xref: /linux/drivers/cpufreq/intel_pstate.c (revision 07fdad3a93756b872da7b53647715c48d0f4a2d0)
1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * intel_pstate.c: Native P state management for Intel processors
4  *
5  * (C) Copyright 2012 Intel Corporation
6  * Author: Dirk Brandewie <dirk.j.brandewie@intel.com>
7  */
8 
9 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
10 
11 #include <linux/kernel.h>
12 #include <linux/kernel_stat.h>
13 #include <linux/module.h>
14 #include <linux/ktime.h>
15 #include <linux/hrtimer.h>
16 #include <linux/tick.h>
17 #include <linux/slab.h>
18 #include <linux/sched/cpufreq.h>
19 #include <linux/sched/smt.h>
20 #include <linux/list.h>
21 #include <linux/cpu.h>
22 #include <linux/cpufreq.h>
23 #include <linux/sysfs.h>
24 #include <linux/types.h>
25 #include <linux/fs.h>
26 #include <linux/acpi.h>
27 #include <linux/vmalloc.h>
28 #include <linux/pm_qos.h>
29 #include <linux/bitfield.h>
30 #include <trace/events/power.h>
31 #include <linux/units.h>
32 
33 #include <asm/cpu.h>
34 #include <asm/div64.h>
35 #include <asm/msr.h>
36 #include <asm/cpu_device_id.h>
37 #include <asm/cpufeature.h>
38 #include <asm/intel-family.h>
39 #include "../drivers/thermal/intel/thermal_interrupt.h"
40 
41 #define INTEL_PSTATE_SAMPLING_INTERVAL	(10 * NSEC_PER_MSEC)
42 
43 #define INTEL_CPUFREQ_TRANSITION_LATENCY	20000
44 #define INTEL_CPUFREQ_TRANSITION_DELAY_HWP	5000
45 #define INTEL_CPUFREQ_TRANSITION_DELAY		500
46 
47 #ifdef CONFIG_ACPI
48 #include <acpi/processor.h>
49 #include <acpi/cppc_acpi.h>
50 #endif
51 
52 #define FRAC_BITS 8
53 #define int_tofp(X) ((int64_t)(X) << FRAC_BITS)
54 #define fp_toint(X) ((X) >> FRAC_BITS)
55 
56 #define ONE_EIGHTH_FP ((int64_t)1 << (FRAC_BITS - 3))
57 
58 #define EXT_BITS 6
59 #define EXT_FRAC_BITS (EXT_BITS + FRAC_BITS)
60 #define fp_ext_toint(X) ((X) >> EXT_FRAC_BITS)
61 #define int_ext_tofp(X) ((int64_t)(X) << EXT_FRAC_BITS)
62 
63 static inline int32_t mul_fp(int32_t x, int32_t y)
64 {
65 	return ((int64_t)x * (int64_t)y) >> FRAC_BITS;
66 }
67 
68 static inline int32_t div_fp(s64 x, s64 y)
69 {
70 	return div64_s64((int64_t)x << FRAC_BITS, y);
71 }
72 
73 static inline int ceiling_fp(int32_t x)
74 {
75 	int mask, ret;
76 
77 	ret = fp_toint(x);
78 	mask = (1 << FRAC_BITS) - 1;
79 	if (x & mask)
80 		ret += 1;
81 	return ret;
82 }
83 
84 static inline u64 mul_ext_fp(u64 x, u64 y)
85 {
86 	return (x * y) >> EXT_FRAC_BITS;
87 }
88 
89 static inline u64 div_ext_fp(u64 x, u64 y)
90 {
91 	return div64_u64(x << EXT_FRAC_BITS, y);
92 }
93 
94 /**
95  * struct sample -	Store performance sample
96  * @core_avg_perf:	Ratio of APERF/MPERF which is the actual average
97  *			performance during last sample period
98  * @busy_scaled:	Scaled busy value which is used to calculate next
99  *			P state. This can be different than core_avg_perf
100  *			to account for cpu idle period
101  * @aperf:		Difference of actual performance frequency clock count
102  *			read from APERF MSR between last and current sample
103  * @mperf:		Difference of maximum performance frequency clock count
104  *			read from MPERF MSR between last and current sample
105  * @tsc:		Difference of time stamp counter between last and
106  *			current sample
107  * @time:		Current time from scheduler
108  *
109  * This structure is used in the cpudata structure to store performance sample
110  * data for choosing next P State.
111  */
112 struct sample {
113 	int32_t core_avg_perf;
114 	int32_t busy_scaled;
115 	u64 aperf;
116 	u64 mperf;
117 	u64 tsc;
118 	u64 time;
119 };
120 
121 /**
122  * struct pstate_data - Store P state data
123  * @current_pstate:	Current requested P state
124  * @min_pstate:		Min P state possible for this platform
125  * @max_pstate:		Max P state possible for this platform
126  * @max_pstate_physical:This is physical Max P state for a processor
127  *			This can be higher than the max_pstate which can
128  *			be limited by platform thermal design power limits
129  * @perf_ctl_scaling:	PERF_CTL P-state to frequency scaling factor
130  * @scaling:		Scaling factor between performance and frequency
131  * @turbo_pstate:	Max Turbo P state possible for this platform
132  * @min_freq:		@min_pstate frequency in cpufreq units
133  * @max_freq:		@max_pstate frequency in cpufreq units
134  * @turbo_freq:		@turbo_pstate frequency in cpufreq units
135  *
136  * Stores the per cpu model P state limits and current P state.
137  */
138 struct pstate_data {
139 	int	current_pstate;
140 	int	min_pstate;
141 	int	max_pstate;
142 	int	max_pstate_physical;
143 	int	perf_ctl_scaling;
144 	int	scaling;
145 	int	turbo_pstate;
146 	unsigned int min_freq;
147 	unsigned int max_freq;
148 	unsigned int turbo_freq;
149 };
150 
151 /**
152  * struct vid_data -	Stores voltage information data
153  * @min:		VID data for this platform corresponding to
154  *			the lowest P state
155  * @max:		VID data corresponding to the highest P State.
156  * @turbo:		VID data for turbo P state
157  * @ratio:		Ratio of (vid max - vid min) /
158  *			(max P state - Min P State)
159  *
160  * Stores the voltage data for DVFS (Dynamic Voltage and Frequency Scaling)
161  * This data is used in Atom platforms, where in addition to target P state,
162  * the voltage data needs to be specified to select next P State.
163  */
164 struct vid_data {
165 	int min;
166 	int max;
167 	int turbo;
168 	int32_t ratio;
169 };
170 
171 /**
172  * struct global_params - Global parameters, mostly tunable via sysfs.
173  * @no_turbo:		Whether or not to use turbo P-states.
174  * @turbo_disabled:	Whether or not turbo P-states are available at all,
175  *			based on the MSR_IA32_MISC_ENABLE value and whether or
176  *			not the maximum reported turbo P-state is different from
177  *			the maximum reported non-turbo one.
178  * @min_perf_pct:	Minimum capacity limit in percent of the maximum turbo
179  *			P-state capacity.
180  * @max_perf_pct:	Maximum capacity limit in percent of the maximum turbo
181  *			P-state capacity.
182  */
183 struct global_params {
184 	bool no_turbo;
185 	bool turbo_disabled;
186 	int max_perf_pct;
187 	int min_perf_pct;
188 };
189 
190 /**
191  * struct cpudata -	Per CPU instance data storage
192  * @cpu:		CPU number for this instance data
193  * @policy:		CPUFreq policy value
194  * @update_util:	CPUFreq utility callback information
195  * @update_util_set:	CPUFreq utility callback is set
196  * @iowait_boost:	iowait-related boost fraction
197  * @last_update:	Time of the last update.
198  * @pstate:		Stores P state limits for this CPU
199  * @vid:		Stores VID limits for this CPU
200  * @last_sample_time:	Last Sample time
201  * @aperf_mperf_shift:	APERF vs MPERF counting frequency difference
202  * @prev_aperf:		Last APERF value read from APERF MSR
203  * @prev_mperf:		Last MPERF value read from MPERF MSR
204  * @prev_tsc:		Last timestamp counter (TSC) value
205  * @sample:		Storage for storing last Sample data
206  * @min_perf_ratio:	Minimum capacity in terms of PERF or HWP ratios
207  * @max_perf_ratio:	Maximum capacity in terms of PERF or HWP ratios
208  * @acpi_perf_data:	Stores ACPI perf information read from _PSS
209  * @valid_pss_table:	Set to true for valid ACPI _PSS entries found
210  * @epp_powersave:	Last saved HWP energy performance preference
211  *			(EPP) or energy performance bias (EPB),
212  *			when policy switched to performance
213  * @epp_policy:		Last saved policy used to set EPP/EPB
214  * @epp_default:	Power on default HWP energy performance
215  *			preference/bias
216  * @epp_cached:		Cached HWP energy-performance preference value
217  * @hwp_req_cached:	Cached value of the last HWP Request MSR
218  * @hwp_cap_cached:	Cached value of the last HWP Capabilities MSR
219  * @last_io_update:	Last time when IO wake flag was set
220  * @capacity_perf:	Highest perf used for scale invariance
221  * @sched_flags:	Store scheduler flags for possible cross CPU update
222  * @hwp_boost_min:	Last HWP boosted min performance
223  * @suspended:		Whether or not the driver has been suspended.
224  * @pd_registered:	Set when a perf domain is registered for this CPU.
225  * @hwp_notify_work:	workqueue for HWP notifications.
226  *
227  * This structure stores per CPU instance data for all CPUs.
228  */
229 struct cpudata {
230 	int cpu;
231 
232 	unsigned int policy;
233 	struct update_util_data update_util;
234 	bool   update_util_set;
235 
236 	struct pstate_data pstate;
237 	struct vid_data vid;
238 
239 	u64	last_update;
240 	u64	last_sample_time;
241 	u64	aperf_mperf_shift;
242 	u64	prev_aperf;
243 	u64	prev_mperf;
244 	u64	prev_tsc;
245 	struct sample sample;
246 	int32_t	min_perf_ratio;
247 	int32_t	max_perf_ratio;
248 #ifdef CONFIG_ACPI
249 	struct acpi_processor_performance acpi_perf_data;
250 	bool valid_pss_table;
251 #endif
252 	unsigned int iowait_boost;
253 	s16 epp_powersave;
254 	s16 epp_policy;
255 	s16 epp_default;
256 	s16 epp_cached;
257 	u64 hwp_req_cached;
258 	u64 hwp_cap_cached;
259 	u64 last_io_update;
260 	unsigned int capacity_perf;
261 	unsigned int sched_flags;
262 	u32 hwp_boost_min;
263 	bool suspended;
264 #ifdef CONFIG_ENERGY_MODEL
265 	bool pd_registered;
266 #endif
267 	struct delayed_work hwp_notify_work;
268 };
269 
270 static struct cpudata **all_cpu_data;
271 
272 /**
273  * struct pstate_funcs - Per CPU model specific callbacks
274  * @get_max:		Callback to get maximum non turbo effective P state
275  * @get_max_physical:	Callback to get maximum non turbo physical P state
276  * @get_min:		Callback to get minimum P state
277  * @get_turbo:		Callback to get turbo P state
278  * @get_scaling:	Callback to get frequency scaling factor
279  * @get_cpu_scaling:	Get frequency scaling factor for a given cpu
280  * @get_aperf_mperf_shift: Callback to get the APERF vs MPERF frequency difference
281  * @get_val:		Callback to convert P state to actual MSR write value
282  * @get_vid:		Callback to get VID data for Atom platforms
283  *
284  * Core and Atom CPU models have different way to get P State limits. This
285  * structure is used to store those callbacks.
286  */
287 struct pstate_funcs {
288 	int (*get_max)(int cpu);
289 	int (*get_max_physical)(int cpu);
290 	int (*get_min)(int cpu);
291 	int (*get_turbo)(int cpu);
292 	int (*get_scaling)(void);
293 	int (*get_cpu_scaling)(int cpu);
294 	int (*get_aperf_mperf_shift)(void);
295 	u64 (*get_val)(struct cpudata*, int pstate);
296 	void (*get_vid)(struct cpudata *);
297 };
298 
299 static struct pstate_funcs pstate_funcs __read_mostly;
300 
301 static bool hwp_active __ro_after_init;
302 static int hwp_mode_bdw __ro_after_init;
303 static bool per_cpu_limits __ro_after_init;
304 static bool hwp_forced __ro_after_init;
305 static bool hwp_boost __read_mostly;
306 static bool hwp_is_hybrid;
307 
308 static struct cpufreq_driver *intel_pstate_driver __read_mostly;
309 
310 #define INTEL_PSTATE_CORE_SCALING	100000
311 #define HYBRID_SCALING_FACTOR_ADL	78741
312 #define HYBRID_SCALING_FACTOR_MTL	80000
313 #define HYBRID_SCALING_FACTOR_LNL	86957
314 
315 static int hybrid_scaling_factor;
316 
317 static inline int core_get_scaling(void)
318 {
319 	return INTEL_PSTATE_CORE_SCALING;
320 }
321 
322 #ifdef CONFIG_ACPI
323 static bool acpi_ppc;
324 #endif
325 
326 static struct global_params global;
327 
328 static DEFINE_MUTEX(intel_pstate_driver_lock);
329 static DEFINE_MUTEX(intel_pstate_limits_lock);
330 
331 #ifdef CONFIG_ACPI
332 
333 static bool intel_pstate_acpi_pm_profile_server(void)
334 {
335 	if (acpi_gbl_FADT.preferred_profile == PM_ENTERPRISE_SERVER ||
336 	    acpi_gbl_FADT.preferred_profile == PM_PERFORMANCE_SERVER)
337 		return true;
338 
339 	return false;
340 }
341 
342 static bool intel_pstate_get_ppc_enable_status(void)
343 {
344 	if (intel_pstate_acpi_pm_profile_server())
345 		return true;
346 
347 	return acpi_ppc;
348 }
349 
350 #ifdef CONFIG_ACPI_CPPC_LIB
351 
352 /* The work item is needed to avoid CPU hotplug locking issues */
353 static void intel_pstste_sched_itmt_work_fn(struct work_struct *work)
354 {
355 	sched_set_itmt_support();
356 }
357 
358 static DECLARE_WORK(sched_itmt_work, intel_pstste_sched_itmt_work_fn);
359 
360 #define CPPC_MAX_PERF	U8_MAX
361 
362 static void intel_pstate_set_itmt_prio(int cpu)
363 {
364 	struct cppc_perf_caps cppc_perf;
365 	static u32 max_highest_perf = 0, min_highest_perf = U32_MAX;
366 	int ret;
367 
368 	ret = cppc_get_perf_caps(cpu, &cppc_perf);
369 	/*
370 	 * If CPPC is not available, fall back to MSR_HWP_CAPABILITIES bits [8:0].
371 	 *
372 	 * Also, on some systems with overclocking enabled, CPPC.highest_perf is
373 	 * hardcoded to 0xff, so CPPC.highest_perf cannot be used to enable ITMT.
374 	 * Fall back to MSR_HWP_CAPABILITIES then too.
375 	 */
376 	if (ret || cppc_perf.highest_perf == CPPC_MAX_PERF)
377 		cppc_perf.highest_perf = HWP_HIGHEST_PERF(READ_ONCE(all_cpu_data[cpu]->hwp_cap_cached));
378 
379 	/*
380 	 * The priorities can be set regardless of whether or not
381 	 * sched_set_itmt_support(true) has been called and it is valid to
382 	 * update them at any time after it has been called.
383 	 */
384 	sched_set_itmt_core_prio(cppc_perf.highest_perf, cpu);
385 
386 	if (max_highest_perf <= min_highest_perf) {
387 		if (cppc_perf.highest_perf > max_highest_perf)
388 			max_highest_perf = cppc_perf.highest_perf;
389 
390 		if (cppc_perf.highest_perf < min_highest_perf)
391 			min_highest_perf = cppc_perf.highest_perf;
392 
393 		if (max_highest_perf > min_highest_perf) {
394 			/*
395 			 * This code can be run during CPU online under the
396 			 * CPU hotplug locks, so sched_set_itmt_support()
397 			 * cannot be called from here.  Queue up a work item
398 			 * to invoke it.
399 			 */
400 			schedule_work(&sched_itmt_work);
401 		}
402 	}
403 }
404 
405 static int intel_pstate_get_cppc_guaranteed(int cpu)
406 {
407 	struct cppc_perf_caps cppc_perf;
408 	int ret;
409 
410 	ret = cppc_get_perf_caps(cpu, &cppc_perf);
411 	if (ret)
412 		return ret;
413 
414 	if (cppc_perf.guaranteed_perf)
415 		return cppc_perf.guaranteed_perf;
416 
417 	return cppc_perf.nominal_perf;
418 }
419 
420 static int intel_pstate_cppc_get_scaling(int cpu)
421 {
422 	struct cppc_perf_caps cppc_perf;
423 
424 	/*
425 	 * Compute the perf-to-frequency scaling factor for the given CPU if
426 	 * possible, unless it would be 0.
427 	 */
428 	if (!cppc_get_perf_caps(cpu, &cppc_perf) &&
429 	    cppc_perf.nominal_perf && cppc_perf.nominal_freq)
430 		return div_u64(cppc_perf.nominal_freq * KHZ_PER_MHZ,
431 			       cppc_perf.nominal_perf);
432 
433 	return core_get_scaling();
434 }
435 
436 #else /* CONFIG_ACPI_CPPC_LIB */
437 static inline void intel_pstate_set_itmt_prio(int cpu)
438 {
439 }
440 #endif /* CONFIG_ACPI_CPPC_LIB */
441 
442 static void intel_pstate_init_acpi_perf_limits(struct cpufreq_policy *policy)
443 {
444 	struct cpudata *cpu;
445 	int ret;
446 	int i;
447 
448 	if (hwp_active) {
449 		intel_pstate_set_itmt_prio(policy->cpu);
450 		return;
451 	}
452 
453 	if (!intel_pstate_get_ppc_enable_status())
454 		return;
455 
456 	cpu = all_cpu_data[policy->cpu];
457 
458 	ret = acpi_processor_register_performance(&cpu->acpi_perf_data,
459 						  policy->cpu);
460 	if (ret)
461 		return;
462 
463 	/*
464 	 * Check if the control value in _PSS is for PERF_CTL MSR, which should
465 	 * guarantee that the states returned by it map to the states in our
466 	 * list directly.
467 	 */
468 	if (cpu->acpi_perf_data.control_register.space_id !=
469 						ACPI_ADR_SPACE_FIXED_HARDWARE)
470 		goto err;
471 
472 	/*
473 	 * If there is only one entry _PSS, simply ignore _PSS and continue as
474 	 * usual without taking _PSS into account
475 	 */
476 	if (cpu->acpi_perf_data.state_count < 2)
477 		goto err;
478 
479 	pr_debug("CPU%u - ACPI _PSS perf data\n", policy->cpu);
480 	for (i = 0; i < cpu->acpi_perf_data.state_count; i++) {
481 		pr_debug("     %cP%d: %u MHz, %u mW, 0x%x\n",
482 			 (i == cpu->acpi_perf_data.state ? '*' : ' '), i,
483 			 (u32) cpu->acpi_perf_data.states[i].core_frequency,
484 			 (u32) cpu->acpi_perf_data.states[i].power,
485 			 (u32) cpu->acpi_perf_data.states[i].control);
486 	}
487 
488 	cpu->valid_pss_table = true;
489 	pr_debug("_PPC limits will be enforced\n");
490 
491 	return;
492 
493  err:
494 	cpu->valid_pss_table = false;
495 	acpi_processor_unregister_performance(policy->cpu);
496 }
497 
498 static void intel_pstate_exit_perf_limits(struct cpufreq_policy *policy)
499 {
500 	struct cpudata *cpu;
501 
502 	cpu = all_cpu_data[policy->cpu];
503 	if (!cpu->valid_pss_table)
504 		return;
505 
506 	acpi_processor_unregister_performance(policy->cpu);
507 }
508 #else /* CONFIG_ACPI */
509 static inline void intel_pstate_init_acpi_perf_limits(struct cpufreq_policy *policy)
510 {
511 }
512 
513 static inline void intel_pstate_exit_perf_limits(struct cpufreq_policy *policy)
514 {
515 }
516 
517 static inline bool intel_pstate_acpi_pm_profile_server(void)
518 {
519 	return false;
520 }
521 #endif /* CONFIG_ACPI */
522 
523 #ifndef CONFIG_ACPI_CPPC_LIB
524 static inline int intel_pstate_get_cppc_guaranteed(int cpu)
525 {
526 	return -ENOTSUPP;
527 }
528 
529 static int intel_pstate_cppc_get_scaling(int cpu)
530 {
531 	return core_get_scaling();
532 }
533 #endif /* CONFIG_ACPI_CPPC_LIB */
534 
535 static int intel_pstate_freq_to_hwp_rel(struct cpudata *cpu, int freq,
536 					unsigned int relation)
537 {
538 	if (freq == cpu->pstate.turbo_freq)
539 		return cpu->pstate.turbo_pstate;
540 
541 	if (freq == cpu->pstate.max_freq)
542 		return cpu->pstate.max_pstate;
543 
544 	switch (relation) {
545 	case CPUFREQ_RELATION_H:
546 		return freq / cpu->pstate.scaling;
547 	case CPUFREQ_RELATION_C:
548 		return DIV_ROUND_CLOSEST(freq, cpu->pstate.scaling);
549 	}
550 
551 	return DIV_ROUND_UP(freq, cpu->pstate.scaling);
552 }
553 
554 static int intel_pstate_freq_to_hwp(struct cpudata *cpu, int freq)
555 {
556 	return intel_pstate_freq_to_hwp_rel(cpu, freq, CPUFREQ_RELATION_L);
557 }
558 
559 /**
560  * intel_pstate_hybrid_hwp_adjust - Calibrate HWP performance levels.
