xref: /linux/drivers/cpufreq/intel_pstate.c (revision f694f30e81c4ade358eb8c75273bac1a48f0cb8f)
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  * @hwp_notify_work:	workqueue for HWP notifications.
225  *
226  * This structure stores per CPU instance data for all CPUs.
227  */
228 struct cpudata {
229 	int cpu;
230 
231 	unsigned int policy;
232 	struct update_util_data update_util;
233 	bool   update_util_set;
234 
235 	struct pstate_data pstate;
236 	struct vid_data vid;
237 
238 	u64	last_update;
239 	u64	last_sample_time;
240 	u64	aperf_mperf_shift;
241 	u64	prev_aperf;
242 	u64	prev_mperf;
243 	u64	prev_tsc;
244 	struct sample sample;
245 	int32_t	min_perf_ratio;
246 	int32_t	max_perf_ratio;
247 #ifdef CONFIG_ACPI
248 	struct acpi_processor_performance acpi_perf_data;
249 	bool valid_pss_table;
250 #endif
251 	unsigned int iowait_boost;
252 	s16 epp_powersave;
253 	s16 epp_policy;
254 	s16 epp_default;
255 	s16 epp_cached;
256 	u64 hwp_req_cached;
257 	u64 hwp_cap_cached;
258 	u64 last_io_update;
259 	unsigned int capacity_perf;
260 	unsigned int sched_flags;
261 	u32 hwp_boost_min;
262 	bool suspended;
263 	struct delayed_work hwp_notify_work;
264 };
265 
266 static struct cpudata **all_cpu_data;
267 
268 /**
269  * struct pstate_funcs - Per CPU model specific callbacks
270  * @get_max:		Callback to get maximum non turbo effective P state
271  * @get_max_physical:	Callback to get maximum non turbo physical P state
272  * @get_min:		Callback to get minimum P state
273  * @get_turbo:		Callback to get turbo P state
274  * @get_scaling:	Callback to get frequency scaling factor
275  * @get_cpu_scaling:	Get frequency scaling factor for a given cpu
276  * @get_aperf_mperf_shift: Callback to get the APERF vs MPERF frequency difference
277  * @get_val:		Callback to convert P state to actual MSR write value
278  * @get_vid:		Callback to get VID data for Atom platforms
279  *
280  * Core and Atom CPU models have different way to get P State limits. This
281  * structure is used to store those callbacks.
282  */
283 struct pstate_funcs {
284 	int (*get_max)(int cpu);
285 	int (*get_max_physical)(int cpu);
286 	int (*get_min)(int cpu);
287 	int (*get_turbo)(int cpu);
288 	int (*get_scaling)(void);
289 	int (*get_cpu_scaling)(int cpu);
290 	int (*get_aperf_mperf_shift)(void);
291 	u64 (*get_val)(struct cpudata*, int pstate);
292 	void (*get_vid)(struct cpudata *);
293 };
294 
295 static struct pstate_funcs pstate_funcs __read_mostly;
296 
297 static bool hwp_active __ro_after_init;
298 static int hwp_mode_bdw __ro_after_init;
299 static bool per_cpu_limits __ro_after_init;
300 static bool hwp_forced __ro_after_init;
301 static bool hwp_boost __read_mostly;
302 static bool hwp_is_hybrid;
303 
304 static struct cpufreq_driver *intel_pstate_driver __read_mostly;
305 
306 #define HYBRID_SCALING_FACTOR_ADL	78741
307 #define HYBRID_SCALING_FACTOR_MTL	80000
308 #define HYBRID_SCALING_FACTOR_LNL	86957
309 
310 static int hybrid_scaling_factor;
311 
312 static inline int core_get_scaling(void)
313 {
314 	return 100000;
315 }
316 
317 #ifdef CONFIG_ACPI
318 static bool acpi_ppc;
319 #endif
320 
321 static struct global_params global;
322 
323 static DEFINE_MUTEX(intel_pstate_driver_lock);
324 static DEFINE_MUTEX(intel_pstate_limits_lock);
325 
326 #ifdef CONFIG_ACPI
327 
328 static bool intel_pstate_acpi_pm_profile_server(void)
329 {
330 	if (acpi_gbl_FADT.preferred_profile == PM_ENTERPRISE_SERVER ||
331 	    acpi_gbl_FADT.preferred_profile == PM_PERFORMANCE_SERVER)
332 		return true;
333 
334 	return false;
335 }
336 
337 static bool intel_pstate_get_ppc_enable_status(void)
338 {
339 	if (intel_pstate_acpi_pm_profile_server())
340 		return true;
341 
342 	return acpi_ppc;
343 }
344 
345 #ifdef CONFIG_ACPI_CPPC_LIB
346 
347 /* The work item is needed to avoid CPU hotplug locking issues */
348 static void intel_pstste_sched_itmt_work_fn(struct work_struct *work)
349 {
350 	sched_set_itmt_support();
351 }
352 
353 static DECLARE_WORK(sched_itmt_work, intel_pstste_sched_itmt_work_fn);
354 
355 #define CPPC_MAX_PERF	U8_MAX
356 
357 static void intel_pstate_set_itmt_prio(int cpu)
358 {
359 	struct cppc_perf_caps cppc_perf;
360 	static u32 max_highest_perf = 0, min_highest_perf = U32_MAX;
361 	int ret;
362 
363 	ret = cppc_get_perf_caps(cpu, &cppc_perf);
364 	/*
365 	 * If CPPC is not available, fall back to MSR_HWP_CAPABILITIES bits [8:0].
366 	 *
367 	 * Also, on some systems with overclocking enabled, CPPC.highest_perf is
368 	 * hardcoded to 0xff, so CPPC.highest_perf cannot be used to enable ITMT.
369 	 * Fall back to MSR_HWP_CAPABILITIES then too.
370 	 */
371 	if (ret || cppc_perf.highest_perf == CPPC_MAX_PERF)
372 		cppc_perf.highest_perf = HWP_HIGHEST_PERF(READ_ONCE(all_cpu_data[cpu]->hwp_cap_cached));
373 
374 	/*
375 	 * The priorities can be set regardless of whether or not
376 	 * sched_set_itmt_support(true) has been called and it is valid to
377 	 * update them at any time after it has been called.
378 	 */
379 	sched_set_itmt_core_prio(cppc_perf.highest_perf, cpu);
380 
381 	if (max_highest_perf <= min_highest_perf) {
382 		if (cppc_perf.highest_perf > max_highest_perf)
383 			max_highest_perf = cppc_perf.highest_perf;
384 
385 		if (cppc_perf.highest_perf < min_highest_perf)
386 			min_highest_perf = cppc_perf.highest_perf;
387 
388 		if (max_highest_perf > min_highest_perf) {
389 			/*
390 			 * This code can be run during CPU online under the
391 			 * CPU hotplug locks, so sched_set_itmt_support()
392 			 * cannot be called from here.  Queue up a work item
393 			 * to invoke it.
394 			 */
395 			schedule_work(&sched_itmt_work);
396 		}
397 	}
398 }
399 
400 static int intel_pstate_get_cppc_guaranteed(int cpu)
401 {
402 	struct cppc_perf_caps cppc_perf;
403 	int ret;
404 
405 	ret = cppc_get_perf_caps(cpu, &cppc_perf);
406 	if (ret)
407 		return ret;
408 
409 	if (cppc_perf.guaranteed_perf)
410 		return cppc_perf.guaranteed_perf;
411 
412 	return cppc_perf.nominal_perf;
413 }
414 
415 static int intel_pstate_cppc_get_scaling(int cpu)
416 {
417 	struct cppc_perf_caps cppc_perf;
418 
419 	/*
420 	 * Compute the perf-to-frequency scaling factor for the given CPU if
421 	 * possible, unless it would be 0.
422 	 */
423 	if (!cppc_get_perf_caps(cpu, &cppc_perf) &&
424 	    cppc_perf.nominal_perf && cppc_perf.nominal_freq)
425 		return div_u64(cppc_perf.nominal_freq * KHZ_PER_MHZ,
426 			       cppc_perf.nominal_perf);
427 
428 	return core_get_scaling();
429 }
430 
431 #else /* CONFIG_ACPI_CPPC_LIB */
432 static inline void intel_pstate_set_itmt_prio(int cpu)
433 {
434 }
435 #endif /* CONFIG_ACPI_CPPC_LIB */
436 
437 static void intel_pstate_init_acpi_perf_limits(struct cpufreq_policy *policy)
438 {
439 	struct cpudata *cpu;
440 	int ret;
441 	int i;
442 
443 	if (hwp_active) {
444 		intel_pstate_set_itmt_prio(policy->cpu);
445 		return;
446 	}
447 
448 	if (!intel_pstate_get_ppc_enable_status())
449 		return;
450 
451 	cpu = all_cpu_data[policy->cpu];
452 
453 	ret = acpi_processor_register_performance(&cpu->acpi_perf_data,
454 						  policy->cpu);
455 	if (ret)
456 		return;
457 
458 	/*
459 	 * Check if the control value in _PSS is for PERF_CTL MSR, which should
460 	 * guarantee that the states returned by it map to the states in our
461 	 * list directly.
462 	 */
463 	if (cpu->acpi_perf_data.control_register.space_id !=
464 						ACPI_ADR_SPACE_FIXED_HARDWARE)
465 		goto err;
466 
467 	/*
468 	 * If there is only one entry _PSS, simply ignore _PSS and continue as
469 	 * usual without taking _PSS into account
470 	 */
471 	if (cpu->acpi_perf_data.state_count < 2)
472 		goto err;
473 
474 	pr_debug("CPU%u - ACPI _PSS perf data\n", policy->cpu);
475 	for (i = 0; i < cpu->acpi_perf_data.state_count; i++) {
476 		pr_debug("     %cP%d: %u MHz, %u mW, 0x%x\n",
477 			 (i == cpu->acpi_perf_data.state ? '*' : ' '), i,
478 			 (u32) cpu->acpi_perf_data.states[i].core_frequency,
479 			 (u32) cpu->acpi_perf_data.states[i].power,
480 			 (u32) cpu->acpi_perf_data.states[i].control);
481 	}
482 
483 	cpu->valid_pss_table = true;
484 	pr_debug("_PPC limits will be enforced\n");
485 
486 	return;
487 
488  err:
489 	cpu->valid_pss_table = false;
490 	acpi_processor_unregister_performance(policy->cpu);
491 }
492 
493 static void intel_pstate_exit_perf_limits(struct cpufreq_policy *policy)
494 {
495 	struct cpudata *cpu;
496 
497 	cpu = all_cpu_data[policy->cpu];
498 	if (!cpu->valid_pss_table)
499 		return;
500 
501 	acpi_processor_unregister_performance(policy->cpu);
502 }
503 #else /* CONFIG_ACPI */
504 static inline void intel_pstate_init_acpi_perf_limits(struct cpufreq_policy *policy)
505 {
506 }
507 
508 static inline void intel_pstate_exit_perf_limits(struct cpufreq_policy *policy)
509 {
510 }
511 
512 static inline bool intel_pstate_acpi_pm_profile_server(void)
513 {
514 	return false;
515 }
516 #endif /* CONFIG_ACPI */
517 
518 #ifndef CONFIG_ACPI_CPPC_LIB
519 static inline int intel_pstate_get_cppc_guaranteed(int cpu)
520 {
521 	return -ENOTSUPP;
522 }
523 
524 static int intel_pstate_cppc_get_scaling(int cpu)
525 {
526 	return core_get_scaling();
527 }
528 #endif /* CONFIG_ACPI_CPPC_LIB */
529 
530 static int intel_pstate_freq_to_hwp_rel(struct cpudata *cpu, int freq,
531 					unsigned int relation)
532 {
533 	if (freq == cpu->pstate.turbo_freq)
534 		return cpu->pstate.turbo_pstate;
535 
536 	if (freq == cpu->pstate.max_freq)
537 		return cpu->pstate.max_pstate;
538 
539 	switch (relation) {
540 	case CPUFREQ_RELATION_H:
541 		return freq / cpu->pstate.scaling;
542 	case CPUFREQ_RELATION_C:
543 		return DIV_ROUND_CLOSEST(freq, cpu->pstate.scaling);
544 	}
545 
546 	return DIV_ROUND_UP(freq, cpu->pstate.scaling);
547 }
548 
549 static int intel_pstate_freq_to_hwp(struct cpudata *cpu, int freq)
550 {
551 	return intel_pstate_freq_to_hwp_rel(cpu, freq, CPUFREQ_RELATION_L);
552 }
553 
554 /**
555  * intel_pstate_hybrid_hwp_adjust - Calibrate HWP performance levels.
556  * @cpu: Target CPU.
557  *
558  * On hybrid processors, HWP may expose more performance levels than there are
559  * P-states accessible through the PERF_CTL interface.  If that happens, the
560  * scaling factor between HWP performance levels and CPU frequency will be less
561  * than the scaling factor between P-state values and CPU frequency.
562  *
563  * In that case, adjust the CPU parameters used in computations accordingly.
564  */
565 static void intel_pstate_hybrid_hwp_adjust(struct cpudata *cpu)
566 {
567 	int perf_ctl_max_phys = cpu->pstate.max_pstate_physical;
568 	int perf_ctl_scaling = cpu->pstate.perf_ctl_scaling;
569 	int perf_ctl_turbo = pstate_funcs.get_turbo(cpu->cpu);
570 	int scaling = cpu->pstate.scaling;
571 	int freq;
572 
573 	pr_debug("CPU%d: perf_ctl_max_phys = %d\n", cpu->cpu, perf_ctl_max_phys);
574 	pr_debug("CPU%d: perf_ctl_turbo = %d\n", cpu->cpu, perf_ctl_turbo);
575 	pr_debug("CPU%d: perf_ctl_scaling = %d\n", cpu->cpu, perf_ctl_scaling);
576 	pr_debug("CPU%d: HWP_CAP guaranteed = %d\n", cpu->cpu, cpu->pstate.max_pstate);
577 	pr_debug("CPU%d: HWP_CAP highest = %d\n", cpu->cpu, cpu->pstate.turbo_pstate);
578 	pr_debug("CPU%d: HWP-to-frequency scaling factor: %d\n", cpu->cpu, scaling);
579 
580 	cpu->pstate.turbo_freq = rounddown(cpu->pstate.turbo_pstate * scaling,
581 					   perf_ctl_scaling);
582 	cpu->pstate.max_freq = rounddown(cpu->pstate.max_pstate * scaling,
583 					 perf_ctl_scaling);
584 
585 	freq = perf_ctl_max_phys * perf_ctl_scaling;
586 	cpu->pstate.max_pstate_physical = intel_pstate_freq_to_hwp(cpu, freq);
587 
588 	freq = cpu->pstate.min_pstate * perf_ctl_scaling;
589 	cpu->pstate.min_freq = freq;
590 	/*
591 	 * Cast the min P-state value retrieved via pstate_funcs.get_min() to
592 	 * the effective range of HWP performance levels.
593 	 */
594 	cpu->pstate.min_pstate = intel_pstate_freq_to_hwp(cpu, freq);
595 }
596 
597 static bool turbo_is_disabled(void)
598 {
599 	u64 misc_en;
600 
601 	rdmsrl(MSR_IA32_MISC_ENABLE, misc_en);
602 
603 	return !!(misc_en & MSR_IA32_MISC_ENABLE_TURBO_DISABLE);
604 }
605 
606 static int min_perf_pct_min(void)
607 {
608 	struct cpudata *cpu = all_cpu_data[0];
609 	int turbo_pstate = cpu->pstate.turbo_pstate;
610 
611 	return turbo_pstate ?
612 		(cpu->pstate.min_pstate * 100 / turbo_pstate) : 0;
613 }
614 
615 static s16 intel_pstate_get_epb(struct cpudata *cpu_data)
616 {
617 	u64 epb;
618 	int ret;
619 
620 	if (!boot_cpu_has(X86_FEATURE_EPB))
621 		return -ENXIO;
622 
623 	ret = rdmsrl_on_cpu(cpu_data->cpu, MSR_IA32_ENERGY_PERF_BIAS, &epb);
624 	if (ret)
625 		return (s16)ret;
626 
627 	return (s16)(epb & 0x0f);
628 }
629 
630 static s16 intel_pstate_get_epp(struct cpudata *cpu_data, u64 hwp_req_data)
631 {
632 	s16 epp;
633 
634 	if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
635 		/*
636 		 * When hwp_req_data is 0, means that caller didn't read
637 		 * MSR_HWP_REQUEST, so need to read and get EPP.
638 		 */
639 		if (!hwp_req_data) {
640 			epp = rdmsrl_on_cpu(cpu_data->cpu, MSR_HWP_REQUEST,
641 					    &hwp_req_data);
642 			if (epp)
643 				return epp;
644 		}
645 		epp = (hwp_req_data >> 24) & 0xff;
646 	} else {
647 		/* When there is no EPP present, HWP uses EPB settings */
648 		epp = intel_pstate_get_epb(cpu_data);
649 	}
650 
651 	return epp;
652 }
653 
654 static int intel_pstate_set_epb(int cpu, s16 pref)
655 {
656 	u64 epb;
657 	int ret;
658 
659 	if (!boot_cpu_has(X86_FEATURE_EPB))
660 		return -ENXIO;
661 
662 	ret = rdmsrl_on_cpu(cpu, MSR_IA32_ENERGY_PERF_BIAS, &epb);
663 	if (ret)
664 		return ret;
665 
666 	epb = (epb & ~0x0f) | pref;
667 	wrmsrl_on_cpu(cpu, MSR_IA32_ENERGY_PERF_BIAS, epb);
668 
669 	return 0;
670 }
671 
672 /*
673  * EPP/EPB display strings corresponding to EPP index in the
674  * energy_perf_strings[]
675  *	index		String
676  *-------------------------------------
677  *	0		default
678  *	1		performance
679  *	2		balance_performance
680  *	3		balance_power
681  *	4		power
682  */
683 
684 enum energy_perf_value_index {
685 	EPP_INDEX_DEFAULT = 0,
686 	EPP_INDEX_PERFORMANCE,
687 	EPP_INDEX_BALANCE_PERFORMANCE,
688 	EPP_INDEX_BALANCE_POWERSAVE,
689 	EPP_INDEX_POWERSAVE,
690 };
691 
692 static const char * const energy_perf_strings[] = {
693 	[EPP_INDEX_DEFAULT] = "default",
694 	[EPP_INDEX_PERFORMANCE] = "performance",
695 	[EPP_INDEX_BALANCE_PERFORMANCE] = "balance_performance",
696 	[EPP_INDEX_BALANCE_POWERSAVE] = "balance_power",
697 	[EPP_INDEX_POWERSAVE] = "power",
698 	NULL
699 };
700 static unsigned int epp_values[] = {
701 	[EPP_INDEX_DEFAULT] = 0, /* Unused index */
702 	[EPP_INDEX_PERFORMANCE] = HWP_EPP_PERFORMANCE,
703 	[EPP_INDEX_BALANCE_PERFORMANCE] = HWP_EPP_BALANCE_PERFORMANCE,
704 	[EPP_INDEX_BALANCE_POWERSAVE] = HWP_EPP_BALANCE_POWERSAVE,
705 	[EPP_INDEX_POWERSAVE] = HWP_EPP_POWERSAVE,
706 };
707 
708 static int intel_pstate_get_energy_pref_index(struct cpudata *cpu_data, int *raw_epp)
709 {
710 	s16 epp;
711 	int index = -EINVAL;
712 
713 	*raw_epp = 0;
714 	epp = intel_pstate_get_epp(cpu_data, 0);
715 	if (epp < 0)
716 		return epp;
717 
718 	if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
719 		if (epp == epp_values[EPP_INDEX_PERFORMANCE])
720 			return EPP_INDEX_PERFORMANCE;
721 		if (epp == epp_values[EPP_INDEX_BALANCE_PERFORMANCE])
722 			return EPP_INDEX_BALANCE_PERFORMANCE;
723 		if (epp == epp_values[EPP_INDEX_BALANCE_POWERSAVE])
724 			return EPP_INDEX_BALANCE_POWERSAVE;
725 		if (epp == epp_values[EPP_INDEX_POWERSAVE])
726 			return EPP_INDEX_POWERSAVE;
727 		*raw_epp = epp;
728 		return 0;
729 	} else if (boot_cpu_has(X86_FEATURE_EPB)) {
730 		/*
731 		 * Range:
732 		 *	0x00-0x03	:	Performance
733 		 *	0x04-0x07	:	Balance performance
734 		 *	0x08-0x0B	:	Balance power
735 		 *	0x0C-0x0F	:	Power
736 		 * The EPB is a 4 bit value, but our ranges restrict the
737 		 * value which can be set. Here only using top two bits
738 		 * effectively.
