = 0;
2470 	cpufreq_add_update_util_hook(cpu_num, &cpu->update_util,
2471 				     (hwp_active ?
2472 				      intel_pstate_update_util_hwp :
2473 				      intel_pstate_update_util));
2474 	cpu->update_util_set = true;
2475 }
2476 
2477 static void intel_pstate_clear_update_util_hook(unsigned int cpu)
2478 {
2479 	struct cpudata *cpu_data = all_cpu_data[cpu];
2480 
2481 	if (!cpu_data->update_util_set)
2482 		return;
2483 
2484 	cpufreq_remove_update_util_hook(cpu);
2485 	cpu_data->update_util_set = false;
2486 	synchronize_rcu();
2487 }
2488 
2489 static int intel_pstate_get_max_freq(struct cpudata *cpu)
2490 {
2491 	return READ_ONCE(global.no_turbo) ?
2492 			cpu->pstate.max_freq : cpu->pstate.turbo_freq;
2493 }
2494 
2495 static void intel_pstate_update_perf_limits(struct cpudata *cpu,
2496 					    unsigned int policy_min,
2497 					    unsigned int policy_max)
2498 {
2499 	int perf_ctl_scaling = cpu->pstate.perf_ctl_scaling;
2500 	int32_t max_policy_perf, min_policy_perf;
2501 
2502 	max_policy_perf = policy_max / perf_ctl_scaling;
2503 	if (policy_max == policy_min) {
2504 		min_policy_perf = max_policy_perf;
2505 	} else {
2506 		min_policy_perf = policy_min / perf_ctl_scaling;
2507 		min_policy_perf = clamp_t(int32_t, min_policy_perf,
2508 					  0, max_policy_perf);
2509 	}
2510 
2511 	/*
2512 	 * HWP needs some special consideration, because HWP_REQUEST uses
2513 	 * abstract values to represent performance rather than pure ratios.
2514 	 */
2515 	if (hwp_active && cpu->pstate.scaling != perf_ctl_scaling) {
2516 		int freq;
2517 
2518 		freq = max_policy_perf * perf_ctl_scaling;
2519 		max_policy_perf = intel_pstate_freq_to_hwp(cpu, freq);
2520 		freq = min_policy_perf * perf_ctl_scaling;
2521 		min_policy_perf = intel_pstate_freq_to_hwp(cpu, freq);
2522 	}
2523 
2524 	pr_debug("cpu:%d min_policy_perf:%d max_policy_perf:%d\n",
2525 		 cpu->cpu, min_policy_perf, max_policy_perf);
2526 
2527 	/* Normalize user input to [min_perf, max_perf] */
2528 	if (per_cpu_limits) {
2529 		cpu->min_perf_ratio = min_policy_perf;
2530 		cpu->max_perf_ratio = max_policy_perf;
2531 	} else {
2532 		int turbo_max = cpu->pstate.turbo_pstate;
2533 		int32_t global_min, global_max;
2534 
2535 		/* Global limits are in percent of the maximum turbo P-state. */
2536 		global_max = DIV_ROUND_UP(turbo_max * global.max_perf_pct, 100);
2537 		global_min = DIV_ROUND_UP(turbo_max * global.min_perf_pct, 100);
2538 		global_min = clamp_t(int32_t, global_min, 0, global_max);
2539 
2540 		pr_debug("cpu:%d global_min:%d global_max:%d\n", cpu->cpu,
2541 			 global_min, global_max);
2542 
2543 		cpu->min_perf_ratio = max(min_policy_perf, global_min);
2544 		cpu->min_perf_ratio = min(cpu->min_perf_ratio, max_policy_perf);
2545 		cpu->max_perf_ratio = min(max_policy_perf, global_max);
2546 		cpu->max_perf_ratio = max(min_policy_perf, cpu->max_perf_ratio);
2547 
2548 		/* Make sure min_perf <= max_perf */
2549 		cpu->min_perf_ratio = min(cpu->min_perf_ratio,
2550 					  cpu->max_perf_ratio);
2551 
2552 	}
2553 	pr_debug("cpu:%d max_perf_ratio:%d min_perf_ratio:%d\n", cpu->cpu,
2554 		 cpu->max_perf_ratio,
2555 		 cpu->min_perf_ratio);
2556 }
2557 
2558 static int intel_pstate_set_policy(struct cpufreq_policy *policy)
2559 {
2560 	struct cpudata *cpu;
2561 
2562 	if (!policy->cpuinfo.max_freq)
2563 		return -ENODEV;
2564 
2565 	pr_debug("set_policy cpuinfo.max %u policy->max %u\n",
2566 		 policy->cpuinfo.max_freq, policy->max);
2567 
2568 	cpu = all_cpu_data[policy->cpu];
2569 	cpu->policy = policy->policy;
2570 
2571 	mutex_lock(&intel_pstate_limits_lock);
2572 
2573 	intel_pstate_update_perf_limits(cpu, policy->min, policy->max);
2574 
2575 	if (cpu->policy == CPUFREQ_POLICY_PERFORMANCE) {
2576 		int pstate = max(cpu->pstate.min_pstate, cpu->max_perf_ratio);
2577 
2578 		/*
2579 		 * NOHZ_FULL CPUs need this as the governor callback may not
2580 		 * be invoked on them.
2581 		 */
2582 		intel_pstate_clear_update_util_hook(policy->cpu);
2583 		intel_pstate_set_pstate(cpu, pstate);
2584 	} else {
2585 		intel_pstate_set_update_util_hook(policy->cpu);
2586 	}
2587 
2588 	if (hwp_active) {
2589 		/*
2590 		 * When hwp_boost was active before and dynamically it
2591 		 * was turned off, in that case we need to clear the
2592 		 * update util hook.
