1 // SPDX-License-Identifier: GPL-2.0-only
2 /*
3  * intel_pstate.c: Native P state management for Intel processors
4  *
5  * (C) Copyright 2012 Intel Corporation
6  * Author: Dirk Brandewie <dirk.j.brandewie@intel.com>
7  */
8 
9 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
10 
11 #include <linux/kernel.h>
12 #include <linux/kernel_stat.h>
13 #include <linux/module.h>
14 #include <linux/ktime.h>
15 #include <linux/hrtimer.h>
16 #include <linux/tick.h>
17 #include <linux/slab.h>
18 #include <linux/sched/cpufreq.h>
19 #include <linux/list.h>
20 #include <linux/cpu.h>
21 #include <linux/cpufreq.h>
22 #include <linux/sysfs.h>
23 #include <linux/types.h>
24 #include <linux/fs.h>
25 #include <linux/acpi.h>
26 #include <linux/vmalloc.h>
27 #include <linux/pm_qos.h>
28 #include <trace/events/power.h>
29 
30 #include <asm/cpu.h>
31 #include <asm/div64.h>
32 #include <asm/msr.h>
33 #include <asm/cpu_device_id.h>
34 #include <asm/cpufeature.h>
35 #include <asm/intel-family.h>
36 #include "../drivers/thermal/intel/thermal_interrupt.h"
37 
38 #define INTEL_PSTATE_SAMPLING_INTERVAL	(10 * NSEC_PER_MSEC)
39 
40 #define INTEL_CPUFREQ_TRANSITION_LATENCY	20000
41 #define INTEL_CPUFREQ_TRANSITION_DELAY_HWP	5000
42 #define INTEL_CPUFREQ_TRANSITION_DELAY		500
43 
44 #ifdef CONFIG_ACPI
45 #include <acpi/processor.h>
46 #include <acpi/cppc_acpi.h>
47 #endif
48 
49 #define FRAC_BITS 8
50 #define int_tofp(X) ((int64_t)(X) << FRAC_BITS)
51 #define fp_toint(X) ((X) >> FRAC_BITS)
52 
53 #define ONE_EIGHTH_FP ((int64_t)1 << (FRAC_BITS - 3))
54 
55 #define EXT_BITS 6
56 #define EXT_FRAC_BITS (EXT_BITS + FRAC_BITS)
57 #define fp_ext_toint(X) ((X) >> EXT_FRAC_BITS)
58 #define int_ext_tofp(X) ((int64_t)(X) << EXT_FRAC_BITS)
59 
60 static inline int32_t mul_fp(int32_t x, int32_t y)
61 {
62 	return ((int64_t)x * (int64_t)y) >> FRAC_BITS;
63 }
64 
65 static inline int32_t div_fp(s64 x, s64 y)
66 {
67 	return div64_s64((int64_t)x << FRAC_BITS, y);
68 }
69 
70 static inline int ceiling_fp(int32_t x)
71 {
72 	int mask, ret;
73 
74 	ret = fp_toint(x);
75 	mask = (1 << FRAC_BITS) - 1;
76 	if (x & mask)
77 		ret += 1;
78 	return ret;
79 }
80 
81 static inline u64 mul_ext_fp(u64 x, u64 y)
82 {
83 	return (x * y) >> EXT_FRAC_BITS;
84 }
85 
86 static inline u64 div_ext_fp(u64 x, u64 y)
87 {
88 	return div64_u64(x << EXT_FRAC_BITS, y);
89 }
90 
91 /**
92  * struct sample -	Store performance sample
93  * @core_avg_perf:	Ratio of APERF/MPERF which is the actual average
94  *			performance during last sample period
95  * @busy_scaled:	Scaled busy value which is used to calculate next
96  *			P state. This can be different than core_avg_perf
97  *			to account for cpu idle period
98  * @aperf:		Difference of actual performance frequency clock count
99  *			read from APERF MSR between last and current sample
100  * @mperf:		Difference of maximum performance frequency clock count
101  *			read from MPERF MSR between last and current sample
102  * @tsc:		Difference of time stamp counter between last and
103  *			current sample
104  * @time:		Current time from scheduler
105  *
106  * This structure is used in the cpudata structure to store performance sample
107  * data for choosing next P State.
108  */
109 struct sample {
110 	int32_t core_avg_perf;
111 	int32_t busy_scaled;
112 	u64 aperf;
113 	u64 mperf;
114 	u64 tsc;
115 	u64 time;
116 };
117 
118 /**
119  * struct pstate_data - Store P state data
120  * @current_pstate:	Current requested P state
121  * @min_pstate:		Min P state possible for this platform
122  * @max_pstate:		Max P state possible for this platform
123  * @max_pstate_physical:This is physical Max P state for a processor
124  *			This can be higher than the max_pstate which can
125  *			be limited by platform thermal design power limits
126  * @perf_ctl_scaling:	PERF_CTL P-state to frequency scaling factor
127  * @scaling:		Scaling factor between performance and frequency
128  * @turbo_pstate:	Max Turbo P state possible for this platform
129  * @min_freq:		@min_pstate frequency in cpufreq units
130  * @max_freq:		@max_pstate frequency in cpufreq units
131  * @turbo_freq:		@turbo_pstate frequency in cpufreq units
132  *
133  * Stores the per cpu model P state limits and current P state.
134  */
135 struct pstate_data {
136 	int	current_pstate;
137 	int	min_pstate;
138 	int	max_pstate;
139 	int	max_pstate_physical;
140 	int	perf_ctl_scaling;
141 	int	scaling;
142 	int	turbo_pstate;
143 	unsigned int min_freq;
144 	unsigned int max_freq;
145 	unsigned int turbo_freq;
146 };
147 
148 /**
149  * struct vid_data -	Stores voltage information data
150  * @min:		VID data for this platform corresponding to
151  *			the lowest P state
152  * @max:		VID data corresponding to the highest P State.
153  * @turbo:		VID data for turbo P state
154  * @ratio:		Ratio of (vid max - vid min) /
155  *			(max P state - Min P State)
156  *
157  * Stores the voltage data for DVFS (Dynamic Voltage and Frequency Scaling)
158  * This data is used in Atom platforms, where in addition to target P state,
159  * the voltage data needs to be specified to select next P State.
160  */
161 struct vid_data {
162 	int min;
163 	int max;
164 	int turbo;
165 	int32_t ratio;
166 };
167 
168 /**
169  * struct global_params - Global parameters, mostly tunable via sysfs.
170  * @no_turbo:		Whether or not to use turbo P-states.
171  * @turbo_disabled:	Whether or not turbo P-states are available at all,
172  *			based on the MSR_IA32_MISC_ENABLE value and whether or
173  *			not the maximum reported turbo P-state is different from
174  *			the maximum reported non-turbo one.
175  * @turbo_disabled_mf:	The @turbo_disabled value reflected by cpuinfo.max_freq.
176  * @min_perf_pct:	Minimum capacity limit in percent of the maximum turbo
177  *			P-state capacity.
178  * @max_perf_pct:	Maximum capacity limit in percent of the maximum turbo
179  *			P-state capacity.
180  */
181 struct global_params {
182 	bool no_turbo;
183 	bool turbo_disabled;
184 	bool turbo_disabled_mf;
185 	int max_perf_pct;
186 	int min_perf_pct;
187 };
188 
189 /**
190  * struct cpudata -	Per CPU instance data storage
191  * @cpu:		CPU number for this instance data
192  * @policy:		CPUFreq policy value
193  * @update_util:	CPUFreq utility callback information
194  * @update_util_set:	CPUFreq utility callback is set
195  * @iowait_boost:	iowait-related boost fraction
196  * @last_update:	Time of the last update.
197  * @pstate:		Stores P state limits for this CPU
198  * @vid:		Stores VID limits for this CPU
199  * @last_sample_time:	Last Sample time
200  * @aperf_mperf_shift:	APERF vs MPERF counting frequency difference
201  * @prev_aperf:		Last APERF value read from APERF MSR
202  * @prev_mperf:		Last MPERF value read from MPERF MSR
203  * @prev_tsc:		Last timestamp counter (TSC) value
204  * @prev_cummulative_iowait: IO Wait time difference from last and
205  *			current sample
206  * @sample:		Storage for storing last Sample data
207  * @min_perf_ratio:	Minimum capacity in terms of PERF or HWP ratios
208  * @max_perf_ratio:	Maximum capacity in terms of PERF or HWP ratios
209  * @acpi_perf_data:	Stores ACPI perf information read from _PSS
210  * @valid_pss_table:	Set to true for valid ACPI _PSS entries found
211  * @epp_powersave:	Last saved HWP energy performance preference
212  *			(EPP) or energy performance bias (EPB),
213  *			when policy switched to performance
214  * @epp_policy:		Last saved policy used to set EPP/EPB
215  * @epp_default:	Power on default HWP energy performance
216  *			preference/bias
217  * @epp_cached		Cached HWP energy-performance preference value
218  * @hwp_req_cached:	Cached value of the last HWP Request MSR
219  * @hwp_cap_cached:	Cached value of the last HWP Capabilities MSR
220  * @last_io_update:	Last time when IO wake flag was set
221  * @sched_flags:	Store scheduler flags for possible cross CPU update
222  * @hwp_boost_min:	Last HWP boosted min performance
223  * @suspended:		Whether or not the driver has been suspended.
224  * @hwp_notify_work:	workqueue for HWP notifications.
225  *
226  * This structure stores per CPU instance data for all CPUs.
227  */
228 struct cpudata {
229 	int cpu;
230 
231 	unsigned int policy;
232 	struct update_util_data update_util;
233 	bool   update_util_set;
234 
235 	struct pstate_data pstate;
236 	struct vid_data vid;
237 
238 	u64	last_update;
239 	u64	last_sample_time;
240 	u64	aperf_mperf_shift;
241 	u64	prev_aperf;
242 	u64	prev_mperf;
243 	u64	prev_tsc;
244 	u64	prev_cummulative_iowait;
245 	struct sample sample;
246 	int32_t	min_perf_ratio;
247 	int32_t	max_perf_ratio;
248 #ifdef CONFIG_ACPI
249 	struct acpi_processor_performance acpi_perf_data;
250 	bool valid_pss_table;
251 #endif
252 	unsigned int iowait_boost;
253 	s16 epp_powersave;
254 	s16 epp_policy;
255 	s16 epp_default;
256 	s16 epp_cached;
257 	u64 hwp_req_cached;
258 	u64 hwp_cap_cached;
259 	u64 last_io_update;
260 	unsigned int sched_flags;
261 	u32 hwp_boost_min;
262 	bool suspended;
263 	struct delayed_work hwp_notify_work;
264 };
265 
266 static struct cpudata **all_cpu_data;
267 
268 /**
269  * struct pstate_funcs - Per CPU model specific callbacks
270  * @get_max:		Callback to get maximum non turbo effective P state
271  * @get_max_physical:	Callback to get maximum non turbo physical P state
272  * @get_min:		Callback to get minimum P state
273  * @get_turbo:		Callback to get turbo P state
274  * @get_scaling:	Callback to get frequency scaling factor
275  * @get_cpu_scaling:	Get frequency scaling factor for a given cpu
276  * @get_aperf_mperf_shift: Callback to get the APERF vs MPERF frequency difference
277  * @get_val:		Callback to convert P state to actual MSR write value
278  * @get_vid:		Callback to get VID data for Atom platforms
279  *
280  * Core and Atom CPU models have different way to get P State limits. This
281  * structure is used to store those callbacks.
282  */
283 struct pstate_funcs {
284 	int (*get_max)(int cpu);
285 	int (*get_max_physical)(int cpu);
286 	int (*get_min)(int cpu);
287 	int (*get_turbo)(int cpu);
288 	int (*get_scaling)(void);
289 	int (*get_cpu_scaling)(int cpu);
290 	int (*get_aperf_mperf_shift)(void);
291 	u64 (*get_val)(struct cpudata*, int pstate);
292 	void (*get_vid)(struct cpudata *);
293 };
294 
295 static struct pstate_funcs pstate_funcs __read_mostly;
296 
297 static int hwp_active __read_mostly;
298 static int hwp_mode_bdw __read_mostly;
299 static bool per_cpu_limits __read_mostly;
300 static bool hwp_boost __read_mostly;
301 static bool hwp_forced __read_mostly;
302 
303 static struct cpufreq_driver *intel_pstate_driver __read_mostly;
304 
305 #define HYBRID_SCALING_FACTOR	78741
306 
307 static inline int core_get_scaling(void)
308 {
309 	return 100000;
310 }
311 
312 #ifdef CONFIG_ACPI
313 static bool acpi_ppc;
314 #endif
315 
316 static struct global_params global;
317 
318 static DEFINE_MUTEX(intel_pstate_driver_lock);
319 static DEFINE_MUTEX(intel_pstate_limits_lock);
320 
321 #ifdef CONFIG_ACPI
322 
323 static bool intel_pstate_acpi_pm_profile_server(void)
324 {
325 	if (acpi_gbl_FADT.preferred_profile == PM_ENTERPRISE_SERVER ||
326 	    acpi_gbl_FADT.preferred_profile == PM_PERFORMANCE_SERVER)
327 		return true;
328 
329 	return false;
330 }
331 
332 static bool intel_pstate_get_ppc_enable_status(void)
333 {
334 	if (intel_pstate_acpi_pm_profile_server())
335 		return true;
336 
337 	return acpi_ppc;
338 }
339 
340 #ifdef CONFIG_ACPI_CPPC_LIB
341 
342 /* The work item is needed to avoid CPU hotplug locking issues */
343 static void intel_pstste_sched_itmt_work_fn(struct work_struct *work)
344 {
345 	sched_set_itmt_support();
346 }
347 
348 static DECLARE_WORK(sched_itmt_work, intel_pstste_sched_itmt_work_fn);
349 
350 #define CPPC_MAX_PERF	U8_MAX
351 
352 static void intel_pstate_set_itmt_prio(int cpu)
353 {
354 	struct cppc_perf_caps cppc_perf;
355 	static u32 max_highest_perf = 0, min_highest_perf = U32_MAX;
356 	int ret;
357 
358 	ret = cppc_get_perf_caps(cpu, &cppc_perf);
359 	if (ret)
360 		return;
361 
362 	/*
363 	 * On some systems with overclocking enabled, CPPC.highest_perf is hardcoded to 0xff.
364 	 * In this case we can't use CPPC.highest_perf to enable ITMT.
365 	 * In this case we can look at MSR_HWP_CAPABILITIES bits [8:0] to decide.
366 	 */
367 	if (cppc_perf.highest_perf == CPPC_MAX_PERF)
368 		cppc_perf.highest_perf = HWP_HIGHEST_PERF(READ_ONCE(all_cpu_data[cpu]->hwp_cap_cached));
369 
370 	/*
371 	 * The priorities can be set regardless of whether or not
372 	 * sched_set_itmt_support(true) has been called and it is valid to
373 	 * update them at any time after it has been called.
374 	 */
375 	sched_set_itmt_core_prio(cppc_perf.highest_perf, cpu);
376 
377 	if (max_highest_perf <= min_highest_perf) {
378 		if (cppc_perf.highest_perf > max_highest_perf)
379 			max_highest_perf = cppc_perf.highest_perf;
380 
381 		if (cppc_perf.highest_perf < min_highest_perf)
382 			min_highest_perf = cppc_perf.highest_perf;
383 
384 		if (max_highest_perf > min_highest_perf) {
385 			/*
386 			 * This code can be run during CPU online under the
387 			 * CPU hotplug locks, so sched_set_itmt_support()
388 			 * cannot be called from here.  Queue up a work item
389 			 * to invoke it.
390 			 */
391 			schedule_work(&sched_itmt_work);
392 		}
393 	}
394 }
395 
396 static int intel_pstate_get_cppc_guaranteed(int cpu)
397 {
398 	struct cppc_perf_caps cppc_perf;
399 	int ret;
400 
401 	ret = cppc_get_perf_caps(cpu, &cppc_perf);
402 	if (ret)
403 		return ret;
404 
405 	if (cppc_perf.guaranteed_perf)
406 		return cppc_perf.guaranteed_perf;
407 
408 	return cppc_perf.nominal_perf;
409 }
410 
411 static int intel_pstate_cppc_get_scaling(int cpu)
412 {
413 	struct cppc_perf_caps cppc_perf;
414 	int ret;
415 
416 	ret = cppc_get_perf_caps(cpu, &cppc_perf);
417 
418 	/*
419 	 * If the nominal frequency and the nominal performance are not
420 	 * zero and the ratio between them is not 100, return the hybrid
421 	 * scaling factor.
422 	 */
423 	if (!ret && cppc_perf.nominal_perf && cppc_perf.nominal_freq &&
424 	    cppc_perf.nominal_perf * 100 != cppc_perf.nominal_freq)
425 		return HYBRID_SCALING_FACTOR;
426 
427 	return core_get_scaling();
428 }
429 
430 #else /* CONFIG_ACPI_CPPC_LIB */
431 static inline void intel_pstate_set_itmt_prio(int cpu)
432 {
433 }
434 #endif /* CONFIG_ACPI_CPPC_LIB */
435 
436 static void intel_pstate_init_acpi_perf_limits(struct cpufreq_policy *policy)
437 {
438 	struct cpudata *cpu;
439 	int ret;
440 	int i;
441 
442 	if (hwp_active) {
443 		intel_pstate_set_itmt_prio(policy->cpu);
444 		return;
445 	}
446 
447 	if (!intel_pstate_get_ppc_enable_status())
448 		return;
449 
450 	cpu = all_cpu_data[policy->cpu];
451 
452 	ret = acpi_processor_register_performance(&cpu->acpi_perf_data,
453 						  policy->cpu);
454 	if (ret)
455 		return;
456 
457 	/*
458 	 * Check if the control value in _PSS is for PERF_CTL MSR, which should
459 	 * guarantee that the states returned by it map to the states in our
460 	 * list directly.
