xref: /linux/drivers/cpufreq/intel_pstate.c (revision e814f3fd16acfb7f9966773953de8f740a1e3202)
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 struct cpudata *hybrid_max_perf_cpu __read_mostly;
940 /*
941  * Protects hybrid_max_perf_cpu, the capacity_perf fields in struct cpudata,
942  * and the x86 arch scale-invariance information from concurrent updates.
943  */
944 static DEFINE_MUTEX(hybrid_capacity_lock);
945 
946 static void hybrid_set_cpu_capacity(struct cpudata *cpu)
947 {
948 	arch_set_cpu_capacity(cpu->cpu, cpu->capacity_perf,
949 			      hybrid_max_perf_cpu->capacity_perf,
950 			      cpu->capacity_perf,
951 			      cpu->pstate.max_pstate_physical);
952 
953 	pr_debug("CPU%d: perf = %u, max. perf = %u, base perf = %d\n", cpu->cpu,
954 		 cpu->capacity_perf, hybrid_max_perf_cpu->capacity_perf,
955 		 cpu->pstate.max_pstate_physical);
956 }
957 
958 static void hybrid_clear_cpu_capacity(unsigned int cpunum)
959 {
960 	arch_set_cpu_capacity(cpunum, 1, 1, 1, 1);
961 }
962 
963 static void hybrid_get_capacity_perf(struct cpudata *cpu)
964 {
965 	if (READ_ONCE(global.no_turbo)) {
966 		cpu->capacity_perf = cpu->pstate.max_pstate_physical;
967 		return;
968 	}
969 
970 	cpu->capacity_perf = HWP_HIGHEST_PERF(READ_ONCE(cpu->hwp_cap_cached));
971 }
972 
973 static void hybrid_set_capacity_of_cpus(void)
974 {
975 	int cpunum;
976 
977 	for_each_online_cpu(cpunum) {
978 		struct cpudata *cpu = all_cpu_data[cpunum];
979 
980 		if (cpu)
981 			hybrid_set_cpu_capacity(cpu);
982 	}
983 }
984 
985 static void hybrid_update_cpu_capacity_scaling(void)
986 {
987 	struct cpudata *max_perf_cpu = NULL;
988 	unsigned int max_cap_perf = 0;
989 	int cpunum;
990 
991 	for_each_online_cpu(cpunum) {
992 		struct cpudata *cpu = all_cpu_data[cpunum];
993 
994 		if (!cpu)
995 			continue;
996 
997 		/*
998 		 * During initialization, CPU performance at full capacity needs
999 		 * to be determined.
1000 		 */
1001 		if (!hybrid_max_perf_cpu)
1002 			hybrid_get_capacity_perf(cpu);
1003 
1004 		/*
1005 		 * If hybrid_max_perf_cpu is not NULL at this point, it is
1006 		 * being replaced, so don't take it into account when looking
1007 		 * for the new one.
1008 		 */
1009 		if (cpu == hybrid_max_perf_cpu)
1010 			continue;
1011 
1012 		if (cpu->capacity_perf > max_cap_perf) {
1013 			max_cap_perf = cpu->capacity_perf;
1014 			max_perf_cpu = cpu;
1015 		}
1016 	}
1017 
1018 	if (max_perf_cpu) {
1019 		hybrid_max_perf_cpu = max_perf_cpu;
1020 		hybrid_set_capacity_of_cpus();
1021 	} else {
1022 		pr_info("Found no CPUs with nonzero maximum performance\n");
1023 		/* Revert to the flat CPU capacity structure. */
1024 		for_each_online_cpu(cpunum)
1025 			hybrid_clear_cpu_capacity(cpunum);
1026 	}
1027 }
1028 
1029 static void __hybrid_refresh_cpu_capacity_scaling(void)
1030 {
1031 	hybrid_max_perf_cpu = NULL;
1032 	hybrid_update_cpu_capacity_scaling();
1033 }
1034 
1035 static void hybrid_refresh_cpu_capacity_scaling(void)
1036 {
1037 	guard(mutex)(&hybrid_capacity_lock);
1038 
1039 	__hybrid_refresh_cpu_capacity_scaling();
1040 }
1041 
1042 static void hybrid_init_cpu_capacity_scaling(bool refresh)
1043 {
1044 	/*
1045 	 * If hybrid_max_perf_cpu is set at this point, the hybrid CPU capacity
1046 	 * scaling has been enabled already and the driver is just changing the
1047 	 * operation mode.
1048 	 */
1049 	if (refresh) {
1050 		hybrid_refresh_cpu_capacity_scaling();
1051 		return;
1052 	}
1053 
1054 	/*
1055 	 * On hybrid systems, use asym capacity instead of ITMT, but because
1056 	 * the capacity of SMT threads is not deterministic even approximately,
1057 	 * do not do that when SMT is in use.
1058 	 */
1059 	if (hwp_is_hybrid && !sched_smt_active() && arch_enable_hybrid_capacity_scale()) {
1060 		hybrid_refresh_cpu_capacity_scaling();
1061 		/*
1062 		 * Disabling ITMT causes sched domains to be rebuilt to disable asym
1063 		 * packing and enable asym capacity.
1064 		 */
1065 		sched_clear_itmt_support();
1066 	}
1067 }
1068 
1069 static bool hybrid_clear_max_perf_cpu(void)
1070 {
1071 	bool ret;
1072 
1073 	guard(mutex)(&hybrid_capacity_lock);
1074 
1075 	ret = !!hybrid_max_perf_cpu;
1076 	hybrid_max_perf_cpu = NULL;
1077 
1078 	return ret;
1079 }
1080 
1081 static void __intel_pstate_get_hwp_cap(struct cpudata *cpu)
1082 {
1083 	u64 cap;
1084 
1085 	rdmsrl_on_cpu(cpu->cpu, MSR_HWP_CAPABILITIES, &cap);
1086 	WRITE_ONCE(cpu->hwp_cap_cached, cap);
1087 	cpu->pstate.max_pstate = HWP_GUARANTEED_PERF(cap);
1088 	cpu->pstate.turbo_pstate = HWP_HIGHEST_PERF(cap);
1089 }
1090 
1091 static void intel_pstate_get_hwp_cap(struct cpudata *cpu)
1092 {
1093 	int scaling = cpu->pstate.scaling;
1094 
1095 	__intel_pstate_get_hwp_cap(cpu);
1096 
1097 	cpu->pstate.max_freq = cpu->pstate.max_pstate * scaling;
1098 	cpu->pstate.turbo_freq = cpu->pstate.turbo_pstate * scaling;
1099 	if (scaling != cpu->pstate.perf_ctl_scaling) {
1100 		int perf_ctl_scaling = cpu->pstate.perf_ctl_scaling;
1101 
1102 		cpu->pstate.max_freq = rounddown(cpu->pstate.max_freq,
1103 						 perf_ctl_scaling);
1104 		cpu->pstate.turbo_freq = rounddown(cpu->pstate.turbo_freq,
1105 						   perf_ctl_scaling);
1106 	}
1107 }
1108 
1109 static void hybrid_update_capacity(struct cpudata *cpu)
1110 {
1111 	unsigned int max_cap_perf;
1112 
1113 	mutex_lock(&hybrid_capacity_lock);
1114 
1115 	if (!hybrid_max_perf_cpu)
1116 		goto unlock;
1117 
1118 	/*
1119 	 * The maximum performance of the CPU may have changed, but assume
1120 	 * that the performance of the other CPUs has not changed.
1121 	 */
1122 	max_cap_perf = hybrid_max_perf_cpu->capacity_perf;
1123 
1124 	intel_pstate_get_hwp_cap(cpu);
1125 
1126 	hybrid_get_capacity_perf(cpu);
1127 	/* Should hybrid_max_perf_cpu be replaced by this CPU? */
1128 	if (cpu->capacity_perf > max_cap_perf) {
1129 		hybrid_max_perf_cpu = cpu;
1130 		hybrid_set_capacity_of_cpus();
1131 		goto unlock;
1132 	}
1133 
1134 	/* If this CPU is hybrid_max_perf_cpu, should it be replaced? */
1135 	if (cpu == hybrid_max_perf_cpu && cpu->capacity_perf < max_cap_perf) {
1136 		hybrid_update_cpu_capacity_scaling();
1137 		goto unlock;
1138 	}
1139 
1140 	hybrid_set_cpu_capacity(cpu);
1141 
1142 unlock:
1143 	mutex_unlock(&hybrid_capacity_lock);
1144 }
1145 
1146 static void intel_pstate_hwp_set(unsigned int cpu)
1147 {
1148 	struct cpudata *cpu_data = all_cpu_data[cpu];
1149 	int max, min;
1150 	u64 value;
1151 	s16 epp;
1152 
1153 	max = cpu_data->max_perf_ratio;
1154 	min = cpu_data->min_perf_ratio;
1155 
1156 	if (cpu_data->policy == CPUFREQ_POLICY_PERFORMANCE)
1157 		min = max;
1158 
1159 	rdmsrl_on_cpu(cpu, MSR_HWP_REQUEST, &value);
1160 
1161 	value &= ~HWP_MIN_PERF(~0L);
1162 	value |= HWP_MIN_PERF(min);
1163 
1164 	value &= ~HWP_MAX_PERF(~0L);
1165 	value |= HWP_MAX_PERF(max);
1166 
1167 	if (cpu_data->epp_policy == cpu_data->policy)
1168 		goto skip_epp;
1169 
1170 	cpu_data->epp_policy = cpu_data->policy;
1171 
1172 	if (cpu_data->policy == CPUFREQ_POLICY_PERFORMANCE) {
1173 		epp = intel_pstate_get_epp(cpu_data, value);
1174 		cpu_data->epp_powersave = epp;
1175 		/* If EPP read was failed, then don't try to write */
1176 		if (epp < 0)
1177 			goto skip_epp;
1178 
1179 		epp = 0;
1180 	} else {
1181 		/* skip setting EPP, when saved value is invalid */
1182 		if (cpu_data->epp_powersave < 0)
1183 			goto skip_epp;
1184 
1185 		/*
1186 		 * No need to restore EPP when it is not zero. This
1187 		 * means:
1188 		 *  - Policy is not changed
1189 		 *  - user has manually changed
1190 		 *  - Error reading EPB
1191 		 */
1192 		epp = intel_pstate_get_epp(cpu_data, value);
1193 		if (epp)
1194 			goto skip_epp;
1195 
1196 		epp = cpu_data->epp_powersave;
1197 	}
1198 	if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
1199 		value &= ~GENMASK_ULL(31, 24);
1200 		value |= (u64)epp << 24;
1201 	} else {
1202 		intel_pstate_set_epb(cpu, epp);
1203 	}
1204 skip_epp:
1205 	WRITE_ONCE(cpu_data->hwp_req_cached, value);
1206 	wrmsrl_on_cpu(cpu, MSR_HWP_REQUEST, value);
1207 }
1208 
1209 static void intel_pstate_disable_hwp_interrupt(struct cpudata *cpudata);
1210 
1211 static void intel_pstate_hwp_offline(struct cpudata *cpu)
1212 {
1213 	u64 value = READ_ONCE(cpu->hwp_req_cached);
1214 	int min_perf;
1215 
1216 	intel_pstate_disable_hwp_interrupt(cpu);
1217 
1218 	if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
1219 		/*
1220 		 * In case the EPP has been set to "performance" by the
1221 		 * active mode "performance" scaling algorithm, replace that
1222 		 * temporary value with the cached EPP one.
1223 		 */
1224 		value &= ~GENMASK_ULL(31, 24);
1225 		value |= HWP_ENERGY_PERF_PREFERENCE(cpu->epp_cached);
1226 		/*
1227 		 * However, make sure that EPP will be set to "performance" when
1228 		 * the CPU is brought back online again and the "performance"
1229 		 * scaling algorithm is still in effect.
1230 		 */
1231 		cpu->epp_policy = CPUFREQ_POLICY_UNKNOWN;
1232 	}
1233 
1234 	/*
1235 	 * Clear the desired perf field in the cached HWP request value to
1236 	 * prevent nonzero desired values from being leaked into the active
1237 	 * mode.
1238 	 */
1239 	value &= ~HWP_DESIRED_PERF(~0L);
1240 	WRITE_ONCE(cpu->hwp_req_cached, value);
1241 
1242 	value &= ~GENMASK_ULL(31, 0);
1243 	min_perf = HWP_LOWEST_PERF(READ_ONCE(cpu->hwp_cap_cached));
1244 
1245 	/* Set hwp_max = hwp_min */
1246 	value |= HWP_MAX_PERF(min_perf);
1247 	value |= HWP_MIN_PERF(min_perf);
1248 
1249 	/* Set EPP to min */
1250 	if (boot_cpu_has(X86_FEATURE_HWP_EPP))
1251 		value |= HWP_ENERGY_PERF_PREFERENCE(HWP_EPP_POWERSAVE);
1252 
1253 	wrmsrl_on_cpu(cpu->cpu, MSR_HWP_REQUEST, value);
1254 
1255 	mutex_lock(&hybrid_capacity_lock);
1256 
1257 	if (!hybrid_max_perf_cpu) {
1258 		mutex_unlock(&hybrid_capacity_lock);
1259 
1260 		return;
1261 	}
1262 
1263 	if (hybrid_max_perf_cpu == cpu)
1264 		hybrid_update_cpu_capacity_scaling();
1265 
1266 	mutex_unlock(&hybrid_capacity_lock);
1267 
1268 	/* Reset the capacity of the CPU going offline to the initial value. */
1269 	hybrid_clear_cpu_capacity(cpu->cpu);
1270 }
1271 
1272 #define POWER_CTL_EE_ENABLE	1
1273 #define POWER_CTL_EE_DISABLE	2
1274 
1275 static int power_ctl_ee_state;
1276 
1277 static void set_power_ctl_ee_state(bool input)
1278 {
1279 	u64 power_ctl;
1280 
1281 	mutex_lock(&intel_pstate_driver_lock);
1282 	rdmsrl(MSR_IA32_POWER_CTL, power_ctl);
1283 	if (input) {
1284 		power_ctl &= ~BIT(MSR_IA32_POWER_CTL_BIT_EE);
1285 		power_ctl_ee_state = POWER_CTL_EE_ENABLE;
1286 	} else {
1287 		power_ctl |= BIT(MSR_IA32_POWER_CTL_BIT_EE);
1288 		power_ctl_ee_state = POWER_CTL_EE_DISABLE;
1289 	}
1290 	wrmsrl(MSR_IA32_POWER_CTL, power_ctl);
1291 	mutex_unlock(&intel_pstate_driver_lock);
1292 }
1293 
1294 static void intel_pstate_hwp_enable(struct cpudata *cpudata);
1295 
1296 static void intel_pstate_hwp_reenable(struct cpudata *cpu)
1297 {
1298 	intel_pstate_hwp_enable(cpu);
1299 	wrmsrl_on_cpu(cpu->cpu, MSR_HWP_REQUEST, READ_ONCE(cpu->hwp_req_cached));
1300 }
1301 
1302 static int intel_pstate_suspend(struct cpufreq_policy *policy)
1303 {
1304 	struct cpudata *cpu = all_cpu_data[policy->cpu];
1305 
1306 	pr_debug("CPU %d suspending\n", cpu->cpu);
1307 
1308 	cpu->suspended = true;
1309 
1310 	/* disable HWP interrupt and cancel any pending work */
1311 	intel_pstate_disable_hwp_interrupt(cpu);
1312 
1313 	return 0;
1314 }
1315 
1316 static int intel_pstate_resume(struct cpufreq_policy *policy)
1317 {
1318 	struct cpudata *cpu = all_cpu_data[policy->cpu];
1319 
1320 	pr_debug("CPU %d resuming\n", cpu->cpu);
1321 
1322 	/* Only restore if the system default is changed */
1323 	if (power_ctl_ee_state == POWER_CTL_EE_ENABLE)
1324 		set_power_ctl_ee_state(true);
1325 	else if (power_ctl_ee_state == POWER_CTL_EE_DISABLE)
1326 		set_power_ctl_ee_state(false);
1327 
1328 	if (cpu->suspended && hwp_active) {
1329 		mutex_lock(&intel_pstate_limits_lock);
1330 
1331 		/* Re-enable HWP, because "online" has not done that. */
1332 		intel_pstate_hwp_reenable(cpu);
1333 
1334 		mutex_unlock(&intel_pstate_limits_lock);
1335 	}
1336 
1337 	cpu->suspended = false;
1338 
1339 	return 0;
1340 }
1341 
1342 static void intel_pstate_update_policies(void)
1343 {
1344 	int cpu;
1345 
1346 	for_each_possible_cpu(cpu)
1347 		cpufreq_update_policy(cpu);
1348 }
1349 
1350 static void __intel_pstate_update_max_freq(struct cpudata *cpudata,
1351 					   struct cpufreq_policy *policy)
1352 {
1353 	if (hwp_active)
1354 		intel_pstate_get_hwp_cap(cpudata);
1355 
1356 	policy->cpuinfo.max_freq = READ_ONCE(global.no_turbo) ?