2593 		 */
2594 		if (!hwp_boost)
2595 			intel_pstate_clear_update_util_hook(policy->cpu);
2596 		intel_pstate_hwp_set(policy->cpu);
2597 	}
2598 	/*
2599 	 * policy->cur is never updated with the intel_pstate driver, but it
2600 	 * is used as a stale frequency value. So, keep it within limits.
2601 	 */
2602 	policy->cur = policy->min;
2603 
2604 	mutex_unlock(&intel_pstate_limits_lock);
2605 
2606 	return 0;
2607 }
2608 
2609 static void intel_pstate_adjust_policy_max(struct cpudata *cpu,
2610 					   struct cpufreq_policy_data *policy)
2611 {
2612 	if (!hwp_active &&
2613 	    cpu->pstate.max_pstate_physical > cpu->pstate.max_pstate &&
2614 	    policy->max < policy->cpuinfo.max_freq &&
2615 	    policy->max > cpu->pstate.max_freq) {
2616 		pr_debug("policy->max > max non turbo frequency\n");
2617 		policy->max = policy->cpuinfo.max_freq;
2618 	}
2619 }
2620 
2621 static void intel_pstate_verify_cpu_policy(struct cpudata *cpu,
2622 					   struct cpufreq_policy_data *policy)
2623 {
2624 	int max_freq;
2625 
2626 	if (hwp_active) {
2627 		intel_pstate_get_hwp_cap(cpu);
2628 		max_freq = READ_ONCE(global.no_turbo) ?
2629 				cpu->pstate.max_freq : cpu->pstate.turbo_freq;
2630 	} else {
2631 		max_freq = intel_pstate_get_max_freq(cpu);
2632 	}
2633 	cpufreq_verify_within_limits(policy, policy->cpuinfo.min_freq, max_freq);
2634 
2635 	intel_pstate_adjust_policy_max(cpu, policy);
2636 }
2637 
2638 static int intel_pstate_verify_policy(struct cpufreq_policy_data *policy)
2639 {
2640 	intel_pstate_verify_cpu_policy(all_cpu_data[policy->cpu], policy);
2641 
2642 	return 0;
2643 }
2644 
2645 static int intel_cpufreq_cpu_offline(struct cpufreq_policy *policy)
2646 {
2647 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2648 
2649 	pr_debug("CPU %d going offline\n", cpu->cpu);
2650 
2651 	if (cpu->suspended)
2652 		return 0;
2653 
2654 	/*
2655 	 * If the CPU is an SMT thread and it goes offline with the performance
2656 	 * settings different from the minimum, it will prevent its sibling
2657 	 * from getting to lower performance levels, so force the minimum
2658 	 * performance on CPU offline to prevent that from happening.
2659 	 */
2660 	if (hwp_active)
2661 		intel_pstate_hwp_offline(cpu);
2662 	else
2663 		intel_pstate_set_min_pstate(cpu);
2664 
2665 	intel_pstate_exit_perf_limits(policy);
2666 
2667 	return 0;
2668 }
2669 
2670 static int intel_pstate_cpu_online(struct cpufreq_policy *policy)
2671 {
2672 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2673 
2674 	pr_debug("CPU %d going online\n", cpu->cpu);
2675 
2676 	intel_pstate_init_acpi_perf_limits(policy);
2677 
2678 	if (hwp_active) {
2679 		/*
2680 		 * Re-enable HWP and clear the "suspended" flag to let "resume"
2681 		 * know that it need not do that.
2682 		 */
2683 		intel_pstate_hwp_reenable(cpu);
2684 		cpu->suspended = false;
2685 	}
2686 
2687 	return 0;
2688 }
2689 
2690 static int intel_pstate_cpu_offline(struct cpufreq_policy *policy)
2691 {
2692 	intel_pstate_clear_update_util_hook(policy->cpu);
2693 
2694 	return intel_cpufreq_cpu_offline(policy);
2695 }
2696 
2697 static int intel_pstate_cpu_exit(struct cpufreq_policy *policy)
2698 {
2699 	pr_debug("CPU %d exiting\n", policy->cpu);
2700 
2701 	policy->fast_switch_possible = false;
2702 
2703 	return 0;
2704 }
2705 
2706 static int __intel_pstate_cpu_init(struct cpufreq_policy *policy)
2707 {
2708 	struct cpudata *cpu;
2709 	int rc;
2710 
2711 	rc = intel_pstate_init_cpu(policy->cpu);
2712 	if (rc)
2713 		return rc;
2714 
2715 	cpu = all_cpu_data[policy->cpu];
2716 
2717 	cpu->max_perf_ratio = 0xFF;
2718 	cpu->min_perf_ratio = 0;
2719 
2720 	/* cpuinfo and default policy values */
2721 	policy->cpuinfo.min_freq = cpu->pstate.min_freq;
2722 	policy->cpuinfo.max_freq = READ_ONCE(global.no_turbo) ?
2723 			cpu->pstate.max_freq : cpu->pstate.turbo_freq;
2724 
2725 	policy->min = policy->cpuinfo.min_freq;
2726 	policy->max = policy->cpuinfo.max_freq;
2727 
2728 	intel_pstate_init_acpi_perf_limits(policy);
2729 
2730 	policy->fast_switch_possible = true;
2731 
2732 	return 0;
2733 }
2734 
2735 static int intel_pstate_cpu_init(struct cpufreq_policy *policy)
2736 {
2737 	int ret = __intel_pstate_cpu_init(policy);
2738 
2739 	if (ret)
2740 		return ret;
2741 
2742 	/*
2743 	 * Set the policy to powersave to provide a valid fallback value in case
2744 	 * the default cpufreq governor is neither powersave nor performance.
