xref: /linux/drivers/cpufreq/intel_pstate.c (revision 7f71507851fc7764b36a3221839607d3a45c2025)
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/sched/smt.h>
20 #include <linux/list.h>
21 #include <linux/cpu.h>
22 #include <linux/cpufreq.h>
23 #include <linux/sysfs.h>
24 #include <linux/types.h>
25 #include <linux/fs.h>
26 #include <linux/acpi.h>
27 #include <linux/vmalloc.h>
28 #include <linux/pm_qos.h>
29 #include <linux/bitfield.h>
30 #include <trace/events/power.h>
31 
32 #include <asm/cpu.h>
33 #include <asm/div64.h>
34 #include <asm/msr.h>
35 #include <asm/cpu_device_id.h>
36 #include <asm/cpufeature.h>
37 #include <asm/intel-family.h>
38 #include "../drivers/thermal/intel/thermal_interrupt.h"
39 
40 #define INTEL_PSTATE_SAMPLING_INTERVAL	(10 * NSEC_PER_MSEC)
41 
42 #define INTEL_CPUFREQ_TRANSITION_LATENCY	20000
43 #define INTEL_CPUFREQ_TRANSITION_DELAY_HWP	5000
44 #define INTEL_CPUFREQ_TRANSITION_DELAY		500
45 
46 #ifdef CONFIG_ACPI
47 #include <acpi/processor.h>
48 #include <acpi/cppc_acpi.h>
49 #endif
50 
51 #define FRAC_BITS 8
52 #define int_tofp(X) ((int64_t)(X) << FRAC_BITS)
53 #define fp_toint(X) ((X) >> FRAC_BITS)
54 
55 #define ONE_EIGHTH_FP ((int64_t)1 << (FRAC_BITS - 3))
56 
57 #define EXT_BITS 6
58 #define EXT_FRAC_BITS (EXT_BITS + FRAC_BITS)
59 #define fp_ext_toint(X) ((X) >> EXT_FRAC_BITS)
60 #define int_ext_tofp(X) ((int64_t)(X) << EXT_FRAC_BITS)
61 
62 static inline int32_t mul_fp(int32_t x, int32_t y)
63 {
64 	return ((int64_t)x * (int64_t)y) >> FRAC_BITS;
65 }
66 
67 static inline int32_t div_fp(s64 x, s64 y)
68 {
69 	return div64_s64((int64_t)x << FRAC_BITS, y);
70 }
71 
72 static inline int ceiling_fp(int32_t x)
73 {
74 	int mask, ret;
75 
76 	ret = fp_toint(x);
77 	mask = (1 << FRAC_BITS) - 1;
78 	if (x & mask)
79 		ret += 1;
80 	return ret;
81 }
82 
83 static inline u64 mul_ext_fp(u64 x, u64 y)
84 {
85 	return (x * y) >> EXT_FRAC_BITS;
86 }
87 
88 static inline u64 div_ext_fp(u64 x, u64 y)
89 {
90 	return div64_u64(x << EXT_FRAC_BITS, y);
91 }
92 
93 /**
94  * struct sample -	Store performance sample
95  * @core_avg_perf:	Ratio of APERF/MPERF which is the actual average
96  *			performance during last sample period
97  * @busy_scaled:	Scaled busy value which is used to calculate next
98  *			P state. This can be different than core_avg_perf
99  *			to account for cpu idle period
100  * @aperf:		Difference of actual performance frequency clock count
101  *			read from APERF MSR between last and current sample
102  * @mperf:		Difference of maximum performance frequency clock count
103  *			read from MPERF MSR between last and current sample
104  * @tsc:		Difference of time stamp counter between last and
105  *			current sample
106  * @time:		Current time from scheduler
107  *
108  * This structure is used in the cpudata structure to store performance sample
109  * data for choosing next P State.
110  */
111 struct sample {
112 	int32_t core_avg_perf;
113 	int32_t busy_scaled;
114 	u64 aperf;
115 	u64 mperf;
116 	u64 tsc;
117 	u64 time;
118 };
119 
120 /**
121  * struct pstate_data - Store P state data
122  * @current_pstate:	Current requested P state
123  * @min_pstate:		Min P state possible for this platform
124  * @max_pstate:		Max P state possible for this platform
125  * @max_pstate_physical:This is physical Max P state for a processor
126  *			This can be higher than the max_pstate which can
127  *			be limited by platform thermal design power limits
128  * @perf_ctl_scaling:	PERF_CTL P-state to frequency scaling factor
129  * @scaling:		Scaling factor between performance and frequency
130  * @turbo_pstate:	Max Turbo P state possible for this platform
131  * @min_freq:		@min_pstate frequency in cpufreq units
132  * @max_freq:		@max_pstate frequency in cpufreq units
133  * @turbo_freq:		@turbo_pstate frequency in cpufreq units
134  *
135  * Stores the per cpu model P state limits and current P state.
136  */
137 struct pstate_data {
138 	int	current_pstate;
139 	int	min_pstate;
140 	int	max_pstate;
141 	int	max_pstate_physical;
142 	int	perf_ctl_scaling;
143 	int	scaling;
144 	int	turbo_pstate;
145 	unsigned int min_freq;
146 	unsigned int max_freq;
147 	unsigned int turbo_freq;
148 };
149 
150 /**
151  * struct vid_data -	Stores voltage information data
152  * @min:		VID data for this platform corresponding to
153  *			the lowest P state
154  * @max:		VID data corresponding to the highest P State.
155  * @turbo:		VID data for turbo P state
156  * @ratio:		Ratio of (vid max - vid min) /
157  *			(max P state - Min P State)
158  *
159  * Stores the voltage data for DVFS (Dynamic Voltage and Frequency Scaling)
160  * This data is used in Atom platforms, where in addition to target P state,
161  * the voltage data needs to be specified to select next P State.
162  */
163 struct vid_data {
164 	int min;
165 	int max;
166 	int turbo;
167 	int32_t ratio;
168 };
169 
170 /**
171  * struct global_params - Global parameters, mostly tunable via sysfs.
172  * @no_turbo:		Whether or not to use turbo P-states.
173  * @turbo_disabled:	Whether or not turbo P-states are available at all,
174  *			based on the MSR_IA32_MISC_ENABLE value and whether or
175  *			not the maximum reported turbo P-state is different from
176  *			the maximum reported non-turbo one.
177  * @min_perf_pct:	Minimum capacity limit in percent of the maximum turbo
178  *			P-state capacity.
179  * @max_perf_pct:	Maximum capacity limit in percent of the maximum turbo
180  *			P-state capacity.
181  */
182 struct global_params {
183 	bool no_turbo;
184 	bool turbo_disabled;
185 	int max_perf_pct;
186 	int min_perf_pct;
187 };
188 
189 /**
190  * struct cpudata -	Per CPU instance data storage
191  * @cpu:		CPU number for this instance data
192  * @policy:		CPUFreq policy value
193  * @update_util:	CPUFreq utility callback information
194  * @update_util_set:	CPUFreq utility callback is set
195  * @iowait_boost:	iowait-related boost fraction
196  * @last_update:	Time of the last update.
197  * @pstate:		Stores P state limits for this CPU
198  * @vid:		Stores VID limits for this CPU
199  * @last_sample_time:	Last Sample time
200  * @aperf_mperf_shift:	APERF vs MPERF counting frequency difference
201  * @prev_aperf:		Last APERF value read from APERF MSR
202  * @prev_mperf:		Last MPERF value read from MPERF MSR
203  * @prev_tsc:		Last timestamp counter (TSC) value
204  * @sample:		Storage for storing last Sample data
205  * @min_perf_ratio:	Minimum capacity in terms of PERF or HWP ratios
206  * @max_perf_ratio:	Maximum capacity in terms of PERF or HWP ratios
207  * @acpi_perf_data:	Stores ACPI perf information read from _PSS
208  * @valid_pss_table:	Set to true for valid ACPI _PSS entries found
209  * @epp_powersave:	Last saved HWP energy performance preference
210  *			(EPP) or energy performance bias (EPB),
211  *			when policy switched to performance
212  * @epp_policy:		Last saved policy used to set EPP/EPB
213  * @epp_default:	Power on default HWP energy performance
214  *			preference/bias
215  * @epp_cached:		Cached HWP energy-performance preference value
216  * @hwp_req_cached:	Cached value of the last HWP Request MSR
217  * @hwp_cap_cached:	Cached value of the last HWP Capabilities MSR
218  * @last_io_update:	Last time when IO wake flag was set
219  * @capacity_perf:	Highest perf used for scale invariance
220  * @sched_flags:	Store scheduler flags for possible cross CPU update
221  * @hwp_boost_min:	Last HWP boosted min performance
222  * @suspended:		Whether or not the driver has been suspended.
223  * @hwp_notify_work:	workqueue for HWP notifications.
224  *
225  * This structure stores per CPU instance data for all CPUs.
226  */
227 struct cpudata {
228 	int cpu;
229 
230 	unsigned int policy;
231 	struct update_util_data update_util;
232 	bool   update_util_set;
233 
234 	struct pstate_data pstate;
235 	struct vid_data vid;
236 
237 	u64	last_update;
238 	u64	last_sample_time;
239 	u64	aperf_mperf_shift;
240 	u64	prev_aperf;
241 	u64	prev_mperf;
242 	u64	prev_tsc;
243 	struct sample sample;
244 	int32_t	min_perf_ratio;
245 	int32_t	max_perf_ratio;
246 #ifdef CONFIG_ACPI
247 	struct acpi_processor_performance acpi_perf_data;
248 	bool valid_pss_table;
249 #endif
250 	unsigned int iowait_boost;
251 	s16 epp_powersave;
252 	s16 epp_policy;
253 	s16 epp_default;
254 	s16 epp_cached;
255 	u64 hwp_req_cached;
256 	u64 hwp_cap_cached;
257 	u64 last_io_update;
258 	unsigned int capacity_perf;
259 	unsigned int sched_flags;
260 	u32 hwp_boost_min;
261 	bool suspended;
262 	struct delayed_work hwp_notify_work;
263 };
264 
265 static struct cpudata **all_cpu_data;
266 
267 /**
268  * struct pstate_funcs - Per CPU model specific callbacks
269  * @get_max:		Callback to get maximum non turbo effective P state
270  * @get_max_physical:	Callback to get maximum non turbo physical P state
271  * @get_min:		Callback to get minimum P state
272  * @get_turbo:		Callback to get turbo P state
273  * @get_scaling:	Callback to get frequency scaling factor
274  * @get_cpu_scaling:	Get frequency scaling factor for a given cpu
275  * @get_aperf_mperf_shift: Callback to get the APERF vs MPERF frequency difference
276  * @get_val:		Callback to convert P state to actual MSR write value
277  * @get_vid:		Callback to get VID data for Atom platforms
278  *
279  * Core and Atom CPU models have different way to get P State limits. This
280  * structure is used to store those callbacks.
281  */
282 struct pstate_funcs {
283 	int (*get_max)(int cpu);
284 	int (*get_max_physical)(int cpu);
285 	int (*get_min)(int cpu);
286 	int (*get_turbo)(int cpu);
287 	int (*get_scaling)(void);
288 	int (*get_cpu_scaling)(int cpu);
289 	int (*get_aperf_mperf_shift)(void);
290 	u64 (*get_val)(struct cpudata*, int pstate);
291 	void (*get_vid)(struct cpudata *);
292 };
293 
294 static struct pstate_funcs pstate_funcs __read_mostly;
295 
296 static bool hwp_active __ro_after_init;
297 static int hwp_mode_bdw __ro_after_init;
298 static bool per_cpu_limits __ro_after_init;
299 static bool hwp_forced __ro_after_init;
300 static bool hwp_boost __read_mostly;
301 static bool hwp_is_hybrid;
302 
303 static struct cpufreq_driver *intel_pstate_driver __read_mostly;
304 
305 #define HYBRID_SCALING_FACTOR		78741
306 #define HYBRID_SCALING_FACTOR_MTL	80000
307 #define HYBRID_SCALING_FACTOR_LNL	86957
308 
309 static int hybrid_scaling_factor = HYBRID_SCALING_FACTOR;
310 
311 static inline int core_get_scaling(void)
312 {
313 	return 100000;
314 }
315 
316 #ifdef CONFIG_ACPI
317 static bool acpi_ppc;
318 #endif
319 
320 static struct global_params global;
321 
322 static DEFINE_MUTEX(intel_pstate_driver_lock);
323 static DEFINE_MUTEX(intel_pstate_limits_lock);
324 
325 #ifdef CONFIG_ACPI
326 
327 static bool intel_pstate_acpi_pm_profile_server(void)
328 {
329 	if (acpi_gbl_FADT.preferred_profile == PM_ENTERPRISE_SERVER ||
330 	    acpi_gbl_FADT.preferred_profile == PM_PERFORMANCE_SERVER)
331 		return true;
332 
333 	return false;
334 }
335 
336 static bool intel_pstate_get_ppc_enable_status(void)
337 {
338 	if (intel_pstate_acpi_pm_profile_server())
339 		return true;
340 
341 	return acpi_ppc;
342 }
343 
344 #ifdef CONFIG_ACPI_CPPC_LIB
345 
346 /* The work item is needed to avoid CPU hotplug locking issues */
347 static void intel_pstste_sched_itmt_work_fn(struct work_struct *work)
348 {
349 	sched_set_itmt_support();
350 }
351 
352 static DECLARE_WORK(sched_itmt_work, intel_pstste_sched_itmt_work_fn);
353 
354 #define CPPC_MAX_PERF	U8_MAX
355 
356 static void intel_pstate_set_itmt_prio(int cpu)
357 {
358 	struct cppc_perf_caps cppc_perf;
359 	static u32 max_highest_perf = 0, min_highest_perf = U32_MAX;
360 	int ret;
361 
362 	ret = cppc_get_perf_caps(cpu, &cppc_perf);
363 	/*
364 	 * If CPPC is not available, fall back to MSR_HWP_CAPABILITIES bits [8:0].
365 	 *
366 	 * Also, on some systems with overclocking enabled, CPPC.highest_perf is
367 	 * hardcoded to 0xff, so CPPC.highest_perf cannot be used to enable ITMT.
368 	 * Fall back to MSR_HWP_CAPABILITIES then too.
369 	 */
370 	if (ret || cppc_perf.highest_perf == CPPC_MAX_PERF)
371 		cppc_perf.highest_perf = HWP_HIGHEST_PERF(READ_ONCE(all_cpu_data[cpu]->hwp_cap_cached));
372 
373 	/*
374 	 * The priorities can be set regardless of whether or not
375 	 * sched_set_itmt_support(true) has been called and it is valid to
376 	 * update them at any time after it has been called.
377 	 */
378 	sched_set_itmt_core_prio(cppc_perf.highest_perf, cpu);
379 
380 	if (max_highest_perf <= min_highest_perf) {
381 		if (cppc_perf.highest_perf > max_highest_perf)
382 			max_highest_perf = cppc_perf.highest_perf;
383 
384 		if (cppc_perf.highest_perf < min_highest_perf)
385 			min_highest_perf = cppc_perf.highest_perf;
386 
387 		if (max_highest_perf > min_highest_perf) {
388 			/*
389 			 * This code can be run during CPU online under the
390 			 * CPU hotplug locks, so sched_set_itmt_support()
391 			 * cannot be called from here.  Queue up a work item
392 			 * to invoke it.
393 			 */
394 			schedule_work(&sched_itmt_work);
395 		}
396 	}
397 }
398 
399 static int intel_pstate_get_cppc_guaranteed(int cpu)
400 {
401 	struct cppc_perf_caps cppc_perf;
402 	int ret;
403 
404 	ret = cppc_get_perf_caps(cpu, &cppc_perf);
405 	if (ret)
406 		return ret;
407 
408 	if (cppc_perf.guaranteed_perf)
409 		return cppc_perf.guaranteed_perf;
410 
411 	return cppc_perf.nominal_perf;
412 }
413 
414 static int intel_pstate_cppc_get_scaling(int cpu)
415 {
416 	struct cppc_perf_caps cppc_perf;
417 	int ret;
418 
419 	ret = cppc_get_perf_caps(cpu, &cppc_perf);
420 
421 	/*
422 	 * If the nominal frequency and the nominal performance are not
423 	 * zero and the ratio between them is not 100, return the hybrid
424 	 * scaling factor.
425 	 */
426 	if (!ret && cppc_perf.nominal_perf && cppc_perf.nominal_freq &&
427 	    cppc_perf.nominal_perf * 100 != cppc_perf.nominal_freq)
428 		return hybrid_scaling_factor;
429 
430 	return core_get_scaling();
431 }
432 
433 #else /* CONFIG_ACPI_CPPC_LIB */
434 static inline void intel_pstate_set_itmt_prio(int cpu)
435 {
436 }
437 #endif /* CONFIG_ACPI_CPPC_LIB */
438 
439 static void intel_pstate_init_acpi_perf_limits(struct cpufreq_policy *policy)
440 {
441 	struct cpudata *cpu;
442 	int ret;
443 	int i;
444 
445 	if (hwp_active) {
446 		intel_pstate_set_itmt_prio(policy->cpu);
447 		return;
448 	}
449 
450 	if (!intel_pstate_get_ppc_enable_status())
451 		return;
452 
453 	cpu = all_cpu_data[policy->cpu];
454 
455 	ret = acpi_processor_register_performance(&cpu->acpi_perf_data,
456 						  policy->cpu);
457 	if (ret)
458 		return;
459 
460 	/*
461 	 * Check if the control value in _PSS is for PERF_CTL MSR, which should
462 	 * guarantee that the states returned by it map to the states in our
463 	 * list directly.
464 	 */
465 	if (cpu->acpi_perf_data.control_register.space_id !=
466 						ACPI_ADR_SPACE_FIXED_HARDWARE)
467 		goto err;
468 
469 	/*
470 	 * If there is only one entry _PSS, simply ignore _PSS and continue as
471 	 * usual without taking _PSS into account
472 	 */
473 	if (cpu->acpi_perf_data.state_count < 2)
474 		goto err;
475 
476 	pr_debug("CPU%u - ACPI _PSS perf data\n", policy->cpu);
477 	for (i = 0; i < cpu->acpi_perf_data.state_count; i++) {
478 		pr_debug("     %cP%d: %u MHz, %u mW, 0x%x\n",
479 			 (i == cpu->acpi_perf_data.state ? '*' : ' '), i,
480 			 (u32) cpu->acpi_perf_data.states[i].core_frequency,
481 			 (u32) cpu->acpi_perf_data.states[i].power,
482 			 (u32) cpu->acpi_perf_data.states[i].control);
483 	}
484 
485 	cpu->valid_pss_table = true;
486 	pr_debug("_PPC limits will be enforced\n");
487 
488 	return;
489 
490  err:
491 	cpu->valid_pss_table = false;
492 	acpi_processor_unregister_performance(policy->cpu);
493 }
494 
495 static void intel_pstate_exit_perf_limits(struct cpufreq_policy *policy)
496 {
497 	struct cpudata *cpu;
498 
499 	cpu = all_cpu_data[policy->cpu];
500 	if (!cpu->valid_pss_table)
501 		return;
502 
503 	acpi_processor_unregister_performance(policy->cpu);
504 }
505 #else /* CONFIG_ACPI */
506 static inline void intel_pstate_init_acpi_perf_limits(struct cpufreq_policy *policy)
507 {
508 }
509 
510 static inline void intel_pstate_exit_perf_limits(struct cpufreq_policy *policy)
511 {
512 }
513 
514 static inline bool intel_pstate_acpi_pm_profile_server(void)
515 {
516 	return false;
517 }
518 #endif /* CONFIG_ACPI */
519 
520 #ifndef CONFIG_ACPI_CPPC_LIB
521 static inline int intel_pstate_get_cppc_guaranteed(int cpu)
522 {
523 	return -ENOTSUPP;
524 }
525 
526 static int intel_pstate_cppc_get_scaling(int cpu)
527 {
528 	return core_get_scaling();
529 }
530 #endif /* CONFIG_ACPI_CPPC_LIB */
531 
532 static int intel_pstate_freq_to_hwp_rel(struct cpudata *cpu, int freq,
533 					unsigned int relation)
534 {
535 	if (freq == cpu->pstate.turbo_freq)
536 		return cpu->pstate.turbo_pstate;
537 
538 	if (freq == cpu->pstate.max_freq)
539 		return cpu->pstate.max_pstate;
540 
541 	switch (relation) {
542 	case CPUFREQ_RELATION_H:
543 		return freq / cpu->pstate.scaling;
544 	case CPUFREQ_RELATION_C:
545 		return DIV_ROUND_CLOSEST(freq, cpu->pstate.scaling);
546 	}
547 
548 	return DIV_ROUND_UP(freq, cpu->pstate.scaling);
549 }
550 
551 static int intel_pstate_freq_to_hwp(struct cpudata *cpu, int freq)
552 {
553 	return intel_pstate_freq_to_hwp_rel(cpu, freq, CPUFREQ_RELATION_L);
554 }
555 
556 /**
557  * intel_pstate_hybrid_hwp_adjust - Calibrate HWP performance levels.