561  * @cpu: Target CPU.
562  *
563  * On hybrid processors, HWP may expose more performance levels than there are
564  * P-states accessible through the PERF_CTL interface.  If that happens, the
565  * scaling factor between HWP performance levels and CPU frequency will be less
566  * than the scaling factor between P-state values and CPU frequency.
567  *
568  * In that case, adjust the CPU parameters used in computations accordingly.
569  */
570 static void intel_pstate_hybrid_hwp_adjust(struct cpudata *cpu)
571 {
572 	int perf_ctl_max_phys = cpu->pstate.max_pstate_physical;
573 	int perf_ctl_scaling = cpu->pstate.perf_ctl_scaling;
574 	int perf_ctl_turbo = pstate_funcs.get_turbo(cpu->cpu);
575 	int scaling = cpu->pstate.scaling;
576 	int freq;
577 
578 	pr_debug("CPU%d: perf_ctl_max_phys = %d\n", cpu->cpu, perf_ctl_max_phys);
579 	pr_debug("CPU%d: perf_ctl_turbo = %d\n", cpu->cpu, perf_ctl_turbo);
580 	pr_debug("CPU%d: perf_ctl_scaling = %d\n", cpu->cpu, perf_ctl_scaling);
581 	pr_debug("CPU%d: HWP_CAP guaranteed = %d\n", cpu->cpu, cpu->pstate.max_pstate);
582 	pr_debug("CPU%d: HWP_CAP highest = %d\n", cpu->cpu, cpu->pstate.turbo_pstate);
583 	pr_debug("CPU%d: HWP-to-frequency scaling factor: %d\n", cpu->cpu, scaling);
584 
585 	cpu->pstate.turbo_freq = rounddown(cpu->pstate.turbo_pstate * scaling,
586 					   perf_ctl_scaling);
587 	cpu->pstate.max_freq = rounddown(cpu->pstate.max_pstate * scaling,
588 					 perf_ctl_scaling);
589 
590 	freq = perf_ctl_max_phys * perf_ctl_scaling;
591 	cpu->pstate.max_pstate_physical = intel_pstate_freq_to_hwp(cpu, freq);
592 
593 	freq = cpu->pstate.min_pstate * perf_ctl_scaling;
594 	cpu->pstate.min_freq = freq;
595 	/*
596 	 * Cast the min P-state value retrieved via pstate_funcs.get_min() to
597 	 * the effective range of HWP performance levels.
598 	 */
599 	cpu->pstate.min_pstate = intel_pstate_freq_to_hwp(cpu, freq);
600 }
601 
602 static bool turbo_is_disabled(void)
603 {
604 	u64 misc_en;
605 
606 	if (!cpu_feature_enabled(X86_FEATURE_IDA))
607 		return true;
608 
609 	rdmsrq(MSR_IA32_MISC_ENABLE, misc_en);
610 
611 	return !!(misc_en & MSR_IA32_MISC_ENABLE_TURBO_DISABLE);
612 }
613 
614 static int min_perf_pct_min(void)
615 {
616 	struct cpudata *cpu = all_cpu_data[0];
617 	int turbo_pstate = cpu->pstate.turbo_pstate;
618 
619 	return turbo_pstate ?
620 		(cpu->pstate.min_pstate * 100 / turbo_pstate) : 0;
621 }
622 
623 static s16 intel_pstate_get_epp(struct cpudata *cpu_data, u64 hwp_req_data)
624 {
625 	s16 epp = -EOPNOTSUPP;
626 
627 	if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
628 		/*
629 		 * When hwp_req_data is 0, means that caller didn't read
630 		 * MSR_HWP_REQUEST, so need to read and get EPP.
631 		 */
632 		if (!hwp_req_data) {
633 			epp = rdmsrq_on_cpu(cpu_data->cpu, MSR_HWP_REQUEST,
634 					    &hwp_req_data);
635 			if (epp)
636 				return epp;
637 		}
638 		epp = (hwp_req_data >> 24) & 0xff;
639 	}
640 
641 	return epp;
642 }
643 
644 /*
645  * EPP display strings corresponding to EPP index in the
646  * energy_perf_strings[]
647  *	index		String
648  *-------------------------------------
649  *	0		default
650  *	1		performance
651  *	2		balance_performance
652  *	3		balance_power
653  *	4		power
654  */
655 
656 enum energy_perf_value_index {
657 	EPP_INDEX_DEFAULT = 0,
658 	EPP_INDEX_PERFORMANCE,
659 	EPP_INDEX_BALANCE_PERFORMANCE,
660 	EPP_INDEX_BALANCE_POWERSAVE,
661 	EPP_INDEX_POWERSAVE,
662 };
663 
664 static const char * const energy_perf_strings[] = {
665 	[EPP_INDEX_DEFAULT] = "default",
666 	[EPP_INDEX_PERFORMANCE] = "performance",
667 	[EPP_INDEX_BALANCE_PERFORMANCE] = "balance_performance",
668 	[EPP_INDEX_BALANCE_POWERSAVE] = "balance_power",
669 	[EPP_INDEX_POWERSAVE] = "power",
670 	NULL
671 };
672 static unsigned int epp_values[] = {
673 	[EPP_INDEX_DEFAULT] = 0, /* Unused index */
674 	[EPP_INDEX_PERFORMANCE] = HWP_EPP_PERFORMANCE,
675 	[EPP_INDEX_BALANCE_PERFORMANCE] = HWP_EPP_BALANCE_PERFORMANCE,
676 	[EPP_INDEX_BALANCE_POWERSAVE] = HWP_EPP_BALANCE_POWERSAVE,
677 	[EPP_INDEX_POWERSAVE] = HWP_EPP_POWERSAVE,
678 };
679 
680 static int intel_pstate_get_energy_pref_index(struct cpudata *cpu_data, int *raw_epp)
681 {
682 	s16 epp;
683 	int index = -EINVAL;
684 
685 	*raw_epp = 0;
686 	epp = intel_pstate_get_epp(cpu_data, 0);
687 	if (epp < 0)
688 		return epp;
689 
690 	if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
691 		if (epp == epp_values[EPP_INDEX_PERFORMANCE])
692 			return EPP_INDEX_PERFORMANCE;
693 		if (epp == epp_values[EPP_INDEX_BALANCE_PERFORMANCE])
694 			return EPP_INDEX_BALANCE_PERFORMANCE;
695 		if (epp == epp_values[EPP_INDEX_BALANCE_POWERSAVE])
696 			return EPP_INDEX_BALANCE_POWERSAVE;
697 		if (epp == epp_values[EPP_INDEX_POWERSAVE])
698 			return EPP_INDEX_POWERSAVE;
699 		*raw_epp = epp;
700 		return 0;
701 	} else if (boot_cpu_has(X86_FEATURE_EPB)) {
702 		/*
703 		 * Range:
704 		 *	0x00-0x03	:	Performance
705 		 *	0x04-0x07	:	Balance performance
706 		 *	0x08-0x0B	:	Balance power
707 		 *	0x0C-0x0F	:	Power
708 		 * The EPB is a 4 bit value, but our ranges restrict the
709 		 * value which can be set. Here only using top two bits
710 		 * effectively.
711 		 */
712 		index = (epp >> 2) + 1;
713 	}
714 
715 	return index;
716 }
717 
718 static int intel_pstate_set_epp(struct cpudata *cpu, u32 epp)
719 {
720 	int ret;
721 
722 	/*
723 	 * Use the cached HWP Request MSR value, because in the active mode the
724 	 * register itself may be updated by intel_pstate_hwp_boost_up() or
725 	 * intel_pstate_hwp_boost_down() at any time.
726 	 */
727 	u64 value = READ_ONCE(cpu->hwp_req_cached);
728 
729 	value &= ~GENMASK_ULL(31, 24);
730 	value |= (u64)epp << 24;
731 	/*
732 	 * The only other updater of hwp_req_cached in the active mode,
733 	 * intel_pstate_hwp_set(), is called under the same lock as this
734 	 * function, so it cannot run in parallel with the update below.
735 	 */
736 	WRITE_ONCE(cpu->hwp_req_cached, value);
737 	ret = wrmsrq_on_cpu(cpu->cpu, MSR_HWP_REQUEST, value);
738 	if (!ret)
739 		cpu->epp_cached = epp;
740 
741 	return ret;
742 }
743 
744 static int intel_pstate_set_energy_pref_index(struct cpudata *cpu_data,
745 					      int pref_index, bool use_raw,
746 					      u32 raw_epp)
747 {
748 	int epp = -EINVAL;
749 	int ret = -EOPNOTSUPP;
750 
751 	if (!pref_index)
752 		epp = cpu_data->epp_default;
753 
754 	if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
755 		if (use_raw)
756 			epp = raw_epp;
757 		else if (epp == -EINVAL)
758 			epp = epp_values[pref_index];
759 
760 		/*
761 		 * To avoid confusion, refuse to set EPP to any values different
762 		 * from 0 (performance) if the current policy is "performance",
763 		 * because those values would be overridden.
764 		 */
765 		if (epp > 0 && cpu_data->policy == CPUFREQ_POLICY_PERFORMANCE)
766 			return -EBUSY;
767 
768 		ret = intel_pstate_set_epp(cpu_data, epp);
769 	}
770 
771 	return ret;
772 }
773 
774 static ssize_t show_energy_performance_available_preferences(
775 				struct cpufreq_policy *policy, char *buf)
776 {
777 	int i = 0;
778 	int ret = 0;
779 
780 	while (energy_perf_strings[i] != NULL)
781 		ret += sprintf(&buf[ret], "%s ", energy_perf_strings[i++]);
782 
783 	ret += sprintf(&buf[ret], "\n");
784 
785 	return ret;
786 }
787 
788 cpufreq_freq_attr_ro(energy_performance_available_preferences);
789 
790 static struct cpufreq_driver intel_pstate;
791 
792 static ssize_t store_energy_performance_preference(
793 		struct cpufreq_policy *policy, const char *buf, size_t count)
794 {
795 	struct cpudata *cpu = all_cpu_data[policy->cpu];
796 	char str_preference[21];
797 	bool raw = false;
798 	ssize_t ret;
799 	u32 epp = 0;
800 
801 	ret = sscanf(buf, "%20s", str_preference);
802 	if (ret != 1)
803 		return -EINVAL;
804 
805 	ret = match_string(energy_perf_strings, -1, str_preference);
806 	if (ret < 0) {
807 		if (!boot_cpu_has(X86_FEATURE_HWP_EPP))
808 			return ret;
809 
810 		ret = kstrtouint(buf, 10, &epp);
811 		if (ret)
812 			return ret;
813 
814 		if (epp > 255)
815 			return -EINVAL;
816 
817 		raw = true;
818 	}
819 
820 	/*
821 	 * This function runs with the policy R/W semaphore held, which
822 	 * guarantees that the driver pointer will not change while it is
823 	 * running.
824 	 */
825 	if (!intel_pstate_driver)
826 		return -EAGAIN;
827 
828 	mutex_lock(&intel_pstate_limits_lock);
829 
830 	if (intel_pstate_driver == &intel_pstate) {
831 		ret = intel_pstate_set_energy_pref_index(cpu, ret, raw, epp);
832 	} else {
833 		/*
834 		 * In the passive mode the governor needs to be stopped on the
835 		 * target CPU before the EPP update and restarted after it,
836 		 * which is super-heavy-weight, so make sure it is worth doing
837 		 * upfront.
838 		 */
839 		if (!raw)
840 			epp = ret ? epp_values[ret] : cpu->epp_default;
841 
842 		if (cpu->epp_cached != epp) {
843 			int err;
844 
845 			cpufreq_stop_governor(policy);
846 			ret = intel_pstate_set_epp(cpu, epp);
847 			err = cpufreq_start_governor(policy);
848 			if (!ret)
849 				ret = err;
850 		} else {
851 			ret = 0;
852 		}
853 	}
854 
855 	mutex_unlock(&intel_pstate_limits_lock);
856 
857 	return ret ?: count;
858 }
859 
860 static ssize_t show_energy_performance_preference(
861 				struct cpufreq_policy *policy, char *buf)
862 {
863 	struct cpudata *cpu_data = all_cpu_data[policy->cpu];
864 	int preference, raw_epp;
865 
866 	preference = intel_pstate_get_energy_pref_index(cpu_data, &raw_epp);
867 	if (preference < 0)
868 		return preference;
869 
870 	if (raw_epp)
871 		return  sprintf(buf, "%d\n", raw_epp);
872 	else
873 		return  sprintf(buf, "%s\n", energy_perf_strings[preference]);
874 }
875 
876 cpufreq_freq_attr_rw(energy_performance_preference);
877 
878 static ssize_t show_base_frequency(struct cpufreq_policy *policy, char *buf)
879 {
880 	struct cpudata *cpu = all_cpu_data[policy->cpu];
881 	int ratio, freq;
882 
883 	ratio = intel_pstate_get_cppc_guaranteed(policy->cpu);
884 	if (ratio <= 0) {
885 		u64 cap;
886 
887 		rdmsrq_on_cpu(policy->cpu, MSR_HWP_CAPABILITIES, &cap);
888 		ratio = HWP_GUARANTEED_PERF(cap);
889 	}
890 
891 	freq = ratio * cpu->pstate.scaling;
892 	if (cpu->pstate.scaling != cpu->pstate.perf_ctl_scaling)
893 		freq = rounddown(freq, cpu->pstate.perf_ctl_scaling);
894 
895 	return sprintf(buf, "%d\n", freq);
896 }
897 
898 cpufreq_freq_attr_ro(base_frequency);
899 
900 enum hwp_cpufreq_attr_index {
901 	HWP_BASE_FREQUENCY_INDEX = 0,
902 	HWP_PERFORMANCE_PREFERENCE_INDEX,
903 	HWP_PERFORMANCE_AVAILABLE_PREFERENCES_INDEX,
904 	HWP_CPUFREQ_ATTR_COUNT,
905 };
906 
907 static struct freq_attr *hwp_cpufreq_attrs[] = {
908 	[HWP_BASE_FREQUENCY_INDEX] = &base_frequency,
909 	[HWP_PERFORMANCE_PREFERENCE_INDEX] = &energy_performance_preference,
910 	[HWP_PERFORMANCE_AVAILABLE_PREFERENCES_INDEX] =
911 				&energy_performance_available_preferences,
912 	[HWP_CPUFREQ_ATTR_COUNT] = NULL,
913 };
914 
915 static bool no_cas __ro_after_init;
916 
917 static struct cpudata *hybrid_max_perf_cpu __read_mostly;
918 /*
919  * Protects hybrid_max_perf_cpu, the capacity_perf fields in struct cpudata,
920  * and the x86 arch scale-invariance information from concurrent updates.
921  */
922 static DEFINE_MUTEX(hybrid_capacity_lock);
923 
924 #ifdef CONFIG_ENERGY_MODEL
925 #define HYBRID_EM_STATE_COUNT	4
926 
927 static int hybrid_active_power(struct device *dev, unsigned long *power,
928 			       unsigned long *freq)
929 {
930 	/*
931 	 * Create "utilization bins" of 0-40%, 40%-60%, 60%-80%, and 80%-100%
932 	 * of the maximum capacity such that two CPUs of the same type will be
933 	 * regarded as equally attractive if the utilization of each of them
934 	 * falls into the same bin, which should prevent tasks from being
935 	 * migrated between them too often.
936 	 *
937 	 * For this purpose, return the "frequency" of 2 for the first
938 	 * performance level and otherwise leave the value set by the caller.
939 	 */
940 	if (!*freq)
941 		*freq = 2;
942 
943 	/* No power information. */
944 	*power = EM_MAX_POWER;
945 
946 	return 0;
947 }
948 
949 static int hybrid_get_cost(struct device *dev, unsigned long freq,
950 			   unsigned long *cost)
951 {
952 	struct pstate_data *pstate = &all_cpu_data[dev->id]->pstate;
953 	struct cpu_cacheinfo *cacheinfo = get_cpu_cacheinfo(dev->id);
954 
955 	/*
956 	 * The smaller the perf-to-frequency scaling factor, the larger the IPC
957 	 * ratio between the given CPU and the least capable CPU in the system.
958 	 * Regard that IPC ratio as the primary cost component and assume that
959 	 * the scaling factors for different CPU types will differ by at least
960 	 * 5% and they will not be above INTEL_PSTATE_CORE_SCALING.
961 	 *
962 	 * Add the freq value to the cost, so that the cost of running on CPUs
963 	 * of the same type in different "utilization bins" is different.
964 	 */
965 	*cost = div_u64(100ULL * INTEL_PSTATE_CORE_SCALING, pstate->scaling) + freq;
966 	/*
967 	 * Increase the cost slightly for CPUs able to access L3 to avoid
968 	 * touching it in case some other CPUs of the same type can do the work
969 	 * without it.
970 	 */
971 	if (cacheinfo) {
972 		unsigned int i;
973 
974 		/* Check if L3 cache is there. */
975 		for (i = 0; i < cacheinfo->num_leaves; i++) {
976 			if (cacheinfo->info_list[i].level == 3) {
977 				*cost += 2;
978 				break;
979 			}
980 		}
981 	}
982 
983 	return 0;
984 }
985 
986 static bool hybrid_register_perf_domain(unsigned int cpu)
987 {
988 	static const struct em_data_callback cb
989 			= EM_ADV_DATA_CB(hybrid_active_power, hybrid_get_cost);
990 	struct cpudata *cpudata = all_cpu_data[cpu];
991 	struct device *cpu_dev;
992 
993 	/*
994 	 * Registering EM perf domains without enabling asymmetric CPU capacity
995 	 * support is not really useful and one domain should not be registered
996 	 * more than once.
