xref: /linux/drivers/cpufreq/intel_pstate.c (revision 7fc2cd2e4b398c57c9cf961cfea05eadbf34c05c)
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 #include <linux/units.h>
32 
33 #include <asm/cpu.h>
34 #include <asm/div64.h>
35 #include <asm/msr.h>
36 #include <asm/cpu_device_id.h>
37 #include <asm/cpufeature.h>
38 #include <asm/intel-family.h>
39 #include "../drivers/thermal/intel/thermal_interrupt.h"
40 
41 #define INTEL_PSTATE_SAMPLING_INTERVAL	(10 * NSEC_PER_MSEC)
42 
43 #define INTEL_CPUFREQ_TRANSITION_LATENCY	20000
44 #define INTEL_CPUFREQ_TRANSITION_DELAY_HWP	5000
45 #define INTEL_CPUFREQ_TRANSITION_DELAY		500
46 
47 #ifdef CONFIG_ACPI
48 #include <acpi/processor.h>
49 #include <acpi/cppc_acpi.h>
50 #endif
51 
52 #define FRAC_BITS 8
53 #define int_tofp(X) ((int64_t)(X) << FRAC_BITS)
54 #define fp_toint(X) ((X) >> FRAC_BITS)
55 
56 #define ONE_EIGHTH_FP ((int64_t)1 << (FRAC_BITS - 3))
57 
58 #define EXT_BITS 6
59 #define EXT_FRAC_BITS (EXT_BITS + FRAC_BITS)
60 #define fp_ext_toint(X) ((X) >> EXT_FRAC_BITS)
61 #define int_ext_tofp(X) ((int64_t)(X) << EXT_FRAC_BITS)
62 
63 static inline int32_t mul_fp(int32_t x, int32_t y)
64 {
65 	return ((int64_t)x * (int64_t)y) >> FRAC_BITS;
66 }
67 
68 static inline int32_t div_fp(s64 x, s64 y)
69 {
70 	return div64_s64((int64_t)x << FRAC_BITS, y);
71 }
72 
73 static inline int ceiling_fp(int32_t x)
74 {
75 	int mask, ret;
76 
77 	ret = fp_toint(x);
78 	mask = (1 << FRAC_BITS) - 1;
79 	if (x & mask)
80 		ret += 1;
81 	return ret;
82 }
83 
84 static inline u64 mul_ext_fp(u64 x, u64 y)
85 {
86 	return (x * y) >> EXT_FRAC_BITS;
87 }
88 
89 static inline u64 div_ext_fp(u64 x, u64 y)
90 {
91 	return div64_u64(x << EXT_FRAC_BITS, y);
92 }
93 
94 /**
95  * struct sample -	Store performance sample
96  * @core_avg_perf:	Ratio of APERF/MPERF which is the actual average
97  *			performance during last sample period
98  * @busy_scaled:	Scaled busy value which is used to calculate next
99  *			P state. This can be different than core_avg_perf
100  *			to account for cpu idle period
101  * @aperf:		Difference of actual performance frequency clock count
102  *			read from APERF MSR between last and current sample
103  * @mperf:		Difference of maximum performance frequency clock count
104  *			read from MPERF MSR between last and current sample
105  * @tsc:		Difference of time stamp counter between last and
106  *			current sample
107  * @time:		Current time from scheduler
108  *
109  * This structure is used in the cpudata structure to store performance sample
110  * data for choosing next P State.
111  */
112 struct sample {
113 	int32_t core_avg_perf;
114 	int32_t busy_scaled;
115 	u64 aperf;
116 	u64 mperf;
117 	u64 tsc;
118 	u64 time;
119 };
120 
121 /**
122  * struct pstate_data - Store P state data
123  * @current_pstate:	Current requested P state
124  * @min_pstate:		Min P state possible for this platform
125  * @max_pstate:		Max P state possible for this platform
126  * @max_pstate_physical:This is physical Max P state for a processor
127  *			This can be higher than the max_pstate which can
128  *			be limited by platform thermal design power limits
129  * @perf_ctl_scaling:	PERF_CTL P-state to frequency scaling factor
130  * @scaling:		Scaling factor between performance and frequency
131  * @turbo_pstate:	Max Turbo P state possible for this platform
132  * @min_freq:		@min_pstate frequency in cpufreq units
133  * @max_freq:		@max_pstate frequency in cpufreq units
134  * @turbo_freq:		@turbo_pstate frequency in cpufreq units
135  *
136  * Stores the per cpu model P state limits and current P state.
137  */
138 struct pstate_data {
139 	int	current_pstate;
140 	int	min_pstate;
141 	int	max_pstate;
142 	int	max_pstate_physical;
143 	int	perf_ctl_scaling;
144 	int	scaling;
145 	int	turbo_pstate;
146 	unsigned int min_freq;
147 	unsigned int max_freq;
148 	unsigned int turbo_freq;
149 };
150 
151 /**
152  * struct vid_data -	Stores voltage information data
153  * @min:		VID data for this platform corresponding to
154  *			the lowest P state
155  * @max:		VID data corresponding to the highest P State.
156  * @turbo:		VID data for turbo P state
157  * @ratio:		Ratio of (vid max - vid min) /
158  *			(max P state - Min P State)
159  *
160  * Stores the voltage data for DVFS (Dynamic Voltage and Frequency Scaling)
161  * This data is used in Atom platforms, where in addition to target P state,
162  * the voltage data needs to be specified to select next P State.
163  */
164 struct vid_data {
165 	int min;
166 	int max;
167 	int turbo;
168 	int32_t ratio;
169 };
170 
171 /**
172  * struct global_params - Global parameters, mostly tunable via sysfs.
173  * @no_turbo:		Whether or not to use turbo P-states.
174  * @turbo_disabled:	Whether or not turbo P-states are available at all,
175  *			based on the MSR_IA32_MISC_ENABLE value and whether or
176  *			not the maximum reported turbo P-state is different from
177  *			the maximum reported non-turbo one.
178  * @min_perf_pct:	Minimum capacity limit in percent of the maximum turbo
179  *			P-state capacity.
180  * @max_perf_pct:	Maximum capacity limit in percent of the maximum turbo
181  *			P-state capacity.
182  */
183 struct global_params {
184 	bool no_turbo;
185 	bool turbo_disabled;
186 	int max_perf_pct;
187 	int min_perf_pct;
188 };
189 
190 /**
191  * struct cpudata -	Per CPU instance data storage
192  * @cpu:		CPU number for this instance data
193  * @policy:		CPUFreq policy value
194  * @update_util:	CPUFreq utility callback information
195  * @update_util_set:	CPUFreq utility callback is set
196  * @iowait_boost:	iowait-related boost fraction
197  * @last_update:	Time of the last update.
198  * @pstate:		Stores P state limits for this CPU
199  * @vid:		Stores VID limits for this CPU
200  * @last_sample_time:	Last Sample time
201  * @aperf_mperf_shift:	APERF vs MPERF counting frequency difference
202  * @prev_aperf:		Last APERF value read from APERF MSR
203  * @prev_mperf:		Last MPERF value read from MPERF MSR
204  * @prev_tsc:		Last timestamp counter (TSC) value
205  * @sample:		Storage for storing last Sample data
206  * @min_perf_ratio:	Minimum capacity in terms of PERF or HWP ratios
207  * @max_perf_ratio:	Maximum capacity in terms of PERF or HWP ratios
208  * @acpi_perf_data:	Stores ACPI perf information read from _PSS
209  * @valid_pss_table:	Set to true for valid ACPI _PSS entries found
210  * @epp_powersave:	Last saved HWP energy performance preference
211  *			(EPP) or energy performance bias (EPB),
212  *			when policy switched to performance
213  * @epp_policy:		Last saved policy used to set EPP/EPB
214  * @epp_default:	Power on default HWP energy performance
215  *			preference/bias
216  * @epp_cached:		Cached HWP energy-performance preference value
217  * @hwp_req_cached:	Cached value of the last HWP Request MSR
218  * @hwp_cap_cached:	Cached value of the last HWP Capabilities MSR
219  * @last_io_update:	Last time when IO wake flag was set
220  * @capacity_perf:	Highest perf used for scale invariance
221  * @sched_flags:	Store scheduler flags for possible cross CPU update
222  * @hwp_boost_min:	Last HWP boosted min performance
223  * @suspended:		Whether or not the driver has been suspended.
224  * @pd_registered:	Set when a perf domain is registered for this CPU.
225  * @hwp_notify_work:	workqueue for HWP notifications.
226  *
227  * This structure stores per CPU instance data for all CPUs.
228  */
229 struct cpudata {
230 	int cpu;
231 
232 	unsigned int policy;
233 	struct update_util_data update_util;
234 	bool   update_util_set;
235 
236 	struct pstate_data pstate;
237 	struct vid_data vid;
238 
239 	u64	last_update;
240 	u64	last_sample_time;
241 	u64	aperf_mperf_shift;
242 	u64	prev_aperf;
243 	u64	prev_mperf;
244 	u64	prev_tsc;
245 	struct sample sample;
246 	int32_t	min_perf_ratio;
247 	int32_t	max_perf_ratio;
248 #ifdef CONFIG_ACPI
249 	struct acpi_processor_performance acpi_perf_data;
250 	bool valid_pss_table;
251 #endif
252 	unsigned int iowait_boost;
253 	s16 epp_powersave;
254 	s16 epp_policy;
255 	s16 epp_default;
256 	s16 epp_cached;
257 	u64 hwp_req_cached;
258 	u64 hwp_cap_cached;
259 	u64 last_io_update;
260 	unsigned int capacity_perf;
261 	unsigned int sched_flags;
262 	u32 hwp_boost_min;
263 	bool suspended;
264 #ifdef CONFIG_ENERGY_MODEL
265 	bool pd_registered;
266 #endif
267 	struct delayed_work hwp_notify_work;
268 };
269 
270 static struct cpudata **all_cpu_data;
271 
272 /**
273  * struct pstate_funcs - Per CPU model specific callbacks
274  * @get_max:		Callback to get maximum non turbo effective P state
275  * @get_max_physical:	Callback to get maximum non turbo physical P state
276  * @get_min:		Callback to get minimum P state
277  * @get_turbo:		Callback to get turbo P state
278  * @get_scaling:	Callback to get frequency scaling factor
279  * @get_cpu_scaling:	Get frequency scaling factor for a given cpu
280  * @get_aperf_mperf_shift: Callback to get the APERF vs MPERF frequency difference
281  * @get_val:		Callback to convert P state to actual MSR write value
282  * @get_vid:		Callback to get VID data for Atom platforms
283  *
284  * Core and Atom CPU models have different way to get P State limits. This
285  * structure is used to store those callbacks.
286  */
287 struct pstate_funcs {
288 	int (*get_max)(int cpu);
289 	int (*get_max_physical)(int cpu);
290 	int (*get_min)(int cpu);
291 	int (*get_turbo)(int cpu);
292 	int (*get_scaling)(void);
293 	int (*get_cpu_scaling)(int cpu);
294 	int (*get_aperf_mperf_shift)(void);
295 	u64 (*get_val)(struct cpudata*, int pstate);
296 	void (*get_vid)(struct cpudata *);
297 };
298 
299 static struct pstate_funcs pstate_funcs __read_mostly;
300 
301 static bool hwp_active __ro_after_init;
302 static int hwp_mode_bdw __ro_after_init;
303 static bool per_cpu_limits __ro_after_init;
304 static bool hwp_forced __ro_after_init;
305 static bool hwp_boost __read_mostly;
306 static bool hwp_is_hybrid;
307 
308 static struct cpufreq_driver *intel_pstate_driver __read_mostly;
309 
310 #define INTEL_PSTATE_CORE_SCALING	100000
311 #define HYBRID_SCALING_FACTOR_ADL	78741
312 #define HYBRID_SCALING_FACTOR_MTL	80000
313 #define HYBRID_SCALING_FACTOR_LNL	86957
314 
315 static int hybrid_scaling_factor;
316 
317 static inline int core_get_scaling(void)
318 {
319 	return INTEL_PSTATE_CORE_SCALING;
320 }
321 
322 #ifdef CONFIG_ACPI
323 static bool acpi_ppc;
324 #endif
325 
326 static struct global_params global;
327 
328 static DEFINE_MUTEX(intel_pstate_driver_lock);
329 static DEFINE_MUTEX(intel_pstate_limits_lock);
330 
331 #ifdef CONFIG_ACPI
332 
333 static bool intel_pstate_acpi_pm_profile_server(void)
334 {
335 	if (acpi_gbl_FADT.preferred_profile == PM_ENTERPRISE_SERVER ||
336 	    acpi_gbl_FADT.preferred_profile == PM_PERFORMANCE_SERVER)
337 		return true;
338 
339 	return false;
340 }
341 
342 static bool intel_pstate_get_ppc_enable_status(void)
343 {
344 	if (intel_pstate_acpi_pm_profile_server())
345 		return true;
346 
347 	return acpi_ppc;
348 }
349 
350 #ifdef CONFIG_ACPI_CPPC_LIB
351 
352 /* The work item is needed to avoid CPU hotplug locking issues */
353 static void intel_pstste_sched_itmt_work_fn(struct work_struct *work)
354 {
355 	sched_set_itmt_support();
356 }
357 
358 static DECLARE_WORK(sched_itmt_work, intel_pstste_sched_itmt_work_fn);
359 
360 #define CPPC_MAX_PERF	U8_MAX
361 
362 static void intel_pstate_set_itmt_prio(int cpu)
363 {
364 	struct cppc_perf_caps cppc_perf;
365 	static u32 max_highest_perf = 0, min_highest_perf = U32_MAX;
366 	int ret;
367 
368 	ret = cppc_get_perf_caps(cpu, &cppc_perf);
369 	/*
370 	 * If CPPC is not available, fall back to MSR_HWP_CAPABILITIES bits [8:0].
371 	 *
372 	 * Also, on some systems with overclocking enabled, CPPC.highest_perf is
373 	 * hardcoded to 0xff, so CPPC.highest_perf cannot be used to enable ITMT.
374 	 * Fall back to MSR_HWP_CAPABILITIES then too.
375 	 */
376 	if (ret || cppc_perf.highest_perf == CPPC_MAX_PERF)
377 		cppc_perf.highest_perf = HWP_HIGHEST_PERF(READ_ONCE(all_cpu_data[cpu]->hwp_cap_cached));
378 
379 	/*
380 	 * The priorities can be set regardless of whether or not
381 	 * sched_set_itmt_support(true) has been called and it is valid to
382 	 * update them at any time after it has been called.
383 	 */
384 	sched_set_itmt_core_prio(cppc_perf.highest_perf, cpu);
385 
386 	if (max_highest_perf <= min_highest_perf) {
387 		if (cppc_perf.highest_perf > max_highest_perf)
388 			max_highest_perf = cppc_perf.highest_perf;
389 
390 		if (cppc_perf.highest_perf < min_highest_perf)
391 			min_highest_perf = cppc_perf.highest_perf;
392 
393 		if (max_highest_perf > min_highest_perf) {
394 			/*
395 			 * This code can be run during CPU online under the
396 			 * CPU hotplug locks, so sched_set_itmt_support()
397 			 * cannot be called from here.  Queue up a work item
398 			 * to invoke it.
399 			 */
400 			schedule_work(&sched_itmt_work);
401 		}
402 	}
403 }
404 
405 static int intel_pstate_get_cppc_guaranteed(int cpu)
406 {
407 	struct cppc_perf_caps cppc_perf;
408 	int ret;
409 
410 	ret = cppc_get_perf_caps(cpu, &cppc_perf);
411 	if (ret)
412 		return ret;
413 
414 	if (cppc_perf.guaranteed_perf)
415 		return cppc_perf.guaranteed_perf;
416 
417 	return cppc_perf.nominal_perf;
418 }
419 
420 static int intel_pstate_cppc_get_scaling(int cpu)
421 {
422 	struct cppc_perf_caps cppc_perf;
423 
424 	/*
425 	 * Compute the perf-to-frequency scaling factor for the given CPU if
426 	 * possible, unless it would be 0.
427 	 */
428 	if (!cppc_get_perf_caps(cpu, &cppc_perf) &&
429 	    cppc_perf.nominal_perf && cppc_perf.nominal_freq)
430 		return div_u64(cppc_perf.nominal_freq * KHZ_PER_MHZ,
431 			       cppc_perf.nominal_perf);
432 
433 	return core_get_scaling();
434 }
435 
436 #else /* CONFIG_ACPI_CPPC_LIB */
437 static inline void intel_pstate_set_itmt_prio(int cpu)
438 {
439 }
440 #endif /* CONFIG_ACPI_CPPC_LIB */
441 
442 static void intel_pstate_init_acpi_perf_limits(struct cpufreq_policy *policy)
443 {
444 	struct cpudata *cpu;
445 	int ret;
446 	int i;
447 
448 	if (hwp_active) {
449 		intel_pstate_set_itmt_prio(policy->cpu);
450 		return;
451 	}
452 
453 	if (!intel_pstate_get_ppc_enable_status())
454 		return;
455 
456 	cpu = all_cpu_data[policy->cpu];
457 
458 	ret = acpi_processor_register_performance(&cpu->acpi_perf_data,
459 						  policy->cpu);
460 	if (ret)
461 		return;
462 
463 	/*
464 	 * Check if the control value in _PSS is for PERF_CTL MSR, which should
465 	 * guarantee that the states returned by it map to the states in our
466 	 * list directly.
467 	 */
468 	if (cpu->acpi_perf_data.control_register.space_id !=
469 						ACPI_ADR_SPACE_FIXED_HARDWARE)
470 		goto err;
471 
472 	/*
473 	 * If there is only one entry _PSS, simply ignore _PSS and continue as
474 	 * usual without taking _PSS into account
475 	 */
476 	if (cpu->acpi_perf_data.state_count < 2)
477 		goto err;
478 
479 	pr_debug("CPU%u - ACPI _PSS perf data\n", policy->cpu);
480 	for (i = 0; i < cpu->acpi_perf_data.state_count; i++) {
481 		pr_debug("     %cP%d: %u MHz, %u mW, 0x%x\n",
482 			 (i == cpu->acpi_perf_data.state ? '*' : ' '), i,
483 			 (u32) cpu->acpi_perf_data.states[i].core_frequency,
484 			 (u32) cpu->acpi_perf_data.states[i].power,
485 			 (u32) cpu->acpi_perf_data.states[i].control);
486 	}
487 
488 	cpu->valid_pss_table = true;
489 	pr_debug("_PPC limits will be enforced\n");
490 
491 	return;
492 
493  err:
494 	cpu->valid_pss_table = false;
495 	acpi_processor_unregister_performance(policy->cpu);
496 }
497 
498 static void intel_pstate_exit_perf_limits(struct cpufreq_policy *policy)
499 {
500 	struct cpudata *cpu;
501 
502 	cpu = all_cpu_data[policy->cpu];
503 	if (!cpu->valid_pss_table)
504 		return;
505 
506 	acpi_processor_unregister_performance(policy->cpu);
507 }
508 #else /* CONFIG_ACPI */
509 static inline void intel_pstate_init_acpi_perf_limits(struct cpufreq_policy *policy)
510 {
511 }
512 
513 static inline void intel_pstate_exit_perf_limits(struct cpufreq_policy *policy)
514 {
515 }
516 
517 static inline bool intel_pstate_acpi_pm_profile_server(void)
518 {
519 	return false;
520 }
521 #endif /* CONFIG_ACPI */
522 
523 #ifndef CONFIG_ACPI_CPPC_LIB
524 static inline int intel_pstate_get_cppc_guaranteed(int cpu)
525 {
526 	return -ENOTSUPP;
527 }
528 
529 static int intel_pstate_cppc_get_scaling(int cpu)
530 {
531 	return core_get_scaling();
532 }
533 #endif /* CONFIG_ACPI_CPPC_LIB */
534 
535 static int intel_pstate_freq_to_hwp_rel(struct cpudata *cpu, int freq,
536 					unsigned int relation)
537 {
538 	if (freq == cpu->pstate.turbo_freq)
539 		return cpu->pstate.turbo_pstate;
540 
541 	if (freq == cpu->pstate.max_freq)
542 		return cpu->pstate.max_pstate;
543 
544 	switch (relation) {
545 	case CPUFREQ_RELATION_H:
546 		return freq / cpu->pstate.scaling;
547 	case CPUFREQ_RELATION_C:
548 		return DIV_ROUND_CLOSEST(freq, cpu->pstate.scaling);
549 	}
550 
551 	return DIV_ROUND_UP(freq, cpu->pstate.scaling);
552 }
553 
554 static int intel_pstate_freq_to_hwp(struct cpudata *cpu, int freq)
555 {
556 	return intel_pstate_freq_to_hwp_rel(cpu, freq, CPUFREQ_RELATION_L);
557 }
558 
559 /**
560  * intel_pstate_hybrid_hwp_adjust - Calibrate HWP performance levels.
