Linux Audio

Check our new training course

Loading...
Note: File does not exist in v3.1.
   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
  58static inline int32_t mul_fp(int32_t x, int32_t y)
  59{
  60	return ((int64_t)x * (int64_t)y) >> FRAC_BITS;
  61}
  62
  63static inline int32_t div_fp(s64 x, s64 y)
  64{
  65	return div64_s64((int64_t)x << FRAC_BITS, y);
  66}
  67
  68static 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
  79static inline u64 mul_ext_fp(u64 x, u64 y)
  80{
  81	return (x * y) >> EXT_FRAC_BITS;
  82}
  83
  84static 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 */
 107struct 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 */
 133struct 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 */
 159struct 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 */
 179struct 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 */
 225struct 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
 262static 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_aperf_mperf_shift: Callback to get the APERF vs MPERF frequency difference
 272 * @get_val:		Callback to convert P state to actual MSR write value
 273 * @get_vid:		Callback to get VID data for Atom platforms
 274 *
 275 * Core and Atom CPU models have different way to get P State limits. This
 276 * structure is used to store those callbacks.
 277 */
 278struct pstate_funcs {
 279	int (*get_max)(void);
 280	int (*get_max_physical)(void);
 281	int (*get_min)(void);
 282	int (*get_turbo)(void);
 283	int (*get_scaling)(void);
 284	int (*get_aperf_mperf_shift)(void);
 285	u64 (*get_val)(struct cpudata*, int pstate);
 286	void (*get_vid)(struct cpudata *);
 287};
 288
 289static struct pstate_funcs pstate_funcs __read_mostly;
 290
 291static int hwp_active __read_mostly;
 292static int hwp_mode_bdw __read_mostly;
 293static bool per_cpu_limits __read_mostly;
 294static bool hwp_boost __read_mostly;
 295
 296static struct cpufreq_driver *intel_pstate_driver __read_mostly;
 297
 298#ifdef CONFIG_ACPI
 299static bool acpi_ppc;
 300#endif
 301
 302static struct global_params global;
 303
 304static DEFINE_MUTEX(intel_pstate_driver_lock);
 305static DEFINE_MUTEX(intel_pstate_limits_lock);
 306
 307#ifdef CONFIG_ACPI
 308
 309static bool intel_pstate_acpi_pm_profile_server(void)
 310{
 311	if (acpi_gbl_FADT.preferred_profile == PM_ENTERPRISE_SERVER ||
 312	    acpi_gbl_FADT.preferred_profile == PM_PERFORMANCE_SERVER)
 313		return true;
 314
 315	return false;
 316}
 317
 318static bool intel_pstate_get_ppc_enable_status(void)
 319{
 320	if (intel_pstate_acpi_pm_profile_server())
 321		return true;
 322
 323	return acpi_ppc;
 324}
 325
 326#ifdef CONFIG_ACPI_CPPC_LIB
 327
 328/* The work item is needed to avoid CPU hotplug locking issues */
 329static void intel_pstste_sched_itmt_work_fn(struct work_struct *work)
 330{
 331	sched_set_itmt_support();
 332}
 333
 334static DECLARE_WORK(sched_itmt_work, intel_pstste_sched_itmt_work_fn);
 335
 336static void intel_pstate_set_itmt_prio(int cpu)
 337{
 338	struct cppc_perf_caps cppc_perf;
 339	static u32 max_highest_perf = 0, min_highest_perf = U32_MAX;
 340	int ret;
 341
 342	ret = cppc_get_perf_caps(cpu, &cppc_perf);
 343	if (ret)
 344		return;
 345
 346	/*
 347	 * The priorities can be set regardless of whether or not
 348	 * sched_set_itmt_support(true) has been called and it is valid to
 349	 * update them at any time after it has been called.
 350	 */
 351	sched_set_itmt_core_prio(cppc_perf.highest_perf, cpu);
 352
 353	if (max_highest_perf <= min_highest_perf) {
 354		if (cppc_perf.highest_perf > max_highest_perf)
 355			max_highest_perf = cppc_perf.highest_perf;
 356
 357		if (cppc_perf.highest_perf < min_highest_perf)
 358			min_highest_perf = cppc_perf.highest_perf;
 359
 360		if (max_highest_perf > min_highest_perf) {
 361			/*
 362			 * This code can be run during CPU online under the
 363			 * CPU hotplug locks, so sched_set_itmt_support()
 364			 * cannot be called from here.  Queue up a work item
 365			 * to invoke it.
 366			 */
 367			schedule_work(&sched_itmt_work);
 368		}
 369	}
 370}
 371
 372static int intel_pstate_get_cppc_guaranteed(int cpu)
 373{
 374	struct cppc_perf_caps cppc_perf;
 375	int ret;
 376
 377	ret = cppc_get_perf_caps(cpu, &cppc_perf);
 378	if (ret)
 379		return ret;
 380
 381	if (cppc_perf.guaranteed_perf)
 382		return cppc_perf.guaranteed_perf;
 383
 384	return cppc_perf.nominal_perf;
 385}
 386
 387#else /* CONFIG_ACPI_CPPC_LIB */
 388static inline void intel_pstate_set_itmt_prio(int cpu)
 389{
 390}
 391#endif /* CONFIG_ACPI_CPPC_LIB */
 392
 393static void intel_pstate_init_acpi_perf_limits(struct cpufreq_policy *policy)
 394{
 395	struct cpudata *cpu;
 396	int ret;
 397	int i;
 398
 399	if (hwp_active) {
 400		intel_pstate_set_itmt_prio(policy->cpu);
 401		return;
 402	}
 403
 404	if (!intel_pstate_get_ppc_enable_status())
 405		return;
 406
 407	cpu = all_cpu_data[policy->cpu];
 408
 409	ret = acpi_processor_register_performance(&cpu->acpi_perf_data,
 410						  policy->cpu);
 411	if (ret)
 412		return;
 413
 414	/*
 415	 * Check if the control value in _PSS is for PERF_CTL MSR, which should
 416	 * guarantee that the states returned by it map to the states in our
 417	 * list directly.
 418	 */
 419	if (cpu->acpi_perf_data.control_register.space_id !=
 420						ACPI_ADR_SPACE_FIXED_HARDWARE)
 421		goto err;
 422
 423	/*
 424	 * If there is only one entry _PSS, simply ignore _PSS and continue as
 425	 * usual without taking _PSS into account
 426	 */
 427	if (cpu->acpi_perf_data.state_count < 2)
 428		goto err;
 429
 430	pr_debug("CPU%u - ACPI _PSS perf data\n", policy->cpu);
 431	for (i = 0; i < cpu->acpi_perf_data.state_count; i++) {
 432		pr_debug("     %cP%d: %u MHz, %u mW, 0x%x\n",
 433			 (i == cpu->acpi_perf_data.state ? '*' : ' '), i,
 434			 (u32) cpu->acpi_perf_data.states[i].core_frequency,
 435			 (u32) cpu->acpi_perf_data.states[i].power,
 436			 (u32) cpu->acpi_perf_data.states[i].control);
 437	}
 438
 439	/*
 440	 * The _PSS table doesn't contain whole turbo frequency range.
 441	 * This just contains +1 MHZ above the max non turbo frequency,
 442	 * with control value corresponding to max turbo ratio. But
 443	 * when cpufreq set policy is called, it will call with this
 444	 * max frequency, which will cause a reduced performance as
 445	 * this driver uses real max turbo frequency as the max
 446	 * frequency. So correct this frequency in _PSS table to
 447	 * correct max turbo frequency based on the turbo state.
 448	 * Also need to convert to MHz as _PSS freq is in MHz.
 449	 */
 450	if (!global.turbo_disabled)
 451		cpu->acpi_perf_data.states[0].core_frequency =
 452					policy->cpuinfo.max_freq / 1000;
 453	cpu->valid_pss_table = true;
 454	pr_debug("_PPC limits will be enforced\n");
 455
 456	return;
 457
 458 err:
 459	cpu->valid_pss_table = false;
 460	acpi_processor_unregister_performance(policy->cpu);
 461}
 462
 463static void intel_pstate_exit_perf_limits(struct cpufreq_policy *policy)
 464{
 465	struct cpudata *cpu;
 466
 467	cpu = all_cpu_data[policy->cpu];
 468	if (!cpu->valid_pss_table)
 469		return;
 470
 471	acpi_processor_unregister_performance(policy->cpu);
 472}
 473
 474static bool intel_pstate_cppc_perf_valid(u32 perf, struct cppc_perf_caps *caps)
 475{
 476	return perf && perf <= caps->highest_perf && perf >= caps->lowest_perf;
 477}
 478
 479static bool intel_pstate_cppc_perf_caps(struct cpudata *cpu,
 480					struct cppc_perf_caps *caps)
 481{
 482	if (cppc_get_perf_caps(cpu->cpu, caps))
 483		return false;
 484
 485	return caps->highest_perf && caps->lowest_perf <= caps->highest_perf;
 486}
 487#else /* CONFIG_ACPI */
 488static inline void intel_pstate_init_acpi_perf_limits(struct cpufreq_policy *policy)
 489{
 490}
 491
 492static inline void intel_pstate_exit_perf_limits(struct cpufreq_policy *policy)
 493{
 494}
 495
 496static inline bool intel_pstate_acpi_pm_profile_server(void)
 497{
 498	return false;
 499}
 500#endif /* CONFIG_ACPI */
 501
 502#ifndef CONFIG_ACPI_CPPC_LIB
 503static inline int intel_pstate_get_cppc_guaranteed(int cpu)
 504{
 505	return -ENOTSUPP;
 506}
 507#endif /* CONFIG_ACPI_CPPC_LIB */
 508
 509static void intel_pstate_hybrid_hwp_perf_ctl_parity(struct cpudata *cpu)
 510{
 511	pr_debug("CPU%d: Using PERF_CTL scaling for HWP\n", cpu->cpu);
 512
 513	cpu->pstate.scaling = cpu->pstate.perf_ctl_scaling;
 514}
 515
 516/**
 517 * intel_pstate_hybrid_hwp_calibrate - Calibrate HWP performance levels.
 518 * @cpu: Target CPU.
 519 *
 520 * On hybrid processors, HWP may expose more performance levels than there are
 521 * P-states accessible through the PERF_CTL interface.  If that happens, the
 522 * scaling factor between HWP performance levels and CPU frequency will be less
 523 * than the scaling factor between P-state values and CPU frequency.
 524 *
 525 * In that case, the scaling factor between HWP performance levels and CPU
 526 * frequency needs to be determined which can be done with the help of the
 527 * observation that certain HWP performance levels should correspond to certain
 528 * P-states, like for example the HWP highest performance should correspond
 529 * to the maximum turbo P-state of the CPU.
 530 */
 531static void intel_pstate_hybrid_hwp_calibrate(struct cpudata *cpu)
 532{
 533	int perf_ctl_max_phys = cpu->pstate.max_pstate_physical;
 534	int perf_ctl_scaling = cpu->pstate.perf_ctl_scaling;
 535	int perf_ctl_turbo = pstate_funcs.get_turbo();
 536	int turbo_freq = perf_ctl_turbo * perf_ctl_scaling;
 537	int perf_ctl_max = pstate_funcs.get_max();
 538	int max_freq = perf_ctl_max * perf_ctl_scaling;
 539	int scaling = INT_MAX;
 540	int freq;
 541
 542	pr_debug("CPU%d: perf_ctl_max_phys = %d\n", cpu->cpu, perf_ctl_max_phys);
 543	pr_debug("CPU%d: perf_ctl_max = %d\n", cpu->cpu, perf_ctl_max);
 544	pr_debug("CPU%d: perf_ctl_turbo = %d\n", cpu->cpu, perf_ctl_turbo);
 545	pr_debug("CPU%d: perf_ctl_scaling = %d\n", cpu->cpu, perf_ctl_scaling);
 546
 547	pr_debug("CPU%d: HWP_CAP guaranteed = %d\n", cpu->cpu, cpu->pstate.max_pstate);
 548	pr_debug("CPU%d: HWP_CAP highest = %d\n", cpu->cpu, cpu->pstate.turbo_pstate);
 549
 550#ifdef CONFIG_ACPI
 551	if (IS_ENABLED(CONFIG_ACPI_CPPC_LIB)) {
 552		struct cppc_perf_caps caps;
 553
 554		if (intel_pstate_cppc_perf_caps(cpu, &caps)) {
 555			if (intel_pstate_cppc_perf_valid(caps.nominal_perf, &caps)) {
 556				pr_debug("CPU%d: Using CPPC nominal\n", cpu->cpu);
 557
 558				/*
 559				 * If the CPPC nominal performance is valid, it
 560				 * can be assumed to correspond to cpu_khz.
 561				 */
 562				if (caps.nominal_perf == perf_ctl_max_phys) {
 563					intel_pstate_hybrid_hwp_perf_ctl_parity(cpu);
 564					return;
 565				}
 566				scaling = DIV_ROUND_UP(cpu_khz, caps.nominal_perf);
 567			} else if (intel_pstate_cppc_perf_valid(caps.guaranteed_perf, &caps)) {
 568				pr_debug("CPU%d: Using CPPC guaranteed\n", cpu->cpu);
 569
 570				/*
 571				 * If the CPPC guaranteed performance is valid,
 572				 * it can be assumed to correspond to max_freq.
 573				 */
 574				if (caps.guaranteed_perf == perf_ctl_max) {
 575					intel_pstate_hybrid_hwp_perf_ctl_parity(cpu);
 576					return;
 577				}
 578				scaling = DIV_ROUND_UP(max_freq, caps.guaranteed_perf);
 579			}
 580		}
 581	}
 582#endif
 583	/*
 584	 * If using the CPPC data to compute the HWP-to-frequency scaling factor
 585	 * doesn't work, use the HWP_CAP gauranteed perf for this purpose with
 586	 * the assumption that it corresponds to max_freq.
 587	 */
 588	if (scaling > perf_ctl_scaling) {
 589		pr_debug("CPU%d: Using HWP_CAP guaranteed\n", cpu->cpu);
 590
 591		if (cpu->pstate.max_pstate == perf_ctl_max) {
 592			intel_pstate_hybrid_hwp_perf_ctl_parity(cpu);
 593			return;
 594		}
 595		scaling = DIV_ROUND_UP(max_freq, cpu->pstate.max_pstate);
 596		if (scaling > perf_ctl_scaling) {
 597			/*
 598			 * This should not happen, because it would mean that
 599			 * the number of HWP perf levels was less than the
 600			 * number of P-states, so use the PERF_CTL scaling in
 601			 * that case.
 602			 */
 603			pr_debug("CPU%d: scaling (%d) out of range\n", cpu->cpu,
 604				scaling);
 605
 606			intel_pstate_hybrid_hwp_perf_ctl_parity(cpu);
 607			return;
 608		}
 609	}
 610
 611	/*
 612	 * If the product of the HWP performance scaling factor obtained above
 613	 * and the HWP_CAP highest performance is greater than the maximum turbo
 614	 * frequency corresponding to the pstate_funcs.get_turbo() return value,
 615	 * the scaling factor is too high, so recompute it so that the HWP_CAP
 616	 * highest performance corresponds to the maximum turbo frequency.