1357 			cpudata->pstate.max_freq : cpudata->pstate.turbo_freq;
1358 
1359 	refresh_frequency_limits(policy);
1360 }
1361 
1362 static void intel_pstate_update_limits(unsigned int cpu)
1363 {
1364 	struct cpufreq_policy *policy = cpufreq_cpu_acquire(cpu);
1365 	struct cpudata *cpudata;
1366 
1367 	if (!policy)
1368 		return;
1369 
1370 	cpudata = all_cpu_data[cpu];
1371 
1372 	__intel_pstate_update_max_freq(cpudata, policy);
1373 
1374 	/* Prevent the driver from being unregistered now. */
1375 	mutex_lock(&intel_pstate_driver_lock);
1376 
1377 	cpufreq_cpu_release(policy);
1378 
1379 	hybrid_update_capacity(cpudata);
1380 
1381 	mutex_unlock(&intel_pstate_driver_lock);
1382 }
1383 
1384 static void intel_pstate_update_limits_for_all(void)
1385 {
1386 	int cpu;
1387 
1388 	for_each_possible_cpu(cpu) {
1389 		struct cpufreq_policy *policy = cpufreq_cpu_acquire(cpu);
1390 
1391 		if (!policy)
1392 			continue;
1393 
1394 		__intel_pstate_update_max_freq(all_cpu_data[cpu], policy);
1395 
1396 		cpufreq_cpu_release(policy);
1397 	}
1398 
1399 	mutex_lock(&hybrid_capacity_lock);
1400 
1401 	if (hybrid_max_perf_cpu)
1402 		__hybrid_refresh_cpu_capacity_scaling();
1403 
1404 	mutex_unlock(&hybrid_capacity_lock);
1405 }
1406 
1407 /************************** sysfs begin ************************/
1408 #define show_one(file_name, object)					\
1409 	static ssize_t show_##file_name					\
1410 	(struct kobject *kobj, struct kobj_attribute *attr, char *buf)	\
1411 	{								\
1412 		return sprintf(buf, "%u\n", global.object);		\
1413 	}
1414 
1415 static ssize_t intel_pstate_show_status(char *buf);
1416 static int intel_pstate_update_status(const char *buf, size_t size);
1417 
1418 static ssize_t show_status(struct kobject *kobj,
1419 			   struct kobj_attribute *attr, char *buf)
1420 {
1421 	ssize_t ret;
1422 
1423 	mutex_lock(&intel_pstate_driver_lock);
1424 	ret = intel_pstate_show_status(buf);
1425 	mutex_unlock(&intel_pstate_driver_lock);
1426 
1427 	return ret;
1428 }
1429 
1430 static ssize_t store_status(struct kobject *a, struct kobj_attribute *b,
1431 			    const char *buf, size_t count)
1432 {
1433 	char *p = memchr(buf, '\n', count);
1434 	int ret;
1435 
1436 	mutex_lock(&intel_pstate_driver_lock);
1437 	ret = intel_pstate_update_status(buf, p ? p - buf : count);
1438 	mutex_unlock(&intel_pstate_driver_lock);
1439 
1440 	return ret < 0 ? ret : count;
1441 }
1442 
1443 static ssize_t show_turbo_pct(struct kobject *kobj,
1444 				struct kobj_attribute *attr, char *buf)
1445 {
1446 	struct cpudata *cpu;
1447 	int total, no_turbo, turbo_pct;
1448 	uint32_t turbo_fp;
1449 
1450 	mutex_lock(&intel_pstate_driver_lock);
1451 
1452 	if (!intel_pstate_driver) {
1453 		mutex_unlock(&intel_pstate_driver_lock);
1454 		return -EAGAIN;
1455 	}
1456 
1457 	cpu = all_cpu_data[0];
1458 
1459 	total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
1460 	no_turbo = cpu->pstate.max_pstate - cpu->pstate.min_pstate + 1;
1461 	turbo_fp = div_fp(no_turbo, total);
1462 	turbo_pct = 100 - fp_toint(mul_fp(turbo_fp, int_tofp(100)));
1463 
1464 	mutex_unlock(&intel_pstate_driver_lock);
1465 
1466 	return sprintf(buf, "%u\n", turbo_pct);
1467 }
1468 
1469 static ssize_t show_num_pstates(struct kobject *kobj,
1470 				struct kobj_attribute *attr, char *buf)
1471 {
1472 	struct cpudata *cpu;
1473 	int total;
1474 
1475 	mutex_lock(&intel_pstate_driver_lock);
1476 
1477 	if (!intel_pstate_driver) {
1478 		mutex_unlock(&intel_pstate_driver_lock);
1479 		return -EAGAIN;
1480 	}
1481 
1482 	cpu = all_cpu_data[0];
1483 	total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
1484 
1485 	mutex_unlock(&intel_pstate_driver_lock);
1486 
1487 	return sprintf(buf, "%u\n", total);
1488 }
1489 
1490 static ssize_t show_no_turbo(struct kobject *kobj,
1491 			     struct kobj_attribute *attr, char *buf)
1492 {
1493 	ssize_t ret;
1494 
1495 	mutex_lock(&intel_pstate_driver_lock);
1496 
1497 	if (!intel_pstate_driver) {
1498 		mutex_unlock(&intel_pstate_driver_lock);
1499 		return -EAGAIN;
1500 	}
1501 
1502 	ret = sprintf(buf, "%u\n", global.no_turbo);
1503 
1504 	mutex_unlock(&intel_pstate_driver_lock);
1505 
1506 	return ret;
1507 }
1508 
1509 static ssize_t store_no_turbo(struct kobject *a, struct kobj_attribute *b,
1510 			      const char *buf, size_t count)
1511 {
1512 	unsigned int input;
1513 	bool no_turbo;
1514 
1515 	if (sscanf(buf, "%u", &input) != 1)
1516 		return -EINVAL;
1517 
1518 	mutex_lock(&intel_pstate_driver_lock);
1519 
1520 	if (!intel_pstate_driver) {
1521 		count = -EAGAIN;
1522 		goto unlock_driver;
1523 	}
1524 
1525 	no_turbo = !!clamp_t(int, input, 0, 1);
1526 
1527 	WRITE_ONCE(global.turbo_disabled, turbo_is_disabled());
1528 	if (global.turbo_disabled && !no_turbo) {
1529 		pr_notice("Turbo disabled by BIOS or unavailable on processor\n");
1530 		count = -EPERM;
1531 		if (global.no_turbo)
1532 			goto unlock_driver;
1533 		else
1534 			no_turbo = 1;
1535 	}
1536 
1537 	if (no_turbo == global.no_turbo) {
1538 		goto unlock_driver;
1539 	}
1540 
1541 	WRITE_ONCE(global.no_turbo, no_turbo);
1542 
1543 	mutex_lock(&intel_pstate_limits_lock);
1544 
1545 	if (no_turbo) {
1546 		struct cpudata *cpu = all_cpu_data[0];
1547 		int pct = cpu->pstate.max_pstate * 100 / cpu->pstate.turbo_pstate;
1548 
1549 		/* Squash the global minimum into the permitted range. */
1550 		if (global.min_perf_pct > pct)
1551 			global.min_perf_pct = pct;
1552 	}
1553 
1554 	mutex_unlock(&intel_pstate_limits_lock);
1555 
1556 	intel_pstate_update_limits_for_all();
1557 	arch_set_max_freq_ratio(no_turbo);
1558 
1559 unlock_driver:
1560 	mutex_unlock(&intel_pstate_driver_lock);
1561 
1562 	return count;
1563 }
1564 
1565 static void update_qos_request(enum freq_qos_req_type type)
1566 {
1567 	struct freq_qos_request *req;
1568 	struct cpufreq_policy *policy;
1569 	int i;
1570 
1571 	for_each_possible_cpu(i) {
1572 		struct cpudata *cpu = all_cpu_data[i];
1573 		unsigned int freq, perf_pct;
1574 
1575 		policy = cpufreq_cpu_get(i);
1576 		if (!policy)
1577 			continue;
1578 
1579 		req = policy->driver_data;
1580 		cpufreq_cpu_put(policy);
1581 
1582 		if (!req)
1583 			continue;
1584 
1585 		if (hwp_active)
1586 			intel_pstate_get_hwp_cap(cpu);
1587 
1588 		if (type == FREQ_QOS_MIN) {
1589 			perf_pct = global.min_perf_pct;
1590 		} else {
1591 			req++;
1592 			perf_pct = global.max_perf_pct;
1593 		}
1594 
1595 		freq = DIV_ROUND_UP(cpu->pstate.turbo_freq * perf_pct, 100);
1596 
1597 		if (freq_qos_update_request(req, freq) < 0)
1598 			pr_warn("Failed to update freq constraint: CPU%d\n", i);
1599 	}
1600 }
1601 
1602 static ssize_t store_max_perf_pct(struct kobject *a, struct kobj_attribute *b,
1603 				  const char *buf, size_t count)
1604 {
1605 	unsigned int input;
1606 	int ret;
1607 
1608 	ret = sscanf(buf, "%u", &input);
1609 	if (ret != 1)
1610 		return -EINVAL;
1611 
1612 	mutex_lock(&intel_pstate_driver_lock);
1613 
1614 	if (!intel_pstate_driver) {
1615 		mutex_unlock(&intel_pstate_driver_lock);
1616 		return -EAGAIN;
1617 	}
1618 
1619 	mutex_lock(&intel_pstate_limits_lock);
1620 
1621 	global.max_perf_pct = clamp_t(int, input, global.min_perf_pct, 100);
1622 
1623 	mutex_unlock(&intel_pstate_limits_lock);
1624 
1625 	if (intel_pstate_driver == &intel_pstate)
1626 		intel_pstate_update_policies();
1627 	else
1628 		update_qos_request(FREQ_QOS_MAX);
1629 
1630 	mutex_unlock(&intel_pstate_driver_lock);
1631 
1632 	return count;
1633 }
1634 
1635 static ssize_t store_min_perf_pct(struct kobject *a, struct kobj_attribute *b,
1636 				  const char *buf, size_t count)
1637 {
1638 	unsigned int input;
1639 	int ret;
1640 
1641 	ret = sscanf(buf, "%u", &input);
1642 	if (ret != 1)
1643 		return -EINVAL;
1644 
1645 	mutex_lock(&intel_pstate_driver_lock);
1646 
1647 	if (!intel_pstate_driver) {
1648 		mutex_unlock(&intel_pstate_driver_lock);
1649 		return -EAGAIN;
1650 	}
1651 
1652 	mutex_lock(&intel_pstate_limits_lock);
1653 
1654 	global.min_perf_pct = clamp_t(int, input,
1655 				      min_perf_pct_min(), global.max_perf_pct);
1656 
1657 	mutex_unlock(&intel_pstate_limits_lock);
1658 
1659 	if (intel_pstate_driver == &intel_pstate)
1660 		intel_pstate_update_policies();
1661 	else
1662 		update_qos_request(FREQ_QOS_MIN);
1663 
1664 	mutex_unlock(&intel_pstate_driver_lock);
1665 
1666 	return count;
1667 }
1668 
1669 static ssize_t show_hwp_dynamic_boost(struct kobject *kobj,
1670 				struct kobj_attribute *attr, char *buf)
1671 {
1672 	return sprintf(buf, "%u\n", hwp_boost);
1673 }
1674 
1675 static ssize_t store_hwp_dynamic_boost(struct kobject *a,
1676 				       struct kobj_attribute *b,
1677 				       const char *buf, size_t count)
1678 {
1679 	unsigned int input;
1680 	int ret;
1681 
1682 	ret = kstrtouint(buf, 10, &input);
1683 	if (ret)
1684 		return ret;
1685 
1686 	mutex_lock(&intel_pstate_driver_lock);
1687 	hwp_boost = !!input;
1688 	intel_pstate_update_policies();
1689 	mutex_unlock(&intel_pstate_driver_lock);
1690 
1691 	return count;
1692 }
1693 
1694 static ssize_t show_energy_efficiency(struct kobject *kobj, struct kobj_attribute *attr,
1695 				      char *buf)
1696 {
1697 	u64 power_ctl;
1698 	int enable;
1699 
1700 	rdmsrl(MSR_IA32_POWER_CTL, power_ctl);
1701 	enable = !!(power_ctl & BIT(MSR_IA32_POWER_CTL_BIT_EE));
1702 	return sprintf(buf, "%d\n", !enable);
1703 }
1704 
1705 static ssize_t store_energy_efficiency(struct kobject *a, struct kobj_attribute *b,
1706 				       const char *buf, size_t count)
1707 {
1708 	bool input;
1709 	int ret;
1710 
1711 	ret = kstrtobool(buf, &input);
1712 	if (ret)
1713 		return ret;
1714 
1715 	set_power_ctl_ee_state(input);
1716 
1717 	return count;
1718 }
1719 
1720 show_one(max_perf_pct, max_perf_pct);
1721 show_one(min_perf_pct, min_perf_pct);
1722 
1723 define_one_global_rw(status);
1724 define_one_global_rw(no_turbo);
1725 define_one_global_rw(max_perf_pct);
1726 define_one_global_rw(min_perf_pct);
1727 define_one_global_ro(turbo_pct);
1728 define_one_global_ro(num_pstates);
1729 define_one_global_rw(hwp_dynamic_boost);
1730 define_one_global_rw(energy_efficiency);
1731 
1732 static struct attribute *intel_pstate_attributes[] = {
1733 	&status.attr,
1734 	&no_turbo.attr,
1735 	NULL
1736 };
1737 
1738 static const struct attribute_group intel_pstate_attr_group = {
1739 	.attrs = intel_pstate_attributes,
1740 };
1741 
1742 static const struct x86_cpu_id intel_pstate_cpu_ee_disable_ids[];
1743 
1744 static struct kobject *intel_pstate_kobject;
1745 
1746 static void __init intel_pstate_sysfs_expose_params(void)
1747 {
1748 	struct device *dev_root = bus_get_dev_root(&cpu_subsys);
1749 	int rc;
1750 
1751 	if (dev_root) {
1752 		intel_pstate_kobject = kobject_create_and_add("intel_pstate", &dev_root->kobj);
1753 		put_device(dev_root);
1754 	}
1755 	if (WARN_ON(!intel_pstate_kobject))
1756 		return;
1757 
1758 	rc = sysfs_create_group(intel_pstate_kobject, &intel_pstate_attr_group);
1759 	if (WARN_ON(rc))
1760 		return;
1761 
1762 	if (!boot_cpu_has(X86_FEATURE_HYBRID_CPU)) {
1763 		rc = sysfs_create_file(intel_pstate_kobject, &turbo_pct.attr);
1764 		WARN_ON(rc);
1765 
1766 		rc = sysfs_create_file(intel_pstate_kobject, &num_pstates.attr);
1767 		WARN_ON(rc);
1768 	}
1769 
1770 	/*
1771 	 * If per cpu limits are enforced there are no global limits, so
1772 	 * return without creating max/min_perf_pct attributes
1773 	 */
1774 	if (per_cpu_limits)
1775 		return;
1776 
1777 	rc = sysfs_create_file(intel_pstate_kobject, &max_perf_pct.attr);
1778 	WARN_ON(rc);
1779 
1780 	rc = sysfs_create_file(intel_pstate_kobject, &min_perf_pct.attr);
1781 	WARN_ON(rc);
1782 
1783 	if (x86_match_cpu(intel_pstate_cpu_ee_disable_ids)) {
1784 		rc = sysfs_create_file(intel_pstate_kobject, &energy_efficiency.attr);
1785 		WARN_ON(rc);
1786 	}
1787 }
1788 
1789 static void __init intel_pstate_sysfs_remove(void)
1790 {
1791 	if (!intel_pstate_kobject)
1792 		return;
1793 
1794 	sysfs_remove_group(intel_pstate_kobject, &intel_pstate_attr_group);
1795 
1796 	if (!boot_cpu_has(X86_FEATURE_HYBRID_CPU)) {
1797 		sysfs_remove_file(intel_pstate_kobject, &num_pstates.attr);
1798 		sysfs_remove_file(intel_pstate_kobject, &turbo_pct.attr);