739 		 */
740 		index = (epp >> 2) + 1;
741 	}
742 
743 	return index;
744 }
745 
746 static int intel_pstate_set_epp(struct cpudata *cpu, u32 epp)
747 {
748 	int ret;
749 
750 	/*
751 	 * Use the cached HWP Request MSR value, because in the active mode the
752 	 * register itself may be updated by intel_pstate_hwp_boost_up() or
753 	 * intel_pstate_hwp_boost_down() at any time.
754 	 */
755 	u64 value = READ_ONCE(cpu->hwp_req_cached);
756 
757 	value &= ~GENMASK_ULL(31, 24);
758 	value |= (u64)epp << 24;
759 	/*
760 	 * The only other updater of hwp_req_cached in the active mode,
761 	 * intel_pstate_hwp_set(), is called under the same lock as this
762 	 * function, so it cannot run in parallel with the update below.
763 	 */
764 	WRITE_ONCE(cpu->hwp_req_cached, value);
765 	ret = wrmsrl_on_cpu(cpu->cpu, MSR_HWP_REQUEST, value);
766 	if (!ret)
767 		cpu->epp_cached = epp;
768 
769 	return ret;
770 }
771 
772 static int intel_pstate_set_energy_pref_index(struct cpudata *cpu_data,
773 					      int pref_index, bool use_raw,
774 					      u32 raw_epp)
775 {
776 	int epp = -EINVAL;
777 	int ret;
778 
779 	if (!pref_index)
780 		epp = cpu_data->epp_default;
781 
782 	if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
783 		if (use_raw)
784 			epp = raw_epp;
785 		else if (epp == -EINVAL)
786 			epp = epp_values[pref_index];
787 
788 		/*
789 		 * To avoid confusion, refuse to set EPP to any values different
790 		 * from 0 (performance) if the current policy is "performance",
791 		 * because those values would be overridden.
792 		 */
793 		if (epp > 0 && cpu_data->policy == CPUFREQ_POLICY_PERFORMANCE)
794 			return -EBUSY;
795 
796 		ret = intel_pstate_set_epp(cpu_data, epp);
797 	} else {
798 		if (epp == -EINVAL)
799 			epp = (pref_index - 1) << 2;
800 		ret = intel_pstate_set_epb(cpu_data->cpu, epp);
801 	}
802 
803 	return ret;
804 }
805 
806 static ssize_t show_energy_performance_available_preferences(
807 				struct cpufreq_policy *policy, char *buf)
808 {
809 	int i = 0;
810 	int ret = 0;
811 
812 	while (energy_perf_strings[i] != NULL)
813 		ret += sprintf(&buf[ret], "%s ", energy_perf_strings[i++]);
814 
815 	ret += sprintf(&buf[ret], "\n");
816 
817 	return ret;
818 }
819 
820 cpufreq_freq_attr_ro(energy_performance_available_preferences);
821 
822 static struct cpufreq_driver intel_pstate;
823 
824 static ssize_t store_energy_performance_preference(
825 		struct cpufreq_policy *policy, const char *buf, size_t count)
826 {
827 	struct cpudata *cpu = all_cpu_data[policy->cpu];
828 	char str_preference[21];
829 	bool raw = false;
830 	ssize_t ret;
831 	u32 epp = 0;
832 
833 	ret = sscanf(buf, "%20s", str_preference);
834 	if (ret != 1)
835 		return -EINVAL;
836 
837 	ret = match_string(energy_perf_strings, -1, str_preference);
838 	if (ret < 0) {
839 		if (!boot_cpu_has(X86_FEATURE_HWP_EPP))
840 			return ret;
841 
842 		ret = kstrtouint(buf, 10, &epp);
843 		if (ret)
844 			return ret;
845 
846 		if (epp > 255)
847 			return -EINVAL;
848 
849 		raw = true;
850 	}
851 
852 	/*
853 	 * This function runs with the policy R/W semaphore held, which
854 	 * guarantees that the driver pointer will not change while it is
855 	 * running.
856 	 */
857 	if (!intel_pstate_driver)
858 		return -EAGAIN;
859 
860 	mutex_lock(&intel_pstate_limits_lock);
861 
862 	if (intel_pstate_driver == &intel_pstate) {
863 		ret = intel_pstate_set_energy_pref_index(cpu, ret, raw, epp);
864 	} else {
865 		/*
866 		 * In the passive mode the governor needs to be stopped on the
867 		 * target CPU before the EPP update and restarted after it,
868 		 * which is super-heavy-weight, so make sure it is worth doing
869 		 * upfront.
870 		 */
871 		if (!raw)
872 			epp = ret ? epp_values[ret] : cpu->epp_default;
873 
874 		if (cpu->epp_cached != epp) {
875 			int err;
876 
877 			cpufreq_stop_governor(policy);
878 			ret = intel_pstate_set_epp(cpu, epp);
879 			err = cpufreq_start_governor(policy);
880 			if (!ret)
881 				ret = err;
882 		} else {
883 			ret = 0;
884 		}
885 	}
886 
887 	mutex_unlock(&intel_pstate_limits_lock);
888 
889 	return ret ?: count;
890 }
891 
892 static ssize_t show_energy_performance_preference(
893 				struct cpufreq_policy *policy, char *buf)
894 {
895 	struct cpudata *cpu_data = all_cpu_data[policy->cpu];
896 	int preference, raw_epp;
897 
898 	preference = intel_pstate_get_energy_pref_index(cpu_data, &raw_epp);
899 	if (preference < 0)
900 		return preference;
901 
902 	if (raw_epp)
903 		return  sprintf(buf, "%d\n", raw_epp);
904 	else
905 		return  sprintf(buf, "%s\n", energy_perf_strings[preference]);
906 }
907 
908 cpufreq_freq_attr_rw(energy_performance_preference);
909 
910 static ssize_t show_base_frequency(struct cpufreq_policy *policy, char *buf)
911 {
912 	struct cpudata *cpu = all_cpu_data[policy->cpu];
913 	int ratio, freq;
914 
915 	ratio = intel_pstate_get_cppc_guaranteed(policy->cpu);
916 	if (ratio <= 0) {
917 		u64 cap;
918 
919 		rdmsrl_on_cpu(policy->cpu, MSR_HWP_CAPABILITIES, &cap);
920 		ratio = HWP_GUARANTEED_PERF(cap);
921 	}
922 
923 	freq = ratio * cpu->pstate.scaling;
924 	if (cpu->pstate.scaling != cpu->pstate.perf_ctl_scaling)
925 		freq = rounddown(freq, cpu->pstate.perf_ctl_scaling);
926 
927 	return sprintf(buf, "%d\n", freq);
928 }
929 
930 cpufreq_freq_attr_ro(base_frequency);
931 
932 static struct freq_attr *hwp_cpufreq_attrs[] = {
933 	&energy_performance_preference,
934 	&energy_performance_available_preferences,
935 	&base_frequency,
936 	NULL,
937 };
938 
939 static bool no_cas __ro_after_init;
940 
941 static struct cpudata *hybrid_max_perf_cpu __read_mostly;
942 /*
943  * Protects hybrid_max_perf_cpu, the capacity_perf fields in struct cpudata,
944  * and the x86 arch scale-invariance information from concurrent updates.
945  */
946 static DEFINE_MUTEX(hybrid_capacity_lock);
947 
948 static void hybrid_set_cpu_capacity(struct cpudata *cpu)
949 {
950 	arch_set_cpu_capacity(cpu->cpu, cpu->capacity_perf,
951 			      hybrid_max_perf_cpu->capacity_perf,
952 			      cpu->capacity_perf,
953 			      cpu->pstate.max_pstate_physical);
954 
955 	pr_debug("CPU%d: perf = %u, max. perf = %u, base perf = %d\n", cpu->cpu,
956 		 cpu->capacity_perf, hybrid_max_perf_cpu->capacity_perf,
957 		 cpu->pstate.max_pstate_physical);
958 }
959 
960 static void hybrid_clear_cpu_capacity(unsigned int cpunum)
961 {
962 	arch_set_cpu_capacity(cpunum, 1, 1, 1, 1);
963 }
964 
965 static void hybrid_get_capacity_perf(struct cpudata *cpu)
966 {
967 	if (READ_ONCE(global.no_turbo)) {
968 		cpu->capacity_perf = cpu->pstate.max_pstate_physical;
969 		return;
970 	}
971 
972 	cpu->capacity_perf = HWP_HIGHEST_PERF(READ_ONCE(cpu->hwp_cap_cached));
973 }
974 
975 static void hybrid_set_capacity_of_cpus(void)
976 {
977 	int cpunum;
978 
979 	for_each_online_cpu(cpunum) {
980 		struct cpudata *cpu = all_cpu_data[cpunum];
981 
982 		if (cpu)
983 			hybrid_set_cpu_capacity(cpu);
984 	}
985 }
986 
987 static void hybrid_update_cpu_capacity_scaling(void)
988 {
989 	struct cpudata *max_perf_cpu = NULL;
990 	unsigned int max_cap_perf = 0;
991 	int cpunum;
992 
993 	for_each_online_cpu(cpunum) {
994 		struct cpudata *cpu = all_cpu_data[cpunum];
995 
996 		if (!cpu)
997 			continue;
998 
999 		/*
1000 		 * During initialization, CPU performance at full capacity needs
1001 		 * to be determined.
1002 		 */
1003 		if (!hybrid_max_perf_cpu)
1004 			hybrid_get_capacity_perf(cpu);
1005 
1006 		/*
1007 		 * If hybrid_max_perf_cpu is not NULL at this point, it is
1008 		 * being replaced, so don't take it into account when looking
1009 		 * for the new one.
1010 		 */
1011 		if (cpu == hybrid_max_perf_cpu)
1012 			continue;
1013 
1014 		if (cpu->capacity_perf > max_cap_perf) {
1015 			max_cap_perf = cpu->capacity_perf;
1016 			max_perf_cpu = cpu;
1017 		}
1018 	}
1019 
1020 	if (max_perf_cpu) {
1021 		hybrid_max_perf_cpu = max_perf_cpu;
1022 		hybrid_set_capacity_of_cpus();
1023 	} else {
1024 		pr_info("Found no CPUs with nonzero maximum performance\n");
1025 		/* Revert to the flat CPU capacity structure. */
1026 		for_each_online_cpu(cpunum)
1027 			hybrid_clear_cpu_capacity(cpunum);
1028 	}
1029 }
1030 
1031 static void __hybrid_refresh_cpu_capacity_scaling(void)
1032 {
1033 	hybrid_max_perf_cpu = NULL;
1034 	hybrid_update_cpu_capacity_scaling();
1035 }
1036 
1037 static void hybrid_refresh_cpu_capacity_scaling(void)
1038 {
1039 	guard(mutex)(&hybrid_capacity_lock);
1040 
1041 	__hybrid_refresh_cpu_capacity_scaling();
1042 }
1043 
1044 static void hybrid_init_cpu_capacity_scaling(bool refresh)
1045 {
1046 	/* Bail out if enabling capacity-aware scheduling is prohibited. */
1047 	if (no_cas)
1048 		return;
1049 
1050 	/*
1051 	 * If hybrid_max_perf_cpu is set at this point, the hybrid CPU capacity
1052 	 * scaling has been enabled already and the driver is just changing the
1053 	 * operation mode.
1054 	 */
1055 	if (refresh) {
1056 		hybrid_refresh_cpu_capacity_scaling();
1057 		return;
1058 	}
1059 
1060 	/*
1061 	 * On hybrid systems, use asym capacity instead of ITMT, but because
1062 	 * the capacity of SMT threads is not deterministic even approximately,
1063 	 * do not do that when SMT is in use.
1064 	 */
1065 	if (hwp_is_hybrid && !sched_smt_active() && arch_enable_hybrid_capacity_scale()) {
1066 		hybrid_refresh_cpu_capacity_scaling();
1067 		/*
1068 		 * Disabling ITMT causes sched domains to be rebuilt to disable asym
1069 		 * packing and enable asym capacity.
1070 		 */
1071 		sched_clear_itmt_support();
1072 	}
1073 }
1074 
1075 static bool hybrid_clear_max_perf_cpu(void)
1076 {
1077 	bool ret;
1078 
1079 	guard(mutex)(&hybrid_capacity_lock);
1080 
1081 	ret = !!hybrid_max_perf_cpu;
1082 	hybrid_max_perf_cpu = NULL;
1083 
1084 	return ret;
1085 }
1086 
1087 static void __intel_pstate_get_hwp_cap(struct cpudata *cpu)
1088 {
1089 	u64 cap;
1090 
1091 	rdmsrl_on_cpu(cpu->cpu, MSR_HWP_CAPABILITIES, &cap);
1092 	WRITE_ONCE(cpu->hwp_cap_cached, cap);
1093 	cpu->pstate.max_pstate = HWP_GUARANTEED_PERF(cap);
1094 	cpu->pstate.turbo_pstate = HWP_HIGHEST_PERF(cap);
1095 }
1096 
1097 static void intel_pstate_get_hwp_cap(struct cpudata *cpu)
1098 {
1099 	int scaling = cpu->pstate.scaling;
1100 
1101 	__intel_pstate_get_hwp_cap(cpu);
1102 
1103 	cpu->pstate.max_freq = cpu->pstate.max_pstate * scaling;
1104 	cpu->pstate.turbo_freq = cpu->pstate.turbo_pstate * scaling;
1105 	if (scaling != cpu->pstate.perf_ctl_scaling) {
1106 		int perf_ctl_scaling = cpu->pstate.perf_ctl_scaling;
1107 
1108 		cpu->pstate.max_freq = rounddown(cpu->pstate.max_freq,
1109 						 perf_ctl_scaling);
1110 		cpu->pstate.turbo_freq = rounddown(cpu->pstate.turbo_freq,
1111 						   perf_ctl_scaling);
1112 	}
1113 }
1114 
1115 static void hybrid_update_capacity(struct cpudata *cpu)
1116 {
1117 	unsigned int max_cap_perf;
1118 
1119 	mutex_lock(&hybrid_capacity_lock);
1120 
1121 	if (!hybrid_max_perf_cpu)
1122 		goto unlock;
1123 
1124 	/*
1125 	 * The maximum performance of the CPU may have changed, but assume
1126 	 * that the performance of the other CPUs has not changed.
1127 	 */
1128 	max_cap_perf = hybrid_max_perf_cpu->capacity_perf;
1129 
1130 	intel_pstate_get_hwp_cap(cpu);
1131 
1132 	hybrid_get_capacity_perf(cpu);
1133 	/* Should hybrid_max_perf_cpu be replaced by this CPU? */
1134 	if (cpu->capacity_perf > max_cap_perf) {
1135 		hybrid_max_perf_cpu = cpu;
1136 		hybrid_set_capacity_of_cpus();
1137 		goto unlock;
1138 	}
1139 
1140 	/* If this CPU is hybrid_max_perf_cpu, should it be replaced? */
1141 	if (cpu == hybrid_max_perf_cpu && cpu->capacity_perf < max_cap_perf) {
1142 		hybrid_update_cpu_capacity_scaling();
1143 		goto unlock;
1144 	}
1145 
1146 	hybrid_set_cpu_capacity(cpu);
1147 
1148 unlock:
1149 	mutex_unlock(&hybrid_capacity_lock);
1150 }
1151 
1152 static void intel_pstate_hwp_set(unsigned int cpu)
1153 {
1154 	struct cpudata *cpu_data = all_cpu_data[cpu];
1155 	int max, min;
1156 	u64 value;
1157 	s16 epp;
1158 
1159 	max = cpu_data->max_perf_ratio;
1160 	min = cpu_data->min_perf_ratio;
1161 
1162 	if (cpu_data->policy == CPUFREQ_POLICY_PERFORMANCE)
1163 		min = max;
1164 
1165 	rdmsrl_on_cpu(cpu, MSR_HWP_REQUEST, &value);
1166 
1167 	value &= ~HWP_MIN_PERF(~0L);
1168 	value |= HWP_MIN_PERF(min);
1169 
1170 	value &= ~HWP_MAX_PERF(~0L);
1171 	value |= HWP_MAX_PERF(max);
1172 
1173 	if (cpu_data->epp_policy == cpu_data->policy)
1174 		goto skip_epp;
1175 
1176 	cpu_data->epp_policy = cpu_data->policy;
1177 
1178 	if (cpu_data->policy == CPUFREQ_POLICY_PERFORMANCE) {
1179 		epp = intel_pstate_get_epp(cpu_data, value);
1180 		cpu_data->epp_powersave = epp;
1181 		/* If EPP read was failed, then don't try to write */
1182 		if (epp < 0)
1183 			goto skip_epp;
1184 
1185 		epp = 0;
1186 	} else {
1187 		/* skip setting EPP, when saved value is invalid */
1188 		if (cpu_data->epp_powersave < 0)
1189 			goto skip_epp;
1190 
1191 		/*
1192 		 * No need to restore EPP when it is not zero. This
1193 		 * means:
1194 		 *  - Policy is not changed
1195 		 *  - user has manually changed
1196 		 *  - Error reading EPB
1197 		 */
1198 		epp = intel_pstate_get_epp(cpu_data, value);
1199 		if (epp)
1200 			goto skip_epp;
1201 
1202 		epp = cpu_data->epp_powersave;
1203 	}
1204 	if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
1205 		value &= ~GENMASK_ULL(31, 24);
1206 		value |= (u64)epp << 24;
1207 	} else {
1208 		intel_pstate_set_epb(cpu, epp);
1209 	}
1210 skip_epp:
1211 	WRITE_ONCE(cpu_data->hwp_req_cached, value);
1212 	wrmsrl_on_cpu(cpu, MSR_HWP_REQUEST, value);
1213 }
1214 
1215 static void intel_pstate_disable_hwp_interrupt(struct cpudata *cpudata);
1216 
1217 static void intel_pstate_hwp_offline(struct cpudata *cpu)
1218 {
1219 	u64 value = READ_ONCE(cpu->hwp_req_cached);
1220 	int min_perf;
1221 
1222 	intel_pstate_disable_hwp_interrupt(cpu);
1223 
1224 	if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
1225 		/*
1226 		 * In case the EPP has been set to "performance" by the
1227 		 * active mode "performance" scaling algorithm, replace that
1228 		 * temporary value with the cached EPP one.