461 	 */
462 	if (cpu->acpi_perf_data.control_register.space_id !=
463 						ACPI_ADR_SPACE_FIXED_HARDWARE)
464 		goto err;
465 
466 	/*
467 	 * If there is only one entry _PSS, simply ignore _PSS and continue as
468 	 * usual without taking _PSS into account
469 	 */
470 	if (cpu->acpi_perf_data.state_count < 2)
471 		goto err;
472 
473 	pr_debug("CPU%u - ACPI _PSS perf data\n", policy->cpu);
474 	for (i = 0; i < cpu->acpi_perf_data.state_count; i++) {
475 		pr_debug("     %cP%d: %u MHz, %u mW, 0x%x\n",
476 			 (i == cpu->acpi_perf_data.state ? '*' : ' '), i,
477 			 (u32) cpu->acpi_perf_data.states[i].core_frequency,
478 			 (u32) cpu->acpi_perf_data.states[i].power,
479 			 (u32) cpu->acpi_perf_data.states[i].control);
480 	}
481 
482 	cpu->valid_pss_table = true;
483 	pr_debug("_PPC limits will be enforced\n");
484 
485 	return;
486 
487  err:
488 	cpu->valid_pss_table = false;
489 	acpi_processor_unregister_performance(policy->cpu);
490 }
491 
492 static void intel_pstate_exit_perf_limits(struct cpufreq_policy *policy)
493 {
494 	struct cpudata *cpu;
495 
496 	cpu = all_cpu_data[policy->cpu];
497 	if (!cpu->valid_pss_table)
498 		return;
499 
500 	acpi_processor_unregister_performance(policy->cpu);
501 }
502 #else /* CONFIG_ACPI */
503 static inline void intel_pstate_init_acpi_perf_limits(struct cpufreq_policy *policy)
504 {
505 }
506 
507 static inline void intel_pstate_exit_perf_limits(struct cpufreq_policy *policy)
508 {
509 }
510 
511 static inline bool intel_pstate_acpi_pm_profile_server(void)
512 {
513 	return false;
514 }
515 #endif /* CONFIG_ACPI */
516 
517 #ifndef CONFIG_ACPI_CPPC_LIB
518 static inline int intel_pstate_get_cppc_guaranteed(int cpu)
519 {
520 	return -ENOTSUPP;
521 }
522 
523 static int intel_pstate_cppc_get_scaling(int cpu)
524 {
525 	return core_get_scaling();
526 }
527 #endif /* CONFIG_ACPI_CPPC_LIB */
528 
529 static int intel_pstate_freq_to_hwp_rel(struct cpudata *cpu, int freq,
530 					unsigned int relation)
531 {
532 	if (freq == cpu->pstate.turbo_freq)
533 		return cpu->pstate.turbo_pstate;
534 
535 	if (freq == cpu->pstate.max_freq)
536 		return cpu->pstate.max_pstate;
537 
538 	switch (relation) {
539 	case CPUFREQ_RELATION_H:
540 		return freq / cpu->pstate.scaling;
541 	case CPUFREQ_RELATION_C:
542 		return DIV_ROUND_CLOSEST(freq, cpu->pstate.scaling);
543 	}
544 
545 	return DIV_ROUND_UP(freq, cpu->pstate.scaling);
546 }
547 
548 static int intel_pstate_freq_to_hwp(struct cpudata *cpu, int freq)
549 {
550 	return intel_pstate_freq_to_hwp_rel(cpu, freq, CPUFREQ_RELATION_L);
551 }
552 
553 /**
554  * intel_pstate_hybrid_hwp_adjust - Calibrate HWP performance levels.
555  * @cpu: Target CPU.
556  *
557  * On hybrid processors, HWP may expose more performance levels than there are
558  * P-states accessible through the PERF_CTL interface.  If that happens, the
559  * scaling factor between HWP performance levels and CPU frequency will be less
560  * than the scaling factor between P-state values and CPU frequency.
561  *
562  * In that case, adjust the CPU parameters used in computations accordingly.
563  */
564 static void intel_pstate_hybrid_hwp_adjust(struct cpudata *cpu)
565 {
566 	int perf_ctl_max_phys = cpu->pstate.max_pstate_physical;
567 	int perf_ctl_scaling = cpu->pstate.perf_ctl_scaling;
568 	int perf_ctl_turbo = pstate_funcs.get_turbo(cpu->cpu);
569 	int scaling = cpu->pstate.scaling;
570 	int freq;
571 
572 	pr_debug("CPU%d: perf_ctl_max_phys = %d\n", cpu->cpu, perf_ctl_max_phys);
573 	pr_debug("CPU%d: perf_ctl_turbo = %d\n", cpu->cpu, perf_ctl_turbo);
574 	pr_debug("CPU%d: perf_ctl_scaling = %d\n", cpu->cpu, perf_ctl_scaling);
575 	pr_debug("CPU%d: HWP_CAP guaranteed = %d\n", cpu->cpu, cpu->pstate.max_pstate);
576 	pr_debug("CPU%d: HWP_CAP highest = %d\n", cpu->cpu, cpu->pstate.turbo_pstate);
577 	pr_debug("CPU%d: HWP-to-frequency scaling factor: %d\n", cpu->cpu, scaling);
578 
579 	cpu->pstate.turbo_freq = rounddown(cpu->pstate.turbo_pstate * scaling,
580 					   perf_ctl_scaling);
581 	cpu->pstate.max_freq = rounddown(cpu->pstate.max_pstate * scaling,
582 					 perf_ctl_scaling);
583 
584 	freq = perf_ctl_max_phys * perf_ctl_scaling;
585 	cpu->pstate.max_pstate_physical = intel_pstate_freq_to_hwp(cpu, freq);
586 
587 	freq = cpu->pstate.min_pstate * perf_ctl_scaling;
588 	cpu->pstate.min_freq = freq;
589 	/*
590 	 * Cast the min P-state value retrieved via pstate_funcs.get_min() to
591 	 * the effective range of HWP performance levels.
592 	 */
593 	cpu->pstate.min_pstate = intel_pstate_freq_to_hwp(cpu, freq);
594 }
595 
596 static inline void update_turbo_state(void)
597 {
598 	u64 misc_en;
599 	struct cpudata *cpu;
600 
601 	cpu = all_cpu_data[0];
602 	rdmsrl(MSR_IA32_MISC_ENABLE, misc_en);
603 	global.turbo_disabled =
604 		(misc_en & MSR_IA32_MISC_ENABLE_TURBO_DISABLE ||
605 		 cpu->pstate.max_pstate == cpu->pstate.turbo_pstate);
606 }
607 
608 static int min_perf_pct_min(void)
609 {
610 	struct cpudata *cpu = all_cpu_data[0];
611 	int turbo_pstate = cpu->pstate.turbo_pstate;
612 
613 	return turbo_pstate ?
614 		(cpu->pstate.min_pstate * 100 / turbo_pstate) : 0;
615 }
616 
617 static s16 intel_pstate_get_epb(struct cpudata *cpu_data)
618 {
619 	u64 epb;
620 	int ret;
621 
622 	if (!boot_cpu_has(X86_FEATURE_EPB))
623 		return -ENXIO;
624 
625 	ret = rdmsrl_on_cpu(cpu_data->cpu, MSR_IA32_ENERGY_PERF_BIAS, &epb);
626 	if (ret)
627 		return (s16)ret;
628 
629 	return (s16)(epb & 0x0f);
630 }
631 
632 static s16 intel_pstate_get_epp(struct cpudata *cpu_data, u64 hwp_req_data)
633 {
634 	s16 epp;
635 
636 	if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
637 		/*
638 		 * When hwp_req_data is 0, means that caller didn't read
639 		 * MSR_HWP_REQUEST, so need to read and get EPP.
640 		 */
641 		if (!hwp_req_data) {
642 			epp = rdmsrl_on_cpu(cpu_data->cpu, MSR_HWP_REQUEST,
643 					    &hwp_req_data);
644 			if (epp)
645 				return epp;
646 		}
647 		epp = (hwp_req_data >> 24) & 0xff;
648 	} else {
649 		/* When there is no EPP present, HWP uses EPB settings */
650 		epp = intel_pstate_get_epb(cpu_data);
651 	}
652 
653 	return epp;
654 }
655 
656 static int intel_pstate_set_epb(int cpu, s16 pref)
657 {
658 	u64 epb;
659 	int ret;
660 
661 	if (!boot_cpu_has(X86_FEATURE_EPB))
662 		return -ENXIO;
663 
664 	ret = rdmsrl_on_cpu(cpu, MSR_IA32_ENERGY_PERF_BIAS, &epb);
665 	if (ret)
666 		return ret;
667 
668 	epb = (epb & ~0x0f) | pref;
669 	wrmsrl_on_cpu(cpu, MSR_IA32_ENERGY_PERF_BIAS, epb);
670 
671 	return 0;
672 }
673 
674 /*
675  * EPP/EPB display strings corresponding to EPP index in the
676  * energy_perf_strings[]
677  *	index		String
678  *-------------------------------------
679  *	0		default
680  *	1		performance
681  *	2		balance_performance
682  *	3		balance_power
683  *	4		power
684  */
685 
686 enum energy_perf_value_index {
687 	EPP_INDEX_DEFAULT = 0,
688 	EPP_INDEX_PERFORMANCE,
689 	EPP_INDEX_BALANCE_PERFORMANCE,
690 	EPP_INDEX_BALANCE_POWERSAVE,
691 	EPP_INDEX_POWERSAVE,
692 };
693 
694 static const char * const energy_perf_strings[] = {
695 	[EPP_INDEX_DEFAULT] = "default",
696 	[EPP_INDEX_PERFORMANCE] = "performance",
697 	[EPP_INDEX_BALANCE_PERFORMANCE] = "balance_performance",
698 	[EPP_INDEX_BALANCE_POWERSAVE] = "balance_power",
699 	[EPP_INDEX_POWERSAVE] = "power",
700 	NULL
701 };
702 static unsigned int epp_values[] = {
703 	[EPP_INDEX_DEFAULT] = 0, /* Unused index */
704 	[EPP_INDEX_PERFORMANCE] = HWP_EPP_PERFORMANCE,
705 	[EPP_INDEX_BALANCE_PERFORMANCE] = HWP_EPP_BALANCE_PERFORMANCE,
706 	[EPP_INDEX_BALANCE_POWERSAVE] = HWP_EPP_BALANCE_POWERSAVE,
707 	[EPP_INDEX_POWERSAVE] = HWP_EPP_POWERSAVE,
708 };
709 
710 static int intel_pstate_get_energy_pref_index(struct cpudata *cpu_data, int *raw_epp)
711 {
712 	s16 epp;
713 	int index = -EINVAL;
714 
715 	*raw_epp = 0;
716 	epp = intel_pstate_get_epp(cpu_data, 0);
717 	if (epp < 0)
718 		return epp;
719 
720 	if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
721 		if (epp == epp_values[EPP_INDEX_PERFORMANCE])
722 			return EPP_INDEX_PERFORMANCE;
723 		if (epp == epp_values[EPP_INDEX_BALANCE_PERFORMANCE])
724 			return EPP_INDEX_BALANCE_PERFORMANCE;
725 		if (epp == epp_values[EPP_INDEX_BALANCE_POWERSAVE])
726 			return EPP_INDEX_BALANCE_POWERSAVE;
727 		if (epp == epp_values[EPP_INDEX_POWERSAVE])
728 			return EPP_INDEX_POWERSAVE;
729 		*raw_epp = epp;
730 		return 0;
731 	} else if (boot_cpu_has(X86_FEATURE_EPB)) {
732 		/*
733 		 * Range:
734 		 *	0x00-0x03	:	Performance
735 		 *	0x04-0x07	:	Balance performance
736 		 *	0x08-0x0B	:	Balance power
737 		 *	0x0C-0x0F	:	Power
738 		 * The EPB is a 4 bit value, but our ranges restrict the
739 		 * value which can be set. Here only using top two bits
740 		 * effectively.
741 		 */
742 		index = (epp >> 2) + 1;
743 	}
744 
745 	return index;
746 }
747 
748 static int intel_pstate_set_epp(struct cpudata *cpu, u32 epp)
749 {
750 	int ret;
751 
752 	/*
753 	 * Use the cached HWP Request MSR value, because in the active mode the
754 	 * register itself may be updated by intel_pstate_hwp_boost_up() or
755 	 * intel_pstate_hwp_boost_down() at any time.
756 	 */
757 	u64 value = READ_ONCE(cpu->hwp_req_cached);
758 
759 	value &= ~GENMASK_ULL(31, 24);
760 	value |= (u64)epp << 24;
761 	/*
762 	 * The only other updater of hwp_req_cached in the active mode,
763 	 * intel_pstate_hwp_set(), is called under the same lock as this
764 	 * function, so it cannot run in parallel with the update below.
765 	 */
766 	WRITE_ONCE(cpu->hwp_req_cached, value);
767 	ret = wrmsrl_on_cpu(cpu->cpu, MSR_HWP_REQUEST, value);
768 	if (!ret)
769 		cpu->epp_cached = epp;
770 
771 	return ret;
772 }
773 
774 static int intel_pstate_set_energy_pref_index(struct cpudata *cpu_data,
775 					      int pref_index, bool use_raw,
776 					      u32 raw_epp)
777 {
778 	int epp = -EINVAL;
779 	int ret;
780 
781 	if (!pref_index)
782 		epp = cpu_data->epp_default;
783 
784 	if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
785 		if (use_raw)
786 			epp = raw_epp;
787 		else if (epp == -EINVAL)
788 			epp = epp_values[pref_index];
789 
790 		/*
791 		 * To avoid confusion, refuse to set EPP to any values different
792 		 * from 0 (performance) if the current policy is "performance",
793 		 * because those values would be overridden.
794 		 */
795 		if (epp > 0 && cpu_data->policy == CPUFREQ_POLICY_PERFORMANCE)
796 			return -EBUSY;
797 
798 		ret = intel_pstate_set_epp(cpu_data, epp);
799 	} else {
800 		if (epp == -EINVAL)
801 			epp = (pref_index - 1) << 2;
802 		ret = intel_pstate_set_epb(cpu_data->cpu, epp);
803 	}
804 
805 	return ret;
806 }
807 
808 static ssize_t show_energy_performance_available_preferences(
809 				struct cpufreq_policy *policy, char *buf)
810 {
811 	int i = 0;
812 	int ret = 0;
813 
814 	while (energy_perf_strings[i] != NULL)
815 		ret += sprintf(&buf[ret], "%s ", energy_perf_strings[i++]);
816 
817 	ret += sprintf(&buf[ret], "\n");
818 
819 	return ret;
820 }
821 
822 cpufreq_freq_attr_ro(energy_performance_available_preferences);
823 
824 static struct cpufreq_driver intel_pstate;
825 
826 static ssize_t store_energy_performance_preference(
827 		struct cpufreq_policy *policy, const char *buf, size_t count)
828 {
829 	struct cpudata *cpu = all_cpu_data[policy->cpu];
830 	char str_preference[21];
831 	bool raw = false;
832 	ssize_t ret;
833 	u32 epp = 0;
834 
835 	ret = sscanf(buf, "%20s", str_preference);
836 	if (ret != 1)
837 		return -EINVAL;
838 
839 	ret = match_string(energy_perf_strings, -1, str_preference);
840 	if (ret < 0) {
841 		if (!boot_cpu_has(X86_FEATURE_HWP_EPP))
842 			return ret;
843 
844 		ret = kstrtouint(buf, 10, &epp);
845 		if (ret)
846 			return ret;
847 
848 		if (epp > 255)
849 			return -EINVAL;
850 
851 		raw = true;
852 	}
853 
854 	/*
855 	 * This function runs with the policy R/W semaphore held, which
856 	 * guarantees that the driver pointer will not change while it is
857 	 * running.
858 	 */
859 	if (!intel_pstate_driver)
860 		return -EAGAIN;
861 
862 	mutex_lock(&intel_pstate_limits_lock);
863 
864 	if (intel_pstate_driver == &intel_pstate) {
865 		ret = intel_pstate_set_energy_pref_index(cpu, ret, raw, epp);
866 	} else {
867 		/*
868 		 * In the passive mode the governor needs to be stopped on the
869 		 * target CPU before the EPP update and restarted after it,
870 		 * which is super-heavy-weight, so make sure it is worth doing
871 		 * upfront.