1799 	}
1800 
1801 	if (!per_cpu_limits) {
1802 		sysfs_remove_file(intel_pstate_kobject, &max_perf_pct.attr);
1803 		sysfs_remove_file(intel_pstate_kobject, &min_perf_pct.attr);
1804 
1805 		if (x86_match_cpu(intel_pstate_cpu_ee_disable_ids))
1806 			sysfs_remove_file(intel_pstate_kobject, &energy_efficiency.attr);
1807 	}
1808 
1809 	kobject_put(intel_pstate_kobject);
1810 }
1811 
1812 static void intel_pstate_sysfs_expose_hwp_dynamic_boost(void)
1813 {
1814 	int rc;
1815 
1816 	if (!hwp_active)
1817 		return;
1818 
1819 	rc = sysfs_create_file(intel_pstate_kobject, &hwp_dynamic_boost.attr);
1820 	WARN_ON_ONCE(rc);
1821 }
1822 
1823 static void intel_pstate_sysfs_hide_hwp_dynamic_boost(void)
1824 {
1825 	if (!hwp_active)
1826 		return;
1827 
1828 	sysfs_remove_file(intel_pstate_kobject, &hwp_dynamic_boost.attr);
1829 }
1830 
1831 /************************** sysfs end ************************/
1832 
1833 static void intel_pstate_notify_work(struct work_struct *work)
1834 {
1835 	struct cpudata *cpudata =
1836 		container_of(to_delayed_work(work), struct cpudata, hwp_notify_work);
1837 	struct cpufreq_policy *policy = cpufreq_cpu_acquire(cpudata->cpu);
1838 
1839 	if (policy) {
1840 		__intel_pstate_update_max_freq(cpudata, policy);
1841 
1842 		cpufreq_cpu_release(policy);
1843 
1844 		/*
1845 		 * The driver will not be unregistered while this function is
1846 		 * running, so update the capacity without acquiring the driver
1847 		 * lock.
1848 		 */
1849 		hybrid_update_capacity(cpudata);
1850 	}
1851 
1852 	wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_STATUS, 0);
1853 }
1854 
1855 static DEFINE_RAW_SPINLOCK(hwp_notify_lock);
1856 static cpumask_t hwp_intr_enable_mask;
1857 
1858 #define HWP_GUARANTEED_PERF_CHANGE_STATUS      BIT(0)
1859 #define HWP_HIGHEST_PERF_CHANGE_STATUS         BIT(3)
1860 
1861 void notify_hwp_interrupt(void)
1862 {
1863 	unsigned int this_cpu = smp_processor_id();
1864 	u64 value, status_mask;
1865 	unsigned long flags;
1866 
1867 	if (!hwp_active || !cpu_feature_enabled(X86_FEATURE_HWP_NOTIFY))
1868 		return;
1869 
1870 	status_mask = HWP_GUARANTEED_PERF_CHANGE_STATUS;
1871 	if (cpu_feature_enabled(X86_FEATURE_HWP_HIGHEST_PERF_CHANGE))
1872 		status_mask |= HWP_HIGHEST_PERF_CHANGE_STATUS;
1873 
1874 	rdmsrl_safe(MSR_HWP_STATUS, &value);
1875 	if (!(value & status_mask))
1876 		return;
1877 
1878 	raw_spin_lock_irqsave(&hwp_notify_lock, flags);
1879 
1880 	if (!cpumask_test_cpu(this_cpu, &hwp_intr_enable_mask))
1881 		goto ack_intr;
1882 
1883 	schedule_delayed_work(&all_cpu_data[this_cpu]->hwp_notify_work,
1884 			      msecs_to_jiffies(10));
1885 
1886 	raw_spin_unlock_irqrestore(&hwp_notify_lock, flags);
1887 
1888 	return;
1889 
1890 ack_intr:
1891 	wrmsrl_safe(MSR_HWP_STATUS, 0);
1892 	raw_spin_unlock_irqrestore(&hwp_notify_lock, flags);
1893 }
1894 
1895 static void intel_pstate_disable_hwp_interrupt(struct cpudata *cpudata)
1896 {
1897 	bool cancel_work;
1898 
1899 	if (!cpu_feature_enabled(X86_FEATURE_HWP_NOTIFY))
1900 		return;
1901 
1902 	/* wrmsrl_on_cpu has to be outside spinlock as this can result in IPC */
1903 	wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_INTERRUPT, 0x00);
1904 
1905 	raw_spin_lock_irq(&hwp_notify_lock);
1906 	cancel_work = cpumask_test_and_clear_cpu(cpudata->cpu, &hwp_intr_enable_mask);
1907 	raw_spin_unlock_irq(&hwp_notify_lock);
1908 
1909 	if (cancel_work)
1910 		cancel_delayed_work_sync(&cpudata->hwp_notify_work);
1911 }
1912 
1913 #define HWP_GUARANTEED_PERF_CHANGE_REQ BIT(0)
1914 #define HWP_HIGHEST_PERF_CHANGE_REQ    BIT(2)
1915 
1916 static void intel_pstate_enable_hwp_interrupt(struct cpudata *cpudata)
1917 {
1918 	/* Enable HWP notification interrupt for performance change */
1919 	if (boot_cpu_has(X86_FEATURE_HWP_NOTIFY)) {
1920 		u64 interrupt_mask = HWP_GUARANTEED_PERF_CHANGE_REQ;
1921 
1922 		raw_spin_lock_irq(&hwp_notify_lock);
1923 		INIT_DELAYED_WORK(&cpudata->hwp_notify_work, intel_pstate_notify_work);
1924 		cpumask_set_cpu(cpudata->cpu, &hwp_intr_enable_mask);
1925 		raw_spin_unlock_irq(&hwp_notify_lock);
1926 
1927 		if (cpu_feature_enabled(X86_FEATURE_HWP_HIGHEST_PERF_CHANGE))
1928 			interrupt_mask |= HWP_HIGHEST_PERF_CHANGE_REQ;
1929 
1930 		/* wrmsrl_on_cpu has to be outside spinlock as this can result in IPC */
1931 		wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_INTERRUPT, interrupt_mask);
1932 		wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_STATUS, 0);
1933 	}
1934 }
1935 
1936 static void intel_pstate_update_epp_defaults(struct cpudata *cpudata)
1937 {
1938 	cpudata->epp_default = intel_pstate_get_epp(cpudata, 0);
1939 
1940 	/*
1941 	 * If the EPP is set by firmware, which means that firmware enabled HWP
1942 	 * - Is equal or less than 0x80 (default balance_perf EPP)
1943 	 * - But less performance oriented than performance EPP
1944 	 *   then use this as new balance_perf EPP.
1945 	 */
1946 	if (hwp_forced && cpudata->epp_default <= HWP_EPP_BALANCE_PERFORMANCE &&
1947 	    cpudata->epp_default > HWP_EPP_PERFORMANCE) {
1948 		epp_values[EPP_INDEX_BALANCE_PERFORMANCE] = cpudata->epp_default;
1949 		return;
1950 	}
1951 
1952 	/*
1953 	 * If this CPU gen doesn't call for change in balance_perf
1954 	 * EPP return.
1955 	 */
1956 	if (epp_values[EPP_INDEX_BALANCE_PERFORMANCE] == HWP_EPP_BALANCE_PERFORMANCE)
1957 		return;
1958 
1959 	/*
1960 	 * Use hard coded value per gen to update the balance_perf
1961 	 * and default EPP.
1962 	 */
1963 	cpudata->epp_default = epp_values[EPP_INDEX_BALANCE_PERFORMANCE];
1964 	intel_pstate_set_epp(cpudata, cpudata->epp_default);
1965 }
1966 
1967 static void intel_pstate_hwp_enable(struct cpudata *cpudata)
1968 {
1969 	/* First disable HWP notification interrupt till we activate again */
1970 	if (boot_cpu_has(X86_FEATURE_HWP_NOTIFY))
1971 		wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_INTERRUPT, 0x00);
1972 
1973 	wrmsrl_on_cpu(cpudata->cpu, MSR_PM_ENABLE, 0x1);
1974 
1975 	intel_pstate_enable_hwp_interrupt(cpudata);
1976 
1977 	if (cpudata->epp_default >= 0)
1978 		return;
1979 
1980 	intel_pstate_update_epp_defaults(cpudata);
1981 }
1982 
1983 static int atom_get_min_pstate(int not_used)
1984 {
1985 	u64 value;
1986 
1987 	rdmsrl(MSR_ATOM_CORE_RATIOS, value);
1988 	return (value >> 8) & 0x7F;
1989 }
1990 
1991 static int atom_get_max_pstate(int not_used)
1992 {
1993 	u64 value;
1994 
1995 	rdmsrl(MSR_ATOM_CORE_RATIOS, value);
1996 	return (value >> 16) & 0x7F;
1997 }
1998 
1999 static int atom_get_turbo_pstate(int not_used)
2000 {
2001 	u64 value;
2002 
2003 	rdmsrl(MSR_ATOM_CORE_TURBO_RATIOS, value);
2004 	return value & 0x7F;
2005 }
2006 
2007 static u64 atom_get_val(struct cpudata *cpudata, int pstate)
2008 {
2009 	u64 val;
2010 	int32_t vid_fp;
2011 	u32 vid;
2012 
2013 	val = (u64)pstate << 8;
2014 	if (READ_ONCE(global.no_turbo) && !READ_ONCE(global.turbo_disabled))
2015 		val |= (u64)1 << 32;
2016 
2017 	vid_fp = cpudata->vid.min + mul_fp(
2018 		int_tofp(pstate - cpudata->pstate.min_pstate),
2019 		cpudata->vid.ratio);
2020 
2021 	vid_fp = clamp_t(int32_t, vid_fp, cpudata->vid.min, cpudata->vid.max);
2022 	vid = ceiling_fp(vid_fp);
2023 
2024 	if (pstate > cpudata->pstate.max_pstate)
2025 		vid = cpudata->vid.turbo;
2026 
2027 	return val | vid;
2028 }
2029 
2030 static int silvermont_get_scaling(void)
2031 {
2032 	u64 value;
2033 	int i;
2034 	/* Defined in Table 35-6 from SDM (Sept 2015) */
2035 	static int silvermont_freq_table[] = {
2036 		83300, 100000, 133300, 116700, 80000};
2037 
2038 	rdmsrl(MSR_FSB_FREQ, value);
2039 	i = value & 0x7;
2040 	WARN_ON(i > 4);
2041 
2042 	return silvermont_freq_table[i];
2043 }
2044 
2045 static int airmont_get_scaling(void)
2046 {
2047 	u64 value;
2048 	int i;
2049 	/* Defined in Table 35-10 from SDM (Sept 2015) */
2050 	static int airmont_freq_table[] = {
2051 		83300, 100000, 133300, 116700, 80000,
2052 		93300, 90000, 88900, 87500};
2053 
2054 	rdmsrl(MSR_FSB_FREQ, value);
2055 	i = value & 0xF;
2056 	WARN_ON(i > 8);
2057 
2058 	return airmont_freq_table[i];
2059 }
2060 
2061 static void atom_get_vid(struct cpudata *cpudata)
2062 {
2063 	u64 value;
2064 
2065 	rdmsrl(MSR_ATOM_CORE_VIDS, value);
2066 	cpudata->vid.min = int_tofp((value >> 8) & 0x7f);
2067 	cpudata->vid.max = int_tofp((value >> 16) & 0x7f);
2068 	cpudata->vid.ratio = div_fp(
2069 		cpudata->vid.max - cpudata->vid.min,
2070 		int_tofp(cpudata->pstate.max_pstate -
2071 			cpudata->pstate.min_pstate));
2072 
2073 	rdmsrl(MSR_ATOM_CORE_TURBO_VIDS, value);
2074 	cpudata->vid.turbo = value & 0x7f;
2075 }
2076 
2077 static int core_get_min_pstate(int cpu)
2078 {
2079 	u64 value;
2080 
2081 	rdmsrl_on_cpu(cpu, MSR_PLATFORM_INFO, &value);
2082 	return (value >> 40) & 0xFF;
2083 }
2084 
2085 static int core_get_max_pstate_physical(int cpu)
2086 {
2087 	u64 value;
2088 
2089 	rdmsrl_on_cpu(cpu, MSR_PLATFORM_INFO, &value);
2090 	return (value >> 8) & 0xFF;
2091 }
2092 
2093 static int core_get_tdp_ratio(int cpu, u64 plat_info)
2094 {
2095 	/* Check how many TDP levels present */
2096 	if (plat_info & 0x600000000) {
2097 		u64 tdp_ctrl;
2098 		u64 tdp_ratio;
2099 		int tdp_msr;
2100 		int err;
2101 
2102 		/* Get the TDP level (0, 1, 2) to get ratios */
2103 		err = rdmsrl_safe_on_cpu(cpu, MSR_CONFIG_TDP_CONTROL, &tdp_ctrl);
2104 		if (err)
2105 			return err;
2106 
2107 		/* TDP MSR are continuous starting at 0x648 */
2108 		tdp_msr = MSR_CONFIG_TDP_NOMINAL + (tdp_ctrl & 0x03);
2109 		err = rdmsrl_safe_on_cpu(cpu, tdp_msr, &tdp_ratio);
2110 		if (err)
2111 			return err;
2112 
2113 		/* For level 1 and 2, bits[23:16] contain the ratio */
2114 		if (tdp_ctrl & 0x03)
2115 			tdp_ratio >>= 16;
2116 
2117 		tdp_ratio &= 0xff; /* ratios are only 8 bits long */
2118 		pr_debug("tdp_ratio %x\n", (int)tdp_ratio);
2119 
2120 		return (int)tdp_ratio;
2121 	}
2122 
2123 	return -ENXIO;
2124 }
2125 
2126 static int core_get_max_pstate(int cpu)
2127 {
2128 	u64 tar;
2129 	u64 plat_info;
2130 	int max_pstate;
2131 	int tdp_ratio;
2132 	int err;
2133 
2134 	rdmsrl_on_cpu(cpu, MSR_PLATFORM_INFO, &plat_info);
2135 	max_pstate = (plat_info >> 8) & 0xFF;
2136 
2137 	tdp_ratio = core_get_tdp_ratio(cpu, plat_info);
2138 	if (tdp_ratio <= 0)
2139 		return max_pstate;
2140 
2141 	if (hwp_active) {
2142 		/* Turbo activation ratio is not used on HWP platforms */
2143 		return tdp_ratio;
2144 	}
2145 
2146 	err = rdmsrl_safe_on_cpu(cpu, MSR_TURBO_ACTIVATION_RATIO, &tar);
2147 	if (!err) {
2148 		int tar_levels;
2149 
2150 		/* Do some sanity checking for safety */
2151 		tar_levels = tar & 0xff;
2152 		if (tdp_ratio - 1 == tar_levels) {
2153 			max_pstate = tar_levels;
2154 			pr_debug("max_pstate=TAC %x\n", max_pstate);
2155 		}
2156 	}
2157 
2158 	return max_pstate;
2159 }
2160 
2161 static int core_get_turbo_pstate(int cpu)
2162 {
2163 	u64 value;
2164 	int nont, ret;
2165 
2166 	rdmsrl_on_cpu(cpu, MSR_TURBO_RATIO_LIMIT, &value);
2167 	nont = core_get_max_pstate(cpu);
2168 	ret = (value) & 255;
2169 	if (ret <= nont)
2170 		ret = nont;
2171 	return ret;
2172 }
2173 
2174 static u64 core_get_val(struct cpudata *cpudata, int pstate)
2175 {
2176 	u64 val;
2177 
2178 	val = (u64)pstate << 8;
2179 	if (READ_ONCE(global.no_turbo) && !READ_ONCE(global.turbo_disabled))
2180 		val |= (u64)1 << 32;
2181 
2182 	return val;
2183 }
2184 
2185 static int knl_get_aperf_mperf_shift(void)
2186 {
2187 	return 10;
2188 }
2189 
2190 static int knl_get_turbo_pstate(int cpu)
2191 {
2192 	u64 value;
2193 	int nont, ret;
2194 
2195 	rdmsrl_on_cpu(cpu, MSR_TURBO_RATIO_LIMIT, &value);
2196 	nont = core_get_max_pstate(cpu);
2197 	ret = (((value) >> 8) & 0xFF);
2198 	if (ret <= nont)
2199 		ret = nont;
2200 	return ret;
2201 }
2202 
2203 static void hybrid_get_type(void *data)
2204 {
2205 	u8 *cpu_type = data;
2206 
2207 	*cpu_type = get_this_hybrid_cpu_type();
2208 }
2209 
2210 static int hwp_get_cpu_scaling(int cpu)
2211 {
2212 	if (hybrid_scaling_factor) {
2213 		u8 cpu_type = 0;
2214 
2215 		smp_call_function_single(cpu, hybrid_get_type, &cpu_type, 1);
2216 
2217 		/*
2218 		 * Return the hybrid scaling factor for P-cores and use the
2219 		 * default core scaling for E-cores.