1229 		 */
1230 		value &= ~GENMASK_ULL(31, 24);
1231 		value |= HWP_ENERGY_PERF_PREFERENCE(cpu->epp_cached);
1232 		/*
1233 		 * However, make sure that EPP will be set to "performance" when
1234 		 * the CPU is brought back online again and the "performance"
1235 		 * scaling algorithm is still in effect.
1236 		 */
1237 		cpu->epp_policy = CPUFREQ_POLICY_UNKNOWN;
1238 	}
1239 
1240 	/*
1241 	 * Clear the desired perf field in the cached HWP request value to
1242 	 * prevent nonzero desired values from being leaked into the active
1243 	 * mode.
1244 	 */
1245 	value &= ~HWP_DESIRED_PERF(~0L);
1246 	WRITE_ONCE(cpu->hwp_req_cached, value);
1247 
1248 	value &= ~GENMASK_ULL(31, 0);
1249 	min_perf = HWP_LOWEST_PERF(READ_ONCE(cpu->hwp_cap_cached));
1250 
1251 	/* Set hwp_max = hwp_min */
1252 	value |= HWP_MAX_PERF(min_perf);
1253 	value |= HWP_MIN_PERF(min_perf);
1254 
1255 	/* Set EPP to min */
1256 	if (boot_cpu_has(X86_FEATURE_HWP_EPP))
1257 		value |= HWP_ENERGY_PERF_PREFERENCE(HWP_EPP_POWERSAVE);
1258 
1259 	wrmsrl_on_cpu(cpu->cpu, MSR_HWP_REQUEST, value);
1260 
1261 	mutex_lock(&hybrid_capacity_lock);
1262 
1263 	if (!hybrid_max_perf_cpu) {
1264 		mutex_unlock(&hybrid_capacity_lock);
1265 
1266 		return;
1267 	}
1268 
1269 	if (hybrid_max_perf_cpu == cpu)
1270 		hybrid_update_cpu_capacity_scaling();
1271 
1272 	mutex_unlock(&hybrid_capacity_lock);
1273 
1274 	/* Reset the capacity of the CPU going offline to the initial value. */
1275 	hybrid_clear_cpu_capacity(cpu->cpu);
1276 }
1277 
1278 #define POWER_CTL_EE_ENABLE	1
1279 #define POWER_CTL_EE_DISABLE	2
1280 
1281 static int power_ctl_ee_state;
1282 
1283 static void set_power_ctl_ee_state(bool input)
1284 {
1285 	u64 power_ctl;
1286 
1287 	mutex_lock(&intel_pstate_driver_lock);
1288 	rdmsrl(MSR_IA32_POWER_CTL, power_ctl);
1289 	if (input) {
1290 		power_ctl &= ~BIT(MSR_IA32_POWER_CTL_BIT_EE);
1291 		power_ctl_ee_state = POWER_CTL_EE_ENABLE;
1292 	} else {
1293 		power_ctl |= BIT(MSR_IA32_POWER_CTL_BIT_EE);
1294 		power_ctl_ee_state = POWER_CTL_EE_DISABLE;
1295 	}
1296 	wrmsrl(MSR_IA32_POWER_CTL, power_ctl);
1297 	mutex_unlock(&intel_pstate_driver_lock);
1298 }
1299 
1300 static void intel_pstate_hwp_enable(struct cpudata *cpudata);
1301 
1302 static void intel_pstate_hwp_reenable(struct cpudata *cpu)
1303 {
1304 	intel_pstate_hwp_enable(cpu);
1305 	wrmsrl_on_cpu(cpu->cpu, MSR_HWP_REQUEST, READ_ONCE(cpu->hwp_req_cached));
1306 }
1307 
1308 static int intel_pstate_suspend(struct cpufreq_policy *policy)
1309 {
1310 	struct cpudata *cpu = all_cpu_data[policy->cpu];
1311 
1312 	pr_debug("CPU %d suspending\n", cpu->cpu);
1313 
1314 	cpu->suspended = true;
1315 
1316 	/* disable HWP interrupt and cancel any pending work */
1317 	intel_pstate_disable_hwp_interrupt(cpu);
1318 
1319 	return 0;
1320 }
1321 
1322 static int intel_pstate_resume(struct cpufreq_policy *policy)
1323 {
1324 	struct cpudata *cpu = all_cpu_data[policy->cpu];
1325 
1326 	pr_debug("CPU %d resuming\n", cpu->cpu);
1327 
1328 	/* Only restore if the system default is changed */
1329 	if (power_ctl_ee_state == POWER_CTL_EE_ENABLE)
1330 		set_power_ctl_ee_state(true);
1331 	else if (power_ctl_ee_state == POWER_CTL_EE_DISABLE)
1332 		set_power_ctl_ee_state(false);
1333 
1334 	if (cpu->suspended && hwp_active) {
1335 		mutex_lock(&intel_pstate_limits_lock);
1336 
1337 		/* Re-enable HWP, because "online" has not done that. */
1338 		intel_pstate_hwp_reenable(cpu);
1339 
1340 		mutex_unlock(&intel_pstate_limits_lock);
1341 	}
1342 
1343 	cpu->suspended = false;
1344 
1345 	return 0;
1346 }
1347 
1348 static void intel_pstate_update_policies(void)
1349 {
1350 	int cpu;
1351 
1352 	for_each_possible_cpu(cpu)
1353 		cpufreq_update_policy(cpu);
1354 }
1355 
1356 static void __intel_pstate_update_max_freq(struct cpudata *cpudata,
1357 					   struct cpufreq_policy *policy)
1358 {
1359 	if (hwp_active)
1360 		intel_pstate_get_hwp_cap(cpudata);
1361 
1362 	policy->cpuinfo.max_freq = READ_ONCE(global.no_turbo) ?
1363 			cpudata->pstate.max_freq : cpudata->pstate.turbo_freq;
1364 
1365 	refresh_frequency_limits(policy);
1366 }
1367 
1368 static void intel_pstate_update_limits(unsigned int cpu)
1369 {
1370 	struct cpufreq_policy *policy = cpufreq_cpu_acquire(cpu);
1371 	struct cpudata *cpudata;
1372 
1373 	if (!policy)
1374 		return;
1375 
1376 	cpudata = all_cpu_data[cpu];
1377 
1378 	__intel_pstate_update_max_freq(cpudata, policy);
1379 
1380 	/* Prevent the driver from being unregistered now. */
1381 	mutex_lock(&intel_pstate_driver_lock);
1382 
1383 	cpufreq_cpu_release(policy);
1384 
1385 	hybrid_update_capacity(cpudata);
1386 
1387 	mutex_unlock(&intel_pstate_driver_lock);
1388 }
1389 
1390 static void intel_pstate_update_limits_for_all(void)
1391 {
1392 	int cpu;
1393 
1394 	for_each_possible_cpu(cpu) {
1395 		struct cpufreq_policy *policy = cpufreq_cpu_acquire(cpu);
1396 
1397 		if (!policy)
1398 			continue;
1399 
1400 		__intel_pstate_update_max_freq(all_cpu_data[cpu], policy);
1401 
1402 		cpufreq_cpu_release(policy);
1403 	}
1404 
1405 	mutex_lock(&hybrid_capacity_lock);
1406 
1407 	if (hybrid_max_perf_cpu)
1408 		__hybrid_refresh_cpu_capacity_scaling();
1409 
1410 	mutex_unlock(&hybrid_capacity_lock);
1411 }
1412 
1413 /************************** sysfs begin ************************/
1414 #define show_one(file_name, object)					\
1415 	static ssize_t show_##file_name					\
1416 	(struct kobject *kobj, struct kobj_attribute *attr, char *buf)	\
1417 	{								\
1418 		return sprintf(buf, "%u\n", global.object);		\
1419 	}
1420 
1421 static ssize_t intel_pstate_show_status(char *buf);
1422 static int intel_pstate_update_status(const char *buf, size_t size);
1423 
1424 static ssize_t show_status(struct kobject *kobj,
1425 			   struct kobj_attribute *attr, char *buf)
1426 {
1427 	ssize_t ret;
1428 
1429 	mutex_lock(&intel_pstate_driver_lock);
1430 	ret = intel_pstate_show_status(buf);
1431 	mutex_unlock(&intel_pstate_driver_lock);
1432 
1433 	return ret;
1434 }
1435 
1436 static ssize_t store_status(struct kobject *a, struct kobj_attribute *b,
1437 			    const char *buf, size_t count)
1438 {
1439 	char *p = memchr(buf, '\n', count);
1440 	int ret;
1441 
1442 	mutex_lock(&intel_pstate_driver_lock);
1443 	ret = intel_pstate_update_status(buf, p ? p - buf : count);
1444 	mutex_unlock(&intel_pstate_driver_lock);
1445 
1446 	return ret < 0 ? ret : count;
1447 }
1448 
1449 static ssize_t show_turbo_pct(struct kobject *kobj,
1450 				struct kobj_attribute *attr, char *buf)
1451 {
1452 	struct cpudata *cpu;
1453 	int total, no_turbo, turbo_pct;
1454 	uint32_t turbo_fp;
1455 
1456 	mutex_lock(&intel_pstate_driver_lock);
1457 
1458 	if (!intel_pstate_driver) {
1459 		mutex_unlock(&intel_pstate_driver_lock);
1460 		return -EAGAIN;
1461 	}
1462 
1463 	cpu = all_cpu_data[0];
1464 
1465 	total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
1466 	no_turbo = cpu->pstate.max_pstate - cpu->pstate.min_pstate + 1;
1467 	turbo_fp = div_fp(no_turbo, total);
1468 	turbo_pct = 100 - fp_toint(mul_fp(turbo_fp, int_tofp(100)));
1469 
1470 	mutex_unlock(&intel_pstate_driver_lock);
1471 
1472 	return sprintf(buf, "%u\n", turbo_pct);
1473 }
1474 
1475 static ssize_t show_num_pstates(struct kobject *kobj,
1476 				struct kobj_attribute *attr, char *buf)
1477 {
1478 	struct cpudata *cpu;
1479 	int total;
1480 
1481 	mutex_lock(&intel_pstate_driver_lock);
1482 
1483 	if (!intel_pstate_driver) {
1484 		mutex_unlock(&intel_pstate_driver_lock);
1485 		return -EAGAIN;
1486 	}
1487 
1488 	cpu = all_cpu_data[0];
1489 	total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
1490 
1491 	mutex_unlock(&intel_pstate_driver_lock);
1492 
1493 	return sprintf(buf, "%u\n", total);
1494 }
1495 
1496 static ssize_t show_no_turbo(struct kobject *kobj,
1497 			     struct kobj_attribute *attr, char *buf)
1498 {
1499 	ssize_t ret;
1500 
1501 	mutex_lock(&intel_pstate_driver_lock);
1502 
1503 	if (!intel_pstate_driver) {
1504 		mutex_unlock(&intel_pstate_driver_lock);
1505 		return -EAGAIN;
1506 	}
1507 
1508 	ret = sprintf(buf, "%u\n", global.no_turbo);
1509 
1510 	mutex_unlock(&intel_pstate_driver_lock);
1511 
1512 	return ret;
1513 }
1514 
1515 static ssize_t store_no_turbo(struct kobject *a, struct kobj_attribute *b,
1516 			      const char *buf, size_t count)
1517 {
1518 	unsigned int input;
1519 	bool no_turbo;
1520 
1521 	if (sscanf(buf, "%u", &input) != 1)
1522 		return -EINVAL;
1523 
1524 	mutex_lock(&intel_pstate_driver_lock);
1525 
1526 	if (!intel_pstate_driver) {
1527 		count = -EAGAIN;
1528 		goto unlock_driver;
1529 	}
1530 
1531 	no_turbo = !!clamp_t(int, input, 0, 1);
1532 
1533 	WRITE_ONCE(global.turbo_disabled, turbo_is_disabled());
1534 	if (global.turbo_disabled && !no_turbo) {
1535 		pr_notice("Turbo disabled by BIOS or unavailable on processor\n");
1536 		count = -EPERM;
1537 		if (global.no_turbo)
1538 			goto unlock_driver;
1539 		else
1540 			no_turbo = 1;
1541 	}
1542 
1543 	if (no_turbo == global.no_turbo) {
1544 		goto unlock_driver;
1545 	}
1546 
1547 	WRITE_ONCE(global.no_turbo, no_turbo);
1548 
1549 	mutex_lock(&intel_pstate_limits_lock);
1550 
1551 	if (no_turbo) {
1552 		struct cpudata *cpu = all_cpu_data[0];
1553 		int pct = cpu->pstate.max_pstate * 100 / cpu->pstate.turbo_pstate;
1554 
1555 		/* Squash the global minimum into the permitted range. */
1556 		if (global.min_perf_pct > pct)
1557 			global.min_perf_pct = pct;
1558 	}
1559 
1560 	mutex_unlock(&intel_pstate_limits_lock);
1561 
1562 	intel_pstate_update_limits_for_all();
1563 	arch_set_max_freq_ratio(no_turbo);
1564 
1565 unlock_driver:
1566 	mutex_unlock(&intel_pstate_driver_lock);
1567 
1568 	return count;
1569 }
1570 
1571 static void update_qos_request(enum freq_qos_req_type type)
1572 {
1573 	struct freq_qos_request *req;
1574 	struct cpufreq_policy *policy;
1575 	int i;
1576 
1577 	for_each_possible_cpu(i) {
1578 		struct cpudata *cpu = all_cpu_data[i];
1579 		unsigned int freq, perf_pct;
1580 
1581 		policy = cpufreq_cpu_get(i);
1582 		if (!policy)
1583 			continue;
1584 
1585 		req = policy->driver_data;
1586 		cpufreq_cpu_put(policy);
1587 
1588 		if (!req)
1589 			continue;
1590 
1591 		if (hwp_active)
1592 			intel_pstate_get_hwp_cap(cpu);
1593 
1594 		if (type == FREQ_QOS_MIN) {
1595 			perf_pct = global.min_perf_pct;
1596 		} else {
1597 			req++;
1598 			perf_pct = global.max_perf_pct;
1599 		}
1600 
1601 		freq = DIV_ROUND_UP(cpu->pstate.turbo_freq * perf_pct, 100);
1602 
1603 		if (freq_qos_update_request(req, freq) < 0)
1604 			pr_warn("Failed to update freq constraint: CPU%d\n", i);
1605 	}
1606 }
1607 
1608 static ssize_t store_max_perf_pct(struct kobject *a, struct kobj_attribute *b,
1609 				  const char *buf, size_t count)
1610 {
1611 	unsigned int input;
1612 	int ret;
1613 
1614 	ret = sscanf(buf, "%u", &input);
1615 	if (ret != 1)
1616 		return -EINVAL;
1617 
1618 	mutex_lock(&intel_pstate_driver_lock);
1619 
1620 	if (!intel_pstate_driver) {
1621 		mutex_unlock(&intel_pstate_driver_lock);
1622 		return -EAGAIN;
1623 	}
1624 
1625 	mutex_lock(&intel_pstate_limits_lock);
1626 
1627 	global.max_perf_pct = clamp_t(int, input, global.min_perf_pct, 100);
1628 
1629 	mutex_unlock(&intel_pstate_limits_lock);
1630 
1631 	if (intel_pstate_driver == &intel_pstate)
1632 		intel_pstate_update_policies();
1633 	else
1634 		update_qos_request(FREQ_QOS_MAX);
1635 
1636 	mutex_unlock(&intel_pstate_driver_lock);
1637 
1638 	return count;
1639 }
1640 
1641 static ssize_t store_min_perf_pct(struct kobject *a, struct kobj_attribute *b,
1642 				  const char *buf, size_t count)
1643 {
1644 	unsigned int input;
1645 	int ret;
1646 
1647 	ret = sscanf(buf, "%u", &input);
1648 	if (ret != 1)
1649 		return -EINVAL;
1650 
1651 	mutex_lock(&intel_pstate_driver_lock);
1652 
1653 	if (!intel_pstate_driver) {
1654 		mutex_unlock(&intel_pstate_driver_lock);
1655 		return -EAGAIN;
1656 	}
1657 
1658 	mutex_lock(&intel_pstate_limits_lock);
1659 
1660 	global.min_perf_pct = clamp_t(int, input,
1661 				      min_perf_pct_min(), global.max_perf_pct);
1662 
1663 	mutex_unlock(&intel_pstate_limits_lock);
1664 
1665 	if (intel_pstate_driver == &intel_pstate)
1666 		intel_pstate_update_policies();
1667 	else
1668 		update_qos_request(FREQ_QOS_MIN);
1669 
1670 	mutex_unlock(&intel_pstate_driver_lock);
1671 
1672 	return count;
1673 }
1674 
1675 static ssize_t show_hwp_dynamic_boost(struct kobject *kobj,
1676 				struct kobj_attribute *attr, char *buf)
1677 {
1678 	return sprintf(buf, "%u\n", hwp_boost);
1679 }
1680 
1681 static ssize_t store_hwp_dynamic_boost(struct kobject *a,
1682 				       struct kobj_attribute *b,
1683 				       const char *buf, size_t count)
1684 {
1685 	unsigned int input;
1686 	int ret;
1687 
1688 	ret = kstrtouint(buf, 10, &input);
1689 	if (ret)
1690 		return ret;
1691 
1692 	mutex_lock(&intel_pstate_driver_lock);
1693 	hwp_boost = !!input;
1694 	intel_pstate_update_policies();
1695 	mutex_unlock(&intel_pstate_driver_lock);
1696 
1697 	return count;
1698 }
1699 
1700 static ssize_t show_energy_efficiency(struct kobject *kobj, struct kobj_attribute *attr,
1701 				      char *buf)
1702 {
1703 	u64 power_ctl;
1704 	int enable;
1705 
1706 	rdmsrl(MSR_IA32_POWER_CTL, power_ctl);
1707 	enable = !!(power_ctl & BIT(MSR_IA32_POWER_CTL_BIT_EE));
1708 	return sprintf(buf, "%d\n", !enable);
1709 }
1710 
1711 static ssize_t store_energy_efficiency(struct kobject *a, struct kobj_attribute *b,
1712 				       const char *buf, size_t count)
1713 {
1714 	bool input;
1715 	int ret;
1716 
1717 	ret = kstrtobool(buf, &input);
1718 	if (ret)
1719 		return ret;
1720 
1721 	set_power_ctl_ee_state(input);
1722 
1723 	return count;
1724 }
1725 
1726 show_one(max_perf_pct, max_perf_pct);
1727 show_one(min_perf_pct, min_perf_pct);
1728 
1729 define_one_global_rw(status);
1730 define_one_global_rw(no_turbo);
1731 define_one_global_rw(max_perf_pct);
1732 define_one_global_rw(min_perf_pct);
1733 define_one_global_ro(turbo_pct);
1734 define_one_global_ro(num_pstates);
1735 define_one_global_rw(hwp_dynamic_boost);
1736 define_one_global_rw(energy_efficiency);