872 		 */
873 		if (!raw)
874 			epp = ret ? epp_values[ret] : cpu->epp_default;
875 
876 		if (cpu->epp_cached != epp) {
877 			int err;
878 
879 			cpufreq_stop_governor(policy);
880 			ret = intel_pstate_set_epp(cpu, epp);
881 			err = cpufreq_start_governor(policy);
882 			if (!ret)
883 				ret = err;
884 		} else {
885 			ret = 0;
886 		}
887 	}
888 
889 	mutex_unlock(&intel_pstate_limits_lock);
890 
891 	return ret ?: count;
892 }
893 
894 static ssize_t show_energy_performance_preference(
895 				struct cpufreq_policy *policy, char *buf)
896 {
897 	struct cpudata *cpu_data = all_cpu_data[policy->cpu];
898 	int preference, raw_epp;
899 
900 	preference = intel_pstate_get_energy_pref_index(cpu_data, &raw_epp);
901 	if (preference < 0)
902 		return preference;
903 
904 	if (raw_epp)
905 		return  sprintf(buf, "%d\n", raw_epp);
906 	else
907 		return  sprintf(buf, "%s\n", energy_perf_strings[preference]);
908 }
909 
910 cpufreq_freq_attr_rw(energy_performance_preference);
911 
912 static ssize_t show_base_frequency(struct cpufreq_policy *policy, char *buf)
913 {
914 	struct cpudata *cpu = all_cpu_data[policy->cpu];
915 	int ratio, freq;
916 
917 	ratio = intel_pstate_get_cppc_guaranteed(policy->cpu);
918 	if (ratio <= 0) {
919 		u64 cap;
920 
921 		rdmsrl_on_cpu(policy->cpu, MSR_HWP_CAPABILITIES, &cap);
922 		ratio = HWP_GUARANTEED_PERF(cap);
923 	}
924 
925 	freq = ratio * cpu->pstate.scaling;
926 	if (cpu->pstate.scaling != cpu->pstate.perf_ctl_scaling)
927 		freq = rounddown(freq, cpu->pstate.perf_ctl_scaling);
928 
929 	return sprintf(buf, "%d\n", freq);
930 }
931 
932 cpufreq_freq_attr_ro(base_frequency);
933 
934 static struct freq_attr *hwp_cpufreq_attrs[] = {
935 	&energy_performance_preference,
936 	&energy_performance_available_preferences,
937 	&base_frequency,
938 	NULL,
939 };
940 
941 static void __intel_pstate_get_hwp_cap(struct cpudata *cpu)
942 {
943 	u64 cap;
944 
945 	rdmsrl_on_cpu(cpu->cpu, MSR_HWP_CAPABILITIES, &cap);
946 	WRITE_ONCE(cpu->hwp_cap_cached, cap);
947 	cpu->pstate.max_pstate = HWP_GUARANTEED_PERF(cap);
948 	cpu->pstate.turbo_pstate = HWP_HIGHEST_PERF(cap);
949 }
950 
951 static void intel_pstate_get_hwp_cap(struct cpudata *cpu)
952 {
953 	int scaling = cpu->pstate.scaling;
954 
955 	__intel_pstate_get_hwp_cap(cpu);
956 
957 	cpu->pstate.max_freq = cpu->pstate.max_pstate * scaling;
958 	cpu->pstate.turbo_freq = cpu->pstate.turbo_pstate * scaling;
959 	if (scaling != cpu->pstate.perf_ctl_scaling) {
960 		int perf_ctl_scaling = cpu->pstate.perf_ctl_scaling;
961 
962 		cpu->pstate.max_freq = rounddown(cpu->pstate.max_freq,
963 						 perf_ctl_scaling);
964 		cpu->pstate.turbo_freq = rounddown(cpu->pstate.turbo_freq,
965 						   perf_ctl_scaling);
966 	}
967 }
968 
969 static void intel_pstate_hwp_set(unsigned int cpu)
970 {
971 	struct cpudata *cpu_data = all_cpu_data[cpu];
972 	int max, min;
973 	u64 value;
974 	s16 epp;
975 
976 	max = cpu_data->max_perf_ratio;
977 	min = cpu_data->min_perf_ratio;
978 
979 	if (cpu_data->policy == CPUFREQ_POLICY_PERFORMANCE)
980 		min = max;
981 
982 	rdmsrl_on_cpu(cpu, MSR_HWP_REQUEST, &value);
983 
984 	value &= ~HWP_MIN_PERF(~0L);
985 	value |= HWP_MIN_PERF(min);
986 
987 	value &= ~HWP_MAX_PERF(~0L);
988 	value |= HWP_MAX_PERF(max);
989 
990 	if (cpu_data->epp_policy == cpu_data->policy)
991 		goto skip_epp;
992 
993 	cpu_data->epp_policy = cpu_data->policy;
994 
995 	if (cpu_data->policy == CPUFREQ_POLICY_PERFORMANCE) {
996 		epp = intel_pstate_get_epp(cpu_data, value);
997 		cpu_data->epp_powersave = epp;
998 		/* If EPP read was failed, then don't try to write */
999 		if (epp < 0)
1000 			goto skip_epp;
1001 
1002 		epp = 0;
1003 	} else {
1004 		/* skip setting EPP, when saved value is invalid */
1005 		if (cpu_data->epp_powersave < 0)
1006 			goto skip_epp;
1007 
1008 		/*
1009 		 * No need to restore EPP when it is not zero. This
1010 		 * means:
1011 		 *  - Policy is not changed
1012 		 *  - user has manually changed
1013 		 *  - Error reading EPB
1014 		 */
1015 		epp = intel_pstate_get_epp(cpu_data, value);
1016 		if (epp)
1017 			goto skip_epp;
1018 
1019 		epp = cpu_data->epp_powersave;
1020 	}
1021 	if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
1022 		value &= ~GENMASK_ULL(31, 24);
1023 		value |= (u64)epp << 24;
1024 	} else {
1025 		intel_pstate_set_epb(cpu, epp);
1026 	}
1027 skip_epp:
1028 	WRITE_ONCE(cpu_data->hwp_req_cached, value);
1029 	wrmsrl_on_cpu(cpu, MSR_HWP_REQUEST, value);
1030 }
1031 
1032 static void intel_pstate_disable_hwp_interrupt(struct cpudata *cpudata);
1033 
1034 static void intel_pstate_hwp_offline(struct cpudata *cpu)
1035 {
1036 	u64 value = READ_ONCE(cpu->hwp_req_cached);
1037 	int min_perf;
1038 
1039 	intel_pstate_disable_hwp_interrupt(cpu);
1040 
1041 	if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
1042 		/*
1043 		 * In case the EPP has been set to "performance" by the
1044 		 * active mode "performance" scaling algorithm, replace that
1045 		 * temporary value with the cached EPP one.
1046 		 */
1047 		value &= ~GENMASK_ULL(31, 24);
1048 		value |= HWP_ENERGY_PERF_PREFERENCE(cpu->epp_cached);
1049 		/*
1050 		 * However, make sure that EPP will be set to "performance" when
1051 		 * the CPU is brought back online again and the "performance"
1052 		 * scaling algorithm is still in effect.
1053 		 */
1054 		cpu->epp_policy = CPUFREQ_POLICY_UNKNOWN;
1055 	}
1056 
1057 	/*
1058 	 * Clear the desired perf field in the cached HWP request value to
1059 	 * prevent nonzero desired values from being leaked into the active
1060 	 * mode.
1061 	 */
1062 	value &= ~HWP_DESIRED_PERF(~0L);
1063 	WRITE_ONCE(cpu->hwp_req_cached, value);
1064 
1065 	value &= ~GENMASK_ULL(31, 0);
1066 	min_perf = HWP_LOWEST_PERF(READ_ONCE(cpu->hwp_cap_cached));
1067 
1068 	/* Set hwp_max = hwp_min */
1069 	value |= HWP_MAX_PERF(min_perf);
1070 	value |= HWP_MIN_PERF(min_perf);
1071 
1072 	/* Set EPP to min */
1073 	if (boot_cpu_has(X86_FEATURE_HWP_EPP))
1074 		value |= HWP_ENERGY_PERF_PREFERENCE(HWP_EPP_POWERSAVE);
1075 
1076 	wrmsrl_on_cpu(cpu->cpu, MSR_HWP_REQUEST, value);
1077 }
1078 
1079 #define POWER_CTL_EE_ENABLE	1
1080 #define POWER_CTL_EE_DISABLE	2
1081 
1082 static int power_ctl_ee_state;
1083 
1084 static void set_power_ctl_ee_state(bool input)
1085 {
1086 	u64 power_ctl;
1087 
1088 	mutex_lock(&intel_pstate_driver_lock);
1089 	rdmsrl(MSR_IA32_POWER_CTL, power_ctl);
1090 	if (input) {
1091 		power_ctl &= ~BIT(MSR_IA32_POWER_CTL_BIT_EE);
1092 		power_ctl_ee_state = POWER_CTL_EE_ENABLE;
1093 	} else {
1094 		power_ctl |= BIT(MSR_IA32_POWER_CTL_BIT_EE);
1095 		power_ctl_ee_state = POWER_CTL_EE_DISABLE;
1096 	}
1097 	wrmsrl(MSR_IA32_POWER_CTL, power_ctl);
1098 	mutex_unlock(&intel_pstate_driver_lock);
1099 }
1100 
1101 static void intel_pstate_hwp_enable(struct cpudata *cpudata);
1102 
1103 static void intel_pstate_hwp_reenable(struct cpudata *cpu)
1104 {
1105 	intel_pstate_hwp_enable(cpu);
1106 	wrmsrl_on_cpu(cpu->cpu, MSR_HWP_REQUEST, READ_ONCE(cpu->hwp_req_cached));
1107 }
1108 
1109 static int intel_pstate_suspend(struct cpufreq_policy *policy)
1110 {
1111 	struct cpudata *cpu = all_cpu_data[policy->cpu];
1112 
1113 	pr_debug("CPU %d suspending\n", cpu->cpu);
1114 
1115 	cpu->suspended = true;
1116 
1117 	/* disable HWP interrupt and cancel any pending work */
1118 	intel_pstate_disable_hwp_interrupt(cpu);
1119 
1120 	return 0;
1121 }
1122 
1123 static int intel_pstate_resume(struct cpufreq_policy *policy)
1124 {
1125 	struct cpudata *cpu = all_cpu_data[policy->cpu];
1126 
1127 	pr_debug("CPU %d resuming\n", cpu->cpu);
1128 
1129 	/* Only restore if the system default is changed */
1130 	if (power_ctl_ee_state == POWER_CTL_EE_ENABLE)
1131 		set_power_ctl_ee_state(true);
1132 	else if (power_ctl_ee_state == POWER_CTL_EE_DISABLE)
1133 		set_power_ctl_ee_state(false);
1134 
1135 	if (cpu->suspended && hwp_active) {
1136 		mutex_lock(&intel_pstate_limits_lock);
1137 
1138 		/* Re-enable HWP, because "online" has not done that. */
1139 		intel_pstate_hwp_reenable(cpu);
1140 
1141 		mutex_unlock(&intel_pstate_limits_lock);
1142 	}
1143 
1144 	cpu->suspended = false;
1145 
1146 	return 0;
1147 }
1148 
1149 static void intel_pstate_update_policies(void)
1150 {
1151 	int cpu;
1152 
1153 	for_each_possible_cpu(cpu)
1154 		cpufreq_update_policy(cpu);
1155 }
1156 
1157 static void __intel_pstate_update_max_freq(struct cpudata *cpudata,
1158 					   struct cpufreq_policy *policy)
1159 {
1160 	policy->cpuinfo.max_freq = global.turbo_disabled_mf ?
1161 			cpudata->pstate.max_freq : cpudata->pstate.turbo_freq;
1162 	refresh_frequency_limits(policy);
1163 }
1164 
1165 static void intel_pstate_update_max_freq(unsigned int cpu)
1166 {
1167 	struct cpufreq_policy *policy = cpufreq_cpu_acquire(cpu);
1168 
1169 	if (!policy)
1170 		return;
1171 
1172 	__intel_pstate_update_max_freq(all_cpu_data[cpu], policy);
1173 
1174 	cpufreq_cpu_release(policy);
1175 }
1176 
1177 static void intel_pstate_update_limits(unsigned int cpu)
1178 {
1179 	mutex_lock(&intel_pstate_driver_lock);
1180 
1181 	update_turbo_state();
1182 	/*
1183 	 * If turbo has been turned on or off globally, policy limits for
1184 	 * all CPUs need to be updated to reflect that.
1185 	 */
1186 	if (global.turbo_disabled_mf != global.turbo_disabled) {
1187 		global.turbo_disabled_mf = global.turbo_disabled;
1188 		arch_set_max_freq_ratio(global.turbo_disabled);
1189 		for_each_possible_cpu(cpu)
1190 			intel_pstate_update_max_freq(cpu);
1191 	} else {
1192 		cpufreq_update_policy(cpu);
1193 	}
1194 
1195 	mutex_unlock(&intel_pstate_driver_lock);
1196 }
1197 
1198 /************************** sysfs begin ************************/
1199 #define show_one(file_name, object)					\
1200 	static ssize_t show_##file_name					\
1201 	(struct kobject *kobj, struct kobj_attribute *attr, char *buf)	\
1202 	{								\
1203 		return sprintf(buf, "%u\n", global.object);		\
1204 	}
1205 
1206 static ssize_t intel_pstate_show_status(char *buf);
1207 static int intel_pstate_update_status(const char *buf, size_t size);
1208 
1209 static ssize_t show_status(struct kobject *kobj,
1210 			   struct kobj_attribute *attr, char *buf)
1211 {
1212 	ssize_t ret;
1213 
1214 	mutex_lock(&intel_pstate_driver_lock);
1215 	ret = intel_pstate_show_status(buf);
1216 	mutex_unlock(&intel_pstate_driver_lock);
1217 
1218 	return ret;
1219 }
1220 
1221 static ssize_t store_status(struct kobject *a, struct kobj_attribute *b,
1222 			    const char *buf, size_t count)
1223 {
1224 	char *p = memchr(buf, '\n', count);
1225 	int ret;
1226 
1227 	mutex_lock(&intel_pstate_driver_lock);
1228 	ret = intel_pstate_update_status(buf, p ? p - buf : count);
1229 	mutex_unlock(&intel_pstate_driver_lock);
1230 
1231 	return ret < 0 ? ret : count;
1232 }
1233 
1234 static ssize_t show_turbo_pct(struct kobject *kobj,
1235 				struct kobj_attribute *attr, char *buf)
1236 {
1237 	struct cpudata *cpu;
1238 	int total, no_turbo, turbo_pct;
1239 	uint32_t turbo_fp;
1240 
1241 	mutex_lock(&intel_pstate_driver_lock);
1242 
1243 	if (!intel_pstate_driver) {
1244 		mutex_unlock(&intel_pstate_driver_lock);
1245 		return -EAGAIN;
1246 	}
1247 
1248 	cpu = all_cpu_data[0];
1249 
1250 	total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
1251 	no_turbo = cpu->pstate.max_pstate - cpu->pstate.min_pstate + 1;
1252 	turbo_fp = div_fp(no_turbo, total);
1253 	turbo_pct = 100 - fp_toint(mul_fp(turbo_fp, int_tofp(100)));
1254 
1255 	mutex_unlock(&intel_pstate_driver_lock);
1256 
1257 	return sprintf(buf, "%u\n", turbo_pct);
1258 }
1259 
1260 static ssize_t show_num_pstates(struct kobject *kobj,
1261 				struct kobj_attribute *attr, char *buf)
1262 {
1263 	struct cpudata *cpu;
1264 	int total;
1265 
1266 	mutex_lock(&intel_pstate_driver_lock);
1267 
1268 	if (!intel_pstate_driver) {
1269 		mutex_unlock(&intel_pstate_driver_lock);
1270 		return -EAGAIN;
1271 	}
1272 
1273 	cpu = all_cpu_data[0];
1274 	total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
1275 
1276 	mutex_unlock(&intel_pstate_driver_lock);
1277 
1278 	return sprintf(buf, "%u\n", total);
1279 }
1280 
1281 static ssize_t show_no_turbo(struct kobject *kobj,
1282 			     struct kobj_attribute *attr, char *buf)
1283 {
1284 	ssize_t ret;
1285 
1286 	mutex_lock(&intel_pstate_driver_lock);
1287 
1288 	if (!intel_pstate_driver) {
1289 		mutex_unlock(&intel_pstate_driver_lock);
1290 		return -EAGAIN;
1291 	}
1292 
1293 	update_turbo_state();
1294 	if (global.turbo_disabled)
1295 		ret = sprintf(buf, "%u\n", global.turbo_disabled);
1296 	else
1297 		ret = sprintf(buf, "%u\n", global.no_turbo);
1298 
1299 	mutex_unlock(&intel_pstate_driver_lock);
1300 
1301 	return ret;
1302 }
1303 
1304 static ssize_t store_no_turbo(struct kobject *a, struct kobj_attribute *b,
1305 			      const char *buf, size_t count)
1306 {
1307 	unsigned int input;
1308 	int ret;
1309 
1310 	ret = sscanf(buf, "%u", &input);
