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/cpu.h>
  31#include <asm/div64.h>
  32#include <asm/msr.h>
  33#include <asm/cpu_device_id.h>
  34#include <asm/cpufeature.h>
  35#include <asm/intel-family.h>
  36#include "../drivers/thermal/intel/thermal_interrupt.h"
  37
  38#define INTEL_PSTATE_SAMPLING_INTERVAL	(10 * NSEC_PER_MSEC)
  39
  40#define INTEL_CPUFREQ_TRANSITION_LATENCY	20000
  41#define INTEL_CPUFREQ_TRANSITION_DELAY_HWP	5000
  42#define INTEL_CPUFREQ_TRANSITION_DELAY		500
  43
  44#ifdef CONFIG_ACPI
  45#include <acpi/processor.h>
  46#include <acpi/cppc_acpi.h>
  47#endif
  48
  49#define FRAC_BITS 8
  50#define int_tofp(X) ((int64_t)(X) << FRAC_BITS)
  51#define fp_toint(X) ((X) >> FRAC_BITS)
  52
  53#define ONE_EIGHTH_FP ((int64_t)1 << (FRAC_BITS - 3))
  54
  55#define EXT_BITS 6
  56#define EXT_FRAC_BITS (EXT_BITS + FRAC_BITS)
  57#define fp_ext_toint(X) ((X) >> EXT_FRAC_BITS)
  58#define int_ext_tofp(X) ((int64_t)(X) << EXT_FRAC_BITS)
  59
  60static inline int32_t mul_fp(int32_t x, int32_t y)
  61{
  62	return ((int64_t)x * (int64_t)y) >> FRAC_BITS;
  63}
  64
  65static inline int32_t div_fp(s64 x, s64 y)
  66{
  67	return div64_s64((int64_t)x << FRAC_BITS, y);
  68}
  69
  70static inline int ceiling_fp(int32_t x)
  71{
  72	int mask, ret;
  73
  74	ret = fp_toint(x);
  75	mask = (1 << FRAC_BITS) - 1;
  76	if (x & mask)
  77		ret += 1;
  78	return ret;
  79}
  80
  81static inline u64 mul_ext_fp(u64 x, u64 y)
  82{
  83	return (x * y) >> EXT_FRAC_BITS;
  84}
  85
  86static inline u64 div_ext_fp(u64 x, u64 y)
  87{
  88	return div64_u64(x << EXT_FRAC_BITS, y);
  89}
  90
  91/**
  92 * struct sample -	Store performance sample
  93 * @core_avg_perf:	Ratio of APERF/MPERF which is the actual average
  94 *			performance during last sample period
  95 * @busy_scaled:	Scaled busy value which is used to calculate next
  96 *			P state. This can be different than core_avg_perf
  97 *			to account for cpu idle period
  98 * @aperf:		Difference of actual performance frequency clock count
  99 *			read from APERF MSR between last and current sample
 100 * @mperf:		Difference of maximum performance frequency clock count
 101 *			read from MPERF MSR between last and current sample
 102 * @tsc:		Difference of time stamp counter between last and
 103 *			current sample
 104 * @time:		Current time from scheduler
 105 *
 106 * This structure is used in the cpudata structure to store performance sample
 107 * data for choosing next P State.
 108 */
 109struct sample {
 110	int32_t core_avg_perf;
 111	int32_t busy_scaled;
 112	u64 aperf;
 113	u64 mperf;
 114	u64 tsc;
 115	u64 time;
 116};
 117
 118/**
 119 * struct pstate_data - Store P state data
 120 * @current_pstate:	Current requested P state
 121 * @min_pstate:		Min P state possible for this platform
 122 * @max_pstate:		Max P state possible for this platform
 123 * @max_pstate_physical:This is physical Max P state for a processor
 124 *			This can be higher than the max_pstate which can
 125 *			be limited by platform thermal design power limits
 126 * @perf_ctl_scaling:	PERF_CTL P-state to frequency scaling factor
 127 * @scaling:		Scaling factor between performance and frequency
 128 * @turbo_pstate:	Max Turbo P state possible for this platform
 129 * @min_freq:		@min_pstate frequency in cpufreq units
 130 * @max_freq:		@max_pstate frequency in cpufreq units
 131 * @turbo_freq:		@turbo_pstate frequency in cpufreq units
 132 *
 133 * Stores the per cpu model P state limits and current P state.
 134 */
 135struct pstate_data {
 136	int	current_pstate;
 137	int	min_pstate;
 138	int	max_pstate;
 139	int	max_pstate_physical;
 140	int	perf_ctl_scaling;
 141	int	scaling;
 142	int	turbo_pstate;
 143	unsigned int min_freq;
 144	unsigned int max_freq;
 145	unsigned int turbo_freq;
 146};
 147
 148/**
 149 * struct vid_data -	Stores voltage information data
 150 * @min:		VID data for this platform corresponding to
 151 *			the lowest P state
 152 * @max:		VID data corresponding to the highest P State.
 153 * @turbo:		VID data for turbo P state
 154 * @ratio:		Ratio of (vid max - vid min) /
 155 *			(max P state - Min P State)
 156 *
 157 * Stores the voltage data for DVFS (Dynamic Voltage and Frequency Scaling)
 158 * This data is used in Atom platforms, where in addition to target P state,
 159 * the voltage data needs to be specified to select next P State.
 160 */
 161struct vid_data {
 162	int min;
 163	int max;
 164	int turbo;
 165	int32_t ratio;
 166};
 167
 168/**
 169 * struct global_params - Global parameters, mostly tunable via sysfs.
 170 * @no_turbo:		Whether or not to use turbo P-states.
 171 * @turbo_disabled:	Whether or not turbo P-states are available at all,
 172 *			based on the MSR_IA32_MISC_ENABLE value and whether or
 173 *			not the maximum reported turbo P-state is different from
 174 *			the maximum reported non-turbo one.
 175 * @turbo_disabled_mf:	The @turbo_disabled value reflected by cpuinfo.max_freq.
