xref: /linux/drivers/cpufreq/intel_pstate.c (revision f69e98a91a01fd7c5755dd710e94a17d6e9f583f)
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 
1326 	mutex_unlock(&intel_pstate_driver_lock);
1327 
1328 	return count;
1329 }
1330 
1331 static void update_qos_request(enum freq_qos_req_type type)
1332 {
1333 	struct freq_qos_request *req;
1334 	struct cpufreq_policy *policy;
1335 	int i;
1336 
1337 	for_each_possible_cpu(i) {
1338 		struct cpudata *cpu = all_cpu_data[i];
1339 		unsigned int freq, perf_pct;
1340 
1341 		policy = cpufreq_cpu_get(i);
1342 		if (!policy)
1343 			continue;
1344 
1345 		req = policy->driver_data;
1346 		cpufreq_cpu_put(policy);
1347 
1348 		if (!req)
1349 			continue;
1350 
1351 		if (hwp_active)
1352 			intel_pstate_get_hwp_cap(cpu);
1353 
1354 		if (type == FREQ_QOS_MIN) {
1355 			perf_pct = global.min_perf_pct;
1356 		} else {
1357 			req++;
1358 			perf_pct = global.max_perf_pct;
1359 		}
1360 
1361 		freq = DIV_ROUND_UP(cpu->pstate.turbo_freq * perf_pct, 100);
1362 
1363 		if (freq_qos_update_request(req, freq) < 0)
1364 			pr_warn("Failed to update freq constraint: CPU%d\n", i);
1365 	}
1366 }
1367 
1368 static ssize_t store_max_perf_pct(struct kobject *a, struct kobj_attribute *b,
1369 				  const char *buf, size_t count)
1370 {
1371 	unsigned int input;
1372 	int ret;
1373 
1374 	ret = sscanf(buf, "%u", &input);
1375 	if (ret != 1)
1376 		return -EINVAL;
1377 
1378 	mutex_lock(&intel_pstate_driver_lock);
1379 
1380 	if (!intel_pstate_driver) {
1381 		mutex_unlock(&intel_pstate_driver_lock);
1382 		return -EAGAIN;
1383 	}
1384 
1385 	mutex_lock(&intel_pstate_limits_lock);
1386 
1387 	global.max_perf_pct = clamp_t(int, input, global.min_perf_pct, 100);
1388 
1389 	mutex_unlock(&intel_pstate_limits_lock);
1390 
1391 	if (intel_pstate_driver == &intel_pstate)
1392 		intel_pstate_update_policies();
1393 	else
1394 		update_qos_request(FREQ_QOS_MAX);
1395 
1396 	mutex_unlock(&intel_pstate_driver_lock);
1397 
1398 	return count;
1399 }
1400 
1401 static ssize_t store_min_perf_pct(struct kobject *a, struct kobj_attribute *b,
1402 				  const char *buf, size_t count)
1403 {
1404 	unsigned int input;
1405 	int ret;
1406 
1407 	ret = sscanf(buf, "%u", &input);
1408 	if (ret != 1)
1409 		return -EINVAL;
1410 
1411 	mutex_lock(&intel_pstate_driver_lock);
1412 
1413 	if (!intel_pstate_driver) {
1414 		mutex_unlock(&intel_pstate_driver_lock);
1415 		return -EAGAIN;
1416 	}
1417 
1418 	mutex_lock(&intel_pstate_limits_lock);
1419 
1420 	global.min_perf_pct = clamp_t(int, input,
1421 				      min_perf_pct_min(), global.max_perf_pct);
1422 
1423 	mutex_unlock(&intel_pstate_limits_lock);
1424 
1425 	if (intel_pstate_driver == &intel_pstate)
1426 		intel_pstate_update_policies();
1427 	else
1428 		update_qos_request(FREQ_QOS_MIN);
1429 
1430 	mutex_unlock(&intel_pstate_driver_lock);
1431 
1432 	return count;
1433 }
1434 
1435 static ssize_t show_hwp_dynamic_boost(struct kobject *kobj,
1436 				struct kobj_attribute *attr, char *buf)
1437 {
1438 	return sprintf(buf, "%u\n", hwp_boost);
1439 }
1440 
1441 static ssize_t store_hwp_dynamic_boost(struct kobject *a,
1442 				       struct kobj_attribute *b,
1443 				       const char *buf, size_t count)
1444 {
1445 	unsigned int input;
1446 	int ret;
1447 
1448 	ret = kstrtouint(buf, 10, &input);
1449 	if (ret)
1450 		return ret;
1451 
1452 	mutex_lock(&intel_pstate_driver_lock);
1453 	hwp_boost = !!input;
1454 	intel_pstate_update_policies();
1455 	mutex_unlock(&intel_pstate_driver_lock);
1456 
1457 	return count;
1458 }
1459 
1460 static ssize_t show_energy_efficiency(struct kobject *kobj, struct kobj_attribute *attr,
1461 				      char *buf)
1462 {
1463 	u64 power_ctl;
1464 	int enable;
1465 
1466 	rdmsrl(MSR_IA32_POWER_CTL, power_ctl);
1467 	enable = !!(power_ctl & BIT(MSR_IA32_POWER_CTL_BIT_EE));
1468 	return sprintf(buf, "%d\n", !enable);
1469 }
1470 
1471 static ssize_t store_energy_efficiency(struct kobject *a, struct kobj_attribute *b,
1472 				       const char *buf, size_t count)
1473 {
1474 	bool input;
1475 	int ret;
1476 
1477 	ret = kstrtobool(buf, &input);
1478 	if (ret)
1479 		return ret;
1480 
1481 	set_power_ctl_ee_state(input);
1482 
1483 	return count;
1484 }
1485 
1486 show_one(max_perf_pct, max_perf_pct);
1487 show_one(min_perf_pct, min_perf_pct);
1488 
1489 define_one_global_rw(status);
1490 define_one_global_rw(no_turbo);
1491 define_one_global_rw(max_perf_pct);
1492 define_one_global_rw(min_perf_pct);
1493 define_one_global_ro(turbo_pct);
1494 define_one_global_ro(num_pstates);
1495 define_one_global_rw(hwp_dynamic_boost);
1496 define_one_global_rw(energy_efficiency);
1497 
1498 static struct attribute *intel_pstate_attributes[] = {
1499 	&status.attr,
1500 	&no_turbo.attr,
1501 	NULL
1502 };
1503 
1504 static const struct attribute_group intel_pstate_attr_group = {
1505 	.attrs = intel_pstate_attributes,
1506 };
1507 
1508 static const struct x86_cpu_id intel_pstate_cpu_ee_disable_ids[];
1509 
1510 static struct kobject *intel_pstate_kobject;
1511 
1512 static void __init intel_pstate_sysfs_expose_params(void)
1513 {
1514 	int rc;
1515 
1516 	intel_pstate_kobject = kobject_create_and_add("intel_pstate",
1517 						&cpu_subsys.dev_root->kobj);
1518 	if (WARN_ON(!intel_pstate_kobject))
1519 		return;
1520 
1521 	rc = sysfs_create_group(intel_pstate_kobject, &intel_pstate_attr_group);
1522 	if (WARN_ON(rc))
1523 		return;
1524 
1525 	if (!boot_cpu_has(X86_FEATURE_HYBRID_CPU)) {
1526 		rc = sysfs_create_file(intel_pstate_kobject, &turbo_pct.attr);
1527 		WARN_ON(rc);
1528 
1529 		rc = sysfs_create_file(intel_pstate_kobject, &num_pstates.attr);
1530 		WARN_ON(rc);
1531 	}
1532 
1533 	/*
1534 	 * If per cpu limits are enforced there are no global limits, so
1535 	 * return without creating max/min_perf_pct attributes
1536 	 */
1537 	if (per_cpu_limits)
1538 		return;
1539 
1540 	rc = sysfs_create_file(intel_pstate_kobject, &max_perf_pct.attr);
1541 	WARN_ON(rc);
1542 
1543 	rc = sysfs_create_file(intel_pstate_kobject, &min_perf_pct.attr);
1544 	WARN_ON(rc);
1545 
1546 	if (x86_match_cpu(intel_pstate_cpu_ee_disable_ids)) {
1547 		rc = sysfs_create_file(intel_pstate_kobject, &energy_efficiency.attr);
1548 		WARN_ON(rc);
1549 	}
1550 }
1551 
1552 static void __init intel_pstate_sysfs_remove(void)
1553 {
1554 	if (!intel_pstate_kobject)
1555 		return;
1556 
1557 	sysfs_remove_group(intel_pstate_kobject, &intel_pstate_attr_group);
1558 
1559 	if (!boot_cpu_has(X86_FEATURE_HYBRID_CPU)) {
1560 		sysfs_remove_file(intel_pstate_kobject, &num_pstates.attr);
1561 		sysfs_remove_file(intel_pstate_kobject, &turbo_pct.attr);
1562 	}
1563 
1564 	if (!per_cpu_limits) {
1565 		sysfs_remove_file(intel_pstate_kobject, &max_perf_pct.attr);
1566 		sysfs_remove_file(intel_pstate_kobject, &min_perf_pct.attr);
1567 
1568 		if (x86_match_cpu(intel_pstate_cpu_ee_disable_ids))
1569 			sysfs_remove_file(intel_pstate_kobject, &energy_efficiency.attr);
1570 	}
1571 
1572 	kobject_put(intel_pstate_kobject);
1573 }
1574 
1575 static void intel_pstate_sysfs_expose_hwp_dynamic_boost(void)
1576 {
1577 	int rc;
1578 
1579 	if (!hwp_active)
1580 		return;
1581 
1582 	rc = sysfs_create_file(intel_pstate_kobject, &hwp_dynamic_boost.attr);
1583 	WARN_ON_ONCE(rc);
1584 }
1585 
1586 static void intel_pstate_sysfs_hide_hwp_dynamic_boost(void)
1587 {
1588 	if (!hwp_active)
1589 		return;
1590 
1591 	sysfs_remove_file(intel_pstate_kobject, &hwp_dynamic_boost.attr);
1592 }
1593 
1594 /************************** sysfs end ************************/
1595 
1596 static void intel_pstate_notify_work(struct work_struct *work)
1597 {
1598 	struct cpudata *cpudata =
1599 		container_of(to_delayed_work(work), struct cpudata, hwp_notify_work);
1600 	struct cpufreq_policy *policy = cpufreq_cpu_acquire(cpudata->cpu);
1601 
1602 	if (policy) {
1603 		intel_pstate_get_hwp_cap(cpudata);
1604 		__intel_pstate_update_max_freq(cpudata, policy);
1605 
1606 		cpufreq_cpu_release(policy);
1607 	}
1608 
1609 	wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_STATUS, 0);
1610 }
1611 
1612 static DEFINE_SPINLOCK(hwp_notify_lock);
1613 static cpumask_t hwp_intr_enable_mask;
1614 
1615 void notify_hwp_interrupt(void)
1616 {
1617 	unsigned int this_cpu = smp_processor_id();
1618 	struct cpudata *cpudata;
1619 	unsigned long flags;
1620 	u64 value;
1621 
1622 	if (!READ_ONCE(hwp_active) || !boot_cpu_has(X86_FEATURE_HWP_NOTIFY))
1623 		return;
1624 
1625 	rdmsrl_safe(MSR_HWP_STATUS, &value);
1626 	if (!(value & 0x01))
1627 		return;
1628 
1629 	spin_lock_irqsave(&hwp_notify_lock, flags);
1630 
1631 	if (!cpumask_test_cpu(this_cpu, &hwp_intr_enable_mask))
1632 		goto ack_intr;
1633 
1634 	/*
1635 	 * Currently we never free all_cpu_data. And we can't reach here
1636 	 * without this allocated. But for safety for future changes, added
1637 	 * check.
