xref: /linux/drivers/cpufreq/intel_pstate.c (revision 1e0731c05c985deb68a97fa44c1adcd3305dda90)
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/cpu.h>
31 #include <asm/div64.h>
32 #include <asm/msr.h>
33 #include <asm/cpu_device_id.h>
34 #include <asm/cpufeature.h>
35 #include <asm/intel-family.h>
36 #include "../drivers/thermal/intel/thermal_interrupt.h"
37 
38 #define INTEL_PSTATE_SAMPLING_INTERVAL	(10 * NSEC_PER_MSEC)
39 
40 #define INTEL_CPUFREQ_TRANSITION_LATENCY	20000
41 #define INTEL_CPUFREQ_TRANSITION_DELAY_HWP	5000
42 #define INTEL_CPUFREQ_TRANSITION_DELAY		500
43 
44 #ifdef CONFIG_ACPI
45 #include <acpi/processor.h>
46 #include <acpi/cppc_acpi.h>
47 #endif
48 
49 #define FRAC_BITS 8
50 #define int_tofp(X) ((int64_t)(X) << FRAC_BITS)
51 #define fp_toint(X) ((X) >> FRAC_BITS)
52 
53 #define ONE_EIGHTH_FP ((int64_t)1 << (FRAC_BITS - 3))
54 
55 #define EXT_BITS 6
56 #define EXT_FRAC_BITS (EXT_BITS + FRAC_BITS)
57 #define fp_ext_toint(X) ((X) >> EXT_FRAC_BITS)
58 #define int_ext_tofp(X) ((int64_t)(X) << EXT_FRAC_BITS)
59 
60 static inline int32_t mul_fp(int32_t x, int32_t y)
61 {
62 	return ((int64_t)x * (int64_t)y) >> FRAC_BITS;
63 }
64 
65 static inline int32_t div_fp(s64 x, s64 y)
66 {
67 	return div64_s64((int64_t)x << FRAC_BITS, y);
68 }
69 
70 static inline int ceiling_fp(int32_t x)
71 {
72 	int mask, ret;
73 
74 	ret = fp_toint(x);
75 	mask = (1 << FRAC_BITS) - 1;
76 	if (x & mask)
77 		ret += 1;
78 	return ret;
79 }
80 
81 static inline u64 mul_ext_fp(u64 x, u64 y)
82 {
83 	return (x * y) >> EXT_FRAC_BITS;
84 }
85 
86 static inline u64 div_ext_fp(u64 x, u64 y)
87 {
88 	return div64_u64(x << EXT_FRAC_BITS, y);
89 }
90 
91 /**
92  * struct sample -	Store performance sample
93  * @core_avg_perf:	Ratio of APERF/MPERF which is the actual average
94  *			performance during last sample period
95  * @busy_scaled:	Scaled busy value which is used to calculate next
96  *			P state. This can be different than core_avg_perf
97  *			to account for cpu idle period
98  * @aperf:		Difference of actual performance frequency clock count
99  *			read from APERF MSR between last and current sample
100  * @mperf:		Difference of maximum performance frequency clock count
101  *			read from MPERF MSR between last and current sample
102  * @tsc:		Difference of time stamp counter between last and
103  *			current sample
104  * @time:		Current time from scheduler
105  *
106  * This structure is used in the cpudata structure to store performance sample
107  * data for choosing next P State.
108  */
109 struct sample {
110 	int32_t core_avg_perf;
111 	int32_t busy_scaled;
112 	u64 aperf;
113 	u64 mperf;
114 	u64 tsc;
115 	u64 time;
116 };
117 
118 /**
119  * struct pstate_data - Store P state data
120  * @current_pstate:	Current requested P state
121  * @min_pstate:		Min P state possible for this platform
122  * @max_pstate:		Max P state possible for this platform
123  * @max_pstate_physical:This is physical Max P state for a processor
124  *			This can be higher than the max_pstate which can
125  *			be limited by platform thermal design power limits
126  * @perf_ctl_scaling:	PERF_CTL P-state to frequency scaling factor
127  * @scaling:		Scaling factor between performance and frequency
128  * @turbo_pstate:	Max Turbo P state possible for this platform
129  * @min_freq:		@min_pstate frequency in cpufreq units
130  * @max_freq:		@max_pstate frequency in cpufreq units
131  * @turbo_freq:		@turbo_pstate frequency in cpufreq units
132  *
133  * Stores the per cpu model P state limits and current P state.
134  */
135 struct pstate_data {
136 	int	current_pstate;
137 	int	min_pstate;
138 	int	max_pstate;
139 	int	max_pstate_physical;
140 	int	perf_ctl_scaling;
141 	int	scaling;
142 	int	turbo_pstate;
143 	unsigned int min_freq;
144 	unsigned int max_freq;
145 	unsigned int turbo_freq;
146 };
147 
148 /**
149  * struct vid_data -	Stores voltage information data
150  * @min:		VID data for this platform corresponding to
151  *			the lowest P state
152  * @max:		VID data corresponding to the highest P State.
153  * @turbo:		VID data for turbo P state
154  * @ratio:		Ratio of (vid max - vid min) /
155  *			(max P state - Min P State)
156  *
157  * Stores the voltage data for DVFS (Dynamic Voltage and Frequency Scaling)
158  * This data is used in Atom platforms, where in addition to target P state,
159  * the voltage data needs to be specified to select next P State.
160  */
161 struct vid_data {
162 	int min;
163 	int max;
164 	int turbo;
165 	int32_t ratio;
166 };
167 
168 /**
169  * struct global_params - Global parameters, mostly tunable via sysfs.
170  * @no_turbo:		Whether or not to use turbo P-states.
171  * @turbo_disabled:	Whether or not turbo P-states are available at all,
172  *			based on the MSR_IA32_MISC_ENABLE value and whether or
173  *			not the maximum reported turbo P-state is different from
174  *			the maximum reported non-turbo one.
175  * @turbo_disabled_mf:	The @turbo_disabled value reflected by cpuinfo.max_freq.
176  * @min_perf_pct:	Minimum capacity limit in percent of the maximum turbo
177  *			P-state capacity.
178  * @max_perf_pct:	Maximum capacity limit in percent of the maximum turbo
179  *			P-state capacity.
180  */
181 struct global_params {
182 	bool no_turbo;
183 	bool turbo_disabled;
184 	bool turbo_disabled_mf;
185 	int max_perf_pct;
186 	int min_perf_pct;
187 };
188 
189 /**
190  * struct cpudata -	Per CPU instance data storage
191  * @cpu:		CPU number for this instance data
192  * @policy:		CPUFreq policy value
193  * @update_util:	CPUFreq utility callback information
194  * @update_util_set:	CPUFreq utility callback is set
195  * @iowait_boost:	iowait-related boost fraction
196  * @last_update:	Time of the last update.
197  * @pstate:		Stores P state limits for this CPU
198  * @vid:		Stores VID limits for this CPU
199  * @last_sample_time:	Last Sample time
200  * @aperf_mperf_shift:	APERF vs MPERF counting frequency difference
201  * @prev_aperf:		Last APERF value read from APERF MSR
202  * @prev_mperf:		Last MPERF value read from MPERF MSR
203  * @prev_tsc:		Last timestamp counter (TSC) value
204  * @prev_cummulative_iowait: IO Wait time difference from last and
205  *			current sample
206  * @sample:		Storage for storing last Sample data
207  * @min_perf_ratio:	Minimum capacity in terms of PERF or HWP ratios
208  * @max_perf_ratio:	Maximum capacity in terms of PERF or HWP ratios
209  * @acpi_perf_data:	Stores ACPI perf information read from _PSS
210  * @valid_pss_table:	Set to true for valid ACPI _PSS entries found
211  * @epp_powersave:	Last saved HWP energy performance preference
212  *			(EPP) or energy performance bias (EPB),
213  *			when policy switched to performance
214  * @epp_policy:		Last saved policy used to set EPP/EPB
215  * @epp_default:	Power on default HWP energy performance
216  *			preference/bias
217  * @epp_cached		Cached HWP energy-performance preference value
218  * @hwp_req_cached:	Cached value of the last HWP Request MSR
219  * @hwp_cap_cached:	Cached value of the last HWP Capabilities MSR
220  * @last_io_update:	Last time when IO wake flag was set
221  * @sched_flags:	Store scheduler flags for possible cross CPU update
222  * @hwp_boost_min:	Last HWP boosted min performance
223  * @suspended:		Whether or not the driver has been suspended.
224  * @hwp_notify_work:	workqueue for HWP notifications.
225  *
226  * This structure stores per CPU instance data for all CPUs.
227  */
228 struct cpudata {
229 	int cpu;
230 
231 	unsigned int policy;
232 	struct update_util_data update_util;
233 	bool   update_util_set;
234 
235 	struct pstate_data pstate;
236 	struct vid_data vid;
237 
238 	u64	last_update;
239 	u64	last_sample_time;
240 	u64	aperf_mperf_shift;
241 	u64	prev_aperf;
242 	u64	prev_mperf;
243 	u64	prev_tsc;
244 	u64	prev_cummulative_iowait;
245 	struct sample sample;
246 	int32_t	min_perf_ratio;
247 	int32_t	max_perf_ratio;
248 #ifdef CONFIG_ACPI
249 	struct acpi_processor_performance acpi_perf_data;
250 	bool valid_pss_table;
251 #endif
252 	unsigned int iowait_boost;
253 	s16 epp_powersave;
254 	s16 epp_policy;
255 	s16 epp_default;
256 	s16 epp_cached;
257 	u64 hwp_req_cached;
258 	u64 hwp_cap_cached;
259 	u64 last_io_update;
260 	unsigned int sched_flags;
261 	u32 hwp_boost_min;
262 	bool suspended;
263 	struct delayed_work hwp_notify_work;
264 };
265 
266 static struct cpudata **all_cpu_data;
267 
268 /**
269  * struct pstate_funcs - Per CPU model specific callbacks
270  * @get_max:		Callback to get maximum non turbo effective P state
271  * @get_max_physical:	Callback to get maximum non turbo physical P state
272  * @get_min:		Callback to get minimum P state
273  * @get_turbo:		Callback to get turbo P state
274  * @get_scaling:	Callback to get frequency scaling factor
275  * @get_cpu_scaling:	Get frequency scaling factor for a given cpu
276  * @get_aperf_mperf_shift: Callback to get the APERF vs MPERF frequency difference
277  * @get_val:		Callback to convert P state to actual MSR write value
278  * @get_vid:		Callback to get VID data for Atom platforms
279  *
280  * Core and Atom CPU models have different way to get P State limits. This
281  * structure is used to store those callbacks.
282  */
283 struct pstate_funcs {
284 	int (*get_max)(int cpu);
285 	int (*get_max_physical)(int cpu);
286 	int (*get_min)(int cpu);
287 	int (*get_turbo)(int cpu);
288 	int (*get_scaling)(void);
289 	int (*get_cpu_scaling)(int cpu);
290 	int (*get_aperf_mperf_shift)(void);
291 	u64 (*get_val)(struct cpudata*, int pstate);
292 	void (*get_vid)(struct cpudata *);
293 };
294 
295 static struct pstate_funcs pstate_funcs __read_mostly;
296 
297 static int hwp_active __read_mostly;
298 static int hwp_mode_bdw __read_mostly;
299 static bool per_cpu_limits __read_mostly;
300 static bool hwp_boost __read_mostly;
301 static bool hwp_forced __read_mostly;
302 
303 static struct cpufreq_driver *intel_pstate_driver __read_mostly;
304 
305 #define HYBRID_SCALING_FACTOR	78741
306 
307 static inline int core_get_scaling(void)
308 {
309 	return 100000;
310 }
311 
312 #ifdef CONFIG_ACPI
313 static bool acpi_ppc;
314 #endif
315 
316 static struct global_params global;
317 
318 static DEFINE_MUTEX(intel_pstate_driver_lock);
319 static DEFINE_MUTEX(intel_pstate_limits_lock);
320 
321 #ifdef CONFIG_ACPI
322 
323 static bool intel_pstate_acpi_pm_profile_server(void)
324 {
325 	if (acpi_gbl_FADT.preferred_profile == PM_ENTERPRISE_SERVER ||
326 	    acpi_gbl_FADT.preferred_profile == PM_PERFORMANCE_SERVER)
327 		return true;
328 
329 	return false;
330 }
331 
332 static bool intel_pstate_get_ppc_enable_status(void)
333 {
334 	if (intel_pstate_acpi_pm_profile_server())
335 		return true;
336 
337 	return acpi_ppc;
338 }
339 
340 #ifdef CONFIG_ACPI_CPPC_LIB
341 
342 /* The work item is needed to avoid CPU hotplug locking issues */
343 static void intel_pstste_sched_itmt_work_fn(struct work_struct *work)
344 {
345 	sched_set_itmt_support();
346 }
347 
348 static DECLARE_WORK(sched_itmt_work, intel_pstste_sched_itmt_work_fn);
349 
350 #define CPPC_MAX_PERF	U8_MAX
351 
352 static void intel_pstate_set_itmt_prio(int cpu)
353 {
354 	struct cppc_perf_caps cppc_perf;
355 	static u32 max_highest_perf = 0, min_highest_perf = U32_MAX;
356 	int ret;
357 
358 	ret = cppc_get_perf_caps(cpu, &cppc_perf);
359 	if (ret)
360 		return;
361 
362 	/*
363 	 * On some systems with overclocking enabled, CPPC.highest_perf is hardcoded to 0xff.
364 	 * In this case we can't use CPPC.highest_perf to enable ITMT.
365 	 * In this case we can look at MSR_HWP_CAPABILITIES bits [8:0] to decide.
366 	 */
367 	if (cppc_perf.highest_perf == CPPC_MAX_PERF)
368 		cppc_perf.highest_perf = HWP_HIGHEST_PERF(READ_ONCE(all_cpu_data[cpu]->hwp_cap_cached));
369 
370 	/*
371 	 * The priorities can be set regardless of whether or not
372 	 * sched_set_itmt_support(true) has been called and it is valid to
373 	 * update them at any time after it has been called.
374 	 */
375 	sched_set_itmt_core_prio(cppc_perf.highest_perf, cpu);
376 
377 	if (max_highest_perf <= min_highest_perf) {
378 		if (cppc_perf.highest_perf > max_highest_perf)
379 			max_highest_perf = cppc_perf.highest_perf;
380 
381 		if (cppc_perf.highest_perf < min_highest_perf)
382 			min_highest_perf = cppc_perf.highest_perf;
383 
384 		if (max_highest_perf > min_highest_perf) {
385 			/*
386 			 * This code can be run during CPU online under the
387 			 * CPU hotplug locks, so sched_set_itmt_support()
388 			 * cannot be called from here.  Queue up a work item
389 			 * to invoke it.
390 			 */
391 			schedule_work(&sched_itmt_work);
392 		}
393 	}
394 }
395 
396 static int intel_pstate_get_cppc_guaranteed(int cpu)
397 {
398 	struct cppc_perf_caps cppc_perf;
399 	int ret;
400 
401 	ret = cppc_get_perf_caps(cpu, &cppc_perf);
402 	if (ret)
403 		return ret;
404 
405 	if (cppc_perf.guaranteed_perf)
406 		return cppc_perf.guaranteed_perf;
407 
408 	return cppc_perf.nominal_perf;
409 }
410 
411 static int intel_pstate_cppc_get_scaling(int cpu)
412 {
413 	struct cppc_perf_caps cppc_perf;
414 	int ret;
415 
416 	ret = cppc_get_perf_caps(cpu, &cppc_perf);
417 
418 	/*
419 	 * If the nominal frequency and the nominal performance are not
420 	 * zero and the ratio between them is not 100, return the hybrid
421 	 * scaling factor.
422 	 */
423 	if (!ret && cppc_perf.nominal_perf && cppc_perf.nominal_freq &&
424 	    cppc_perf.nominal_perf * 100 != cppc_perf.nominal_freq)
425 		return HYBRID_SCALING_FACTOR;
426 
427 	return core_get_scaling();
428 }
429 
430 #else /* CONFIG_ACPI_CPPC_LIB */
431 static inline void intel_pstate_set_itmt_prio(int cpu)
432 {
433 }
434 #endif /* CONFIG_ACPI_CPPC_LIB */
435 
436 static void intel_pstate_init_acpi_perf_limits(struct cpufreq_policy *policy)
437 {
438 	struct cpudata *cpu;
439 	int ret;
440 	int i;
441 
442 	if (hwp_active) {
443 		intel_pstate_set_itmt_prio(policy->cpu);
444 		return;
445 	}
446 
447 	if (!intel_pstate_get_ppc_enable_status())
448 		return;
449 
450 	cpu = all_cpu_data[policy->cpu];
451 
452 	ret = acpi_processor_register_performance(&cpu->acpi_perf_data,
453 						  policy->cpu);
454 	if (ret)
455 		return;
456 
457 	/*
458 	 * Check if the control value in _PSS is for PERF_CTL MSR, which should
459 	 * guarantee that the states returned by it map to the states in our
460 	 * list directly.