 176 * @min_perf_pct:	Minimum capacity limit in percent of the maximum turbo
 177 *			P-state capacity.
 178 * @max_perf_pct:	Maximum capacity limit in percent of the maximum turbo
 179 *			P-state capacity.
 180 */
 181struct global_params {
 182	bool no_turbo;
 183	bool turbo_disabled;
 184	bool turbo_disabled_mf;
 185	int max_perf_pct;
 186	int min_perf_pct;
 187};
 188
 189/**
 190 * struct cpudata -	Per CPU instance data storage
 191 * @cpu:		CPU number for this instance data
 192 * @policy:		CPUFreq policy value
 193 * @update_util:	CPUFreq utility callback information
 194 * @update_util_set:	CPUFreq utility callback is set
 195 * @iowait_boost:	iowait-related boost fraction
 196 * @last_update:	Time of the last update.
 197 * @pstate:		Stores P state limits for this CPU
 198 * @vid:		Stores VID limits for this CPU
 199 * @last_sample_time:	Last Sample time
 200 * @aperf_mperf_shift:	APERF vs MPERF counting frequency difference
 201 * @prev_aperf:		Last APERF value read from APERF MSR
 202 * @prev_mperf:		Last MPERF value read from MPERF MSR
 203 * @prev_tsc:		Last timestamp counter (TSC) value
 204 * @prev_cummulative_iowait: IO Wait time difference from last and
 205 *			current sample
 206 * @sample:		Storage for storing last Sample data
 207 * @min_perf_ratio:	Minimum capacity in terms of PERF or HWP ratios
 208 * @max_perf_ratio:	Maximum capacity in terms of PERF or HWP ratios
 209 * @acpi_perf_data:	Stores ACPI perf information read from _PSS
 210 * @valid_pss_table:	Set to true for valid ACPI _PSS entries found
 211 * @epp_powersave:	Last saved HWP energy performance preference
 212 *			(EPP) or energy performance bias (EPB),
 213 *			when policy switched to performance
 214 * @epp_policy:		Last saved policy used to set EPP/EPB
 215 * @epp_default:	Power on default HWP energy performance
 216 *			preference/bias
 217 * @epp_cached		Cached HWP energy-performance preference value
 218 * @hwp_req_cached:	Cached value of the last HWP Request MSR
 219 * @hwp_cap_cached:	Cached value of the last HWP Capabilities MSR
 220 * @last_io_update:	Last time when IO wake flag was set
 221 * @sched_flags:	Store scheduler flags for possible cross CPU update
 222 * @hwp_boost_min:	Last HWP boosted min performance
 223 * @suspended:		Whether or not the driver has been suspended.
 224 * @hwp_notify_work:	workqueue for HWP notifications.
 225 *
 226 * This structure stores per CPU instance data for all CPUs.
 227 */
 228struct cpudata {
 229	int cpu;
 230
 231	unsigned int policy;
 232	struct update_util_data update_util;
 233	bool   update_util_set;
 234
 235	struct pstate_data pstate;
 236	struct vid_data vid;
 237
 238	u64	last_update;
 239	u64	last_sample_time;
 240	u64	aperf_mperf_shift;
 241	u64	prev_aperf;
 242	u64	prev_mperf;
 243	u64	prev_tsc;
 244	u64	prev_cummulative_iowait;
 245	struct sample sample;
 246	int32_t	min_perf_ratio;
 247	int32_t	max_perf_ratio;
 248#ifdef CONFIG_ACPI
 249	struct acpi_processor_performance acpi_perf_data;
 250	bool valid_pss_table;
 251#endif
 252	unsigned int iowait_boost;
 253	s16 epp_powersave;
 254	s16 epp_policy;
 255	s16 epp_default;
 256	s16 epp_cached;
 257	u64 hwp_req_cached;
 258	u64 hwp_cap_cached;
 259	u64 last_io_update;
 260	unsigned int sched_flags;
 261	u32 hwp_boost_min;
 262	bool suspended;
 263	struct delayed_work hwp_notify_work;
 264};
 265
 266static struct cpudata **all_cpu_data;
 267
 268/**
 269 * struct pstate_funcs - Per CPU model specific callbacks
 270 * @get_max:		Callback to get maximum non turbo effective P state
 271 * @get_max_physical:	Callback to get maximum non turbo physical P state
 272 * @get_min:		Callback to get minimum P state
 273 * @get_turbo:		Callback to get turbo P state
 274 * @get_scaling:	Callback to get frequency scaling factor
 275 * @get_cpu_scaling:	Get frequency scaling factor for a given cpu
 276 * @get_aperf_mperf_shift: Callback to get the APERF vs MPERF frequency difference
 277 * @get_val:		Callback to convert P state to actual MSR write value
 278 * @get_vid:		Callback to get VID data for Atom platforms
 279 *
 280 * Core and Atom CPU models have different way to get P State limits. This
 281 * structure is used to store those callbacks.