558  * @cpu: Target CPU.
559  *
560  * On hybrid processors, HWP may expose more performance levels than there are
561  * P-states accessible through the PERF_CTL interface.  If that happens, the
562  * scaling factor between HWP performance levels and CPU frequency will be less
563  * than the scaling factor between P-state values and CPU frequency.
564  *
565  * In that case, adjust the CPU parameters used in computations accordingly.
566  */
567 static void intel_pstate_hybrid_hwp_adjust(struct cpudata *cpu)
568 {
569 	int perf_ctl_max_phys = cpu->pstate.max_pstate_physical;
570 	int perf_ctl_scaling = cpu->pstate.perf_ctl_scaling;
571 	int perf_ctl_turbo = pstate_funcs.get_turbo(cpu->cpu);
572 	int scaling = cpu->pstate.scaling;
573 	int freq;
574 
575 	pr_debug("CPU%d: perf_ctl_max_phys = %d\n", cpu->cpu, perf_ctl_max_phys);
576 	pr_debug("CPU%d: perf_ctl_turbo = %d\n", cpu->cpu, perf_ctl_turbo);
577 	pr_debug("CPU%d: perf_ctl_scaling = %d\n", cpu->cpu, perf_ctl_scaling);
578 	pr_debug("CPU%d: HWP_CAP guaranteed = %d\n", cpu->cpu, cpu->pstate.max_pstate);
579 	pr_debug("CPU%d: HWP_CAP highest = %d\n", cpu->cpu, cpu->pstate.turbo_pstate);
580 	pr_debug("CPU%d: HWP-to-frequency scaling factor: %d\n", cpu->cpu, scaling);
581 
582 	cpu->pstate.turbo_freq = rounddown(cpu->pstate.turbo_pstate * scaling,
583 					   perf_ctl_scaling);
584 	cpu->pstate.max_freq = rounddown(cpu->pstate.max_pstate * scaling,
585 					 perf_ctl_scaling);
586 
587 	freq = perf_ctl_max_phys * perf_ctl_scaling;
588 	cpu->pstate.max_pstate_physical = intel_pstate_freq_to_hwp(cpu, freq);
589 
590 	freq = cpu->pstate.min_pstate * perf_ctl_scaling;
591 	cpu->pstate.min_freq = freq;
592 	/*
593 	 * Cast the min P-state value retrieved via pstate_funcs.get_min() to
594 	 * the effective range of HWP performance levels.
595 	 */
596 	cpu->pstate.min_pstate = intel_pstate_freq_to_hwp(cpu, freq);
597 }
598 
599 static bool turbo_is_disabled(void)
600 {
601 	u64 misc_en;
602 
603 	rdmsrl(MSR_IA32_MISC_ENABLE, misc_en);
604 
605 	return !!(misc_en & MSR_IA32_MISC_ENABLE_TURBO_DISABLE);
606 }
607 
608 static int min_perf_pct_min(void)
609 {
610 	struct cpudata *cpu = all_cpu_data[0];
611 	int turbo_pstate = cpu->pstate.turbo_pstate;
612 
613 	return turbo_pstate ?
614 		(cpu->pstate.min_pstate * 100 / turbo_pstate) : 0;
615 }
616 
617 static s16 intel_pstate_get_epb(struct cpudata *cpu_data)
618 {
619 	u64 epb;
620 	int ret;
621 
622 	if (!boot_cpu_has(X86_FEATURE_EPB))
623 		return -ENXIO;
624 
625 	ret = rdmsrl_on_cpu(cpu_data->cpu, MSR_IA32_ENERGY_PERF_BIAS, &epb);
626 	if (ret)
627 		return (s16)ret;
628 
629 	return (s16)(epb & 0x0f);
630 }
631 
632 static s16 intel_pstate_get_epp(struct cpudata *cpu_data, u64 hwp_req_data)
633 {
634 	s16 epp;
635 
636 	if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
637 		/*
638 		 * When hwp_req_data is 0, means that caller didn't read
639 		 * MSR_HWP_REQUEST, so need to read and get EPP.
640 		 */
641 		if (!hwp_req_data) {
642 			epp = rdmsrl_on_cpu(cpu_data->cpu, MSR_HWP_REQUEST,
643 					    &hwp_req_data);
644 			if (epp)
645 				return epp;
646 		}
647 		epp = (hwp_req_data >> 24) & 0xff;
648 	} else {
649 		/* When there is no EPP present, HWP uses EPB settings */
650 		epp = intel_pstate_get_epb(cpu_data);
651 	}
652 
653 	return epp;
654 }
655 
656 static int intel_pstate_set_epb(int cpu, s16 pref)
657 {
658 	u64 epb;
659 	int ret;
660 
661 	if (!boot_cpu_has(X86_FEATURE_EPB))
662 		return -ENXIO;
663 
664 	ret = rdmsrl_on_cpu(cpu, MSR_IA32_ENERGY_PERF_BIAS, &epb);
665 	if (ret)
666 		return ret;
667 
668 	epb = (epb & ~0x0f) | pref;
669 	wrmsrl_on_cpu(cpu, MSR_IA32_ENERGY_PERF_BIAS, epb);
670 
671 	return 0;
672 }
673 
674 /*
675  * EPP/EPB display strings corresponding to EPP index in the
676  * energy_perf_strings[]
677  *	index		String
678  *-------------------------------------
679  *	0		default
680  *	1		performance
681  *	2		balance_performance
682  *	3		balance_power
683  *	4		power
684  */
685 
686 enum energy_perf_value_index {
687 	EPP_INDEX_DEFAULT = 0,
688 	EPP_INDEX_PERFORMANCE,
689 	EPP_INDEX_BALANCE_PERFORMANCE,
690 	EPP_INDEX_BALANCE_POWERSAVE,
691 	EPP_INDEX_POWERSAVE,
692 };
693 
694 static const char * const energy_perf_strings[] = {
695 	[EPP_INDEX_DEFAULT] = "default",
696 	[EPP_INDEX_PERFORMANCE] = "performance",
697 	[EPP_INDEX_BALANCE_PERFORMANCE] = "balance_performance",
698 	[EPP_INDEX_BALANCE_POWERSAVE] = "balance_power",
699 	[EPP_INDEX_POWERSAVE] = "power",
700 	NULL
701 };
702 static unsigned int epp_values[] = {
703 	[EPP_INDEX_DEFAULT] = 0, /* Unused index */
704 	[EPP_INDEX_PERFORMANCE] = HWP_EPP_PERFORMANCE,
705 	[EPP_INDEX_BALANCE_PERFORMANCE] = HWP_EPP_BALANCE_PERFORMANCE,
706 	[EPP_INDEX_BALANCE_POWERSAVE] = HWP_EPP_BALANCE_POWERSAVE,
707 	[EPP_INDEX_POWERSAVE] = HWP_EPP_POWERSAVE,
708 };
709 
710 static int intel_pstate_get_energy_pref_index(struct cpudata *cpu_data, int *raw_epp)
711 {
712 	s16 epp;
713 	int index = -EINVAL;
714 
715 	*raw_epp = 0;
716 	epp = intel_pstate_get_epp(cpu_data, 0);
717 	if (epp < 0)
718 		return epp;
719 
720 	if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
721 		if (epp == epp_values[EPP_INDEX_PERFORMANCE])
722 			return EPP_INDEX_PERFORMANCE;
723 		if (epp == epp_values[EPP_INDEX_BALANCE_PERFORMANCE])
724 			return EPP_INDEX_BALANCE_PERFORMANCE;
725 		if (epp == epp_values[EPP_INDEX_BALANCE_POWERSAVE])
726 			return EPP_INDEX_BALANCE_POWERSAVE;
727 		if (epp == epp_values[EPP_INDEX_POWERSAVE])
728 			return EPP_INDEX_POWERSAVE;
729 		*raw_epp = epp;
730 		return 0;
731 	} else if (boot_cpu_has(X86_FEATURE_EPB)) {
732 		/*
733 		 * Range:
734 		 *	0x00-0x03	:	Performance
735 		 *	0x04-0x07	:	Balance performance
736 		 *	0x08-0x0B	:	Balance power
737 		 *	0x0C-0x0F	:	Power
738 		 * The EPB is a 4 bit value, but our ranges restrict the
739 		 * value which can be set. Here only using top two bits
740 		 * effectively.
741 		 */
742 		index = (epp >> 2) + 1;
743 	}
744 
745 	return index;
746 }
747 
748 static int intel_pstate_set_epp(struct cpudata *cpu, u32 epp)
749 {
750 	int ret;
751 
752 	/*
753 	 * Use the cached HWP Request MSR value, because in the active mode the
754 	 * register itself may be updated by intel_pstate_hwp_boost_up() or
755 	 * intel_pstate_hwp_boost_down() at any time.
756 	 */
757 	u64 value = READ_ONCE(cpu->hwp_req_cached);
758 
759 	value &= ~GENMASK_ULL(31, 24);
760 	value |= (u64)epp << 24;
761 	/*
762 	 * The only other updater of hwp_req_cached in the active mode,
763 	 * intel_pstate_hwp_set(), is called under the same lock as this
764 	 * function, so it cannot run in parallel with the update below.
765 	 */
766 	WRITE_ONCE(cpu->hwp_req_cached, value);
767 	ret = wrmsrl_on_cpu(cpu->cpu, MSR_HWP_REQUEST, value);
768 	if (!ret)
769 		cpu->epp_cached = epp;
770 
771 	return ret;
772 }
773 
774 static int intel_pstate_set_energy_pref_index(struct cpudata *cpu_data,
775 					      int pref_index, bool use_raw,
776 					      u32 raw_epp)
777 {
778 	int epp = -EINVAL;
779 	int ret;
780 
781 	if (!pref_index)
782 		epp = cpu_data->epp_default;
783 
784 	if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
785 		if (use_raw)
786 			epp = raw_epp;
787 		else if (epp == -EINVAL)
788 			epp = epp_values[pref_index];
789 
790 		/*
791 		 * To avoid confusion, refuse to set EPP to any values different
792 		 * from 0 (performance) if the current policy is "performance",
793 		 * because those values would be overridden.
794 		 */
795 		if (epp > 0 && cpu_data->policy == CPUFREQ_POLICY_PERFORMANCE)
796 			return -EBUSY;
797 
798 		ret = intel_pstate_set_epp(cpu_data, epp);
799 	} else {
800 		if (epp == -EINVAL)
801 			epp = (pref_index - 1) << 2;
802 		ret = intel_pstate_set_epb(cpu_data->cpu, epp);
803 	}
804 
805 	return ret;
806 }
807 
808 static ssize_t show_energy_performance_available_preferences(
809 				struct cpufreq_policy *policy, char *buf)
810 {
811 	int i = 0;
812 	int ret = 0;
813 
814 	while (energy_perf_strings[i] != NULL)
815 		ret += sprintf(&buf[ret], "%s ", energy_perf_strings[i++]);
816 
817 	ret += sprintf(&buf[ret], "\n");
818 
819 	return ret;
820 }
821 
822 cpufreq_freq_attr_ro(energy_performance_available_preferences);
823 
824 static struct cpufreq_driver intel_pstate;
825 
826 static ssize_t store_energy_performance_preference(
827 		struct cpufreq_policy *policy, const char *buf, size_t count)
828 {
829 	struct cpudata *cpu = all_cpu_data[policy->cpu];
830 	char str_preference[21];
831 	bool raw = false;
832 	ssize_t ret;
833 	u32 epp = 0;
834 
835 	ret = sscanf(buf, "%20s", str_preference);
836 	if (ret != 1)
837 		return -EINVAL;
838 
839 	ret = match_string(energy_perf_strings, -1, str_preference);
840 	if (ret < 0) {
841 		if (!boot_cpu_has(X86_FEATURE_HWP_EPP))
842 			return ret;
843 
844 		ret = kstrtouint(buf, 10, &epp);
845 		if (ret)
846 			return ret;
847 
848 		if (epp > 255)
849 			return -EINVAL;
850 
851 		raw = true;
852 	}
853 
854 	/*
855 	 * This function runs with the policy R/W semaphore held, which
856 	 * guarantees that the driver pointer will not change while it is
857 	 * running.
858 	 */
859 	if (!intel_pstate_driver)
860 		return -EAGAIN;
861 
862 	mutex_lock(&intel_pstate_limits_lock);
863 
864 	if (intel_pstate_driver == &intel_pstate) {
865 		ret = intel_pstate_set_energy_pref_index(cpu, ret, raw, epp);
866 	} else {
867 		/*
868 		 * In the passive mode the governor needs to be stopped on the
869 		 * target CPU before the EPP update and restarted after it,
870 		 * which is super-heavy-weight, so make sure it is worth doing
871 		 * upfront.
872 		 */
873 		if (!raw)
874 			epp = ret ? epp_values[ret] : cpu->epp_default;
875 
876 		if (cpu->epp_cached != epp) {
877 			int err;
878 
879 			cpufreq_stop_governor(policy);
880 			ret = intel_pstate_set_epp(cpu, epp);
881 			err = cpufreq_start_governor(policy);
882 			if (!ret)
883 				ret = err;
884 		} else {
885 			ret = 0;
886 		}
887 	}
888 
889 	mutex_unlock(&intel_pstate_limits_lock);
890 
891 	return ret ?: count;
892 }
893 
894 static ssize_t show_energy_performance_preference(
895 				struct cpufreq_policy *policy, char *buf)
896 {
897 	struct cpudata *cpu_data = all_cpu_data[policy->cpu];
898 	int preference, raw_epp;
899 
900 	preference = intel_pstate_get_energy_pref_index(cpu_data, &raw_epp);
901 	if (preference < 0)
902 		return preference;
903 
904 	if (raw_epp)
905 		return  sprintf(buf, "%d\n", raw_epp);
906 	else
907 		return  sprintf(buf, "%s\n", energy_perf_strings[preference]);
908 }
909 
910 cpufreq_freq_attr_rw(energy_performance_preference);
911 
912 static ssize_t show_base_frequency(struct cpufreq_policy *policy, char *buf)
913 {
914 	struct cpudata *cpu = all_cpu_data[policy->cpu];
915 	int ratio, freq;
916 
917 	ratio = intel_pstate_get_cppc_guaranteed(policy->cpu);
918 	if (ratio <= 0) {
919 		u64 cap;
920 
921 		rdmsrl_on_cpu(policy->cpu, MSR_HWP_CAPABILITIES, &cap);
922 		ratio = HWP_GUARANTEED_PERF(cap);
923 	}
924 
925 	freq = ratio * cpu->pstate.scaling;
926 	if (cpu->pstate.scaling != cpu->pstate.perf_ctl_scaling)
927 		freq = rounddown(freq, cpu->pstate.perf_ctl_scaling);
928 
929 	return sprintf(buf, "%d\n", freq);
930 }
931 
932 cpufreq_freq_attr_ro(base_frequency);
933 
934 static struct freq_attr *hwp_cpufreq_attrs[] = {
935 	&energy_performance_preference,
936 	&energy_performance_available_preferences,
937 	&base_frequency,
938 	NULL,
939 };
940 
941 static struct cpudata *hybrid_max_perf_cpu __read_mostly;
942 /*
943  * Protects hybrid_max_perf_cpu, the capacity_perf fields in struct cpudata,
944  * and the x86 arch scale-invariance information from concurrent updates.
945  */
946 static DEFINE_MUTEX(hybrid_capacity_lock);
947 
948 static void hybrid_set_cpu_capacity(struct cpudata *cpu)
949 {
950 	arch_set_cpu_capacity(cpu->cpu, cpu->capacity_perf,
951 			      hybrid_max_perf_cpu->capacity_perf,
952 			      cpu->capacity_perf,
953 			      cpu->pstate.max_pstate_physical);
954 
955 	pr_debug("CPU%d: perf = %u, max. perf = %u, base perf = %d\n", cpu->cpu,
956 		 cpu->capacity_perf, hybrid_max_perf_cpu->capacity_perf,
957 		 cpu->pstate.max_pstate_physical);
958 }
959 
960 static void hybrid_clear_cpu_capacity(unsigned int cpunum)
961 {
962 	arch_set_cpu_capacity(cpunum, 1, 1, 1, 1);
963 }
964 
965 static void hybrid_get_capacity_perf(struct cpudata *cpu)
966 {
967 	if (READ_ONCE(global.no_turbo)) {
968 		cpu->capacity_perf = cpu->pstate.max_pstate_physical;
969 		return;
970 	}
971 
972 	cpu->capacity_perf = HWP_HIGHEST_PERF(READ_ONCE(cpu->hwp_cap_cached));
973 }
974 
975 static void hybrid_set_capacity_of_cpus(void)
976 {
977 	int cpunum;
978 
979 	for_each_online_cpu(cpunum) {
980 		struct cpudata *cpu = all_cpu_data[cpunum];
981 
982 		if (cpu)
983 			hybrid_set_cpu_capacity(cpu);
984 	}
985 }
986 
987 static void hybrid_update_cpu_capacity_scaling(void)
988 {
989 	struct cpudata *max_perf_cpu = NULL;
990 	unsigned int max_cap_perf = 0;
991 	int cpunum;
992 
993 	for_each_online_cpu(cpunum) {
994 		struct cpudata *cpu = all_cpu_data[cpunum];
995 
996 		if (!cpu)
997 			continue;
998 
999 		/*
1000 		 * During initialization, CPU performance at full capacity needs
1001 		 * to be determined.
1002 		 */
1003 		if (!hybrid_max_perf_cpu)
1004 			hybrid_get_capacity_perf(cpu);
1005 
1006 		/*
1007 		 * If hybrid_max_perf_cpu is not NULL at this point, it is
1008 		 * being replaced, so don't take it into account when looking
1009 		 * for the new one.