1737 
1738 static struct attribute *intel_pstate_attributes[] = {
1739 	&status.attr,
1740 	&no_turbo.attr,
1741 	NULL
1742 };
1743 
1744 static const struct attribute_group intel_pstate_attr_group = {
1745 	.attrs = intel_pstate_attributes,
1746 };
1747 
1748 static const struct x86_cpu_id intel_pstate_cpu_ee_disable_ids[];
1749 
1750 static struct kobject *intel_pstate_kobject;
1751 
1752 static void __init intel_pstate_sysfs_expose_params(void)
1753 {
1754 	struct device *dev_root = bus_get_dev_root(&cpu_subsys);
1755 	int rc;
1756 
1757 	if (dev_root) {
1758 		intel_pstate_kobject = kobject_create_and_add("intel_pstate", &dev_root->kobj);
1759 		put_device(dev_root);
1760 	}
1761 	if (WARN_ON(!intel_pstate_kobject))
1762 		return;
1763 
1764 	rc = sysfs_create_group(intel_pstate_kobject, &intel_pstate_attr_group);
1765 	if (WARN_ON(rc))
1766 		return;
1767 
1768 	if (!boot_cpu_has(X86_FEATURE_HYBRID_CPU)) {
1769 		rc = sysfs_create_file(intel_pstate_kobject, &turbo_pct.attr);
1770 		WARN_ON(rc);
1771 
1772 		rc = sysfs_create_file(intel_pstate_kobject, &num_pstates.attr);
1773 		WARN_ON(rc);
1774 	}
1775 
1776 	/*
1777 	 * If per cpu limits are enforced there are no global limits, so
1778 	 * return without creating max/min_perf_pct attributes
1779 	 */
1780 	if (per_cpu_limits)
1781 		return;
1782 
1783 	rc = sysfs_create_file(intel_pstate_kobject, &max_perf_pct.attr);
1784 	WARN_ON(rc);
1785 
1786 	rc = sysfs_create_file(intel_pstate_kobject, &min_perf_pct.attr);
1787 	WARN_ON(rc);
1788 
1789 	if (x86_match_cpu(intel_pstate_cpu_ee_disable_ids)) {
1790 		rc = sysfs_create_file(intel_pstate_kobject, &energy_efficiency.attr);
1791 		WARN_ON(rc);
1792 	}
1793 }
1794 
1795 static void __init intel_pstate_sysfs_remove(void)
1796 {
1797 	if (!intel_pstate_kobject)
1798 		return;
1799 
1800 	sysfs_remove_group(intel_pstate_kobject, &intel_pstate_attr_group);
1801 
1802 	if (!boot_cpu_has(X86_FEATURE_HYBRID_CPU)) {
1803 		sysfs_remove_file(intel_pstate_kobject, &num_pstates.attr);
1804 		sysfs_remove_file(intel_pstate_kobject, &turbo_pct.attr);
1805 	}
1806 
1807 	if (!per_cpu_limits) {
1808 		sysfs_remove_file(intel_pstate_kobject, &max_perf_pct.attr);
1809 		sysfs_remove_file(intel_pstate_kobject, &min_perf_pct.attr);
1810 
1811 		if (x86_match_cpu(intel_pstate_cpu_ee_disable_ids))
1812 			sysfs_remove_file(intel_pstate_kobject, &energy_efficiency.attr);
1813 	}
1814 
1815 	kobject_put(intel_pstate_kobject);
1816 }
1817 
1818 static void intel_pstate_sysfs_expose_hwp_dynamic_boost(void)
1819 {
1820 	int rc;
1821 
1822 	if (!hwp_active)
1823 		return;
1824 
1825 	rc = sysfs_create_file(intel_pstate_kobject, &hwp_dynamic_boost.attr);
1826 	WARN_ON_ONCE(rc);
1827 }
1828 
1829 static void intel_pstate_sysfs_hide_hwp_dynamic_boost(void)
1830 {
1831 	if (!hwp_active)
1832 		return;
1833 
1834 	sysfs_remove_file(intel_pstate_kobject, &hwp_dynamic_boost.attr);
1835 }
1836 
1837 /************************** sysfs end ************************/
1838 
1839 static void intel_pstate_notify_work(struct work_struct *work)
1840 {
1841 	struct cpudata *cpudata =
1842 		container_of(to_delayed_work(work), struct cpudata, hwp_notify_work);
1843 	struct cpufreq_policy *policy = cpufreq_cpu_acquire(cpudata->cpu);
1844 
1845 	if (policy) {
1846 		__intel_pstate_update_max_freq(cpudata, policy);
1847 
1848 		cpufreq_cpu_release(policy);
1849 
1850 		/*
1851 		 * The driver will not be unregistered while this function is
1852 		 * running, so update the capacity without acquiring the driver
1853 		 * lock.
1854 		 */
1855 		hybrid_update_capacity(cpudata);
1856 	}
1857 
1858 	wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_STATUS, 0);
1859 }
1860 
1861 static DEFINE_RAW_SPINLOCK(hwp_notify_lock);
1862 static cpumask_t hwp_intr_enable_mask;
1863 
1864 #define HWP_GUARANTEED_PERF_CHANGE_STATUS      BIT(0)
1865 #define HWP_HIGHEST_PERF_CHANGE_STATUS         BIT(3)
1866 
1867 void notify_hwp_interrupt(void)
1868 {
1869 	unsigned int this_cpu = smp_processor_id();
1870 	u64 value, status_mask;
1871 	unsigned long flags;
1872 
1873 	if (!hwp_active || !cpu_feature_enabled(X86_FEATURE_HWP_NOTIFY))
1874 		return;
1875 
1876 	status_mask = HWP_GUARANTEED_PERF_CHANGE_STATUS;
1877 	if (cpu_feature_enabled(X86_FEATURE_HWP_HIGHEST_PERF_CHANGE))
1878 		status_mask |= HWP_HIGHEST_PERF_CHANGE_STATUS;
1879 
1880 	rdmsrl_safe(MSR_HWP_STATUS, &value);
1881 	if (!(value & status_mask))
1882 		return;
1883 
1884 	raw_spin_lock_irqsave(&hwp_notify_lock, flags);
1885 
1886 	if (!cpumask_test_cpu(this_cpu, &hwp_intr_enable_mask))
1887 		goto ack_intr;
1888 
1889 	schedule_delayed_work(&all_cpu_data[this_cpu]->hwp_notify_work,
1890 			      msecs_to_jiffies(10));
1891 
1892 	raw_spin_unlock_irqrestore(&hwp_notify_lock, flags);
1893 
1894 	return;
1895 
1896 ack_intr:
1897 	wrmsrl_safe(MSR_HWP_STATUS, 0);
1898 	raw_spin_unlock_irqrestore(&hwp_notify_lock, flags);
1899 }
1900 
1901 static void intel_pstate_disable_hwp_interrupt(struct cpudata *cpudata)
1902 {
1903 	bool cancel_work;
1904 
1905 	if (!cpu_feature_enabled(X86_FEATURE_HWP_NOTIFY))
1906 		return;
1907 
1908 	/* wrmsrl_on_cpu has to be outside spinlock as this can result in IPC */
1909 	wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_INTERRUPT, 0x00);
1910 
1911 	raw_spin_lock_irq(&hwp_notify_lock);
1912 	cancel_work = cpumask_test_and_clear_cpu(cpudata->cpu, &hwp_intr_enable_mask);
1913 	raw_spin_unlock_irq(&hwp_notify_lock);
1914 
1915 	if (cancel_work)
1916 		cancel_delayed_work_sync(&cpudata->hwp_notify_work);
1917 }
1918 
1919 #define HWP_GUARANTEED_PERF_CHANGE_REQ BIT(0)
1920 #define HWP_HIGHEST_PERF_CHANGE_REQ    BIT(2)
1921 
1922 static void intel_pstate_enable_hwp_interrupt(struct cpudata *cpudata)
1923 {
1924 	/* Enable HWP notification interrupt for performance change */
1925 	if (boot_cpu_has(X86_FEATURE_HWP_NOTIFY)) {
1926 		u64 interrupt_mask = HWP_GUARANTEED_PERF_CHANGE_REQ;
1927 
1928 		raw_spin_lock_irq(&hwp_notify_lock);
1929 		INIT_DELAYED_WORK(&cpudata->hwp_notify_work, intel_pstate_notify_work);
1930 		cpumask_set_cpu(cpudata->cpu, &hwp_intr_enable_mask);
1931 		raw_spin_unlock_irq(&hwp_notify_lock);
1932 
1933 		if (cpu_feature_enabled(X86_FEATURE_HWP_HIGHEST_PERF_CHANGE))
1934 			interrupt_mask |= HWP_HIGHEST_PERF_CHANGE_REQ;
1935 
1936 		/* wrmsrl_on_cpu has to be outside spinlock as this can result in IPC */
1937 		wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_INTERRUPT, interrupt_mask);
1938 		wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_STATUS, 0);
1939 	}
1940 }
1941 
1942 static void intel_pstate_update_epp_defaults(struct cpudata *cpudata)
1943 {
1944 	cpudata->epp_default = intel_pstate_get_epp(cpudata, 0);
1945 
1946 	/*
1947 	 * If the EPP is set by firmware, which means that firmware enabled HWP
1948 	 * - Is equal or less than 0x80 (default balance_perf EPP)
1949 	 * - But less performance oriented than performance EPP
1950 	 *   then use this as new balance_perf EPP.
1951 	 */
1952 	if (hwp_forced && cpudata->epp_default <= HWP_EPP_BALANCE_PERFORMANCE &&
1953 	    cpudata->epp_default > HWP_EPP_PERFORMANCE) {
1954 		epp_values[EPP_INDEX_BALANCE_PERFORMANCE] = cpudata->epp_default;
1955 		return;
1956 	}
1957 
1958 	/*
1959 	 * If this CPU gen doesn't call for change in balance_perf
1960 	 * EPP return.
1961 	 */
1962 	if (epp_values[EPP_INDEX_BALANCE_PERFORMANCE] == HWP_EPP_BALANCE_PERFORMANCE)
1963 		return;
1964 
1965 	/*
1966 	 * Use hard coded value per gen to update the balance_perf
1967 	 * and default EPP.
1968 	 */
1969 	cpudata->epp_default = epp_values[EPP_INDEX_BALANCE_PERFORMANCE];
1970 	intel_pstate_set_epp(cpudata, cpudata->epp_default);
1971 }
1972 
1973 static void intel_pstate_hwp_enable(struct cpudata *cpudata)
1974 {
1975 	/* First disable HWP notification interrupt till we activate again */
1976 	if (boot_cpu_has(X86_FEATURE_HWP_NOTIFY))
1977 		wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_INTERRUPT, 0x00);
1978 
1979 	wrmsrl_on_cpu(cpudata->cpu, MSR_PM_ENABLE, 0x1);
1980 
1981 	intel_pstate_enable_hwp_interrupt(cpudata);
1982 
1983 	if (cpudata->epp_default >= 0)
1984 		return;
1985 
1986 	intel_pstate_update_epp_defaults(cpudata);
1987 }
1988 
1989 static int atom_get_min_pstate(int not_used)
1990 {
1991 	u64 value;
1992 
1993 	rdmsrl(MSR_ATOM_CORE_RATIOS, value);
1994 	return (value >> 8) & 0x7F;
1995 }
1996 
1997 static int atom_get_max_pstate(int not_used)
1998 {
1999 	u64 value;
2000 
2001 	rdmsrl(MSR_ATOM_CORE_RATIOS, value);
2002 	return (value >> 16) & 0x7F;
2003 }
2004 
2005 static int atom_get_turbo_pstate(int not_used)
2006 {
2007 	u64 value;
2008 
2009 	rdmsrl(MSR_ATOM_CORE_TURBO_RATIOS, value);
2010 	return value & 0x7F;
2011 }
2012 
2013 static u64 atom_get_val(struct cpudata *cpudata, int pstate)
2014 {
2015 	u64 val;
2016 	int32_t vid_fp;
2017 	u32 vid;
2018 
2019 	val = (u64)pstate << 8;
2020 	if (READ_ONCE(global.no_turbo) && !READ_ONCE(global.turbo_disabled))
2021 		val |= (u64)1 << 32;
2022 
2023 	vid_fp = cpudata->vid.min + mul_fp(
2024 		int_tofp(pstate - cpudata->pstate.min_pstate),
2025 		cpudata->vid.ratio);
2026 
2027 	vid_fp = clamp_t(int32_t, vid_fp, cpudata->vid.min, cpudata->vid.max);
2028 	vid = ceiling_fp(vid_fp);
2029 
2030 	if (pstate > cpudata->pstate.max_pstate)
2031 		vid = cpudata->vid.turbo;
2032 
2033 	return val | vid;
2034 }
2035 
2036 static int silvermont_get_scaling(void)
2037 {
2038 	u64 value;
2039 	int i;
2040 	/* Defined in Table 35-6 from SDM (Sept 2015) */
2041 	static int silvermont_freq_table[] = {
2042 		83300, 100000, 133300, 116700, 80000};
2043 
2044 	rdmsrl(MSR_FSB_FREQ, value);
2045 	i = value & 0x7;
2046 	WARN_ON(i > 4);
2047 
2048 	return silvermont_freq_table[i];
2049 }
2050 
2051 static int airmont_get_scaling(void)
2052 {
2053 	u64 value;
2054 	int i;
2055 	/* Defined in Table 35-10 from SDM (Sept 2015) */
2056 	static int airmont_freq_table[] = {
2057 		83300, 100000, 133300, 116700, 80000,
2058 		93300, 90000, 88900, 87500};
2059 
2060 	rdmsrl(MSR_FSB_FREQ, value);
2061 	i = value & 0xF;
2062 	WARN_ON(i > 8);
2063 
2064 	return airmont_freq_table[i];
2065 }
2066 
2067 static void atom_get_vid(struct cpudata *cpudata)
2068 {
2069 	u64 value;
2070 
2071 	rdmsrl(MSR_ATOM_CORE_VIDS, value);
2072 	cpudata->vid.min = int_tofp((value >> 8) & 0x7f);
2073 	cpudata->vid.max = int_tofp((value >> 16) & 0x7f);
2074 	cpudata->vid.ratio = div_fp(
2075 		cpudata->vid.max - cpudata->vid.min,
2076 		int_tofp(cpudata->pstate.max_pstate -
2077 			cpudata->pstate.min_pstate));
2078 
2079 	rdmsrl(MSR_ATOM_CORE_TURBO_VIDS, value);
2080 	cpudata->vid.turbo = value & 0x7f;
2081 }
2082 
2083 static int core_get_min_pstate(int cpu)
2084 {
2085 	u64 value;
2086 
2087 	rdmsrl_on_cpu(cpu, MSR_PLATFORM_INFO, &value);
2088 	return (value >> 40) & 0xFF;
2089 }
2090 
2091 static int core_get_max_pstate_physical(int cpu)
2092 {
2093 	u64 value;
2094 
2095 	rdmsrl_on_cpu(cpu, MSR_PLATFORM_INFO, &value);
2096 	return (value >> 8) & 0xFF;
2097 }
2098 
2099 static int core_get_tdp_ratio(int cpu, u64 plat_info)
2100 {
2101 	/* Check how many TDP levels present */
2102 	if (plat_info & 0x600000000) {
2103 		u64 tdp_ctrl;
2104 		u64 tdp_ratio;
2105 		int tdp_msr;
2106 		int err;
2107 
2108 		/* Get the TDP level (0, 1, 2) to get ratios */
2109 		err = rdmsrl_safe_on_cpu(cpu, MSR_CONFIG_TDP_CONTROL, &tdp_ctrl);
2110 		if (err)
2111 			return err;
2112 
2113 		/* TDP MSR are continuous starting at 0x648 */
2114 		tdp_msr = MSR_CONFIG_TDP_NOMINAL + (tdp_ctrl & 0x03);
2115 		err = rdmsrl_safe_on_cpu(cpu, tdp_msr, &tdp_ratio);
2116 		if (err)
2117 			return err;
2118 
2119 		/* For level 1 and 2, bits[23:16] contain the ratio */
2120 		if (tdp_ctrl & 0x03)
2121 			tdp_ratio >>= 16;
2122 
2123 		tdp_ratio &= 0xff; /* ratios are only 8 bits long */
2124 		pr_debug("tdp_ratio %x\n", (int)tdp_ratio);
2125 
2126 		return (int)tdp_ratio;
2127 	}
2128 
2129 	return -ENXIO;
2130 }
2131 
2132 static int core_get_max_pstate(int cpu)
2133 {
2134 	u64 tar;
2135 	u64 plat_info;
2136 	int max_pstate;
2137 	int tdp_ratio;
2138 	int err;
2139 
2140 	rdmsrl_on_cpu(cpu, MSR_PLATFORM_INFO, &plat_info);
2141 	max_pstate = (plat_info >> 8) & 0xFF;
2142 
2143 	tdp_ratio = core_get_tdp_ratio(cpu, plat_info);
2144 	if (tdp_ratio <= 0)
2145 		return max_pstate;
2146 
2147 	if (hwp_active) {
2148 		/* Turbo activation ratio is not used on HWP platforms */
2149 		return tdp_ratio;
2150 	}
2151 
2152 	err = rdmsrl_safe_on_cpu(cpu, MSR_TURBO_ACTIVATION_RATIO, &tar);
2153 	if (!err) {
2154 		int tar_levels;
2155 
2156 		/* Do some sanity checking for safety */
2157 		tar_levels = tar & 0xff;
2158 		if (tdp_ratio - 1 == tar_levels) {
2159 			max_pstate = tar_levels;
2160 			pr_debug("max_pstate=TAC %x\n", max_pstate);
2161 		}
2162 	}
2163 
2164 	return max_pstate;
2165 }
2166 
2167 static int core_get_turbo_pstate(int cpu)
2168 {
2169 	u64 value;
2170 	int nont, ret;
2171 
2172 	rdmsrl_on_cpu(cpu, MSR_TURBO_RATIO_LIMIT, &value);
2173 	nont = core_get_max_pstate(cpu);
2174 	ret = (value) & 255;
2175 	if (ret <= nont)
2176 		ret = nont;
2177 	return ret;
2178 }
2179 
2180 static u64 core_get_val(struct cpudata *cpudata, int pstate)
2181 {
2182 	u64 val;
2183 
2184 	val = (u64)pstate << 8;
2185 	if (READ_ONCE(global.no_turbo) && !READ_ONCE(global.turbo_disabled))
2186 		val |= (u64)1 << 32;
2187 
2188 	return val;
2189 }
2190 
2191 static int knl_get_aperf_mperf_shift(void)
2192 {
2193 	return 10;
2194 }
2195 
2196 static int knl_get_turbo_pstate(int cpu)
2197 {
2198 	u64 value;
2199 	int nont, ret;
2200 
2201 	rdmsrl_on_cpu(cpu, MSR_TURBO_RATIO_LIMIT, &value);
2202 	nont = core_get_max_pstate(cpu);
2203 	ret = (((value) >> 8) & 0xFF);
2204 	if (ret <= nont)
2205 		ret = nont;
2206 	return ret;
2207 }
2208 
2209 static int hwp_get_cpu_scaling(int cpu)
2210 {
2211 	if (hybrid_scaling_factor) {
2212 		struct cpuinfo_x86 *c = &cpu_data(smp_processor_id());
2213 		u8 cpu_type = c->topo.intel_type;
2214 
2215 		/*
2216 		 * Return the hybrid scaling factor for P-cores and use the
2217 		 * default core scaling for E-cores.