 617	 */
 618	if (turbo_freq < cpu->pstate.turbo_pstate * scaling) {
 619		pr_debug("CPU%d: scaling too high (%d)\n", cpu->cpu, scaling);
 620
 621		cpu->pstate.turbo_freq = turbo_freq;
 622		scaling = DIV_ROUND_UP(turbo_freq, cpu->pstate.turbo_pstate);
 623	}
 624
 625	cpu->pstate.scaling = scaling;
 626
 627	pr_debug("CPU%d: HWP-to-frequency scaling factor: %d\n", cpu->cpu, scaling);
 628
 629	cpu->pstate.max_freq = rounddown(cpu->pstate.max_pstate * scaling,
 630					 perf_ctl_scaling);
 631
 632	freq = perf_ctl_max_phys * perf_ctl_scaling;
 633	cpu->pstate.max_pstate_physical = DIV_ROUND_UP(freq, scaling);
 634
 635	cpu->pstate.min_freq = cpu->pstate.min_pstate * perf_ctl_scaling;
 636	/*
 637	 * Cast the min P-state value retrieved via pstate_funcs.get_min() to
 638	 * the effective range of HWP performance levels.
 639	 */
 640	cpu->pstate.min_pstate = DIV_ROUND_UP(cpu->pstate.min_freq, scaling);
 641}
 642
 643static inline void update_turbo_state(void)
 644{
 645	u64 misc_en;
 646	struct cpudata *cpu;
 647
 648	cpu = all_cpu_data[0];
 649	rdmsrl(MSR_IA32_MISC_ENABLE, misc_en);
 650	global.turbo_disabled =
 651		(misc_en & MSR_IA32_MISC_ENABLE_TURBO_DISABLE ||
 652		 cpu->pstate.max_pstate == cpu->pstate.turbo_pstate);
 653}
 654
 655static int min_perf_pct_min(void)
 656{
 657	struct cpudata *cpu = all_cpu_data[0];
 658	int turbo_pstate = cpu->pstate.turbo_pstate;
 659
 660	return turbo_pstate ?
 661		(cpu->pstate.min_pstate * 100 / turbo_pstate) : 0;
 662}
 663
 664static s16 intel_pstate_get_epb(struct cpudata *cpu_data)
 665{
 666	u64 epb;
 667	int ret;
 668
 669	if (!boot_cpu_has(X86_FEATURE_EPB))
 670		return -ENXIO;
 671
 672	ret = rdmsrl_on_cpu(cpu_data->cpu, MSR_IA32_ENERGY_PERF_BIAS, &epb);
 673	if (ret)
 674		return (s16)ret;
 675
 676	return (s16)(epb & 0x0f);
 677}
 678
 679static s16 intel_pstate_get_epp(struct cpudata *cpu_data, u64 hwp_req_data)
 680{
 681	s16 epp;
 682
 683	if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
 684		/*
 685		 * When hwp_req_data is 0, means that caller didn't read
 686		 * MSR_HWP_REQUEST, so need to read and get EPP.
 687		 */
 688		if (!hwp_req_data) {
 689			epp = rdmsrl_on_cpu(cpu_data->cpu, MSR_HWP_REQUEST,
 690					    &hwp_req_data);
 691			if (epp)
 692				return epp;
 693		}
 694		epp = (hwp_req_data >> 24) & 0xff;
 695	} else {
 696		/* When there is no EPP present, HWP uses EPB settings */
 697		epp = intel_pstate_get_epb(cpu_data);
 698	}
 699
 700	return epp;
 701}
 702
 703static int intel_pstate_set_epb(int cpu, s16 pref)
 704{
 705	u64 epb;
 706	int ret;
 707
 708	if (!boot_cpu_has(X86_FEATURE_EPB))
 709		return -ENXIO;
 710
 711	ret = rdmsrl_on_cpu(cpu, MSR_IA32_ENERGY_PERF_BIAS, &epb);
 712	if (ret)
 713		return ret;
 714
 715	epb = (epb & ~0x0f) | pref;
 716	wrmsrl_on_cpu(cpu, MSR_IA32_ENERGY_PERF_BIAS, epb);
 717
 718	return 0;
 719}
 720
 721/*
 722 * EPP/EPB display strings corresponding to EPP index in the
 723 * energy_perf_strings[]
 724 *	index		String
 725 *-------------------------------------
 726 *	0		default
 727 *	1		performance
 728 *	2		balance_performance
 729 *	3		balance_power
 730 *	4		power
 731 */
 732static const char * const energy_perf_strings[] = {
 733	"default",
 734	"performance",
 735	"balance_performance",
 736	"balance_power",
 737	"power",
 738	NULL
 739};
 740static const unsigned int epp_values[] = {
 741	HWP_EPP_PERFORMANCE,
 742	HWP_EPP_BALANCE_PERFORMANCE,
 743	HWP_EPP_BALANCE_POWERSAVE,
 744	HWP_EPP_POWERSAVE
 745};
 746
 747static int intel_pstate_get_energy_pref_index(struct cpudata *cpu_data, int *raw_epp)
 748{
 749	s16 epp;
 750	int index = -EINVAL;
 751
 752	*raw_epp = 0;
 753	epp = intel_pstate_get_epp(cpu_data, 0);
 754	if (epp < 0)
 755		return epp;
 756
 757	if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
 758		if (epp == HWP_EPP_PERFORMANCE)
 759			return 1;
 760		if (epp == HWP_EPP_BALANCE_PERFORMANCE)
 761			return 2;
 762		if (epp == HWP_EPP_BALANCE_POWERSAVE)
 763			return 3;
 764		if (epp == HWP_EPP_POWERSAVE)
 765			return 4;
 766		*raw_epp = epp;
 767		return 0;
 768	} else if (boot_cpu_has(X86_FEATURE_EPB)) {
 769		/*
 770		 * Range:
 771		 *	0x00-0x03	:	Performance
 772		 *	0x04-0x07	:	Balance performance
 773		 *	0x08-0x0B	:	Balance power
 774		 *	0x0C-0x0F	:	Power
 775		 * The EPB is a 4 bit value, but our ranges restrict the
 776		 * value which can be set. Here only using top two bits
 777		 * effectively.
 778		 */
 779		index = (epp >> 2) + 1;
 780	}
 781
 782	return index;
 783}
 784
 785static int intel_pstate_set_epp(struct cpudata *cpu, u32 epp)
 786{
 787	int ret;
 788
 789	/*
 790	 * Use the cached HWP Request MSR value, because in the active mode the
 791	 * register itself may be updated by intel_pstate_hwp_boost_up() or
 792	 * intel_pstate_hwp_boost_down() at any time.
 793	 */
 794	u64 value = READ_ONCE(cpu->hwp_req_cached);
 795
 796	value &= ~GENMASK_ULL(31, 24);
 797	value |= (u64)epp << 24;
 798	/*
 799	 * The only other updater of hwp_req_cached in the active mode,
 800	 * intel_pstate_hwp_set(), is called under the same lock as this
 801	 * function, so it cannot run in parallel with the update below.
 802	 */
 803	WRITE_ONCE(cpu->hwp_req_cached, value);
 804	ret = wrmsrl_on_cpu(cpu->cpu, MSR_HWP_REQUEST, value);
 805	if (!ret)
 806		cpu->epp_cached = epp;
 807
 808	return ret;
 809}
 810
 811static int intel_pstate_set_energy_pref_index(struct cpudata *cpu_data,
 812					      int pref_index, bool use_raw,
 813					      u32 raw_epp)
 814{
 815	int epp = -EINVAL;
 816	int ret;
 817
 818	if (!pref_index)
 819		epp = cpu_data->epp_default;
 820
 821	if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
 822		if (use_raw)
 823			epp = raw_epp;
 824		else if (epp == -EINVAL)
 825			epp = epp_values[pref_index - 1];
 826
 827		/*
 828		 * To avoid confusion, refuse to set EPP to any values different
 829		 * from 0 (performance) if the current policy is "performance",
 830		 * because those values would be overridden.
 831		 */
 832		if (epp > 0 && cpu_data->policy == CPUFREQ_POLICY_PERFORMANCE)
 833			return -EBUSY;
 834
 835		ret = intel_pstate_set_epp(cpu_data, epp);
 836	} else {
 837		if (epp == -EINVAL)
 838			epp = (pref_index - 1) << 2;
 839		ret = intel_pstate_set_epb(cpu_data->cpu, epp);
 840	}
 841
 842	return ret;
 843}
 844
 845static ssize_t show_energy_performance_available_preferences(
 846				struct cpufreq_policy *policy, char *buf)
 847{
 848	int i = 0;
 849	int ret = 0;
 850
 851	while (energy_perf_strings[i] != NULL)
 852		ret += sprintf(&buf[ret], "%s ", energy_perf_strings[i++]);
 853
 854	ret += sprintf(&buf[ret], "\n");
 855
 856	return ret;
 857}
 858
 859cpufreq_freq_attr_ro(energy_performance_available_preferences);
 860
 861static struct cpufreq_driver intel_pstate;
 862
 863static ssize_t store_energy_performance_preference(
 864		struct cpufreq_policy *policy, const char *buf, size_t count)
 865{
 866	struct cpudata *cpu = all_cpu_data[policy->cpu];
 867	char str_preference[21];
 868	bool raw = false;
 869	ssize_t ret;
 870	u32 epp = 0;
 871
 872	ret = sscanf(buf, "%20s", str_preference);
 873	if (ret != 1)
 874		return -EINVAL;
 875
 876	ret = match_string(energy_perf_strings, -1, str_preference);
 877	if (ret < 0) {
 878		if (!boot_cpu_has(X86_FEATURE_HWP_EPP))
 879			return ret;
 880
 881		ret = kstrtouint(buf, 10, &epp);
 882		if (ret)
 883			return ret;
 884
 885		if (epp > 255)
 886			return -EINVAL;
 887
 888		raw = true;
 889	}
 890
 891	/*
 892	 * This function runs with the policy R/W semaphore held, which
 893	 * guarantees that the driver pointer will not change while it is
 894	 * running.
 895	 */
 896	if (!intel_pstate_driver)
 897		return -EAGAIN;
 898
 899	mutex_lock(&intel_pstate_limits_lock);
 900
 901	if (intel_pstate_driver == &intel_pstate) {
 902		ret = intel_pstate_set_energy_pref_index(cpu, ret, raw, epp);
 903	} else {
 904		/*
 905		 * In the passive mode the governor needs to be stopped on the
 906		 * target CPU before the EPP update and restarted after it,
 907		 * which is super-heavy-weight, so make sure it is worth doing
 908		 * upfront.
 909		 */
 910		if (!raw)
 911			epp = ret ? epp_values[ret - 1] : cpu->epp_default;
 912
 913		if (cpu->epp_cached != epp) {
 914			int err;
 915
 916			cpufreq_stop_governor(policy);
 917			ret = intel_pstate_set_epp(cpu, epp);
 918			err = cpufreq_start_governor(policy);
 919			if (!ret)
 920				ret = err;
 921		}
 922	}
 923
 924	mutex_unlock(&intel_pstate_limits_lock);
 925
 926	return ret ?: count;
 927}
 928
 929static ssize_t show_energy_performance_preference(
 930				struct cpufreq_policy *policy, char *buf)
 931{
 932	struct cpudata *cpu_data = all_cpu_data[policy->cpu];
 933	int preference, raw_epp;
 934
 935	preference = intel_pstate_get_energy_pref_index(cpu_data, &raw_epp);
 936	if (preference < 0)
 937		return preference;
 938
 939	if (raw_epp)
 940		return  sprintf(buf, "%d\n", raw_epp);
 941	else
 942		return  sprintf(buf, "%s\n", energy_perf_strings[preference]);
 943}
 944
 945cpufreq_freq_attr_rw(energy_performance_preference);
 946
 947static ssize_t show_base_frequency(struct cpufreq_policy *policy, char *buf)
 948{
 949	struct cpudata *cpu = all_cpu_data[policy->cpu];
 950	int ratio, freq;
 951
 952	ratio = intel_pstate_get_cppc_guaranteed(policy->cpu);
 953	if (ratio <= 0) {
 954		u64 cap;
 955
 956		rdmsrl_on_cpu(policy->cpu, MSR_HWP_CAPABILITIES, &cap);
 957		ratio = HWP_GUARANTEED_PERF(cap);
 958	}
 959
 960	freq = ratio * cpu->pstate.scaling;
 961	if (cpu->pstate.scaling != cpu->pstate.perf_ctl_scaling)
 962		freq = rounddown(freq, cpu->pstate.perf_ctl_scaling);
 963
 964	return sprintf(buf, "%d\n", freq);
 965}
 966
 967cpufreq_freq_attr_ro(base_frequency);
 968
 969static struct freq_attr *hwp_cpufreq_attrs[] = {
 970	&energy_performance_preference,
 971	&energy_performance_available_preferences,
 972	&base_frequency,
 973	NULL,
 974};
 975
 976static void __intel_pstate_get_hwp_cap(struct cpudata *cpu)
 977{
 978	u64 cap;
 979
 980	rdmsrl_on_cpu(cpu->cpu, MSR_HWP_CAPABILITIES, &cap);
 981	WRITE_ONCE(cpu->hwp_cap_cached, cap);
 982	cpu->pstate.max_pstate = HWP_GUARANTEED_PERF(cap);
 983	cpu->pstate.turbo_pstate = HWP_HIGHEST_PERF(cap);
 984}
 985
 986static void intel_pstate_get_hwp_cap(struct cpudata *cpu)
 987{
 988	int scaling = cpu->pstate.scaling;
 989
 990	__intel_pstate_get_hwp_cap(cpu);
 991
 992	cpu->pstate.max_freq = cpu->pstate.max_pstate * scaling;
 993	cpu->pstate.turbo_freq = cpu->pstate.turbo_pstate * scaling;
 994	if (scaling != cpu->pstate.perf_ctl_scaling) {
 995		int perf_ctl_scaling = cpu->pstate.perf_ctl_scaling;
 996
 997		cpu->pstate.max_freq = rounddown(cpu->pstate.max_freq,
 998						 perf_ctl_scaling);
 999		cpu->pstate.turbo_freq = rounddown(cpu->pstate.turbo_freq,
1000						   perf_ctl_scaling);
1001	}
1002}
1003
1004static void intel_pstate_hwp_set(unsigned int cpu)
1005{
1006	struct cpudata *cpu_data = all_cpu_data[cpu];
1007	int max, min;
1008	u64 value;
1009	s16 epp;
1010
1011	max = cpu_data->max_perf_ratio;
1012	min = cpu_data->min_perf_ratio;
1013
1014	if (cpu_data->policy == CPUFREQ_POLICY_PERFORMANCE)
1015		min = max;
1016
1017	rdmsrl_on_cpu(cpu, MSR_HWP_REQUEST, &value);
1018
1019	value &= ~HWP_MIN_PERF(~0L);
1020	value |= HWP_MIN_PERF(min);
1021
1022	value &= ~HWP_MAX_PERF(~0L);
1023	value |= HWP_MAX_PERF(max);
1024
1025	if (cpu_data->epp_policy == cpu_data->policy)
1026		goto skip_epp;
1027
1028	cpu_data->epp_policy = cpu_data->policy;
1029
1030	if (cpu_data->policy == CPUFREQ_POLICY_PERFORMANCE) {
1031		epp = intel_pstate_get_epp(cpu_data, value);
1032		cpu_data->epp_powersave = epp;
1033		/* If EPP read was failed, then don't try to write */
1034		if (epp < 0)
1035			goto skip_epp;
1036
1037		epp = 0;
1038	} else {
1039		/* skip setting EPP, when saved value is invalid */
1040		if (cpu_data->epp_powersave < 0)
1041			goto skip_epp;
1042
1043		/*
1044		 * No need to restore EPP when it is not zero. This
1045		 * means:
1046		 *  - Policy is not changed
1047		 *  - user has manually changed
1048		 *  - Error reading EPB
1049		 */
1050		epp = intel_pstate_get_epp(cpu_data, value);
1051		if (epp)
1052			goto skip_epp;
1053
1054		epp = cpu_data->epp_powersave;
1055	}
1056	if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
1057		value &= ~GENMASK_ULL(31, 24);
1058		value |= (u64)epp << 24;
1059	} else {
1060		intel_pstate_set_epb(cpu, epp);
1061	}
1062skip_epp:
1063	WRITE_ONCE(cpu_data->hwp_req_cached, value);
1064	wrmsrl_on_cpu(cpu, MSR_HWP_REQUEST, value);
1065}
1066
1067static void intel_pstate_hwp_offline(struct cpudata *cpu)
1068{
1069	u64 value = READ_ONCE(cpu->hwp_req_cached);
1070	int min_perf;
1071
1072	if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
1073		/*
1074		 * In case the EPP has been set to "performance" by the
1075		 * active mode "performance" scaling algorithm, replace that
1076		 * temporary value with the cached EPP one.