1638 	 */
1639 	if (unlikely(!READ_ONCE(all_cpu_data)))
1640 		goto ack_intr;
1641 
1642 	/*
1643 	 * The free is done during cleanup, when cpufreq registry is failed.
1644 	 * We wouldn't be here if it fails on init or switch status. But for
1645 	 * future changes, added check.
1646 	 */
1647 	cpudata = READ_ONCE(all_cpu_data[this_cpu]);
1648 	if (unlikely(!cpudata))
1649 		goto ack_intr;
1650 
1651 	schedule_delayed_work(&cpudata->hwp_notify_work, msecs_to_jiffies(10));
1652 
1653 	spin_unlock_irqrestore(&hwp_notify_lock, flags);
1654 
1655 	return;
1656 
1657 ack_intr:
1658 	wrmsrl_safe(MSR_HWP_STATUS, 0);
1659 	spin_unlock_irqrestore(&hwp_notify_lock, flags);
1660 }
1661 
1662 static void intel_pstate_disable_hwp_interrupt(struct cpudata *cpudata)
1663 {
1664 	unsigned long flags;
1665 
1666 	if (!boot_cpu_has(X86_FEATURE_HWP_NOTIFY))
1667 		return;
1668 
1669 	/* wrmsrl_on_cpu has to be outside spinlock as this can result in IPC */
1670 	wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_INTERRUPT, 0x00);
1671 
1672 	spin_lock_irqsave(&hwp_notify_lock, flags);
1673 	if (cpumask_test_and_clear_cpu(cpudata->cpu, &hwp_intr_enable_mask))
1674 		cancel_delayed_work(&cpudata->hwp_notify_work);
1675 	spin_unlock_irqrestore(&hwp_notify_lock, flags);
1676 }
1677 
1678 static void intel_pstate_enable_hwp_interrupt(struct cpudata *cpudata)
1679 {
1680 	/* Enable HWP notification interrupt for guaranteed performance change */
1681 	if (boot_cpu_has(X86_FEATURE_HWP_NOTIFY)) {
1682 		unsigned long flags;
1683 
1684 		spin_lock_irqsave(&hwp_notify_lock, flags);
1685 		INIT_DELAYED_WORK(&cpudata->hwp_notify_work, intel_pstate_notify_work);
1686 		cpumask_set_cpu(cpudata->cpu, &hwp_intr_enable_mask);
1687 		spin_unlock_irqrestore(&hwp_notify_lock, flags);
1688 
1689 		/* wrmsrl_on_cpu has to be outside spinlock as this can result in IPC */
1690 		wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_INTERRUPT, 0x01);
1691 		wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_STATUS, 0);
1692 	}
1693 }
1694 
1695 static void intel_pstate_update_epp_defaults(struct cpudata *cpudata)
1696 {
1697 	cpudata->epp_default = intel_pstate_get_epp(cpudata, 0);
1698 
1699 	/*
1700 	 * If this CPU gen doesn't call for change in balance_perf
1701 	 * EPP return.
1702 	 */
1703 	if (epp_values[EPP_INDEX_BALANCE_PERFORMANCE] == HWP_EPP_BALANCE_PERFORMANCE)
1704 		return;
1705 
1706 	/*
1707 	 * If powerup EPP is something other than chipset default 0x80 and
1708 	 * - is more performance oriented than 0x80 (default balance_perf EPP)
1709 	 * - But less performance oriented than performance EPP
1710 	 *   then use this as new balance_perf EPP.
1711 	 */
1712 	if (cpudata->epp_default < HWP_EPP_BALANCE_PERFORMANCE &&
1713 	    cpudata->epp_default > HWP_EPP_PERFORMANCE) {
1714 		epp_values[EPP_INDEX_BALANCE_PERFORMANCE] = cpudata->epp_default;
1715 		return;
1716 	}
1717 
1718 	/*
1719 	 * Use hard coded value per gen to update the balance_perf
1720 	 * and default EPP.
1721 	 */
1722 	cpudata->epp_default = epp_values[EPP_INDEX_BALANCE_PERFORMANCE];
1723 	intel_pstate_set_epp(cpudata, cpudata->epp_default);
1724 }
1725 
1726 static void intel_pstate_hwp_enable(struct cpudata *cpudata)
1727 {
1728 	/* First disable HWP notification interrupt till we activate again */
1729 	if (boot_cpu_has(X86_FEATURE_HWP_NOTIFY))
1730 		wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_INTERRUPT, 0x00);
1731 
1732 	wrmsrl_on_cpu(cpudata->cpu, MSR_PM_ENABLE, 0x1);
1733 
1734 	intel_pstate_enable_hwp_interrupt(cpudata);
1735 
1736 	if (cpudata->epp_default >= 0)
1737 		return;
1738 
1739 	intel_pstate_update_epp_defaults(cpudata);
1740 }
1741 
1742 static int atom_get_min_pstate(void)
1743 {
1744 	u64 value;
1745 
1746 	rdmsrl(MSR_ATOM_CORE_RATIOS, value);
1747 	return (value >> 8) & 0x7F;
1748 }
1749 
1750 static int atom_get_max_pstate(void)
1751 {
1752 	u64 value;
1753 
1754 	rdmsrl(MSR_ATOM_CORE_RATIOS, value);
1755 	return (value >> 16) & 0x7F;
1756 }
1757 
1758 static int atom_get_turbo_pstate(void)
1759 {
1760 	u64 value;
1761 
1762 	rdmsrl(MSR_ATOM_CORE_TURBO_RATIOS, value);
1763 	return value & 0x7F;
1764 }
1765 
1766 static u64 atom_get_val(struct cpudata *cpudata, int pstate)
1767 {
1768 	u64 val;
1769 	int32_t vid_fp;
1770 	u32 vid;
1771 
1772 	val = (u64)pstate << 8;
1773 	if (global.no_turbo && !global.turbo_disabled)
1774 		val |= (u64)1 << 32;
1775 
1776 	vid_fp = cpudata->vid.min + mul_fp(
1777 		int_tofp(pstate - cpudata->pstate.min_pstate),
1778 		cpudata->vid.ratio);
1779 
1780 	vid_fp = clamp_t(int32_t, vid_fp, cpudata->vid.min, cpudata->vid.max);
1781 	vid = ceiling_fp(vid_fp);
1782 
1783 	if (pstate > cpudata->pstate.max_pstate)
1784 		vid = cpudata->vid.turbo;
1785 
1786 	return val | vid;
1787 }
1788 
1789 static int silvermont_get_scaling(void)
1790 {
1791 	u64 value;
1792 	int i;
1793 	/* Defined in Table 35-6 from SDM (Sept 2015) */
1794 	static int silvermont_freq_table[] = {
1795 		83300, 100000, 133300, 116700, 80000};
1796 
1797 	rdmsrl(MSR_FSB_FREQ, value);
1798 	i = value & 0x7;
1799 	WARN_ON(i > 4);
1800 
1801 	return silvermont_freq_table[i];
1802 }
1803 
1804 static int airmont_get_scaling(void)
1805 {
1806 	u64 value;
1807 	int i;
1808 	/* Defined in Table 35-10 from SDM (Sept 2015) */
1809 	static int airmont_freq_table[] = {
1810 		83300, 100000, 133300, 116700, 80000,
1811 		93300, 90000, 88900, 87500};
1812 
1813 	rdmsrl(MSR_FSB_FREQ, value);
1814 	i = value & 0xF;
1815 	WARN_ON(i > 8);
1816 
1817 	return airmont_freq_table[i];
1818 }
1819 
1820 static void atom_get_vid(struct cpudata *cpudata)
1821 {
1822 	u64 value;
1823 
1824 	rdmsrl(MSR_ATOM_CORE_VIDS, value);
1825 	cpudata->vid.min = int_tofp((value >> 8) & 0x7f);
1826 	cpudata->vid.max = int_tofp((value >> 16) & 0x7f);
1827 	cpudata->vid.ratio = div_fp(
1828 		cpudata->vid.max - cpudata->vid.min,
1829 		int_tofp(cpudata->pstate.max_pstate -
1830 			cpudata->pstate.min_pstate));
1831 
1832 	rdmsrl(MSR_ATOM_CORE_TURBO_VIDS, value);
1833 	cpudata->vid.turbo = value & 0x7f;
1834 }
1835 
1836 static int core_get_min_pstate(void)
1837 {
1838 	u64 value;
1839 
1840 	rdmsrl(MSR_PLATFORM_INFO, value);
1841 	return (value >> 40) & 0xFF;
1842 }
1843 
1844 static int core_get_max_pstate_physical(void)
1845 {
1846 	u64 value;
1847 
1848 	rdmsrl(MSR_PLATFORM_INFO, value);
1849 	return (value >> 8) & 0xFF;
1850 }
1851 
1852 static int core_get_tdp_ratio(u64 plat_info)
1853 {
1854 	/* Check how many TDP levels present */
1855 	if (plat_info & 0x600000000) {
1856 		u64 tdp_ctrl;
1857 		u64 tdp_ratio;
1858 		int tdp_msr;
1859 		int err;
1860 
1861 		/* Get the TDP level (0, 1, 2) to get ratios */
1862 		err = rdmsrl_safe(MSR_CONFIG_TDP_CONTROL, &tdp_ctrl);
1863 		if (err)
1864 			return err;
1865 
1866 		/* TDP MSR are continuous starting at 0x648 */
1867 		tdp_msr = MSR_CONFIG_TDP_NOMINAL + (tdp_ctrl & 0x03);
1868 		err = rdmsrl_safe(tdp_msr, &tdp_ratio);
1869 		if (err)
1870 			return err;
1871 
1872 		/* For level 1 and 2, bits[23:16] contain the ratio */
1873 		if (tdp_ctrl & 0x03)
1874 			tdp_ratio >>= 16;
1875 
1876 		tdp_ratio &= 0xff; /* ratios are only 8 bits long */
1877 		pr_debug("tdp_ratio %x\n", (int)tdp_ratio);
1878 
1879 		return (int)tdp_ratio;
1880 	}
1881 
1882 	return -ENXIO;
1883 }
1884 
1885 static int core_get_max_pstate(void)
1886 {
1887 	u64 tar;
1888 	u64 plat_info;
1889 	int max_pstate;
1890 	int tdp_ratio;
1891 	int err;
1892 
1893 	rdmsrl(MSR_PLATFORM_INFO, plat_info);
1894 	max_pstate = (plat_info >> 8) & 0xFF;
1895 
1896 	tdp_ratio = core_get_tdp_ratio(plat_info);
1897 	if (tdp_ratio <= 0)
1898 		return max_pstate;
1899 
1900 	if (hwp_active) {
1901 		/* Turbo activation ratio is not used on HWP platforms */
1902 		return tdp_ratio;
1903 	}
1904 
1905 	err = rdmsrl_safe(MSR_TURBO_ACTIVATION_RATIO, &tar);
1906 	if (!err) {
1907 		int tar_levels;
1908 
1909 		/* Do some sanity checking for safety */
1910 		tar_levels = tar & 0xff;
1911 		if (tdp_ratio - 1 == tar_levels) {
1912 			max_pstate = tar_levels;
1913 			pr_debug("max_pstate=TAC %x\n", max_pstate);
1914 		}
1915 	}
1916 
1917 	return max_pstate;
1918 }
1919 
1920 static int core_get_turbo_pstate(void)
1921 {
1922 	u64 value;
1923 	int nont, ret;
1924 
1925 	rdmsrl(MSR_TURBO_RATIO_LIMIT, value);
1926 	nont = core_get_max_pstate();
1927 	ret = (value) & 255;
1928 	if (ret <= nont)
1929 		ret = nont;
1930 	return ret;
1931 }
1932 
1933 static inline int core_get_scaling(void)
1934 {
1935 	return 100000;
1936 }
1937 
1938 static u64 core_get_val(struct cpudata *cpudata, int pstate)
1939 {
1940 	u64 val;
1941 
1942 	val = (u64)pstate << 8;
1943 	if (global.no_turbo && !global.turbo_disabled)
1944 		val |= (u64)1 << 32;
1945 
1946 	return val;
1947 }
1948 
1949 static int knl_get_aperf_mperf_shift(void)
1950 {
1951 	return 10;
1952 }
1953 
1954 static int knl_get_turbo_pstate(void)
1955 {
1956 	u64 value;
1957 	int nont, ret;
1958 
1959 	rdmsrl(MSR_TURBO_RATIO_LIMIT, value);
1960 	nont = core_get_max_pstate();
1961 	ret = (((value) >> 8) & 0xFF);
1962 	if (ret <= nont)
1963 		ret = nont;
1964 	return ret;
1965 }
1966 
1967 #ifdef CONFIG_ACPI_CPPC_LIB
1968 static u32 hybrid_ref_perf;
1969 
1970 static int hybrid_get_cpu_scaling(int cpu)
1971 {
1972 	return DIV_ROUND_UP(core_get_scaling() * hybrid_ref_perf,
1973 			    intel_pstate_cppc_nominal(cpu));
1974 }
1975 
1976 static void intel_pstate_cppc_set_cpu_scaling(void)
1977 {
1978 	u32 min_nominal_perf = U32_MAX;
1979 	int cpu;
1980 
1981 	for_each_present_cpu(cpu) {
1982 		u32 nominal_perf = intel_pstate_cppc_nominal(cpu);
1983 
1984 		if (nominal_perf && nominal_perf < min_nominal_perf)
1985 			min_nominal_perf = nominal_perf;
1986 	}
1987 
1988 	if (min_nominal_perf < U32_MAX) {
1989 		hybrid_ref_perf = min_nominal_perf;
1990 		pstate_funcs.get_cpu_scaling = hybrid_get_cpu_scaling;
1991 	}
1992 }
1993 #else
1994 static inline void intel_pstate_cppc_set_cpu_scaling(void)
1995 {
1996 }
1997 #endif /* CONFIG_ACPI_CPPC_LIB */
1998 
1999 static void intel_pstate_set_pstate(struct cpudata *cpu, int pstate)
2000 {
2001 	trace_cpu_frequency(pstate * cpu->pstate.scaling, cpu->cpu);
2002 	cpu->pstate.current_pstate = pstate;
2003 	/*
2004 	 * Generally, there is no guarantee that this code will always run on
2005 	 * the CPU being updated, so force the register update to run on the
2006 	 * right CPU.