997 	 */
998 	if (!hybrid_max_perf_cpu || cpudata->pd_registered)
999 		return false;
1000 
1001 	cpu_dev = get_cpu_device(cpu);
1002 	if (!cpu_dev)
1003 		return false;
1004 
1005 	if (em_dev_register_pd_no_update(cpu_dev, HYBRID_EM_STATE_COUNT, &cb,
1006 					 cpumask_of(cpu), false))
1007 		return false;
1008 
1009 	cpudata->pd_registered = true;
1010 
1011 	return true;
1012 }
1013 
1014 static void hybrid_register_all_perf_domains(void)
1015 {
1016 	unsigned int cpu;
1017 
1018 	for_each_online_cpu(cpu)
1019 		hybrid_register_perf_domain(cpu);
1020 }
1021 
1022 static void hybrid_update_perf_domain(struct cpudata *cpu)
1023 {
1024 	if (cpu->pd_registered)
1025 		em_adjust_cpu_capacity(cpu->cpu);
1026 }
1027 #else /* !CONFIG_ENERGY_MODEL */
1028 static inline bool hybrid_register_perf_domain(unsigned int cpu) { return false; }
1029 static inline void hybrid_register_all_perf_domains(void) {}
1030 static inline void hybrid_update_perf_domain(struct cpudata *cpu) {}
1031 #endif /* CONFIG_ENERGY_MODEL */
1032 
1033 static void hybrid_set_cpu_capacity(struct cpudata *cpu)
1034 {
1035 	arch_set_cpu_capacity(cpu->cpu, cpu->capacity_perf,
1036 			      hybrid_max_perf_cpu->capacity_perf,
1037 			      cpu->capacity_perf,
1038 			      cpu->pstate.max_pstate_physical);
1039 	hybrid_update_perf_domain(cpu);
1040 
1041 	topology_set_cpu_scale(cpu->cpu, arch_scale_cpu_capacity(cpu->cpu));
1042 
1043 	pr_debug("CPU%d: perf = %u, max. perf = %u, base perf = %d\n", cpu->cpu,
1044 		 cpu->capacity_perf, hybrid_max_perf_cpu->capacity_perf,
1045 		 cpu->pstate.max_pstate_physical);
1046 }
1047 
1048 static void hybrid_clear_cpu_capacity(unsigned int cpunum)
1049 {
1050 	arch_set_cpu_capacity(cpunum, 1, 1, 1, 1);
1051 }
1052 
1053 static void hybrid_get_capacity_perf(struct cpudata *cpu)
1054 {
1055 	if (READ_ONCE(global.no_turbo)) {
1056 		cpu->capacity_perf = cpu->pstate.max_pstate_physical;
1057 		return;
1058 	}
1059 
1060 	cpu->capacity_perf = HWP_HIGHEST_PERF(READ_ONCE(cpu->hwp_cap_cached));
1061 }
1062 
1063 static void hybrid_set_capacity_of_cpus(void)
1064 {
1065 	int cpunum;
1066 
1067 	for_each_online_cpu(cpunum) {
1068 		struct cpudata *cpu = all_cpu_data[cpunum];
1069 
1070 		if (cpu)
1071 			hybrid_set_cpu_capacity(cpu);
1072 	}
1073 }
1074 
1075 static void hybrid_update_cpu_capacity_scaling(void)
1076 {
1077 	struct cpudata *max_perf_cpu = NULL;
1078 	unsigned int max_cap_perf = 0;
1079 	int cpunum;
1080 
1081 	for_each_online_cpu(cpunum) {
1082 		struct cpudata *cpu = all_cpu_data[cpunum];
1083 
1084 		if (!cpu)
1085 			continue;
1086 
1087 		/*
1088 		 * During initialization, CPU performance at full capacity needs
1089 		 * to be determined.
1090 		 */
1091 		if (!hybrid_max_perf_cpu)
1092 			hybrid_get_capacity_perf(cpu);
1093 
1094 		/*
1095 		 * If hybrid_max_perf_cpu is not NULL at this point, it is
1096 		 * being replaced, so don't take it into account when looking
1097 		 * for the new one.
1098 		 */
1099 		if (cpu == hybrid_max_perf_cpu)
1100 			continue;
1101 
1102 		if (cpu->capacity_perf > max_cap_perf) {
1103 			max_cap_perf = cpu->capacity_perf;
1104 			max_perf_cpu = cpu;
1105 		}
1106 	}
1107 
1108 	if (max_perf_cpu) {
1109 		hybrid_max_perf_cpu = max_perf_cpu;
1110 		hybrid_set_capacity_of_cpus();
1111 	} else {
1112 		pr_info("Found no CPUs with nonzero maximum performance\n");
1113 		/* Revert to the flat CPU capacity structure. */
1114 		for_each_online_cpu(cpunum)
1115 			hybrid_clear_cpu_capacity(cpunum);
1116 	}
1117 }
1118 
1119 static void __hybrid_refresh_cpu_capacity_scaling(void)
1120 {
1121 	hybrid_max_perf_cpu = NULL;
1122 	hybrid_update_cpu_capacity_scaling();
1123 }
1124 
1125 static void hybrid_refresh_cpu_capacity_scaling(void)
1126 {
1127 	guard(mutex)(&hybrid_capacity_lock);
1128 
1129 	__hybrid_refresh_cpu_capacity_scaling();
1130 	/*
1131 	 * Perf domains are not registered before setting hybrid_max_perf_cpu,
1132 	 * so register them all after setting up CPU capacity scaling.
1133 	 */
1134 	hybrid_register_all_perf_domains();
1135 }
1136 
1137 static void hybrid_init_cpu_capacity_scaling(bool refresh)
1138 {
1139 	/* Bail out if enabling capacity-aware scheduling is prohibited. */
1140 	if (no_cas)
1141 		return;
1142 
1143 	/*
1144 	 * If hybrid_max_perf_cpu is set at this point, the hybrid CPU capacity
1145 	 * scaling has been enabled already and the driver is just changing the
1146 	 * operation mode.
1147 	 */
1148 	if (refresh) {
1149 		hybrid_refresh_cpu_capacity_scaling();
1150 		return;
1151 	}
1152 
1153 	/*
1154 	 * On hybrid systems, use asym capacity instead of ITMT, but because
1155 	 * the capacity of SMT threads is not deterministic even approximately,
1156 	 * do not do that when SMT is in use.
1157 	 */
1158 	if (hwp_is_hybrid && !sched_smt_active() && arch_enable_hybrid_capacity_scale()) {
1159 		hybrid_refresh_cpu_capacity_scaling();
1160 		/*
1161 		 * Disabling ITMT causes sched domains to be rebuilt to disable asym
1162 		 * packing and enable asym capacity and EAS.
1163 		 */
1164 		sched_clear_itmt_support();
1165 	}
1166 }
1167 
1168 static bool hybrid_clear_max_perf_cpu(void)
1169 {
1170 	bool ret;
1171 
1172 	guard(mutex)(&hybrid_capacity_lock);
1173 
1174 	ret = !!hybrid_max_perf_cpu;
1175 	hybrid_max_perf_cpu = NULL;
1176 
1177 	return ret;
1178 }
1179 
1180 static void __intel_pstate_get_hwp_cap(struct cpudata *cpu)
1181 {
1182 	u64 cap;
1183 
1184 	rdmsrq_on_cpu(cpu->cpu, MSR_HWP_CAPABILITIES, &cap);
1185 	WRITE_ONCE(cpu->hwp_cap_cached, cap);
1186 	cpu->pstate.max_pstate = HWP_GUARANTEED_PERF(cap);
1187 	cpu->pstate.turbo_pstate = HWP_HIGHEST_PERF(cap);
1188 }
1189 
1190 static void intel_pstate_get_hwp_cap(struct cpudata *cpu)
1191 {
1192 	int scaling = cpu->pstate.scaling;
1193 
1194 	__intel_pstate_get_hwp_cap(cpu);
1195 
1196 	cpu->pstate.max_freq = cpu->pstate.max_pstate * scaling;
1197 	cpu->pstate.turbo_freq = cpu->pstate.turbo_pstate * scaling;
1198 	if (scaling != cpu->pstate.perf_ctl_scaling) {
1199 		int perf_ctl_scaling = cpu->pstate.perf_ctl_scaling;
1200 
1201 		cpu->pstate.max_freq = rounddown(cpu->pstate.max_freq,
1202 						 perf_ctl_scaling);
1203 		cpu->pstate.turbo_freq = rounddown(cpu->pstate.turbo_freq,
1204 						   perf_ctl_scaling);
1205 	}
1206 }
1207 
1208 static void hybrid_update_capacity(struct cpudata *cpu)
1209 {
1210 	unsigned int max_cap_perf;
1211 
1212 	mutex_lock(&hybrid_capacity_lock);
1213 
1214 	if (!hybrid_max_perf_cpu)
1215 		goto unlock;
1216 
1217 	/*
1218 	 * The maximum performance of the CPU may have changed, but assume
1219 	 * that the performance of the other CPUs has not changed.
1220 	 */
1221 	max_cap_perf = hybrid_max_perf_cpu->capacity_perf;
1222 
1223 	intel_pstate_get_hwp_cap(cpu);
1224 
1225 	hybrid_get_capacity_perf(cpu);
1226 	/* Should hybrid_max_perf_cpu be replaced by this CPU? */
1227 	if (cpu->capacity_perf > max_cap_perf) {
1228 		hybrid_max_perf_cpu = cpu;
1229 		hybrid_set_capacity_of_cpus();
1230 		goto unlock;
1231 	}
1232 
1233 	/* If this CPU is hybrid_max_perf_cpu, should it be replaced? */
1234 	if (cpu == hybrid_max_perf_cpu && cpu->capacity_perf < max_cap_perf) {
1235 		hybrid_update_cpu_capacity_scaling();
1236 		goto unlock;
1237 	}
1238 
1239 	hybrid_set_cpu_capacity(cpu);
1240 	/*
1241 	 * If the CPU was offline to start with and it is going online for the
1242 	 * first time, a perf domain needs to be registered for it if hybrid
1243 	 * capacity scaling has been enabled already.  In that case, sched
1244 	 * domains need to be rebuilt to take the new perf domain into account.
1245 	 */
1246 	if (hybrid_register_perf_domain(cpu->cpu))
1247 		em_rebuild_sched_domains();
1248 
1249 unlock:
1250 	mutex_unlock(&hybrid_capacity_lock);
1251 }
1252 
1253 static void intel_pstate_hwp_set(unsigned int cpu)
1254 {
1255 	struct cpudata *cpu_data = all_cpu_data[cpu];
1256 	int max, min;
1257 	u64 value;
1258 	s16 epp;
1259 
1260 	max = cpu_data->max_perf_ratio;
1261 	min = cpu_data->min_perf_ratio;
1262 
1263 	if (cpu_data->policy == CPUFREQ_POLICY_PERFORMANCE)
1264 		min = max;
1265 
1266 	rdmsrq_on_cpu(cpu, MSR_HWP_REQUEST, &value);
1267 
1268 	value &= ~HWP_MIN_PERF(~0L);
1269 	value |= HWP_MIN_PERF(min);
1270 
1271 	value &= ~HWP_MAX_PERF(~0L);
1272 	value |= HWP_MAX_PERF(max);
1273 
1274 	if (cpu_data->epp_policy == cpu_data->policy)
1275 		goto skip_epp;
1276 
1277 	cpu_data->epp_policy = cpu_data->policy;
1278 
1279 	if (cpu_data->policy == CPUFREQ_POLICY_PERFORMANCE) {
1280 		epp = intel_pstate_get_epp(cpu_data, value);
1281 		cpu_data->epp_powersave = epp;
1282 		/* If EPP read was failed, then don't try to write */
1283 		if (epp < 0)
1284 			goto skip_epp;
1285 
1286 		epp = 0;
1287 	} else {
1288 		/* skip setting EPP, when saved value is invalid */
1289 		if (cpu_data->epp_powersave < 0)
1290 			goto skip_epp;
1291 
1292 		/*
1293 		 * No need to restore EPP when it is not zero. This
1294 		 * means:
1295 		 *  - Policy is not changed
1296 		 *  - user has manually changed
1297 		 *  - Error reading EPB
1298 		 */
1299 		epp = intel_pstate_get_epp(cpu_data, value);
1300 		if (epp)
1301 			goto skip_epp;
1302 
1303 		epp = cpu_data->epp_powersave;
1304 	}
1305 	if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
1306 		value &= ~GENMASK_ULL(31, 24);
1307 		value |= (u64)epp << 24;
1308 	}
1309 
1310 skip_epp:
1311 	WRITE_ONCE(cpu_data->hwp_req_cached, value);
1312 	wrmsrq_on_cpu(cpu, MSR_HWP_REQUEST, value);
1313 }
1314 
1315 static void intel_pstate_disable_hwp_interrupt(struct cpudata *cpudata);
1316 
1317 static void intel_pstate_hwp_offline(struct cpudata *cpu)
1318 {
1319 	u64 value = READ_ONCE(cpu->hwp_req_cached);
1320 	int min_perf;
1321 
1322 	intel_pstate_disable_hwp_interrupt(cpu);
1323 
1324 	if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
1325 		/*
1326 		 * In case the EPP has been set to "performance" by the
1327 		 * active mode "performance" scaling algorithm, replace that
1328 		 * temporary value with the cached EPP one.
1329 		 */
1330 		value &= ~GENMASK_ULL(31, 24);
1331 		value |= HWP_ENERGY_PERF_PREFERENCE(cpu->epp_cached);
1332 		/*
1333 		 * However, make sure that EPP will be set to "performance" when
1334 		 * the CPU is brought back online again and the "performance"
1335 		 * scaling algorithm is still in effect.
1336 		 */
1337 		cpu->epp_policy = CPUFREQ_POLICY_UNKNOWN;
1338 	}
1339 
1340 	/*
1341 	 * Clear the desired perf field in the cached HWP request value to
1342 	 * prevent nonzero desired values from being leaked into the active
1343 	 * mode.
1344 	 */
1345 	value &= ~HWP_DESIRED_PERF(~0L);
1346 	WRITE_ONCE(cpu->hwp_req_cached, value);
1347 
1348 	value &= ~GENMASK_ULL(31, 0);
1349 	min_perf = HWP_LOWEST_PERF(READ_ONCE(cpu->hwp_cap_cached));
1350 
1351 	/* Set hwp_max = hwp_min */
1352 	value |= HWP_MAX_PERF(min_perf);
1353 	value |= HWP_MIN_PERF(min_perf);
1354 
1355 	/* Set EPP to min */
1356 	if (boot_cpu_has(X86_FEATURE_HWP_EPP))
1357 		value |= HWP_ENERGY_PERF_PREFERENCE(HWP_EPP_POWERSAVE);
1358 
1359 	wrmsrq_on_cpu(cpu->cpu, MSR_HWP_REQUEST, value);
1360 
1361 	mutex_lock(&hybrid_capacity_lock);
1362 
1363 	if (!hybrid_max_perf_cpu) {
1364 		mutex_unlock(&hybrid_capacity_lock);
1365 
1366 		return;
1367 	}
1368 
1369 	if (hybrid_max_perf_cpu == cpu)
1370 		hybrid_update_cpu_capacity_scaling();
1371 
1372 	mutex_unlock(&hybrid_capacity_lock);
1373 
1374 	/* Reset the capacity of the CPU going offline to the initial value. */
1375 	hybrid_clear_cpu_capacity(cpu->cpu);
1376 }
1377 
1378 #define POWER_CTL_EE_ENABLE	1
1379 #define POWER_CTL_EE_DISABLE	2
1380 
1381 /* Enable bit for Dynamic Efficiency Control (DEC) */
1382 #define POWER_CTL_DEC_ENABLE	27
1383 
1384 static int power_ctl_ee_state;
1385 
1386 static void set_power_ctl_ee_state(bool input)
1387 {
1388 	u64 power_ctl;
1389 
1390 	mutex_lock(&intel_pstate_driver_lock);
1391 	rdmsrq(MSR_IA32_POWER_CTL, power_ctl);
1392 	if (input) {
1393 		power_ctl &= ~BIT(MSR_IA32_POWER_CTL_BIT_EE);
1394 		power_ctl_ee_state = POWER_CTL_EE_ENABLE;
1395 	} else {
1396 		power_ctl |= BIT(MSR_IA32_POWER_CTL_BIT_EE);
1397 		power_ctl_ee_state = POWER_CTL_EE_DISABLE;
1398 	}
1399 	wrmsrq(MSR_IA32_POWER_CTL, power_ctl);
1400 	mutex_unlock(&intel_pstate_driver_lock);
1401 }
1402 
1403 static void intel_pstate_hwp_enable(struct cpudata *cpudata);
1404 
1405 static void intel_pstate_hwp_reenable(struct cpudata *cpu)
1406 {
1407 	intel_pstate_hwp_enable(cpu);
1408 	wrmsrq_on_cpu(cpu->cpu, MSR_HWP_REQUEST, READ_ONCE(cpu->hwp_req_cached));
1409 }
1410 
1411 static int intel_pstate_suspend(struct cpufreq_policy *policy)
1412 {
1413 	struct cpudata *cpu = all_cpu_data[policy->cpu];
1414 
1415 	pr_debug("CPU %d suspending\n", cpu->cpu);
1416 
1417 	cpu->suspended = true;
1418 
1419 	/* disable HWP interrupt and cancel any pending work */
1420 	intel_pstate_disable_hwp_interrupt(cpu);
1421 
1422 	return 0;
1423 }
1424 
1425 static int intel_pstate_resume(struct cpufreq_policy *policy)
1426 {
1427 	struct cpudata *cpu = all_cpu_data[policy->cpu];
1428 
1429 	pr_debug("CPU %d resuming\n", cpu->cpu);
1430 
1431 	/* Only restore if the system default is changed */
1432 	if (power_ctl_ee_state == POWER_CTL_EE_ENABLE)
1433 		set_power_ctl_ee_state(true);
1434 	else if (power_ctl_ee_state == POWER_CTL_EE_DISABLE)
1435 		set_power_ctl_ee_state(false);
1436 
1437 	if (cpu->suspended && hwp_active) {
1438 		mutex_lock(&intel_pstate_limits_lock);
1439 
1440 		/* Re-enable HWP, because "online" has not done that. */
1441 		intel_pstate_hwp_reenable(cpu);
1442 
1443 		mutex_unlock(&intel_pstate_limits_lock);
1444 	}
1445 
1446 	cpu->suspended = false;
1447 
1448 	return 0;
1449 }
1450 
1451 static void intel_pstate_update_policies(void)
1452 {
1453 	int cpu;
1454 
1455 	for_each_possible_cpu(cpu)
1456 		cpufreq_update_policy(cpu);
1457 }
1458 
1459 static void __intel_pstate_update_max_freq(struct cpufreq_policy *policy,
1460 					   struct cpudata *cpudata)
1461 {
1462 	guard(cpufreq_policy_write)(policy);
1463 
1464 	if (hwp_active)
1465 		intel_pstate_get_hwp_cap(cpudata);
1466 
1467 	policy->cpuinfo.max_freq = READ_ONCE(global.no_turbo) ?