561  * @cpu: Target CPU.
562  *
563  * On hybrid processors, HWP may expose more performance levels than there are
564  * P-states accessible through the PERF_CTL interface.  If that happens, the
565  * scaling factor between HWP performance levels and CPU frequency will be less
566  * than the scaling factor between P-state values and CPU frequency.
567  *
568  * In that case, adjust the CPU parameters used in computations accordingly.
569  */
570 static void intel_pstate_hybrid_hwp_adjust(struct cpudata *cpu)
571 {
572 	int perf_ctl_max_phys = cpu->pstate.max_pstate_physical;
573 	int perf_ctl_scaling = cpu->pstate.perf_ctl_scaling;
574 	int perf_ctl_turbo = pstate_funcs.get_turbo(cpu->cpu);
575 	int scaling = cpu->pstate.scaling;
576 	int freq;
577 
578 	pr_debug("CPU%d: PERF_CTL max_phys = %d\n", cpu->cpu, perf_ctl_max_phys);
579 	pr_debug("CPU%d: PERF_CTL turbo = %d\n", cpu->cpu, perf_ctl_turbo);
580 	pr_debug("CPU%d: PERF_CTL scaling = %d\n", cpu->cpu, perf_ctl_scaling);
581 	pr_debug("CPU%d: HWP_CAP guaranteed = %d\n", cpu->cpu, cpu->pstate.max_pstate);
582 	pr_debug("CPU%d: HWP_CAP highest = %d\n", cpu->cpu, cpu->pstate.turbo_pstate);
583 	pr_debug("CPU%d: HWP-to-frequency scaling factor: %d\n", cpu->cpu, scaling);
584 
585 	if (scaling == perf_ctl_scaling)
586 		return;
587 
588 	hwp_is_hybrid = true;
589 
590 	cpu->pstate.turbo_freq = rounddown(cpu->pstate.turbo_pstate * scaling,
591 					   perf_ctl_scaling);
592 	cpu->pstate.max_freq = rounddown(cpu->pstate.max_pstate * scaling,
593 					 perf_ctl_scaling);
594 
595 	freq = perf_ctl_max_phys * perf_ctl_scaling;
596 	cpu->pstate.max_pstate_physical = intel_pstate_freq_to_hwp(cpu, freq);
597 
598 	freq = cpu->pstate.min_pstate * perf_ctl_scaling;
599 	cpu->pstate.min_freq = freq;
600 	/*
601 	 * Cast the min P-state value retrieved via pstate_funcs.get_min() to
602 	 * the effective range of HWP performance levels.
603 	 */
604 	cpu->pstate.min_pstate = intel_pstate_freq_to_hwp(cpu, freq);
605 }
606 
607 static bool turbo_is_disabled(void)
608 {
609 	u64 misc_en;
610 
611 	rdmsrq(MSR_IA32_MISC_ENABLE, misc_en);
612 
613 	return !!(misc_en & MSR_IA32_MISC_ENABLE_TURBO_DISABLE);
614 }
615 
616 static int min_perf_pct_min(void)
617 {
618 	struct cpudata *cpu = all_cpu_data[0];
619 	int turbo_pstate = cpu->pstate.turbo_pstate;
620 
621 	return turbo_pstate ?
622 		(cpu->pstate.min_pstate * 100 / turbo_pstate) : 0;
623 }
624 
625 static s16 intel_pstate_get_epp(struct cpudata *cpu_data, u64 hwp_req_data)
626 {
627 	s16 epp = -EOPNOTSUPP;
628 
629 	if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
630 		/*
631 		 * When hwp_req_data is 0, means that caller didn't read
632 		 * MSR_HWP_REQUEST, so need to read and get EPP.
633 		 */
634 		if (!hwp_req_data) {
635 			epp = rdmsrq_on_cpu(cpu_data->cpu, MSR_HWP_REQUEST,
636 					    &hwp_req_data);
637 			if (epp)
638 				return epp;
639 		}
640 		epp = (hwp_req_data >> 24) & 0xff;
641 	}
642 
643 	return epp;
644 }
645 
646 /*
647  * EPP display strings corresponding to EPP index in the
648  * energy_perf_strings[]
649  *	index		String
650  *-------------------------------------
651  *	0		default
652  *	1		performance
653  *	2		balance_performance
654  *	3		balance_power
655  *	4		power
656  */
657 
658 enum energy_perf_value_index {
659 	EPP_INDEX_DEFAULT = 0,
660 	EPP_INDEX_PERFORMANCE,
661 	EPP_INDEX_BALANCE_PERFORMANCE,
662 	EPP_INDEX_BALANCE_POWERSAVE,
663 	EPP_INDEX_POWERSAVE,
664 };
665 
666 static const char * const energy_perf_strings[] = {
667 	[EPP_INDEX_DEFAULT] = "default",
668 	[EPP_INDEX_PERFORMANCE] = "performance",
669 	[EPP_INDEX_BALANCE_PERFORMANCE] = "balance_performance",
670 	[EPP_INDEX_BALANCE_POWERSAVE] = "balance_power",
671 	[EPP_INDEX_POWERSAVE] = "power",
672 	NULL
673 };
674 static unsigned int epp_values[] = {
675 	[EPP_INDEX_DEFAULT] = 0, /* Unused index */
676 	[EPP_INDEX_PERFORMANCE] = HWP_EPP_PERFORMANCE,
677 	[EPP_INDEX_BALANCE_PERFORMANCE] = HWP_EPP_BALANCE_PERFORMANCE,
678 	[EPP_INDEX_BALANCE_POWERSAVE] = HWP_EPP_BALANCE_POWERSAVE,
679 	[EPP_INDEX_POWERSAVE] = HWP_EPP_POWERSAVE,
680 };
681 
682 static int intel_pstate_get_energy_pref_index(struct cpudata *cpu_data, int *raw_epp)
683 {
684 	s16 epp;
685 	int index = -EINVAL;
686 
687 	*raw_epp = 0;
688 	epp = intel_pstate_get_epp(cpu_data, 0);
689 	if (epp < 0)
690 		return epp;
691 
692 	if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
693 		if (epp == epp_values[EPP_INDEX_PERFORMANCE])
694 			return EPP_INDEX_PERFORMANCE;
695 		if (epp == epp_values[EPP_INDEX_BALANCE_PERFORMANCE])
696 			return EPP_INDEX_BALANCE_PERFORMANCE;
697 		if (epp == epp_values[EPP_INDEX_BALANCE_POWERSAVE])
698 			return EPP_INDEX_BALANCE_POWERSAVE;
699 		if (epp == epp_values[EPP_INDEX_POWERSAVE])
700 			return EPP_INDEX_POWERSAVE;
701 		*raw_epp = epp;
702 		return 0;
703 	} else if (boot_cpu_has(X86_FEATURE_EPB)) {
704 		/*
705 		 * Range:
706 		 *	0x00-0x03	:	Performance
707 		 *	0x04-0x07	:	Balance performance
708 		 *	0x08-0x0B	:	Balance power
709 		 *	0x0C-0x0F	:	Power
710 		 * The EPB is a 4 bit value, but our ranges restrict the
711 		 * value which can be set. Here only using top two bits
712 		 * effectively.
713 		 */
714 		index = (epp >> 2) + 1;
715 	}
716 
717 	return index;
718 }
719 
720 static int intel_pstate_set_epp(struct cpudata *cpu, u32 epp)
721 {
722 	int ret;
723 
724 	/*
725 	 * Use the cached HWP Request MSR value, because in the active mode the
726 	 * register itself may be updated by intel_pstate_hwp_boost_up() or
727 	 * intel_pstate_hwp_boost_down() at any time.
728 	 */
729 	u64 value = READ_ONCE(cpu->hwp_req_cached);
730 
731 	value &= ~GENMASK_ULL(31, 24);
732 	value |= (u64)epp << 24;
733 	/*
734 	 * The only other updater of hwp_req_cached in the active mode,
735 	 * intel_pstate_hwp_set(), is called under the same lock as this
736 	 * function, so it cannot run in parallel with the update below.
737 	 */
738 	WRITE_ONCE(cpu->hwp_req_cached, value);
739 	ret = wrmsrq_on_cpu(cpu->cpu, MSR_HWP_REQUEST, value);
740 	if (!ret)
741 		cpu->epp_cached = epp;
742 
743 	return ret;
744 }
745 
746 static int intel_pstate_set_energy_pref_index(struct cpudata *cpu_data,
747 					      int pref_index, bool use_raw,
748 					      u32 raw_epp)
749 {
750 	int epp = -EINVAL;
751 	int ret = -EOPNOTSUPP;
752 
753 	if (!pref_index)
754 		epp = cpu_data->epp_default;
755 
756 	if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
757 		if (use_raw)
758 			epp = raw_epp;
759 		else if (epp == -EINVAL)
760 			epp = epp_values[pref_index];
761 
762 		/*
763 		 * To avoid confusion, refuse to set EPP to any values different
764 		 * from 0 (performance) if the current policy is "performance",
765 		 * because those values would be overridden.
766 		 */
767 		if (epp > 0 && cpu_data->policy == CPUFREQ_POLICY_PERFORMANCE)
768 			return -EBUSY;
769 
770 		ret = intel_pstate_set_epp(cpu_data, epp);
771 	}
772 
773 	return ret;
774 }
775 
776 static ssize_t show_energy_performance_available_preferences(
777 				struct cpufreq_policy *policy, char *buf)
778 {
779 	int i = 0;
780 	int ret = 0;
781 
782 	while (energy_perf_strings[i] != NULL)
783 		ret += sprintf(&buf[ret], "%s ", energy_perf_strings[i++]);
784 
785 	ret += sprintf(&buf[ret], "\n");
786 
787 	return ret;
788 }
789 
790 cpufreq_freq_attr_ro(energy_performance_available_preferences);
791 
792 static struct cpufreq_driver intel_pstate;
793 
794 static ssize_t store_energy_performance_preference(
795 		struct cpufreq_policy *policy, const char *buf, size_t count)
796 {
797 	struct cpudata *cpu = all_cpu_data[policy->cpu];
798 	char str_preference[21];
799 	bool raw = false;
800 	ssize_t ret;
801 	u32 epp = 0;
802 
803 	ret = sscanf(buf, "%20s", str_preference);
804 	if (ret != 1)
805 		return -EINVAL;
806 
807 	ret = match_string(energy_perf_strings, -1, str_preference);
808 	if (ret < 0) {
809 		if (!boot_cpu_has(X86_FEATURE_HWP_EPP))
810 			return ret;
811 
812 		ret = kstrtouint(buf, 10, &epp);
813 		if (ret)
814 			return ret;
815 
816 		if (epp > 255)
817 			return -EINVAL;
818 
819 		raw = true;
820 	}
821 
822 	/*
823 	 * This function runs with the policy R/W semaphore held, which
824 	 * guarantees that the driver pointer will not change while it is
825 	 * running.
826 	 */
827 	if (!intel_pstate_driver)
828 		return -EAGAIN;
829 
830 	mutex_lock(&intel_pstate_limits_lock);
831 
832 	if (intel_pstate_driver == &intel_pstate) {
833 		ret = intel_pstate_set_energy_pref_index(cpu, ret, raw, epp);
834 	} else {
835 		/*
836 		 * In the passive mode the governor needs to be stopped on the
837 		 * target CPU before the EPP update and restarted after it,
838 		 * which is super-heavy-weight, so make sure it is worth doing
839 		 * upfront.
840 		 */
841 		if (!raw)
842 			epp = ret ? epp_values[ret] : cpu->epp_default;
843 
844 		if (cpu->epp_cached != epp) {
845 			int err;
846 
847 			cpufreq_stop_governor(policy);
848 			ret = intel_pstate_set_epp(cpu, epp);
849 			err = cpufreq_start_governor(policy);
850 			if (!ret)
851 				ret = err;
852 		} else {
853 			ret = 0;
854 		}
855 	}
856 
857 	mutex_unlock(&intel_pstate_limits_lock);
858 
859 	return ret ?: count;
860 }
861 
862 static ssize_t show_energy_performance_preference(
863 				struct cpufreq_policy *policy, char *buf)
864 {
865 	struct cpudata *cpu_data = all_cpu_data[policy->cpu];
866 	int preference, raw_epp;
867 
868 	preference = intel_pstate_get_energy_pref_index(cpu_data, &raw_epp);
869 	if (preference < 0)
870 		return preference;
871 
872 	if (raw_epp)
873 		return  sprintf(buf, "%d\n", raw_epp);
874 	else
875 		return  sprintf(buf, "%s\n", energy_perf_strings[preference]);
876 }
877 
878 cpufreq_freq_attr_rw(energy_performance_preference);
879 
880 static ssize_t show_base_frequency(struct cpufreq_policy *policy, char *buf)
881 {
882 	struct cpudata *cpu = all_cpu_data[policy->cpu];
883 	int ratio, freq;
884 
885 	ratio = intel_pstate_get_cppc_guaranteed(policy->cpu);
886 	if (ratio <= 0) {
887 		u64 cap;
888 
889 		rdmsrq_on_cpu(policy->cpu, MSR_HWP_CAPABILITIES, &cap);
890 		ratio = HWP_GUARANTEED_PERF(cap);
891 	}
892 
893 	freq = ratio * cpu->pstate.scaling;
894 	if (cpu->pstate.scaling != cpu->pstate.perf_ctl_scaling)
895 		freq = rounddown(freq, cpu->pstate.perf_ctl_scaling);
896 
897 	return sprintf(buf, "%d\n", freq);
898 }
899 
900 cpufreq_freq_attr_ro(base_frequency);
901 
902 enum hwp_cpufreq_attr_index {
903 	HWP_BASE_FREQUENCY_INDEX = 0,
904 	HWP_PERFORMANCE_PREFERENCE_INDEX,
905 	HWP_PERFORMANCE_AVAILABLE_PREFERENCES_INDEX,
906 	HWP_CPUFREQ_ATTR_COUNT,
907 };
908 
909 static struct freq_attr *hwp_cpufreq_attrs[] = {
910 	[HWP_BASE_FREQUENCY_INDEX] = &base_frequency,
911 	[HWP_PERFORMANCE_PREFERENCE_INDEX] = &energy_performance_preference,
912 	[HWP_PERFORMANCE_AVAILABLE_PREFERENCES_INDEX] =
913 				&energy_performance_available_preferences,
914 	[HWP_CPUFREQ_ATTR_COUNT] = NULL,
915 };
916 
917 static u8 hybrid_get_cpu_type(unsigned int cpu)
918 {
919 	return cpu_data(cpu).topo.intel_type;
920 }
921 
922 static bool no_cas __ro_after_init;
923 
924 static struct cpudata *hybrid_max_perf_cpu __read_mostly;
925 /*
926  * Protects hybrid_max_perf_cpu, the capacity_perf fields in struct cpudata,
927  * and the x86 arch scale-invariance information from concurrent updates.
928  */
929 static DEFINE_MUTEX(hybrid_capacity_lock);
930 
931 #ifdef CONFIG_ENERGY_MODEL
932 #define HYBRID_EM_STATE_COUNT	4
933 
934 static int hybrid_active_power(struct device *dev, unsigned long *power,
935 			       unsigned long *freq)
936 {
937 	/*
938 	 * Create four "states" corresponding to 40%, 60%, 80%, and 100% of the
939 	 * full capacity.
940 	 *
941 	 * For this purpose, return the "frequency" of 2 for the first
942 	 * performance level and otherwise leave the value set by the caller.
943 	 */
944 	if (!*freq)
945 		*freq = 2;
946 
947 	/* No power information. */
948 	*power = EM_MAX_POWER;
949 
950 	return 0;
951 }
952 
953 static bool hybrid_has_l3(unsigned int cpu)
954 {
955 	struct cpu_cacheinfo *cacheinfo = get_cpu_cacheinfo(cpu);
956 	unsigned int i;
957 
958 	if (!cacheinfo)
959 		return false;
960 
961 	for (i = 0; i < cacheinfo->num_leaves; i++) {
962 		if (cacheinfo->info_list[i].level == 3)
963 			return true;
964 	}
965 
966 	return false;
967 }
968 
969 static int hybrid_get_cost(struct device *dev, unsigned long freq,
970 			   unsigned long *cost)
971 {
972 	/* Facilitate load balancing between CPUs of the same type. */
973 	*cost = freq;
974 	/*
975 	 * Adjust the cost depending on CPU type.
976 	 *
977 	 * The idea is to start loading up LPE-cores before E-cores and start
978 	 * to populate E-cores when LPE-cores are utilized above 60% of the
979 	 * capacity.  Similarly, P-cores start to be populated when E-cores are
980 	 * utilized above 60% of the capacity.
981 	 */
982 	if (hybrid_get_cpu_type(dev->id) == INTEL_CPU_TYPE_ATOM) {
983 		if (hybrid_has_l3(dev->id)) /* E-core */
984 			*cost += 1;
985 	} else { /* P-core */
986 		*cost += 2;
987 	}
988 
989 	return 0;
990 }
991 
992 static bool hybrid_register_perf_domain(unsigned int cpu)
993 {
994 	static const struct em_data_callback cb
995 			= EM_ADV_DATA_CB(hybrid_active_power, hybrid_get_cost);
996 	struct cpudata *cpudata = all_cpu_data[cpu];
997 	struct device *cpu_dev;
998 
999 	/*
1000 	 * Registering EM perf domains without enabling asymmetric CPU capacity
1001 	 * support is not really useful and one domain should not be registered
1002 	 * more than once.
1003 	 */
1004 	if (!hybrid_max_perf_cpu || cpudata->pd_registered)
1005 		return false;
1006 
1007 	cpu_dev = get_cpu_device(cpu);
1008 	if (!cpu_dev)
1009 		return false;
1010 
1011 	if (em_dev_register_pd_no_update(cpu_dev, HYBRID_EM_STATE_COUNT, &cb,
1012 					 cpumask_of(cpu), false))
1013 		return false;
1014 
1015 	cpudata->pd_registered = true;
1016 
1017 	return true;
1018 }
1019 
1020 static void hybrid_register_all_perf_domains(void)
1021 {
1022 	unsigned int cpu;
1023 
1024 	for_each_online_cpu(cpu)
1025 		hybrid_register_perf_domain(cpu);
1026 }
1027 
1028 static void hybrid_update_perf_domain(struct cpudata *cpu)
1029 {
1030 	if (cpu->pd_registered)
1031 		em_adjust_cpu_capacity(cpu->cpu);
1032 }
1033 #else /* !CONFIG_ENERGY_MODEL */
1034 static inline bool hybrid_register_perf_domain(unsigned int cpu) { return false; }
1035 static inline void hybrid_register_all_perf_domains(void) {}
1036 static inline void hybrid_update_perf_domain(struct cpudata *cpu) {}
1037 #endif /* CONFIG_ENERGY_MODEL */
1038 
1039 static void hybrid_set_cpu_capacity(struct cpudata *cpu)
1040 {
1041 	arch_set_cpu_capacity(cpu->cpu, cpu->capacity_perf,
1042 			      hybrid_max_perf_cpu->capacity_perf,
1043 			      cpu->capacity_perf,
1044 			      cpu->pstate.max_pstate_physical);
1045 	hybrid_update_perf_domain(cpu);
1046 
1047 	topology_set_cpu_scale(cpu->cpu, arch_scale_cpu_capacity(cpu->cpu));
1048 
1049 	pr_debug("CPU%d: capacity perf = %u, base perf = %u, sys max perf = %u\n",
1050 		 cpu->cpu, cpu->capacity_perf, cpu->pstate.max_pstate_physical,
1051 		 hybrid_max_perf_cpu->capacity_perf);
1052 }
1053 
1054 static void hybrid_clear_cpu_capacity(unsigned int cpunum)
1055 {
1056 	arch_set_cpu_capacity(cpunum, 1, 1, 1, 1);
1057 }
1058 
1059 static void hybrid_get_capacity_perf(struct cpudata *cpu)
1060 {
1061 	if (READ_ONCE(global.no_turbo)) {
1062 		cpu->capacity_perf = cpu->pstate.max_pstate_physical;
1063 		return;
1064 	}
1065 
1066 	cpu->capacity_perf = HWP_HIGHEST_PERF(READ_ONCE(cpu->hwp_cap_cached));
1067 }
1068 
1069 static void hybrid_set_capacity_of_cpus(void)
1070 {
1071 	int cpunum;
1072 
1073 	for_each_online_cpu(cpunum) {
1074 		struct cpudata *cpu = all_cpu_data[cpunum];
1075 
1076 		if (cpu)
1077 			hybrid_set_cpu_capacity(cpu);
1078 	}
1079 }
1080 
1081 static void hybrid_update_cpu_capacity_scaling(void)
1082 {
1083 	struct cpudata *max_perf_cpu = NULL;
1084 	unsigned int max_cap_perf = 0;
1085 	int cpunum;
1086 
1087 	for_each_online_cpu(cpunum) {
1088 		struct cpudata *cpu = all_cpu_data[cpunum];
1089 
1090 		if (!cpu)
1091 			continue;
1092 
1093 		/*
1094 		 * During initialization, CPU performance at full capacity needs
1095 		 * to be determined.
