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