1077		 */
1078		value &= ~GENMASK_ULL(31, 24);
1079		value |= HWP_ENERGY_PERF_PREFERENCE(cpu->epp_cached);
1080		WRITE_ONCE(cpu->hwp_req_cached, value);
1081	}
1082
1083	value &= ~GENMASK_ULL(31, 0);
1084	min_perf = HWP_LOWEST_PERF(READ_ONCE(cpu->hwp_cap_cached));
1085
1086	/* Set hwp_max = hwp_min */
1087	value |= HWP_MAX_PERF(min_perf);
1088	value |= HWP_MIN_PERF(min_perf);
1089
1090	/* Set EPP to min */
1091	if (boot_cpu_has(X86_FEATURE_HWP_EPP))
1092		value |= HWP_ENERGY_PERF_PREFERENCE(HWP_EPP_POWERSAVE);
1093
1094	wrmsrl_on_cpu(cpu->cpu, MSR_HWP_REQUEST, value);
1095}
1096
1097#define POWER_CTL_EE_ENABLE	1
1098#define POWER_CTL_EE_DISABLE	2
1099
1100static int power_ctl_ee_state;
1101
1102static void set_power_ctl_ee_state(bool input)
1103{
1104	u64 power_ctl;
1105
1106	mutex_lock(&intel_pstate_driver_lock);
1107	rdmsrl(MSR_IA32_POWER_CTL, power_ctl);
1108	if (input) {
1109		power_ctl &= ~BIT(MSR_IA32_POWER_CTL_BIT_EE);
1110		power_ctl_ee_state = POWER_CTL_EE_ENABLE;
1111	} else {
1112		power_ctl |= BIT(MSR_IA32_POWER_CTL_BIT_EE);
1113		power_ctl_ee_state = POWER_CTL_EE_DISABLE;
1114	}
1115	wrmsrl(MSR_IA32_POWER_CTL, power_ctl);
1116	mutex_unlock(&intel_pstate_driver_lock);
1117}
1118
1119static void intel_pstate_hwp_enable(struct cpudata *cpudata);
1120
1121static void intel_pstate_hwp_reenable(struct cpudata *cpu)
1122{
1123	intel_pstate_hwp_enable(cpu);
1124	wrmsrl_on_cpu(cpu->cpu, MSR_HWP_REQUEST, READ_ONCE(cpu->hwp_req_cached));
1125}
1126
1127static int intel_pstate_suspend(struct cpufreq_policy *policy)
1128{
1129	struct cpudata *cpu = all_cpu_data[policy->cpu];
1130
1131	pr_debug("CPU %d suspending\n", cpu->cpu);
1132
1133	cpu->suspended = true;
1134
1135	return 0;
1136}
1137
1138static int intel_pstate_resume(struct cpufreq_policy *policy)
1139{
1140	struct cpudata *cpu = all_cpu_data[policy->cpu];
1141
1142	pr_debug("CPU %d resuming\n", cpu->cpu);
1143
1144	/* Only restore if the system default is changed */
1145	if (power_ctl_ee_state == POWER_CTL_EE_ENABLE)
1146		set_power_ctl_ee_state(true);
1147	else if (power_ctl_ee_state == POWER_CTL_EE_DISABLE)
1148		set_power_ctl_ee_state(false);
1149
1150	if (cpu->suspended && hwp_active) {
1151		mutex_lock(&intel_pstate_limits_lock);
1152
1153		/* Re-enable HWP, because "online" has not done that. */
1154		intel_pstate_hwp_reenable(cpu);
1155
1156		mutex_unlock(&intel_pstate_limits_lock);
1157	}
1158
1159	cpu->suspended = false;
1160
1161	return 0;
1162}
1163
1164static void intel_pstate_update_policies(void)
1165{
1166	int cpu;
1167
1168	for_each_possible_cpu(cpu)
1169		cpufreq_update_policy(cpu);
1170}
1171
1172static void intel_pstate_update_max_freq(unsigned int cpu)
1173{
1174	struct cpufreq_policy *policy = cpufreq_cpu_acquire(cpu);
1175	struct cpudata *cpudata;
1176
1177	if (!policy)
1178		return;
1179
1180	cpudata = all_cpu_data[cpu];
1181	policy->cpuinfo.max_freq = global.turbo_disabled_mf ?
1182			cpudata->pstate.max_freq : cpudata->pstate.turbo_freq;
1183
1184	refresh_frequency_limits(policy);
1185
1186	cpufreq_cpu_release(policy);
1187}
1188
1189static void intel_pstate_update_limits(unsigned int cpu)
1190{
1191	mutex_lock(&intel_pstate_driver_lock);
1192
1193	update_turbo_state();
1194	/*
1195	 * If turbo has been turned on or off globally, policy limits for
1196	 * all CPUs need to be updated to reflect that.
1197	 */
1198	if (global.turbo_disabled_mf != global.turbo_disabled) {
1199		global.turbo_disabled_mf = global.turbo_disabled;
1200		arch_set_max_freq_ratio(global.turbo_disabled);
1201		for_each_possible_cpu(cpu)
1202			intel_pstate_update_max_freq(cpu);
1203	} else {
1204		cpufreq_update_policy(cpu);
1205	}
1206
1207	mutex_unlock(&intel_pstate_driver_lock);
1208}
1209
1210/************************** sysfs begin ************************/
1211#define show_one(file_name, object)					\
1212	static ssize_t show_##file_name					\
1213	(struct kobject *kobj, struct kobj_attribute *attr, char *buf)	\
1214	{								\
1215		return sprintf(buf, "%u\n", global.object);		\
1216	}
1217
1218static ssize_t intel_pstate_show_status(char *buf);
1219static int intel_pstate_update_status(const char *buf, size_t size);
1220
1221static ssize_t show_status(struct kobject *kobj,
1222			   struct kobj_attribute *attr, char *buf)
1223{
1224	ssize_t ret;
1225
1226	mutex_lock(&intel_pstate_driver_lock);
1227	ret = intel_pstate_show_status(buf);
1228	mutex_unlock(&intel_pstate_driver_lock);
1229
1230	return ret;
1231}
1232
1233static ssize_t store_status(struct kobject *a, struct kobj_attribute *b,
1234			    const char *buf, size_t count)
1235{
1236	char *p = memchr(buf, '\n', count);
1237	int ret;
1238
1239	mutex_lock(&intel_pstate_driver_lock);
1240	ret = intel_pstate_update_status(buf, p ? p - buf : count);
1241	mutex_unlock(&intel_pstate_driver_lock);
1242
1243	return ret < 0 ? ret : count;
1244}
1245
1246static ssize_t show_turbo_pct(struct kobject *kobj,
1247				struct kobj_attribute *attr, char *buf)
1248{
1249	struct cpudata *cpu;
1250	int total, no_turbo, turbo_pct;
1251	uint32_t turbo_fp;
1252
1253	mutex_lock(&intel_pstate_driver_lock);
1254
1255	if (!intel_pstate_driver) {
1256		mutex_unlock(&intel_pstate_driver_lock);
1257		return -EAGAIN;
1258	}
1259
1260	cpu = all_cpu_data[0];
1261
1262	total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
1263	no_turbo = cpu->pstate.max_pstate - cpu->pstate.min_pstate + 1;
1264	turbo_fp = div_fp(no_turbo, total);
1265	turbo_pct = 100 - fp_toint(mul_fp(turbo_fp, int_tofp(100)));
1266
1267	mutex_unlock(&intel_pstate_driver_lock);
1268
1269	return sprintf(buf, "%u\n", turbo_pct);
1270}
1271
1272static ssize_t show_num_pstates(struct kobject *kobj,
1273				struct kobj_attribute *attr, char *buf)
1274{
1275	struct cpudata *cpu;
1276	int total;
1277
1278	mutex_lock(&intel_pstate_driver_lock);
1279
1280	if (!intel_pstate_driver) {
1281		mutex_unlock(&intel_pstate_driver_lock);
1282		return -EAGAIN;
1283	}
1284
1285	cpu = all_cpu_data[0];
1286	total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
1287
1288	mutex_unlock(&intel_pstate_driver_lock);
1289
1290	return sprintf(buf, "%u\n", total);
1291}
1292
1293static ssize_t show_no_turbo(struct kobject *kobj,
1294			     struct kobj_attribute *attr, char *buf)
1295{
1296	ssize_t ret;
1297
1298	mutex_lock(&intel_pstate_driver_lock);
1299
1300	if (!intel_pstate_driver) {
1301		mutex_unlock(&intel_pstate_driver_lock);
1302		return -EAGAIN;
1303	}
1304
1305	update_turbo_state();
1306	if (global.turbo_disabled)
1307		ret = sprintf(buf, "%u\n", global.turbo_disabled);
1308	else
1309		ret = sprintf(buf, "%u\n", global.no_turbo);
1310
1311	mutex_unlock(&intel_pstate_driver_lock);
1312
1313	return ret;
1314}
1315
1316static ssize_t store_no_turbo(struct kobject *a, struct kobj_attribute *b,
1317			      const char *buf, size_t count)
1318{
1319	unsigned int input;
1320	int ret;
1321
1322	ret = sscanf(buf, "%u", &input);
1323	if (ret != 1)
1324		return -EINVAL;
1325
1326	mutex_lock(&intel_pstate_driver_lock);
1327
1328	if (!intel_pstate_driver) {
1329		mutex_unlock(&intel_pstate_driver_lock);
1330		return -EAGAIN;
1331	}
1332
1333	mutex_lock(&intel_pstate_limits_lock);
1334
1335	update_turbo_state();
1336	if (global.turbo_disabled) {
1337		pr_notice_once("Turbo disabled by BIOS or unavailable on processor\n");
1338		mutex_unlock(&intel_pstate_limits_lock);
1339		mutex_unlock(&intel_pstate_driver_lock);
1340		return -EPERM;
1341	}
1342
1343	global.no_turbo = clamp_t(int, input, 0, 1);
1344
1345	if (global.no_turbo) {
1346		struct cpudata *cpu = all_cpu_data[0];
1347		int pct = cpu->pstate.max_pstate * 100 / cpu->pstate.turbo_pstate;
1348
1349		/* Squash the global minimum into the permitted range. */
1350		if (global.min_perf_pct > pct)
1351			global.min_perf_pct = pct;
1352	}
1353
1354	mutex_unlock(&intel_pstate_limits_lock);
1355
1356	intel_pstate_update_policies();
1357
1358	mutex_unlock(&intel_pstate_driver_lock);
1359
1360	return count;
1361}
1362
1363static void update_qos_request(enum freq_qos_req_type type)
1364{
1365	struct freq_qos_request *req;
1366	struct cpufreq_policy *policy;
1367	int i;
1368
1369	for_each_possible_cpu(i) {
1370		struct cpudata *cpu = all_cpu_data[i];
1371		unsigned int freq, perf_pct;
1372
1373		policy = cpufreq_cpu_get(i);
1374		if (!policy)
1375			continue;
1376
1377		req = policy->driver_data;
1378		cpufreq_cpu_put(policy);
1379
1380		if (!req)
1381			continue;
1382
1383		if (hwp_active)
1384			intel_pstate_get_hwp_cap(cpu);
1385
1386		if (type == FREQ_QOS_MIN) {
1387			perf_pct = global.min_perf_pct;
1388		} else {
1389			req++;
1390			perf_pct = global.max_perf_pct;
1391		}
1392
1393		freq = DIV_ROUND_UP(cpu->pstate.turbo_freq * perf_pct, 100);
1394
1395		if (freq_qos_update_request(req, freq) < 0)
1396			pr_warn("Failed to update freq constraint: CPU%d\n", i);
1397	}
1398}
1399
1400static ssize_t store_max_perf_pct(struct kobject *a, struct kobj_attribute *b,
1401				  const char *buf, size_t count)
1402{
1403	unsigned int input;
1404	int ret;
1405
1406	ret = sscanf(buf, "%u", &input);
1407	if (ret != 1)
1408		return -EINVAL;
1409
1410	mutex_lock(&intel_pstate_driver_lock);
1411
1412	if (!intel_pstate_driver) {
1413		mutex_unlock(&intel_pstate_driver_lock);
1414		return -EAGAIN;
1415	}
1416
1417	mutex_lock(&intel_pstate_limits_lock);
1418
1419	global.max_perf_pct = clamp_t(int, input, global.min_perf_pct, 100);
1420
1421	mutex_unlock(&intel_pstate_limits_lock);
1422
1423	if (intel_pstate_driver == &intel_pstate)
1424		intel_pstate_update_policies();
1425	else
1426		update_qos_request(FREQ_QOS_MAX);
1427
1428	mutex_unlock(&intel_pstate_driver_lock);
1429
1430	return count;
1431}
1432
1433static ssize_t store_min_perf_pct(struct kobject *a, struct kobj_attribute *b,
1434				  const char *buf, size_t count)
1435{
1436	unsigned int input;
1437	int ret;
1438
1439	ret = sscanf(buf, "%u", &input);
1440	if (ret != 1)
1441		return -EINVAL;
1442
1443	mutex_lock(&intel_pstate_driver_lock);
1444
1445	if (!intel_pstate_driver) {
1446		mutex_unlock(&intel_pstate_driver_lock);
1447		return -EAGAIN;
1448	}
1449
1450	mutex_lock(&intel_pstate_limits_lock);
1451
1452	global.min_perf_pct = clamp_t(int, input,
1453				      min_perf_pct_min(), global.max_perf_pct);
1454
1455	mutex_unlock(&intel_pstate_limits_lock);
1456
1457	if (intel_pstate_driver == &intel_pstate)
1458		intel_pstate_update_policies();
1459	else
1460		update_qos_request(FREQ_QOS_MIN);
1461
1462	mutex_unlock(&intel_pstate_driver_lock);
1463
1464	return count;
1465}
1466