1468 			cpudata->pstate.max_freq : cpudata->pstate.turbo_freq;
1469 
1470 	refresh_frequency_limits(policy);
1471 }
1472 
1473 static bool intel_pstate_update_max_freq(struct cpudata *cpudata)
1474 {
1475 	struct cpufreq_policy *policy __free(put_cpufreq_policy) = cpufreq_cpu_get(cpudata->cpu);
1476 	if (!policy)
1477 		return false;
1478 
1479 	__intel_pstate_update_max_freq(policy, cpudata);
1480 
1481 	return true;
1482 }
1483 
1484 static void intel_pstate_update_limits(struct cpufreq_policy *policy)
1485 {
1486 	struct cpudata *cpudata = all_cpu_data[policy->cpu];
1487 
1488 	__intel_pstate_update_max_freq(policy, cpudata);
1489 
1490 	hybrid_update_capacity(cpudata);
1491 }
1492 
1493 static void intel_pstate_update_limits_for_all(void)
1494 {
1495 	int cpu;
1496 
1497 	for_each_possible_cpu(cpu)
1498 		intel_pstate_update_max_freq(all_cpu_data[cpu]);
1499 
1500 	mutex_lock(&hybrid_capacity_lock);
1501 
1502 	if (hybrid_max_perf_cpu)
1503 		__hybrid_refresh_cpu_capacity_scaling();
1504 
1505 	mutex_unlock(&hybrid_capacity_lock);
1506 }
1507 
1508 /************************** sysfs begin ************************/
1509 #define show_one(file_name, object)					\
1510 	static ssize_t show_##file_name					\
1511 	(struct kobject *kobj, struct kobj_attribute *attr, char *buf)	\
1512 	{								\
1513 		return sprintf(buf, "%u\n", global.object);		\
1514 	}
1515 
1516 static ssize_t intel_pstate_show_status(char *buf);
1517 static int intel_pstate_update_status(const char *buf, size_t size);
1518 
1519 static ssize_t show_status(struct kobject *kobj,
1520 			   struct kobj_attribute *attr, char *buf)
1521 {
1522 	ssize_t ret;
1523 
1524 	mutex_lock(&intel_pstate_driver_lock);
1525 	ret = intel_pstate_show_status(buf);
1526 	mutex_unlock(&intel_pstate_driver_lock);
1527 
1528 	return ret;
1529 }
1530 
1531 static ssize_t store_status(struct kobject *a, struct kobj_attribute *b,
1532 			    const char *buf, size_t count)
1533 {
1534 	char *p = memchr(buf, '\n', count);
1535 	int ret;
1536 
1537 	mutex_lock(&intel_pstate_driver_lock);
1538 	ret = intel_pstate_update_status(buf, p ? p - buf : count);
1539 	mutex_unlock(&intel_pstate_driver_lock);
1540 
1541 	return ret < 0 ? ret : count;
1542 }
1543 
1544 static ssize_t show_turbo_pct(struct kobject *kobj,
1545 				struct kobj_attribute *attr, char *buf)
1546 {
1547 	struct cpudata *cpu;
1548 	int total, no_turbo, turbo_pct;
1549 	uint32_t turbo_fp;
1550 
1551 	mutex_lock(&intel_pstate_driver_lock);
1552 
1553 	if (!intel_pstate_driver) {
1554 		mutex_unlock(&intel_pstate_driver_lock);
1555 		return -EAGAIN;
1556 	}
1557 
1558 	cpu = all_cpu_data[0];
1559 
1560 	total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
1561 	no_turbo = cpu->pstate.max_pstate - cpu->pstate.min_pstate + 1;
1562 	turbo_fp = div_fp(no_turbo, total);
1563 	turbo_pct = 100 - fp_toint(mul_fp(turbo_fp, int_tofp(100)));
1564 
1565 	mutex_unlock(&intel_pstate_driver_lock);
1566 
1567 	return sprintf(buf, "%u\n", turbo_pct);
1568 }
1569 
1570 static ssize_t show_num_pstates(struct kobject *kobj,
1571 				struct kobj_attribute *attr, char *buf)
1572 {
1573 	struct cpudata *cpu;
1574 	int total;
1575 
1576 	mutex_lock(&intel_pstate_driver_lock);
1577 
1578 	if (!intel_pstate_driver) {
1579 		mutex_unlock(&intel_pstate_driver_lock);
1580 		return -EAGAIN;
1581 	}
1582 
1583 	cpu = all_cpu_data[0];
1584 	total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
1585 
1586 	mutex_unlock(&intel_pstate_driver_lock);
1587 
1588 	return sprintf(buf, "%u\n", total);
1589 }
1590 
1591 static ssize_t show_no_turbo(struct kobject *kobj,
1592 			     struct kobj_attribute *attr, char *buf)
1593 {
1594 	ssize_t ret;
1595 
1596 	mutex_lock(&intel_pstate_driver_lock);
1597 
1598 	if (!intel_pstate_driver) {
1599 		mutex_unlock(&intel_pstate_driver_lock);
1600 		return -EAGAIN;
1601 	}
1602 
1603 	ret = sprintf(buf, "%u\n", global.no_turbo);
1604 
1605 	mutex_unlock(&intel_pstate_driver_lock);
1606 
1607 	return ret;
1608 }
1609 
1610 static ssize_t store_no_turbo(struct kobject *a, struct kobj_attribute *b,
1611 			      const char *buf, size_t count)
1612 {
1613 	unsigned int input;
1614 	bool no_turbo;
1615 
1616 	if (sscanf(buf, "%u", &input) != 1)
1617 		return -EINVAL;
1618 
1619 	mutex_lock(&intel_pstate_driver_lock);
1620 
1621 	if (!intel_pstate_driver) {
1622 		count = -EAGAIN;
1623 		goto unlock_driver;
1624 	}
1625 
1626 	no_turbo = !!clamp_t(int, input, 0, 1);
1627 
1628 	WRITE_ONCE(global.turbo_disabled, turbo_is_disabled());
1629 	if (global.turbo_disabled && !no_turbo) {
1630 		pr_notice("Turbo disabled by BIOS or unavailable on processor\n");
1631 		count = -EPERM;
1632 		if (global.no_turbo)
1633 			goto unlock_driver;
1634 		else
1635 			no_turbo = 1;
1636 	}
1637 
1638 	if (no_turbo == global.no_turbo) {
1639 		goto unlock_driver;
1640 	}
1641 
1642 	WRITE_ONCE(global.no_turbo, no_turbo);
1643 
1644 	mutex_lock(&intel_pstate_limits_lock);
1645 
1646 	if (no_turbo) {
1647 		struct cpudata *cpu = all_cpu_data[0];
1648 		int pct = cpu->pstate.max_pstate * 100 / cpu->pstate.turbo_pstate;
1649 
1650 		/* Squash the global minimum into the permitted range. */
1651 		if (global.min_perf_pct > pct)
1652 			global.min_perf_pct = pct;
1653 	}
1654 
1655 	mutex_unlock(&intel_pstate_limits_lock);
1656 
1657 	intel_pstate_update_limits_for_all();
1658 	arch_set_max_freq_ratio(no_turbo);
1659 
1660 unlock_driver:
1661 	mutex_unlock(&intel_pstate_driver_lock);
1662 
1663 	return count;
1664 }
1665 
1666 static void update_cpu_qos_request(int cpu, enum freq_qos_req_type type)
1667 {
1668 	struct cpudata *cpudata = all_cpu_data[cpu];
1669 	unsigned int freq = cpudata->pstate.turbo_freq;
1670 	struct freq_qos_request *req;
1671 
1672 	struct cpufreq_policy *policy __free(put_cpufreq_policy) = cpufreq_cpu_get(cpu);
1673 	if (!policy)
1674 		return;
1675 
1676 	req = policy->driver_data;
1677 	if (!req)
1678 		return;
1679 
1680 	if (hwp_active)
1681 		intel_pstate_get_hwp_cap(cpudata);
1682 
1683 	if (type == FREQ_QOS_MIN) {
1684 		freq = DIV_ROUND_UP(freq * global.min_perf_pct, 100);
1685 	} else {
1686 		req++;
1687 		freq = (freq * global.max_perf_pct) / 100;
1688 	}
1689 
1690 	if (freq_qos_update_request(req, freq) < 0)
1691 		pr_warn("Failed to update freq constraint: CPU%d\n", cpu);
1692 }
1693 
1694 static void update_qos_requests(enum freq_qos_req_type type)
1695 {
1696 	int i;
1697 
1698 	for_each_possible_cpu(i)
1699 		update_cpu_qos_request(i, type);
1700 }
1701 
1702 static ssize_t store_max_perf_pct(struct kobject *a, struct kobj_attribute *b,
1703 				  const char *buf, size_t count)
1704 {
1705 	unsigned int input;
1706 	int ret;
1707 
1708 	ret = sscanf(buf, "%u", &input);
1709 	if (ret != 1)
1710 		return -EINVAL;
1711 
1712 	mutex_lock(&intel_pstate_driver_lock);
1713 
1714 	if (!intel_pstate_driver) {
1715 		mutex_unlock(&intel_pstate_driver_lock);
1716 		return -EAGAIN;
1717 	}
1718 
1719 	mutex_lock(&intel_pstate_limits_lock);
1720 
1721 	global.max_perf_pct = clamp_t(int, input, global.min_perf_pct, 100);
1722 
1723 	mutex_unlock(&intel_pstate_limits_lock);
1724 
1725 	if (intel_pstate_driver == &intel_pstate)
1726 		intel_pstate_update_policies();
1727 	else
1728 		update_qos_requests(FREQ_QOS_MAX);
1729 
1730 	mutex_unlock(&intel_pstate_driver_lock);
1731 
1732 	return count;
1733 }
1734 
1735 static ssize_t store_min_perf_pct(struct kobject *a, struct kobj_attribute *b,
1736 				  const char *buf, size_t count)
1737 {
1738 	unsigned int input;
1739 	int ret;
1740 
1741 	ret = sscanf(buf, "%u", &input);
1742 	if (ret != 1)
1743 		return -EINVAL;
1744 
1745 	mutex_lock(&intel_pstate_driver_lock);
1746 
1747 	if (!intel_pstate_driver) {
1748 		mutex_unlock(&intel_pstate_driver_lock);
1749 		return -EAGAIN;
1750 	}
1751 
1752 	mutex_lock(&intel_pstate_limits_lock);
1753 
1754 	global.min_perf_pct = clamp_t(int, input,
1755 				      min_perf_pct_min(), global.max_perf_pct);
1756 
1757 	mutex_unlock(&intel_pstate_limits_lock);
1758 
1759 	if (intel_pstate_driver == &intel_pstate)
1760 		intel_pstate_update_policies();
1761 	else
1762 		update_qos_requests(FREQ_QOS_MIN);
1763 
1764 	mutex_unlock(&intel_pstate_driver_lock);
1765 
1766 	return count;
1767 }
1768 
1769 static ssize_t show_hwp_dynamic_boost(struct kobject *kobj,
1770 				struct kobj_attribute *attr, char *buf)
1771 {
1772 	return sprintf(buf, "%u\n", hwp_boost);
1773 }
1774 
1775 static ssize_t store_hwp_dynamic_boost(struct kobject *a,
1776 				       struct kobj_attribute *b,
1777 				       const char *buf, size_t count)
1778 {
1779 	unsigned int input;
1780 	int ret;
1781 
1782 	ret = kstrtouint(buf, 10, &input);
1783 	if (ret)
1784 		return ret;
1785 
1786 	mutex_lock(&intel_pstate_driver_lock);
1787 	hwp_boost = !!input;
1788 	intel_pstate_update_policies();
1789 	mutex_unlock(&intel_pstate_driver_lock);
1790 
1791 	return count;
1792 }
1793 
1794 static ssize_t show_energy_efficiency(struct kobject *kobj, struct kobj_attribute *attr,
1795 				      char *buf)
1796 {
1797 	u64 power_ctl;
1798 	int enable;
1799 
1800 	rdmsrq(MSR_IA32_POWER_CTL, power_ctl);
1801 	enable = !!(power_ctl & BIT(MSR_IA32_POWER_CTL_BIT_EE));
1802 	return sprintf(buf, "%d\n", !enable);
1803 }
1804 
1805 static ssize_t store_energy_efficiency(struct kobject *a, struct kobj_attribute *b,
1806 				       const char *buf, size_t count)
1807 {
1808 	bool input;
1809 	int ret;
1810 
1811 	ret = kstrtobool(buf, &input);
1812 	if (ret)
1813 		return ret;
1814 
1815 	set_power_ctl_ee_state(input);
1816 
1817 	return count;
1818 }
1819 
1820 show_one(max_perf_pct, max_perf_pct);
1821 show_one(min_perf_pct, min_perf_pct);
1822 
1823 define_one_global_rw(status);
1824 define_one_global_rw(no_turbo);
1825 define_one_global_rw(max_perf_pct);
1826 define_one_global_rw(min_perf_pct);
1827 define_one_global_ro(turbo_pct);
1828 define_one_global_ro(num_pstates);
1829 define_one_global_rw(hwp_dynamic_boost);
1830 define_one_global_rw(energy_efficiency);
1831 
1832 static struct attribute *intel_pstate_attributes[] = {
1833 	&status.attr,
1834 	&no_turbo.attr,
1835 	NULL
1836 };
1837 
1838 static const struct attribute_group intel_pstate_attr_group = {
1839 	.attrs = intel_pstate_attributes,
1840 };
1841 
1842 static const struct x86_cpu_id intel_pstate_cpu_ee_disable_ids[];
1843 
1844 static struct kobject *intel_pstate_kobject;
1845 
1846 static void __init intel_pstate_sysfs_expose_params(void)
1847 {
1848 	struct device *dev_root = bus_get_dev_root(&cpu_subsys);
1849 	int rc;
1850 
1851 	if (dev_root) {
1852 		intel_pstate_kobject = kobject_create_and_add("intel_pstate", &dev_root->kobj);
1853 		put_device(dev_root);
1854 	}
1855 	if (WARN_ON(!intel_pstate_kobject))
1856 		return;
1857 
1858 	rc = sysfs_create_group(intel_pstate_kobject, &intel_pstate_attr_group);
1859 	if (WARN_ON(rc))
1860 		return;
1861 
1862 	if (!boot_cpu_has(X86_FEATURE_HYBRID_CPU)) {
1863 		rc = sysfs_create_file(intel_pstate_kobject, &turbo_pct.attr);
1864 		WARN_ON(rc);
1865 
1866 		rc = sysfs_create_file(intel_pstate_kobject, &num_pstates.attr);
1867 		WARN_ON(rc);
1868 	}
1869 
1870 	/*
1871 	 * If per cpu limits are enforced there are no global limits, so
1872 	 * return without creating max/min_perf_pct attributes
1873 	 */
1874 	if (per_cpu_limits)
1875 		return;
1876 
1877 	rc = sysfs_create_file(intel_pstate_kobject, &max_perf_pct.attr);
1878 	WARN_ON(rc);
1879 
1880 	rc = sysfs_create_file(intel_pstate_kobject, &min_perf_pct.attr);
1881 	WARN_ON(rc);
1882 
1883 	if (x86_match_cpu(intel_pstate_cpu_ee_disable_ids)) {
1884 		rc = sysfs_create_file(intel_pstate_kobject, &energy_efficiency.attr);
1885 		WARN_ON(rc);
1886 	}
1887 }
1888 
1889 static void __init intel_pstate_sysfs_remove(void)
1890 {
1891 	if (!intel_pstate_kobject)
1892 		return;
1893 
1894 	sysfs_remove_group(intel_pstate_kobject, &intel_pstate_attr_group);
1895 
1896 	if (!boot_cpu_has(X86_FEATURE_HYBRID_CPU)) {
1897 		sysfs_remove_file(intel_pstate_kobject, &num_pstates.attr);
1898 		sysfs_remove_file(intel_pstate_kobject, &turbo_pct.attr);
1899 	}
1900 
1901 	if (!per_cpu_limits) {
1902 		sysfs_remove_file(intel_pstate_kobject, &max_perf_pct.attr);
1903 		sysfs_remove_file(intel_pstate_kobject, &min_perf_pct.attr);
1904 
1905 		if (x86_match_cpu(intel_pstate_cpu_ee_disable_ids))
1906 			sysfs_remove_file(intel_pstate_kobject, &energy_efficiency.attr);
1907 	}
1908 
1909 	kobject_put(intel_pstate_kobject);
1910 }
1911 
1912 static void intel_pstate_sysfs_expose_hwp_dynamic_boost(void)
1913 {
1914 	int rc;
1915 
1916 	if (!hwp_active)
1917 		return;
1918 
1919 	rc = sysfs_create_file(intel_pstate_kobject, &hwp_dynamic_boost.attr);
1920 	WARN_ON_ONCE(rc);
1921 }
1922 
1923 static void intel_pstate_sysfs_hide_hwp_dynamic_boost(void)
1924 {
1925 	if (!hwp_active)
1926 		return;
1927 
1928 	sysfs_remove_file(intel_pstate_kobject, &hwp_dynamic_boost.attr);
1929 }
1930 
1931 /************************** sysfs end ************************/
1932 
1933 static void intel_pstate_notify_work(struct work_struct *work)
1934 {
1935 	struct cpudata *cpudata =
1936 		container_of(to_delayed_work(work), struct cpudata, hwp_notify_work);
1937 
1938 	if (intel_pstate_update_max_freq(cpudata)) {
1939 		/*
1940 		 * The driver will not be unregistered while this function is
1941 		 * running, so update the capacity without acquiring the driver
1942 		 * lock.
1943 		 */
1944 		hybrid_update_capacity(cpudata);
1945 	}
1946 
1947 	wrmsrq_on_cpu(cpudata->cpu, MSR_HWP_STATUS, 0);
1948 }
1949 
1950 static DEFINE_RAW_SPINLOCK(hwp_notify_lock);
1951 static cpumask_t hwp_intr_enable_mask;
1952 
1953 #define HWP_GUARANTEED_PERF_CHANGE_STATUS      BIT(0)
1954 #define HWP_HIGHEST_PERF_CHANGE_STATUS         BIT(3)
1955 
1956 void notify_hwp_interrupt(void)
1957 {
1958 	unsigned int this_cpu = smp_processor_id();
1959 	u64 value, status_mask;
1960 	unsigned long flags;
1961 
1962 	if (!hwp_active || !cpu_feature_enabled(X86_FEATURE_HWP_NOTIFY))
1963 		return;
1964 
1965 	status_mask = HWP_GUARANTEED_PERF_CHANGE_STATUS;
1966 	if (cpu_feature_enabled(X86_FEATURE_HWP_HIGHEST_PERF_CHANGE))
1967 		status_mask |= HWP_HIGHEST_PERF_CHANGE_STATUS;
1968 
1969 	rdmsrq_safe(MSR_HWP_STATUS, &value);
1970 	if (!(value & status_mask))
1971 		return;
1972 
1973 	raw_spin_lock_irqsave(&hwp_notify_lock, flags);
1974 
1975 	if (!cpumask_test_cpu(this_cpu, &hwp_intr_enable_mask))
1976 		goto ack_intr;
1977 
1978 	schedule_delayed_work(&all_cpu_data[this_cpu]->hwp_notify_work,
1979 			      msecs_to_jiffies(10));
1980 
1981 	raw_spin_unlock_irqrestore(&hwp_notify_lock, flags);
1982 
1983 	return;
1984 
1985 ack_intr:
1986 	wrmsrq_safe(MSR_HWP_STATUS, 0);
1987 	raw_spin_unlock_irqrestore(&hwp_notify_lock, flags);
1988 }
1989 
1990 static void intel_pstate_disable_hwp_interrupt(struct cpudata *cpudata)
1991 {
1992 	bool cancel_work;
1993 
1994 	if (!cpu_feature_enabled(X86_FEATURE_HWP_NOTIFY))
1995 		return;
1996 
1997 	/* wrmsrq_on_cpu has to be outside spinlock as this can result in IPC */
1998 	wrmsrq_on_cpu(cpudata->cpu, MSR_HWP_INTERRUPT, 0x00);
1999 
2000 	raw_spin_lock_irq(&hwp_notify_lock);
2001 	cancel_work = cpumask_test_and_clear_cpu(cpudata->cpu, &hwp_intr_enable_mask);
2002 	raw_spin_unlock_irq(&hwp_notify_lock);
2003 
2004 	if (cancel_work)
2005 		cancel_delayed_work_sync(&cpudata->hwp_notify_work);
2006 }
2007 
2008 #define HWP_GUARANTEED_PERF_CHANGE_REQ BIT(0)
2009 #define HWP_HIGHEST_PERF_CHANGE_REQ    BIT(2)
2010 
2011 static void intel_pstate_enable_hwp_interrupt(struct cpudata *cpudata)
2012 {
2013 	/* Enable HWP notification interrupt for performance change */
2014 	if (boot_cpu_has(X86_FEATURE_HWP_NOTIFY)) {
2015 		u64 interrupt_mask = HWP_GUARANTEED_PERF_CHANGE_REQ;
2016 
2017 		raw_spin_lock_irq(&hwp_notify_lock);
2018 		INIT_DELAYED_WORK(&cpudata->hwp_notify_work, intel_pstate_notify_work);
2019 		cpumask_set_cpu(cpudata->cpu, &hwp_intr_enable_mask);
2020 		raw_spin_unlock_irq(&hwp_notify_lock);
2021 
2022 		if (cpu_feature_enabled(X86_FEATURE_HWP_HIGHEST_PERF_CHANGE))
2023 			interrupt_mask |= HWP_HIGHEST_PERF_CHANGE_REQ;
2024 
2025 		/* wrmsrq_on_cpu has to be outside spinlock as this can result in IPC */
2026 		wrmsrq_on_cpu(cpudata->cpu, MSR_HWP_INTERRUPT, interrupt_mask);
2027 		wrmsrq_on_cpu(cpudata->cpu, MSR_HWP_STATUS, 0);
2028 	}
2029 }
2030 
2031 static void intel_pstate_update_epp_defaults(struct cpudata *cpudata)
2032 {
2033 	cpudata->epp_default = intel_pstate_get_epp(cpudata, 0);
2034 
2035 	/*
2036 	 * If the EPP is set by firmware, which means that firmware enabled HWP
2037 	 * - Is equal or less than 0x80 (default balance_perf EPP)
2038 	 * - But less performance oriented than performance EPP
2039 	 *   then use this as new balance_perf EPP.
2040 	 */
2041 	if (hwp_forced && cpudata->epp_default <= HWP_EPP_BALANCE_PERFORMANCE &&
2042 	    cpudata->epp_default > HWP_EPP_PERFORMANCE) {
2043 		epp_values[EPP_INDEX_BALANCE_PERFORMANCE] = cpudata->epp_default;
2044 		return;
2045 	}
2046 
2047 	/*
2048 	 * If this CPU gen doesn't call for change in balance_perf
2049 	 * EPP return.
2050 	 */
2051 	if (epp_values[EPP_INDEX_BALANCE_PERFORMANCE] == HWP_EPP_BALANCE_PERFORMANCE)
2052 		return;
2053 
2054 	/*
2055 	 * Use hard coded value per gen to update the balance_perf
2056 	 * and default EPP.