2218 		 */
2219 		if (cpu_type == INTEL_CPU_TYPE_CORE)
2220 			return hybrid_scaling_factor;
2221 
2222 		if (cpu_type == INTEL_CPU_TYPE_ATOM)
2223 			return core_get_scaling();
2224 	}
2225 
2226 	/* Use core scaling on non-hybrid systems. */
2227 	if (!cpu_feature_enabled(X86_FEATURE_HYBRID_CPU))
2228 		return core_get_scaling();
2229 
2230 	/*
2231 	 * The system is hybrid, but the hybrid scaling factor is not known or
2232 	 * the CPU type is not one of the above, so use CPPC to compute the
2233 	 * scaling factor for this CPU.
2234 	 */
2235 	return intel_pstate_cppc_get_scaling(cpu);
2236 }
2237 
2238 static void intel_pstate_set_pstate(struct cpudata *cpu, int pstate)
2239 {
2240 	trace_cpu_frequency(pstate * cpu->pstate.scaling, cpu->cpu);
2241 	cpu->pstate.current_pstate = pstate;
2242 	/*
2243 	 * Generally, there is no guarantee that this code will always run on
2244 	 * the CPU being updated, so force the register update to run on the
2245 	 * right CPU.
2246 	 */
2247 	wrmsrl_on_cpu(cpu->cpu, MSR_IA32_PERF_CTL,
2248 		      pstate_funcs.get_val(cpu, pstate));
2249 }
2250 
2251 static void intel_pstate_set_min_pstate(struct cpudata *cpu)
2252 {
2253 	intel_pstate_set_pstate(cpu, cpu->pstate.min_pstate);
2254 }
2255 
2256 static void intel_pstate_get_cpu_pstates(struct cpudata *cpu)
2257 {
2258 	int perf_ctl_max_phys = pstate_funcs.get_max_physical(cpu->cpu);
2259 	int perf_ctl_scaling = pstate_funcs.get_scaling();
2260 
2261 	cpu->pstate.min_pstate = pstate_funcs.get_min(cpu->cpu);
2262 	cpu->pstate.max_pstate_physical = perf_ctl_max_phys;
2263 	cpu->pstate.perf_ctl_scaling = perf_ctl_scaling;
2264 
2265 	if (hwp_active && !hwp_mode_bdw) {
2266 		__intel_pstate_get_hwp_cap(cpu);
2267 
2268 		if (pstate_funcs.get_cpu_scaling) {
2269 			cpu->pstate.scaling = pstate_funcs.get_cpu_scaling(cpu->cpu);
2270 			if (cpu->pstate.scaling != perf_ctl_scaling) {
2271 				intel_pstate_hybrid_hwp_adjust(cpu);
2272 				hwp_is_hybrid = true;
2273 			}
2274 		} else {
2275 			cpu->pstate.scaling = perf_ctl_scaling;
2276 		}
2277 		/*
2278 		 * If the CPU is going online for the first time and it was
2279 		 * offline initially, asym capacity scaling needs to be updated.
2280 		 */
2281 		hybrid_update_capacity(cpu);
2282 	} else {
2283 		cpu->pstate.scaling = perf_ctl_scaling;
2284 		cpu->pstate.max_pstate = pstate_funcs.get_max(cpu->cpu);
2285 		cpu->pstate.turbo_pstate = pstate_funcs.get_turbo(cpu->cpu);
2286 	}
2287 
2288 	if (cpu->pstate.scaling == perf_ctl_scaling) {
2289 		cpu->pstate.min_freq = cpu->pstate.min_pstate * perf_ctl_scaling;
2290 		cpu->pstate.max_freq = cpu->pstate.max_pstate * perf_ctl_scaling;
2291 		cpu->pstate.turbo_freq = cpu->pstate.turbo_pstate * perf_ctl_scaling;
2292 	}
2293 
2294 	if (pstate_funcs.get_aperf_mperf_shift)
2295 		cpu->aperf_mperf_shift = pstate_funcs.get_aperf_mperf_shift();
2296 
2297 	if (pstate_funcs.get_vid)
2298 		pstate_funcs.get_vid(cpu);
2299 
2300 	intel_pstate_set_min_pstate(cpu);
2301 }
2302 
2303 /*
2304  * Long hold time will keep high perf limits for long time,
2305  * which negatively impacts perf/watt for some workloads,
2306  * like specpower. 3ms is based on experiements on some
2307  * workoads.
2308  */
2309 static int hwp_boost_hold_time_ns = 3 * NSEC_PER_MSEC;
2310 
2311 static inline void intel_pstate_hwp_boost_up(struct cpudata *cpu)
2312 {
2313 	u64 hwp_req = READ_ONCE(cpu->hwp_req_cached);
2314 	u64 hwp_cap = READ_ONCE(cpu->hwp_cap_cached);
2315 	u32 max_limit = (hwp_req & 0xff00) >> 8;
2316 	u32 min_limit = (hwp_req & 0xff);
2317 	u32 boost_level1;
2318 
2319 	/*
2320 	 * Cases to consider (User changes via sysfs or boot time):
2321 	 * If, P0 (Turbo max) = P1 (Guaranteed max) = min:
2322 	 *	No boost, return.
2323 	 * If, P0 (Turbo max) > P1 (Guaranteed max) = min:
2324 	 *     Should result in one level boost only for P0.
2325 	 * If, P0 (Turbo max) = P1 (Guaranteed max) > min:
2326 	 *     Should result in two level boost:
2327 	 *         (min + p1)/2 and P1.
2328 	 * If, P0 (Turbo max) > P1 (Guaranteed max) > min:
2329 	 *     Should result in three level boost:
2330 	 *        (min + p1)/2, P1 and P0.
2331 	 */
2332 
2333 	/* If max and min are equal or already at max, nothing to boost */
2334 	if (max_limit == min_limit || cpu->hwp_boost_min >= max_limit)
2335 		return;
2336 
2337 	if (!cpu->hwp_boost_min)
2338 		cpu->hwp_boost_min = min_limit;
2339 
2340 	/* level at half way mark between min and guranteed */
2341 	boost_level1 = (HWP_GUARANTEED_PERF(hwp_cap) + min_limit) >> 1;
2342 
2343 	if (cpu->hwp_boost_min < boost_level1)
2344 		cpu->hwp_boost_min = boost_level1;
2345 	else if (cpu->hwp_boost_min < HWP_GUARANTEED_PERF(hwp_cap))
2346 		cpu->hwp_boost_min = HWP_GUARANTEED_PERF(hwp_cap);
2347 	else if (cpu->hwp_boost_min == HWP_GUARANTEED_PERF(hwp_cap) &&
2348 		 max_limit != HWP_GUARANTEED_PERF(hwp_cap))
2349 		cpu->hwp_boost_min = max_limit;
2350 	else
2351 		return;
2352 
2353 	hwp_req = (hwp_req & ~GENMASK_ULL(7, 0)) | cpu->hwp_boost_min;
2354 	wrmsrl(MSR_HWP_REQUEST, hwp_req);
2355 	cpu->last_update = cpu->sample.time;
2356 }
2357 
2358 static inline void intel_pstate_hwp_boost_down(struct cpudata *cpu)
2359 {
2360 	if (cpu->hwp_boost_min) {
2361 		bool expired;
2362 
2363 		/* Check if we are idle for hold time to boost down */
2364 		expired = time_after64(cpu->sample.time, cpu->last_update +
2365 				       hwp_boost_hold_time_ns);
2366 		if (expired) {
2367 			wrmsrl(MSR_HWP_REQUEST, cpu->hwp_req_cached);
2368 			cpu->hwp_boost_min = 0;
2369 		}
2370 	}
2371 	cpu->last_update = cpu->sample.time;
2372 }
2373 
2374 static inline void intel_pstate_update_util_hwp_local(struct cpudata *cpu,
2375 						      u64 time)
2376 {
2377 	cpu->sample.time = time;
2378 
2379 	if (cpu->sched_flags & SCHED_CPUFREQ_IOWAIT) {
2380 		bool do_io = false;
2381 
2382 		cpu->sched_flags = 0;
2383 		/*
2384 		 * Set iowait_boost flag and update time. Since IO WAIT flag
2385 		 * is set all the time, we can't just conclude that there is
2386 		 * some IO bound activity is scheduled on this CPU with just
2387 		 * one occurrence. If we receive at least two in two
2388 		 * consecutive ticks, then we treat as boost candidate.
2389 		 */
2390 		if (time_before64(time, cpu->last_io_update + 2 * TICK_NSEC))
2391 			do_io = true;
2392 
2393 		cpu->last_io_update = time;
2394 
2395 		if (do_io)
2396 			intel_pstate_hwp_boost_up(cpu);
2397 
2398 	} else {
2399 		intel_pstate_hwp_boost_down(cpu);
2400 	}
2401 }
2402 
2403 static inline void intel_pstate_update_util_hwp(struct update_util_data *data,
2404 						u64 time, unsigned int flags)
2405 {
2406 	struct cpudata *cpu = container_of(data, struct cpudata, update_util);
2407 
2408 	cpu->sched_flags |= flags;
2409 
2410 	if (smp_processor_id() == cpu->cpu)
2411 		intel_pstate_update_util_hwp_local(cpu, time);
2412 }
2413 
2414 static inline void intel_pstate_calc_avg_perf(struct cpudata *cpu)
2415 {
2416 	struct sample *sample = &cpu->sample;
2417 
2418 	sample->core_avg_perf = div_ext_fp(sample->aperf, sample->mperf);
2419 }
2420 
2421 static inline bool intel_pstate_sample(struct cpudata *cpu, u64 time)
2422 {
2423 	u64 aperf, mperf;
2424 	unsigned long flags;
2425 	u64 tsc;
2426 
2427 	local_irq_save(flags);
2428 	rdmsrl(MSR_IA32_APERF, aperf);
2429 	rdmsrl(MSR_IA32_MPERF, mperf);
2430 	tsc = rdtsc();
2431 	if (cpu->prev_mperf == mperf || cpu->prev_tsc == tsc) {
2432 		local_irq_restore(flags);
2433 		return false;
2434 	}
2435 	local_irq_restore(flags);
2436 
2437 	cpu->last_sample_time = cpu->sample.time;
2438 	cpu->sample.time = time;
2439 	cpu->sample.aperf = aperf;
2440 	cpu->sample.mperf = mperf;
2441 	cpu->sample.tsc =  tsc;
2442 	cpu->sample.aperf -= cpu->prev_aperf;
2443 	cpu->sample.mperf -= cpu->prev_mperf;
2444 	cpu->sample.tsc -= cpu->prev_tsc;
2445 
2446 	cpu->prev_aperf = aperf;
2447 	cpu->prev_mperf = mperf;
2448 	cpu->prev_tsc = tsc;
2449 	/*
2450 	 * First time this function is invoked in a given cycle, all of the
2451 	 * previous sample data fields are equal to zero or stale and they must
2452 	 * be populated with meaningful numbers for things to work, so assume
2453 	 * that sample.time will always be reset before setting the utilization
2454 	 * update hook and make the caller skip the sample then.
2455 	 */
2456 	if (cpu->last_sample_time) {
2457 		intel_pstate_calc_avg_perf(cpu);
2458 		return true;
2459 	}
2460 	return false;
2461 }
2462 
2463 static inline int32_t get_avg_frequency(struct cpudata *cpu)
2464 {
2465 	return mul_ext_fp(cpu->sample.core_avg_perf, cpu_khz);
2466 }
2467 
2468 static inline int32_t get_avg_pstate(struct cpudata *cpu)
2469 {
2470 	return mul_ext_fp(cpu->pstate.max_pstate_physical,
2471 			  cpu->sample.core_avg_perf);
2472 }
2473 
2474 static inline int32_t get_target_pstate(struct cpudata *cpu)
2475 {
2476 	struct sample *sample = &cpu->sample;
2477 	int32_t busy_frac;
2478 	int target, avg_pstate;
2479 
2480 	busy_frac = div_fp(sample->mperf << cpu->aperf_mperf_shift,
2481 			   sample->tsc);
2482 
2483 	if (busy_frac < cpu->iowait_boost)
2484 		busy_frac = cpu->iowait_boost;
2485 
2486 	sample->busy_scaled = busy_frac * 100;
2487 
2488 	target = READ_ONCE(global.no_turbo) ?
2489 			cpu->pstate.max_pstate : cpu->pstate.turbo_pstate;
2490 	target += target >> 2;
2491 	target = mul_fp(target, busy_frac);
2492 	if (target < cpu->pstate.min_pstate)
2493 		target = cpu->pstate.min_pstate;
2494 
2495 	/*
2496 	 * If the average P-state during the previous cycle was higher than the
2497 	 * current target, add 50% of the difference to the target to reduce
2498 	 * possible performance oscillations and offset possible performance
2499 	 * loss related to moving the workload from one CPU to another within
2500 	 * a package/module.