1010 		 */
1011 		if (cpu == hybrid_max_perf_cpu)
1012 			continue;
1013 
1014 		if (cpu->capacity_perf > max_cap_perf) {
1015 			max_cap_perf = cpu->capacity_perf;
1016 			max_perf_cpu = cpu;
1017 		}
1018 	}
1019 
1020 	if (max_perf_cpu) {
1021 		hybrid_max_perf_cpu = max_perf_cpu;
1022 		hybrid_set_capacity_of_cpus();
1023 	} else {
1024 		pr_info("Found no CPUs with nonzero maximum performance\n");
1025 		/* Revert to the flat CPU capacity structure. */
1026 		for_each_online_cpu(cpunum)
1027 			hybrid_clear_cpu_capacity(cpunum);
1028 	}
1029 }
1030 
1031 static void __hybrid_refresh_cpu_capacity_scaling(void)
1032 {
1033 	hybrid_max_perf_cpu = NULL;
1034 	hybrid_update_cpu_capacity_scaling();
1035 }
1036 
1037 static void hybrid_refresh_cpu_capacity_scaling(void)
1038 {
1039 	guard(mutex)(&hybrid_capacity_lock);
1040 
1041 	__hybrid_refresh_cpu_capacity_scaling();
1042 }
1043 
1044 static void hybrid_init_cpu_capacity_scaling(bool refresh)
1045 {
1046 	/*
1047 	 * If hybrid_max_perf_cpu is set at this point, the hybrid CPU capacity
1048 	 * scaling has been enabled already and the driver is just changing the
1049 	 * operation mode.
1050 	 */
1051 	if (refresh) {
1052 		hybrid_refresh_cpu_capacity_scaling();
1053 		return;
1054 	}
1055 
1056 	/*
1057 	 * On hybrid systems, use asym capacity instead of ITMT, but because
1058 	 * the capacity of SMT threads is not deterministic even approximately,
1059 	 * do not do that when SMT is in use.
1060 	 */
1061 	if (hwp_is_hybrid && !sched_smt_active() && arch_enable_hybrid_capacity_scale()) {
1062 		hybrid_refresh_cpu_capacity_scaling();
1063 		/*
1064 		 * Disabling ITMT causes sched domains to be rebuilt to disable asym
1065 		 * packing and enable asym capacity.
1066 		 */
1067 		sched_clear_itmt_support();
1068 	}
1069 }
1070 
1071 static bool hybrid_clear_max_perf_cpu(void)
1072 {
1073 	bool ret;
1074 
1075 	guard(mutex)(&hybrid_capacity_lock);
1076 
1077 	ret = !!hybrid_max_perf_cpu;
1078 	hybrid_max_perf_cpu = NULL;
1079 
1080 	return ret;
1081 }
1082 
1083 static void __intel_pstate_get_hwp_cap(struct cpudata *cpu)
1084 {
1085 	u64 cap;
1086 
1087 	rdmsrl_on_cpu(cpu->cpu, MSR_HWP_CAPABILITIES, &cap);
1088 	WRITE_ONCE(cpu->hwp_cap_cached, cap);
1089 	cpu->pstate.max_pstate = HWP_GUARANTEED_PERF(cap);
1090 	cpu->pstate.turbo_pstate = HWP_HIGHEST_PERF(cap);
1091 }
1092 
1093 static void intel_pstate_get_hwp_cap(struct cpudata *cpu)
1094 {
1095 	int scaling = cpu->pstate.scaling;
1096 
1097 	__intel_pstate_get_hwp_cap(cpu);
1098 
1099 	cpu->pstate.max_freq = cpu->pstate.max_pstate * scaling;
1100 	cpu->pstate.turbo_freq = cpu->pstate.turbo_pstate * scaling;
1101 	if (scaling != cpu->pstate.perf_ctl_scaling) {
1102 		int perf_ctl_scaling = cpu->pstate.perf_ctl_scaling;
1103 
1104 		cpu->pstate.max_freq = rounddown(cpu->pstate.max_freq,
1105 						 perf_ctl_scaling);
1106 		cpu->pstate.turbo_freq = rounddown(cpu->pstate.turbo_freq,
1107 						   perf_ctl_scaling);
1108 	}
1109 }
1110 
1111 static void hybrid_update_capacity(struct cpudata *cpu)
1112 {
1113 	unsigned int max_cap_perf;
1114 
1115 	mutex_lock(&hybrid_capacity_lock);
1116 
1117 	if (!hybrid_max_perf_cpu)
1118 		goto unlock;
1119 
1120 	/*
1121 	 * The maximum performance of the CPU may have changed, but assume
1122 	 * that the performance of the other CPUs has not changed.
1123 	 */
1124 	max_cap_perf = hybrid_max_perf_cpu->capacity_perf;
1125 
1126 	intel_pstate_get_hwp_cap(cpu);
1127 
1128 	hybrid_get_capacity_perf(cpu);
1129 	/* Should hybrid_max_perf_cpu be replaced by this CPU? */
1130 	if (cpu->capacity_perf > max_cap_perf) {
1131 		hybrid_max_perf_cpu = cpu;
1132 		hybrid_set_capacity_of_cpus();
1133 		goto unlock;
1134 	}
1135 
1136 	/* If this CPU is hybrid_max_perf_cpu, should it be replaced? */
1137 	if (cpu == hybrid_max_perf_cpu && cpu->capacity_perf < max_cap_perf) {
1138 		hybrid_update_cpu_capacity_scaling();
1139 		goto unlock;
1140 	}
1141 
1142 	hybrid_set_cpu_capacity(cpu);
1143 
1144 unlock:
1145 	mutex_unlock(&hybrid_capacity_lock);
1146 }
1147 
1148 static void intel_pstate_hwp_set(unsigned int cpu)
1149 {
1150 	struct cpudata *cpu_data = all_cpu_data[cpu];
1151 	int max, min;
1152 	u64 value;
1153 	s16 epp;
1154 
1155 	max = cpu_data->max_perf_ratio;
1156 	min = cpu_data->min_perf_ratio;
1157 
1158 	if (cpu_data->policy == CPUFREQ_POLICY_PERFORMANCE)
1159 		min = max;
1160 
1161 	rdmsrl_on_cpu(cpu, MSR_HWP_REQUEST, &value);
1162 
1163 	value &= ~HWP_MIN_PERF(~0L);
1164 	value |= HWP_MIN_PERF(min);
1165 
1166 	value &= ~HWP_MAX_PERF(~0L);
1167 	value |= HWP_MAX_PERF(max);
1168 
1169 	if (cpu_data->epp_policy == cpu_data->policy)
1170 		goto skip_epp;
1171 
1172 	cpu_data->epp_policy = cpu_data->policy;
1173 
1174 	if (cpu_data->policy == CPUFREQ_POLICY_PERFORMANCE) {
1175 		epp = intel_pstate_get_epp(cpu_data, value);
1176 		cpu_data->epp_powersave = epp;
1177 		/* If EPP read was failed, then don't try to write */
1178 		if (epp < 0)
1179 			goto skip_epp;
1180 
1181 		epp = 0;
1182 	} else {
1183 		/* skip setting EPP, when saved value is invalid */
1184 		if (cpu_data->epp_powersave < 0)
1185 			goto skip_epp;
1186 
1187 		/*
1188 		 * No need to restore EPP when it is not zero. This
1189 		 * means:
1190 		 *  - Policy is not changed
1191 		 *  - user has manually changed
1192 		 *  - Error reading EPB
1193 		 */
1194 		epp = intel_pstate_get_epp(cpu_data, value);
1195 		if (epp)
1196 			goto skip_epp;
1197 
1198 		epp = cpu_data->epp_powersave;
1199 	}
1200 	if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
1201 		value &= ~GENMASK_ULL(31, 24);
1202 		value |= (u64)epp << 24;
1203 	} else {
1204 		intel_pstate_set_epb(cpu, epp);
1205 	}
1206 skip_epp:
1207 	WRITE_ONCE(cpu_data->hwp_req_cached, value);
1208 	wrmsrl_on_cpu(cpu, MSR_HWP_REQUEST, value);
1209 }
1210 
1211 static void intel_pstate_disable_hwp_interrupt(struct cpudata *cpudata);
1212 
1213 static void intel_pstate_hwp_offline(struct cpudata *cpu)
1214 {
1215 	u64 value = READ_ONCE(cpu->hwp_req_cached);
1216 	int min_perf;
1217 
1218 	intel_pstate_disable_hwp_interrupt(cpu);
1219 
1220 	if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
1221 		/*
1222 		 * In case the EPP has been set to "performance" by the
1223 		 * active mode "performance" scaling algorithm, replace that
1224 		 * temporary value with the cached EPP one.
1225 		 */
1226 		value &= ~GENMASK_ULL(31, 24);
1227 		value |= HWP_ENERGY_PERF_PREFERENCE(cpu->epp_cached);
1228 		/*
1229 		 * However, make sure that EPP will be set to "performance" when
1230 		 * the CPU is brought back online again and the "performance"
1231 		 * scaling algorithm is still in effect.
1232 		 */
1233 		cpu->epp_policy = CPUFREQ_POLICY_UNKNOWN;
1234 	}
1235 
1236 	/*
1237 	 * Clear the desired perf field in the cached HWP request value to
1238 	 * prevent nonzero desired values from being leaked into the active
1239 	 * mode.
1240 	 */
1241 	value &= ~HWP_DESIRED_PERF(~0L);
1242 	WRITE_ONCE(cpu->hwp_req_cached, value);
1243 
1244 	value &= ~GENMASK_ULL(31, 0);
1245 	min_perf = HWP_LOWEST_PERF(READ_ONCE(cpu->hwp_cap_cached));
1246 
1247 	/* Set hwp_max = hwp_min */
1248 	value |= HWP_MAX_PERF(min_perf);
1249 	value |= HWP_MIN_PERF(min_perf);
1250 
1251 	/* Set EPP to min */
1252 	if (boot_cpu_has(X86_FEATURE_HWP_EPP))
1253 		value |= HWP_ENERGY_PERF_PREFERENCE(HWP_EPP_POWERSAVE);
1254 
1255 	wrmsrl_on_cpu(cpu->cpu, MSR_HWP_REQUEST, value);
1256 
1257 	mutex_lock(&hybrid_capacity_lock);
1258 
1259 	if (!hybrid_max_perf_cpu) {
1260 		mutex_unlock(&hybrid_capacity_lock);
1261 
1262 		return;
1263 	}
1264 
1265 	if (hybrid_max_perf_cpu == cpu)
1266 		hybrid_update_cpu_capacity_scaling();
1267 
1268 	mutex_unlock(&hybrid_capacity_lock);
1269 
1270 	/* Reset the capacity of the CPU going offline to the initial value. */
1271 	hybrid_clear_cpu_capacity(cpu->cpu);
1272 }
1273 
1274 #define POWER_CTL_EE_ENABLE	1
1275 #define POWER_CTL_EE_DISABLE	2
1276 
1277 static int power_ctl_ee_state;
1278 
1279 static void set_power_ctl_ee_state(bool input)
1280 {
1281 	u64 power_ctl;
1282 
1283 	mutex_lock(&intel_pstate_driver_lock);
1284 	rdmsrl(MSR_IA32_POWER_CTL, power_ctl);
1285 	if (input) {
1286 		power_ctl &= ~BIT(MSR_IA32_POWER_CTL_BIT_EE);
1287 		power_ctl_ee_state = POWER_CTL_EE_ENABLE;
1288 	} else {
1289 		power_ctl |= BIT(MSR_IA32_POWER_CTL_BIT_EE);
1290 		power_ctl_ee_state = POWER_CTL_EE_DISABLE;
1291 	}
1292 	wrmsrl(MSR_IA32_POWER_CTL, power_ctl);
1293 	mutex_unlock(&intel_pstate_driver_lock);
1294 }
1295 
1296 static void intel_pstate_hwp_enable(struct cpudata *cpudata);
1297 
1298 static void intel_pstate_hwp_reenable(struct cpudata *cpu)
1299 {
1300 	intel_pstate_hwp_enable(cpu);
1301 	wrmsrl_on_cpu(cpu->cpu, MSR_HWP_REQUEST, READ_ONCE(cpu->hwp_req_cached));
1302 }
1303 
1304 static int intel_pstate_suspend(struct cpufreq_policy *policy)
1305 {
1306 	struct cpudata *cpu = all_cpu_data[policy->cpu];
1307 
1308 	pr_debug("CPU %d suspending\n", cpu->cpu);
1309 
1310 	cpu->suspended = true;
1311 
1312 	/* disable HWP interrupt and cancel any pending work */
1313 	intel_pstate_disable_hwp_interrupt(cpu);
1314 
1315 	return 0;
1316 }
1317 
1318 static int intel_pstate_resume(struct cpufreq_policy *policy)
1319 {
1320 	struct cpudata *cpu = all_cpu_data[policy->cpu];
1321 
1322 	pr_debug("CPU %d resuming\n", cpu->cpu);
1323 
1324 	/* Only restore if the system default is changed */
1325 	if (power_ctl_ee_state == POWER_CTL_EE_ENABLE)
1326 		set_power_ctl_ee_state(true);
1327 	else if (power_ctl_ee_state == POWER_CTL_EE_DISABLE)
1328 		set_power_ctl_ee_state(false);
1329 
1330 	if (cpu->suspended && hwp_active) {
1331 		mutex_lock(&intel_pstate_limits_lock);
1332 
1333 		/* Re-enable HWP, because "online" has not done that. */
1334 		intel_pstate_hwp_reenable(cpu);
1335 
1336 		mutex_unlock(&intel_pstate_limits_lock);
1337 	}
1338 
1339 	cpu->suspended = false;
1340 
1341 	return 0;
1342 }
1343 
1344 static void intel_pstate_update_policies(void)
1345 {
1346 	int cpu;
1347 
1348 	for_each_possible_cpu(cpu)
1349 		cpufreq_update_policy(cpu);
1350 }
1351 
1352 static void __intel_pstate_update_max_freq(struct cpudata *cpudata,
1353 					   struct cpufreq_policy *policy)
1354 {
1355 	if (hwp_active)
1356 		intel_pstate_get_hwp_cap(cpudata);
1357 
1358 	policy->cpuinfo.max_freq = READ_ONCE(global.no_turbo) ?