2007 	 */
2008 	wrmsrl_on_cpu(cpu->cpu, MSR_IA32_PERF_CTL,
2009 		      pstate_funcs.get_val(cpu, pstate));
2010 }
2011 
2012 static void intel_pstate_set_min_pstate(struct cpudata *cpu)
2013 {
2014 	intel_pstate_set_pstate(cpu, cpu->pstate.min_pstate);
2015 }
2016 
2017 static void intel_pstate_max_within_limits(struct cpudata *cpu)
2018 {
2019 	int pstate = max(cpu->pstate.min_pstate, cpu->max_perf_ratio);
2020 
2021 	update_turbo_state();
2022 	intel_pstate_set_pstate(cpu, pstate);
2023 }
2024 
2025 static void intel_pstate_get_cpu_pstates(struct cpudata *cpu)
2026 {
2027 	int perf_ctl_max_phys = pstate_funcs.get_max_physical();
2028 	int perf_ctl_scaling = pstate_funcs.get_scaling();
2029 
2030 	cpu->pstate.min_pstate = pstate_funcs.get_min();
2031 	cpu->pstate.max_pstate_physical = perf_ctl_max_phys;
2032 	cpu->pstate.perf_ctl_scaling = perf_ctl_scaling;
2033 
2034 	if (hwp_active && !hwp_mode_bdw) {
2035 		__intel_pstate_get_hwp_cap(cpu);
2036 
2037 		if (pstate_funcs.get_cpu_scaling) {
2038 			cpu->pstate.scaling = pstate_funcs.get_cpu_scaling(cpu->cpu);
2039 			if (cpu->pstate.scaling != perf_ctl_scaling)
2040 				intel_pstate_hybrid_hwp_adjust(cpu);
2041 		} else {
2042 			cpu->pstate.scaling = perf_ctl_scaling;
2043 		}
2044 	} else {
2045 		cpu->pstate.scaling = perf_ctl_scaling;
2046 		cpu->pstate.max_pstate = pstate_funcs.get_max();
2047 		cpu->pstate.turbo_pstate = pstate_funcs.get_turbo();
2048 	}
2049 
2050 	if (cpu->pstate.scaling == perf_ctl_scaling) {
2051 		cpu->pstate.min_freq = cpu->pstate.min_pstate * perf_ctl_scaling;
2052 		cpu->pstate.max_freq = cpu->pstate.max_pstate * perf_ctl_scaling;
2053 		cpu->pstate.turbo_freq = cpu->pstate.turbo_pstate * perf_ctl_scaling;
2054 	}
2055 
2056 	if (pstate_funcs.get_aperf_mperf_shift)
2057 		cpu->aperf_mperf_shift = pstate_funcs.get_aperf_mperf_shift();
2058 
2059 	if (pstate_funcs.get_vid)
2060 		pstate_funcs.get_vid(cpu);
2061 
2062 	intel_pstate_set_min_pstate(cpu);
2063 }
2064 
2065 /*
2066  * Long hold time will keep high perf limits for long time,
2067  * which negatively impacts perf/watt for some workloads,
2068  * like specpower. 3ms is based on experiements on some
2069  * workoads.
2070  */
2071 static int hwp_boost_hold_time_ns = 3 * NSEC_PER_MSEC;
2072 
2073 static inline void intel_pstate_hwp_boost_up(struct cpudata *cpu)
2074 {
2075 	u64 hwp_req = READ_ONCE(cpu->hwp_req_cached);
2076 	u64 hwp_cap = READ_ONCE(cpu->hwp_cap_cached);
2077 	u32 max_limit = (hwp_req & 0xff00) >> 8;
2078 	u32 min_limit = (hwp_req & 0xff);
2079 	u32 boost_level1;
2080 
2081 	/*
2082 	 * Cases to consider (User changes via sysfs or boot time):
2083 	 * If, P0 (Turbo max) = P1 (Guaranteed max) = min:
2084 	 *	No boost, return.
2085 	 * If, P0 (Turbo max) > P1 (Guaranteed max) = min:
2086 	 *     Should result in one level boost only for P0.
2087 	 * If, P0 (Turbo max) = P1 (Guaranteed max) > min:
2088 	 *     Should result in two level boost:
2089 	 *         (min + p1)/2 and P1.
2090 	 * If, P0 (Turbo max) > P1 (Guaranteed max) > min:
2091 	 *     Should result in three level boost:
2092 	 *        (min + p1)/2, P1 and P0.
2093 	 */
2094 
2095 	/* If max and min are equal or already at max, nothing to boost */
2096 	if (max_limit == min_limit || cpu->hwp_boost_min >= max_limit)
2097 		return;
2098 
2099 	if (!cpu->hwp_boost_min)
2100 		cpu->hwp_boost_min = min_limit;
2101 
2102 	/* level at half way mark between min and guranteed */
2103 	boost_level1 = (HWP_GUARANTEED_PERF(hwp_cap) + min_limit) >> 1;
2104 
2105 	if (cpu->hwp_boost_min < boost_level1)
2106 		cpu->hwp_boost_min = boost_level1;
2107 	else if (cpu->hwp_boost_min < HWP_GUARANTEED_PERF(hwp_cap))
2108 		cpu->hwp_boost_min = HWP_GUARANTEED_PERF(hwp_cap);
2109 	else if (cpu->hwp_boost_min == HWP_GUARANTEED_PERF(hwp_cap) &&
2110 		 max_limit != HWP_GUARANTEED_PERF(hwp_cap))
2111 		cpu->hwp_boost_min = max_limit;
2112 	else
2113 		return;
2114 
2115 	hwp_req = (hwp_req & ~GENMASK_ULL(7, 0)) | cpu->hwp_boost_min;
2116 	wrmsrl(MSR_HWP_REQUEST, hwp_req);
2117 	cpu->last_update = cpu->sample.time;
2118 }
2119 
2120 static inline void intel_pstate_hwp_boost_down(struct cpudata *cpu)
2121 {
2122 	if (cpu->hwp_boost_min) {
2123 		bool expired;
2124 
2125 		/* Check if we are idle for hold time to boost down */
2126 		expired = time_after64(cpu->sample.time, cpu->last_update +
2127 				       hwp_boost_hold_time_ns);
2128 		if (expired) {
2129 			wrmsrl(MSR_HWP_REQUEST, cpu->hwp_req_cached);
2130 			cpu->hwp_boost_min = 0;
2131 		}
2132 	}
2133 	cpu->last_update = cpu->sample.time;
2134 }
2135 
2136 static inline void intel_pstate_update_util_hwp_local(struct cpudata *cpu,
2137 						      u64 time)
2138 {
2139 	cpu->sample.time = time;
2140 
2141 	if (cpu->sched_flags & SCHED_CPUFREQ_IOWAIT) {
2142 		bool do_io = false;
2143 
2144 		cpu->sched_flags = 0;
2145 		/*
2146 		 * Set iowait_boost flag and update time. Since IO WAIT flag
2147 		 * is set all the time, we can't just conclude that there is
2148 		 * some IO bound activity is scheduled on this CPU with just
2149 		 * one occurrence. If we receive at least two in two
2150 		 * consecutive ticks, then we treat as boost candidate.