2501 	 */
2502 	avg_pstate = get_avg_pstate(cpu);
2503 	if (avg_pstate > target)
2504 		target += (avg_pstate - target) >> 1;
2505 
2506 	return target;
2507 }
2508 
2509 static int intel_pstate_prepare_request(struct cpudata *cpu, int pstate)
2510 {
2511 	int min_pstate = max(cpu->pstate.min_pstate, cpu->min_perf_ratio);
2512 	int max_pstate = max(min_pstate, cpu->max_perf_ratio);
2513 
2514 	return clamp_t(int, pstate, min_pstate, max_pstate);
2515 }
2516 
2517 static void intel_pstate_update_pstate(struct cpudata *cpu, int pstate)
2518 {
2519 	if (pstate == cpu->pstate.current_pstate)
2520 		return;
2521 
2522 	cpu->pstate.current_pstate = pstate;
2523 	wrmsrl(MSR_IA32_PERF_CTL, pstate_funcs.get_val(cpu, pstate));
2524 }
2525 
2526 static void intel_pstate_adjust_pstate(struct cpudata *cpu)
2527 {
2528 	int from = cpu->pstate.current_pstate;
2529 	struct sample *sample;
2530 	int target_pstate;
2531 
2532 	target_pstate = get_target_pstate(cpu);
2533 	target_pstate = intel_pstate_prepare_request(cpu, target_pstate);
2534 	trace_cpu_frequency(target_pstate * cpu->pstate.scaling, cpu->cpu);
2535 	intel_pstate_update_pstate(cpu, target_pstate);
2536 
2537 	sample = &cpu->sample;
2538 	trace_pstate_sample(mul_ext_fp(100, sample->core_avg_perf),
2539 		fp_toint(sample->busy_scaled),
2540 		from,
2541 		cpu->pstate.current_pstate,
2542 		sample->mperf,
2543 		sample->aperf,
2544 		sample->tsc,
2545 		get_avg_frequency(cpu),
2546 		fp_toint(cpu->iowait_boost * 100));
2547 }
2548 
2549 static void intel_pstate_update_util(struct update_util_data *data, u64 time,
2550 				     unsigned int flags)
2551 {
2552 	struct cpudata *cpu = container_of(data, struct cpudata, update_util);
2553 	u64 delta_ns;
2554 
2555 	/* Don't allow remote callbacks */
2556 	if (smp_processor_id() != cpu->cpu)
2557 		return;
2558 
2559 	delta_ns = time - cpu->last_update;
2560 	if (flags & SCHED_CPUFREQ_IOWAIT) {
2561 		/* Start over if the CPU may have been idle. */
2562 		if (delta_ns > TICK_NSEC) {
2563 			cpu->iowait_boost = ONE_EIGHTH_FP;
2564 		} else if (cpu->iowait_boost >= ONE_EIGHTH_FP) {
2565 			cpu->iowait_boost <<= 1;
2566 			if (cpu->iowait_boost > int_tofp(1))
2567 				cpu->iowait_boost = int_tofp(1);
2568 		} else {
2569 			cpu->iowait_boost = ONE_EIGHTH_FP;
2570 		}
2571 	} else if (cpu->iowait_boost) {
2572 		/* Clear iowait_boost if the CPU may have been idle. */
2573 		if (delta_ns > TICK_NSEC)
2574 			cpu->iowait_boost = 0;
2575 		else
2576 			cpu->iowait_boost >>= 1;
2577 	}
2578 	cpu->last_update = time;
2579 	delta_ns = time - cpu->sample.time;
2580 	if ((s64)delta_ns < INTEL_PSTATE_SAMPLING_INTERVAL)
2581 		return;
2582 
2583 	if (intel_pstate_sample(cpu, time))
2584 		intel_pstate_adjust_pstate(cpu);
2585 }
2586 
2587 static struct pstate_funcs core_funcs = {
2588 	.get_max = core_get_max_pstate,
2589 	.get_max_physical = core_get_max_pstate_physical,
2590 	.get_min = core_get_min_pstate,
2591 	.get_turbo = core_get_turbo_pstate,
2592 	.get_scaling = core_get_scaling,
2593 	.get_val = core_get_val,
2594 };
2595 
2596 static const struct pstate_funcs silvermont_funcs = {
2597 	.get_max = atom_get_max_pstate,
2598 	.get_max_physical = atom_get_max_pstate,
2599 	.get_min = atom_get_min_pstate,
2600 	.get_turbo = atom_get_turbo_pstate,
2601 	.get_val = atom_get_val,
2602 	.get_scaling = silvermont_get_scaling,
2603 	.get_vid = atom_get_vid,
2604 };
2605 
2606 static const struct pstate_funcs airmont_funcs = {
2607 	.get_max = atom_get_max_pstate,
2608 	.get_max_physical = atom_get_max_pstate,
2609 	.get_min = atom_get_min_pstate,
2610 	.get_turbo = atom_get_turbo_pstate,
2611 	.get_val = atom_get_val,
2612 	.get_scaling = airmont_get_scaling,
2613 	.get_vid = atom_get_vid,
2614 };
2615 
2616 static const struct pstate_funcs knl_funcs = {
2617 	.get_max = core_get_max_pstate,
2618 	.get_max_physical = core_get_max_pstate_physical,
2619 	.get_min = core_get_min_pstate,
2620 	.get_turbo = knl_get_turbo_pstate,
2621 	.get_aperf_mperf_shift = knl_get_aperf_mperf_shift,
2622 	.get_scaling = core_get_scaling,
2623 	.get_val = core_get_val,
2624 };
2625 
2626 #define X86_MATCH(vfm, policy)					 \
2627 	X86_MATCH_VFM_FEATURE(vfm, X86_FEATURE_APERFMPERF, &policy)
2628 
2629 static const struct x86_cpu_id intel_pstate_cpu_ids[] = {
2630 	X86_MATCH(INTEL_SANDYBRIDGE,		core_funcs),
2631 	X86_MATCH(INTEL_SANDYBRIDGE_X,		core_funcs),
2632 	X86_MATCH(INTEL_ATOM_SILVERMONT,	silvermont_funcs),
2633 	X86_MATCH(INTEL_IVYBRIDGE,		core_funcs),
2634 	X86_MATCH(INTEL_HASWELL,		core_funcs),
2635 	X86_MATCH(INTEL_BROADWELL,		core_funcs),
2636 	X86_MATCH(INTEL_IVYBRIDGE_X,		core_funcs),
2637 	X86_MATCH(INTEL_HASWELL_X,		core_funcs),
2638 	X86_MATCH(INTEL_HASWELL_L,		core_funcs),
2639 	X86_MATCH(INTEL_HASWELL_G,		core_funcs),
2640 	X86_MATCH(INTEL_BROADWELL_G,		core_funcs),
2641 	X86_MATCH(INTEL_ATOM_AIRMONT,		airmont_funcs),
2642 	X86_MATCH(INTEL_SKYLAKE_L,		core_funcs),
2643 	X86_MATCH(INTEL_BROADWELL_X,		core_funcs),
2644 	X86_MATCH(INTEL_SKYLAKE,		core_funcs),
2645 	X86_MATCH(INTEL_BROADWELL_D,		core_funcs),
2646 	X86_MATCH(INTEL_XEON_PHI_KNL,		knl_funcs),
2647 	X86_MATCH(INTEL_XEON_PHI_KNM,		knl_funcs),
2648 	X86_MATCH(INTEL_ATOM_GOLDMONT,		core_funcs),
2649 	X86_MATCH(INTEL_ATOM_GOLDMONT_PLUS,	core_funcs),
2650 	X86_MATCH(INTEL_SKYLAKE_X,		core_funcs),
2651 	X86_MATCH(INTEL_COMETLAKE,		core_funcs),
2652 	X86_MATCH(INTEL_ICELAKE_X,		core_funcs),
2653 	X86_MATCH(INTEL_TIGERLAKE,		core_funcs),
2654 	X86_MATCH(INTEL_SAPPHIRERAPIDS_X,	core_funcs),
2655 	X86_MATCH(INTEL_EMERALDRAPIDS_X,	core_funcs),
2656 	{}
2657 };
2658 MODULE_DEVICE_TABLE(x86cpu, intel_pstate_cpu_ids);
2659 
2660 #ifdef CONFIG_ACPI
2661 static const struct x86_cpu_id intel_pstate_cpu_oob_ids[] __initconst = {
2662 	X86_MATCH(INTEL_BROADWELL_D,		core_funcs),
2663 	X86_MATCH(INTEL_BROADWELL_X,		core_funcs),
2664 	X86_MATCH(INTEL_SKYLAKE_X,		core_funcs),
2665 	X86_MATCH(INTEL_ICELAKE_X,		core_funcs),
2666 	X86_MATCH(INTEL_SAPPHIRERAPIDS_X,	core_funcs),
2667 	X86_MATCH(INTEL_EMERALDRAPIDS_X,	core_funcs),
2668 	X86_MATCH(INTEL_GRANITERAPIDS_D,	core_funcs),
2669 	X86_MATCH(INTEL_GRANITERAPIDS_X,	core_funcs),
2670 	X86_MATCH(INTEL_ATOM_CRESTMONT,		core_funcs),
2671 	X86_MATCH(INTEL_ATOM_CRESTMONT_X,	core_funcs),
2672 	{}
2673 };
2674 #endif
2675 
2676 static const struct x86_cpu_id intel_pstate_cpu_ee_disable_ids[] = {
2677 	X86_MATCH(INTEL_KABYLAKE,		core_funcs),
2678 	{}
2679 };
2680 
2681 static int intel_pstate_init_cpu(unsigned int cpunum)
2682 {
2683 	struct cpudata *cpu;
2684 
2685 	cpu = all_cpu_data[cpunum];
2686 
2687 	if (!cpu) {
2688 		cpu = kzalloc(sizeof(*cpu), GFP_KERNEL);
2689 		if (!cpu)
2690 			return -ENOMEM;
2691 
2692 		WRITE_ONCE(all_cpu_data[cpunum], cpu);
2693 
2694 		cpu->cpu = cpunum;
2695 
2696 		cpu->epp_default = -EINVAL;
2697 
2698 		if (hwp_active) {
2699 			intel_pstate_hwp_enable(cpu);
2700 
2701 			if (intel_pstate_acpi_pm_profile_server())
2702 				hwp_boost = true;
2703 		}
2704 	} else if (hwp_active) {
2705 		/*
2706 		 * Re-enable HWP in case this happens after a resume from ACPI
2707 		 * S3 if the CPU was offline during the whole system/resume
2708 		 * cycle.
2709 		 */
2710 		intel_pstate_hwp_reenable(cpu);
2711 	}
2712 
2713 	cpu->epp_powersave = -EINVAL;
2714 	cpu->epp_policy = CPUFREQ_POLICY_UNKNOWN;
2715 
2716 	intel_pstate_get_cpu_pstates(cpu);
2717 
2718 	pr_debug("controlling: cpu %d\n", cpunum);
2719 
2720 	return 0;
2721 }
2722 
2723 static void intel_pstate_set_update_util_hook(unsigned int cpu_num)
2724 {
2725 	struct cpudata *cpu = all_cpu_data[cpu_num];
2726 
2727 	if (hwp_active && !hwp_boost)
2728 		return;
2729 
2730 	if (cpu->update_util_set)
2731 		return;
2732 
2733 	/* Prevent intel_pstate_update_util() from using stale data. */
2734 	cpu->sample.time = 0;
2735 	cpufreq_add_update_util_hook(cpu_num, &cpu->update_util,
2736 				     (hwp_active ?
2737 				      intel_pstate_update_util_hwp :
2738 				      intel_pstate_update_util));
2739 	cpu->update_util_set = true;
2740 }
2741 
2742 static void intel_pstate_clear_update_util_hook(unsigned int cpu)
2743 {
2744 	struct cpudata *cpu_data = all_cpu_data[cpu];
2745 
2746 	if (!cpu_data->update_util_set)
2747 		return;
2748 
2749 	cpufreq_remove_update_util_hook(cpu);
2750 	cpu_data->update_util_set = false;
2751 	synchronize_rcu();
2752 }
2753 
2754 static int intel_pstate_get_max_freq(struct cpudata *cpu)
2755 {
2756 	return READ_ONCE(global.no_turbo) ?
2757 			cpu->pstate.max_freq : cpu->pstate.turbo_freq;
2758 }
2759 
2760 static void intel_pstate_update_perf_limits(struct cpudata *cpu,
2761 					    unsigned int policy_min,
2762 					    unsigned int policy_max)
2763 {
2764 	int perf_ctl_scaling = cpu->pstate.perf_ctl_scaling;
2765 	int32_t max_policy_perf, min_policy_perf;
2766 
2767 	max_policy_perf = policy_max / perf_ctl_scaling;
2768 	if (policy_max == policy_min) {
2769 		min_policy_perf = max_policy_perf;
2770 	} else {
2771 		min_policy_perf = policy_min / perf_ctl_scaling;
2772 		min_policy_perf = clamp_t(int32_t, min_policy_perf,
2773 					  0, max_policy_perf);
2774 	}
2775 
2776 	/*
2777 	 * HWP needs some special consideration, because HWP_REQUEST uses
2778 	 * abstract values to represent performance rather than pure ratios.
2779 	 */
2780 	if (hwp_active && cpu->pstate.scaling != perf_ctl_scaling) {
2781 		int freq;
2782 
2783 		freq = max_policy_perf * perf_ctl_scaling;
2784 		max_policy_perf = intel_pstate_freq_to_hwp(cpu, freq);
2785 		freq = min_policy_perf * perf_ctl_scaling;
2786 		min_policy_perf = intel_pstate_freq_to_hwp(cpu, freq);
2787 	}
2788 
2789 	pr_debug("cpu:%d min_policy_perf:%d max_policy_perf:%d\n",
2790 		 cpu->cpu, min_policy_perf, max_policy_perf);
2791 
2792 	/* Normalize user input to [min_perf, max_perf] */
2793 	if (per_cpu_limits) {
2794 		cpu->min_perf_ratio = min_policy_perf;
2795 		cpu->max_perf_ratio = max_policy_perf;
2796 	} else {
2797 		int turbo_max = cpu->pstate.turbo_pstate;
2798 		int32_t global_min, global_max;
2799 
2800 		/* Global limits are in percent of the maximum turbo P-state. */
2801 		global_max = DIV_ROUND_UP(turbo_max * global.max_perf_pct, 100);
2802 		global_min = DIV_ROUND_UP(turbo_max * global.min_perf_pct, 100);
2803 		global_min = clamp_t(int32_t, global_min, 0, global_max);
2804 
2805 		pr_debug("cpu:%d global_min:%d global_max:%d\n", cpu->cpu,
2806 			 global_min, global_max);
2807 
2808 		cpu->min_perf_ratio = max(min_policy_perf, global_min);
2809 		cpu->min_perf_ratio = min(cpu->min_perf_ratio, max_policy_perf);
2810 		cpu->max_perf_ratio = min(max_policy_perf, global_max);
2811 		cpu->max_perf_ratio = max(min_policy_perf, cpu->max_perf_ratio);
2812 
2813 		/* Make sure min_perf <= max_perf */
2814 		cpu->min_perf_ratio = min(cpu->min_perf_ratio,
2815 					  cpu->max_perf_ratio);
2816 
2817 	}
2818 	pr_debug("cpu:%d max_perf_ratio:%d min_perf_ratio:%d\n", cpu->cpu,
2819 		 cpu->max_perf_ratio,
2820 		 cpu->min_perf_ratio);
2821 }
2822 
2823 static int intel_pstate_set_policy(struct cpufreq_policy *policy)
2824 {
2825 	struct cpudata *cpu;
2826 
2827 	if (!policy->cpuinfo.max_freq)
2828 		return -ENODEV;
2829 
2830 	pr_debug("set_policy cpuinfo.max %u policy->max %u\n",
2831 		 policy->cpuinfo.max_freq, policy->max);
2832 
2833 	cpu = all_cpu_data[policy->cpu];
2834 	cpu->policy = policy->policy;
2835 
2836 	mutex_lock(&intel_pstate_limits_lock);
2837 
2838 	intel_pstate_update_perf_limits(cpu, policy->min, policy->max);
2839 
2840 	if (cpu->policy == CPUFREQ_POLICY_PERFORMANCE) {
2841 		int pstate = max(cpu->pstate.min_pstate, cpu->max_perf_ratio);
2842 
2843 		/*
2844 		 * NOHZ_FULL CPUs need this as the governor callback may not
2845 		 * be invoked on them.
2846 		 */
2847 		intel_pstate_clear_update_util_hook(policy->cpu);
2848 		intel_pstate_set_pstate(cpu, pstate);
2849 	} else {
2850 		intel_pstate_set_update_util_hook(policy->cpu);
2851 	}
2852 
2853 	if (hwp_active) {
2854 		/*
2855 		 * When hwp_boost was active before and dynamically it
2856 		 * was turned off, in that case we need to clear the
2857 		 * update util hook.
2858 		 */
2859 		if (!hwp_boost)
2860 			intel_pstate_clear_update_util_hook(policy->cpu);
2861 		intel_pstate_hwp_set(policy->cpu);
2862 	}
2863 	/*
2864 	 * policy->cur is never updated with the intel_pstate driver, but it
2865 	 * is used as a stale frequency value. So, keep it within limits.
2866 	 */
2867 	policy->cur = policy->min;
2868 
2869 	mutex_unlock(&intel_pstate_limits_lock);
2870 
2871 	return 0;
2872 }
2873 
2874 static void intel_pstate_adjust_policy_max(struct cpudata *cpu,
2875 					   struct cpufreq_policy_data *policy)
2876 {
2877 	if (!hwp_active &&
2878 	    cpu->pstate.max_pstate_physical > cpu->pstate.max_pstate &&
2879 	    policy->max < policy->cpuinfo.max_freq &&
2880 	    policy->max > cpu->pstate.max_freq) {
2881 		pr_debug("policy->max > max non turbo frequency\n");
2882 		policy->max = policy->cpuinfo.max_freq;
2883 	}
2884 }
2885 
2886 static void intel_pstate_verify_cpu_policy(struct cpudata *cpu,
2887 					   struct cpufreq_policy_data *policy)
2888 {
2889 	int max_freq;
2890 
2891 	if (hwp_active) {
2892 		intel_pstate_get_hwp_cap(cpu);
2893 		max_freq = READ_ONCE(global.no_turbo) ?
2894 				cpu->pstate.max_freq : cpu->pstate.turbo_freq;
2895 	} else {
2896 		max_freq = intel_pstate_get_max_freq(cpu);
2897 	}
2898 	cpufreq_verify_within_limits(policy, policy->cpuinfo.min_freq, max_freq);
2899 
2900 	intel_pstate_adjust_policy_max(cpu, policy);
2901 }
2902 
2903 static int intel_pstate_verify_policy(struct cpufreq_policy_data *policy)
2904 {
2905 	intel_pstate_verify_cpu_policy(all_cpu_data[policy->cpu], policy);
2906 
2907 	return 0;
2908 }
2909 
2910 static int intel_cpufreq_cpu_offline(struct cpufreq_policy *policy)
2911 {
2912 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2913 
2914 	pr_debug("CPU %d going offline\n", cpu->cpu);
2915 
2916 	if (cpu->suspended)
2917 		return 0;
2918 
2919 	/*
2920 	 * If the CPU is an SMT thread and it goes offline with the performance
2921 	 * settings different from the minimum, it will prevent its sibling
2922 	 * from getting to lower performance levels, so force the minimum
2923 	 * performance on CPU offline to prevent that from happening.
2924 	 */
2925 	if (hwp_active)
2926 		intel_pstate_hwp_offline(cpu);
2927 	else
2928 		intel_pstate_set_min_pstate(cpu);
2929 
2930 	intel_pstate_exit_perf_limits(policy);
2931 
2932 	return 0;
2933 }
2934 
2935 static int intel_pstate_cpu_online(struct cpufreq_policy *policy)
2936 {
2937 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2938 
2939 	pr_debug("CPU %d going online\n", cpu->cpu);
2940 
2941 	intel_pstate_init_acpi_perf_limits(policy);
2942 
2943 	if (hwp_active) {
2944 		/*
2945 		 * Re-enable HWP and clear the "suspended" flag to let "resume"
2946 		 * know that it need not do that.
2947 		 */
2948 		intel_pstate_hwp_reenable(cpu);
2949 		cpu->suspended = false;
2950 
2951 		hybrid_update_capacity(cpu);
2952 	}
2953 
2954 	return 0;
2955 }
2956 
2957 static int intel_pstate_cpu_offline(struct cpufreq_policy *policy)
2958 {
2959 	intel_pstate_clear_update_util_hook(policy->cpu);
2960 
2961 	return intel_cpufreq_cpu_offline(policy);
2962 }
2963 
2964 static void intel_pstate_cpu_exit(struct cpufreq_policy *policy)
2965 {
2966 	pr_debug("CPU %d exiting\n", policy->cpu);
2967 
2968 	policy->fast_switch_possible = false;
2969 }
2970 
2971 static int __intel_pstate_cpu_init(struct cpufreq_policy *policy)
2972 {
2973 	struct cpudata *cpu;
2974 	int rc;
2975 
2976 	rc = intel_pstate_init_cpu(policy->cpu);
2977 	if (rc)
2978 		return rc;
2979 
2980 	cpu = all_cpu_data[policy->cpu];
2981 
2982 	cpu->max_perf_ratio = 0xFF;
2983 	cpu->min_perf_ratio = 0;
2984 
2985 	/* cpuinfo and default policy values */
2986 	policy->cpuinfo.min_freq = cpu->pstate.min_freq;
2987 	policy->cpuinfo.max_freq = READ_ONCE(global.no_turbo) ?