1096 		 */
1097 		if (!hybrid_max_perf_cpu)
1098 			hybrid_get_capacity_perf(cpu);
1099 
1100 		/*
1101 		 * If hybrid_max_perf_cpu is not NULL at this point, it is
1102 		 * being replaced, so don't take it into account when looking
1103 		 * for the new one.
1104 		 */
1105 		if (cpu == hybrid_max_perf_cpu)
1106 			continue;
1107 
1108 		if (cpu->capacity_perf > max_cap_perf) {
1109 			max_cap_perf = cpu->capacity_perf;
1110 			max_perf_cpu = cpu;
1111 		}
1112 	}
1113 
1114 	if (max_perf_cpu) {
1115 		hybrid_max_perf_cpu = max_perf_cpu;
1116 		hybrid_set_capacity_of_cpus();
1117 	} else {
1118 		pr_info("Found no CPUs with nonzero maximum performance\n");
1119 		/* Revert to the flat CPU capacity structure. */
1120 		for_each_online_cpu(cpunum)
1121 			hybrid_clear_cpu_capacity(cpunum);
1122 	}
1123 }
1124 
1125 static void __hybrid_refresh_cpu_capacity_scaling(void)
1126 {
1127 	hybrid_max_perf_cpu = NULL;
1128 	hybrid_update_cpu_capacity_scaling();
1129 }
1130 
1131 static void hybrid_refresh_cpu_capacity_scaling(void)
1132 {
1133 	guard(mutex)(&hybrid_capacity_lock);
1134 
1135 	__hybrid_refresh_cpu_capacity_scaling();
1136 	/*
1137 	 * Perf domains are not registered before setting hybrid_max_perf_cpu,
1138 	 * so register them all after setting up CPU capacity scaling.
1139 	 */
1140 	hybrid_register_all_perf_domains();
1141 }
1142 
1143 static void hybrid_init_cpu_capacity_scaling(bool refresh)
1144 {
1145 	/* Bail out if enabling capacity-aware scheduling is prohibited. */
1146 	if (no_cas)
1147 		return;
1148 
1149 	/*
1150 	 * If hybrid_max_perf_cpu is set at this point, the hybrid CPU capacity
1151 	 * scaling has been enabled already and the driver is just changing the
1152 	 * operation mode.
1153 	 */
1154 	if (refresh) {
1155 		hybrid_refresh_cpu_capacity_scaling();
1156 		return;
1157 	}
1158 
1159 	/*
1160 	 * On hybrid systems, use asym capacity instead of ITMT, but because
1161 	 * the capacity of SMT threads is not deterministic even approximately,
1162 	 * do not do that when SMT is in use.
1163 	 */
1164 	if (hwp_is_hybrid && !sched_smt_active() && arch_enable_hybrid_capacity_scale()) {
1165 		hybrid_refresh_cpu_capacity_scaling();
1166 		/*
1167 		 * Disabling ITMT causes sched domains to be rebuilt to disable asym
1168 		 * packing and enable asym capacity and EAS.
1169 		 */
1170 		sched_clear_itmt_support();
1171 	}
1172 }
1173 
1174 static bool hybrid_clear_max_perf_cpu(void)
1175 {
1176 	bool ret;
1177 
1178 	guard(mutex)(&hybrid_capacity_lock);
1179 
1180 	ret = !!hybrid_max_perf_cpu;
1181 	hybrid_max_perf_cpu = NULL;
1182 
1183 	return ret;
1184 }
1185 
1186 static void __intel_pstate_get_hwp_cap(struct cpudata *cpu)
1187 {
1188 	u64 cap;
1189 
1190 	rdmsrq_on_cpu(cpu->cpu, MSR_HWP_CAPABILITIES, &cap);
1191 	WRITE_ONCE(cpu->hwp_cap_cached, cap);
1192 	cpu->pstate.max_pstate = HWP_GUARANTEED_PERF(cap);
1193 	cpu->pstate.turbo_pstate = HWP_HIGHEST_PERF(cap);
1194 }
1195 
1196 static void intel_pstate_get_hwp_cap(struct cpudata *cpu)
1197 {
1198 	int scaling = cpu->pstate.scaling;
1199 
1200 	__intel_pstate_get_hwp_cap(cpu);
1201 
1202 	cpu->pstate.max_freq = cpu->pstate.max_pstate * scaling;
1203 	cpu->pstate.turbo_freq = cpu->pstate.turbo_pstate * scaling;
1204 	if (scaling != cpu->pstate.perf_ctl_scaling) {
1205 		int perf_ctl_scaling = cpu->pstate.perf_ctl_scaling;
1206 
1207 		cpu->pstate.max_freq = rounddown(cpu->pstate.max_freq,
1208 						 perf_ctl_scaling);
1209 		cpu->pstate.turbo_freq = rounddown(cpu->pstate.turbo_freq,
1210 						   perf_ctl_scaling);
1211 	}
1212 }
1213 
1214 static void hybrid_update_capacity(struct cpudata *cpu)
1215 {
1216 	unsigned int max_cap_perf;
1217 
1218 	mutex_lock(&hybrid_capacity_lock);
1219 
1220 	if (!hybrid_max_perf_cpu)
1221 		goto unlock;
1222 
1223 	/*
1224 	 * The maximum performance of the CPU may have changed, but assume
1225 	 * that the performance of the other CPUs has not changed.
1226 	 */
1227 	max_cap_perf = hybrid_max_perf_cpu->capacity_perf;
1228 
1229 	intel_pstate_get_hwp_cap(cpu);
1230 
1231 	hybrid_get_capacity_perf(cpu);
1232 	/* Should hybrid_max_perf_cpu be replaced by this CPU? */
1233 	if (cpu->capacity_perf > max_cap_perf) {
1234 		hybrid_max_perf_cpu = cpu;
1235 		hybrid_set_capacity_of_cpus();
1236 		goto unlock;
1237 	}
1238 
1239 	/* If this CPU is hybrid_max_perf_cpu, should it be replaced? */
1240 	if (cpu == hybrid_max_perf_cpu && cpu->capacity_perf < max_cap_perf) {
1241 		hybrid_update_cpu_capacity_scaling();
1242 		goto unlock;
1243 	}
1244 
1245 	hybrid_set_cpu_capacity(cpu);
1246 	/*
1247 	 * If the CPU was offline to start with and it is going online for the
1248 	 * first time, a perf domain needs to be registered for it if hybrid
1249 	 * capacity scaling has been enabled already.  In that case, sched
1250 	 * domains need to be rebuilt to take the new perf domain into account.
1251 	 */
1252 	if (hybrid_register_perf_domain(cpu->cpu))
1253 		em_rebuild_sched_domains();
1254 
1255 unlock:
1256 	mutex_unlock(&hybrid_capacity_lock);
1257 }
1258 
1259 static void intel_pstate_hwp_set(unsigned int cpu)
1260 {
1261 	struct cpudata *cpu_data = all_cpu_data[cpu];
1262 	int max, min;
1263 	u64 value;
1264 	s16 epp;
1265 
1266 	max = cpu_data->max_perf_ratio;
1267 	min = cpu_data->min_perf_ratio;
1268 
1269 	if (cpu_data->policy == CPUFREQ_POLICY_PERFORMANCE)
1270 		min = max;
1271 
1272 	rdmsrq_on_cpu(cpu, MSR_HWP_REQUEST, &value);
1273 
1274 	value &= ~HWP_MIN_PERF(~0L);
1275 	value |= HWP_MIN_PERF(min);
1276 
1277 	value &= ~HWP_MAX_PERF(~0L);
1278 	value |= HWP_MAX_PERF(max);
1279 
1280 	if (cpu_data->epp_policy == cpu_data->policy)
1281 		goto skip_epp;
1282 
1283 	cpu_data->epp_policy = cpu_data->policy;
1284 
1285 	if (cpu_data->policy == CPUFREQ_POLICY_PERFORMANCE) {
1286 		epp = intel_pstate_get_epp(cpu_data, value);
1287 		cpu_data->epp_powersave = epp;
1288 		/* If EPP read was failed, then don't try to write */
1289 		if (epp < 0)
1290 			goto skip_epp;
1291 
1292 		epp = 0;
1293 	} else {
1294 		/* skip setting EPP, when saved value is invalid */
1295 		if (cpu_data->epp_powersave < 0)
1296 			goto skip_epp;
1297 
1298 		/*
1299 		 * No need to restore EPP when it is not zero. This
1300 		 * means:
1301 		 *  - Policy is not changed
1302 		 *  - user has manually changed
1303 		 *  - Error reading EPB
1304 		 */
1305 		epp = intel_pstate_get_epp(cpu_data, value);
1306 		if (epp)
1307 			goto skip_epp;
1308 
1309 		epp = cpu_data->epp_powersave;
1310 	}
1311 	if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
1312 		value &= ~GENMASK_ULL(31, 24);
1313 		value |= (u64)epp << 24;
1314 	}
1315 
1316 skip_epp:
1317 	WRITE_ONCE(cpu_data->hwp_req_cached, value);
1318 	wrmsrq_on_cpu(cpu, MSR_HWP_REQUEST, value);
1319 }
1320 
1321 static void intel_pstate_disable_hwp_interrupt(struct cpudata *cpudata);
1322 
1323 static void intel_pstate_hwp_offline(struct cpudata *cpu)
1324 {
1325 	u64 value = READ_ONCE(cpu->hwp_req_cached);
1326 	int min_perf;
1327 
1328 	intel_pstate_disable_hwp_interrupt(cpu);
1329 
1330 	if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
1331 		/*
1332 		 * In case the EPP has been set to "performance" by the
1333 		 * active mode "performance" scaling algorithm, replace that
1334 		 * temporary value with the cached EPP one.
1335 		 */
1336 		value &= ~GENMASK_ULL(31, 24);
1337 		value |= HWP_ENERGY_PERF_PREFERENCE(cpu->epp_cached);
1338 		/*
1339 		 * However, make sure that EPP will be set to "performance" when
1340 		 * the CPU is brought back online again and the "performance"
1341 		 * scaling algorithm is still in effect.
1342 		 */
1343 		cpu->epp_policy = CPUFREQ_POLICY_UNKNOWN;
1344 	}
1345 
1346 	/*
1347 	 * Clear the desired perf field in the cached HWP request value to
1348 	 * prevent nonzero desired values from being leaked into the active
1349 	 * mode.
1350 	 */
1351 	value &= ~HWP_DESIRED_PERF(~0L);
1352 	WRITE_ONCE(cpu->hwp_req_cached, value);
1353 
1354 	value &= ~GENMASK_ULL(31, 0);
1355 	min_perf = HWP_LOWEST_PERF(READ_ONCE(cpu->hwp_cap_cached));
1356 
1357 	/* Set hwp_max = hwp_min */
1358 	value |= HWP_MAX_PERF(min_perf);
1359 	value |= HWP_MIN_PERF(min_perf);
1360 
1361 	/* Set EPP to min */
1362 	if (boot_cpu_has(X86_FEATURE_HWP_EPP))
1363 		value |= HWP_ENERGY_PERF_PREFERENCE(HWP_EPP_POWERSAVE);
1364 
1365 	wrmsrq_on_cpu(cpu->cpu, MSR_HWP_REQUEST, value);
1366 
1367 	mutex_lock(&hybrid_capacity_lock);
1368 
1369 	if (!hybrid_max_perf_cpu) {
1370 		mutex_unlock(&hybrid_capacity_lock);
1371 
1372 		return;
1373 	}
1374 
1375 	if (hybrid_max_perf_cpu == cpu)
1376 		hybrid_update_cpu_capacity_scaling();
1377 
1378 	mutex_unlock(&hybrid_capacity_lock);
1379 
1380 	/* Reset the capacity of the CPU going offline to the initial value. */
1381 	hybrid_clear_cpu_capacity(cpu->cpu);
1382 }
1383 
1384 #define POWER_CTL_EE_ENABLE	1
1385 #define POWER_CTL_EE_DISABLE	2
1386 
1387 /* Enable bit for Dynamic Efficiency Control (DEC) */
1388 #define POWER_CTL_DEC_ENABLE	27
1389 
1390 static int power_ctl_ee_state;
1391 
1392 static void set_power_ctl_ee_state(bool input)
1393 {
1394 	u64 power_ctl;
1395 
1396 	guard(mutex)(&intel_pstate_driver_lock);
1397 
1398 	rdmsrq(MSR_IA32_POWER_CTL, power_ctl);
1399 	if (input) {
1400 		power_ctl &= ~BIT(MSR_IA32_POWER_CTL_BIT_EE);
1401 		power_ctl_ee_state = POWER_CTL_EE_ENABLE;
1402 	} else {
1403 		power_ctl |= BIT(MSR_IA32_POWER_CTL_BIT_EE);
1404 		power_ctl_ee_state = POWER_CTL_EE_DISABLE;
1405 	}
1406 	wrmsrq(MSR_IA32_POWER_CTL, power_ctl);
1407 }
1408 
1409 static void intel_pstate_hwp_enable(struct cpudata *cpudata);
1410 
1411 static void intel_pstate_hwp_reenable(struct cpudata *cpu)
1412 {
1413 	intel_pstate_hwp_enable(cpu);
1414 	wrmsrq_on_cpu(cpu->cpu, MSR_HWP_REQUEST, READ_ONCE(cpu->hwp_req_cached));
1415 }
1416 
1417 static int intel_pstate_suspend(struct cpufreq_policy *policy)
1418 {
1419 	struct cpudata *cpu = all_cpu_data[policy->cpu];
1420 
1421 	pr_debug("CPU %d suspending\n", cpu->cpu);
1422 
1423 	cpu->suspended = true;
1424 
1425 	/* disable HWP interrupt and cancel any pending work */
1426 	intel_pstate_disable_hwp_interrupt(cpu);
1427 
1428 	return 0;
1429 }
1430 
1431 static int intel_pstate_resume(struct cpufreq_policy *policy)
1432 {
1433 	struct cpudata *cpu = all_cpu_data[policy->cpu];
1434 
1435 	pr_debug("CPU %d resuming\n", cpu->cpu);
1436 
1437 	/* Only restore if the system default is changed */
1438 	if (power_ctl_ee_state == POWER_CTL_EE_ENABLE)
1439 		set_power_ctl_ee_state(true);
1440 	else if (power_ctl_ee_state == POWER_CTL_EE_DISABLE)
1441 		set_power_ctl_ee_state(false);
1442 
1443 	if (cpu->suspended && hwp_active) {
1444 		mutex_lock(&intel_pstate_limits_lock);
1445 
1446 		/* Re-enable HWP, because "online" has not done that. */
1447 		intel_pstate_hwp_reenable(cpu);
1448 
1449 		mutex_unlock(&intel_pstate_limits_lock);
1450 	}
1451 
1452 	cpu->suspended = false;
1453 
1454 	return 0;
1455 }
1456 
1457 static void intel_pstate_update_policies(void)
1458 {
1459 	int cpu;
1460 
1461 	for_each_possible_cpu(cpu)
1462 		cpufreq_update_policy(cpu);
1463 }
1464 
1465 static void __intel_pstate_update_max_freq(struct cpufreq_policy *policy,
1466 					   struct cpudata *cpudata)
1467 {
1468 	guard(cpufreq_policy_write)(policy);
1469 
1470 	if (hwp_active)
1471 		intel_pstate_get_hwp_cap(cpudata);
1472 
1473 	policy->cpuinfo.max_freq = READ_ONCE(global.no_turbo) ?