1311 	if (ret != 1)
1312 		return -EINVAL;
1313 
1314 	mutex_lock(&intel_pstate_driver_lock);
1315 
1316 	if (!intel_pstate_driver) {
1317 		mutex_unlock(&intel_pstate_driver_lock);
1318 		return -EAGAIN;
1319 	}
1320 
1321 	mutex_lock(&intel_pstate_limits_lock);
1322 
1323 	update_turbo_state();
1324 	if (global.turbo_disabled) {
1325 		pr_notice_once("Turbo disabled by BIOS or unavailable on processor\n");
1326 		mutex_unlock(&intel_pstate_limits_lock);
1327 		mutex_unlock(&intel_pstate_driver_lock);
1328 		return -EPERM;
1329 	}
1330 
1331 	global.no_turbo = clamp_t(int, input, 0, 1);
1332 
1333 	if (global.no_turbo) {
1334 		struct cpudata *cpu = all_cpu_data[0];
1335 		int pct = cpu->pstate.max_pstate * 100 / cpu->pstate.turbo_pstate;
1336 
1337 		/* Squash the global minimum into the permitted range. */
1338 		if (global.min_perf_pct > pct)
1339 			global.min_perf_pct = pct;
1340 	}
1341 
1342 	mutex_unlock(&intel_pstate_limits_lock);
1343 
1344 	intel_pstate_update_policies();
1345 	arch_set_max_freq_ratio(global.no_turbo);
1346 
1347 	mutex_unlock(&intel_pstate_driver_lock);
1348 
1349 	return count;
1350 }
1351 
1352 static void update_qos_request(enum freq_qos_req_type type)
1353 {
1354 	struct freq_qos_request *req;
1355 	struct cpufreq_policy *policy;
1356 	int i;
1357 
1358 	for_each_possible_cpu(i) {
1359 		struct cpudata *cpu = all_cpu_data[i];
1360 		unsigned int freq, perf_pct;
1361 
1362 		policy = cpufreq_cpu_get(i);
1363 		if (!policy)
1364 			continue;
1365 
1366 		req = policy->driver_data;
1367 		cpufreq_cpu_put(policy);
1368 
1369 		if (!req)
1370 			continue;
1371 
1372 		if (hwp_active)
1373 			intel_pstate_get_hwp_cap(cpu);
1374 
1375 		if (type == FREQ_QOS_MIN) {
1376 			perf_pct = global.min_perf_pct;
1377 		} else {
1378 			req++;
1379 			perf_pct = global.max_perf_pct;
1380 		}
1381 
1382 		freq = DIV_ROUND_UP(cpu->pstate.turbo_freq * perf_pct, 100);
1383 
1384 		if (freq_qos_update_request(req, freq) < 0)
1385 			pr_warn("Failed to update freq constraint: CPU%d\n", i);
1386 	}
1387 }
1388 
1389 static ssize_t store_max_perf_pct(struct kobject *a, struct kobj_attribute *b,
1390 				  const char *buf, size_t count)
1391 {
1392 	unsigned int input;
1393 	int ret;
1394 
1395 	ret = sscanf(buf, "%u", &input);
1396 	if (ret != 1)
1397 		return -EINVAL;
1398 
1399 	mutex_lock(&intel_pstate_driver_lock);
1400 
1401 	if (!intel_pstate_driver) {
1402 		mutex_unlock(&intel_pstate_driver_lock);
1403 		return -EAGAIN;
1404 	}
1405 
1406 	mutex_lock(&intel_pstate_limits_lock);
1407 
1408 	global.max_perf_pct = clamp_t(int, input, global.min_perf_pct, 100);
1409 
1410 	mutex_unlock(&intel_pstate_limits_lock);
1411 
1412 	if (intel_pstate_driver == &intel_pstate)
1413 		intel_pstate_update_policies();
1414 	else
1415 		update_qos_request(FREQ_QOS_MAX);
1416 
1417 	mutex_unlock(&intel_pstate_driver_lock);
1418 
1419 	return count;
1420 }
1421 
1422 static ssize_t store_min_perf_pct(struct kobject *a, struct kobj_attribute *b,
1423 				  const char *buf, size_t count)
1424 {
1425 	unsigned int input;
1426 	int ret;
1427 
1428 	ret = sscanf(buf, "%u", &input);
1429 	if (ret != 1)
1430 		return -EINVAL;
1431 
1432 	mutex_lock(&intel_pstate_driver_lock);
1433 
1434 	if (!intel_pstate_driver) {
1435 		mutex_unlock(&intel_pstate_driver_lock);
1436 		return -EAGAIN;
1437 	}
1438 
1439 	mutex_lock(&intel_pstate_limits_lock);
1440 
1441 	global.min_perf_pct = clamp_t(int, input,
1442 				      min_perf_pct_min(), global.max_perf_pct);
1443 
1444 	mutex_unlock(&intel_pstate_limits_lock);
1445 
1446 	if (intel_pstate_driver == &intel_pstate)
1447 		intel_pstate_update_policies();
1448 	else
1449 		update_qos_request(FREQ_QOS_MIN);
1450 
1451 	mutex_unlock(&intel_pstate_driver_lock);
1452 
1453 	return count;
1454 }
1455 
1456 static ssize_t show_hwp_dynamic_boost(struct kobject *kobj,
1457 				struct kobj_attribute *attr, char *buf)
1458 {
1459 	return sprintf(buf, "%u\n", hwp_boost);
1460 }
1461 
1462 static ssize_t store_hwp_dynamic_boost(struct kobject *a,
1463 				       struct kobj_attribute *b,
1464 				       const char *buf, size_t count)
1465 {
1466 	unsigned int input;
1467 	int ret;
1468 
1469 	ret = kstrtouint(buf, 10, &input);
1470 	if (ret)
1471 		return ret;
1472 
1473 	mutex_lock(&intel_pstate_driver_lock);
1474 	hwp_boost = !!input;
1475 	intel_pstate_update_policies();
1476 	mutex_unlock(&intel_pstate_driver_lock);
1477 
1478 	return count;
1479 }
1480 
1481 static ssize_t show_energy_efficiency(struct kobject *kobj, struct kobj_attribute *attr,
1482 				      char *buf)
1483 {
1484 	u64 power_ctl;
1485 	int enable;
1486 
1487 	rdmsrl(MSR_IA32_POWER_CTL, power_ctl);
1488 	enable = !!(power_ctl & BIT(MSR_IA32_POWER_CTL_BIT_EE));
1489 	return sprintf(buf, "%d\n", !enable);
1490 }
1491 
1492 static ssize_t store_energy_efficiency(struct kobject *a, struct kobj_attribute *b,
1493 				       const char *buf, size_t count)
1494 {
1495 	bool input;
1496 	int ret;
1497 
1498 	ret = kstrtobool(buf, &input);
1499 	if (ret)
1500 		return ret;
1501 
1502 	set_power_ctl_ee_state(input);
1503 
1504 	return count;
1505 }
1506 
1507 show_one(max_perf_pct, max_perf_pct);
1508 show_one(min_perf_pct, min_perf_pct);
1509 
1510 define_one_global_rw(status);
1511 define_one_global_rw(no_turbo);
1512 define_one_global_rw(max_perf_pct);
1513 define_one_global_rw(min_perf_pct);
1514 define_one_global_ro(turbo_pct);
1515 define_one_global_ro(num_pstates);
1516 define_one_global_rw(hwp_dynamic_boost);
1517 define_one_global_rw(energy_efficiency);
1518 
1519 static struct attribute *intel_pstate_attributes[] = {
1520 	&status.attr,
1521 	&no_turbo.attr,
1522 	NULL
1523 };
1524 
1525 static const struct attribute_group intel_pstate_attr_group = {
1526 	.attrs = intel_pstate_attributes,
1527 };
1528 
1529 static const struct x86_cpu_id intel_pstate_cpu_ee_disable_ids[];
1530 
1531 static struct kobject *intel_pstate_kobject;
1532 
1533 static void __init intel_pstate_sysfs_expose_params(void)
1534 {
1535 	struct device *dev_root = bus_get_dev_root(&cpu_subsys);
1536 	int rc;
1537 
1538 	if (dev_root) {
1539 		intel_pstate_kobject = kobject_create_and_add("intel_pstate", &dev_root->kobj);
1540 		put_device(dev_root);
1541 	}
1542 	if (WARN_ON(!intel_pstate_kobject))
1543 		return;
1544 
1545 	rc = sysfs_create_group(intel_pstate_kobject, &intel_pstate_attr_group);
1546 	if (WARN_ON(rc))
1547 		return;
1548 
1549 	if (!boot_cpu_has(X86_FEATURE_HYBRID_CPU)) {
1550 		rc = sysfs_create_file(intel_pstate_kobject, &turbo_pct.attr);
1551 		WARN_ON(rc);
1552 
1553 		rc = sysfs_create_file(intel_pstate_kobject, &num_pstates.attr);
1554 		WARN_ON(rc);
1555 	}
1556 
1557 	/*
1558 	 * If per cpu limits are enforced there are no global limits, so
1559 	 * return without creating max/min_perf_pct attributes
1560 	 */
1561 	if (per_cpu_limits)
1562 		return;
1563 
1564 	rc = sysfs_create_file(intel_pstate_kobject, &max_perf_pct.attr);
1565 	WARN_ON(rc);
1566 
1567 	rc = sysfs_create_file(intel_pstate_kobject, &min_perf_pct.attr);
1568 	WARN_ON(rc);
1569 
1570 	if (x86_match_cpu(intel_pstate_cpu_ee_disable_ids)) {
1571 		rc = sysfs_create_file(intel_pstate_kobject, &energy_efficiency.attr);
1572 		WARN_ON(rc);
1573 	}
1574 }
1575 
1576 static void __init intel_pstate_sysfs_remove(void)
1577 {
1578 	if (!intel_pstate_kobject)
1579 		return;
1580 
1581 	sysfs_remove_group(intel_pstate_kobject, &intel_pstate_attr_group);
1582 
1583 	if (!boot_cpu_has(X86_FEATURE_HYBRID_CPU)) {
1584 		sysfs_remove_file(intel_pstate_kobject, &num_pstates.attr);
1585 		sysfs_remove_file(intel_pstate_kobject, &turbo_pct.attr);
1586 	}
1587 
1588 	if (!per_cpu_limits) {
1589 		sysfs_remove_file(intel_pstate_kobject, &max_perf_pct.attr);
1590 		sysfs_remove_file(intel_pstate_kobject, &min_perf_pct.attr);
1591 
1592 		if (x86_match_cpu(intel_pstate_cpu_ee_disable_ids))
1593 			sysfs_remove_file(intel_pstate_kobject, &energy_efficiency.attr);
1594 	}
1595 
1596 	kobject_put(intel_pstate_kobject);
1597 }
1598 
1599 static void intel_pstate_sysfs_expose_hwp_dynamic_boost(void)
1600 {
1601 	int rc;
1602 
1603 	if (!hwp_active)
1604 		return;
1605 
1606 	rc = sysfs_create_file(intel_pstate_kobject, &hwp_dynamic_boost.attr);
1607 	WARN_ON_ONCE(rc);
1608 }
1609 
1610 static void intel_pstate_sysfs_hide_hwp_dynamic_boost(void)
1611 {
1612 	if (!hwp_active)
1613 		return;
1614 
1615 	sysfs_remove_file(intel_pstate_kobject, &hwp_dynamic_boost.attr);
1616 }
1617 
1618 /************************** sysfs end ************************/
1619 
1620 static void intel_pstate_notify_work(struct work_struct *work)
1621 {
1622 	struct cpudata *cpudata =
1623 		container_of(to_delayed_work(work), struct cpudata, hwp_notify_work);
1624 	struct cpufreq_policy *policy = cpufreq_cpu_acquire(cpudata->cpu);
1625 
1626 	if (policy) {
1627 		intel_pstate_get_hwp_cap(cpudata);
1628 		__intel_pstate_update_max_freq(cpudata, policy);
1629 
1630 		cpufreq_cpu_release(policy);
1631 	}
1632 
1633 	wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_STATUS, 0);
1634 }
1635 
1636 static DEFINE_SPINLOCK(hwp_notify_lock);
1637 static cpumask_t hwp_intr_enable_mask;
1638 
1639 void notify_hwp_interrupt(void)
1640 {
1641 	unsigned int this_cpu = smp_processor_id();
1642 	struct cpudata *cpudata;
1643 	unsigned long flags;
1644 	u64 value;
1645 
1646 	if (!READ_ONCE(hwp_active) || !boot_cpu_has(X86_FEATURE_HWP_NOTIFY))
1647 		return;
1648 
1649 	rdmsrl_safe(MSR_HWP_STATUS, &value);
1650 	if (!(value & 0x01))
1651 		return;
1652 
1653 	spin_lock_irqsave(&hwp_notify_lock, flags);
1654 
1655 	if (!cpumask_test_cpu(this_cpu, &hwp_intr_enable_mask))
1656 		goto ack_intr;
1657 
1658 	/*
1659 	 * Currently we never free all_cpu_data. And we can't reach here
1660 	 * without this allocated. But for safety for future changes, added
1661 	 * check.
1662 	 */
1663 	if (unlikely(!READ_ONCE(all_cpu_data)))
1664 		goto ack_intr;
1665 
1666 	/*
1667 	 * The free is done during cleanup, when cpufreq registry is failed.
1668 	 * We wouldn't be here if it fails on init or switch status. But for
1669 	 * future changes, added check.
1670 	 */
1671 	cpudata = READ_ONCE(all_cpu_data[this_cpu]);
1672 	if (unlikely(!cpudata))
1673 		goto ack_intr;
1674 
1675 	schedule_delayed_work(&cpudata->hwp_notify_work, msecs_to_jiffies(10));
1676 
1677 	spin_unlock_irqrestore(&hwp_notify_lock, flags);
1678 
1679 	return;
1680 
1681 ack_intr:
1682 	wrmsrl_safe(MSR_HWP_STATUS, 0);
1683 	spin_unlock_irqrestore(&hwp_notify_lock, flags);
1684 }
1685 
1686 static void intel_pstate_disable_hwp_interrupt(struct cpudata *cpudata)
1687 {
1688 	unsigned long flags;
1689 
1690 	if (!boot_cpu_has(X86_FEATURE_HWP_NOTIFY))
1691 		return;
1692 
1693 	/* wrmsrl_on_cpu has to be outside spinlock as this can result in IPC */
1694 	wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_INTERRUPT, 0x00);
1695 
1696 	spin_lock_irqsave(&hwp_notify_lock, flags);
1697 	if (cpumask_test_and_clear_cpu(cpudata->cpu, &hwp_intr_enable_mask))
1698 		cancel_delayed_work(&cpudata->hwp_notify_work);
1699 	spin_unlock_irqrestore(&hwp_notify_lock, flags);
1700 }
1701 
1702 static void intel_pstate_enable_hwp_interrupt(struct cpudata *cpudata)
1703 {
1704 	/* Enable HWP notification interrupt for guaranteed performance change */
1705 	if (boot_cpu_has(X86_FEATURE_HWP_NOTIFY)) {
1706 		unsigned long flags;
1707 
1708 		spin_lock_irqsave(&hwp_notify_lock, flags);
1709 		INIT_DELAYED_WORK(&cpudata->hwp_notify_work, intel_pstate_notify_work);
1710 		cpumask_set_cpu(cpudata->cpu, &hwp_intr_enable_mask);
1711 		spin_unlock_irqrestore(&hwp_notify_lock, flags);
1712 
1713 		/* wrmsrl_on_cpu has to be outside spinlock as this can result in IPC */
1714 		wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_INTERRUPT, 0x01);
1715 		wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_STATUS, 0);
1716 	}
1717 }
1718 
1719 static void intel_pstate_update_epp_defaults(struct cpudata *cpudata)
1720 {
1721 	cpudata->epp_default = intel_pstate_get_epp(cpudata, 0);
1722 
1723 	/*
1724 	 * If this CPU gen doesn't call for change in balance_perf
1725 	 * EPP return.
1726 	 */
1727 	if (epp_values[EPP_INDEX_BALANCE_PERFORMANCE] == HWP_EPP_BALANCE_PERFORMANCE)
1728 		return;
1729 
1730 	/*
1731 	 * If the EPP is set by firmware, which means that firmware enabled HWP
1732 	 * - Is equal or less than 0x80 (default balance_perf EPP)
1733 	 * - But less performance oriented than performance EPP
1734 	 *   then use this as new balance_perf EPP.
1735 	 */
1736 	if (hwp_forced && cpudata->epp_default <= HWP_EPP_BALANCE_PERFORMANCE &&
1737 	    cpudata->epp_default > HWP_EPP_PERFORMANCE) {
1738 		epp_values[EPP_INDEX_BALANCE_PERFORMANCE] = cpudata->epp_default;
1739 		return;
1740 	}
1741 
1742 	/*
1743 	 * Use hard coded value per gen to update the balance_perf
1744 	 * and default EPP.