2057 	 */
2058 	cpudata->epp_default = epp_values[EPP_INDEX_BALANCE_PERFORMANCE];
2059 	intel_pstate_set_epp(cpudata, cpudata->epp_default);
2060 }
2061 
2062 static void intel_pstate_hwp_enable(struct cpudata *cpudata)
2063 {
2064 	/* First disable HWP notification interrupt till we activate again */
2065 	if (boot_cpu_has(X86_FEATURE_HWP_NOTIFY))
2066 		wrmsrq_on_cpu(cpudata->cpu, MSR_HWP_INTERRUPT, 0x00);
2067 
2068 	wrmsrq_on_cpu(cpudata->cpu, MSR_PM_ENABLE, 0x1);
2069 
2070 	intel_pstate_enable_hwp_interrupt(cpudata);
2071 
2072 	if (cpudata->epp_default >= 0)
2073 		return;
2074 
2075 	intel_pstate_update_epp_defaults(cpudata);
2076 }
2077 
2078 static int atom_get_min_pstate(int not_used)
2079 {
2080 	u64 value;
2081 
2082 	rdmsrq(MSR_ATOM_CORE_RATIOS, value);
2083 	return (value >> 8) & 0x7F;
2084 }
2085 
2086 static int atom_get_max_pstate(int not_used)
2087 {
2088 	u64 value;
2089 
2090 	rdmsrq(MSR_ATOM_CORE_RATIOS, value);
2091 	return (value >> 16) & 0x7F;
2092 }
2093 
2094 static int atom_get_turbo_pstate(int not_used)
2095 {
2096 	u64 value;
2097 
2098 	rdmsrq(MSR_ATOM_CORE_TURBO_RATIOS, value);
2099 	return value & 0x7F;
2100 }
2101 
2102 static u64 atom_get_val(struct cpudata *cpudata, int pstate)
2103 {
2104 	u64 val;
2105 	int32_t vid_fp;
2106 	u32 vid;
2107 
2108 	val = (u64)pstate << 8;
2109 	if (READ_ONCE(global.no_turbo) && !READ_ONCE(global.turbo_disabled))
2110 		val |= (u64)1 << 32;
2111 
2112 	vid_fp = cpudata->vid.min + mul_fp(
2113 		int_tofp(pstate - cpudata->pstate.min_pstate),
2114 		cpudata->vid.ratio);
2115 
2116 	vid_fp = clamp_t(int32_t, vid_fp, cpudata->vid.min, cpudata->vid.max);
2117 	vid = ceiling_fp(vid_fp);
2118 
2119 	if (pstate > cpudata->pstate.max_pstate)
2120 		vid = cpudata->vid.turbo;
2121 
2122 	return val | vid;
2123 }
2124 
2125 static int silvermont_get_scaling(void)
2126 {
2127 	u64 value;
2128 	int i;
2129 	/* Defined in Table 35-6 from SDM (Sept 2015) */
2130 	static int silvermont_freq_table[] = {
2131 		83300, 100000, 133300, 116700, 80000};
2132 
2133 	rdmsrq(MSR_FSB_FREQ, value);
2134 	i = value & 0x7;
2135 	WARN_ON(i > 4);
2136 
2137 	return silvermont_freq_table[i];
2138 }
2139 
2140 static int airmont_get_scaling(void)
2141 {
2142 	u64 value;
2143 	int i;
2144 	/* Defined in Table 35-10 from SDM (Sept 2015) */
2145 	static int airmont_freq_table[] = {
2146 		83300, 100000, 133300, 116700, 80000,
2147 		93300, 90000, 88900, 87500};
2148 
2149 	rdmsrq(MSR_FSB_FREQ, value);
2150 	i = value & 0xF;
2151 	WARN_ON(i > 8);
2152 
2153 	return airmont_freq_table[i];
2154 }
2155 
2156 static void atom_get_vid(struct cpudata *cpudata)
2157 {
2158 	u64 value;
2159 
2160 	rdmsrq(MSR_ATOM_CORE_VIDS, value);
2161 	cpudata->vid.min = int_tofp((value >> 8) & 0x7f);
2162 	cpudata->vid.max = int_tofp((value >> 16) & 0x7f);
2163 	cpudata->vid.ratio = div_fp(
2164 		cpudata->vid.max - cpudata->vid.min,
2165 		int_tofp(cpudata->pstate.max_pstate -
2166 			cpudata->pstate.min_pstate));
2167 
2168 	rdmsrq(MSR_ATOM_CORE_TURBO_VIDS, value);
2169 	cpudata->vid.turbo = value & 0x7f;
2170 }
2171 
2172 static int core_get_min_pstate(int cpu)
2173 {
2174 	u64 value;
2175 
2176 	rdmsrq_on_cpu(cpu, MSR_PLATFORM_INFO, &value);
2177 	return (value >> 40) & 0xFF;
2178 }
2179 
2180 static int core_get_max_pstate_physical(int cpu)
2181 {
2182 	u64 value;
2183 
2184 	rdmsrq_on_cpu(cpu, MSR_PLATFORM_INFO, &value);
2185 	return (value >> 8) & 0xFF;
2186 }
2187 
2188 static int core_get_tdp_ratio(int cpu, u64 plat_info)
2189 {
2190 	/* Check how many TDP levels present */
2191 	if (plat_info & 0x600000000) {
2192 		u64 tdp_ctrl;
2193 		u64 tdp_ratio;
2194 		int tdp_msr;
2195 		int err;
2196 
2197 		/* Get the TDP level (0, 1, 2) to get ratios */
2198 		err = rdmsrq_safe_on_cpu(cpu, MSR_CONFIG_TDP_CONTROL, &tdp_ctrl);
2199 		if (err)
2200 			return err;
2201 
2202 		/* TDP MSR are continuous starting at 0x648 */
2203 		tdp_msr = MSR_CONFIG_TDP_NOMINAL + (tdp_ctrl & 0x03);
2204 		err = rdmsrq_safe_on_cpu(cpu, tdp_msr, &tdp_ratio);
2205 		if (err)
2206 			return err;
2207 
2208 		/* For level 1 and 2, bits[23:16] contain the ratio */
2209 		if (tdp_ctrl & 0x03)
2210 			tdp_ratio >>= 16;
2211 
2212 		tdp_ratio &= 0xff; /* ratios are only 8 bits long */
2213 		pr_debug("tdp_ratio %x\n", (int)tdp_ratio);
2214 
2215 		return (int)tdp_ratio;
2216 	}
2217 
2218 	return -ENXIO;
2219 }
2220 
2221 static int core_get_max_pstate(int cpu)
2222 {
2223 	u64 tar;
2224 	u64 plat_info;
2225 	int max_pstate;
2226 	int tdp_ratio;
2227 	int err;
2228 
2229 	rdmsrq_on_cpu(cpu, MSR_PLATFORM_INFO, &plat_info);
2230 	max_pstate = (plat_info >> 8) & 0xFF;
2231 
2232 	tdp_ratio = core_get_tdp_ratio(cpu, plat_info);
2233 	if (tdp_ratio <= 0)
2234 		return max_pstate;
2235 
2236 	if (hwp_active) {
2237 		/* Turbo activation ratio is not used on HWP platforms */
2238 		return tdp_ratio;
2239 	}
2240 
2241 	err = rdmsrq_safe_on_cpu(cpu, MSR_TURBO_ACTIVATION_RATIO, &tar);
2242 	if (!err) {
2243 		int tar_levels;
2244 
2245 		/* Do some sanity checking for safety */
2246 		tar_levels = tar & 0xff;
2247 		if (tdp_ratio - 1 == tar_levels) {
2248 			max_pstate = tar_levels;
2249 			pr_debug("max_pstate=TAC %x\n", max_pstate);
2250 		}
2251 	}
2252 
2253 	return max_pstate;
2254 }
2255 
2256 static int core_get_turbo_pstate(int cpu)
2257 {
2258 	u64 value;
2259 	int nont, ret;
2260 
2261 	rdmsrq_on_cpu(cpu, MSR_TURBO_RATIO_LIMIT, &value);
2262 	nont = core_get_max_pstate(cpu);
2263 	ret = (value) & 255;
2264 	if (ret <= nont)
2265 		ret = nont;
2266 	return ret;
2267 }
2268 
2269 static u64 core_get_val(struct cpudata *cpudata, int pstate)
2270 {
2271 	u64 val;
2272 
2273 	val = (u64)pstate << 8;
2274 	if (READ_ONCE(global.no_turbo) && !READ_ONCE(global.turbo_disabled))
2275 		val |= (u64)1 << 32;
2276 
2277 	return val;
2278 }
2279 
2280 static int knl_get_aperf_mperf_shift(void)
2281 {
2282 	return 10;
2283 }
2284 
2285 static int knl_get_turbo_pstate(int cpu)
2286 {
2287 	u64 value;
2288 	int nont, ret;
2289 
2290 	rdmsrq_on_cpu(cpu, MSR_TURBO_RATIO_LIMIT, &value);
2291 	nont = core_get_max_pstate(cpu);
2292 	ret = (((value) >> 8) & 0xFF);
2293 	if (ret <= nont)
2294 		ret = nont;
2295 	return ret;
2296 }
2297 
2298 static int hwp_get_cpu_scaling(int cpu)
2299 {
2300 	if (hybrid_scaling_factor) {
2301 		struct cpuinfo_x86 *c = &cpu_data(cpu);
2302 		u8 cpu_type = c->topo.intel_type;
2303 
2304 		/*
2305 		 * Return the hybrid scaling factor for P-cores and use the
2306 		 * default core scaling for E-cores.
2307 		 */
2308 		if (cpu_type == INTEL_CPU_TYPE_CORE)
2309 			return hybrid_scaling_factor;
2310 
2311 		if (cpu_type == INTEL_CPU_TYPE_ATOM)
2312 			return core_get_scaling();
2313 	}
2314 
2315 	/* Use core scaling on non-hybrid systems. */
2316 	if (!cpu_feature_enabled(X86_FEATURE_HYBRID_CPU))
2317 		return core_get_scaling();
2318 
2319 	/*
2320 	 * The system is hybrid, but the hybrid scaling factor is not known or
2321 	 * the CPU type is not one of the above, so use CPPC to compute the
2322 	 * scaling factor for this CPU.
2323 	 */
2324 	return intel_pstate_cppc_get_scaling(cpu);
2325 }
2326 
2327 static void intel_pstate_set_pstate(struct cpudata *cpu, int pstate)
2328 {
2329 	trace_cpu_frequency(pstate * cpu->pstate.scaling, cpu->cpu);
2330 	cpu->pstate.current_pstate = pstate;
2331 	/*
2332 	 * Generally, there is no guarantee that this code will always run on
2333 	 * the CPU being updated, so force the register update to run on the
2334 	 * right CPU.
2335 	 */
2336 	wrmsrq_on_cpu(cpu->cpu, MSR_IA32_PERF_CTL,
2337 		      pstate_funcs.get_val(cpu, pstate));
2338 }
2339 
2340 static void intel_pstate_set_min_pstate(struct cpudata *cpu)
2341 {
2342 	intel_pstate_set_pstate(cpu, cpu->pstate.min_pstate);
2343 }
2344 
2345 static void intel_pstate_get_cpu_pstates(struct cpudata *cpu)
2346 {
2347 	int perf_ctl_max_phys = pstate_funcs.get_max_physical(cpu->cpu);
2348 	int perf_ctl_scaling = pstate_funcs.get_scaling();
2349 
2350 	cpu->pstate.min_pstate = pstate_funcs.get_min(cpu->cpu);
2351 	cpu->pstate.max_pstate_physical = perf_ctl_max_phys;
2352 	cpu->pstate.perf_ctl_scaling = perf_ctl_scaling;
2353 
2354 	if (hwp_active && !hwp_mode_bdw) {
2355 		__intel_pstate_get_hwp_cap(cpu);
2356 
2357 		if (pstate_funcs.get_cpu_scaling) {
2358 			cpu->pstate.scaling = pstate_funcs.get_cpu_scaling(cpu->cpu);
2359 			if (cpu->pstate.scaling != perf_ctl_scaling) {
2360 				intel_pstate_hybrid_hwp_adjust(cpu);
2361 				hwp_is_hybrid = true;
2362 			}
2363 		} else {
2364 			cpu->pstate.scaling = perf_ctl_scaling;
2365 		}
2366 		/*
2367 		 * If the CPU is going online for the first time and it was
2368 		 * offline initially, asym capacity scaling needs to be updated.
2369 		 */
2370 		hybrid_update_capacity(cpu);
2371 	} else {
2372 		cpu->pstate.scaling = perf_ctl_scaling;
2373 		cpu->pstate.max_pstate = pstate_funcs.get_max(cpu->cpu);
2374 		cpu->pstate.turbo_pstate = pstate_funcs.get_turbo(cpu->cpu);
2375 	}
2376 
2377 	if (cpu->pstate.scaling == perf_ctl_scaling) {
2378 		cpu->pstate.min_freq = cpu->pstate.min_pstate * perf_ctl_scaling;
2379 		cpu->pstate.max_freq = cpu->pstate.max_pstate * perf_ctl_scaling;
2380 		cpu->pstate.turbo_freq = cpu->pstate.turbo_pstate * perf_ctl_scaling;
2381 	}
2382 
2383 	if (pstate_funcs.get_aperf_mperf_shift)
2384 		cpu->aperf_mperf_shift = pstate_funcs.get_aperf_mperf_shift();
2385 
2386 	if (pstate_funcs.get_vid)
2387 		pstate_funcs.get_vid(cpu);
2388 
2389 	intel_pstate_set_min_pstate(cpu);
2390 }
2391 
2392 /*
2393  * Long hold time will keep high perf limits for long time,
2394  * which negatively impacts perf/watt for some workloads,
2395  * like specpower. 3ms is based on experiements on some
2396  * workoads.
2397  */
2398 static int hwp_boost_hold_time_ns = 3 * NSEC_PER_MSEC;
2399 
2400 static inline void intel_pstate_hwp_boost_up(struct cpudata *cpu)
2401 {
2402 	u64 hwp_req = READ_ONCE(cpu->hwp_req_cached);
2403 	u64 hwp_cap = READ_ONCE(cpu->hwp_cap_cached);
2404 	u32 max_limit = (hwp_req & 0xff00) >> 8;
2405 	u32 min_limit = (hwp_req & 0xff);
2406 	u32 boost_level1;
2407 
2408 	/*
2409 	 * Cases to consider (User changes via sysfs or boot time):
2410 	 * If, P0 (Turbo max) = P1 (Guaranteed max) = min:
2411 	 *	No boost, return.
2412 	 * If, P0 (Turbo max) > P1 (Guaranteed max) = min:
2413 	 *     Should result in one level boost only for P0.
2414 	 * If, P0 (Turbo max) = P1 (Guaranteed max) > min:
2415 	 *     Should result in two level boost:
2416 	 *         (min + p1)/2 and P1.
2417 	 * If, P0 (Turbo max) > P1 (Guaranteed max) > min:
2418 	 *     Should result in three level boost:
2419 	 *        (min + p1)/2, P1 and P0.
2420 	 */
2421 
2422 	/* If max and min are equal or already at max, nothing to boost */
2423 	if (max_limit == min_limit || cpu->hwp_boost_min >= max_limit)
2424 		return;
2425 
2426 	if (!cpu->hwp_boost_min)
2427 		cpu->hwp_boost_min = min_limit;
2428 
2429 	/* level at half way mark between min and guranteed */
2430 	boost_level1 = (HWP_GUARANTEED_PERF(hwp_cap) + min_limit) >> 1;
2431 
2432 	if (cpu->hwp_boost_min < boost_level1)
2433 		cpu->hwp_boost_min = boost_level1;
2434 	else if (cpu->hwp_boost_min < HWP_GUARANTEED_PERF(hwp_cap))
2435 		cpu->hwp_boost_min = HWP_GUARANTEED_PERF(hwp_cap);
2436 	else if (cpu->hwp_boost_min == HWP_GUARANTEED_PERF(hwp_cap) &&
2437 		 max_limit != HWP_GUARANTEED_PERF(hwp_cap))
2438 		cpu->hwp_boost_min = max_limit;
2439 	else
2440 		return;
2441 
2442 	hwp_req = (hwp_req & ~GENMASK_ULL(7, 0)) | cpu->hwp_boost_min;
2443 	wrmsrq(MSR_HWP_REQUEST, hwp_req);
2444 	cpu->last_update = cpu->sample.time;
2445 }
2446 
2447 static inline void intel_pstate_hwp_boost_down(struct cpudata *cpu)
2448 {
2449 	if (cpu->hwp_boost_min) {
2450 		bool expired;
2451 
2452 		/* Check if we are idle for hold time to boost down */
2453 		expired = time_after64(cpu->sample.time, cpu->last_update +
2454 				       hwp_boost_hold_time_ns);
2455 		if (expired) {
2456 			wrmsrq(MSR_HWP_REQUEST, cpu->hwp_req_cached);
2457 			cpu->hwp_boost_min = 0;
2458 		}
2459 	}
2460 	cpu->last_update = cpu->sample.time;
2461 }
2462 
2463 static inline void intel_pstate_update_util_hwp_local(struct cpudata *cpu,
2464 						      u64 time)
2465 {
2466 	cpu->sample.time = time;
2467 
2468 	if (cpu->sched_flags & SCHED_CPUFREQ_IOWAIT) {
2469 		bool do_io = false;
2470 
2471 		cpu->sched_flags = 0;
2472 		/*
2473 		 * Set iowait_boost flag and update time. Since IO WAIT flag
2474 		 * is set all the time, we can't just conclude that there is
2475 		 * some IO bound activity is scheduled on this CPU with just
2476 		 * one occurrence. If we receive at least two in two
2477 		 * consecutive ticks, then we treat as boost candidate.
2478 		 */
2479 		if (time_before64(time, cpu->last_io_update + 2 * TICK_NSEC))
2480 			do_io = true;
2481 
2482 		cpu->last_io_update = time;
2483 
2484 		if (do_io)
2485 			intel_pstate_hwp_boost_up(cpu);
2486 
2487 	} else {
2488 		intel_pstate_hwp_boost_down(cpu);
2489 	}
2490 }
2491 
2492 static inline void intel_pstate_update_util_hwp(struct update_util_data *data,
2493 						u64 time, unsigned int flags)
2494 {
2495 	struct cpudata *cpu = container_of(data, struct cpudata, update_util);
2496 
2497 	cpu->sched_flags |= flags;
2498 
2499 	if (smp_processor_id() == cpu->cpu)
2500 		intel_pstate_update_util_hwp_local(cpu, time);
2501 }
2502 
2503 static inline void intel_pstate_calc_avg_perf(struct cpudata *cpu)
2504 {
2505 	struct sample *sample = &cpu->sample;
2506 
2507 	sample->core_avg_perf = div_ext_fp(sample->aperf, sample->mperf);
2508 }
2509 
2510 static inline bool intel_pstate_sample(struct cpudata *cpu, u64 time)
2511 {
2512 	u64 aperf, mperf;
2513 	unsigned long flags;
2514 	u64 tsc;
2515 
2516 	local_irq_save(flags);
2517 	rdmsrq(MSR_IA32_APERF, aperf);
2518 	rdmsrq(MSR_IA32_MPERF, mperf);
2519 	tsc = rdtsc();
2520 	if (cpu->prev_mperf == mperf || cpu->prev_tsc == tsc) {
2521 		local_irq_restore(flags);
2522 		return false;
2523 	}
2524 	local_irq_restore(flags);
2525 
2526 	cpu->last_sample_time = cpu->sample.time;
2527 	cpu->sample.time = time;
2528 	cpu->sample.aperf = aperf;
2529 	cpu->sample.mperf = mperf;
2530 	cpu->sample.tsc =  tsc;
2531 	cpu->sample.aperf -= cpu->prev_aperf;
2532 	cpu->sample.mperf -= cpu->prev_mperf;
2533 	cpu->sample.tsc -= cpu->prev_tsc;
2534 
2535 	cpu->prev_aperf = aperf;
2536 	cpu->prev_mperf = mperf;
2537 	cpu->prev_tsc = tsc;
2538 	/*
2539 	 * First time this function is invoked in a given cycle, all of the
2540 	 * previous sample data fields are equal to zero or stale and they must
2541 	 * be populated with meaningful numbers for things to work, so assume
2542 	 * that sample.time will always be reset before setting the utilization
2543 	 * update hook and make the caller skip the sample then.
