xref: /linux/drivers/cpufreq/intel_pstate.c (revision 288440de9e5fdb4a3ff73864850f080c1250fc81)
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/list.h>
20 #include <linux/cpu.h>
21 #include <linux/cpufreq.h>
22 #include <linux/sysfs.h>
23 #include <linux/types.h>
24 #include <linux/fs.h>
25 #include <linux/acpi.h>
26 #include <linux/vmalloc.h>
27 #include <linux/pm_qos.h>
28 #include <trace/events/power.h>
29 
30 #include <asm/div64.h>
31 #include <asm/msr.h>
32 #include <asm/cpu_device_id.h>
33 #include <asm/cpufeature.h>
34 #include <asm/intel-family.h>
35 #include "../drivers/thermal/intel/thermal_interrupt.h"
36 
37 #define INTEL_PSTATE_SAMPLING_INTERVAL	(10 * NSEC_PER_MSEC)
38 
39 #define INTEL_CPUFREQ_TRANSITION_LATENCY	20000
40 #define INTEL_CPUFREQ_TRANSITION_DELAY_HWP	5000
41 #define INTEL_CPUFREQ_TRANSITION_DELAY		500
42 
43 #ifdef CONFIG_ACPI
44 #include <acpi/processor.h>
45 #include <acpi/cppc_acpi.h>
46 #endif
47 
48 #define FRAC_BITS 8
49 #define int_tofp(X) ((int64_t)(X) << FRAC_BITS)
50 #define fp_toint(X) ((X) >> FRAC_BITS)
51 
52 #define ONE_EIGHTH_FP ((int64_t)1 << (FRAC_BITS - 3))
53 
54 #define EXT_BITS 6
55 #define EXT_FRAC_BITS (EXT_BITS + FRAC_BITS)
56 #define fp_ext_toint(X) ((X) >> EXT_FRAC_BITS)
57 #define int_ext_tofp(X) ((int64_t)(X) << EXT_FRAC_BITS)
58 
59 static inline int32_t mul_fp(int32_t x, int32_t y)
60 {
61 	return ((int64_t)x * (int64_t)y) >> FRAC_BITS;
62 }
63 
64 static inline int32_t div_fp(s64 x, s64 y)
65 {
66 	return div64_s64((int64_t)x << FRAC_BITS, y);
67 }
68 
69 static inline int ceiling_fp(int32_t x)
70 {
71 	int mask, ret;
72 
73 	ret = fp_toint(x);
74 	mask = (1 << FRAC_BITS) - 1;
75 	if (x & mask)
76 		ret += 1;
77 	return ret;
78 }
79 
80 static inline u64 mul_ext_fp(u64 x, u64 y)
81 {
82 	return (x * y) >> EXT_FRAC_BITS;
83 }
84 
85 static inline u64 div_ext_fp(u64 x, u64 y)
86 {
87 	return div64_u64(x << EXT_FRAC_BITS, y);
88 }
89 
90 /**
91  * struct sample -	Store performance sample
92  * @core_avg_perf:	Ratio of APERF/MPERF which is the actual average
93  *			performance during last sample period
94  * @busy_scaled:	Scaled busy value which is used to calculate next
95  *			P state. This can be different than core_avg_perf
96  *			to account for cpu idle period
97  * @aperf:		Difference of actual performance frequency clock count
98  *			read from APERF MSR between last and current sample
99  * @mperf:		Difference of maximum performance frequency clock count
100  *			read from MPERF MSR between last and current sample
101  * @tsc:		Difference of time stamp counter between last and
102  *			current sample
103  * @time:		Current time from scheduler
104  *
105  * This structure is used in the cpudata structure to store performance sample
106  * data for choosing next P State.
107  */
108 struct sample {
109 	int32_t core_avg_perf;
110 	int32_t busy_scaled;
111 	u64 aperf;
112 	u64 mperf;
113 	u64 tsc;
114 	u64 time;
115 };
116 
117 /**
118  * struct pstate_data - Store P state data
119  * @current_pstate:	Current requested P state
120  * @min_pstate:		Min P state possible for this platform
121  * @max_pstate:		Max P state possible for this platform
122  * @max_pstate_physical:This is physical Max P state for a processor
123  *			This can be higher than the max_pstate which can
124  *			be limited by platform thermal design power limits
125  * @perf_ctl_scaling:	PERF_CTL P-state to frequency scaling factor
126  * @scaling:		Scaling factor between performance and frequency
127  * @turbo_pstate:	Max Turbo P state possible for this platform
128  * @min_freq:		@min_pstate frequency in cpufreq units
129  * @max_freq:		@max_pstate frequency in cpufreq units
130  * @turbo_freq:		@turbo_pstate frequency in cpufreq units
131  *
132  * Stores the per cpu model P state limits and current P state.
133  */
134 struct pstate_data {
135 	int	current_pstate;
136 	int	min_pstate;
137 	int	max_pstate;
138 	int	max_pstate_physical;
139 	int	perf_ctl_scaling;
140 	int	scaling;
141 	int	turbo_pstate;
142 	unsigned int min_freq;
143 	unsigned int max_freq;
144 	unsigned int turbo_freq;
145 };
146 
147 /**
148  * struct vid_data -	Stores voltage information data
149  * @min:		VID data for this platform corresponding to
150  *			the lowest P state
151  * @max:		VID data corresponding to the highest P State.
152  * @turbo:		VID data for turbo P state
153  * @ratio:		Ratio of (vid max - vid min) /
154  *			(max P state - Min P State)
155  *
156  * Stores the voltage data for DVFS (Dynamic Voltage and Frequency Scaling)
157  * This data is used in Atom platforms, where in addition to target P state,
158  * the voltage data needs to be specified to select next P State.
159  */
160 struct vid_data {
161 	int min;
162 	int max;
163 	int turbo;
164 	int32_t ratio;
165 };
166 
167 /**
168  * struct global_params - Global parameters, mostly tunable via sysfs.
169  * @no_turbo:		Whether or not to use turbo P-states.
170  * @turbo_disabled:	Whether or not turbo P-states are available at all,
171  *			based on the MSR_IA32_MISC_ENABLE value and whether or
172  *			not the maximum reported turbo P-state is different from
173  *			the maximum reported non-turbo one.
174  * @turbo_disabled_mf:	The @turbo_disabled value reflected by cpuinfo.max_freq.
175  * @min_perf_pct:	Minimum capacity limit in percent of the maximum turbo
176  *			P-state capacity.
177  * @max_perf_pct:	Maximum capacity limit in percent of the maximum turbo
178  *			P-state capacity.
179  */
180 struct global_params {
181 	bool no_turbo;
182 	bool turbo_disabled;
183 	bool turbo_disabled_mf;
184 	int max_perf_pct;
185 	int min_perf_pct;
186 };
187 
188 /**
189  * struct cpudata -	Per CPU instance data storage
190  * @cpu:		CPU number for this instance data
191  * @policy:		CPUFreq policy value
192  * @update_util:	CPUFreq utility callback information
193  * @update_util_set:	CPUFreq utility callback is set
194  * @iowait_boost:	iowait-related boost fraction
195  * @last_update:	Time of the last update.
196  * @pstate:		Stores P state limits for this CPU
197  * @vid:		Stores VID limits for this CPU
198  * @last_sample_time:	Last Sample time
199  * @aperf_mperf_shift:	APERF vs MPERF counting frequency difference
200  * @prev_aperf:		Last APERF value read from APERF MSR
201  * @prev_mperf:		Last MPERF value read from MPERF MSR
202  * @prev_tsc:		Last timestamp counter (TSC) value
203  * @prev_cummulative_iowait: IO Wait time difference from last and
204  *			current sample
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  * @sched_flags:	Store scheduler flags for possible cross CPU update
221  * @hwp_boost_min:	Last HWP boosted min performance
222  * @suspended:		Whether or not the driver has been suspended.
223  * @hwp_notify_work:	workqueue for HWP notifications.
224  *
225  * This structure stores per CPU instance data for all CPUs.
226  */
227 struct cpudata {
228 	int cpu;
229 
230 	unsigned int policy;
231 	struct update_util_data update_util;
232 	bool   update_util_set;
233 
234 	struct pstate_data pstate;
235 	struct vid_data vid;
236 
237 	u64	last_update;
238 	u64	last_sample_time;
239 	u64	aperf_mperf_shift;
240 	u64	prev_aperf;
241 	u64	prev_mperf;
242 	u64	prev_tsc;
243 	u64	prev_cummulative_iowait;
244 	struct sample sample;
245 	int32_t	min_perf_ratio;
246 	int32_t	max_perf_ratio;
247 #ifdef CONFIG_ACPI
248 	struct acpi_processor_performance acpi_perf_data;
249 	bool valid_pss_table;
250 #endif
251 	unsigned int iowait_boost;
252 	s16 epp_powersave;
253 	s16 epp_policy;
254 	s16 epp_default;
255 	s16 epp_cached;
256 	u64 hwp_req_cached;
257 	u64 hwp_cap_cached;
258 	u64 last_io_update;
259 	unsigned int sched_flags;
260 	u32 hwp_boost_min;
261 	bool suspended;
262 	struct delayed_work hwp_notify_work;
263 };
264 
265 static struct cpudata **all_cpu_data;
266 
267 /**
268  * struct pstate_funcs - Per CPU model specific callbacks
269  * @get_max:		Callback to get maximum non turbo effective P state
270  * @get_max_physical:	Callback to get maximum non turbo physical P state
271  * @get_min:		Callback to get minimum P state
272  * @get_turbo:		Callback to get turbo P state
273  * @get_scaling:	Callback to get frequency scaling factor
274  * @get_cpu_scaling:	Get frequency scaling factor for a given cpu
275  * @get_aperf_mperf_shift: Callback to get the APERF vs MPERF frequency difference
276  * @get_val:		Callback to convert P state to actual MSR write value
277  * @get_vid:		Callback to get VID data for Atom platforms
278  *
279  * Core and Atom CPU models have different way to get P State limits. This
280  * structure is used to store those callbacks.
281  */
282 struct pstate_funcs {
283 	int (*get_max)(void);
284 	int (*get_max_physical)(void);
285 	int (*get_min)(void);
286 	int (*get_turbo)(void);
287 	int (*get_scaling)(void);
288 	int (*get_cpu_scaling)(int cpu);
289 	int (*get_aperf_mperf_shift)(void);
290 	u64 (*get_val)(struct cpudata*, int pstate);
291 	void (*get_vid)(struct cpudata *);
292 };
293 
294 static struct pstate_funcs pstate_funcs __read_mostly;
295 
296 static int hwp_active __read_mostly;
297 static int hwp_mode_bdw __read_mostly;
298 static bool per_cpu_limits __read_mostly;
299 static bool hwp_boost __read_mostly;
300 
301 static struct cpufreq_driver *intel_pstate_driver __read_mostly;
302 
303 #ifdef CONFIG_ACPI
304 static bool acpi_ppc;
305 #endif
306 
307 static struct global_params global;
308 
309 static DEFINE_MUTEX(intel_pstate_driver_lock);
310 static DEFINE_MUTEX(intel_pstate_limits_lock);
311 
312 #ifdef CONFIG_ACPI
313 
314 static bool intel_pstate_acpi_pm_profile_server(void)
315 {
316 	if (acpi_gbl_FADT.preferred_profile == PM_ENTERPRISE_SERVER ||
317 	    acpi_gbl_FADT.preferred_profile == PM_PERFORMANCE_SERVER)
318 		return true;
319 
320 	return false;
321 }
322 
323 static bool intel_pstate_get_ppc_enable_status(void)
324 {
325 	if (intel_pstate_acpi_pm_profile_server())
326 		return true;
327 
328 	return acpi_ppc;
329 }
330 
331 #ifdef CONFIG_ACPI_CPPC_LIB
332 
333 /* The work item is needed to avoid CPU hotplug locking issues */
334 static void intel_pstste_sched_itmt_work_fn(struct work_struct *work)
335 {
336 	sched_set_itmt_support();
337 }
338 
339 static DECLARE_WORK(sched_itmt_work, intel_pstste_sched_itmt_work_fn);
340 
341 #define CPPC_MAX_PERF	U8_MAX
342 
343 static void intel_pstate_set_itmt_prio(int cpu)
344 {
345 	struct cppc_perf_caps cppc_perf;
346 	static u32 max_highest_perf = 0, min_highest_perf = U32_MAX;
347 	int ret;
348 
349 	ret = cppc_get_perf_caps(cpu, &cppc_perf);
350 	if (ret)
351 		return;
352 
353 	/*
354 	 * On some systems with overclocking enabled, CPPC.highest_perf is hardcoded to 0xff.
355 	 * In this case we can't use CPPC.highest_perf to enable ITMT.
356 	 * In this case we can look at MSR_HWP_CAPABILITIES bits [8:0] to decide.
357 	 */
358 	if (cppc_perf.highest_perf == CPPC_MAX_PERF)
359 		cppc_perf.highest_perf = HWP_HIGHEST_PERF(READ_ONCE(all_cpu_data[cpu]->hwp_cap_cached));
360 
361 	/*
362 	 * The priorities can be set regardless of whether or not
363 	 * sched_set_itmt_support(true) has been called and it is valid to
364 	 * update them at any time after it has been called.
365 	 */
366 	sched_set_itmt_core_prio(cppc_perf.highest_perf, cpu);
367 
368 	if (max_highest_perf <= min_highest_perf) {
369 		if (cppc_perf.highest_perf > max_highest_perf)
370 			max_highest_perf = cppc_perf.highest_perf;
371 
372 		if (cppc_perf.highest_perf < min_highest_perf)
373 			min_highest_perf = cppc_perf.highest_perf;
374 
375 		if (max_highest_perf > min_highest_perf) {
376 			/*
377 			 * This code can be run during CPU online under the
378 			 * CPU hotplug locks, so sched_set_itmt_support()
379 			 * cannot be called from here.  Queue up a work item
380 			 * to invoke it.
381 			 */
382 			schedule_work(&sched_itmt_work);
383 		}
384 	}
385 }
386 
387 static int intel_pstate_get_cppc_guaranteed(int cpu)
388 {
389 	struct cppc_perf_caps cppc_perf;
390 	int ret;
391 
392 	ret = cppc_get_perf_caps(cpu, &cppc_perf);
393 	if (ret)
394 		return ret;
395 
396 	if (cppc_perf.guaranteed_perf)
397 		return cppc_perf.guaranteed_perf;
398 
399 	return cppc_perf.nominal_perf;
400 }
401 
402 static u32 intel_pstate_cppc_nominal(int cpu)
403 {
404 	u64 nominal_perf;
405 
406 	if (cppc_get_nominal_perf(cpu, &nominal_perf))
407 		return 0;
408 
409 	return nominal_perf;
410 }
411 #else /* CONFIG_ACPI_CPPC_LIB */
412 static inline void intel_pstate_set_itmt_prio(int cpu)
413 {
414 }
415 #endif /* CONFIG_ACPI_CPPC_LIB */
416 
417 static void intel_pstate_init_acpi_perf_limits(struct cpufreq_policy *policy)
418 {
419 	struct cpudata *cpu;
420 	int ret;
421 	int i;
422 
423 	if (hwp_active) {
424 		intel_pstate_set_itmt_prio(policy->cpu);
425 		return;
426 	}
427 
428 	if (!intel_pstate_get_ppc_enable_status())
429 		return;
430 
431 	cpu = all_cpu_data[policy->cpu];
432 
433 	ret = acpi_processor_register_performance(&cpu->acpi_perf_data,
434 						  policy->cpu);
435 	if (ret)
436 		return;
437 
438 	/*
439 	 * Check if the control value in _PSS is for PERF_CTL MSR, which should
440 	 * guarantee that the states returned by it map to the states in our
441 	 * list directly.
442 	 */
443 	if (cpu->acpi_perf_data.control_register.space_id !=
444 						ACPI_ADR_SPACE_FIXED_HARDWARE)
445 		goto err;
446 
447 	/*
448 	 * If there is only one entry _PSS, simply ignore _PSS and continue as
449 	 * usual without taking _PSS into account
450 	 */
451 	if (cpu->acpi_perf_data.state_count < 2)
452 		goto err;
453 
454 	pr_debug("CPU%u - ACPI _PSS perf data\n", policy->cpu);
455 	for (i = 0; i < cpu->acpi_perf_data.state_count; i++) {
456 		pr_debug("     %cP%d: %u MHz, %u mW, 0x%x\n",
457 			 (i == cpu->acpi_perf_data.state ? '*' : ' '), i,
458 			 (u32) cpu->acpi_perf_data.states[i].core_frequency,
459 			 (u32) cpu->acpi_perf_data.states[i].power,
460 			 (u32) cpu->acpi_perf_data.states[i].control);
461 	}
462 
463 	/*
464 	 * The _PSS table doesn't contain whole turbo frequency range.
465 	 * This just contains +1 MHZ above the max non turbo frequency,
466 	 * with control value corresponding to max turbo ratio. But
467 	 * when cpufreq set policy is called, it will call with this
468 	 * max frequency, which will cause a reduced performance as
469 	 * this driver uses real max turbo frequency as the max
470 	 * frequency. So correct this frequency in _PSS table to
471 	 * correct max turbo frequency based on the turbo state.
472 	 * Also need to convert to MHz as _PSS freq is in MHz.