1359 			cpudata->pstate.max_freq : cpudata->pstate.turbo_freq;
1360 
1361 	refresh_frequency_limits(policy);
1362 }
1363 
1364 static void intel_pstate_update_limits(unsigned int cpu)
1365 {
1366 	struct cpufreq_policy *policy = cpufreq_cpu_acquire(cpu);
1367 	struct cpudata *cpudata;
1368 
1369 	if (!policy)
1370 		return;
1371 
1372 	cpudata = all_cpu_data[cpu];
1373 
1374 	__intel_pstate_update_max_freq(cpudata, policy);
1375 
1376 	/* Prevent the driver from being unregistered now. */
1377 	mutex_lock(&intel_pstate_driver_lock);
1378 
1379 	cpufreq_cpu_release(policy);
1380 
1381 	hybrid_update_capacity(cpudata);
1382 
1383 	mutex_unlock(&intel_pstate_driver_lock);
1384 }
1385 
1386 static void intel_pstate_update_limits_for_all(void)
1387 {
1388 	int cpu;
1389 
1390 	for_each_possible_cpu(cpu) {
1391 		struct cpufreq_policy *policy = cpufreq_cpu_acquire(cpu);
1392 
1393 		if (!policy)
1394 			continue;
1395 
1396 		__intel_pstate_update_max_freq(all_cpu_data[cpu], policy);
1397 
1398 		cpufreq_cpu_release(policy);
1399 	}
1400 
1401 	mutex_lock(&hybrid_capacity_lock);
1402 
1403 	if (hybrid_max_perf_cpu)
1404 		__hybrid_refresh_cpu_capacity_scaling();
1405 
1406 	mutex_unlock(&hybrid_capacity_lock);
1407 }
1408 
1409 /************************** sysfs begin ************************/
1410 #define show_one(file_name, object)					\
1411 	static ssize_t show_##file_name					\
1412 	(struct kobject *kobj, struct kobj_attribute *attr, char *buf)	\
1413 	{								\
1414 		return sprintf(buf, "%u\n", global.object);		\
1415 	}
1416 
1417 static ssize_t intel_pstate_show_status(char *buf);
1418 static int intel_pstate_update_status(const char *buf, size_t size);
1419 
1420 static ssize_t show_status(struct kobject *kobj,
1421 			   struct kobj_attribute *attr, char *buf)
1422 {
1423 	ssize_t ret;
1424 
1425 	mutex_lock(&intel_pstate_driver_lock);
1426 	ret = intel_pstate_show_status(buf);
1427 	mutex_unlock(&intel_pstate_driver_lock);
1428 
1429 	return ret;
1430 }
1431 
1432 static ssize_t store_status(struct kobject *a, struct kobj_attribute *b,
1433 			    const char *buf, size_t count)
1434 {
1435 	char *p = memchr(buf, '\n', count);
1436 	int ret;
1437 
1438 	mutex_lock(&intel_pstate_driver_lock);
1439 	ret = intel_pstate_update_status(buf, p ? p - buf : count);
1440 	mutex_unlock(&intel_pstate_driver_lock);
1441 
1442 	return ret < 0 ? ret : count;
1443 }
1444 
1445 static ssize_t show_turbo_pct(struct kobject *kobj,
1446 				struct kobj_attribute *attr, char *buf)
1447 {
1448 	struct cpudata *cpu;
1449 	int total, no_turbo, turbo_pct;
1450 	uint32_t turbo_fp;
1451 
1452 	mutex_lock(&intel_pstate_driver_lock);
1453 
1454 	if (!intel_pstate_driver) {
1455 		mutex_unlock(&intel_pstate_driver_lock);
1456 		return -EAGAIN;
1457 	}
1458 
1459 	cpu = all_cpu_data[0];
1460 
1461 	total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
1462 	no_turbo = cpu->pstate.max_pstate - cpu->pstate.min_pstate + 1;
1463 	turbo_fp = div_fp(no_turbo, total);
1464 	turbo_pct = 100 - fp_toint(mul_fp(turbo_fp, int_tofp(100)));
1465 
1466 	mutex_unlock(&intel_pstate_driver_lock);
1467 
1468 	return sprintf(buf, "%u\n", turbo_pct);
1469 }
1470 
1471 static ssize_t show_num_pstates(struct kobject *kobj,
1472 				struct kobj_attribute *attr, char *buf)
1473 {
1474 	struct cpudata *cpu;
1475 	int total;
1476 
1477 	mutex_lock(&intel_pstate_driver_lock);
1478 
1479 	if (!intel_pstate_driver) {
1480 		mutex_unlock(&intel_pstate_driver_lock);
1481 		return -EAGAIN;
1482 	}
1483 
1484 	cpu = all_cpu_data[0];
1485 	total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
1486 
1487 	mutex_unlock(&intel_pstate_driver_lock);
1488 
1489 	return sprintf(buf, "%u\n", total);
1490 }
1491 
1492 static ssize_t show_no_turbo(struct kobject *kobj,
1493 			     struct kobj_attribute *attr, char *buf)
1494 {
1495 	ssize_t ret;
1496 
1497 	mutex_lock(&intel_pstate_driver_lock);
1498 
1499 	if (!intel_pstate_driver) {
1500 		mutex_unlock(&intel_pstate_driver_lock);
1501 		return -EAGAIN;
1502 	}
1503 
1504 	ret = sprintf(buf, "%u\n", global.no_turbo);
1505 
1506 	mutex_unlock(&intel_pstate_driver_lock);
1507 
1508 	return ret;
1509 }
1510 
1511 static ssize_t store_no_turbo(struct kobject *a, struct kobj_attribute *b,
1512 			      const char *buf, size_t count)
1513 {
1514 	unsigned int input;
1515 	bool no_turbo;
1516 
1517 	if (sscanf(buf, "%u", &input) != 1)
1518 		return -EINVAL;
1519 
1520 	mutex_lock(&intel_pstate_driver_lock);
1521 
1522 	if (!intel_pstate_driver) {
1523 		count = -EAGAIN;
1524 		goto unlock_driver;
1525 	}
1526 
1527 	no_turbo = !!clamp_t(int, input, 0, 1);
1528 
1529 	WRITE_ONCE(global.turbo_disabled, turbo_is_disabled());
1530 	if (global.turbo_disabled && !no_turbo) {
1531 		pr_notice("Turbo disabled by BIOS or unavailable on processor\n");
1532 		count = -EPERM;
1533 		if (global.no_turbo)
1534 			goto unlock_driver;
1535 		else
1536 			no_turbo = 1;
1537 	}
1538 
1539 	if (no_turbo == global.no_turbo) {
1540 		goto unlock_driver;
1541 	}
1542 
1543 	WRITE_ONCE(global.no_turbo, no_turbo);
1544 
1545 	mutex_lock(&intel_pstate_limits_lock);
1546 
1547 	if (no_turbo) {
1548 		struct cpudata *cpu = all_cpu_data[0];
1549 		int pct = cpu->pstate.max_pstate * 100 / cpu->pstate.turbo_pstate;
1550 
1551 		/* Squash the global minimum into the permitted range. */
1552 		if (global.min_perf_pct > pct)
1553 			global.min_perf_pct = pct;
1554 	}
1555 
1556 	mutex_unlock(&intel_pstate_limits_lock);
1557 
1558 	intel_pstate_update_limits_for_all();
1559 	arch_set_max_freq_ratio(no_turbo);
1560 
1561 unlock_driver:
1562 	mutex_unlock(&intel_pstate_driver_lock);
1563 
1564 	return count;
1565 }
1566 
1567 static void update_qos_request(enum freq_qos_req_type type)
1568 {
1569 	struct freq_qos_request *req;
1570 	struct cpufreq_policy *policy;
1571 	int i;
1572 
1573 	for_each_possible_cpu(i) {
1574 		struct cpudata *cpu = all_cpu_data[i];
1575 		unsigned int freq, perf_pct;
1576 
1577 		policy = cpufreq_cpu_get(i);
1578 		if (!policy)
1579 			continue;
1580 
1581 		req = policy->driver_data;
1582 		cpufreq_cpu_put(policy);
1583 
1584 		if (!req)
1585 			continue;
1586 
1587 		if (hwp_active)
1588 			intel_pstate_get_hwp_cap(cpu);
1589 
1590 		if (type == FREQ_QOS_MIN) {
1591 			perf_pct = global.min_perf_pct;
1592 		} else {
1593 			req++;
1594 			perf_pct = global.max_perf_pct;
1595 		}
1596 
1597 		freq = DIV_ROUND_UP(cpu->pstate.turbo_freq * perf_pct, 100);
1598 
1599 		if (freq_qos_update_request(req, freq) < 0)
1600 			pr_warn("Failed to update freq constraint: CPU%d\n", i);
1601 	}
1602 }
1603 
1604 static ssize_t store_max_perf_pct(struct kobject *a, struct kobj_attribute *b,
1605 				  const char *buf, size_t count)
1606 {
1607 	unsigned int input;
1608 	int ret;
1609 
1610 	ret = sscanf(buf, "%u", &input);
1611 	if (ret != 1)
1612 		return -EINVAL;
1613 
1614 	mutex_lock(&intel_pstate_driver_lock);
1615 
1616 	if (!intel_pstate_driver) {
1617 		mutex_unlock(&intel_pstate_driver_lock);
1618 		return -EAGAIN;
1619 	}
1620 
1621 	mutex_lock(&intel_pstate_limits_lock);
1622 
1623 	global.max_perf_pct = clamp_t(int, input, global.min_perf_pct, 100);
1624 
1625 	mutex_unlock(&intel_pstate_limits_lock);
1626 
1627 	if (intel_pstate_driver == &intel_pstate)
1628 		intel_pstate_update_policies();
1629 	else
1630 		update_qos_request(FREQ_QOS_MAX);
1631 
1632 	mutex_unlock(&intel_pstate_driver_lock);
1633 
1634 	return count;
1635 }
1636 
1637 static ssize_t store_min_perf_pct(struct kobject *a, struct kobj_attribute *b,
1638 				  const char *buf, size_t count)
1639 {
1640 	unsigned int input;
1641 	int ret;
1642 
1643 	ret = sscanf(buf, "%u", &input);
1644 	if (ret != 1)
1645 		return -EINVAL;
1646 
1647 	mutex_lock(&intel_pstate_driver_lock);
1648 
1649 	if (!intel_pstate_driver) {
1650 		mutex_unlock(&intel_pstate_driver_lock);
1651 		return -EAGAIN;
1652 	}
1653 
1654 	mutex_lock(&intel_pstate_limits_lock);
1655 
1656 	global.min_perf_pct = clamp_t(int, input,
1657 				      min_perf_pct_min(), global.max_perf_pct);
1658 
1659 	mutex_unlock(&intel_pstate_limits_lock);
1660 
1661 	if (intel_pstate_driver == &intel_pstate)
1662 		intel_pstate_update_policies();
1663 	else
1664 		update_qos_request(FREQ_QOS_MIN);
1665 
1666 	mutex_unlock(&intel_pstate_driver_lock);
1667 
1668 	return count;
1669 }
1670 
1671 static ssize_t show_hwp_dynamic_boost(struct kobject *kobj,
1672 				struct kobj_attribute *attr, char *buf)
1673 {
1674 	return sprintf(buf, "%u\n", hwp_boost);
1675 }
1676 
1677 static ssize_t store_hwp_dynamic_boost(struct kobject *a,
1678 				       struct kobj_attribute *b,
1679 				       const char *buf, size_t count)
1680 {
1681 	unsigned int input;
1682 	int ret;
1683 
1684 	ret = kstrtouint(buf, 10, &input);
1685 	if (ret)
1686 		return ret;
1687 
1688 	mutex_lock(&intel_pstate_driver_lock);
1689 	hwp_boost = !!input;
1690 	intel_pstate_update_policies();
1691 	mutex_unlock(&intel_pstate_driver_lock);
1692 
1693 	return count;
1694 }
1695 
1696 static ssize_t show_energy_efficiency(struct kobject *kobj, struct kobj_attribute *attr,
1697 				      char *buf)
1698 {
1699 	u64 power_ctl;
1700 	int enable;
1701 
1702 	rdmsrl(MSR_IA32_POWER_CTL, power_ctl);
1703 	enable = !!(power_ctl & BIT(MSR_IA32_POWER_CTL_BIT_EE));
1704 	return sprintf(buf, "%d\n", !enable);
1705 }
1706 
1707 static ssize_t store_energy_efficiency(struct kobject *a, struct kobj_attribute *b,
1708 				       const char *buf, size_t count)
1709 {
1710 	bool input;
1711 	int ret;
1712 
1713 	ret = kstrtobool(buf, &input);
1714 	if (ret)
1715 		return ret;
1716 
1717 	set_power_ctl_ee_state(input);
1718 
1719 	return count;
1720 }
1721 
1722 show_one(max_perf_pct, max_perf_pct);
1723 show_one(min_perf_pct, min_perf_pct);
1724 
1725 define_one_global_rw(status);
1726 define_one_global_rw(no_turbo);
1727 define_one_global_rw(max_perf_pct);
1728 define_one_global_rw(min_perf_pct);
1729 define_one_global_ro(turbo_pct);
1730 define_one_global_ro(num_pstates);
1731 define_one_global_rw(hwp_dynamic_boost);
1732 define_one_global_rw(energy_efficiency);
1733 
1734 static struct attribute *intel_pstate_attributes[] = {
1735 	&status.attr,
1736 	&no_turbo.attr,
1737 	NULL
1738 };
1739 
1740 static const struct attribute_group intel_pstate_attr_group = {
1741 	.attrs = intel_pstate_attributes,
1742 };
1743 
1744 static const struct x86_cpu_id intel_pstate_cpu_ee_disable_ids[];
1745 
1746 static struct kobject *intel_pstate_kobject;
1747 
1748 static void __init intel_pstate_sysfs_expose_params(void)
1749 {
1750 	struct device *dev_root = bus_get_dev_root(&cpu_subsys);
1751 	int rc;
1752 
1753 	if (dev_root) {
1754 		intel_pstate_kobject = kobject_create_and_add("intel_pstate", &dev_root->kobj);
1755 		put_device(dev_root);
1756 	}
1757 	if (WARN_ON(!intel_pstate_kobject))
1758 		return;
1759 
1760 	rc = sysfs_create_group(intel_pstate_kobject, &intel_pstate_attr_group);
1761 	if (WARN_ON(rc))
1762 		return;
1763 
1764 	if (!boot_cpu_has(X86_FEATURE_HYBRID_CPU)) {
1765 		rc = sysfs_create_file(intel_pstate_kobject, &turbo_pct.attr);
1766 		WARN_ON(rc);
1767 
1768 		rc = sysfs_create_file(intel_pstate_kobject, &num_pstates.attr);
1769 		WARN_ON(rc);
1770 	}
1771 
1772 	/*
1773 	 * If per cpu limits are enforced there are no global limits, so
1774 	 * return without creating max/min_perf_pct attributes
1775 	 */
1776 	if (per_cpu_limits)
1777 		return;
1778 
1779 	rc = sysfs_create_file(intel_pstate_kobject, &max_perf_pct.attr);
1780 	WARN_ON(rc);
1781 
1782 	rc = sysfs_create_file(intel_pstate_kobject, &min_perf_pct.attr);
1783 	WARN_ON(rc);
1784 
1785 	if (x86_match_cpu(intel_pstate_cpu_ee_disable_ids)) {
1786 		rc = sysfs_create_file(intel_pstate_kobject, &energy_efficiency.attr);
1787 		WARN_ON(rc);
1788 	}
1789 }
1790 
1791 static void __init intel_pstate_sysfs_remove(void)
1792 {
1793 	if (!intel_pstate_kobject)
1794 		return;
1795 
1796 	sysfs_remove_group(intel_pstate_kobject, &intel_pstate_attr_group);
1797 
1798 	if (!boot_cpu_has(X86_FEATURE_HYBRID_CPU)) {
1799 		sysfs_remove_file(intel_pstate_kobject, &num_pstates.attr);
1800 		sysfs_remove_file(intel_pstate_kobject, &turbo_pct.attr);
1801 	}
1802 
1803 	if (!per_cpu_limits) {
1804 		sysfs_remove_file(intel_pstate_kobject, &max_perf_pct.attr);
1805 		sysfs_remove_file(intel_pstate_kobject, &min_perf_pct.attr);
1806 
1807 		if (x86_match_cpu(intel_pstate_cpu_ee_disable_ids))
1808 			sysfs_remove_file(intel_pstate_kobject, &energy_efficiency.attr);
1809 	}
1810 
1811 	kobject_put(intel_pstate_kobject);
1812 }
1813 
1814 static void intel_pstate_sysfs_expose_hwp_dynamic_boost(void)
1815 {
1816 	int rc;
1817 
1818 	if (!hwp_active)
1819 		return;
1820 
1821 	rc = sysfs_create_file(intel_pstate_kobject, &hwp_dynamic_boost.attr);
1822 	WARN_ON_ONCE(rc);
1823 }
1824 
1825 static void intel_pstate_sysfs_hide_hwp_dynamic_boost(void)
1826 {
1827 	if (!hwp_active)
1828 		return;
1829 
1830 	sysfs_remove_file(intel_pstate_kobject, &hwp_dynamic_boost.attr);
1831 }
1832 
1833 /************************** sysfs end ************************/
1834 
1835 static void intel_pstate_notify_work(struct work_struct *work)
1836 {
1837 	struct cpudata *cpudata =
1838 		container_of(to_delayed_work(work), struct cpudata, hwp_notify_work);
1839 	struct cpufreq_policy *policy = cpufreq_cpu_acquire(cpudata->cpu);
1840 
1841 	if (policy) {
1842 		__intel_pstate_update_max_freq(cpudata, policy);
1843 
1844 		cpufreq_cpu_release(policy);
1845 
1846 		/*
1847 		 * The driver will not be unregistered while this function is
1848 		 * running, so update the capacity without acquiring the driver
1849 		 * lock.
1850 		 */
1851 		hybrid_update_capacity(cpudata);
1852 	}
1853 
1854 	wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_STATUS, 0);
1855 }
1856 
1857 static DEFINE_RAW_SPINLOCK(hwp_notify_lock);
1858 static cpumask_t hwp_intr_enable_mask;
1859 
1860 #define HWP_GUARANTEED_PERF_CHANGE_STATUS      BIT(0)
1861 #define HWP_HIGHEST_PERF_CHANGE_STATUS         BIT(3)
1862 
1863 void notify_hwp_interrupt(void)
1864 {
1865 	unsigned int this_cpu = smp_processor_id();
1866 	u64 value, status_mask;
1867 	unsigned long flags;
1868 
1869 	if (!hwp_active || !cpu_feature_enabled(X86_FEATURE_HWP_NOTIFY))
1870 		return;
1871 
1872 	status_mask = HWP_GUARANTEED_PERF_CHANGE_STATUS;
1873 	if (cpu_feature_enabled(X86_FEATURE_HWP_HIGHEST_PERF_CHANGE))
1874 		status_mask |= HWP_HIGHEST_PERF_CHANGE_STATUS;
1875 
1876 	rdmsrl_safe(MSR_HWP_STATUS, &value);
1877 	if (!(value & status_mask))
1878 		return;
1879 
1880 	raw_spin_lock_irqsave(&hwp_notify_lock, flags);
1881 
1882 	if (!cpumask_test_cpu(this_cpu, &hwp_intr_enable_mask))
1883 		goto ack_intr;
1884 
1885 	schedule_delayed_work(&all_cpu_data[this_cpu]->hwp_notify_work,
1886 			      msecs_to_jiffies(10));
1887 
1888 	raw_spin_unlock_irqrestore(&hwp_notify_lock, flags);
1889 
1890 	return;
1891 
1892 ack_intr:
1893 	wrmsrl_safe(MSR_HWP_STATUS, 0);
1894 	raw_spin_unlock_irqrestore(&hwp_notify_lock, flags);
1895 }
1896 
1897 static void intel_pstate_disable_hwp_interrupt(struct cpudata *cpudata)
1898 {
1899 	bool cancel_work;
1900 
1901 	if (!cpu_feature_enabled(X86_FEATURE_HWP_NOTIFY))
1902 		return;
1903 
1904 	/* wrmsrl_on_cpu has to be outside spinlock as this can result in IPC */
1905 	wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_INTERRUPT, 0x00);
1906 
1907 	raw_spin_lock_irq(&hwp_notify_lock);
1908 	cancel_work = cpumask_test_and_clear_cpu(cpudata->cpu, &hwp_intr_enable_mask);
1909 	raw_spin_unlock_irq(&hwp_notify_lock);
1910 
1911 	if (cancel_work)
1912 		cancel_delayed_work_sync(&cpudata->hwp_notify_work);
1913 }
1914 
1915 #define HWP_GUARANTEED_PERF_CHANGE_REQ BIT(0)
1916 #define HWP_HIGHEST_PERF_CHANGE_REQ    BIT(2)
1917 
1918 static void intel_pstate_enable_hwp_interrupt(struct cpudata *cpudata)
1919 {
1920 	/* Enable HWP notification interrupt for performance change */
1921 	if (boot_cpu_has(X86_FEATURE_HWP_NOTIFY)) {
1922 		u64 interrupt_mask = HWP_GUARANTEED_PERF_CHANGE_REQ;
1923 
1924 		raw_spin_lock_irq(&hwp_notify_lock);
1925 		INIT_DELAYED_WORK(&cpudata->hwp_notify_work, intel_pstate_notify_work);
1926 		cpumask_set_cpu(cpudata->cpu, &hwp_intr_enable_mask);
1927 		raw_spin_unlock_irq(&hwp_notify_lock);
1928 
1929 		if (cpu_feature_enabled(X86_FEATURE_HWP_HIGHEST_PERF_CHANGE))
1930 			interrupt_mask |= HWP_HIGHEST_PERF_CHANGE_REQ;
1931 
1932 		/* wrmsrl_on_cpu has to be outside spinlock as this can result in IPC */
1933 		wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_INTERRUPT, interrupt_mask);
1934 		wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_STATUS, 0);
1935 	}
1936 }
1937 
1938 static void intel_pstate_update_epp_defaults(struct cpudata *cpudata)
1939 {
1940 	cpudata->epp_default = intel_pstate_get_epp(cpudata, 0);
1941 
1942 	/*
1943 	 * If the EPP is set by firmware, which means that firmware enabled HWP
1944 	 * - Is equal or less than 0x80 (default balance_perf EPP)
1945 	 * - But less performance oriented than performance EPP
1946 	 *   then use this as new balance_perf EPP.
1947 	 */
1948 	if (hwp_forced && cpudata->epp_default <= HWP_EPP_BALANCE_PERFORMANCE &&
1949 	    cpudata->epp_default > HWP_EPP_PERFORMANCE) {
1950 		epp_values[EPP_INDEX_BALANCE_PERFORMANCE] = cpudata->epp_default;
1951 		return;
1952 	}
1953 
1954 	/*
1955 	 * If this CPU gen doesn't call for change in balance_perf
1956 	 * EPP return.
1957 	 */
1958 	if (epp_values[EPP_INDEX_BALANCE_PERFORMANCE] == HWP_EPP_BALANCE_PERFORMANCE)
1959 		return;
1960 
1961 	/*
1962 	 * Use hard coded value per gen to update the balance_perf
1963 	 * and default EPP.