2220 		 */
2221 		if (cpu_type == 0x40)
2222 			return hybrid_scaling_factor;
2223 
2224 		if (cpu_type == 0x20)
2225 			return core_get_scaling();
2226 	}
2227 
2228 	/* Use core scaling on non-hybrid systems. */
2229 	if (!cpu_feature_enabled(X86_FEATURE_HYBRID_CPU))
2230 		return core_get_scaling();
2231 
2232 	/*
2233 	 * The system is hybrid, but the hybrid scaling factor is not known or
2234 	 * the CPU type is not one of the above, so use CPPC to compute the
2235 	 * scaling factor for this CPU.
2236 	 */
2237 	return intel_pstate_cppc_get_scaling(cpu);
2238 }
2239 
2240 static void intel_pstate_set_pstate(struct cpudata *cpu, int pstate)
2241 {
2242 	trace_cpu_frequency(pstate * cpu->pstate.scaling, cpu->cpu);
2243 	cpu->pstate.current_pstate = pstate;
2244 	/*
2245 	 * Generally, there is no guarantee that this code will always run on
2246 	 * the CPU being updated, so force the register update to run on the
2247 	 * right CPU.
2248 	 */
2249 	wrmsrl_on_cpu(cpu->cpu, MSR_IA32_PERF_CTL,
2250 		      pstate_funcs.get_val(cpu, pstate));
2251 }
2252 
2253 static void intel_pstate_set_min_pstate(struct cpudata *cpu)
2254 {
2255 	intel_pstate_set_pstate(cpu, cpu->pstate.min_pstate);
2256 }
2257 
2258 static void intel_pstate_get_cpu_pstates(struct cpudata *cpu)
2259 {
2260 	int perf_ctl_max_phys = pstate_funcs.get_max_physical(cpu->cpu);
2261 	int perf_ctl_scaling = pstate_funcs.get_scaling();
2262 
2263 	cpu->pstate.min_pstate = pstate_funcs.get_min(cpu->cpu);
2264 	cpu->pstate.max_pstate_physical = perf_ctl_max_phys;
2265 	cpu->pstate.perf_ctl_scaling = perf_ctl_scaling;
2266 
2267 	if (hwp_active && !hwp_mode_bdw) {
2268 		__intel_pstate_get_hwp_cap(cpu);
2269 
2270 		if (pstate_funcs.get_cpu_scaling) {
2271 			cpu->pstate.scaling = pstate_funcs.get_cpu_scaling(cpu->cpu);
2272 			if (cpu->pstate.scaling != perf_ctl_scaling) {
2273 				intel_pstate_hybrid_hwp_adjust(cpu);
2274 				hwp_is_hybrid = true;
2275 			}
2276 		} else {
2277 			cpu->pstate.scaling = perf_ctl_scaling;
2278 		}
2279 		/*
2280 		 * If the CPU is going online for the first time and it was
2281 		 * offline initially, asym capacity scaling needs to be updated.
2282 		 */
2283 		hybrid_update_capacity(cpu);
2284 	} else {
2285 		cpu->pstate.scaling = perf_ctl_scaling;
2286 		cpu->pstate.max_pstate = pstate_funcs.get_max(cpu->cpu);
2287 		cpu->pstate.turbo_pstate = pstate_funcs.get_turbo(cpu->cpu);
2288 	}
2289 
2290 	if (cpu->pstate.scaling == perf_ctl_scaling) {
2291 		cpu->pstate.min_freq = cpu->pstate.min_pstate * perf_ctl_scaling;
2292 		cpu->pstate.max_freq = cpu->pstate.max_pstate * perf_ctl_scaling;
2293 		cpu->pstate.turbo_freq = cpu->pstate.turbo_pstate * perf_ctl_scaling;
2294 	}
2295 
2296 	if (pstate_funcs.get_aperf_mperf_shift)
2297 		cpu->aperf_mperf_shift = pstate_funcs.get_aperf_mperf_shift();
2298 
2299 	if (pstate_funcs.get_vid)
2300 		pstate_funcs.get_vid(cpu);
2301 
2302 	intel_pstate_set_min_pstate(cpu);
2303 }
2304 
2305 /*
2306  * Long hold time will keep high perf limits for long time,
2307  * which negatively impacts perf/watt for some workloads,
2308  * like specpower. 3ms is based on experiements on some
2309  * workoads.
2310  */
2311 static int hwp_boost_hold_time_ns = 3 * NSEC_PER_MSEC;
2312 
2313 static inline void intel_pstate_hwp_boost_up(struct cpudata *cpu)
2314 {
2315 	u64 hwp_req = READ_ONCE(cpu->hwp_req_cached);
2316 	u64 hwp_cap = READ_ONCE(cpu->hwp_cap_cached);
2317 	u32 max_limit = (hwp_req & 0xff00) >> 8;
2318 	u32 min_limit = (hwp_req & 0xff);
2319 	u32 boost_level1;
2320 
2321 	/*
2322 	 * Cases to consider (User changes via sysfs or boot time):
2323 	 * If, P0 (Turbo max) = P1 (Guaranteed max) = min:
2324 	 *	No boost, return.
2325 	 * If, P0 (Turbo max) > P1 (Guaranteed max) = min:
2326 	 *     Should result in one level boost only for P0.
2327 	 * If, P0 (Turbo max) = P1 (Guaranteed max) > min:
2328 	 *     Should result in two level boost:
2329 	 *         (min + p1)/2 and P1.
2330 	 * If, P0 (Turbo max) > P1 (Guaranteed max) > min:
2331 	 *     Should result in three level boost:
2332 	 *        (min + p1)/2, P1 and P0.
2333 	 */
2334 
2335 	/* If max and min are equal or already at max, nothing to boost */
2336 	if (max_limit == min_limit || cpu->hwp_boost_min >= max_limit)
2337 		return;
2338 
2339 	if (!cpu->hwp_boost_min)
2340 		cpu->hwp_boost_min = min_limit;
2341 
2342 	/* level at half way mark between min and guranteed */
2343 	boost_level1 = (HWP_GUARANTEED_PERF(hwp_cap) + min_limit) >> 1;
2344 
2345 	if (cpu->hwp_boost_min < boost_level1)
2346 		cpu->hwp_boost_min = boost_level1;
2347 	else if (cpu->hwp_boost_min < HWP_GUARANTEED_PERF(hwp_cap))
2348 		cpu->hwp_boost_min = HWP_GUARANTEED_PERF(hwp_cap);
2349 	else if (cpu->hwp_boost_min == HWP_GUARANTEED_PERF(hwp_cap) &&
2350 		 max_limit != HWP_GUARANTEED_PERF(hwp_cap))
2351 		cpu->hwp_boost_min = max_limit;
2352 	else
2353 		return;
2354 
2355 	hwp_req = (hwp_req & ~GENMASK_ULL(7, 0)) | cpu->hwp_boost_min;
2356 	wrmsrl(MSR_HWP_REQUEST, hwp_req);
2357 	cpu->last_update = cpu->sample.time;
2358 }
2359 
2360 static inline void intel_pstate_hwp_boost_down(struct cpudata *cpu)
2361 {
2362 	if (cpu->hwp_boost_min) {
2363 		bool expired;
2364 
2365 		/* Check if we are idle for hold time to boost down */
2366 		expired = time_after64(cpu->sample.time, cpu->last_update +
2367 				       hwp_boost_hold_time_ns);
2368 		if (expired) {
2369 			wrmsrl(MSR_HWP_REQUEST, cpu->hwp_req_cached);
2370 			cpu->hwp_boost_min = 0;
2371 		}
2372 	}
2373 	cpu->last_update = cpu->sample.time;
2374 }
2375 
2376 static inline void intel_pstate_update_util_hwp_local(struct cpudata *cpu,
2377 						      u64 time)
2378 {
2379 	cpu->sample.time = time;
2380 
2381 	if (cpu->sched_flags & SCHED_CPUFREQ_IOWAIT) {
2382 		bool do_io = false;
2383 
2384 		cpu->sched_flags = 0;
2385 		/*
2386 		 * Set iowait_boost flag and update time. Since IO WAIT flag
2387 		 * is set all the time, we can't just conclude that there is
2388 		 * some IO bound activity is scheduled on this CPU with just
2389 		 * one occurrence. If we receive at least two in two
2390 		 * consecutive ticks, then we treat as boost candidate.
2391 		 */
2392 		if (time_before64(time, cpu->last_io_update + 2 * TICK_NSEC))
2393 			do_io = true;
2394 
2395 		cpu->last_io_update = time;
2396 
2397 		if (do_io)
2398 			intel_pstate_hwp_boost_up(cpu);
2399 
2400 	} else {
2401 		intel_pstate_hwp_boost_down(cpu);
2402 	}
2403 }
2404 
2405 static inline void intel_pstate_update_util_hwp(struct update_util_data *data,
2406 						u64 time, unsigned int flags)
2407 {
2408 	struct cpudata *cpu = container_of(data, struct cpudata, update_util);
2409 
2410 	cpu->sched_flags |= flags;
2411 
2412 	if (smp_processor_id() == cpu->cpu)
2413 		intel_pstate_update_util_hwp_local(cpu, time);
2414 }
2415 
2416 static inline void intel_pstate_calc_avg_perf(struct cpudata *cpu)
2417 {
2418 	struct sample *sample = &cpu->sample;
2419 
2420 	sample->core_avg_perf = div_ext_fp(sample->aperf, sample->mperf);
2421 }
2422 
2423 static inline bool intel_pstate_sample(struct cpudata *cpu, u64 time)
2424 {
2425 	u64 aperf, mperf;
2426 	unsigned long flags;
2427 	u64 tsc;
2428 
2429 	local_irq_save(flags);
2430 	rdmsrl(MSR_IA32_APERF, aperf);
2431 	rdmsrl(MSR_IA32_MPERF, mperf);
2432 	tsc = rdtsc();
2433 	if (cpu->prev_mperf == mperf || cpu->prev_tsc == tsc) {
2434 		local_irq_restore(flags);
2435 		return false;
2436 	}
2437 	local_irq_restore(flags);
2438 
2439 	cpu->last_sample_time = cpu->sample.time;
2440 	cpu->sample.time = time;
2441 	cpu->sample.aperf = aperf;
2442 	cpu->sample.mperf = mperf;
2443 	cpu->sample.tsc =  tsc;
2444 	cpu->sample.aperf -= cpu->prev_aperf;
2445 	cpu->sample.mperf -= cpu->prev_mperf;
2446 	cpu->sample.tsc -= cpu->prev_tsc;
2447 
2448 	cpu->prev_aperf = aperf;
2449 	cpu->prev_mperf = mperf;
2450 	cpu->prev_tsc = tsc;
2451 	/*
2452 	 * First time this function is invoked in a given cycle, all of the
2453 	 * previous sample data fields are equal to zero or stale and they must
2454 	 * be populated with meaningful numbers for things to work, so assume
2455 	 * that sample.time will always be reset before setting the utilization
2456 	 * update hook and make the caller skip the sample then.