1474 			cpudata->pstate.max_freq : cpudata->pstate.turbo_freq;
1475 
1476 	refresh_frequency_limits(policy);
1477 }
1478 
1479 static bool intel_pstate_update_max_freq(struct cpudata *cpudata)
1480 {
1481 	struct cpufreq_policy *policy __free(put_cpufreq_policy) = cpufreq_cpu_get(cpudata->cpu);
1482 	if (!policy)
1483 		return false;
1484 
1485 	__intel_pstate_update_max_freq(policy, cpudata);
1486 
1487 	return true;
1488 }
1489 
1490 static void intel_pstate_update_limits(struct cpufreq_policy *policy)
1491 {
1492 	struct cpudata *cpudata = all_cpu_data[policy->cpu];
1493 
1494 	__intel_pstate_update_max_freq(policy, cpudata);
1495 
1496 	hybrid_update_capacity(cpudata);
1497 }
1498 
1499 static void intel_pstate_update_limits_for_all(void)
1500 {
1501 	int cpu;
1502 
1503 	for_each_possible_cpu(cpu)
1504 		intel_pstate_update_max_freq(all_cpu_data[cpu]);
1505 
1506 	mutex_lock(&hybrid_capacity_lock);
1507 
1508 	if (hybrid_max_perf_cpu)
1509 		__hybrid_refresh_cpu_capacity_scaling();
1510 
1511 	mutex_unlock(&hybrid_capacity_lock);
1512 }
1513 
1514 /************************** sysfs begin ************************/
1515 #define show_one(file_name, object)					\
1516 	static ssize_t show_##file_name					\
1517 	(struct kobject *kobj, struct kobj_attribute *attr, char *buf)	\
1518 	{								\
1519 		return sprintf(buf, "%u\n", global.object);		\
1520 	}
1521 
1522 static ssize_t intel_pstate_show_status(char *buf);
1523 static int intel_pstate_update_status(const char *buf, size_t size);
1524 
1525 static ssize_t show_status(struct kobject *kobj,
1526 			   struct kobj_attribute *attr, char *buf)
1527 {
1528 	guard(mutex)(&intel_pstate_driver_lock);
1529 
1530 	return intel_pstate_show_status(buf);
1531 }
1532 
1533 static ssize_t store_status(struct kobject *a, struct kobj_attribute *b,
1534 			    const char *buf, size_t count)
1535 {
1536 	char *p = memchr(buf, '\n', count);
1537 	int ret;
1538 
1539 	guard(mutex)(&intel_pstate_driver_lock);
1540 
1541 	ret = intel_pstate_update_status(buf, p ? p - buf : count);
1542 	if (ret < 0)
1543 		return ret;
1544 
1545 	return count;
1546 }
1547 
1548 static ssize_t show_turbo_pct(struct kobject *kobj,
1549 				struct kobj_attribute *attr, char *buf)
1550 {
1551 	struct cpudata *cpu;
1552 	int total, no_turbo, turbo_pct;
1553 	uint32_t turbo_fp;
1554 
1555 	guard(mutex)(&intel_pstate_driver_lock);
1556 
1557 	if (!intel_pstate_driver)
1558 		return -EAGAIN;
1559 
1560 	cpu = all_cpu_data[0];
1561 
1562 	total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
1563 	no_turbo = cpu->pstate.max_pstate - cpu->pstate.min_pstate + 1;
1564 	turbo_fp = div_fp(no_turbo, total);
1565 	turbo_pct = 100 - fp_toint(mul_fp(turbo_fp, int_tofp(100)));
1566 
1567 	return sprintf(buf, "%u\n", turbo_pct);
1568 }
1569 
1570 static ssize_t show_num_pstates(struct kobject *kobj,
1571 				struct kobj_attribute *attr, char *buf)
1572 {
1573 	struct cpudata *cpu;
1574 	int total;
1575 
1576 	guard(mutex)(&intel_pstate_driver_lock);
1577 
1578 	if (!intel_pstate_driver)
1579 		return -EAGAIN;
1580 
1581 	cpu = all_cpu_data[0];
1582 	total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
1583 
1584 	return sprintf(buf, "%u\n", total);
1585 }
1586 
1587 static ssize_t show_no_turbo(struct kobject *kobj,
1588 			     struct kobj_attribute *attr, char *buf)
1589 {
1590 	guard(mutex)(&intel_pstate_driver_lock);
1591 
1592 	if (!intel_pstate_driver)
1593 		return -EAGAIN;
1594 
1595 	return sprintf(buf, "%u\n", global.no_turbo);
1596 }
1597 
1598 static ssize_t store_no_turbo(struct kobject *a, struct kobj_attribute *b,
1599 			      const char *buf, size_t count)
1600 {
1601 	unsigned int input;
1602 	bool no_turbo;
1603 
1604 	if (sscanf(buf, "%u", &input) != 1)
1605 		return -EINVAL;
1606 
1607 	guard(mutex)(&intel_pstate_driver_lock);
1608 
1609 	if (!intel_pstate_driver)
1610 		return -EAGAIN;
1611 
1612 	no_turbo = !!clamp_t(int, input, 0, 1);
1613 
1614 	WRITE_ONCE(global.turbo_disabled, turbo_is_disabled());
1615 	if (global.turbo_disabled && !no_turbo) {
1616 		pr_notice("Turbo disabled by BIOS or unavailable on processor\n");
1617 		if (global.no_turbo)
1618 			return -EPERM;
1619 
1620 		no_turbo = 1;
1621 	}
1622 
1623 	if (no_turbo == global.no_turbo)
1624 		return count;
1625 
1626 	WRITE_ONCE(global.no_turbo, no_turbo);
1627 
1628 	mutex_lock(&intel_pstate_limits_lock);
1629 
1630 	if (no_turbo) {
1631 		struct cpudata *cpu = all_cpu_data[0];
1632 		int pct = cpu->pstate.max_pstate * 100 / cpu->pstate.turbo_pstate;
1633 
1634 		/* Squash the global minimum into the permitted range. */
1635 		if (global.min_perf_pct > pct)
1636 			global.min_perf_pct = pct;
1637 	}
1638 
1639 	mutex_unlock(&intel_pstate_limits_lock);
1640 
1641 	intel_pstate_update_limits_for_all();
1642 	arch_set_max_freq_ratio(no_turbo);
1643 
1644 	return count;
1645 }
1646 
1647 static void update_cpu_qos_request(int cpu, enum freq_qos_req_type type)
1648 {
1649 	struct cpudata *cpudata = all_cpu_data[cpu];
1650 	unsigned int freq = cpudata->pstate.turbo_freq;
1651 	struct freq_qos_request *req;
1652 
1653 	struct cpufreq_policy *policy __free(put_cpufreq_policy) = cpufreq_cpu_get(cpu);
1654 	if (!policy)
1655 		return;
1656 
1657 	req = policy->driver_data;
1658 	if (!req)
1659 		return;
1660 
1661 	if (hwp_active)
1662 		intel_pstate_get_hwp_cap(cpudata);
1663 
1664 	if (type == FREQ_QOS_MIN) {
1665 		freq = DIV_ROUND_UP(freq * global.min_perf_pct, 100);
1666 	} else {
1667 		req++;
1668 		freq = (freq * global.max_perf_pct) / 100;
1669 	}
1670 
1671 	if (freq_qos_update_request(req, freq) < 0)
1672 		pr_warn("Failed to update freq constraint: CPU%d\n", cpu);
1673 }
1674 
1675 static void update_qos_requests(enum freq_qos_req_type type)
1676 {
1677 	int i;
1678 
1679 	for_each_possible_cpu(i)
1680 		update_cpu_qos_request(i, type);
1681 }
1682 
1683 static ssize_t store_max_perf_pct(struct kobject *a, struct kobj_attribute *b,
1684 				  const char *buf, size_t count)
1685 {
1686 	unsigned int input;
1687 	int ret;
1688 
1689 	ret = sscanf(buf, "%u", &input);
1690 	if (ret != 1)
1691 		return -EINVAL;
1692 
1693 	guard(mutex)(&intel_pstate_driver_lock);
1694 
1695 	if (!intel_pstate_driver)
1696 		return -EAGAIN;
1697 
1698 	mutex_lock(&intel_pstate_limits_lock);
1699 
1700 	global.max_perf_pct = clamp_t(int, input, global.min_perf_pct, 100);
1701 
1702 	mutex_unlock(&intel_pstate_limits_lock);
1703 
1704 	if (intel_pstate_driver == &intel_pstate)
1705 		intel_pstate_update_policies();
1706 	else
1707 		update_qos_requests(FREQ_QOS_MAX);
1708 
1709 	return count;
1710 }
1711 
1712 static ssize_t store_min_perf_pct(struct kobject *a, struct kobj_attribute *b,
1713 				  const char *buf, size_t count)
1714 {
1715 	unsigned int input;
1716 	int ret;
1717 
1718 	ret = sscanf(buf, "%u", &input);
1719 	if (ret != 1)
1720 		return -EINVAL;
1721 
1722 	guard(mutex)(&intel_pstate_driver_lock);
1723 
1724 	if (!intel_pstate_driver)
1725 		return -EAGAIN;
1726 
1727 	mutex_lock(&intel_pstate_limits_lock);
1728 
1729 	global.min_perf_pct = clamp_t(int, input,
1730 				      min_perf_pct_min(), global.max_perf_pct);
1731 
1732 	mutex_unlock(&intel_pstate_limits_lock);
1733 
1734 	if (intel_pstate_driver == &intel_pstate)
1735 		intel_pstate_update_policies();
1736 	else
1737 		update_qos_requests(FREQ_QOS_MIN);
1738 
1739 	return count;
1740 }
1741 
1742 static ssize_t show_hwp_dynamic_boost(struct kobject *kobj,
1743 				struct kobj_attribute *attr, char *buf)
1744 {
1745 	return sprintf(buf, "%u\n", hwp_boost);
1746 }
1747 
1748 static ssize_t store_hwp_dynamic_boost(struct kobject *a,
1749 				       struct kobj_attribute *b,
1750 				       const char *buf, size_t count)
1751 {
1752 	unsigned int input;
1753 	int ret;
1754 
1755 	ret = kstrtouint(buf, 10, &input);
1756 	if (ret)
1757 		return ret;
1758 
1759 	guard(mutex)(&intel_pstate_driver_lock);
1760 
1761 	hwp_boost = !!input;
1762 	intel_pstate_update_policies();
1763 
1764 	return count;
1765 }
1766 
1767 static ssize_t show_energy_efficiency(struct kobject *kobj, struct kobj_attribute *attr,
1768 				      char *buf)
1769 {
1770 	u64 power_ctl;
1771 	int enable;
1772 
1773 	rdmsrq(MSR_IA32_POWER_CTL, power_ctl);
1774 	enable = !!(power_ctl & BIT(MSR_IA32_POWER_CTL_BIT_EE));
1775 	return sprintf(buf, "%d\n", !enable);
1776 }
1777 
1778 static ssize_t store_energy_efficiency(struct kobject *a, struct kobj_attribute *b,
1779 				       const char *buf, size_t count)
1780 {
1781 	bool input;
1782 	int ret;
1783 
1784 	ret = kstrtobool(buf, &input);
1785 	if (ret)
1786 		return ret;
1787 
1788 	set_power_ctl_ee_state(input);
1789 
1790 	return count;
1791 }
1792 
1793 show_one(max_perf_pct, max_perf_pct);
1794 show_one(min_perf_pct, min_perf_pct);
1795 
1796 define_one_global_rw(status);
1797 define_one_global_rw(no_turbo);
1798 define_one_global_rw(max_perf_pct);
1799 define_one_global_rw(min_perf_pct);
1800 define_one_global_ro(turbo_pct);
1801 define_one_global_ro(num_pstates);
1802 define_one_global_rw(hwp_dynamic_boost);
1803 define_one_global_rw(energy_efficiency);
1804 
1805 static struct attribute *intel_pstate_attributes[] = {
1806 	&status.attr,
1807 	&no_turbo.attr,
1808 	NULL
1809 };
1810 
1811 static const struct attribute_group intel_pstate_attr_group = {
1812 	.attrs = intel_pstate_attributes,
1813 };
1814 
1815 static const struct x86_cpu_id intel_pstate_cpu_ee_disable_ids[];
1816 
1817 static struct kobject *intel_pstate_kobject;
1818 
1819 static void __init intel_pstate_sysfs_expose_params(void)
1820 {
1821 	struct device *dev_root = bus_get_dev_root(&cpu_subsys);
1822 	int rc;
1823 
1824 	if (dev_root) {
1825 		intel_pstate_kobject = kobject_create_and_add("intel_pstate", &dev_root->kobj);
1826 		put_device(dev_root);
1827 	}
1828 	if (WARN_ON(!intel_pstate_kobject))
1829 		return;
1830 
1831 	rc = sysfs_create_group(intel_pstate_kobject, &intel_pstate_attr_group);
1832 	if (WARN_ON(rc))
1833 		return;
1834 
1835 	if (!boot_cpu_has(X86_FEATURE_HYBRID_CPU)) {
1836 		rc = sysfs_create_file(intel_pstate_kobject, &turbo_pct.attr);
1837 		WARN_ON(rc);
1838 
1839 		rc = sysfs_create_file(intel_pstate_kobject, &num_pstates.attr);
1840 		WARN_ON(rc);
1841 	}
1842 
1843 	/*
1844 	 * If per cpu limits are enforced there are no global limits, so
1845 	 * return without creating max/min_perf_pct attributes
1846 	 */
1847 	if (per_cpu_limits)
1848 		return;
1849 
1850 	rc = sysfs_create_file(intel_pstate_kobject, &max_perf_pct.attr);
1851 	WARN_ON(rc);
1852 
1853 	rc = sysfs_create_file(intel_pstate_kobject, &min_perf_pct.attr);
1854 	WARN_ON(rc);
1855 
1856 	if (x86_match_cpu(intel_pstate_cpu_ee_disable_ids)) {
1857 		rc = sysfs_create_file(intel_pstate_kobject, &energy_efficiency.attr);
1858 		WARN_ON(rc);
1859 	}
1860 }
1861 
1862 static void __init intel_pstate_sysfs_remove(void)
1863 {
1864 	if (!intel_pstate_kobject)
1865 		return;
1866 
1867 	sysfs_remove_group(intel_pstate_kobject, &intel_pstate_attr_group);
1868 
1869 	if (!boot_cpu_has(X86_FEATURE_HYBRID_CPU)) {
1870 		sysfs_remove_file(intel_pstate_kobject, &num_pstates.attr);
1871 		sysfs_remove_file(intel_pstate_kobject, &turbo_pct.attr);
1872 	}
1873 
1874 	if (!per_cpu_limits) {
1875 		sysfs_remove_file(intel_pstate_kobject, &max_perf_pct.attr);
1876 		sysfs_remove_file(intel_pstate_kobject, &min_perf_pct.attr);
1877 
1878 		if (x86_match_cpu(intel_pstate_cpu_ee_disable_ids))
1879 			sysfs_remove_file(intel_pstate_kobject, &energy_efficiency.attr);
1880 	}
1881 
1882 	kobject_put(intel_pstate_kobject);
1883 }
1884 
1885 static void intel_pstate_sysfs_expose_hwp_dynamic_boost(void)
1886 {
1887 	int rc;
1888 
1889 	if (!hwp_active)
1890 		return;
1891 
1892 	rc = sysfs_create_file(intel_pstate_kobject, &hwp_dynamic_boost.attr);
1893 	WARN_ON_ONCE(rc);
1894 }
1895 
1896 static void intel_pstate_sysfs_hide_hwp_dynamic_boost(void)
1897 {
1898 	if (!hwp_active)
1899 		return;
1900 
1901 	sysfs_remove_file(intel_pstate_kobject, &hwp_dynamic_boost.attr);
1902 }
1903 
1904 /************************** sysfs end ************************/
1905 
1906 static void intel_pstate_notify_work(struct work_struct *work)
1907 {
1908 	struct cpudata *cpudata =
1909 		container_of(to_delayed_work(work), struct cpudata, hwp_notify_work);
1910 
1911 	if (intel_pstate_update_max_freq(cpudata)) {
1912 		/*
1913 		 * The driver will not be unregistered while this function is
1914 		 * running, so update the capacity without acquiring the driver
1915 		 * lock.
1916 		 */
1917 		hybrid_update_capacity(cpudata);
1918 	}
1919 
1920 	wrmsrq_on_cpu(cpudata->cpu, MSR_HWP_STATUS, 0);
1921 }
1922 
1923 static DEFINE_RAW_SPINLOCK(hwp_notify_lock);
1924 static cpumask_t hwp_intr_enable_mask;
1925 
1926 #define HWP_GUARANTEED_PERF_CHANGE_STATUS      BIT(0)
1927 #define HWP_HIGHEST_PERF_CHANGE_STATUS         BIT(3)
1928 
1929 void notify_hwp_interrupt(void)
1930 {
1931 	unsigned int this_cpu = smp_processor_id();
1932 	u64 value, status_mask;
1933 	unsigned long flags;
1934 
1935 	if (!hwp_active || !cpu_feature_enabled(X86_FEATURE_HWP_NOTIFY))
1936 		return;
1937 
1938 	status_mask = HWP_GUARANTEED_PERF_CHANGE_STATUS;
1939 	if (cpu_feature_enabled(X86_FEATURE_HWP_HIGHEST_PERF_CHANGE))
1940 		status_mask |= HWP_HIGHEST_PERF_CHANGE_STATUS;
1941 
1942 	rdmsrq_safe(MSR_HWP_STATUS, &value);
1943 	if (!(value & status_mask))
1944 		return;
1945 
1946 	raw_spin_lock_irqsave(&hwp_notify_lock, flags);
1947 
1948 	if (!cpumask_test_cpu(this_cpu, &hwp_intr_enable_mask))
1949 		goto ack_intr;
1950 
1951 	schedule_delayed_work(&all_cpu_data[this_cpu]->hwp_notify_work,
1952 			      msecs_to_jiffies(10));
1953 
1954 	raw_spin_unlock_irqrestore(&hwp_notify_lock, flags);
1955 
1956 	return;
1957 
1958 ack_intr:
1959 	wrmsrq_safe(MSR_HWP_STATUS, 0);
1960 	raw_spin_unlock_irqrestore(&hwp_notify_lock, flags);
1961 }
1962 
1963 static void intel_pstate_disable_hwp_interrupt(struct cpudata *cpudata)
1964 {
1965 	bool cancel_work;
1966 
1967 	if (!cpu_feature_enabled(X86_FEATURE_HWP_NOTIFY))
1968 		return;
1969 
1970 	/* wrmsrq_on_cpu has to be outside spinlock as this can result in IPC */
1971 	wrmsrq_on_cpu(cpudata->cpu, MSR_HWP_INTERRUPT, 0x00);
1972 
1973 	raw_spin_lock_irq(&hwp_notify_lock);
1974 	cancel_work = cpumask_test_and_clear_cpu(cpudata->cpu, &hwp_intr_enable_mask);
1975 	raw_spin_unlock_irq(&hwp_notify_lock);
1976 
1977 	if (cancel_work)
1978 		cancel_delayed_work_sync(&cpudata->hwp_notify_work);
1979 }
1980 
1981 #define HWP_GUARANTEED_PERF_CHANGE_REQ BIT(0)
1982 #define HWP_HIGHEST_PERF_CHANGE_REQ    BIT(2)
1983 
1984 static void intel_pstate_enable_hwp_interrupt(struct cpudata *cpudata)
1985 {
1986 	/* Enable HWP notification interrupt for performance change */
1987 	if (boot_cpu_has(X86_FEATURE_HWP_NOTIFY)) {
1988 		u64 interrupt_mask = HWP_GUARANTEED_PERF_CHANGE_REQ;
1989 
1990 		raw_spin_lock_irq(&hwp_notify_lock);
1991 		INIT_DELAYED_WORK(&cpudata->hwp_notify_work, intel_pstate_notify_work);
1992 		cpumask_set_cpu(cpudata->cpu, &hwp_intr_enable_mask);
1993 		raw_spin_unlock_irq(&hwp_notify_lock);
1994 
1995 		if (cpu_feature_enabled(X86_FEATURE_HWP_HIGHEST_PERF_CHANGE))
1996 			interrupt_mask |= HWP_HIGHEST_PERF_CHANGE_REQ;
1997 
1998 		/* wrmsrq_on_cpu has to be outside spinlock as this can result in IPC */
1999 		wrmsrq_on_cpu(cpudata->cpu, MSR_HWP_INTERRUPT, interrupt_mask);
2000 		wrmsrq_on_cpu(cpudata->cpu, MSR_HWP_STATUS, 0);
2001 	}
2002 }
2003 
2004 static void intel_pstate_update_epp_defaults(struct cpudata *cpudata)
2005 {
2006 	cpudata->epp_default = intel_pstate_get_epp(cpudata, 0);
2007 
2008 	/*
2009 	 * If the EPP is set by firmware, which means that firmware enabled HWP
2010 	 * - Is equal or less than 0x80 (default balance_perf EPP)
2011 	 * - But less performance oriented than performance EPP
2012 	 *   then use this as new balance_perf EPP.
2013 	 */
2014 	if (hwp_forced && cpudata->epp_default <= HWP_EPP_BALANCE_PERFORMANCE &&
2015 	    cpudata->epp_default > HWP_EPP_PERFORMANCE) {
2016 		epp_values[EPP_INDEX_BALANCE_PERFORMANCE] = cpudata->epp_default;
2017 		return;
2018 	}
2019 
2020 	/*
2021 	 * If this CPU gen doesn't call for change in balance_perf
2022 	 * EPP return.
2023 	 */
2024 	if (epp_values[EPP_INDEX_BALANCE_PERFORMANCE] == HWP_EPP_BALANCE_PERFORMANCE)
2025 		return;
2026 
2027 	/*
2028 	 * Use hard coded value per gen to update the balance_perf
2029 	 * and default EPP.