 282 */
 283struct pstate_funcs {
 284	int (*get_max)(int cpu);
 285	int (*get_max_physical)(int cpu);
 286	int (*get_min)(int cpu);
 287	int (*get_turbo)(int cpu);
 288	int (*get_scaling)(void);
 289	int (*get_cpu_scaling)(int cpu);
 290	int (*get_aperf_mperf_shift)(void);
 291	u64 (*get_val)(struct cpudata*, int pstate);
 292	void (*get_vid)(struct cpudata *);
 293};
 294
 295static struct pstate_funcs pstate_funcs __read_mostly;
 296
 297static int hwp_active __read_mostly;
 298static int hwp_mode_bdw __read_mostly;
 299static bool per_cpu_limits __read_mostly;
 300static bool hwp_boost __read_mostly;
 301static bool hwp_forced __read_mostly;
 302
 303static struct cpufreq_driver *intel_pstate_driver __read_mostly;
 304
 305#define HYBRID_SCALING_FACTOR		78741
 306#define HYBRID_SCALING_FACTOR_MTL	80000
 307
 308static int hybrid_scaling_factor = HYBRID_SCALING_FACTOR;
 309
 310static inline int core_get_scaling(void)
 311{
 312	return 100000;
 313}
 314
 315#ifdef CONFIG_ACPI
 316static bool acpi_ppc;
 317#endif
 318
 319static struct global_params global;
 320
 321static DEFINE_MUTEX(intel_pstate_driver_lock);
 322static DEFINE_MUTEX(intel_pstate_limits_lock);
 323
 324#ifdef CONFIG_ACPI
 325
 326static bool intel_pstate_acpi_pm_profile_server(void)
 327{
 328	if (acpi_gbl_FADT.preferred_profile == PM_ENTERPRISE_SERVER ||
 329	    acpi_gbl_FADT.preferred_profile == PM_PERFORMANCE_SERVER)
 330		return true;
 331
 332	return false;
 333}
 334
 335static bool intel_pstate_get_ppc_enable_status(void)
 336{
 337	if (intel_pstate_acpi_pm_profile_server())
 338		return true;
 339
 340	return acpi_ppc;
 341}
 342
 343#ifdef CONFIG_ACPI_CPPC_LIB
 344
 345/* The work item is needed to avoid CPU hotplug locking issues */
 346static void intel_pstste_sched_itmt_work_fn(struct work_struct *work)
 347{
 348	sched_set_itmt_support();
 349}
 350
 351static DECLARE_WORK(sched_itmt_work, intel_pstste_sched_itmt_work_fn);
 352
 353#define CPPC_MAX_PERF	U8_MAX
 354
 355static void intel_pstate_set_itmt_prio(int cpu)
 356{
 357	struct cppc_perf_caps cppc_perf;
 358	static u32 max_highest_perf = 0, min_highest_perf = U32_MAX;
 359	int ret;
 360
 361	ret = cppc_get_perf_caps(cpu, &cppc_perf);
 362	if (ret)
 363		return;
 364
 365	/*
 366	 * On some systems with overclocking enabled, CPPC.highest_perf is hardcoded to 0xff.
 367	 * In this case we can't use CPPC.highest_perf to enable ITMT.
 368	 * In this case we can look at MSR_HWP_CAPABILITIES bits [8:0] to decide.
 369	 */
 370	if (cppc_perf.highest_perf == CPPC_MAX_PERF)
 371		cppc_perf.highest_perf = HWP_HIGHEST_PERF(READ_ONCE(all_cpu_data[cpu]->hwp_cap_cached));
 372
 373	/*
 374	 * The priorities can be set regardless of whether or not
 375	 * sched_set_itmt_support(true) has been called and it is valid to
 376	 * update them at any time after it has been called.
 377	 */
 378	sched_set_itmt_core_prio(cppc_perf.highest_perf, cpu);
 379
 380	if (max_highest_perf <= min_highest_perf) {
 381		if (cppc_perf.highest_perf > max_highest_perf)
 382			max_highest_perf = cppc_perf.highest_perf;
 383
 384		if (cppc_perf.highest_perf < min_highest_perf)
 385			min_highest_perf = cppc_perf.highest_perf;
 386
 387		if (max_highest_perf > min_highest_perf) {
 388			/*
 389			 * This code can be run during CPU online under the
 390			 * CPU hotplug locks, so sched_set_itmt_support()
 391			 * cannot be called from here.  Queue up a work item
 392			 * to invoke it.
 393			 */
 394			schedule_work(&sched_itmt_work);
 395		}
 396	}
 397}
 398
 399static int intel_pstate_get_cppc_guaranteed(int cpu)
 400{
 401	struct cppc_perf_caps cppc_perf;
 402	int ret;
 403
 404	ret = cppc_get_perf_caps(cpu, &cppc_perf);
 405	if (ret)
 406		return ret;
 407
 408	if (cppc_perf.guaranteed_perf)
 409		return cppc_perf.guaranteed_perf;
 410
 411	return cppc_perf.nominal_perf;
 412}
 413
 414static int intel_pstate_cppc_get_scaling(int cpu)
 415{
 416	struct cppc_perf_caps cppc_perf;
 417	int ret;
 418
 419	ret = cppc_get_perf_caps(cpu, &cppc_perf);
 420
 421	/*
 422	 * If the nominal frequency and the nominal performance are not
 423	 * zero and the ratio between them is not 100, return the hybrid
 424	 * scaling factor.
 425	 */
 426	if (!ret && cppc_perf.nominal_perf && cppc_perf.nominal_freq &&
 427	    cppc_perf.nominal_perf * 100 != cppc_perf.nominal_freq)
 428		return hybrid_scaling_factor;
 429
 430	return core_get_scaling();
 431}
 432
 433#else /* CONFIG_ACPI_CPPC_LIB */
 434static inline void intel_pstate_set_itmt_prio(int cpu)
 435{
 436}
 437#endif /* CONFIG_ACPI_CPPC_LIB */
 438
 439static void intel_pstate_init_acpi_perf_limits(struct cpufreq_policy *policy)
 440{
 441	struct cpudata *cpu;
 442	int ret;
 443	int i;
 444
 445	if (hwp_active) {
 446		intel_pstate_set_itmt_prio(policy->cpu);
 447		return;
 448	}
 449
 450	if (!intel_pstate_get_ppc_enable_status())
 451		return;
 452
 453	cpu = all_cpu_data[policy->cpu];
 454
 455	ret = acpi_processor_register_performance(&cpu->acpi_perf_data,
 456						  policy->cpu);
 457	if (ret)
 458		return;
 459
 460	/*
 461	 * Check if the control value in _PSS is for PERF_CTL MSR, which should
 462	 * guarantee that the states returned by it map to the states in our
 463	 * list directly.