473 	 */
474 	if (!global.turbo_disabled)
475 		cpu->acpi_perf_data.states[0].core_frequency =
476 					policy->cpuinfo.max_freq / 1000;
477 	cpu->valid_pss_table = true;
478 	pr_debug("_PPC limits will be enforced\n");
479 
480 	return;
481 
482  err:
483 	cpu->valid_pss_table = false;
484 	acpi_processor_unregister_performance(policy->cpu);
485 }
486 
487 static void intel_pstate_exit_perf_limits(struct cpufreq_policy *policy)
488 {
489 	struct cpudata *cpu;
490 
491 	cpu = all_cpu_data[policy->cpu];
492 	if (!cpu->valid_pss_table)
493 		return;
494 
495 	acpi_processor_unregister_performance(policy->cpu);
496 }
497 #else /* CONFIG_ACPI */
498 static inline void intel_pstate_init_acpi_perf_limits(struct cpufreq_policy *policy)
499 {
500 }
501 
502 static inline void intel_pstate_exit_perf_limits(struct cpufreq_policy *policy)
503 {
504 }
505 
506 static inline bool intel_pstate_acpi_pm_profile_server(void)
507 {
508 	return false;
509 }
510 #endif /* CONFIG_ACPI */
511 
512 #ifndef CONFIG_ACPI_CPPC_LIB
513 static inline int intel_pstate_get_cppc_guaranteed(int cpu)
514 {
515 	return -ENOTSUPP;
516 }
517 #endif /* CONFIG_ACPI_CPPC_LIB */
518 
519 /**
520  * intel_pstate_hybrid_hwp_adjust - Calibrate HWP performance levels.
521  * @cpu: Target CPU.
522  *
523  * On hybrid processors, HWP may expose more performance levels than there are
524  * P-states accessible through the PERF_CTL interface.  If that happens, the
525  * scaling factor between HWP performance levels and CPU frequency will be less
526  * than the scaling factor between P-state values and CPU frequency.
527  *
528  * In that case, adjust the CPU parameters used in computations accordingly.
529  */
530 static void intel_pstate_hybrid_hwp_adjust(struct cpudata *cpu)
531 {
532 	int perf_ctl_max_phys = cpu->pstate.max_pstate_physical;
533 	int perf_ctl_scaling = cpu->pstate.perf_ctl_scaling;
534 	int perf_ctl_turbo = pstate_funcs.get_turbo();
535 	int turbo_freq = perf_ctl_turbo * perf_ctl_scaling;
536 	int scaling = cpu->pstate.scaling;
537 
538 	pr_debug("CPU%d: perf_ctl_max_phys = %d\n", cpu->cpu, perf_ctl_max_phys);
539 	pr_debug("CPU%d: perf_ctl_max = %d\n", cpu->cpu, pstate_funcs.get_max());
540 	pr_debug("CPU%d: perf_ctl_turbo = %d\n", cpu->cpu, perf_ctl_turbo);
541 	pr_debug("CPU%d: perf_ctl_scaling = %d\n", cpu->cpu, perf_ctl_scaling);
542 	pr_debug("CPU%d: HWP_CAP guaranteed = %d\n", cpu->cpu, cpu->pstate.max_pstate);
543 	pr_debug("CPU%d: HWP_CAP highest = %d\n", cpu->cpu, cpu->pstate.turbo_pstate);
544 	pr_debug("CPU%d: HWP-to-frequency scaling factor: %d\n", cpu->cpu, scaling);
545 
546 	/*
547 	 * If the product of the HWP performance scaling factor and the HWP_CAP
548 	 * highest performance is greater than the maximum turbo frequency
549 	 * corresponding to the pstate_funcs.get_turbo() return value, the
550 	 * scaling factor is too high, so recompute it to make the HWP_CAP
551 	 * highest performance correspond to the maximum turbo frequency.
552 	 */
553 	cpu->pstate.turbo_freq = cpu->pstate.turbo_pstate * scaling;
554 	if (turbo_freq < cpu->pstate.turbo_freq) {
555 		cpu->pstate.turbo_freq = turbo_freq;
556 		scaling = DIV_ROUND_UP(turbo_freq, cpu->pstate.turbo_pstate);
557 		cpu->pstate.scaling = scaling;
558 
559 		pr_debug("CPU%d: refined HWP-to-frequency scaling factor: %d\n",
560 			 cpu->cpu, scaling);
561 	}
562 
563 	cpu->pstate.max_freq = rounddown(cpu->pstate.max_pstate * scaling,
564 					 perf_ctl_scaling);
565 
566 	cpu->pstate.max_pstate_physical =
567 			DIV_ROUND_UP(perf_ctl_max_phys * perf_ctl_scaling,
568 				     scaling);
569 
570 	cpu->pstate.min_freq = cpu->pstate.min_pstate * perf_ctl_scaling;
571 	/*
572 	 * Cast the min P-state value retrieved via pstate_funcs.get_min() to
573 	 * the effective range of HWP performance levels.
574 	 */
575 	cpu->pstate.min_pstate = DIV_ROUND_UP(cpu->pstate.min_freq, scaling);
576 }
577 
578 static inline void update_turbo_state(void)
579 {
580 	u64 misc_en;
581 	struct cpudata *cpu;
582 
583 	cpu = all_cpu_data[0];
584 	rdmsrl(MSR_IA32_MISC_ENABLE, misc_en);
585 	global.turbo_disabled =
586 		(misc_en & MSR_IA32_MISC_ENABLE_TURBO_DISABLE ||
587 		 cpu->pstate.max_pstate == cpu->pstate.turbo_pstate);
588 }
589 
590 static int min_perf_pct_min(void)
591 {
592 	struct cpudata *cpu = all_cpu_data[0];
593 	int turbo_pstate = cpu->pstate.turbo_pstate;
594 
595 	return turbo_pstate ?
596 		(cpu->pstate.min_pstate * 100 / turbo_pstate) : 0;
597 }
598 
599 static s16 intel_pstate_get_epb(struct cpudata *cpu_data)
600 {
601 	u64 epb;
602 	int ret;
603 
604 	if (!boot_cpu_has(X86_FEATURE_EPB))
605 		return -ENXIO;
606 
607 	ret = rdmsrl_on_cpu(cpu_data->cpu, MSR_IA32_ENERGY_PERF_BIAS, &epb);
608 	if (ret)
609 		return (s16)ret;
610 
611 	return (s16)(epb & 0x0f);
612 }
613 
614 static s16 intel_pstate_get_epp(struct cpudata *cpu_data, u64 hwp_req_data)
615 {
616 	s16 epp;
617 
618 	if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
619 		/*
620 		 * When hwp_req_data is 0, means that caller didn't read
621 		 * MSR_HWP_REQUEST, so need to read and get EPP.
622 		 */
623 		if (!hwp_req_data) {
624 			epp = rdmsrl_on_cpu(cpu_data->cpu, MSR_HWP_REQUEST,
625 					    &hwp_req_data);
626 			if (epp)
627 				return epp;
628 		}
629 		epp = (hwp_req_data >> 24) & 0xff;
630 	} else {
631 		/* When there is no EPP present, HWP uses EPB settings */
632 		epp = intel_pstate_get_epb(cpu_data);
633 	}
634 
635 	return epp;
636 }
637 
638 static int intel_pstate_set_epb(int cpu, s16 pref)
639 {
640 	u64 epb;
641 	int ret;
642 
643 	if (!boot_cpu_has(X86_FEATURE_EPB))
644 		return -ENXIO;
645 
646 	ret = rdmsrl_on_cpu(cpu, MSR_IA32_ENERGY_PERF_BIAS, &epb);
647 	if (ret)
648 		return ret;
649 
650 	epb = (epb & ~0x0f) | pref;
651 	wrmsrl_on_cpu(cpu, MSR_IA32_ENERGY_PERF_BIAS, epb);
652 
653 	return 0;
654 }
655 
656 /*
657  * EPP/EPB display strings corresponding to EPP index in the
658  * energy_perf_strings[]
659  *	index		String
660  *-------------------------------------
661  *	0		default
662  *	1		performance
663  *	2		balance_performance
664  *	3		balance_power
665  *	4		power
666  */
667 
668 enum energy_perf_value_index {
669 	EPP_INDEX_DEFAULT = 0,
670 	EPP_INDEX_PERFORMANCE,
671 	EPP_INDEX_BALANCE_PERFORMANCE,
672 	EPP_INDEX_BALANCE_POWERSAVE,
673 	EPP_INDEX_POWERSAVE,
674 };
675 
676 static const char * const energy_perf_strings[] = {
677 	[EPP_INDEX_DEFAULT] = "default",
678 	[EPP_INDEX_PERFORMANCE] = "performance",
679 	[EPP_INDEX_BALANCE_PERFORMANCE] = "balance_performance",
680 	[EPP_INDEX_BALANCE_POWERSAVE] = "balance_power",
681 	[EPP_INDEX_POWERSAVE] = "power",
682 	NULL
683 };
684 static unsigned int epp_values[] = {
685 	[EPP_INDEX_DEFAULT] = 0, /* Unused index */
686 	[EPP_INDEX_PERFORMANCE] = HWP_EPP_PERFORMANCE,
687 	[EPP_INDEX_BALANCE_PERFORMANCE] = HWP_EPP_BALANCE_PERFORMANCE,
688 	[EPP_INDEX_BALANCE_POWERSAVE] = HWP_EPP_BALANCE_POWERSAVE,
689 	[EPP_INDEX_POWERSAVE] = HWP_EPP_POWERSAVE,
690 };
691 
692 static int intel_pstate_get_energy_pref_index(struct cpudata *cpu_data, int *raw_epp)
693 {
694 	s16 epp;
695 	int index = -EINVAL;
696 
697 	*raw_epp = 0;
698 	epp = intel_pstate_get_epp(cpu_data, 0);
699 	if (epp < 0)
700 		return epp;
701 
702 	if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
703 		if (epp == epp_values[EPP_INDEX_PERFORMANCE])
704 			return EPP_INDEX_PERFORMANCE;
705 		if (epp == epp_values[EPP_INDEX_BALANCE_PERFORMANCE])
706 			return EPP_INDEX_BALANCE_PERFORMANCE;
707 		if (epp == epp_values[EPP_INDEX_BALANCE_POWERSAVE])
708 			return EPP_INDEX_BALANCE_POWERSAVE;
709 		if (epp == epp_values[EPP_INDEX_POWERSAVE])
710 			return EPP_INDEX_POWERSAVE;
711 		*raw_epp = epp;
712 		return 0;
713 	} else if (boot_cpu_has(X86_FEATURE_EPB)) {
714 		/*
715 		 * Range:
716 		 *	0x00-0x03	:	Performance
717 		 *	0x04-0x07	:	Balance performance
718 		 *	0x08-0x0B	:	Balance power
719 		 *	0x0C-0x0F	:	Power
720 		 * The EPB is a 4 bit value, but our ranges restrict the
721 		 * value which can be set. Here only using top two bits
722 		 * effectively.
723 		 */
724 		index = (epp >> 2) + 1;
725 	}
726 
727 	return index;
728 }
729 
730 static int intel_pstate_set_epp(struct cpudata *cpu, u32 epp)
731 {
732 	int ret;
733 
734 	/*
735 	 * Use the cached HWP Request MSR value, because in the active mode the
736 	 * register itself may be updated by intel_pstate_hwp_boost_up() or
737 	 * intel_pstate_hwp_boost_down() at any time.
738 	 */
739 	u64 value = READ_ONCE(cpu->hwp_req_cached);
740 
741 	value &= ~GENMASK_ULL(31, 24);
742 	value |= (u64)epp << 24;
743 	/*
744 	 * The only other updater of hwp_req_cached in the active mode,
745 	 * intel_pstate_hwp_set(), is called under the same lock as this
746 	 * function, so it cannot run in parallel with the update below.
747 	 */
748 	WRITE_ONCE(cpu->hwp_req_cached, value);
749 	ret = wrmsrl_on_cpu(cpu->cpu, MSR_HWP_REQUEST, value);
750 	if (!ret)
751 		cpu->epp_cached = epp;
752 
753 	return ret;
754 }
755 
756 static int intel_pstate_set_energy_pref_index(struct cpudata *cpu_data,
757 					      int pref_index, bool use_raw,
758 					      u32 raw_epp)
759 {
760 	int epp = -EINVAL;
761 	int ret;
762 
763 	if (!pref_index)
764 		epp = cpu_data->epp_default;
765 
766 	if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
767 		if (use_raw)
768 			epp = raw_epp;
769 		else if (epp == -EINVAL)
770 			epp = epp_values[pref_index];
771 
772 		/*
773 		 * To avoid confusion, refuse to set EPP to any values different
774 		 * from 0 (performance) if the current policy is "performance",
775 		 * because those values would be overridden.
776 		 */
777 		if (epp > 0 && cpu_data->policy == CPUFREQ_POLICY_PERFORMANCE)
778 			return -EBUSY;
779 
780 		ret = intel_pstate_set_epp(cpu_data, epp);
781 	} else {
782 		if (epp == -EINVAL)
783 			epp = (pref_index - 1) << 2;
784 		ret = intel_pstate_set_epb(cpu_data->cpu, epp);
785 	}
786 
787 	return ret;
788 }
789 
790 static ssize_t show_energy_performance_available_preferences(
791 				struct cpufreq_policy *policy, char *buf)
792 {
793 	int i = 0;
794 	int ret = 0;
795 
796 	while (energy_perf_strings[i] != NULL)
797 		ret += sprintf(&buf[ret], "%s ", energy_perf_strings[i++]);
798 
799 	ret += sprintf(&buf[ret], "\n");
800 
801 	return ret;
802 }
803 
804 cpufreq_freq_attr_ro(energy_performance_available_preferences);
805 
806 static struct cpufreq_driver intel_pstate;
807 
808 static ssize_t store_energy_performance_preference(
809 		struct cpufreq_policy *policy, const char *buf, size_t count)
810 {
811 	struct cpudata *cpu = all_cpu_data[policy->cpu];
812 	char str_preference[21];
813 	bool raw = false;
814 	ssize_t ret;
815 	u32 epp = 0;
816 
817 	ret = sscanf(buf, "%20s", str_preference);
818 	if (ret != 1)
819 		return -EINVAL;
820 
821 	ret = match_string(energy_perf_strings, -1, str_preference);
822 	if (ret < 0) {
823 		if (!boot_cpu_has(X86_FEATURE_HWP_EPP))
824 			return ret;
825 
826 		ret = kstrtouint(buf, 10, &epp);
827 		if (ret)
828 			return ret;
829 
830 		if (epp > 255)
831 			return -EINVAL;
832 
833 		raw = true;
834 	}
835 
836 	/*
837 	 * This function runs with the policy R/W semaphore held, which
838 	 * guarantees that the driver pointer will not change while it is
839 	 * running.
840 	 */
841 	if (!intel_pstate_driver)
842 		return -EAGAIN;
843 
844 	mutex_lock(&intel_pstate_limits_lock);
845 
846 	if (intel_pstate_driver == &intel_pstate) {
847 		ret = intel_pstate_set_energy_pref_index(cpu, ret, raw, epp);
848 	} else {
849 		/*
850 		 * In the passive mode the governor needs to be stopped on the
851 		 * target CPU before the EPP update and restarted after it,
852 		 * which is super-heavy-weight, so make sure it is worth doing
853 		 * upfront.