1467static ssize_t show_hwp_dynamic_boost(struct kobject *kobj,
1468				struct kobj_attribute *attr, char *buf)
1469{
1470	return sprintf(buf, "%u\n", hwp_boost);
1471}
1472
1473static ssize_t store_hwp_dynamic_boost(struct kobject *a,
1474				       struct kobj_attribute *b,
1475				       const char *buf, size_t count)
1476{
1477	unsigned int input;
1478	int ret;
1479
1480	ret = kstrtouint(buf, 10, &input);
1481	if (ret)
1482		return ret;
1483
1484	mutex_lock(&intel_pstate_driver_lock);
1485	hwp_boost = !!input;
1486	intel_pstate_update_policies();
1487	mutex_unlock(&intel_pstate_driver_lock);
1488
1489	return count;
1490}
1491
1492static ssize_t show_energy_efficiency(struct kobject *kobj, struct kobj_attribute *attr,
1493				      char *buf)
1494{
1495	u64 power_ctl;
1496	int enable;
1497
1498	rdmsrl(MSR_IA32_POWER_CTL, power_ctl);
1499	enable = !!(power_ctl & BIT(MSR_IA32_POWER_CTL_BIT_EE));
1500	return sprintf(buf, "%d\n", !enable);
1501}
1502
1503static ssize_t store_energy_efficiency(struct kobject *a, struct kobj_attribute *b,
1504				       const char *buf, size_t count)
1505{
1506	bool input;
1507	int ret;
1508
1509	ret = kstrtobool(buf, &input);
1510	if (ret)
1511		return ret;
1512
1513	set_power_ctl_ee_state(input);
1514
1515	return count;
1516}
1517
1518show_one(max_perf_pct, max_perf_pct);
1519show_one(min_perf_pct, min_perf_pct);
1520
1521define_one_global_rw(status);
1522define_one_global_rw(no_turbo);
1523define_one_global_rw(max_perf_pct);
1524define_one_global_rw(min_perf_pct);
1525define_one_global_ro(turbo_pct);
1526define_one_global_ro(num_pstates);
1527define_one_global_rw(hwp_dynamic_boost);
1528define_one_global_rw(energy_efficiency);
1529
1530static struct attribute *intel_pstate_attributes[] = {
1531	&status.attr,
1532	&no_turbo.attr,
1533	NULL
1534};
1535
1536static const struct attribute_group intel_pstate_attr_group = {
1537	.attrs = intel_pstate_attributes,
1538};
1539
1540static const struct x86_cpu_id intel_pstate_cpu_ee_disable_ids[];
1541
1542static struct kobject *intel_pstate_kobject;
1543
1544static void __init intel_pstate_sysfs_expose_params(void)
1545{
1546	int rc;
1547
1548	intel_pstate_kobject = kobject_create_and_add("intel_pstate",
1549						&cpu_subsys.dev_root->kobj);
1550	if (WARN_ON(!intel_pstate_kobject))
1551		return;
1552
1553	rc = sysfs_create_group(intel_pstate_kobject, &intel_pstate_attr_group);
1554	if (WARN_ON(rc))
1555		return;
1556
1557	if (!boot_cpu_has(X86_FEATURE_HYBRID_CPU)) {
1558		rc = sysfs_create_file(intel_pstate_kobject, &turbo_pct.attr);
1559		WARN_ON(rc);
1560
1561		rc = sysfs_create_file(intel_pstate_kobject, &num_pstates.attr);
1562		WARN_ON(rc);
1563	}
1564
1565	/*
1566	 * If per cpu limits are enforced there are no global limits, so
1567	 * return without creating max/min_perf_pct attributes
1568	 */
1569	if (per_cpu_limits)
1570		return;
1571
1572	rc = sysfs_create_file(intel_pstate_kobject, &max_perf_pct.attr);
1573	WARN_ON(rc);
1574
1575	rc = sysfs_create_file(intel_pstate_kobject, &min_perf_pct.attr);
1576	WARN_ON(rc);
1577
1578	if (x86_match_cpu(intel_pstate_cpu_ee_disable_ids)) {
1579		rc = sysfs_create_file(intel_pstate_kobject, &energy_efficiency.attr);
1580		WARN_ON(rc);
1581	}
1582}
1583
1584static void __init intel_pstate_sysfs_remove(void)
1585{
1586	if (!intel_pstate_kobject)
1587		return;
1588
1589	sysfs_remove_group(intel_pstate_kobject, &intel_pstate_attr_group);
1590
1591	if (!boot_cpu_has(X86_FEATURE_HYBRID_CPU)) {
1592		sysfs_remove_file(intel_pstate_kobject, &num_pstates.attr);
1593		sysfs_remove_file(intel_pstate_kobject, &turbo_pct.attr);
1594	}
1595
1596	if (!per_cpu_limits) {
1597		sysfs_remove_file(intel_pstate_kobject, &max_perf_pct.attr);
1598		sysfs_remove_file(intel_pstate_kobject, &min_perf_pct.attr);
1599
1600		if (x86_match_cpu(intel_pstate_cpu_ee_disable_ids))
1601			sysfs_remove_file(intel_pstate_kobject, &energy_efficiency.attr);
1602	}
1603
1604	kobject_put(intel_pstate_kobject);
1605}
1606
1607static void intel_pstate_sysfs_expose_hwp_dynamic_boost(void)
1608{
1609	int rc;
1610
1611	if (!hwp_active)
1612		return;
1613
1614	rc = sysfs_create_file(intel_pstate_kobject, &hwp_dynamic_boost.attr);
1615	WARN_ON_ONCE(rc);
1616}
1617
1618static void intel_pstate_sysfs_hide_hwp_dynamic_boost(void)
1619{
1620	if (!hwp_active)
1621		return;
1622
1623	sysfs_remove_file(intel_pstate_kobject, &hwp_dynamic_boost.attr);
1624}
1625
1626/************************** sysfs end ************************/
1627
1628static void intel_pstate_hwp_enable(struct cpudata *cpudata)
1629{
1630	/* First disable HWP notification interrupt as we don't process them */
1631	if (boot_cpu_has(X86_FEATURE_HWP_NOTIFY))
1632		wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_INTERRUPT, 0x00);
1633
1634	wrmsrl_on_cpu(cpudata->cpu, MSR_PM_ENABLE, 0x1);
1635	if (cpudata->epp_default == -EINVAL)
1636		cpudata->epp_default = intel_pstate_get_epp(cpudata, 0);
1637}
1638
1639static int atom_get_min_pstate(void)
1640{
1641	u64 value;
1642
1643	rdmsrl(MSR_ATOM_CORE_RATIOS, value);
1644	return (value >> 8) & 0x7F;
1645}
1646
1647static int atom_get_max_pstate(void)
1648{
1649	u64 value;
1650
1651	rdmsrl(MSR_ATOM_CORE_RATIOS, value);
1652	return (value >> 16) & 0x7F;
1653}
1654
1655static int atom_get_turbo_pstate(void)
1656{
1657	u64 value;
1658
1659	rdmsrl(MSR_ATOM_CORE_TURBO_RATIOS, value);
1660	return value & 0x7F;
1661}
1662
1663static u64 atom_get_val(struct cpudata *cpudata, int pstate)
1664{
1665	u64 val;
1666	int32_t vid_fp;
1667	u32 vid;
1668
1669	val = (u64)pstate << 8;
1670	if (global.no_turbo && !global.turbo_disabled)
1671		val |= (u64)1 << 32;
1672
1673	vid_fp = cpudata->vid.min + mul_fp(
1674		int_tofp(pstate - cpudata->pstate.min_pstate),
1675		cpudata->vid.ratio);
1676
1677	vid_fp = clamp_t(int32_t, vid_fp, cpudata->vid.min, cpudata->vid.max);
1678	vid = ceiling_fp(vid_fp);
1679
1680	if (pstate > cpudata->pstate.max_pstate)
1681		vid = cpudata->vid.turbo;
1682
1683	return val | vid;
1684}
1685
1686static int silvermont_get_scaling(void)
1687{
1688	u64 value;
1689	int i;
1690	/* Defined in Table 35-6 from SDM (Sept 2015) */
1691	static int silvermont_freq_table[] = {
1692		83300, 100000, 133300, 116700, 80000};
1693
1694	rdmsrl(MSR_FSB_FREQ, value);
1695	i = value & 0x7;
1696	WARN_ON(i > 4);
1697
1698	return silvermont_freq_table[i];
1699}
1700
1701static int airmont_get_scaling(void)
1702{
1703	u64 value;
1704	int i;
1705	/* Defined in Table 35-10 from SDM (Sept 2015) */
1706	static int airmont_freq_table[] = {
1707		83300, 100000, 133300, 116700, 80000,
1708		93300, 90000, 88900, 87500};
1709
1710	rdmsrl(MSR_FSB_FREQ, value);
1711	i = value & 0xF;
1712	WARN_ON(i > 8);
1713
1714	return airmont_freq_table[i];
1715}
1716
1717static void atom_get_vid(struct cpudata *cpudata)
1718{
1719	u64 value;
1720
1721	rdmsrl(MSR_ATOM_CORE_VIDS, value);
1722	cpudata->vid.min = int_tofp((value >> 8) & 0x7f);
1723	cpudata->vid.max = int_tofp((value >> 16) & 0x7f);
1724	cpudata->vid.ratio = div_fp(
1725		cpudata->vid.max - cpudata->vid.min,
1726		int_tofp(cpudata->pstate.max_pstate -
1727			cpudata->pstate.min_pstate));
1728
1729	rdmsrl(MSR_ATOM_CORE_TURBO_VIDS, value);
1730	cpudata->vid.turbo = value & 0x7f;
1731}
1732
1733static int core_get_min_pstate(void)
1734{
1735	u64 value;
1736
1737	rdmsrl(MSR_PLATFORM_INFO, value);
1738	return (value >> 40) & 0xFF;
1739}
1740
1741static int core_get_max_pstate_physical(void)
1742{
1743	u64 value;
1744
1745	rdmsrl(MSR_PLATFORM_INFO, value);
1746	return (value >> 8) & 0xFF;
1747}
1748
1749static int core_get_tdp_ratio(u64 plat_info)
1750{
1751	/* Check how many TDP levels present */
1752	if (plat_info & 0x600000000) {
1753		u64 tdp_ctrl;
1754		u64 tdp_ratio;
1755		int tdp_msr;
1756		int err;
1757
1758		/* Get the TDP level (0, 1, 2) to get ratios */
1759		err = rdmsrl_safe(MSR_CONFIG_TDP_CONTROL, &tdp_ctrl);
1760		if (err)
1761			return err;
1762
1763		/* TDP MSR are continuous starting at 0x648 */
1764		tdp_msr = MSR_CONFIG_TDP_NOMINAL + (tdp_ctrl & 0x03);
1765		err = rdmsrl_safe(tdp_msr, &tdp_ratio);
1766		if (err)
1767			return err;
1768
1769		/* For level 1 and 2, bits[23:16] contain the ratio */
1770		if (tdp_ctrl & 0x03)
1771			tdp_ratio >>= 16;
1772
1773		tdp_ratio &= 0xff; /* ratios are only 8 bits long */
1774		pr_debug("tdp_ratio %x\n", (int)tdp_ratio);
1775
1776		return (int)tdp_ratio;
1777	}
1778
1779	return -ENXIO;
1780}
1781
1782static int core_get_max_pstate(void)
1783{
1784	u64 tar;
1785	u64 plat_info;
1786	int max_pstate;
1787	int tdp_ratio;
1788	int err;
1789
1790	rdmsrl(MSR_PLATFORM_INFO, plat_info);
1791	max_pstate = (plat_info >> 8) & 0xFF;
1792
1793	tdp_ratio = core_get_tdp_ratio(plat_info);
1794	if (tdp_ratio <= 0)
1795		return max_pstate;
1796
1797	if (hwp_active) {
1798		/* Turbo activation ratio is not used on HWP platforms */
1799		return tdp_ratio;
1800	}
1801
1802	err = rdmsrl_safe(MSR_TURBO_ACTIVATION_RATIO, &tar);
1803	if (!err) {
1804		int tar_levels;
1805
1806		/* Do some sanity checking for safety */
1807		tar_levels = tar & 0xff;
1808		if (tdp_ratio - 1 == tar_levels) {
1809			max_pstate = tar_levels;
1810			pr_debug("max_pstate=TAC %x\n", max_pstate);
1811		}
1812	}
1813
1814	return max_pstate;
1815}
1816
1817static int core_get_turbo_pstate(void)
1818{
1819	u64 value;
1820	int nont, ret;
1821
1822	rdmsrl(MSR_TURBO_RATIO_LIMIT, value);
1823	nont = core_get_max_pstate();
1824	ret = (value) & 255;
1825	if (ret <= nont)
1826		ret = nont;
1827	return ret;
1828}
1829
1830static inline int core_get_scaling(void)
1831{
1832	return 100000;
1833}
1834
1835static u64 core_get_val(struct cpudata *cpudata, int pstate)
1836{
1837	u64 val;
1838
1839	val = (u64)pstate << 8;
1840	if (global.no_turbo && !global.turbo_disabled)
1841		val |= (u64)1 << 32;
1842
1843	return val;
1844}
1845
1846static int knl_get_aperf_mperf_shift(void)
1847{
1848	return 10;
1849}
1850
1851static int knl_get_turbo_pstate(void)
1852{
1853	u64 value;
1854	int nont, ret;
1855
1856	rdmsrl(MSR_TURBO_RATIO_LIMIT, value);
1857	nont = core_get_max_pstate();
1858	ret = (((value) >> 8) & 0xFF);
1859	if (ret <= nont)
1860		ret = nont;
1861	return ret;
1862}
1863
1864static void intel_pstate_set_pstate(struct cpudata *cpu, int pstate)
1865{
1866	trace_cpu_frequency(pstate * cpu->pstate.scaling, cpu->cpu);
1867	cpu->pstate.current_pstate = pstate;
1868	/*
1869	 * Generally, there is no guarantee that this code will always run on
1870	 * the CPU being updated, so force the register update to run on the
1871	 * right CPU.