2544 	 */
2545 	if (likely(cpu->last_sample_time)) {
2546 		intel_pstate_calc_avg_perf(cpu);
2547 		return true;
2548 	}
2549 	return false;
2550 }
2551 
2552 static inline int32_t get_avg_frequency(struct cpudata *cpu)
2553 {
2554 	return mul_ext_fp(cpu->sample.core_avg_perf, cpu_khz);
2555 }
2556 
2557 static inline int32_t get_avg_pstate(struct cpudata *cpu)
2558 {
2559 	return mul_ext_fp(cpu->pstate.max_pstate_physical,
2560 			  cpu->sample.core_avg_perf);
2561 }
2562 
2563 static inline int32_t get_target_pstate(struct cpudata *cpu)
2564 {
2565 	struct sample *sample = &cpu->sample;
2566 	int32_t busy_frac;
2567 	int target, avg_pstate;
2568 
2569 	busy_frac = div_fp(sample->mperf << cpu->aperf_mperf_shift,
2570 			   sample->tsc);
2571 
2572 	if (busy_frac < cpu->iowait_boost)
2573 		busy_frac = cpu->iowait_boost;
2574 
2575 	sample->busy_scaled = busy_frac * 100;
2576 
2577 	target = READ_ONCE(global.no_turbo) ?
2578 			cpu->pstate.max_pstate : cpu->pstate.turbo_pstate;
2579 	target += target >> 2;
2580 	target = mul_fp(target, busy_frac);
2581 	if (target < cpu->pstate.min_pstate)
2582 		target = cpu->pstate.min_pstate;
2583 
2584 	/*
2585 	 * If the average P-state during the previous cycle was higher than the
2586 	 * current target, add 50% of the difference to the target to reduce
2587 	 * possible performance oscillations and offset possible performance
2588 	 * loss related to moving the workload from one CPU to another within
2589 	 * a package/module.
2590 	 */
2591 	avg_pstate = get_avg_pstate(cpu);
2592 	if (avg_pstate > target)
2593 		target += (avg_pstate - target) >> 1;
2594 
2595 	return target;
2596 }
2597 
2598 static int intel_pstate_prepare_request(struct cpudata *cpu, int pstate)
2599 {
2600 	int min_pstate = max(cpu->pstate.min_pstate, cpu->min_perf_ratio);
2601 	int max_pstate = max(min_pstate, cpu->max_perf_ratio);
2602 
2603 	return clamp_t(int, pstate, min_pstate, max_pstate);
2604 }
2605 
2606 static void intel_pstate_update_pstate(struct cpudata *cpu, int pstate)
2607 {
2608 	if (pstate == cpu->pstate.current_pstate)
2609 		return;
2610 
2611 	cpu->pstate.current_pstate = pstate;
2612 	wrmsrq(MSR_IA32_PERF_CTL, pstate_funcs.get_val(cpu, pstate));
2613 }
2614 
2615 static void intel_pstate_adjust_pstate(struct cpudata *cpu)
2616 {
2617 	int from = cpu->pstate.current_pstate;
2618 	struct sample *sample;
2619 	int target_pstate;
2620 
2621 	target_pstate = get_target_pstate(cpu);
2622 	target_pstate = intel_pstate_prepare_request(cpu, target_pstate);
2623 	trace_cpu_frequency(target_pstate * cpu->pstate.scaling, cpu->cpu);
2624 	intel_pstate_update_pstate(cpu, target_pstate);
2625 
2626 	sample = &cpu->sample;
2627 	trace_pstate_sample(mul_ext_fp(100, sample->core_avg_perf),
2628 		fp_toint(sample->busy_scaled),
2629 		from,
2630 		cpu->pstate.current_pstate,
2631 		sample->mperf,
2632 		sample->aperf,
2633 		sample->tsc,
2634 		get_avg_frequency(cpu),
2635 		fp_toint(cpu->iowait_boost * 100));
2636 }
2637 
2638 static void intel_pstate_update_util(struct update_util_data *data, u64 time,
2639 				     unsigned int flags)
2640 {
2641 	struct cpudata *cpu = container_of(data, struct cpudata, update_util);
2642 	u64 delta_ns;
2643 
2644 	/* Don't allow remote callbacks */
2645 	if (smp_processor_id() != cpu->cpu)
2646 		return;
2647 
2648 	delta_ns = time - cpu->last_update;
2649 	if (flags & SCHED_CPUFREQ_IOWAIT) {
2650 		/* Start over if the CPU may have been idle. */
2651 		if (delta_ns > TICK_NSEC) {
2652 			cpu->iowait_boost = ONE_EIGHTH_FP;
2653 		} else if (cpu->iowait_boost >= ONE_EIGHTH_FP) {
2654 			cpu->iowait_boost <<= 1;
2655 			if (cpu->iowait_boost > int_tofp(1))
2656 				cpu->iowait_boost = int_tofp(1);
2657 		} else {
2658 			cpu->iowait_boost = ONE_EIGHTH_FP;
2659 		}
2660 	} else if (cpu->iowait_boost) {
2661 		/* Clear iowait_boost if the CPU may have been idle. */
2662 		if (delta_ns > TICK_NSEC)
2663 			cpu->iowait_boost = 0;
2664 		else
2665 			cpu->iowait_boost >>= 1;
2666 	}
2667 	cpu->last_update = time;
2668 	delta_ns = time - cpu->sample.time;
2669 	if ((s64)delta_ns < INTEL_PSTATE_SAMPLING_INTERVAL)
2670 		return;
2671 
2672 	if (intel_pstate_sample(cpu, time))
2673 		intel_pstate_adjust_pstate(cpu);
2674 }
2675 
2676 static struct pstate_funcs core_funcs = {
2677 	.get_max = core_get_max_pstate,
2678 	.get_max_physical = core_get_max_pstate_physical,
2679 	.get_min = core_get_min_pstate,
2680 	.get_turbo = core_get_turbo_pstate,
2681 	.get_scaling = core_get_scaling,
2682 	.get_val = core_get_val,
2683 };
2684 
2685 static const struct pstate_funcs silvermont_funcs = {
2686 	.get_max = atom_get_max_pstate,
2687 	.get_max_physical = atom_get_max_pstate,
2688 	.get_min = atom_get_min_pstate,
2689 	.get_turbo = atom_get_turbo_pstate,
2690 	.get_val = atom_get_val,
2691 	.get_scaling = silvermont_get_scaling,
2692 	.get_vid = atom_get_vid,
2693 };
2694 
2695 static const struct pstate_funcs airmont_funcs = {
2696 	.get_max = atom_get_max_pstate,
2697 	.get_max_physical = atom_get_max_pstate,
2698 	.get_min = atom_get_min_pstate,
2699 	.get_turbo = atom_get_turbo_pstate,
2700 	.get_val = atom_get_val,
2701 	.get_scaling = airmont_get_scaling,
2702 	.get_vid = atom_get_vid,
2703 };
2704 
2705 static const struct pstate_funcs knl_funcs = {
2706 	.get_max = core_get_max_pstate,
2707 	.get_max_physical = core_get_max_pstate_physical,
2708 	.get_min = core_get_min_pstate,
2709 	.get_turbo = knl_get_turbo_pstate,
2710 	.get_aperf_mperf_shift = knl_get_aperf_mperf_shift,
2711 	.get_scaling = core_get_scaling,
2712 	.get_val = core_get_val,
2713 };
2714 
2715 #define X86_MATCH(vfm, policy)					 \
2716 	X86_MATCH_VFM_FEATURE(vfm, X86_FEATURE_APERFMPERF, &policy)
2717 
2718 static const struct x86_cpu_id intel_pstate_cpu_ids[] = {
2719 	X86_MATCH(INTEL_SANDYBRIDGE,		core_funcs),
2720 	X86_MATCH(INTEL_SANDYBRIDGE_X,		core_funcs),
2721 	X86_MATCH(INTEL_ATOM_SILVERMONT,	silvermont_funcs),
2722 	X86_MATCH(INTEL_IVYBRIDGE,		core_funcs),
2723 	X86_MATCH(INTEL_HASWELL,		core_funcs),
2724 	X86_MATCH(INTEL_BROADWELL,		core_funcs),
2725 	X86_MATCH(INTEL_IVYBRIDGE_X,		core_funcs),
2726 	X86_MATCH(INTEL_HASWELL_X,		core_funcs),
2727 	X86_MATCH(INTEL_HASWELL_L,		core_funcs),
2728 	X86_MATCH(INTEL_HASWELL_G,		core_funcs),
2729 	X86_MATCH(INTEL_BROADWELL_G,		core_funcs),
2730 	X86_MATCH(INTEL_ATOM_AIRMONT,		airmont_funcs),
2731 	X86_MATCH(INTEL_SKYLAKE_L,		core_funcs),
2732 	X86_MATCH(INTEL_BROADWELL_X,		core_funcs),
2733 	X86_MATCH(INTEL_SKYLAKE,		core_funcs),
2734 	X86_MATCH(INTEL_BROADWELL_D,		core_funcs),
2735 	X86_MATCH(INTEL_XEON_PHI_KNL,		knl_funcs),
2736 	X86_MATCH(INTEL_XEON_PHI_KNM,		knl_funcs),
2737 	X86_MATCH(INTEL_ATOM_GOLDMONT,		core_funcs),
2738 	X86_MATCH(INTEL_ATOM_GOLDMONT_PLUS,	core_funcs),
2739 	X86_MATCH(INTEL_SKYLAKE_X,		core_funcs),
2740 	X86_MATCH(INTEL_COMETLAKE,		core_funcs),
2741 	X86_MATCH(INTEL_ICELAKE_X,		core_funcs),
2742 	X86_MATCH(INTEL_TIGERLAKE,		core_funcs),
2743 	X86_MATCH(INTEL_SAPPHIRERAPIDS_X,	core_funcs),
2744 	X86_MATCH(INTEL_EMERALDRAPIDS_X,	core_funcs),
2745 	X86_MATCH(INTEL_GRANITERAPIDS_D,	core_funcs),
2746 	X86_MATCH(INTEL_GRANITERAPIDS_X,	core_funcs),
2747 	{}
2748 };
2749 MODULE_DEVICE_TABLE(x86cpu, intel_pstate_cpu_ids);
2750 
2751 #ifdef CONFIG_ACPI
2752 static const struct x86_cpu_id intel_pstate_cpu_oob_ids[] __initconst = {
2753 	X86_MATCH(INTEL_BROADWELL_D,		core_funcs),
2754 	X86_MATCH(INTEL_BROADWELL_X,		core_funcs),
2755 	X86_MATCH(INTEL_SKYLAKE_X,		core_funcs),
2756 	X86_MATCH(INTEL_ICELAKE_X,		core_funcs),
2757 	X86_MATCH(INTEL_SAPPHIRERAPIDS_X,	core_funcs),
2758 	X86_MATCH(INTEL_EMERALDRAPIDS_X,	core_funcs),
2759 	X86_MATCH(INTEL_GRANITERAPIDS_D,	core_funcs),
2760 	X86_MATCH(INTEL_GRANITERAPIDS_X,	core_funcs),
2761 	X86_MATCH(INTEL_ATOM_CRESTMONT,		core_funcs),
2762 	X86_MATCH(INTEL_ATOM_CRESTMONT_X,	core_funcs),
2763 	X86_MATCH(INTEL_ATOM_DARKMONT_X,	core_funcs),
2764 	{}
2765 };
2766 #endif
2767 
2768 static const struct x86_cpu_id intel_pstate_cpu_ee_disable_ids[] = {
2769 	X86_MATCH(INTEL_KABYLAKE,		core_funcs),
2770 	{}
2771 };
2772 
2773 static int intel_pstate_init_cpu(unsigned int cpunum)
2774 {
2775 	struct cpudata *cpu;
2776 
2777 	cpu = all_cpu_data[cpunum];
2778 
2779 	if (!cpu) {
2780 		cpu = kzalloc(sizeof(*cpu), GFP_KERNEL);
2781 		if (!cpu)
2782 			return -ENOMEM;
2783 
2784 		WRITE_ONCE(all_cpu_data[cpunum], cpu);
2785 
2786 		cpu->cpu = cpunum;
2787 
2788 		cpu->epp_default = -EINVAL;
2789 
2790 		if (hwp_active) {
2791 			intel_pstate_hwp_enable(cpu);
2792 
2793 			if (intel_pstate_acpi_pm_profile_server())
2794 				hwp_boost = true;
2795 		}
2796 	} else if (hwp_active) {
2797 		/*
2798 		 * Re-enable HWP in case this happens after a resume from ACPI
2799 		 * S3 if the CPU was offline during the whole system/resume
2800 		 * cycle.
2801 		 */
2802 		intel_pstate_hwp_reenable(cpu);
2803 	}
2804 
2805 	cpu->epp_powersave = -EINVAL;
2806 	cpu->epp_policy = CPUFREQ_POLICY_UNKNOWN;
2807 
2808 	intel_pstate_get_cpu_pstates(cpu);
2809 
2810 	pr_debug("controlling: cpu %d\n", cpunum);
2811 
2812 	return 0;
2813 }
2814 
2815 static void intel_pstate_set_update_util_hook(unsigned int cpu_num)
2816 {
2817 	struct cpudata *cpu = all_cpu_data[cpu_num];
2818 
2819 	if (hwp_active && !hwp_boost)
2820 		return;
2821 
2822 	if (cpu->update_util_set)
2823 		return;
2824 
2825 	/* Prevent intel_pstate_update_util() from using stale data. */
2826 	cpu->sample.time = 0;
2827 	cpufreq_add_update_util_hook(cpu_num, &cpu->update_util,
2828 				     (hwp_active ?
2829 				      intel_pstate_update_util_hwp :
2830 				      intel_pstate_update_util));
2831 	cpu->update_util_set = true;
2832 }
2833 
2834 static void intel_pstate_clear_update_util_hook(unsigned int cpu)
2835 {
2836 	struct cpudata *cpu_data = all_cpu_data[cpu];
2837 
2838 	if (!cpu_data->update_util_set)
2839 		return;
2840 
2841 	cpufreq_remove_update_util_hook(cpu);
2842 	cpu_data->update_util_set = false;
2843 	synchronize_rcu();
2844 }
2845 
2846 static int intel_pstate_get_max_freq(struct cpudata *cpu)
2847 {
2848 	return READ_ONCE(global.no_turbo) ?
2849 			cpu->pstate.max_freq : cpu->pstate.turbo_freq;
2850 }
2851 
2852 static void intel_pstate_update_perf_limits(struct cpudata *cpu,
2853 					    unsigned int policy_min,
2854 					    unsigned int policy_max)
2855 {
2856 	int perf_ctl_scaling = cpu->pstate.perf_ctl_scaling;
2857 	int32_t max_policy_perf, min_policy_perf;
2858 
2859 	max_policy_perf = policy_max / perf_ctl_scaling;
2860 	if (policy_max == policy_min) {
2861 		min_policy_perf = max_policy_perf;
2862 	} else {
2863 		min_policy_perf = policy_min / perf_ctl_scaling;
2864 		min_policy_perf = clamp_t(int32_t, min_policy_perf,
2865 					  0, max_policy_perf);
2866 	}
2867 
2868 	/*
2869 	 * HWP needs some special consideration, because HWP_REQUEST uses
2870 	 * abstract values to represent performance rather than pure ratios.
2871 	 */
2872 	if (hwp_active && cpu->pstate.scaling != perf_ctl_scaling) {
2873 		int freq;
2874 
2875 		freq = max_policy_perf * perf_ctl_scaling;
2876 		max_policy_perf = intel_pstate_freq_to_hwp(cpu, freq);
2877 		freq = min_policy_perf * perf_ctl_scaling;
2878 		min_policy_perf = intel_pstate_freq_to_hwp(cpu, freq);
2879 	}
2880 
2881 	pr_debug("cpu:%d min_policy_perf:%d max_policy_perf:%d\n",
2882 		 cpu->cpu, min_policy_perf, max_policy_perf);
2883 
2884 	/* Normalize user input to [min_perf, max_perf] */
2885 	if (per_cpu_limits) {
2886 		cpu->min_perf_ratio = min_policy_perf;
2887 		cpu->max_perf_ratio = max_policy_perf;
2888 	} else {
2889 		int turbo_max = cpu->pstate.turbo_pstate;
2890 		int32_t global_min, global_max;
2891 
2892 		/* Global limits are in percent of the maximum turbo P-state. */
2893 		global_max = DIV_ROUND_UP(turbo_max * global.max_perf_pct, 100);
2894 		global_min = DIV_ROUND_UP(turbo_max * global.min_perf_pct, 100);
2895 		global_min = clamp_t(int32_t, global_min, 0, global_max);
2896 
2897 		pr_debug("cpu:%d global_min:%d global_max:%d\n", cpu->cpu,
2898 			 global_min, global_max);
2899 
2900 		cpu->min_perf_ratio = max(min_policy_perf, global_min);
2901 		cpu->min_perf_ratio = min(cpu->min_perf_ratio, max_policy_perf);
2902 		cpu->max_perf_ratio = min(max_policy_perf, global_max);
2903 		cpu->max_perf_ratio = max(min_policy_perf, cpu->max_perf_ratio);
2904 
2905 		/* Make sure min_perf <= max_perf */
2906 		cpu->min_perf_ratio = min(cpu->min_perf_ratio,
2907 					  cpu->max_perf_ratio);
2908 
2909 	}
2910 	pr_debug("cpu:%d max_perf_ratio:%d min_perf_ratio:%d\n", cpu->cpu,
2911 		 cpu->max_perf_ratio,
2912 		 cpu->min_perf_ratio);
2913 }
2914 
2915 static int intel_pstate_set_policy(struct cpufreq_policy *policy)
2916 {
2917 	struct cpudata *cpu;
2918 
2919 	if (!policy->cpuinfo.max_freq)
2920 		return -ENODEV;
2921 
2922 	pr_debug("set_policy cpuinfo.max %u policy->max %u\n",
2923 		 policy->cpuinfo.max_freq, policy->max);
2924 
2925 	cpu = all_cpu_data[policy->cpu];
2926 	cpu->policy = policy->policy;
2927 
2928 	mutex_lock(&intel_pstate_limits_lock);
2929 
2930 	intel_pstate_update_perf_limits(cpu, policy->min, policy->max);
2931 
2932 	if (cpu->policy == CPUFREQ_POLICY_PERFORMANCE) {
2933 		int pstate = max(cpu->pstate.min_pstate, cpu->max_perf_ratio);
2934 
2935 		/*
2936 		 * NOHZ_FULL CPUs need this as the governor callback may not
2937 		 * be invoked on them.
2938 		 */
2939 		intel_pstate_clear_update_util_hook(policy->cpu);
2940 		intel_pstate_set_pstate(cpu, pstate);
2941 	} else {
2942 		intel_pstate_set_update_util_hook(policy->cpu);
2943 	}
2944 
2945 	if (hwp_active) {
2946 		/*
2947 		 * When hwp_boost was active before and dynamically it
2948 		 * was turned off, in that case we need to clear the
2949 		 * update util hook.
2950 		 */
2951 		if (!hwp_boost)
2952 			intel_pstate_clear_update_util_hook(policy->cpu);
2953 		intel_pstate_hwp_set(policy->cpu);
2954 	}
2955 	/*
2956 	 * policy->cur is never updated with the intel_pstate driver, but it
2957 	 * is used as a stale frequency value. So, keep it within limits.
2958 	 */
2959 	policy->cur = policy->min;
2960 
2961 	mutex_unlock(&intel_pstate_limits_lock);
2962 
2963 	return 0;
2964 }
2965 
2966 static void intel_pstate_adjust_policy_max(struct cpudata *cpu,
2967 					   struct cpufreq_policy_data *policy)
2968 {
2969 	if (!hwp_active &&
2970 	    cpu->pstate.max_pstate_physical > cpu->pstate.max_pstate &&
2971 	    policy->max < policy->cpuinfo.max_freq &&
2972 	    policy->max > cpu->pstate.max_freq) {
2973 		pr_debug("policy->max > max non turbo frequency\n");
2974 		policy->max = policy->cpuinfo.max_freq;
2975 	}
2976 }
2977 
2978 static void intel_pstate_verify_cpu_policy(struct cpudata *cpu,
2979 					   struct cpufreq_policy_data *policy)
2980 {
2981 	int max_freq;
2982 
2983 	if (hwp_active) {
2984 		intel_pstate_get_hwp_cap(cpu);
2985 		max_freq = READ_ONCE(global.no_turbo) ?