1964 	 */
1965 	cpudata->epp_default = epp_values[EPP_INDEX_BALANCE_PERFORMANCE];
1966 	intel_pstate_set_epp(cpudata, cpudata->epp_default);
1967 }
1968 
1969 static void intel_pstate_hwp_enable(struct cpudata *cpudata)
1970 {
1971 	/* First disable HWP notification interrupt till we activate again */
1972 	if (boot_cpu_has(X86_FEATURE_HWP_NOTIFY))
1973 		wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_INTERRUPT, 0x00);
1974 
1975 	wrmsrl_on_cpu(cpudata->cpu, MSR_PM_ENABLE, 0x1);
1976 
1977 	intel_pstate_enable_hwp_interrupt(cpudata);
1978 
1979 	if (cpudata->epp_default >= 0)
1980 		return;
1981 
1982 	intel_pstate_update_epp_defaults(cpudata);
1983 }
1984 
1985 static int atom_get_min_pstate(int not_used)
1986 {
1987 	u64 value;
1988 
1989 	rdmsrl(MSR_ATOM_CORE_RATIOS, value);
1990 	return (value >> 8) & 0x7F;
1991 }
1992 
1993 static int atom_get_max_pstate(int not_used)
1994 {
1995 	u64 value;
1996 
1997 	rdmsrl(MSR_ATOM_CORE_RATIOS, value);
1998 	return (value >> 16) & 0x7F;
1999 }
2000 
2001 static int atom_get_turbo_pstate(int not_used)
2002 {
2003 	u64 value;
2004 
2005 	rdmsrl(MSR_ATOM_CORE_TURBO_RATIOS, value);
2006 	return value & 0x7F;
2007 }
2008 
2009 static u64 atom_get_val(struct cpudata *cpudata, int pstate)
2010 {
2011 	u64 val;
2012 	int32_t vid_fp;
2013 	u32 vid;
2014 
2015 	val = (u64)pstate << 8;
2016 	if (READ_ONCE(global.no_turbo) && !READ_ONCE(global.turbo_disabled))
2017 		val |= (u64)1 << 32;
2018 
2019 	vid_fp = cpudata->vid.min + mul_fp(
2020 		int_tofp(pstate - cpudata->pstate.min_pstate),
2021 		cpudata->vid.ratio);
2022 
2023 	vid_fp = clamp_t(int32_t, vid_fp, cpudata->vid.min, cpudata->vid.max);
2024 	vid = ceiling_fp(vid_fp);
2025 
2026 	if (pstate > cpudata->pstate.max_pstate)
2027 		vid = cpudata->vid.turbo;
2028 
2029 	return val | vid;
2030 }
2031 
2032 static int silvermont_get_scaling(void)
2033 {
2034 	u64 value;
2035 	int i;
2036 	/* Defined in Table 35-6 from SDM (Sept 2015) */
2037 	static int silvermont_freq_table[] = {
2038 		83300, 100000, 133300, 116700, 80000};
2039 
2040 	rdmsrl(MSR_FSB_FREQ, value);
2041 	i = value & 0x7;
2042 	WARN_ON(i > 4);
2043 
2044 	return silvermont_freq_table[i];
2045 }
2046 
2047 static int airmont_get_scaling(void)
2048 {
2049 	u64 value;
2050 	int i;
2051 	/* Defined in Table 35-10 from SDM (Sept 2015) */
2052 	static int airmont_freq_table[] = {
2053 		83300, 100000, 133300, 116700, 80000,
2054 		93300, 90000, 88900, 87500};
2055 
2056 	rdmsrl(MSR_FSB_FREQ, value);
2057 	i = value & 0xF;
2058 	WARN_ON(i > 8);
2059 
2060 	return airmont_freq_table[i];
2061 }
2062 
2063 static void atom_get_vid(struct cpudata *cpudata)
2064 {
2065 	u64 value;
2066 
2067 	rdmsrl(MSR_ATOM_CORE_VIDS, value);
2068 	cpudata->vid.min = int_tofp((value >> 8) & 0x7f);
2069 	cpudata->vid.max = int_tofp((value >> 16) & 0x7f);
2070 	cpudata->vid.ratio = div_fp(
2071 		cpudata->vid.max - cpudata->vid.min,
2072 		int_tofp(cpudata->pstate.max_pstate -
2073 			cpudata->pstate.min_pstate));
2074 
2075 	rdmsrl(MSR_ATOM_CORE_TURBO_VIDS, value);
2076 	cpudata->vid.turbo = value & 0x7f;
2077 }
2078 
2079 static int core_get_min_pstate(int cpu)
2080 {
2081 	u64 value;
2082 
2083 	rdmsrl_on_cpu(cpu, MSR_PLATFORM_INFO, &value);
2084 	return (value >> 40) & 0xFF;
2085 }
2086 
2087 static int core_get_max_pstate_physical(int cpu)
2088 {
2089 	u64 value;
2090 
2091 	rdmsrl_on_cpu(cpu, MSR_PLATFORM_INFO, &value);
2092 	return (value >> 8) & 0xFF;
2093 }
2094 
2095 static int core_get_tdp_ratio(int cpu, u64 plat_info)
2096 {
2097 	/* Check how many TDP levels present */
2098 	if (plat_info & 0x600000000) {
2099 		u64 tdp_ctrl;
2100 		u64 tdp_ratio;
2101 		int tdp_msr;
2102 		int err;
2103 
2104 		/* Get the TDP level (0, 1, 2) to get ratios */
2105 		err = rdmsrl_safe_on_cpu(cpu, MSR_CONFIG_TDP_CONTROL, &tdp_ctrl);
2106 		if (err)
2107 			return err;
2108 
2109 		/* TDP MSR are continuous starting at 0x648 */
2110 		tdp_msr = MSR_CONFIG_TDP_NOMINAL + (tdp_ctrl & 0x03);
2111 		err = rdmsrl_safe_on_cpu(cpu, tdp_msr, &tdp_ratio);
2112 		if (err)
2113 			return err;
2114 
2115 		/* For level 1 and 2, bits[23:16] contain the ratio */
2116 		if (tdp_ctrl & 0x03)
2117 			tdp_ratio >>= 16;
2118 
2119 		tdp_ratio &= 0xff; /* ratios are only 8 bits long */
2120 		pr_debug("tdp_ratio %x\n", (int)tdp_ratio);
2121 
2122 		return (int)tdp_ratio;
2123 	}
2124 
2125 	return -ENXIO;
2126 }
2127 
2128 static int core_get_max_pstate(int cpu)
2129 {
2130 	u64 tar;
2131 	u64 plat_info;
2132 	int max_pstate;
2133 	int tdp_ratio;
2134 	int err;
2135 
2136 	rdmsrl_on_cpu(cpu, MSR_PLATFORM_INFO, &plat_info);
2137 	max_pstate = (plat_info >> 8) & 0xFF;
2138 
2139 	tdp_ratio = core_get_tdp_ratio(cpu, plat_info);
2140 	if (tdp_ratio <= 0)
2141 		return max_pstate;
2142 
2143 	if (hwp_active) {
2144 		/* Turbo activation ratio is not used on HWP platforms */
2145 		return tdp_ratio;
2146 	}
2147 
2148 	err = rdmsrl_safe_on_cpu(cpu, MSR_TURBO_ACTIVATION_RATIO, &tar);
2149 	if (!err) {
2150 		int tar_levels;
2151 
2152 		/* Do some sanity checking for safety */
2153 		tar_levels = tar & 0xff;
2154 		if (tdp_ratio - 1 == tar_levels) {
2155 			max_pstate = tar_levels;
2156 			pr_debug("max_pstate=TAC %x\n", max_pstate);
2157 		}
2158 	}
2159 
2160 	return max_pstate;
2161 }
2162 
2163 static int core_get_turbo_pstate(int cpu)
2164 {
2165 	u64 value;
2166 	int nont, ret;
2167 
2168 	rdmsrl_on_cpu(cpu, MSR_TURBO_RATIO_LIMIT, &value);
2169 	nont = core_get_max_pstate(cpu);
2170 	ret = (value) & 255;
2171 	if (ret <= nont)
2172 		ret = nont;
2173 	return ret;
2174 }
2175 
2176 static u64 core_get_val(struct cpudata *cpudata, int pstate)
2177 {
2178 	u64 val;
2179 
2180 	val = (u64)pstate << 8;
2181 	if (READ_ONCE(global.no_turbo) && !READ_ONCE(global.turbo_disabled))
2182 		val |= (u64)1 << 32;
2183 
2184 	return val;
2185 }
2186 
2187 static int knl_get_aperf_mperf_shift(void)
2188 {
2189 	return 10;
2190 }
2191 
2192 static int knl_get_turbo_pstate(int cpu)
2193 {
2194 	u64 value;
2195 	int nont, ret;
2196 
2197 	rdmsrl_on_cpu(cpu, MSR_TURBO_RATIO_LIMIT, &value);
2198 	nont = core_get_max_pstate(cpu);
2199 	ret = (((value) >> 8) & 0xFF);
2200 	if (ret <= nont)
2201 		ret = nont;
2202 	return ret;
2203 }
2204 
2205 static void hybrid_get_type(void *data)
2206 {
2207 	u8 *cpu_type = data;
2208 
2209 	*cpu_type = get_this_hybrid_cpu_type();
2210 }
2211 
2212 static int hwp_get_cpu_scaling(int cpu)
2213 {
2214 	u8 cpu_type = 0;
2215 
2216 	smp_call_function_single(cpu, hybrid_get_type, &cpu_type, 1);
2217 	/* P-cores have a smaller perf level-to-freqency scaling factor. */
2218 	if (cpu_type == 0x40)
2219 		return hybrid_scaling_factor;
2220 
2221 	/* Use default core scaling for E-cores */
2222 	if (cpu_type == 0x20)
2223 		return core_get_scaling();
2224 
2225 	/*
2226 	 * If reached here, this system is either non-hybrid (like Tiger
2227 	 * Lake) or hybrid-capable (like Alder Lake or Raptor Lake) with
2228 	 * no E cores (in which case CPUID for hybrid support is 0).
2229 	 *
2230 	 * The CPPC nominal_frequency field is 0 for non-hybrid systems,
2231 	 * so the default core scaling will be used for them.
2232 	 */
2233 	return intel_pstate_cppc_get_scaling(cpu);
2234 }
2235 
2236 static void intel_pstate_set_pstate(struct cpudata *cpu, int pstate)
2237 {
2238 	trace_cpu_frequency(pstate * cpu->pstate.scaling, cpu->cpu);
2239 	cpu->pstate.current_pstate = pstate;
2240 	/*
2241 	 * Generally, there is no guarantee that this code will always run on
2242 	 * the CPU being updated, so force the register update to run on the
2243 	 * right CPU.
2244 	 */
2245 	wrmsrl_on_cpu(cpu->cpu, MSR_IA32_PERF_CTL,
2246 		      pstate_funcs.get_val(cpu, pstate));
2247 }
2248 
2249 static void intel_pstate_set_min_pstate(struct cpudata *cpu)
2250 {
2251 	intel_pstate_set_pstate(cpu, cpu->pstate.min_pstate);
2252 }
2253 
2254 static void intel_pstate_get_cpu_pstates(struct cpudata *cpu)
2255 {
2256 	int perf_ctl_max_phys = pstate_funcs.get_max_physical(cpu->cpu);
2257 	int perf_ctl_scaling = pstate_funcs.get_scaling();
2258 
2259 	cpu->pstate.min_pstate = pstate_funcs.get_min(cpu->cpu);
2260 	cpu->pstate.max_pstate_physical = perf_ctl_max_phys;
2261 	cpu->pstate.perf_ctl_scaling = perf_ctl_scaling;
2262 
2263 	if (hwp_active && !hwp_mode_bdw) {
2264 		__intel_pstate_get_hwp_cap(cpu);
2265 
2266 		if (pstate_funcs.get_cpu_scaling) {
2267 			cpu->pstate.scaling = pstate_funcs.get_cpu_scaling(cpu->cpu);
2268 			if (cpu->pstate.scaling != perf_ctl_scaling) {
2269 				intel_pstate_hybrid_hwp_adjust(cpu);
2270 				hwp_is_hybrid = true;
2271 			}
2272 		} else {
2273 			cpu->pstate.scaling = perf_ctl_scaling;
2274 		}
2275 		/*
2276 		 * If the CPU is going online for the first time and it was
2277 		 * offline initially, asym capacity scaling needs to be updated.
2278 		 */
2279 		hybrid_update_capacity(cpu);
2280 	} else {
2281 		cpu->pstate.scaling = perf_ctl_scaling;
2282 		cpu->pstate.max_pstate = pstate_funcs.get_max(cpu->cpu);
2283 		cpu->pstate.turbo_pstate = pstate_funcs.get_turbo(cpu->cpu);
2284 	}
2285 
2286 	if (cpu->pstate.scaling == perf_ctl_scaling) {
2287 		cpu->pstate.min_freq = cpu->pstate.min_pstate * perf_ctl_scaling;
2288 		cpu->pstate.max_freq = cpu->pstate.max_pstate * perf_ctl_scaling;
2289 		cpu->pstate.turbo_freq = cpu->pstate.turbo_pstate * perf_ctl_scaling;
2290 	}
2291 
2292 	if (pstate_funcs.get_aperf_mperf_shift)
2293 		cpu->aperf_mperf_shift = pstate_funcs.get_aperf_mperf_shift();
2294 
2295 	if (pstate_funcs.get_vid)
2296 		pstate_funcs.get_vid(cpu);
2297 
2298 	intel_pstate_set_min_pstate(cpu);
2299 }
2300 
2301 /*
2302  * Long hold time will keep high perf limits for long time,
2303  * which negatively impacts perf/watt for some workloads,
2304  * like specpower. 3ms is based on experiements on some
2305  * workoads.
2306  */
2307 static int hwp_boost_hold_time_ns = 3 * NSEC_PER_MSEC;
2308 
2309 static inline void intel_pstate_hwp_boost_up(struct cpudata *cpu)
2310 {
2311 	u64 hwp_req = READ_ONCE(cpu->hwp_req_cached);
2312 	u64 hwp_cap = READ_ONCE(cpu->hwp_cap_cached);
2313 	u32 max_limit = (hwp_req & 0xff00) >> 8;
2314 	u32 min_limit = (hwp_req & 0xff);
2315 	u32 boost_level1;
2316 
2317 	/*
2318 	 * Cases to consider (User changes via sysfs or boot time):
2319 	 * If, P0 (Turbo max) = P1 (Guaranteed max) = min:
2320 	 *	No boost, return.
2321 	 * If, P0 (Turbo max) > P1 (Guaranteed max) = min:
2322 	 *     Should result in one level boost only for P0.
2323 	 * If, P0 (Turbo max) = P1 (Guaranteed max) > min:
2324 	 *     Should result in two level boost:
2325 	 *         (min + p1)/2 and P1.
2326 	 * If, P0 (Turbo max) > P1 (Guaranteed max) > min:
2327 	 *     Should result in three level boost:
2328 	 *        (min + p1)/2, P1 and P0.
2329 	 */
2330 
2331 	/* If max and min are equal or already at max, nothing to boost */
2332 	if (max_limit == min_limit || cpu->hwp_boost_min >= max_limit)
2333 		return;
2334 
2335 	if (!cpu->hwp_boost_min)
2336 		cpu->hwp_boost_min = min_limit;
2337 
2338 	/* level at half way mark between min and guranteed */
2339 	boost_level1 = (HWP_GUARANTEED_PERF(hwp_cap) + min_limit) >> 1;
2340 
2341 	if (cpu->hwp_boost_min < boost_level1)
2342 		cpu->hwp_boost_min = boost_level1;
2343 	else if (cpu->hwp_boost_min < HWP_GUARANTEED_PERF(hwp_cap))
2344 		cpu->hwp_boost_min = HWP_GUARANTEED_PERF(hwp_cap);
2345 	else if (cpu->hwp_boost_min == HWP_GUARANTEED_PERF(hwp_cap) &&
2346 		 max_limit != HWP_GUARANTEED_PERF(hwp_cap))
2347 		cpu->hwp_boost_min = max_limit;
2348 	else
2349 		return;
2350 
2351 	hwp_req = (hwp_req & ~GENMASK_ULL(7, 0)) | cpu->hwp_boost_min;
2352 	wrmsrl(MSR_HWP_REQUEST, hwp_req);
2353 	cpu->last_update = cpu->sample.time;
2354 }
2355 
2356 static inline void intel_pstate_hwp_boost_down(struct cpudata *cpu)
2357 {
2358 	if (cpu->hwp_boost_min) {
2359 		bool expired;
2360 
2361 		/* Check if we are idle for hold time to boost down */
2362 		expired = time_after64(cpu->sample.time, cpu->last_update +
2363 				       hwp_boost_hold_time_ns);
2364 		if (expired) {
2365 			wrmsrl(MSR_HWP_REQUEST, cpu->hwp_req_cached);
2366 			cpu->hwp_boost_min = 0;
2367 		}
2368 	}
2369 	cpu->last_update = cpu->sample.time;
2370 }
2371 
2372 static inline void intel_pstate_update_util_hwp_local(struct cpudata *cpu,
2373 						      u64 time)
2374 {
2375 	cpu->sample.time = time;
2376 
2377 	if (cpu->sched_flags & SCHED_CPUFREQ_IOWAIT) {
2378 		bool do_io = false;
2379 
2380 		cpu->sched_flags = 0;
2381 		/*
2382 		 * Set iowait_boost flag and update time. Since IO WAIT flag
2383 		 * is set all the time, we can't just conclude that there is
2384 		 * some IO bound activity is scheduled on this CPU with just
2385 		 * one occurrence. If we receive at least two in two
2386 		 * consecutive ticks, then we treat as boost candidate.
2387 		 */
2388 		if (time_before64(time, cpu->last_io_update + 2 * TICK_NSEC))
2389 			do_io = true;
2390 
2391 		cpu->last_io_update = time;
2392 
2393 		if (do_io)
2394 			intel_pstate_hwp_boost_up(cpu);
2395 
2396 	} else {
2397 		intel_pstate_hwp_boost_down(cpu);
2398 	}
2399 }
2400 
2401 static inline void intel_pstate_update_util_hwp(struct update_util_data *data,
2402 						u64 time, unsigned int flags)
2403 {
2404 	struct cpudata *cpu = container_of(data, struct cpudata, update_util);
2405 
2406 	cpu->sched_flags |= flags;
2407 
2408 	if (smp_processor_id() == cpu->cpu)
2409 		intel_pstate_update_util_hwp_local(cpu, time);
2410 }
2411 
2412 static inline void intel_pstate_calc_avg_perf(struct cpudata *cpu)
2413 {
2414 	struct sample *sample = &cpu->sample;
2415 
2416 	sample->core_avg_perf = div_ext_fp(sample->aperf, sample->mperf);
2417 }
2418 
2419 static inline bool intel_pstate_sample(struct cpudata *cpu, u64 time)
2420 {
2421 	u64 aperf, mperf;
2422 	unsigned long flags;
2423 	u64 tsc;
2424 
2425 	local_irq_save(flags);
2426 	rdmsrl(MSR_IA32_APERF, aperf);
2427 	rdmsrl(MSR_IA32_MPERF, mperf);
2428 	tsc = rdtsc();
2429 	if (cpu->prev_mperf == mperf || cpu->prev_tsc == tsc) {
2430 		local_irq_restore(flags);
2431 		return false;
2432 	}
2433 	local_irq_restore(flags);
2434 
2435 	cpu->last_sample_time = cpu->sample.time;
2436 	cpu->sample.time = time;
2437 	cpu->sample.aperf = aperf;
2438 	cpu->sample.mperf = mperf;
2439 	cpu->sample.tsc =  tsc;
2440 	cpu->sample.aperf -= cpu->prev_aperf;
2441 	cpu->sample.mperf -= cpu->prev_mperf;
2442 	cpu->sample.tsc -= cpu->prev_tsc;
2443 
2444 	cpu->prev_aperf = aperf;
2445 	cpu->prev_mperf = mperf;
2446 	cpu->prev_tsc = tsc;
2447 	/*
2448 	 * First time this function is invoked in a given cycle, all of the
2449 	 * previous sample data fields are equal to zero or stale and they must
2450 	 * be populated with meaningful numbers for things to work, so assume
2451 	 * that sample.time will always be reset before setting the utilization
2452 	 * update hook and make the caller skip the sample then.
2453 	 */
2454 	if (cpu->last_sample_time) {
2455 		intel_pstate_calc_avg_perf(cpu);
2456 		return true;
2457 	}
2458 	return false;
2459 }
2460 
2461 static inline int32_t get_avg_frequency(struct cpudata *cpu)
2462 {
2463 	return mul_ext_fp(cpu->sample.core_avg_perf, cpu_khz);
2464 }
2465 
2466 static inline int32_t get_avg_pstate(struct cpudata *cpu)
2467 {
2468 	return mul_ext_fp(cpu->pstate.max_pstate_physical,
2469 			  cpu->sample.core_avg_perf);
2470 }
2471 
2472 static inline int32_t get_target_pstate(struct cpudata *cpu)
2473 {
2474 	struct sample *sample = &cpu->sample;
2475 	int32_t busy_frac;
2476 	int target, avg_pstate;
2477 
2478 	busy_frac = div_fp(sample->mperf << cpu->aperf_mperf_shift,
2479 			   sample->tsc);
2480 
2481 	if (busy_frac < cpu->iowait_boost)
2482 		busy_frac = cpu->iowait_boost;
2483 
2484 	sample->busy_scaled = busy_frac * 100;
2485 
2486 	target = READ_ONCE(global.no_turbo) ?