1745 	 */
1746 	cpudata->epp_default = epp_values[EPP_INDEX_BALANCE_PERFORMANCE];
1747 	intel_pstate_set_epp(cpudata, cpudata->epp_default);
1748 }
1749 
1750 static void intel_pstate_hwp_enable(struct cpudata *cpudata)
1751 {
1752 	/* First disable HWP notification interrupt till we activate again */
1753 	if (boot_cpu_has(X86_FEATURE_HWP_NOTIFY))
1754 		wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_INTERRUPT, 0x00);
1755 
1756 	wrmsrl_on_cpu(cpudata->cpu, MSR_PM_ENABLE, 0x1);
1757 
1758 	intel_pstate_enable_hwp_interrupt(cpudata);
1759 
1760 	if (cpudata->epp_default >= 0)
1761 		return;
1762 
1763 	intel_pstate_update_epp_defaults(cpudata);
1764 }
1765 
1766 static int atom_get_min_pstate(int not_used)
1767 {
1768 	u64 value;
1769 
1770 	rdmsrl(MSR_ATOM_CORE_RATIOS, value);
1771 	return (value >> 8) & 0x7F;
1772 }
1773 
1774 static int atom_get_max_pstate(int not_used)
1775 {
1776 	u64 value;
1777 
1778 	rdmsrl(MSR_ATOM_CORE_RATIOS, value);
1779 	return (value >> 16) & 0x7F;
1780 }
1781 
1782 static int atom_get_turbo_pstate(int not_used)
1783 {
1784 	u64 value;
1785 
1786 	rdmsrl(MSR_ATOM_CORE_TURBO_RATIOS, value);
1787 	return value & 0x7F;
1788 }
1789 
1790 static u64 atom_get_val(struct cpudata *cpudata, int pstate)
1791 {
1792 	u64 val;
1793 	int32_t vid_fp;
1794 	u32 vid;
1795 
1796 	val = (u64)pstate << 8;
1797 	if (global.no_turbo && !global.turbo_disabled)
1798 		val |= (u64)1 << 32;
1799 
1800 	vid_fp = cpudata->vid.min + mul_fp(
1801 		int_tofp(pstate - cpudata->pstate.min_pstate),
1802 		cpudata->vid.ratio);
1803 
1804 	vid_fp = clamp_t(int32_t, vid_fp, cpudata->vid.min, cpudata->vid.max);
1805 	vid = ceiling_fp(vid_fp);
1806 
1807 	if (pstate > cpudata->pstate.max_pstate)
1808 		vid = cpudata->vid.turbo;
1809 
1810 	return val | vid;
1811 }
1812 
1813 static int silvermont_get_scaling(void)
1814 {
1815 	u64 value;
1816 	int i;
1817 	/* Defined in Table 35-6 from SDM (Sept 2015) */
1818 	static int silvermont_freq_table[] = {
1819 		83300, 100000, 133300, 116700, 80000};
1820 
1821 	rdmsrl(MSR_FSB_FREQ, value);
1822 	i = value & 0x7;
1823 	WARN_ON(i > 4);
1824 
1825 	return silvermont_freq_table[i];
1826 }
1827 
1828 static int airmont_get_scaling(void)
1829 {
1830 	u64 value;
1831 	int i;
1832 	/* Defined in Table 35-10 from SDM (Sept 2015) */
1833 	static int airmont_freq_table[] = {
1834 		83300, 100000, 133300, 116700, 80000,
1835 		93300, 90000, 88900, 87500};
1836 
1837 	rdmsrl(MSR_FSB_FREQ, value);
1838 	i = value & 0xF;
1839 	WARN_ON(i > 8);
1840 
1841 	return airmont_freq_table[i];
1842 }
1843 
1844 static void atom_get_vid(struct cpudata *cpudata)
1845 {
1846 	u64 value;
1847 
1848 	rdmsrl(MSR_ATOM_CORE_VIDS, value);
1849 	cpudata->vid.min = int_tofp((value >> 8) & 0x7f);
1850 	cpudata->vid.max = int_tofp((value >> 16) & 0x7f);
1851 	cpudata->vid.ratio = div_fp(
1852 		cpudata->vid.max - cpudata->vid.min,
1853 		int_tofp(cpudata->pstate.max_pstate -
1854 			cpudata->pstate.min_pstate));
1855 
1856 	rdmsrl(MSR_ATOM_CORE_TURBO_VIDS, value);
1857 	cpudata->vid.turbo = value & 0x7f;
1858 }
1859 
1860 static int core_get_min_pstate(int cpu)
1861 {
1862 	u64 value;
1863 
1864 	rdmsrl_on_cpu(cpu, MSR_PLATFORM_INFO, &value);
1865 	return (value >> 40) & 0xFF;
1866 }
1867 
1868 static int core_get_max_pstate_physical(int cpu)
1869 {
1870 	u64 value;
1871 
1872 	rdmsrl_on_cpu(cpu, MSR_PLATFORM_INFO, &value);
1873 	return (value >> 8) & 0xFF;
1874 }
1875 
1876 static int core_get_tdp_ratio(int cpu, u64 plat_info)
1877 {
1878 	/* Check how many TDP levels present */
1879 	if (plat_info & 0x600000000) {
1880 		u64 tdp_ctrl;
1881 		u64 tdp_ratio;
1882 		int tdp_msr;
1883 		int err;
1884 
1885 		/* Get the TDP level (0, 1, 2) to get ratios */
1886 		err = rdmsrl_safe_on_cpu(cpu, MSR_CONFIG_TDP_CONTROL, &tdp_ctrl);
1887 		if (err)
1888 			return err;
1889 
1890 		/* TDP MSR are continuous starting at 0x648 */
1891 		tdp_msr = MSR_CONFIG_TDP_NOMINAL + (tdp_ctrl & 0x03);
1892 		err = rdmsrl_safe_on_cpu(cpu, tdp_msr, &tdp_ratio);
1893 		if (err)
1894 			return err;
1895 
1896 		/* For level 1 and 2, bits[23:16] contain the ratio */
1897 		if (tdp_ctrl & 0x03)
1898 			tdp_ratio >>= 16;
1899 
1900 		tdp_ratio &= 0xff; /* ratios are only 8 bits long */
1901 		pr_debug("tdp_ratio %x\n", (int)tdp_ratio);
1902 
1903 		return (int)tdp_ratio;
1904 	}
1905 
1906 	return -ENXIO;
1907 }
1908 
1909 static int core_get_max_pstate(int cpu)
1910 {
1911 	u64 tar;
1912 	u64 plat_info;
1913 	int max_pstate;
1914 	int tdp_ratio;
1915 	int err;
1916 
1917 	rdmsrl_on_cpu(cpu, MSR_PLATFORM_INFO, &plat_info);
1918 	max_pstate = (plat_info >> 8) & 0xFF;
1919 
1920 	tdp_ratio = core_get_tdp_ratio(cpu, plat_info);
1921 	if (tdp_ratio <= 0)
1922 		return max_pstate;
1923 
1924 	if (hwp_active) {
1925 		/* Turbo activation ratio is not used on HWP platforms */
1926 		return tdp_ratio;
1927 	}
1928 
1929 	err = rdmsrl_safe_on_cpu(cpu, MSR_TURBO_ACTIVATION_RATIO, &tar);
1930 	if (!err) {
1931 		int tar_levels;
1932 
1933 		/* Do some sanity checking for safety */
1934 		tar_levels = tar & 0xff;
1935 		if (tdp_ratio - 1 == tar_levels) {
1936 			max_pstate = tar_levels;
1937 			pr_debug("max_pstate=TAC %x\n", max_pstate);
1938 		}
1939 	}
1940 
1941 	return max_pstate;
1942 }
1943 
1944 static int core_get_turbo_pstate(int cpu)
1945 {
1946 	u64 value;
1947 	int nont, ret;
1948 
1949 	rdmsrl_on_cpu(cpu, MSR_TURBO_RATIO_LIMIT, &value);
1950 	nont = core_get_max_pstate(cpu);
1951 	ret = (value) & 255;
1952 	if (ret <= nont)
1953 		ret = nont;
1954 	return ret;
1955 }
1956 
1957 static u64 core_get_val(struct cpudata *cpudata, int pstate)
1958 {
1959 	u64 val;
1960 
1961 	val = (u64)pstate << 8;
1962 	if (global.no_turbo && !global.turbo_disabled)
1963 		val |= (u64)1 << 32;
1964 
1965 	return val;
1966 }
1967 
1968 static int knl_get_aperf_mperf_shift(void)
1969 {
1970 	return 10;
1971 }
1972 
1973 static int knl_get_turbo_pstate(int cpu)
1974 {
1975 	u64 value;
1976 	int nont, ret;
1977 
1978 	rdmsrl_on_cpu(cpu, MSR_TURBO_RATIO_LIMIT, &value);
1979 	nont = core_get_max_pstate(cpu);
1980 	ret = (((value) >> 8) & 0xFF);
1981 	if (ret <= nont)
1982 		ret = nont;
1983 	return ret;
1984 }
1985 
1986 static void hybrid_get_type(void *data)
1987 {
1988 	u8 *cpu_type = data;
1989 
1990 	*cpu_type = get_this_hybrid_cpu_type();
1991 }
1992 
1993 static int hwp_get_cpu_scaling(int cpu)
1994 {
1995 	u8 cpu_type = 0;
1996 
1997 	smp_call_function_single(cpu, hybrid_get_type, &cpu_type, 1);
1998 	/* P-cores have a smaller perf level-to-freqency scaling factor. */
1999 	if (cpu_type == 0x40)
2000 		return HYBRID_SCALING_FACTOR;
2001 
2002 	/* Use default core scaling for E-cores */
2003 	if (cpu_type == 0x20)
2004 		return core_get_scaling();
2005 
2006 	/*
2007 	 * If reached here, this system is either non-hybrid (like Tiger
2008 	 * Lake) or hybrid-capable (like Alder Lake or Raptor Lake) with
2009 	 * no E cores (in which case CPUID for hybrid support is 0).
2010 	 *
2011 	 * The CPPC nominal_frequency field is 0 for non-hybrid systems,
2012 	 * so the default core scaling will be used for them.
2013 	 */
2014 	return intel_pstate_cppc_get_scaling(cpu);
2015 }
2016 
2017 static void intel_pstate_set_pstate(struct cpudata *cpu, int pstate)
2018 {
2019 	trace_cpu_frequency(pstate * cpu->pstate.scaling, cpu->cpu);
2020 	cpu->pstate.current_pstate = pstate;
2021 	/*
2022 	 * Generally, there is no guarantee that this code will always run on
2023 	 * the CPU being updated, so force the register update to run on the
2024 	 * right CPU.
2025 	 */
2026 	wrmsrl_on_cpu(cpu->cpu, MSR_IA32_PERF_CTL,
2027 		      pstate_funcs.get_val(cpu, pstate));
2028 }
2029 
2030 static void intel_pstate_set_min_pstate(struct cpudata *cpu)
2031 {
2032 	intel_pstate_set_pstate(cpu, cpu->pstate.min_pstate);
2033 }
2034 
2035 static void intel_pstate_max_within_limits(struct cpudata *cpu)
2036 {
2037 	int pstate = max(cpu->pstate.min_pstate, cpu->max_perf_ratio);
2038 
2039 	update_turbo_state();
2040 	intel_pstate_set_pstate(cpu, pstate);
2041 }
2042 
2043 static void intel_pstate_get_cpu_pstates(struct cpudata *cpu)
2044 {
2045 	int perf_ctl_max_phys = pstate_funcs.get_max_physical(cpu->cpu);
2046 	int perf_ctl_scaling = pstate_funcs.get_scaling();
2047 
2048 	cpu->pstate.min_pstate = pstate_funcs.get_min(cpu->cpu);
2049 	cpu->pstate.max_pstate_physical = perf_ctl_max_phys;
2050 	cpu->pstate.perf_ctl_scaling = perf_ctl_scaling;
2051 
2052 	if (hwp_active && !hwp_mode_bdw) {
2053 		__intel_pstate_get_hwp_cap(cpu);
2054 
2055 		if (pstate_funcs.get_cpu_scaling) {
2056 			cpu->pstate.scaling = pstate_funcs.get_cpu_scaling(cpu->cpu);
2057 			if (cpu->pstate.scaling != perf_ctl_scaling)
2058 				intel_pstate_hybrid_hwp_adjust(cpu);
2059 		} else {
2060 			cpu->pstate.scaling = perf_ctl_scaling;
2061 		}
2062 	} else {
2063 		cpu->pstate.scaling = perf_ctl_scaling;
2064 		cpu->pstate.max_pstate = pstate_funcs.get_max(cpu->cpu);
2065 		cpu->pstate.turbo_pstate = pstate_funcs.get_turbo(cpu->cpu);
2066 	}
2067 
2068 	if (cpu->pstate.scaling == perf_ctl_scaling) {
2069 		cpu->pstate.min_freq = cpu->pstate.min_pstate * perf_ctl_scaling;
2070 		cpu->pstate.max_freq = cpu->pstate.max_pstate * perf_ctl_scaling;
2071 		cpu->pstate.turbo_freq = cpu->pstate.turbo_pstate * perf_ctl_scaling;
2072 	}
2073 
2074 	if (pstate_funcs.get_aperf_mperf_shift)
2075 		cpu->aperf_mperf_shift = pstate_funcs.get_aperf_mperf_shift();
2076 
2077 	if (pstate_funcs.get_vid)
2078 		pstate_funcs.get_vid(cpu);
2079 
2080 	intel_pstate_set_min_pstate(cpu);
2081 }
2082 
2083 /*
2084  * Long hold time will keep high perf limits for long time,
2085  * which negatively impacts perf/watt for some workloads,
2086  * like specpower. 3ms is based on experiements on some
2087  * workoads.
2088  */
2089 static int hwp_boost_hold_time_ns = 3 * NSEC_PER_MSEC;
2090 
2091 static inline void intel_pstate_hwp_boost_up(struct cpudata *cpu)
2092 {
2093 	u64 hwp_req = READ_ONCE(cpu->hwp_req_cached);
2094 	u64 hwp_cap = READ_ONCE(cpu->hwp_cap_cached);
2095 	u32 max_limit = (hwp_req & 0xff00) >> 8;
2096 	u32 min_limit = (hwp_req & 0xff);
2097 	u32 boost_level1;
2098 
2099 	/*
2100 	 * Cases to consider (User changes via sysfs or boot time):
2101 	 * If, P0 (Turbo max) = P1 (Guaranteed max) = min:
2102 	 *	No boost, return.
2103 	 * If, P0 (Turbo max) > P1 (Guaranteed max) = min:
2104 	 *     Should result in one level boost only for P0.
2105 	 * If, P0 (Turbo max) = P1 (Guaranteed max) > min:
2106 	 *     Should result in two level boost:
2107 	 *         (min + p1)/2 and P1.
2108 	 * If, P0 (Turbo max) > P1 (Guaranteed max) > min:
2109 	 *     Should result in three level boost:
2110 	 *        (min + p1)/2, P1 and P0.
2111 	 */
2112 
2113 	/* If max and min are equal or already at max, nothing to boost */
2114 	if (max_limit == min_limit || cpu->hwp_boost_min >= max_limit)
2115 		return;
2116 
2117 	if (!cpu->hwp_boost_min)
2118 		cpu->hwp_boost_min = min_limit;
2119 
2120 	/* level at half way mark between min and guranteed */
2121 	boost_level1 = (HWP_GUARANTEED_PERF(hwp_cap) + min_limit) >> 1;
2122 
2123 	if (cpu->hwp_boost_min < boost_level1)
2124 		cpu->hwp_boost_min = boost_level1;
2125 	else if (cpu->hwp_boost_min < HWP_GUARANTEED_PERF(hwp_cap))
2126 		cpu->hwp_boost_min = HWP_GUARANTEED_PERF(hwp_cap);
2127 	else if (cpu->hwp_boost_min == HWP_GUARANTEED_PERF(hwp_cap) &&
2128 		 max_limit != HWP_GUARANTEED_PERF(hwp_cap))
2129 		cpu->hwp_boost_min = max_limit;
2130 	else
2131 		return;
2132 
2133 	hwp_req = (hwp_req & ~GENMASK_ULL(7, 0)) | cpu->hwp_boost_min;
2134 	wrmsrl(MSR_HWP_REQUEST, hwp_req);
2135 	cpu->last_update = cpu->sample.time;
2136 }
2137 
2138 static inline void intel_pstate_hwp_boost_down(struct cpudata *cpu)
2139 {
2140 	if (cpu->hwp_boost_min) {
2141 		bool expired;
2142 
2143 		/* Check if we are idle for hold time to boost down */
2144 		expired = time_after64(cpu->sample.time, cpu->last_update +
2145 				       hwp_boost_hold_time_ns);
2146 		if (expired) {
2147 			wrmsrl(MSR_HWP_REQUEST, cpu->hwp_req_cached);
2148 			cpu->hwp_boost_min = 0;
2149 		}
2150 	}
2151 	cpu->last_update = cpu->sample.time;
2152 }
2153 
2154 static inline void intel_pstate_update_util_hwp_local(struct cpudata *cpu,
2155 						      u64 time)
2156 {
2157 	cpu->sample.time = time;
2158 
2159 	if (cpu->sched_flags & SCHED_CPUFREQ_IOWAIT) {
2160 		bool do_io = false;
2161 
2162 		cpu->sched_flags = 0;
2163 		/*
2164 		 * Set iowait_boost flag and update time. Since IO WAIT flag
2165 		 * is set all the time, we can't just conclude that there is
2166 		 * some IO bound activity is scheduled on this CPU with just
2167 		 * one occurrence. If we receive at least two in two
2168 		 * consecutive ticks, then we treat as boost candidate.
2169 		 */
2170 		if (time_before64(time, cpu->last_io_update + 2 * TICK_NSEC))
2171 			do_io = true;
2172 
2173 		cpu->last_io_update = time;
2174 
2175 		if (do_io)
2176 			intel_pstate_hwp_boost_up(cpu);
2177 
2178 	} else {
2179 		intel_pstate_hwp_boost_down(cpu);
2180 	}
2181 }
2182 
2183 static inline void intel_pstate_update_util_hwp(struct update_util_data *data,
2184 						u64 time, unsigned int flags)
2185 {
2186 	struct cpudata *cpu = container_of(data, struct cpudata, update_util);
2187 
2188 	cpu->sched_flags |= flags;
2189 
2190 	if (smp_processor_id() == cpu->cpu)
2191 		intel_pstate_update_util_hwp_local(cpu, time);
2192 }
2193 
2194 static inline void intel_pstate_calc_avg_perf(struct cpudata *cpu)
2195 {
2196 	struct sample *sample = &cpu->sample;
2197 
2198 	sample->core_avg_perf = div_ext_fp(sample->aperf, sample->mperf);
2199 }
2200 
2201 static inline bool intel_pstate_sample(struct cpudata *cpu, u64 time)
2202 {
2203 	u64 aperf, mperf;
2204 	unsigned long flags;
2205 	u64 tsc;
2206 
2207 	local_irq_save(flags);
2208 	rdmsrl(MSR_IA32_APERF, aperf);
2209 	rdmsrl(MSR_IA32_MPERF, mperf);
2210 	tsc = rdtsc();
2211 	if (cpu->prev_mperf == mperf || cpu->prev_tsc == tsc) {
2212 		local_irq_restore(flags);
2213 		return false;
2214 	}
2215 	local_irq_restore(flags);
2216 
2217 	cpu->last_sample_time = cpu->sample.time;
2218 	cpu->sample.time = time;
2219 	cpu->sample.aperf = aperf;
2220 	cpu->sample.mperf = mperf;
2221 	cpu->sample.tsc =  tsc;
2222 	cpu->sample.aperf -= cpu->prev_aperf;
2223 	cpu->sample.mperf -= cpu->prev_mperf;
2224 	cpu->sample.tsc -= cpu->prev_tsc;
2225 
2226 	cpu->prev_aperf = aperf;
2227 	cpu->prev_mperf = mperf;
2228 	cpu->prev_tsc = tsc;
2229 	/*
2230 	 * First time this function is invoked in a given cycle, all of the
2231 	 * previous sample data fields are equal to zero or stale and they must
2232 	 * be populated with meaningful numbers for things to work, so assume
2233 	 * that sample.time will always be reset before setting the utilization
2234 	 * update hook and make the caller skip the sample then.
2235 	 */
2236 	if (cpu->last_sample_time) {
2237 		intel_pstate_calc_avg_perf(cpu);
2238 		return true;
2239 	}
2240 	return false;
2241 }
2242 
2243 static inline int32_t get_avg_frequency(struct cpudata *cpu)
2244 {
2245 	return mul_ext_fp(cpu->sample.core_avg_perf, cpu_khz);
2246 }
2247 
2248 static inline int32_t get_avg_pstate(struct cpudata *cpu)
2249 {
2250 	return mul_ext_fp(cpu->pstate.max_pstate_physical,
2251 			  cpu->sample.core_avg_perf);
2252 }
2253 
2254 static inline int32_t get_target_pstate(struct cpudata *cpu)
2255 {
2256 	struct sample *sample = &cpu->sample;
2257 	int32_t busy_frac;
2258 	int target, avg_pstate;
2259 
2260 	busy_frac = div_fp(sample->mperf << cpu->aperf_mperf_shift,
2261 			   sample->tsc);
2262 
2263 	if (busy_frac < cpu->iowait_boost)
2264 		busy_frac = cpu->iowait_boost;
2265 
2266 	sample->busy_scaled = busy_frac * 100;
2267 
2268 	target = global.no_turbo || global.turbo_disabled ?