2457 	 */
2458 	if (cpu->last_sample_time) {
2459 		intel_pstate_calc_avg_perf(cpu);
2460 		return true;
2461 	}
2462 	return false;
2463 }
2464 
2465 static inline int32_t get_avg_frequency(struct cpudata *cpu)
2466 {
2467 	return mul_ext_fp(cpu->sample.core_avg_perf, cpu_khz);
2468 }
2469 
2470 static inline int32_t get_avg_pstate(struct cpudata *cpu)
2471 {
2472 	return mul_ext_fp(cpu->pstate.max_pstate_physical,
2473 			  cpu->sample.core_avg_perf);
2474 }
2475 
2476 static inline int32_t get_target_pstate(struct cpudata *cpu)
2477 {
2478 	struct sample *sample = &cpu->sample;
2479 	int32_t busy_frac;
2480 	int target, avg_pstate;
2481 
2482 	busy_frac = div_fp(sample->mperf << cpu->aperf_mperf_shift,
2483 			   sample->tsc);
2484 
2485 	if (busy_frac < cpu->iowait_boost)
2486 		busy_frac = cpu->iowait_boost;
2487 
2488 	sample->busy_scaled = busy_frac * 100;
2489 
2490 	target = READ_ONCE(global.no_turbo) ?
2491 			cpu->pstate.max_pstate : cpu->pstate.turbo_pstate;
2492 	target += target >> 2;
2493 	target = mul_fp(target, busy_frac);
2494 	if (target < cpu->pstate.min_pstate)
2495 		target = cpu->pstate.min_pstate;
2496 
2497 	/*
2498 	 * If the average P-state during the previous cycle was higher than the
2499 	 * current target, add 50% of the difference to the target to reduce
2500 	 * possible performance oscillations and offset possible performance
2501 	 * loss related to moving the workload from one CPU to another within
2502 	 * a package/module.
2503 	 */
2504 	avg_pstate = get_avg_pstate(cpu);
2505 	if (avg_pstate > target)
2506 		target += (avg_pstate - target) >> 1;
2507 
2508 	return target;
2509 }
2510 
2511 static int intel_pstate_prepare_request(struct cpudata *cpu, int pstate)
2512 {
2513 	int min_pstate = max(cpu->pstate.min_pstate, cpu->min_perf_ratio);
2514 	int max_pstate = max(min_pstate, cpu->max_perf_ratio);
2515 
2516 	return clamp_t(int, pstate, min_pstate, max_pstate);
2517 }
2518 
2519 static void intel_pstate_update_pstate(struct cpudata *cpu, int pstate)
2520 {
2521 	if (pstate == cpu->pstate.current_pstate)
2522 		return;
2523 
2524 	cpu->pstate.current_pstate = pstate;
2525 	wrmsrl(MSR_IA32_PERF_CTL, pstate_funcs.get_val(cpu, pstate));
2526 }
2527 
2528 static void intel_pstate_adjust_pstate(struct cpudata *cpu)
2529 {
2530 	int from = cpu->pstate.current_pstate;
2531 	struct sample *sample;
2532 	int target_pstate;
2533 
2534 	target_pstate = get_target_pstate(cpu);
2535 	target_pstate = intel_pstate_prepare_request(cpu, target_pstate);
2536 	trace_cpu_frequency(target_pstate * cpu->pstate.scaling, cpu->cpu);
2537 	intel_pstate_update_pstate(cpu, target_pstate);
2538 
2539 	sample = &cpu->sample;
2540 	trace_pstate_sample(mul_ext_fp(100, sample->core_avg_perf),
2541 		fp_toint(sample->busy_scaled),
2542 		from,
2543 		cpu->pstate.current_pstate,
2544 		sample->mperf,
2545 		sample->aperf,
2546 		sample->tsc,
2547 		get_avg_frequency(cpu),
2548 		fp_toint(cpu->iowait_boost * 100));
2549 }
2550 
2551 static void intel_pstate_update_util(struct update_util_data *data, u64 time,
2552 				     unsigned int flags)
2553 {
2554 	struct cpudata *cpu = container_of(data, struct cpudata, update_util);
2555 	u64 delta_ns;
2556 
2557 	/* Don't allow remote callbacks */
2558 	if (smp_processor_id() != cpu->cpu)
2559 		return;
2560 
2561 	delta_ns = time - cpu->last_update;
2562 	if (flags & SCHED_CPUFREQ_IOWAIT) {
2563 		/* Start over if the CPU may have been idle. */
2564 		if (delta_ns > TICK_NSEC) {
2565 			cpu->iowait_boost = ONE_EIGHTH_FP;
2566 		} else if (cpu->iowait_boost >= ONE_EIGHTH_FP) {
2567 			cpu->iowait_boost <<= 1;
2568 			if (cpu->iowait_boost > int_tofp(1))
2569 				cpu->iowait_boost = int_tofp(1);
2570 		} else {
2571 			cpu->iowait_boost = ONE_EIGHTH_FP;
2572 		}
2573 	} else if (cpu->iowait_boost) {
2574 		/* Clear iowait_boost if the CPU may have been idle. */
2575 		if (delta_ns > TICK_NSEC)
2576 			cpu->iowait_boost = 0;
2577 		else
2578 			cpu->iowait_boost >>= 1;
2579 	}
2580 	cpu->last_update = time;
2581 	delta_ns = time - cpu->sample.time;
2582 	if ((s64)delta_ns < INTEL_PSTATE_SAMPLING_INTERVAL)
2583 		return;
2584 
2585 	if (intel_pstate_sample(cpu, time))
2586 		intel_pstate_adjust_pstate(cpu);
2587 }
2588 
2589 static struct pstate_funcs core_funcs = {
2590 	.get_max = core_get_max_pstate,
2591 	.get_max_physical = core_get_max_pstate_physical,
2592 	.get_min = core_get_min_pstate,
2593 	.get_turbo = core_get_turbo_pstate,
2594 	.get_scaling = core_get_scaling,
2595 	.get_val = core_get_val,
2596 };
2597 
2598 static const struct pstate_funcs silvermont_funcs = {
2599 	.get_max = atom_get_max_pstate,
2600 	.get_max_physical = atom_get_max_pstate,
2601 	.get_min = atom_get_min_pstate,
2602 	.get_turbo = atom_get_turbo_pstate,
2603 	.get_val = atom_get_val,
2604 	.get_scaling = silvermont_get_scaling,
2605 	.get_vid = atom_get_vid,
2606 };
2607 
2608 static const struct pstate_funcs airmont_funcs = {
2609 	.get_max = atom_get_max_pstate,
2610 	.get_max_physical = atom_get_max_pstate,
2611 	.get_min = atom_get_min_pstate,
2612 	.get_turbo = atom_get_turbo_pstate,
2613 	.get_val = atom_get_val,
2614 	.get_scaling = airmont_get_scaling,
2615 	.get_vid = atom_get_vid,
2616 };
2617 
2618 static const struct pstate_funcs knl_funcs = {
2619 	.get_max = core_get_max_pstate,
2620 	.get_max_physical = core_get_max_pstate_physical,
2621 	.get_min = core_get_min_pstate,
2622 	.get_turbo = knl_get_turbo_pstate,
2623 	.get_aperf_mperf_shift = knl_get_aperf_mperf_shift,
2624 	.get_scaling = core_get_scaling,
2625 	.get_val = core_get_val,
2626 };
2627 
2628 #define X86_MATCH(vfm, policy)					 \
2629 	X86_MATCH_VFM_FEATURE(vfm, X86_FEATURE_APERFMPERF, &policy)
2630 
2631 static const struct x86_cpu_id intel_pstate_cpu_ids[] = {
2632 	X86_MATCH(INTEL_SANDYBRIDGE,		core_funcs),
2633 	X86_MATCH(INTEL_SANDYBRIDGE_X,		core_funcs),
2634 	X86_MATCH(INTEL_ATOM_SILVERMONT,	silvermont_funcs),
2635 	X86_MATCH(INTEL_IVYBRIDGE,		core_funcs),
2636 	X86_MATCH(INTEL_HASWELL,		core_funcs),
2637 	X86_MATCH(INTEL_BROADWELL,		core_funcs),
2638 	X86_MATCH(INTEL_IVYBRIDGE_X,		core_funcs),
2639 	X86_MATCH(INTEL_HASWELL_X,		core_funcs),
2640 	X86_MATCH(INTEL_HASWELL_L,		core_funcs),
2641 	X86_MATCH(INTEL_HASWELL_G,		core_funcs),
2642 	X86_MATCH(INTEL_BROADWELL_G,		core_funcs),
2643 	X86_MATCH(INTEL_ATOM_AIRMONT,		airmont_funcs),
2644 	X86_MATCH(INTEL_SKYLAKE_L,		core_funcs),
2645 	X86_MATCH(INTEL_BROADWELL_X,		core_funcs),
2646 	X86_MATCH(INTEL_SKYLAKE,		core_funcs),
2647 	X86_MATCH(INTEL_BROADWELL_D,		core_funcs),
2648 	X86_MATCH(INTEL_XEON_PHI_KNL,		knl_funcs),
2649 	X86_MATCH(INTEL_XEON_PHI_KNM,		knl_funcs),
2650 	X86_MATCH(INTEL_ATOM_GOLDMONT,		core_funcs),
2651 	X86_MATCH(INTEL_ATOM_GOLDMONT_PLUS,	core_funcs),
2652 	X86_MATCH(INTEL_SKYLAKE_X,		core_funcs),
2653 	X86_MATCH(INTEL_COMETLAKE,		core_funcs),
2654 	X86_MATCH(INTEL_ICELAKE_X,		core_funcs),
2655 	X86_MATCH(INTEL_TIGERLAKE,		core_funcs),
2656 	X86_MATCH(INTEL_SAPPHIRERAPIDS_X,	core_funcs),
2657 	X86_MATCH(INTEL_EMERALDRAPIDS_X,	core_funcs),
2658 	{}
2659 };
2660 MODULE_DEVICE_TABLE(x86cpu, intel_pstate_cpu_ids);
2661 
2662 #ifdef CONFIG_ACPI
2663 static const struct x86_cpu_id intel_pstate_cpu_oob_ids[] __initconst = {
2664 	X86_MATCH(INTEL_BROADWELL_D,		core_funcs),
2665 	X86_MATCH(INTEL_BROADWELL_X,		core_funcs),
2666 	X86_MATCH(INTEL_SKYLAKE_X,		core_funcs),
2667 	X86_MATCH(INTEL_ICELAKE_X,		core_funcs),
2668 	X86_MATCH(INTEL_SAPPHIRERAPIDS_X,	core_funcs),
2669 	X86_MATCH(INTEL_EMERALDRAPIDS_X,	core_funcs),
2670 	X86_MATCH(INTEL_GRANITERAPIDS_D,	core_funcs),
2671 	X86_MATCH(INTEL_GRANITERAPIDS_X,	core_funcs),
2672 	X86_MATCH(INTEL_ATOM_CRESTMONT,		core_funcs),
2673 	X86_MATCH(INTEL_ATOM_CRESTMONT_X,	core_funcs),
2674 	{}
2675 };
2676 #endif
2677 
2678 static const struct x86_cpu_id intel_pstate_cpu_ee_disable_ids[] = {
2679 	X86_MATCH(INTEL_KABYLAKE,		core_funcs),
2680 	{}
2681 };
2682 
2683 static int intel_pstate_init_cpu(unsigned int cpunum)
2684 {
2685 	struct cpudata *cpu;
2686 
2687 	cpu = all_cpu_data[cpunum];
2688 
2689 	if (!cpu) {
2690 		cpu = kzalloc(sizeof(*cpu), GFP_KERNEL);
2691 		if (!cpu)
2692 			return -ENOMEM;
2693 
2694 		WRITE_ONCE(all_cpu_data[cpunum], cpu);
2695 
2696 		cpu->cpu = cpunum;
2697 
2698 		cpu->epp_default = -EINVAL;
2699 
2700 		if (hwp_active) {
2701 			intel_pstate_hwp_enable(cpu);
2702 
2703 			if (intel_pstate_acpi_pm_profile_server())
2704 				hwp_boost = true;
2705 		}
2706 	} else if (hwp_active) {
2707 		/*
2708 		 * Re-enable HWP in case this happens after a resume from ACPI
2709 		 * S3 if the CPU was offline during the whole system/resume
2710 		 * cycle.
2711 		 */
2712 		intel_pstate_hwp_reenable(cpu);
2713 	}
2714 
2715 	cpu->epp_powersave = -EINVAL;
2716 	cpu->epp_policy = CPUFREQ_POLICY_UNKNOWN;
2717 
2718 	intel_pstate_get_cpu_pstates(cpu);
2719 
2720 	pr_debug("controlling: cpu %d\n", cpunum);
2721 
2722 	return 0;
2723 }
2724 
2725 static void intel_pstate_set_update_util_hook(unsigned int cpu_num)
2726 {
2727 	struct cpudata *cpu = all_cpu_data[cpu_num];
2728 
2729 	if (hwp_active && !hwp_boost)
2730 		return;
2731 
2732 	if (cpu->update_util_set)
2733 		return;
2734 
2735 	/* Prevent intel_pstate_update_util() from using stale data. */
2736 	cpu->sample.time = 0;
2737 	cpufreq_add_update_util_hook(cpu_num, &cpu->update_util,
2738 				     (hwp_active ?
2739 				      intel_pstate_update_util_hwp :
2740 				      intel_pstate_update_util));
2741 	cpu->update_util_set = true;
2742 }
2743 
2744 static void intel_pstate_clear_update_util_hook(unsigned int cpu)
2745 {
2746 	struct cpudata *cpu_data = all_cpu_data[cpu];
2747 
2748 	if (!cpu_data->update_util_set)
2749 		return;
2750 
2751 	cpufreq_remove_update_util_hook(cpu);
2752 	cpu_data->update_util_set = false;
2753 	synchronize_rcu();
2754 }
2755 
2756 static int intel_pstate_get_max_freq(struct cpudata *cpu)
2757 {
2758 	return READ_ONCE(global.no_turbo) ?
2759 			cpu->pstate.max_freq : cpu->pstate.turbo_freq;
2760 }
2761 
2762 static void intel_pstate_update_perf_limits(struct cpudata *cpu,
2763 					    unsigned int policy_min,
2764 					    unsigned int policy_max)
2765 {
2766 	int perf_ctl_scaling = cpu->pstate.perf_ctl_scaling;
2767 	int32_t max_policy_perf, min_policy_perf;
2768 
2769 	max_policy_perf = policy_max / perf_ctl_scaling;
2770 	if (policy_max == policy_min) {
2771 		min_policy_perf = max_policy_perf;
2772 	} else {
2773 		min_policy_perf = policy_min / perf_ctl_scaling;
2774 		min_policy_perf = clamp_t(int32_t, min_policy_perf,
2775 					  0, max_policy_perf);
2776 	}
2777 
2778 	/*
2779 	 * HWP needs some special consideration, because HWP_REQUEST uses
2780 	 * abstract values to represent performance rather than pure ratios.