2487 			cpu->pstate.max_pstate : cpu->pstate.turbo_pstate;
2488 	target += target >> 2;
2489 	target = mul_fp(target, busy_frac);
2490 	if (target < cpu->pstate.min_pstate)
2491 		target = cpu->pstate.min_pstate;
2492 
2493 	/*
2494 	 * If the average P-state during the previous cycle was higher than the
2495 	 * current target, add 50% of the difference to the target to reduce
2496 	 * possible performance oscillations and offset possible performance
2497 	 * loss related to moving the workload from one CPU to another within
2498 	 * a package/module.
2499 	 */
2500 	avg_pstate = get_avg_pstate(cpu);
2501 	if (avg_pstate > target)
2502 		target += (avg_pstate - target) >> 1;
2503 
2504 	return target;
2505 }
2506 
2507 static int intel_pstate_prepare_request(struct cpudata *cpu, int pstate)
2508 {
2509 	int min_pstate = max(cpu->pstate.min_pstate, cpu->min_perf_ratio);
2510 	int max_pstate = max(min_pstate, cpu->max_perf_ratio);
2511 
2512 	return clamp_t(int, pstate, min_pstate, max_pstate);
2513 }
2514 
2515 static void intel_pstate_update_pstate(struct cpudata *cpu, int pstate)
2516 {
2517 	if (pstate == cpu->pstate.current_pstate)
2518 		return;
2519 
2520 	cpu->pstate.current_pstate = pstate;
2521 	wrmsrl(MSR_IA32_PERF_CTL, pstate_funcs.get_val(cpu, pstate));
2522 }
2523 
2524 static void intel_pstate_adjust_pstate(struct cpudata *cpu)
2525 {
2526 	int from = cpu->pstate.current_pstate;
2527 	struct sample *sample;
2528 	int target_pstate;
2529 
2530 	target_pstate = get_target_pstate(cpu);
2531 	target_pstate = intel_pstate_prepare_request(cpu, target_pstate);
2532 	trace_cpu_frequency(target_pstate * cpu->pstate.scaling, cpu->cpu);
2533 	intel_pstate_update_pstate(cpu, target_pstate);
2534 
2535 	sample = &cpu->sample;
2536 	trace_pstate_sample(mul_ext_fp(100, sample->core_avg_perf),
2537 		fp_toint(sample->busy_scaled),
2538 		from,
2539 		cpu->pstate.current_pstate,
2540 		sample->mperf,
2541 		sample->aperf,
2542 		sample->tsc,
2543 		get_avg_frequency(cpu),
2544 		fp_toint(cpu->iowait_boost * 100));
2545 }
2546 
2547 static void intel_pstate_update_util(struct update_util_data *data, u64 time,
2548 				     unsigned int flags)
2549 {
2550 	struct cpudata *cpu = container_of(data, struct cpudata, update_util);
2551 	u64 delta_ns;
2552 
2553 	/* Don't allow remote callbacks */
2554 	if (smp_processor_id() != cpu->cpu)
2555 		return;
2556 
2557 	delta_ns = time - cpu->last_update;
2558 	if (flags & SCHED_CPUFREQ_IOWAIT) {
2559 		/* Start over if the CPU may have been idle. */
2560 		if (delta_ns > TICK_NSEC) {
2561 			cpu->iowait_boost = ONE_EIGHTH_FP;
2562 		} else if (cpu->iowait_boost >= ONE_EIGHTH_FP) {
2563 			cpu->iowait_boost <<= 1;
2564 			if (cpu->iowait_boost > int_tofp(1))
2565 				cpu->iowait_boost = int_tofp(1);
2566 		} else {
2567 			cpu->iowait_boost = ONE_EIGHTH_FP;
2568 		}
2569 	} else if (cpu->iowait_boost) {
2570 		/* Clear iowait_boost if the CPU may have been idle. */
2571 		if (delta_ns > TICK_NSEC)
2572 			cpu->iowait_boost = 0;
2573 		else
2574 			cpu->iowait_boost >>= 1;
2575 	}
2576 	cpu->last_update = time;
2577 	delta_ns = time - cpu->sample.time;
2578 	if ((s64)delta_ns < INTEL_PSTATE_SAMPLING_INTERVAL)
2579 		return;
2580 
2581 	if (intel_pstate_sample(cpu, time))
2582 		intel_pstate_adjust_pstate(cpu);
2583 }
2584 
2585 static struct pstate_funcs core_funcs = {
2586 	.get_max = core_get_max_pstate,
2587 	.get_max_physical = core_get_max_pstate_physical,
2588 	.get_min = core_get_min_pstate,
2589 	.get_turbo = core_get_turbo_pstate,
2590 	.get_scaling = core_get_scaling,
2591 	.get_val = core_get_val,
2592 };
2593 
2594 static const struct pstate_funcs silvermont_funcs = {
2595 	.get_max = atom_get_max_pstate,
2596 	.get_max_physical = atom_get_max_pstate,
2597 	.get_min = atom_get_min_pstate,
2598 	.get_turbo = atom_get_turbo_pstate,
2599 	.get_val = atom_get_val,
2600 	.get_scaling = silvermont_get_scaling,
2601 	.get_vid = atom_get_vid,
2602 };
2603 
2604 static const struct pstate_funcs airmont_funcs = {
2605 	.get_max = atom_get_max_pstate,
2606 	.get_max_physical = atom_get_max_pstate,
2607 	.get_min = atom_get_min_pstate,
2608 	.get_turbo = atom_get_turbo_pstate,
2609 	.get_val = atom_get_val,
2610 	.get_scaling = airmont_get_scaling,
2611 	.get_vid = atom_get_vid,
2612 };
2613 
2614 static const struct pstate_funcs knl_funcs = {
2615 	.get_max = core_get_max_pstate,
2616 	.get_max_physical = core_get_max_pstate_physical,
2617 	.get_min = core_get_min_pstate,
2618 	.get_turbo = knl_get_turbo_pstate,
2619 	.get_aperf_mperf_shift = knl_get_aperf_mperf_shift,
2620 	.get_scaling = core_get_scaling,
2621 	.get_val = core_get_val,
2622 };
2623 
2624 #define X86_MATCH(vfm, policy)					 \
2625 	X86_MATCH_VFM_FEATURE(vfm, X86_FEATURE_APERFMPERF, &policy)
2626 
2627 static const struct x86_cpu_id intel_pstate_cpu_ids[] = {
2628 	X86_MATCH(INTEL_SANDYBRIDGE,		core_funcs),
2629 	X86_MATCH(INTEL_SANDYBRIDGE_X,		core_funcs),
2630 	X86_MATCH(INTEL_ATOM_SILVERMONT,	silvermont_funcs),
2631 	X86_MATCH(INTEL_IVYBRIDGE,		core_funcs),
2632 	X86_MATCH(INTEL_HASWELL,		core_funcs),
2633 	X86_MATCH(INTEL_BROADWELL,		core_funcs),
2634 	X86_MATCH(INTEL_IVYBRIDGE_X,		core_funcs),
2635 	X86_MATCH(INTEL_HASWELL_X,		core_funcs),
2636 	X86_MATCH(INTEL_HASWELL_L,		core_funcs),
2637 	X86_MATCH(INTEL_HASWELL_G,		core_funcs),
2638 	X86_MATCH(INTEL_BROADWELL_G,		core_funcs),
2639 	X86_MATCH(INTEL_ATOM_AIRMONT,		airmont_funcs),
2640 	X86_MATCH(INTEL_SKYLAKE_L,		core_funcs),
2641 	X86_MATCH(INTEL_BROADWELL_X,		core_funcs),
2642 	X86_MATCH(INTEL_SKYLAKE,		core_funcs),
2643 	X86_MATCH(INTEL_BROADWELL_D,		core_funcs),
2644 	X86_MATCH(INTEL_XEON_PHI_KNL,		knl_funcs),
2645 	X86_MATCH(INTEL_XEON_PHI_KNM,		knl_funcs),
2646 	X86_MATCH(INTEL_ATOM_GOLDMONT,		core_funcs),
2647 	X86_MATCH(INTEL_ATOM_GOLDMONT_PLUS,	core_funcs),
2648 	X86_MATCH(INTEL_SKYLAKE_X,		core_funcs),
2649 	X86_MATCH(INTEL_COMETLAKE,		core_funcs),
2650 	X86_MATCH(INTEL_ICELAKE_X,		core_funcs),
2651 	X86_MATCH(INTEL_TIGERLAKE,		core_funcs),
2652 	X86_MATCH(INTEL_SAPPHIRERAPIDS_X,	core_funcs),
2653 	X86_MATCH(INTEL_EMERALDRAPIDS_X,	core_funcs),
2654 	{}
2655 };
2656 MODULE_DEVICE_TABLE(x86cpu, intel_pstate_cpu_ids);
2657 
2658 #ifdef CONFIG_ACPI
2659 static const struct x86_cpu_id intel_pstate_cpu_oob_ids[] __initconst = {
2660 	X86_MATCH(INTEL_BROADWELL_D,		core_funcs),
2661 	X86_MATCH(INTEL_BROADWELL_X,		core_funcs),
2662 	X86_MATCH(INTEL_SKYLAKE_X,		core_funcs),
2663 	X86_MATCH(INTEL_ICELAKE_X,		core_funcs),
2664 	X86_MATCH(INTEL_SAPPHIRERAPIDS_X,	core_funcs),
2665 	X86_MATCH(INTEL_EMERALDRAPIDS_X,	core_funcs),
2666 	X86_MATCH(INTEL_GRANITERAPIDS_D,	core_funcs),
2667 	X86_MATCH(INTEL_GRANITERAPIDS_X,	core_funcs),
2668 	X86_MATCH(INTEL_ATOM_CRESTMONT,		core_funcs),
2669 	X86_MATCH(INTEL_ATOM_CRESTMONT_X,	core_funcs),
2670 	{}
2671 };
2672 #endif
2673 
2674 static const struct x86_cpu_id intel_pstate_cpu_ee_disable_ids[] = {
2675 	X86_MATCH(INTEL_KABYLAKE,		core_funcs),
2676 	{}
2677 };
2678 
2679 static int intel_pstate_init_cpu(unsigned int cpunum)
2680 {
2681 	struct cpudata *cpu;
2682 
2683 	cpu = all_cpu_data[cpunum];
2684 
2685 	if (!cpu) {
2686 		cpu = kzalloc(sizeof(*cpu), GFP_KERNEL);
2687 		if (!cpu)
2688 			return -ENOMEM;
2689 
2690 		WRITE_ONCE(all_cpu_data[cpunum], cpu);
2691 
2692 		cpu->cpu = cpunum;
2693 
2694 		cpu->epp_default = -EINVAL;
2695 
2696 		if (hwp_active) {
2697 			intel_pstate_hwp_enable(cpu);
2698 
2699 			if (intel_pstate_acpi_pm_profile_server())
2700 				hwp_boost = true;
2701 		}
2702 	} else if (hwp_active) {
2703 		/*
2704 		 * Re-enable HWP in case this happens after a resume from ACPI
2705 		 * S3 if the CPU was offline during the whole system/resume
2706 		 * cycle.
2707 		 */
2708 		intel_pstate_hwp_reenable(cpu);
2709 	}
2710 
2711 	cpu->epp_powersave = -EINVAL;
2712 	cpu->epp_policy = 0;
2713 
2714 	intel_pstate_get_cpu_pstates(cpu);
2715 
2716 	pr_debug("controlling: cpu %d\n", cpunum);
2717 
2718 	return 0;
2719 }
2720 
2721 static void intel_pstate_set_update_util_hook(unsigned int cpu_num)
2722 {
2723 	struct cpudata *cpu = all_cpu_data[cpu_num];
2724 
2725 	if (hwp_active && !hwp_boost)
2726 		return;
2727 
2728 	if (cpu->update_util_set)
2729 		return;
2730 
2731 	/* Prevent intel_pstate_update_util() from using stale data. */
2732 	cpu->sample.time = 0;
2733 	cpufreq_add_update_util_hook(cpu_num, &cpu->update_util,
2734 				     (hwp_active ?
2735 				      intel_pstate_update_util_hwp :
2736 				      intel_pstate_update_util));
2737 	cpu->update_util_set = true;
2738 }
2739 
2740 static void intel_pstate_clear_update_util_hook(unsigned int cpu)
2741 {
2742 	struct cpudata *cpu_data = all_cpu_data[cpu];
2743 
2744 	if (!cpu_data->update_util_set)
2745 		return;
2746 
2747 	cpufreq_remove_update_util_hook(cpu);
2748 	cpu_data->update_util_set = false;
2749 	synchronize_rcu();
2750 }
2751 
2752 static int intel_pstate_get_max_freq(struct cpudata *cpu)
2753 {
2754 	return READ_ONCE(global.no_turbo) ?
2755 			cpu->pstate.max_freq : cpu->pstate.turbo_freq;
2756 }
2757 
2758 static void intel_pstate_update_perf_limits(struct cpudata *cpu,
2759 					    unsigned int policy_min,
2760 					    unsigned int policy_max)
2761 {
2762 	int perf_ctl_scaling = cpu->pstate.perf_ctl_scaling;
2763 	int32_t max_policy_perf, min_policy_perf;
2764 
2765 	max_policy_perf = policy_max / perf_ctl_scaling;
2766 	if (policy_max == policy_min) {
2767 		min_policy_perf = max_policy_perf;
2768 	} else {
2769 		min_policy_perf = policy_min / perf_ctl_scaling;
2770 		min_policy_perf = clamp_t(int32_t, min_policy_perf,
2771 					  0, max_policy_perf);
2772 	}
2773 
2774 	/*
2775 	 * HWP needs some special consideration, because HWP_REQUEST uses
2776 	 * abstract values to represent performance rather than pure ratios.
2777 	 */
2778 	if (hwp_active && cpu->pstate.scaling != perf_ctl_scaling) {
2779 		int freq;
2780 
2781 		freq = max_policy_perf * perf_ctl_scaling;
2782 		max_policy_perf = intel_pstate_freq_to_hwp(cpu, freq);
2783 		freq = min_policy_perf * perf_ctl_scaling;
2784 		min_policy_perf = intel_pstate_freq_to_hwp(cpu, freq);
2785 	}
2786 
2787 	pr_debug("cpu:%d min_policy_perf:%d max_policy_perf:%d\n",
2788 		 cpu->cpu, min_policy_perf, max_policy_perf);
2789 
2790 	/* Normalize user input to [min_perf, max_perf] */
2791 	if (per_cpu_limits) {
2792 		cpu->min_perf_ratio = min_policy_perf;
2793 		cpu->max_perf_ratio = max_policy_perf;
2794 	} else {
2795 		int turbo_max = cpu->pstate.turbo_pstate;
2796 		int32_t global_min, global_max;
2797 
2798 		/* Global limits are in percent of the maximum turbo P-state. */
2799 		global_max = DIV_ROUND_UP(turbo_max * global.max_perf_pct, 100);
2800 		global_min = DIV_ROUND_UP(turbo_max * global.min_perf_pct, 100);
2801 		global_min = clamp_t(int32_t, global_min, 0, global_max);
2802 
2803 		pr_debug("cpu:%d global_min:%d global_max:%d\n", cpu->cpu,
2804 			 global_min, global_max);
2805 
2806 		cpu->min_perf_ratio = max(min_policy_perf, global_min);
2807 		cpu->min_perf_ratio = min(cpu->min_perf_ratio, max_policy_perf);
2808 		cpu->max_perf_ratio = min(max_policy_perf, global_max);
2809 		cpu->max_perf_ratio = max(min_policy_perf, cpu->max_perf_ratio);
2810 
2811 		/* Make sure min_perf <= max_perf */
2812 		cpu->min_perf_ratio = min(cpu->min_perf_ratio,
2813 					  cpu->max_perf_ratio);
2814 
2815 	}
2816 	pr_debug("cpu:%d max_perf_ratio:%d min_perf_ratio:%d\n", cpu->cpu,
2817 		 cpu->max_perf_ratio,
2818 		 cpu->min_perf_ratio);
2819 }
2820 
2821 static int intel_pstate_set_policy(struct cpufreq_policy *policy)
2822 {
2823 	struct cpudata *cpu;
2824 
2825 	if (!policy->cpuinfo.max_freq)
2826 		return -ENODEV;
2827 
2828 	pr_debug("set_policy cpuinfo.max %u policy->max %u\n",
2829 		 policy->cpuinfo.max_freq, policy->max);
2830 
2831 	cpu = all_cpu_data[policy->cpu];
2832 	cpu->policy = policy->policy;
2833 
2834 	mutex_lock(&intel_pstate_limits_lock);
2835 
2836 	intel_pstate_update_perf_limits(cpu, policy->min, policy->max);
2837 
2838 	if (cpu->policy == CPUFREQ_POLICY_PERFORMANCE) {
2839 		int pstate = max(cpu->pstate.min_pstate, cpu->max_perf_ratio);
2840 
2841 		/*
2842 		 * NOHZ_FULL CPUs need this as the governor callback may not
2843 		 * be invoked on them.
2844 		 */
2845 		intel_pstate_clear_update_util_hook(policy->cpu);
2846 		intel_pstate_set_pstate(cpu, pstate);
2847 	} else {
2848 		intel_pstate_set_update_util_hook(policy->cpu);
2849 	}
2850 
2851 	if (hwp_active) {
2852 		/*
2853 		 * When hwp_boost was active before and dynamically it
2854 		 * was turned off, in that case we need to clear the
2855 		 * update util hook.
2856 		 */
2857 		if (!hwp_boost)
2858 			intel_pstate_clear_update_util_hook(policy->cpu);
2859 		intel_pstate_hwp_set(policy->cpu);
2860 	}
2861 	/*
2862 	 * policy->cur is never updated with the intel_pstate driver, but it
2863 	 * is used as a stale frequency value. So, keep it within limits.
2864 	 */
2865 	policy->cur = policy->min;
2866 
2867 	mutex_unlock(&intel_pstate_limits_lock);
2868 
2869 	return 0;
2870 }
2871 
2872 static void intel_pstate_adjust_policy_max(struct cpudata *cpu,
2873 					   struct cpufreq_policy_data *policy)
2874 {
2875 	if (!hwp_active &&
2876 	    cpu->pstate.max_pstate_physical > cpu->pstate.max_pstate &&
2877 	    policy->max < policy->cpuinfo.max_freq &&
2878 	    policy->max > cpu->pstate.max_freq) {
2879 		pr_debug("policy->max > max non turbo frequency\n");
2880 		policy->max = policy->cpuinfo.max_freq;
2881 	}
2882 }
2883 
2884 static void intel_pstate_verify_cpu_policy(struct cpudata *cpu,
2885 					   struct cpufreq_policy_data *policy)
2886 {
2887 	int max_freq;
2888 
2889 	if (hwp_active) {
2890 		intel_pstate_get_hwp_cap(cpu);
2891 		max_freq = READ_ONCE(global.no_turbo) ?