461 	 */
462 	if (cpu->acpi_perf_data.control_register.space_id !=
463 						ACPI_ADR_SPACE_FIXED_HARDWARE)
464 		goto err;
465 
466 	/*
467 	 * If there is only one entry _PSS, simply ignore _PSS and continue as
468 	 * usual without taking _PSS into account
469 	 */
470 	if (cpu->acpi_perf_data.state_count < 2)
471 		goto err;
472 
473 	pr_debug("CPU%u - ACPI _PSS perf data\n", policy->cpu);
474 	for (i = 0; i < cpu->acpi_perf_data.state_count; i++) {
475 		pr_debug("     %cP%d: %u MHz, %u mW, 0x%x\n",
476 			 (i == cpu->acpi_perf_data.state ? '*' : ' '), i,
477 			 (u32) cpu->acpi_perf_data.states[i].core_frequency,
478 			 (u32) cpu->acpi_perf_data.states[i].power,
479 			 (u32) cpu->acpi_perf_data.states[i].control);
480 	}
481 
482 	cpu->valid_pss_table = true;
483 	pr_debug("_PPC limits will be enforced\n");
484 
485 	return;
486 
487  err:
488 	cpu->valid_pss_table = false;
489 	acpi_processor_unregister_performance(policy->cpu);
490 }
491 
492 static void intel_pstate_exit_perf_limits(struct cpufreq_policy *policy)
493 {
494 	struct cpudata *cpu;
495 
496 	cpu = all_cpu_data[policy->cpu];
497 	if (!cpu->valid_pss_table)
498 		return;
499 
500 	acpi_processor_unregister_performance(policy->cpu);
501 }
502 #else /* CONFIG_ACPI */
503 static inline void intel_pstate_init_acpi_perf_limits(struct cpufreq_policy *policy)
504 {
505 }
506 
507 static inline void intel_pstate_exit_perf_limits(struct cpufreq_policy *policy)
508 {
509 }
510 
511 static inline bool intel_pstate_acpi_pm_profile_server(void)
512 {
513 	return false;
514 }
515 #endif /* CONFIG_ACPI */
516 
517 #ifndef CONFIG_ACPI_CPPC_LIB
518 static inline int intel_pstate_get_cppc_guaranteed(int cpu)
519 {
520 	return -ENOTSUPP;
521 }
522 
523 static int intel_pstate_cppc_get_scaling(int cpu)
524 {
525 	return core_get_scaling();
526 }
527 #endif /* CONFIG_ACPI_CPPC_LIB */
528 
529 /**
530  * intel_pstate_hybrid_hwp_adjust - Calibrate HWP performance levels.
531  * @cpu: Target CPU.
532  *
533  * On hybrid processors, HWP may expose more performance levels than there are
534  * P-states accessible through the PERF_CTL interface.  If that happens, the
535  * scaling factor between HWP performance levels and CPU frequency will be less
536  * than the scaling factor between P-state values and CPU frequency.
537  *
538  * In that case, adjust the CPU parameters used in computations accordingly.
539  */
540 static void intel_pstate_hybrid_hwp_adjust(struct cpudata *cpu)
541 {
542 	int perf_ctl_max_phys = cpu->pstate.max_pstate_physical;
543 	int perf_ctl_scaling = cpu->pstate.perf_ctl_scaling;
544 	int perf_ctl_turbo = pstate_funcs.get_turbo(cpu->cpu);
545 	int scaling = cpu->pstate.scaling;
546 
547 	pr_debug("CPU%d: perf_ctl_max_phys = %d\n", cpu->cpu, perf_ctl_max_phys);
548 	pr_debug("CPU%d: perf_ctl_turbo = %d\n", cpu->cpu, perf_ctl_turbo);
549 	pr_debug("CPU%d: perf_ctl_scaling = %d\n", cpu->cpu, perf_ctl_scaling);
550 	pr_debug("CPU%d: HWP_CAP guaranteed = %d\n", cpu->cpu, cpu->pstate.max_pstate);
551 	pr_debug("CPU%d: HWP_CAP highest = %d\n", cpu->cpu, cpu->pstate.turbo_pstate);
552 	pr_debug("CPU%d: HWP-to-frequency scaling factor: %d\n", cpu->cpu, scaling);
553 
554 	cpu->pstate.turbo_freq = rounddown(cpu->pstate.turbo_pstate * scaling,
555 					   perf_ctl_scaling);
556 	cpu->pstate.max_freq = rounddown(cpu->pstate.max_pstate * scaling,
557 					 perf_ctl_scaling);
558 
559 	cpu->pstate.max_pstate_physical =
560 			DIV_ROUND_UP(perf_ctl_max_phys * perf_ctl_scaling,
561 				     scaling);
562 
563 	cpu->pstate.min_freq = cpu->pstate.min_pstate * perf_ctl_scaling;
564 	/*
565 	 * Cast the min P-state value retrieved via pstate_funcs.get_min() to
566 	 * the effective range of HWP performance levels.
567 	 */
568 	cpu->pstate.min_pstate = DIV_ROUND_UP(cpu->pstate.min_freq, scaling);
569 }
570 
571 static inline void update_turbo_state(void)
572 {
573 	u64 misc_en;
574 	struct cpudata *cpu;
575 
576 	cpu = all_cpu_data[0];
577 	rdmsrl(MSR_IA32_MISC_ENABLE, misc_en);
578 	global.turbo_disabled =
579 		(misc_en & MSR_IA32_MISC_ENABLE_TURBO_DISABLE ||
580 		 cpu->pstate.max_pstate == cpu->pstate.turbo_pstate);
581 }
582 
583 static int min_perf_pct_min(void)
584 {
585 	struct cpudata *cpu = all_cpu_data[0];
586 	int turbo_pstate = cpu->pstate.turbo_pstate;
587 
588 	return turbo_pstate ?
589 		(cpu->pstate.min_pstate * 100 / turbo_pstate) : 0;
590 }
591 
592 static s16 intel_pstate_get_epb(struct cpudata *cpu_data)
593 {
594 	u64 epb;
595 	int ret;
596 
597 	if (!boot_cpu_has(X86_FEATURE_EPB))
598 		return -ENXIO;
599 
600 	ret = rdmsrl_on_cpu(cpu_data->cpu, MSR_IA32_ENERGY_PERF_BIAS, &epb);
601 	if (ret)
602 		return (s16)ret;
603 
604 	return (s16)(epb & 0x0f);
605 }
606 
607 static s16 intel_pstate_get_epp(struct cpudata *cpu_data, u64 hwp_req_data)
608 {
609 	s16 epp;
610 
611 	if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
612 		/*
613 		 * When hwp_req_data is 0, means that caller didn't read
614 		 * MSR_HWP_REQUEST, so need to read and get EPP.
615 		 */
616 		if (!hwp_req_data) {
617 			epp = rdmsrl_on_cpu(cpu_data->cpu, MSR_HWP_REQUEST,
618 					    &hwp_req_data);
619 			if (epp)
620 				return epp;
621 		}
622 		epp = (hwp_req_data >> 24) & 0xff;
623 	} else {
624 		/* When there is no EPP present, HWP uses EPB settings */
625 		epp = intel_pstate_get_epb(cpu_data);
626 	}
627 
628 	return epp;
629 }
630 
631 static int intel_pstate_set_epb(int cpu, s16 pref)
632 {
633 	u64 epb;
634 	int ret;
635 
636 	if (!boot_cpu_has(X86_FEATURE_EPB))
637 		return -ENXIO;
638 
639 	ret = rdmsrl_on_cpu(cpu, MSR_IA32_ENERGY_PERF_BIAS, &epb);
640 	if (ret)
641 		return ret;
642 
643 	epb = (epb & ~0x0f) | pref;
644 	wrmsrl_on_cpu(cpu, MSR_IA32_ENERGY_PERF_BIAS, epb);
645 
646 	return 0;
647 }
648 
649 /*
650  * EPP/EPB display strings corresponding to EPP index in the
651  * energy_perf_strings[]
652  *	index		String
653  *-------------------------------------
654  *	0		default
655  *	1		performance
656  *	2		balance_performance
657  *	3		balance_power
658  *	4		power
659  */
660 
661 enum energy_perf_value_index {
662 	EPP_INDEX_DEFAULT = 0,
663 	EPP_INDEX_PERFORMANCE,
664 	EPP_INDEX_BALANCE_PERFORMANCE,
665 	EPP_INDEX_BALANCE_POWERSAVE,
666 	EPP_INDEX_POWERSAVE,
667 };
668 
669 static const char * const energy_perf_strings[] = {
670 	[EPP_INDEX_DEFAULT] = "default",
671 	[EPP_INDEX_PERFORMANCE] = "performance",
672 	[EPP_INDEX_BALANCE_PERFORMANCE] = "balance_performance",
673 	[EPP_INDEX_BALANCE_POWERSAVE] = "balance_power",
674 	[EPP_INDEX_POWERSAVE] = "power",
675 	NULL
676 };
677 static unsigned int epp_values[] = {
678 	[EPP_INDEX_DEFAULT] = 0, /* Unused index */
679 	[EPP_INDEX_PERFORMANCE] = HWP_EPP_PERFORMANCE,
680 	[EPP_INDEX_BALANCE_PERFORMANCE] = HWP_EPP_BALANCE_PERFORMANCE,
681 	[EPP_INDEX_BALANCE_POWERSAVE] = HWP_EPP_BALANCE_POWERSAVE,
682 	[EPP_INDEX_POWERSAVE] = HWP_EPP_POWERSAVE,
683 };
684 
685 static int intel_pstate_get_energy_pref_index(struct cpudata *cpu_data, int *raw_epp)
686 {
687 	s16 epp;
688 	int index = -EINVAL;
689 
690 	*raw_epp = 0;
691 	epp = intel_pstate_get_epp(cpu_data, 0);
692 	if (epp < 0)
693 		return epp;
694 
695 	if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
696 		if (epp == epp_values[EPP_INDEX_PERFORMANCE])
697 			return EPP_INDEX_PERFORMANCE;
698 		if (epp == epp_values[EPP_INDEX_BALANCE_PERFORMANCE])
699 			return EPP_INDEX_BALANCE_PERFORMANCE;
700 		if (epp == epp_values[EPP_INDEX_BALANCE_POWERSAVE])
701 			return EPP_INDEX_BALANCE_POWERSAVE;
702 		if (epp == epp_values[EPP_INDEX_POWERSAVE])
703 			return EPP_INDEX_POWERSAVE;
704 		*raw_epp = epp;
705 		return 0;
706 	} else if (boot_cpu_has(X86_FEATURE_EPB)) {
707 		/*
708 		 * Range:
709 		 *	0x00-0x03	:	Performance
710 		 *	0x04-0x07	:	Balance performance
711 		 *	0x08-0x0B	:	Balance power
712 		 *	0x0C-0x0F	:	Power
713 		 * The EPB is a 4 bit value, but our ranges restrict the
714 		 * value which can be set. Here only using top two bits
715 		 * effectively.
716 		 */
717 		index = (epp >> 2) + 1;
718 	}
719 
720 	return index;
721 }
722 
723 static int intel_pstate_set_epp(struct cpudata *cpu, u32 epp)
724 {
725 	int ret;
726 
727 	/*
728 	 * Use the cached HWP Request MSR value, because in the active mode the
729 	 * register itself may be updated by intel_pstate_hwp_boost_up() or
730 	 * intel_pstate_hwp_boost_down() at any time.
731 	 */
732 	u64 value = READ_ONCE(cpu->hwp_req_cached);
733 
734 	value &= ~GENMASK_ULL(31, 24);
735 	value |= (u64)epp << 24;
736 	/*
737 	 * The only other updater of hwp_req_cached in the active mode,
738 	 * intel_pstate_hwp_set(), is called under the same lock as this
739 	 * function, so it cannot run in parallel with the update below.
740 	 */
741 	WRITE_ONCE(cpu->hwp_req_cached, value);
742 	ret = wrmsrl_on_cpu(cpu->cpu, MSR_HWP_REQUEST, value);
743 	if (!ret)
744 		cpu->epp_cached = epp;
745 
746 	return ret;
747 }
748 
749 static int intel_pstate_set_energy_pref_index(struct cpudata *cpu_data,
750 					      int pref_index, bool use_raw,
751 					      u32 raw_epp)
752 {
753 	int epp = -EINVAL;
754 	int ret;
755 
756 	if (!pref_index)
757 		epp = cpu_data->epp_default;
758 
759 	if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
760 		if (use_raw)
761 			epp = raw_epp;
762 		else if (epp == -EINVAL)
763 			epp = epp_values[pref_index];
764 
765 		/*
766 		 * To avoid confusion, refuse to set EPP to any values different
767 		 * from 0 (performance) if the current policy is "performance",
768 		 * because those values would be overridden.
769 		 */
770 		if (epp > 0 && cpu_data->policy == CPUFREQ_POLICY_PERFORMANCE)
771 			return -EBUSY;
772 
773 		ret = intel_pstate_set_epp(cpu_data, epp);
774 	} else {
775 		if (epp == -EINVAL)
776 			epp = (pref_index - 1) << 2;
777 		ret = intel_pstate_set_epb(cpu_data->cpu, epp);
778 	}
779 
780 	return ret;
781 }
782 
783 static ssize_t show_energy_performance_available_preferences(
784 				struct cpufreq_policy *policy, char *buf)
785 {
786 	int i = 0;
787 	int ret = 0;
788 
789 	while (energy_perf_strings[i] != NULL)
790 		ret += sprintf(&buf[ret], "%s ", energy_perf_strings[i++]);
791 
792 	ret += sprintf(&buf[ret], "\n");
793 
794 	return ret;
795 }
796 
797 cpufreq_freq_attr_ro(energy_performance_available_preferences);
798 
799 static struct cpufreq_driver intel_pstate;
800 
801 static ssize_t store_energy_performance_preference(
802 		struct cpufreq_policy *policy, const char *buf, size_t count)
803 {
804 	struct cpudata *cpu = all_cpu_data[policy->cpu];
805 	char str_preference[21];
806 	bool raw = false;
807 	ssize_t ret;
808 	u32 epp = 0;
809 
810 	ret = sscanf(buf, "%20s", str_preference);
811 	if (ret != 1)
812 		return -EINVAL;
813 
814 	ret = match_string(energy_perf_strings, -1, str_preference);
815 	if (ret < 0) {
816 		if (!boot_cpu_has(X86_FEATURE_HWP_EPP))
817 			return ret;
818 
819 		ret = kstrtouint(buf, 10, &epp);
820 		if (ret)
821 			return ret;
822 
823 		if (epp > 255)
824 			return -EINVAL;
825 
826 		raw = true;
827 	}
828 
829 	/*
830 	 * This function runs with the policy R/W semaphore held, which
831 	 * guarantees that the driver pointer will not change while it is
832 	 * running.
833 	 */
834 	if (!intel_pstate_driver)
835 		return -EAGAIN;
836 
837 	mutex_lock(&intel_pstate_limits_lock);
838 
839 	if (intel_pstate_driver == &intel_pstate) {
840 		ret = intel_pstate_set_energy_pref_index(cpu, ret, raw, epp);
841 	} else {
842 		/*
843 		 * In the passive mode the governor needs to be stopped on the
844 		 * target CPU before the EPP update and restarted after it,
845 		 * which is super-heavy-weight, so make sure it is worth doing
846 		 * upfront.