1872	 */
1873	wrmsrl_on_cpu(cpu->cpu, MSR_IA32_PERF_CTL,
1874		      pstate_funcs.get_val(cpu, pstate));
1875}
1876
1877static void intel_pstate_set_min_pstate(struct cpudata *cpu)
1878{
1879	intel_pstate_set_pstate(cpu, cpu->pstate.min_pstate);
1880}
1881
1882static void intel_pstate_max_within_limits(struct cpudata *cpu)
1883{
1884	int pstate = max(cpu->pstate.min_pstate, cpu->max_perf_ratio);
1885
1886	update_turbo_state();
1887	intel_pstate_set_pstate(cpu, pstate);
1888}
1889
1890static void intel_pstate_get_cpu_pstates(struct cpudata *cpu)
1891{
1892	bool hybrid_cpu = boot_cpu_has(X86_FEATURE_HYBRID_CPU);
1893	int perf_ctl_max_phys = pstate_funcs.get_max_physical();
1894	int perf_ctl_scaling = hybrid_cpu ? cpu_khz / perf_ctl_max_phys :
1895					    pstate_funcs.get_scaling();
1896
1897	cpu->pstate.min_pstate = pstate_funcs.get_min();
1898	cpu->pstate.max_pstate_physical = perf_ctl_max_phys;
1899	cpu->pstate.perf_ctl_scaling = perf_ctl_scaling;
1900
1901	if (hwp_active && !hwp_mode_bdw) {
1902		__intel_pstate_get_hwp_cap(cpu);
1903
1904		if (hybrid_cpu)
1905			intel_pstate_hybrid_hwp_calibrate(cpu);
1906		else
1907			cpu->pstate.scaling = perf_ctl_scaling;
1908	} else {
1909		cpu->pstate.scaling = perf_ctl_scaling;
1910		cpu->pstate.max_pstate = pstate_funcs.get_max();
1911		cpu->pstate.turbo_pstate = pstate_funcs.get_turbo();
1912	}
1913
1914	if (cpu->pstate.scaling == perf_ctl_scaling) {
1915		cpu->pstate.min_freq = cpu->pstate.min_pstate * perf_ctl_scaling;
1916		cpu->pstate.max_freq = cpu->pstate.max_pstate * perf_ctl_scaling;
1917		cpu->pstate.turbo_freq = cpu->pstate.turbo_pstate * perf_ctl_scaling;
1918	}
1919
1920	if (pstate_funcs.get_aperf_mperf_shift)
1921		cpu->aperf_mperf_shift = pstate_funcs.get_aperf_mperf_shift();
1922
1923	if (pstate_funcs.get_vid)
1924		pstate_funcs.get_vid(cpu);
1925
1926	intel_pstate_set_min_pstate(cpu);
1927}
1928
1929/*
1930 * Long hold time will keep high perf limits for long time,
1931 * which negatively impacts perf/watt for some workloads,
1932 * like specpower. 3ms is based on experiements on some
1933 * workoads.
1934 */
1935static int hwp_boost_hold_time_ns = 3 * NSEC_PER_MSEC;
1936
1937static inline void intel_pstate_hwp_boost_up(struct cpudata *cpu)
1938{
1939	u64 hwp_req = READ_ONCE(cpu->hwp_req_cached);
1940	u64 hwp_cap = READ_ONCE(cpu->hwp_cap_cached);
1941	u32 max_limit = (hwp_req & 0xff00) >> 8;
1942	u32 min_limit = (hwp_req & 0xff);
1943	u32 boost_level1;
1944
1945	/*
1946	 * Cases to consider (User changes via sysfs or boot time):
1947	 * If, P0 (Turbo max) = P1 (Guaranteed max) = min:
1948	 *	No boost, return.
1949	 * If, P0 (Turbo max) > P1 (Guaranteed max) = min:
1950	 *     Should result in one level boost only for P0.
1951	 * If, P0 (Turbo max) = P1 (Guaranteed max) > min:
1952	 *     Should result in two level boost:
1953	 *         (min + p1)/2 and P1.
1954	 * If, P0 (Turbo max) > P1 (Guaranteed max) > min:
1955	 *     Should result in three level boost:
1956	 *        (min + p1)/2, P1 and P0.
1957	 */
1958
1959	/* If max and min are equal or already at max, nothing to boost */
1960	if (max_limit == min_limit || cpu->hwp_boost_min >= max_limit)
1961		return;
1962
1963	if (!cpu->hwp_boost_min)
1964		cpu->hwp_boost_min = min_limit;
1965
1966	/* level at half way mark between min and guranteed */
1967	boost_level1 = (HWP_GUARANTEED_PERF(hwp_cap) + min_limit) >> 1;
1968
1969	if (cpu->hwp_boost_min < boost_level1)
1970		cpu->hwp_boost_min = boost_level1;
1971	else if (cpu->hwp_boost_min < HWP_GUARANTEED_PERF(hwp_cap))
1972		cpu->hwp_boost_min = HWP_GUARANTEED_PERF(hwp_cap);
1973	else if (cpu->hwp_boost_min == HWP_GUARANTEED_PERF(hwp_cap) &&
1974		 max_limit != HWP_GUARANTEED_PERF(hwp_cap))
1975		cpu->hwp_boost_min = max_limit;
1976	else
1977		return;
1978
1979	hwp_req = (hwp_req & ~GENMASK_ULL(7, 0)) | cpu->hwp_boost_min;
1980	wrmsrl(MSR_HWP_REQUEST, hwp_req);
1981	cpu->last_update = cpu->sample.time;
1982}
1983
1984static inline void intel_pstate_hwp_boost_down(struct cpudata *cpu)
1985{
1986	if (cpu->hwp_boost_min) {
1987		bool expired;
1988
1989		/* Check if we are idle for hold time to boost down */
1990		expired = time_after64(cpu->sample.time, cpu->last_update +
1991				       hwp_boost_hold_time_ns);
1992		if (expired) {
1993			wrmsrl(MSR_HWP_REQUEST, cpu->hwp_req_cached);
1994			cpu->hwp_boost_min = 0;
1995		}
1996	}
1997	cpu->last_update = cpu->sample.time;
1998}
1999
2000static inline void intel_pstate_update_util_hwp_local(struct cpudata *cpu,
2001						      u64 time)
2002{
2003	cpu->sample.time = time;
2004
2005	if (cpu->sched_flags & SCHED_CPUFREQ_IOWAIT) {
2006		bool do_io = false;
2007
2008		cpu->sched_flags = 0;
2009		/*
2010		 * Set iowait_boost flag and update time. Since IO WAIT flag
2011		 * is set all the time, we can't just conclude that there is
2012		 * some IO bound activity is scheduled on this CPU with just
2013		 * one occurrence. If we receive at least two in two
2014		 * consecutive ticks, then we treat as boost candidate.
2015		 */
2016		if (time_before64(time, cpu->last_io_update + 2 * TICK_NSEC))
2017			do_io = true;
2018
2019		cpu->last_io_update = time;
2020
2021		if (do_io)
2022			intel_pstate_hwp_boost_up(cpu);
2023
2024	} else {
2025		intel_pstate_hwp_boost_down(cpu);
2026	}
2027}
2028
2029static inline void intel_pstate_update_util_hwp(struct update_util_data *data,
2030						u64 time, unsigned int flags)
2031{
2032	struct cpudata *cpu = container_of(data, struct cpudata, update_util);
2033
2034	cpu->sched_flags |= flags;
2035
2036	if (smp_processor_id() == cpu->cpu)
2037		intel_pstate_update_util_hwp_local(cpu, time);
2038}
2039
2040static inline void intel_pstate_calc_avg_perf(struct cpudata *cpu)
2041{
2042	struct sample *sample = &cpu->sample;
2043
2044	sample->core_avg_perf = div_ext_fp(sample->aperf, sample->mperf);
2045}
2046
2047static inline bool intel_pstate_sample(struct cpudata *cpu, u64 time)
2048{
2049	u64 aperf, mperf;
2050	unsigned long flags;
2051	u64 tsc;
2052
2053	local_irq_save(flags);
2054	rdmsrl(MSR_IA32_APERF, aperf);
2055	rdmsrl(MSR_IA32_MPERF, mperf);
2056	tsc = rdtsc();
2057	if (cpu->prev_mperf == mperf || cpu->prev_tsc == tsc) {
2058		local_irq_restore(flags);
2059		return false;
2060	}
2061	local_irq_restore(flags);
2062
2063	cpu->last_sample_time = cpu->sample.time;
2064	cpu->sample.time = time;
2065	cpu->sample.aperf = aperf;
2066	cpu->sample.mperf = mperf;
2067	cpu->sample.tsc =  tsc;
2068	cpu->sample.aperf -= cpu->prev_aperf;
2069	cpu->sample.mperf -= cpu->prev_mperf;
2070	cpu->sample.tsc -= cpu->prev_tsc;
2071
2072	cpu->prev_aperf = aperf;
2073	cpu->prev_mperf = mperf;
2074	cpu->prev_tsc = tsc;
2075	/*
2076	 * First time this function is invoked in a given cycle, all of the
2077	 * previous sample data fields are equal to zero or stale and they must
2078	 * be populated with meaningful numbers for things to work, so assume
2079	 * that sample.time will always be reset before setting the utilization
2080	 * update hook and make the caller skip the sample then.
2081	 */
2082	if (cpu->last_sample_time) {
2083		intel_pstate_calc_avg_perf(cpu);
2084		return true;
2085	}
2086	return false;
2087}
2088
2089static inline int32_t get_avg_frequency(struct cpudata *cpu)
2090{
2091	return mul_ext_fp(cpu->sample.core_avg_perf, cpu_khz);
2092}
2093
2094static inline int32_t get_avg_pstate(struct cpudata *cpu)
2095{
2096	return mul_ext_fp(cpu->pstate.max_pstate_physical,
2097			  cpu->sample.core_avg_perf);
2098}
2099
2100static inline int32_t get_target_pstate(struct cpudata *cpu)
2101{
2102	struct sample *sample = &cpu->sample;
2103	int32_t busy_frac;
2104	int target, avg_pstate;
2105
2106	busy_frac = div_fp(sample->mperf << cpu->aperf_mperf_shift,
2107			   sample->tsc);
2108
2109	if (busy_frac < cpu->iowait_boost)
2110		busy_frac = cpu->iowait_boost;
2111
2112	sample->busy_scaled = busy_frac * 100;
2113
2114	target = global.no_turbo || global.turbo_disabled ?
2115			cpu->pstate.max_pstate : cpu->pstate.turbo_pstate;
2116	target += target >> 2;
2117	target = mul_fp(target, busy_frac);
2118	if (target < cpu->pstate.min_pstate)
2119		target = cpu->pstate.min_pstate;
2120
2121	/*
2122	 * If the average P-state during the previous cycle was higher than the
2123	 * current target, add 50% of the difference to the target to reduce
2124	 * possible performance oscillations and offset possible performance
2125	 * loss related to moving the workload from one CPU to another within
2126	 * a package/module.