2892 				cpu->pstate.max_freq : cpu->pstate.turbo_freq;
2893 	} else {
2894 		max_freq = intel_pstate_get_max_freq(cpu);
2895 	}
2896 	cpufreq_verify_within_limits(policy, policy->cpuinfo.min_freq, max_freq);
2897 
2898 	intel_pstate_adjust_policy_max(cpu, policy);
2899 }
2900 
2901 static int intel_pstate_verify_policy(struct cpufreq_policy_data *policy)
2902 {
2903 	intel_pstate_verify_cpu_policy(all_cpu_data[policy->cpu], policy);
2904 
2905 	return 0;
2906 }
2907 
2908 static int intel_cpufreq_cpu_offline(struct cpufreq_policy *policy)
2909 {
2910 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2911 
2912 	pr_debug("CPU %d going offline\n", cpu->cpu);
2913 
2914 	if (cpu->suspended)
2915 		return 0;
2916 
2917 	/*
2918 	 * If the CPU is an SMT thread and it goes offline with the performance
2919 	 * settings different from the minimum, it will prevent its sibling
2920 	 * from getting to lower performance levels, so force the minimum
2921 	 * performance on CPU offline to prevent that from happening.
2922 	 */
2923 	if (hwp_active)
2924 		intel_pstate_hwp_offline(cpu);
2925 	else
2926 		intel_pstate_set_min_pstate(cpu);
2927 
2928 	intel_pstate_exit_perf_limits(policy);
2929 
2930 	return 0;
2931 }
2932 
2933 static int intel_pstate_cpu_online(struct cpufreq_policy *policy)
2934 {
2935 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2936 
2937 	pr_debug("CPU %d going online\n", cpu->cpu);
2938 
2939 	intel_pstate_init_acpi_perf_limits(policy);
2940 
2941 	if (hwp_active) {
2942 		/*
2943 		 * Re-enable HWP and clear the "suspended" flag to let "resume"
2944 		 * know that it need not do that.
2945 		 */
2946 		intel_pstate_hwp_reenable(cpu);
2947 		cpu->suspended = false;
2948 
2949 		hybrid_update_capacity(cpu);
2950 	}
2951 
2952 	return 0;
2953 }
2954 
2955 static int intel_pstate_cpu_offline(struct cpufreq_policy *policy)
2956 {
2957 	intel_pstate_clear_update_util_hook(policy->cpu);
2958 
2959 	return intel_cpufreq_cpu_offline(policy);
2960 }
2961 
2962 static void intel_pstate_cpu_exit(struct cpufreq_policy *policy)
2963 {
2964 	pr_debug("CPU %d exiting\n", policy->cpu);
2965 
2966 	policy->fast_switch_possible = false;
2967 }
2968 
2969 static int __intel_pstate_cpu_init(struct cpufreq_policy *policy)
2970 {
2971 	struct cpudata *cpu;
2972 	int rc;
2973 
2974 	rc = intel_pstate_init_cpu(policy->cpu);
2975 	if (rc)
2976 		return rc;
2977 
2978 	cpu = all_cpu_data[policy->cpu];
2979 
2980 	cpu->max_perf_ratio = 0xFF;
2981 	cpu->min_perf_ratio = 0;
2982 
2983 	/* cpuinfo and default policy values */
2984 	policy->cpuinfo.min_freq = cpu->pstate.min_freq;
2985 	policy->cpuinfo.max_freq = READ_ONCE(global.no_turbo) ?
2986 			cpu->pstate.max_freq : cpu->pstate.turbo_freq;
2987 
2988 	policy->min = policy->cpuinfo.min_freq;
2989 	policy->max = policy->cpuinfo.max_freq;
2990 
2991 	intel_pstate_init_acpi_perf_limits(policy);
2992 
2993 	policy->fast_switch_possible = true;
2994 
2995 	return 0;
2996 }
2997 
2998 static int intel_pstate_cpu_init(struct cpufreq_policy *policy)
2999 {
3000 	int ret = __intel_pstate_cpu_init(policy);
3001 
3002 	if (ret)
3003 		return ret;
3004 
3005 	/*
3006 	 * Set the policy to powersave to provide a valid fallback value in case
3007 	 * the default cpufreq governor is neither powersave nor performance.
3008 	 */
3009 	policy->policy = CPUFREQ_POLICY_POWERSAVE;
3010 
3011 	if (hwp_active) {
3012 		struct cpudata *cpu = all_cpu_data[policy->cpu];
3013 
3014 		cpu->epp_cached = intel_pstate_get_epp(cpu, 0);
3015 	}
3016 
3017 	return 0;
3018 }
3019 
3020 static struct cpufreq_driver intel_pstate = {
3021 	.flags		= CPUFREQ_CONST_LOOPS,
3022 	.verify		= intel_pstate_verify_policy,
3023 	.setpolicy	= intel_pstate_set_policy,
3024 	.suspend	= intel_pstate_suspend,
3025 	.resume		= intel_pstate_resume,
3026 	.init		= intel_pstate_cpu_init,
3027 	.exit		= intel_pstate_cpu_exit,
3028 	.offline	= intel_pstate_cpu_offline,
3029 	.online		= intel_pstate_cpu_online,
3030 	.update_limits	= intel_pstate_update_limits,
3031 	.name		= "intel_pstate",
3032 };
3033 
3034 static int intel_cpufreq_verify_policy(struct cpufreq_policy_data *policy)
3035 {
3036 	struct cpudata *cpu = all_cpu_data[policy->cpu];
3037 
3038 	intel_pstate_verify_cpu_policy(cpu, policy);
3039 	intel_pstate_update_perf_limits(cpu, policy->min, policy->max);
3040 
3041 	return 0;
3042 }
3043 
3044 /* Use of trace in passive mode:
3045  *
3046  * In passive mode the trace core_busy field (also known as the
3047  * performance field, and lablelled as such on the graphs; also known as
3048  * core_avg_perf) is not needed and so is re-assigned to indicate if the
3049  * driver call was via the normal or fast switch path. Various graphs
3050  * output from the intel_pstate_tracer.py utility that include core_busy
3051  * (or performance or core_avg_perf) have a fixed y-axis from 0 to 100%,
3052  * so we use 10 to indicate the normal path through the driver, and
3053  * 90 to indicate the fast switch path through the driver.
3054  * The scaled_busy field is not used, and is set to 0.
3055  */
3056 
3057 #define	INTEL_PSTATE_TRACE_TARGET 10
3058 #define	INTEL_PSTATE_TRACE_FAST_SWITCH 90
3059 
3060 static void intel_cpufreq_trace(struct cpudata *cpu, unsigned int trace_type, int old_pstate)
3061 {
3062 	struct sample *sample;
3063 
3064 	if (!trace_pstate_sample_enabled())
3065 		return;
3066 
3067 	if (!intel_pstate_sample(cpu, ktime_get()))
3068 		return;
3069 
3070 	sample = &cpu->sample;
3071 	trace_pstate_sample(trace_type,
3072 		0,
3073 		old_pstate,
3074 		cpu->pstate.current_pstate,
3075 		sample->mperf,
3076 		sample->aperf,
3077 		sample->tsc,
3078 		get_avg_frequency(cpu),
3079 		fp_toint(cpu->iowait_boost * 100));
3080 }
3081 
3082 static void intel_cpufreq_hwp_update(struct cpudata *cpu, u32 min, u32 max,
3083 				     u32 desired, bool fast_switch)
3084 {
3085 	u64 prev = READ_ONCE(cpu->hwp_req_cached), value = prev;
3086 
3087 	value &= ~HWP_MIN_PERF(~0L);
3088 	value |= HWP_MIN_PERF(min);
3089 
3090 	value &= ~HWP_MAX_PERF(~0L);
3091 	value |= HWP_MAX_PERF(max);
3092 
3093 	value &= ~HWP_DESIRED_PERF(~0L);
3094 	value |= HWP_DESIRED_PERF(desired);
3095 
3096 	if (value == prev)
3097 		return;
3098 
3099 	WRITE_ONCE(cpu->hwp_req_cached, value);
3100 	if (fast_switch)
3101 		wrmsrl(MSR_HWP_REQUEST, value);
3102 	else
3103 		wrmsrl_on_cpu(cpu->cpu, MSR_HWP_REQUEST, value);
3104 }
3105 
3106 static void intel_cpufreq_perf_ctl_update(struct cpudata *cpu,
3107 					  u32 target_pstate, bool fast_switch)
3108 {
3109 	if (fast_switch)
3110 		wrmsrl(MSR_IA32_PERF_CTL,
3111 		       pstate_funcs.get_val(cpu, target_pstate));
3112 	else
3113 		wrmsrl_on_cpu(cpu->cpu, MSR_IA32_PERF_CTL,
3114 			      pstate_funcs.get_val(cpu, target_pstate));
3115 }
3116 
3117 static int intel_cpufreq_update_pstate(struct cpufreq_policy *policy,
3118 				       int target_pstate, bool fast_switch)
3119 {
3120 	struct cpudata *cpu = all_cpu_data[policy->cpu];
3121 	int old_pstate = cpu->pstate.current_pstate;
3122 
3123 	target_pstate = intel_pstate_prepare_request(cpu, target_pstate);
3124 	if (hwp_active) {
3125 		int max_pstate = policy->strict_target ?
3126 					target_pstate : cpu->max_perf_ratio;
3127 
3128 		intel_cpufreq_hwp_update(cpu, target_pstate, max_pstate, 0,
3129 					 fast_switch);
3130 	} else if (target_pstate != old_pstate) {
3131 		intel_cpufreq_perf_ctl_update(cpu, target_pstate, fast_switch);
3132 	}
3133 
3134 	cpu->pstate.current_pstate = target_pstate;
3135 
3136 	intel_cpufreq_trace(cpu, fast_switch ? INTEL_PSTATE_TRACE_FAST_SWITCH :
3137 			    INTEL_PSTATE_TRACE_TARGET, old_pstate);
3138 
3139 	return target_pstate;
3140 }
3141 
3142 static int intel_cpufreq_target(struct cpufreq_policy *policy,
3143 				unsigned int target_freq,
3144 				unsigned int relation)
3145 {
3146 	struct cpudata *cpu = all_cpu_data[policy->cpu];
3147 	struct cpufreq_freqs freqs;
3148 	int target_pstate;
3149 
3150 	freqs.old = policy->cur;
3151 	freqs.new = target_freq;
3152 
3153 	cpufreq_freq_transition_begin(policy, &freqs);
3154 
3155 	target_pstate = intel_pstate_freq_to_hwp_rel(cpu, freqs.new, relation);
3156 	target_pstate = intel_cpufreq_update_pstate(policy, target_pstate, false);
3157 
3158 	freqs.new = target_pstate * cpu->pstate.scaling;
3159 
3160 	cpufreq_freq_transition_end(policy, &freqs, false);
3161 
3162 	return 0;
3163 }
3164 
3165 static unsigned int intel_cpufreq_fast_switch(struct cpufreq_policy *policy,
3166 					      unsigned int target_freq)
3167 {
3168 	struct cpudata *cpu = all_cpu_data[policy->cpu];
3169 	int target_pstate;
3170 
3171 	target_pstate = intel_pstate_freq_to_hwp(cpu, target_freq);
3172 
3173 	target_pstate = intel_cpufreq_update_pstate(policy, target_pstate, true);
3174 
3175 	return target_pstate * cpu->pstate.scaling;
3176 }
3177 
3178 static void intel_cpufreq_adjust_perf(unsigned int cpunum,
3179 				      unsigned long min_perf,
3180 				      unsigned long target_perf,
3181 				      unsigned long capacity)
3182 {
3183 	struct cpudata *cpu = all_cpu_data[cpunum];
3184 	u64 hwp_cap = READ_ONCE(cpu->hwp_cap_cached);
3185 	int old_pstate = cpu->pstate.current_pstate;
3186 	int cap_pstate, min_pstate, max_pstate, target_pstate;
3187 
3188 	cap_pstate = READ_ONCE(global.no_turbo) ?
3189 					HWP_GUARANTEED_PERF(hwp_cap) :
3190 					HWP_HIGHEST_PERF(hwp_cap);
3191 
3192 	/* Optimization: Avoid unnecessary divisions. */
3193 
3194 	target_pstate = cap_pstate;
3195 	if (target_perf < capacity)
3196 		target_pstate = DIV_ROUND_UP(cap_pstate * target_perf, capacity);
3197 
3198 	min_pstate = cap_pstate;
3199 	if (min_perf < capacity)
3200 		min_pstate = DIV_ROUND_UP(cap_pstate * min_perf, capacity);
3201 
3202 	if (min_pstate < cpu->pstate.min_pstate)
3203 		min_pstate = cpu->pstate.min_pstate;
3204 
3205 	if (min_pstate < cpu->min_perf_ratio)
3206 		min_pstate = cpu->min_perf_ratio;
3207 
3208 	if (min_pstate > cpu->max_perf_ratio)
3209 		min_pstate = cpu->max_perf_ratio;
3210 
3211 	max_pstate = min(cap_pstate, cpu->max_perf_ratio);
3212 	if (max_pstate < min_pstate)
3213 		max_pstate = min_pstate;
3214 
3215 	target_pstate = clamp_t(int, target_pstate, min_pstate, max_pstate);
3216 
3217 	intel_cpufreq_hwp_update(cpu, min_pstate, max_pstate, target_pstate, true);
3218 
3219 	cpu->pstate.current_pstate = target_pstate;
3220 	intel_cpufreq_trace(cpu, INTEL_PSTATE_TRACE_FAST_SWITCH, old_pstate);
3221 }
3222 
3223 static int intel_cpufreq_cpu_init(struct cpufreq_policy *policy)
3224 {
3225 	struct freq_qos_request *req;
3226 	struct cpudata *cpu;
3227 	struct device *dev;
3228 	int ret, freq;
3229 
3230 	dev = get_cpu_device(policy->cpu);
3231 	if (!dev)
3232 		return -ENODEV;
3233 
3234 	ret = __intel_pstate_cpu_init(policy);
3235 	if (ret)
3236 		return ret;
3237 
3238 	policy->cpuinfo.transition_latency = INTEL_CPUFREQ_TRANSITION_LATENCY;
3239 	/* This reflects the intel_pstate_get_cpu_pstates() setting. */
3240 	policy->cur = policy->cpuinfo.min_freq;
3241 
3242 	req = kcalloc(2, sizeof(*req), GFP_KERNEL);
3243 	if (!req) {
3244 		ret = -ENOMEM;
3245 		goto pstate_exit;
3246 	}
3247 
3248 	cpu = all_cpu_data[policy->cpu];
3249 
3250 	if (hwp_active) {
3251 		u64 value;
3252 
3253 		policy->transition_delay_us = INTEL_CPUFREQ_TRANSITION_DELAY_HWP;
3254 
3255 		intel_pstate_get_hwp_cap(cpu);
3256 
3257 		rdmsrl_on_cpu(cpu->cpu, MSR_HWP_REQUEST, &value);
3258 		WRITE_ONCE(cpu->hwp_req_cached, value);
3259 
3260 		cpu->epp_cached = intel_pstate_get_epp(cpu, value);
3261 	} else {
3262 		policy->transition_delay_us = INTEL_CPUFREQ_TRANSITION_DELAY;
3263 	}
3264 
3265 	freq = DIV_ROUND_UP(cpu->pstate.turbo_freq * global.min_perf_pct, 100);
3266 
3267 	ret = freq_qos_add_request(&policy->constraints, req, FREQ_QOS_MIN,
3268 				   freq);
3269 	if (ret < 0) {
3270 		dev_err(dev, "Failed to add min-freq constraint (%d)\n", ret);
3271 		goto free_req;
3272 	}
3273 
3274 	freq = DIV_ROUND_UP(cpu->pstate.turbo_freq * global.max_perf_pct, 100);
3275 
3276 	ret = freq_qos_add_request(&policy->constraints, req + 1, FREQ_QOS_MAX,
3277 				   freq);
3278 	if (ret < 0) {
3279 		dev_err(dev, "Failed to add max-freq constraint (%d)\n", ret);
3280 		goto remove_min_req;
3281 	}
3282 
3283 	policy->driver_data = req;
3284 
3285 	return 0;
3286 
3287 remove_min_req:
3288 	freq_qos_remove_request(req);
3289 free_req:
3290 	kfree(req);
3291 pstate_exit:
3292 	intel_pstate_exit_perf_limits(policy);
3293 
3294 	return ret;
3295 }
3296 
3297 static void intel_cpufreq_cpu_exit(struct cpufreq_policy *policy)
3298 {
3299 	struct freq_qos_request *req;
3300 
3301 	req = policy->driver_data;
3302 
3303 	freq_qos_remove_request(req + 1);
3304 	freq_qos_remove_request(req);
3305 	kfree(req);
3306 
3307 	intel_pstate_cpu_exit(policy);
3308 }
3309 
3310 static int intel_cpufreq_suspend(struct cpufreq_policy *policy)
3311 {
3312 	intel_pstate_suspend(policy);
3313 
3314 	if (hwp_active) {
3315 		struct cpudata *cpu = all_cpu_data[policy->cpu];
3316 		u64 value = READ_ONCE(cpu->hwp_req_cached);
3317 
3318 		/*
3319 		 * Clear the desired perf field in MSR_HWP_REQUEST in case
3320 		 * intel_cpufreq_adjust_perf() is in use and the last value
3321 		 * written by it may not be suitable.
3322 		 */
3323 		value &= ~HWP_DESIRED_PERF(~0L);
3324 		wrmsrl_on_cpu(cpu->cpu, MSR_HWP_REQUEST, value);
3325 		WRITE_ONCE(cpu->hwp_req_cached, value);
3326 	}
3327 
3328 	return 0;
3329 }
3330 
3331 static struct cpufreq_driver intel_cpufreq = {
3332 	.flags		= CPUFREQ_CONST_LOOPS,
3333 	.verify		= intel_cpufreq_verify_policy,
3334 	.target		= intel_cpufreq_target,
3335 	.fast_switch	= intel_cpufreq_fast_switch,
3336 	.init		= intel_cpufreq_cpu_init,
3337 	.exit		= intel_cpufreq_cpu_exit,
3338 	.offline	= intel_cpufreq_cpu_offline,
3339 	.online		= intel_pstate_cpu_online,
3340 	.suspend	= intel_cpufreq_suspend,
3341 	.resume		= intel_pstate_resume,
3342 	.update_limits	= intel_pstate_update_limits,
3343 	.name		= "intel_cpufreq",
3344 };
3345 
3346 static struct cpufreq_driver *default_driver;
3347 
3348 static void intel_pstate_driver_cleanup(void)
3349 {
3350 	unsigned int cpu;
3351 
3352 	cpus_read_lock();
3353 	for_each_online_cpu(cpu) {
3354 		if (all_cpu_data[cpu]) {
3355 			if (intel_pstate_driver == &intel_pstate)
3356 				intel_pstate_clear_update_util_hook(cpu);
3357 
3358 			kfree(all_cpu_data[cpu]);
3359 			WRITE_ONCE(all_cpu_data[cpu], NULL);
3360 		}
3361 	}
3362 	cpus_read_unlock();
3363 
3364 	intel_pstate_driver = NULL;
3365 }
3366 
3367 static int intel_pstate_register_driver(struct cpufreq_driver *driver)
3368 {
3369 	bool refresh_cpu_cap_scaling;
3370 	int ret;
3371 
3372 	if (driver == &intel_pstate)
3373 		intel_pstate_sysfs_expose_hwp_dynamic_boost();
3374 
3375 	memset(&global, 0, sizeof(global));
3376 	global.max_perf_pct = 100;
3377 	global.turbo_disabled = turbo_is_disabled();
3378 	global.no_turbo = global.turbo_disabled;
3379 
3380 	arch_set_max_freq_ratio(global.turbo_disabled);
3381 
3382 	refresh_cpu_cap_scaling = hybrid_clear_max_perf_cpu();
3383 
3384 	intel_pstate_driver = driver;
3385 	ret = cpufreq_register_driver(intel_pstate_driver);
3386 	if (ret) {
3387 		intel_pstate_driver_cleanup();
3388 		return ret;
3389 	}
3390 
3391 	global.min_perf_pct = min_perf_pct_min();
3392 
3393 	hybrid_init_cpu_capacity_scaling(refresh_cpu_cap_scaling);
3394 
3395 	return 0;
3396 }
3397 
3398 static ssize_t intel_pstate_show_status(char *buf)
3399 {
3400 	if (!intel_pstate_driver)
3401 		return sprintf(buf, "off\n");
3402 
3403 	return sprintf(buf, "%s\n", intel_pstate_driver == &intel_pstate ?