2745 	 */
2746 	policy->policy = CPUFREQ_POLICY_POWERSAVE;
2747 
2748 	if (hwp_active) {
2749 		struct cpudata *cpu = all_cpu_data[policy->cpu];
2750 
2751 		cpu->epp_cached = intel_pstate_get_epp(cpu, 0);
2752 	}
2753 
2754 	return 0;
2755 }
2756 
2757 static struct cpufreq_driver intel_pstate = {
2758 	.flags		= CPUFREQ_CONST_LOOPS,
2759 	.verify		= intel_pstate_verify_policy,
2760 	.setpolicy	= intel_pstate_set_policy,
2761 	.suspend	= intel_pstate_suspend,
2762 	.resume		= intel_pstate_resume,
2763 	.init		= intel_pstate_cpu_init,
2764 	.exit		= intel_pstate_cpu_exit,
2765 	.offline	= intel_pstate_cpu_offline,
2766 	.online		= intel_pstate_cpu_online,
2767 	.update_limits	= intel_pstate_update_limits,
2768 	.name		= "intel_pstate",
2769 };
2770 
2771 static int intel_cpufreq_verify_policy(struct cpufreq_policy_data *policy)
2772 {
2773 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2774 
2775 	intel_pstate_verify_cpu_policy(cpu, policy);
2776 	intel_pstate_update_perf_limits(cpu, policy->min, policy->max);
2777 
2778 	return 0;
2779 }
2780 
2781 /* Use of trace in passive mode:
2782  *
2783  * In passive mode the trace core_busy field (also known as the
2784  * performance field, and lablelled as such on the graphs; also known as
2785  * core_avg_perf) is not needed and so is re-assigned to indicate if the
2786  * driver call was via the normal or fast switch path. Various graphs
2787  * output from the intel_pstate_tracer.py utility that include core_busy
2788  * (or performance or core_avg_perf) have a fixed y-axis from 0 to 100%,
2789  * so we use 10 to indicate the normal path through the driver, and
2790  * 90 to indicate the fast switch path through the driver.
2791  * The scaled_busy field is not used, and is set to 0.
2792  */
2793 
2794 #define	INTEL_PSTATE_TRACE_TARGET 10
2795 #define	INTEL_PSTATE_TRACE_FAST_SWITCH 90
2796 
2797 static void intel_cpufreq_trace(struct cpudata *cpu, unsigned int trace_type, int old_pstate)
2798 {
2799 	struct sample *sample;
2800 
2801 	if (!trace_pstate_sample_enabled())
2802 		return;
2803 
2804 	if (!intel_pstate_sample(cpu, ktime_get()))
2805 		return;
2806 
2807 	sample = &cpu->sample;
2808 	trace_pstate_sample(trace_type,
2809 		0,
2810 		old_pstate,
2811 		cpu->pstate.current_pstate,
2812 		sample->mperf,
2813 		sample->aperf,
2814 		sample->tsc,
2815 		get_avg_frequency(cpu),
2816 		fp_toint(cpu->iowait_boost * 100));
2817 }
2818 
2819 static void intel_cpufreq_hwp_update(struct cpudata *cpu, u32 min, u32 max,
2820 				     u32 desired, bool fast_switch)
2821 {
2822 	u64 prev = READ_ONCE(cpu->hwp_req_cached), value = prev;
2823 
2824 	value &= ~HWP_MIN_PERF(~0L);
2825 	value |= HWP_MIN_PERF(min);
2826 
2827 	value &= ~HWP_MAX_PERF(~0L);
2828 	value |= HWP_MAX_PERF(max);
2829 
2830 	value &= ~HWP_DESIRED_PERF(~0L);
2831 	value |= HWP_DESIRED_PERF(desired);
2832 
2833 	if (value == prev)
2834 		return;
2835 
2836 	WRITE_ONCE(cpu->hwp_req_cached, value);
2837 	if (fast_switch)
2838 		wrmsrl(MSR_HWP_REQUEST, value);
2839 	else
2840 		wrmsrl_on_cpu(cpu->cpu, MSR_HWP_REQUEST, value);
2841 }
2842 
2843 static void intel_cpufreq_perf_ctl_update(struct cpudata *cpu,
2844 					  u32 target_pstate, bool fast_switch)
2845 {
2846 	if (fast_switch)
2847 		wrmsrl(MSR_IA32_PERF_CTL,
2848 		       pstate_funcs.get_val(cpu, target_pstate));
2849 	else
2850 		wrmsrl_on_cpu(cpu->cpu, MSR_IA32_PERF_CTL,
2851 			      pstate_funcs.get_val(cpu, target_pstate));
2852 }
2853 
2854 static int intel_cpufreq_update_pstate(struct cpufreq_policy *policy,
2855 				       int target_pstate, bool fast_switch)
2856 {
2857 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2858 	int old_pstate = cpu->pstate.current_pstate;
2859 
2860 	target_pstate = intel_pstate_prepare_request(cpu, target_pstate);
2861 	if (hwp_active) {
2862 		int max_pstate = policy->strict_target ?