2151 		 */
2152 		if (time_before64(time, cpu->last_io_update + 2 * TICK_NSEC))
2153 			do_io = true;
2154 
2155 		cpu->last_io_update = time;
2156 
2157 		if (do_io)
2158 			intel_pstate_hwp_boost_up(cpu);
2159 
2160 	} else {
2161 		intel_pstate_hwp_boost_down(cpu);
2162 	}
2163 }
2164 
2165 static inline void intel_pstate_update_util_hwp(struct update_util_data *data,
2166 						u64 time, unsigned int flags)
2167 {
2168 	struct cpudata *cpu = container_of(data, struct cpudata, update_util);
2169 
2170 	cpu->sched_flags |= flags;
2171 
2172 	if (smp_processor_id() == cpu->cpu)
2173 		intel_pstate_update_util_hwp_local(cpu, time);
2174 }
2175 
2176 static inline void intel_pstate_calc_avg_perf(struct cpudata *cpu)
2177 {
2178 	struct sample *sample = &cpu->sample;
2179 
2180 	sample->core_avg_perf = div_ext_fp(sample->aperf, sample->mperf);
2181 }
2182 
2183 static inline bool intel_pstate_sample(struct cpudata *cpu, u64 time)
2184 {
2185 	u64 aperf, mperf;
2186 	unsigned long flags;
2187 	u64 tsc;
2188 
2189 	local_irq_save(flags);
2190 	rdmsrl(MSR_IA32_APERF, aperf);
2191 	rdmsrl(MSR_IA32_MPERF, mperf);
2192 	tsc = rdtsc();
2193 	if (cpu->prev_mperf == mperf || cpu->prev_tsc == tsc) {
2194 		local_irq_restore(flags);
2195 		return false;
2196 	}
2197 	local_irq_restore(flags);
2198 
2199 	cpu->last_sample_time = cpu->sample.time;
2200 	cpu->sample.time = time;
2201 	cpu->sample.aperf = aperf;
2202 	cpu->sample.mperf = mperf;
2203 	cpu->sample.tsc =  tsc;
2204 	cpu->sample.aperf -= cpu->prev_aperf;
2205 	cpu->sample.mperf -= cpu->prev_mperf;
2206 	cpu->sample.tsc -= cpu->prev_tsc;
2207 
2208 	cpu->prev_aperf = aperf;
2209 	cpu->prev_mperf = mperf;
2210 	cpu->prev_tsc = tsc;
2211 	/*
2212 	 * First time this function is invoked in a given cycle, all of the
2213 	 * previous sample data fields are equal to zero or stale and they must
2214 	 * be populated with meaningful numbers for things to work, so assume
2215 	 * that sample.time will always be reset before setting the utilization
2216 	 * update hook and make the caller skip the sample then.
2217 	 */
2218 	if (cpu->last_sample_time) {
2219 		intel_pstate_calc_avg_perf(cpu);
2220 		return true;
2221 	}
2222 	return false;
2223 }
2224 
2225 static inline int32_t get_avg_frequency(struct cpudata *cpu)
2226 {
2227 	return mul_ext_fp(cpu->sample.core_avg_perf, cpu_khz);
2228 }
2229 
2230 static inline int32_t get_avg_pstate(struct cpudata *cpu)
2231 {
2232 	return mul_ext_fp(cpu->pstate.max_pstate_physical,
2233 			  cpu->sample.core_avg_perf);
2234 }
2235 
2236 static inline int32_t get_target_pstate(struct cpudata *cpu)
2237 {
2238 	struct sample *sample = &cpu->sample;
2239 	int32_t busy_frac;
2240 	int target, avg_pstate;
2241 
2242 	busy_frac = div_fp(sample->mperf << cpu->aperf_mperf_shift,
2243 			   sample->tsc);
2244 
2245 	if (busy_frac < cpu->iowait_boost)
2246 		busy_frac = cpu->iowait_boost;
2247 
2248 	sample->busy_scaled = busy_frac * 100;
2249 
2250 	target = global.no_turbo || global.turbo_disabled ?
2251 			cpu->pstate.max_pstate : cpu->pstate.turbo_pstate;
2252 	target += target >> 2;
2253 	target = mul_fp(target, busy_frac);
2254 	if (target < cpu->pstate.min_pstate)
2255 		target = cpu->pstate.min_pstate;
2256 
2257 	/*
2258 	 * If the average P-state during the previous cycle was higher than the
2259 	 * current target, add 50% of the difference to the target to reduce
2260 	 * possible performance oscillations and offset possible performance
2261 	 * loss related to moving the workload from one CPU to another within
2262 	 * a package/module.
2263 	 */
2264 	avg_pstate = get_avg_pstate(cpu);
2265 	if (avg_pstate > target)
2266 		target += (avg_pstate - target) >> 1;
2267 
2268 	return target;
2269 }
2270 
2271 static int intel_pstate_prepare_request(struct cpudata *cpu, int pstate)
2272 {
2273 	int min_pstate = max(cpu->pstate.min_pstate, cpu->min_perf_ratio);
2274 	int max_pstate = max(min_pstate, cpu->max_perf_ratio);
2275 
2276 	return clamp_t(int, pstate, min_pstate, max_pstate);
2277 }
2278 
2279 static void intel_pstate_update_pstate(struct cpudata *cpu, int pstate)
2280 {
2281 	if (pstate == cpu->pstate.current_pstate)
2282 		return;
2283 
2284 	cpu->pstate.current_pstate = pstate;
2285 	wrmsrl(MSR_IA32_PERF_CTL, pstate_funcs.get_val(cpu, pstate));
2286 }
2287 
2288 static void intel_pstate_adjust_pstate(struct cpudata *cpu)
2289 {
2290 	int from = cpu->pstate.current_pstate;
2291 	struct sample *sample;
2292 	int target_pstate;
2293 
2294 	update_turbo_state();
2295 
2296 	target_pstate = get_target_pstate(cpu);
2297 	target_pstate = intel_pstate_prepare_request(cpu, target_pstate);
2298 	trace_cpu_frequency(target_pstate * cpu->pstate.scaling, cpu->cpu);
2299 	intel_pstate_update_pstate(cpu, target_pstate);
2300 
2301 	sample = &cpu->sample;
2302 	trace_pstate_sample(mul_ext_fp(100, sample->core_avg_perf),
2303 		fp_toint(sample->busy_scaled),
2304 		from,
2305 		cpu->pstate.current_pstate,
2306 		sample->mperf,
2307 		sample->aperf,
2308 		sample->tsc,
2309 		get_avg_frequency(cpu),
2310 		fp_toint(cpu->iowait_boost * 100));
2311 }
2312 
2313 static void intel_pstate_update_util(struct update_util_data *data, u64 time,
2314 				     unsigned int flags)
2315 {
2316 	struct cpudata *cpu = container_of(data, struct cpudata, update_util);
2317 	u64 delta_ns;
2318 
2319 	/* Don't allow remote callbacks */
2320 	if (smp_processor_id() != cpu->cpu)
2321 		return;
2322 
2323 	delta_ns = time - cpu->last_update;
2324 	if (flags & SCHED_CPUFREQ_IOWAIT) {
2325 		/* Start over if the CPU may have been idle. */
2326 		if (delta_ns > TICK_NSEC) {
2327 			cpu->iowait_boost = ONE_EIGHTH_FP;
2328 		} else if (cpu->iowait_boost >= ONE_EIGHTH_FP) {
2329 			cpu->iowait_boost <<= 1;
2330 			if (cpu->iowait_boost > int_tofp(1))
2331 				cpu->iowait_boost = int_tofp(1);
2332 		} else {
2333 			cpu->iowait_boost = ONE_EIGHTH_FP;
2334 		}
2335 	} else if (cpu->iowait_boost) {
2336 		/* Clear iowait_boost if the CPU may have been idle. */
2337 		if (delta_ns > TICK_NSEC)
2338 			cpu->iowait_boost = 0;
2339 		else
2340 			cpu->iowait_boost >>= 1;
2341 	}
2342 	cpu->last_update = time;
2343 	delta_ns = time - cpu->sample.time;
2344 	if ((s64)delta_ns < INTEL_PSTATE_SAMPLING_INTERVAL)
2345 		return;
2346 
2347 	if (intel_pstate_sample(cpu, time))
2348 		intel_pstate_adjust_pstate(cpu);
2349 }
2350 
2351 static struct pstate_funcs core_funcs = {
2352 	.get_max = core_get_max_pstate,
2353 	.get_max_physical = core_get_max_pstate_physical,
2354 	.get_min = core_get_min_pstate,
2355 	.get_turbo = core_get_turbo_pstate,
2356 	.get_scaling = core_get_scaling,
2357 	.get_val = core_get_val,
2358 };
2359 
2360 static const struct pstate_funcs silvermont_funcs = {
2361 	.get_max = atom_get_max_pstate,
2362 	.get_max_physical = atom_get_max_pstate,
2363 	.get_min = atom_get_min_pstate,
2364 	.get_turbo = atom_get_turbo_pstate,
2365 	.get_val = atom_get_val,
2366 	.get_scaling = silvermont_get_scaling,
2367 	.get_vid = atom_get_vid,
2368 };
2369 
2370 static const struct pstate_funcs airmont_funcs = {
2371 	.get_max = atom_get_max_pstate,
2372 	.get_max_physical = atom_get_max_pstate,
2373 	.get_min = atom_get_min_pstate,
2374 	.get_turbo = atom_get_turbo_pstate,
2375 	.get_val = atom_get_val,
2376 	.get_scaling = airmont_get_scaling,
2377 	.get_vid = atom_get_vid,
2378 };
2379 
2380 static const struct pstate_funcs knl_funcs = {
2381 	.get_max = core_get_max_pstate,
2382 	.get_max_physical = core_get_max_pstate_physical,
2383 	.get_min = core_get_min_pstate,
2384 	.get_turbo = knl_get_turbo_pstate,
2385 	.get_aperf_mperf_shift = knl_get_aperf_mperf_shift,
2386 	.get_scaling = core_get_scaling,
2387 	.get_val = core_get_val,
2388 };
2389 
2390 #define X86_MATCH(model, policy)					 \
2391 	X86_MATCH_VENDOR_FAM_MODEL_FEATURE(INTEL, 6, INTEL_FAM6_##model, \
2392 					   X86_FEATURE_APERFMPERF, &policy)
2393 
2394 static const struct x86_cpu_id intel_pstate_cpu_ids[] = {
2395 	X86_MATCH(SANDYBRIDGE,		core_funcs),
2396 	X86_MATCH(SANDYBRIDGE_X,	core_funcs),
2397 	X86_MATCH(ATOM_SILVERMONT,	silvermont_funcs),
2398 	X86_MATCH(IVYBRIDGE,		core_funcs),
2399 	X86_MATCH(HASWELL,		core_funcs),
2400 	X86_MATCH(BROADWELL,		core_funcs),
2401 	X86_MATCH(IVYBRIDGE_X,		core_funcs),
2402 	X86_MATCH(HASWELL_X,		core_funcs),
2403 	X86_MATCH(HASWELL_L,		core_funcs),
2404 	X86_MATCH(HASWELL_G,		core_funcs),
2405 	X86_MATCH(BROADWELL_G,		core_funcs),
2406 	X86_MATCH(ATOM_AIRMONT,		airmont_funcs),
2407 	X86_MATCH(SKYLAKE_L,		core_funcs),
2408 	X86_MATCH(BROADWELL_X,		core_funcs),
2409 	X86_MATCH(SKYLAKE,		core_funcs),
2410 	X86_MATCH(BROADWELL_D,		core_funcs),
2411 	X86_MATCH(XEON_PHI_KNL,		knl_funcs),
2412 	X86_MATCH(XEON_PHI_KNM,		knl_funcs),
2413 	X86_MATCH(ATOM_GOLDMONT,	core_funcs),
2414 	X86_MATCH(ATOM_GOLDMONT_PLUS,	core_funcs),
2415 	X86_MATCH(SKYLAKE_X,		core_funcs),
2416 	X86_MATCH(COMETLAKE,		core_funcs),
2417 	X86_MATCH(ICELAKE_X,		core_funcs),
2418 	{}
2419 };
2420 MODULE_DEVICE_TABLE(x86cpu, intel_pstate_cpu_ids);
2421 
2422 static const struct x86_cpu_id intel_pstate_cpu_oob_ids[] __initconst = {
2423 	X86_MATCH(BROADWELL_D,		core_funcs),
2424 	X86_MATCH(BROADWELL_X,		core_funcs),
2425 	X86_MATCH(SKYLAKE_X,		core_funcs),
2426 	X86_MATCH(ICELAKE_X,		core_funcs),
2427 	{}
2428 };
2429 
2430 static const struct x86_cpu_id intel_pstate_cpu_ee_disable_ids[] = {
2431 	X86_MATCH(KABYLAKE,		core_funcs),
2432 	{}
2433 };
2434 
2435 static const struct x86_cpu_id intel_pstate_hwp_boost_ids[] = {
2436 	X86_MATCH(SKYLAKE_X,		core_funcs),
2437 	X86_MATCH(SKYLAKE,		core_funcs),
2438 	{}
2439 };
2440 
2441 static int intel_pstate_init_cpu(unsigned int cpunum)
2442 {
2443 	struct cpudata *cpu;
2444 
2445 	cpu = all_cpu_data[cpunum];
2446 
2447 	if (!cpu) {
2448 		cpu = kzalloc(sizeof(*cpu), GFP_KERNEL);
2449 		if (!cpu)
2450 			return -ENOMEM;
2451 
2452 		WRITE_ONCE(all_cpu_data[cpunum], cpu);
2453 
2454 		cpu->cpu = cpunum;
2455 
2456 		cpu->epp_default = -EINVAL;
2457 
2458 		if (hwp_active) {
2459 			const struct x86_cpu_id *id;
2460 
2461 			intel_pstate_hwp_enable(cpu);
2462 
2463 			id = x86_match_cpu(intel_pstate_hwp_boost_ids);
2464 			if (id && intel_pstate_acpi_pm_profile_server())
2465 				hwp_boost = true;
2466 		}
2467 	} else if (hwp_active) {
2468 		/*
2469 		 * Re-enable HWP in case this happens after a resume from ACPI
2470 		 * S3 if the CPU was offline during the whole system/resume
2471 		 * cycle.