2269 			cpu->pstate.max_pstate : cpu->pstate.turbo_pstate;
2270 	target += target >> 2;
2271 	target = mul_fp(target, busy_frac);
2272 	if (target < cpu->pstate.min_pstate)
2273 		target = cpu->pstate.min_pstate;
2274 
2275 	/*
2276 	 * If the average P-state during the previous cycle was higher than the
2277 	 * current target, add 50% of the difference to the target to reduce
2278 	 * possible performance oscillations and offset possible performance
2279 	 * loss related to moving the workload from one CPU to another within
2280 	 * a package/module.
2281 	 */
2282 	avg_pstate = get_avg_pstate(cpu);
2283 	if (avg_pstate > target)
2284 		target += (avg_pstate - target) >> 1;
2285 
2286 	return target;
2287 }
2288 
2289 static int intel_pstate_prepare_request(struct cpudata *cpu, int pstate)
2290 {
2291 	int min_pstate = max(cpu->pstate.min_pstate, cpu->min_perf_ratio);
2292 	int max_pstate = max(min_pstate, cpu->max_perf_ratio);
2293 
2294 	return clamp_t(int, pstate, min_pstate, max_pstate);
2295 }
2296 
2297 static void intel_pstate_update_pstate(struct cpudata *cpu, int pstate)
2298 {
2299 	if (pstate == cpu->pstate.current_pstate)
2300 		return;
2301 
2302 	cpu->pstate.current_pstate = pstate;
2303 	wrmsrl(MSR_IA32_PERF_CTL, pstate_funcs.get_val(cpu, pstate));
2304 }
2305 
2306 static void intel_pstate_adjust_pstate(struct cpudata *cpu)
2307 {
2308 	int from = cpu->pstate.current_pstate;
2309 	struct sample *sample;
2310 	int target_pstate;
2311 
2312 	update_turbo_state();
2313 
2314 	target_pstate = get_target_pstate(cpu);
2315 	target_pstate = intel_pstate_prepare_request(cpu, target_pstate);
2316 	trace_cpu_frequency(target_pstate * cpu->pstate.scaling, cpu->cpu);
2317 	intel_pstate_update_pstate(cpu, target_pstate);
2318 
2319 	sample = &cpu->sample;
2320 	trace_pstate_sample(mul_ext_fp(100, sample->core_avg_perf),
2321 		fp_toint(sample->busy_scaled),
2322 		from,
2323 		cpu->pstate.current_pstate,
2324 		sample->mperf,
2325 		sample->aperf,
2326 		sample->tsc,
2327 		get_avg_frequency(cpu),
2328 		fp_toint(cpu->iowait_boost * 100));
2329 }
2330 
2331 static void intel_pstate_update_util(struct update_util_data *data, u64 time,
2332 				     unsigned int flags)
2333 {
2334 	struct cpudata *cpu = container_of(data, struct cpudata, update_util);
2335 	u64 delta_ns;
2336 
2337 	/* Don't allow remote callbacks */
2338 	if (smp_processor_id() != cpu->cpu)
2339 		return;
2340 
2341 	delta_ns = time - cpu->last_update;
2342 	if (flags & SCHED_CPUFREQ_IOWAIT) {
2343 		/* Start over if the CPU may have been idle. */
2344 		if (delta_ns > TICK_NSEC) {
2345 			cpu->iowait_boost = ONE_EIGHTH_FP;
2346 		} else if (cpu->iowait_boost >= ONE_EIGHTH_FP) {
2347 			cpu->iowait_boost <<= 1;
2348 			if (cpu->iowait_boost > int_tofp(1))
2349 				cpu->iowait_boost = int_tofp(1);
2350 		} else {
2351 			cpu->iowait_boost = ONE_EIGHTH_FP;
2352 		}
2353 	} else if (cpu->iowait_boost) {
2354 		/* Clear iowait_boost if the CPU may have been idle. */
2355 		if (delta_ns > TICK_NSEC)
2356 			cpu->iowait_boost = 0;
2357 		else
2358 			cpu->iowait_boost >>= 1;
2359 	}
2360 	cpu->last_update = time;
2361 	delta_ns = time - cpu->sample.time;
2362 	if ((s64)delta_ns < INTEL_PSTATE_SAMPLING_INTERVAL)
2363 		return;
2364 
2365 	if (intel_pstate_sample(cpu, time))
2366 		intel_pstate_adjust_pstate(cpu);
2367 }
2368 
2369 static struct pstate_funcs core_funcs = {
2370 	.get_max = core_get_max_pstate,
2371 	.get_max_physical = core_get_max_pstate_physical,
2372 	.get_min = core_get_min_pstate,
2373 	.get_turbo = core_get_turbo_pstate,
2374 	.get_scaling = core_get_scaling,
2375 	.get_val = core_get_val,
2376 };
2377 
2378 static const struct pstate_funcs silvermont_funcs = {
2379 	.get_max = atom_get_max_pstate,
2380 	.get_max_physical = atom_get_max_pstate,
2381 	.get_min = atom_get_min_pstate,
2382 	.get_turbo = atom_get_turbo_pstate,
2383 	.get_val = atom_get_val,
2384 	.get_scaling = silvermont_get_scaling,
2385 	.get_vid = atom_get_vid,
2386 };
2387 
2388 static const struct pstate_funcs airmont_funcs = {
2389 	.get_max = atom_get_max_pstate,
2390 	.get_max_physical = atom_get_max_pstate,
2391 	.get_min = atom_get_min_pstate,
2392 	.get_turbo = atom_get_turbo_pstate,
2393 	.get_val = atom_get_val,
2394 	.get_scaling = airmont_get_scaling,
2395 	.get_vid = atom_get_vid,
2396 };
2397 
2398 static const struct pstate_funcs knl_funcs = {
2399 	.get_max = core_get_max_pstate,
2400 	.get_max_physical = core_get_max_pstate_physical,
2401 	.get_min = core_get_min_pstate,
2402 	.get_turbo = knl_get_turbo_pstate,
2403 	.get_aperf_mperf_shift = knl_get_aperf_mperf_shift,
2404 	.get_scaling = core_get_scaling,
2405 	.get_val = core_get_val,
2406 };
2407 
2408 #define X86_MATCH(model, policy)					 \
2409 	X86_MATCH_VENDOR_FAM_MODEL_FEATURE(INTEL, 6, INTEL_FAM6_##model, \
2410 					   X86_FEATURE_APERFMPERF, &policy)
2411 
2412 static const struct x86_cpu_id intel_pstate_cpu_ids[] = {
2413 	X86_MATCH(SANDYBRIDGE,		core_funcs),
2414 	X86_MATCH(SANDYBRIDGE_X,	core_funcs),
2415 	X86_MATCH(ATOM_SILVERMONT,	silvermont_funcs),
2416 	X86_MATCH(IVYBRIDGE,		core_funcs),
2417 	X86_MATCH(HASWELL,		core_funcs),
2418 	X86_MATCH(BROADWELL,		core_funcs),
2419 	X86_MATCH(IVYBRIDGE_X,		core_funcs),
2420 	X86_MATCH(HASWELL_X,		core_funcs),
2421 	X86_MATCH(HASWELL_L,		core_funcs),
2422 	X86_MATCH(HASWELL_G,		core_funcs),
2423 	X86_MATCH(BROADWELL_G,		core_funcs),
2424 	X86_MATCH(ATOM_AIRMONT,		airmont_funcs),
2425 	X86_MATCH(SKYLAKE_L,		core_funcs),
2426 	X86_MATCH(BROADWELL_X,		core_funcs),
2427 	X86_MATCH(SKYLAKE,		core_funcs),
2428 	X86_MATCH(BROADWELL_D,		core_funcs),
2429 	X86_MATCH(XEON_PHI_KNL,		knl_funcs),
2430 	X86_MATCH(XEON_PHI_KNM,		knl_funcs),
2431 	X86_MATCH(ATOM_GOLDMONT,	core_funcs),
2432 	X86_MATCH(ATOM_GOLDMONT_PLUS,	core_funcs),
2433 	X86_MATCH(SKYLAKE_X,		core_funcs),
2434 	X86_MATCH(COMETLAKE,		core_funcs),
2435 	X86_MATCH(ICELAKE_X,		core_funcs),
2436 	X86_MATCH(TIGERLAKE,		core_funcs),
2437 	X86_MATCH(SAPPHIRERAPIDS_X,	core_funcs),
2438 	{}
2439 };
2440 MODULE_DEVICE_TABLE(x86cpu, intel_pstate_cpu_ids);
2441 
2442 static const struct x86_cpu_id intel_pstate_cpu_oob_ids[] __initconst = {
2443 	X86_MATCH(BROADWELL_D,		core_funcs),
2444 	X86_MATCH(BROADWELL_X,		core_funcs),
2445 	X86_MATCH(SKYLAKE_X,		core_funcs),
2446 	X86_MATCH(ICELAKE_X,		core_funcs),
2447 	X86_MATCH(SAPPHIRERAPIDS_X,	core_funcs),
2448 	{}
2449 };
2450 
2451 static const struct x86_cpu_id intel_pstate_cpu_ee_disable_ids[] = {
2452 	X86_MATCH(KABYLAKE,		core_funcs),
2453 	{}
2454 };
2455 
2456 static int intel_pstate_init_cpu(unsigned int cpunum)
2457 {
2458 	struct cpudata *cpu;
2459 
2460 	cpu = all_cpu_data[cpunum];
2461 
2462 	if (!cpu) {
2463 		cpu = kzalloc(sizeof(*cpu), GFP_KERNEL);
2464 		if (!cpu)
2465 			return -ENOMEM;
2466 
2467 		WRITE_ONCE(all_cpu_data[cpunum], cpu);
2468 
2469 		cpu->cpu = cpunum;
2470 
2471 		cpu->epp_default = -EINVAL;
2472 
2473 		if (hwp_active) {
2474 			intel_pstate_hwp_enable(cpu);
2475 
2476 			if (intel_pstate_acpi_pm_profile_server())
2477 				hwp_boost = true;
2478 		}
2479 	} else if (hwp_active) {
2480 		/*
2481 		 * Re-enable HWP in case this happens after a resume from ACPI
2482 		 * S3 if the CPU was offline during the whole system/resume
2483 		 * cycle.
2484 		 */
2485 		intel_pstate_hwp_reenable(cpu);
2486 	}
2487 
2488 	cpu->epp_powersave = -EINVAL;
2489 	cpu->epp_policy = 0;
2490 
2491 	intel_pstate_get_cpu_pstates(cpu);
2492 
2493 	pr_debug("controlling: cpu %d\n", cpunum);
2494 
2495 	return 0;
2496 }
2497 
2498 static void intel_pstate_set_update_util_hook(unsigned int cpu_num)
2499 {
2500 	struct cpudata *cpu = all_cpu_data[cpu_num];
2501 
2502 	if (hwp_active && !hwp_boost)
2503 		return;
2504 
2505 	if (cpu->update_util_set)
2506 		return;
2507 
2508 	/* Prevent intel_pstate_update_util() from using stale data. */
2509 	cpu->sample.time = 0;
2510 	cpufreq_add_update_util_hook(cpu_num, &cpu->update_util,
2511 				     (hwp_active ?
2512 				      intel_pstate_update_util_hwp :
2513 				      intel_pstate_update_util));
2514 	cpu->update_util_set = true;
2515 }
2516 
2517 static void intel_pstate_clear_update_util_hook(unsigned int cpu)
2518 {
2519 	struct cpudata *cpu_data = all_cpu_data[cpu];
2520 
2521 	if (!cpu_data->update_util_set)
2522 		return;
2523 
2524 	cpufreq_remove_update_util_hook(cpu);
2525 	cpu_data->update_util_set = false;
2526 	synchronize_rcu();
2527 }
2528 
2529 static int intel_pstate_get_max_freq(struct cpudata *cpu)
2530 {
2531 	return global.turbo_disabled || global.no_turbo ?
2532 			cpu->pstate.max_freq : cpu->pstate.turbo_freq;
2533 }
2534 
2535 static void intel_pstate_update_perf_limits(struct cpudata *cpu,
2536 					    unsigned int policy_min,
2537 					    unsigned int policy_max)
2538 {
2539 	int perf_ctl_scaling = cpu->pstate.perf_ctl_scaling;
2540 	int32_t max_policy_perf, min_policy_perf;
2541 
2542 	max_policy_perf = policy_max / perf_ctl_scaling;
2543 	if (policy_max == policy_min) {
2544 		min_policy_perf = max_policy_perf;
2545 	} else {
2546 		min_policy_perf = policy_min / perf_ctl_scaling;
2547 		min_policy_perf = clamp_t(int32_t, min_policy_perf,
2548 					  0, max_policy_perf);
2549 	}
2550 
2551 	/*
2552 	 * HWP needs some special consideration, because HWP_REQUEST uses
2553 	 * abstract values to represent performance rather than pure ratios.
2554 	 */
2555 	if (hwp_active && cpu->pstate.scaling != perf_ctl_scaling) {
2556 		int freq;
2557 
2558 		freq = max_policy_perf * perf_ctl_scaling;
2559 		max_policy_perf = intel_pstate_freq_to_hwp(cpu, freq);
2560 		freq = min_policy_perf * perf_ctl_scaling;
2561 		min_policy_perf = intel_pstate_freq_to_hwp(cpu, freq);
2562 	}
2563 
2564 	pr_debug("cpu:%d min_policy_perf:%d max_policy_perf:%d\n",
2565 		 cpu->cpu, min_policy_perf, max_policy_perf);
2566 
2567 	/* Normalize user input to [min_perf, max_perf] */
2568 	if (per_cpu_limits) {
2569 		cpu->min_perf_ratio = min_policy_perf;
2570 		cpu->max_perf_ratio = max_policy_perf;
2571 	} else {
2572 		int turbo_max = cpu->pstate.turbo_pstate;
2573 		int32_t global_min, global_max;
2574 
2575 		/* Global limits are in percent of the maximum turbo P-state. */
2576 		global_max = DIV_ROUND_UP(turbo_max * global.max_perf_pct, 100);
2577 		global_min = DIV_ROUND_UP(turbo_max * global.min_perf_pct, 100);
2578 		global_min = clamp_t(int32_t, global_min, 0, global_max);
2579 
2580 		pr_debug("cpu:%d global_min:%d global_max:%d\n", cpu->cpu,
2581 			 global_min, global_max);
2582 
2583 		cpu->min_perf_ratio = max(min_policy_perf, global_min);
2584 		cpu->min_perf_ratio = min(cpu->min_perf_ratio, max_policy_perf);
2585 		cpu->max_perf_ratio = min(max_policy_perf, global_max);
2586 		cpu->max_perf_ratio = max(min_policy_perf, cpu->max_perf_ratio);
2587 
2588 		/* Make sure min_perf <= max_perf */
2589 		cpu->min_perf_ratio = min(cpu->min_perf_ratio,
2590 					  cpu->max_perf_ratio);
2591 
2592 	}
2593 	pr_debug("cpu:%d max_perf_ratio:%d min_perf_ratio:%d\n", cpu->cpu,
2594 		 cpu->max_perf_ratio,
2595 		 cpu->min_perf_ratio);
2596 }
2597 
2598 static int intel_pstate_set_policy(struct cpufreq_policy *policy)
2599 {
2600 	struct cpudata *cpu;
2601 
2602 	if (!policy->cpuinfo.max_freq)
2603 		return -ENODEV;
2604 
2605 	pr_debug("set_policy cpuinfo.max %u policy->max %u\n",
2606 		 policy->cpuinfo.max_freq, policy->max);
2607 
2608 	cpu = all_cpu_data[policy->cpu];
2609 	cpu->policy = policy->policy;
2610 
2611 	mutex_lock(&intel_pstate_limits_lock);
2612 
2613 	intel_pstate_update_perf_limits(cpu, policy->min, policy->max);
2614 
2615 	if (cpu->policy == CPUFREQ_POLICY_PERFORMANCE) {
2616 		/*
2617 		 * NOHZ_FULL CPUs need this as the governor callback may not
2618 		 * be invoked on them.
2619 		 */
2620 		intel_pstate_clear_update_util_hook(policy->cpu);
2621 		intel_pstate_max_within_limits(cpu);
2622 	} else {
2623 		intel_pstate_set_update_util_hook(policy->cpu);
2624 	}
2625 
2626 	if (hwp_active) {
2627 		/*
2628 		 * When hwp_boost was active before and dynamically it
2629 		 * was turned off, in that case we need to clear the
2630 		 * update util hook.
2631 		 */
2632 		if (!hwp_boost)
2633 			intel_pstate_clear_update_util_hook(policy->cpu);
2634 		intel_pstate_hwp_set(policy->cpu);
2635 	}
2636 	/*
2637 	 * policy->cur is never updated with the intel_pstate driver, but it
2638 	 * is used as a stale frequency value. So, keep it within limits.
2639 	 */
2640 	policy->cur = policy->min;
2641 
2642 	mutex_unlock(&intel_pstate_limits_lock);
2643 
2644 	return 0;
2645 }
2646 
2647 static void intel_pstate_adjust_policy_max(struct cpudata *cpu,
2648 					   struct cpufreq_policy_data *policy)
2649 {
2650 	if (!hwp_active &&
2651 	    cpu->pstate.max_pstate_physical > cpu->pstate.max_pstate &&
2652 	    policy->max < policy->cpuinfo.max_freq &&
2653 	    policy->max > cpu->pstate.max_freq) {
2654 		pr_debug("policy->max > max non turbo frequency\n");
2655 		policy->max = policy->cpuinfo.max_freq;
2656 	}
2657 }
2658 
2659 static void intel_pstate_verify_cpu_policy(struct cpudata *cpu,
2660 					   struct cpufreq_policy_data *policy)
2661 {
2662 	int max_freq;
2663 
2664 	update_turbo_state();
2665 	if (hwp_active) {
2666 		intel_pstate_get_hwp_cap(cpu);
2667 		max_freq = global.no_turbo || global.turbo_disabled ?