854 		 */
855 		if (!raw)
856 			epp = ret ? epp_values[ret] : cpu->epp_default;
857 
858 		if (cpu->epp_cached != epp) {
859 			int err;
860 
861 			cpufreq_stop_governor(policy);
862 			ret = intel_pstate_set_epp(cpu, epp);
863 			err = cpufreq_start_governor(policy);
864 			if (!ret)
865 				ret = err;
866 		}
867 	}
868 
869 	mutex_unlock(&intel_pstate_limits_lock);
870 
871 	return ret ?: count;
872 }
873 
874 static ssize_t show_energy_performance_preference(
875 				struct cpufreq_policy *policy, char *buf)
876 {
877 	struct cpudata *cpu_data = all_cpu_data[policy->cpu];
878 	int preference, raw_epp;
879 
880 	preference = intel_pstate_get_energy_pref_index(cpu_data, &raw_epp);
881 	if (preference < 0)
882 		return preference;
883 
884 	if (raw_epp)
885 		return  sprintf(buf, "%d\n", raw_epp);
886 	else
887 		return  sprintf(buf, "%s\n", energy_perf_strings[preference]);
888 }
889 
890 cpufreq_freq_attr_rw(energy_performance_preference);
891 
892 static ssize_t show_base_frequency(struct cpufreq_policy *policy, char *buf)
893 {
894 	struct cpudata *cpu = all_cpu_data[policy->cpu];
895 	int ratio, freq;
896 
897 	ratio = intel_pstate_get_cppc_guaranteed(policy->cpu);
898 	if (ratio <= 0) {
899 		u64 cap;
900 
901 		rdmsrl_on_cpu(policy->cpu, MSR_HWP_CAPABILITIES, &cap);
902 		ratio = HWP_GUARANTEED_PERF(cap);
903 	}
904 
905 	freq = ratio * cpu->pstate.scaling;
906 	if (cpu->pstate.scaling != cpu->pstate.perf_ctl_scaling)
907 		freq = rounddown(freq, cpu->pstate.perf_ctl_scaling);
908 
909 	return sprintf(buf, "%d\n", freq);
910 }
911 
912 cpufreq_freq_attr_ro(base_frequency);
913 
914 static struct freq_attr *hwp_cpufreq_attrs[] = {
915 	&energy_performance_preference,
916 	&energy_performance_available_preferences,
917 	&base_frequency,
918 	NULL,
919 };
920 
921 static void __intel_pstate_get_hwp_cap(struct cpudata *cpu)
922 {
923 	u64 cap;
924 
925 	rdmsrl_on_cpu(cpu->cpu, MSR_HWP_CAPABILITIES, &cap);
926 	WRITE_ONCE(cpu->hwp_cap_cached, cap);
927 	cpu->pstate.max_pstate = HWP_GUARANTEED_PERF(cap);
928 	cpu->pstate.turbo_pstate = HWP_HIGHEST_PERF(cap);
929 }
930 
931 static void intel_pstate_get_hwp_cap(struct cpudata *cpu)
932 {
933 	int scaling = cpu->pstate.scaling;
934 
935 	__intel_pstate_get_hwp_cap(cpu);
936 
937 	cpu->pstate.max_freq = cpu->pstate.max_pstate * scaling;
938 	cpu->pstate.turbo_freq = cpu->pstate.turbo_pstate * scaling;
939 	if (scaling != cpu->pstate.perf_ctl_scaling) {
940 		int perf_ctl_scaling = cpu->pstate.perf_ctl_scaling;
941 
942 		cpu->pstate.max_freq = rounddown(cpu->pstate.max_freq,
943 						 perf_ctl_scaling);
944 		cpu->pstate.turbo_freq = rounddown(cpu->pstate.turbo_freq,
945 						   perf_ctl_scaling);
946 	}
947 }
948 
949 static void intel_pstate_hwp_set(unsigned int cpu)
950 {
951 	struct cpudata *cpu_data = all_cpu_data[cpu];
952 	int max, min;
953 	u64 value;
954 	s16 epp;
955 
956 	max = cpu_data->max_perf_ratio;
957 	min = cpu_data->min_perf_ratio;
958 
959 	if (cpu_data->policy == CPUFREQ_POLICY_PERFORMANCE)
960 		min = max;
961 
962 	rdmsrl_on_cpu(cpu, MSR_HWP_REQUEST, &value);
963 
964 	value &= ~HWP_MIN_PERF(~0L);
965 	value |= HWP_MIN_PERF(min);
966 
967 	value &= ~HWP_MAX_PERF(~0L);
968 	value |= HWP_MAX_PERF(max);
969 
970 	if (cpu_data->epp_policy == cpu_data->policy)
971 		goto skip_epp;
972 
973 	cpu_data->epp_policy = cpu_data->policy;
974 
975 	if (cpu_data->policy == CPUFREQ_POLICY_PERFORMANCE) {
976 		epp = intel_pstate_get_epp(cpu_data, value);
977 		cpu_data->epp_powersave = epp;
978 		/* If EPP read was failed, then don't try to write */
979 		if (epp < 0)
980 			goto skip_epp;
981 
982 		epp = 0;
983 	} else {
984 		/* skip setting EPP, when saved value is invalid */
985 		if (cpu_data->epp_powersave < 0)
986 			goto skip_epp;
987 
988 		/*
989 		 * No need to restore EPP when it is not zero. This
990 		 * means:
991 		 *  - Policy is not changed
992 		 *  - user has manually changed
993 		 *  - Error reading EPB
994 		 */
995 		epp = intel_pstate_get_epp(cpu_data, value);
996 		if (epp)
997 			goto skip_epp;
998 
999 		epp = cpu_data->epp_powersave;
1000 	}
1001 	if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
1002 		value &= ~GENMASK_ULL(31, 24);
1003 		value |= (u64)epp << 24;
1004 	} else {
1005 		intel_pstate_set_epb(cpu, epp);
1006 	}
1007 skip_epp:
1008 	WRITE_ONCE(cpu_data->hwp_req_cached, value);
1009 	wrmsrl_on_cpu(cpu, MSR_HWP_REQUEST, value);
1010 }
1011 
1012 static void intel_pstate_disable_hwp_interrupt(struct cpudata *cpudata);
1013 
1014 static void intel_pstate_hwp_offline(struct cpudata *cpu)
1015 {
1016 	u64 value = READ_ONCE(cpu->hwp_req_cached);
1017 	int min_perf;
1018 
1019 	intel_pstate_disable_hwp_interrupt(cpu);
1020 
1021 	if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
1022 		/*
1023 		 * In case the EPP has been set to "performance" by the
1024 		 * active mode "performance" scaling algorithm, replace that
1025 		 * temporary value with the cached EPP one.
1026 		 */
1027 		value &= ~GENMASK_ULL(31, 24);
1028 		value |= HWP_ENERGY_PERF_PREFERENCE(cpu->epp_cached);
1029 		/*
1030 		 * However, make sure that EPP will be set to "performance" when
1031 		 * the CPU is brought back online again and the "performance"
1032 		 * scaling algorithm is still in effect.
1033 		 */
1034 		cpu->epp_policy = CPUFREQ_POLICY_UNKNOWN;
1035 	}
1036 
1037 	/*
1038 	 * Clear the desired perf field in the cached HWP request value to
1039 	 * prevent nonzero desired values from being leaked into the active
1040 	 * mode.
1041 	 */
1042 	value &= ~HWP_DESIRED_PERF(~0L);
1043 	WRITE_ONCE(cpu->hwp_req_cached, value);
1044 
1045 	value &= ~GENMASK_ULL(31, 0);
1046 	min_perf = HWP_LOWEST_PERF(READ_ONCE(cpu->hwp_cap_cached));
1047 
1048 	/* Set hwp_max = hwp_min */
1049 	value |= HWP_MAX_PERF(min_perf);
1050 	value |= HWP_MIN_PERF(min_perf);
1051 
1052 	/* Set EPP to min */
1053 	if (boot_cpu_has(X86_FEATURE_HWP_EPP))
1054 		value |= HWP_ENERGY_PERF_PREFERENCE(HWP_EPP_POWERSAVE);
1055 
1056 	wrmsrl_on_cpu(cpu->cpu, MSR_HWP_REQUEST, value);
1057 }
1058 
1059 #define POWER_CTL_EE_ENABLE	1
1060 #define POWER_CTL_EE_DISABLE	2
1061 
1062 static int power_ctl_ee_state;
1063 
1064 static void set_power_ctl_ee_state(bool input)
1065 {
1066 	u64 power_ctl;
1067 
1068 	mutex_lock(&intel_pstate_driver_lock);
1069 	rdmsrl(MSR_IA32_POWER_CTL, power_ctl);
1070 	if (input) {
1071 		power_ctl &= ~BIT(MSR_IA32_POWER_CTL_BIT_EE);
1072 		power_ctl_ee_state = POWER_CTL_EE_ENABLE;
1073 	} else {
1074 		power_ctl |= BIT(MSR_IA32_POWER_CTL_BIT_EE);
1075 		power_ctl_ee_state = POWER_CTL_EE_DISABLE;
1076 	}
1077 	wrmsrl(MSR_IA32_POWER_CTL, power_ctl);
1078 	mutex_unlock(&intel_pstate_driver_lock);
1079 }
1080 
1081 static void intel_pstate_hwp_enable(struct cpudata *cpudata);
1082 
1083 static void intel_pstate_hwp_reenable(struct cpudata *cpu)
1084 {
1085 	intel_pstate_hwp_enable(cpu);
1086 	wrmsrl_on_cpu(cpu->cpu, MSR_HWP_REQUEST, READ_ONCE(cpu->hwp_req_cached));
1087 }
1088 
1089 static int intel_pstate_suspend(struct cpufreq_policy *policy)
1090 {
1091 	struct cpudata *cpu = all_cpu_data[policy->cpu];
1092 
1093 	pr_debug("CPU %d suspending\n", cpu->cpu);
1094 
1095 	cpu->suspended = true;
1096 
1097 	/* disable HWP interrupt and cancel any pending work */
1098 	intel_pstate_disable_hwp_interrupt(cpu);
1099 
1100 	return 0;
1101 }
1102 
1103 static int intel_pstate_resume(struct cpufreq_policy *policy)
1104 {
1105 	struct cpudata *cpu = all_cpu_data[policy->cpu];
1106 
1107 	pr_debug("CPU %d resuming\n", cpu->cpu);
1108 
1109 	/* Only restore if the system default is changed */
1110 	if (power_ctl_ee_state == POWER_CTL_EE_ENABLE)
1111 		set_power_ctl_ee_state(true);
1112 	else if (power_ctl_ee_state == POWER_CTL_EE_DISABLE)
1113 		set_power_ctl_ee_state(false);
1114 
1115 	if (cpu->suspended && hwp_active) {
1116 		mutex_lock(&intel_pstate_limits_lock);
1117 
1118 		/* Re-enable HWP, because "online" has not done that. */
1119 		intel_pstate_hwp_reenable(cpu);
1120 
1121 		mutex_unlock(&intel_pstate_limits_lock);
1122 	}
1123 
1124 	cpu->suspended = false;
1125 
1126 	return 0;
1127 }
1128 
1129 static void intel_pstate_update_policies(void)
1130 {
1131 	int cpu;
1132 
1133 	for_each_possible_cpu(cpu)
1134 		cpufreq_update_policy(cpu);
1135 }
1136 
1137 static void __intel_pstate_update_max_freq(struct cpudata *cpudata,
1138 					   struct cpufreq_policy *policy)
1139 {
1140 	policy->cpuinfo.max_freq = global.turbo_disabled_mf ?
1141 			cpudata->pstate.max_freq : cpudata->pstate.turbo_freq;
1142 	refresh_frequency_limits(policy);
1143 }
1144 
1145 static void intel_pstate_update_max_freq(unsigned int cpu)
1146 {
1147 	struct cpufreq_policy *policy = cpufreq_cpu_acquire(cpu);
1148 
1149 	if (!policy)
1150 		return;
1151 
1152 	__intel_pstate_update_max_freq(all_cpu_data[cpu], policy);
1153 
1154 	cpufreq_cpu_release(policy);
1155 }
1156 
1157 static void intel_pstate_update_limits(unsigned int cpu)
1158 {
1159 	mutex_lock(&intel_pstate_driver_lock);
1160 
1161 	update_turbo_state();
1162 	/*
1163 	 * If turbo has been turned on or off globally, policy limits for
1164 	 * all CPUs need to be updated to reflect that.
1165 	 */
1166 	if (global.turbo_disabled_mf != global.turbo_disabled) {
1167 		global.turbo_disabled_mf = global.turbo_disabled;
1168 		arch_set_max_freq_ratio(global.turbo_disabled);
1169 		for_each_possible_cpu(cpu)
1170 			intel_pstate_update_max_freq(cpu);
1171 	} else {
1172 		cpufreq_update_policy(cpu);
1173 	}
1174 
1175 	mutex_unlock(&intel_pstate_driver_lock);
1176 }
1177 
1178 /************************** sysfs begin ************************/
1179 #define show_one(file_name, object)					\
1180 	static ssize_t show_##file_name					\
1181 	(struct kobject *kobj, struct kobj_attribute *attr, char *buf)	\
1182 	{								\
1183 		return sprintf(buf, "%u\n", global.object);		\
1184 	}
1185 
1186 static ssize_t intel_pstate_show_status(char *buf);
1187 static int intel_pstate_update_status(const char *buf, size_t size);
1188 
1189 static ssize_t show_status(struct kobject *kobj,
1190 			   struct kobj_attribute *attr, char *buf)
1191 {
1192 	ssize_t ret;
1193 
1194 	mutex_lock(&intel_pstate_driver_lock);
1195 	ret = intel_pstate_show_status(buf);
1196 	mutex_unlock(&intel_pstate_driver_lock);
1197 
1198 	return ret;
1199 }
1200 
1201 static ssize_t store_status(struct kobject *a, struct kobj_attribute *b,
1202 			    const char *buf, size_t count)
1203 {
1204 	char *p = memchr(buf, '\n', count);
1205 	int ret;
1206 
1207 	mutex_lock(&intel_pstate_driver_lock);
1208 	ret = intel_pstate_update_status(buf, p ? p - buf : count);
1209 	mutex_unlock(&intel_pstate_driver_lock);
1210 
1211 	return ret < 0 ? ret : count;
1212 }
1213 
1214 static ssize_t show_turbo_pct(struct kobject *kobj,
1215 				struct kobj_attribute *attr, char *buf)
1216 {
1217 	struct cpudata *cpu;
1218 	int total, no_turbo, turbo_pct;
1219 	uint32_t turbo_fp;
1220 
1221 	mutex_lock(&intel_pstate_driver_lock);
1222 
1223 	if (!intel_pstate_driver) {
1224 		mutex_unlock(&intel_pstate_driver_lock);
1225 		return -EAGAIN;
1226 	}
1227 
1228 	cpu = all_cpu_data[0];
1229 
1230 	total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
1231 	no_turbo = cpu->pstate.max_pstate - cpu->pstate.min_pstate + 1;
1232 	turbo_fp = div_fp(no_turbo, total);
1233 	turbo_pct = 100 - fp_toint(mul_fp(turbo_fp, int_tofp(100)));
1234 
1235 	mutex_unlock(&intel_pstate_driver_lock);
1236 
1237 	return sprintf(buf, "%u\n", turbo_pct);
1238 }
1239 
1240 static ssize_t show_num_pstates(struct kobject *kobj,
1241 				struct kobj_attribute *attr, char *buf)
1242 {
1243 	struct cpudata *cpu;
1244 	int total;
1245 
1246 	mutex_lock(&intel_pstate_driver_lock);
1247 
1248 	if (!intel_pstate_driver) {
1249 		mutex_unlock(&intel_pstate_driver_lock);
1250 		return -EAGAIN;
1251 	}
1252 
1253 	cpu = all_cpu_data[0];
1254 	total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
1255 
1256 	mutex_unlock(&intel_pstate_driver_lock);
1257 
1258 	return sprintf(buf, "%u\n", total);
1259 }
1260 
1261 static ssize_t show_no_turbo(struct kobject *kobj,
1262 			     struct kobj_attribute *attr, char *buf)
1263 {
1264 	ssize_t ret;
1265 
1266 	mutex_lock(&intel_pstate_driver_lock);
1267 
1268 	if (!intel_pstate_driver) {
1269 		mutex_unlock(&intel_pstate_driver_lock);
1270 		return -EAGAIN;
1271 	}
1272 
1273 	update_turbo_state();
1274 	if (global.turbo_disabled)
1275 		ret = sprintf(buf, "%u\n", global.turbo_disabled);
1276 	else
1277 		ret = sprintf(buf, "%u\n", global.no_turbo);
1278 
1279 	mutex_unlock(&intel_pstate_driver_lock);
1280 
1281 	return ret;
1282 }
1283 
1284 static ssize_t store_no_turbo(struct kobject *a, struct kobj_attribute *b,
1285 			      const char *buf, size_t count)
1286 {
1287 	unsigned int input;
1288 	int ret;