2472 		 */
2473 		intel_pstate_hwp_reenable(cpu);
2474 	}
2475 
2476 	cpu->epp_powersave = -EINVAL;
2477 	cpu->epp_policy = 0;
2478 
2479 	intel_pstate_get_cpu_pstates(cpu);
2480 
2481 	pr_debug("controlling: cpu %d\n", cpunum);
2482 
2483 	return 0;
2484 }
2485 
2486 static void intel_pstate_set_update_util_hook(unsigned int cpu_num)
2487 {
2488 	struct cpudata *cpu = all_cpu_data[cpu_num];
2489 
2490 	if (hwp_active && !hwp_boost)
2491 		return;
2492 
2493 	if (cpu->update_util_set)
2494 		return;
2495 
2496 	/* Prevent intel_pstate_update_util() from using stale data. */
2497 	cpu->sample.time = 0;
2498 	cpufreq_add_update_util_hook(cpu_num, &cpu->update_util,
2499 				     (hwp_active ?
2500 				      intel_pstate_update_util_hwp :
2501 				      intel_pstate_update_util));
2502 	cpu->update_util_set = true;
2503 }
2504 
2505 static void intel_pstate_clear_update_util_hook(unsigned int cpu)
2506 {
2507 	struct cpudata *cpu_data = all_cpu_data[cpu];
2508 
2509 	if (!cpu_data->update_util_set)
2510 		return;
2511 
2512 	cpufreq_remove_update_util_hook(cpu);
2513 	cpu_data->update_util_set = false;
2514 	synchronize_rcu();
2515 }
2516 
2517 static int intel_pstate_get_max_freq(struct cpudata *cpu)
2518 {
2519 	return global.turbo_disabled || global.no_turbo ?
2520 			cpu->pstate.max_freq : cpu->pstate.turbo_freq;
2521 }
2522 
2523 static void intel_pstate_update_perf_limits(struct cpudata *cpu,
2524 					    unsigned int policy_min,
2525 					    unsigned int policy_max)
2526 {
2527 	int perf_ctl_scaling = cpu->pstate.perf_ctl_scaling;
2528 	int32_t max_policy_perf, min_policy_perf;
2529 
2530 	max_policy_perf = policy_max / perf_ctl_scaling;
2531 	if (policy_max == policy_min) {
2532 		min_policy_perf = max_policy_perf;
2533 	} else {
2534 		min_policy_perf = policy_min / perf_ctl_scaling;
2535 		min_policy_perf = clamp_t(int32_t, min_policy_perf,
2536 					  0, max_policy_perf);
2537 	}
2538 
2539 	/*
2540 	 * HWP needs some special consideration, because HWP_REQUEST uses
2541 	 * abstract values to represent performance rather than pure ratios.
2542 	 */
2543 	if (hwp_active && cpu->pstate.scaling != perf_ctl_scaling) {
2544 		int scaling = cpu->pstate.scaling;
2545 		int freq;
2546 
2547 		freq = max_policy_perf * perf_ctl_scaling;
2548 		max_policy_perf = DIV_ROUND_UP(freq, scaling);
2549 		freq = min_policy_perf * perf_ctl_scaling;
2550 		min_policy_perf = DIV_ROUND_UP(freq, scaling);
2551 	}
2552 
2553 	pr_debug("cpu:%d min_policy_perf:%d max_policy_perf:%d\n",
2554 		 cpu->cpu, min_policy_perf, max_policy_perf);
2555 
2556 	/* Normalize user input to [min_perf, max_perf] */
2557 	if (per_cpu_limits) {
2558 		cpu->min_perf_ratio = min_policy_perf;
2559 		cpu->max_perf_ratio = max_policy_perf;
2560 	} else {
2561 		int turbo_max = cpu->pstate.turbo_pstate;
2562 		int32_t global_min, global_max;
2563 
2564 		/* Global limits are in percent of the maximum turbo P-state. */
2565 		global_max = DIV_ROUND_UP(turbo_max * global.max_perf_pct, 100);
2566 		global_min = DIV_ROUND_UP(turbo_max * global.min_perf_pct, 100);
2567 		global_min = clamp_t(int32_t, global_min, 0, global_max);
2568 
2569 		pr_debug("cpu:%d global_min:%d global_max:%d\n", cpu->cpu,
2570 			 global_min, global_max);
2571 
2572 		cpu->min_perf_ratio = max(min_policy_perf, global_min);
2573 		cpu->min_perf_ratio = min(cpu->min_perf_ratio, max_policy_perf);
2574 		cpu->max_perf_ratio = min(max_policy_perf, global_max);
2575 		cpu->max_perf_ratio = max(min_policy_perf, cpu->max_perf_ratio);
2576 
2577 		/* Make sure min_perf <= max_perf */
2578 		cpu->min_perf_ratio = min(cpu->min_perf_ratio,
2579 					  cpu->max_perf_ratio);
2580 
2581 	}
2582 	pr_debug("cpu:%d max_perf_ratio:%d min_perf_ratio:%d\n", cpu->cpu,
2583 		 cpu->max_perf_ratio,
2584 		 cpu->min_perf_ratio);
2585 }
2586 
2587 static int intel_pstate_set_policy(struct cpufreq_policy *policy)
2588 {
2589 	struct cpudata *cpu;
2590 
2591 	if (!policy->cpuinfo.max_freq)
2592 		return -ENODEV;
2593 
2594 	pr_debug("set_policy cpuinfo.max %u policy->max %u\n",
2595 		 policy->cpuinfo.max_freq, policy->max);
2596 
2597 	cpu = all_cpu_data[policy->cpu];
2598 	cpu->policy = policy->policy;
2599 
2600 	mutex_lock(&intel_pstate_limits_lock);
2601 
2602 	intel_pstate_update_perf_limits(cpu, policy->min, policy->max);
2603 
2604 	if (cpu->policy == CPUFREQ_POLICY_PERFORMANCE) {
2605 		/*
2606 		 * NOHZ_FULL CPUs need this as the governor callback may not
2607 		 * be invoked on them.
2608 		 */
2609 		intel_pstate_clear_update_util_hook(policy->cpu);
2610 		intel_pstate_max_within_limits(cpu);
2611 	} else {
2612 		intel_pstate_set_update_util_hook(policy->cpu);
2613 	}
2614 
2615 	if (hwp_active) {
2616 		/*
2617 		 * When hwp_boost was active before and dynamically it
2618 		 * was turned off, in that case we need to clear the
2619 		 * update util hook.
2620 		 */
2621 		if (!hwp_boost)
2622 			intel_pstate_clear_update_util_hook(policy->cpu);
2623 		intel_pstate_hwp_set(policy->cpu);
2624 	}
2625 
2626 	mutex_unlock(&intel_pstate_limits_lock);
2627 
2628 	return 0;
2629 }
2630 
2631 static void intel_pstate_adjust_policy_max(struct cpudata *cpu,
2632 					   struct cpufreq_policy_data *policy)
2633 {
2634 	if (!hwp_active &&
2635 	    cpu->pstate.max_pstate_physical > cpu->pstate.max_pstate &&
2636 	    policy->max < policy->cpuinfo.max_freq &&
2637 	    policy->max > cpu->pstate.max_freq) {
2638 		pr_debug("policy->max > max non turbo frequency\n");
2639 		policy->max = policy->cpuinfo.max_freq;
2640 	}
2641 }
2642 
2643 static void intel_pstate_verify_cpu_policy(struct cpudata *cpu,
2644 					   struct cpufreq_policy_data *policy)
2645 {
2646 	int max_freq;
2647 
2648 	update_turbo_state();
2649 	if (hwp_active) {
2650 		intel_pstate_get_hwp_cap(cpu);
2651 		max_freq = global.no_turbo || global.turbo_disabled ?
2652 				cpu->pstate.max_freq : cpu->pstate.turbo_freq;
2653 	} else {
2654 		max_freq = intel_pstate_get_max_freq(cpu);
2655 	}
2656 	cpufreq_verify_within_limits(policy, policy->cpuinfo.min_freq, max_freq);
2657 
2658 	intel_pstate_adjust_policy_max(cpu, policy);
2659 }
2660 
2661 static int intel_pstate_verify_policy(struct cpufreq_policy_data *policy)
2662 {
2663 	intel_pstate_verify_cpu_policy(all_cpu_data[policy->cpu], policy);
2664 
2665 	return 0;
2666 }
2667 
2668 static int intel_cpufreq_cpu_offline(struct cpufreq_policy *policy)
2669 {
2670 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2671 
2672 	pr_debug("CPU %d going offline\n", cpu->cpu);
2673 
2674 	if (cpu->suspended)
2675 		return 0;
2676 
2677 	/*
2678 	 * If the CPU is an SMT thread and it goes offline with the performance
2679 	 * settings different from the minimum, it will prevent its sibling
2680 	 * from getting to lower performance levels, so force the minimum
2681 	 * performance on CPU offline to prevent that from happening.
2682 	 */
2683 	if (hwp_active)
2684 		intel_pstate_hwp_offline(cpu);
2685 	else
2686 		intel_pstate_set_min_pstate(cpu);
2687 
2688 	intel_pstate_exit_perf_limits(policy);
2689 
2690 	return 0;
2691 }
2692 
2693 static int intel_pstate_cpu_online(struct cpufreq_policy *policy)
2694 {
2695 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2696 
2697 	pr_debug("CPU %d going online\n", cpu->cpu);
2698 
2699 	intel_pstate_init_acpi_perf_limits(policy);
2700 
2701 	if (hwp_active) {
2702 		/*
2703 		 * Re-enable HWP and clear the "suspended" flag to let "resume"
2704 		 * know that it need not do that.
