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