 464	 */
 465	if (cpu->acpi_perf_data.control_register.space_id !=
 466						ACPI_ADR_SPACE_FIXED_HARDWARE)
 467		goto err;
 468
 469	/*
 470	 * If there is only one entry _PSS, simply ignore _PSS and continue as
 471	 * usual without taking _PSS into account
 472	 */
 473	if (cpu->acpi_perf_data.state_count < 2)
 474		goto err;
 475
 476	pr_debug("CPU%u - ACPI _PSS perf data\n", policy->cpu);
 477	for (i = 0; i < cpu->acpi_perf_data.state_count; i++) {
 478		pr_debug("     %cP%d: %u MHz, %u mW, 0x%x\n",
 479			 (i == cpu->acpi_perf_data.state ? '*' : ' '), i,
 480			 (u32) cpu->acpi_perf_data.states[i].core_frequency,
 481			 (u32) cpu->acpi_perf_data.states[i].power,
 482			 (u32) cpu->acpi_perf_data.states[i].control);
 483	}
 484
 485	cpu->valid_pss_table = true;
 486	pr_debug("_PPC limits will be enforced\n");
 487
 488	return;
 489
 490 err:
 491	cpu->valid_pss_table = false;
 492	acpi_processor_unregister_performance(policy->cpu);
 493}
 494
 495static void intel_pstate_exit_perf_limits(struct cpufreq_policy *policy)
 496{
 497	struct cpudata *cpu;
 498
 499	cpu = all_cpu_data[policy->cpu];
 500	if (!cpu->valid_pss_table)
 501		return;
 502
 503	acpi_processor_unregister_performance(policy->cpu);
 504}
 505#else /* CONFIG_ACPI */
 506static inline void intel_pstate_init_acpi_perf_limits(struct cpufreq_policy *policy)
 507{
 508}
 509
 510static inline void intel_pstate_exit_perf_limits(struct cpufreq_policy *policy)
 511{
 512}
 513
 514static inline bool intel_pstate_acpi_pm_profile_server(void)
 515{
 516	return false;
 517}
 518#endif /* CONFIG_ACPI */
 519
 520#ifndef CONFIG_ACPI_CPPC_LIB
 521static inline int intel_pstate_get_cppc_guaranteed(int cpu)
 522{
 523	return -ENOTSUPP;
 524}
 525
 526static int intel_pstate_cppc_get_scaling(int cpu)
 527{
 528	return core_get_scaling();
 529}
 530#endif /* CONFIG_ACPI_CPPC_LIB */
 531
 532static int intel_pstate_freq_to_hwp_rel(struct cpudata *cpu, int freq,
 533					unsigned int relation)
 534{
 535	if (freq == cpu->pstate.turbo_freq)
 536		return cpu->pstate.turbo_pstate;
 537
 538	if (freq == cpu->pstate.max_freq)
 539		return cpu->pstate.max_pstate;
 540
 541	switch (relation) {
 542	case CPUFREQ_RELATION_H:
 543		return freq / cpu->pstate.scaling;
 544	case CPUFREQ_RELATION_C:
 545		return DIV_ROUND_CLOSEST(freq, cpu->pstate.scaling);
 546	}
 547
 548	return DIV_ROUND_UP(freq, cpu->pstate.scaling);
 549}
 550
 551static int intel_pstate_freq_to_hwp(struct cpudata *cpu, int freq)
 552{
 553	return intel_pstate_freq_to_hwp_rel(cpu, freq, CPUFREQ_RELATION_L);
 554}
 555
 556/**
 557 * intel_pstate_hybrid_hwp_adjust - Calibrate HWP performance levels.
 558 * @cpu: Target CPU.
 559 *
 560 * On hybrid processors, HWP may expose more performance levels than there are
 561 * P-states accessible through the PERF_CTL interface.  If that happens, the
 562 * scaling factor between HWP performance levels and CPU frequency will be less
 563 * than the scaling factor between P-state values and CPU frequency.
 564 *
 565 * In that case, adjust the CPU parameters used in computations accordingly.
 566 */
 567static void intel_pstate_hybrid_hwp_adjust(struct cpudata *cpu)
 568{
 569	int perf_ctl_max_phys = cpu->pstate.max_pstate_physical;
 570	int perf_ctl_scaling = cpu->pstate.perf_ctl_scaling;
 571	int perf_ctl_turbo = pstate_funcs.get_turbo(cpu->cpu);
 572	int scaling = cpu->pstate.scaling;
 573	int freq;
 574
 575	pr_debug("CPU%d: perf_ctl_max_phys = %d\n", cpu->cpu, perf_ctl_max_phys);
 576	pr_debug("CPU%d: perf_ctl_turbo = %d\n", cpu->cpu, perf_ctl_turbo);
 577	pr_debug("CPU%d: perf_ctl_scaling = %d\n", cpu->cpu, perf_ctl_scaling);
 578	pr_debug("CPU%d: HWP_CAP guaranteed = %d\n", cpu->cpu, cpu->pstate.max_pstate);
 579	pr_debug("CPU%d: HWP_CAP highest = %d\n", cpu->cpu, cpu->pstate.turbo_pstate);
 580	pr_debug("CPU%d: HWP-to-frequency scaling factor: %d\n", cpu->cpu, scaling);
 581
 582	cpu->pstate.turbo_freq = rounddown(cpu->pstate.turbo_pstate * scaling,
 583					   perf_ctl_scaling);
 584	cpu->pstate.max_freq = rounddown(cpu->pstate.max_pstate * scaling,
 585					 perf_ctl_scaling);
 586
 587	freq = perf_ctl_max_phys * perf_ctl_scaling;
 588	cpu->pstate.max_pstate_physical = intel_pstate_freq_to_hwp(cpu, freq);
 589
 590	freq = cpu->pstate.min_pstate * perf_ctl_scaling;
 591	cpu->pstate.min_freq = freq;
 592	/*
 593	 * Cast the min P-state value retrieved via pstate_funcs.get_min() to
 594	 * the effective range of HWP performance levels.
 595	 */
 596	cpu->pstate.min_pstate = intel_pstate_freq_to_hwp(cpu, freq);
 597}
 598
 599static inline void update_turbo_state(void)
 600{
 601	u64 misc_en;
 602
 603	rdmsrl(MSR_IA32_MISC_ENABLE, misc_en);
 604	global.turbo_disabled = misc_en & MSR_IA32_MISC_ENABLE_TURBO_DISABLE;
 605}
 606
 607static int min_perf_pct_min(void)
 608{
 609	struct cpudata *cpu = all_cpu_data[0];
 610	int turbo_pstate = cpu->pstate.turbo_pstate;
 611
 612	return turbo_pstate ?