2781 	 */
2782 	if (hwp_active && cpu->pstate.scaling != perf_ctl_scaling) {
2783 		int freq;
2784 
2785 		freq = max_policy_perf * perf_ctl_scaling;
2786 		max_policy_perf = intel_pstate_freq_to_hwp(cpu, freq);
2787 		freq = min_policy_perf * perf_ctl_scaling;
2788 		min_policy_perf = intel_pstate_freq_to_hwp(cpu, freq);
2789 	}
2790 
2791 	pr_debug("cpu:%d min_policy_perf:%d max_policy_perf:%d\n",
2792 		 cpu->cpu, min_policy_perf, max_policy_perf);
2793 
2794 	/* Normalize user input to [min_perf, max_perf] */
2795 	if (per_cpu_limits) {
2796 		cpu->min_perf_ratio = min_policy_perf;
2797 		cpu->max_perf_ratio = max_policy_perf;
2798 	} else {
2799 		int turbo_max = cpu->pstate.turbo_pstate;
2800 		int32_t global_min, global_max;
2801 
2802 		/* Global limits are in percent of the maximum turbo P-state. */
2803 		global_max = DIV_ROUND_UP(turbo_max * global.max_perf_pct, 100);
2804 		global_min = DIV_ROUND_UP(turbo_max * global.min_perf_pct, 100);
2805 		global_min = clamp_t(int32_t, global_min, 0, global_max);
2806 
2807 		pr_debug("cpu:%d global_min:%d global_max:%d\n", cpu->cpu,
2808 			 global_min, global_max);
2809 
2810 		cpu->min_perf_ratio = max(min_policy_perf, global_min);
2811 		cpu->min_perf_ratio = min(cpu->min_perf_ratio, max_policy_perf);
2812 		cpu->max_perf_ratio = min(max_policy_perf, global_max);
2813 		cpu->max_perf_ratio = max(min_policy_perf, cpu->max_perf_ratio);
2814 
2815 		/* Make sure min_perf <= max_perf */
2816 		cpu->min_perf_ratio = min(cpu->min_perf_ratio,
2817 					  cpu->max_perf_ratio);
2818 
2819 	}
2820 	pr_debug("cpu:%d max_perf_ratio:%d min_perf_ratio:%d\n", cpu->cpu,
2821 		 cpu->max_perf_ratio,
2822 		 cpu->min_perf_ratio);
2823 }
2824 
2825 static int intel_pstate_set_policy(struct cpufreq_policy *policy)
2826 {
2827 	struct cpudata *cpu;
2828 
2829 	if (!policy->cpuinfo.max_freq)
2830 		return -ENODEV;
2831 
2832 	pr_debug("set_policy cpuinfo.max %u policy->max %u\n",
2833 		 policy->cpuinfo.max_freq, policy->max);
2834 
2835 	cpu = all_cpu_data[policy->cpu];
2836 	cpu->policy = policy->policy;
2837 
2838 	mutex_lock(&intel_pstate_limits_lock);
2839 
2840 	intel_pstate_update_perf_limits(cpu, policy->min, policy->max);
2841 
2842 	if (cpu->policy == CPUFREQ_POLICY_PERFORMANCE) {
2843 		int pstate = max(cpu->pstate.min_pstate, cpu->max_perf_ratio);
2844 
2845 		/*
2846 		 * NOHZ_FULL CPUs need this as the governor callback may not
2847 		 * be invoked on them.
2848 		 */
2849 		intel_pstate_clear_update_util_hook(policy->cpu);
2850 		intel_pstate_set_pstate(cpu, pstate);
2851 	} else {
2852 		intel_pstate_set_update_util_hook(policy->cpu);
2853 	}
2854 
2855 	if (hwp_active) {
2856 		/*
2857 		 * When hwp_boost was active before and dynamically it
2858 		 * was turned off, in that case we need to clear the
2859 		 * update util hook.
2860 		 */
2861 		if (!hwp_boost)
2862 			intel_pstate_clear_update_util_hook(policy->cpu);
2863 		intel_pstate_hwp_set(policy->cpu);
2864 	}
2865 	/*
2866 	 * policy->cur is never updated with the intel_pstate driver, but it
2867 	 * is used as a stale frequency value. So, keep it within limits.
2868 	 */
2869 	policy->cur = policy->min;
2870 
2871 	mutex_unlock(&intel_pstate_limits_lock);
2872 
2873 	return 0;
2874 }
2875 
2876 static void intel_pstate_adjust_policy_max(struct cpudata *cpu,
2877 					   struct cpufreq_policy_data *policy)
2878 {
2879 	if (!hwp_active &&
2880 	    cpu->pstate.max_pstate_physical > cpu->pstate.max_pstate &&
2881 	    policy->max < policy->cpuinfo.max_freq &&
2882 	    policy->max > cpu->pstate.max_freq) {
2883 		pr_debug("policy->max > max non turbo frequency\n");
2884 		policy->max = policy->cpuinfo.max_freq;
2885 	}
2886 }
2887 
2888 static void intel_pstate_verify_cpu_policy(struct cpudata *cpu,
2889 					   struct cpufreq_policy_data *policy)
2890 {
2891 	int max_freq;
2892 
2893 	if (hwp_active) {
2894 		intel_pstate_get_hwp_cap(cpu);
2895 		max_freq = READ_ONCE(global.no_turbo) ?
2896 				cpu->pstate.max_freq : cpu->pstate.turbo_freq;
2897 	} else {
2898 		max_freq = intel_pstate_get_max_freq(cpu);
2899 	}
2900 	cpufreq_verify_within_limits(policy, policy->cpuinfo.min_freq, max_freq);
2901 
2902 	intel_pstate_adjust_policy_max(cpu, policy);
2903 }
2904 
2905 static int intel_pstate_verify_policy(struct cpufreq_policy_data *policy)
2906 {
2907 	intel_pstate_verify_cpu_policy(all_cpu_data[policy->cpu], policy);
2908 
2909 	return 0;
2910 }
2911 
2912 static int intel_cpufreq_cpu_offline(struct cpufreq_policy *policy)
2913 {
2914 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2915 
2916 	pr_debug("CPU %d going offline\n", cpu->cpu);
2917 
2918 	if (cpu->suspended)
2919 		return 0;
2920 
2921 	/*
2922 	 * If the CPU is an SMT thread and it goes offline with the performance
2923 	 * settings different from the minimum, it will prevent its sibling
2924 	 * from getting to lower performance levels, so force the minimum
2925 	 * performance on CPU offline to prevent that from happening.
2926 	 */
2927 	if (hwp_active)
2928 		intel_pstate_hwp_offline(cpu);
2929 	else
2930 		intel_pstate_set_min_pstate(cpu);
2931 
2932 	intel_pstate_exit_perf_limits(policy);
2933 
2934 	return 0;
2935 }
2936 
2937 static int intel_pstate_cpu_online(struct cpufreq_policy *policy)
2938 {
2939 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2940 
2941 	pr_debug("CPU %d going online\n", cpu->cpu);
2942 
2943 	intel_pstate_init_acpi_perf_limits(policy);
2944 
2945 	if (hwp_active) {
2946 		/*
2947 		 * Re-enable HWP and clear the "suspended" flag to let "resume"
2948 		 * know that it need not do that.
2949 		 */
2950 		intel_pstate_hwp_reenable(cpu);
2951 		cpu->suspended = false;
2952 
2953 		hybrid_update_capacity(cpu);
2954 	}
2955 
2956 	return 0;
2957 }
2958 
2959 static int intel_pstate_cpu_offline(struct cpufreq_policy *policy)
2960 {
2961 	intel_pstate_clear_update_util_hook(policy->cpu);
2962 
2963 	return intel_cpufreq_cpu_offline(policy);
2964 }
2965 
2966 static void intel_pstate_cpu_exit(struct cpufreq_policy *policy)
2967 {
2968 	pr_debug("CPU %d exiting\n", policy->cpu);
2969 
2970 	policy->fast_switch_possible = false;
2971 }
2972 
2973 static int __intel_pstate_cpu_init(struct cpufreq_policy *policy)
2974 {
2975 	struct cpudata *cpu;
2976 	int rc;
2977 
2978 	rc = intel_pstate_init_cpu(policy->cpu);
2979 	if (rc)
2980 		return rc;
2981 
2982 	cpu = all_cpu_data[policy->cpu];
2983 
2984 	cpu->max_perf_ratio = 0xFF;
2985 	cpu->min_perf_ratio = 0;
2986 
2987 	/* cpuinfo and default policy values */
2988 	policy->cpuinfo.min_freq = cpu->pstate.min_freq;
2989 	policy->cpuinfo.max_freq = READ_ONCE(global.no_turbo) ?
2990 			cpu->pstate.max_freq : cpu->pstate.turbo_freq;
2991 
2992 	policy->min = policy->cpuinfo.min_freq;
2993 	policy->max = policy->cpuinfo.max_freq;
2994 
2995 	intel_pstate_init_acpi_perf_limits(policy);
2996 
2997 	policy->fast_switch_possible = true;
2998 
2999 	return 0;
3000 }
3001 
3002 static int intel_pstate_cpu_init(struct cpufreq_policy *policy)
3003 {
3004 	int ret = __intel_pstate_cpu_init(policy);
3005 
3006 	if (ret)
3007 		return ret;
3008 
3009 	/*
3010 	 * Set the policy to powersave to provide a valid fallback value in case
3011 	 * the default cpufreq governor is neither powersave nor performance.
3012 	 */
3013 	policy->policy = CPUFREQ_POLICY_POWERSAVE;
3014 
3015 	if (hwp_active) {
3016 		struct cpudata *cpu = all_cpu_data[policy->cpu];
3017 
3018 		cpu->epp_cached = intel_pstate_get_epp(cpu, 0);
3019 	}
3020 
3021 	return 0;
3022 }
3023 
3024 static struct cpufreq_driver intel_pstate = {
3025 	.flags		= CPUFREQ_CONST_LOOPS,
3026 	.verify		= intel_pstate_verify_policy,
3027 	.setpolicy	= intel_pstate_set_policy,
3028 	.suspend	= intel_pstate_suspend,
3029 	.resume		= intel_pstate_resume,
3030 	.init		= intel_pstate_cpu_init,
3031 	.exit		= intel_pstate_cpu_exit,
3032 	.offline	= intel_pstate_cpu_offline,
3033 	.online		= intel_pstate_cpu_online,
3034 	.update_limits	= intel_pstate_update_limits,
3035 	.name		= "intel_pstate",
3036 };
3037 
3038 static int intel_cpufreq_verify_policy(struct cpufreq_policy_data *policy)
3039 {
3040 	struct cpudata *cpu = all_cpu_data[policy->cpu];
3041 
3042 	intel_pstate_verify_cpu_policy(cpu, policy);
3043 	intel_pstate_update_perf_limits(cpu, policy->min, policy->max);
3044 
3045 	return 0;
3046 }
3047 
3048 /* Use of trace in passive mode:
3049  *
3050  * In passive mode the trace core_busy field (also known as the
3051  * performance field, and lablelled as such on the graphs; also known as
3052  * core_avg_perf) is not needed and so is re-assigned to indicate if the
3053  * driver call was via the normal or fast switch path. Various graphs
3054  * output from the intel_pstate_tracer.py utility that include core_busy
3055  * (or performance or core_avg_perf) have a fixed y-axis from 0 to 100%,
3056  * so we use 10 to indicate the normal path through the driver, and
3057  * 90 to indicate the fast switch path through the driver.
3058  * The scaled_busy field is not used, and is set to 0.
3059  */
3060 
3061 #define	INTEL_PSTATE_TRACE_TARGET 10
3062 #define	INTEL_PSTATE_TRACE_FAST_SWITCH 90
3063 
3064 static void intel_cpufreq_trace(struct cpudata *cpu, unsigned int trace_type, int old_pstate)
3065 {
3066 	struct sample *sample;
3067 
3068 	if (!trace_pstate_sample_enabled())
3069 		return;
3070 
3071 	if (!intel_pstate_sample(cpu, ktime_get()))
3072 		return;
3073 
3074 	sample = &cpu->sample;
3075 	trace_pstate_sample(trace_type,
3076 		0,
3077 		old_pstate,
3078 		cpu->pstate.current_pstate,
3079 		sample->mperf,
3080 		sample->aperf,
3081 		sample->tsc,
3082 		get_avg_frequency(cpu),
3083 		fp_toint(cpu->iowait_boost * 100));
3084 }
3085 
3086 static void intel_cpufreq_hwp_update(struct cpudata *cpu, u32 min, u32 max,
3087 				     u32 desired, bool fast_switch)
3088 {
3089 	u64 prev = READ_ONCE(cpu->hwp_req_cached), value = prev;
3090 
3091 	value &= ~HWP_MIN_PERF(~0L);
3092 	value |= HWP_MIN_PERF(min);
3093 
3094 	value &= ~HWP_MAX_PERF(~0L);
3095 	value |= HWP_MAX_PERF(max);
3096 
3097 	value &= ~HWP_DESIRED_PERF(~0L);
3098 	value |= HWP_DESIRED_PERF(desired);
3099 
3100 	if (value == prev)
3101 		return;
3102 
3103 	WRITE_ONCE(cpu->hwp_req_cached, value);
3104 	if (fast_switch)
3105 		wrmsrl(MSR_HWP_REQUEST, value);
3106 	else
3107 		wrmsrl_on_cpu(cpu->cpu, MSR_HWP_REQUEST, value);
3108 }
3109 
3110 static void intel_cpufreq_perf_ctl_update(struct cpudata *cpu,
3111 					  u32 target_pstate, bool fast_switch)
3112 {
3113 	if (fast_switch)
3114 		wrmsrl(MSR_IA32_PERF_CTL,
3115 		       pstate_funcs.get_val(cpu, target_pstate));
3116 	else
3117 		wrmsrl_on_cpu(cpu->cpu, MSR_IA32_PERF_CTL,
3118 			      pstate_funcs.get_val(cpu, target_pstate));
3119 }
3120 
3121 static int intel_cpufreq_update_pstate(struct cpufreq_policy *policy,
3122 				       int target_pstate, bool fast_switch)
3123 {
3124 	struct cpudata *cpu = all_cpu_data[policy->cpu];
3125 	int old_pstate = cpu->pstate.current_pstate;
3126 
3127 	target_pstate = intel_pstate_prepare_request(cpu, target_pstate);
3128 	if (hwp_active) {
3129 		int max_pstate = policy->strict_target ?