847 		 */
848 		if (!raw)
849 			epp = ret ? epp_values[ret] : cpu->epp_default;
850 
851 		if (cpu->epp_cached != epp) {
852 			int err;
853 
854 			cpufreq_stop_governor(policy);
855 			ret = intel_pstate_set_epp(cpu, epp);
856 			err = cpufreq_start_governor(policy);
857 			if (!ret)
858 				ret = err;
859 		} else {
860 			ret = 0;
861 		}
862 	}
863 
864 	mutex_unlock(&intel_pstate_limits_lock);
865 
866 	return ret ?: count;
867 }
868 
869 static ssize_t show_energy_performance_preference(
870 				struct cpufreq_policy *policy, char *buf)
871 {
872 	struct cpudata *cpu_data = all_cpu_data[policy->cpu];
873 	int preference, raw_epp;
874 
875 	preference = intel_pstate_get_energy_pref_index(cpu_data, &raw_epp);
876 	if (preference < 0)
877 		return preference;
878 
879 	if (raw_epp)
880 		return  sprintf(buf, "%d\n", raw_epp);
881 	else
882 		return  sprintf(buf, "%s\n", energy_perf_strings[preference]);
883 }
884 
885 cpufreq_freq_attr_rw(energy_performance_preference);
886 
887 static ssize_t show_base_frequency(struct cpufreq_policy *policy, char *buf)
888 {
889 	struct cpudata *cpu = all_cpu_data[policy->cpu];
890 	int ratio, freq;
891 
892 	ratio = intel_pstate_get_cppc_guaranteed(policy->cpu);
893 	if (ratio <= 0) {
894 		u64 cap;
895 
896 		rdmsrl_on_cpu(policy->cpu, MSR_HWP_CAPABILITIES, &cap);
897 		ratio = HWP_GUARANTEED_PERF(cap);
898 	}
899 
900 	freq = ratio * cpu->pstate.scaling;
901 	if (cpu->pstate.scaling != cpu->pstate.perf_ctl_scaling)
902 		freq = rounddown(freq, cpu->pstate.perf_ctl_scaling);
903 
904 	return sprintf(buf, "%d\n", freq);
905 }
906 
907 cpufreq_freq_attr_ro(base_frequency);
908 
909 static struct freq_attr *hwp_cpufreq_attrs[] = {
910 	&energy_performance_preference,
911 	&energy_performance_available_preferences,
912 	&base_frequency,
913 	NULL,
914 };
915 
916 static void __intel_pstate_get_hwp_cap(struct cpudata *cpu)
917 {
918 	u64 cap;
919 
920 	rdmsrl_on_cpu(cpu->cpu, MSR_HWP_CAPABILITIES, &cap);
921 	WRITE_ONCE(cpu->hwp_cap_cached, cap);
922 	cpu->pstate.max_pstate = HWP_GUARANTEED_PERF(cap);
923 	cpu->pstate.turbo_pstate = HWP_HIGHEST_PERF(cap);
924 }
925 
926 static void intel_pstate_get_hwp_cap(struct cpudata *cpu)
927 {
928 	int scaling = cpu->pstate.scaling;
929 
930 	__intel_pstate_get_hwp_cap(cpu);
931 
932 	cpu->pstate.max_freq = cpu->pstate.max_pstate * scaling;
933 	cpu->pstate.turbo_freq = cpu->pstate.turbo_pstate * scaling;
934 	if (scaling != cpu->pstate.perf_ctl_scaling) {
935 		int perf_ctl_scaling = cpu->pstate.perf_ctl_scaling;
936 
937 		cpu->pstate.max_freq = rounddown(cpu->pstate.max_freq,
938 						 perf_ctl_scaling);
939 		cpu->pstate.turbo_freq = rounddown(cpu->pstate.turbo_freq,
940 						   perf_ctl_scaling);
941 	}
942 }
943 
944 static void intel_pstate_hwp_set(unsigned int cpu)
945 {
946 	struct cpudata *cpu_data = all_cpu_data[cpu];
947 	int max, min;
948 	u64 value;
949 	s16 epp;
950 
951 	max = cpu_data->max_perf_ratio;
952 	min = cpu_data->min_perf_ratio;
953 
954 	if (cpu_data->policy == CPUFREQ_POLICY_PERFORMANCE)
955 		min = max;
956 
957 	rdmsrl_on_cpu(cpu, MSR_HWP_REQUEST, &value);
958 
959 	value &= ~HWP_MIN_PERF(~0L);
960 	value |= HWP_MIN_PERF(min);
961 
962 	value &= ~HWP_MAX_PERF(~0L);
963 	value |= HWP_MAX_PERF(max);
964 
965 	if (cpu_data->epp_policy == cpu_data->policy)
966 		goto skip_epp;
967 
968 	cpu_data->epp_policy = cpu_data->policy;
969 
970 	if (cpu_data->policy == CPUFREQ_POLICY_PERFORMANCE) {
971 		epp = intel_pstate_get_epp(cpu_data, value);
972 		cpu_data->epp_powersave = epp;
973 		/* If EPP read was failed, then don't try to write */
974 		if (epp < 0)
975 			goto skip_epp;
976 
977 		epp = 0;
978 	} else {
979 		/* skip setting EPP, when saved value is invalid */
980 		if (cpu_data->epp_powersave < 0)
981 			goto skip_epp;
982 
983 		/*
984 		 * No need to restore EPP when it is not zero. This
985 		 * means:
986 		 *  - Policy is not changed
987 		 *  - user has manually changed
988 		 *  - Error reading EPB
989 		 */
990 		epp = intel_pstate_get_epp(cpu_data, value);
991 		if (epp)
992 			goto skip_epp;
993 
994 		epp = cpu_data->epp_powersave;
995 	}
996 	if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
997 		value &= ~GENMASK_ULL(31, 24);
998 		value |= (u64)epp << 24;
999 	} else {
1000 		intel_pstate_set_epb(cpu, epp);
1001 	}
1002 skip_epp:
1003 	WRITE_ONCE(cpu_data->hwp_req_cached, value);
1004 	wrmsrl_on_cpu(cpu, MSR_HWP_REQUEST, value);
1005 }
1006 
1007 static void intel_pstate_disable_hwp_interrupt(struct cpudata *cpudata);
1008 
1009 static void intel_pstate_hwp_offline(struct cpudata *cpu)
1010 {
1011 	u64 value = READ_ONCE(cpu->hwp_req_cached);
1012 	int min_perf;
1013 
1014 	intel_pstate_disable_hwp_interrupt(cpu);
1015 
1016 	if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
1017 		/*
1018 		 * In case the EPP has been set to "performance" by the
1019 		 * active mode "performance" scaling algorithm, replace that
1020 		 * temporary value with the cached EPP one.
1021 		 */
1022 		value &= ~GENMASK_ULL(31, 24);
1023 		value |= HWP_ENERGY_PERF_PREFERENCE(cpu->epp_cached);
1024 		/*
1025 		 * However, make sure that EPP will be set to "performance" when
1026 		 * the CPU is brought back online again and the "performance"
1027 		 * scaling algorithm is still in effect.
1028 		 */
1029 		cpu->epp_policy = CPUFREQ_POLICY_UNKNOWN;
1030 	}
1031 
1032 	/*
1033 	 * Clear the desired perf field in the cached HWP request value to
1034 	 * prevent nonzero desired values from being leaked into the active
1035 	 * mode.
1036 	 */
1037 	value &= ~HWP_DESIRED_PERF(~0L);
1038 	WRITE_ONCE(cpu->hwp_req_cached, value);
1039 
1040 	value &= ~GENMASK_ULL(31, 0);
1041 	min_perf = HWP_LOWEST_PERF(READ_ONCE(cpu->hwp_cap_cached));
1042 
1043 	/* Set hwp_max = hwp_min */
1044 	value |= HWP_MAX_PERF(min_perf);
1045 	value |= HWP_MIN_PERF(min_perf);
1046 
1047 	/* Set EPP to min */
1048 	if (boot_cpu_has(X86_FEATURE_HWP_EPP))
1049 		value |= HWP_ENERGY_PERF_PREFERENCE(HWP_EPP_POWERSAVE);
1050 
1051 	wrmsrl_on_cpu(cpu->cpu, MSR_HWP_REQUEST, value);
1052 }
1053 
1054 #define POWER_CTL_EE_ENABLE	1
1055 #define POWER_CTL_EE_DISABLE	2
1056 
1057 static int power_ctl_ee_state;
1058 
1059 static void set_power_ctl_ee_state(bool input)
1060 {
1061 	u64 power_ctl;
1062 
1063 	mutex_lock(&intel_pstate_driver_lock);
1064 	rdmsrl(MSR_IA32_POWER_CTL, power_ctl);
1065 	if (input) {
1066 		power_ctl &= ~BIT(MSR_IA32_POWER_CTL_BIT_EE);
1067 		power_ctl_ee_state = POWER_CTL_EE_ENABLE;
1068 	} else {
1069 		power_ctl |= BIT(MSR_IA32_POWER_CTL_BIT_EE);
1070 		power_ctl_ee_state = POWER_CTL_EE_DISABLE;
1071 	}
1072 	wrmsrl(MSR_IA32_POWER_CTL, power_ctl);
1073 	mutex_unlock(&intel_pstate_driver_lock);
1074 }
1075 
1076 static void intel_pstate_hwp_enable(struct cpudata *cpudata);
1077 
1078 static void intel_pstate_hwp_reenable(struct cpudata *cpu)
1079 {
1080 	intel_pstate_hwp_enable(cpu);
1081 	wrmsrl_on_cpu(cpu->cpu, MSR_HWP_REQUEST, READ_ONCE(cpu->hwp_req_cached));
1082 }
1083 
1084 static int intel_pstate_suspend(struct cpufreq_policy *policy)
1085 {
1086 	struct cpudata *cpu = all_cpu_data[policy->cpu];
1087 
1088 	pr_debug("CPU %d suspending\n", cpu->cpu);
1089 
1090 	cpu->suspended = true;
1091 
1092 	/* disable HWP interrupt and cancel any pending work */
1093 	intel_pstate_disable_hwp_interrupt(cpu);
1094 
1095 	return 0;
1096 }
1097 
1098 static int intel_pstate_resume(struct cpufreq_policy *policy)
1099 {
1100 	struct cpudata *cpu = all_cpu_data[policy->cpu];
1101 
1102 	pr_debug("CPU %d resuming\n", cpu->cpu);
1103 
1104 	/* Only restore if the system default is changed */
1105 	if (power_ctl_ee_state == POWER_CTL_EE_ENABLE)
1106 		set_power_ctl_ee_state(true);
1107 	else if (power_ctl_ee_state == POWER_CTL_EE_DISABLE)
1108 		set_power_ctl_ee_state(false);
1109 
1110 	if (cpu->suspended && hwp_active) {
1111 		mutex_lock(&intel_pstate_limits_lock);
1112 
1113 		/* Re-enable HWP, because "online" has not done that. */
1114 		intel_pstate_hwp_reenable(cpu);
1115 
1116 		mutex_unlock(&intel_pstate_limits_lock);
1117 	}
1118 
1119 	cpu->suspended = false;
1120 
1121 	return 0;
1122 }
1123 
1124 static void intel_pstate_update_policies(void)
1125 {
1126 	int cpu;
1127 
1128 	for_each_possible_cpu(cpu)
1129 		cpufreq_update_policy(cpu);
1130 }
1131 
1132 static void __intel_pstate_update_max_freq(struct cpudata *cpudata,
1133 					   struct cpufreq_policy *policy)
1134 {
1135 	policy->cpuinfo.max_freq = global.turbo_disabled_mf ?
1136 			cpudata->pstate.max_freq : cpudata->pstate.turbo_freq;
1137 	refresh_frequency_limits(policy);
1138 }
1139 
1140 static void intel_pstate_update_max_freq(unsigned int cpu)
1141 {
1142 	struct cpufreq_policy *policy = cpufreq_cpu_acquire(cpu);
1143 
1144 	if (!policy)
1145 		return;
1146 
1147 	__intel_pstate_update_max_freq(all_cpu_data[cpu], policy);
1148 
1149 	cpufreq_cpu_release(policy);
1150 }
1151 
1152 static void intel_pstate_update_limits(unsigned int cpu)
1153 {
1154 	mutex_lock(&intel_pstate_driver_lock);
1155 
1156 	update_turbo_state();
1157 	/*
1158 	 * If turbo has been turned on or off globally, policy limits for
1159 	 * all CPUs need to be updated to reflect that.
1160 	 */
1161 	if (global.turbo_disabled_mf != global.turbo_disabled) {
1162 		global.turbo_disabled_mf = global.turbo_disabled;
1163 		arch_set_max_freq_ratio(global.turbo_disabled);
1164 		for_each_possible_cpu(cpu)
1165 			intel_pstate_update_max_freq(cpu);
1166 	} else {
1167 		cpufreq_update_policy(cpu);
1168 	}
1169 
1170 	mutex_unlock(&intel_pstate_driver_lock);
1171 }
1172 
1173 /************************** sysfs begin ************************/
1174 #define show_one(file_name, object)					\
1175 	static ssize_t show_##file_name					\
1176 	(struct kobject *kobj, struct kobj_attribute *attr, char *buf)	\
1177 	{								\
1178 		return sprintf(buf, "%u\n", global.object);		\
1179 	}
1180 
1181 static ssize_t intel_pstate_show_status(char *buf);
1182 static int intel_pstate_update_status(const char *buf, size_t size);
1183 
1184 static ssize_t show_status(struct kobject *kobj,
1185 			   struct kobj_attribute *attr, char *buf)
1186 {
1187 	ssize_t ret;
1188 
1189 	mutex_lock(&intel_pstate_driver_lock);
1190 	ret = intel_pstate_show_status(buf);
1191 	mutex_unlock(&intel_pstate_driver_lock);
1192 
1193 	return ret;
1194 }
1195 
1196 static ssize_t store_status(struct kobject *a, struct kobj_attribute *b,
1197 			    const char *buf, size_t count)
1198 {
1199 	char *p = memchr(buf, '\n', count);
1200 	int ret;
1201 
1202 	mutex_lock(&intel_pstate_driver_lock);
1203 	ret = intel_pstate_update_status(buf, p ? p - buf : count);
1204 	mutex_unlock(&intel_pstate_driver_lock);
1205 
1206 	return ret < 0 ? ret : count;
1207 }
1208 
1209 static ssize_t show_turbo_pct(struct kobject *kobj,
1210 				struct kobj_attribute *attr, char *buf)
1211 {
1212 	struct cpudata *cpu;
1213 	int total, no_turbo, turbo_pct;
1214 	uint32_t turbo_fp;
1215 
1216 	mutex_lock(&intel_pstate_driver_lock);
1217 
1218 	if (!intel_pstate_driver) {
1219 		mutex_unlock(&intel_pstate_driver_lock);
1220 		return -EAGAIN;
1221 	}
1222 
1223 	cpu = all_cpu_data[0];
1224 
1225 	total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
1226 	no_turbo = cpu->pstate.max_pstate - cpu->pstate.min_pstate + 1;
1227 	turbo_fp = div_fp(no_turbo, total);
1228 	turbo_pct = 100 - fp_toint(mul_fp(turbo_fp, int_tofp(100)));
1229 
1230 	mutex_unlock(&intel_pstate_driver_lock);
1231 
1232 	return sprintf(buf, "%u\n", turbo_pct);
1233 }
1234 
1235 static ssize_t show_num_pstates(struct kobject *kobj,
1236 				struct kobj_attribute *attr, char *buf)
1237 {
1238 	struct cpudata *cpu;
1239 	int total;
1240 
1241 	mutex_lock(&intel_pstate_driver_lock);
1242 
1243 	if (!intel_pstate_driver) {
1244 		mutex_unlock(&intel_pstate_driver_lock);
1245 		return -EAGAIN;
1246 	}
1247 
1248 	cpu = all_cpu_data[0];
1249 	total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
1250 
1251 	mutex_unlock(&intel_pstate_driver_lock);
1252 
1253 	return sprintf(buf, "%u\n", total);
1254 }
1255 
1256 static ssize_t show_no_turbo(struct kobject *kobj,
1257 			     struct kobj_attribute *attr, char *buf)
1258 {
1259 	ssize_t ret;
1260 
1261 	mutex_lock(&intel_pstate_driver_lock);
1262 
1263 	if (!intel_pstate_driver) {
1264 		mutex_unlock(&intel_pstate_driver_lock);
1265 		return -EAGAIN;
1266 	}
1267 
1268 	update_turbo_state();
1269 	if (global.turbo_disabled)
1270 		ret = sprintf(buf, "%u\n", global.turbo_disabled);
1271 	else
1272 		ret = sprintf(buf, "%u\n", global.no_turbo);
1273 
1274 	mutex_unlock(&intel_pstate_driver_lock);
1275 
1276 	return ret;
1277 }
1278 
1279 static ssize_t store_no_turbo(struct kobject *a, struct kobj_attribute *b,
1280 			      const char *buf, size_t count)
1281 {
1282 	unsigned int input;
1283 	int ret;
1284 
1285 	ret = sscanf(buf, "%u", &input);