2705 		 */
2706 		intel_pstate_hwp_reenable(cpu);
2707 		cpu->suspended = false;
2708 	}
2709 
2710 	return 0;
2711 }
2712 
2713 static int intel_pstate_cpu_offline(struct cpufreq_policy *policy)
2714 {
2715 	intel_pstate_clear_update_util_hook(policy->cpu);
2716 
2717 	return intel_cpufreq_cpu_offline(policy);
2718 }
2719 
2720 static int intel_pstate_cpu_exit(struct cpufreq_policy *policy)
2721 {
2722 	pr_debug("CPU %d exiting\n", policy->cpu);
2723 
2724 	policy->fast_switch_possible = false;
2725 
2726 	return 0;
2727 }
2728 
2729 static int __intel_pstate_cpu_init(struct cpufreq_policy *policy)
2730 {
2731 	struct cpudata *cpu;
2732 	int rc;
2733 
2734 	rc = intel_pstate_init_cpu(policy->cpu);
2735 	if (rc)
2736 		return rc;
2737 
2738 	cpu = all_cpu_data[policy->cpu];
2739 
2740 	cpu->max_perf_ratio = 0xFF;
2741 	cpu->min_perf_ratio = 0;
2742 
2743 	/* cpuinfo and default policy values */
2744 	policy->cpuinfo.min_freq = cpu->pstate.min_freq;
2745 	update_turbo_state();
2746 	global.turbo_disabled_mf = global.turbo_disabled;
2747 	policy->cpuinfo.max_freq = global.turbo_disabled ?
2748 			cpu->pstate.max_freq : cpu->pstate.turbo_freq;
2749 
2750 	policy->min = policy->cpuinfo.min_freq;
2751 	policy->max = policy->cpuinfo.max_freq;
2752 
2753 	intel_pstate_init_acpi_perf_limits(policy);
2754 
2755 	policy->fast_switch_possible = true;
2756 
2757 	return 0;
2758 }
2759 
2760 static int intel_pstate_cpu_init(struct cpufreq_policy *policy)
2761 {
2762 	int ret = __intel_pstate_cpu_init(policy);
2763 
2764 	if (ret)
2765 		return ret;
2766 
2767 	/*
2768 	 * Set the policy to powersave to provide a valid fallback value in case
2769 	 * the default cpufreq governor is neither powersave nor performance.
2770 	 */
2771 	policy->policy = CPUFREQ_POLICY_POWERSAVE;
2772 
2773 	if (hwp_active) {
2774 		struct cpudata *cpu = all_cpu_data[policy->cpu];
2775 
2776 		cpu->epp_cached = intel_pstate_get_epp(cpu, 0);
2777 	}
2778 
2779 	return 0;
2780 }
2781 
2782 static struct cpufreq_driver intel_pstate = {
2783 	.flags		= CPUFREQ_CONST_LOOPS,
2784 	.verify		= intel_pstate_verify_policy,
2785 	.setpolicy	= intel_pstate_set_policy,
2786 	.suspend	= intel_pstate_suspend,
2787 	.resume		= intel_pstate_resume,
2788 	.init		= intel_pstate_cpu_init,
2789 	.exit		= intel_pstate_cpu_exit,
2790 	.offline	= intel_pstate_cpu_offline,
2791 	.online		= intel_pstate_cpu_online,
2792 	.update_limits	= intel_pstate_update_limits,
2793 	.name		= "intel_pstate",
2794 };
2795 
2796 static int intel_cpufreq_verify_policy(struct cpufreq_policy_data *policy)
2797 {
2798 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2799 
2800 	intel_pstate_verify_cpu_policy(cpu, policy);
2801 	intel_pstate_update_perf_limits(cpu, policy->min, policy->max);
2802 
2803 	return 0;
2804 }
2805 
2806 /* Use of trace in passive mode:
2807  *
2808  * In passive mode the trace core_busy field (also known as the
2809  * performance field, and lablelled as such on the graphs; also known as
2810  * core_avg_perf) is not needed and so is re-assigned to indicate if the
2811  * driver call was via the normal or fast switch path. Various graphs
2812  * output from the intel_pstate_tracer.py utility that include core_busy
2813  * (or performance or core_avg_perf) have a fixed y-axis from 0 to 100%,
2814  * so we use 10 to indicate the normal path through the driver, and
2815  * 90 to indicate the fast switch path through the driver.
2816  * The scaled_busy field is not used, and is set to 0.
2817  */
2818 
2819 #define	INTEL_PSTATE_TRACE_TARGET 10
2820 #define	INTEL_PSTATE_TRACE_FAST_SWITCH 90
2821 
2822 static void intel_cpufreq_trace(struct cpudata *cpu, unsigned int trace_type, int old_pstate)
2823 {
2824 	struct sample *sample;
2825 
2826 	if (!trace_pstate_sample_enabled())
2827 		return;
2828 
2829 	if (!intel_pstate_sample(cpu, ktime_get()))
2830 		return;
2831 
2832 	sample = &cpu->sample;
2833 	trace_pstate_sample(trace_type,
2834 		0,
2835 		old_pstate,
2836 		cpu->pstate.current_pstate,
2837 		sample->mperf,
2838 		sample->aperf,
2839 		sample->tsc,
2840 		get_avg_frequency(cpu),
2841 		fp_toint(cpu->iowait_boost * 100));
2842 }
2843 
2844 static void intel_cpufreq_hwp_update(struct cpudata *cpu, u32 min, u32 max,
2845 				     u32 desired, bool fast_switch)
2846 {
2847 	u64 prev = READ_ONCE(cpu->hwp_req_cached), value = prev;
2848 
2849 	value &= ~HWP_MIN_PERF(~0L);
2850 	value |= HWP_MIN_PERF(min);
2851 
2852 	value &= ~HWP_MAX_PERF(~0L);
2853 	value |= HWP_MAX_PERF(max);
2854 
2855 	value &= ~HWP_DESIRED_PERF(~0L);
2856 	value |= HWP_DESIRED_PERF(desired);
2857 
2858 	if (value == prev)
2859 		return;
2860 
2861 	WRITE_ONCE(cpu->hwp_req_cached, value);
2862 	if (fast_switch)
2863 		wrmsrl(MSR_HWP_REQUEST, value);
2864 	else
2865 		wrmsrl_on_cpu(cpu->cpu, MSR_HWP_REQUEST, value);
2866 }
2867 
2868 static void intel_cpufreq_perf_ctl_update(struct cpudata *cpu,
2869 					  u32 target_pstate, bool fast_switch)
2870 {
2871 	if (fast_switch)
2872 		wrmsrl(MSR_IA32_PERF_CTL,
2873 		       pstate_funcs.get_val(cpu, target_pstate));
2874 	else
2875 		wrmsrl_on_cpu(cpu->cpu, MSR_IA32_PERF_CTL,
2876 			      pstate_funcs.get_val(cpu, target_pstate));
2877 }
2878 
2879 static int intel_cpufreq_update_pstate(struct cpufreq_policy *policy,
2880 				       int target_pstate, bool fast_switch)
2881 {
2882 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2883 	int old_pstate = cpu->pstate.current_pstate;
2884 
2885 	target_pstate = intel_pstate_prepare_request(cpu, target_pstate);
2886 	if (hwp_active) {
2887 		int max_pstate = policy->strict_target ?
2888 					target_pstate : cpu->max_perf_ratio;
2889 
2890 		intel_cpufreq_hwp_update(cpu, target_pstate, max_pstate, 0,
2891 					 fast_switch);
2892 	} else if (target_pstate != old_pstate) {
2893 		intel_cpufreq_perf_ctl_update(cpu, target_pstate, fast_switch);
2894 	}
2895 
2896 	cpu->pstate.current_pstate = target_pstate;
2897 
2898 	intel_cpufreq_trace(cpu, fast_switch ? INTEL_PSTATE_TRACE_FAST_SWITCH :
2899 			    INTEL_PSTATE_TRACE_TARGET, old_pstate);
2900 
2901 	return target_pstate;
2902 }
2903 
2904 static int intel_cpufreq_target(struct cpufreq_policy *policy,
2905 				unsigned int target_freq,
2906 				unsigned int relation)
2907 {
2908 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2909 	struct cpufreq_freqs freqs;
2910 	int target_pstate;
2911 
2912 	update_turbo_state();
2913 
2914 	freqs.old = policy->cur;
2915 	freqs.new = target_freq;
2916 
2917 	cpufreq_freq_transition_begin(policy, &freqs);
2918 
2919 	switch (relation) {
2920 	case CPUFREQ_RELATION_L:
2921 		target_pstate = DIV_ROUND_UP(freqs.new, cpu->pstate.scaling);
2922 		break;
2923 	case CPUFREQ_RELATION_H:
2924 		target_pstate = freqs.new / cpu->pstate.scaling;
2925 		break;
2926 	default:
2927 		target_pstate = DIV_ROUND_CLOSEST(freqs.new, cpu->pstate.scaling);
2928 		break;
2929 	}
2930 
2931 	target_pstate = intel_cpufreq_update_pstate(policy, target_pstate, false);
2932 
2933 	freqs.new = target_pstate * cpu->pstate.scaling;
2934 
2935 	cpufreq_freq_transition_end(policy, &freqs, false);
2936 
2937 	return 0;
2938 }
2939 
2940 static unsigned int intel_cpufreq_fast_switch(struct cpufreq_policy *policy,
2941 					      unsigned int target_freq)
2942 {
2943 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2944 	int target_pstate;
2945 
2946 	update_turbo_state();
2947 
2948 	target_pstate = DIV_ROUND_UP(target_freq, cpu->pstate.scaling);
2949 
2950 	target_pstate = intel_cpufreq_update_pstate(policy, target_pstate, true);
2951 
2952 	return target_pstate * cpu->pstate.scaling;
2953 }
2954 
2955 static void intel_cpufreq_adjust_perf(unsigned int cpunum,
2956 				      unsigned long min_perf,
2957 				      unsigned long target_perf,
2958 				      unsigned long capacity)
2959 {
2960 	struct cpudata *cpu = all_cpu_data[cpunum];
2961 	u64 hwp_cap = READ_ONCE(cpu->hwp_cap_cached);
2962 	int old_pstate = cpu->pstate.current_pstate;
2963 	int cap_pstate, min_pstate, max_pstate, target_pstate;
2964 
2965 	update_turbo_state();
2966 	cap_pstate = global.turbo_disabled ? HWP_GUARANTEED_PERF(hwp_cap) :
2967 					     HWP_HIGHEST_PERF(hwp_cap);
2968 
2969 	/* Optimization: Avoid unnecessary divisions. */
2970 
2971 	target_pstate = cap_pstate;
2972 	if (target_perf < capacity)
2973 		target_pstate = DIV_ROUND_UP(cap_pstate * target_perf, capacity);
2974 
2975 	min_pstate = cap_pstate;
2976 	if (min_perf < capacity)
2977 		min_pstate = DIV_ROUND_UP(cap_pstate * min_perf, capacity);
2978 
2979 	if (min_pstate < cpu->pstate.min_pstate)
2980 		min_pstate = cpu->pstate.min_pstate;
2981 
2982 	if (min_pstate < cpu->min_perf_ratio)
2983 		min_pstate = cpu->min_perf_ratio;
2984 
2985 	max_pstate = min(cap_pstate, cpu->max_perf_ratio);
2986 	if (max_pstate < min_pstate)
2987 		max_pstate = min_pstate;
2988 
2989 	target_pstate = clamp_t(int, target_pstate, min_pstate, max_pstate);
2990 
2991 	intel_cpufreq_hwp_update(cpu, min_pstate, max_pstate, target_pstate, true);
2992 
2993 	cpu->pstate.current_pstate = target_pstate;
2994 	intel_cpufreq_trace(cpu, INTEL_PSTATE_TRACE_FAST_SWITCH, old_pstate);
2995 }
2996 
2997 static int intel_cpufreq_cpu_init(struct cpufreq_policy *policy)
2998 {
2999 	struct freq_qos_request *req;
3000 	struct cpudata *cpu;
3001 	struct device *dev;
3002 	int ret, freq;
3003 
3004 	dev = get_cpu_device(policy->cpu);
3005 	if (!dev)
3006 		return -ENODEV;
3007 
3008 	ret = __intel_pstate_cpu_init(policy);
3009 	if (ret)
3010 		return ret;
3011 
3012 	policy->cpuinfo.transition_latency = INTEL_CPUFREQ_TRANSITION_LATENCY;
3013 	/* This reflects the intel_pstate_get_cpu_pstates() setting. */
3014 	policy->cur = policy->cpuinfo.min_freq;
3015 
3016 	req = kcalloc(2, sizeof(*req), GFP_KERNEL);
3017 	if (!req) {
3018 		ret = -ENOMEM;
3019 		goto pstate_exit;
3020 	}
3021 
3022 	cpu = all_cpu_data[policy->cpu];
3023 
3024 	if (hwp_active) {
3025 		u64 value;
3026 
3027 		policy->transition_delay_us = INTEL_CPUFREQ_TRANSITION_DELAY_HWP;
3028 
3029 		intel_pstate_get_hwp_cap(cpu);
3030 