3130 					target_pstate : cpu->max_perf_ratio;
3131 
3132 		intel_cpufreq_hwp_update(cpu, target_pstate, max_pstate, 0,
3133 					 fast_switch);
3134 	} else if (target_pstate != old_pstate) {
3135 		intel_cpufreq_perf_ctl_update(cpu, target_pstate, fast_switch);
3136 	}
3137 
3138 	cpu->pstate.current_pstate = target_pstate;
3139 
3140 	intel_cpufreq_trace(cpu, fast_switch ? INTEL_PSTATE_TRACE_FAST_SWITCH :
3141 			    INTEL_PSTATE_TRACE_TARGET, old_pstate);
3142 
3143 	return target_pstate;
3144 }
3145 
3146 static int intel_cpufreq_target(struct cpufreq_policy *policy,
3147 				unsigned int target_freq,
3148 				unsigned int relation)
3149 {
3150 	struct cpudata *cpu = all_cpu_data[policy->cpu];
3151 	struct cpufreq_freqs freqs;
3152 	int target_pstate;
3153 
3154 	freqs.old = policy->cur;
3155 	freqs.new = target_freq;
3156 
3157 	cpufreq_freq_transition_begin(policy, &freqs);
3158 
3159 	target_pstate = intel_pstate_freq_to_hwp_rel(cpu, freqs.new, relation);
3160 	target_pstate = intel_cpufreq_update_pstate(policy, target_pstate, false);
3161 
3162 	freqs.new = target_pstate * cpu->pstate.scaling;
3163 
3164 	cpufreq_freq_transition_end(policy, &freqs, false);
3165 
3166 	return 0;
3167 }
3168 
3169 static unsigned int intel_cpufreq_fast_switch(struct cpufreq_policy *policy,
3170 					      unsigned int target_freq)
3171 {
3172 	struct cpudata *cpu = all_cpu_data[policy->cpu];
3173 	int target_pstate;
3174 
3175 	target_pstate = intel_pstate_freq_to_hwp(cpu, target_freq);
3176 
3177 	target_pstate = intel_cpufreq_update_pstate(policy, target_pstate, true);
3178 
3179 	return target_pstate * cpu->pstate.scaling;
3180 }
3181 
3182 static void intel_cpufreq_adjust_perf(unsigned int cpunum,
3183 				      unsigned long min_perf,
3184 				      unsigned long target_perf,
3185 				      unsigned long capacity)
3186 {
3187 	struct cpudata *cpu = all_cpu_data[cpunum];
3188 	u64 hwp_cap = READ_ONCE(cpu->hwp_cap_cached);
3189 	int old_pstate = cpu->pstate.current_pstate;
3190 	int cap_pstate, min_pstate, max_pstate, target_pstate;
3191 
3192 	cap_pstate = READ_ONCE(global.no_turbo) ?
3193 					HWP_GUARANTEED_PERF(hwp_cap) :
3194 					HWP_HIGHEST_PERF(hwp_cap);
3195 
3196 	/* Optimization: Avoid unnecessary divisions. */
3197 
3198 	target_pstate = cap_pstate;
3199 	if (target_perf < capacity)
3200 		target_pstate = DIV_ROUND_UP(cap_pstate * target_perf, capacity);
3201 
3202 	min_pstate = cap_pstate;
3203 	if (min_perf < capacity)
3204 		min_pstate = DIV_ROUND_UP(cap_pstate * min_perf, capacity);
3205 
3206 	if (min_pstate < cpu->pstate.min_pstate)
3207 		min_pstate = cpu->pstate.min_pstate;
3208 
3209 	if (min_pstate < cpu->min_perf_ratio)
3210 		min_pstate = cpu->min_perf_ratio;
3211 
3212 	if (min_pstate > cpu->max_perf_ratio)
3213 		min_pstate = cpu->max_perf_ratio;
3214 
3215 	max_pstate = min(cap_pstate, cpu->max_perf_ratio);
3216 	if (max_pstate < min_pstate)
3217 		max_pstate = min_pstate;
3218 
3219 	target_pstate = clamp_t(int, target_pstate, min_pstate, max_pstate);
3220 
3221 	intel_cpufreq_hwp_update(cpu, min_pstate, max_pstate, target_pstate, true);
3222 
3223 	cpu->pstate.current_pstate = target_pstate;
3224 	intel_cpufreq_trace(cpu, INTEL_PSTATE_TRACE_FAST_SWITCH, old_pstate);
3225 }
3226 
3227 static int intel_cpufreq_cpu_init(struct cpufreq_policy *policy)
3228 {
3229 	struct freq_qos_request *req;
3230 	struct cpudata *cpu;
3231 	struct device *dev;
3232 	int ret, freq;
3233 
3234 	dev = get_cpu_device(policy->cpu);
3235 	if (!dev)
3236 		return -ENODEV;
3237 
3238 	ret = __intel_pstate_cpu_init(policy);
3239 	if (ret)
3240 		return ret;
3241 
3242 	policy->cpuinfo.transition_latency = INTEL_CPUFREQ_TRANSITION_LATENCY;
3243 	/* This reflects the intel_pstate_get_cpu_pstates() setting. */
3244 	policy->cur = policy->cpuinfo.min_freq;
3245 
3246 	req = kcalloc(2, sizeof(*req), GFP_KERNEL);
3247 	if (!req) {
3248 		ret = -ENOMEM;
3249 		goto pstate_exit;
3250 	}
3251 
3252 	cpu = all_cpu_data[policy->cpu];
3253 
3254 	if (hwp_active) {
3255 		u64 value;
3256 
3257 		policy->transition_delay_us = INTEL_CPUFREQ_TRANSITION_DELAY_HWP;
3258 
3259 		intel_pstate_get_hwp_cap(cpu);
3260 
3261 		rdmsrl_on_cpu(cpu->cpu, MSR_HWP_REQUEST, &value);
3262 		WRITE_ONCE(cpu->hwp_req_cached, value);
3263 
3264 		cpu->epp_cached = intel_pstate_get_epp(cpu, value);
3265 	} else {
3266 		policy->transition_delay_us = INTEL_CPUFREQ_TRANSITION_DELAY;
3267 	}
3268 
3269 	freq = DIV_ROUND_UP(cpu->pstate.turbo_freq * global.min_perf_pct, 100);
3270 
3271 	ret = freq_qos_add_request(&policy->constraints, req, FREQ_QOS_MIN,
3272 				   freq);
3273 	if (ret < 0) {
3274 		dev_err(dev, "Failed to add min-freq constraint (%d)\n", ret);
3275 		goto free_req;
3276 	}
3277 
3278 	freq = DIV_ROUND_UP(cpu->pstate.turbo_freq * global.max_perf_pct, 100);
3279 
3280 	ret = freq_qos_add_request(&policy->constraints, req + 1, FREQ_QOS_MAX,
3281 				   freq);
3282 	if (ret < 0) {
3283 		dev_err(dev, "Failed to add max-freq constraint (%d)\n", ret);
3284 		goto remove_min_req;
3285 	}
3286 
3287 	policy->driver_data = req;
3288 
3289 	return 0;
3290 
3291 remove_min_req:
3292 	freq_qos_remove_request(req);
3293 free_req:
3294 	kfree(req);
3295 pstate_exit:
3296 	intel_pstate_exit_perf_limits(policy);
3297 
3298 	return ret;
3299 }
3300 
3301 static void intel_cpufreq_cpu_exit(struct cpufreq_policy *policy)
3302 {
3303 	struct freq_qos_request *req;
3304 
3305 	req = policy->driver_data;
3306 
3307 	freq_qos_remove_request(req + 1);
3308 	freq_qos_remove_request(req);
3309 	kfree(req);
3310 
3311 	intel_pstate_cpu_exit(policy);
3312 }
3313 
3314 static int intel_cpufreq_suspend(struct cpufreq_policy *policy)
3315 {
3316 	intel_pstate_suspend(policy);
3317 
3318 	if (hwp_active) {
3319 		struct cpudata *cpu = all_cpu_data[policy->cpu];
3320 		u64 value = READ_ONCE(cpu->hwp_req_cached);
3321 
3322 		/*
3323 		 * Clear the desired perf field in MSR_HWP_REQUEST in case
3324 		 * intel_cpufreq_adjust_perf() is in use and the last value
3325 		 * written by it may not be suitable.
3326 		 */
3327 		value &= ~HWP_DESIRED_PERF(~0L);
3328 		wrmsrl_on_cpu(cpu->cpu, MSR_HWP_REQUEST, value);
3329 		WRITE_ONCE(cpu->hwp_req_cached, value);
3330 	}
3331 
3332 	return 0;
3333 }
3334 
3335 static struct cpufreq_driver intel_cpufreq = {
3336 	.flags		= CPUFREQ_CONST_LOOPS,
3337 	.verify		= intel_cpufreq_verify_policy,
3338 	.target		= intel_cpufreq_target,
3339 	.fast_switch	= intel_cpufreq_fast_switch,
3340 	.init		= intel_cpufreq_cpu_init,
3341 	.exit		= intel_cpufreq_cpu_exit,
3342 	.offline	= intel_cpufreq_cpu_offline,
3343 	.online		= intel_pstate_cpu_online,
3344 	.suspend	= intel_cpufreq_suspend,
3345 	.resume		= intel_pstate_resume,
3346 	.update_limits	= intel_pstate_update_limits,
3347 	.name		= "intel_cpufreq",
3348 };
3349 
3350 static struct cpufreq_driver *default_driver;
3351 
3352 static void intel_pstate_driver_cleanup(void)
3353 {
3354 	unsigned int cpu;
3355 
3356 	cpus_read_lock();
3357 	for_each_online_cpu(cpu) {
3358 		if (all_cpu_data[cpu]) {
3359 			if (intel_pstate_driver == &intel_pstate)
3360 				intel_pstate_clear_update_util_hook(cpu);
3361 
3362 			kfree(all_cpu_data[cpu]);
3363 			WRITE_ONCE(all_cpu_data[cpu], NULL);
3364 		}
3365 	}
3366 	cpus_read_unlock();
3367 
3368 	intel_pstate_driver = NULL;
3369 }
3370 
3371 static int intel_pstate_register_driver(struct cpufreq_driver *driver)
3372 {
3373 	bool refresh_cpu_cap_scaling;
3374 	int ret;
3375 
3376 	if (driver == &intel_pstate)
3377 		intel_pstate_sysfs_expose_hwp_dynamic_boost();
3378 
3379 	memset(&global, 0, sizeof(global));
3380 	global.max_perf_pct = 100;
3381 	global.turbo_disabled = turbo_is_disabled();
3382 	global.no_turbo = global.turbo_disabled;
3383 
3384 	arch_set_max_freq_ratio(global.turbo_disabled);
3385 
3386 	refresh_cpu_cap_scaling = hybrid_clear_max_perf_cpu();
3387 
3388 	intel_pstate_driver = driver;
3389 	ret = cpufreq_register_driver(intel_pstate_driver);
3390 	if (ret) {
3391 		intel_pstate_driver_cleanup();
3392 		return ret;
3393 	}
3394 
3395 	global.min_perf_pct = min_perf_pct_min();
3396 
3397 	hybrid_init_cpu_capacity_scaling(refresh_cpu_cap_scaling);
3398 
3399 	return 0;
3400 }
3401 
3402 static ssize_t intel_pstate_show_status(char *buf)
3403 {
3404 	if (!intel_pstate_driver)
3405 		return sprintf(buf, "off\n");
3406 
3407 	return sprintf(buf, "%s\n", intel_pstate_driver == &intel_pstate ?
3408 					"active" : "passive");
3409 }
3410 
3411 static int intel_pstate_update_status(const char *buf, size_t size)
3412 {
3413 	if (size == 3 && !strncmp(buf, "off", size)) {
3414 		if (!intel_pstate_driver)
3415 			return -EINVAL;
3416 
3417 		if (hwp_active)
3418 			return -EBUSY;
3419 
3420 		cpufreq_unregister_driver(intel_pstate_driver);
3421 		intel_pstate_driver_cleanup();
3422 		return 0;
3423 	}
3424 
3425 	if (size == 6 && !strncmp(buf, "active", size)) {
3426 		if (intel_pstate_driver) {
3427 			if (intel_pstate_driver == &intel_pstate)
3428 				return 0;
3429 
3430 			cpufreq_unregister_driver(intel_pstate_driver);
3431 		}
3432 
3433 		return intel_pstate_register_driver(&intel_pstate);
3434 	}
3435 
3436 	if (size == 7 && !strncmp(buf, "passive", size)) {
3437 		if (intel_pstate_driver) {
3438 			if (intel_pstate_driver == &intel_cpufreq)
3439 				return 0;
3440 
3441 			cpufreq_unregister_driver(intel_pstate_driver);
3442 			intel_pstate_sysfs_hide_hwp_dynamic_boost();
3443 		}
3444 
3445 		return intel_pstate_register_driver(&intel_cpufreq);
3446 	}
3447 
3448 	return -EINVAL;
3449 }
3450 
3451 static int no_load __initdata;
3452 static int no_hwp __initdata;
3453 static int hwp_only __initdata;
3454 static unsigned int force_load __initdata;
3455 
3456 static int __init intel_pstate_msrs_not_valid(void)
3457 {
3458 	if (!pstate_funcs.get_max(0) ||
3459 	    !pstate_funcs.get_min(0) ||
3460 	    !pstate_funcs.get_turbo(0))
3461 		return -ENODEV;
3462 
3463 	return 0;
3464 }
3465 
3466 static void __init copy_cpu_funcs(struct pstate_funcs *funcs)
3467 {
3468 	pstate_funcs.get_max   = funcs->get_max;
3469 	pstate_funcs.get_max_physical = funcs->get_max_physical;
3470 	pstate_funcs.get_min   = funcs->get_min;
3471 	pstate_funcs.get_turbo = funcs->get_turbo;
3472 	pstate_funcs.get_scaling = funcs->get_scaling;
3473 	pstate_funcs.get_val   = funcs->get_val;
3474 	pstate_funcs.get_vid   = funcs->get_vid;
3475 	pstate_funcs.get_aperf_mperf_shift = funcs->get_aperf_mperf_shift;
3476 }
3477 
3478 #ifdef CONFIG_ACPI
3479 
3480 static bool __init intel_pstate_no_acpi_pss(void)
3481 {
3482 	int i;
3483 
3484 	for_each_possible_cpu(i) {
3485 		acpi_status status;
3486 		union acpi_object *pss;
3487 		struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
3488 		struct acpi_processor *pr = per_cpu(processors, i);
3489 
3490 		if (!pr)
3491 			continue;
3492 
3493 		status = acpi_evaluate_object(pr->handle, "_PSS", NULL, &buffer);
3494 		if (ACPI_FAILURE(status))
3495 			continue;
3496 
3497 		pss = buffer.pointer;
3498 		if (pss && pss->type == ACPI_TYPE_PACKAGE) {
3499 			kfree(pss);
3500 			return false;
3501 		}
3502 
3503 		kfree(pss);
3504 	}
3505 
3506 	pr_debug("ACPI _PSS not found\n");
3507 	return true;
3508 }
3509 
3510 static bool __init intel_pstate_no_acpi_pcch(void)
3511 {
3512 	acpi_status status;
3513 	acpi_handle handle;
3514 
3515 	status = acpi_get_handle(NULL, "\\_SB", &handle);
3516 	if (ACPI_FAILURE(status))
3517 		goto not_found;
3518 
3519 	if (acpi_has_method(handle, "PCCH"))
3520 		return false;
3521 
3522 not_found:
3523 	pr_debug("ACPI PCCH not found\n");
3524 	return true;
3525 }
3526 
3527 static bool __init intel_pstate_has_acpi_ppc(void)
3528 {
3529 	int i;
3530 
3531 	for_each_possible_cpu(i) {
3532 		struct acpi_processor *pr = per_cpu(processors, i);
3533 
3534 		if (!pr)
3535 			continue;
3536 		if (acpi_has_method(pr->handle, "_PPC"))
3537 			return true;
3538 	}
3539 	pr_debug("ACPI _PPC not found\n");
3540 	return false;
3541 }
3542 
3543 enum {
3544 	PSS,
3545 	PPC,
3546 };
3547 
3548 /* Hardware vendor-specific info that has its own power management modes */
3549 static struct acpi_platform_list plat_info[] __initdata = {
3550 	{"HP    ", "ProLiant", 0, ACPI_SIG_FADT, all_versions, NULL, PSS},
3551 	{"ORACLE", "X4-2    ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3552 	{"ORACLE", "X4-2L   ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3553 	{"ORACLE", "X4-2B   ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3554 	{"ORACLE", "X3-2    ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3555 	{"ORACLE", "X3-2L   ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3556 	{"ORACLE", "X3-2B   ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3557 	{"ORACLE", "X4470M2 ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3558 	{"ORACLE", "X4270M3 ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3559 	{"ORACLE", "X4270M2 ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3560 	{"ORACLE", "X4170M2 ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3561 	{"ORACLE", "X4170 M3", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3562 	{"ORACLE", "X4275 M3", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3563 	{"ORACLE", "X6-2    ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3564 	{"ORACLE", "Sudbury ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3565 	{ } /* End */
3566 };
3567 
3568 #define BITMASK_OOB	(BIT(8) | BIT(18))
3569 
3570 static bool __init intel_pstate_platform_pwr_mgmt_exists(void)
3571 {
3572 	const struct x86_cpu_id *id;
3573 	u64 misc_pwr;
3574 	int idx;
3575 
3576 	id = x86_match_cpu(intel_pstate_cpu_oob_ids);
3577 	if (id) {
3578 		rdmsrl(MSR_MISC_PWR_MGMT, misc_pwr);
3579 		if (misc_pwr & BITMASK_OOB) {
3580 			pr_debug("Bit 8 or 18 in the MISC_PWR_MGMT MSR set\n");
3581 			pr_debug("P states are controlled in Out of Band mode by the firmware/hardware\n");
3582 			return true;
3583 		}
3584 	}
3585 
3586 	idx = acpi_match_platform_list(plat_info);
3587 	if (idx < 0)
3588 		return false;
3589 
3590 	switch (plat_info[idx].data) {
3591 	case PSS:
3592 		if (!intel_pstate_no_acpi_pss())
3593 			return false;
3594 
3595 		return intel_pstate_no_acpi_pcch();
3596 	case PPC:
3597 		return intel_pstate_has_acpi_ppc() && !force_load;
3598 	}
3599 
3600 	return false;
3601 }
3602 
3603 static void intel_pstate_request_control_from_smm(void)
3604 {
3605 	/*
3606 	 * It may be unsafe to request P-states control from SMM if _PPC support
3607 	 * has not been enabled.