1286 	if (ret != 1)
1287 		return -EINVAL;
1288 
1289 	mutex_lock(&intel_pstate_driver_lock);
1290 
1291 	if (!intel_pstate_driver) {
1292 		mutex_unlock(&intel_pstate_driver_lock);
1293 		return -EAGAIN;
1294 	}
1295 
1296 	mutex_lock(&intel_pstate_limits_lock);
1297 
1298 	update_turbo_state();
1299 	if (global.turbo_disabled) {
1300 		pr_notice_once("Turbo disabled by BIOS or unavailable on processor\n");
1301 		mutex_unlock(&intel_pstate_limits_lock);
1302 		mutex_unlock(&intel_pstate_driver_lock);
1303 		return -EPERM;
1304 	}
1305 
1306 	global.no_turbo = clamp_t(int, input, 0, 1);
1307 
1308 	if (global.no_turbo) {
1309 		struct cpudata *cpu = all_cpu_data[0];
1310 		int pct = cpu->pstate.max_pstate * 100 / cpu->pstate.turbo_pstate;
1311 
1312 		/* Squash the global minimum into the permitted range. */
1313 		if (global.min_perf_pct > pct)
1314 			global.min_perf_pct = pct;
1315 	}
1316 
1317 	mutex_unlock(&intel_pstate_limits_lock);
1318 
1319 	intel_pstate_update_policies();
1320 	arch_set_max_freq_ratio(global.no_turbo);
1321 
1322 	mutex_unlock(&intel_pstate_driver_lock);
1323 
1324 	return count;
1325 }
1326 
1327 static void update_qos_request(enum freq_qos_req_type type)
1328 {
1329 	struct freq_qos_request *req;
1330 	struct cpufreq_policy *policy;
1331 	int i;
1332 
1333 	for_each_possible_cpu(i) {
1334 		struct cpudata *cpu = all_cpu_data[i];
1335 		unsigned int freq, perf_pct;
1336 
1337 		policy = cpufreq_cpu_get(i);
1338 		if (!policy)
1339 			continue;
1340 
1341 		req = policy->driver_data;
1342 		cpufreq_cpu_put(policy);
1343 
1344 		if (!req)
1345 			continue;
1346 
1347 		if (hwp_active)
1348 			intel_pstate_get_hwp_cap(cpu);
1349 
1350 		if (type == FREQ_QOS_MIN) {
1351 			perf_pct = global.min_perf_pct;
1352 		} else {
1353 			req++;
1354 			perf_pct = global.max_perf_pct;
1355 		}
1356 
1357 		freq = DIV_ROUND_UP(cpu->pstate.turbo_freq * perf_pct, 100);
1358 
1359 		if (freq_qos_update_request(req, freq) < 0)
1360 			pr_warn("Failed to update freq constraint: CPU%d\n", i);
1361 	}
1362 }
1363 
1364 static ssize_t store_max_perf_pct(struct kobject *a, struct kobj_attribute *b,
1365 				  const char *buf, size_t count)
1366 {
1367 	unsigned int input;
1368 	int ret;
1369 
1370 	ret = sscanf(buf, "%u", &input);
1371 	if (ret != 1)
1372 		return -EINVAL;
1373 
1374 	mutex_lock(&intel_pstate_driver_lock);
1375 
1376 	if (!intel_pstate_driver) {
1377 		mutex_unlock(&intel_pstate_driver_lock);
1378 		return -EAGAIN;
1379 	}
1380 
1381 	mutex_lock(&intel_pstate_limits_lock);
1382 
1383 	global.max_perf_pct = clamp_t(int, input, global.min_perf_pct, 100);
1384 
1385 	mutex_unlock(&intel_pstate_limits_lock);
1386 
1387 	if (intel_pstate_driver == &intel_pstate)
1388 		intel_pstate_update_policies();
1389 	else
1390 		update_qos_request(FREQ_QOS_MAX);
1391 
1392 	mutex_unlock(&intel_pstate_driver_lock);
1393 
1394 	return count;
1395 }
1396 
1397 static ssize_t store_min_perf_pct(struct kobject *a, struct kobj_attribute *b,
1398 				  const char *buf, size_t count)
1399 {
1400 	unsigned int input;
1401 	int ret;
1402 
1403 	ret = sscanf(buf, "%u", &input);
1404 	if (ret != 1)
1405 		return -EINVAL;
1406 
1407 	mutex_lock(&intel_pstate_driver_lock);
1408 
1409 	if (!intel_pstate_driver) {
1410 		mutex_unlock(&intel_pstate_driver_lock);
1411 		return -EAGAIN;
1412 	}
1413 
1414 	mutex_lock(&intel_pstate_limits_lock);
1415 
1416 	global.min_perf_pct = clamp_t(int, input,
1417 				      min_perf_pct_min(), global.max_perf_pct);
1418 
1419 	mutex_unlock(&intel_pstate_limits_lock);
1420 
1421 	if (intel_pstate_driver == &intel_pstate)
1422 		intel_pstate_update_policies();
1423 	else
1424 		update_qos_request(FREQ_QOS_MIN);
1425 
1426 	mutex_unlock(&intel_pstate_driver_lock);
1427 
1428 	return count;
1429 }
1430 
1431 static ssize_t show_hwp_dynamic_boost(struct kobject *kobj,
1432 				struct kobj_attribute *attr, char *buf)
1433 {
1434 	return sprintf(buf, "%u\n", hwp_boost);
1435 }
1436 
1437 static ssize_t store_hwp_dynamic_boost(struct kobject *a,
1438 				       struct kobj_attribute *b,
1439 				       const char *buf, size_t count)
1440 {
1441 	unsigned int input;
1442 	int ret;
1443 
1444 	ret = kstrtouint(buf, 10, &input);
1445 	if (ret)
1446 		return ret;
1447 
1448 	mutex_lock(&intel_pstate_driver_lock);
1449 	hwp_boost = !!input;
1450 	intel_pstate_update_policies();
1451 	mutex_unlock(&intel_pstate_driver_lock);
1452 
1453 	return count;
1454 }
1455 
1456 static ssize_t show_energy_efficiency(struct kobject *kobj, struct kobj_attribute *attr,
1457 				      char *buf)
1458 {
1459 	u64 power_ctl;
1460 	int enable;
1461 
1462 	rdmsrl(MSR_IA32_POWER_CTL, power_ctl);
1463 	enable = !!(power_ctl & BIT(MSR_IA32_POWER_CTL_BIT_EE));
1464 	return sprintf(buf, "%d\n", !enable);
1465 }
1466 
1467 static ssize_t store_energy_efficiency(struct kobject *a, struct kobj_attribute *b,
1468 				       const char *buf, size_t count)
1469 {
1470 	bool input;
1471 	int ret;
1472 
1473 	ret = kstrtobool(buf, &input);
1474 	if (ret)
1475 		return ret;
1476 
1477 	set_power_ctl_ee_state(input);
1478 
1479 	return count;
1480 }
1481 
1482 show_one(max_perf_pct, max_perf_pct);
1483 show_one(min_perf_pct, min_perf_pct);
1484 
1485 define_one_global_rw(status);
1486 define_one_global_rw(no_turbo);
1487 define_one_global_rw(max_perf_pct);
1488 define_one_global_rw(min_perf_pct);
1489 define_one_global_ro(turbo_pct);
1490 define_one_global_ro(num_pstates);
1491 define_one_global_rw(hwp_dynamic_boost);
1492 define_one_global_rw(energy_efficiency);
1493 
1494 static struct attribute *intel_pstate_attributes[] = {
1495 	&status.attr,
1496 	&no_turbo.attr,
1497 	NULL
1498 };
1499 
1500 static const struct attribute_group intel_pstate_attr_group = {
1501 	.attrs = intel_pstate_attributes,
1502 };
1503 
1504 static const struct x86_cpu_id intel_pstate_cpu_ee_disable_ids[];
1505 
1506 static struct kobject *intel_pstate_kobject;
1507 
1508 static void __init intel_pstate_sysfs_expose_params(void)
1509 {
1510 	struct device *dev_root = bus_get_dev_root(&cpu_subsys);
1511 	int rc;
1512 
1513 	if (dev_root) {
1514 		intel_pstate_kobject = kobject_create_and_add("intel_pstate", &dev_root->kobj);
1515 		put_device(dev_root);
1516 	}
1517 	if (WARN_ON(!intel_pstate_kobject))
1518 		return;
1519 
1520 	rc = sysfs_create_group(intel_pstate_kobject, &intel_pstate_attr_group);
1521 	if (WARN_ON(rc))
1522 		return;
1523 
1524 	if (!boot_cpu_has(X86_FEATURE_HYBRID_CPU)) {
1525 		rc = sysfs_create_file(intel_pstate_kobject, &turbo_pct.attr);
1526 		WARN_ON(rc);
1527 
1528 		rc = sysfs_create_file(intel_pstate_kobject, &num_pstates.attr);
1529 		WARN_ON(rc);
1530 	}
1531 
1532 	/*
1533 	 * If per cpu limits are enforced there are no global limits, so
1534 	 * return without creating max/min_perf_pct attributes
1535 	 */
1536 	if (per_cpu_limits)
1537 		return;
1538 
1539 	rc = sysfs_create_file(intel_pstate_kobject, &max_perf_pct.attr);
1540 	WARN_ON(rc);
1541 
1542 	rc = sysfs_create_file(intel_pstate_kobject, &min_perf_pct.attr);
1543 	WARN_ON(rc);
1544 
1545 	if (x86_match_cpu(intel_pstate_cpu_ee_disable_ids)) {
1546 		rc = sysfs_create_file(intel_pstate_kobject, &energy_efficiency.attr);
1547 		WARN_ON(rc);
1548 	}
1549 }
1550 
1551 static void __init intel_pstate_sysfs_remove(void)
1552 {
1553 	if (!intel_pstate_kobject)
1554 		return;
1555 
1556 	sysfs_remove_group(intel_pstate_kobject, &intel_pstate_attr_group);
1557 
1558 	if (!boot_cpu_has(X86_FEATURE_HYBRID_CPU)) {
1559 		sysfs_remove_file(intel_pstate_kobject, &num_pstates.attr);
1560 		sysfs_remove_file(intel_pstate_kobject, &turbo_pct.attr);
1561 	}
1562 
1563 	if (!per_cpu_limits) {
1564 		sysfs_remove_file(intel_pstate_kobject, &max_perf_pct.attr);
1565 		sysfs_remove_file(intel_pstate_kobject, &min_perf_pct.attr);
1566 
1567 		if (x86_match_cpu(intel_pstate_cpu_ee_disable_ids))
1568 			sysfs_remove_file(intel_pstate_kobject, &energy_efficiency.attr);
1569 	}
1570 
1571 	kobject_put(intel_pstate_kobject);
1572 }
1573 
1574 static void intel_pstate_sysfs_expose_hwp_dynamic_boost(void)
1575 {
1576 	int rc;
1577 
1578 	if (!hwp_active)
1579 		return;
1580 
1581 	rc = sysfs_create_file(intel_pstate_kobject, &hwp_dynamic_boost.attr);
1582 	WARN_ON_ONCE(rc);
1583 }
1584 
1585 static void intel_pstate_sysfs_hide_hwp_dynamic_boost(void)
1586 {
1587 	if (!hwp_active)
1588 		return;
1589 
1590 	sysfs_remove_file(intel_pstate_kobject, &hwp_dynamic_boost.attr);
1591 }
1592 
1593 /************************** sysfs end ************************/
1594 
1595 static void intel_pstate_notify_work(struct work_struct *work)
1596 {
1597 	struct cpudata *cpudata =
1598 		container_of(to_delayed_work(work), struct cpudata, hwp_notify_work);
1599 	struct cpufreq_policy *policy = cpufreq_cpu_acquire(cpudata->cpu);
1600 
1601 	if (policy) {
1602 		intel_pstate_get_hwp_cap(cpudata);
1603 		__intel_pstate_update_max_freq(cpudata, policy);
1604 
1605 		cpufreq_cpu_release(policy);
1606 	}
1607 
1608 	wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_STATUS, 0);
1609 }
1610 
1611 static DEFINE_SPINLOCK(hwp_notify_lock);
1612 static cpumask_t hwp_intr_enable_mask;
1613 
1614 void notify_hwp_interrupt(void)
1615 {
1616 	unsigned int this_cpu = smp_processor_id();
1617 	struct cpudata *cpudata;
1618 	unsigned long flags;
1619 	u64 value;
1620 
1621 	if (!READ_ONCE(hwp_active) || !boot_cpu_has(X86_FEATURE_HWP_NOTIFY))
1622 		return;
1623 
1624 	rdmsrl_safe(MSR_HWP_STATUS, &value);
1625 	if (!(value & 0x01))
1626 		return;
1627 
1628 	spin_lock_irqsave(&hwp_notify_lock, flags);
1629 
1630 	if (!cpumask_test_cpu(this_cpu, &hwp_intr_enable_mask))
1631 		goto ack_intr;
1632 
1633 	/*
1634 	 * Currently we never free all_cpu_data. And we can't reach here
1635 	 * without this allocated. But for safety for future changes, added
1636 	 * check.
1637 	 */
1638 	if (unlikely(!READ_ONCE(all_cpu_data)))
1639 		goto ack_intr;
1640 
1641 	/*
1642 	 * The free is done during cleanup, when cpufreq registry is failed.
1643 	 * We wouldn't be here if it fails on init or switch status. But for
1644 	 * future changes, added check.
1645 	 */
1646 	cpudata = READ_ONCE(all_cpu_data[this_cpu]);
1647 	if (unlikely(!cpudata))
1648 		goto ack_intr;
1649 
1650 	schedule_delayed_work(&cpudata->hwp_notify_work, msecs_to_jiffies(10));
1651 
1652 	spin_unlock_irqrestore(&hwp_notify_lock, flags);
1653 
1654 	return;
1655 
1656 ack_intr:
1657 	wrmsrl_safe(MSR_HWP_STATUS, 0);
1658 	spin_unlock_irqrestore(&hwp_notify_lock, flags);
1659 }
1660 
1661 static void intel_pstate_disable_hwp_interrupt(struct cpudata *cpudata)
1662 {
1663 	unsigned long flags;
1664 
1665 	if (!boot_cpu_has(X86_FEATURE_HWP_NOTIFY))
1666 		return;
1667 
1668 	/* wrmsrl_on_cpu has to be outside spinlock as this can result in IPC */
1669 	wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_INTERRUPT, 0x00);
1670 
1671 	spin_lock_irqsave(&hwp_notify_lock, flags);
1672 	if (cpumask_test_and_clear_cpu(cpudata->cpu, &hwp_intr_enable_mask))
1673 		cancel_delayed_work(&cpudata->hwp_notify_work);
1674 	spin_unlock_irqrestore(&hwp_notify_lock, flags);
1675 }
1676 
1677 static void intel_pstate_enable_hwp_interrupt(struct cpudata *cpudata)
1678 {
1679 	/* Enable HWP notification interrupt for guaranteed performance change */
1680 	if (boot_cpu_has(X86_FEATURE_HWP_NOTIFY)) {
1681 		unsigned long flags;
1682 
1683 		spin_lock_irqsave(&hwp_notify_lock, flags);
1684 		INIT_DELAYED_WORK(&cpudata->hwp_notify_work, intel_pstate_notify_work);
1685 		cpumask_set_cpu(cpudata->cpu, &hwp_intr_enable_mask);
1686 		spin_unlock_irqrestore(&hwp_notify_lock, flags);
1687 
1688 		/* wrmsrl_on_cpu has to be outside spinlock as this can result in IPC */
1689 		wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_INTERRUPT, 0x01);
1690 		wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_STATUS, 0);
1691 	}
1692 }
1693 
1694 static void intel_pstate_update_epp_defaults(struct cpudata *cpudata)
1695 {
1696 	cpudata->epp_default = intel_pstate_get_epp(cpudata, 0);
1697 
1698 	/*
1699 	 * If this CPU gen doesn't call for change in balance_perf
1700 	 * EPP return.
1701 	 */
1702 	if (epp_values[EPP_INDEX_BALANCE_PERFORMANCE] == HWP_EPP_BALANCE_PERFORMANCE)
1703 		return;
1704 
1705 	/*
1706 	 * If the EPP is set by firmware, which means that firmware enabled HWP
1707 	 * - Is equal or less than 0x80 (default balance_perf EPP)
1708 	 * - But less performance oriented than performance EPP
1709 	 *   then use this as new balance_perf EPP.
1710 	 */
1711 	if (hwp_forced && cpudata->epp_default <= HWP_EPP_BALANCE_PERFORMANCE &&
1712 	    cpudata->epp_default > HWP_EPP_PERFORMANCE) {
1713 		epp_values[EPP_INDEX_BALANCE_PERFORMANCE] = cpudata->epp_default;
1714 		return;
1715 	}
1716 
1717 	/*
1718 	 * Use hard coded value per gen to update the balance_perf
1719 	 * and default EPP.