 613		(cpu->pstate.min_pstate * 100 / turbo_pstate) : 0;
 614}
 615
 616static s16 intel_pstate_get_epb(struct cpudata *cpu_data)
 617{
 618	u64 epb;
 619	int ret;
 620
 621	if (!boot_cpu_has(X86_FEATURE_EPB))
 622		return -ENXIO;
 623
 624	ret = rdmsrl_on_cpu(cpu_data->cpu, MSR_IA32_ENERGY_PERF_BIAS, &epb);
 625	if (ret)
 626		return (s16)ret;
 627
 628	return (s16)(epb & 0x0f);
 629}
 630
 631static s16 intel_pstate_get_epp(struct cpudata *cpu_data, u64 hwp_req_data)
 632{
 633	s16 epp;
 634
 635	if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
 636		/*
 637		 * When hwp_req_data is 0, means that caller didn't read
 638		 * MSR_HWP_REQUEST, so need to read and get EPP.
 639		 */
 640		if (!hwp_req_data) {
 641			epp = rdmsrl_on_cpu(cpu_data->cpu, MSR_HWP_REQUEST,
 642					    &hwp_req_data);
 643			if (epp)
 644				return epp;
 645		}
 646		epp = (hwp_req_data >> 24) & 0xff;
 647	} else {
 648		/* When there is no EPP present, HWP uses EPB settings */
 649		epp = intel_pstate_get_epb(cpu_data);
 650	}
 651
 652	return epp;
 653}
 654
 655static int intel_pstate_set_epb(int cpu, s16 pref)
 656{
 657	u64 epb;
 658	int ret;
 659
 660	if (!boot_cpu_has(X86_FEATURE_EPB))
 661		return -ENXIO;
 662
 663	ret = rdmsrl_on_cpu(cpu, MSR_IA32_ENERGY_PERF_BIAS, &epb);
 664	if (ret)
 665		return ret;
 666
 667	epb = (epb & ~0x0f) | pref;
 668	wrmsrl_on_cpu(cpu, MSR_IA32_ENERGY_PERF_BIAS, epb);
 669
 670	return 0;
 671}
 672
 673/*
 674 * EPP/EPB display strings corresponding to EPP index in the
 675 * energy_perf_strings[]
 676 *	index		String
 677 *-------------------------------------
 678 *	0		default
 679 *	1		performance
 680 *	2		balance_performance
 681 *	3		balance_power
 682 *	4		power
 683 */
 684
 685enum energy_perf_value_index {
 686	EPP_INDEX_DEFAULT = 0,
 687	EPP_INDEX_PERFORMANCE,
 688	EPP_INDEX_BALANCE_PERFORMANCE,
 689	EPP_INDEX_BALANCE_POWERSAVE,
 690	EPP_INDEX_POWERSAVE,
 691};
 692
 693static const char * const energy_perf_strings[] = {
 694	[EPP_INDEX_DEFAULT] = "default",
 695	[EPP_INDEX_PERFORMANCE] = "performance",
 696	[EPP_INDEX_BALANCE_PERFORMANCE] = "balance_performance",
 697	[EPP_INDEX_BALANCE_POWERSAVE] = "balance_power",
 698	[EPP_INDEX_POWERSAVE] = "power",
 699	NULL
 700};
 701static unsigned int epp_values[] = {
 702	[EPP_INDEX_DEFAULT] = 0, /* Unused index */
 703	[EPP_INDEX_PERFORMANCE] = HWP_EPP_PERFORMANCE,
 704	[EPP_INDEX_BALANCE_PERFORMANCE] = HWP_EPP_BALANCE_PERFORMANCE,
 705	[EPP_INDEX_BALANCE_POWERSAVE] = HWP_EPP_BALANCE_POWERSAVE,
 706	[EPP_INDEX_POWERSAVE] = HWP_EPP_POWERSAVE,
 707};
 708
 709static int intel_pstate_get_energy_pref_index(struct cpudata *cpu_data, int *raw_epp)
 710{
 711	s16 epp;
 712	int index = -EINVAL;
 713
 714	*raw_epp = 0;
 715	epp = intel_pstate_get_epp(cpu_data, 0);
 716	if (epp < 0)
 717		return epp;
 718
 719	if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
 720		if (epp == epp_values[EPP_INDEX_PERFORMANCE])
 721			return EPP_INDEX_PERFORMANCE;
 722		if (epp == epp_values[EPP_INDEX_BALANCE_PERFORMANCE])
 723			return EPP_INDEX_BALANCE_PERFORMANCE;
 724		if (epp == epp_values[EPP_INDEX_BALANCE_POWERSAVE])
 725			return EPP_INDEX_BALANCE_POWERSAVE;
 726		if (epp == epp_values[EPP_INDEX_POWERSAVE])
 727			return EPP_INDEX_POWERSAVE;
 728		*raw_epp = epp;
 729		return 0;
 730	} else if (boot_cpu_has(X86_FEATURE_EPB)) {
 731		/*
 732		 * Range:
 733		 *	0x00-0x03	:	Performance
 734		 *	0x04-0x07	:	Balance performance
 735		 *	0x08-0x0B	:	Balance power
 736		 *	0x0C-0x0F	:	Power
 737		 * The EPB is a 4 bit value, but our ranges restrict the
 738		 * value which can be set. Here only using top two bits
 739		 * effectively.
 740		 */
 741		index = (epp >> 2) + 1;
 742	}
 743
 744	return index;
 745}
 746
 747static int intel_pstate_set_epp(struct cpudata *cpu, u32 epp)
 748{
 749	int ret;
 750
 751	/*
 752	 * Use the cached HWP Request MSR value, because in the active mode the
 753	 * register itself may be updated by intel_pstate_hwp_boost_up() or
 754	 * intel_pstate_hwp_boost_down() at any time.