2127	 */
2128	avg_pstate = get_avg_pstate(cpu);
2129	if (avg_pstate > target)
2130		target += (avg_pstate - target) >> 1;
2131
2132	return target;
2133}
2134
2135static int intel_pstate_prepare_request(struct cpudata *cpu, int pstate)
2136{
2137	int min_pstate = max(cpu->pstate.min_pstate, cpu->min_perf_ratio);
2138	int max_pstate = max(min_pstate, cpu->max_perf_ratio);
2139
2140	return clamp_t(int, pstate, min_pstate, max_pstate);
2141}
2142
2143static void intel_pstate_update_pstate(struct cpudata *cpu, int pstate)
2144{
2145	if (pstate == cpu->pstate.current_pstate)
2146		return;
2147
2148	cpu->pstate.current_pstate = pstate;
2149	wrmsrl(MSR_IA32_PERF_CTL, pstate_funcs.get_val(cpu, pstate));
2150}
2151
2152static void intel_pstate_adjust_pstate(struct cpudata *cpu)
2153{
2154	int from = cpu->pstate.current_pstate;
2155	struct sample *sample;
2156	int target_pstate;
2157
2158	update_turbo_state();
2159
2160	target_pstate = get_target_pstate(cpu);
2161	target_pstate = intel_pstate_prepare_request(cpu, target_pstate);
2162	trace_cpu_frequency(target_pstate * cpu->pstate.scaling, cpu->cpu);
2163	intel_pstate_update_pstate(cpu, target_pstate);
2164
2165	sample = &cpu->sample;
2166	trace_pstate_sample(mul_ext_fp(100, sample->core_avg_perf),
2167		fp_toint(sample->busy_scaled),
2168		from,
2169		cpu->pstate.current_pstate,
2170		sample->mperf,
2171		sample->aperf,
2172		sample->tsc,
2173		get_avg_frequency(cpu),
2174		fp_toint(cpu->iowait_boost * 100));
2175}
2176
2177static void intel_pstate_update_util(struct update_util_data *data, u64 time,
2178				     unsigned int flags)
2179{
2180	struct cpudata *cpu = container_of(data, struct cpudata, update_util);
2181	u64 delta_ns;
2182
2183	/* Don't allow remote callbacks */
2184	if (smp_processor_id() != cpu->cpu)
2185		return;
2186
2187	delta_ns = time - cpu->last_update;
2188	if (flags & SCHED_CPUFREQ_IOWAIT) {
2189		/* Start over if the CPU may have been idle. */
2190		if (delta_ns > TICK_NSEC) {
2191			cpu->iowait_boost = ONE_EIGHTH_FP;
2192		} else if (cpu->iowait_boost >= ONE_EIGHTH_FP) {
2193			cpu->iowait_boost <<= 1;
2194			if (cpu->iowait_boost > int_tofp(1))
2195				cpu->iowait_boost = int_tofp(1);
2196		} else {
2197			cpu->iowait_boost = ONE_EIGHTH_FP;
2198		}
2199	} else if (cpu->iowait_boost) {
2200		/* Clear iowait_boost if the CPU may have been idle. */
2201		if (delta_ns > TICK_NSEC)
2202			cpu->iowait_boost = 0;
2203		else
2204			cpu->iowait_boost >>= 1;
2205	}
2206	cpu->last_update = time;
2207	delta_ns = time - cpu->sample.time;
2208	if ((s64)delta_ns < INTEL_PSTATE_SAMPLING_INTERVAL)
2209		return;
2210
2211	if (intel_pstate_sample(cpu, time))
2212		intel_pstate_adjust_pstate(cpu);
2213}
2214
2215static struct pstate_funcs core_funcs = {
2216	.get_max = core_get_max_pstate,
2217	.get_max_physical = core_get_max_pstate_physical,
2218	.get_min = core_get_min_pstate,
2219	.get_turbo = core_get_turbo_pstate,
2220	.get_scaling = core_get_scaling,
2221	.get_val = core_get_val,
2222};
2223
2224static const struct pstate_funcs silvermont_funcs = {
2225	.get_max = atom_get_max_pstate,
2226	.get_max_physical = atom_get_max_pstate,
2227	.get_min = atom_get_min_pstate,
2228	.get_turbo = atom_get_turbo_pstate,
2229	.get_val = atom_get_val,
2230	.get_scaling = silvermont_get_scaling,
2231	.get_vid = atom_get_vid,
2232};
2233
2234static const struct pstate_funcs airmont_funcs = {
2235	.get_max = atom_get_max_pstate,
2236	.get_max_physical = atom_get_max_pstate,
2237	.get_min = atom_get_min_pstate,
2238	.get_turbo = atom_get_turbo_pstate,
2239	.get_val = atom_get_val,
2240	.get_scaling = airmont_get_scaling,
2241	.get_vid = atom_get_vid,
2242};
2243
2244static const struct pstate_funcs knl_funcs = {
2245	.get_max = core_get_max_pstate,
2246	.get_max_physical = core_get_max_pstate_physical,
2247	.get_min = core_get_min_pstate,
2248	.get_turbo = knl_get_turbo_pstate,
2249	.get_aperf_mperf_shift = knl_get_aperf_mperf_shift,
2250	.get_scaling = core_get_scaling,
2251	.get_val = core_get_val,
2252};
2253
2254#define X86_MATCH(model, policy)					 \
2255	X86_MATCH_VENDOR_FAM_MODEL_FEATURE(INTEL, 6, INTEL_FAM6_##model, \
2256					   X86_FEATURE_APERFMPERF, &policy)
2257
2258static const struct x86_cpu_id intel_pstate_cpu_ids[] = {
2259	X86_MATCH(SANDYBRIDGE,		core_funcs),
2260	X86_MATCH(SANDYBRIDGE_X,	core_funcs),
2261	X86_MATCH(ATOM_SILVERMONT,	silvermont_funcs),
2262	X86_MATCH(IVYBRIDGE,		core_funcs),
2263	X86_MATCH(HASWELL,		core_funcs),
2264	X86_MATCH(BROADWELL,		core_funcs),
2265	X86_MATCH(IVYBRIDGE_X,		core_funcs),
2266	X86_MATCH(HASWELL_X,		core_funcs),
2267	X86_MATCH(HASWELL_L,		core_funcs),
2268	X86_MATCH(HASWELL_G,		core_funcs),
2269	X86_MATCH(BROADWELL_G,		core_funcs),
2270	X86_MATCH(ATOM_AIRMONT,		airmont_funcs),
2271	X86_MATCH(SKYLAKE_L,		core_funcs),
2272	X86_MATCH(BROADWELL_X,		core_funcs),
2273	X86_MATCH(SKYLAKE,		core_funcs),
2274	X86_MATCH(BROADWELL_D,		core_funcs),
2275	X86_MATCH(XEON_PHI_KNL,		knl_funcs),
2276	X86_MATCH(XEON_PHI_KNM,		knl_funcs),
2277	X86_MATCH(ATOM_GOLDMONT,	core_funcs),
2278	X86_MATCH(ATOM_GOLDMONT_PLUS,	core_funcs),
2279	X86_MATCH(SKYLAKE_X,		core_funcs),
2280	X86_MATCH(COMETLAKE,		core_funcs),
2281	X86_MATCH(ICELAKE_X,		core_funcs),
2282	{}
2283};
2284MODULE_DEVICE_TABLE(x86cpu, intel_pstate_cpu_ids);
2285
2286static const struct x86_cpu_id intel_pstate_cpu_oob_ids[] __initconst = {
2287	X86_MATCH(BROADWELL_D,		core_funcs),
2288	X86_MATCH(BROADWELL_X,		core_funcs),
2289	X86_MATCH(SKYLAKE_X,		core_funcs),
2290	{}
2291};
2292
2293static const struct x86_cpu_id intel_pstate_cpu_ee_disable_ids[] = {
2294	X86_MATCH(KABYLAKE,		core_funcs),
2295	{}
2296};
2297
2298static const struct x86_cpu_id intel_pstate_hwp_boost_ids[] = {
2299	X86_MATCH(SKYLAKE_X,		core_funcs),
2300	X86_MATCH(SKYLAKE,		core_funcs),
2301	{}
2302};
2303
2304static int intel_pstate_init_cpu(unsigned int cpunum)
2305{
2306	struct cpudata *cpu;
2307
2308	cpu = all_cpu_data[cpunum];
2309
2310	if (!cpu) {
2311		cpu = kzalloc(sizeof(*cpu), GFP_KERNEL);
2312		if (!cpu)
2313			return -ENOMEM;
2314
2315		all_cpu_data[cpunum] = cpu;
2316
2317		cpu->cpu = cpunum;
2318
2319		cpu->epp_default = -EINVAL;
2320
2321		if (hwp_active) {
2322			const struct x86_cpu_id *id;
2323
2324			intel_pstate_hwp_enable(cpu);
2325
2326			id = x86_match_cpu(intel_pstate_hwp_boost_ids);
2327			if (id && intel_pstate_acpi_pm_profile_server())
2328				hwp_boost = true;
2329		}
2330	} else if (hwp_active) {
2331		/*
2332		 * Re-enable HWP in case this happens after a resume from ACPI
2333		 * S3 if the CPU was offline during the whole system/resume
2334		 * cycle.
2335		 */
2336		intel_pstate_hwp_reenable(cpu);
2337	}
2338
2339	cpu->epp_powersave = -EINVAL;
2340	cpu->epp_policy = 0;
2341
2342	intel_pstate_get_cpu_pstates(cpu);
2343
2344	pr_debug("controlling: cpu %d\n", cpunum);
2345
2346	return 0;
2347}
2348
2349static void intel_pstate_set_update_util_hook(unsigned int cpu_num)
2350{
2351	struct cpudata *cpu = all_cpu_data[cpu_num];
2352
2353	if (hwp_active && !hwp_boost)
2354		return;
2355
2356	if (cpu->update_util_set)
2357		return;
2358
2359	/* Prevent intel_pstate_update_util() from using stale data. */
2360	cpu->sample.time = 0;
2361	cpufreq_add_update_util_hook(cpu_num, &cpu->update_util,
2362				     (hwp_active ?
2363				      intel_pstate_update_util_hwp :
2364				      intel_pstate_update_util));
2365	cpu->update_util_set = true;
2366}
2367
2368static void intel_pstate_clear_update_util_hook(unsigned int cpu)
2369{
2370	struct cpudata *cpu_data = all_cpu_data[cpu];
2371
2372	if (!cpu_data->update_util_set)
2373		return;
2374
2375	cpufreq_remove_update_util_hook(cpu);
2376	cpu_data->update_util_set = false;
2377	synchronize_rcu();
2378}
2379
2380static int intel_pstate_get_max_freq(struct cpudata *cpu)
2381{
2382	return global.turbo_disabled || global.no_turbo ?
2383			cpu->pstate.max_freq : cpu->pstate.turbo_freq;
2384}
2385
2386static void intel_pstate_update_perf_limits(struct cpudata *cpu,
2387					    unsigned int policy_min,
2388					    unsigned int policy_max)
2389{
2390	int perf_ctl_scaling = cpu->pstate.perf_ctl_scaling;
2391	int32_t max_policy_perf, min_policy_perf;
2392
2393	max_policy_perf = policy_max / perf_ctl_scaling;
2394	if (policy_max == policy_min) {
2395		min_policy_perf = max_policy_perf;
2396	} else {
2397		min_policy_perf = policy_min / perf_ctl_scaling;
2398		min_policy_perf = clamp_t(int32_t, min_policy_perf,
2399					  0, max_policy_perf);
2400	}
2401
2402	/*
2403	 * HWP needs some special consideration, because HWP_REQUEST uses
2404	 * abstract values to represent performance rather than pure ratios.
2405	 */
2406	if (hwp_active) {
2407		intel_pstate_get_hwp_cap(cpu);
2408
2409		if (cpu->pstate.scaling != perf_ctl_scaling) {
2410			int scaling = cpu->pstate.scaling;
2411			int freq;
2412
2413			freq = max_policy_perf * perf_ctl_scaling;
2414			max_policy_perf = DIV_ROUND_UP(freq, scaling);
2415			freq = min_policy_perf * perf_ctl_scaling;
2416			min_policy_perf = DIV_ROUND_UP(freq, scaling);
2417		}
2418	}
2419
2420	pr_debug("cpu:%d min_policy_perf:%d max_policy_perf:%d\n",
2421		 cpu->cpu, min_policy_perf, max_policy_perf);
2422
2423	/* Normalize user input to [min_perf, max_perf] */
2424	if (per_cpu_limits) {
2425		cpu->min_perf_ratio = min_policy_perf;
2426		cpu->max_perf_ratio = max_policy_perf;
2427	} else {
2428		int turbo_max = cpu->pstate.turbo_pstate;
2429		int32_t global_min, global_max;
2430
2431		/* Global limits are in percent of the maximum turbo P-state. */
2432		global_max = DIV_ROUND_UP(turbo_max * global.max_perf_pct, 100);
2433		global_min = DIV_ROUND_UP(turbo_max * global.min_perf_pct, 100);
2434		global_min = clamp_t(int32_t, global_min, 0, global_max);
2435
2436		pr_debug("cpu:%d global_min:%d global_max:%d\n", cpu->cpu,
2437			 global_min, global_max);
2438
2439		cpu->min_perf_ratio = max(min_policy_perf, global_min);
2440		cpu->min_perf_ratio = min(cpu->min_perf_ratio, max_policy_perf);
2441		cpu->max_perf_ratio = min(max_policy_perf, global_max);
2442		cpu->max_perf_ratio = max(min_policy_perf, cpu->max_perf_ratio);
2443
2444		/* Make sure min_perf <= max_perf */
2445		cpu->min_perf_ratio = min(cpu->min_perf_ratio,
2446					  cpu->max_perf_ratio);
2447
2448	}
2449	pr_debug("cpu:%d max_perf_ratio:%d min_perf_ratio:%d\n", cpu->cpu,
2450		 cpu->max_perf_ratio,
2451		 cpu->min_perf_ratio);
2452}
2453
2454static int intel_pstate_set_policy(struct cpufreq_policy *policy)
2455{
2456	struct cpudata *cpu;
2457
2458	if (!policy->cpuinfo.max_freq)
2459		return -ENODEV;
2460
2461	pr_debug("set_policy cpuinfo.max %u policy->max %u\n",
2462		 policy->cpuinfo.max_freq, policy->max);
2463
2464	cpu = all_cpu_data[policy->cpu];
2465	cpu->policy = policy->policy;
2466
2467	mutex_lock(&intel_pstate_limits_lock);
2468
2469	intel_pstate_update_perf_limits(cpu, policy->min, policy->max);
2470
2471	if (cpu->policy == CPUFREQ_POLICY_PERFORMANCE) {
2472		/*
2473		 * NOHZ_FULL CPUs need this as the governor callback may not
2474		 * be invoked on them.
2475		 */
2476		intel_pstate_clear_update_util_hook(policy->cpu);
2477		intel_pstate_max_within_limits(cpu);
2478	} else {
2479		intel_pstate_set_update_util_hook(policy->cpu);
2480	}
2481
2482	if (hwp_active) {
2483		/*
2484		 * When hwp_boost was active before and dynamically it
2485		 * was turned off, in that case we need to clear the
2486		 * update util hook.
2487		 */
2488		if (!hwp_boost)
2489			intel_pstate_clear_update_util_hook(policy->cpu);
2490		intel_pstate_hwp_set(policy->cpu);
2491	}
2492
2493	mutex_unlock(&intel_pstate_limits_lock);
2494
2495	return 0;
2496}
2497
2498static void intel_pstate_adjust_policy_max(struct cpudata *cpu,
2499					   struct cpufreq_policy_data *policy)
2500{
2501	if (!hwp_active &&
2502	    cpu->pstate.max_pstate_physical > cpu->pstate.max_pstate &&
2503	    policy->max < policy->cpuinfo.max_freq &&
2504	    policy->max > cpu->pstate.max_freq) {
2505		pr_debug("policy->max > max non turbo frequency\n");
2506		policy->max = policy->cpuinfo.max_freq;
2507	}
2508}
2509
2510static void intel_pstate_verify_cpu_policy(struct cpudata *cpu,
2511					   struct cpufreq_policy_data *policy)
2512{
2513	int max_freq;
2514
2515	update_turbo_state();
2516	if (hwp_active) {
2517		intel_pstate_get_hwp_cap(cpu);
2518		max_freq = global.no_turbo || global.turbo_disabled ?
2519				cpu->pstate.max_freq : cpu->pstate.turbo_freq;
2520	} else {
2521		max_freq = intel_pstate_get_max_freq(cpu);
2522	}
2523	cpufreq_verify_within_limits(policy, policy->cpuinfo.min_freq, max_freq);
2524
2525	intel_pstate_adjust_policy_max(cpu, policy);
2526}
2527
2528static int intel_pstate_verify_policy(struct cpufreq_policy_data *policy)
2529{
2530	intel_pstate_verify_cpu_policy(all_cpu_data[policy->cpu], policy);
2531
2532	return 0;
2533}
2534
2535static int intel_cpufreq_cpu_offline(struct cpufreq_policy *policy)
2536{
2537	struct cpudata *cpu = all_cpu_data[policy->cpu];
2538
2539	pr_debug("CPU %d going offline\n", cpu->cpu);
2540
2541	if (cpu->suspended)
2542		return 0;
2543
2544	/*
2545	 * If the CPU is an SMT thread and it goes offline with the performance
2546	 * settings different from the minimum, it will prevent its sibling
2547	 * from getting to lower performance levels, so force the minimum
2548	 * performance on CPU offline to prevent that from happening.
2549	 */
2550	if (hwp_active)
2551		intel_pstate_hwp_offline(cpu);
2552	else
2553		intel_pstate_set_min_pstate(cpu);
2554
2555	intel_pstate_exit_perf_limits(policy);
2556
2557	return 0;
2558}
2559
2560static int intel_pstate_cpu_online(struct cpufreq_policy *policy)
2561{
2562	struct cpudata *cpu = all_cpu_data[policy->cpu];
2563
2564	pr_debug("CPU %d going online\n", cpu->cpu);
2565
2566	intel_pstate_init_acpi_perf_limits(policy);
2567
2568	if (hwp_active) {
2569		/*
2570		 * Re-enable HWP and clear the "suspended" flag to let "resume"
2571		 * know that it need not do that.