1720 	 */
1721 	cpudata->epp_default = epp_values[EPP_INDEX_BALANCE_PERFORMANCE];
1722 	intel_pstate_set_epp(cpudata, cpudata->epp_default);
1723 }
1724 
1725 static void intel_pstate_hwp_enable(struct cpudata *cpudata)
1726 {
1727 	/* First disable HWP notification interrupt till we activate again */
1728 	if (boot_cpu_has(X86_FEATURE_HWP_NOTIFY))
1729 		wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_INTERRUPT, 0x00);
1730 
1731 	wrmsrl_on_cpu(cpudata->cpu, MSR_PM_ENABLE, 0x1);
1732 
1733 	intel_pstate_enable_hwp_interrupt(cpudata);
1734 
1735 	if (cpudata->epp_default >= 0)
1736 		return;
1737 
1738 	intel_pstate_update_epp_defaults(cpudata);
1739 }
1740 
1741 static int atom_get_min_pstate(int not_used)
1742 {
1743 	u64 value;
1744 
1745 	rdmsrl(MSR_ATOM_CORE_RATIOS, value);
1746 	return (value >> 8) & 0x7F;
1747 }
1748 
1749 static int atom_get_max_pstate(int not_used)
1750 {
1751 	u64 value;
1752 
1753 	rdmsrl(MSR_ATOM_CORE_RATIOS, value);
1754 	return (value >> 16) & 0x7F;
1755 }
1756 
1757 static int atom_get_turbo_pstate(int not_used)
1758 {
1759 	u64 value;
1760 
1761 	rdmsrl(MSR_ATOM_CORE_TURBO_RATIOS, value);
1762 	return value & 0x7F;
1763 }
1764 
1765 static u64 atom_get_val(struct cpudata *cpudata, int pstate)
1766 {
1767 	u64 val;
1768 	int32_t vid_fp;
1769 	u32 vid;
1770 
1771 	val = (u64)pstate << 8;
1772 	if (global.no_turbo && !global.turbo_disabled)
1773 		val |= (u64)1 << 32;
1774 
1775 	vid_fp = cpudata->vid.min + mul_fp(
1776 		int_tofp(pstate - cpudata->pstate.min_pstate),
1777 		cpudata->vid.ratio);
1778 
1779 	vid_fp = clamp_t(int32_t, vid_fp, cpudata->vid.min, cpudata->vid.max);
1780 	vid = ceiling_fp(vid_fp);
1781 
1782 	if (pstate > cpudata->pstate.max_pstate)
1783 		vid = cpudata->vid.turbo;
1784 
1785 	return val | vid;
1786 }
1787 
1788 static int silvermont_get_scaling(void)
1789 {
1790 	u64 value;
1791 	int i;
1792 	/* Defined in Table 35-6 from SDM (Sept 2015) */
1793 	static int silvermont_freq_table[] = {
1794 		83300, 100000, 133300, 116700, 80000};
1795 
1796 	rdmsrl(MSR_FSB_FREQ, value);
1797 	i = value & 0x7;
1798 	WARN_ON(i > 4);
1799 
1800 	return silvermont_freq_table[i];
1801 }
1802 
1803 static int airmont_get_scaling(void)
1804 {
1805 	u64 value;
1806 	int i;
1807 	/* Defined in Table 35-10 from SDM (Sept 2015) */
1808 	static int airmont_freq_table[] = {
1809 		83300, 100000, 133300, 116700, 80000,
1810 		93300, 90000, 88900, 87500};
1811 
1812 	rdmsrl(MSR_FSB_FREQ, value);
1813 	i = value & 0xF;
1814 	WARN_ON(i > 8);
1815 
1816 	return airmont_freq_table[i];
1817 }
1818 
1819 static void atom_get_vid(struct cpudata *cpudata)
1820 {
1821 	u64 value;
1822 
1823 	rdmsrl(MSR_ATOM_CORE_VIDS, value);
1824 	cpudata->vid.min = int_tofp((value >> 8) & 0x7f);
1825 	cpudata->vid.max = int_tofp((value >> 16) & 0x7f);
1826 	cpudata->vid.ratio = div_fp(
1827 		cpudata->vid.max - cpudata->vid.min,
1828 		int_tofp(cpudata->pstate.max_pstate -
1829 			cpudata->pstate.min_pstate));
1830 
1831 	rdmsrl(MSR_ATOM_CORE_TURBO_VIDS, value);
1832 	cpudata->vid.turbo = value & 0x7f;
1833 }
1834 
1835 static int core_get_min_pstate(int cpu)
1836 {
1837 	u64 value;
1838 
1839 	rdmsrl_on_cpu(cpu, MSR_PLATFORM_INFO, &value);
1840 	return (value >> 40) & 0xFF;
1841 }
1842 
1843 static int core_get_max_pstate_physical(int cpu)
1844 {
1845 	u64 value;
1846 
1847 	rdmsrl_on_cpu(cpu, MSR_PLATFORM_INFO, &value);
1848 	return (value >> 8) & 0xFF;
1849 }
1850 
1851 static int core_get_tdp_ratio(int cpu, u64 plat_info)
1852 {
1853 	/* Check how many TDP levels present */
1854 	if (plat_info & 0x600000000) {
1855 		u64 tdp_ctrl;
1856 		u64 tdp_ratio;
1857 		int tdp_msr;
1858 		int err;
1859 
1860 		/* Get the TDP level (0, 1, 2) to get ratios */
1861 		err = rdmsrl_safe_on_cpu(cpu, MSR_CONFIG_TDP_CONTROL, &tdp_ctrl);
1862 		if (err)
1863 			return err;
1864 
1865 		/* TDP MSR are continuous starting at 0x648 */
1866 		tdp_msr = MSR_CONFIG_TDP_NOMINAL + (tdp_ctrl & 0x03);
1867 		err = rdmsrl_safe_on_cpu(cpu, tdp_msr, &tdp_ratio);
1868 		if (err)
1869 			return err;
1870 
1871 		/* For level 1 and 2, bits[23:16] contain the ratio */
1872 		if (tdp_ctrl & 0x03)
1873 			tdp_ratio >>= 16;
1874 
1875 		tdp_ratio &= 0xff; /* ratios are only 8 bits long */
1876 		pr_debug("tdp_ratio %x\n", (int)tdp_ratio);
1877 
1878 		return (int)tdp_ratio;
1879 	}
1880 
1881 	return -ENXIO;
1882 }
1883 
1884 static int core_get_max_pstate(int cpu)
1885 {
1886 	u64 tar;
1887 	u64 plat_info;
1888 	int max_pstate;
1889 	int tdp_ratio;
1890 	int err;
1891 
1892 	rdmsrl_on_cpu(cpu, MSR_PLATFORM_INFO, &plat_info);
1893 	max_pstate = (plat_info >> 8) & 0xFF;
1894 
1895 	tdp_ratio = core_get_tdp_ratio(cpu, plat_info);
1896 	if (tdp_ratio <= 0)
1897 		return max_pstate;
1898 
1899 	if (hwp_active) {
1900 		/* Turbo activation ratio is not used on HWP platforms */
1901 		return tdp_ratio;
1902 	}
1903 
1904 	err = rdmsrl_safe_on_cpu(cpu, MSR_TURBO_ACTIVATION_RATIO, &tar);
1905 	if (!err) {
1906 		int tar_levels;
1907 
1908 		/* Do some sanity checking for safety */
1909 		tar_levels = tar & 0xff;
1910 		if (tdp_ratio - 1 == tar_levels) {
1911 			max_pstate = tar_levels;
1912 			pr_debug("max_pstate=TAC %x\n", max_pstate);
1913 		}
1914 	}
1915 
1916 	return max_pstate;
1917 }
1918 
1919 static int core_get_turbo_pstate(int cpu)
1920 {
1921 	u64 value;
1922 	int nont, ret;
1923 
1924 	rdmsrl_on_cpu(cpu, MSR_TURBO_RATIO_LIMIT, &value);
1925 	nont = core_get_max_pstate(cpu);
1926 	ret = (value) & 255;
1927 	if (ret <= nont)
1928 		ret = nont;
1929 	return ret;
1930 }
1931 
1932 static u64 core_get_val(struct cpudata *cpudata, int pstate)
1933 {
1934 	u64 val;
1935 
1936 	val = (u64)pstate << 8;
1937 	if (global.no_turbo && !global.turbo_disabled)
1938 		val |= (u64)1 << 32;
1939 
1940 	return val;
1941 }
1942 
1943 static int knl_get_aperf_mperf_shift(void)
1944 {
1945 	return 10;
1946 }
1947 
1948 static int knl_get_turbo_pstate(int cpu)
1949 {
1950 	u64 value;
1951 	int nont, ret;
1952 
1953 	rdmsrl_on_cpu(cpu, MSR_TURBO_RATIO_LIMIT, &value);
1954 	nont = core_get_max_pstate(cpu);
1955 	ret = (((value) >> 8) & 0xFF);
1956 	if (ret <= nont)
1957 		ret = nont;
1958 	return ret;
1959 }
1960 
1961 static void hybrid_get_type(void *data)
1962 {
1963 	u8 *cpu_type = data;
1964 
1965 	*cpu_type = get_this_hybrid_cpu_type();
1966 }
1967 
1968 static int hwp_get_cpu_scaling(int cpu)
1969 {
1970 	u8 cpu_type = 0;
1971 
1972 	smp_call_function_single(cpu, hybrid_get_type, &cpu_type, 1);
1973 	/* P-cores have a smaller perf level-to-freqency scaling factor. */
1974 	if (cpu_type == 0x40)
1975 		return HYBRID_SCALING_FACTOR;
1976 
1977 	/* Use default core scaling for E-cores */
1978 	if (cpu_type == 0x20)
1979 		return core_get_scaling();
1980 
1981 	/*
1982 	 * If reached here, this system is either non-hybrid (like Tiger
1983 	 * Lake) or hybrid-capable (like Alder Lake or Raptor Lake) with
1984 	 * no E cores (in which case CPUID for hybrid support is 0).
1985 	 *
1986 	 * The CPPC nominal_frequency field is 0 for non-hybrid systems,
1987 	 * so the default core scaling will be used for them.
1988 	 */
1989 	return intel_pstate_cppc_get_scaling(cpu);
1990 }
1991 
1992 static void intel_pstate_set_pstate(struct cpudata *cpu, int pstate)
1993 {
1994 	trace_cpu_frequency(pstate * cpu->pstate.scaling, cpu->cpu);
1995 	cpu->pstate.current_pstate = pstate;
1996 	/*
1997 	 * Generally, there is no guarantee that this code will always run on
1998 	 * the CPU being updated, so force the register update to run on the
1999 	 * right CPU.
2000 	 */
2001 	wrmsrl_on_cpu(cpu->cpu, MSR_IA32_PERF_CTL,
2002 		      pstate_funcs.get_val(cpu, pstate));
2003 }
2004 
2005 static void intel_pstate_set_min_pstate(struct cpudata *cpu)
2006 {
2007 	intel_pstate_set_pstate(cpu, cpu->pstate.min_pstate);
2008 }
2009 
2010 static void intel_pstate_max_within_limits(struct cpudata *cpu)
2011 {
2012 	int pstate = max(cpu->pstate.min_pstate, cpu->max_perf_ratio);
2013 
2014 	update_turbo_state();
2015 	intel_pstate_set_pstate(cpu, pstate);
2016 }
2017 
2018 static void intel_pstate_get_cpu_pstates(struct cpudata *cpu)
2019 {
2020 	int perf_ctl_max_phys = pstate_funcs.get_max_physical(cpu->cpu);
2021 	int perf_ctl_scaling = pstate_funcs.get_scaling();
2022 
2023 	cpu->pstate.min_pstate = pstate_funcs.get_min(cpu->cpu);
2024 	cpu->pstate.max_pstate_physical = perf_ctl_max_phys;
2025 	cpu->pstate.perf_ctl_scaling = perf_ctl_scaling;
2026 
2027 	if (hwp_active && !hwp_mode_bdw) {
2028 		__intel_pstate_get_hwp_cap(cpu);
2029 
2030 		if (pstate_funcs.get_cpu_scaling) {
2031 			cpu->pstate.scaling = pstate_funcs.get_cpu_scaling(cpu->cpu);
2032 			if (cpu->pstate.scaling != perf_ctl_scaling)
2033 				intel_pstate_hybrid_hwp_adjust(cpu);
2034 		} else {
2035 			cpu->pstate.scaling = perf_ctl_scaling;
2036 		}
2037 	} else {
2038 		cpu->pstate.scaling = perf_ctl_scaling;
2039 		cpu->pstate.max_pstate = pstate_funcs.get_max(cpu->cpu);
2040 		cpu->pstate.turbo_pstate = pstate_funcs.get_turbo(cpu->cpu);
2041 	}
2042 
2043 	if (cpu->pstate.scaling == perf_ctl_scaling) {
2044 		cpu->pstate.min_freq = cpu->pstate.min_pstate * perf_ctl_scaling;
2045 		cpu->pstate.max_freq = cpu->pstate.max_pstate * perf_ctl_scaling;
2046 		cpu->pstate.turbo_freq = cpu->pstate.turbo_pstate * perf_ctl_scaling;
2047 	}
2048 
2049 	if (pstate_funcs.get_aperf_mperf_shift)
2050 		cpu->aperf_mperf_shift = pstate_funcs.get_aperf_mperf_shift();
2051 
2052 	if (pstate_funcs.get_vid)
2053 		pstate_funcs.get_vid(cpu);
2054 
2055 	intel_pstate_set_min_pstate(cpu);
2056 }
2057 
2058 /*
2059  * Long hold time will keep high perf limits for long time,
2060  * which negatively impacts perf/watt for some workloads,
2061  * like specpower. 3ms is based on experiements on some
2062  * workoads.
2063  */
2064 static int hwp_boost_hold_time_ns = 3 * NSEC_PER_MSEC;
2065 
2066 static inline void intel_pstate_hwp_boost_up(struct cpudata *cpu)
2067 {
2068 	u64 hwp_req = READ_ONCE(cpu->hwp_req_cached);
2069 	u64 hwp_cap = READ_ONCE(cpu->hwp_cap_cached);
2070 	u32 max_limit = (hwp_req & 0xff00) >> 8;
2071 	u32 min_limit = (hwp_req & 0xff);
2072 	u32 boost_level1;
2073 
2074 	/*
2075 	 * Cases to consider (User changes via sysfs or boot time):
2076 	 * If, P0 (Turbo max) = P1 (Guaranteed max) = min:
2077 	 *	No boost, return.
2078 	 * If, P0 (Turbo max) > P1 (Guaranteed max) = min:
2079 	 *     Should result in one level boost only for P0.
2080 	 * If, P0 (Turbo max) = P1 (Guaranteed max) > min:
2081 	 *     Should result in two level boost:
2082 	 *         (min + p1)/2 and P1.
2083 	 * If, P0 (Turbo max) > P1 (Guaranteed max) > min:
2084 	 *     Should result in three level boost:
2085 	 *        (min + p1)/2, P1 and P0.
2086 	 */
2087 
2088 	/* If max and min are equal or already at max, nothing to boost */
2089 	if (max_limit == min_limit || cpu->hwp_boost_min >= max_limit)
2090 		return;
2091 
2092 	if (!cpu->hwp_boost_min)
2093 		cpu->hwp_boost_min = min_limit;
2094 
2095 	/* level at half way mark between min and guranteed */
2096 	boost_level1 = (HWP_GUARANTEED_PERF(hwp_cap) + min_limit) >> 1;
2097 
2098 	if (cpu->hwp_boost_min < boost_level1)
2099 		cpu->hwp_boost_min = boost_level1;
2100 	else if (cpu->hwp_boost_min < HWP_GUARANTEED_PERF(hwp_cap))
2101 		cpu->hwp_boost_min = HWP_GUARANTEED_PERF(hwp_cap);
2102 	else if (cpu->hwp_boost_min == HWP_GUARANTEED_PERF(hwp_cap) &&
2103 		 max_limit != HWP_GUARANTEED_PERF(hwp_cap))
2104 		cpu->hwp_boost_min = max_limit;
2105 	else
2106 		return;
2107 
2108 	hwp_req = (hwp_req & ~GENMASK_ULL(7, 0)) | cpu->hwp_boost_min;
2109 	wrmsrl(MSR_HWP_REQUEST, hwp_req);
2110 	cpu->last_update = cpu->sample.time;
2111 }
2112 
2113 static inline void intel_pstate_hwp_boost_down(struct cpudata *cpu)
2114 {
2115 	if (cpu->hwp_boost_min) {
2116 		bool expired;
2117 
2118 		/* Check if we are idle for hold time to boost down */
2119 		expired = time_after64(cpu->sample.time, cpu->last_update +
2120 				       hwp_boost_hold_time_ns);
2121 		if (expired) {
2122 			wrmsrl(MSR_HWP_REQUEST, cpu->hwp_req_cached);
2123 			cpu->hwp_boost_min = 0;
2124 		}
2125 	}
2126 	cpu->last_update = cpu->sample.time;
2127 }
2128 
2129 static inline void intel_pstate_update_util_hwp_local(struct cpudata *cpu,
2130 						      u64 time)
2131 {
2132 	cpu->sample.time = time;
2133 
2134 	if (cpu->sched_flags & SCHED_CPUFREQ_IOWAIT) {
2135 		bool do_io = false;
2136 
2137 		cpu->sched_flags = 0;
2138 		/*
2139 		 * Set iowait_boost flag and update time. Since IO WAIT flag
2140 		 * is set all the time, we can't just conclude that there is
2141 		 * some IO bound activity is scheduled on this CPU with just
2142 		 * one occurrence. If we receive at least two in two
2143 		 * consecutive ticks, then we treat as boost candidate.
2144 		 */
2145 		if (time_before64(time, cpu->last_io_update + 2 * TICK_NSEC))
2146 			do_io = true;
2147 
2148 		cpu->last_io_update = time;
2149 
2150 		if (do_io)
2151 			intel_pstate_hwp_boost_up(cpu);
2152 
2153 	} else {
2154 		intel_pstate_hwp_boost_down(cpu);
2155 	}
2156 }
2157 
2158 static inline void intel_pstate_update_util_hwp(struct update_util_data *data,
2159 						u64 time, unsigned int flags)
2160 {
2161 	struct cpudata *cpu = container_of(data, struct cpudata, update_util);
2162 
2163 	cpu->sched_flags |= flags;
2164 
2165 	if (smp_processor_id() == cpu->cpu)
2166 		intel_pstate_update_util_hwp_local(cpu, time);
2167 }
2168 
2169 static inline void intel_pstate_calc_avg_perf(struct cpudata *cpu)
2170 {
2171 	struct sample *sample = &cpu->sample;
2172 
2173 	sample->core_avg_perf = div_ext_fp(sample->aperf, sample->mperf);
2174 }
2175 
2176 static inline bool intel_pstate_sample(struct cpudata *cpu, u64 time)
2177 {
2178 	u64 aperf, mperf;
2179 	unsigned long flags;
2180 	u64 tsc;
2181 
2182 	local_irq_save(flags);
2183 	rdmsrl(MSR_IA32_APERF, aperf);
2184 	rdmsrl(MSR_IA32_MPERF, mperf);
2185 	tsc = rdtsc();
2186 	if (cpu->prev_mperf == mperf || cpu->prev_tsc == tsc) {
2187 		local_irq_restore(flags);
2188 		return false;
2189 	}
2190 	local_irq_restore(flags);
2191 
2192 	cpu->last_sample_time = cpu->sample.time;
2193 	cpu->sample.time = time;
2194 	cpu->sample.aperf = aperf;
2195 	cpu->sample.mperf = mperf;
2196 	cpu->sample.tsc =  tsc;
2197 	cpu->sample.aperf -= cpu->prev_aperf;
2198 	cpu->sample.mperf -= cpu->prev_mperf;
2199 	cpu->sample.tsc -= cpu->prev_tsc;
2200 
2201 	cpu->prev_aperf = aperf;
2202 	cpu->prev_mperf = mperf;
2203 	cpu->prev_tsc = tsc;
2204 	/*
2205 	 * First time this function is invoked in a given cycle, all of the
2206 	 * previous sample data fields are equal to zero or stale and they must
2207 	 * be populated with meaningful numbers for things to work, so assume
2208 	 * that sample.time will always be reset before setting the utilization
2209 	 * update hook and make the caller skip the sample then.
2210 	 */
2211 	if (cpu->last_sample_time) {
2212 		intel_pstate_calc_avg_perf(cpu);
2213 		return true;
2214 	}
2215 	return false;
2216 }
2217 
2218 static inline int32_t get_avg_frequency(struct cpudata *cpu)
2219 {
2220 	return mul_ext_fp(cpu->sample.core_avg_perf, cpu_khz);
2221 }
2222 
2223 static inline int32_t get_avg_pstate(struct cpudata *cpu)
2224 {
2225 	return mul_ext_fp(cpu->pstate.max_pstate_physical,
2226 			  cpu->sample.core_avg_perf);
2227 }
2228 
2229 static inline int32_t get_target_pstate(struct cpudata *cpu)
2230 {
2231 	struct sample *sample = &cpu->sample;
2232 	int32_t busy_frac;
2233 	int target, avg_pstate;
2234 
2235 	busy_frac = div_fp(sample->mperf << cpu->aperf_mperf_shift,
2236 			   sample->tsc);
2237 
2238 	if (busy_frac < cpu->iowait_boost)
2239 		busy_frac = cpu->iowait_boost;
2240 
2241 	sample->busy_scaled = busy_frac * 100;
2242 
2243 	target = global.no_turbo || global.turbo_disabled ?
2244 			cpu->pstate.max_pstate : cpu->pstate.turbo_pstate;
2245 	target += target >> 2;
2246 	target = mul_fp(target, busy_frac);
2247 	if (target < cpu->pstate.min_pstate)
2248 		target = cpu->pstate.min_pstate;
2249 
2250 	/*
2251 	 * If the average P-state during the previous cycle was higher than the
2252 	 * current target, add 50% of the difference to the target to reduce
2253 	 * possible performance oscillations and offset possible performance
2254 	 * loss related to moving the workload from one CPU to another within
2255 	 * a package/module.