2863 					target_pstate : cpu->max_perf_ratio;
2864 
2865 		intel_cpufreq_hwp_update(cpu, target_pstate, max_pstate, 0,
2866 					 fast_switch);
2867 	} else if (target_pstate != old_pstate) {
2868 		intel_cpufreq_perf_ctl_update(cpu, target_pstate, fast_switch);
2869 	}
2870 
2871 	cpu->pstate.current_pstate = target_pstate;
2872 
2873 	intel_cpufreq_trace(cpu, fast_switch ? INTEL_PSTATE_TRACE_FAST_SWITCH :
2874 			    INTEL_PSTATE_TRACE_TARGET, old_pstate);
2875 
2876 	return target_pstate;
2877 }
2878 
2879 static int intel_cpufreq_target(struct cpufreq_policy *policy,
2880 				unsigned int target_freq,
2881 				unsigned int relation)
2882 {
2883 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2884 	struct cpufreq_freqs freqs;
2885 	int target_pstate;
2886 
2887 	freqs.old = policy->cur;
2888 	freqs.new = target_freq;
2889 
2890 	cpufreq_freq_transition_begin(policy, &freqs);
2891 
2892 	target_pstate = intel_pstate_freq_to_hwp_rel(cpu, freqs.new, relation);
2893 	target_pstate = intel_cpufreq_update_pstate(policy, target_pstate, false);
2894 
2895 	freqs.new = target_pstate * cpu->pstate.scaling;
2896 
2897 	cpufreq_freq_transition_end(policy, &freqs, false);
2898 
2899 	return 0;
2900 }
2901 
2902 static unsigned int intel_cpufreq_fast_switch(struct cpufreq_policy *policy,
2903 					      unsigned int target_freq)
2904 {
2905 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2906 	int target_pstate;
2907 
2908 	target_pstate = intel_pstate_freq_to_hwp(cpu, target_freq);
2909 
2910 	target_pstate = intel_cpufreq_update_pstate(policy, target_pstate, true);
2911 
2912 	return target_pstate * cpu->pstate.scaling;
2913 }
2914 
2915 static void intel_cpufreq_adjust_perf(unsigned int cpunum,
2916 				      unsigned long min_perf,
2917 				      unsigned long target_perf,
2918 				      unsigned long capacity)
2919 {
2920 	struct cpudata *cpu = all_cpu_data[cpunum];
2921 	u64 hwp_cap = READ_ONCE(cpu->hwp_cap_cached);
2922 	int old_pstate = cpu->pstate.current_pstate;
2923 	int cap_pstate, min_pstate, max_pstate, target_pstate;
2924 
2925 	cap_pstate = READ_ONCE(global.no_turbo) ?
2926 					HWP_GUARANTEED_PERF(hwp_cap) :
2927 					HWP_HIGHEST_PERF(hwp_cap);
2928 
2929 	/* Optimization: Avoid unnecessary divisions. */
2930 
2931 	target_pstate = cap_pstate;
2932 	if (target_perf < capacity)
2933 		target_pstate = DIV_ROUND_UP(cap_pstate * target_perf, capacity);
2934 
2935 	min_pstate = cap_pstate;
2936 	if (min_perf < capacity)
2937 		min_pstate = DIV_ROUND_UP(cap_pstate * min_perf, capacity);
2938 
2939 	if (min_pstate < cpu->pstate.min_pstate)
2940 		min_pstate = cpu->pstate.min_pstate;
2941 
2942 	if (min_pstate < cpu->min_perf_ratio)
2943 		min_pstate = cpu->min_perf_ratio;
2944 
2945 	if (min_pstate > cpu->max_perf_ratio)
2946 		min_pstate = cpu->max_perf_ratio;
2947 
2948 	max_pstate = min(cap_pstate, cpu->max_perf_ratio);
2949 	if (max_pstate < min_pstate)
2950 		max_pstate = min_pstate;
2951 
2952 	target_pstate = clamp_t(int, target_pstate, min_pstate, max_pstate);
2953 
2954 	intel_cpufreq_hwp_update(cpu, min_pstate, max_pstate, target_pstate, true);
2955 
2956 	cpu->pstate.current_pstate = target_pstate;
2957 	intel_cpufreq_trace(cpu, INTEL_PSTATE_TRACE_FAST_SWITCH, old_pstate);
2958 }
2959 
2960 static int intel_cpufreq_cpu_init(struct cpufreq_policy *policy)
2961 {
2962 	struct freq_qos_request *req;
2963 	struct cpudata *cpu;
2964 	struct device *dev;
2965 	int ret, freq;
2966 
2967 	dev = get_cpu_device(policy->cpu);
2968 	if (!dev)
2969 		return -ENODEV;
2970 
2971 	ret = __intel_pstate_cpu_init(policy);
2972 	if (ret)
2973 		return ret;
2974 
2975 	policy->cpuinfo.transition_latency = INTEL_CPUFREQ_TRANSITION_LATENCY;
2976 	/* This reflects the intel_pstate_get_cpu_pstates() setting. */
2977 	policy->cur = policy->cpuinfo.min_freq;
2978 
2979 	req = kcalloc(2, sizeof(*req), GFP_KERNEL);
2980 	if (!req) {
2981 		ret = -ENOMEM;
2982 		goto pstate_exit;
2983 	}
2984 
2985 	cpu = all_cpu_data[policy->cpu];
2986 
2987 	if (hwp_active) {
2988 		u64 value;
2989 
2990 		policy->transition_delay_us = INTEL_CPUFREQ_TRANSITION_DELAY_HWP;
2991 
2992 		intel_pstate_get_hwp_cap(cpu);
2993 
2994 		rdmsrl_on_cpu(cpu->cpu, MSR_HWP_REQUEST, &value);
2995 		WRITE_ONCE(cpu->hwp_req_cached, value);
2996 
2997 		cpu->epp_cached = intel_pstate_get_epp(cpu, value);
2998 	} else {
2999 		policy->transition_delay_us = INTEL_CPUFREQ_TRANSITION_DELAY;
3000 	}
3001 
3002 	freq = DIV_ROUND_UP(cpu->pstate.turbo_freq * global.min_perf_pct, 100);
3003 
3004 	ret = freq_qos_add_request(&policy->constraints, req, FREQ_QOS_MIN,
3005 				   freq);
3006 	if (ret < 0) {
3007 		dev_err(dev, "Failed to add min-freq constraint (%d)\n", ret);
3008 		goto free_req;
3009 	}
3010 
3011 	freq = DIV_ROUND_UP(cpu->pstate.turbo_freq * global.max_perf_pct, 100);
3012 
3013 	ret = freq_qos_add_request(&policy->constraints, req + 1, FREQ_QOS_MAX,
3014 				   freq);
3015 	if (ret < 0) {
3016 		dev_err(dev, "Failed to add max-freq constraint (%d)\n", ret);
3017 		goto remove_min_req;
3018 	}
3019 
3020 	policy->driver_data = req;
3021 
3022 	return 0;
3023 
3024 remove_min_req:
3025 	freq_qos_remove_request(req);
3026 free_req:
3027 	kfree(req);
3028 pstate_exit:
3029 	intel_pstate_exit_perf_limits(policy);
3030 
3031 	return ret;
3032 }
3033 
3034 static int intel_cpufreq_cpu_exit(struct cpufreq_policy *policy)
3035 {
3036 	struct freq_qos_request *req;
3037 
3038 	req = policy->driver_data;
3039 
3040 	freq_qos_remove_request(req + 1);
3041 	freq_qos_remove_request(req);
3042 	kfree(req);
3043 
3044 	return intel_pstate_cpu_exit(policy);
3045 }
3046 
3047 static int intel_cpufreq_suspend(struct cpufreq_policy *policy)
3048 {
3049 	intel_pstate_suspend(policy);
3050 
3051 	if (hwp_active) {
3052 		struct cpudata *cpu = all_cpu_data[policy->cpu];
3053 		u64 value = READ_ONCE(cpu->hwp_req_cached);
3054 
3055 		/*
3056 		 * Clear the desired perf field in MSR_HWP_REQUEST in case
3057 		 * intel_cpufreq_adjust_perf() is in use and the last value
3058 		 * written by it may not be suitable.