2030 	 */
2031 	cpudata->epp_default = epp_values[EPP_INDEX_BALANCE_PERFORMANCE];
2032 	intel_pstate_set_epp(cpudata, cpudata->epp_default);
2033 }
2034 
2035 static void intel_pstate_hwp_enable(struct cpudata *cpudata)
2036 {
2037 	/* First disable HWP notification interrupt till we activate again */
2038 	if (boot_cpu_has(X86_FEATURE_HWP_NOTIFY))
2039 		wrmsrq_on_cpu(cpudata->cpu, MSR_HWP_INTERRUPT, 0x00);
2040 
2041 	wrmsrq_on_cpu(cpudata->cpu, MSR_PM_ENABLE, 0x1);
2042 
2043 	intel_pstate_enable_hwp_interrupt(cpudata);
2044 
2045 	if (cpudata->epp_default >= 0)
2046 		return;
2047 
2048 	intel_pstate_update_epp_defaults(cpudata);
2049 }
2050 
2051 static u64 get_perf_ctl_val(int pstate)
2052 {
2053 	u64 val;
2054 
2055 	val = (u64)pstate << 8;
2056 	if (READ_ONCE(global.no_turbo) && !READ_ONCE(global.turbo_disabled) &&
2057 	    cpu_feature_enabled(X86_FEATURE_IDA))
2058 		val |= (u64)1 << 32;
2059 
2060 	return val;
2061 }
2062 
2063 static int atom_get_min_pstate(int not_used)
2064 {
2065 	u64 value;
2066 
2067 	rdmsrq(MSR_ATOM_CORE_RATIOS, value);
2068 	return (value >> 8) & 0x7F;
2069 }
2070 
2071 static int atom_get_max_pstate(int not_used)
2072 {
2073 	u64 value;
2074 
2075 	rdmsrq(MSR_ATOM_CORE_RATIOS, value);
2076 	return (value >> 16) & 0x7F;
2077 }
2078 
2079 static int atom_get_turbo_pstate(int not_used)
2080 {
2081 	u64 value;
2082 
2083 	rdmsrq(MSR_ATOM_CORE_TURBO_RATIOS, value);
2084 	return value & 0x7F;
2085 }
2086 
2087 static u64 atom_get_val(struct cpudata *cpudata, int pstate)
2088 {
2089 	u64 val = get_perf_ctl_val(pstate);
2090 	int32_t vid_fp;
2091 	u32 vid;
2092 
2093 	vid_fp = cpudata->vid.min + mul_fp(
2094 		int_tofp(pstate - cpudata->pstate.min_pstate),
2095 		cpudata->vid.ratio);
2096 
2097 	vid_fp = clamp_t(int32_t, vid_fp, cpudata->vid.min, cpudata->vid.max);
2098 	vid = ceiling_fp(vid_fp);
2099 
2100 	if (pstate > cpudata->pstate.max_pstate)
2101 		vid = cpudata->vid.turbo;
2102 
2103 	return val | vid;
2104 }
2105 
2106 static int silvermont_get_scaling(void)
2107 {
2108 	u64 value;
2109 	int i;
2110 	/* Defined in Table 35-6 from SDM (Sept 2015) */
2111 	static int silvermont_freq_table[] = {
2112 		83300, 100000, 133300, 116700, 80000};
2113 
2114 	rdmsrq(MSR_FSB_FREQ, value);
2115 	i = value & 0x7;
2116 	WARN_ON(i > 4);
2117 
2118 	return silvermont_freq_table[i];
2119 }
2120 
2121 static int airmont_get_scaling(void)
2122 {
2123 	u64 value;
2124 	int i;
2125 	/* Defined in Table 35-10 from SDM (Sept 2015) */
2126 	static int airmont_freq_table[] = {
2127 		83300, 100000, 133300, 116700, 80000,
2128 		93300, 90000, 88900, 87500};
2129 
2130 	rdmsrq(MSR_FSB_FREQ, value);
2131 	i = value & 0xF;
2132 	WARN_ON(i > 8);
2133 
2134 	return airmont_freq_table[i];
2135 }
2136 
2137 static void atom_get_vid(struct cpudata *cpudata)
2138 {
2139 	u64 value;
2140 
2141 	rdmsrq(MSR_ATOM_CORE_VIDS, value);
2142 	cpudata->vid.min = int_tofp((value >> 8) & 0x7f);
2143 	cpudata->vid.max = int_tofp((value >> 16) & 0x7f);
2144 	cpudata->vid.ratio = div_fp(
2145 		cpudata->vid.max - cpudata->vid.min,
2146 		int_tofp(cpudata->pstate.max_pstate -
2147 			cpudata->pstate.min_pstate));
2148 
2149 	rdmsrq(MSR_ATOM_CORE_TURBO_VIDS, value);
2150 	cpudata->vid.turbo = value & 0x7f;
2151 }
2152 
2153 static int core_get_min_pstate(int cpu)
2154 {
2155 	u64 value;
2156 
2157 	rdmsrq_on_cpu(cpu, MSR_PLATFORM_INFO, &value);
2158 	return (value >> 40) & 0xFF;
2159 }
2160 
2161 static int core_get_max_pstate_physical(int cpu)
2162 {
2163 	u64 value;
2164 
2165 	rdmsrq_on_cpu(cpu, MSR_PLATFORM_INFO, &value);
2166 	return (value >> 8) & 0xFF;
2167 }
2168 
2169 static int core_get_tdp_ratio(int cpu, u64 plat_info)
2170 {
2171 	/* Check how many TDP levels present */
2172 	if (plat_info & 0x600000000) {
2173 		u64 tdp_ctrl;
2174 		u64 tdp_ratio;
2175 		int tdp_msr;
2176 		int err;
2177 
2178 		/* Get the TDP level (0, 1, 2) to get ratios */
2179 		err = rdmsrq_safe_on_cpu(cpu, MSR_CONFIG_TDP_CONTROL, &tdp_ctrl);
2180 		if (err)
2181 			return err;
2182 
2183 		/* TDP MSR are continuous starting at 0x648 */
2184 		tdp_msr = MSR_CONFIG_TDP_NOMINAL + (tdp_ctrl & 0x03);
2185 		err = rdmsrq_safe_on_cpu(cpu, tdp_msr, &tdp_ratio);
2186 		if (err)
2187 			return err;
2188 
2189 		/* For level 1 and 2, bits[23:16] contain the ratio */
2190 		if (tdp_ctrl & 0x03)
2191 			tdp_ratio >>= 16;
2192 
2193 		tdp_ratio &= 0xff; /* ratios are only 8 bits long */
2194 		pr_debug("tdp_ratio %x\n", (int)tdp_ratio);
2195 
2196 		return (int)tdp_ratio;
2197 	}
2198 
2199 	return -ENXIO;
2200 }
2201 
2202 static int core_get_max_pstate(int cpu)
2203 {
2204 	u64 tar;
2205 	u64 plat_info;
2206 	int max_pstate;
2207 	int tdp_ratio;
2208 	int err;
2209 
2210 	rdmsrq_on_cpu(cpu, MSR_PLATFORM_INFO, &plat_info);
2211 	max_pstate = (plat_info >> 8) & 0xFF;
2212 
2213 	tdp_ratio = core_get_tdp_ratio(cpu, plat_info);
2214 	if (tdp_ratio <= 0)
2215 		return max_pstate;
2216 
2217 	if (hwp_active) {
2218 		/* Turbo activation ratio is not used on HWP platforms */
2219 		return tdp_ratio;
2220 	}
2221 
2222 	err = rdmsrq_safe_on_cpu(cpu, MSR_TURBO_ACTIVATION_RATIO, &tar);
2223 	if (!err) {
2224 		int tar_levels;
2225 
2226 		/* Do some sanity checking for safety */
2227 		tar_levels = tar & 0xff;
2228 		if (tdp_ratio - 1 == tar_levels) {
2229 			max_pstate = tar_levels;
2230 			pr_debug("max_pstate=TAC %x\n", max_pstate);
2231 		}
2232 	}
2233 
2234 	return max_pstate;
2235 }
2236 
2237 static int core_get_turbo_pstate(int cpu)
2238 {
2239 	u64 value;
2240 	int nont, ret;
2241 
2242 	rdmsrq_on_cpu(cpu, MSR_TURBO_RATIO_LIMIT, &value);
2243 	nont = core_get_max_pstate(cpu);
2244 	ret = (value) & 255;
2245 	if (ret <= nont)
2246 		ret = nont;
2247 	return ret;
2248 }
2249 
2250 static u64 core_get_val(struct cpudata *cpudata, int pstate)
2251 {
2252 	return get_perf_ctl_val(pstate);
2253 }
2254 
2255 static int knl_get_aperf_mperf_shift(void)
2256 {
2257 	return 10;
2258 }
2259 
2260 static int knl_get_turbo_pstate(int cpu)
2261 {
2262 	u64 value;
2263 	int nont, ret;
2264 
2265 	rdmsrq_on_cpu(cpu, MSR_TURBO_RATIO_LIMIT, &value);
2266 	nont = core_get_max_pstate(cpu);
2267 	ret = (((value) >> 8) & 0xFF);
2268 	if (ret <= nont)
2269 		ret = nont;
2270 	return ret;
2271 }
2272 
2273 static int hwp_get_cpu_scaling(int cpu)
2274 {
2275 	if (hybrid_scaling_factor) {
2276 		/*
2277 		 * Return the hybrid scaling factor for P-cores and use the
2278 		 * default core scaling for E-cores.
2279 		 */
2280 		if (hybrid_get_cpu_type(cpu) == INTEL_CPU_TYPE_CORE)
2281 			return hybrid_scaling_factor;
2282 
2283 		return core_get_scaling();
2284 	}
2285 
2286 	/* Use core scaling on non-hybrid systems. */
2287 	if (!cpu_feature_enabled(X86_FEATURE_HYBRID_CPU))
2288 		return core_get_scaling();
2289 
2290 	/*
2291 	 * The system is hybrid, but the hybrid scaling factor is not known or
2292 	 * the CPU type is not one of the above, so use CPPC to compute the
2293 	 * scaling factor for this CPU.
2294 	 */
2295 	return intel_pstate_cppc_get_scaling(cpu);
2296 }
2297 
2298 static void intel_pstate_set_pstate(struct cpudata *cpu, int pstate)
2299 {
2300 	trace_cpu_frequency(pstate * cpu->pstate.scaling, cpu->cpu);
2301 	cpu->pstate.current_pstate = pstate;
2302 	/*
2303 	 * Generally, there is no guarantee that this code will always run on
2304 	 * the CPU being updated, so force the register update to run on the
2305 	 * right CPU.
2306 	 */
2307 	wrmsrq_on_cpu(cpu->cpu, MSR_IA32_PERF_CTL,
2308 		      pstate_funcs.get_val(cpu, pstate));
2309 }
2310 
2311 static void intel_pstate_set_min_pstate(struct cpudata *cpu)
2312 {
2313 	intel_pstate_set_pstate(cpu, cpu->pstate.min_pstate);
2314 }
2315 
2316 static void intel_pstate_get_cpu_pstates(struct cpudata *cpu)
2317 {
2318 	int perf_ctl_scaling = pstate_funcs.get_scaling();
2319 
2320 	cpu->pstate.max_pstate_physical = pstate_funcs.get_max_physical(cpu->cpu);
2321 	cpu->pstate.min_pstate = pstate_funcs.get_min(cpu->cpu);
2322 	cpu->pstate.perf_ctl_scaling = perf_ctl_scaling;
2323 
2324 	if (hwp_active && !hwp_mode_bdw) {
2325 		__intel_pstate_get_hwp_cap(cpu);
2326 
2327 		if (pstate_funcs.get_cpu_scaling) {
2328 			cpu->pstate.scaling = pstate_funcs.get_cpu_scaling(cpu->cpu);
2329 			intel_pstate_hybrid_hwp_adjust(cpu);
2330 		} else {
2331 			cpu->pstate.scaling = perf_ctl_scaling;
2332 		}
2333 		/*
2334 		 * If the CPU is going online for the first time and it was
2335 		 * offline initially, asym capacity scaling needs to be updated.
2336 		 */
2337 		hybrid_update_capacity(cpu);
2338 	} else {
2339 		cpu->pstate.scaling = perf_ctl_scaling;
2340 		cpu->pstate.max_pstate = pstate_funcs.get_max(cpu->cpu);
2341 		cpu->pstate.turbo_pstate = pstate_funcs.get_turbo(cpu->cpu);
2342 	}
2343 
2344 	if (cpu->pstate.scaling == perf_ctl_scaling) {
2345 		cpu->pstate.min_freq = cpu->pstate.min_pstate * perf_ctl_scaling;
2346 		cpu->pstate.max_freq = cpu->pstate.max_pstate * perf_ctl_scaling;
2347 		cpu->pstate.turbo_freq = cpu->pstate.turbo_pstate * perf_ctl_scaling;
2348 	}
2349 
2350 	if (pstate_funcs.get_aperf_mperf_shift)
2351 		cpu->aperf_mperf_shift = pstate_funcs.get_aperf_mperf_shift();
2352 
2353 	if (pstate_funcs.get_vid)
2354 		pstate_funcs.get_vid(cpu);
2355 
2356 	intel_pstate_set_min_pstate(cpu);
2357 }
2358 
2359 /*
2360  * Long hold time will keep high perf limits for long time,
2361  * which negatively impacts perf/watt for some workloads,
2362  * like specpower. 3ms is based on experiements on some
2363  * workoads.
2364  */
2365 static int hwp_boost_hold_time_ns = 3 * NSEC_PER_MSEC;
2366 
2367 static inline void intel_pstate_hwp_boost_up(struct cpudata *cpu)
2368 {
2369 	u64 hwp_req = READ_ONCE(cpu->hwp_req_cached);
2370 	u64 hwp_cap = READ_ONCE(cpu->hwp_cap_cached);
2371 	u32 max_limit = (hwp_req & 0xff00) >> 8;
2372 	u32 min_limit = (hwp_req & 0xff);
2373 	u32 boost_level1;
2374 
2375 	/*
2376 	 * Cases to consider (User changes via sysfs or boot time):
2377 	 * If, P0 (Turbo max) = P1 (Guaranteed max) = min:
2378 	 *	No boost, return.
2379 	 * If, P0 (Turbo max) > P1 (Guaranteed max) = min:
2380 	 *     Should result in one level boost only for P0.
2381 	 * If, P0 (Turbo max) = P1 (Guaranteed max) > min:
2382 	 *     Should result in two level boost:
2383 	 *         (min + p1)/2 and P1.
2384 	 * If, P0 (Turbo max) > P1 (Guaranteed max) > min:
2385 	 *     Should result in three level boost:
2386 	 *        (min + p1)/2, P1 and P0.
2387 	 */
2388 
2389 	/* If max and min are equal or already at max, nothing to boost */
2390 	if (max_limit == min_limit || cpu->hwp_boost_min >= max_limit)
2391 		return;
2392 
2393 	if (!cpu->hwp_boost_min)
2394 		cpu->hwp_boost_min = min_limit;
2395 
2396 	/* level at half way mark between min and guranteed */
2397 	boost_level1 = (HWP_GUARANTEED_PERF(hwp_cap) + min_limit) >> 1;
2398 
2399 	if (cpu->hwp_boost_min < boost_level1)
2400 		cpu->hwp_boost_min = boost_level1;
2401 	else if (cpu->hwp_boost_min < HWP_GUARANTEED_PERF(hwp_cap))
2402 		cpu->hwp_boost_min = HWP_GUARANTEED_PERF(hwp_cap);
2403 	else if (cpu->hwp_boost_min == HWP_GUARANTEED_PERF(hwp_cap) &&
2404 		 max_limit != HWP_GUARANTEED_PERF(hwp_cap))
2405 		cpu->hwp_boost_min = max_limit;
2406 	else
2407 		return;
2408 
2409 	hwp_req = (hwp_req & ~GENMASK_ULL(7, 0)) | cpu->hwp_boost_min;
2410 	wrmsrq(MSR_HWP_REQUEST, hwp_req);
2411 	cpu->last_update = cpu->sample.time;
2412 }
2413 
2414 static inline void intel_pstate_hwp_boost_down(struct cpudata *cpu)
2415 {
2416 	if (cpu->hwp_boost_min) {
2417 		bool expired;
2418 
2419 		/* Check if we are idle for hold time to boost down */
2420 		expired = time_after64(cpu->sample.time, cpu->last_update +
2421 				       hwp_boost_hold_time_ns);
2422 		if (expired) {
2423 			wrmsrq(MSR_HWP_REQUEST, cpu->hwp_req_cached);
2424 			cpu->hwp_boost_min = 0;
2425 		}
2426 	}
2427 	cpu->last_update = cpu->sample.time;
2428 }
2429 
2430 static inline void intel_pstate_update_util_hwp_local(struct cpudata *cpu,
2431 						      u64 time)
2432 {
2433 	cpu->sample.time = time;
2434 
2435 	if (cpu->sched_flags & SCHED_CPUFREQ_IOWAIT) {
2436 		bool do_io = false;
2437 
2438 		cpu->sched_flags = 0;
2439 		/*
2440 		 * Set iowait_boost flag and update time. Since IO WAIT flag
2441 		 * is set all the time, we can't just conclude that there is
2442 		 * some IO bound activity is scheduled on this CPU with just
2443 		 * one occurrence. If we receive at least two in two
2444 		 * consecutive ticks, then we treat as boost candidate.
2445 		 */
2446 		if (time_before64(time, cpu->last_io_update + 2 * TICK_NSEC))
2447 			do_io = true;
2448 
2449 		cpu->last_io_update = time;
2450 
2451 		if (do_io)
2452 			intel_pstate_hwp_boost_up(cpu);
2453 
2454 	} else {
2455 		intel_pstate_hwp_boost_down(cpu);
2456 	}
2457 }
2458 
2459 static inline void intel_pstate_update_util_hwp(struct update_util_data *data,
2460 						u64 time, unsigned int flags)
2461 {
2462 	struct cpudata *cpu = container_of(data, struct cpudata, update_util);
2463 
2464 	cpu->sched_flags |= flags;
2465 
2466 	if (smp_processor_id() == cpu->cpu)
2467 		intel_pstate_update_util_hwp_local(cpu, time);
2468 }
2469 
2470 static inline void intel_pstate_calc_avg_perf(struct cpudata *cpu)
2471 {
2472 	struct sample *sample = &cpu->sample;
2473 
2474 	sample->core_avg_perf = div_ext_fp(sample->aperf, sample->mperf);
2475 }
2476 
2477 static inline bool intel_pstate_sample(struct cpudata *cpu, u64 time)
2478 {
2479 	u64 aperf, mperf;
2480 	unsigned long flags;
2481 	u64 tsc;
2482 
2483 	local_irq_save(flags);
2484 	rdmsrq(MSR_IA32_APERF, aperf);
2485 	rdmsrq(MSR_IA32_MPERF, mperf);
2486 	tsc = rdtsc();
2487 	if (cpu->prev_mperf == mperf || cpu->prev_tsc == tsc) {
2488 		local_irq_restore(flags);
2489 		return false;
2490 	}
2491 	local_irq_restore(flags);
2492 
2493 	cpu->last_sample_time = cpu->sample.time;
2494 	cpu->sample.time = time;
2495 	cpu->sample.aperf = aperf;
2496 	cpu->sample.mperf = mperf;
2497 	cpu->sample.tsc =  tsc;
2498 	cpu->sample.aperf -= cpu->prev_aperf;
2499 	cpu->sample.mperf -= cpu->prev_mperf;
2500 	cpu->sample.tsc -= cpu->prev_tsc;
2501 
2502 	cpu->prev_aperf = aperf;
2503 	cpu->prev_mperf = mperf;
2504 	cpu->prev_tsc = tsc;
2505 	/*
2506 	 * First time this function is invoked in a given cycle, all of the
2507 	 * previous sample data fields are equal to zero or stale and they must
2508 	 * be populated with meaningful numbers for things to work, so assume
2509 	 * that sample.time will always be reset before setting the utilization
2510 	 * update hook and make the caller skip the sample then.
2511 	 */
2512 	if (likely(cpu->last_sample_time)) {
2513 		intel_pstate_calc_avg_perf(cpu);
2514 		return true;
2515 	}
2516 	return false;
2517 }
2518 
2519 static inline int32_t get_avg_frequency(struct cpudata *cpu)
2520 {
2521 	return mul_ext_fp(cpu->sample.core_avg_perf, cpu_khz);
2522 }
2523 
2524 static inline int32_t get_avg_pstate(struct cpudata *cpu)
2525 {
2526 	return mul_ext_fp(cpu->pstate.max_pstate_physical,
2527 			  cpu->sample.core_avg_perf);
2528 }
2529 
2530 static inline int32_t get_target_pstate(struct cpudata *cpu)
2531 {
2532 	struct sample *sample = &cpu->sample;
2533 	int32_t busy_frac;
2534 	int target, avg_pstate;
2535 
2536 	busy_frac = div_fp(sample->mperf << cpu->aperf_mperf_shift,
2537 			   sample->tsc);
2538 
2539 	if (busy_frac < cpu->iowait_boost)
2540 		busy_frac = cpu->iowait_boost;
2541 
2542 	sample->busy_scaled = busy_frac * 100;
2543 
2544 	target = READ_ONCE(global.no_turbo) ?