2256 	 */
2257 	avg_pstate = get_avg_pstate(cpu);
2258 	if (avg_pstate > target)
2259 		target += (avg_pstate - target) >> 1;
2260 
2261 	return target;
2262 }
2263 
2264 static int intel_pstate_prepare_request(struct cpudata *cpu, int pstate)
2265 {
2266 	int min_pstate = max(cpu->pstate.min_pstate, cpu->min_perf_ratio);
2267 	int max_pstate = max(min_pstate, cpu->max_perf_ratio);
2268 
2269 	return clamp_t(int, pstate, min_pstate, max_pstate);
2270 }
2271 
2272 static void intel_pstate_update_pstate(struct cpudata *cpu, int pstate)
2273 {
2274 	if (pstate == cpu->pstate.current_pstate)
2275 		return;
2276 
2277 	cpu->pstate.current_pstate = pstate;
2278 	wrmsrl(MSR_IA32_PERF_CTL, pstate_funcs.get_val(cpu, pstate));
2279 }
2280 
2281 static void intel_pstate_adjust_pstate(struct cpudata *cpu)
2282 {
2283 	int from = cpu->pstate.current_pstate;
2284 	struct sample *sample;
2285 	int target_pstate;
2286 
2287 	update_turbo_state();
2288 
2289 	target_pstate = get_target_pstate(cpu);
2290 	target_pstate = intel_pstate_prepare_request(cpu, target_pstate);
2291 	trace_cpu_frequency(target_pstate * cpu->pstate.scaling, cpu->cpu);
2292 	intel_pstate_update_pstate(cpu, target_pstate);
2293 
2294 	sample = &cpu->sample;
2295 	trace_pstate_sample(mul_ext_fp(100, sample->core_avg_perf),
2296 		fp_toint(sample->busy_scaled),
2297 		from,
2298 		cpu->pstate.current_pstate,
2299 		sample->mperf,
2300 		sample->aperf,
2301 		sample->tsc,
2302 		get_avg_frequency(cpu),
2303 		fp_toint(cpu->iowait_boost * 100));
2304 }
2305 
2306 static void intel_pstate_update_util(struct update_util_data *data, u64 time,
2307 				     unsigned int flags)
2308 {
2309 	struct cpudata *cpu = container_of(data, struct cpudata, update_util);
2310 	u64 delta_ns;
2311 
2312 	/* Don't allow remote callbacks */
2313 	if (smp_processor_id() != cpu->cpu)
2314 		return;
2315 
2316 	delta_ns = time - cpu->last_update;
2317 	if (flags & SCHED_CPUFREQ_IOWAIT) {
2318 		/* Start over if the CPU may have been idle. */
2319 		if (delta_ns > TICK_NSEC) {
2320 			cpu->iowait_boost = ONE_EIGHTH_FP;
2321 		} else if (cpu->iowait_boost >= ONE_EIGHTH_FP) {
2322 			cpu->iowait_boost <<= 1;
2323 			if (cpu->iowait_boost > int_tofp(1))
2324 				cpu->iowait_boost = int_tofp(1);
2325 		} else {
2326 			cpu->iowait_boost = ONE_EIGHTH_FP;
2327 		}
2328 	} else if (cpu->iowait_boost) {
2329 		/* Clear iowait_boost if the CPU may have been idle. */
2330 		if (delta_ns > TICK_NSEC)
2331 			cpu->iowait_boost = 0;
2332 		else
2333 			cpu->iowait_boost >>= 1;
2334 	}
2335 	cpu->last_update = time;
2336 	delta_ns = time - cpu->sample.time;
2337 	if ((s64)delta_ns < INTEL_PSTATE_SAMPLING_INTERVAL)
2338 		return;
2339 
2340 	if (intel_pstate_sample(cpu, time))
2341 		intel_pstate_adjust_pstate(cpu);
2342 }
2343 
2344 static struct pstate_funcs core_funcs = {
2345 	.get_max = core_get_max_pstate,
2346 	.get_max_physical = core_get_max_pstate_physical,
2347 	.get_min = core_get_min_pstate,
2348 	.get_turbo = core_get_turbo_pstate,
2349 	.get_scaling = core_get_scaling,
2350 	.get_val = core_get_val,
2351 };
2352 
2353 static const struct pstate_funcs silvermont_funcs = {
2354 	.get_max = atom_get_max_pstate,
2355 	.get_max_physical = atom_get_max_pstate,
2356 	.get_min = atom_get_min_pstate,
2357 	.get_turbo = atom_get_turbo_pstate,
2358 	.get_val = atom_get_val,
2359 	.get_scaling = silvermont_get_scaling,
2360 	.get_vid = atom_get_vid,
2361 };
2362 
2363 static const struct pstate_funcs airmont_funcs = {
2364 	.get_max = atom_get_max_pstate,
2365 	.get_max_physical = atom_get_max_pstate,
2366 	.get_min = atom_get_min_pstate,
2367 	.get_turbo = atom_get_turbo_pstate,
2368 	.get_val = atom_get_val,
2369 	.get_scaling = airmont_get_scaling,
2370 	.get_vid = atom_get_vid,
2371 };
2372 
2373 static const struct pstate_funcs knl_funcs = {
2374 	.get_max = core_get_max_pstate,
2375 	.get_max_physical = core_get_max_pstate_physical,
2376 	.get_min = core_get_min_pstate,
2377 	.get_turbo = knl_get_turbo_pstate,
2378 	.get_aperf_mperf_shift = knl_get_aperf_mperf_shift,
2379 	.get_scaling = core_get_scaling,
2380 	.get_val = core_get_val,
2381 };
2382 
2383 #define X86_MATCH(model, policy)					 \
2384 	X86_MATCH_VENDOR_FAM_MODEL_FEATURE(INTEL, 6, INTEL_FAM6_##model, \
2385 					   X86_FEATURE_APERFMPERF, &policy)
2386 
2387 static const struct x86_cpu_id intel_pstate_cpu_ids[] = {
2388 	X86_MATCH(SANDYBRIDGE,		core_funcs),
2389 	X86_MATCH(SANDYBRIDGE_X,	core_funcs),
2390 	X86_MATCH(ATOM_SILVERMONT,	silvermont_funcs),
2391 	X86_MATCH(IVYBRIDGE,		core_funcs),
2392 	X86_MATCH(HASWELL,		core_funcs),
2393 	X86_MATCH(BROADWELL,		core_funcs),
2394 	X86_MATCH(IVYBRIDGE_X,		core_funcs),
2395 	X86_MATCH(HASWELL_X,		core_funcs),
2396 	X86_MATCH(HASWELL_L,		core_funcs),
2397 	X86_MATCH(HASWELL_G,		core_funcs),
2398 	X86_MATCH(BROADWELL_G,		core_funcs),
2399 	X86_MATCH(ATOM_AIRMONT,		airmont_funcs),
2400 	X86_MATCH(SKYLAKE_L,		core_funcs),
2401 	X86_MATCH(BROADWELL_X,		core_funcs),
2402 	X86_MATCH(SKYLAKE,		core_funcs),
2403 	X86_MATCH(BROADWELL_D,		core_funcs),
2404 	X86_MATCH(XEON_PHI_KNL,		knl_funcs),
2405 	X86_MATCH(XEON_PHI_KNM,		knl_funcs),
2406 	X86_MATCH(ATOM_GOLDMONT,	core_funcs),
2407 	X86_MATCH(ATOM_GOLDMONT_PLUS,	core_funcs),
2408 	X86_MATCH(SKYLAKE_X,		core_funcs),
2409 	X86_MATCH(COMETLAKE,		core_funcs),
2410 	X86_MATCH(ICELAKE_X,		core_funcs),
2411 	X86_MATCH(TIGERLAKE,		core_funcs),
2412 	X86_MATCH(SAPPHIRERAPIDS_X,	core_funcs),
2413 	{}
2414 };
2415 MODULE_DEVICE_TABLE(x86cpu, intel_pstate_cpu_ids);
2416 
2417 static const struct x86_cpu_id intel_pstate_cpu_oob_ids[] __initconst = {
2418 	X86_MATCH(BROADWELL_D,		core_funcs),
2419 	X86_MATCH(BROADWELL_X,		core_funcs),
2420 	X86_MATCH(SKYLAKE_X,		core_funcs),
2421 	X86_MATCH(ICELAKE_X,		core_funcs),
2422 	X86_MATCH(SAPPHIRERAPIDS_X,	core_funcs),
2423 	{}
2424 };
2425 
2426 static const struct x86_cpu_id intel_pstate_cpu_ee_disable_ids[] = {
2427 	X86_MATCH(KABYLAKE,		core_funcs),
2428 	{}
2429 };
2430 
2431 static int intel_pstate_init_cpu(unsigned int cpunum)
2432 {
2433 	struct cpudata *cpu;
2434 
2435 	cpu = all_cpu_data[cpunum];
2436 
2437 	if (!cpu) {
2438 		cpu = kzalloc(sizeof(*cpu), GFP_KERNEL);
2439 		if (!cpu)
2440 			return -ENOMEM;
2441 
2442 		WRITE_ONCE(all_cpu_data[cpunum], cpu);
2443 
2444 		cpu->cpu = cpunum;
2445 
2446 		cpu->epp_default = -EINVAL;
2447 
2448 		if (hwp_active) {
2449 			intel_pstate_hwp_enable(cpu);
2450 
2451 			if (intel_pstate_acpi_pm_profile_server())
2452 				hwp_boost = true;
2453 		}
2454 	} else if (hwp_active) {
2455 		/*
2456 		 * Re-enable HWP in case this happens after a resume from ACPI
2457 		 * S3 if the CPU was offline during the whole system/resume
2458 		 * cycle.
2459 		 */
2460 		intel_pstate_hwp_reenable(cpu);
2461 	}
2462 
2463 	cpu->epp_powersave = -EINVAL;
2464 	cpu->epp_policy = 0;
2465 
2466 	intel_pstate_get_cpu_pstates(cpu);
2467 
2468 	pr_debug("controlling: cpu %d\n", cpunum);
2469 
2470 	return 0;
2471 }
2472 
2473 static void intel_pstate_set_update_util_hook(unsigned int cpu_num)
2474 {
2475 	struct cpudata *cpu = all_cpu_data[cpu_num];
2476 
2477 	if (hwp_active && !hwp_boost)
2478 		return;
2479 
2480 	if (cpu->update_util_set)
2481 		return;
2482 
2483 	/* Prevent intel_pstate_update_util() from using stale data. */
2484 	cpu->sample.time = 0;
2485 	cpufreq_add_update_util_hook(cpu_num, &cpu->update_util,
2486 				     (hwp_active ?
2487 				      intel_pstate_update_util_hwp :
2488 				      intel_pstate_update_util));
2489 	cpu->update_util_set = true;
2490 }
2491 
2492 static void intel_pstate_clear_update_util_hook(unsigned int cpu)
2493 {
2494 	struct cpudata *cpu_data = all_cpu_data[cpu];
2495 
2496 	if (!cpu_data->update_util_set)
2497 		return;
2498 
2499 	cpufreq_remove_update_util_hook(cpu);
2500 	cpu_data->update_util_set = false;
2501 	synchronize_rcu();
2502 }
2503 
2504 static int intel_pstate_get_max_freq(struct cpudata *cpu)
2505 {
2506 	return global.turbo_disabled || global.no_turbo ?
2507 			cpu->pstate.max_freq : cpu->pstate.turbo_freq;
2508 }
2509 
2510 static void intel_pstate_update_perf_limits(struct cpudata *cpu,
2511 					    unsigned int policy_min,
2512 					    unsigned int policy_max)
2513 {
2514 	int perf_ctl_scaling = cpu->pstate.perf_ctl_scaling;
2515 	int32_t max_policy_perf, min_policy_perf;
2516 
2517 	max_policy_perf = policy_max / perf_ctl_scaling;
2518 	if (policy_max == policy_min) {
2519 		min_policy_perf = max_policy_perf;
2520 	} else {
2521 		min_policy_perf = policy_min / perf_ctl_scaling;
2522 		min_policy_perf = clamp_t(int32_t, min_policy_perf,
2523 					  0, max_policy_perf);
2524 	}
2525 
2526 	/*
2527 	 * HWP needs some special consideration, because HWP_REQUEST uses
2528 	 * abstract values to represent performance rather than pure ratios.
2529 	 */
2530 	if (hwp_active && cpu->pstate.scaling != perf_ctl_scaling) {
2531 		int scaling = cpu->pstate.scaling;
2532 		int freq;
2533 
2534 		freq = max_policy_perf * perf_ctl_scaling;
2535 		max_policy_perf = DIV_ROUND_UP(freq, scaling);
2536 		freq = min_policy_perf * perf_ctl_scaling;
2537 		min_policy_perf = DIV_ROUND_UP(freq, scaling);
2538 	}
2539 
2540 	pr_debug("cpu:%d min_policy_perf:%d max_policy_perf:%d\n",
2541 		 cpu->cpu, min_policy_perf, max_policy_perf);
2542 
2543 	/* Normalize user input to [min_perf, max_perf] */
2544 	if (per_cpu_limits) {
2545 		cpu->min_perf_ratio = min_policy_perf;
2546 		cpu->max_perf_ratio = max_policy_perf;
2547 	} else {
2548 		int turbo_max = cpu->pstate.turbo_pstate;
2549 		int32_t global_min, global_max;
2550 
2551 		/* Global limits are in percent of the maximum turbo P-state. */
2552 		global_max = DIV_ROUND_UP(turbo_max * global.max_perf_pct, 100);
2553 		global_min = DIV_ROUND_UP(turbo_max * global.min_perf_pct, 100);
2554 		global_min = clamp_t(int32_t, global_min, 0, global_max);
2555 
2556 		pr_debug("cpu:%d global_min:%d global_max:%d\n", cpu->cpu,
2557 			 global_min, global_max);
2558 
2559 		cpu->min_perf_ratio = max(min_policy_perf, global_min);
2560 		cpu->min_perf_ratio = min(cpu->min_perf_ratio, max_policy_perf);
2561 		cpu->max_perf_ratio = min(max_policy_perf, global_max);
2562 		cpu->max_perf_ratio = max(min_policy_perf, cpu->max_perf_ratio);
2563 
2564 		/* Make sure min_perf <= max_perf */
2565 		cpu->min_perf_ratio = min(cpu->min_perf_ratio,
2566 					  cpu->max_perf_ratio);
2567 
2568 	}
2569 	pr_debug("cpu:%d max_perf_ratio:%d min_perf_ratio:%d\n", cpu->cpu,
2570 		 cpu->max_perf_ratio,
2571 		 cpu->min_perf_ratio);
2572 }
2573 
2574 static int intel_pstate_set_policy(struct cpufreq_policy *policy)
2575 {
2576 	struct cpudata *cpu;
2577 
2578 	if (!policy->cpuinfo.max_freq)
2579 		return -ENODEV;
2580 
2581 	pr_debug("set_policy cpuinfo.max %u policy->max %u\n",
2582 		 policy->cpuinfo.max_freq, policy->max);
2583 
2584 	cpu = all_cpu_data[policy->cpu];
2585 	cpu->policy = policy->policy;
2586 
2587 	mutex_lock(&intel_pstate_limits_lock);
2588 
2589 	intel_pstate_update_perf_limits(cpu, policy->min, policy->max);
2590 
2591 	if (cpu->policy == CPUFREQ_POLICY_PERFORMANCE) {
2592 		/*
2593 		 * NOHZ_FULL CPUs need this as the governor callback may not
2594 		 * be invoked on them.
2595 		 */
2596 		intel_pstate_clear_update_util_hook(policy->cpu);
2597 		intel_pstate_max_within_limits(cpu);
2598 	} else {
2599 		intel_pstate_set_update_util_hook(policy->cpu);
2600 	}
2601 
2602 	if (hwp_active) {
2603 		/*
2604 		 * When hwp_boost was active before and dynamically it
2605 		 * was turned off, in that case we need to clear the
2606 		 * update util hook.
2607 		 */
2608 		if (!hwp_boost)
2609 			intel_pstate_clear_update_util_hook(policy->cpu);
2610 		intel_pstate_hwp_set(policy->cpu);
2611 	}
2612 	/*
2613 	 * policy->cur is never updated with the intel_pstate driver, but it
2614 	 * is used as a stale frequency value. So, keep it within limits.
2615 	 */
2616 	policy->cur = policy->min;
2617 
2618 	mutex_unlock(&intel_pstate_limits_lock);
2619 
2620 	return 0;
2621 }
2622 
2623 static void intel_pstate_adjust_policy_max(struct cpudata *cpu,
2624 					   struct cpufreq_policy_data *policy)
2625 {
2626 	if (!hwp_active &&
2627 	    cpu->pstate.max_pstate_physical > cpu->pstate.max_pstate &&
2628 	    policy->max < policy->cpuinfo.max_freq &&
2629 	    policy->max > cpu->pstate.max_freq) {
2630 		pr_debug("policy->max > max non turbo frequency\n");
2631 		policy->max = policy->cpuinfo.max_freq;
2632 	}
2633 }
2634 
2635 static void intel_pstate_verify_cpu_policy(struct cpudata *cpu,
2636 					   struct cpufreq_policy_data *policy)
2637 {
2638 	int max_freq;
2639 
2640 	update_turbo_state();
2641 	if (hwp_active) {
2642 		intel_pstate_get_hwp_cap(cpu);
2643 		max_freq = global.no_turbo || global.turbo_disabled ?