1289 
1290 	ret = sscanf(buf, "%u", &input);
1291 	if (ret != 1)
1292 		return -EINVAL;
1293 
1294 	mutex_lock(&intel_pstate_driver_lock);
1295 
1296 	if (!intel_pstate_driver) {
1297 		mutex_unlock(&intel_pstate_driver_lock);
1298 		return -EAGAIN;
1299 	}
1300 
1301 	mutex_lock(&intel_pstate_limits_lock);
1302 
1303 	update_turbo_state();
1304 	if (global.turbo_disabled) {
1305 		pr_notice_once("Turbo disabled by BIOS or unavailable on processor\n");
1306 		mutex_unlock(&intel_pstate_limits_lock);
1307 		mutex_unlock(&intel_pstate_driver_lock);
1308 		return -EPERM;
1309 	}
1310 
1311 	global.no_turbo = clamp_t(int, input, 0, 1);
1312 
1313 	if (global.no_turbo) {
1314 		struct cpudata *cpu = all_cpu_data[0];
1315 		int pct = cpu->pstate.max_pstate * 100 / cpu->pstate.turbo_pstate;
1316 
1317 		/* Squash the global minimum into the permitted range. */
1318 		if (global.min_perf_pct > pct)
1319 			global.min_perf_pct = pct;
1320 	}
1321 
1322 	mutex_unlock(&intel_pstate_limits_lock);
1323 
1324 	intel_pstate_update_policies();
1325 	arch_set_max_freq_ratio(global.no_turbo);
1326 
1327 	mutex_unlock(&intel_pstate_driver_lock);
1328 
1329 	return count;
1330 }
1331 
1332 static void update_qos_request(enum freq_qos_req_type type)
1333 {
1334 	struct freq_qos_request *req;
1335 	struct cpufreq_policy *policy;
1336 	int i;
1337 
1338 	for_each_possible_cpu(i) {
1339 		struct cpudata *cpu = all_cpu_data[i];
1340 		unsigned int freq, perf_pct;
1341 
1342 		policy = cpufreq_cpu_get(i);
1343 		if (!policy)
1344 			continue;
1345 
1346 		req = policy->driver_data;
1347 		cpufreq_cpu_put(policy);
1348 
1349 		if (!req)
1350 			continue;
1351 
1352 		if (hwp_active)
1353 			intel_pstate_get_hwp_cap(cpu);
1354 
1355 		if (type == FREQ_QOS_MIN) {
1356 			perf_pct = global.min_perf_pct;
1357 		} else {
1358 			req++;
1359 			perf_pct = global.max_perf_pct;
1360 		}
1361 
1362 		freq = DIV_ROUND_UP(cpu->pstate.turbo_freq * perf_pct, 100);
1363 
1364 		if (freq_qos_update_request(req, freq) < 0)
1365 			pr_warn("Failed to update freq constraint: CPU%d\n", i);
1366 	}
1367 }
1368 
1369 static ssize_t store_max_perf_pct(struct kobject *a, struct kobj_attribute *b,
1370 				  const char *buf, size_t count)
1371 {
1372 	unsigned int input;
1373 	int ret;
1374 
1375 	ret = sscanf(buf, "%u", &input);
1376 	if (ret != 1)
1377 		return -EINVAL;
1378 
1379 	mutex_lock(&intel_pstate_driver_lock);
1380 
1381 	if (!intel_pstate_driver) {
1382 		mutex_unlock(&intel_pstate_driver_lock);
1383 		return -EAGAIN;
1384 	}
1385 
1386 	mutex_lock(&intel_pstate_limits_lock);
1387 
1388 	global.max_perf_pct = clamp_t(int, input, global.min_perf_pct, 100);
1389 
1390 	mutex_unlock(&intel_pstate_limits_lock);
1391 
1392 	if (intel_pstate_driver == &intel_pstate)
1393 		intel_pstate_update_policies();
1394 	else
1395 		update_qos_request(FREQ_QOS_MAX);
1396 
1397 	mutex_unlock(&intel_pstate_driver_lock);
1398 
1399 	return count;
1400 }
1401 
1402 static ssize_t store_min_perf_pct(struct kobject *a, struct kobj_attribute *b,
1403 				  const char *buf, size_t count)
1404 {
1405 	unsigned int input;
1406 	int ret;
1407 
1408 	ret = sscanf(buf, "%u", &input);
1409 	if (ret != 1)
1410 		return -EINVAL;
1411 
1412 	mutex_lock(&intel_pstate_driver_lock);
1413 
1414 	if (!intel_pstate_driver) {
1415 		mutex_unlock(&intel_pstate_driver_lock);
1416 		return -EAGAIN;
1417 	}
1418 
1419 	mutex_lock(&intel_pstate_limits_lock);
1420 
1421 	global.min_perf_pct = clamp_t(int, input,
1422 				      min_perf_pct_min(), global.max_perf_pct);
1423 
1424 	mutex_unlock(&intel_pstate_limits_lock);
1425 
1426 	if (intel_pstate_driver == &intel_pstate)
1427 		intel_pstate_update_policies();
1428 	else
1429 		update_qos_request(FREQ_QOS_MIN);
1430 
1431 	mutex_unlock(&intel_pstate_driver_lock);
1432 
1433 	return count;
1434 }
1435 
1436 static ssize_t show_hwp_dynamic_boost(struct kobject *kobj,
1437 				struct kobj_attribute *attr, char *buf)
1438 {
1439 	return sprintf(buf, "%u\n", hwp_boost);
1440 }
1441 
1442 static ssize_t store_hwp_dynamic_boost(struct kobject *a,
1443 				       struct kobj_attribute *b,
1444 				       const char *buf, size_t count)
1445 {
1446 	unsigned int input;
1447 	int ret;
1448 
1449 	ret = kstrtouint(buf, 10, &input);
1450 	if (ret)
1451 		return ret;
1452 
1453 	mutex_lock(&intel_pstate_driver_lock);
1454 	hwp_boost = !!input;
1455 	intel_pstate_update_policies();
1456 	mutex_unlock(&intel_pstate_driver_lock);
1457 
1458 	return count;
1459 }
1460 
1461 static ssize_t show_energy_efficiency(struct kobject *kobj, struct kobj_attribute *attr,
1462 				      char *buf)
1463 {
1464 	u64 power_ctl;
1465 	int enable;
1466 
1467 	rdmsrl(MSR_IA32_POWER_CTL, power_ctl);
1468 	enable = !!(power_ctl & BIT(MSR_IA32_POWER_CTL_BIT_EE));
1469 	return sprintf(buf, "%d\n", !enable);
1470 }
1471 
1472 static ssize_t store_energy_efficiency(struct kobject *a, struct kobj_attribute *b,
1473 				       const char *buf, size_t count)
1474 {
1475 	bool input;
1476 	int ret;
1477 
1478 	ret = kstrtobool(buf, &input);
1479 	if (ret)
1480 		return ret;
1481 
1482 	set_power_ctl_ee_state(input);
1483 
1484 	return count;
1485 }
1486 
1487 show_one(max_perf_pct, max_perf_pct);
1488 show_one(min_perf_pct, min_perf_pct);
1489 
1490 define_one_global_rw(status);
1491 define_one_global_rw(no_turbo);
1492 define_one_global_rw(max_perf_pct);
1493 define_one_global_rw(min_perf_pct);
1494 define_one_global_ro(turbo_pct);
1495 define_one_global_ro(num_pstates);
1496 define_one_global_rw(hwp_dynamic_boost);
1497 define_one_global_rw(energy_efficiency);
1498 
1499 static struct attribute *intel_pstate_attributes[] = {
1500 	&status.attr,
1501 	&no_turbo.attr,
1502 	NULL
1503 };
1504 
1505 static const struct attribute_group intel_pstate_attr_group = {
1506 	.attrs = intel_pstate_attributes,
1507 };
1508 
1509 static const struct x86_cpu_id intel_pstate_cpu_ee_disable_ids[];
1510 
1511 static struct kobject *intel_pstate_kobject;
1512 
1513 static void __init intel_pstate_sysfs_expose_params(void)
1514 {
1515 	int rc;
1516 
1517 	intel_pstate_kobject = kobject_create_and_add("intel_pstate",
1518 						&cpu_subsys.dev_root->kobj);
1519 	if (WARN_ON(!intel_pstate_kobject))
1520 		return;
1521 
1522 	rc = sysfs_create_group(intel_pstate_kobject, &intel_pstate_attr_group);
1523 	if (WARN_ON(rc))
1524 		return;
1525 
1526 	if (!boot_cpu_has(X86_FEATURE_HYBRID_CPU)) {
1527 		rc = sysfs_create_file(intel_pstate_kobject, &turbo_pct.attr);
1528 		WARN_ON(rc);
1529 
1530 		rc = sysfs_create_file(intel_pstate_kobject, &num_pstates.attr);
1531 		WARN_ON(rc);
1532 	}
1533 
1534 	/*
1535 	 * If per cpu limits are enforced there are no global limits, so
1536 	 * return without creating max/min_perf_pct attributes
1537 	 */
1538 	if (per_cpu_limits)
1539 		return;
1540 
1541 	rc = sysfs_create_file(intel_pstate_kobject, &max_perf_pct.attr);
1542 	WARN_ON(rc);
1543 
1544 	rc = sysfs_create_file(intel_pstate_kobject, &min_perf_pct.attr);
1545 	WARN_ON(rc);
1546 
1547 	if (x86_match_cpu(intel_pstate_cpu_ee_disable_ids)) {
1548 		rc = sysfs_create_file(intel_pstate_kobject, &energy_efficiency.attr);
1549 		WARN_ON(rc);
1550 	}
1551 }
1552 
1553 static void __init intel_pstate_sysfs_remove(void)
1554 {
1555 	if (!intel_pstate_kobject)
1556 		return;
1557 
1558 	sysfs_remove_group(intel_pstate_kobject, &intel_pstate_attr_group);
1559 
1560 	if (!boot_cpu_has(X86_FEATURE_HYBRID_CPU)) {
1561 		sysfs_remove_file(intel_pstate_kobject, &num_pstates.attr);
1562 		sysfs_remove_file(intel_pstate_kobject, &turbo_pct.attr);
1563 	}
1564 
1565 	if (!per_cpu_limits) {
1566 		sysfs_remove_file(intel_pstate_kobject, &max_perf_pct.attr);
1567 		sysfs_remove_file(intel_pstate_kobject, &min_perf_pct.attr);
1568 
1569 		if (x86_match_cpu(intel_pstate_cpu_ee_disable_ids))
1570 			sysfs_remove_file(intel_pstate_kobject, &energy_efficiency.attr);
1571 	}
1572 
1573 	kobject_put(intel_pstate_kobject);
1574 }
1575 
1576 static void intel_pstate_sysfs_expose_hwp_dynamic_boost(void)
1577 {
1578 	int rc;
1579 
1580 	if (!hwp_active)
1581 		return;
1582 
1583 	rc = sysfs_create_file(intel_pstate_kobject, &hwp_dynamic_boost.attr);
1584 	WARN_ON_ONCE(rc);
1585 }
1586 
1587 static void intel_pstate_sysfs_hide_hwp_dynamic_boost(void)
1588 {
1589 	if (!hwp_active)
1590 		return;
1591 
1592 	sysfs_remove_file(intel_pstate_kobject, &hwp_dynamic_boost.attr);
1593 }
1594 
1595 /************************** sysfs end ************************/
1596 
1597 static void intel_pstate_notify_work(struct work_struct *work)
1598 {
1599 	struct cpudata *cpudata =
1600 		container_of(to_delayed_work(work), struct cpudata, hwp_notify_work);
1601 	struct cpufreq_policy *policy = cpufreq_cpu_acquire(cpudata->cpu);
1602 
1603 	if (policy) {
1604 		intel_pstate_get_hwp_cap(cpudata);
1605 		__intel_pstate_update_max_freq(cpudata, policy);
1606 
1607 		cpufreq_cpu_release(policy);
1608 	}
1609 
1610 	wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_STATUS, 0);
1611 }
1612 
1613 static DEFINE_SPINLOCK(hwp_notify_lock);
1614 static cpumask_t hwp_intr_enable_mask;
1615 
1616 void notify_hwp_interrupt(void)
1617 {
1618 	unsigned int this_cpu = smp_processor_id();
1619 	struct cpudata *cpudata;
1620 	unsigned long flags;
1621 	u64 value;
1622 
1623 	if (!READ_ONCE(hwp_active) || !boot_cpu_has(X86_FEATURE_HWP_NOTIFY))
1624 		return;
1625 
1626 	rdmsrl_safe(MSR_HWP_STATUS, &value);
1627 	if (!(value & 0x01))
1628 		return;
1629 
1630 	spin_lock_irqsave(&hwp_notify_lock, flags);
1631 
1632 	if (!cpumask_test_cpu(this_cpu, &hwp_intr_enable_mask))
1633 		goto ack_intr;
1634 
1635 	/*
1636 	 * Currently we never free all_cpu_data. And we can't reach here
1637 	 * without this allocated. But for safety for future changes, added
1638 	 * check.
1639 	 */
1640 	if (unlikely(!READ_ONCE(all_cpu_data)))
1641 		goto ack_intr;
1642 
1643 	/*
1644 	 * The free is done during cleanup, when cpufreq registry is failed.
1645 	 * We wouldn't be here if it fails on init or switch status. But for
1646 	 * future changes, added check.
1647 	 */
1648 	cpudata = READ_ONCE(all_cpu_data[this_cpu]);
1649 	if (unlikely(!cpudata))
1650 		goto ack_intr;
1651 
1652 	schedule_delayed_work(&cpudata->hwp_notify_work, msecs_to_jiffies(10));
1653 
1654 	spin_unlock_irqrestore(&hwp_notify_lock, flags);
1655 
1656 	return;
1657 
1658 ack_intr:
1659 	wrmsrl_safe(MSR_HWP_STATUS, 0);
1660 	spin_unlock_irqrestore(&hwp_notify_lock, flags);
1661 }
1662 
1663 static void intel_pstate_disable_hwp_interrupt(struct cpudata *cpudata)
1664 {
1665 	unsigned long flags;
1666 
1667 	if (!boot_cpu_has(X86_FEATURE_HWP_NOTIFY))
1668 		return;
1669 
1670 	/* wrmsrl_on_cpu has to be outside spinlock as this can result in IPC */
1671 	wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_INTERRUPT, 0x00);
1672 
1673 	spin_lock_irqsave(&hwp_notify_lock, flags);
1674 	if (cpumask_test_and_clear_cpu(cpudata->cpu, &hwp_intr_enable_mask))
1675 		cancel_delayed_work(&cpudata->hwp_notify_work);
1676 	spin_unlock_irqrestore(&hwp_notify_lock, flags);
1677 }
1678 
1679 static void intel_pstate_enable_hwp_interrupt(struct cpudata *cpudata)
1680 {
1681 	/* Enable HWP notification interrupt for guaranteed performance change */
1682 	if (boot_cpu_has(X86_FEATURE_HWP_NOTIFY)) {
1683 		unsigned long flags;
1684 
1685 		spin_lock_irqsave(&hwp_notify_lock, flags);
1686 		INIT_DELAYED_WORK(&cpudata->hwp_notify_work, intel_pstate_notify_work);
1687 		cpumask_set_cpu(cpudata->cpu, &hwp_intr_enable_mask);
1688 		spin_unlock_irqrestore(&hwp_notify_lock, flags);
1689 
1690 		/* wrmsrl_on_cpu has to be outside spinlock as this can result in IPC */
1691 		wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_INTERRUPT, 0x01);
1692 		wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_STATUS, 0);
1693 	}
1694 }
1695 
1696 static void intel_pstate_update_epp_defaults(struct cpudata *cpudata)
1697 {
1698 	cpudata->epp_default = intel_pstate_get_epp(cpudata, 0);
1699 
1700 	/*
1701 	 * If this CPU gen doesn't call for change in balance_perf
1702 	 * EPP return.
1703 	 */
1704 	if (epp_values[EPP_INDEX_BALANCE_PERFORMANCE] == HWP_EPP_BALANCE_PERFORMANCE)
1705 		return;
1706 
1707 	/*
1708 	 * If powerup EPP is something other than chipset default 0x80 and
1709 	 * - is more performance oriented than 0x80 (default balance_perf EPP)
1710 	 * - But less performance oriented than performance EPP
1711 	 *   then use this as new balance_perf EPP.
1712 	 */
1713 	if (cpudata->epp_default < HWP_EPP_BALANCE_PERFORMANCE &&
1714 	    cpudata->epp_default > HWP_EPP_PERFORMANCE) {
1715 		epp_values[EPP_INDEX_BALANCE_PERFORMANCE] = cpudata->epp_default;
1716 		return;
1717 	}
1718 
1719 	/*
1720 	 * Use hard coded value per gen to update the balance_perf
1721 	 * and default EPP.