2988 			cpu->pstate.max_freq : cpu->pstate.turbo_freq;
2989 
2990 	policy->min = policy->cpuinfo.min_freq;
2991 	policy->max = policy->cpuinfo.max_freq;
2992 
2993 	intel_pstate_init_acpi_perf_limits(policy);
2994 
2995 	policy->fast_switch_possible = true;
2996 
2997 	return 0;
2998 }
2999 
3000 static int intel_pstate_cpu_init(struct cpufreq_policy *policy)
3001 {
3002 	int ret = __intel_pstate_cpu_init(policy);
3003 
3004 	if (ret)
3005 		return ret;
3006 
3007 	/*
3008 	 * Set the policy to powersave to provide a valid fallback value in case
3009 	 * the default cpufreq governor is neither powersave nor performance.
3010 	 */
3011 	policy->policy = CPUFREQ_POLICY_POWERSAVE;
3012 
3013 	if (hwp_active) {
3014 		struct cpudata *cpu = all_cpu_data[policy->cpu];
3015 
3016 		cpu->epp_cached = intel_pstate_get_epp(cpu, 0);
3017 	}
3018 
3019 	return 0;
3020 }
3021 
3022 static struct cpufreq_driver intel_pstate = {
3023 	.flags		= CPUFREQ_CONST_LOOPS,
3024 	.verify		= intel_pstate_verify_policy,
3025 	.setpolicy	= intel_pstate_set_policy,
3026 	.suspend	= intel_pstate_suspend,
3027 	.resume		= intel_pstate_resume,
3028 	.init		= intel_pstate_cpu_init,
3029 	.exit		= intel_pstate_cpu_exit,
3030 	.offline	= intel_pstate_cpu_offline,
3031 	.online		= intel_pstate_cpu_online,
3032 	.update_limits	= intel_pstate_update_limits,
3033 	.name		= "intel_pstate",
3034 };
3035 
3036 static int intel_cpufreq_verify_policy(struct cpufreq_policy_data *policy)
3037 {
3038 	struct cpudata *cpu = all_cpu_data[policy->cpu];
3039 
3040 	intel_pstate_verify_cpu_policy(cpu, policy);
3041 	intel_pstate_update_perf_limits(cpu, policy->min, policy->max);
3042 
3043 	return 0;
3044 }
3045 
3046 /* Use of trace in passive mode:
3047  *
3048  * In passive mode the trace core_busy field (also known as the
3049  * performance field, and lablelled as such on the graphs; also known as
3050  * core_avg_perf) is not needed and so is re-assigned to indicate if the
3051  * driver call was via the normal or fast switch path. Various graphs
3052  * output from the intel_pstate_tracer.py utility that include core_busy
3053  * (or performance or core_avg_perf) have a fixed y-axis from 0 to 100%,
3054  * so we use 10 to indicate the normal path through the driver, and
3055  * 90 to indicate the fast switch path through the driver.
3056  * The scaled_busy field is not used, and is set to 0.
3057  */
3058 
3059 #define	INTEL_PSTATE_TRACE_TARGET 10
3060 #define	INTEL_PSTATE_TRACE_FAST_SWITCH 90
3061 
3062 static void intel_cpufreq_trace(struct cpudata *cpu, unsigned int trace_type, int old_pstate)
3063 {
3064 	struct sample *sample;
3065 
3066 	if (!trace_pstate_sample_enabled())
3067 		return;
3068 
3069 	if (!intel_pstate_sample(cpu, ktime_get()))
3070 		return;
3071 
3072 	sample = &cpu->sample;
3073 	trace_pstate_sample(trace_type,
3074 		0,
3075 		old_pstate,
3076 		cpu->pstate.current_pstate,
3077 		sample->mperf,
3078 		sample->aperf,
3079 		sample->tsc,
3080 		get_avg_frequency(cpu),
3081 		fp_toint(cpu->iowait_boost * 100));
3082 }
3083 
3084 static void intel_cpufreq_hwp_update(struct cpudata *cpu, u32 min, u32 max,
3085 				     u32 desired, bool fast_switch)
3086 {
3087 	u64 prev = READ_ONCE(cpu->hwp_req_cached), value = prev;
3088 
3089 	value &= ~HWP_MIN_PERF(~0L);
3090 	value |= HWP_MIN_PERF(min);
3091 
3092 	value &= ~HWP_MAX_PERF(~0L);
3093 	value |= HWP_MAX_PERF(max);
3094 
3095 	value &= ~HWP_DESIRED_PERF(~0L);
3096 	value |= HWP_DESIRED_PERF(desired);
3097 
3098 	if (value == prev)
3099 		return;
3100 
3101 	WRITE_ONCE(cpu->hwp_req_cached, value);
3102 	if (fast_switch)
3103 		wrmsrl(MSR_HWP_REQUEST, value);
3104 	else
3105 		wrmsrl_on_cpu(cpu->cpu, MSR_HWP_REQUEST, value);
3106 }
3107 
3108 static void intel_cpufreq_perf_ctl_update(struct cpudata *cpu,
3109 					  u32 target_pstate, bool fast_switch)
3110 {
3111 	if (fast_switch)
3112 		wrmsrl(MSR_IA32_PERF_CTL,
3113 		       pstate_funcs.get_val(cpu, target_pstate));
3114 	else
3115 		wrmsrl_on_cpu(cpu->cpu, MSR_IA32_PERF_CTL,
3116 			      pstate_funcs.get_val(cpu, target_pstate));
3117 }
3118 
3119 static int intel_cpufreq_update_pstate(struct cpufreq_policy *policy,
3120 				       int target_pstate, bool fast_switch)
3121 {
3122 	struct cpudata *cpu = all_cpu_data[policy->cpu];
3123 	int old_pstate = cpu->pstate.current_pstate;
3124 
3125 	target_pstate = intel_pstate_prepare_request(cpu, target_pstate);
3126 	if (hwp_active) {
3127 		int max_pstate = policy->strict_target ?
3128 					target_pstate : cpu->max_perf_ratio;
3129 
3130 		intel_cpufreq_hwp_update(cpu, target_pstate, max_pstate, 0,
3131 					 fast_switch);
3132 	} else if (target_pstate != old_pstate) {
3133 		intel_cpufreq_perf_ctl_update(cpu, target_pstate, fast_switch);
3134 	}
3135 
3136 	cpu->pstate.current_pstate = target_pstate;
3137 
3138 	intel_cpufreq_trace(cpu, fast_switch ? INTEL_PSTATE_TRACE_FAST_SWITCH :
3139 			    INTEL_PSTATE_TRACE_TARGET, old_pstate);
3140 
3141 	return target_pstate;
3142 }
3143 
3144 static int intel_cpufreq_target(struct cpufreq_policy *policy,
3145 				unsigned int target_freq,
3146 				unsigned int relation)
3147 {
3148 	struct cpudata *cpu = all_cpu_data[policy->cpu];
3149 	struct cpufreq_freqs freqs;
3150 	int target_pstate;
3151 
3152 	freqs.old = policy->cur;
3153 	freqs.new = target_freq;
3154 
3155 	cpufreq_freq_transition_begin(policy, &freqs);
3156 
3157 	target_pstate = intel_pstate_freq_to_hwp_rel(cpu, freqs.new, relation);
3158 	target_pstate = intel_cpufreq_update_pstate(policy, target_pstate, false);
3159 
3160 	freqs.new = target_pstate * cpu->pstate.scaling;
3161 
3162 	cpufreq_freq_transition_end(policy, &freqs, false);
3163 
3164 	return 0;
3165 }
3166 
3167 static unsigned int intel_cpufreq_fast_switch(struct cpufreq_policy *policy,
3168 					      unsigned int target_freq)
3169 {
3170 	struct cpudata *cpu = all_cpu_data[policy->cpu];
3171 	int target_pstate;
3172 
3173 	target_pstate = intel_pstate_freq_to_hwp(cpu, target_freq);
3174 
3175 	target_pstate = intel_cpufreq_update_pstate(policy, target_pstate, true);
3176 
3177 	return target_pstate * cpu->pstate.scaling;
3178 }
3179 
3180 static void intel_cpufreq_adjust_perf(unsigned int cpunum,
3181 				      unsigned long min_perf,
3182 				      unsigned long target_perf,
3183 				      unsigned long capacity)
3184 {
3185 	struct cpudata *cpu = all_cpu_data[cpunum];
3186 	u64 hwp_cap = READ_ONCE(cpu->hwp_cap_cached);
3187 	int old_pstate = cpu->pstate.current_pstate;
3188 	int cap_pstate, min_pstate, max_pstate, target_pstate;
3189 
3190 	cap_pstate = READ_ONCE(global.no_turbo) ?
3191 					HWP_GUARANTEED_PERF(hwp_cap) :
3192 					HWP_HIGHEST_PERF(hwp_cap);
3193 
3194 	/* Optimization: Avoid unnecessary divisions. */
3195 
3196 	target_pstate = cap_pstate;
3197 	if (target_perf < capacity)
3198 		target_pstate = DIV_ROUND_UP(cap_pstate * target_perf, capacity);
3199 
3200 	min_pstate = cap_pstate;
3201 	if (min_perf < capacity)
3202 		min_pstate = DIV_ROUND_UP(cap_pstate * min_perf, capacity);
3203 
3204 	if (min_pstate < cpu->pstate.min_pstate)
3205 		min_pstate = cpu->pstate.min_pstate;
3206 
3207 	if (min_pstate < cpu->min_perf_ratio)
3208 		min_pstate = cpu->min_perf_ratio;
3209 
3210 	if (min_pstate > cpu->max_perf_ratio)
3211 		min_pstate = cpu->max_perf_ratio;
3212 
3213 	max_pstate = min(cap_pstate, cpu->max_perf_ratio);
3214 	if (max_pstate < min_pstate)
3215 		max_pstate = min_pstate;
3216 
3217 	target_pstate = clamp_t(int, target_pstate, min_pstate, max_pstate);
3218 
3219 	intel_cpufreq_hwp_update(cpu, min_pstate, max_pstate, target_pstate, true);
3220 
3221 	cpu->pstate.current_pstate = target_pstate;
3222 	intel_cpufreq_trace(cpu, INTEL_PSTATE_TRACE_FAST_SWITCH, old_pstate);
3223 }
3224 
3225 static int intel_cpufreq_cpu_init(struct cpufreq_policy *policy)
3226 {
3227 	struct freq_qos_request *req;
3228 	struct cpudata *cpu;
3229 	struct device *dev;
3230 	int ret, freq;
3231 
3232 	dev = get_cpu_device(policy->cpu);
3233 	if (!dev)
3234 		return -ENODEV;
3235 
3236 	ret = __intel_pstate_cpu_init(policy);
3237 	if (ret)
3238 		return ret;
3239 
3240 	policy->cpuinfo.transition_latency = INTEL_CPUFREQ_TRANSITION_LATENCY;
3241 	/* This reflects the intel_pstate_get_cpu_pstates() setting. */
3242 	policy->cur = policy->cpuinfo.min_freq;
3243 
3244 	req = kcalloc(2, sizeof(*req), GFP_KERNEL);
3245 	if (!req) {
3246 		ret = -ENOMEM;
3247 		goto pstate_exit;
3248 	}
3249 
3250 	cpu = all_cpu_data[policy->cpu];
3251 
3252 	if (hwp_active) {
3253 		u64 value;
3254 
3255 		policy->transition_delay_us = INTEL_CPUFREQ_TRANSITION_DELAY_HWP;
3256 
3257 		intel_pstate_get_hwp_cap(cpu);
3258 
3259 		rdmsrl_on_cpu(cpu->cpu, MSR_HWP_REQUEST, &value);
3260 		WRITE_ONCE(cpu->hwp_req_cached, value);
3261 
3262 		cpu->epp_cached = intel_pstate_get_epp(cpu, value);
3263 	} else {
3264 		policy->transition_delay_us = INTEL_CPUFREQ_TRANSITION_DELAY;
3265 	}
3266 
3267 	freq = DIV_ROUND_UP(cpu->pstate.turbo_freq * global.min_perf_pct, 100);
3268 
3269 	ret = freq_qos_add_request(&policy->constraints, req, FREQ_QOS_MIN,
3270 				   freq);
3271 	if (ret < 0) {
3272 		dev_err(dev, "Failed to add min-freq constraint (%d)\n", ret);
3273 		goto free_req;
3274 	}
3275 
3276 	freq = DIV_ROUND_UP(cpu->pstate.turbo_freq * global.max_perf_pct, 100);
3277 
3278 	ret = freq_qos_add_request(&policy->constraints, req + 1, FREQ_QOS_MAX,
3279 				   freq);
3280 	if (ret < 0) {
3281 		dev_err(dev, "Failed to add max-freq constraint (%d)\n", ret);
3282 		goto remove_min_req;
3283 	}
3284 
3285 	policy->driver_data = req;
3286 
3287 	return 0;
3288 
3289 remove_min_req:
3290 	freq_qos_remove_request(req);
3291 free_req:
3292 	kfree(req);
3293 pstate_exit:
3294 	intel_pstate_exit_perf_limits(policy);
3295 
3296 	return ret;
3297 }
3298 
3299 static void intel_cpufreq_cpu_exit(struct cpufreq_policy *policy)
3300 {
3301 	struct freq_qos_request *req;
3302 
3303 	req = policy->driver_data;
3304 
3305 	freq_qos_remove_request(req + 1);
3306 	freq_qos_remove_request(req);
3307 	kfree(req);
3308 
3309 	intel_pstate_cpu_exit(policy);
3310 }
3311 
3312 static int intel_cpufreq_suspend(struct cpufreq_policy *policy)
3313 {
3314 	intel_pstate_suspend(policy);
3315 
3316 	if (hwp_active) {
3317 		struct cpudata *cpu = all_cpu_data[policy->cpu];
3318 		u64 value = READ_ONCE(cpu->hwp_req_cached);
3319 
3320 		/*
3321 		 * Clear the desired perf field in MSR_HWP_REQUEST in case
3322 		 * intel_cpufreq_adjust_perf() is in use and the last value
3323 		 * written by it may not be suitable.
3324 		 */
3325 		value &= ~HWP_DESIRED_PERF(~0L);
3326 		wrmsrl_on_cpu(cpu->cpu, MSR_HWP_REQUEST, value);
3327 		WRITE_ONCE(cpu->hwp_req_cached, value);
3328 	}
3329 
3330 	return 0;
3331 }
3332 
3333 static struct cpufreq_driver intel_cpufreq = {
3334 	.flags		= CPUFREQ_CONST_LOOPS,
3335 	.verify		= intel_cpufreq_verify_policy,
3336 	.target		= intel_cpufreq_target,
3337 	.fast_switch	= intel_cpufreq_fast_switch,
3338 	.init		= intel_cpufreq_cpu_init,
3339 	.exit		= intel_cpufreq_cpu_exit,
3340 	.offline	= intel_cpufreq_cpu_offline,
3341 	.online		= intel_pstate_cpu_online,
3342 	.suspend	= intel_cpufreq_suspend,
3343 	.resume		= intel_pstate_resume,
3344 	.update_limits	= intel_pstate_update_limits,
3345 	.name		= "intel_cpufreq",
3346 };
3347 
3348 static struct cpufreq_driver *default_driver;
3349 
3350 static void intel_pstate_driver_cleanup(void)
3351 {
3352 	unsigned int cpu;
3353 
3354 	cpus_read_lock();
3355 	for_each_online_cpu(cpu) {
3356 		if (all_cpu_data[cpu]) {
3357 			if (intel_pstate_driver == &intel_pstate)
3358 				intel_pstate_clear_update_util_hook(cpu);
3359 
3360 			kfree(all_cpu_data[cpu]);
3361 			WRITE_ONCE(all_cpu_data[cpu], NULL);
3362 		}
3363 	}
3364 	cpus_read_unlock();
3365 
3366 	intel_pstate_driver = NULL;
3367 }
3368 
3369 static int intel_pstate_register_driver(struct cpufreq_driver *driver)
3370 {
3371 	bool refresh_cpu_cap_scaling;
3372 	int ret;
3373 
3374 	if (driver == &intel_pstate)
3375 		intel_pstate_sysfs_expose_hwp_dynamic_boost();
3376 
3377 	memset(&global, 0, sizeof(global));
3378 	global.max_perf_pct = 100;
3379 	global.turbo_disabled = turbo_is_disabled();
3380 	global.no_turbo = global.turbo_disabled;
3381 
3382 	arch_set_max_freq_ratio(global.turbo_disabled);
3383 
3384 	refresh_cpu_cap_scaling = hybrid_clear_max_perf_cpu();
3385 
3386 	intel_pstate_driver = driver;
3387 	ret = cpufreq_register_driver(intel_pstate_driver);
3388 	if (ret) {
3389 		intel_pstate_driver_cleanup();
3390 		return ret;
3391 	}
3392 
3393 	global.min_perf_pct = min_perf_pct_min();
3394 
3395 	hybrid_init_cpu_capacity_scaling(refresh_cpu_cap_scaling);
3396 
3397 	return 0;
3398 }
3399 
3400 static ssize_t intel_pstate_show_status(char *buf)
3401 {
3402 	if (!intel_pstate_driver)
3403 		return sprintf(buf, "off\n");
3404 
3405 	return sprintf(buf, "%s\n", intel_pstate_driver == &intel_pstate ?