3059 		 */
3060 		value &= ~HWP_DESIRED_PERF(~0L);
3061 		wrmsrl_on_cpu(cpu->cpu, MSR_HWP_REQUEST, value);
3062 		WRITE_ONCE(cpu->hwp_req_cached, value);
3063 	}
3064 
3065 	return 0;
3066 }
3067 
3068 static struct cpufreq_driver intel_cpufreq = {
3069 	.flags		= CPUFREQ_CONST_LOOPS,
3070 	.verify		= intel_cpufreq_verify_policy,
3071 	.target		= intel_cpufreq_target,
3072 	.fast_switch	= intel_cpufreq_fast_switch,
3073 	.init		= intel_cpufreq_cpu_init,
3074 	.exit		= intel_cpufreq_cpu_exit,
3075 	.offline	= intel_cpufreq_cpu_offline,
3076 	.online		= intel_pstate_cpu_online,
3077 	.suspend	= intel_cpufreq_suspend,
3078 	.resume		= intel_pstate_resume,
3079 	.update_limits	= intel_pstate_update_limits,
3080 	.name		= "intel_cpufreq",
3081 };
3082 
3083 static struct cpufreq_driver *default_driver;
3084 
3085 static void intel_pstate_driver_cleanup(void)
3086 {
3087 	unsigned int cpu;
3088 
3089 	cpus_read_lock();
3090 	for_each_online_cpu(cpu) {
3091 		if (all_cpu_data[cpu]) {
3092 			if (intel_pstate_driver == &intel_pstate)
3093 				intel_pstate_clear_update_util_hook(cpu);
3094 
3095 			kfree(all_cpu_data[cpu]);
3096 			WRITE_ONCE(all_cpu_data[cpu], NULL);
3097 		}
3098 	}
3099 	cpus_read_unlock();
3100 
3101 	intel_pstate_driver = NULL;
3102 }
3103 
3104 static int intel_pstate_register_driver(struct cpufreq_driver *driver)
3105 {
3106 	int ret;
3107 
3108 	if (driver == &intel_pstate)
3109 		intel_pstate_sysfs_expose_hwp_dynamic_boost();
3110 
3111 	memset(&global, 0, sizeof(global));
3112 	global.max_perf_pct = 100;
3113 	global.turbo_disabled = turbo_is_disabled();
3114 	global.no_turbo = global.turbo_disabled;
3115 
3116 	arch_set_max_freq_ratio(global.turbo_disabled);
3117 
3118 	intel_pstate_driver = driver;
3119 	ret = cpufreq_register_driver(intel_pstate_driver);
3120 	if (ret) {
3121 		intel_pstate_driver_cleanup();
3122 		return ret;
3123 	}
3124 
3125 	global.min_perf_pct = min_perf_pct_min();
3126 
3127 	return 0;
3128 }
3129 
3130 static ssize_t intel_pstate_show_status(char *buf)
3131 {
3132 	if (!intel_pstate_driver)
3133 		return sprintf(buf, "off\n");
3134 
3135 	return sprintf(buf, "%s\n", intel_pstate_driver == &intel_pstate ?