1722 	 */
1723 	cpudata->epp_default = epp_values[EPP_INDEX_BALANCE_PERFORMANCE];
1724 	intel_pstate_set_epp(cpudata, cpudata->epp_default);
1725 }
1726 
1727 static void intel_pstate_hwp_enable(struct cpudata *cpudata)
1728 {
1729 	/* First disable HWP notification interrupt till we activate again */
1730 	if (boot_cpu_has(X86_FEATURE_HWP_NOTIFY))
1731 		wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_INTERRUPT, 0x00);
1732 
1733 	wrmsrl_on_cpu(cpudata->cpu, MSR_PM_ENABLE, 0x1);
1734 
1735 	intel_pstate_enable_hwp_interrupt(cpudata);
1736 
1737 	if (cpudata->epp_default >= 0)
1738 		return;
1739 
1740 	intel_pstate_update_epp_defaults(cpudata);
1741 }
1742 
1743 static int atom_get_min_pstate(void)
1744 {
1745 	u64 value;
1746 
1747 	rdmsrl(MSR_ATOM_CORE_RATIOS, value);
1748 	return (value >> 8) & 0x7F;
1749 }
1750 
1751 static int atom_get_max_pstate(void)
1752 {
1753 	u64 value;
1754 
1755 	rdmsrl(MSR_ATOM_CORE_RATIOS, value);
1756 	return (value >> 16) & 0x7F;
1757 }
1758 
1759 static int atom_get_turbo_pstate(void)
1760 {
1761 	u64 value;
1762 
1763 	rdmsrl(MSR_ATOM_CORE_TURBO_RATIOS, value);
1764 	return value & 0x7F;
1765 }
1766 
1767 static u64 atom_get_val(struct cpudata *cpudata, int pstate)
1768 {
1769 	u64 val;
1770 	int32_t vid_fp;
1771 	u32 vid;
1772 
1773 	val = (u64)pstate << 8;
1774 	if (global.no_turbo && !global.turbo_disabled)
1775 		val |= (u64)1 << 32;
1776 
1777 	vid_fp = cpudata->vid.min + mul_fp(
1778 		int_tofp(pstate - cpudata->pstate.min_pstate),
1779 		cpudata->vid.ratio);
1780 
1781 	vid_fp = clamp_t(int32_t, vid_fp, cpudata->vid.min, cpudata->vid.max);
1782 	vid = ceiling_fp(vid_fp);
1783 
1784 	if (pstate > cpudata->pstate.max_pstate)
1785 		vid = cpudata->vid.turbo;
1786 
1787 	return val | vid;
1788 }
1789 
1790 static int silvermont_get_scaling(void)
1791 {
1792 	u64 value;
1793 	int i;
1794 	/* Defined in Table 35-6 from SDM (Sept 2015) */
1795 	static int silvermont_freq_table[] = {
1796 		83300, 100000, 133300, 116700, 80000};
1797 
1798 	rdmsrl(MSR_FSB_FREQ, value);
1799 	i = value & 0x7;
1800 	WARN_ON(i > 4);
1801 
1802 	return silvermont_freq_table[i];
1803 }
1804 
1805 static int airmont_get_scaling(void)
1806 {
1807 	u64 value;
1808 	int i;
1809 	/* Defined in Table 35-10 from SDM (Sept 2015) */
1810 	static int airmont_freq_table[] = {
1811 		83300, 100000, 133300, 116700, 80000,
1812 		93300, 90000, 88900, 87500};
1813 
1814 	rdmsrl(MSR_FSB_FREQ, value);
1815 	i = value & 0xF;
1816 	WARN_ON(i > 8);
1817 
1818 	return airmont_freq_table[i];
1819 }
1820 
1821 static void atom_get_vid(struct cpudata *cpudata)
1822 {
1823 	u64 value;
1824 
1825 	rdmsrl(MSR_ATOM_CORE_VIDS, value);
1826 	cpudata->vid.min = int_tofp((value >> 8) & 0x7f);
1827 	cpudata->vid.max = int_tofp((value >> 16) & 0x7f);
1828 	cpudata->vid.ratio = div_fp(
1829 		cpudata->vid.max - cpudata->vid.min,
1830 		int_tofp(cpudata->pstate.max_pstate -
1831 			cpudata->pstate.min_pstate));
1832 
1833 	rdmsrl(MSR_ATOM_CORE_TURBO_VIDS, value);
1834 	cpudata->vid.turbo = value & 0x7f;
1835 }
1836 
1837 static int core_get_min_pstate(void)
1838 {
1839 	u64 value;
1840 
1841 	rdmsrl(MSR_PLATFORM_INFO, value);
1842 	return (value >> 40) & 0xFF;
1843 }
1844 
1845 static int core_get_max_pstate_physical(void)
1846 {
1847 	u64 value;
1848 
1849 	rdmsrl(MSR_PLATFORM_INFO, value);
1850 	return (value >> 8) & 0xFF;
1851 }
1852 
1853 static int core_get_tdp_ratio(u64 plat_info)
1854 {
1855 	/* Check how many TDP levels present */
1856 	if (plat_info & 0x600000000) {
1857 		u64 tdp_ctrl;
1858 		u64 tdp_ratio;
1859 		int tdp_msr;
1860 		int err;
1861 
1862 		/* Get the TDP level (0, 1, 2) to get ratios */
1863 		err = rdmsrl_safe(MSR_CONFIG_TDP_CONTROL, &tdp_ctrl);
1864 		if (err)
1865 			return err;
1866 
1867 		/* TDP MSR are continuous starting at 0x648 */
1868 		tdp_msr = MSR_CONFIG_TDP_NOMINAL + (tdp_ctrl & 0x03);
1869 		err = rdmsrl_safe(tdp_msr, &tdp_ratio);
1870 		if (err)
1871 			return err;
1872 
1873 		/* For level 1 and 2, bits[23:16] contain the ratio */
1874 		if (tdp_ctrl & 0x03)
1875 			tdp_ratio >>= 16;
1876 
1877 		tdp_ratio &= 0xff; /* ratios are only 8 bits long */
1878 		pr_debug("tdp_ratio %x\n", (int)tdp_ratio);
1879 
1880 		return (int)tdp_ratio;
1881 	}
1882 
1883 	return -ENXIO;
1884 }
1885 
1886 static int core_get_max_pstate(void)
1887 {
1888 	u64 tar;
1889 	u64 plat_info;
1890 	int max_pstate;
1891 	int tdp_ratio;
1892 	int err;
1893 
1894 	rdmsrl(MSR_PLATFORM_INFO, plat_info);
1895 	max_pstate = (plat_info >> 8) & 0xFF;
1896 
1897 	tdp_ratio = core_get_tdp_ratio(plat_info);
1898 	if (tdp_ratio <= 0)
1899 		return max_pstate;
1900 
1901 	if (hwp_active) {
1902 		/* Turbo activation ratio is not used on HWP platforms */
1903 		return tdp_ratio;
1904 	}
1905 
1906 	err = rdmsrl_safe(MSR_TURBO_ACTIVATION_RATIO, &tar);
1907 	if (!err) {
1908 		int tar_levels;
1909 
1910 		/* Do some sanity checking for safety */
1911 		tar_levels = tar & 0xff;
1912 		if (tdp_ratio - 1 == tar_levels) {
1913 			max_pstate = tar_levels;
1914 			pr_debug("max_pstate=TAC %x\n", max_pstate);
1915 		}
1916 	}
1917 
1918 	return max_pstate;
1919 }
1920 
1921 static int core_get_turbo_pstate(void)
1922 {
1923 	u64 value;
1924 	int nont, ret;
1925 
1926 	rdmsrl(MSR_TURBO_RATIO_LIMIT, value);
1927 	nont = core_get_max_pstate();
1928 	ret = (value) & 255;
1929 	if (ret <= nont)
1930 		ret = nont;
1931 	return ret;
1932 }
1933 
1934 static inline int core_get_scaling(void)
1935 {
1936 	return 100000;
1937 }
1938 
1939 static u64 core_get_val(struct cpudata *cpudata, int pstate)
1940 {
1941 	u64 val;
1942 
1943 	val = (u64)pstate << 8;
1944 	if (global.no_turbo && !global.turbo_disabled)
1945 		val |= (u64)1 << 32;
1946 
1947 	return val;
1948 }
1949 
1950 static int knl_get_aperf_mperf_shift(void)
1951 {
1952 	return 10;
1953 }
1954 
1955 static int knl_get_turbo_pstate(void)
1956 {
1957 	u64 value;
1958 	int nont, ret;
1959 
1960 	rdmsrl(MSR_TURBO_RATIO_LIMIT, value);
1961 	nont = core_get_max_pstate();
1962 	ret = (((value) >> 8) & 0xFF);
1963 	if (ret <= nont)
1964 		ret = nont;
1965 	return ret;
1966 }
1967 
1968 #ifdef CONFIG_ACPI_CPPC_LIB
1969 static u32 hybrid_ref_perf;
1970 
1971 static int hybrid_get_cpu_scaling(int cpu)
1972 {
1973 	return DIV_ROUND_UP(core_get_scaling() * hybrid_ref_perf,
1974 			    intel_pstate_cppc_nominal(cpu));
1975 }
1976 
1977 static void intel_pstate_cppc_set_cpu_scaling(void)
1978 {
1979 	u32 min_nominal_perf = U32_MAX;
1980 	int cpu;
1981 
1982 	for_each_present_cpu(cpu) {
1983 		u32 nominal_perf = intel_pstate_cppc_nominal(cpu);
1984 
1985 		if (nominal_perf && nominal_perf < min_nominal_perf)
1986 			min_nominal_perf = nominal_perf;
1987 	}
1988 
1989 	if (min_nominal_perf < U32_MAX) {
1990 		hybrid_ref_perf = min_nominal_perf;
1991 		pstate_funcs.get_cpu_scaling = hybrid_get_cpu_scaling;
1992 	}
1993 }
1994 #else
1995 static inline void intel_pstate_cppc_set_cpu_scaling(void)
1996 {
1997 }
1998 #endif /* CONFIG_ACPI_CPPC_LIB */
1999 
2000 static void intel_pstate_set_pstate(struct cpudata *cpu, int pstate)
2001 {
2002 	trace_cpu_frequency(pstate * cpu->pstate.scaling, cpu->cpu);
2003 	cpu->pstate.current_pstate = pstate;
2004 	/*
2005 	 * Generally, there is no guarantee that this code will always run on
2006 	 * the CPU being updated, so force the register update to run on the
2007 	 * right CPU.
2008 	 */
2009 	wrmsrl_on_cpu(cpu->cpu, MSR_IA32_PERF_CTL,
2010 		      pstate_funcs.get_val(cpu, pstate));
2011 }
2012 
2013 static void intel_pstate_set_min_pstate(struct cpudata *cpu)
2014 {
2015 	intel_pstate_set_pstate(cpu, cpu->pstate.min_pstate);
2016 }
2017 
2018 static void intel_pstate_max_within_limits(struct cpudata *cpu)
2019 {
2020 	int pstate = max(cpu->pstate.min_pstate, cpu->max_perf_ratio);
2021 
2022 	update_turbo_state();
2023 	intel_pstate_set_pstate(cpu, pstate);
2024 }
2025 
2026 static void intel_pstate_get_cpu_pstates(struct cpudata *cpu)
2027 {
2028 	int perf_ctl_max_phys = pstate_funcs.get_max_physical();
2029 	int perf_ctl_scaling = pstate_funcs.get_scaling();
2030 
2031 	cpu->pstate.min_pstate = pstate_funcs.get_min();
2032 	cpu->pstate.max_pstate_physical = perf_ctl_max_phys;
2033 	cpu->pstate.perf_ctl_scaling = perf_ctl_scaling;
2034 
2035 	if (hwp_active && !hwp_mode_bdw) {
2036 		__intel_pstate_get_hwp_cap(cpu);
2037 
2038 		if (pstate_funcs.get_cpu_scaling) {
2039 			cpu->pstate.scaling = pstate_funcs.get_cpu_scaling(cpu->cpu);
2040 			if (cpu->pstate.scaling != perf_ctl_scaling)
2041 				intel_pstate_hybrid_hwp_adjust(cpu);
2042 		} else {
2043 			cpu->pstate.scaling = perf_ctl_scaling;
2044 		}
2045 	} else {
2046 		cpu->pstate.scaling = perf_ctl_scaling;
2047 		cpu->pstate.max_pstate = pstate_funcs.get_max();
2048 		cpu->pstate.turbo_pstate = pstate_funcs.get_turbo();
2049 	}
2050 
2051 	if (cpu->pstate.scaling == perf_ctl_scaling) {
2052 		cpu->pstate.min_freq = cpu->pstate.min_pstate * perf_ctl_scaling;
2053 		cpu->pstate.max_freq = cpu->pstate.max_pstate * perf_ctl_scaling;
2054 		cpu->pstate.turbo_freq = cpu->pstate.turbo_pstate * perf_ctl_scaling;
2055 	}
2056 
2057 	if (pstate_funcs.get_aperf_mperf_shift)
2058 		cpu->aperf_mperf_shift = pstate_funcs.get_aperf_mperf_shift();
2059 
2060 	if (pstate_funcs.get_vid)
2061 		pstate_funcs.get_vid(cpu);
2062 
2063 	intel_pstate_set_min_pstate(cpu);
2064 }
2065 
2066 /*
2067  * Long hold time will keep high perf limits for long time,
2068  * which negatively impacts perf/watt for some workloads,
2069  * like specpower. 3ms is based on experiements on some
2070  * workoads.
2071  */
2072 static int hwp_boost_hold_time_ns = 3 * NSEC_PER_MSEC;
2073 
2074 static inline void intel_pstate_hwp_boost_up(struct cpudata *cpu)
2075 {
2076 	u64 hwp_req = READ_ONCE(cpu->hwp_req_cached);
2077 	u64 hwp_cap = READ_ONCE(cpu->hwp_cap_cached);
2078 	u32 max_limit = (hwp_req & 0xff00) >> 8;
2079 	u32 min_limit = (hwp_req & 0xff);
2080 	u32 boost_level1;
2081 
2082 	/*
2083 	 * Cases to consider (User changes via sysfs or boot time):
2084 	 * If, P0 (Turbo max) = P1 (Guaranteed max) = min:
2085 	 *	No boost, return.
2086 	 * If, P0 (Turbo max) > P1 (Guaranteed max) = min:
2087 	 *     Should result in one level boost only for P0.
2088 	 * If, P0 (Turbo max) = P1 (Guaranteed max) > min:
2089 	 *     Should result in two level boost:
2090 	 *         (min + p1)/2 and P1.
2091 	 * If, P0 (Turbo max) > P1 (Guaranteed max) > min:
2092 	 *     Should result in three level boost:
2093 	 *        (min + p1)/2, P1 and P0.
2094 	 */
2095 
2096 	/* If max and min are equal or already at max, nothing to boost */
2097 	if (max_limit == min_limit || cpu->hwp_boost_min >= max_limit)
2098 		return;
2099 
2100 	if (!cpu->hwp_boost_min)
2101 		cpu->hwp_boost_min = min_limit;
2102 
2103 	/* level at half way mark between min and guranteed */
2104 	boost_level1 = (HWP_GUARANTEED_PERF(hwp_cap) + min_limit) >> 1;
2105 
2106 	if (cpu->hwp_boost_min < boost_level1)
2107 		cpu->hwp_boost_min = boost_level1;
2108 	else if (cpu->hwp_boost_min < HWP_GUARANTEED_PERF(hwp_cap))
2109 		cpu->hwp_boost_min = HWP_GUARANTEED_PERF(hwp_cap);
2110 	else if (cpu->hwp_boost_min == HWP_GUARANTEED_PERF(hwp_cap) &&
2111 		 max_limit != HWP_GUARANTEED_PERF(hwp_cap))
2112 		cpu->hwp_boost_min = max_limit;
2113 	else
2114 		return;
2115 
2116 	hwp_req = (hwp_req & ~GENMASK_ULL(7, 0)) | cpu->hwp_boost_min;
2117 	wrmsrl(MSR_HWP_REQUEST, hwp_req);
2118 	cpu->last_update = cpu->sample.time;
2119 }
2120 
2121 static inline void intel_pstate_hwp_boost_down(struct cpudata *cpu)
2122 {
2123 	if (cpu->hwp_boost_min) {
2124 		bool expired;
2125 
2126 		/* Check if we are idle for hold time to boost down */
2127 		expired = time_after64(cpu->sample.time, cpu->last_update +
2128 				       hwp_boost_hold_time_ns);
2129 		if (expired) {
2130 			wrmsrl(MSR_HWP_REQUEST, cpu->hwp_req_cached);
2131 			cpu->hwp_boost_min = 0;
2132 		}
2133 	}
2134 	cpu->last_update = cpu->sample.time;
2135 }
2136 
2137 static inline void intel_pstate_update_util_hwp_local(struct cpudata *cpu,
2138 						      u64 time)
2139 {
2140 	cpu->sample.time = time;
2141 
2142 	if (cpu->sched_flags & SCHED_CPUFREQ_IOWAIT) {
2143 		bool do_io = false;
2144 
2145 		cpu->sched_flags = 0;
2146 		/*
2147 		 * Set iowait_boost flag and update time. Since IO WAIT flag
2148 		 * is set all the time, we can't just conclude that there is
2149 		 * some IO bound activity is scheduled on this CPU with just
2150 		 * one occurrence. If we receive at least two in two
2151 		 * consecutive ticks, then we treat as boost candidate.
2152 		 */
2153 		if (time_before64(time, cpu->last_io_update + 2 * TICK_NSEC))
2154 			do_io = true;
2155 
2156 		cpu->last_io_update = time;
2157 
2158 		if (do_io)
2159 			intel_pstate_hwp_boost_up(cpu);
2160 
2161 	} else {
2162 		intel_pstate_hwp_boost_down(cpu);
2163 	}
2164 }
2165 
2166 static inline void intel_pstate_update_util_hwp(struct update_util_data *data,
2167 						u64 time, unsigned int flags)
2168 {
2169 	struct cpudata *cpu = container_of(data, struct cpudata, update_util);
2170 
2171 	cpu->sched_flags |= flags;
2172 
2173 	if (smp_processor_id() == cpu->cpu)
2174 		intel_pstate_update_util_hwp_local(cpu, time);
2175 }
2176 
2177 static inline void intel_pstate_calc_avg_perf(struct cpudata *cpu)
2178 {
2179 	struct sample *sample = &cpu->sample;
2180 
2181 	sample->core_avg_perf = div_ext_fp(sample->aperf, sample->mperf);
2182 }
2183 
2184 static inline bool intel_pstate_sample(struct cpudata *cpu, u64 time)
2185 {
2186 	u64 aperf, mperf;
2187 	unsigned long flags;
2188 	u64 tsc;
2189 
2190 	local_irq_save(flags);
2191 	rdmsrl(MSR_IA32_APERF, aperf);
2192 	rdmsrl(MSR_IA32_MPERF, mperf);
2193 	tsc = rdtsc();
2194 	if (cpu->prev_mperf == mperf || cpu->prev_tsc == tsc) {
2195 		local_irq_restore(flags);
2196 		return false;
2197 	}
2198 	local_irq_restore(flags);
2199 
2200 	cpu->last_sample_time = cpu->sample.time;
2201 	cpu->sample.time = time;
2202 	cpu->sample.aperf = aperf;
2203 	cpu->sample.mperf = mperf;
2204 	cpu->sample.tsc =  tsc;
2205 	cpu->sample.aperf -= cpu->prev_aperf;
2206 	cpu->sample.mperf -= cpu->prev_mperf;
2207 	cpu->sample.tsc -= cpu->prev_tsc;
2208 
2209 	cpu->prev_aperf = aperf;
2210 	cpu->prev_mperf = mperf;
2211 	cpu->prev_tsc = tsc;
2212 	/*
2213 	 * First time this function is invoked in a given cycle, all of the
2214 	 * previous sample data fields are equal to zero or stale and they must
2215 	 * be populated with meaningful numbers for things to work, so assume
2216 	 * that sample.time will always be reset before setting the utilization
2217 	 * update hook and make the caller skip the sample then.
2218 	 */
2219 	if (cpu->last_sample_time) {
2220 		intel_pstate_calc_avg_perf(cpu);
2221 		return true;
2222 	}
2223 	return false;
2224 }
2225 
2226 static inline int32_t get_avg_frequency(struct cpudata *cpu)
2227 {
2228 	return mul_ext_fp(cpu->sample.core_avg_perf, cpu_khz);
2229 }
2230 
2231 static inline int32_t get_avg_pstate(struct cpudata *cpu)
2232 {
2233 	return mul_ext_fp(cpu->pstate.max_pstate_physical,
2234 			  cpu->sample.core_avg_perf);
2235 }
2236 
2237 static inline int32_t get_target_pstate(struct cpudata *cpu)
2238 {
2239 	struct sample *sample = &cpu->sample;
2240 	int32_t busy_frac;
2241 	int target, avg_pstate;
2242 
2243 	busy_frac = div_fp(sample->mperf << cpu->aperf_mperf_shift,
2244 			   sample->tsc);
2245 
2246 	if (busy_frac < cpu->iowait_boost)
2247 		busy_frac = cpu->iowait_boost;
2248 
2249 	sample->busy_scaled = busy_frac * 100;
2250 
2251 	target = global.no_turbo || global.turbo_disabled ?
2252 			cpu->pstate.max_pstate : cpu->pstate.turbo_pstate;
2253 	target += target >> 2;
2254 	target = mul_fp(target, busy_frac);
2255 	if (target < cpu->pstate.min_pstate)
2256 		target = cpu->pstate.min_pstate;
2257 
2258 	/*
2259 	 * If the average P-state during the previous cycle was higher than the
2260 	 * current target, add 50% of the difference to the target to reduce
2261 	 * possible performance oscillations and offset possible performance
2262 	 * loss related to moving the workload from one CPU to another within
2263 	 * a package/module.