3031 		rdmsrl_on_cpu(cpu->cpu, MSR_HWP_REQUEST, &value);
3032 		WRITE_ONCE(cpu->hwp_req_cached, value);
3033 
3034 		cpu->epp_cached = intel_pstate_get_epp(cpu, value);
3035 	} else {
3036 		policy->transition_delay_us = INTEL_CPUFREQ_TRANSITION_DELAY;
3037 	}
3038 
3039 	freq = DIV_ROUND_UP(cpu->pstate.turbo_freq * global.min_perf_pct, 100);
3040 
3041 	ret = freq_qos_add_request(&policy->constraints, req, FREQ_QOS_MIN,
3042 				   freq);
3043 	if (ret < 0) {
3044 		dev_err(dev, "Failed to add min-freq constraint (%d)\n", ret);
3045 		goto free_req;
3046 	}
3047 
3048 	freq = DIV_ROUND_UP(cpu->pstate.turbo_freq * global.max_perf_pct, 100);
3049 
3050 	ret = freq_qos_add_request(&policy->constraints, req + 1, FREQ_QOS_MAX,
3051 				   freq);
3052 	if (ret < 0) {
3053 		dev_err(dev, "Failed to add max-freq constraint (%d)\n", ret);
3054 		goto remove_min_req;
3055 	}
3056 
3057 	policy->driver_data = req;
3058 
3059 	return 0;
3060 
3061 remove_min_req:
3062 	freq_qos_remove_request(req);
3063 free_req:
3064 	kfree(req);
3065 pstate_exit:
3066 	intel_pstate_exit_perf_limits(policy);
3067 
3068 	return ret;
3069 }
3070 
3071 static int intel_cpufreq_cpu_exit(struct cpufreq_policy *policy)
3072 {
3073 	struct freq_qos_request *req;
3074 
3075 	req = policy->driver_data;
3076 
3077 	freq_qos_remove_request(req + 1);
3078 	freq_qos_remove_request(req);
3079 	kfree(req);
3080 
3081 	return intel_pstate_cpu_exit(policy);
3082 }
3083 
3084 static int intel_cpufreq_suspend(struct cpufreq_policy *policy)
3085 {
3086 	intel_pstate_suspend(policy);
3087 
3088 	if (hwp_active) {
3089 		struct cpudata *cpu = all_cpu_data[policy->cpu];
3090 		u64 value = READ_ONCE(cpu->hwp_req_cached);
3091 
3092 		/*
3093 		 * Clear the desired perf field in MSR_HWP_REQUEST in case
3094 		 * intel_cpufreq_adjust_perf() is in use and the last value
3095 		 * written by it may not be suitable.
3096 		 */
3097 		value &= ~HWP_DESIRED_PERF(~0L);
3098 		wrmsrl_on_cpu(cpu->cpu, MSR_HWP_REQUEST, value);
3099 		WRITE_ONCE(cpu->hwp_req_cached, value);
3100 	}
3101 
3102 	return 0;
3103 }
3104 
3105 static struct cpufreq_driver intel_cpufreq = {
3106 	.flags		= CPUFREQ_CONST_LOOPS,
3107 	.verify		= intel_cpufreq_verify_policy,
3108 	.target		= intel_cpufreq_target,
3109 	.fast_switch	= intel_cpufreq_fast_switch,
3110 	.init		= intel_cpufreq_cpu_init,
3111 	.exit		= intel_cpufreq_cpu_exit,
3112 	.offline	= intel_cpufreq_cpu_offline,
3113 	.online		= intel_pstate_cpu_online,
3114 	.suspend	= intel_cpufreq_suspend,
3115 	.resume		= intel_pstate_resume,
3116 	.update_limits	= intel_pstate_update_limits,
3117 	.name		= "intel_cpufreq",
3118 };
3119 
3120 static struct cpufreq_driver *default_driver;
3121 
3122 static void intel_pstate_driver_cleanup(void)
3123 {
3124 	unsigned int cpu;
3125 
3126 	cpus_read_lock();
3127 	for_each_online_cpu(cpu) {
3128 		if (all_cpu_data[cpu]) {
3129 			if (intel_pstate_driver == &intel_pstate)
3130 				intel_pstate_clear_update_util_hook(cpu);
3131 
3132 			spin_lock(&hwp_notify_lock);
3133 			kfree(all_cpu_data[cpu]);
3134 			WRITE_ONCE(all_cpu_data[cpu], NULL);
3135 			spin_unlock(&hwp_notify_lock);
3136 		}
3137 	}
3138 	cpus_read_unlock();
3139 
3140 	intel_pstate_driver = NULL;
3141 }
3142 
3143 static int intel_pstate_register_driver(struct cpufreq_driver *driver)
3144 {
3145 	int ret;
3146 
3147 	if (driver == &intel_pstate)
3148 		intel_pstate_sysfs_expose_hwp_dynamic_boost();
3149 
3150 	memset(&global, 0, sizeof(global));
3151 	global.max_perf_pct = 100;
3152 
3153 	intel_pstate_driver = driver;
3154 	ret = cpufreq_register_driver(intel_pstate_driver);
3155 	if (ret) {
3156 		intel_pstate_driver_cleanup();
3157 		return ret;
3158 	}
3159 
3160 	global.min_perf_pct = min_perf_pct_min();
3161 
3162 	return 0;
3163 }
3164 
3165 static ssize_t intel_pstate_show_status(char *buf)
3166 {
3167 	if (!intel_pstate_driver)
3168 		return sprintf(buf, "off\n");
3169 
3170 	return sprintf(buf, "%s\n", intel_pstate_driver == &intel_pstate ?
3171 					"active" : "passive");
3172 }
3173 
3174 static int intel_pstate_update_status(const char *buf, size_t size)
3175 {
3176 	if (size == 3 && !strncmp(buf, "off", size)) {
3177 		if (!intel_pstate_driver)
3178 			return -EINVAL;
3179 
3180 		if (hwp_active)
3181 			return -EBUSY;
3182 
3183 		cpufreq_unregister_driver(intel_pstate_driver);
3184 		intel_pstate_driver_cleanup();
3185 		return 0;
3186 	}
3187 
3188 	if (size == 6 && !strncmp(buf, "active", size)) {
3189 		if (intel_pstate_driver) {
3190 			if (intel_pstate_driver == &intel_pstate)
3191 				return 0;
3192 
3193 			cpufreq_unregister_driver(intel_pstate_driver);
3194 		}
3195 
3196 		return intel_pstate_register_driver(&intel_pstate);
3197 	}
3198 
3199 	if (size == 7 && !strncmp(buf, "passive", size)) {
3200 		if (intel_pstate_driver) {
3201 			if (intel_pstate_driver == &intel_cpufreq)
3202 				return 0;
3203 
3204 			cpufreq_unregister_driver(intel_pstate_driver);
3205 			intel_pstate_sysfs_hide_hwp_dynamic_boost();
3206 		}
3207 
3208 		return intel_pstate_register_driver(&intel_cpufreq);
3209 	}
3210 
3211 	return -EINVAL;
3212 }
3213 
3214 static int no_load __initdata;
3215 static int no_hwp __initdata;
3216 static int hwp_only __initdata;
3217 static unsigned int force_load __initdata;
3218 
3219 static int __init intel_pstate_msrs_not_valid(void)
3220 {
3221 	if (!pstate_funcs.get_max() ||
3222 	    !pstate_funcs.get_min() ||
3223 	    !pstate_funcs.get_turbo())
3224 		return -ENODEV;
3225 
3226 	return 0;
3227 }
3228 
3229 static void __init copy_cpu_funcs(struct pstate_funcs *funcs)
3230 {
3231 	pstate_funcs.get_max   = funcs->get_max;
3232 	pstate_funcs.get_max_physical = funcs->get_max_physical;
3233 	pstate_funcs.get_min   = funcs->get_min;
3234 	pstate_funcs.get_turbo = funcs->get_turbo;
3235 	pstate_funcs.get_scaling = funcs->get_scaling;
3236 	pstate_funcs.get_val   = funcs->get_val;
3237 	pstate_funcs.get_vid   = funcs->get_vid;
3238 	pstate_funcs.get_aperf_mperf_shift = funcs->get_aperf_mperf_shift;
3239 }
3240 
3241 #ifdef CONFIG_ACPI
3242 
3243 static bool __init intel_pstate_no_acpi_pss(void)
3244 {
3245 	int i;
3246 
3247 	for_each_possible_cpu(i) {
3248 		acpi_status status;
3249 		union acpi_object *pss;
3250 		struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
3251 		struct acpi_processor *pr = per_cpu(processors, i);
3252 
3253 		if (!pr)
3254 			continue;
3255 
3256 		status = acpi_evaluate_object(pr->handle, "_PSS", NULL, &buffer);
3257 		if (ACPI_FAILURE(status))
3258 			continue;
3259 
3260 		pss = buffer.pointer;
3261 		if (pss && pss->type == ACPI_TYPE_PACKAGE) {
3262 			kfree(pss);
3263 			return false;
3264 		}
3265 
3266 		kfree(pss);
3267 	}
3268 
3269 	pr_debug("ACPI _PSS not found\n");
3270 	return true;
3271 }
3272 
3273 static bool __init intel_pstate_no_acpi_pcch(void)
3274 {
3275 	acpi_status status;
3276 	acpi_handle handle;
3277 
3278 	status = acpi_get_handle(NULL, "\\_SB", &handle);
3279 	if (ACPI_FAILURE(status))
3280 		goto not_found;
3281 
3282 	if (acpi_has_method(handle, "PCCH"))
3283 		return false;
3284 
3285 not_found:
3286 	pr_debug("ACPI PCCH not found\n");
3287 	return true;
3288 }
3289 
3290 static bool __init intel_pstate_has_acpi_ppc(void)
3291 {
3292 	int i;
3293 
3294 	for_each_possible_cpu(i) {
3295 		struct acpi_processor *pr = per_cpu(processors, i);
3296 
3297 		if (!pr)
3298 			continue;
3299 		if (acpi_has_method(pr->handle, "_PPC"))
3300 			return true;
3301 	}
3302 	pr_debug("ACPI _PPC not found\n");
3303 	return false;
3304 }
3305 
3306 enum {
3307 	PSS,
3308 	PPC,
3309 };
3310 
3311 /* Hardware vendor-specific info that has its own power management modes */
3312 static struct acpi_platform_list plat_info[] __initdata = {
3313 	{"HP    ", "ProLiant", 0, ACPI_SIG_FADT, all_versions, NULL, PSS},
3314 	{"ORACLE", "X4-2    ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3315 	{"ORACLE", "X4-2L   ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3316 	{"ORACLE", "X4-2B   ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3317 	{"ORACLE", "X3-2    ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3318 	{"ORACLE", "X3-2L   ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3319 	{"ORACLE", "X3-2B   ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3320 	{"ORACLE", "X4470M2 ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3321 	{"ORACLE", "X4270M3 ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3322 	{"ORACLE", "X4270M2 ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3323 	{"ORACLE", "X4170M2 ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3324 	{"ORACLE", "X4170 M3", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3325 	{"ORACLE", "X4275 M3", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3326 	{"ORACLE", "X6-2    ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3327 	{"ORACLE", "Sudbury ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3328 	{ } /* End */
3329 };
3330 
3331 #define BITMASK_OOB	(BIT(8) | BIT(18))
3332 
3333 static bool __init intel_pstate_platform_pwr_mgmt_exists(void)
3334 {
3335 	const struct x86_cpu_id *id;
3336 	u64 misc_pwr;
3337 	int idx;
3338 
3339 	id = x86_match_cpu(intel_pstate_cpu_oob_ids);
3340 	if (id) {
3341 		rdmsrl(MSR_MISC_PWR_MGMT, misc_pwr);
3342 		if (misc_pwr & BITMASK_OOB) {
3343 			pr_debug("Bit 8 or 18 in the MISC_PWR_MGMT MSR set\n");
3344 			pr_debug("P states are controlled in Out of Band mode by the firmware/hardware\n");
3345 			return true;
3346 		}
3347 	}
3348 
3349 	idx = acpi_match_platform_list(plat_info);
3350 	if (idx < 0)
3351 		return false;
3352 
3353 	switch (plat_info[idx].data) {
3354 	case PSS:
3355 		if (!intel_pstate_no_acpi_pss())
3356 			return false;
3357 
3358 		return intel_pstate_no_acpi_pcch();
3359 	case PPC:
3360 		return intel_pstate_has_acpi_ppc() && !force_load;
3361 	}
3362 
3363 	return false;
3364 }
3365 
3366 static void intel_pstate_request_control_from_smm(void)
3367 {
3368 	/*
3369 	 * It may be unsafe to request P-states control from SMM if _PPC support
3370 	 * has not been enabled.