2545 			cpu->pstate.max_pstate : cpu->pstate.turbo_pstate;
2546 	target += target >> 2;
2547 	target = mul_fp(target, busy_frac);
2548 	if (target < cpu->pstate.min_pstate)
2549 		target = cpu->pstate.min_pstate;
2550 
2551 	/*
2552 	 * If the average P-state during the previous cycle was higher than the
2553 	 * current target, add 50% of the difference to the target to reduce
2554 	 * possible performance oscillations and offset possible performance
2555 	 * loss related to moving the workload from one CPU to another within
2556 	 * a package/module.
2557 	 */
2558 	avg_pstate = get_avg_pstate(cpu);
2559 	if (avg_pstate > target)
2560 		target += (avg_pstate - target) >> 1;
2561 
2562 	return target;
2563 }
2564 
2565 static int intel_pstate_prepare_request(struct cpudata *cpu, int pstate)
2566 {
2567 	int min_pstate = max(cpu->pstate.min_pstate, cpu->min_perf_ratio);
2568 	int max_pstate = max(min_pstate, cpu->max_perf_ratio);
2569 
2570 	return clamp_t(int, pstate, min_pstate, max_pstate);
2571 }
2572 
2573 static void intel_pstate_update_pstate(struct cpudata *cpu, int pstate)
2574 {
2575 	if (pstate == cpu->pstate.current_pstate)
2576 		return;
2577 
2578 	cpu->pstate.current_pstate = pstate;
2579 	wrmsrq(MSR_IA32_PERF_CTL, pstate_funcs.get_val(cpu, pstate));
2580 }
2581 
2582 static void intel_pstate_adjust_pstate(struct cpudata *cpu)
2583 {
2584 	int from = cpu->pstate.current_pstate;
2585 	struct sample *sample;
2586 	int target_pstate;
2587 
2588 	target_pstate = get_target_pstate(cpu);
2589 	target_pstate = intel_pstate_prepare_request(cpu, target_pstate);
2590 	trace_cpu_frequency(target_pstate * cpu->pstate.scaling, cpu->cpu);
2591 	intel_pstate_update_pstate(cpu, target_pstate);
2592 
2593 	sample = &cpu->sample;
2594 	trace_pstate_sample(mul_ext_fp(100, sample->core_avg_perf),
2595 		fp_toint(sample->busy_scaled),
2596 		from,
2597 		cpu->pstate.current_pstate,
2598 		sample->mperf,
2599 		sample->aperf,
2600 		sample->tsc,
2601 		get_avg_frequency(cpu),
2602 		fp_toint(cpu->iowait_boost * 100));
2603 }
2604 
2605 static void intel_pstate_update_util(struct update_util_data *data, u64 time,
2606 				     unsigned int flags)
2607 {
2608 	struct cpudata *cpu = container_of(data, struct cpudata, update_util);
2609 	u64 delta_ns;
2610 
2611 	/* Don't allow remote callbacks */
2612 	if (smp_processor_id() != cpu->cpu)
2613 		return;
2614 
2615 	delta_ns = time - cpu->last_update;
2616 	if (flags & SCHED_CPUFREQ_IOWAIT) {
2617 		/* Start over if the CPU may have been idle. */
2618 		if (delta_ns > TICK_NSEC) {
2619 			cpu->iowait_boost = ONE_EIGHTH_FP;
2620 		} else if (cpu->iowait_boost >= ONE_EIGHTH_FP) {
2621 			cpu->iowait_boost <<= 1;
2622 			if (cpu->iowait_boost > int_tofp(1))
2623 				cpu->iowait_boost = int_tofp(1);
2624 		} else {
2625 			cpu->iowait_boost = ONE_EIGHTH_FP;
2626 		}
2627 	} else if (cpu->iowait_boost) {
2628 		/* Clear iowait_boost if the CPU may have been idle. */
2629 		if (delta_ns > TICK_NSEC)
2630 			cpu->iowait_boost = 0;
2631 		else
2632 			cpu->iowait_boost >>= 1;
2633 	}
2634 	cpu->last_update = time;
2635 	delta_ns = time - cpu->sample.time;
2636 	if ((s64)delta_ns < INTEL_PSTATE_SAMPLING_INTERVAL)
2637 		return;
2638 
2639 	if (intel_pstate_sample(cpu, time))
2640 		intel_pstate_adjust_pstate(cpu);
2641 }
2642 
2643 static struct pstate_funcs core_funcs = {
2644 	.get_max = core_get_max_pstate,
2645 	.get_max_physical = core_get_max_pstate_physical,
2646 	.get_min = core_get_min_pstate,
2647 	.get_turbo = core_get_turbo_pstate,
2648 	.get_scaling = core_get_scaling,
2649 	.get_val = core_get_val,
2650 };
2651 
2652 static const struct pstate_funcs silvermont_funcs = {
2653 	.get_max = atom_get_max_pstate,
2654 	.get_max_physical = atom_get_max_pstate,
2655 	.get_min = atom_get_min_pstate,
2656 	.get_turbo = atom_get_turbo_pstate,
2657 	.get_val = atom_get_val,
2658 	.get_scaling = silvermont_get_scaling,
2659 	.get_vid = atom_get_vid,
2660 };
2661 
2662 static const struct pstate_funcs airmont_funcs = {
2663 	.get_max = atom_get_max_pstate,
2664 	.get_max_physical = atom_get_max_pstate,
2665 	.get_min = atom_get_min_pstate,
2666 	.get_turbo = atom_get_turbo_pstate,
2667 	.get_val = atom_get_val,
2668 	.get_scaling = airmont_get_scaling,
2669 	.get_vid = atom_get_vid,
2670 };
2671 
2672 static const struct pstate_funcs knl_funcs = {
2673 	.get_max = core_get_max_pstate,
2674 	.get_max_physical = core_get_max_pstate_physical,
2675 	.get_min = core_get_min_pstate,
2676 	.get_turbo = knl_get_turbo_pstate,
2677 	.get_aperf_mperf_shift = knl_get_aperf_mperf_shift,
2678 	.get_scaling = core_get_scaling,
2679 	.get_val = core_get_val,
2680 };
2681 
2682 #define X86_MATCH(vfm, policy)					 \
2683 	X86_MATCH_VFM_FEATURE(vfm, X86_FEATURE_APERFMPERF, &policy)
2684 
2685 static const struct x86_cpu_id intel_pstate_cpu_ids[] = {
2686 	X86_MATCH(INTEL_SANDYBRIDGE,		core_funcs),
2687 	X86_MATCH(INTEL_SANDYBRIDGE_X,		core_funcs),
2688 	X86_MATCH(INTEL_ATOM_SILVERMONT,	silvermont_funcs),
2689 	X86_MATCH(INTEL_IVYBRIDGE,		core_funcs),
2690 	X86_MATCH(INTEL_HASWELL,		core_funcs),
2691 	X86_MATCH(INTEL_BROADWELL,		core_funcs),
2692 	X86_MATCH(INTEL_IVYBRIDGE_X,		core_funcs),
2693 	X86_MATCH(INTEL_HASWELL_X,		core_funcs),
2694 	X86_MATCH(INTEL_HASWELL_L,		core_funcs),
2695 	X86_MATCH(INTEL_HASWELL_G,		core_funcs),
2696 	X86_MATCH(INTEL_BROADWELL_G,		core_funcs),
2697 	X86_MATCH(INTEL_ATOM_AIRMONT,		airmont_funcs),
2698 	X86_MATCH(INTEL_SKYLAKE_L,		core_funcs),
2699 	X86_MATCH(INTEL_BROADWELL_X,		core_funcs),
2700 	X86_MATCH(INTEL_SKYLAKE,		core_funcs),
2701 	X86_MATCH(INTEL_BROADWELL_D,		core_funcs),
2702 	X86_MATCH(INTEL_XEON_PHI_KNL,		knl_funcs),
2703 	X86_MATCH(INTEL_XEON_PHI_KNM,		knl_funcs),
2704 	X86_MATCH(INTEL_ATOM_GOLDMONT,		core_funcs),
2705 	X86_MATCH(INTEL_ATOM_GOLDMONT_PLUS,	core_funcs),
2706 	X86_MATCH(INTEL_SKYLAKE_X,		core_funcs),
2707 	X86_MATCH(INTEL_COMETLAKE,		core_funcs),
2708 	X86_MATCH(INTEL_ICELAKE_X,		core_funcs),
2709 	X86_MATCH(INTEL_TIGERLAKE,		core_funcs),
2710 	X86_MATCH(INTEL_SAPPHIRERAPIDS_X,	core_funcs),
2711 	X86_MATCH(INTEL_EMERALDRAPIDS_X,	core_funcs),
2712 	X86_MATCH(INTEL_GRANITERAPIDS_D,	core_funcs),
2713 	X86_MATCH(INTEL_GRANITERAPIDS_X,	core_funcs),
2714 	{}
2715 };
2716 MODULE_DEVICE_TABLE(x86cpu, intel_pstate_cpu_ids);
2717 
2718 #ifdef CONFIG_ACPI
2719 static const struct x86_cpu_id intel_pstate_cpu_oob_ids[] __initconst = {
2720 	X86_MATCH(INTEL_BROADWELL_D,		core_funcs),
2721 	X86_MATCH(INTEL_BROADWELL_X,		core_funcs),
2722 	X86_MATCH(INTEL_SKYLAKE_X,		core_funcs),
2723 	X86_MATCH(INTEL_ICELAKE_X,		core_funcs),
2724 	X86_MATCH(INTEL_SAPPHIRERAPIDS_X,	core_funcs),
2725 	X86_MATCH(INTEL_EMERALDRAPIDS_X,	core_funcs),
2726 	X86_MATCH(INTEL_GRANITERAPIDS_D,	core_funcs),
2727 	X86_MATCH(INTEL_GRANITERAPIDS_X,	core_funcs),
2728 	X86_MATCH(INTEL_ATOM_CRESTMONT,		core_funcs),
2729 	X86_MATCH(INTEL_ATOM_CRESTMONT_X,	core_funcs),
2730 	X86_MATCH(INTEL_ATOM_DARKMONT_X,	core_funcs),
2731 	X86_MATCH(INTEL_DIAMONDRAPIDS_X,	core_funcs),
2732 	{}
2733 };
2734 #endif
2735 
2736 static const struct x86_cpu_id intel_pstate_cpu_ee_disable_ids[] = {
2737 	X86_MATCH(INTEL_KABYLAKE,		core_funcs),
2738 	{}
2739 };
2740 
2741 static int intel_pstate_init_cpu(unsigned int cpunum)
2742 {
2743 	struct cpudata *cpu;
2744 
2745 	cpu = all_cpu_data[cpunum];
2746 
2747 	if (!cpu) {
2748 		cpu = kzalloc(sizeof(*cpu), GFP_KERNEL);
2749 		if (!cpu)
2750 			return -ENOMEM;
2751 
2752 		WRITE_ONCE(all_cpu_data[cpunum], cpu);
2753 
2754 		cpu->cpu = cpunum;
2755 
2756 		cpu->epp_default = -EINVAL;
2757 
2758 		if (hwp_active) {
2759 			intel_pstate_hwp_enable(cpu);
2760 
2761 			if (intel_pstate_acpi_pm_profile_server())
2762 				hwp_boost = true;
2763 		}
2764 	} else if (hwp_active) {
2765 		/*
2766 		 * Re-enable HWP in case this happens after a resume from ACPI
2767 		 * S3 if the CPU was offline during the whole system/resume
2768 		 * cycle.
2769 		 */
2770 		intel_pstate_hwp_reenable(cpu);
2771 	}
2772 
2773 	cpu->epp_powersave = -EINVAL;
2774 	cpu->epp_policy = CPUFREQ_POLICY_UNKNOWN;
2775 
2776 	intel_pstate_get_cpu_pstates(cpu);
2777 
2778 	pr_debug("controlling: cpu %d\n", cpunum);
2779 
2780 	return 0;
2781 }
2782 
2783 static void intel_pstate_set_update_util_hook(unsigned int cpu_num)
2784 {
2785 	struct cpudata *cpu = all_cpu_data[cpu_num];
2786 
2787 	if (hwp_active && !hwp_boost)
2788 		return;
2789 
2790 	if (cpu->update_util_set)
2791 		return;
2792 
2793 	/* Prevent intel_pstate_update_util() from using stale data. */
2794 	cpu->sample.time = 0;
2795 	cpufreq_add_update_util_hook(cpu_num, &cpu->update_util,
2796 				     (hwp_active ?
2797 				      intel_pstate_update_util_hwp :
2798 				      intel_pstate_update_util));
2799 	cpu->update_util_set = true;
2800 }
2801 
2802 static void intel_pstate_clear_update_util_hook(unsigned int cpu)
2803 {
2804 	struct cpudata *cpu_data = all_cpu_data[cpu];
2805 
2806 	if (!cpu_data->update_util_set)
2807 		return;
2808 
2809 	cpufreq_remove_update_util_hook(cpu);
2810 	cpu_data->update_util_set = false;
2811 	synchronize_rcu();
2812 }
2813 
2814 static int intel_pstate_get_max_freq(struct cpudata *cpu)
2815 {
2816 	return READ_ONCE(global.no_turbo) ?
2817 			cpu->pstate.max_freq : cpu->pstate.turbo_freq;
2818 }
2819 
2820 static void intel_pstate_update_perf_limits(struct cpudata *cpu,
2821 					    unsigned int policy_min,
2822 					    unsigned int policy_max)
2823 {
2824 	int perf_ctl_scaling = cpu->pstate.perf_ctl_scaling;
2825 	int32_t max_policy_perf, min_policy_perf;
2826 
2827 	max_policy_perf = policy_max / perf_ctl_scaling;
2828 	if (policy_max == policy_min) {
2829 		min_policy_perf = max_policy_perf;
2830 	} else {
2831 		min_policy_perf = policy_min / perf_ctl_scaling;
2832 		min_policy_perf = clamp_t(int32_t, min_policy_perf,
2833 					  0, max_policy_perf);
2834 	}
2835 
2836 	/*
2837 	 * HWP needs some special consideration, because HWP_REQUEST uses
2838 	 * abstract values to represent performance rather than pure ratios.
2839 	 */
2840 	if (hwp_active && cpu->pstate.scaling != perf_ctl_scaling) {
2841 		int freq;
2842 
2843 		freq = max_policy_perf * perf_ctl_scaling;
2844 		max_policy_perf = intel_pstate_freq_to_hwp(cpu, freq);
2845 		freq = min_policy_perf * perf_ctl_scaling;
2846 		min_policy_perf = intel_pstate_freq_to_hwp(cpu, freq);
2847 	}
2848 
2849 	pr_debug("cpu:%d min_policy_perf:%d max_policy_perf:%d\n",
2850 		 cpu->cpu, min_policy_perf, max_policy_perf);
2851 
2852 	/* Normalize user input to [min_perf, max_perf] */
2853 	if (per_cpu_limits) {
2854 		cpu->min_perf_ratio = min_policy_perf;
2855 		cpu->max_perf_ratio = max_policy_perf;
2856 	} else {
2857 		int turbo_max = cpu->pstate.turbo_pstate;
2858 		int32_t global_min, global_max;
2859 
2860 		/* Global limits are in percent of the maximum turbo P-state. */
2861 		global_max = DIV_ROUND_UP(turbo_max * global.max_perf_pct, 100);
2862 		global_min = DIV_ROUND_UP(turbo_max * global.min_perf_pct, 100);
2863 		global_min = clamp_t(int32_t, global_min, 0, global_max);
2864 
2865 		pr_debug("cpu:%d global_min:%d global_max:%d\n", cpu->cpu,
2866 			 global_min, global_max);
2867 
2868 		cpu->min_perf_ratio = max(min_policy_perf, global_min);
2869 		cpu->min_perf_ratio = min(cpu->min_perf_ratio, max_policy_perf);
2870 		cpu->max_perf_ratio = min(max_policy_perf, global_max);
2871 		cpu->max_perf_ratio = max(min_policy_perf, cpu->max_perf_ratio);
2872 
2873 		/* Make sure min_perf <= max_perf */
2874 		cpu->min_perf_ratio = min(cpu->min_perf_ratio,
2875 					  cpu->max_perf_ratio);
2876 
2877 	}
2878 	pr_debug("cpu:%d max_perf_ratio:%d min_perf_ratio:%d\n", cpu->cpu,
2879 		 cpu->max_perf_ratio,
2880 		 cpu->min_perf_ratio);
2881 }
2882 
2883 static int intel_pstate_set_policy(struct cpufreq_policy *policy)
2884 {
2885 	struct cpudata *cpu;
2886 
2887 	if (!policy->cpuinfo.max_freq)
2888 		return -ENODEV;
2889 
2890 	pr_debug("set_policy cpuinfo.max %u policy->max %u\n",
2891 		 policy->cpuinfo.max_freq, policy->max);
2892 
2893 	cpu = all_cpu_data[policy->cpu];
2894 	cpu->policy = policy->policy;
2895 
2896 	mutex_lock(&intel_pstate_limits_lock);
2897 
2898 	intel_pstate_update_perf_limits(cpu, policy->min, policy->max);
2899 
2900 	if (cpu->policy == CPUFREQ_POLICY_PERFORMANCE) {
2901 		int pstate = max(cpu->pstate.min_pstate, cpu->max_perf_ratio);
2902 
2903 		/*
2904 		 * NOHZ_FULL CPUs need this as the governor callback may not
2905 		 * be invoked on them.
2906 		 */
2907 		intel_pstate_clear_update_util_hook(policy->cpu);
2908 		intel_pstate_set_pstate(cpu, pstate);
2909 	} else {
2910 		intel_pstate_set_update_util_hook(policy->cpu);
2911 	}
2912 
2913 	if (hwp_active) {
2914 		/*
2915 		 * When hwp_boost was active before and dynamically it
2916 		 * was turned off, in that case we need to clear the
2917 		 * update util hook.
2918 		 */
2919 		if (!hwp_boost)
2920 			intel_pstate_clear_update_util_hook(policy->cpu);
2921 		intel_pstate_hwp_set(policy->cpu);
2922 	}
2923 	/*
2924 	 * policy->cur is never updated with the intel_pstate driver, but it
2925 	 * is used as a stale frequency value. So, keep it within limits.
2926 	 */
2927 	policy->cur = policy->min;
2928 
2929 	mutex_unlock(&intel_pstate_limits_lock);
2930 
2931 	return 0;
2932 }
2933 
2934 static void intel_pstate_adjust_policy_max(struct cpudata *cpu,
2935 					   struct cpufreq_policy_data *policy)
2936 {
2937 	if (!hwp_active &&
2938 	    cpu->pstate.max_pstate_physical > cpu->pstate.max_pstate &&
2939 	    policy->max < policy->cpuinfo.max_freq &&
2940 	    policy->max > cpu->pstate.max_freq) {
2941 		pr_debug("policy->max > max non turbo frequency\n");
2942 		policy->max = policy->cpuinfo.max_freq;
2943 	}
2944 }
2945 
2946 static void intel_pstate_verify_cpu_policy(struct cpudata *cpu,
2947 					   struct cpufreq_policy_data *policy)
2948 {
2949 	int max_freq;
2950 
2951 	if (hwp_active) {
2952 		intel_pstate_get_hwp_cap(cpu);
2953 		max_freq = READ_ONCE(global.no_turbo) ?
2954 				cpu->pstate.max_freq : cpu->pstate.turbo_freq;
2955 	} else {
2956 		max_freq = intel_pstate_get_max_freq(cpu);
2957 	}
2958 	cpufreq_verify_within_limits(policy, policy->cpuinfo.min_freq, max_freq);
2959 
2960 	intel_pstate_adjust_policy_max(cpu, policy);
2961 }
2962 
2963 static int intel_pstate_verify_policy(struct cpufreq_policy_data *policy)
2964 {
2965 	intel_pstate_verify_cpu_policy(all_cpu_data[policy->cpu], policy);
2966 
2967 	return 0;
2968 }
2969 
2970 static int intel_cpufreq_cpu_offline(struct cpufreq_policy *policy)
2971 {
2972 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2973 
2974 	pr_debug("CPU %d going offline\n", cpu->cpu);
2975 
2976 	if (cpu->suspended)
2977 		return 0;
2978 
2979 	/*
2980 	 * If the CPU is an SMT thread and it goes offline with the performance
2981 	 * settings different from the minimum, it will prevent its sibling
2982 	 * from getting to lower performance levels, so force the minimum
2983 	 * performance on CPU offline to prevent that from happening.