2264 	 */
2265 	avg_pstate = get_avg_pstate(cpu);
2266 	if (avg_pstate > target)
2267 		target += (avg_pstate - target) >> 1;
2268 
2269 	return target;
2270 }
2271 
2272 static int intel_pstate_prepare_request(struct cpudata *cpu, int pstate)
2273 {
2274 	int min_pstate = max(cpu->pstate.min_pstate, cpu->min_perf_ratio);
2275 	int max_pstate = max(min_pstate, cpu->max_perf_ratio);
2276 
2277 	return clamp_t(int, pstate, min_pstate, max_pstate);
2278 }
2279 
2280 static void intel_pstate_update_pstate(struct cpudata *cpu, int pstate)
2281 {
2282 	if (pstate == cpu->pstate.current_pstate)
2283 		return;
2284 
2285 	cpu->pstate.current_pstate = pstate;
2286 	wrmsrl(MSR_IA32_PERF_CTL, pstate_funcs.get_val(cpu, pstate));
2287 }
2288 
2289 static void intel_pstate_adjust_pstate(struct cpudata *cpu)
2290 {
2291 	int from = cpu->pstate.current_pstate;
2292 	struct sample *sample;
2293 	int target_pstate;
2294 
2295 	update_turbo_state();
2296 
2297 	target_pstate = get_target_pstate(cpu);
2298 	target_pstate = intel_pstate_prepare_request(cpu, target_pstate);
2299 	trace_cpu_frequency(target_pstate * cpu->pstate.scaling, cpu->cpu);
2300 	intel_pstate_update_pstate(cpu, target_pstate);
2301 
2302 	sample = &cpu->sample;
2303 	trace_pstate_sample(mul_ext_fp(100, sample->core_avg_perf),
2304 		fp_toint(sample->busy_scaled),
2305 		from,
2306 		cpu->pstate.current_pstate,
2307 		sample->mperf,
2308 		sample->aperf,
2309 		sample->tsc,
2310 		get_avg_frequency(cpu),
2311 		fp_toint(cpu->iowait_boost * 100));
2312 }
2313 
2314 static void intel_pstate_update_util(struct update_util_data *data, u64 time,
2315 				     unsigned int flags)
2316 {
2317 	struct cpudata *cpu = container_of(data, struct cpudata, update_util);
2318 	u64 delta_ns;
2319 
2320 	/* Don't allow remote callbacks */
2321 	if (smp_processor_id() != cpu->cpu)
2322 		return;
2323 
2324 	delta_ns = time - cpu->last_update;
2325 	if (flags & SCHED_CPUFREQ_IOWAIT) {
2326 		/* Start over if the CPU may have been idle. */
2327 		if (delta_ns > TICK_NSEC) {
2328 			cpu->iowait_boost = ONE_EIGHTH_FP;
2329 		} else if (cpu->iowait_boost >= ONE_EIGHTH_FP) {
2330 			cpu->iowait_boost <<= 1;
2331 			if (cpu->iowait_boost > int_tofp(1))
2332 				cpu->iowait_boost = int_tofp(1);
2333 		} else {
2334 			cpu->iowait_boost = ONE_EIGHTH_FP;
2335 		}
2336 	} else if (cpu->iowait_boost) {
2337 		/* Clear iowait_boost if the CPU may have been idle. */
2338 		if (delta_ns > TICK_NSEC)
2339 			cpu->iowait_boost = 0;
2340 		else
2341 			cpu->iowait_boost >>= 1;
2342 	}
2343 	cpu->last_update = time;
2344 	delta_ns = time - cpu->sample.time;
2345 	if ((s64)delta_ns < INTEL_PSTATE_SAMPLING_INTERVAL)
2346 		return;
2347 
2348 	if (intel_pstate_sample(cpu, time))
2349 		intel_pstate_adjust_pstate(cpu);
2350 }
2351 
2352 static struct pstate_funcs core_funcs = {
2353 	.get_max = core_get_max_pstate,
2354 	.get_max_physical = core_get_max_pstate_physical,
2355 	.get_min = core_get_min_pstate,
2356 	.get_turbo = core_get_turbo_pstate,
2357 	.get_scaling = core_get_scaling,
2358 	.get_val = core_get_val,
2359 };
2360 
2361 static const struct pstate_funcs silvermont_funcs = {
2362 	.get_max = atom_get_max_pstate,
2363 	.get_max_physical = atom_get_max_pstate,
2364 	.get_min = atom_get_min_pstate,
2365 	.get_turbo = atom_get_turbo_pstate,
2366 	.get_val = atom_get_val,
2367 	.get_scaling = silvermont_get_scaling,
2368 	.get_vid = atom_get_vid,
2369 };
2370 
2371 static const struct pstate_funcs airmont_funcs = {
2372 	.get_max = atom_get_max_pstate,
2373 	.get_max_physical = atom_get_max_pstate,
2374 	.get_min = atom_get_min_pstate,
2375 	.get_turbo = atom_get_turbo_pstate,
2376 	.get_val = atom_get_val,
2377 	.get_scaling = airmont_get_scaling,
2378 	.get_vid = atom_get_vid,
2379 };
2380 
2381 static const struct pstate_funcs knl_funcs = {
2382 	.get_max = core_get_max_pstate,
2383 	.get_max_physical = core_get_max_pstate_physical,
2384 	.get_min = core_get_min_pstate,
2385 	.get_turbo = knl_get_turbo_pstate,
2386 	.get_aperf_mperf_shift = knl_get_aperf_mperf_shift,
2387 	.get_scaling = core_get_scaling,
2388 	.get_val = core_get_val,
2389 };
2390 
2391 #define X86_MATCH(model, policy)					 \
2392 	X86_MATCH_VENDOR_FAM_MODEL_FEATURE(INTEL, 6, INTEL_FAM6_##model, \
2393 					   X86_FEATURE_APERFMPERF, &policy)
2394 
2395 static const struct x86_cpu_id intel_pstate_cpu_ids[] = {
2396 	X86_MATCH(SANDYBRIDGE,		core_funcs),
2397 	X86_MATCH(SANDYBRIDGE_X,	core_funcs),
2398 	X86_MATCH(ATOM_SILVERMONT,	silvermont_funcs),
2399 	X86_MATCH(IVYBRIDGE,		core_funcs),
2400 	X86_MATCH(HASWELL,		core_funcs),
2401 	X86_MATCH(BROADWELL,		core_funcs),
2402 	X86_MATCH(IVYBRIDGE_X,		core_funcs),
2403 	X86_MATCH(HASWELL_X,		core_funcs),
2404 	X86_MATCH(HASWELL_L,		core_funcs),
2405 	X86_MATCH(HASWELL_G,		core_funcs),
2406 	X86_MATCH(BROADWELL_G,		core_funcs),
2407 	X86_MATCH(ATOM_AIRMONT,		airmont_funcs),
2408 	X86_MATCH(SKYLAKE_L,		core_funcs),
2409 	X86_MATCH(BROADWELL_X,		core_funcs),
2410 	X86_MATCH(SKYLAKE,		core_funcs),
2411 	X86_MATCH(BROADWELL_D,		core_funcs),
2412 	X86_MATCH(XEON_PHI_KNL,		knl_funcs),
2413 	X86_MATCH(XEON_PHI_KNM,		knl_funcs),
2414 	X86_MATCH(ATOM_GOLDMONT,	core_funcs),
2415 	X86_MATCH(ATOM_GOLDMONT_PLUS,	core_funcs),
2416 	X86_MATCH(SKYLAKE_X,		core_funcs),
2417 	X86_MATCH(COMETLAKE,		core_funcs),
2418 	X86_MATCH(ICELAKE_X,		core_funcs),
2419 	{}
2420 };
2421 MODULE_DEVICE_TABLE(x86cpu, intel_pstate_cpu_ids);
2422 
2423 static const struct x86_cpu_id intel_pstate_cpu_oob_ids[] __initconst = {
2424 	X86_MATCH(BROADWELL_D,		core_funcs),
2425 	X86_MATCH(BROADWELL_X,		core_funcs),
2426 	X86_MATCH(SKYLAKE_X,		core_funcs),
2427 	X86_MATCH(ICELAKE_X,		core_funcs),
2428 	X86_MATCH(SAPPHIRERAPIDS_X,	core_funcs),
2429 	{}
2430 };
2431 
2432 static const struct x86_cpu_id intel_pstate_cpu_ee_disable_ids[] = {
2433 	X86_MATCH(KABYLAKE,		core_funcs),
2434 	{}
2435 };
2436 
2437 static const struct x86_cpu_id intel_pstate_hwp_boost_ids[] = {
2438 	X86_MATCH(SKYLAKE_X,		core_funcs),
2439 	X86_MATCH(SKYLAKE,		core_funcs),
2440 	{}
2441 };
2442 
2443 static int intel_pstate_init_cpu(unsigned int cpunum)
2444 {
2445 	struct cpudata *cpu;
2446 
2447 	cpu = all_cpu_data[cpunum];
2448 
2449 	if (!cpu) {
2450 		cpu = kzalloc(sizeof(*cpu), GFP_KERNEL);
2451 		if (!cpu)
2452 			return -ENOMEM;
2453 
2454 		WRITE_ONCE(all_cpu_data[cpunum], cpu);
2455 
2456 		cpu->cpu = cpunum;
2457 
2458 		cpu->epp_default = -EINVAL;
2459 
2460 		if (hwp_active) {
2461 			const struct x86_cpu_id *id;
2462 
2463 			intel_pstate_hwp_enable(cpu);
2464 
2465 			id = x86_match_cpu(intel_pstate_hwp_boost_ids);
2466 			if (id && intel_pstate_acpi_pm_profile_server())
2467 				hwp_boost = true;
2468 		}
2469 	} else if (hwp_active) {
2470 		/*
2471 		 * Re-enable HWP in case this happens after a resume from ACPI
2472 		 * S3 if the CPU was offline during the whole system/resume
2473 		 * cycle.
2474 		 */
2475 		intel_pstate_hwp_reenable(cpu);
2476 	}
2477 
2478 	cpu->epp_powersave = -EINVAL;
2479 	cpu->epp_policy = 0;
2480 
2481 	intel_pstate_get_cpu_pstates(cpu);
2482 
2483 	pr_debug("controlling: cpu %d\n", cpunum);
2484 
2485 	return 0;
2486 }
2487 
2488 static void intel_pstate_set_update_util_hook(unsigned int cpu_num)
2489 {
2490 	struct cpudata *cpu = all_cpu_data[cpu_num];
2491 
2492 	if (hwp_active && !hwp_boost)
2493 		return;
2494 
2495 	if (cpu->update_util_set)
2496 		return;
2497 
2498 	/* Prevent intel_pstate_update_util() from using stale data. */
2499 	cpu->sample.time = 0;
2500 	cpufreq_add_update_util_hook(cpu_num, &cpu->update_util,
2501 				     (hwp_active ?
2502 				      intel_pstate_update_util_hwp :
2503 				      intel_pstate_update_util));
2504 	cpu->update_util_set = true;
2505 }
2506 
2507 static void intel_pstate_clear_update_util_hook(unsigned int cpu)
2508 {
2509 	struct cpudata *cpu_data = all_cpu_data[cpu];
2510 
2511 	if (!cpu_data->update_util_set)
2512 		return;
2513 
2514 	cpufreq_remove_update_util_hook(cpu);
2515 	cpu_data->update_util_set = false;
2516 	synchronize_rcu();
2517 }
2518 
2519 static int intel_pstate_get_max_freq(struct cpudata *cpu)
2520 {
2521 	return global.turbo_disabled || global.no_turbo ?
2522 			cpu->pstate.max_freq : cpu->pstate.turbo_freq;
2523 }
2524 
2525 static void intel_pstate_update_perf_limits(struct cpudata *cpu,
2526 					    unsigned int policy_min,
2527 					    unsigned int policy_max)
2528 {
2529 	int perf_ctl_scaling = cpu->pstate.perf_ctl_scaling;
2530 	int32_t max_policy_perf, min_policy_perf;
2531 
2532 	max_policy_perf = policy_max / perf_ctl_scaling;
2533 	if (policy_max == policy_min) {
2534 		min_policy_perf = max_policy_perf;
2535 	} else {
2536 		min_policy_perf = policy_min / perf_ctl_scaling;
2537 		min_policy_perf = clamp_t(int32_t, min_policy_perf,
2538 					  0, max_policy_perf);
2539 	}
2540 
2541 	/*
2542 	 * HWP needs some special consideration, because HWP_REQUEST uses
2543 	 * abstract values to represent performance rather than pure ratios.
2544 	 */
2545 	if (hwp_active && cpu->pstate.scaling != perf_ctl_scaling) {
2546 		int scaling = cpu->pstate.scaling;
2547 		int freq;
2548 
2549 		freq = max_policy_perf * perf_ctl_scaling;
2550 		max_policy_perf = DIV_ROUND_UP(freq, scaling);
2551 		freq = min_policy_perf * perf_ctl_scaling;
2552 		min_policy_perf = DIV_ROUND_UP(freq, scaling);
2553 	}
2554 
2555 	pr_debug("cpu:%d min_policy_perf:%d max_policy_perf:%d\n",
2556 		 cpu->cpu, min_policy_perf, max_policy_perf);
2557 
2558 	/* Normalize user input to [min_perf, max_perf] */
2559 	if (per_cpu_limits) {
2560 		cpu->min_perf_ratio = min_policy_perf;
2561 		cpu->max_perf_ratio = max_policy_perf;
2562 	} else {
2563 		int turbo_max = cpu->pstate.turbo_pstate;
2564 		int32_t global_min, global_max;
2565 
2566 		/* Global limits are in percent of the maximum turbo P-state. */
2567 		global_max = DIV_ROUND_UP(turbo_max * global.max_perf_pct, 100);
2568 		global_min = DIV_ROUND_UP(turbo_max * global.min_perf_pct, 100);
2569 		global_min = clamp_t(int32_t, global_min, 0, global_max);
2570 
2571 		pr_debug("cpu:%d global_min:%d global_max:%d\n", cpu->cpu,
2572 			 global_min, global_max);
2573 
2574 		cpu->min_perf_ratio = max(min_policy_perf, global_min);
2575 		cpu->min_perf_ratio = min(cpu->min_perf_ratio, max_policy_perf);
2576 		cpu->max_perf_ratio = min(max_policy_perf, global_max);
2577 		cpu->max_perf_ratio = max(min_policy_perf, cpu->max_perf_ratio);
2578 
2579 		/* Make sure min_perf <= max_perf */
2580 		cpu->min_perf_ratio = min(cpu->min_perf_ratio,
2581 					  cpu->max_perf_ratio);
2582 
2583 	}
2584 	pr_debug("cpu:%d max_perf_ratio:%d min_perf_ratio:%d\n", cpu->cpu,
2585 		 cpu->max_perf_ratio,
2586 		 cpu->min_perf_ratio);
2587 }
2588 
2589 static int intel_pstate_set_policy(struct cpufreq_policy *policy)
2590 {
2591 	struct cpudata *cpu;
2592 
2593 	if (!policy->cpuinfo.max_freq)
2594 		return -ENODEV;
2595 
2596 	pr_debug("set_policy cpuinfo.max %u policy->max %u\n",
2597 		 policy->cpuinfo.max_freq, policy->max);
2598 
2599 	cpu = all_cpu_data[policy->cpu];
2600 	cpu->policy = policy->policy;
2601 
2602 	mutex_lock(&intel_pstate_limits_lock);
2603 
2604 	intel_pstate_update_perf_limits(cpu, policy->min, policy->max);
2605 
2606 	if (cpu->policy == CPUFREQ_POLICY_PERFORMANCE) {
2607 		/*
2608 		 * NOHZ_FULL CPUs need this as the governor callback may not
2609 		 * be invoked on them.
2610 		 */
2611 		intel_pstate_clear_update_util_hook(policy->cpu);
2612 		intel_pstate_max_within_limits(cpu);
2613 	} else {
2614 		intel_pstate_set_update_util_hook(policy->cpu);
2615 	}
2616 
2617 	if (hwp_active) {
2618 		/*
2619 		 * When hwp_boost was active before and dynamically it
2620 		 * was turned off, in that case we need to clear the
2621 		 * update util hook.
2622 		 */
2623 		if (!hwp_boost)
2624 			intel_pstate_clear_update_util_hook(policy->cpu);
2625 		intel_pstate_hwp_set(policy->cpu);
2626 	}
2627 
2628 	mutex_unlock(&intel_pstate_limits_lock);
2629 
2630 	return 0;
2631 }
2632 
2633 static void intel_pstate_adjust_policy_max(struct cpudata *cpu,
2634 					   struct cpufreq_policy_data *policy)
2635 {
2636 	if (!hwp_active &&
2637 	    cpu->pstate.max_pstate_physical > cpu->pstate.max_pstate &&
2638 	    policy->max < policy->cpuinfo.max_freq &&
2639 	    policy->max > cpu->pstate.max_freq) {
2640 		pr_debug("policy->max > max non turbo frequency\n");
2641 		policy->max = policy->cpuinfo.max_freq;
2642 	}
2643 }
2644 
2645 static void intel_pstate_verify_cpu_policy(struct cpudata *cpu,
2646 					   struct cpufreq_policy_data *policy)
2647 {
2648 	int max_freq;
2649 
2650 	update_turbo_state();
2651 	if (hwp_active) {
2652 		intel_pstate_get_hwp_cap(cpu);
2653 		max_freq = global.no_turbo || global.turbo_disabled ?