3404 					"active" : "passive");
3405 }
3406 
3407 static int intel_pstate_update_status(const char *buf, size_t size)
3408 {
3409 	if (size == 3 && !strncmp(buf, "off", size)) {
3410 		if (!intel_pstate_driver)
3411 			return -EINVAL;
3412 
3413 		if (hwp_active)
3414 			return -EBUSY;
3415 
3416 		cpufreq_unregister_driver(intel_pstate_driver);
3417 		intel_pstate_driver_cleanup();
3418 		return 0;
3419 	}
3420 
3421 	if (size == 6 && !strncmp(buf, "active", size)) {
3422 		if (intel_pstate_driver) {
3423 			if (intel_pstate_driver == &intel_pstate)
3424 				return 0;
3425 
3426 			cpufreq_unregister_driver(intel_pstate_driver);
3427 		}
3428 
3429 		return intel_pstate_register_driver(&intel_pstate);
3430 	}
3431 
3432 	if (size == 7 && !strncmp(buf, "passive", size)) {
3433 		if (intel_pstate_driver) {
3434 			if (intel_pstate_driver == &intel_cpufreq)
3435 				return 0;
3436 
3437 			cpufreq_unregister_driver(intel_pstate_driver);
3438 			intel_pstate_sysfs_hide_hwp_dynamic_boost();
3439 		}
3440 
3441 		return intel_pstate_register_driver(&intel_cpufreq);
3442 	}
3443 
3444 	return -EINVAL;
3445 }
3446 
3447 static int no_load __initdata;
3448 static int no_hwp __initdata;
3449 static int hwp_only __initdata;
3450 static unsigned int force_load __initdata;
3451 
3452 static int __init intel_pstate_msrs_not_valid(void)
3453 {
3454 	if (!pstate_funcs.get_max(0) ||
3455 	    !pstate_funcs.get_min(0) ||
3456 	    !pstate_funcs.get_turbo(0))
3457 		return -ENODEV;
3458 
3459 	return 0;
3460 }
3461 
3462 static void __init copy_cpu_funcs(struct pstate_funcs *funcs)
3463 {
3464 	pstate_funcs.get_max   = funcs->get_max;
3465 	pstate_funcs.get_max_physical = funcs->get_max_physical;
3466 	pstate_funcs.get_min   = funcs->get_min;
3467 	pstate_funcs.get_turbo = funcs->get_turbo;
3468 	pstate_funcs.get_scaling = funcs->get_scaling;
3469 	pstate_funcs.get_val   = funcs->get_val;
3470 	pstate_funcs.get_vid   = funcs->get_vid;
3471 	pstate_funcs.get_aperf_mperf_shift = funcs->get_aperf_mperf_shift;
3472 }
3473 
3474 #ifdef CONFIG_ACPI
3475 
3476 static bool __init intel_pstate_no_acpi_pss(void)
3477 {
3478 	int i;
3479 
3480 	for_each_possible_cpu(i) {
3481 		acpi_status status;
3482 		union acpi_object *pss;
3483 		struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
3484 		struct acpi_processor *pr = per_cpu(processors, i);
3485 
3486 		if (!pr)
3487 			continue;
3488 
3489 		status = acpi_evaluate_object(pr->handle, "_PSS", NULL, &buffer);
3490 		if (ACPI_FAILURE(status))
3491 			continue;
3492 
3493 		pss = buffer.pointer;
3494 		if (pss && pss->type == ACPI_TYPE_PACKAGE) {
3495 			kfree(pss);
3496 			return false;
3497 		}
3498 
3499 		kfree(pss);
3500 	}
3501 
3502 	pr_debug("ACPI _PSS not found\n");
3503 	return true;
3504 }
3505 
3506 static bool __init intel_pstate_no_acpi_pcch(void)
3507 {
3508 	acpi_status status;
3509 	acpi_handle handle;
3510 
3511 	status = acpi_get_handle(NULL, "\\_SB", &handle);
3512 	if (ACPI_FAILURE(status))
3513 		goto not_found;
3514 
3515 	if (acpi_has_method(handle, "PCCH"))
3516 		return false;
3517 
3518 not_found:
3519 	pr_debug("ACPI PCCH not found\n");
3520 	return true;
3521 }
3522 
3523 static bool __init intel_pstate_has_acpi_ppc(void)
3524 {
3525 	int i;
3526 
3527 	for_each_possible_cpu(i) {
3528 		struct acpi_processor *pr = per_cpu(processors, i);
3529 
3530 		if (!pr)
3531 			continue;
3532 		if (acpi_has_method(pr->handle, "_PPC"))
3533 			return true;
3534 	}
3535 	pr_debug("ACPI _PPC not found\n");
3536 	return false;
3537 }
3538 
3539 enum {
3540 	PSS,
3541 	PPC,
3542 };
3543 
3544 /* Hardware vendor-specific info that has its own power management modes */
3545 static struct acpi_platform_list plat_info[] __initdata = {
3546 	{"HP    ", "ProLiant", 0, ACPI_SIG_FADT, all_versions, NULL, PSS},
3547 	{"ORACLE", "X4-2    ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3548 	{"ORACLE", "X4-2L   ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3549 	{"ORACLE", "X4-2B   ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3550 	{"ORACLE", "X3-2    ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3551 	{"ORACLE", "X3-2L   ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3552 	{"ORACLE", "X3-2B   ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3553 	{"ORACLE", "X4470M2 ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3554 	{"ORACLE", "X4270M3 ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3555 	{"ORACLE", "X4270M2 ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3556 	{"ORACLE", "X4170M2 ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3557 	{"ORACLE", "X4170 M3", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3558 	{"ORACLE", "X4275 M3", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3559 	{"ORACLE", "X6-2    ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3560 	{"ORACLE", "Sudbury ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3561 	{ } /* End */
3562 };
3563 
3564 #define BITMASK_OOB	(BIT(8) | BIT(18))
3565 
3566 static bool __init intel_pstate_platform_pwr_mgmt_exists(void)
3567 {
3568 	const struct x86_cpu_id *id;
3569 	u64 misc_pwr;
3570 	int idx;
3571 
3572 	id = x86_match_cpu(intel_pstate_cpu_oob_ids);
3573 	if (id) {
3574 		rdmsrl(MSR_MISC_PWR_MGMT, misc_pwr);
3575 		if (misc_pwr & BITMASK_OOB) {
3576 			pr_debug("Bit 8 or 18 in the MISC_PWR_MGMT MSR set\n");
3577 			pr_debug("P states are controlled in Out of Band mode by the firmware/hardware\n");
3578 			return true;
3579 		}
3580 	}
3581 
3582 	idx = acpi_match_platform_list(plat_info);
3583 	if (idx < 0)
3584 		return false;
3585 
3586 	switch (plat_info[idx].data) {
3587 	case PSS:
3588 		if (!intel_pstate_no_acpi_pss())
3589 			return false;
3590 
3591 		return intel_pstate_no_acpi_pcch();
3592 	case PPC:
3593 		return intel_pstate_has_acpi_ppc() && !force_load;
3594 	}
3595 
3596 	return false;
3597 }
3598 
3599 static void intel_pstate_request_control_from_smm(void)
3600 {
3601 	/*
3602 	 * It may be unsafe to request P-states control from SMM if _PPC support
3603 	 * has not been enabled.
3604 	 */
3605 	if (acpi_ppc)
3606 		acpi_processor_pstate_control();
3607 }
3608 #else /* CONFIG_ACPI not enabled */
3609 static inline bool intel_pstate_platform_pwr_mgmt_exists(void) { return false; }
3610 static inline bool intel_pstate_has_acpi_ppc(void) { return false; }
3611 static inline void intel_pstate_request_control_from_smm(void) {}
3612 #endif /* CONFIG_ACPI */
3613 
3614 #define INTEL_PSTATE_HWP_BROADWELL	0x01
3615 
3616 #define X86_MATCH_HWP(vfm, hwp_mode)				\
3617 	X86_MATCH_VFM_FEATURE(vfm, X86_FEATURE_HWP, hwp_mode)
3618 
3619 static const struct x86_cpu_id hwp_support_ids[] __initconst = {
3620 	X86_MATCH_HWP(INTEL_BROADWELL_X,	INTEL_PSTATE_HWP_BROADWELL),
3621 	X86_MATCH_HWP(INTEL_BROADWELL_D,	INTEL_PSTATE_HWP_BROADWELL),
3622 	X86_MATCH_HWP(INTEL_ANY,		0),
3623 	{}
3624 };
3625 
3626 static bool intel_pstate_hwp_is_enabled(void)
3627 {
3628 	u64 value;
3629 
3630 	rdmsrl(MSR_PM_ENABLE, value);
3631 	return !!(value & 0x1);
3632 }
3633 
3634 #define POWERSAVE_MASK			GENMASK(7, 0)
3635 #define BALANCE_POWER_MASK		GENMASK(15, 8)
3636 #define BALANCE_PERFORMANCE_MASK	GENMASK(23, 16)
3637 #define PERFORMANCE_MASK		GENMASK(31, 24)
3638 
3639 #define HWP_SET_EPP_VALUES(powersave, balance_power, balance_perf, performance) \
3640 	(FIELD_PREP_CONST(POWERSAVE_MASK, powersave) |\
3641 	 FIELD_PREP_CONST(BALANCE_POWER_MASK, balance_power) |\
3642 	 FIELD_PREP_CONST(BALANCE_PERFORMANCE_MASK, balance_perf) |\
3643 	 FIELD_PREP_CONST(PERFORMANCE_MASK, performance))
3644 
3645 #define HWP_SET_DEF_BALANCE_PERF_EPP(balance_perf) \
3646 	(HWP_SET_EPP_VALUES(HWP_EPP_POWERSAVE, HWP_EPP_BALANCE_POWERSAVE,\
3647 	 balance_perf, HWP_EPP_PERFORMANCE))
3648 
3649 static const struct x86_cpu_id intel_epp_default[] = {
3650 	/*
3651 	 * Set EPP value as 102, this is the max suggested EPP
3652 	 * which can result in one core turbo frequency for
3653 	 * AlderLake Mobile CPUs.
3654 	 */
3655 	X86_MATCH_VFM(INTEL_ALDERLAKE_L, HWP_SET_DEF_BALANCE_PERF_EPP(102)),
3656 	X86_MATCH_VFM(INTEL_SAPPHIRERAPIDS_X, HWP_SET_DEF_BALANCE_PERF_EPP(32)),
3657 	X86_MATCH_VFM(INTEL_EMERALDRAPIDS_X, HWP_SET_DEF_BALANCE_PERF_EPP(32)),
3658 	X86_MATCH_VFM(INTEL_GRANITERAPIDS_X, HWP_SET_DEF_BALANCE_PERF_EPP(32)),
3659 	X86_MATCH_VFM(INTEL_GRANITERAPIDS_D, HWP_SET_DEF_BALANCE_PERF_EPP(32)),
3660 	X86_MATCH_VFM(INTEL_METEORLAKE_L, HWP_SET_EPP_VALUES(HWP_EPP_POWERSAVE,
3661 		      179, 64, 16)),
3662 	X86_MATCH_VFM(INTEL_ARROWLAKE, HWP_SET_EPP_VALUES(HWP_EPP_POWERSAVE,
3663 		      179, 64, 16)),
3664 	{}
3665 };
3666 
3667 static const struct x86_cpu_id intel_hybrid_scaling_factor[] = {
3668 	X86_MATCH_VFM(INTEL_METEORLAKE_L, HYBRID_SCALING_FACTOR_MTL),
3669 	X86_MATCH_VFM(INTEL_ARROWLAKE, HYBRID_SCALING_FACTOR_MTL),
3670 	X86_MATCH_VFM(INTEL_LUNARLAKE_M, HYBRID_SCALING_FACTOR_LNL),
3671 	{}
3672 };
3673 
3674 static int __init intel_pstate_init(void)
3675 {
3676 	static struct cpudata **_all_cpu_data;
3677 	const struct x86_cpu_id *id;
3678 	int rc;
3679 
3680 	if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL)
3681 		return -ENODEV;
3682 
3683 	id = x86_match_cpu(hwp_support_ids);
3684 	if (id) {
3685 		hwp_forced = intel_pstate_hwp_is_enabled();
3686 
3687 		if (hwp_forced)
3688 			pr_info("HWP enabled by BIOS\n");
3689 		else if (no_load)
3690 			return -ENODEV;
3691 
3692 		copy_cpu_funcs(&core_funcs);
3693 		/*
3694 		 * Avoid enabling HWP for processors without EPP support,
3695 		 * because that means incomplete HWP implementation which is a
3696 		 * corner case and supporting it is generally problematic.
3697 		 *
3698 		 * If HWP is enabled already, though, there is no choice but to
3699 		 * deal with it.
3700 		 */
3701 		if ((!no_hwp && boot_cpu_has(X86_FEATURE_HWP_EPP)) || hwp_forced) {
3702 			hwp_active = true;
3703 			hwp_mode_bdw = id->driver_data;
3704 			intel_pstate.attr = hwp_cpufreq_attrs;
3705 			intel_cpufreq.attr = hwp_cpufreq_attrs;
3706 			intel_cpufreq.flags |= CPUFREQ_NEED_UPDATE_LIMITS;
3707 			intel_cpufreq.adjust_perf = intel_cpufreq_adjust_perf;
3708 			if (!default_driver)
3709 				default_driver = &intel_pstate;
3710 
3711 			pstate_funcs.get_cpu_scaling = hwp_get_cpu_scaling;
3712 
3713 			goto hwp_cpu_matched;
3714 		}
3715 		pr_info("HWP not enabled\n");
3716 	} else {
3717 		if (no_load)
3718 			return -ENODEV;
3719 
3720 		id = x86_match_cpu(intel_pstate_cpu_ids);
3721 		if (!id) {
3722 			pr_info("CPU model not supported\n");
3723 			return -ENODEV;
3724 		}
3725 
3726 		copy_cpu_funcs((struct pstate_funcs *)id->driver_data);
3727 	}
3728 
3729 	if (intel_pstate_msrs_not_valid()) {
3730 		pr_info("Invalid MSRs\n");
3731 		return -ENODEV;
3732 	}
3733 	/* Without HWP start in the passive mode. */
3734 	if (!default_driver)
3735 		default_driver = &intel_cpufreq;
3736 
3737 hwp_cpu_matched:
3738 	/*
3739 	 * The Intel pstate driver will be ignored if the platform
3740 	 * firmware has its own power management modes.
3741 	 */
3742 	if (intel_pstate_platform_pwr_mgmt_exists()) {
3743 		pr_info("P-states controlled by the platform\n");
3744 		return -ENODEV;
3745 	}
3746 
3747 	if (!hwp_active && hwp_only)
3748 		return -ENOTSUPP;
3749 
3750 	pr_info("Intel P-state driver initializing\n");
3751 
3752 	_all_cpu_data = vzalloc(array_size(sizeof(void *), num_possible_cpus()));
3753 	if (!_all_cpu_data)
3754 		return -ENOMEM;
3755 
3756 	WRITE_ONCE(all_cpu_data, _all_cpu_data);
3757 
3758 	intel_pstate_request_control_from_smm();
3759 
3760 	intel_pstate_sysfs_expose_params();
3761 
3762 	if (hwp_active) {
3763 		const struct x86_cpu_id *id = x86_match_cpu(intel_epp_default);
3764 		const struct x86_cpu_id *hybrid_id = x86_match_cpu(intel_hybrid_scaling_factor);
3765 
3766 		if (id) {
3767 			epp_values[EPP_INDEX_POWERSAVE] =
3768 					FIELD_GET(POWERSAVE_MASK, id->driver_data);
3769 			epp_values[EPP_INDEX_BALANCE_POWERSAVE] =
3770 					FIELD_GET(BALANCE_POWER_MASK, id->driver_data);
3771 			epp_values[EPP_INDEX_BALANCE_PERFORMANCE] =
3772 					FIELD_GET(BALANCE_PERFORMANCE_MASK, id->driver_data);
3773 			epp_values[EPP_INDEX_PERFORMANCE] =
3774 					FIELD_GET(PERFORMANCE_MASK, id->driver_data);
3775 			pr_debug("Updated EPPs powersave:%x balanced power:%x balanced perf:%x performance:%x\n",
3776 				 epp_values[EPP_INDEX_POWERSAVE],
3777 				 epp_values[EPP_INDEX_BALANCE_POWERSAVE],
3778 				 epp_values[EPP_INDEX_BALANCE_PERFORMANCE],
3779 				 epp_values[EPP_INDEX_PERFORMANCE]);
3780 		}
3781 
3782 		if (hybrid_id) {
3783 			hybrid_scaling_factor = hybrid_id->driver_data;
3784 			pr_debug("hybrid scaling factor: %d\n", hybrid_scaling_factor);
3785 		}
3786 
3787 	}
3788 
3789 	mutex_lock(&intel_pstate_driver_lock);
3790 	rc = intel_pstate_register_driver(default_driver);
3791 	mutex_unlock(&intel_pstate_driver_lock);
3792 	if (rc) {
3793 		intel_pstate_sysfs_remove();
3794 		return rc;
3795 	}
3796 
3797 	if (hwp_active) {
3798 		const struct x86_cpu_id *id;
3799 
3800 		id = x86_match_cpu(intel_pstate_cpu_ee_disable_ids);
3801 		if (id) {
3802 			set_power_ctl_ee_state(false);
3803 			pr_info("Disabling energy efficiency optimization\n");
3804 		}
3805 
3806 		pr_info("HWP enabled\n");
3807 	} else if (boot_cpu_has(X86_FEATURE_HYBRID_CPU)) {
3808 		pr_warn("Problematic setup: Hybrid processor with disabled HWP\n");
3809 	}
3810 
3811 	return 0;
3812 }
3813 device_initcall(intel_pstate_init);
3814 
3815 static int __init intel_pstate_setup(char *str)
3816 {
3817 	if (!str)
3818 		return -EINVAL;
3819 
3820 	if (!strcmp(str, "disable"))
3821 		no_load = 1;
3822 	else if (!strcmp(str, "active"))
3823 		default_driver = &intel_pstate;
3824 	else if (!strcmp(str, "passive"))
3825 		default_driver = &intel_cpufreq;
3826 
3827 	if (!strcmp(str, "no_hwp"))
3828 		no_hwp = 1;
3829 
3830 	if (!strcmp(str, "force"))
3831 		force_load = 1;
3832 	if (!strcmp(str, "hwp_only"))
3833 		hwp_only = 1;
3834 	if (!strcmp(str, "per_cpu_perf_limits"))
3835 		per_cpu_limits = true;
3836 
3837 #ifdef CONFIG_ACPI
3838 	if (!strcmp(str, "support_acpi_ppc"))
3839 		acpi_ppc = true;
3840 #endif
3841 
3842 	return 0;
3843 }
3844 early_param("intel_pstate", intel_pstate_setup);
3845 
3846 MODULE_AUTHOR("Dirk Brandewie <dirk.j.brandewie@intel.com>");
3847 MODULE_DESCRIPTION("'intel_pstate' - P state driver Intel Core processors");
3848