3136 					"active" : "passive");
3137 }
3138 
3139 static int intel_pstate_update_status(const char *buf, size_t size)
3140 {
3141 	if (size == 3 && !strncmp(buf, "off", size)) {
3142 		if (!intel_pstate_driver)
3143 			return -EINVAL;
3144 
3145 		if (hwp_active)
3146 			return -EBUSY;
3147 
3148 		cpufreq_unregister_driver(intel_pstate_driver);
3149 		intel_pstate_driver_cleanup();
3150 		return 0;
3151 	}
3152 
3153 	if (size == 6 && !strncmp(buf, "active", size)) {
3154 		if (intel_pstate_driver) {
3155 			if (intel_pstate_driver == &intel_pstate)
3156 				return 0;
3157 
3158 			cpufreq_unregister_driver(intel_pstate_driver);
3159 		}
3160 
3161 		return intel_pstate_register_driver(&intel_pstate);
3162 	}
3163 
3164 	if (size == 7 && !strncmp(buf, "passive", size)) {
3165 		if (intel_pstate_driver) {
3166 			if (intel_pstate_driver == &intel_cpufreq)
3167 				return 0;
3168 
3169 			cpufreq_unregister_driver(intel_pstate_driver);
3170 			intel_pstate_sysfs_hide_hwp_dynamic_boost();
3171 		}
3172 
3173 		return intel_pstate_register_driver(&intel_cpufreq);
3174 	}
3175 
3176 	return -EINVAL;
3177 }
3178 
3179 static int no_load __initdata;
3180 static int no_hwp __initdata;
3181 static int hwp_only __initdata;
3182 static unsigned int force_load __initdata;
3183 
3184 static int __init intel_pstate_msrs_not_valid(void)
3185 {
3186 	if (!pstate_funcs.get_max(0) ||
3187 	    !pstate_funcs.get_min(0) ||
3188 	    !pstate_funcs.get_turbo(0))
3189 		return -ENODEV;
3190 
3191 	return 0;
3192 }
3193 
3194 static void __init copy_cpu_funcs(struct pstate_funcs *funcs)
3195 {
3196 	pstate_funcs.get_max   = funcs->get_max;
3197 	pstate_funcs.get_max_physical = funcs->get_max_physical;
3198 	pstate_funcs.get_min   = funcs->get_min;
3199 	pstate_funcs.get_turbo = funcs->get_turbo;
3200 	pstate_funcs.get_scaling = funcs->get_scaling;
3201 	pstate_funcs.get_val   = funcs->get_val;
3202 	pstate_funcs.get_vid   = funcs->get_vid;
3203 	pstate_funcs.get_aperf_mperf_shift = funcs->get_aperf_mperf_shift;
3204 }
3205 
3206 #ifdef CONFIG_ACPI
3207 
3208 static bool __init intel_pstate_no_acpi_pss(void)
3209 {
3210 	int i;
3211 
3212 	for_each_possible_cpu(i) {
3213 		acpi_status status;
3214 		union acpi_object *pss;
3215 		struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
3216 		struct acpi_processor *pr = per_cpu(processors, i);
3217 
3218 		if (!pr)
3219 			continue;
3220 
3221 		status = acpi_evaluate_object(pr->handle, "_PSS", NULL, &buffer);
3222 		if (ACPI_FAILURE(status))
3223 			continue;
3224 
3225 		pss = buffer.pointer;
3226 		if (pss && pss->type == ACPI_TYPE_PACKAGE) {
3227 			kfree(pss);
3228 			return false;
3229 		}
3230 
3231 		kfree(pss);
3232 	}
3233 
3234 	pr_debug("ACPI _PSS not found\n");
3235 	return true;
3236 }
3237 
3238 static bool __init intel_pstate_no_acpi_pcch(void)
3239 {
3240 	acpi_status status;
3241 	acpi_handle handle;
3242 
3243 	status = acpi_get_handle(NULL, "\\_SB", &handle);
3244 	if (ACPI_FAILURE(status))
3245 		goto not_found;
3246 
3247 	if (acpi_has_method(handle, "PCCH"))
3248 		return false;
3249 
3250 not_found:
3251 	pr_debug("ACPI PCCH not found\n");
3252 	return true;
3253 }
3254 
3255 static bool __init intel_pstate_has_acpi_ppc(void)
3256 {
3257 	int i;
3258 
3259 	for_each_possible_cpu(i) {
3260 		struct acpi_processor *pr = per_cpu(processors, i);
3261 
3262 		if (!pr)
3263 			continue;
3264 		if (acpi_has_method(pr->handle, "_PPC"))
3265 			return true;
3266 	}
3267 	pr_debug("ACPI _PPC not found\n");
3268 	return false;
3269 }
3270 
3271 enum {
3272 	PSS,
3273 	PPC,
3274 };
3275 
3276 /* Hardware vendor-specific info that has its own power management modes */
3277 static struct acpi_platform_list plat_info[] __initdata = {
3278 	{"HP    ", "ProLiant", 0, ACPI_SIG_FADT, all_versions, NULL, PSS},
3279 	{"ORACLE", "X4-2    ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3280 	{"ORACLE", "X4-2L   ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3281 	{"ORACLE", "X4-2B   ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3282 	{"ORACLE", "X3-2    ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3283 	{"ORACLE", "X3-2L   ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3284 	{"ORACLE", "X3-2B   ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3285 	{"ORACLE", "X4470M2 ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3286 	{"ORACLE", "X4270M3 ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3287 	{"ORACLE", "X4270M2 ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3288 	{"ORACLE", "X4170M2 ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3289 	{"ORACLE", "X4170 M3", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3290 	{"ORACLE", "X4275 M3", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3291 	{"ORACLE", "X6-2    ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3292 	{"ORACLE", "Sudbury ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3293 	{ } /* End */
3294 };
3295 
3296 #define BITMASK_OOB	(BIT(8) | BIT(18))
3297 
3298 static bool __init intel_pstate_platform_pwr_mgmt_exists(void)
3299 {
3300 	const struct x86_cpu_id *id;
3301 	u64 misc_pwr;
3302 	int idx;
3303 
3304 	id = x86_match_cpu(intel_pstate_cpu_oob_ids);
3305 	if (id) {
3306 		rdmsrl(MSR_MISC_PWR_MGMT, misc_pwr);
3307 		if (misc_pwr & BITMASK_OOB) {
3308 			pr_debug("Bit 8 or 18 in the MISC_PWR_MGMT MSR set\n");
3309 			pr_debug("P states are controlled in Out of Band mode by the firmware/hardware\n");
3310 			return true;
3311 		}
3312 	}
3313 
3314 	idx = acpi_match_platform_list(plat_info);
3315 	if (idx < 0)
3316 		return false;
3317 
3318 	switch (plat_info[idx].data) {
3319 	case PSS:
3320 		if (!intel_pstate_no_acpi_pss())
3321 			return false;
3322 
3323 		return intel_pstate_no_acpi_pcch();
3324 	case PPC:
3325 		return intel_pstate_has_acpi_ppc() && !force_load;
3326 	}
3327 
3328 	return false;
3329 }
3330 
3331 static void intel_pstate_request_control_from_smm(void)
3332 {
3333 	/*
3334 	 * It may be unsafe to request P-states control from SMM if _PPC support
3335 	 * has not been enabled.