3371 	 */
3372 	if (acpi_ppc)
3373 		acpi_processor_pstate_control();
3374 }
3375 #else /* CONFIG_ACPI not enabled */
3376 static inline bool intel_pstate_platform_pwr_mgmt_exists(void) { return false; }
3377 static inline bool intel_pstate_has_acpi_ppc(void) { return false; }
3378 static inline void intel_pstate_request_control_from_smm(void) {}
3379 #endif /* CONFIG_ACPI */
3380 
3381 #define INTEL_PSTATE_HWP_BROADWELL	0x01
3382 
3383 #define X86_MATCH_HWP(model, hwp_mode)					\
3384 	X86_MATCH_VENDOR_FAM_MODEL_FEATURE(INTEL, 6, INTEL_FAM6_##model, \
3385 					   X86_FEATURE_HWP, hwp_mode)
3386 
3387 static const struct x86_cpu_id hwp_support_ids[] __initconst = {
3388 	X86_MATCH_HWP(BROADWELL_X,	INTEL_PSTATE_HWP_BROADWELL),
3389 	X86_MATCH_HWP(BROADWELL_D,	INTEL_PSTATE_HWP_BROADWELL),
3390 	X86_MATCH_HWP(ANY,		0),
3391 	{}
3392 };
3393 
3394 static bool intel_pstate_hwp_is_enabled(void)
3395 {
3396 	u64 value;
3397 
3398 	rdmsrl(MSR_PM_ENABLE, value);
3399 	return !!(value & 0x1);
3400 }
3401 
3402 static const struct x86_cpu_id intel_epp_balance_perf[] = {
3403 	/*
3404 	 * Set EPP value as 102, this is the max suggested EPP
3405 	 * which can result in one core turbo frequency for
3406 	 * AlderLake Mobile CPUs.
3407 	 */
3408 	X86_MATCH_INTEL_FAM6_MODEL(ALDERLAKE_L, 102),
3409 	{}
3410 };
3411 
3412 static int __init intel_pstate_init(void)
3413 {
3414 	static struct cpudata **_all_cpu_data;
3415 	const struct x86_cpu_id *id;
3416 	int rc;
3417 
3418 	if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL)
3419 		return -ENODEV;
3420 
3421 	id = x86_match_cpu(hwp_support_ids);
3422 	if (id) {
3423 		bool hwp_forced = intel_pstate_hwp_is_enabled();
3424 
3425 		if (hwp_forced)
3426 			pr_info("HWP enabled by BIOS\n");
3427 		else if (no_load)
3428 			return -ENODEV;
3429 
3430 		copy_cpu_funcs(&core_funcs);
3431 		/*
3432 		 * Avoid enabling HWP for processors without EPP support,
3433 		 * because that means incomplete HWP implementation which is a
3434 		 * corner case and supporting it is generally problematic.
3435 		 *
3436 		 * If HWP is enabled already, though, there is no choice but to
3437 		 * deal with it.
3438 		 */
3439 		if ((!no_hwp && boot_cpu_has(X86_FEATURE_HWP_EPP)) || hwp_forced) {
3440 			WRITE_ONCE(hwp_active, 1);
3441 			hwp_mode_bdw = id->driver_data;
3442 			intel_pstate.attr = hwp_cpufreq_attrs;
3443 			intel_cpufreq.attr = hwp_cpufreq_attrs;
3444 			intel_cpufreq.flags |= CPUFREQ_NEED_UPDATE_LIMITS;
3445 			intel_cpufreq.adjust_perf = intel_cpufreq_adjust_perf;
3446 			if (!default_driver)
3447 				default_driver = &intel_pstate;
3448 
3449 			if (boot_cpu_has(X86_FEATURE_HYBRID_CPU))
3450 				intel_pstate_cppc_set_cpu_scaling();
3451 
3452 			goto hwp_cpu_matched;
3453 		}
3454 		pr_info("HWP not enabled\n");
3455 	} else {
3456 		if (no_load)
3457 			return -ENODEV;
3458 
3459 		id = x86_match_cpu(intel_pstate_cpu_ids);
3460 		if (!id) {
3461 			pr_info("CPU model not supported\n");
3462 			return -ENODEV;
3463 		}
3464 
3465 		copy_cpu_funcs((struct pstate_funcs *)id->driver_data);
3466 	}
3467 
3468 	if (intel_pstate_msrs_not_valid()) {
3469 		pr_info("Invalid MSRs\n");
3470 		return -ENODEV;
3471 	}
3472 	/* Without HWP start in the passive mode. */
3473 	if (!default_driver)
3474 		default_driver = &intel_cpufreq;
3475 
3476 hwp_cpu_matched:
3477 	/*
3478 	 * The Intel pstate driver will be ignored if the platform
3479 	 * firmware has its own power management modes.
3480 	 */
3481 	if (intel_pstate_platform_pwr_mgmt_exists()) {
3482 		pr_info("P-states controlled by the platform\n");
3483 		return -ENODEV;
3484 	}
3485 
3486 	if (!hwp_active && hwp_only)
3487 		return -ENOTSUPP;
3488 
3489 	pr_info("Intel P-state driver initializing\n");
3490 
3491 	_all_cpu_data = vzalloc(array_size(sizeof(void *), num_possible_cpus()));
3492 	if (!_all_cpu_data)
3493 		return -ENOMEM;
3494 
3495 	WRITE_ONCE(all_cpu_data, _all_cpu_data);
3496 
3497 	intel_pstate_request_control_from_smm();
3498 
3499 	intel_pstate_sysfs_expose_params();
3500 
3501 	if (hwp_active) {
3502 		const struct x86_cpu_id *id = x86_match_cpu(intel_epp_balance_perf);
3503 
3504 		if (id)
3505 			epp_values[EPP_INDEX_BALANCE_PERFORMANCE] = id->driver_data;
3506 	}
3507 
3508 	mutex_lock(&intel_pstate_driver_lock);
3509 	rc = intel_pstate_register_driver(default_driver);
3510 	mutex_unlock(&intel_pstate_driver_lock);
3511 	if (rc) {
3512 		intel_pstate_sysfs_remove();
3513 		return rc;
3514 	}
3515 
3516 	if (hwp_active) {
3517 		const struct x86_cpu_id *id;
3518 
3519 		id = x86_match_cpu(intel_pstate_cpu_ee_disable_ids);
3520 		if (id) {
3521 			set_power_ctl_ee_state(false);
3522 			pr_info("Disabling energy efficiency optimization\n");
3523 		}
3524 
3525 		pr_info("HWP enabled\n");
3526 	} else if (boot_cpu_has(X86_FEATURE_HYBRID_CPU)) {
3527 		pr_warn("Problematic setup: Hybrid processor with disabled HWP\n");
3528 	}
3529 
3530 	return 0;
3531 }
3532 device_initcall(intel_pstate_init);
3533 
3534 static int __init intel_pstate_setup(char *str)
3535 {
3536 	if (!str)
3537 		return -EINVAL;
3538 
3539 	if (!strcmp(str, "disable"))
3540 		no_load = 1;
3541 	else if (!strcmp(str, "active"))
3542 		default_driver = &intel_pstate;
3543 	else if (!strcmp(str, "passive"))
3544 		default_driver = &intel_cpufreq;
3545 
3546 	if (!strcmp(str, "no_hwp"))
3547 		no_hwp = 1;
3548 
3549 	if (!strcmp(str, "force"))
3550 		force_load = 1;
3551 	if (!strcmp(str, "hwp_only"))
3552 		hwp_only = 1;
3553 	if (!strcmp(str, "per_cpu_perf_limits"))
3554 		per_cpu_limits = true;
3555 
3556 #ifdef CONFIG_ACPI
3557 	if (!strcmp(str, "support_acpi_ppc"))
3558 		acpi_ppc = true;
3559 #endif
3560 
3561 	return 0;
3562 }
3563 early_param("intel_pstate", intel_pstate_setup);
3564 
3565 MODULE_AUTHOR("Dirk Brandewie <dirk.j.brandewie@intel.com>");
3566 MODULE_DESCRIPTION("'intel_pstate' - P state driver Intel Core processors");
3567 MODULE_LICENSE("GPL");
3568