2572		 */
2573		intel_pstate_hwp_reenable(cpu);
2574		cpu->suspended = false;
2575	}
2576
2577	return 0;
2578}
2579
2580static int intel_pstate_cpu_offline(struct cpufreq_policy *policy)
2581{
2582	intel_pstate_clear_update_util_hook(policy->cpu);
2583
2584	return intel_cpufreq_cpu_offline(policy);
2585}
2586
2587static int intel_pstate_cpu_exit(struct cpufreq_policy *policy)
2588{
2589	pr_debug("CPU %d exiting\n", policy->cpu);
2590
2591	policy->fast_switch_possible = false;
2592
2593	return 0;
2594}
2595
2596static int __intel_pstate_cpu_init(struct cpufreq_policy *policy)
2597{
2598	struct cpudata *cpu;
2599	int rc;
2600
2601	rc = intel_pstate_init_cpu(policy->cpu);
2602	if (rc)
2603		return rc;
2604
2605	cpu = all_cpu_data[policy->cpu];
2606
2607	cpu->max_perf_ratio = 0xFF;
2608	cpu->min_perf_ratio = 0;
2609
2610	/* cpuinfo and default policy values */
2611	policy->cpuinfo.min_freq = cpu->pstate.min_freq;
2612	update_turbo_state();
2613	global.turbo_disabled_mf = global.turbo_disabled;
2614	policy->cpuinfo.max_freq = global.turbo_disabled ?
2615			cpu->pstate.max_freq : cpu->pstate.turbo_freq;
2616
2617	policy->min = policy->cpuinfo.min_freq;
2618	policy->max = policy->cpuinfo.max_freq;
2619
2620	intel_pstate_init_acpi_perf_limits(policy);
2621
2622	policy->fast_switch_possible = true;
2623
2624	return 0;
2625}
2626
2627static int intel_pstate_cpu_init(struct cpufreq_policy *policy)
2628{
2629	int ret = __intel_pstate_cpu_init(policy);
2630
2631	if (ret)
2632		return ret;
2633
2634	/*
2635	 * Set the policy to powersave to provide a valid fallback value in case
2636	 * the default cpufreq governor is neither powersave nor performance.
2637	 */
2638	policy->policy = CPUFREQ_POLICY_POWERSAVE;
2639
2640	if (hwp_active) {
2641		struct cpudata *cpu = all_cpu_data[policy->cpu];
2642
2643		cpu->epp_cached = intel_pstate_get_epp(cpu, 0);
2644	}
2645
2646	return 0;
2647}
2648
2649static struct cpufreq_driver intel_pstate = {
2650	.flags		= CPUFREQ_CONST_LOOPS,
2651	.verify		= intel_pstate_verify_policy,
2652	.setpolicy	= intel_pstate_set_policy,
2653	.suspend	= intel_pstate_suspend,
2654	.resume		= intel_pstate_resume,
2655	.init		= intel_pstate_cpu_init,
2656	.exit		= intel_pstate_cpu_exit,
2657	.offline	= intel_pstate_cpu_offline,
2658	.online		= intel_pstate_cpu_online,
2659	.update_limits	= intel_pstate_update_limits,
2660	.name		= "intel_pstate",
2661};
2662
2663static int intel_cpufreq_verify_policy(struct cpufreq_policy_data *policy)
2664{
2665	struct cpudata *cpu = all_cpu_data[policy->cpu];
2666
2667	intel_pstate_verify_cpu_policy(cpu, policy);
2668	intel_pstate_update_perf_limits(cpu, policy->min, policy->max);
2669
2670	return 0;
2671}
2672
2673/* Use of trace in passive mode:
2674 *
2675 * In passive mode the trace core_busy field (also known as the
2676 * performance field, and lablelled as such on the graphs; also known as
2677 * core_avg_perf) is not needed and so is re-assigned to indicate if the
2678 * driver call was via the normal or fast switch path. Various graphs
2679 * output from the intel_pstate_tracer.py utility that include core_busy
2680 * (or performance or core_avg_perf) have a fixed y-axis from 0 to 100%,
2681 * so we use 10 to indicate the normal path through the driver, and
2682 * 90 to indicate the fast switch path through the driver.
2683 * The scaled_busy field is not used, and is set to 0.
2684 */
2685
2686#define	INTEL_PSTATE_TRACE_TARGET 10
2687#define	INTEL_PSTATE_TRACE_FAST_SWITCH 90
2688
2689static void intel_cpufreq_trace(struct cpudata *cpu, unsigned int trace_type, int old_pstate)
2690{
2691	struct sample *sample;
2692
2693	if (!trace_pstate_sample_enabled())
2694		return;
2695
2696	if (!intel_pstate_sample(cpu, ktime_get()))
2697		return;
2698
2699	sample = &cpu->sample;
2700	trace_pstate_sample(trace_type,
2701		0,
2702		old_pstate,
2703		cpu->pstate.current_pstate,
2704		sample->mperf,
2705		sample->aperf,
2706		sample->tsc,
2707		get_avg_frequency(cpu),
2708		fp_toint(cpu->iowait_boost * 100));
2709}
2710
2711static void intel_cpufreq_hwp_update(struct cpudata *cpu, u32 min, u32 max,
2712				     u32 desired, bool fast_switch)
2713{
2714	u64 prev = READ_ONCE(cpu->hwp_req_cached), value = prev;
2715
2716	value &= ~HWP_MIN_PERF(~0L);
2717	value |= HWP_MIN_PERF(min);
2718
2719	value &= ~HWP_MAX_PERF(~0L);
2720	value |= HWP_MAX_PERF(max);
2721
2722	value &= ~HWP_DESIRED_PERF(~0L);
2723	value |= HWP_DESIRED_PERF(desired);
2724
2725	if (value == prev)
2726		return;
2727
2728	WRITE_ONCE(cpu->hwp_req_cached, value);
2729	if (fast_switch)
2730		wrmsrl(MSR_HWP_REQUEST, value);
2731	else
2732		wrmsrl_on_cpu(cpu->cpu, MSR_HWP_REQUEST, value);
2733}
2734
2735static void intel_cpufreq_perf_ctl_update(struct cpudata *cpu,
2736					  u32 target_pstate, bool fast_switch)
2737{
2738	if (fast_switch)
2739		wrmsrl(MSR_IA32_PERF_CTL,
2740		       pstate_funcs.get_val(cpu, target_pstate));
2741	else
2742		wrmsrl_on_cpu(cpu->cpu, MSR_IA32_PERF_CTL,
2743			      pstate_funcs.get_val(cpu, target_pstate));
2744}
2745
2746static int intel_cpufreq_update_pstate(struct cpufreq_policy *policy,
2747				       int target_pstate, bool fast_switch)
2748{
2749	struct cpudata *cpu = all_cpu_data[policy->cpu];
2750	int old_pstate = cpu->pstate.current_pstate;
2751
2752	target_pstate = intel_pstate_prepare_request(cpu, target_pstate);
2753	if (hwp_active) {
2754		int max_pstate = policy->strict_target ?
2755					target_pstate : cpu->max_perf_ratio;
2756
2757		intel_cpufreq_hwp_update(cpu, target_pstate, max_pstate, 0,
2758					 fast_switch);
2759	} else if (target_pstate != old_pstate) {
2760		intel_cpufreq_perf_ctl_update(cpu, target_pstate, fast_switch);
2761	}
2762
2763	cpu->pstate.current_pstate = target_pstate;
2764
2765	intel_cpufreq_trace(cpu, fast_switch ? INTEL_PSTATE_TRACE_FAST_SWITCH :
2766			    INTEL_PSTATE_TRACE_TARGET, old_pstate);
2767
2768	return target_pstate;
2769}
2770
2771static int intel_cpufreq_target(struct cpufreq_policy *policy,
2772				unsigned int target_freq,
2773				unsigned int relation)
2774{
2775	struct cpudata *cpu = all_cpu_data[policy->cpu];
2776	struct cpufreq_freqs freqs;
2777	int target_pstate;
2778
2779	update_turbo_state();
2780
2781	freqs.old = policy->cur;
2782	freqs.new = target_freq;
2783
2784	cpufreq_freq_transition_begin(policy, &freqs);
2785
2786	switch (relation) {
2787	case CPUFREQ_RELATION_L:
2788		target_pstate = DIV_ROUND_UP(freqs.new, cpu->pstate.scaling);
2789		break;
2790	case CPUFREQ_RELATION_H:
2791		target_pstate = freqs.new / cpu->pstate.scaling;
2792		break;
2793	default:
2794		target_pstate = DIV_ROUND_CLOSEST(freqs.new, cpu->pstate.scaling);
2795		break;
2796	}
2797
2798	target_pstate = intel_cpufreq_update_pstate(policy, target_pstate, false);
2799
2800	freqs.new = target_pstate * cpu->pstate.scaling;
2801
2802	cpufreq_freq_transition_end(policy, &freqs, false);
2803
2804	return 0;
2805}
2806
2807static unsigned int intel_cpufreq_fast_switch(struct cpufreq_policy *policy,
2808					      unsigned int target_freq)
2809{
2810	struct cpudata *cpu = all_cpu_data[policy->cpu];
2811	int target_pstate;
2812
2813	update_turbo_state();
2814
2815	target_pstate = DIV_ROUND_UP(target_freq, cpu->pstate.scaling);
2816
2817	target_pstate = intel_cpufreq_update_pstate(policy, target_pstate, true);
2818
2819	return target_pstate * cpu->pstate.scaling;
2820}
2821
2822static void intel_cpufreq_adjust_perf(unsigned int cpunum,
2823				      unsigned long min_perf,
2824				      unsigned long target_perf,
2825				      unsigned long capacity)
2826{
2827	struct cpudata *cpu = all_cpu_data[cpunum];
2828	u64 hwp_cap = READ_ONCE(cpu->hwp_cap_cached);
2829	int old_pstate = cpu->pstate.current_pstate;
2830	int cap_pstate, min_pstate, max_pstate, target_pstate;
2831
2832	update_turbo_state();
2833	cap_pstate = global.turbo_disabled ? HWP_GUARANTEED_PERF(hwp_cap) :
2834					     HWP_HIGHEST_PERF(hwp_cap);
2835
2836	/* Optimization: Avoid unnecessary divisions. */
2837
2838	target_pstate = cap_pstate;
2839	if (target_perf < capacity)
2840		target_pstate = DIV_ROUND_UP(cap_pstate * target_perf, capacity);
2841
2842	min_pstate = cap_pstate;
2843	if (min_perf < capacity)
2844		min_pstate = DIV_ROUND_UP(cap_pstate * min_perf, capacity);
2845
2846	if (min_pstate < cpu->pstate.min_pstate)
2847		min_pstate = cpu->pstate.min_pstate;
2848
2849	if (min_pstate < cpu->min_perf_ratio)
2850		min_pstate = cpu->min_perf_ratio;
2851
2852	max_pstate = min(cap_pstate, cpu->max_perf_ratio);
2853	if (max_pstate < min_pstate)
2854		max_pstate = min_pstate;
2855
2856	target_pstate = clamp_t(int, target_pstate, min_pstate, max_pstate);
2857
2858	intel_cpufreq_hwp_update(cpu, min_pstate, max_pstate, target_pstate, true);
2859
2860	cpu->pstate.current_pstate = target_pstate;
2861	intel_cpufreq_trace(cpu, INTEL_PSTATE_TRACE_FAST_SWITCH, old_pstate);
2862}
2863
2864static int intel_cpufreq_cpu_init(struct cpufreq_policy *policy)
2865{
2866	struct freq_qos_request *req;
2867	struct cpudata *cpu;
2868	struct device *dev;
2869	int ret, freq;
2870
2871	dev = get_cpu_device(policy->cpu);
2872	if (!dev)
2873		return -ENODEV;
2874
2875	ret = __intel_pstate_cpu_init(policy);
2876	if (ret)
2877		return ret;
2878
2879	policy->cpuinfo.transition_latency = INTEL_CPUFREQ_TRANSITION_LATENCY;
2880	/* This reflects the intel_pstate_get_cpu_pstates() setting. */
2881	policy->cur = policy->cpuinfo.min_freq;
2882
2883	req = kcalloc(2, sizeof(*req), GFP_KERNEL);
2884	if (!req) {
2885		ret = -ENOMEM;
2886		goto pstate_exit;
2887	}
2888
2889	cpu = all_cpu_data[policy->cpu];
2890
2891	if (hwp_active) {
2892		u64 value;
2893
2894		policy->transition_delay_us = INTEL_CPUFREQ_TRANSITION_DELAY_HWP;
2895
2896		intel_pstate_get_hwp_cap(cpu);
2897
2898		rdmsrl_on_cpu(cpu->cpu, MSR_HWP_REQUEST, &value);
2899		WRITE_ONCE(cpu->hwp_req_cached, value);
2900
2901		cpu->epp_cached = intel_pstate_get_epp(cpu, value);
2902	} else {
2903		policy->transition_delay_us = INTEL_CPUFREQ_TRANSITION_DELAY;
2904	}
2905
2906	freq = DIV_ROUND_UP(cpu->pstate.turbo_freq * global.min_perf_pct, 100);
2907
2908	ret = freq_qos_add_request(&policy->constraints, req, FREQ_QOS_MIN,
2909				   freq);
2910	if (ret < 0) {
2911		dev_err(dev, "Failed to add min-freq constraint (%d)\n", ret);
2912		goto free_req;
2913	}
2914
2915	freq = DIV_ROUND_UP(cpu->pstate.turbo_freq * global.max_perf_pct, 100);
2916
2917	ret = freq_qos_add_request(&policy->constraints, req + 1, FREQ_QOS_MAX,
2918				   freq);
2919	if (ret < 0) {
2920		dev_err(dev, "Failed to add max-freq constraint (%d)\n", ret);
2921		goto remove_min_req;
2922	}
2923
2924	policy->driver_data = req;
2925
2926	return 0;
2927
2928remove_min_req:
2929	freq_qos_remove_request(req);
2930free_req:
2931	kfree(req);
2932pstate_exit:
2933	intel_pstate_exit_perf_limits(policy);
2934
2935	return ret;
2936}
2937
2938static int intel_cpufreq_cpu_exit(struct cpufreq_policy *policy)
2939{
2940	struct freq_qos_request *req;
2941
2942	req = policy->driver_data;
2943
2944	freq_qos_remove_request(req + 1);
2945	freq_qos_remove_request(req);
2946	kfree(req);
2947
2948	return intel_pstate_cpu_exit(policy);
2949}
2950
2951static struct cpufreq_driver intel_cpufreq = {
2952	.flags		= CPUFREQ_CONST_LOOPS,
2953	.verify		= intel_cpufreq_verify_policy,
2954	.target		= intel_cpufreq_target,
2955	.fast_switch	= intel_cpufreq_fast_switch,
2956	.init		= intel_cpufreq_cpu_init,
2957	.exit		= intel_cpufreq_cpu_exit,
2958	.offline	= intel_cpufreq_cpu_offline,
2959	.online		= intel_pstate_cpu_online,
2960	.suspend	= intel_pstate_suspend,
2961	.resume		= intel_pstate_resume,
2962	.update_limits	= intel_pstate_update_limits,
2963	.name		= "intel_cpufreq",
2964};
2965
2966static struct cpufreq_driver *default_driver;
2967
2968static void intel_pstate_driver_cleanup(void)
2969{
2970	unsigned int cpu;
2971
2972	get_online_cpus();
2973	for_each_online_cpu(cpu) {
2974		if (all_cpu_data[cpu]) {
2975			if (intel_pstate_driver == &intel_pstate)
2976				intel_pstate_clear_update_util_hook(cpu);
2977
2978			kfree(all_cpu_data[cpu]);
2979			all_cpu_data[cpu] = NULL;
2980		}
2981	}
2982	put_online_cpus();
2983
2984	intel_pstate_driver = NULL;
2985}
2986
2987static int intel_pstate_register_driver(struct cpufreq_driver *driver)
2988{
2989	int ret;
2990
2991	if (driver == &intel_pstate)
2992		intel_pstate_sysfs_expose_hwp_dynamic_boost();
2993
2994	memset(&global, 0, sizeof(global));
2995	global.max_perf_pct = 100;
2996
2997	intel_pstate_driver = driver;
2998	ret = cpufreq_register_driver(intel_pstate_driver);
2999	if (ret) {
3000		intel_pstate_driver_cleanup();
3001		return ret;
3002	}
3003
3004	global.min_perf_pct = min_perf_pct_min();
3005
3006	return 0;
3007}
3008
3009static ssize_t intel_pstate_show_status(char *buf)
3010{
3011	if (!intel_pstate_driver)
3012		return sprintf(buf, "off\n");
3013
3014	return sprintf(buf, "%s\n", intel_pstate_driver == &intel_pstate ?