2644 				cpu->pstate.max_freq : cpu->pstate.turbo_freq;
2645 	} else {
2646 		max_freq = intel_pstate_get_max_freq(cpu);
2647 	}
2648 	cpufreq_verify_within_limits(policy, policy->cpuinfo.min_freq, max_freq);
2649 
2650 	intel_pstate_adjust_policy_max(cpu, policy);
2651 }
2652 
2653 static int intel_pstate_verify_policy(struct cpufreq_policy_data *policy)
2654 {
2655 	intel_pstate_verify_cpu_policy(all_cpu_data[policy->cpu], policy);
2656 
2657 	return 0;
2658 }
2659 
2660 static int intel_cpufreq_cpu_offline(struct cpufreq_policy *policy)
2661 {
2662 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2663 
2664 	pr_debug("CPU %d going offline\n", cpu->cpu);
2665 
2666 	if (cpu->suspended)
2667 		return 0;
2668 
2669 	/*
2670 	 * If the CPU is an SMT thread and it goes offline with the performance
2671 	 * settings different from the minimum, it will prevent its sibling
2672 	 * from getting to lower performance levels, so force the minimum
2673 	 * performance on CPU offline to prevent that from happening.
2674 	 */
2675 	if (hwp_active)
2676 		intel_pstate_hwp_offline(cpu);
2677 	else
2678 		intel_pstate_set_min_pstate(cpu);
2679 
2680 	intel_pstate_exit_perf_limits(policy);
2681 
2682 	return 0;
2683 }
2684 
2685 static int intel_pstate_cpu_online(struct cpufreq_policy *policy)
2686 {
2687 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2688 
2689 	pr_debug("CPU %d going online\n", cpu->cpu);
2690 
2691 	intel_pstate_init_acpi_perf_limits(policy);
2692 
2693 	if (hwp_active) {
2694 		/*
2695 		 * Re-enable HWP and clear the "suspended" flag to let "resume"
2696 		 * know that it need not do that.
2697 		 */
2698 		intel_pstate_hwp_reenable(cpu);
2699 		cpu->suspended = false;
2700 	}
2701 
2702 	return 0;
2703 }
2704 
2705 static int intel_pstate_cpu_offline(struct cpufreq_policy *policy)
2706 {
2707 	intel_pstate_clear_update_util_hook(policy->cpu);
2708 
2709 	return intel_cpufreq_cpu_offline(policy);
2710 }
2711 
2712 static int intel_pstate_cpu_exit(struct cpufreq_policy *policy)
2713 {
2714 	pr_debug("CPU %d exiting\n", policy->cpu);
2715 
2716 	policy->fast_switch_possible = false;
2717 
2718 	return 0;
2719 }
2720 
2721 static int __intel_pstate_cpu_init(struct cpufreq_policy *policy)
2722 {
2723 	struct cpudata *cpu;
2724 	int rc;
2725 
2726 	rc = intel_pstate_init_cpu(policy->cpu);
2727 	if (rc)
2728 		return rc;
2729 
2730 	cpu = all_cpu_data[policy->cpu];
2731 
2732 	cpu->max_perf_ratio = 0xFF;
2733 	cpu->min_perf_ratio = 0;
2734 
2735 	/* cpuinfo and default policy values */
2736 	policy->cpuinfo.min_freq = cpu->pstate.min_freq;
2737 	update_turbo_state();
2738 	global.turbo_disabled_mf = global.turbo_disabled;
2739 	policy->cpuinfo.max_freq = global.turbo_disabled ?
2740 			cpu->pstate.max_freq : cpu->pstate.turbo_freq;
2741 
2742 	policy->min = policy->cpuinfo.min_freq;
2743 	policy->max = policy->cpuinfo.max_freq;
2744 
2745 	intel_pstate_init_acpi_perf_limits(policy);
2746 
2747 	policy->fast_switch_possible = true;
2748 
2749 	return 0;
2750 }
2751 
2752 static int intel_pstate_cpu_init(struct cpufreq_policy *policy)
2753 {
2754 	int ret = __intel_pstate_cpu_init(policy);
2755 
2756 	if (ret)
2757 		return ret;
2758 
2759 	/*
2760 	 * Set the policy to powersave to provide a valid fallback value in case
2761 	 * the default cpufreq governor is neither powersave nor performance.
2762 	 */
2763 	policy->policy = CPUFREQ_POLICY_POWERSAVE;
2764 
2765 	if (hwp_active) {
2766 		struct cpudata *cpu = all_cpu_data[policy->cpu];
2767 
2768 		cpu->epp_cached = intel_pstate_get_epp(cpu, 0);
2769 	}
2770 
2771 	return 0;
2772 }
2773 
2774 static struct cpufreq_driver intel_pstate = {
2775 	.flags		= CPUFREQ_CONST_LOOPS,
2776 	.verify		= intel_pstate_verify_policy,
2777 	.setpolicy	= intel_pstate_set_policy,
2778 	.suspend	= intel_pstate_suspend,
2779 	.resume		= intel_pstate_resume,
2780 	.init		= intel_pstate_cpu_init,
2781 	.exit		= intel_pstate_cpu_exit,
2782 	.offline	= intel_pstate_cpu_offline,
2783 	.online		= intel_pstate_cpu_online,
2784 	.update_limits	= intel_pstate_update_limits,
2785 	.name		= "intel_pstate",
2786 };
2787 
2788 static int intel_cpufreq_verify_policy(struct cpufreq_policy_data *policy)
2789 {
2790 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2791 
2792 	intel_pstate_verify_cpu_policy(cpu, policy);
2793 	intel_pstate_update_perf_limits(cpu, policy->min, policy->max);
2794 
2795 	return 0;
2796 }
2797 
2798 /* Use of trace in passive mode:
2799  *
2800  * In passive mode the trace core_busy field (also known as the
2801  * performance field, and lablelled as such on the graphs; also known as
2802  * core_avg_perf) is not needed and so is re-assigned to indicate if the
2803  * driver call was via the normal or fast switch path. Various graphs
2804  * output from the intel_pstate_tracer.py utility that include core_busy
2805  * (or performance or core_avg_perf) have a fixed y-axis from 0 to 100%,
2806  * so we use 10 to indicate the normal path through the driver, and
2807  * 90 to indicate the fast switch path through the driver.
2808  * The scaled_busy field is not used, and is set to 0.
2809  */
2810 
2811 #define	INTEL_PSTATE_TRACE_TARGET 10
2812 #define	INTEL_PSTATE_TRACE_FAST_SWITCH 90
2813 
2814 static void intel_cpufreq_trace(struct cpudata *cpu, unsigned int trace_type, int old_pstate)
2815 {
2816 	struct sample *sample;
2817 
2818 	if (!trace_pstate_sample_enabled())
2819 		return;
2820 
2821 	if (!intel_pstate_sample(cpu, ktime_get()))
2822 		return;
2823 
2824 	sample = &cpu->sample;
2825 	trace_pstate_sample(trace_type,
2826 		0,
2827 		old_pstate,
2828 		cpu->pstate.current_pstate,
2829 		sample->mperf,
2830 		sample->aperf,
2831 		sample->tsc,
2832 		get_avg_frequency(cpu),
2833 		fp_toint(cpu->iowait_boost * 100));
2834 }
2835 
2836 static void intel_cpufreq_hwp_update(struct cpudata *cpu, u32 min, u32 max,
2837 				     u32 desired, bool fast_switch)
2838 {
2839 	u64 prev = READ_ONCE(cpu->hwp_req_cached), value = prev;
2840 
2841 	value &= ~HWP_MIN_PERF(~0L);
2842 	value |= HWP_MIN_PERF(min);
2843 
2844 	value &= ~HWP_MAX_PERF(~0L);
2845 	value |= HWP_MAX_PERF(max);
2846 
2847 	value &= ~HWP_DESIRED_PERF(~0L);
2848 	value |= HWP_DESIRED_PERF(desired);
2849 
2850 	if (value == prev)
2851 		return;
2852 
2853 	WRITE_ONCE(cpu->hwp_req_cached, value);
2854 	if (fast_switch)
2855 		wrmsrl(MSR_HWP_REQUEST, value);
2856 	else
2857 		wrmsrl_on_cpu(cpu->cpu, MSR_HWP_REQUEST, value);
2858 }
2859 
2860 static void intel_cpufreq_perf_ctl_update(struct cpudata *cpu,
2861 					  u32 target_pstate, bool fast_switch)
2862 {
2863 	if (fast_switch)
2864 		wrmsrl(MSR_IA32_PERF_CTL,
2865 		       pstate_funcs.get_val(cpu, target_pstate));
2866 	else
2867 		wrmsrl_on_cpu(cpu->cpu, MSR_IA32_PERF_CTL,
2868 			      pstate_funcs.get_val(cpu, target_pstate));
2869 }
2870 
2871 static int intel_cpufreq_update_pstate(struct cpufreq_policy *policy,
2872 				       int target_pstate, bool fast_switch)
2873 {
2874 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2875 	int old_pstate = cpu->pstate.current_pstate;
2876 
2877 	target_pstate = intel_pstate_prepare_request(cpu, target_pstate);
2878 	if (hwp_active) {
2879 		int max_pstate = policy->strict_target ?
2880 					target_pstate : cpu->max_perf_ratio;
2881 
2882 		intel_cpufreq_hwp_update(cpu, target_pstate, max_pstate, 0,
2883 					 fast_switch);
2884 	} else if (target_pstate != old_pstate) {
2885 		intel_cpufreq_perf_ctl_update(cpu, target_pstate, fast_switch);
2886 	}
2887 
2888 	cpu->pstate.current_pstate = target_pstate;
2889 
2890 	intel_cpufreq_trace(cpu, fast_switch ? INTEL_PSTATE_TRACE_FAST_SWITCH :
2891 			    INTEL_PSTATE_TRACE_TARGET, old_pstate);
2892 
2893 	return target_pstate;
2894 }
2895 
2896 static int intel_cpufreq_target(struct cpufreq_policy *policy,
2897 				unsigned int target_freq,
2898 				unsigned int relation)
2899 {
2900 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2901 	struct cpufreq_freqs freqs;
2902 	int target_pstate;
2903 
2904 	update_turbo_state();
2905 
2906 	freqs.old = policy->cur;
2907 	freqs.new = target_freq;
2908 
2909 	cpufreq_freq_transition_begin(policy, &freqs);
2910 
2911 	switch (relation) {
2912 	case CPUFREQ_RELATION_L:
2913 		target_pstate = DIV_ROUND_UP(freqs.new, cpu->pstate.scaling);
2914 		break;
2915 	case CPUFREQ_RELATION_H:
2916 		target_pstate = freqs.new / cpu->pstate.scaling;
2917 		break;
2918 	default:
2919 		target_pstate = DIV_ROUND_CLOSEST(freqs.new, cpu->pstate.scaling);
2920 		break;
2921 	}
2922 
2923 	target_pstate = intel_cpufreq_update_pstate(policy, target_pstate, false);
2924 
2925 	freqs.new = target_pstate * cpu->pstate.scaling;
2926 
2927 	cpufreq_freq_transition_end(policy, &freqs, false);
2928 
2929 	return 0;
2930 }
2931 
2932 static unsigned int intel_cpufreq_fast_switch(struct cpufreq_policy *policy,
2933 					      unsigned int target_freq)
2934 {
2935 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2936 	int target_pstate;
2937 
2938 	update_turbo_state();
2939 
2940 	target_pstate = DIV_ROUND_UP(target_freq, cpu->pstate.scaling);
2941 
2942 	target_pstate = intel_cpufreq_update_pstate(policy, target_pstate, true);
2943 
2944 	return target_pstate * cpu->pstate.scaling;
2945 }
2946 
2947 static void intel_cpufreq_adjust_perf(unsigned int cpunum,
2948 				      unsigned long min_perf,
2949 				      unsigned long target_perf,
2950 				      unsigned long capacity)
2951 {
2952 	struct cpudata *cpu = all_cpu_data[cpunum];
2953 	u64 hwp_cap = READ_ONCE(cpu->hwp_cap_cached);
2954 	int old_pstate = cpu->pstate.current_pstate;
2955 	int cap_pstate, min_pstate, max_pstate, target_pstate;
2956 
2957 	update_turbo_state();
2958 	cap_pstate = global.turbo_disabled ? HWP_GUARANTEED_PERF(hwp_cap) :
2959 					     HWP_HIGHEST_PERF(hwp_cap);
2960 
2961 	/* Optimization: Avoid unnecessary divisions. */
2962 
2963 	target_pstate = cap_pstate;
2964 	if (target_perf < capacity)
2965 		target_pstate = DIV_ROUND_UP(cap_pstate * target_perf, capacity);
2966 
2967 	min_pstate = cap_pstate;
2968 	if (min_perf < capacity)
2969 		min_pstate = DIV_ROUND_UP(cap_pstate * min_perf, capacity);
2970 
2971 	if (min_pstate < cpu->pstate.min_pstate)
2972 		min_pstate = cpu->pstate.min_pstate;
2973 
2974 	if (min_pstate < cpu->min_perf_ratio)
2975 		min_pstate = cpu->min_perf_ratio;
2976 
2977 	max_pstate = min(cap_pstate, cpu->max_perf_ratio);
2978 	if (max_pstate < min_pstate)
2979 		max_pstate = min_pstate;
2980 
2981 	target_pstate = clamp_t(int, target_pstate, min_pstate, max_pstate);
2982 
2983 	intel_cpufreq_hwp_update(cpu, min_pstate, max_pstate, target_pstate, true);
2984 
2985 	cpu->pstate.current_pstate = target_pstate;
2986 	intel_cpufreq_trace(cpu, INTEL_PSTATE_TRACE_FAST_SWITCH, old_pstate);
2987 }
2988 
2989 static int intel_cpufreq_cpu_init(struct cpufreq_policy *policy)
2990 {
2991 	struct freq_qos_request *req;
2992 	struct cpudata *cpu;
2993 	struct device *dev;
2994 	int ret, freq;
2995 
2996 	dev = get_cpu_device(policy->cpu);
2997 	if (!dev)
2998 		return -ENODEV;
2999 
3000 	ret = __intel_pstate_cpu_init(policy);
3001 	if (ret)
3002 		return ret;
3003 
3004 	policy->cpuinfo.transition_latency = INTEL_CPUFREQ_TRANSITION_LATENCY;
3005 	/* This reflects the intel_pstate_get_cpu_pstates() setting. */
3006 	policy->cur = policy->cpuinfo.min_freq;
3007 
3008 	req = kcalloc(2, sizeof(*req), GFP_KERNEL);
3009 	if (!req) {
3010 		ret = -ENOMEM;
3011 		goto pstate_exit;
3012 	}
3013 
3014 	cpu = all_cpu_data[policy->cpu];
3015 
3016 	if (hwp_active) {
3017 		u64 value;
3018 
3019 		policy->transition_delay_us = INTEL_CPUFREQ_TRANSITION_DELAY_HWP;
3020 
3021 		intel_pstate_get_hwp_cap(cpu);
3022 
3023 		rdmsrl_on_cpu(cpu->cpu, MSR_HWP_REQUEST, &value);
3024 		WRITE_ONCE(cpu->hwp_req_cached, value);
3025 
3026 		cpu->epp_cached = intel_pstate_get_epp(cpu, value);
3027 	} else {
3028 		policy->transition_delay_us = INTEL_CPUFREQ_TRANSITION_DELAY;
3029 	}
3030 
3031 	freq = DIV_ROUND_UP(cpu->pstate.turbo_freq * global.min_perf_pct, 100);
3032 
3033 	ret = freq_qos_add_request(&policy->constraints, req, FREQ_QOS_MIN,
3034 				   freq);
3035 	if (ret < 0) {
3036 		dev_err(dev, "Failed to add min-freq constraint (%d)\n", ret);
3037 		goto free_req;
3038 	}
3039 
3040 	freq = DIV_ROUND_UP(cpu->pstate.turbo_freq * global.max_perf_pct, 100);
3041 
3042 	ret = freq_qos_add_request(&policy->constraints, req + 1, FREQ_QOS_MAX,
3043 				   freq);
3044 	if (ret < 0) {
3045 		dev_err(dev, "Failed to add max-freq constraint (%d)\n", ret);
3046 		goto remove_min_req;
3047 	}
3048 
3049 	policy->driver_data = req;
3050 
3051 	return 0;
3052 
3053 remove_min_req:
3054 	freq_qos_remove_request(req);
3055 free_req:
3056 	kfree(req);
3057 pstate_exit:
3058 	intel_pstate_exit_perf_limits(policy);
3059 
3060 	return ret;
3061 }
3062 
3063 static int intel_cpufreq_cpu_exit(struct cpufreq_policy *policy)
3064 {
3065 	struct freq_qos_request *req;
3066 
3067 	req = policy->driver_data;
3068 
3069 	freq_qos_remove_request(req + 1);
3070 	freq_qos_remove_request(req);
3071 	kfree(req);
3072 
3073 	return intel_pstate_cpu_exit(policy);
3074 }
3075 
3076 static int intel_cpufreq_suspend(struct cpufreq_policy *policy)
3077 {
3078 	intel_pstate_suspend(policy);
3079 
3080 	if (hwp_active) {
3081 		struct cpudata *cpu = all_cpu_data[policy->cpu];
3082 		u64 value = READ_ONCE(cpu->hwp_req_cached);
3083 
3084 		/*
3085 		 * Clear the desired perf field in MSR_HWP_REQUEST in case
3086 		 * intel_cpufreq_adjust_perf() is in use and the last value
3087 		 * written by it may not be suitable.