2654 				cpu->pstate.max_freq : cpu->pstate.turbo_freq;
2655 	} else {
2656 		max_freq = intel_pstate_get_max_freq(cpu);
2657 	}
2658 	cpufreq_verify_within_limits(policy, policy->cpuinfo.min_freq, max_freq);
2659 
2660 	intel_pstate_adjust_policy_max(cpu, policy);
2661 }
2662 
2663 static int intel_pstate_verify_policy(struct cpufreq_policy_data *policy)
2664 {
2665 	intel_pstate_verify_cpu_policy(all_cpu_data[policy->cpu], policy);
2666 
2667 	return 0;
2668 }
2669 
2670 static int intel_cpufreq_cpu_offline(struct cpufreq_policy *policy)
2671 {
2672 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2673 
2674 	pr_debug("CPU %d going offline\n", cpu->cpu);
2675 
2676 	if (cpu->suspended)
2677 		return 0;
2678 
2679 	/*
2680 	 * If the CPU is an SMT thread and it goes offline with the performance
2681 	 * settings different from the minimum, it will prevent its sibling
2682 	 * from getting to lower performance levels, so force the minimum
2683 	 * performance on CPU offline to prevent that from happening.
2684 	 */
2685 	if (hwp_active)
2686 		intel_pstate_hwp_offline(cpu);
2687 	else
2688 		intel_pstate_set_min_pstate(cpu);
2689 
2690 	intel_pstate_exit_perf_limits(policy);
2691 
2692 	return 0;
2693 }
2694 
2695 static int intel_pstate_cpu_online(struct cpufreq_policy *policy)
2696 {
2697 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2698 
2699 	pr_debug("CPU %d going online\n", cpu->cpu);
2700 
2701 	intel_pstate_init_acpi_perf_limits(policy);
2702 
2703 	if (hwp_active) {
2704 		/*
2705 		 * Re-enable HWP and clear the "suspended" flag to let "resume"
2706 		 * know that it need not do that.
2707 		 */
2708 		intel_pstate_hwp_reenable(cpu);
2709 		cpu->suspended = false;
2710 	}
2711 
2712 	return 0;
2713 }
2714 
2715 static int intel_pstate_cpu_offline(struct cpufreq_policy *policy)
2716 {
2717 	intel_pstate_clear_update_util_hook(policy->cpu);
2718 
2719 	return intel_cpufreq_cpu_offline(policy);
2720 }
2721 
2722 static int intel_pstate_cpu_exit(struct cpufreq_policy *policy)
2723 {
2724 	pr_debug("CPU %d exiting\n", policy->cpu);
2725 
2726 	policy->fast_switch_possible = false;
2727 
2728 	return 0;
2729 }
2730 
2731 static int __intel_pstate_cpu_init(struct cpufreq_policy *policy)
2732 {
2733 	struct cpudata *cpu;
2734 	int rc;
2735 
2736 	rc = intel_pstate_init_cpu(policy->cpu);
2737 	if (rc)
2738 		return rc;
2739 
2740 	cpu = all_cpu_data[policy->cpu];
2741 
2742 	cpu->max_perf_ratio = 0xFF;
2743 	cpu->min_perf_ratio = 0;
2744 
2745 	/* cpuinfo and default policy values */
2746 	policy->cpuinfo.min_freq = cpu->pstate.min_freq;
2747 	update_turbo_state();
2748 	global.turbo_disabled_mf = global.turbo_disabled;
2749 	policy->cpuinfo.max_freq = global.turbo_disabled ?
2750 			cpu->pstate.max_freq : cpu->pstate.turbo_freq;
2751 
2752 	policy->min = policy->cpuinfo.min_freq;
2753 	policy->max = policy->cpuinfo.max_freq;
2754 
2755 	intel_pstate_init_acpi_perf_limits(policy);
2756 
2757 	policy->fast_switch_possible = true;
2758 
2759 	return 0;
2760 }
2761 
2762 static int intel_pstate_cpu_init(struct cpufreq_policy *policy)
2763 {
2764 	int ret = __intel_pstate_cpu_init(policy);
2765 
2766 	if (ret)
2767 		return ret;
2768 
2769 	/*
2770 	 * Set the policy to powersave to provide a valid fallback value in case
2771 	 * the default cpufreq governor is neither powersave nor performance.
2772 	 */
2773 	policy->policy = CPUFREQ_POLICY_POWERSAVE;
2774 
2775 	if (hwp_active) {
2776 		struct cpudata *cpu = all_cpu_data[policy->cpu];
2777 
2778 		cpu->epp_cached = intel_pstate_get_epp(cpu, 0);
2779 	}
2780 
2781 	return 0;
2782 }
2783 
2784 static struct cpufreq_driver intel_pstate = {
2785 	.flags		= CPUFREQ_CONST_LOOPS,
2786 	.verify		= intel_pstate_verify_policy,
2787 	.setpolicy	= intel_pstate_set_policy,
2788 	.suspend	= intel_pstate_suspend,
2789 	.resume		= intel_pstate_resume,
2790 	.init		= intel_pstate_cpu_init,
2791 	.exit		= intel_pstate_cpu_exit,
2792 	.offline	= intel_pstate_cpu_offline,
2793 	.online		= intel_pstate_cpu_online,
2794 	.update_limits	= intel_pstate_update_limits,
2795 	.name		= "intel_pstate",
2796 };
2797 
2798 static int intel_cpufreq_verify_policy(struct cpufreq_policy_data *policy)
2799 {
2800 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2801 
2802 	intel_pstate_verify_cpu_policy(cpu, policy);
2803 	intel_pstate_update_perf_limits(cpu, policy->min, policy->max);
2804 
2805 	return 0;
2806 }
2807 
2808 /* Use of trace in passive mode:
2809  *
2810  * In passive mode the trace core_busy field (also known as the
2811  * performance field, and lablelled as such on the graphs; also known as
2812  * core_avg_perf) is not needed and so is re-assigned to indicate if the
2813  * driver call was via the normal or fast switch path. Various graphs
2814  * output from the intel_pstate_tracer.py utility that include core_busy
2815  * (or performance or core_avg_perf) have a fixed y-axis from 0 to 100%,
2816  * so we use 10 to indicate the normal path through the driver, and
2817  * 90 to indicate the fast switch path through the driver.
2818  * The scaled_busy field is not used, and is set to 0.
2819  */
2820 
2821 #define	INTEL_PSTATE_TRACE_TARGET 10
2822 #define	INTEL_PSTATE_TRACE_FAST_SWITCH 90
2823 
2824 static void intel_cpufreq_trace(struct cpudata *cpu, unsigned int trace_type, int old_pstate)
2825 {
2826 	struct sample *sample;
2827 
2828 	if (!trace_pstate_sample_enabled())
2829 		return;
2830 
2831 	if (!intel_pstate_sample(cpu, ktime_get()))
2832 		return;
2833 
2834 	sample = &cpu->sample;
2835 	trace_pstate_sample(trace_type,
2836 		0,
2837 		old_pstate,
2838 		cpu->pstate.current_pstate,
2839 		sample->mperf,
2840 		sample->aperf,
2841 		sample->tsc,
2842 		get_avg_frequency(cpu),
2843 		fp_toint(cpu->iowait_boost * 100));
2844 }
2845 
2846 static void intel_cpufreq_hwp_update(struct cpudata *cpu, u32 min, u32 max,
2847 				     u32 desired, bool fast_switch)
2848 {
2849 	u64 prev = READ_ONCE(cpu->hwp_req_cached), value = prev;
2850 
2851 	value &= ~HWP_MIN_PERF(~0L);
2852 	value |= HWP_MIN_PERF(min);
2853 
2854 	value &= ~HWP_MAX_PERF(~0L);
2855 	value |= HWP_MAX_PERF(max);
2856 
2857 	value &= ~HWP_DESIRED_PERF(~0L);
2858 	value |= HWP_DESIRED_PERF(desired);
2859 
2860 	if (value == prev)
2861 		return;
2862 
2863 	WRITE_ONCE(cpu->hwp_req_cached, value);
2864 	if (fast_switch)
2865 		wrmsrl(MSR_HWP_REQUEST, value);
2866 	else
2867 		wrmsrl_on_cpu(cpu->cpu, MSR_HWP_REQUEST, value);
2868 }
2869 
2870 static void intel_cpufreq_perf_ctl_update(struct cpudata *cpu,
2871 					  u32 target_pstate, bool fast_switch)
2872 {
2873 	if (fast_switch)
2874 		wrmsrl(MSR_IA32_PERF_CTL,
2875 		       pstate_funcs.get_val(cpu, target_pstate));
2876 	else
2877 		wrmsrl_on_cpu(cpu->cpu, MSR_IA32_PERF_CTL,
2878 			      pstate_funcs.get_val(cpu, target_pstate));
2879 }
2880 
2881 static int intel_cpufreq_update_pstate(struct cpufreq_policy *policy,
2882 				       int target_pstate, bool fast_switch)
2883 {
2884 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2885 	int old_pstate = cpu->pstate.current_pstate;
2886 
2887 	target_pstate = intel_pstate_prepare_request(cpu, target_pstate);
2888 	if (hwp_active) {
2889 		int max_pstate = policy->strict_target ?
2890 					target_pstate : cpu->max_perf_ratio;
2891 
2892 		intel_cpufreq_hwp_update(cpu, target_pstate, max_pstate, 0,
2893 					 fast_switch);
2894 	} else if (target_pstate != old_pstate) {
2895 		intel_cpufreq_perf_ctl_update(cpu, target_pstate, fast_switch);
2896 	}
2897 
2898 	cpu->pstate.current_pstate = target_pstate;
2899 
2900 	intel_cpufreq_trace(cpu, fast_switch ? INTEL_PSTATE_TRACE_FAST_SWITCH :
2901 			    INTEL_PSTATE_TRACE_TARGET, old_pstate);
2902 
2903 	return target_pstate;
2904 }
2905 
2906 static int intel_cpufreq_target(struct cpufreq_policy *policy,
2907 				unsigned int target_freq,
2908 				unsigned int relation)
2909 {
2910 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2911 	struct cpufreq_freqs freqs;
2912 	int target_pstate;
2913 
2914 	update_turbo_state();
2915 
2916 	freqs.old = policy->cur;
2917 	freqs.new = target_freq;
2918 
2919 	cpufreq_freq_transition_begin(policy, &freqs);
2920 
2921 	switch (relation) {
2922 	case CPUFREQ_RELATION_L:
2923 		target_pstate = DIV_ROUND_UP(freqs.new, cpu->pstate.scaling);
2924 		break;
2925 	case CPUFREQ_RELATION_H:
2926 		target_pstate = freqs.new / cpu->pstate.scaling;
2927 		break;
2928 	default:
2929 		target_pstate = DIV_ROUND_CLOSEST(freqs.new, cpu->pstate.scaling);
2930 		break;
2931 	}
2932 
2933 	target_pstate = intel_cpufreq_update_pstate(policy, target_pstate, false);
2934 
2935 	freqs.new = target_pstate * cpu->pstate.scaling;
2936 
2937 	cpufreq_freq_transition_end(policy, &freqs, false);
2938 
2939 	return 0;
2940 }
2941 
2942 static unsigned int intel_cpufreq_fast_switch(struct cpufreq_policy *policy,
2943 					      unsigned int target_freq)
2944 {
2945 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2946 	int target_pstate;
2947 
2948 	update_turbo_state();
2949 
2950 	target_pstate = DIV_ROUND_UP(target_freq, cpu->pstate.scaling);
2951 
2952 	target_pstate = intel_cpufreq_update_pstate(policy, target_pstate, true);
2953 
2954 	return target_pstate * cpu->pstate.scaling;
2955 }
2956 
2957 static void intel_cpufreq_adjust_perf(unsigned int cpunum,
2958 				      unsigned long min_perf,
2959 				      unsigned long target_perf,
2960 				      unsigned long capacity)
2961 {
2962 	struct cpudata *cpu = all_cpu_data[cpunum];
2963 	u64 hwp_cap = READ_ONCE(cpu->hwp_cap_cached);
2964 	int old_pstate = cpu->pstate.current_pstate;
2965 	int cap_pstate, min_pstate, max_pstate, target_pstate;
2966 
2967 	update_turbo_state();
2968 	cap_pstate = global.turbo_disabled ? HWP_GUARANTEED_PERF(hwp_cap) :
2969 					     HWP_HIGHEST_PERF(hwp_cap);
2970 
2971 	/* Optimization: Avoid unnecessary divisions. */
2972 
2973 	target_pstate = cap_pstate;
2974 	if (target_perf < capacity)
2975 		target_pstate = DIV_ROUND_UP(cap_pstate * target_perf, capacity);
2976 
2977 	min_pstate = cap_pstate;
2978 	if (min_perf < capacity)
2979 		min_pstate = DIV_ROUND_UP(cap_pstate * min_perf, capacity);
2980 
2981 	if (min_pstate < cpu->pstate.min_pstate)
2982 		min_pstate = cpu->pstate.min_pstate;
2983 
2984 	if (min_pstate < cpu->min_perf_ratio)
2985 		min_pstate = cpu->min_perf_ratio;
2986 
2987 	max_pstate = min(cap_pstate, cpu->max_perf_ratio);
2988 	if (max_pstate < min_pstate)
2989 		max_pstate = min_pstate;
2990 
2991 	target_pstate = clamp_t(int, target_pstate, min_pstate, max_pstate);
2992 
2993 	intel_cpufreq_hwp_update(cpu, min_pstate, max_pstate, target_pstate, true);
2994 
2995 	cpu->pstate.current_pstate = target_pstate;
2996 	intel_cpufreq_trace(cpu, INTEL_PSTATE_TRACE_FAST_SWITCH, old_pstate);
2997 }
2998 
2999 static int intel_cpufreq_cpu_init(struct cpufreq_policy *policy)
3000 {
3001 	struct freq_qos_request *req;
3002 	struct cpudata *cpu;
3003 	struct device *dev;
3004 	int ret, freq;
3005 
3006 	dev = get_cpu_device(policy->cpu);
3007 	if (!dev)
3008 		return -ENODEV;
3009 
3010 	ret = __intel_pstate_cpu_init(policy);
3011 	if (ret)
3012 		return ret;
3013 
3014 	policy->cpuinfo.transition_latency = INTEL_CPUFREQ_TRANSITION_LATENCY;
3015 	/* This reflects the intel_pstate_get_cpu_pstates() setting. */
3016 	policy->cur = policy->cpuinfo.min_freq;
3017 
3018 	req = kcalloc(2, sizeof(*req), GFP_KERNEL);
3019 	if (!req) {
3020 		ret = -ENOMEM;
3021 		goto pstate_exit;
3022 	}
3023 
3024 	cpu = all_cpu_data[policy->cpu];
3025 
3026 	if (hwp_active) {
3027 		u64 value;
3028 
3029 		policy->transition_delay_us = INTEL_CPUFREQ_TRANSITION_DELAY_HWP;
3030 
3031 		intel_pstate_get_hwp_cap(cpu);
3032 
3033 		rdmsrl_on_cpu(cpu->cpu, MSR_HWP_REQUEST, &value);
3034 		WRITE_ONCE(cpu->hwp_req_cached, value);
3035 
3036 		cpu->epp_cached = intel_pstate_get_epp(cpu, value);
3037 	} else {
3038 		policy->transition_delay_us = INTEL_CPUFREQ_TRANSITION_DELAY;
3039 	}
3040 
3041 	freq = DIV_ROUND_UP(cpu->pstate.turbo_freq * global.min_perf_pct, 100);
3042 
3043 	ret = freq_qos_add_request(&policy->constraints, req, FREQ_QOS_MIN,
3044 				   freq);
3045 	if (ret < 0) {
3046 		dev_err(dev, "Failed to add min-freq constraint (%d)\n", ret);
3047 		goto free_req;
3048 	}
3049 
3050 	freq = DIV_ROUND_UP(cpu->pstate.turbo_freq * global.max_perf_pct, 100);
3051 
3052 	ret = freq_qos_add_request(&policy->constraints, req + 1, FREQ_QOS_MAX,
3053 				   freq);
3054 	if (ret < 0) {
3055 		dev_err(dev, "Failed to add max-freq constraint (%d)\n", ret);
3056 		goto remove_min_req;
3057 	}
3058 
3059 	policy->driver_data = req;
3060 
3061 	return 0;
3062 
3063 remove_min_req:
3064 	freq_qos_remove_request(req);
3065 free_req:
3066 	kfree(req);
3067 pstate_exit:
3068 	intel_pstate_exit_perf_limits(policy);
3069 
3070 	return ret;
3071 }
3072 
3073 static int intel_cpufreq_cpu_exit(struct cpufreq_policy *policy)
3074 {
3075 	struct freq_qos_request *req;
3076 
3077 	req = policy->driver_data;
3078 
3079 	freq_qos_remove_request(req + 1);
3080 	freq_qos_remove_request(req);
3081 	kfree(req);
3082 
3083 	return intel_pstate_cpu_exit(policy);
3084 }
3085 
3086 static int intel_cpufreq_suspend(struct cpufreq_policy *policy)
3087 {
3088 	intel_pstate_suspend(policy);
3089 
3090 	if (hwp_active) {
3091 		struct cpudata *cpu = all_cpu_data[policy->cpu];
3092 		u64 value = READ_ONCE(cpu->hwp_req_cached);
3093 
3094 		/*
3095 		 * Clear the desired perf field in MSR_HWP_REQUEST in case
3096 		 * intel_cpufreq_adjust_perf() is in use and the last value
3097 		 * written by it may not be suitable.