2668 				cpu->pstate.max_freq : cpu->pstate.turbo_freq;
2669 	} else {
2670 		max_freq = intel_pstate_get_max_freq(cpu);
2671 	}
2672 	cpufreq_verify_within_limits(policy, policy->cpuinfo.min_freq, max_freq);
2673 
2674 	intel_pstate_adjust_policy_max(cpu, policy);
2675 }
2676 
2677 static int intel_pstate_verify_policy(struct cpufreq_policy_data *policy)
2678 {
2679 	intel_pstate_verify_cpu_policy(all_cpu_data[policy->cpu], policy);
2680 
2681 	return 0;
2682 }
2683 
2684 static int intel_cpufreq_cpu_offline(struct cpufreq_policy *policy)
2685 {
2686 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2687 
2688 	pr_debug("CPU %d going offline\n", cpu->cpu);
2689 
2690 	if (cpu->suspended)
2691 		return 0;
2692 
2693 	/*
2694 	 * If the CPU is an SMT thread and it goes offline with the performance
2695 	 * settings different from the minimum, it will prevent its sibling
2696 	 * from getting to lower performance levels, so force the minimum
2697 	 * performance on CPU offline to prevent that from happening.
2698 	 */
2699 	if (hwp_active)
2700 		intel_pstate_hwp_offline(cpu);
2701 	else
2702 		intel_pstate_set_min_pstate(cpu);
2703 
2704 	intel_pstate_exit_perf_limits(policy);
2705 
2706 	return 0;
2707 }
2708 
2709 static int intel_pstate_cpu_online(struct cpufreq_policy *policy)
2710 {
2711 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2712 
2713 	pr_debug("CPU %d going online\n", cpu->cpu);
2714 
2715 	intel_pstate_init_acpi_perf_limits(policy);
2716 
2717 	if (hwp_active) {
2718 		/*
2719 		 * Re-enable HWP and clear the "suspended" flag to let "resume"
2720 		 * know that it need not do that.
2721 		 */
2722 		intel_pstate_hwp_reenable(cpu);
2723 		cpu->suspended = false;
2724 	}
2725 
2726 	return 0;
2727 }
2728 
2729 static int intel_pstate_cpu_offline(struct cpufreq_policy *policy)
2730 {
2731 	intel_pstate_clear_update_util_hook(policy->cpu);
2732 
2733 	return intel_cpufreq_cpu_offline(policy);
2734 }
2735 
2736 static int intel_pstate_cpu_exit(struct cpufreq_policy *policy)
2737 {
2738 	pr_debug("CPU %d exiting\n", policy->cpu);
2739 
2740 	policy->fast_switch_possible = false;
2741 
2742 	return 0;
2743 }
2744 
2745 static int __intel_pstate_cpu_init(struct cpufreq_policy *policy)
2746 {
2747 	struct cpudata *cpu;
2748 	int rc;
2749 
2750 	rc = intel_pstate_init_cpu(policy->cpu);
2751 	if (rc)
2752 		return rc;
2753 
2754 	cpu = all_cpu_data[policy->cpu];
2755 
2756 	cpu->max_perf_ratio = 0xFF;
2757 	cpu->min_perf_ratio = 0;
2758 
2759 	/* cpuinfo and default policy values */
2760 	policy->cpuinfo.min_freq = cpu->pstate.min_freq;
2761 	update_turbo_state();
2762 	global.turbo_disabled_mf = global.turbo_disabled;
2763 	policy->cpuinfo.max_freq = global.turbo_disabled ?
2764 			cpu->pstate.max_freq : cpu->pstate.turbo_freq;
2765 
2766 	policy->min = policy->cpuinfo.min_freq;
2767 	policy->max = policy->cpuinfo.max_freq;
2768 
2769 	intel_pstate_init_acpi_perf_limits(policy);
2770 
2771 	policy->fast_switch_possible = true;
2772 
2773 	return 0;
2774 }
2775 
2776 static int intel_pstate_cpu_init(struct cpufreq_policy *policy)
2777 {
2778 	int ret = __intel_pstate_cpu_init(policy);
2779 
2780 	if (ret)
2781 		return ret;
2782 
2783 	/*
2784 	 * Set the policy to powersave to provide a valid fallback value in case
2785 	 * the default cpufreq governor is neither powersave nor performance.
2786 	 */
2787 	policy->policy = CPUFREQ_POLICY_POWERSAVE;
2788 
2789 	if (hwp_active) {
2790 		struct cpudata *cpu = all_cpu_data[policy->cpu];
2791 
2792 		cpu->epp_cached = intel_pstate_get_epp(cpu, 0);
2793 	}
2794 
2795 	return 0;
2796 }
2797 
2798 static struct cpufreq_driver intel_pstate = {
2799 	.flags		= CPUFREQ_CONST_LOOPS,
2800 	.verify		= intel_pstate_verify_policy,
2801 	.setpolicy	= intel_pstate_set_policy,
2802 	.suspend	= intel_pstate_suspend,
2803 	.resume		= intel_pstate_resume,
2804 	.init		= intel_pstate_cpu_init,
2805 	.exit		= intel_pstate_cpu_exit,
2806 	.offline	= intel_pstate_cpu_offline,
2807 	.online		= intel_pstate_cpu_online,
2808 	.update_limits	= intel_pstate_update_limits,
2809 	.name		= "intel_pstate",
2810 };
2811 
2812 static int intel_cpufreq_verify_policy(struct cpufreq_policy_data *policy)
2813 {
2814 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2815 
2816 	intel_pstate_verify_cpu_policy(cpu, policy);
2817 	intel_pstate_update_perf_limits(cpu, policy->min, policy->max);
2818 
2819 	return 0;
2820 }
2821 
2822 /* Use of trace in passive mode:
2823  *
2824  * In passive mode the trace core_busy field (also known as the
2825  * performance field, and lablelled as such on the graphs; also known as
2826  * core_avg_perf) is not needed and so is re-assigned to indicate if the
2827  * driver call was via the normal or fast switch path. Various graphs
2828  * output from the intel_pstate_tracer.py utility that include core_busy
2829  * (or performance or core_avg_perf) have a fixed y-axis from 0 to 100%,
2830  * so we use 10 to indicate the normal path through the driver, and
2831  * 90 to indicate the fast switch path through the driver.
2832  * The scaled_busy field is not used, and is set to 0.
2833  */
2834 
2835 #define	INTEL_PSTATE_TRACE_TARGET 10
2836 #define	INTEL_PSTATE_TRACE_FAST_SWITCH 90
2837 
2838 static void intel_cpufreq_trace(struct cpudata *cpu, unsigned int trace_type, int old_pstate)
2839 {
2840 	struct sample *sample;
2841 
2842 	if (!trace_pstate_sample_enabled())
2843 		return;
2844 
2845 	if (!intel_pstate_sample(cpu, ktime_get()))
2846 		return;
2847 
2848 	sample = &cpu->sample;
2849 	trace_pstate_sample(trace_type,
2850 		0,
2851 		old_pstate,
2852 		cpu->pstate.current_pstate,
2853 		sample->mperf,
2854 		sample->aperf,
2855 		sample->tsc,
2856 		get_avg_frequency(cpu),
2857 		fp_toint(cpu->iowait_boost * 100));
2858 }
2859 
2860 static void intel_cpufreq_hwp_update(struct cpudata *cpu, u32 min, u32 max,
2861 				     u32 desired, bool fast_switch)
2862 {
2863 	u64 prev = READ_ONCE(cpu->hwp_req_cached), value = prev;
2864 
2865 	value &= ~HWP_MIN_PERF(~0L);
2866 	value |= HWP_MIN_PERF(min);
2867 
2868 	value &= ~HWP_MAX_PERF(~0L);
2869 	value |= HWP_MAX_PERF(max);
2870 
2871 	value &= ~HWP_DESIRED_PERF(~0L);
2872 	value |= HWP_DESIRED_PERF(desired);
2873 
2874 	if (value == prev)
2875 		return;
2876 
2877 	WRITE_ONCE(cpu->hwp_req_cached, value);
2878 	if (fast_switch)
2879 		wrmsrl(MSR_HWP_REQUEST, value);
2880 	else
2881 		wrmsrl_on_cpu(cpu->cpu, MSR_HWP_REQUEST, value);
2882 }
2883 
2884 static void intel_cpufreq_perf_ctl_update(struct cpudata *cpu,
2885 					  u32 target_pstate, bool fast_switch)
2886 {
2887 	if (fast_switch)
2888 		wrmsrl(MSR_IA32_PERF_CTL,
2889 		       pstate_funcs.get_val(cpu, target_pstate));
2890 	else
2891 		wrmsrl_on_cpu(cpu->cpu, MSR_IA32_PERF_CTL,
2892 			      pstate_funcs.get_val(cpu, target_pstate));
2893 }
2894 
2895 static int intel_cpufreq_update_pstate(struct cpufreq_policy *policy,
2896 				       int target_pstate, bool fast_switch)
2897 {
2898 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2899 	int old_pstate = cpu->pstate.current_pstate;
2900 
2901 	target_pstate = intel_pstate_prepare_request(cpu, target_pstate);
2902 	if (hwp_active) {
2903 		int max_pstate = policy->strict_target ?
2904 					target_pstate : cpu->max_perf_ratio;
2905 
2906 		intel_cpufreq_hwp_update(cpu, target_pstate, max_pstate, 0,
2907 					 fast_switch);
2908 	} else if (target_pstate != old_pstate) {
2909 		intel_cpufreq_perf_ctl_update(cpu, target_pstate, fast_switch);
2910 	}
2911 
2912 	cpu->pstate.current_pstate = target_pstate;
2913 
2914 	intel_cpufreq_trace(cpu, fast_switch ? INTEL_PSTATE_TRACE_FAST_SWITCH :
2915 			    INTEL_PSTATE_TRACE_TARGET, old_pstate);
2916 
2917 	return target_pstate;
2918 }
2919 
2920 static int intel_cpufreq_target(struct cpufreq_policy *policy,
2921 				unsigned int target_freq,
2922 				unsigned int relation)
2923 {
2924 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2925 	struct cpufreq_freqs freqs;
2926 	int target_pstate;
2927 
2928 	update_turbo_state();
2929 
2930 	freqs.old = policy->cur;
2931 	freqs.new = target_freq;
2932 
2933 	cpufreq_freq_transition_begin(policy, &freqs);
2934 
2935 	target_pstate = intel_pstate_freq_to_hwp_rel(cpu, freqs.new, relation);
2936 	target_pstate = intel_cpufreq_update_pstate(policy, target_pstate, false);
2937 
2938 	freqs.new = target_pstate * cpu->pstate.scaling;
2939 
2940 	cpufreq_freq_transition_end(policy, &freqs, false);
2941 
2942 	return 0;
2943 }
2944 
2945 static unsigned int intel_cpufreq_fast_switch(struct cpufreq_policy *policy,
2946 					      unsigned int target_freq)
2947 {
2948 	struct cpudata *cpu = all_cpu_data[policy->cpu];
2949 	int target_pstate;
2950 
2951 	update_turbo_state();
2952 
2953 	target_pstate = intel_pstate_freq_to_hwp(cpu, target_freq);
2954 
2955 	target_pstate = intel_cpufreq_update_pstate(policy, target_pstate, true);
2956 
2957 	return target_pstate * cpu->pstate.scaling;
2958 }
2959 
2960 static void intel_cpufreq_adjust_perf(unsigned int cpunum,
2961 				      unsigned long min_perf,
2962 				      unsigned long target_perf,
2963 				      unsigned long capacity)
2964 {
2965 	struct cpudata *cpu = all_cpu_data[cpunum];
2966 	u64 hwp_cap = READ_ONCE(cpu->hwp_cap_cached);
2967 	int old_pstate = cpu->pstate.current_pstate;
2968 	int cap_pstate, min_pstate, max_pstate, target_pstate;
2969 
2970 	update_turbo_state();
2971 	cap_pstate = global.turbo_disabled ? HWP_GUARANTEED_PERF(hwp_cap) :
2972 					     HWP_HIGHEST_PERF(hwp_cap);
2973 
2974 	/* Optimization: Avoid unnecessary divisions. */
2975 
2976 	target_pstate = cap_pstate;
2977 	if (target_perf < capacity)
2978 		target_pstate = DIV_ROUND_UP(cap_pstate * target_perf, capacity);
2979 
2980 	min_pstate = cap_pstate;
2981 	if (min_perf < capacity)
2982 		min_pstate = DIV_ROUND_UP(cap_pstate * min_perf, capacity);
2983 
2984 	if (min_pstate < cpu->pstate.min_pstate)
2985 		min_pstate = cpu->pstate.min_pstate;
2986 
2987 	if (min_pstate < cpu->min_perf_ratio)
2988 		min_pstate = cpu->min_perf_ratio;
2989 
2990 	max_pstate = min(cap_pstate, cpu->max_perf_ratio);
2991 	if (max_pstate < min_pstate)
2992 		max_pstate = min_pstate;
2993 
2994 	target_pstate = clamp_t(int, target_pstate, min_pstate, max_pstate);
2995 
2996 	intel_cpufreq_hwp_update(cpu, min_pstate, max_pstate, target_pstate, true);
2997 
2998 	cpu->pstate.current_pstate = target_pstate;
2999 	intel_cpufreq_trace(cpu, INTEL_PSTATE_TRACE_FAST_SWITCH, old_pstate);
3000 }
3001 
3002 static int intel_cpufreq_cpu_init(struct cpufreq_policy *policy)
3003 {
3004 	struct freq_qos_request *req;
3005 	struct cpudata *cpu;
3006 	struct device *dev;
3007 	int ret, freq;
3008 
3009 	dev = get_cpu_device(policy->cpu);
3010 	if (!dev)
3011 		return -ENODEV;
3012 
3013 	ret = __intel_pstate_cpu_init(policy);
3014 	if (ret)
3015 		return ret;
3016 
3017 	policy->cpuinfo.transition_latency = INTEL_CPUFREQ_TRANSITION_LATENCY;
3018 	/* This reflects the intel_pstate_get_cpu_pstates() setting. */
3019 	policy->cur = policy->cpuinfo.min_freq;
3020 
3021 	req = kcalloc(2, sizeof(*req), GFP_KERNEL);
3022 	if (!req) {
3023 		ret = -ENOMEM;
3024 		goto pstate_exit;
3025 	}
3026 
3027 	cpu = all_cpu_data[policy->cpu];
3028 
3029 	if (hwp_active) {
3030 		u64 value;
3031 
3032 		policy->transition_delay_us = INTEL_CPUFREQ_TRANSITION_DELAY_HWP;
3033 
3034 		intel_pstate_get_hwp_cap(cpu);
3035 
3036 		rdmsrl_on_cpu(cpu->cpu, MSR_HWP_REQUEST, &value);
3037 		WRITE_ONCE(cpu->hwp_req_cached, value);
3038 
3039 		cpu->epp_cached = intel_pstate_get_epp(cpu, value);
3040 	} else {
3041 		policy->transition_delay_us = INTEL_CPUFREQ_TRANSITION_DELAY;
3042 	}
3043 
3044 	freq = DIV_ROUND_UP(cpu->pstate.turbo_freq * global.min_perf_pct, 100);
3045 
3046 	ret = freq_qos_add_request(&policy->constraints, req, FREQ_QOS_MIN,
3047 				   freq);
3048 	if (ret < 0) {
3049 		dev_err(dev, "Failed to add min-freq constraint (%d)\n", ret);
3050 		goto free_req;
3051 	}
3052 
3053 	freq = DIV_ROUND_UP(cpu->pstate.turbo_freq * global.max_perf_pct, 100);
3054 
3055 	ret = freq_qos_add_request(&policy->constraints, req + 1, FREQ_QOS_MAX,
3056 				   freq);
3057 	if (ret < 0) {
3058 		dev_err(dev, "Failed to add max-freq constraint (%d)\n", ret);
3059 		goto remove_min_req;
3060 	}
3061 
3062 	policy->driver_data = req;
3063 
3064 	return 0;
3065 
3066 remove_min_req:
3067 	freq_qos_remove_request(req);
3068 free_req:
3069 	kfree(req);
3070 pstate_exit:
3071 	intel_pstate_exit_perf_limits(policy);
3072 
3073 	return ret;
3074 }
3075 
3076 static int intel_cpufreq_cpu_exit(struct cpufreq_policy *policy)
3077 {
3078 	struct freq_qos_request *req;
3079 
3080 	req = policy->driver_data;
3081 
3082 	freq_qos_remove_request(req + 1);
3083 	freq_qos_remove_request(req);
3084 	kfree(req);
3085 
3086 	return intel_pstate_cpu_exit(policy);
3087 }
3088 
3089 static int intel_cpufreq_suspend(struct cpufreq_policy *policy)
3090 {
3091 	intel_pstate_suspend(policy);
3092 
3093 	if (hwp_active) {
3094 		struct cpudata *cpu = all_cpu_data[policy->cpu];
3095 		u64 value = READ_ONCE(cpu->hwp_req_cached);
3096 
3097 		/*
3098 		 * Clear the desired perf field in MSR_HWP_REQUEST in case
3099 		 * intel_cpufreq_adjust_perf() is in use and the last value
3100 		 * written by it may not be suitable.