 755	 */
 756	u64 value = READ_ONCE(cpu->hwp_req_cached);
 757
 758	value &= ~GENMASK_ULL(31, 24);
 759	value |= (u64)epp << 24;
 760	/*
 761	 * The only other updater of hwp_req_cached in the active mode,
 762	 * intel_pstate_hwp_set(), is called under the same lock as this
 763	 * function, so it cannot run in parallel with the update below.
 764	 */
 765	WRITE_ONCE(cpu->hwp_req_cached, value);
 766	ret = wrmsrl_on_cpu(cpu->cpu, MSR_HWP_REQUEST, value);
 767	if (!ret)
 768		cpu->epp_cached = epp;
 769
 770	return ret;
 771}
 772
 773static int intel_pstate_set_energy_pref_index(struct cpudata *cpu_data,
 774					      int pref_index, bool use_raw,
 775					      u32 raw_epp)
 776{
 777	int epp = -EINVAL;
 778	int ret;
 779
 780	if (!pref_index)
 781		epp = cpu_data->epp_default;
 782
 783	if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
 784		if (use_raw)
 785			epp = raw_epp;
 786		else if (epp == -EINVAL)
 787			epp = epp_values[pref_index];
 788
 789		/*
 790		 * To avoid confusion, refuse to set EPP to any values different
 791		 * from 0 (performance) if the current policy is "performance",
 792		 * because those values would be overridden.
 793		 */
 794		if (epp > 0 && cpu_data->policy == CPUFREQ_POLICY_PERFORMANCE)
 795			return -EBUSY;
 796
 797		ret = intel_pstate_set_epp(cpu_data, epp);
 798	} else {
 799		if (epp == -EINVAL)
 800			epp = (pref_index - 1) << 2;
 801		ret = intel_pstate_set_epb(cpu_data->cpu, epp);
 802	}
 803
 804	return ret;
 805}
 806
 807static ssize_t show_energy_performance_available_preferences(
 808				struct cpufreq_policy *policy, char *buf)
 809{
 810	int i = 0;
 811	int ret = 0;
 812
 813	while (energy_perf_strings[i] != NULL)
 814		ret += sprintf(&buf[ret], "%s ", energy_perf_strings[i++]);
 815
 816	ret += sprintf(&buf[ret], "\n");
 817
 818	return ret;
 819}
 820
 821cpufreq_freq_attr_ro(energy_performance_available_preferences);
 822
 823static struct cpufreq_driver intel_pstate;
 824
 825static ssize_t store_energy_performance_preference(
 826		struct cpufreq_policy *policy, const char *buf, size_t count)
 827{
 828	struct cpudata *cpu = all_cpu_data[policy->cpu];
 829	char str_preference[21];
 830	bool raw = false;
 831	ssize_t ret;
 832	u32 epp = 0;
 833
 834	ret = sscanf(buf, "%20s", str_preference);
 835	if (ret != 1)
 836		return -EINVAL;
 837
 838	ret = match_string(energy_perf_strings, -1, str_preference);
 839	if (ret < 0) {
 840		if (!boot_cpu_has(X86_FEATURE_HWP_EPP))
 841			return ret;
 842
 843		ret = kstrtouint(buf, 10, &epp);
 844		if (ret)
 845			return ret;
 846
 847		if (epp > 255)
 848			return -EINVAL;
 849
 850		raw = true;
 851	}
 852
 853	/*
 854	 * This function runs with the policy R/W semaphore held, which
 855	 * guarantees that the driver pointer will not change while it is
 856	 * running.
 857	 */
 858	if (!intel_pstate_driver)
 859		return -EAGAIN;
 860
 861	mutex_lock(&intel_pstate_limits_lock);
 862
 863	if (intel_pstate_driver == &intel_pstate) {
 864		ret = intel_pstate_set_energy_pref_index(cpu, ret, raw, epp);
 865	} else {
 866		/*
 867		 * In the passive mode the governor needs to be stopped on the
 868		 * target CPU before the EPP update and restarted after it,
 869		 * which is super-heavy-weight, so make sure it is worth doing
 870		 * upfront.
 871		 */
 872		if (!raw)
 873			epp = ret ? epp_values[ret] : cpu->epp_default;
 874
 875		if (cpu->epp_cached != epp) {
 876			int err;
 877
 878			cpufreq_stop_governor(policy);
 879			ret = intel_pstate_set_epp(cpu, epp);
 880			err = cpufreq_start_governor(policy);
 881			if (!ret)
 882				ret = err;
 883		} else {
 884			ret = 0;
 885		}
 886	}
 887
 888	mutex_unlock(&intel_pstate_limits_lock);
 889
 890	return ret ?: count;
 891}
 892
 893static ssize_t show_energy_performance_preference(
 894				struct cpufreq_policy *policy, char *buf)
 895{
 896	struct cpudata *cpu_data = all_cpu_data[policy->cpu];
 897	int preference, raw_epp;
 898
 899	preference = intel_pstate_get_energy_pref_index(cpu_data, &raw_epp);
 900	if (preference < 0)
 901		return preference;
 902
 903	if (raw_epp)
 904		return  sprintf(buf, "%d\n", raw_epp);
 905	else
 906		return  sprintf(buf, "%s\n", energy_perf_strings[preference]);
 907}
 908
 909cpufreq_freq_attr_rw(energy_performance_preference);
 910
 911static ssize_t show_base_frequency(struct cpufreq_policy *policy, char *buf)
 912{