3336 	 */
3337 	if (acpi_ppc)
3338 		acpi_processor_pstate_control();
3339 }
3340 #else /* CONFIG_ACPI not enabled */
3341 static inline bool intel_pstate_platform_pwr_mgmt_exists(void) { return false; }
3342 static inline bool intel_pstate_has_acpi_ppc(void) { return false; }
3343 static inline void intel_pstate_request_control_from_smm(void) {}
3344 #endif /* CONFIG_ACPI */
3345 
3346 #define INTEL_PSTATE_HWP_BROADWELL	0x01
3347 
3348 #define X86_MATCH_HWP(model, hwp_mode)					\
3349 	X86_MATCH_VENDOR_FAM_MODEL_FEATURE(INTEL, 6, INTEL_FAM6_##model, \
3350 					   X86_FEATURE_HWP, hwp_mode)
3351 
3352 static const struct x86_cpu_id hwp_support_ids[] __initconst = {
3353 	X86_MATCH_HWP(BROADWELL_X,	INTEL_PSTATE_HWP_BROADWELL),
3354 	X86_MATCH_HWP(BROADWELL_D,	INTEL_PSTATE_HWP_BROADWELL),
3355 	X86_MATCH_HWP(ANY,		0),
3356 	{}
3357 };
3358 
3359 static bool intel_pstate_hwp_is_enabled(void)
3360 {
3361 	u64 value;
3362 
3363 	rdmsrl(MSR_PM_ENABLE, value);
3364 	return !!(value & 0x1);
3365 }
3366 
3367 #define POWERSAVE_MASK			GENMASK(7, 0)
3368 #define BALANCE_POWER_MASK		GENMASK(15, 8)
3369 #define BALANCE_PERFORMANCE_MASK	GENMASK(23, 16)
3370 #define PERFORMANCE_MASK		GENMASK(31, 24)
3371 
3372 #define HWP_SET_EPP_VALUES(powersave, balance_power, balance_perf, performance) \
3373 	(FIELD_PREP_CONST(POWERSAVE_MASK, powersave) |\
3374 	 FIELD_PREP_CONST(BALANCE_POWER_MASK, balance_power) |\
3375 	 FIELD_PREP_CONST(BALANCE_PERFORMANCE_MASK, balance_perf) |\
3376 	 FIELD_PREP_CONST(PERFORMANCE_MASK, performance))
3377 
3378 #define HWP_SET_DEF_BALANCE_PERF_EPP(balance_perf) \
3379 	(HWP_SET_EPP_VALUES(HWP_EPP_POWERSAVE, HWP_EPP_BALANCE_POWERSAVE,\
3380 	 balance_perf, HWP_EPP_PERFORMANCE))
3381 
3382 static const struct x86_cpu_id intel_epp_default[] = {
3383 	/*
3384 	 * Set EPP value as 102, this is the max suggested EPP
3385 	 * which can result in one core turbo frequency for
3386 	 * AlderLake Mobile CPUs.
3387 	 */
3388 	X86_MATCH_INTEL_FAM6_MODEL(ALDERLAKE_L, HWP_SET_DEF_BALANCE_PERF_EPP(102)),
3389 	X86_MATCH_INTEL_FAM6_MODEL(SAPPHIRERAPIDS_X, HWP_SET_DEF_BALANCE_PERF_EPP(32)),
3390 	X86_MATCH_INTEL_FAM6_MODEL(METEORLAKE_L, HWP_SET_EPP_VALUES(HWP_EPP_POWERSAVE,
3391 							HWP_EPP_BALANCE_POWERSAVE, 115, 16)),
3392 	{}
3393 };
3394 
3395 static const struct x86_cpu_id intel_hybrid_scaling_factor[] = {
3396 	X86_MATCH_INTEL_FAM6_MODEL(METEORLAKE_L, HYBRID_SCALING_FACTOR_MTL),
3397 	{}
3398 };
3399 
3400 static int __init intel_pstate_init(void)
3401 {
3402 	static struct cpudata **_all_cpu_data;
3403 	const struct x86_cpu_id *id;
3404 	int rc;
3405 
3406 	if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL)
3407 		return -ENODEV;
3408 
3409 	id = x86_match_cpu(hwp_support_ids);
3410 	if (id) {
3411 		hwp_forced = intel_pstate_hwp_is_enabled();
3412 
3413 		if (hwp_forced)
3414 			pr_info("HWP enabled by BIOS\n");
3415 		else if (no_load)
3416 			return -ENODEV;
3417 
3418 		copy_cpu_funcs(&core_funcs);
3419 		/*
3420 		 * Avoid enabling HWP for processors without EPP support,
3421 		 * because that means incomplete HWP implementation which is a
3422 		 * corner case and supporting it is generally problematic.
3423 		 *
3424 		 * If HWP is enabled already, though, there is no choice but to
3425 		 * deal with it.