2984 	 */
2985 	if (hwp_active)
2986 		intel_pstate_hwp_offline(cpu);
2987 	else
2988 		intel_pstate_set_min_pstate(cpu);
2989 
2990 	intel_pstate_exit_perf_limits(policy);
2991 
2992 	return 0;
2993 }
2994 
2995 static int intel_pstate_cpu_online(struct cpufreq_policy *policy)
2996 {
2997 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2998 
2999 	pr_debug("CPU %d going online\n", cpu->cpu);
3000 
3001 	intel_pstate_init_acpi_perf_limits(policy);
3002 
3003 	if (hwp_active) {
3004 		/*
3005 		 * Re-enable HWP and clear the "suspended" flag to let "resume"
3006 		 * know that it need not do that.
3007 		 */
3008 		intel_pstate_hwp_reenable(cpu);
3009 		cpu->suspended = false;
3010 
3011 		hybrid_update_capacity(cpu);
3012 	}
3013 
3014 	return 0;
3015 }
3016 
3017 static int intel_pstate_cpu_offline(struct cpufreq_policy *policy)
3018 {
3019 	intel_pstate_clear_update_util_hook(policy->cpu);
3020 
3021 	return intel_cpufreq_cpu_offline(policy);
3022 }
3023 
3024 static void intel_pstate_cpu_exit(struct cpufreq_policy *policy)
3025 {
3026 	pr_debug("CPU %d exiting\n", policy->cpu);
3027 
3028 	policy->fast_switch_possible = false;
3029 }
3030 
3031 static int __intel_pstate_cpu_init(struct cpufreq_policy *policy)
3032 {
3033 	struct cpudata *cpu;
3034 	int rc;
3035 
3036 	rc = intel_pstate_init_cpu(policy->cpu);
3037 	if (rc)
3038 		return rc;
3039 
3040 	cpu = all_cpu_data[policy->cpu];
3041 
3042 	cpu->max_perf_ratio = 0xFF;
3043 	cpu->min_perf_ratio = 0;
3044 
3045 	/* cpuinfo and default policy values */
3046 	policy->cpuinfo.min_freq = cpu->pstate.min_freq;
3047 	policy->cpuinfo.max_freq = READ_ONCE(global.no_turbo) ?
3048 			cpu->pstate.max_freq : cpu->pstate.turbo_freq;
3049 
3050 	policy->min = policy->cpuinfo.min_freq;
3051 	policy->max = policy->cpuinfo.max_freq;
3052 
3053 	intel_pstate_init_acpi_perf_limits(policy);
3054 
3055 	policy->fast_switch_possible = true;
3056 
3057 	return 0;
3058 }
3059 
3060 static int intel_pstate_cpu_init(struct cpufreq_policy *policy)
3061 {
3062 	int ret = __intel_pstate_cpu_init(policy);
3063 
3064 	if (ret)
3065 		return ret;
3066 
3067 	/*
3068 	 * Set the policy to powersave to provide a valid fallback value in case
3069 	 * the default cpufreq governor is neither powersave nor performance.
3070 	 */
3071 	policy->policy = CPUFREQ_POLICY_POWERSAVE;
3072 
3073 	if (hwp_active) {
3074 		struct cpudata *cpu = all_cpu_data[policy->cpu];
3075 
3076 		cpu->epp_cached = intel_pstate_get_epp(cpu, 0);
3077 	}
3078 
3079 	return 0;
3080 }
3081 
3082 static struct cpufreq_driver intel_pstate = {
3083 	.flags		= CPUFREQ_CONST_LOOPS,
3084 	.verify		= intel_pstate_verify_policy,
3085 	.setpolicy	= intel_pstate_set_policy,
3086 	.suspend	= intel_pstate_suspend,
3087 	.resume		= intel_pstate_resume,
3088 	.init		= intel_pstate_cpu_init,
3089 	.exit		= intel_pstate_cpu_exit,
3090 	.offline	= intel_pstate_cpu_offline,
3091 	.online		= intel_pstate_cpu_online,
3092 	.update_limits	= intel_pstate_update_limits,
3093 	.name		= "intel_pstate",
3094 };
3095 
3096 static int intel_cpufreq_verify_policy(struct cpufreq_policy_data *policy)
3097 {
3098 	struct cpudata *cpu = all_cpu_data[policy->cpu];
3099 
3100 	intel_pstate_verify_cpu_policy(cpu, policy);
3101 	intel_pstate_update_perf_limits(cpu, policy->min, policy->max);
3102 
3103 	return 0;
3104 }
3105 
3106 /* Use of trace in passive mode:
3107  *
3108  * In passive mode the trace core_busy field (also known as the
3109  * performance field, and lablelled as such on the graphs; also known as
3110  * core_avg_perf) is not needed and so is re-assigned to indicate if the
3111  * driver call was via the normal or fast switch path. Various graphs
3112  * output from the intel_pstate_tracer.py utility that include core_busy
3113  * (or performance or core_avg_perf) have a fixed y-axis from 0 to 100%,
3114  * so we use 10 to indicate the normal path through the driver, and
3115  * 90 to indicate the fast switch path through the driver.
3116  * The scaled_busy field is not used, and is set to 0.
3117  */
3118 
3119 #define	INTEL_PSTATE_TRACE_TARGET 10
3120 #define	INTEL_PSTATE_TRACE_FAST_SWITCH 90
3121 
3122 static void intel_cpufreq_trace(struct cpudata *cpu, unsigned int trace_type, int old_pstate)
3123 {
3124 	struct sample *sample;
3125 
3126 	if (!trace_pstate_sample_enabled())
3127 		return;
3128 
3129 	if (!intel_pstate_sample(cpu, ktime_get()))
3130 		return;
3131 
3132 	sample = &cpu->sample;
3133 	trace_pstate_sample(trace_type,
3134 		0,
3135 		old_pstate,
3136 		cpu->pstate.current_pstate,
3137 		sample->mperf,
3138 		sample->aperf,
3139 		sample->tsc,
3140 		get_avg_frequency(cpu),
3141 		fp_toint(cpu->iowait_boost * 100));
3142 }
3143 
3144 static void intel_cpufreq_hwp_update(struct cpudata *cpu, u32 min, u32 max,
3145 				     u32 desired, bool fast_switch)
3146 {
3147 	u64 prev = READ_ONCE(cpu->hwp_req_cached), value = prev;
3148 
3149 	value &= ~HWP_MIN_PERF(~0L);
3150 	value |= HWP_MIN_PERF(min);
3151 
3152 	value &= ~HWP_MAX_PERF(~0L);
3153 	value |= HWP_MAX_PERF(max);
3154 
3155 	value &= ~HWP_DESIRED_PERF(~0L);
3156 	value |= HWP_DESIRED_PERF(desired);
3157 
3158 	if (value == prev)
3159 		return;
3160 
3161 	WRITE_ONCE(cpu->hwp_req_cached, value);
3162 	if (fast_switch)
3163 		wrmsrq(MSR_HWP_REQUEST, value);
3164 	else
3165 		wrmsrq_on_cpu(cpu->cpu, MSR_HWP_REQUEST, value);
3166 }
3167 
3168 static void intel_cpufreq_perf_ctl_update(struct cpudata *cpu,
3169 					  u32 target_pstate, bool fast_switch)
3170 {
3171 	if (fast_switch)
3172 		wrmsrq(MSR_IA32_PERF_CTL,
3173 		       pstate_funcs.get_val(cpu, target_pstate));
3174 	else
3175 		wrmsrq_on_cpu(cpu->cpu, MSR_IA32_PERF_CTL,
3176 			      pstate_funcs.get_val(cpu, target_pstate));
3177 }
3178 
3179 static int intel_cpufreq_update_pstate(struct cpufreq_policy *policy,
3180 				       int target_pstate, bool fast_switch)
3181 {
3182 	struct cpudata *cpu = all_cpu_data[policy->cpu];
3183 	int old_pstate = cpu->pstate.current_pstate;
3184 
3185 	target_pstate = intel_pstate_prepare_request(cpu, target_pstate);
3186 	if (hwp_active) {
3187 		int max_pstate = policy->strict_target ?
3188 					target_pstate : cpu->max_perf_ratio;
3189 
3190 		intel_cpufreq_hwp_update(cpu, target_pstate, max_pstate,
3191 					 target_pstate, fast_switch);
3192 	} else if (target_pstate != old_pstate) {
3193 		intel_cpufreq_perf_ctl_update(cpu, target_pstate, fast_switch);
3194 	}
3195 
3196 	cpu->pstate.current_pstate = target_pstate;
3197 
3198 	intel_cpufreq_trace(cpu, fast_switch ? INTEL_PSTATE_TRACE_FAST_SWITCH :
3199 			    INTEL_PSTATE_TRACE_TARGET, old_pstate);
3200 
3201 	return target_pstate;
3202 }
3203 
3204 static int intel_cpufreq_target(struct cpufreq_policy *policy,
3205 				unsigned int target_freq,
3206 				unsigned int relation)
3207 {
3208 	struct cpudata *cpu = all_cpu_data[policy->cpu];
3209 	struct cpufreq_freqs freqs;
3210 	int target_pstate;
3211 
3212 	freqs.old = policy->cur;
3213 	freqs.new = target_freq;
3214 
3215 	cpufreq_freq_transition_begin(policy, &freqs);
3216 
3217 	target_pstate = intel_pstate_freq_to_hwp_rel(cpu, freqs.new, relation);
3218 	target_pstate = intel_cpufreq_update_pstate(policy, target_pstate, false);
3219 
3220 	freqs.new = target_pstate * cpu->pstate.scaling;
3221 
3222 	cpufreq_freq_transition_end(policy, &freqs, false);
3223 
3224 	return 0;
3225 }
3226 
3227 static unsigned int intel_cpufreq_fast_switch(struct cpufreq_policy *policy,
3228 					      unsigned int target_freq)
3229 {
3230 	struct cpudata *cpu = all_cpu_data[policy->cpu];
3231 	int target_pstate;
3232 
3233 	target_pstate = intel_pstate_freq_to_hwp(cpu, target_freq);
3234 
3235 	target_pstate = intel_cpufreq_update_pstate(policy, target_pstate, true);
3236 
3237 	return target_pstate * cpu->pstate.scaling;
3238 }
3239 
3240 static void intel_cpufreq_adjust_perf(unsigned int cpunum,
3241 				      unsigned long min_perf,
3242 				      unsigned long target_perf,
3243 				      unsigned long capacity)
3244 {
3245 	struct cpudata *cpu = all_cpu_data[cpunum];
3246 	u64 hwp_cap = READ_ONCE(cpu->hwp_cap_cached);
3247 	int old_pstate = cpu->pstate.current_pstate;
3248 	int cap_pstate, min_pstate, max_pstate, target_pstate;
3249 
3250 	cap_pstate = READ_ONCE(global.no_turbo) ?
3251 					HWP_GUARANTEED_PERF(hwp_cap) :
3252 					HWP_HIGHEST_PERF(hwp_cap);
3253 
3254 	/* Optimization: Avoid unnecessary divisions. */
3255 
3256 	target_pstate = cap_pstate;
3257 	if (target_perf < capacity)
3258 		target_pstate = DIV_ROUND_UP(cap_pstate * target_perf, capacity);
3259 
3260 	min_pstate = cap_pstate;
3261 	if (min_perf < capacity)
3262 		min_pstate = DIV_ROUND_UP(cap_pstate * min_perf, capacity);
3263 
3264 	if (min_pstate < cpu->pstate.min_pstate)
3265 		min_pstate = cpu->pstate.min_pstate;
3266 
3267 	if (min_pstate < cpu->min_perf_ratio)
3268 		min_pstate = cpu->min_perf_ratio;
3269 
3270 	if (min_pstate > cpu->max_perf_ratio)
3271 		min_pstate = cpu->max_perf_ratio;
3272 
3273 	max_pstate = min(cap_pstate, cpu->max_perf_ratio);
3274 	if (max_pstate < min_pstate)
3275 		max_pstate = min_pstate;
3276 
3277 	target_pstate = clamp_t(int, target_pstate, min_pstate, max_pstate);
3278 
3279 	intel_cpufreq_hwp_update(cpu, min_pstate, max_pstate, target_pstate, true);
3280 
3281 	cpu->pstate.current_pstate = target_pstate;
3282 	intel_cpufreq_trace(cpu, INTEL_PSTATE_TRACE_FAST_SWITCH, old_pstate);
3283 }
3284 
3285 static int intel_cpufreq_cpu_init(struct cpufreq_policy *policy)
3286 {
3287 	struct freq_qos_request *req;
3288 	struct cpudata *cpu;
3289 	struct device *dev;
3290 	int ret, freq;
3291 
3292 	dev = get_cpu_device(policy->cpu);
3293 	if (!dev)
3294 		return -ENODEV;
3295 
3296 	ret = __intel_pstate_cpu_init(policy);
3297 	if (ret)
3298 		return ret;
3299 
3300 	policy->cpuinfo.transition_latency = INTEL_CPUFREQ_TRANSITION_LATENCY;
3301 	/* This reflects the intel_pstate_get_cpu_pstates() setting. */
3302 	policy->cur = policy->cpuinfo.min_freq;
3303 
3304 	req = kcalloc(2, sizeof(*req), GFP_KERNEL);
3305 	if (!req) {
3306 		ret = -ENOMEM;
3307 		goto pstate_exit;
3308 	}
3309 
3310 	cpu = all_cpu_data[policy->cpu];
3311 
3312 	if (hwp_active) {
3313 		u64 value;
3314 
3315 		policy->transition_delay_us = INTEL_CPUFREQ_TRANSITION_DELAY_HWP;
3316 
3317 		intel_pstate_get_hwp_cap(cpu);
3318 
3319 		rdmsrq_on_cpu(cpu->cpu, MSR_HWP_REQUEST, &value);
3320 		WRITE_ONCE(cpu->hwp_req_cached, value);
3321 
3322 		cpu->epp_cached = intel_pstate_get_epp(cpu, value);
3323 	} else {
3324 		policy->transition_delay_us = INTEL_CPUFREQ_TRANSITION_DELAY;
3325 	}
3326 
3327 	freq = DIV_ROUND_UP(cpu->pstate.turbo_freq * global.min_perf_pct, 100);
3328 
3329 	ret = freq_qos_add_request(&policy->constraints, req, FREQ_QOS_MIN,
3330 				   freq);
3331 	if (ret < 0) {
3332 		dev_err(dev, "Failed to add min-freq constraint (%d)\n", ret);
3333 		goto free_req;
3334 	}
3335 
3336 	freq = DIV_ROUND_UP(cpu->pstate.turbo_freq * global.max_perf_pct, 100);
3337 
3338 	ret = freq_qos_add_request(&policy->constraints, req + 1, FREQ_QOS_MAX,
3339 				   freq);
3340 	if (ret < 0) {
3341 		dev_err(dev, "Failed to add max-freq constraint (%d)\n", ret);
3342 		goto remove_min_req;
3343 	}
3344 
3345 	policy->driver_data = req;
3346 
3347 	return 0;
3348 
3349 remove_min_req:
3350 	freq_qos_remove_request(req);
3351 free_req:
3352 	kfree(req);
3353 pstate_exit:
3354 	intel_pstate_exit_perf_limits(policy);
3355 
3356 	return ret;
3357 }
3358 
3359 static void intel_cpufreq_cpu_exit(struct cpufreq_policy *policy)
3360 {
3361 	struct freq_qos_request *req;
3362 
3363 	req = policy->driver_data;
3364 
3365 	freq_qos_remove_request(req + 1);
3366 	freq_qos_remove_request(req);
3367 	kfree(req);
3368 
3369 	intel_pstate_cpu_exit(policy);
3370 }
3371 
3372 static int intel_cpufreq_suspend(struct cpufreq_policy *policy)
3373 {
3374 	intel_pstate_suspend(policy);
3375 
3376 	if (hwp_active) {
3377 		struct cpudata *cpu = all_cpu_data[policy->cpu];
3378 		u64 value = READ_ONCE(cpu->hwp_req_cached);
3379 
3380 		/*
3381 		 * Clear the desired perf field in MSR_HWP_REQUEST in case
3382 		 * intel_cpufreq_adjust_perf() is in use and the last value
3383 		 * written by it may not be suitable.
3384 		 */
3385 		value &= ~HWP_DESIRED_PERF(~0L);
3386 		wrmsrq_on_cpu(cpu->cpu, MSR_HWP_REQUEST, value);
3387 		WRITE_ONCE(cpu->hwp_req_cached, value);
3388 	}
3389 
3390 	return 0;
3391 }
3392 
3393 static struct cpufreq_driver intel_cpufreq = {
3394 	.flags		= CPUFREQ_CONST_LOOPS,
3395 	.verify		= intel_cpufreq_verify_policy,
3396 	.target		= intel_cpufreq_target,
3397 	.fast_switch	= intel_cpufreq_fast_switch,
3398 	.init		= intel_cpufreq_cpu_init,
3399 	.exit		= intel_cpufreq_cpu_exit,
3400 	.offline	= intel_cpufreq_cpu_offline,
3401 	.online		= intel_pstate_cpu_online,
3402 	.suspend	= intel_cpufreq_suspend,
3403 	.resume		= intel_pstate_resume,
3404 	.update_limits	= intel_pstate_update_limits,
3405 	.name		= "intel_cpufreq",
3406 };
3407 
3408 static struct cpufreq_driver *default_driver;
3409 
3410 static void intel_pstate_driver_cleanup(void)
3411 {
3412 	unsigned int cpu;
3413 
3414 	cpus_read_lock();
3415 	for_each_online_cpu(cpu) {
3416 		if (all_cpu_data[cpu]) {
3417 			if (intel_pstate_driver == &intel_pstate)
3418 				intel_pstate_clear_update_util_hook(cpu);
3419 
3420 			kfree(all_cpu_data[cpu]);
3421 			WRITE_ONCE(all_cpu_data[cpu], NULL);
3422 		}
3423 	}
3424 	cpus_read_unlock();
3425 
3426 	intel_pstate_driver = NULL;
3427 }
3428 
3429 static int intel_pstate_register_driver(struct cpufreq_driver *driver)
3430 {
3431 	bool refresh_cpu_cap_scaling;
3432 	int ret;
3433 
3434 	if (driver == &intel_pstate)
3435 		intel_pstate_sysfs_expose_hwp_dynamic_boost();
3436 
3437 	memset(&global, 0, sizeof(global));
3438 	global.max_perf_pct = 100;
3439 	global.turbo_disabled = turbo_is_disabled();
3440 	global.no_turbo = global.turbo_disabled;
3441 
3442 	arch_set_max_freq_ratio(global.turbo_disabled);
3443 
3444 	refresh_cpu_cap_scaling = hybrid_clear_max_perf_cpu();
3445 
3446 	intel_pstate_driver = driver;
3447 	ret = cpufreq_register_driver(intel_pstate_driver);
3448 	if (ret) {
3449 		intel_pstate_driver_cleanup();
3450 		return ret;
3451 	}
3452 
3453 	global.min_perf_pct = min_perf_pct_min();
3454 
3455 	hybrid_init_cpu_capacity_scaling(refresh_cpu_cap_scaling);
3456 
3457 	return 0;
3458 }
3459 
3460 static ssize_t intel_pstate_show_status(char *buf)
3461 {
3462 	if (!intel_pstate_driver)
3463 		return sprintf(buf, "off\n");
3464 
3465 	return sprintf(buf, "%s\n", intel_pstate_driver == &intel_pstate ?