3015					"active" : "passive");
3016}
3017
3018static int intel_pstate_update_status(const char *buf, size_t size)
3019{
3020	if (size == 3 && !strncmp(buf, "off", size)) {
3021		if (!intel_pstate_driver)
3022			return -EINVAL;
3023
3024		if (hwp_active)
3025			return -EBUSY;
3026
3027		cpufreq_unregister_driver(intel_pstate_driver);
3028		intel_pstate_driver_cleanup();
3029		return 0;
3030	}
3031
3032	if (size == 6 && !strncmp(buf, "active", size)) {
3033		if (intel_pstate_driver) {
3034			if (intel_pstate_driver == &intel_pstate)
3035				return 0;
3036
3037			cpufreq_unregister_driver(intel_pstate_driver);
3038		}
3039
3040		return intel_pstate_register_driver(&intel_pstate);
3041	}
3042
3043	if (size == 7 && !strncmp(buf, "passive", size)) {
3044		if (intel_pstate_driver) {
3045			if (intel_pstate_driver == &intel_cpufreq)
3046				return 0;
3047
3048			cpufreq_unregister_driver(intel_pstate_driver);
3049			intel_pstate_sysfs_hide_hwp_dynamic_boost();
3050		}
3051
3052		return intel_pstate_register_driver(&intel_cpufreq);
3053	}
3054
3055	return -EINVAL;
3056}
3057
3058static int no_load __initdata;
3059static int no_hwp __initdata;
3060static int hwp_only __initdata;
3061static unsigned int force_load __initdata;
3062
3063static int __init intel_pstate_msrs_not_valid(void)
3064{
3065	if (!pstate_funcs.get_max() ||
3066	    !pstate_funcs.get_min() ||
3067	    !pstate_funcs.get_turbo())
3068		return -ENODEV;
3069
3070	return 0;
3071}
3072
3073static void __init copy_cpu_funcs(struct pstate_funcs *funcs)
3074{
3075	pstate_funcs.get_max   = funcs->get_max;
3076	pstate_funcs.get_max_physical = funcs->get_max_physical;
3077	pstate_funcs.get_min   = funcs->get_min;
3078	pstate_funcs.get_turbo = funcs->get_turbo;
3079	pstate_funcs.get_scaling = funcs->get_scaling;
3080	pstate_funcs.get_val   = funcs->get_val;
3081	pstate_funcs.get_vid   = funcs->get_vid;
3082	pstate_funcs.get_aperf_mperf_shift = funcs->get_aperf_mperf_shift;
3083}
3084
3085#ifdef CONFIG_ACPI
3086
3087static bool __init intel_pstate_no_acpi_pss(void)
3088{
3089	int i;
3090
3091	for_each_possible_cpu(i) {
3092		acpi_status status;
3093		union acpi_object *pss;
3094		struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
3095		struct acpi_processor *pr = per_cpu(processors, i);
3096
3097		if (!pr)
3098			continue;
3099
3100		status = acpi_evaluate_object(pr->handle, "_PSS", NULL, &buffer);
3101		if (ACPI_FAILURE(status))
3102			continue;
3103
3104		pss = buffer.pointer;
3105		if (pss && pss->type == ACPI_TYPE_PACKAGE) {
3106			kfree(pss);
3107			return false;
3108		}
3109
3110		kfree(pss);
3111	}
3112
3113	pr_debug("ACPI _PSS not found\n");
3114	return true;
3115}
3116
3117static bool __init intel_pstate_no_acpi_pcch(void)
3118{
3119	acpi_status status;
3120	acpi_handle handle;
3121
3122	status = acpi_get_handle(NULL, "\\_SB", &handle);
3123	if (ACPI_FAILURE(status))
3124		goto not_found;
3125
3126	if (acpi_has_method(handle, "PCCH"))
3127		return false;
3128
3129not_found:
3130	pr_debug("ACPI PCCH not found\n");
3131	return true;
3132}
3133
3134static bool __init intel_pstate_has_acpi_ppc(void)
3135{
3136	int i;
3137
3138	for_each_possible_cpu(i) {
3139		struct acpi_processor *pr = per_cpu(processors, i);
3140
3141		if (!pr)
3142			continue;
3143		if (acpi_has_method(pr->handle, "_PPC"))
3144			return true;
3145	}
3146	pr_debug("ACPI _PPC not found\n");
3147	return false;
3148}
3149
3150enum {
3151	PSS,
3152	PPC,
3153};
3154
3155/* Hardware vendor-specific info that has its own power management modes */
3156static struct acpi_platform_list plat_info[] __initdata = {
3157	{"HP    ", "ProLiant", 0, ACPI_SIG_FADT, all_versions, NULL, PSS},
3158	{"ORACLE", "X4-2    ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3159	{"ORACLE", "X4-2L   ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3160	{"ORACLE", "X4-2B   ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3161	{"ORACLE", "X3-2    ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3162	{"ORACLE", "X3-2L   ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3163	{"ORACLE", "X3-2B   ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3164	{"ORACLE", "X4470M2 ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3165	{"ORACLE", "X4270M3 ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3166	{"ORACLE", "X4270M2 ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3167	{"ORACLE", "X4170M2 ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3168	{"ORACLE", "X4170 M3", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3169	{"ORACLE", "X4275 M3", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3170	{"ORACLE", "X6-2    ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3171	{"ORACLE", "Sudbury ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3172	{ } /* End */
3173};
3174
3175#define BITMASK_OOB	(BIT(8) | BIT(18))
3176
3177static bool __init intel_pstate_platform_pwr_mgmt_exists(void)
3178{
3179	const struct x86_cpu_id *id;
3180	u64 misc_pwr;
3181	int idx;
3182
3183	id = x86_match_cpu(intel_pstate_cpu_oob_ids);
3184	if (id) {
3185		rdmsrl(MSR_MISC_PWR_MGMT, misc_pwr);
3186		if (misc_pwr & BITMASK_OOB) {
3187			pr_debug("Bit 8 or 18 in the MISC_PWR_MGMT MSR set\n");
3188			pr_debug("P states are controlled in Out of Band mode by the firmware/hardware\n");
3189			return true;
3190		}
3191	}
3192
3193	idx = acpi_match_platform_list(plat_info);
3194	if (idx < 0)
3195		return false;
3196
3197	switch (plat_info[idx].data) {
3198	case PSS:
3199		if (!intel_pstate_no_acpi_pss())
3200			return false;
3201
3202		return intel_pstate_no_acpi_pcch();
3203	case PPC:
3204		return intel_pstate_has_acpi_ppc() && !force_load;
3205	}
3206
3207	return false;
3208}
3209
3210static void intel_pstate_request_control_from_smm(void)
3211{
3212	/*
3213	 * It may be unsafe to request P-states control from SMM if _PPC support
3214	 * has not been enabled.
3215	 */
3216	if (acpi_ppc)
3217		acpi_processor_pstate_control();
3218}
3219#else /* CONFIG_ACPI not enabled */
3220static inline bool intel_pstate_platform_pwr_mgmt_exists(void) { return false; }
3221static inline bool intel_pstate_has_acpi_ppc(void) { return false; }
3222static inline void intel_pstate_request_control_from_smm(void) {}
3223#endif /* CONFIG_ACPI */
3224
3225#define INTEL_PSTATE_HWP_BROADWELL	0x01
3226
3227#define X86_MATCH_HWP(model, hwp_mode)					\
3228	X86_MATCH_VENDOR_FAM_MODEL_FEATURE(INTEL, 6, INTEL_FAM6_##model, \
3229					   X86_FEATURE_HWP, hwp_mode)
3230
3231static const struct x86_cpu_id hwp_support_ids[] __initconst = {
3232	X86_MATCH_HWP(BROADWELL_X,	INTEL_PSTATE_HWP_BROADWELL),
3233	X86_MATCH_HWP(BROADWELL_D,	INTEL_PSTATE_HWP_BROADWELL),
3234	X86_MATCH_HWP(ANY,		0),
3235	{}
3236};
3237
3238static bool intel_pstate_hwp_is_enabled(void)
3239{
3240	u64 value;
3241
3242	rdmsrl(MSR_PM_ENABLE, value);
3243	return !!(value & 0x1);
3244}
3245
3246static int __init intel_pstate_init(void)
3247{
3248	const struct x86_cpu_id *id;
3249	int rc;
3250
3251	if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL)
3252		return -ENODEV;
3253
3254	id = x86_match_cpu(hwp_support_ids);
3255	if (id) {
3256		bool hwp_forced = intel_pstate_hwp_is_enabled();
3257
3258		if (hwp_forced)
3259			pr_info("HWP enabled by BIOS\n");
3260		else if (no_load)
3261			return -ENODEV;
3262
3263		copy_cpu_funcs(&core_funcs);
3264		/*
3265		 * Avoid enabling HWP for processors without EPP support,
3266		 * because that means incomplete HWP implementation which is a
3267		 * corner case and supporting it is generally problematic.
3268		 *
3269		 * If HWP is enabled already, though, there is no choice but to
3270		 * deal with it.
3271		 */
3272		if ((!no_hwp && boot_cpu_has(X86_FEATURE_HWP_EPP)) || hwp_forced) {
3273			hwp_active++;
3274			hwp_mode_bdw = id->driver_data;
3275			intel_pstate.attr = hwp_cpufreq_attrs;
3276			intel_cpufreq.attr = hwp_cpufreq_attrs;
3277			intel_cpufreq.flags |= CPUFREQ_NEED_UPDATE_LIMITS;
3278			intel_cpufreq.adjust_perf = intel_cpufreq_adjust_perf;
3279			if (!default_driver)
3280				default_driver = &intel_pstate;
3281
3282			goto hwp_cpu_matched;
3283		}
3284		pr_info("HWP not enabled\n");
3285	} else {
3286		if (no_load)
3287			return -ENODEV;
3288
3289		id = x86_match_cpu(intel_pstate_cpu_ids);
3290		if (!id) {
3291			pr_info("CPU model not supported\n");
3292			return -ENODEV;
3293		}
3294
3295		copy_cpu_funcs((struct pstate_funcs *)id->driver_data);
3296	}
3297
3298	if (intel_pstate_msrs_not_valid()) {
3299		pr_info("Invalid MSRs\n");
3300		return -ENODEV;
3301	}
3302	/* Without HWP start in the passive mode. */
3303	if (!default_driver)
3304		default_driver = &intel_cpufreq;
3305
3306hwp_cpu_matched:
3307	/*
3308	 * The Intel pstate driver will be ignored if the platform
3309	 * firmware has its own power management modes.
3310	 */
3311	if (intel_pstate_platform_pwr_mgmt_exists()) {
3312		pr_info("P-states controlled by the platform\n");
3313		return -ENODEV;
3314	}
3315
3316	if (!hwp_active && hwp_only)
3317		return -ENOTSUPP;
3318
3319	pr_info("Intel P-state driver initializing\n");
3320
3321	all_cpu_data = vzalloc(array_size(sizeof(void *), num_possible_cpus()));
3322	if (!all_cpu_data)
3323		return -ENOMEM;
3324
3325	intel_pstate_request_control_from_smm();
3326
3327	intel_pstate_sysfs_expose_params();
3328
3329	mutex_lock(&intel_pstate_driver_lock);
3330	rc = intel_pstate_register_driver(default_driver);
3331	mutex_unlock(&intel_pstate_driver_lock);
3332	if (rc) {
3333		intel_pstate_sysfs_remove();
3334		return rc;
3335	}
3336
3337	if (hwp_active) {
3338		const struct x86_cpu_id *id;
3339
3340		id = x86_match_cpu(intel_pstate_cpu_ee_disable_ids);
3341		if (id) {
3342			set_power_ctl_ee_state(false);
3343			pr_info("Disabling energy efficiency optimization\n");
3344		}
3345
3346		pr_info("HWP enabled\n");
3347	} else if (boot_cpu_has(X86_FEATURE_HYBRID_CPU)) {
3348		pr_warn("Problematic setup: Hybrid processor with disabled HWP\n");
3349	}
3350
3351	return 0;
3352}
3353device_initcall(intel_pstate_init);
3354
3355static int __init intel_pstate_setup(char *str)
3356{
3357	if (!str)
3358		return -EINVAL;
3359
3360	if (!strcmp(str, "disable"))
3361		no_load = 1;
3362	else if (!strcmp(str, "active"))
3363		default_driver = &intel_pstate;
3364	else if (!strcmp(str, "passive"))
3365		default_driver = &intel_cpufreq;
3366
3367	if (!strcmp(str, "no_hwp"))
3368		no_hwp = 1;
3369
3370	if (!strcmp(str, "force"))
3371		force_load = 1;
3372	if (!strcmp(str, "hwp_only"))
3373		hwp_only = 1;
3374	if (!strcmp(str, "per_cpu_perf_limits"))
3375		per_cpu_limits = true;
3376
3377#ifdef CONFIG_ACPI
3378	if (!strcmp(str, "support_acpi_ppc"))
3379		acpi_ppc = true;
3380#endif
3381
3382	return 0;
3383}
3384early_param("intel_pstate", intel_pstate_setup);
3385
3386MODULE_AUTHOR("Dirk Brandewie <dirk.j.brandewie@intel.com>");
3387MODULE_DESCRIPTION("'intel_pstate' - P state driver Intel Core processors");
3388MODULE_LICENSE("GPL");