3426 		 */
3427 		if ((!no_hwp && boot_cpu_has(X86_FEATURE_HWP_EPP)) || hwp_forced) {
3428 			hwp_active = true;
3429 			hwp_mode_bdw = id->driver_data;
3430 			intel_pstate.attr = hwp_cpufreq_attrs;
3431 			intel_cpufreq.attr = hwp_cpufreq_attrs;
3432 			intel_cpufreq.flags |= CPUFREQ_NEED_UPDATE_LIMITS;
3433 			intel_cpufreq.adjust_perf = intel_cpufreq_adjust_perf;
3434 			if (!default_driver)
3435 				default_driver = &intel_pstate;
3436 
3437 			pstate_funcs.get_cpu_scaling = hwp_get_cpu_scaling;
3438 
3439 			goto hwp_cpu_matched;
3440 		}
3441 		pr_info("HWP not enabled\n");
3442 	} else {
3443 		if (no_load)
3444 			return -ENODEV;
3445 
3446 		id = x86_match_cpu(intel_pstate_cpu_ids);
3447 		if (!id) {
3448 			pr_info("CPU model not supported\n");
3449 			return -ENODEV;
3450 		}
3451 
3452 		copy_cpu_funcs((struct pstate_funcs *)id->driver_data);
3453 	}
3454 
3455 	if (intel_pstate_msrs_not_valid()) {
3456 		pr_info("Invalid MSRs\n");
3457 		return -ENODEV;
3458 	}
3459 	/* Without HWP start in the passive mode. */
3460 	if (!default_driver)
3461 		default_driver = &intel_cpufreq;
3462 
3463 hwp_cpu_matched:
3464 	/*
3465 	 * The Intel pstate driver will be ignored if the platform
3466 	 * firmware has its own power management modes.
3467 	 */
3468 	if (intel_pstate_platform_pwr_mgmt_exists()) {
3469 		pr_info("P-states controlled by the platform\n");
3470 		return -ENODEV;
3471 	}
3472 
3473 	if (!hwp_active && hwp_only)
3474 		return -ENOTSUPP;
3475 
3476 	pr_info("Intel P-state driver initializing\n");
3477 
3478 	_all_cpu_data = vzalloc(array_size(sizeof(void *), num_possible_cpus()));
3479 	if (!_all_cpu_data)
3480 		return -ENOMEM;
3481 
3482 	WRITE_ONCE(all_cpu_data, _all_cpu_data);
3483 
3484 	intel_pstate_request_control_from_smm();
3485 
3486 	intel_pstate_sysfs_expose_params();
3487 
3488 	if (hwp_active) {
3489 		const struct x86_cpu_id *id = x86_match_cpu(intel_epp_default);
3490 		const struct x86_cpu_id *hybrid_id = x86_match_cpu(intel_hybrid_scaling_factor);
3491 
3492 		if (id) {
3493 			epp_values[EPP_INDEX_POWERSAVE] =
3494 					FIELD_GET(POWERSAVE_MASK, id->driver_data);
3495 			epp_values[EPP_INDEX_BALANCE_POWERSAVE] =
3496 					FIELD_GET(BALANCE_POWER_MASK, id->driver_data);
3497 			epp_values[EPP_INDEX_BALANCE_PERFORMANCE] =
3498 					FIELD_GET(BALANCE_PERFORMANCE_MASK, id->driver_data);
3499 			epp_values[EPP_INDEX_PERFORMANCE] =
3500 					FIELD_GET(PERFORMANCE_MASK, id->driver_data);
3501 			pr_debug("Updated EPPs powersave:%x balanced power:%x balanced perf:%x performance:%x\n",
3502 				 epp_values[EPP_INDEX_POWERSAVE],
3503 				 epp_values[EPP_INDEX_BALANCE_POWERSAVE],
3504 				 epp_values[EPP_INDEX_BALANCE_PERFORMANCE],
3505 				 epp_values[EPP_INDEX_PERFORMANCE]);
3506 		}
3507 
3508 		if (hybrid_id) {
3509 			hybrid_scaling_factor = hybrid_id->driver_data;
3510 			pr_debug("hybrid scaling factor: %d\n", hybrid_scaling_factor);
3511 		}
3512 
3513 	}
3514 
3515 	mutex_lock(&intel_pstate_driver_lock);
3516 	rc = intel_pstate_register_driver(default_driver);
3517 	mutex_unlock(&intel_pstate_driver_lock);
3518 	if (rc) {
3519 		intel_pstate_sysfs_remove();
3520 		return rc;
3521 	}
3522 
3523 	if (hwp_active) {
3524 		const struct x86_cpu_id *id;
3525 
3526 		id = x86_match_cpu(intel_pstate_cpu_ee_disable_ids);
3527 		if (id) {
3528 			set_power_ctl_ee_state(false);
3529 			pr_info("Disabling energy efficiency optimization\n");
3530 		}
3531 
3532 		pr_info("HWP enabled\n");
3533 	} else if (boot_cpu_has(X86_FEATURE_HYBRID_CPU)) {
3534 		pr_warn("Problematic setup: Hybrid processor with disabled HWP\n");
3535 	}
3536 
3537 	return 0;
3538 }
3539 device_initcall(intel_pstate_init);
3540 
3541 static int __init intel_pstate_setup(char *str)
3542 {
3543 	if (!str)
3544 		return -EINVAL;
3545 
3546 	if (!strcmp(str, "disable"))
3547 		no_load = 1;
3548 	else if (!strcmp(str, "active"))
3549 		default_driver = &intel_pstate;
3550 	else if (!strcmp(str, "passive"))
3551 		default_driver = &intel_cpufreq;
3552 
3553 	if (!strcmp(str, "no_hwp"))
3554 		no_hwp = 1;
3555 
3556 	if (!strcmp(str, "force"))
3557 		force_load = 1;
3558 	if (!strcmp(str, "hwp_only"))
3559 		hwp_only = 1;
3560 	if (!strcmp(str, "per_cpu_perf_limits"))
3561 		per_cpu_limits = true;
3562 
3563 #ifdef CONFIG_ACPI
3564 	if (!strcmp(str, "support_acpi_ppc"))
3565 		acpi_ppc = true;
3566 #endif
3567 
3568 	return 0;
3569 }
3570 early_param("intel_pstate", intel_pstate_setup);
3571 
3572 MODULE_AUTHOR("Dirk Brandewie <dirk.j.brandewie@intel.com>");
3573 MODULE_DESCRIPTION("'intel_pstate' - P state driver Intel Core processors");
3574