3466 					"active" : "passive");
3467 }
3468 
3469 static int intel_pstate_update_status(const char *buf, size_t size)
3470 {
3471 	if (size == 3 && !strncmp(buf, "off", size)) {
3472 		if (!intel_pstate_driver)
3473 			return -EINVAL;
3474 
3475 		if (hwp_active)
3476 			return -EBUSY;
3477 
3478 		cpufreq_unregister_driver(intel_pstate_driver);
3479 		intel_pstate_driver_cleanup();
3480 		return 0;
3481 	}
3482 
3483 	if (size == 6 && !strncmp(buf, "active", size)) {
3484 		if (intel_pstate_driver) {
3485 			if (intel_pstate_driver == &intel_pstate)
3486 				return 0;
3487 
3488 			cpufreq_unregister_driver(intel_pstate_driver);
3489 		}
3490 
3491 		return intel_pstate_register_driver(&intel_pstate);
3492 	}
3493 
3494 	if (size == 7 && !strncmp(buf, "passive", size)) {
3495 		if (intel_pstate_driver) {
3496 			if (intel_pstate_driver == &intel_cpufreq)
3497 				return 0;
3498 
3499 			cpufreq_unregister_driver(intel_pstate_driver);
3500 			intel_pstate_sysfs_hide_hwp_dynamic_boost();
3501 		}
3502 
3503 		return intel_pstate_register_driver(&intel_cpufreq);
3504 	}
3505 
3506 	return -EINVAL;
3507 }
3508 
3509 static int no_load __initdata;
3510 static int no_hwp __initdata;
3511 static int hwp_only __initdata;
3512 static unsigned int force_load __initdata;
3513 
3514 static int __init intel_pstate_msrs_not_valid(void)
3515 {
3516 	if (!pstate_funcs.get_max(0) ||
3517 	    !pstate_funcs.get_min(0) ||
3518 	    !pstate_funcs.get_turbo(0))
3519 		return -ENODEV;
3520 
3521 	return 0;
3522 }
3523 
3524 static void __init copy_cpu_funcs(struct pstate_funcs *funcs)
3525 {
3526 	pstate_funcs.get_max   = funcs->get_max;
3527 	pstate_funcs.get_max_physical = funcs->get_max_physical;
3528 	pstate_funcs.get_min   = funcs->get_min;
3529 	pstate_funcs.get_turbo = funcs->get_turbo;
3530 	pstate_funcs.get_scaling = funcs->get_scaling;
3531 	pstate_funcs.get_val   = funcs->get_val;
3532 	pstate_funcs.get_vid   = funcs->get_vid;
3533 	pstate_funcs.get_aperf_mperf_shift = funcs->get_aperf_mperf_shift;
3534 }
3535 
3536 #ifdef CONFIG_ACPI
3537 
3538 static bool __init intel_pstate_no_acpi_pss(void)
3539 {
3540 	int i;
3541 
3542 	for_each_possible_cpu(i) {
3543 		acpi_status status;
3544 		union acpi_object *pss;
3545 		struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
3546 		struct acpi_processor *pr = per_cpu(processors, i);
3547 
3548 		if (!pr)
3549 			continue;
3550 
3551 		status = acpi_evaluate_object(pr->handle, "_PSS", NULL, &buffer);
3552 		if (ACPI_FAILURE(status))
3553 			continue;
3554 
3555 		pss = buffer.pointer;
3556 		if (pss && pss->type == ACPI_TYPE_PACKAGE) {
3557 			kfree(pss);
3558 			return false;
3559 		}
3560 
3561 		kfree(pss);
3562 	}
3563 
3564 	pr_debug("ACPI _PSS not found\n");
3565 	return true;
3566 }
3567 
3568 static bool __init intel_pstate_no_acpi_pcch(void)
3569 {
3570 	acpi_status status;
3571 	acpi_handle handle;
3572 
3573 	status = acpi_get_handle(NULL, "\\_SB", &handle);
3574 	if (ACPI_FAILURE(status))
3575 		goto not_found;
3576 
3577 	if (acpi_has_method(handle, "PCCH"))
3578 		return false;
3579 
3580 not_found:
3581 	pr_debug("ACPI PCCH not found\n");
3582 	return true;
3583 }
3584 
3585 static bool __init intel_pstate_has_acpi_ppc(void)
3586 {
3587 	int i;
3588 
3589 	for_each_possible_cpu(i) {
3590 		struct acpi_processor *pr = per_cpu(processors, i);
3591 
3592 		if (!pr)
3593 			continue;
3594 		if (acpi_has_method(pr->handle, "_PPC"))
3595 			return true;
3596 	}
3597 	pr_debug("ACPI _PPC not found\n");
3598 	return false;
3599 }
3600 
3601 enum {
3602 	PSS,
3603 	PPC,
3604 };
3605 
3606 /* Hardware vendor-specific info that has its own power management modes */
3607 static struct acpi_platform_list plat_info[] __initdata = {
3608 	{"HP    ", "ProLiant", 0, ACPI_SIG_FADT, all_versions, NULL, PSS},
3609 	{"ORACLE", "X4-2    ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3610 	{"ORACLE", "X4-2L   ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3611 	{"ORACLE", "X4-2B   ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3612 	{"ORACLE", "X3-2    ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3613 	{"ORACLE", "X3-2L   ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3614 	{"ORACLE", "X3-2B   ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3615 	{"ORACLE", "X4470M2 ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3616 	{"ORACLE", "X4270M3 ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3617 	{"ORACLE", "X4270M2 ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3618 	{"ORACLE", "X4170M2 ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3619 	{"ORACLE", "X4170 M3", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3620 	{"ORACLE", "X4275 M3", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3621 	{"ORACLE", "X6-2    ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3622 	{"ORACLE", "Sudbury ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3623 	{ } /* End */
3624 };
3625 
3626 #define BITMASK_OOB	(BIT(8) | BIT(18))
3627 
3628 static bool __init intel_pstate_platform_pwr_mgmt_exists(void)
3629 {
3630 	const struct x86_cpu_id *id;
3631 	u64 misc_pwr;
3632 	int idx;
3633 
3634 	id = x86_match_cpu(intel_pstate_cpu_oob_ids);
3635 	if (id) {
3636 		rdmsrq(MSR_MISC_PWR_MGMT, misc_pwr);
3637 		if (misc_pwr & BITMASK_OOB) {
3638 			pr_debug("Bit 8 or 18 in the MISC_PWR_MGMT MSR set\n");
3639 			pr_debug("P states are controlled in Out of Band mode by the firmware/hardware\n");
3640 			return true;
3641 		}
3642 	}
3643 
3644 	idx = acpi_match_platform_list(plat_info);
3645 	if (idx < 0)
3646 		return false;
3647 
3648 	switch (plat_info[idx].data) {
3649 	case PSS:
3650 		if (!intel_pstate_no_acpi_pss())
3651 			return false;
3652 
3653 		return intel_pstate_no_acpi_pcch();
3654 	case PPC:
3655 		return intel_pstate_has_acpi_ppc() && !force_load;
3656 	}
3657 
3658 	return false;
3659 }
3660 
3661 static void intel_pstate_request_control_from_smm(void)
3662 {
3663 	/*
3664 	 * It may be unsafe to request P-states control from SMM if _PPC support
3665 	 * has not been enabled.
3666 	 */
3667 	if (acpi_ppc)
3668 		acpi_processor_pstate_control();
3669 }
3670 #else /* CONFIG_ACPI not enabled */
3671 static inline bool intel_pstate_platform_pwr_mgmt_exists(void) { return false; }
3672 static inline bool intel_pstate_has_acpi_ppc(void) { return false; }
3673 static inline void intel_pstate_request_control_from_smm(void) {}
3674 #endif /* CONFIG_ACPI */
3675 
3676 #define INTEL_PSTATE_HWP_BROADWELL	0x01
3677 
3678 #define X86_MATCH_HWP(vfm, hwp_mode)				\
3679 	X86_MATCH_VFM_FEATURE(vfm, X86_FEATURE_HWP, hwp_mode)
3680 
3681 static const struct x86_cpu_id hwp_support_ids[] __initconst = {
3682 	X86_MATCH_HWP(INTEL_BROADWELL_X,	INTEL_PSTATE_HWP_BROADWELL),
3683 	X86_MATCH_HWP(INTEL_BROADWELL_D,	INTEL_PSTATE_HWP_BROADWELL),
3684 	X86_MATCH_HWP(INTEL_ANY,		0),
3685 	{}
3686 };
3687 
3688 static bool intel_pstate_hwp_is_enabled(void)
3689 {
3690 	u64 value;
3691 
3692 	rdmsrq(MSR_PM_ENABLE, value);
3693 	return !!(value & 0x1);
3694 }
3695 
3696 #define POWERSAVE_MASK			GENMASK(7, 0)
3697 #define BALANCE_POWER_MASK		GENMASK(15, 8)
3698 #define BALANCE_PERFORMANCE_MASK	GENMASK(23, 16)
3699 #define PERFORMANCE_MASK		GENMASK(31, 24)
3700 
3701 #define HWP_SET_EPP_VALUES(powersave, balance_power, balance_perf, performance) \
3702 	(FIELD_PREP_CONST(POWERSAVE_MASK, powersave) |\
3703 	 FIELD_PREP_CONST(BALANCE_POWER_MASK, balance_power) |\
3704 	 FIELD_PREP_CONST(BALANCE_PERFORMANCE_MASK, balance_perf) |\
3705 	 FIELD_PREP_CONST(PERFORMANCE_MASK, performance))
3706 
3707 #define HWP_SET_DEF_BALANCE_PERF_EPP(balance_perf) \
3708 	(HWP_SET_EPP_VALUES(HWP_EPP_POWERSAVE, HWP_EPP_BALANCE_POWERSAVE,\
3709 	 balance_perf, HWP_EPP_PERFORMANCE))
3710 
3711 static const struct x86_cpu_id intel_epp_default[] = {
3712 	/*
3713 	 * Set EPP value as 102, this is the max suggested EPP
3714 	 * which can result in one core turbo frequency for
3715 	 * AlderLake Mobile CPUs.
3716 	 */
3717 	X86_MATCH_VFM(INTEL_ALDERLAKE_L, HWP_SET_DEF_BALANCE_PERF_EPP(102)),
3718 	X86_MATCH_VFM(INTEL_SAPPHIRERAPIDS_X, HWP_SET_DEF_BALANCE_PERF_EPP(32)),
3719 	X86_MATCH_VFM(INTEL_EMERALDRAPIDS_X, HWP_SET_DEF_BALANCE_PERF_EPP(32)),
3720 	X86_MATCH_VFM(INTEL_GRANITERAPIDS_X, HWP_SET_DEF_BALANCE_PERF_EPP(32)),
3721 	X86_MATCH_VFM(INTEL_GRANITERAPIDS_D, HWP_SET_DEF_BALANCE_PERF_EPP(32)),
3722 	X86_MATCH_VFM(INTEL_METEORLAKE_L, HWP_SET_EPP_VALUES(HWP_EPP_POWERSAVE,
3723 		      179, 64, 16)),
3724 	X86_MATCH_VFM(INTEL_ARROWLAKE, HWP_SET_EPP_VALUES(HWP_EPP_POWERSAVE,
3725 		      179, 64, 16)),
3726 	{}
3727 };
3728 
3729 static const struct x86_cpu_id intel_hybrid_scaling_factor[] = {
3730 	X86_MATCH_VFM(INTEL_ALDERLAKE, HYBRID_SCALING_FACTOR_ADL),
3731 	X86_MATCH_VFM(INTEL_ALDERLAKE_L, HYBRID_SCALING_FACTOR_ADL),
3732 	X86_MATCH_VFM(INTEL_RAPTORLAKE, HYBRID_SCALING_FACTOR_ADL),
3733 	X86_MATCH_VFM(INTEL_RAPTORLAKE_P, HYBRID_SCALING_FACTOR_ADL),
3734 	X86_MATCH_VFM(INTEL_RAPTORLAKE_S, HYBRID_SCALING_FACTOR_ADL),
3735 	X86_MATCH_VFM(INTEL_METEORLAKE_L, HYBRID_SCALING_FACTOR_MTL),
3736 	X86_MATCH_VFM(INTEL_LUNARLAKE_M, HYBRID_SCALING_FACTOR_LNL),
3737 	{}
3738 };
3739 
3740 static bool hwp_check_epp(void)
3741 {
3742 	if (boot_cpu_has(X86_FEATURE_HWP_EPP))
3743 		return true;
3744 
3745 	/* Without EPP support, don't expose EPP-related sysfs attributes. */
3746 	hwp_cpufreq_attrs[HWP_PERFORMANCE_PREFERENCE_INDEX] = NULL;
3747 	hwp_cpufreq_attrs[HWP_PERFORMANCE_AVAILABLE_PREFERENCES_INDEX] = NULL;
3748 
3749 	return false;
3750 }
3751 
3752 static bool hwp_check_dec(void)
3753 {
3754 	u64 power_ctl;
3755 
3756 	rdmsrq(MSR_IA32_POWER_CTL, power_ctl);
3757 	return !!(power_ctl & BIT(POWER_CTL_DEC_ENABLE));
3758 }
3759 
3760 static int __init intel_pstate_init(void)
3761 {
3762 	static struct cpudata **_all_cpu_data;
3763 	const struct x86_cpu_id *id;
3764 	int rc;
3765 
3766 	if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL)
3767 		return -ENODEV;
3768 
3769 	/*
3770 	 * The Intel pstate driver will be ignored if the platform
3771 	 * firmware has its own power management modes.
3772 	 */
3773 	if (intel_pstate_platform_pwr_mgmt_exists()) {
3774 		pr_info("P-states controlled by the platform\n");
3775 		return -ENODEV;
3776 	}
3777 
3778 	id = x86_match_cpu(hwp_support_ids);
3779 	if (id) {
3780 		bool epp_present = hwp_check_epp();
3781 
3782 		/*
3783 		 * If HWP is enabled already, there is no choice but to deal
3784 		 * with it.
3785 		 */
3786 		hwp_forced = intel_pstate_hwp_is_enabled();
3787 		if (hwp_forced) {
3788 			pr_info("HWP enabled by BIOS\n");
3789 			no_hwp = 0;
3790 		} else if (no_load) {
3791 			return -ENODEV;
3792 		} else if (!epp_present && !hwp_check_dec()) {
3793 			/*
3794 			 * Avoid enabling HWP for processors without EPP support
3795 			 * unless the Dynamic Efficiency Control (DEC) enable
3796 			 * bit (MSR_IA32_POWER_CTL, bit 27) is set because that
3797 			 * means incomplete HWP implementation which is a corner
3798 			 * case and supporting it is generally problematic.
3799 			 */
3800 			no_hwp = 1;
3801 		}
3802 
3803 		copy_cpu_funcs(&core_funcs);
3804 
3805 		if (!no_hwp) {
3806 			hwp_active = true;
3807 			hwp_mode_bdw = id->driver_data;
3808 			intel_pstate.attr = hwp_cpufreq_attrs;
3809 			intel_cpufreq.attr = hwp_cpufreq_attrs;
3810 			intel_cpufreq.flags |= CPUFREQ_NEED_UPDATE_LIMITS;
3811 			intel_cpufreq.adjust_perf = intel_cpufreq_adjust_perf;
3812 			if (!default_driver)
3813 				default_driver = &intel_pstate;
3814 
3815 			pstate_funcs.get_cpu_scaling = hwp_get_cpu_scaling;
3816 
3817 			goto hwp_cpu_matched;
3818 		}
3819 		pr_info("HWP not enabled\n");
3820 	} else {
3821 		if (no_load)
3822 			return -ENODEV;
3823 
3824 		id = x86_match_cpu(intel_pstate_cpu_ids);
3825 		if (!id) {
3826 			pr_info("CPU model not supported\n");
3827 			return -ENODEV;
3828 		}
3829 
3830 		copy_cpu_funcs((struct pstate_funcs *)id->driver_data);
3831 	}
3832 
3833 	if (intel_pstate_msrs_not_valid()) {
3834 		pr_info("Invalid MSRs\n");
3835 		return -ENODEV;
3836 	}
3837 	/* Without HWP start in the passive mode. */
3838 	if (!default_driver)
3839 		default_driver = &intel_cpufreq;
3840 
3841 hwp_cpu_matched:
3842 	if (!hwp_active && hwp_only)
3843 		return -ENOTSUPP;
3844 
3845 	pr_info("Intel P-state driver initializing\n");
3846 
3847 	_all_cpu_data = vzalloc(array_size(sizeof(void *), num_possible_cpus()));
3848 	if (!_all_cpu_data)
3849 		return -ENOMEM;
3850 
3851 	WRITE_ONCE(all_cpu_data, _all_cpu_data);
3852 
3853 	intel_pstate_request_control_from_smm();
3854 
3855 	intel_pstate_sysfs_expose_params();
3856 
3857 	if (hwp_active) {
3858 		const struct x86_cpu_id *id = x86_match_cpu(intel_epp_default);
3859 		const struct x86_cpu_id *hybrid_id = x86_match_cpu(intel_hybrid_scaling_factor);
3860 
3861 		if (id) {
3862 			epp_values[EPP_INDEX_POWERSAVE] =
3863 					FIELD_GET(POWERSAVE_MASK, id->driver_data);
3864 			epp_values[EPP_INDEX_BALANCE_POWERSAVE] =
3865 					FIELD_GET(BALANCE_POWER_MASK, id->driver_data);
3866 			epp_values[EPP_INDEX_BALANCE_PERFORMANCE] =
3867 					FIELD_GET(BALANCE_PERFORMANCE_MASK, id->driver_data);
3868 			epp_values[EPP_INDEX_PERFORMANCE] =
3869 					FIELD_GET(PERFORMANCE_MASK, id->driver_data);
3870 			pr_debug("Updated EPPs powersave:%x balanced power:%x balanced perf:%x performance:%x\n",
3871 				 epp_values[EPP_INDEX_POWERSAVE],
3872 				 epp_values[EPP_INDEX_BALANCE_POWERSAVE],
3873 				 epp_values[EPP_INDEX_BALANCE_PERFORMANCE],
3874 				 epp_values[EPP_INDEX_PERFORMANCE]);
3875 		}
3876 
3877 		if (hybrid_id) {
3878 			hybrid_scaling_factor = hybrid_id->driver_data;
3879 			pr_debug("hybrid scaling factor: %d\n", hybrid_scaling_factor);
3880 		}
3881 
3882 	}
3883 
3884 	scoped_guard(mutex, &intel_pstate_driver_lock) {
3885 		rc = intel_pstate_register_driver(default_driver);
3886 	}
3887 	if (rc) {
3888 		intel_pstate_sysfs_remove();
3889 		return rc;
3890 	}
3891 
3892 	if (hwp_active) {
3893 		const struct x86_cpu_id *id;
3894 
3895 		id = x86_match_cpu(intel_pstate_cpu_ee_disable_ids);
3896 		if (id) {
3897 			set_power_ctl_ee_state(false);
3898 			pr_info("Disabling energy efficiency optimization\n");
3899 		}
3900 
3901 		pr_info("HWP enabled\n");
3902 	} else if (boot_cpu_has(X86_FEATURE_HYBRID_CPU)) {
3903 		pr_warn("Problematic setup: Hybrid processor with disabled HWP\n");
3904 	}
3905 
3906 	return 0;
3907 }
3908 device_initcall(intel_pstate_init);
3909 
3910 static int __init intel_pstate_setup(char *str)
3911 {
3912 	if (!str)
3913 		return -EINVAL;
3914 
3915 	if (!strcmp(str, "disable"))
3916 		no_load = 1;
3917 	else if (!strcmp(str, "active"))
3918 		default_driver = &intel_pstate;
3919 	else if (!strcmp(str, "passive"))
3920 		default_driver = &intel_cpufreq;
3921 
3922 	if (!strcmp(str, "no_hwp"))
3923 		no_hwp = 1;
3924 
3925 	if (!strcmp(str, "no_cas"))
3926 		no_cas = true;
3927 
3928 	if (!strcmp(str, "force"))
3929 		force_load = 1;
3930 	if (!strcmp(str, "hwp_only"))
3931 		hwp_only = 1;
3932 	if (!strcmp(str, "per_cpu_perf_limits"))
3933 		per_cpu_limits = true;
3934 
3935 #ifdef CONFIG_ACPI
3936 	if (!strcmp(str, "support_acpi_ppc"))
3937 		acpi_ppc = true;
3938 #endif
3939 
3940 	return 0;
3941 }
3942 early_param("intel_pstate", intel_pstate_setup);
3943 
3944 MODULE_AUTHOR("Dirk Brandewie <dirk.j.brandewie@intel.com>");
3945 MODULE_DESCRIPTION("'intel_pstate' - P state driver Intel Core processors");
3946