 913	struct cpudata *cpu = all_cpu_data[policy->cpu];
 914	int ratio, freq;
 915
 916	ratio = intel_pstate_get_cppc_guaranteed(policy->cpu);
 917	if (ratio <= 0) {
 918		u64 cap;
 919
 920		rdmsrl_on_cpu(policy->cpu, MSR_HWP_CAPABILITIES, &cap);
 921		ratio = HWP_GUARANTEED_PERF(cap);
 922	}
 923
 924	freq = ratio * cpu->pstate.scaling;
 925	if (cpu->pstate.scaling != cpu->pstate.perf_ctl_scaling)
 926		freq = rounddown(freq, cpu->pstate.perf_ctl_scaling);
 927
 928	return sprintf(buf, "%d\n", freq);
 929}
 930
 931cpufreq_freq_attr_ro(base_frequency);
 932
 933static struct freq_attr *hwp_cpufreq_attrs[] = {
 934	&energy_performance_preference,
 935	&energy_performance_available_preferences,
 936	&base_frequency,
 937	NULL,
 938};
 939
 940static void __intel_pstate_get_hwp_cap(struct cpudata *cpu)
 941{
 942	u64 cap;
 943
 944	rdmsrl_on_cpu(cpu->cpu, MSR_HWP_CAPABILITIES, &cap);
 945	WRITE_ONCE(cpu->hwp_cap_cached, cap);
 946	cpu->pstate.max_pstate = HWP_GUARANTEED_PERF(cap);
 947	cpu->pstate.turbo_pstate = HWP_HIGHEST_PERF(cap);
 948}
 949
 950static void intel_pstate_get_hwp_cap(struct cpudata *cpu)
 951{
 952	int scaling = cpu->pstate.scaling;
 953
 954	__intel_pstate_get_hwp_cap(cpu);
 955
 956	cpu->pstate.max_freq = cpu->pstate.max_pstate * scaling;
 957	cpu->pstate.turbo_freq = cpu->pstate.turbo_pstate * scaling;
 958	if (scaling != cpu->pstate.perf_ctl_scaling) {
 959		int perf_ctl_scaling = cpu->pstate.perf_ctl_scaling;
 960
 961		cpu->pstate.max_freq = rounddown(cpu->pstate.max_freq,
 962						 perf_ctl_scaling);
 963		cpu->pstate.turbo_freq = rounddown(cpu->pstate.turbo_freq,
 964						   perf_ctl_scaling);
 965	}
 966}
 967
 968static void intel_pstate_hwp_set(unsigned int cpu)
 969{
 970	struct cpudata *cpu_data = all_cpu_data[cpu];
 971	int max, min;
 972	u64 value;
 973	s16 epp;
 974
 975	max = cpu_data->max_perf_ratio;
 976	min = cpu_data->min_perf_ratio;
 977
 978	if (cpu_data->policy == CPUFREQ_POLICY_PERFORMANCE)
 979		min = max;
 980
 981	rdmsrl_on_cpu(cpu, MSR_HWP_REQUEST, &value);
 982
 983	value &= ~HWP_MIN_PERF(~0L);
 984	value |= HWP_MIN_PERF(min);
 985
 986	value &= ~HWP_MAX_PERF(~0L);
 987	value |= HWP_MAX_PERF(max);
 988
 989	if (cpu_data->epp_policy == cpu_data->policy)
 990		goto skip_epp;
 991
 992	cpu_data->epp_policy = cpu_data->policy;
 993
 994	if (cpu_data->policy == CPUFREQ_POLICY_PERFORMANCE) {
 995		epp = intel_pstate_get_epp(cpu_data, value);
 996		cpu_data->epp_powersave = epp;
 997		/* If EPP read was failed, then don't try to write */
 998		if (epp < 0)
 999			goto skip_epp;
1000
1001		epp = 0;
1002	} else {
1003		/* skip setting EPP, when saved value is invalid */
1004		if (cpu_data->epp_powersave < 0)
1005			goto skip_epp;
1006
1007		/*
1008		 * No need to restore EPP when it is not zero. This
1009		 * means:
1010		 *  - Policy is not changed
1011		 *  - user has manually changed
1012		 *  - Error reading EPB
1013		 */
1014		epp = intel_pstate_get_epp(cpu_data, value);
1015		if (epp)
1016			goto skip_epp;
1017
1018		epp = cpu_data->epp_powersave;
1019	}
1020	if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
1021		value &= ~GENMASK_ULL(31, 24);
1022		value |= (u64)epp << 24;
1023	} else {
1024		intel_pstate_set_epb(cpu, epp);
1025	}
1026skip_epp:
1027	WRITE_ONCE(cpu_data->hwp_req_cached, value);
1028	wrmsrl_on_cpu(cpu, MSR_HWP_REQUEST, value);
1029}
1030
1031static void intel_pstate_disable_hwp_interrupt(struct cpudata *cpudata);
1032
1033static void intel_pstate_hwp_offline(struct cpudata *cpu)
1034{
1035	u64 value = READ_ONCE(cpu->hwp_req_cached);
1036	int min_perf;
1037
1038	intel_pstate_disable_hwp_interrupt(cpu);
1039
1040	if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
1041		/*
1042		 * In case the EPP has been set to "performance" by the
1043		 * active mode "performance" scaling algorithm, replace that
1044		 * temporary value with the cached EPP one.
1045		 */
1046		value &= ~GENMASK_ULL(31, 24);
1047		value |= HWP_ENERGY_PERF_PREFERENCE(cpu->epp_cached);
1048		/*
1049		 * However, make sure that EPP will be set to "performance" when
1050		 * the CPU is brought back online again and the "performance"
1051		 * scaling algorithm is still in effect.
1052		 */
1053		cpu->epp_policy = CPUFREQ_POLICY_UNKNOWN;
1054	}
1055
1056	/*
1057	 * Clear the desired perf field in the cached HWP request value to
1058	 * prevent nonzero desired values from being leaked into the active
1059	 * mode.