2986 				cpu->pstate.max_freq : cpu->pstate.turbo_freq;
2987 	} else {
2988 		max_freq = intel_pstate_get_max_freq(cpu);
2989 	}
2990 	cpufreq_verify_within_limits(policy, policy->cpuinfo.min_freq, max_freq);
2991 
2992 	intel_pstate_adjust_policy_max(cpu, policy);
2993 }
2994 
2995 static int intel_pstate_verify_policy(struct cpufreq_policy_data *policy)
2996 {
2997 	intel_pstate_verify_cpu_policy(all_cpu_data[policy->cpu], policy);
2998 
2999 	return 0;
3000 }
3001 
3002 static int intel_cpufreq_cpu_offline(struct cpufreq_policy *policy)
3003 {
3004 	struct cpudata *cpu = all_cpu_data[policy->cpu];
3005 
3006 	pr_debug("CPU %d going offline\n", cpu->cpu);
3007 
3008 	if (cpu->suspended)
3009 		return 0;
3010 
3011 	/*
3012 	 * If the CPU is an SMT thread and it goes offline with the performance
3013 	 * settings different from the minimum, it will prevent its sibling
3014 	 * from getting to lower performance levels, so force the minimum
3015 	 * performance on CPU offline to prevent that from happening.
3016 	 */
3017 	if (hwp_active)
3018 		intel_pstate_hwp_offline(cpu);
3019 	else
3020 		intel_pstate_set_min_pstate(cpu);
3021 
3022 	intel_pstate_exit_perf_limits(policy);
3023 
3024 	return 0;
3025 }
3026 
3027 static int intel_pstate_cpu_online(struct cpufreq_policy *policy)
3028 {
3029 	struct cpudata *cpu = all_cpu_data[policy->cpu];
3030 
3031 	pr_debug("CPU %d going online\n", cpu->cpu);
3032 
3033 	intel_pstate_init_acpi_perf_limits(policy);
3034 
3035 	if (hwp_active) {
3036 		/*
3037 		 * Re-enable HWP and clear the "suspended" flag to let "resume"
3038 		 * know that it need not do that.
3039 		 */
3040 		intel_pstate_hwp_reenable(cpu);
3041 		cpu->suspended = false;
3042 
3043 		hybrid_update_capacity(cpu);
3044 	}
3045 
3046 	return 0;
3047 }
3048 
3049 static int intel_pstate_cpu_offline(struct cpufreq_policy *policy)
3050 {
3051 	intel_pstate_clear_update_util_hook(policy->cpu);
3052 
3053 	return intel_cpufreq_cpu_offline(policy);
3054 }
3055 
3056 static void intel_pstate_cpu_exit(struct cpufreq_policy *policy)
3057 {
3058 	pr_debug("CPU %d exiting\n", policy->cpu);
3059 
3060 	policy->fast_switch_possible = false;
3061 }
3062 
3063 static int __intel_pstate_cpu_init(struct cpufreq_policy *policy)
3064 {
3065 	struct cpudata *cpu;
3066 	int rc;
3067 
3068 	rc = intel_pstate_init_cpu(policy->cpu);
3069 	if (rc)
3070 		return rc;
3071 
3072 	cpu = all_cpu_data[policy->cpu];
3073 
3074 	cpu->max_perf_ratio = 0xFF;
3075 	cpu->min_perf_ratio = 0;
3076 
3077 	/* cpuinfo and default policy values */
3078 	policy->cpuinfo.min_freq = cpu->pstate.min_freq;
3079 	policy->cpuinfo.max_freq = READ_ONCE(global.no_turbo) ?
3080 			cpu->pstate.max_freq : cpu->pstate.turbo_freq;
3081 
3082 	policy->min = policy->cpuinfo.min_freq;
3083 	policy->max = policy->cpuinfo.max_freq;
3084 
3085 	intel_pstate_init_acpi_perf_limits(policy);
3086 
3087 	policy->fast_switch_possible = true;
3088 
3089 	return 0;
3090 }
3091 
3092 static int intel_pstate_cpu_init(struct cpufreq_policy *policy)
3093 {
3094 	int ret = __intel_pstate_cpu_init(policy);
3095 
3096 	if (ret)
3097 		return ret;
3098 
3099 	/*
3100 	 * Set the policy to powersave to provide a valid fallback value in case
3101 	 * the default cpufreq governor is neither powersave nor performance.
3102 	 */
3103 	policy->policy = CPUFREQ_POLICY_POWERSAVE;
3104 
3105 	if (hwp_active) {
3106 		struct cpudata *cpu = all_cpu_data[policy->cpu];
3107 
3108 		cpu->epp_cached = intel_pstate_get_epp(cpu, 0);
3109 	}
3110 
3111 	return 0;
3112 }
3113 
3114 static struct cpufreq_driver intel_pstate = {
3115 	.flags		= CPUFREQ_CONST_LOOPS,
3116 	.verify		= intel_pstate_verify_policy,
3117 	.setpolicy	= intel_pstate_set_policy,
3118 	.suspend	= intel_pstate_suspend,
3119 	.resume		= intel_pstate_resume,
3120 	.init		= intel_pstate_cpu_init,
3121 	.exit		= intel_pstate_cpu_exit,
3122 	.offline	= intel_pstate_cpu_offline,
3123 	.online		= intel_pstate_cpu_online,
3124 	.update_limits	= intel_pstate_update_limits,
3125 	.name		= "intel_pstate",
3126 };
3127 
3128 static int intel_cpufreq_verify_policy(struct cpufreq_policy_data *policy)
3129 {
3130 	struct cpudata *cpu = all_cpu_data[policy->cpu];
3131 
3132 	intel_pstate_verify_cpu_policy(cpu, policy);
3133 	intel_pstate_update_perf_limits(cpu, policy->min, policy->max);
3134 
3135 	return 0;
3136 }
3137 
3138 /* Use of trace in passive mode:
3139  *
3140  * In passive mode the trace core_busy field (also known as the
3141  * performance field, and lablelled as such on the graphs; also known as
3142  * core_avg_perf) is not needed and so is re-assigned to indicate if the
3143  * driver call was via the normal or fast switch path. Various graphs
3144  * output from the intel_pstate_tracer.py utility that include core_busy
3145  * (or performance or core_avg_perf) have a fixed y-axis from 0 to 100%,
3146  * so we use 10 to indicate the normal path through the driver, and
3147  * 90 to indicate the fast switch path through the driver.
3148  * The scaled_busy field is not used, and is set to 0.
3149  */
3150 
3151 #define	INTEL_PSTATE_TRACE_TARGET 10
3152 #define	INTEL_PSTATE_TRACE_FAST_SWITCH 90
3153 
3154 static void intel_cpufreq_trace(struct cpudata *cpu, unsigned int trace_type, int old_pstate)
3155 {
3156 	struct sample *sample;
3157 
3158 	if (!trace_pstate_sample_enabled())
3159 		return;
3160 
3161 	if (!intel_pstate_sample(cpu, ktime_get()))
3162 		return;
3163 
3164 	sample = &cpu->sample;
3165 	trace_pstate_sample(trace_type,
3166 		0,
3167 		old_pstate,
3168 		cpu->pstate.current_pstate,
3169 		sample->mperf,
3170 		sample->aperf,
3171 		sample->tsc,
3172 		get_avg_frequency(cpu),
3173 		fp_toint(cpu->iowait_boost * 100));
3174 }
3175 
3176 static void intel_cpufreq_hwp_update(struct cpudata *cpu, u32 min, u32 max,
3177 				     u32 desired, bool fast_switch)
3178 {
3179 	u64 prev = READ_ONCE(cpu->hwp_req_cached), value = prev;
3180 
3181 	value &= ~HWP_MIN_PERF(~0L);
3182 	value |= HWP_MIN_PERF(min);
3183 
3184 	value &= ~HWP_MAX_PERF(~0L);
3185 	value |= HWP_MAX_PERF(max);
3186 
3187 	value &= ~HWP_DESIRED_PERF(~0L);
3188 	value |= HWP_DESIRED_PERF(desired);
3189 
3190 	if (value == prev)
3191 		return;
3192 
3193 	WRITE_ONCE(cpu->hwp_req_cached, value);
3194 	if (fast_switch)
3195 		wrmsrq(MSR_HWP_REQUEST, value);
3196 	else
3197 		wrmsrq_on_cpu(cpu->cpu, MSR_HWP_REQUEST, value);
3198 }
3199 
3200 static void intel_cpufreq_perf_ctl_update(struct cpudata *cpu,
3201 					  u32 target_pstate, bool fast_switch)
3202 {
3203 	if (fast_switch)
3204 		wrmsrq(MSR_IA32_PERF_CTL,
3205 		       pstate_funcs.get_val(cpu, target_pstate));
3206 	else
3207 		wrmsrq_on_cpu(cpu->cpu, MSR_IA32_PERF_CTL,
3208 			      pstate_funcs.get_val(cpu, target_pstate));
3209 }
3210 
3211 static int intel_cpufreq_update_pstate(struct cpufreq_policy *policy,
3212 				       int target_pstate, bool fast_switch)
3213 {
3214 	struct cpudata *cpu = all_cpu_data[policy->cpu];
3215 	int old_pstate = cpu->pstate.current_pstate;
3216 
3217 	target_pstate = intel_pstate_prepare_request(cpu, target_pstate);
3218 	if (hwp_active) {
3219 		int max_pstate = policy->strict_target ?
3220 					target_pstate : cpu->max_perf_ratio;
3221 
3222 		intel_cpufreq_hwp_update(cpu, target_pstate, max_pstate,
3223 					 target_pstate, fast_switch);
3224 	} else if (target_pstate != old_pstate) {
3225 		intel_cpufreq_perf_ctl_update(cpu, target_pstate, fast_switch);
3226 	}
3227 
3228 	cpu->pstate.current_pstate = target_pstate;
3229 
3230 	intel_cpufreq_trace(cpu, fast_switch ? INTEL_PSTATE_TRACE_FAST_SWITCH :
3231 			    INTEL_PSTATE_TRACE_TARGET, old_pstate);
3232 
3233 	return target_pstate;
3234 }
3235 
3236 static int intel_cpufreq_target(struct cpufreq_policy *policy,
3237 				unsigned int target_freq,
3238 				unsigned int relation)
3239 {
3240 	struct cpudata *cpu = all_cpu_data[policy->cpu];
3241 	struct cpufreq_freqs freqs;
3242 	int target_pstate;
3243 
3244 	freqs.old = policy->cur;
3245 	freqs.new = target_freq;
3246 
3247 	cpufreq_freq_transition_begin(policy, &freqs);
3248 
3249 	target_pstate = intel_pstate_freq_to_hwp_rel(cpu, freqs.new, relation);
3250 	target_pstate = intel_cpufreq_update_pstate(policy, target_pstate, false);
3251 
3252 	freqs.new = target_pstate * cpu->pstate.scaling;
3253 
3254 	cpufreq_freq_transition_end(policy, &freqs, false);
3255 
3256 	return 0;
3257 }
3258 
3259 static unsigned int intel_cpufreq_fast_switch(struct cpufreq_policy *policy,
3260 					      unsigned int target_freq)
3261 {
3262 	struct cpudata *cpu = all_cpu_data[policy->cpu];
3263 	int target_pstate;
3264 
3265 	target_pstate = intel_pstate_freq_to_hwp(cpu, target_freq);
3266 
3267 	target_pstate = intel_cpufreq_update_pstate(policy, target_pstate, true);
3268 
3269 	return target_pstate * cpu->pstate.scaling;
3270 }
3271 
3272 static void intel_cpufreq_adjust_perf(unsigned int cpunum,
3273 				      unsigned long min_perf,
3274 				      unsigned long target_perf,
3275 				      unsigned long capacity)
3276 {
3277 	struct cpudata *cpu = all_cpu_data[cpunum];
3278 	u64 hwp_cap = READ_ONCE(cpu->hwp_cap_cached);
3279 	int old_pstate = cpu->pstate.current_pstate;
3280 	int cap_pstate, min_pstate, max_pstate, target_pstate;
3281 
3282 	cap_pstate = READ_ONCE(global.no_turbo) ?
3283 					HWP_GUARANTEED_PERF(hwp_cap) :
3284 					HWP_HIGHEST_PERF(hwp_cap);
3285 
3286 	/* Optimization: Avoid unnecessary divisions. */
3287 
3288 	target_pstate = cap_pstate;
3289 	if (target_perf < capacity)
3290 		target_pstate = DIV_ROUND_UP(cap_pstate * target_perf, capacity);
3291 
3292 	min_pstate = cap_pstate;
3293 	if (min_perf < capacity)
3294 		min_pstate = DIV_ROUND_UP(cap_pstate * min_perf, capacity);
3295 
3296 	if (min_pstate < cpu->pstate.min_pstate)
3297 		min_pstate = cpu->pstate.min_pstate;
3298 
3299 	if (min_pstate < cpu->min_perf_ratio)
3300 		min_pstate = cpu->min_perf_ratio;
3301 
3302 	if (min_pstate > cpu->max_perf_ratio)
3303 		min_pstate = cpu->max_perf_ratio;
3304 
3305 	max_pstate = min(cap_pstate, cpu->max_perf_ratio);
3306 	if (max_pstate < min_pstate)
3307 		max_pstate = min_pstate;
3308 
3309 	target_pstate = clamp_t(int, target_pstate, min_pstate, max_pstate);
3310 
3311 	intel_cpufreq_hwp_update(cpu, min_pstate, max_pstate, target_pstate, true);
3312 
3313 	cpu->pstate.current_pstate = target_pstate;
3314 	intel_cpufreq_trace(cpu, INTEL_PSTATE_TRACE_FAST_SWITCH, old_pstate);
3315 }
3316 
3317 static int intel_cpufreq_cpu_init(struct cpufreq_policy *policy)
3318 {
3319 	struct freq_qos_request *req;
3320 	struct cpudata *cpu;
3321 	struct device *dev;
3322 	int ret, freq;
3323 
3324 	dev = get_cpu_device(policy->cpu);
3325 	if (!dev)
3326 		return -ENODEV;
3327 
3328 	ret = __intel_pstate_cpu_init(policy);
3329 	if (ret)
3330 		return ret;
3331 
3332 	policy->cpuinfo.transition_latency = INTEL_CPUFREQ_TRANSITION_LATENCY;
3333 	/* This reflects the intel_pstate_get_cpu_pstates() setting. */
3334 	policy->cur = policy->cpuinfo.min_freq;
3335 
3336 	req = kcalloc(2, sizeof(*req), GFP_KERNEL);
3337 	if (!req) {
3338 		ret = -ENOMEM;
3339 		goto pstate_exit;
3340 	}
3341 
3342 	cpu = all_cpu_data[policy->cpu];
3343 
3344 	if (hwp_active) {
3345 		u64 value;
3346 
3347 		policy->transition_delay_us = INTEL_CPUFREQ_TRANSITION_DELAY_HWP;
3348 
3349 		intel_pstate_get_hwp_cap(cpu);
3350 
3351 		rdmsrq_on_cpu(cpu->cpu, MSR_HWP_REQUEST, &value);
3352 		WRITE_ONCE(cpu->hwp_req_cached, value);
3353 
3354 		cpu->epp_cached = intel_pstate_get_epp(cpu, value);
3355 	} else {
3356 		policy->transition_delay_us = INTEL_CPUFREQ_TRANSITION_DELAY;
3357 	}
3358 
3359 	freq = DIV_ROUND_UP(cpu->pstate.turbo_freq * global.min_perf_pct, 100);
3360 
3361 	ret = freq_qos_add_request(&policy->constraints, req, FREQ_QOS_MIN,
3362 				   freq);
3363 	if (ret < 0) {
3364 		dev_err(dev, "Failed to add min-freq constraint (%d)\n", ret);
3365 		goto free_req;
3366 	}
3367 
3368 	freq = DIV_ROUND_UP(cpu->pstate.turbo_freq * global.max_perf_pct, 100);
3369 
3370 	ret = freq_qos_add_request(&policy->constraints, req + 1, FREQ_QOS_MAX,
3371 				   freq);
3372 	if (ret < 0) {
3373 		dev_err(dev, "Failed to add max-freq constraint (%d)\n", ret);
3374 		goto remove_min_req;
3375 	}
3376 
3377 	policy->driver_data = req;
3378 
3379 	return 0;
3380 
3381 remove_min_req:
3382 	freq_qos_remove_request(req);
3383 free_req:
3384 	kfree(req);
3385 pstate_exit:
3386 	intel_pstate_exit_perf_limits(policy);
3387 
3388 	return ret;
3389 }
3390 
3391 static void intel_cpufreq_cpu_exit(struct cpufreq_policy *policy)
3392 {
3393 	struct freq_qos_request *req;
3394 
3395 	req = policy->driver_data;
3396 
3397 	freq_qos_remove_request(req + 1);
3398 	freq_qos_remove_request(req);
3399 	kfree(req);
3400 
3401 	intel_pstate_cpu_exit(policy);
3402 }
3403 
3404 static int intel_cpufreq_suspend(struct cpufreq_policy *policy)
3405 {
3406 	intel_pstate_suspend(policy);
3407 
3408 	if (hwp_active) {
3409 		struct cpudata *cpu = all_cpu_data[policy->cpu];
3410 		u64 value = READ_ONCE(cpu->hwp_req_cached);
3411 
3412 		/*
3413 		 * Clear the desired perf field in MSR_HWP_REQUEST in case
3414 		 * intel_cpufreq_adjust_perf() is in use and the last value
3415 		 * written by it may not be suitable.
3416 		 */
3417 		value &= ~HWP_DESIRED_PERF(~0L);
3418 		wrmsrq_on_cpu(cpu->cpu, MSR_HWP_REQUEST, value);
3419 		WRITE_ONCE(cpu->hwp_req_cached, value);
3420 	}
3421 
3422 	return 0;
3423 }
3424 
3425 static struct cpufreq_driver intel_cpufreq = {
3426 	.flags		= CPUFREQ_CONST_LOOPS,
3427 	.verify		= intel_cpufreq_verify_policy,
3428 	.target		= intel_cpufreq_target,
3429 	.fast_switch	= intel_cpufreq_fast_switch,
3430 	.init		= intel_cpufreq_cpu_init,
3431 	.exit		= intel_cpufreq_cpu_exit,
3432 	.offline	= intel_cpufreq_cpu_offline,
3433 	.online		= intel_pstate_cpu_online,
3434 	.suspend	= intel_cpufreq_suspend,
3435 	.resume		= intel_pstate_resume,
3436 	.update_limits	= intel_pstate_update_limits,
3437 	.name		= "intel_cpufreq",
3438 };
3439 
3440 static struct cpufreq_driver *default_driver;
3441 
3442 static void intel_pstate_driver_cleanup(void)
3443 {
3444 	unsigned int cpu;
3445 
3446 	cpus_read_lock();
3447 	for_each_online_cpu(cpu) {
3448 		if (all_cpu_data[cpu]) {
3449 			if (intel_pstate_driver == &intel_pstate)
3450 				intel_pstate_clear_update_util_hook(cpu);
3451 
3452 			kfree(all_cpu_data[cpu]);
3453 			WRITE_ONCE(all_cpu_data[cpu], NULL);
3454 		}
3455 	}
3456 	cpus_read_unlock();
3457 
3458 	intel_pstate_driver = NULL;
3459 }
3460 
3461 static int intel_pstate_register_driver(struct cpufreq_driver *driver)
3462 {
3463 	bool refresh_cpu_cap_scaling;
3464 	int ret;
3465 
3466 	if (driver == &intel_pstate)
3467 		intel_pstate_sysfs_expose_hwp_dynamic_boost();
3468 
3469 	memset(&global, 0, sizeof(global));
3470 	global.max_perf_pct = 100;
3471 	global.turbo_disabled = turbo_is_disabled();
3472 	global.no_turbo = global.turbo_disabled;
3473 
3474 	arch_set_max_freq_ratio(global.turbo_disabled);
3475 
3476 	refresh_cpu_cap_scaling = hybrid_clear_max_perf_cpu();
3477 
3478 	intel_pstate_driver = driver;
3479 	ret = cpufreq_register_driver(intel_pstate_driver);
3480 	if (ret) {
3481 		intel_pstate_driver_cleanup();
3482 		return ret;
3483 	}
3484 
3485 	global.min_perf_pct = min_perf_pct_min();
3486 
3487 	hybrid_init_cpu_capacity_scaling(refresh_cpu_cap_scaling);
3488 
3489 	return 0;
3490 }
3491 
3492 static ssize_t intel_pstate_show_status(char *buf)
3493 {
3494 	if (!intel_pstate_driver)
3495 		return sprintf(buf, "off\n");
3496 
3497 	return sprintf(buf, "%s\n", intel_pstate_driver == &intel_pstate ?