3101 		 */
3102 		value &= ~HWP_DESIRED_PERF(~0L);
3103 		wrmsrl_on_cpu(cpu->cpu, MSR_HWP_REQUEST, value);
3104 		WRITE_ONCE(cpu->hwp_req_cached, value);
3105 	}
3106 
3107 	return 0;
3108 }
3109 
3110 static struct cpufreq_driver intel_cpufreq = {
3111 	.flags		= CPUFREQ_CONST_LOOPS,
3112 	.verify		= intel_cpufreq_verify_policy,
3113 	.target		= intel_cpufreq_target,
3114 	.fast_switch	= intel_cpufreq_fast_switch,
3115 	.init		= intel_cpufreq_cpu_init,
3116 	.exit		= intel_cpufreq_cpu_exit,
3117 	.offline	= intel_cpufreq_cpu_offline,
3118 	.online		= intel_pstate_cpu_online,
3119 	.suspend	= intel_cpufreq_suspend,
3120 	.resume		= intel_pstate_resume,
3121 	.update_limits	= intel_pstate_update_limits,
3122 	.name		= "intel_cpufreq",
3123 };
3124 
3125 static struct cpufreq_driver *default_driver;
3126 
3127 static void intel_pstate_driver_cleanup(void)
3128 {
3129 	unsigned int cpu;
3130 
3131 	cpus_read_lock();
3132 	for_each_online_cpu(cpu) {
3133 		if (all_cpu_data[cpu]) {
3134 			if (intel_pstate_driver == &intel_pstate)
3135 				intel_pstate_clear_update_util_hook(cpu);
3136 
3137 			spin_lock(&hwp_notify_lock);
3138 			kfree(all_cpu_data[cpu]);
3139 			WRITE_ONCE(all_cpu_data[cpu], NULL);
3140 			spin_unlock(&hwp_notify_lock);
3141 		}
3142 	}
3143 	cpus_read_unlock();
3144 
3145 	intel_pstate_driver = NULL;
3146 }
3147 
3148 static int intel_pstate_register_driver(struct cpufreq_driver *driver)
3149 {
3150 	int ret;
3151 
3152 	if (driver == &intel_pstate)
3153 		intel_pstate_sysfs_expose_hwp_dynamic_boost();
3154 
3155 	memset(&global, 0, sizeof(global));
3156 	global.max_perf_pct = 100;
3157 
3158 	intel_pstate_driver = driver;
3159 	ret = cpufreq_register_driver(intel_pstate_driver);
3160 	if (ret) {
3161 		intel_pstate_driver_cleanup();
3162 		return ret;
3163 	}
3164 
3165 	global.min_perf_pct = min_perf_pct_min();
3166 
3167 	return 0;
3168 }
3169 
3170 static ssize_t intel_pstate_show_status(char *buf)
3171 {
3172 	if (!intel_pstate_driver)
3173 		return sprintf(buf, "off\n");
3174 
3175 	return sprintf(buf, "%s\n", intel_pstate_driver == &intel_pstate ?
3176 					"active" : "passive");
3177 }
3178 
3179 static int intel_pstate_update_status(const char *buf, size_t size)
3180 {
3181 	if (size == 3 && !strncmp(buf, "off", size)) {
3182 		if (!intel_pstate_driver)
3183 			return -EINVAL;
3184 
3185 		if (hwp_active)
3186 			return -EBUSY;
3187 
3188 		cpufreq_unregister_driver(intel_pstate_driver);
3189 		intel_pstate_driver_cleanup();
3190 		return 0;
3191 	}
3192 
3193 	if (size == 6 && !strncmp(buf, "active", size)) {
3194 		if (intel_pstate_driver) {
3195 			if (intel_pstate_driver == &intel_pstate)
3196 				return 0;
3197 
3198 			cpufreq_unregister_driver(intel_pstate_driver);
3199 		}
3200 
3201 		return intel_pstate_register_driver(&intel_pstate);
3202 	}
3203 
3204 	if (size == 7 && !strncmp(buf, "passive", size)) {
3205 		if (intel_pstate_driver) {
3206 			if (intel_pstate_driver == &intel_cpufreq)
3207 				return 0;
3208 
3209 			cpufreq_unregister_driver(intel_pstate_driver);
3210 			intel_pstate_sysfs_hide_hwp_dynamic_boost();
3211 		}
3212 
3213 		return intel_pstate_register_driver(&intel_cpufreq);
3214 	}
3215 
3216 	return -EINVAL;
3217 }
3218 
3219 static int no_load __initdata;
3220 static int no_hwp __initdata;
3221 static int hwp_only __initdata;
3222 static unsigned int force_load __initdata;
3223 
3224 static int __init intel_pstate_msrs_not_valid(void)
3225 {
3226 	if (!pstate_funcs.get_max(0) ||
3227 	    !pstate_funcs.get_min(0) ||
3228 	    !pstate_funcs.get_turbo(0))
3229 		return -ENODEV;
3230 
3231 	return 0;
3232 }
3233 
3234 static void __init copy_cpu_funcs(struct pstate_funcs *funcs)
3235 {
3236 	pstate_funcs.get_max   = funcs->get_max;
3237 	pstate_funcs.get_max_physical = funcs->get_max_physical;
3238 	pstate_funcs.get_min   = funcs->get_min;
3239 	pstate_funcs.get_turbo = funcs->get_turbo;
3240 	pstate_funcs.get_scaling = funcs->get_scaling;
3241 	pstate_funcs.get_val   = funcs->get_val;
3242 	pstate_funcs.get_vid   = funcs->get_vid;
3243 	pstate_funcs.get_aperf_mperf_shift = funcs->get_aperf_mperf_shift;
3244 }
3245 
3246 #ifdef CONFIG_ACPI
3247 
3248 static bool __init intel_pstate_no_acpi_pss(void)
3249 {
3250 	int i;
3251 
3252 	for_each_possible_cpu(i) {
3253 		acpi_status status;
3254 		union acpi_object *pss;
3255 		struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
3256 		struct acpi_processor *pr = per_cpu(processors, i);
3257 
3258 		if (!pr)
3259 			continue;
3260 
3261 		status = acpi_evaluate_object(pr->handle, "_PSS", NULL, &buffer);
3262 		if (ACPI_FAILURE(status))
3263 			continue;
3264 
3265 		pss = buffer.pointer;
3266 		if (pss && pss->type == ACPI_TYPE_PACKAGE) {
3267 			kfree(pss);
3268 			return false;
3269 		}
3270 
3271 		kfree(pss);
3272 	}
3273 
3274 	pr_debug("ACPI _PSS not found\n");
3275 	return true;
3276 }
3277 
3278 static bool __init intel_pstate_no_acpi_pcch(void)
3279 {
3280 	acpi_status status;
3281 	acpi_handle handle;
3282 
3283 	status = acpi_get_handle(NULL, "\\_SB", &handle);
3284 	if (ACPI_FAILURE(status))
3285 		goto not_found;
3286 
3287 	if (acpi_has_method(handle, "PCCH"))
3288 		return false;
3289 
3290 not_found:
3291 	pr_debug("ACPI PCCH not found\n");
3292 	return true;
3293 }
3294 
3295 static bool __init intel_pstate_has_acpi_ppc(void)
3296 {
3297 	int i;
3298 
3299 	for_each_possible_cpu(i) {
3300 		struct acpi_processor *pr = per_cpu(processors, i);
3301 
3302 		if (!pr)
3303 			continue;
3304 		if (acpi_has_method(pr->handle, "_PPC"))
3305 			return true;
3306 	}
3307 	pr_debug("ACPI _PPC not found\n");
3308 	return false;
3309 }
3310 
3311 enum {
3312 	PSS,
3313 	PPC,
3314 };
3315 
3316 /* Hardware vendor-specific info that has its own power management modes */
3317 static struct acpi_platform_list plat_info[] __initdata = {
3318 	{"HP    ", "ProLiant", 0, ACPI_SIG_FADT, all_versions, NULL, PSS},
3319 	{"ORACLE", "X4-2    ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3320 	{"ORACLE", "X4-2L   ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3321 	{"ORACLE", "X4-2B   ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3322 	{"ORACLE", "X3-2    ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3323 	{"ORACLE", "X3-2L   ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3324 	{"ORACLE", "X3-2B   ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3325 	{"ORACLE", "X4470M2 ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3326 	{"ORACLE", "X4270M3 ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3327 	{"ORACLE", "X4270M2 ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3328 	{"ORACLE", "X4170M2 ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3329 	{"ORACLE", "X4170 M3", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3330 	{"ORACLE", "X4275 M3", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3331 	{"ORACLE", "X6-2    ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3332 	{"ORACLE", "Sudbury ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3333 	{ } /* End */
3334 };
3335 
3336 #define BITMASK_OOB	(BIT(8) | BIT(18))
3337 
3338 static bool __init intel_pstate_platform_pwr_mgmt_exists(void)
3339 {
3340 	const struct x86_cpu_id *id;
3341 	u64 misc_pwr;
3342 	int idx;
3343 
3344 	id = x86_match_cpu(intel_pstate_cpu_oob_ids);
3345 	if (id) {
3346 		rdmsrl(MSR_MISC_PWR_MGMT, misc_pwr);
3347 		if (misc_pwr & BITMASK_OOB) {
3348 			pr_debug("Bit 8 or 18 in the MISC_PWR_MGMT MSR set\n");
3349 			pr_debug("P states are controlled in Out of Band mode by the firmware/hardware\n");
3350 			return true;
3351 		}
3352 	}
3353 
3354 	idx = acpi_match_platform_list(plat_info);
3355 	if (idx < 0)
3356 		return false;
3357 
3358 	switch (plat_info[idx].data) {
3359 	case PSS:
3360 		if (!intel_pstate_no_acpi_pss())
3361 			return false;
3362 
3363 		return intel_pstate_no_acpi_pcch();
3364 	case PPC:
3365 		return intel_pstate_has_acpi_ppc() && !force_load;
3366 	}
3367 
3368 	return false;
3369 }
3370 
3371 static void intel_pstate_request_control_from_smm(void)
3372 {
3373 	/*
3374 	 * It may be unsafe to request P-states control from SMM if _PPC support
3375 	 * has not been enabled.
3376 	 */
3377 	if (acpi_ppc)
3378 		acpi_processor_pstate_control();
3379 }
3380 #else /* CONFIG_ACPI not enabled */
3381 static inline bool intel_pstate_platform_pwr_mgmt_exists(void) { return false; }
3382 static inline bool intel_pstate_has_acpi_ppc(void) { return false; }
3383 static inline void intel_pstate_request_control_from_smm(void) {}
3384 #endif /* CONFIG_ACPI */
3385 
3386 #define INTEL_PSTATE_HWP_BROADWELL	0x01
3387 
3388 #define X86_MATCH_HWP(model, hwp_mode)					\
3389 	X86_MATCH_VENDOR_FAM_MODEL_FEATURE(INTEL, 6, INTEL_FAM6_##model, \
3390 					   X86_FEATURE_HWP, hwp_mode)
3391 
3392 static const struct x86_cpu_id hwp_support_ids[] __initconst = {
3393 	X86_MATCH_HWP(BROADWELL_X,	INTEL_PSTATE_HWP_BROADWELL),
3394 	X86_MATCH_HWP(BROADWELL_D,	INTEL_PSTATE_HWP_BROADWELL),
3395 	X86_MATCH_HWP(ANY,		0),
3396 	{}
3397 };
3398 
3399 static bool intel_pstate_hwp_is_enabled(void)
3400 {
3401 	u64 value;
3402 
3403 	rdmsrl(MSR_PM_ENABLE, value);
3404 	return !!(value & 0x1);
3405 }
3406 
3407 static const struct x86_cpu_id intel_epp_balance_perf[] = {
3408 	/*
3409 	 * Set EPP value as 102, this is the max suggested EPP
3410 	 * which can result in one core turbo frequency for
3411 	 * AlderLake Mobile CPUs.
3412 	 */
3413 	X86_MATCH_INTEL_FAM6_MODEL(ALDERLAKE_L, 102),
3414 	X86_MATCH_INTEL_FAM6_MODEL(SAPPHIRERAPIDS_X, 32),
3415 	{}
3416 };
3417 
3418 static int __init intel_pstate_init(void)
3419 {
3420 	static struct cpudata **_all_cpu_data;
3421 	const struct x86_cpu_id *id;
3422 	int rc;
3423 
3424 	if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL)
3425 		return -ENODEV;
3426 
3427 	id = x86_match_cpu(hwp_support_ids);
3428 	if (id) {
3429 		hwp_forced = intel_pstate_hwp_is_enabled();
3430 
3431 		if (hwp_forced)
3432 			pr_info("HWP enabled by BIOS\n");
3433 		else if (no_load)
3434 			return -ENODEV;
3435 
3436 		copy_cpu_funcs(&core_funcs);
3437 		/*
3438 		 * Avoid enabling HWP for processors without EPP support,
3439 		 * because that means incomplete HWP implementation which is a
3440 		 * corner case and supporting it is generally problematic.
3441 		 *
3442 		 * If HWP is enabled already, though, there is no choice but to
3443 		 * deal with it.
3444 		 */
3445 		if ((!no_hwp && boot_cpu_has(X86_FEATURE_HWP_EPP)) || hwp_forced) {
3446 			WRITE_ONCE(hwp_active, 1);
3447 			hwp_mode_bdw = id->driver_data;
3448 			intel_pstate.attr = hwp_cpufreq_attrs;
3449 			intel_cpufreq.attr = hwp_cpufreq_attrs;
3450 			intel_cpufreq.flags |= CPUFREQ_NEED_UPDATE_LIMITS;
3451 			intel_cpufreq.adjust_perf = intel_cpufreq_adjust_perf;
3452 			if (!default_driver)
3453 				default_driver = &intel_pstate;
3454 
3455 			pstate_funcs.get_cpu_scaling = hwp_get_cpu_scaling;
3456 
3457 			goto hwp_cpu_matched;
3458 		}
3459 		pr_info("HWP not enabled\n");
3460 	} else {
3461 		if (no_load)
3462 			return -ENODEV;
3463 
3464 		id = x86_match_cpu(intel_pstate_cpu_ids);
3465 		if (!id) {
3466 			pr_info("CPU model not supported\n");
3467 			return -ENODEV;
3468 		}
3469 
3470 		copy_cpu_funcs((struct pstate_funcs *)id->driver_data);
3471 	}
3472 
3473 	if (intel_pstate_msrs_not_valid()) {
3474 		pr_info("Invalid MSRs\n");
3475 		return -ENODEV;
3476 	}
3477 	/* Without HWP start in the passive mode. */
3478 	if (!default_driver)
3479 		default_driver = &intel_cpufreq;
3480 
3481 hwp_cpu_matched:
3482 	/*
3483 	 * The Intel pstate driver will be ignored if the platform
3484 	 * firmware has its own power management modes.
3485 	 */
3486 	if (intel_pstate_platform_pwr_mgmt_exists()) {
3487 		pr_info("P-states controlled by the platform\n");
3488 		return -ENODEV;
3489 	}
3490 
3491 	if (!hwp_active && hwp_only)
3492 		return -ENOTSUPP;
3493 
3494 	pr_info("Intel P-state driver initializing\n");
3495 
3496 	_all_cpu_data = vzalloc(array_size(sizeof(void *), num_possible_cpus()));
3497 	if (!_all_cpu_data)
3498 		return -ENOMEM;
3499 
3500 	WRITE_ONCE(all_cpu_data, _all_cpu_data);
3501 
3502 	intel_pstate_request_control_from_smm();
3503 
3504 	intel_pstate_sysfs_expose_params();
3505 
3506 	if (hwp_active) {
3507 		const struct x86_cpu_id *id = x86_match_cpu(intel_epp_balance_perf);
3508 
3509 		if (id)
3510 			epp_values[EPP_INDEX_BALANCE_PERFORMANCE] = id->driver_data;
3511 	}
3512 
3513 	mutex_lock(&intel_pstate_driver_lock);
3514 	rc = intel_pstate_register_driver(default_driver);
3515 	mutex_unlock(&intel_pstate_driver_lock);
3516 	if (rc) {
3517 		intel_pstate_sysfs_remove();
3518 		return rc;
3519 	}
3520 
3521 	if (hwp_active) {
3522 		const struct x86_cpu_id *id;
3523 
3524 		id = x86_match_cpu(intel_pstate_cpu_ee_disable_ids);
3525 		if (id) {
3526 			set_power_ctl_ee_state(false);
3527 			pr_info("Disabling energy efficiency optimization\n");
3528 		}
3529 
3530 		pr_info("HWP enabled\n");
3531 	} else if (boot_cpu_has(X86_FEATURE_HYBRID_CPU)) {
3532 		pr_warn("Problematic setup: Hybrid processor with disabled HWP\n");
3533 	}
3534 
3535 	return 0;
3536 }
3537 device_initcall(intel_pstate_init);
3538 
3539 static int __init intel_pstate_setup(char *str)
3540 {
3541 	if (!str)
3542 		return -EINVAL;
3543 
3544 	if (!strcmp(str, "disable"))
3545 		no_load = 1;
3546 	else if (!strcmp(str, "active"))
3547 		default_driver = &intel_pstate;
3548 	else if (!strcmp(str, "passive"))
3549 		default_driver = &intel_cpufreq;
3550 
3551 	if (!strcmp(str, "no_hwp"))
3552 		no_hwp = 1;
3553 
3554 	if (!strcmp(str, "force"))
3555 		force_load = 1;
3556 	if (!strcmp(str, "hwp_only"))
3557 		hwp_only = 1;
3558 	if (!strcmp(str, "per_cpu_perf_limits"))
3559 		per_cpu_limits = true;
3560 
3561 #ifdef CONFIG_ACPI
3562 	if (!strcmp(str, "support_acpi_ppc"))
3563 		acpi_ppc = true;
3564 #endif
3565 
3566 	return 0;
3567 }
3568 early_param("intel_pstate", intel_pstate_setup);
3569 
3570 MODULE_AUTHOR("Dirk Brandewie <dirk.j.brandewie@intel.com>");
3571 MODULE_DESCRIPTION("'intel_pstate' - P state driver Intel Core processors");
3572