3608 	 */
3609 	if (acpi_ppc)
3610 		acpi_processor_pstate_control();
3611 }
3612 #else /* CONFIG_ACPI not enabled */
3613 static inline bool intel_pstate_platform_pwr_mgmt_exists(void) { return false; }
3614 static inline bool intel_pstate_has_acpi_ppc(void) { return false; }
3615 static inline void intel_pstate_request_control_from_smm(void) {}
3616 #endif /* CONFIG_ACPI */
3617 
3618 #define INTEL_PSTATE_HWP_BROADWELL	0x01
3619 
3620 #define X86_MATCH_HWP(vfm, hwp_mode)				\
3621 	X86_MATCH_VFM_FEATURE(vfm, X86_FEATURE_HWP, hwp_mode)
3622 
3623 static const struct x86_cpu_id hwp_support_ids[] __initconst = {
3624 	X86_MATCH_HWP(INTEL_BROADWELL_X,	INTEL_PSTATE_HWP_BROADWELL),
3625 	X86_MATCH_HWP(INTEL_BROADWELL_D,	INTEL_PSTATE_HWP_BROADWELL),
3626 	X86_MATCH_HWP(INTEL_ANY,		0),
3627 	{}
3628 };
3629 
3630 static bool intel_pstate_hwp_is_enabled(void)
3631 {
3632 	u64 value;
3633 
3634 	rdmsrl(MSR_PM_ENABLE, value);
3635 	return !!(value & 0x1);
3636 }
3637 
3638 #define POWERSAVE_MASK			GENMASK(7, 0)
3639 #define BALANCE_POWER_MASK		GENMASK(15, 8)
3640 #define BALANCE_PERFORMANCE_MASK	GENMASK(23, 16)
3641 #define PERFORMANCE_MASK		GENMASK(31, 24)
3642 
3643 #define HWP_SET_EPP_VALUES(powersave, balance_power, balance_perf, performance) \
3644 	(FIELD_PREP_CONST(POWERSAVE_MASK, powersave) |\
3645 	 FIELD_PREP_CONST(BALANCE_POWER_MASK, balance_power) |\
3646 	 FIELD_PREP_CONST(BALANCE_PERFORMANCE_MASK, balance_perf) |\
3647 	 FIELD_PREP_CONST(PERFORMANCE_MASK, performance))
3648 
3649 #define HWP_SET_DEF_BALANCE_PERF_EPP(balance_perf) \
3650 	(HWP_SET_EPP_VALUES(HWP_EPP_POWERSAVE, HWP_EPP_BALANCE_POWERSAVE,\
3651 	 balance_perf, HWP_EPP_PERFORMANCE))
3652 
3653 static const struct x86_cpu_id intel_epp_default[] = {
3654 	/*
3655 	 * Set EPP value as 102, this is the max suggested EPP
3656 	 * which can result in one core turbo frequency for
3657 	 * AlderLake Mobile CPUs.
3658 	 */
3659 	X86_MATCH_VFM(INTEL_ALDERLAKE_L, HWP_SET_DEF_BALANCE_PERF_EPP(102)),
3660 	X86_MATCH_VFM(INTEL_SAPPHIRERAPIDS_X, HWP_SET_DEF_BALANCE_PERF_EPP(32)),
3661 	X86_MATCH_VFM(INTEL_EMERALDRAPIDS_X, HWP_SET_DEF_BALANCE_PERF_EPP(32)),
3662 	X86_MATCH_VFM(INTEL_GRANITERAPIDS_X, HWP_SET_DEF_BALANCE_PERF_EPP(32)),
3663 	X86_MATCH_VFM(INTEL_GRANITERAPIDS_D, HWP_SET_DEF_BALANCE_PERF_EPP(32)),
3664 	X86_MATCH_VFM(INTEL_METEORLAKE_L, HWP_SET_EPP_VALUES(HWP_EPP_POWERSAVE,
3665 		      179, 64, 16)),
3666 	X86_MATCH_VFM(INTEL_ARROWLAKE, HWP_SET_EPP_VALUES(HWP_EPP_POWERSAVE,
3667 		      179, 64, 16)),
3668 	{}
3669 };
3670 
3671 static const struct x86_cpu_id intel_hybrid_scaling_factor[] = {
3672 	X86_MATCH_VFM(INTEL_ALDERLAKE, HYBRID_SCALING_FACTOR_ADL),
3673 	X86_MATCH_VFM(INTEL_ALDERLAKE_L, HYBRID_SCALING_FACTOR_ADL),
3674 	X86_MATCH_VFM(INTEL_RAPTORLAKE, HYBRID_SCALING_FACTOR_ADL),
3675 	X86_MATCH_VFM(INTEL_RAPTORLAKE_P, HYBRID_SCALING_FACTOR_ADL),
3676 	X86_MATCH_VFM(INTEL_RAPTORLAKE_S, HYBRID_SCALING_FACTOR_ADL),
3677 	X86_MATCH_VFM(INTEL_METEORLAKE_L, HYBRID_SCALING_FACTOR_MTL),
3678 	X86_MATCH_VFM(INTEL_LUNARLAKE_M, HYBRID_SCALING_FACTOR_LNL),
3679 	{}
3680 };
3681 
3682 static int __init intel_pstate_init(void)
3683 {
3684 	static struct cpudata **_all_cpu_data;
3685 	const struct x86_cpu_id *id;
3686 	int rc;
3687 
3688 	if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL)
3689 		return -ENODEV;
3690 
3691 	id = x86_match_cpu(hwp_support_ids);
3692 	if (id) {
3693 		hwp_forced = intel_pstate_hwp_is_enabled();
3694 
3695 		if (hwp_forced)
3696 			pr_info("HWP enabled by BIOS\n");
3697 		else if (no_load)
3698 			return -ENODEV;
3699 
3700 		copy_cpu_funcs(&core_funcs);
3701 		/*
3702 		 * Avoid enabling HWP for processors without EPP support,
3703 		 * because that means incomplete HWP implementation which is a
3704 		 * corner case and supporting it is generally problematic.
3705 		 *
3706 		 * If HWP is enabled already, though, there is no choice but to
3707 		 * deal with it.
3708 		 */
3709 		if ((!no_hwp && boot_cpu_has(X86_FEATURE_HWP_EPP)) || hwp_forced) {
3710 			hwp_active = true;
3711 			hwp_mode_bdw = id->driver_data;
3712 			intel_pstate.attr = hwp_cpufreq_attrs;
3713 			intel_cpufreq.attr = hwp_cpufreq_attrs;
3714 			intel_cpufreq.flags |= CPUFREQ_NEED_UPDATE_LIMITS;
3715 			intel_cpufreq.adjust_perf = intel_cpufreq_adjust_perf;
3716 			if (!default_driver)
3717 				default_driver = &intel_pstate;
3718 
3719 			pstate_funcs.get_cpu_scaling = hwp_get_cpu_scaling;
3720 
3721 			goto hwp_cpu_matched;
3722 		}
3723 		pr_info("HWP not enabled\n");
3724 	} else {
3725 		if (no_load)
3726 			return -ENODEV;
3727 
3728 		id = x86_match_cpu(intel_pstate_cpu_ids);
3729 		if (!id) {
3730 			pr_info("CPU model not supported\n");
3731 			return -ENODEV;
3732 		}
3733 
3734 		copy_cpu_funcs((struct pstate_funcs *)id->driver_data);
3735 	}
3736 
3737 	if (intel_pstate_msrs_not_valid()) {
3738 		pr_info("Invalid MSRs\n");
3739 		return -ENODEV;
3740 	}
3741 	/* Without HWP start in the passive mode. */
3742 	if (!default_driver)
3743 		default_driver = &intel_cpufreq;
3744 
3745 hwp_cpu_matched:
3746 	/*
3747 	 * The Intel pstate driver will be ignored if the platform
3748 	 * firmware has its own power management modes.
3749 	 */
3750 	if (intel_pstate_platform_pwr_mgmt_exists()) {
3751 		pr_info("P-states controlled by the platform\n");
3752 		return -ENODEV;
3753 	}
3754 
3755 	if (!hwp_active && hwp_only)
3756 		return -ENOTSUPP;
3757 
3758 	pr_info("Intel P-state driver initializing\n");
3759 
3760 	_all_cpu_data = vzalloc(array_size(sizeof(void *), num_possible_cpus()));
3761 	if (!_all_cpu_data)
3762 		return -ENOMEM;
3763 
3764 	WRITE_ONCE(all_cpu_data, _all_cpu_data);
3765 
3766 	intel_pstate_request_control_from_smm();
3767 
3768 	intel_pstate_sysfs_expose_params();
3769 
3770 	if (hwp_active) {
3771 		const struct x86_cpu_id *id = x86_match_cpu(intel_epp_default);
3772 		const struct x86_cpu_id *hybrid_id = x86_match_cpu(intel_hybrid_scaling_factor);
3773 
3774 		if (id) {
3775 			epp_values[EPP_INDEX_POWERSAVE] =
3776 					FIELD_GET(POWERSAVE_MASK, id->driver_data);
3777 			epp_values[EPP_INDEX_BALANCE_POWERSAVE] =
3778 					FIELD_GET(BALANCE_POWER_MASK, id->driver_data);
3779 			epp_values[EPP_INDEX_BALANCE_PERFORMANCE] =
3780 					FIELD_GET(BALANCE_PERFORMANCE_MASK, id->driver_data);
3781 			epp_values[EPP_INDEX_PERFORMANCE] =
3782 					FIELD_GET(PERFORMANCE_MASK, id->driver_data);
3783 			pr_debug("Updated EPPs powersave:%x balanced power:%x balanced perf:%x performance:%x\n",
3784 				 epp_values[EPP_INDEX_POWERSAVE],
3785 				 epp_values[EPP_INDEX_BALANCE_POWERSAVE],
3786 				 epp_values[EPP_INDEX_BALANCE_PERFORMANCE],
3787 				 epp_values[EPP_INDEX_PERFORMANCE]);
3788 		}
3789 
3790 		if (hybrid_id) {
3791 			hybrid_scaling_factor = hybrid_id->driver_data;
3792 			pr_debug("hybrid scaling factor: %d\n", hybrid_scaling_factor);
3793 		}
3794 
3795 	}
3796 
3797 	mutex_lock(&intel_pstate_driver_lock);
3798 	rc = intel_pstate_register_driver(default_driver);
3799 	mutex_unlock(&intel_pstate_driver_lock);
3800 	if (rc) {
3801 		intel_pstate_sysfs_remove();
3802 		return rc;
3803 	}
3804 
3805 	if (hwp_active) {
3806 		const struct x86_cpu_id *id;
3807 
3808 		id = x86_match_cpu(intel_pstate_cpu_ee_disable_ids);
3809 		if (id) {
3810 			set_power_ctl_ee_state(false);
3811 			pr_info("Disabling energy efficiency optimization\n");
3812 		}
3813 
3814 		pr_info("HWP enabled\n");
3815 	} else if (boot_cpu_has(X86_FEATURE_HYBRID_CPU)) {
3816 		pr_warn("Problematic setup: Hybrid processor with disabled HWP\n");
3817 	}
3818 
3819 	return 0;
3820 }
3821 device_initcall(intel_pstate_init);
3822 
3823 static int __init intel_pstate_setup(char *str)
3824 {
3825 	if (!str)
3826 		return -EINVAL;
3827 
3828 	if (!strcmp(str, "disable"))
3829 		no_load = 1;
3830 	else if (!strcmp(str, "active"))
3831 		default_driver = &intel_pstate;
3832 	else if (!strcmp(str, "passive"))
3833 		default_driver = &intel_cpufreq;
3834 
3835 	if (!strcmp(str, "no_hwp"))
3836 		no_hwp = 1;
3837 
3838 	if (!strcmp(str, "force"))
3839 		force_load = 1;
3840 	if (!strcmp(str, "hwp_only"))
3841 		hwp_only = 1;
3842 	if (!strcmp(str, "per_cpu_perf_limits"))
3843 		per_cpu_limits = true;
3844 
3845 #ifdef CONFIG_ACPI
3846 	if (!strcmp(str, "support_acpi_ppc"))
3847 		acpi_ppc = true;
3848 #endif
3849 
3850 	return 0;
3851 }
3852 early_param("intel_pstate", intel_pstate_setup);
3853 
3854 MODULE_AUTHOR("Dirk Brandewie <dirk.j.brandewie@intel.com>");
3855 MODULE_DESCRIPTION("'intel_pstate' - P state driver Intel Core processors");
3856