3098 		 */
3099 		value &= ~HWP_DESIRED_PERF(~0L);
3100 		wrmsrl_on_cpu(cpu->cpu, MSR_HWP_REQUEST, value);
3101 		WRITE_ONCE(cpu->hwp_req_cached, value);
3102 	}
3103 
3104 	return 0;
3105 }
3106 
3107 static struct cpufreq_driver intel_cpufreq = {
3108 	.flags		= CPUFREQ_CONST_LOOPS,
3109 	.verify		= intel_cpufreq_verify_policy,
3110 	.target		= intel_cpufreq_target,
3111 	.fast_switch	= intel_cpufreq_fast_switch,
3112 	.init		= intel_cpufreq_cpu_init,
3113 	.exit		= intel_cpufreq_cpu_exit,
3114 	.offline	= intel_cpufreq_cpu_offline,
3115 	.online		= intel_pstate_cpu_online,
3116 	.suspend	= intel_cpufreq_suspend,
3117 	.resume		= intel_pstate_resume,
3118 	.update_limits	= intel_pstate_update_limits,
3119 	.name		= "intel_cpufreq",
3120 };
3121 
3122 static struct cpufreq_driver *default_driver;
3123 
3124 static void intel_pstate_driver_cleanup(void)
3125 {
3126 	unsigned int cpu;
3127 
3128 	cpus_read_lock();
3129 	for_each_online_cpu(cpu) {
3130 		if (all_cpu_data[cpu]) {
3131 			if (intel_pstate_driver == &intel_pstate)
3132 				intel_pstate_clear_update_util_hook(cpu);
3133 
3134 			spin_lock(&hwp_notify_lock);
3135 			kfree(all_cpu_data[cpu]);
3136 			WRITE_ONCE(all_cpu_data[cpu], NULL);
3137 			spin_unlock(&hwp_notify_lock);
3138 		}
3139 	}
3140 	cpus_read_unlock();
3141 
3142 	intel_pstate_driver = NULL;
3143 }
3144 
3145 static int intel_pstate_register_driver(struct cpufreq_driver *driver)
3146 {
3147 	int ret;
3148 
3149 	if (driver == &intel_pstate)
3150 		intel_pstate_sysfs_expose_hwp_dynamic_boost();
3151 
3152 	memset(&global, 0, sizeof(global));
3153 	global.max_perf_pct = 100;
3154 
3155 	intel_pstate_driver = driver;
3156 	ret = cpufreq_register_driver(intel_pstate_driver);
3157 	if (ret) {
3158 		intel_pstate_driver_cleanup();
3159 		return ret;
3160 	}
3161 
3162 	global.min_perf_pct = min_perf_pct_min();
3163 
3164 	return 0;
3165 }
3166 
3167 static ssize_t intel_pstate_show_status(char *buf)
3168 {
3169 	if (!intel_pstate_driver)
3170 		return sprintf(buf, "off\n");
3171 
3172 	return sprintf(buf, "%s\n", intel_pstate_driver == &intel_pstate ?
3173 					"active" : "passive");
3174 }
3175 
3176 static int intel_pstate_update_status(const char *buf, size_t size)
3177 {
3178 	if (size == 3 && !strncmp(buf, "off", size)) {
3179 		if (!intel_pstate_driver)
3180 			return -EINVAL;
3181 
3182 		if (hwp_active)
3183 			return -EBUSY;
3184 
3185 		cpufreq_unregister_driver(intel_pstate_driver);
3186 		intel_pstate_driver_cleanup();
3187 		return 0;
3188 	}
3189 
3190 	if (size == 6 && !strncmp(buf, "active", size)) {
3191 		if (intel_pstate_driver) {
3192 			if (intel_pstate_driver == &intel_pstate)
3193 				return 0;
3194 
3195 			cpufreq_unregister_driver(intel_pstate_driver);
3196 		}
3197 
3198 		return intel_pstate_register_driver(&intel_pstate);
3199 	}
3200 
3201 	if (size == 7 && !strncmp(buf, "passive", size)) {
3202 		if (intel_pstate_driver) {
3203 			if (intel_pstate_driver == &intel_cpufreq)
3204 				return 0;
3205 
3206 			cpufreq_unregister_driver(intel_pstate_driver);
3207 			intel_pstate_sysfs_hide_hwp_dynamic_boost();
3208 		}
3209 
3210 		return intel_pstate_register_driver(&intel_cpufreq);
3211 	}
3212 
3213 	return -EINVAL;
3214 }
3215 
3216 static int no_load __initdata;
3217 static int no_hwp __initdata;
3218 static int hwp_only __initdata;
3219 static unsigned int force_load __initdata;
3220 
3221 static int __init intel_pstate_msrs_not_valid(void)
3222 {
3223 	if (!pstate_funcs.get_max() ||
3224 	    !pstate_funcs.get_min() ||
3225 	    !pstate_funcs.get_turbo())
3226 		return -ENODEV;
3227 
3228 	return 0;
3229 }
3230 
3231 static void __init copy_cpu_funcs(struct pstate_funcs *funcs)
3232 {
3233 	pstate_funcs.get_max   = funcs->get_max;
3234 	pstate_funcs.get_max_physical = funcs->get_max_physical;
3235 	pstate_funcs.get_min   = funcs->get_min;
3236 	pstate_funcs.get_turbo = funcs->get_turbo;
3237 	pstate_funcs.get_scaling = funcs->get_scaling;
3238 	pstate_funcs.get_val   = funcs->get_val;
3239 	pstate_funcs.get_vid   = funcs->get_vid;
3240 	pstate_funcs.get_aperf_mperf_shift = funcs->get_aperf_mperf_shift;
3241 }
3242 
3243 #ifdef CONFIG_ACPI
3244 
3245 static bool __init intel_pstate_no_acpi_pss(void)
3246 {
3247 	int i;
3248 
3249 	for_each_possible_cpu(i) {
3250 		acpi_status status;
3251 		union acpi_object *pss;
3252 		struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
3253 		struct acpi_processor *pr = per_cpu(processors, i);
3254 
3255 		if (!pr)
3256 			continue;
3257 
3258 		status = acpi_evaluate_object(pr->handle, "_PSS", NULL, &buffer);
3259 		if (ACPI_FAILURE(status))
3260 			continue;
3261 
3262 		pss = buffer.pointer;
3263 		if (pss && pss->type == ACPI_TYPE_PACKAGE) {
3264 			kfree(pss);
3265 			return false;
3266 		}
3267 
3268 		kfree(pss);
3269 	}
3270 
3271 	pr_debug("ACPI _PSS not found\n");
3272 	return true;
3273 }
3274 
3275 static bool __init intel_pstate_no_acpi_pcch(void)
3276 {
3277 	acpi_status status;
3278 	acpi_handle handle;
3279 
3280 	status = acpi_get_handle(NULL, "\\_SB", &handle);
3281 	if (ACPI_FAILURE(status))
3282 		goto not_found;
3283 
3284 	if (acpi_has_method(handle, "PCCH"))
3285 		return false;
3286 
3287 not_found:
3288 	pr_debug("ACPI PCCH not found\n");
3289 	return true;
3290 }
3291 
3292 static bool __init intel_pstate_has_acpi_ppc(void)
3293 {
3294 	int i;
3295 
3296 	for_each_possible_cpu(i) {
3297 		struct acpi_processor *pr = per_cpu(processors, i);
3298 
3299 		if (!pr)
3300 			continue;
3301 		if (acpi_has_method(pr->handle, "_PPC"))
3302 			return true;
3303 	}
3304 	pr_debug("ACPI _PPC not found\n");
3305 	return false;
3306 }
3307 
3308 enum {
3309 	PSS,
3310 	PPC,
3311 };
3312 
3313 /* Hardware vendor-specific info that has its own power management modes */
3314 static struct acpi_platform_list plat_info[] __initdata = {
3315 	{"HP    ", "ProLiant", 0, ACPI_SIG_FADT, all_versions, NULL, PSS},
3316 	{"ORACLE", "X4-2    ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3317 	{"ORACLE", "X4-2L   ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3318 	{"ORACLE", "X4-2B   ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3319 	{"ORACLE", "X3-2    ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3320 	{"ORACLE", "X3-2L   ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3321 	{"ORACLE", "X3-2B   ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3322 	{"ORACLE", "X4470M2 ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3323 	{"ORACLE", "X4270M3 ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3324 	{"ORACLE", "X4270M2 ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3325 	{"ORACLE", "X4170M2 ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3326 	{"ORACLE", "X4170 M3", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3327 	{"ORACLE", "X4275 M3", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3328 	{"ORACLE", "X6-2    ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3329 	{"ORACLE", "Sudbury ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3330 	{ } /* End */
3331 };
3332 
3333 #define BITMASK_OOB	(BIT(8) | BIT(18))
3334 
3335 static bool __init intel_pstate_platform_pwr_mgmt_exists(void)
3336 {
3337 	const struct x86_cpu_id *id;
3338 	u64 misc_pwr;
3339 	int idx;
3340 
3341 	id = x86_match_cpu(intel_pstate_cpu_oob_ids);
3342 	if (id) {
3343 		rdmsrl(MSR_MISC_PWR_MGMT, misc_pwr);
3344 		if (misc_pwr & BITMASK_OOB) {
3345 			pr_debug("Bit 8 or 18 in the MISC_PWR_MGMT MSR set\n");
3346 			pr_debug("P states are controlled in Out of Band mode by the firmware/hardware\n");
3347 			return true;
3348 		}
3349 	}
3350 
3351 	idx = acpi_match_platform_list(plat_info);
3352 	if (idx < 0)
3353 		return false;
3354 
3355 	switch (plat_info[idx].data) {
3356 	case PSS:
3357 		if (!intel_pstate_no_acpi_pss())
3358 			return false;
3359 
3360 		return intel_pstate_no_acpi_pcch();
3361 	case PPC:
3362 		return intel_pstate_has_acpi_ppc() && !force_load;
3363 	}
3364 
3365 	return false;
3366 }
3367 
3368 static void intel_pstate_request_control_from_smm(void)
3369 {
3370 	/*
3371 	 * It may be unsafe to request P-states control from SMM if _PPC support
3372 	 * has not been enabled.
3373 	 */
3374 	if (acpi_ppc)
3375 		acpi_processor_pstate_control();
3376 }
3377 #else /* CONFIG_ACPI not enabled */
3378 static inline bool intel_pstate_platform_pwr_mgmt_exists(void) { return false; }
3379 static inline bool intel_pstate_has_acpi_ppc(void) { return false; }
3380 static inline void intel_pstate_request_control_from_smm(void) {}
3381 #endif /* CONFIG_ACPI */
3382 
3383 #define INTEL_PSTATE_HWP_BROADWELL	0x01
3384 
3385 #define X86_MATCH_HWP(model, hwp_mode)					\
3386 	X86_MATCH_VENDOR_FAM_MODEL_FEATURE(INTEL, 6, INTEL_FAM6_##model, \
3387 					   X86_FEATURE_HWP, hwp_mode)
3388 
3389 static const struct x86_cpu_id hwp_support_ids[] __initconst = {
3390 	X86_MATCH_HWP(BROADWELL_X,	INTEL_PSTATE_HWP_BROADWELL),
3391 	X86_MATCH_HWP(BROADWELL_D,	INTEL_PSTATE_HWP_BROADWELL),
3392 	X86_MATCH_HWP(ANY,		0),
3393 	{}
3394 };
3395 
3396 static bool intel_pstate_hwp_is_enabled(void)
3397 {
3398 	u64 value;
3399 
3400 	rdmsrl(MSR_PM_ENABLE, value);
3401 	return !!(value & 0x1);
3402 }
3403 
3404 static const struct x86_cpu_id intel_epp_balance_perf[] = {
3405 	/*
3406 	 * Set EPP value as 102, this is the max suggested EPP
3407 	 * which can result in one core turbo frequency for
3408 	 * AlderLake Mobile CPUs.
3409 	 */
3410 	X86_MATCH_INTEL_FAM6_MODEL(ALDERLAKE_L, 102),
3411 	{}
3412 };
3413 
3414 static int __init intel_pstate_init(void)
3415 {
3416 	static struct cpudata **_all_cpu_data;
3417 	const struct x86_cpu_id *id;
3418 	int rc;
3419 
3420 	if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL)
3421 		return -ENODEV;
3422 
3423 	id = x86_match_cpu(hwp_support_ids);
3424 	if (id) {
3425 		bool hwp_forced = intel_pstate_hwp_is_enabled();
3426 
3427 		if (hwp_forced)
3428 			pr_info("HWP enabled by BIOS\n");
3429 		else if (no_load)
3430 			return -ENODEV;
3431 
3432 		copy_cpu_funcs(&core_funcs);
3433 		/*
3434 		 * Avoid enabling HWP for processors without EPP support,
3435 		 * because that means incomplete HWP implementation which is a
3436 		 * corner case and supporting it is generally problematic.
3437 		 *
3438 		 * If HWP is enabled already, though, there is no choice but to
3439 		 * deal with it.
3440 		 */
3441 		if ((!no_hwp && boot_cpu_has(X86_FEATURE_HWP_EPP)) || hwp_forced) {
3442 			WRITE_ONCE(hwp_active, 1);
3443 			hwp_mode_bdw = id->driver_data;
3444 			intel_pstate.attr = hwp_cpufreq_attrs;
3445 			intel_cpufreq.attr = hwp_cpufreq_attrs;
3446 			intel_cpufreq.flags |= CPUFREQ_NEED_UPDATE_LIMITS;
3447 			intel_cpufreq.adjust_perf = intel_cpufreq_adjust_perf;
3448 			if (!default_driver)
3449 				default_driver = &intel_pstate;
3450 
3451 			if (boot_cpu_has(X86_FEATURE_HYBRID_CPU))
3452 				intel_pstate_cppc_set_cpu_scaling();
3453 
3454 			goto hwp_cpu_matched;
3455 		}
3456 		pr_info("HWP not enabled\n");
3457 	} else {
3458 		if (no_load)
3459 			return -ENODEV;
3460 
3461 		id = x86_match_cpu(intel_pstate_cpu_ids);
3462 		if (!id) {
3463 			pr_info("CPU model not supported\n");
3464 			return -ENODEV;
3465 		}
3466 
3467 		copy_cpu_funcs((struct pstate_funcs *)id->driver_data);
3468 	}
3469 
3470 	if (intel_pstate_msrs_not_valid()) {
3471 		pr_info("Invalid MSRs\n");
3472 		return -ENODEV;
3473 	}
3474 	/* Without HWP start in the passive mode. */
3475 	if (!default_driver)
3476 		default_driver = &intel_cpufreq;
3477 
3478 hwp_cpu_matched:
3479 	/*
3480 	 * The Intel pstate driver will be ignored if the platform
3481 	 * firmware has its own power management modes.
3482 	 */
3483 	if (intel_pstate_platform_pwr_mgmt_exists()) {
3484 		pr_info("P-states controlled by the platform\n");
3485 		return -ENODEV;
3486 	}
3487 
3488 	if (!hwp_active && hwp_only)
3489 		return -ENOTSUPP;
3490 
3491 	pr_info("Intel P-state driver initializing\n");
3492 
3493 	_all_cpu_data = vzalloc(array_size(sizeof(void *), num_possible_cpus()));
3494 	if (!_all_cpu_data)
3495 		return -ENOMEM;
3496 
3497 	WRITE_ONCE(all_cpu_data, _all_cpu_data);
3498 
3499 	intel_pstate_request_control_from_smm();
3500 
3501 	intel_pstate_sysfs_expose_params();
3502 
3503 	if (hwp_active) {
3504 		const struct x86_cpu_id *id = x86_match_cpu(intel_epp_balance_perf);
3505 
3506 		if (id)
3507 			epp_values[EPP_INDEX_BALANCE_PERFORMANCE] = id->driver_data;
3508 	}
3509 
3510 	mutex_lock(&intel_pstate_driver_lock);
3511 	rc = intel_pstate_register_driver(default_driver);
3512 	mutex_unlock(&intel_pstate_driver_lock);
3513 	if (rc) {
3514 		intel_pstate_sysfs_remove();
3515 		return rc;
3516 	}
3517 
3518 	if (hwp_active) {
3519 		const struct x86_cpu_id *id;
3520 
3521 		id = x86_match_cpu(intel_pstate_cpu_ee_disable_ids);
3522 		if (id) {
3523 			set_power_ctl_ee_state(false);
3524 			pr_info("Disabling energy efficiency optimization\n");
3525 		}
3526 
3527 		pr_info("HWP enabled\n");
3528 	} else if (boot_cpu_has(X86_FEATURE_HYBRID_CPU)) {
3529 		pr_warn("Problematic setup: Hybrid processor with disabled HWP\n");
3530 	}
3531 
3532 	return 0;
3533 }
3534 device_initcall(intel_pstate_init);
3535 
3536 static int __init intel_pstate_setup(char *str)
3537 {
3538 	if (!str)
3539 		return -EINVAL;
3540 
3541 	if (!strcmp(str, "disable"))
3542 		no_load = 1;
3543 	else if (!strcmp(str, "active"))
3544 		default_driver = &intel_pstate;
3545 	else if (!strcmp(str, "passive"))
3546 		default_driver = &intel_cpufreq;
3547 
3548 	if (!strcmp(str, "no_hwp"))
3549 		no_hwp = 1;
3550 
3551 	if (!strcmp(str, "force"))
3552 		force_load = 1;
3553 	if (!strcmp(str, "hwp_only"))
3554 		hwp_only = 1;
3555 	if (!strcmp(str, "per_cpu_perf_limits"))
3556 		per_cpu_limits = true;
3557 
3558 #ifdef CONFIG_ACPI
3559 	if (!strcmp(str, "support_acpi_ppc"))
3560 		acpi_ppc = true;
3561 #endif
3562 
3563 	return 0;
3564 }
3565 early_param("intel_pstate", intel_pstate_setup);
3566 
3567 MODULE_AUTHOR("Dirk Brandewie <dirk.j.brandewie@intel.com>");
3568 MODULE_DESCRIPTION("'intel_pstate' - P state driver Intel Core processors");
3569 MODULE_LICENSE("GPL");
3570