3406 					"active" : "passive");
3407 }
3408 
3409 static int intel_pstate_update_status(const char *buf, size_t size)
3410 {
3411 	if (size == 3 && !strncmp(buf, "off", size)) {
3412 		if (!intel_pstate_driver)
3413 			return -EINVAL;
3414 
3415 		if (hwp_active)
3416 			return -EBUSY;
3417 
3418 		cpufreq_unregister_driver(intel_pstate_driver);
3419 		intel_pstate_driver_cleanup();
3420 		return 0;
3421 	}
3422 
3423 	if (size == 6 && !strncmp(buf, "active", size)) {
3424 		if (intel_pstate_driver) {
3425 			if (intel_pstate_driver == &intel_pstate)
3426 				return 0;
3427 
3428 			cpufreq_unregister_driver(intel_pstate_driver);
3429 		}
3430 
3431 		return intel_pstate_register_driver(&intel_pstate);
3432 	}
3433 
3434 	if (size == 7 && !strncmp(buf, "passive", size)) {
3435 		if (intel_pstate_driver) {
3436 			if (intel_pstate_driver == &intel_cpufreq)
3437 				return 0;
3438 
3439 			cpufreq_unregister_driver(intel_pstate_driver);
3440 			intel_pstate_sysfs_hide_hwp_dynamic_boost();
3441 		}
3442 
3443 		return intel_pstate_register_driver(&intel_cpufreq);
3444 	}
3445 
3446 	return -EINVAL;
3447 }
3448 
3449 static int no_load __initdata;
3450 static int no_hwp __initdata;
3451 static int hwp_only __initdata;
3452 static unsigned int force_load __initdata;
3453 
3454 static int __init intel_pstate_msrs_not_valid(void)
3455 {
3456 	if (!pstate_funcs.get_max(0) ||
3457 	    !pstate_funcs.get_min(0) ||
3458 	    !pstate_funcs.get_turbo(0))
3459 		return -ENODEV;
3460 
3461 	return 0;
3462 }
3463 
3464 static void __init copy_cpu_funcs(struct pstate_funcs *funcs)
3465 {
3466 	pstate_funcs.get_max   = funcs->get_max;
3467 	pstate_funcs.get_max_physical = funcs->get_max_physical;
3468 	pstate_funcs.get_min   = funcs->get_min;
3469 	pstate_funcs.get_turbo = funcs->get_turbo;
3470 	pstate_funcs.get_scaling = funcs->get_scaling;
3471 	pstate_funcs.get_val   = funcs->get_val;
3472 	pstate_funcs.get_vid   = funcs->get_vid;
3473 	pstate_funcs.get_aperf_mperf_shift = funcs->get_aperf_mperf_shift;
3474 }
3475 
3476 #ifdef CONFIG_ACPI
3477 
3478 static bool __init intel_pstate_no_acpi_pss(void)
3479 {
3480 	int i;
3481 
3482 	for_each_possible_cpu(i) {
3483 		acpi_status status;
3484 		union acpi_object *pss;
3485 		struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
3486 		struct acpi_processor *pr = per_cpu(processors, i);
3487 
3488 		if (!pr)
3489 			continue;
3490 
3491 		status = acpi_evaluate_object(pr->handle, "_PSS", NULL, &buffer);
3492 		if (ACPI_FAILURE(status))
3493 			continue;
3494 
3495 		pss = buffer.pointer;
3496 		if (pss && pss->type == ACPI_TYPE_PACKAGE) {
3497 			kfree(pss);
3498 			return false;
3499 		}
3500 
3501 		kfree(pss);
3502 	}
3503 
3504 	pr_debug("ACPI _PSS not found\n");
3505 	return true;
3506 }
3507 
3508 static bool __init intel_pstate_no_acpi_pcch(void)
3509 {
3510 	acpi_status status;
3511 	acpi_handle handle;
3512 
3513 	status = acpi_get_handle(NULL, "\\_SB", &handle);
3514 	if (ACPI_FAILURE(status))
3515 		goto not_found;
3516 
3517 	if (acpi_has_method(handle, "PCCH"))
3518 		return false;
3519 
3520 not_found:
3521 	pr_debug("ACPI PCCH not found\n");
3522 	return true;
3523 }
3524 
3525 static bool __init intel_pstate_has_acpi_ppc(void)
3526 {
3527 	int i;
3528 
3529 	for_each_possible_cpu(i) {
3530 		struct acpi_processor *pr = per_cpu(processors, i);
3531 
3532 		if (!pr)
3533 			continue;
3534 		if (acpi_has_method(pr->handle, "_PPC"))
3535 			return true;
3536 	}
3537 	pr_debug("ACPI _PPC not found\n");
3538 	return false;
3539 }
3540 
3541 enum {
3542 	PSS,
3543 	PPC,
3544 };
3545 
3546 /* Hardware vendor-specific info that has its own power management modes */
3547 static struct acpi_platform_list plat_info[] __initdata = {
3548 	{"HP    ", "ProLiant", 0, ACPI_SIG_FADT, all_versions, NULL, PSS},
3549 	{"ORACLE", "X4-2    ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3550 	{"ORACLE", "X4-2L   ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3551 	{"ORACLE", "X4-2B   ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3552 	{"ORACLE", "X3-2    ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3553 	{"ORACLE", "X3-2L   ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3554 	{"ORACLE", "X3-2B   ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3555 	{"ORACLE", "X4470M2 ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3556 	{"ORACLE", "X4270M3 ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3557 	{"ORACLE", "X4270M2 ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3558 	{"ORACLE", "X4170M2 ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3559 	{"ORACLE", "X4170 M3", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3560 	{"ORACLE", "X4275 M3", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3561 	{"ORACLE", "X6-2    ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3562 	{"ORACLE", "Sudbury ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3563 	{ } /* End */
3564 };
3565 
3566 #define BITMASK_OOB	(BIT(8) | BIT(18))
3567 
3568 static bool __init intel_pstate_platform_pwr_mgmt_exists(void)
3569 {
3570 	const struct x86_cpu_id *id;
3571 	u64 misc_pwr;
3572 	int idx;
3573 
3574 	id = x86_match_cpu(intel_pstate_cpu_oob_ids);
3575 	if (id) {
3576 		rdmsrl(MSR_MISC_PWR_MGMT, misc_pwr);
3577 		if (misc_pwr & BITMASK_OOB) {
3578 			pr_debug("Bit 8 or 18 in the MISC_PWR_MGMT MSR set\n");
3579 			pr_debug("P states are controlled in Out of Band mode by the firmware/hardware\n");
3580 			return true;
3581 		}
3582 	}
3583 
3584 	idx = acpi_match_platform_list(plat_info);
3585 	if (idx < 0)
3586 		return false;
3587 
3588 	switch (plat_info[idx].data) {
3589 	case PSS:
3590 		if (!intel_pstate_no_acpi_pss())
3591 			return false;
3592 
3593 		return intel_pstate_no_acpi_pcch();
3594 	case PPC:
3595 		return intel_pstate_has_acpi_ppc() && !force_load;
3596 	}
3597 
3598 	return false;
3599 }
3600 
3601 static void intel_pstate_request_control_from_smm(void)
3602 {
3603 	/*
3604 	 * It may be unsafe to request P-states control from SMM if _PPC support
3605 	 * has not been enabled.
3606 	 */
3607 	if (acpi_ppc)
3608 		acpi_processor_pstate_control();
3609 }
3610 #else /* CONFIG_ACPI not enabled */
3611 static inline bool intel_pstate_platform_pwr_mgmt_exists(void) { return false; }
3612 static inline bool intel_pstate_has_acpi_ppc(void) { return false; }
3613 static inline void intel_pstate_request_control_from_smm(void) {}
3614 #endif /* CONFIG_ACPI */
3615 
3616 #define INTEL_PSTATE_HWP_BROADWELL	0x01
3617 
3618 #define X86_MATCH_HWP(vfm, hwp_mode)				\
3619 	X86_MATCH_VFM_FEATURE(vfm, X86_FEATURE_HWP, hwp_mode)
3620 
3621 static const struct x86_cpu_id hwp_support_ids[] __initconst = {
3622 	X86_MATCH_HWP(INTEL_BROADWELL_X,	INTEL_PSTATE_HWP_BROADWELL),
3623 	X86_MATCH_HWP(INTEL_BROADWELL_D,	INTEL_PSTATE_HWP_BROADWELL),
3624 	X86_MATCH_HWP(INTEL_ANY,		0),
3625 	{}
3626 };
3627 
3628 static bool intel_pstate_hwp_is_enabled(void)
3629 {
3630 	u64 value;
3631 
3632 	rdmsrl(MSR_PM_ENABLE, value);
3633 	return !!(value & 0x1);
3634 }
3635 
3636 #define POWERSAVE_MASK			GENMASK(7, 0)
3637 #define BALANCE_POWER_MASK		GENMASK(15, 8)
3638 #define BALANCE_PERFORMANCE_MASK	GENMASK(23, 16)
3639 #define PERFORMANCE_MASK		GENMASK(31, 24)
3640 
3641 #define HWP_SET_EPP_VALUES(powersave, balance_power, balance_perf, performance) \
3642 	(FIELD_PREP_CONST(POWERSAVE_MASK, powersave) |\
3643 	 FIELD_PREP_CONST(BALANCE_POWER_MASK, balance_power) |\
3644 	 FIELD_PREP_CONST(BALANCE_PERFORMANCE_MASK, balance_perf) |\
3645 	 FIELD_PREP_CONST(PERFORMANCE_MASK, performance))
3646 
3647 #define HWP_SET_DEF_BALANCE_PERF_EPP(balance_perf) \
3648 	(HWP_SET_EPP_VALUES(HWP_EPP_POWERSAVE, HWP_EPP_BALANCE_POWERSAVE,\
3649 	 balance_perf, HWP_EPP_PERFORMANCE))
3650 
3651 static const struct x86_cpu_id intel_epp_default[] = {
3652 	/*
3653 	 * Set EPP value as 102, this is the max suggested EPP
3654 	 * which can result in one core turbo frequency for
3655 	 * AlderLake Mobile CPUs.
3656 	 */
3657 	X86_MATCH_VFM(INTEL_ALDERLAKE_L, HWP_SET_DEF_BALANCE_PERF_EPP(102)),
3658 	X86_MATCH_VFM(INTEL_SAPPHIRERAPIDS_X, HWP_SET_DEF_BALANCE_PERF_EPP(32)),
3659 	X86_MATCH_VFM(INTEL_EMERALDRAPIDS_X, HWP_SET_DEF_BALANCE_PERF_EPP(32)),
3660 	X86_MATCH_VFM(INTEL_GRANITERAPIDS_X, HWP_SET_DEF_BALANCE_PERF_EPP(32)),
3661 	X86_MATCH_VFM(INTEL_GRANITERAPIDS_D, HWP_SET_DEF_BALANCE_PERF_EPP(32)),
3662 	X86_MATCH_VFM(INTEL_METEORLAKE_L, HWP_SET_EPP_VALUES(HWP_EPP_POWERSAVE,
3663 		      179, 64, 16)),
3664 	X86_MATCH_VFM(INTEL_ARROWLAKE, HWP_SET_EPP_VALUES(HWP_EPP_POWERSAVE,
3665 		      179, 64, 16)),
3666 	{}
3667 };
3668 
3669 static const struct x86_cpu_id intel_hybrid_scaling_factor[] = {
3670 	X86_MATCH_VFM(INTEL_ALDERLAKE, HYBRID_SCALING_FACTOR_ADL),
3671 	X86_MATCH_VFM(INTEL_ALDERLAKE_L, HYBRID_SCALING_FACTOR_ADL),
3672 	X86_MATCH_VFM(INTEL_RAPTORLAKE, HYBRID_SCALING_FACTOR_ADL),
3673 	X86_MATCH_VFM(INTEL_RAPTORLAKE_P, HYBRID_SCALING_FACTOR_ADL),
3674 	X86_MATCH_VFM(INTEL_RAPTORLAKE_S, HYBRID_SCALING_FACTOR_ADL),
3675 	X86_MATCH_VFM(INTEL_METEORLAKE_L, HYBRID_SCALING_FACTOR_MTL),
3676 	X86_MATCH_VFM(INTEL_LUNARLAKE_M, HYBRID_SCALING_FACTOR_LNL),
3677 	{}
3678 };
3679 
3680 static int __init intel_pstate_init(void)
3681 {
3682 	static struct cpudata **_all_cpu_data;
3683 	const struct x86_cpu_id *id;
3684 	int rc;
3685 
3686 	if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL)
3687 		return -ENODEV;
3688 
3689 	/*
3690 	 * The Intel pstate driver will be ignored if the platform
3691 	 * firmware has its own power management modes.
3692 	 */
3693 	if (intel_pstate_platform_pwr_mgmt_exists()) {
3694 		pr_info("P-states controlled by the platform\n");
3695 		return -ENODEV;
3696 	}
3697 
3698 	id = x86_match_cpu(hwp_support_ids);
3699 	if (id) {
3700 		hwp_forced = intel_pstate_hwp_is_enabled();
3701 
3702 		if (hwp_forced)
3703 			pr_info("HWP enabled by BIOS\n");
3704 		else if (no_load)
3705 			return -ENODEV;
3706 
3707 		copy_cpu_funcs(&core_funcs);
3708 		/*
3709 		 * Avoid enabling HWP for processors without EPP support,
3710 		 * because that means incomplete HWP implementation which is a
3711 		 * corner case and supporting it is generally problematic.
3712 		 *
3713 		 * If HWP is enabled already, though, there is no choice but to
3714 		 * deal with it.
3715 		 */
3716 		if ((!no_hwp && boot_cpu_has(X86_FEATURE_HWP_EPP)) || hwp_forced) {
3717 			hwp_active = true;
3718 			hwp_mode_bdw = id->driver_data;
3719 			intel_pstate.attr = hwp_cpufreq_attrs;
3720 			intel_cpufreq.attr = hwp_cpufreq_attrs;
3721 			intel_cpufreq.flags |= CPUFREQ_NEED_UPDATE_LIMITS;
3722 			intel_cpufreq.adjust_perf = intel_cpufreq_adjust_perf;
3723 			if (!default_driver)
3724 				default_driver = &intel_pstate;
3725 
3726 			pstate_funcs.get_cpu_scaling = hwp_get_cpu_scaling;
3727 
3728 			goto hwp_cpu_matched;
3729 		}
3730 		pr_info("HWP not enabled\n");
3731 	} else {
3732 		if (no_load)
3733 			return -ENODEV;
3734 
3735 		id = x86_match_cpu(intel_pstate_cpu_ids);
3736 		if (!id) {
3737 			pr_info("CPU model not supported\n");
3738 			return -ENODEV;
3739 		}
3740 
3741 		copy_cpu_funcs((struct pstate_funcs *)id->driver_data);
3742 	}
3743 
3744 	if (intel_pstate_msrs_not_valid()) {
3745 		pr_info("Invalid MSRs\n");
3746 		return -ENODEV;
3747 	}
3748 	/* Without HWP start in the passive mode. */
3749 	if (!default_driver)
3750 		default_driver = &intel_cpufreq;
3751 
3752 hwp_cpu_matched:
3753 	if (!hwp_active && hwp_only)
3754 		return -ENOTSUPP;
3755 
3756 	pr_info("Intel P-state driver initializing\n");
3757 
3758 	_all_cpu_data = vzalloc(array_size(sizeof(void *), num_possible_cpus()));
3759 	if (!_all_cpu_data)
3760 		return -ENOMEM;
3761 
3762 	WRITE_ONCE(all_cpu_data, _all_cpu_data);
3763 
3764 	intel_pstate_request_control_from_smm();
3765 
3766 	intel_pstate_sysfs_expose_params();
3767 
3768 	if (hwp_active) {
3769 		const struct x86_cpu_id *id = x86_match_cpu(intel_epp_default);
3770 		const struct x86_cpu_id *hybrid_id = x86_match_cpu(intel_hybrid_scaling_factor);
3771 
3772 		if (id) {
3773 			epp_values[EPP_INDEX_POWERSAVE] =
3774 					FIELD_GET(POWERSAVE_MASK, id->driver_data);
3775 			epp_values[EPP_INDEX_BALANCE_POWERSAVE] =
3776 					FIELD_GET(BALANCE_POWER_MASK, id->driver_data);
3777 			epp_values[EPP_INDEX_BALANCE_PERFORMANCE] =
3778 					FIELD_GET(BALANCE_PERFORMANCE_MASK, id->driver_data);
3779 			epp_values[EPP_INDEX_PERFORMANCE] =
3780 					FIELD_GET(PERFORMANCE_MASK, id->driver_data);
3781 			pr_debug("Updated EPPs powersave:%x balanced power:%x balanced perf:%x performance:%x\n",
3782 				 epp_values[EPP_INDEX_POWERSAVE],
3783 				 epp_values[EPP_INDEX_BALANCE_POWERSAVE],
3784 				 epp_values[EPP_INDEX_BALANCE_PERFORMANCE],
3785 				 epp_values[EPP_INDEX_PERFORMANCE]);
3786 		}
3787 
3788 		if (hybrid_id) {
3789 			hybrid_scaling_factor = hybrid_id->driver_data;
3790 			pr_debug("hybrid scaling factor: %d\n", hybrid_scaling_factor);
3791 		}
3792 
3793 	}
3794 
3795 	mutex_lock(&intel_pstate_driver_lock);
3796 	rc = intel_pstate_register_driver(default_driver);
3797 	mutex_unlock(&intel_pstate_driver_lock);
3798 	if (rc) {
3799 		intel_pstate_sysfs_remove();
3800 		return rc;
3801 	}
3802 
3803 	if (hwp_active) {
3804 		const struct x86_cpu_id *id;
3805 
3806 		id = x86_match_cpu(intel_pstate_cpu_ee_disable_ids);
3807 		if (id) {
3808 			set_power_ctl_ee_state(false);
3809 			pr_info("Disabling energy efficiency optimization\n");
3810 		}
3811 
3812 		pr_info("HWP enabled\n");
3813 	} else if (boot_cpu_has(X86_FEATURE_HYBRID_CPU)) {
3814 		pr_warn("Problematic setup: Hybrid processor with disabled HWP\n");
3815 	}
3816 
3817 	return 0;
3818 }
3819 device_initcall(intel_pstate_init);
3820 
3821 static int __init intel_pstate_setup(char *str)
3822 {
3823 	if (!str)
3824 		return -EINVAL;
3825 
3826 	if (!strcmp(str, "disable"))
3827 		no_load = 1;
3828 	else if (!strcmp(str, "active"))
3829 		default_driver = &intel_pstate;
3830 	else if (!strcmp(str, "passive"))
3831 		default_driver = &intel_cpufreq;
3832 
3833 	if (!strcmp(str, "no_hwp"))
3834 		no_hwp = 1;
3835 
3836 	if (!strcmp(str, "no_cas"))
3837 		no_cas = true;
3838 
3839 	if (!strcmp(str, "force"))
3840 		force_load = 1;
3841 	if (!strcmp(str, "hwp_only"))
3842 		hwp_only = 1;
3843 	if (!strcmp(str, "per_cpu_perf_limits"))
3844 		per_cpu_limits = true;
3845 
3846 #ifdef CONFIG_ACPI
3847 	if (!strcmp(str, "support_acpi_ppc"))
3848 		acpi_ppc = true;
3849 #endif
3850 
3851 	return 0;
3852 }
3853 early_param("intel_pstate", intel_pstate_setup);
3854 
3855 MODULE_AUTHOR("Dirk Brandewie <dirk.j.brandewie@intel.com>");
3856 MODULE_DESCRIPTION("'intel_pstate' - P state driver Intel Core processors");
3857