3088 		 */
3089 		value &= ~HWP_DESIRED_PERF(~0L);
3090 		wrmsrl_on_cpu(cpu->cpu, MSR_HWP_REQUEST, value);
3091 		WRITE_ONCE(cpu->hwp_req_cached, value);
3092 	}
3093 
3094 	return 0;
3095 }
3096 
3097 static struct cpufreq_driver intel_cpufreq = {
3098 	.flags		= CPUFREQ_CONST_LOOPS,
3099 	.verify		= intel_cpufreq_verify_policy,
3100 	.target		= intel_cpufreq_target,
3101 	.fast_switch	= intel_cpufreq_fast_switch,
3102 	.init		= intel_cpufreq_cpu_init,
3103 	.exit		= intel_cpufreq_cpu_exit,
3104 	.offline	= intel_cpufreq_cpu_offline,
3105 	.online		= intel_pstate_cpu_online,
3106 	.suspend	= intel_cpufreq_suspend,
3107 	.resume		= intel_pstate_resume,
3108 	.update_limits	= intel_pstate_update_limits,
3109 	.name		= "intel_cpufreq",
3110 };
3111 
3112 static struct cpufreq_driver *default_driver;
3113 
3114 static void intel_pstate_driver_cleanup(void)
3115 {
3116 	unsigned int cpu;
3117 
3118 	cpus_read_lock();
3119 	for_each_online_cpu(cpu) {
3120 		if (all_cpu_data[cpu]) {
3121 			if (intel_pstate_driver == &intel_pstate)
3122 				intel_pstate_clear_update_util_hook(cpu);
3123 
3124 			spin_lock(&hwp_notify_lock);
3125 			kfree(all_cpu_data[cpu]);
3126 			WRITE_ONCE(all_cpu_data[cpu], NULL);
3127 			spin_unlock(&hwp_notify_lock);
3128 		}
3129 	}
3130 	cpus_read_unlock();
3131 
3132 	intel_pstate_driver = NULL;
3133 }
3134 
3135 static int intel_pstate_register_driver(struct cpufreq_driver *driver)
3136 {
3137 	int ret;
3138 
3139 	if (driver == &intel_pstate)
3140 		intel_pstate_sysfs_expose_hwp_dynamic_boost();
3141 
3142 	memset(&global, 0, sizeof(global));
3143 	global.max_perf_pct = 100;
3144 
3145 	intel_pstate_driver = driver;
3146 	ret = cpufreq_register_driver(intel_pstate_driver);
3147 	if (ret) {
3148 		intel_pstate_driver_cleanup();
3149 		return ret;
3150 	}
3151 
3152 	global.min_perf_pct = min_perf_pct_min();
3153 
3154 	return 0;
3155 }
3156 
3157 static ssize_t intel_pstate_show_status(char *buf)
3158 {
3159 	if (!intel_pstate_driver)
3160 		return sprintf(buf, "off\n");
3161 
3162 	return sprintf(buf, "%s\n", intel_pstate_driver == &intel_pstate ?
3163 					"active" : "passive");
3164 }
3165 
3166 static int intel_pstate_update_status(const char *buf, size_t size)
3167 {
3168 	if (size == 3 && !strncmp(buf, "off", size)) {
3169 		if (!intel_pstate_driver)
3170 			return -EINVAL;
3171 
3172 		if (hwp_active)
3173 			return -EBUSY;
3174 
3175 		cpufreq_unregister_driver(intel_pstate_driver);
3176 		intel_pstate_driver_cleanup();
3177 		return 0;
3178 	}
3179 
3180 	if (size == 6 && !strncmp(buf, "active", size)) {
3181 		if (intel_pstate_driver) {
3182 			if (intel_pstate_driver == &intel_pstate)
3183 				return 0;
3184 
3185 			cpufreq_unregister_driver(intel_pstate_driver);
3186 		}
3187 
3188 		return intel_pstate_register_driver(&intel_pstate);
3189 	}
3190 
3191 	if (size == 7 && !strncmp(buf, "passive", size)) {
3192 		if (intel_pstate_driver) {
3193 			if (intel_pstate_driver == &intel_cpufreq)
3194 				return 0;
3195 
3196 			cpufreq_unregister_driver(intel_pstate_driver);
3197 			intel_pstate_sysfs_hide_hwp_dynamic_boost();
3198 		}
3199 
3200 		return intel_pstate_register_driver(&intel_cpufreq);
3201 	}
3202 
3203 	return -EINVAL;
3204 }
3205 
3206 static int no_load __initdata;
3207 static int no_hwp __initdata;
3208 static int hwp_only __initdata;
3209 static unsigned int force_load __initdata;
3210 
3211 static int __init intel_pstate_msrs_not_valid(void)
3212 {
3213 	if (!pstate_funcs.get_max(0) ||
3214 	    !pstate_funcs.get_min(0) ||
3215 	    !pstate_funcs.get_turbo(0))
3216 		return -ENODEV;
3217 
3218 	return 0;
3219 }
3220 
3221 static void __init copy_cpu_funcs(struct pstate_funcs *funcs)
3222 {
3223 	pstate_funcs.get_max   = funcs->get_max;
3224 	pstate_funcs.get_max_physical = funcs->get_max_physical;
3225 	pstate_funcs.get_min   = funcs->get_min;
3226 	pstate_funcs.get_turbo = funcs->get_turbo;
3227 	pstate_funcs.get_scaling = funcs->get_scaling;
3228 	pstate_funcs.get_val   = funcs->get_val;
3229 	pstate_funcs.get_vid   = funcs->get_vid;
3230 	pstate_funcs.get_aperf_mperf_shift = funcs->get_aperf_mperf_shift;
3231 }
3232 
3233 #ifdef CONFIG_ACPI
3234 
3235 static bool __init intel_pstate_no_acpi_pss(void)
3236 {
3237 	int i;
3238 
3239 	for_each_possible_cpu(i) {
3240 		acpi_status status;
3241 		union acpi_object *pss;
3242 		struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
3243 		struct acpi_processor *pr = per_cpu(processors, i);
3244 
3245 		if (!pr)
3246 			continue;
3247 
3248 		status = acpi_evaluate_object(pr->handle, "_PSS", NULL, &buffer);
3249 		if (ACPI_FAILURE(status))
3250 			continue;
3251 
3252 		pss = buffer.pointer;
3253 		if (pss && pss->type == ACPI_TYPE_PACKAGE) {
3254 			kfree(pss);
3255 			return false;
3256 		}
3257 
3258 		kfree(pss);
3259 	}
3260 
3261 	pr_debug("ACPI _PSS not found\n");
3262 	return true;
3263 }
3264 
3265 static bool __init intel_pstate_no_acpi_pcch(void)
3266 {
3267 	acpi_status status;
3268 	acpi_handle handle;
3269 
3270 	status = acpi_get_handle(NULL, "\\_SB", &handle);
3271 	if (ACPI_FAILURE(status))
3272 		goto not_found;
3273 
3274 	if (acpi_has_method(handle, "PCCH"))
3275 		return false;
3276 
3277 not_found:
3278 	pr_debug("ACPI PCCH not found\n");
3279 	return true;
3280 }
3281 
3282 static bool __init intel_pstate_has_acpi_ppc(void)
3283 {
3284 	int i;
3285 
3286 	for_each_possible_cpu(i) {
3287 		struct acpi_processor *pr = per_cpu(processors, i);
3288 
3289 		if (!pr)
3290 			continue;
3291 		if (acpi_has_method(pr->handle, "_PPC"))
3292 			return true;
3293 	}
3294 	pr_debug("ACPI _PPC not found\n");
3295 	return false;
3296 }
3297 
3298 enum {
3299 	PSS,
3300 	PPC,
3301 };
3302 
3303 /* Hardware vendor-specific info that has its own power management modes */
3304 static struct acpi_platform_list plat_info[] __initdata = {
3305 	{"HP    ", "ProLiant", 0, ACPI_SIG_FADT, all_versions, NULL, PSS},
3306 	{"ORACLE", "X4-2    ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3307 	{"ORACLE", "X4-2L   ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3308 	{"ORACLE", "X4-2B   ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3309 	{"ORACLE", "X3-2    ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3310 	{"ORACLE", "X3-2L   ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3311 	{"ORACLE", "X3-2B   ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3312 	{"ORACLE", "X4470M2 ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3313 	{"ORACLE", "X4270M3 ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3314 	{"ORACLE", "X4270M2 ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3315 	{"ORACLE", "X4170M2 ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3316 	{"ORACLE", "X4170 M3", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3317 	{"ORACLE", "X4275 M3", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3318 	{"ORACLE", "X6-2    ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3319 	{"ORACLE", "Sudbury ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3320 	{ } /* End */
3321 };
3322 
3323 #define BITMASK_OOB	(BIT(8) | BIT(18))
3324 
3325 static bool __init intel_pstate_platform_pwr_mgmt_exists(void)
3326 {
3327 	const struct x86_cpu_id *id;
3328 	u64 misc_pwr;
3329 	int idx;
3330 
3331 	id = x86_match_cpu(intel_pstate_cpu_oob_ids);
3332 	if (id) {
3333 		rdmsrl(MSR_MISC_PWR_MGMT, misc_pwr);
3334 		if (misc_pwr & BITMASK_OOB) {
3335 			pr_debug("Bit 8 or 18 in the MISC_PWR_MGMT MSR set\n");
3336 			pr_debug("P states are controlled in Out of Band mode by the firmware/hardware\n");
3337 			return true;
3338 		}
3339 	}
3340 
3341 	idx = acpi_match_platform_list(plat_info);
3342 	if (idx < 0)
3343 		return false;
3344 
3345 	switch (plat_info[idx].data) {
3346 	case PSS:
3347 		if (!intel_pstate_no_acpi_pss())
3348 			return false;
3349 
3350 		return intel_pstate_no_acpi_pcch();
3351 	case PPC:
3352 		return intel_pstate_has_acpi_ppc() && !force_load;
3353 	}
3354 
3355 	return false;
3356 }
3357 
3358 static void intel_pstate_request_control_from_smm(void)
3359 {
3360 	/*
3361 	 * It may be unsafe to request P-states control from SMM if _PPC support
3362 	 * has not been enabled.
3363 	 */
3364 	if (acpi_ppc)
3365 		acpi_processor_pstate_control();
3366 }
3367 #else /* CONFIG_ACPI not enabled */
3368 static inline bool intel_pstate_platform_pwr_mgmt_exists(void) { return false; }
3369 static inline bool intel_pstate_has_acpi_ppc(void) { return false; }
3370 static inline void intel_pstate_request_control_from_smm(void) {}
3371 #endif /* CONFIG_ACPI */
3372 
3373 #define INTEL_PSTATE_HWP_BROADWELL	0x01
3374 
3375 #define X86_MATCH_HWP(model, hwp_mode)					\
3376 	X86_MATCH_VENDOR_FAM_MODEL_FEATURE(INTEL, 6, INTEL_FAM6_##model, \
3377 					   X86_FEATURE_HWP, hwp_mode)
3378 
3379 static const struct x86_cpu_id hwp_support_ids[] __initconst = {
3380 	X86_MATCH_HWP(BROADWELL_X,	INTEL_PSTATE_HWP_BROADWELL),
3381 	X86_MATCH_HWP(BROADWELL_D,	INTEL_PSTATE_HWP_BROADWELL),
3382 	X86_MATCH_HWP(ANY,		0),
3383 	{}
3384 };
3385 
3386 static bool intel_pstate_hwp_is_enabled(void)
3387 {
3388 	u64 value;
3389 
3390 	rdmsrl(MSR_PM_ENABLE, value);
3391 	return !!(value & 0x1);
3392 }
3393 
3394 static const struct x86_cpu_id intel_epp_balance_perf[] = {
3395 	/*
3396 	 * Set EPP value as 102, this is the max suggested EPP
3397 	 * which can result in one core turbo frequency for
3398 	 * AlderLake Mobile CPUs.
3399 	 */
3400 	X86_MATCH_INTEL_FAM6_MODEL(ALDERLAKE_L, 102),
3401 	X86_MATCH_INTEL_FAM6_MODEL(SAPPHIRERAPIDS_X, 32),
3402 	{}
3403 };
3404 
3405 static int __init intel_pstate_init(void)
3406 {
3407 	static struct cpudata **_all_cpu_data;
3408 	const struct x86_cpu_id *id;
3409 	int rc;
3410 
3411 	if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL)
3412 		return -ENODEV;
3413 
3414 	id = x86_match_cpu(hwp_support_ids);
3415 	if (id) {
3416 		hwp_forced = intel_pstate_hwp_is_enabled();
3417 
3418 		if (hwp_forced)
3419 			pr_info("HWP enabled by BIOS\n");
3420 		else if (no_load)
3421 			return -ENODEV;
3422 
3423 		copy_cpu_funcs(&core_funcs);
3424 		/*
3425 		 * Avoid enabling HWP for processors without EPP support,
3426 		 * because that means incomplete HWP implementation which is a
3427 		 * corner case and supporting it is generally problematic.
3428 		 *
3429 		 * If HWP is enabled already, though, there is no choice but to
3430 		 * deal with it.
3431 		 */
3432 		if ((!no_hwp && boot_cpu_has(X86_FEATURE_HWP_EPP)) || hwp_forced) {
3433 			WRITE_ONCE(hwp_active, 1);
3434 			hwp_mode_bdw = id->driver_data;
3435 			intel_pstate.attr = hwp_cpufreq_attrs;
3436 			intel_cpufreq.attr = hwp_cpufreq_attrs;
3437 			intel_cpufreq.flags |= CPUFREQ_NEED_UPDATE_LIMITS;
3438 			intel_cpufreq.adjust_perf = intel_cpufreq_adjust_perf;
3439 			if (!default_driver)
3440 				default_driver = &intel_pstate;
3441 
3442 			pstate_funcs.get_cpu_scaling = hwp_get_cpu_scaling;
3443 
3444 			goto hwp_cpu_matched;
3445 		}
3446 		pr_info("HWP not enabled\n");
3447 	} else {
3448 		if (no_load)
3449 			return -ENODEV;
3450 
3451 		id = x86_match_cpu(intel_pstate_cpu_ids);
3452 		if (!id) {
3453 			pr_info("CPU model not supported\n");
3454 			return -ENODEV;
3455 		}
3456 
3457 		copy_cpu_funcs((struct pstate_funcs *)id->driver_data);
3458 	}
3459 
3460 	if (intel_pstate_msrs_not_valid()) {
3461 		pr_info("Invalid MSRs\n");
3462 		return -ENODEV;
3463 	}
3464 	/* Without HWP start in the passive mode. */
3465 	if (!default_driver)
3466 		default_driver = &intel_cpufreq;
3467 
3468 hwp_cpu_matched:
3469 	/*
3470 	 * The Intel pstate driver will be ignored if the platform
3471 	 * firmware has its own power management modes.
3472 	 */
3473 	if (intel_pstate_platform_pwr_mgmt_exists()) {
3474 		pr_info("P-states controlled by the platform\n");
3475 		return -ENODEV;
3476 	}
3477 
3478 	if (!hwp_active && hwp_only)
3479 		return -ENOTSUPP;
3480 
3481 	pr_info("Intel P-state driver initializing\n");
3482 
3483 	_all_cpu_data = vzalloc(array_size(sizeof(void *), num_possible_cpus()));
3484 	if (!_all_cpu_data)
3485 		return -ENOMEM;
3486 
3487 	WRITE_ONCE(all_cpu_data, _all_cpu_data);
3488 
3489 	intel_pstate_request_control_from_smm();
3490 
3491 	intel_pstate_sysfs_expose_params();
3492 
3493 	if (hwp_active) {
3494 		const struct x86_cpu_id *id = x86_match_cpu(intel_epp_balance_perf);
3495 
3496 		if (id)
3497 			epp_values[EPP_INDEX_BALANCE_PERFORMANCE] = id->driver_data;
3498 	}
3499 
3500 	mutex_lock(&intel_pstate_driver_lock);
3501 	rc = intel_pstate_register_driver(default_driver);
3502 	mutex_unlock(&intel_pstate_driver_lock);
3503 	if (rc) {
3504 		intel_pstate_sysfs_remove();
3505 		return rc;
3506 	}
3507 
3508 	if (hwp_active) {
3509 		const struct x86_cpu_id *id;
3510 
3511 		id = x86_match_cpu(intel_pstate_cpu_ee_disable_ids);
3512 		if (id) {
3513 			set_power_ctl_ee_state(false);
3514 			pr_info("Disabling energy efficiency optimization\n");
3515 		}
3516 
3517 		pr_info("HWP enabled\n");
3518 	} else if (boot_cpu_has(X86_FEATURE_HYBRID_CPU)) {
3519 		pr_warn("Problematic setup: Hybrid processor with disabled HWP\n");
3520 	}
3521 
3522 	return 0;
3523 }
3524 device_initcall(intel_pstate_init);
3525 
3526 static int __init intel_pstate_setup(char *str)
3527 {
3528 	if (!str)
3529 		return -EINVAL;
3530 
3531 	if (!strcmp(str, "disable"))
3532 		no_load = 1;
3533 	else if (!strcmp(str, "active"))
3534 		default_driver = &intel_pstate;
3535 	else if (!strcmp(str, "passive"))
3536 		default_driver = &intel_cpufreq;
3537 
3538 	if (!strcmp(str, "no_hwp"))
3539 		no_hwp = 1;
3540 
3541 	if (!strcmp(str, "force"))
3542 		force_load = 1;
3543 	if (!strcmp(str, "hwp_only"))
3544 		hwp_only = 1;
3545 	if (!strcmp(str, "per_cpu_perf_limits"))
3546 		per_cpu_limits = true;
3547 
3548 #ifdef CONFIG_ACPI
3549 	if (!strcmp(str, "support_acpi_ppc"))
3550 		acpi_ppc = true;
3551 #endif
3552 
3553 	return 0;
3554 }
3555 early_param("intel_pstate", intel_pstate_setup);
3556 
3557 MODULE_AUTHOR("Dirk Brandewie <dirk.j.brandewie@intel.com>");
3558 MODULE_DESCRIPTION("'intel_pstate' - P state driver Intel Core processors");
3559