3498 					"active" : "passive");
3499 }
3500 
3501 static int intel_pstate_update_status(const char *buf, size_t size)
3502 {
3503 	if (size == 3 && !strncmp(buf, "off", size)) {
3504 		if (!intel_pstate_driver)
3505 			return -EINVAL;
3506 
3507 		if (hwp_active)
3508 			return -EBUSY;
3509 
3510 		cpufreq_unregister_driver(intel_pstate_driver);
3511 		intel_pstate_driver_cleanup();
3512 		return 0;
3513 	}
3514 
3515 	if (size == 6 && !strncmp(buf, "active", size)) {
3516 		if (intel_pstate_driver) {
3517 			if (intel_pstate_driver == &intel_pstate)
3518 				return 0;
3519 
3520 			cpufreq_unregister_driver(intel_pstate_driver);
3521 		}
3522 
3523 		return intel_pstate_register_driver(&intel_pstate);
3524 	}
3525 
3526 	if (size == 7 && !strncmp(buf, "passive", size)) {
3527 		if (intel_pstate_driver) {
3528 			if (intel_pstate_driver == &intel_cpufreq)
3529 				return 0;
3530 
3531 			cpufreq_unregister_driver(intel_pstate_driver);
3532 			intel_pstate_sysfs_hide_hwp_dynamic_boost();
3533 		}
3534 
3535 		return intel_pstate_register_driver(&intel_cpufreq);
3536 	}
3537 
3538 	return -EINVAL;
3539 }
3540 
3541 static int no_load __initdata;
3542 static int no_hwp __initdata;
3543 static int hwp_only __initdata;
3544 static unsigned int force_load __initdata;
3545 
3546 static int __init intel_pstate_msrs_not_valid(void)
3547 {
3548 	if (!pstate_funcs.get_max(0) ||
3549 	    !pstate_funcs.get_min(0) ||
3550 	    !pstate_funcs.get_turbo(0))
3551 		return -ENODEV;
3552 
3553 	return 0;
3554 }
3555 
3556 static void __init copy_cpu_funcs(struct pstate_funcs *funcs)
3557 {
3558 	pstate_funcs.get_max   = funcs->get_max;
3559 	pstate_funcs.get_max_physical = funcs->get_max_physical;
3560 	pstate_funcs.get_min   = funcs->get_min;
3561 	pstate_funcs.get_turbo = funcs->get_turbo;
3562 	pstate_funcs.get_scaling = funcs->get_scaling;
3563 	pstate_funcs.get_val   = funcs->get_val;
3564 	pstate_funcs.get_vid   = funcs->get_vid;
3565 	pstate_funcs.get_aperf_mperf_shift = funcs->get_aperf_mperf_shift;
3566 }
3567 
3568 #ifdef CONFIG_ACPI
3569 
3570 static bool __init intel_pstate_no_acpi_pss(void)
3571 {
3572 	int i;
3573 
3574 	for_each_possible_cpu(i) {
3575 		acpi_status status;
3576 		union acpi_object *pss;
3577 		struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
3578 		struct acpi_processor *pr = per_cpu(processors, i);
3579 
3580 		if (!pr)
3581 			continue;
3582 
3583 		status = acpi_evaluate_object(pr->handle, "_PSS", NULL, &buffer);
3584 		if (ACPI_FAILURE(status))
3585 			continue;
3586 
3587 		pss = buffer.pointer;
3588 		if (pss && pss->type == ACPI_TYPE_PACKAGE) {
3589 			kfree(pss);
3590 			return false;
3591 		}
3592 
3593 		kfree(pss);
3594 	}
3595 
3596 	pr_debug("ACPI _PSS not found\n");
3597 	return true;
3598 }
3599 
3600 static bool __init intel_pstate_no_acpi_pcch(void)
3601 {
3602 	acpi_status status;
3603 	acpi_handle handle;
3604 
3605 	status = acpi_get_handle(NULL, "\\_SB", &handle);
3606 	if (ACPI_FAILURE(status))
3607 		goto not_found;
3608 
3609 	if (acpi_has_method(handle, "PCCH"))
3610 		return false;
3611 
3612 not_found:
3613 	pr_debug("ACPI PCCH not found\n");
3614 	return true;
3615 }
3616 
3617 static bool __init intel_pstate_has_acpi_ppc(void)
3618 {
3619 	int i;
3620 
3621 	for_each_possible_cpu(i) {
3622 		struct acpi_processor *pr = per_cpu(processors, i);
3623 
3624 		if (!pr)
3625 			continue;
3626 		if (acpi_has_method(pr->handle, "_PPC"))
3627 			return true;
3628 	}
3629 	pr_debug("ACPI _PPC not found\n");
3630 	return false;
3631 }
3632 
3633 enum {
3634 	PSS,
3635 	PPC,
3636 };
3637 
3638 /* Hardware vendor-specific info that has its own power management modes */
3639 static struct acpi_platform_list plat_info[] __initdata = {
3640 	{"HP    ", "ProLiant", 0, ACPI_SIG_FADT, all_versions, NULL, PSS},
3641 	{"ORACLE", "X4-2    ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3642 	{"ORACLE", "X4-2L   ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3643 	{"ORACLE", "X4-2B   ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3644 	{"ORACLE", "X3-2    ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3645 	{"ORACLE", "X3-2L   ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3646 	{"ORACLE", "X3-2B   ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3647 	{"ORACLE", "X4470M2 ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3648 	{"ORACLE", "X4270M3 ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3649 	{"ORACLE", "X4270M2 ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3650 	{"ORACLE", "X4170M2 ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3651 	{"ORACLE", "X4170 M3", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3652 	{"ORACLE", "X4275 M3", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3653 	{"ORACLE", "X6-2    ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3654 	{"ORACLE", "Sudbury ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3655 	{ } /* End */
3656 };
3657 
3658 #define BITMASK_OOB	(BIT(8) | BIT(18))
3659 
3660 static bool __init intel_pstate_platform_pwr_mgmt_exists(void)
3661 {
3662 	const struct x86_cpu_id *id;
3663 	u64 misc_pwr;
3664 	int idx;
3665 
3666 	id = x86_match_cpu(intel_pstate_cpu_oob_ids);
3667 	if (id) {
3668 		rdmsrq(MSR_MISC_PWR_MGMT, misc_pwr);
3669 		if (misc_pwr & BITMASK_OOB) {
3670 			pr_debug("Bit 8 or 18 in the MISC_PWR_MGMT MSR set\n");
3671 			pr_debug("P states are controlled in Out of Band mode by the firmware/hardware\n");
3672 			return true;
3673 		}
3674 	}
3675 
3676 	idx = acpi_match_platform_list(plat_info);
3677 	if (idx < 0)
3678 		return false;
3679 
3680 	switch (plat_info[idx].data) {
3681 	case PSS:
3682 		if (!intel_pstate_no_acpi_pss())
3683 			return false;
3684 
3685 		return intel_pstate_no_acpi_pcch();
3686 	case PPC:
3687 		return intel_pstate_has_acpi_ppc() && !force_load;
3688 	}
3689 
3690 	return false;
3691 }
3692 
3693 static void intel_pstate_request_control_from_smm(void)
3694 {
3695 	/*
3696 	 * It may be unsafe to request P-states control from SMM if _PPC support
3697 	 * has not been enabled.
3698 	 */
3699 	if (acpi_ppc)
3700 		acpi_processor_pstate_control();
3701 }
3702 #else /* CONFIG_ACPI not enabled */
3703 static inline bool intel_pstate_platform_pwr_mgmt_exists(void) { return false; }
3704 static inline bool intel_pstate_has_acpi_ppc(void) { return false; }
3705 static inline void intel_pstate_request_control_from_smm(void) {}
3706 #endif /* CONFIG_ACPI */
3707 
3708 #define INTEL_PSTATE_HWP_BROADWELL	0x01
3709 
3710 #define X86_MATCH_HWP(vfm, hwp_mode)				\
3711 	X86_MATCH_VFM_FEATURE(vfm, X86_FEATURE_HWP, hwp_mode)
3712 
3713 static const struct x86_cpu_id hwp_support_ids[] __initconst = {
3714 	X86_MATCH_HWP(INTEL_BROADWELL_X,	INTEL_PSTATE_HWP_BROADWELL),
3715 	X86_MATCH_HWP(INTEL_BROADWELL_D,	INTEL_PSTATE_HWP_BROADWELL),
3716 	X86_MATCH_HWP(INTEL_ANY,		0),
3717 	{}
3718 };
3719 
3720 static bool intel_pstate_hwp_is_enabled(void)
3721 {
3722 	u64 value;
3723 
3724 	rdmsrq(MSR_PM_ENABLE, value);
3725 	return !!(value & 0x1);
3726 }
3727 
3728 #define POWERSAVE_MASK			GENMASK(7, 0)
3729 #define BALANCE_POWER_MASK		GENMASK(15, 8)
3730 #define BALANCE_PERFORMANCE_MASK	GENMASK(23, 16)
3731 #define PERFORMANCE_MASK		GENMASK(31, 24)
3732 
3733 #define HWP_SET_EPP_VALUES(powersave, balance_power, balance_perf, performance) \
3734 	(FIELD_PREP_CONST(POWERSAVE_MASK, powersave) |\
3735 	 FIELD_PREP_CONST(BALANCE_POWER_MASK, balance_power) |\
3736 	 FIELD_PREP_CONST(BALANCE_PERFORMANCE_MASK, balance_perf) |\
3737 	 FIELD_PREP_CONST(PERFORMANCE_MASK, performance))
3738 
3739 #define HWP_SET_DEF_BALANCE_PERF_EPP(balance_perf) \
3740 	(HWP_SET_EPP_VALUES(HWP_EPP_POWERSAVE, HWP_EPP_BALANCE_POWERSAVE,\
3741 	 balance_perf, HWP_EPP_PERFORMANCE))
3742 
3743 static const struct x86_cpu_id intel_epp_default[] = {
3744 	/*
3745 	 * Set EPP value as 102, this is the max suggested EPP
3746 	 * which can result in one core turbo frequency for
3747 	 * AlderLake Mobile CPUs.
3748 	 */
3749 	X86_MATCH_VFM(INTEL_ALDERLAKE_L, HWP_SET_DEF_BALANCE_PERF_EPP(102)),
3750 	X86_MATCH_VFM(INTEL_SAPPHIRERAPIDS_X, HWP_SET_DEF_BALANCE_PERF_EPP(32)),
3751 	X86_MATCH_VFM(INTEL_EMERALDRAPIDS_X, HWP_SET_DEF_BALANCE_PERF_EPP(32)),
3752 	X86_MATCH_VFM(INTEL_GRANITERAPIDS_X, HWP_SET_DEF_BALANCE_PERF_EPP(32)),
3753 	X86_MATCH_VFM(INTEL_GRANITERAPIDS_D, HWP_SET_DEF_BALANCE_PERF_EPP(32)),
3754 	X86_MATCH_VFM(INTEL_METEORLAKE_L, HWP_SET_EPP_VALUES(HWP_EPP_POWERSAVE,
3755 		      179, 64, 16)),
3756 	X86_MATCH_VFM(INTEL_ARROWLAKE, HWP_SET_EPP_VALUES(HWP_EPP_POWERSAVE,
3757 		      179, 64, 16)),
3758 	{}
3759 };
3760 
3761 static const struct x86_cpu_id intel_hybrid_scaling_factor[] = {
3762 	X86_MATCH_VFM(INTEL_ALDERLAKE, HYBRID_SCALING_FACTOR_ADL),
3763 	X86_MATCH_VFM(INTEL_ALDERLAKE_L, HYBRID_SCALING_FACTOR_ADL),
3764 	X86_MATCH_VFM(INTEL_RAPTORLAKE, HYBRID_SCALING_FACTOR_ADL),
3765 	X86_MATCH_VFM(INTEL_RAPTORLAKE_P, HYBRID_SCALING_FACTOR_ADL),
3766 	X86_MATCH_VFM(INTEL_RAPTORLAKE_S, HYBRID_SCALING_FACTOR_ADL),
3767 	X86_MATCH_VFM(INTEL_METEORLAKE_L, HYBRID_SCALING_FACTOR_MTL),
3768 	X86_MATCH_VFM(INTEL_LUNARLAKE_M, HYBRID_SCALING_FACTOR_LNL),
3769 	{}
3770 };
3771 
3772 static bool hwp_check_epp(void)
3773 {
3774 	if (boot_cpu_has(X86_FEATURE_HWP_EPP))
3775 		return true;
3776 
3777 	/* Without EPP support, don't expose EPP-related sysfs attributes. */
3778 	hwp_cpufreq_attrs[HWP_PERFORMANCE_PREFERENCE_INDEX] = NULL;
3779 	hwp_cpufreq_attrs[HWP_PERFORMANCE_AVAILABLE_PREFERENCES_INDEX] = NULL;
3780 
3781 	return false;
3782 }
3783 
3784 static bool hwp_check_dec(void)
3785 {
3786 	u64 power_ctl;
3787 
3788 	rdmsrq(MSR_IA32_POWER_CTL, power_ctl);
3789 	return !!(power_ctl & BIT(POWER_CTL_DEC_ENABLE));
3790 }
3791 
3792 static int __init intel_pstate_init(void)
3793 {
3794 	static struct cpudata **_all_cpu_data;
3795 	const struct x86_cpu_id *id;
3796 	int rc;
3797 
3798 	if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL)
3799 		return -ENODEV;
3800 
3801 	/*
3802 	 * The Intel pstate driver will be ignored if the platform
3803 	 * firmware has its own power management modes.
3804 	 */
3805 	if (intel_pstate_platform_pwr_mgmt_exists()) {
3806 		pr_info("P-states controlled by the platform\n");
3807 		return -ENODEV;
3808 	}
3809 
3810 	id = x86_match_cpu(hwp_support_ids);
3811 	if (id) {
3812 		bool epp_present = hwp_check_epp();
3813 
3814 		/*
3815 		 * If HWP is enabled already, there is no choice but to deal
3816 		 * with it.
3817 		 */
3818 		hwp_forced = intel_pstate_hwp_is_enabled();
3819 		if (hwp_forced) {
3820 			pr_info("HWP enabled by BIOS\n");
3821 			no_hwp = 0;
3822 		} else if (no_load) {
3823 			return -ENODEV;
3824 		} else if (!epp_present && !hwp_check_dec()) {
3825 			/*
3826 			 * Avoid enabling HWP for processors without EPP support
3827 			 * unless the Dynamic Efficiency Control (DEC) enable
3828 			 * bit (MSR_IA32_POWER_CTL, bit 27) is set because that
3829 			 * means incomplete HWP implementation which is a corner
3830 			 * case and supporting it is generally problematic.
3831 			 */
3832 			no_hwp = 1;
3833 		}
3834 
3835 		copy_cpu_funcs(&core_funcs);
3836 
3837 		if (!no_hwp) {
3838 			hwp_active = true;
3839 			hwp_mode_bdw = id->driver_data;
3840 			intel_pstate.attr = hwp_cpufreq_attrs;
3841 			intel_cpufreq.attr = hwp_cpufreq_attrs;
3842 			intel_cpufreq.flags |= CPUFREQ_NEED_UPDATE_LIMITS;
3843 			intel_cpufreq.adjust_perf = intel_cpufreq_adjust_perf;
3844 			if (!default_driver)
3845 				default_driver = &intel_pstate;
3846 
3847 			pstate_funcs.get_cpu_scaling = hwp_get_cpu_scaling;
3848 
3849 			goto hwp_cpu_matched;
3850 		}
3851 		pr_info("HWP not enabled\n");
3852 	} else {
3853 		if (no_load)
3854 			return -ENODEV;
3855 
3856 		id = x86_match_cpu(intel_pstate_cpu_ids);
3857 		if (!id) {
3858 			pr_info("CPU model not supported\n");
3859 			return -ENODEV;
3860 		}
3861 
3862 		copy_cpu_funcs((struct pstate_funcs *)id->driver_data);
3863 	}
3864 
3865 	if (intel_pstate_msrs_not_valid()) {
3866 		pr_info("Invalid MSRs\n");
3867 		return -ENODEV;
3868 	}
3869 	/* Without HWP start in the passive mode. */
3870 	if (!default_driver)
3871 		default_driver = &intel_cpufreq;
3872 
3873 hwp_cpu_matched:
3874 	if (!hwp_active && hwp_only)
3875 		return -ENOTSUPP;
3876 
3877 	pr_info("Intel P-state driver initializing\n");
3878 
3879 	_all_cpu_data = vzalloc(array_size(sizeof(void *), num_possible_cpus()));
3880 	if (!_all_cpu_data)
3881 		return -ENOMEM;
3882 
3883 	WRITE_ONCE(all_cpu_data, _all_cpu_data);
3884 
3885 	intel_pstate_request_control_from_smm();
3886 
3887 	intel_pstate_sysfs_expose_params();
3888 
3889 	if (hwp_active) {
3890 		const struct x86_cpu_id *id = x86_match_cpu(intel_epp_default);
3891 		const struct x86_cpu_id *hybrid_id = x86_match_cpu(intel_hybrid_scaling_factor);
3892 
3893 		if (id) {
3894 			epp_values[EPP_INDEX_POWERSAVE] =
3895 					FIELD_GET(POWERSAVE_MASK, id->driver_data);
3896 			epp_values[EPP_INDEX_BALANCE_POWERSAVE] =
3897 					FIELD_GET(BALANCE_POWER_MASK, id->driver_data);
3898 			epp_values[EPP_INDEX_BALANCE_PERFORMANCE] =
3899 					FIELD_GET(BALANCE_PERFORMANCE_MASK, id->driver_data);
3900 			epp_values[EPP_INDEX_PERFORMANCE] =
3901 					FIELD_GET(PERFORMANCE_MASK, id->driver_data);
3902 			pr_debug("Updated EPPs powersave:%x balanced power:%x balanced perf:%x performance:%x\n",
3903 				 epp_values[EPP_INDEX_POWERSAVE],
3904 				 epp_values[EPP_INDEX_BALANCE_POWERSAVE],
3905 				 epp_values[EPP_INDEX_BALANCE_PERFORMANCE],
3906 				 epp_values[EPP_INDEX_PERFORMANCE]);
3907 		}
3908 
3909 		if (hybrid_id) {
3910 			hybrid_scaling_factor = hybrid_id->driver_data;
3911 			pr_debug("hybrid scaling factor: %d\n", hybrid_scaling_factor);
3912 		}
3913 
3914 	}
3915 
3916 	mutex_lock(&intel_pstate_driver_lock);
3917 	rc = intel_pstate_register_driver(default_driver);
3918 	mutex_unlock(&intel_pstate_driver_lock);
3919 	if (rc) {
3920 		intel_pstate_sysfs_remove();
3921 		return rc;
3922 	}
3923 
3924 	if (hwp_active) {
3925 		const struct x86_cpu_id *id;
3926 
3927 		id = x86_match_cpu(intel_pstate_cpu_ee_disable_ids);
3928 		if (id) {
3929 			set_power_ctl_ee_state(false);
3930 			pr_info("Disabling energy efficiency optimization\n");
3931 		}
3932 
3933 		pr_info("HWP enabled\n");
3934 	} else if (boot_cpu_has(X86_FEATURE_HYBRID_CPU)) {
3935 		pr_warn("Problematic setup: Hybrid processor with disabled HWP\n");
3936 	}
3937 
3938 	return 0;
3939 }
3940 device_initcall(intel_pstate_init);
3941 
3942 static int __init intel_pstate_setup(char *str)
3943 {
3944 	if (!str)
3945 		return -EINVAL;
3946 
3947 	if (!strcmp(str, "disable"))
3948 		no_load = 1;
3949 	else if (!strcmp(str, "active"))
3950 		default_driver = &intel_pstate;
3951 	else if (!strcmp(str, "passive"))
3952 		default_driver = &intel_cpufreq;
3953 
3954 	if (!strcmp(str, "no_hwp"))
3955 		no_hwp = 1;
3956 
3957 	if (!strcmp(str, "no_cas"))
3958 		no_cas = true;
3959 
3960 	if (!strcmp(str, "force"))
3961 		force_load = 1;
3962 	if (!strcmp(str, "hwp_only"))
3963 		hwp_only = 1;
3964 	if (!strcmp(str, "per_cpu_perf_limits"))
3965 		per_cpu_limits = true;
3966 
3967 #ifdef CONFIG_ACPI
3968 	if (!strcmp(str, "support_acpi_ppc"))
3969 		acpi_ppc = true;
3970 #endif
3971 
3972 	return 0;
3973 }
3974 early_param("intel_pstate", intel_pstate_setup);
3975 
3976 MODULE_AUTHOR("Dirk Brandewie <dirk.j.brandewie@intel.com>");
3977 MODULE_DESCRIPTION("'intel_pstate' - P state driver Intel Core processors");
3978