1060	 */
1061	value &= ~HWP_DESIRED_PERF(~0L);
1062	WRITE_ONCE(cpu->hwp_req_cached, value);
1063
1064	value &= ~GENMASK_ULL(31, 0);
1065	min_perf = HWP_LOWEST_PERF(READ_ONCE(cpu->hwp_cap_cached));
1066
1067	/* Set hwp_max = hwp_min */
1068	value |= HWP_MAX_PERF(min_perf);
1069	value |= HWP_MIN_PERF(min_perf);
1070
1071	/* Set EPP to min */
1072	if (boot_cpu_has(X86_FEATURE_HWP_EPP))
1073		value |= HWP_ENERGY_PERF_PREFERENCE(HWP_EPP_POWERSAVE);
1074
1075	wrmsrl_on_cpu(cpu->cpu, MSR_HWP_REQUEST, value);
1076}
1077
1078#define POWER_CTL_EE_ENABLE	1
1079#define POWER_CTL_EE_DISABLE	2
1080
1081static int power_ctl_ee_state;
1082
1083static void set_power_ctl_ee_state(bool input)
1084{
1085	u64 power_ctl;
1086
1087	mutex_lock(&intel_pstate_driver_lock);
1088	rdmsrl(MSR_IA32_POWER_CTL, power_ctl);
1089	if (input) {
1090		power_ctl &= ~BIT(MSR_IA32_POWER_CTL_BIT_EE);
1091		power_ctl_ee_state = POWER_CTL_EE_ENABLE;
1092	} else {
1093		power_ctl |= BIT(MSR_IA32_POWER_CTL_BIT_EE);
1094		power_ctl_ee_state = POWER_CTL_EE_DISABLE;
1095	}
1096	wrmsrl(MSR_IA32_POWER_CTL, power_ctl);
1097	mutex_unlock(&intel_pstate_driver_lock);
1098}
1099
1100static void intel_pstate_hwp_enable(struct cpudata *cpudata);
1101
1102static void intel_pstate_hwp_reenable(struct cpudata *cpu)
1103{
1104	intel_pstate_hwp_enable(cpu);
1105	wrmsrl_on_cpu(cpu->cpu, MSR_HWP_REQUEST, READ_ONCE(cpu->hwp_req_cached));
1106}
1107
1108static int intel_pstate_suspend(struct cpufreq_policy *policy)
1109{
1110	struct cpudata *cpu = all_cpu_data[policy->cpu];
1111
1112	pr_debug("CPU %d suspending\n", cpu->cpu);
1113
1114	cpu->suspended = true;
1115
1116	/* disable HWP interrupt and cancel any pending work */
1117	intel_pstate_disable_hwp_interrupt(cpu);
1118
1119	return 0;
1120}
1121
1122static int intel_pstate_resume(struct cpufreq_policy *policy)
1123{
1124	struct cpudata *cpu = all_cpu_data[policy->cpu];
1125
1126	pr_debug("CPU %d resuming\n", cpu->cpu);
1127
1128	/* Only restore if the system default is changed */
1129	if (power_ctl_ee_state == POWER_CTL_EE_ENABLE)
1130		set_power_ctl_ee_state(true);
1131	else if (power_ctl_ee_state == POWER_CTL_EE_DISABLE)
1132		set_power_ctl_ee_state(false);
1133
1134	if (cpu->suspended && hwp_active) {
1135		mutex_lock(&intel_pstate_limits_lock);
1136
1137		/* Re-enable HWP, because "online" has not done that. */
1138		intel_pstate_hwp_reenable(cpu);
1139
1140		mutex_unlock(&intel_pstate_limits_lock);
1141	}
1142
1143	cpu->suspended = false;
1144
1145	return 0;
1146}
1147
1148static void intel_pstate_update_policies(void)
1149{
1150	int cpu;
1151
1152	for_each_possible_cpu(cpu)
1153		cpufreq_update_policy(cpu);
1154}
1155
1156static void __intel_pstate_update_max_freq(struct cpudata *cpudata,
1157					   struct cpufreq_policy *policy)
1158{
1159	policy->cpuinfo.max_freq = global.turbo_disabled_mf ?
1160			cpudata->pstate.max_freq : cpudata->pstate.turbo_freq;
1161	refresh_frequency_limits(policy);
1162}
1163
1164static void intel_pstate_update_max_freq(unsigned int cpu)
1165{
1166	struct cpufreq_policy *policy = cpufreq_cpu_acquire(cpu);
1167
1168	if (!policy)
1169		return;
1170
1171	__intel_pstate_update_max_freq(all_cpu_data[cpu], policy);
1172
1173	cpufreq_cpu_release(policy);
1174}
1175
1176static void intel_pstate_update_limits(unsigned int cpu)
1177{
1178	mutex_lock(&intel_pstate_driver_lock);
1179
1180	update_turbo_state();
1181	/*
1182	 * If turbo has been turned on or off globally, policy limits for
1183	 * all CPUs need to be updated to reflect that.
1184	 */
1185	if (global.turbo_disabled_mf != global.turbo_disabled) {
1186		global.turbo_disabled_mf = global.turbo_disabled;
1187		arch_set_max_freq_ratio(global.turbo_disabled);
1188		for_each_possible_cpu(cpu)
1189			intel_pstate_update_max_freq(cpu);
1190	} else {
1191		cpufreq_update_policy(cpu);
1192	}
1193
1194	mutex_unlock(&intel_pstate_driver_lock);
1195}
1196
1197/************************** sysfs begin ************************/
1198#define show_one(file_name, object)					\
1199	static ssize_t show_##file_name					\
1200	(struct kobject *kobj, struct kobj_attribute *attr, char *buf)	\
1201	{								\
1202		return sprintf(buf, "%u\n", global.object);		\
1203	}
1204
1205static ssize_t intel_pstate_show_status(char *buf);
1206static int intel_pstate_update_status(const char *buf, size_t size);
1207
1208static ssize_t show_status(struct kobject *kobj,
1209			   struct kobj_attribute *attr, char *buf)
1210{
1211	ssize_t ret;
1212
1213	mutex_lock(&intel_pstate_driver_lock);
1214	ret = intel_pstate_show_status(buf);
1215	mutex_unlock(&intel_pstate_driver_lock);
1216
1217	return ret;
1218}
1219
1220static ssize_t store_status(struct kobject *a, struct kobj_attribute *b,
1221			    const char *buf, size_t count)
1222{
1223	char *p = memchr(buf, '\n', count);
1224	int ret;
1225
1226	mutex_lock(&intel_pstate_driver_lock);
1227	ret = intel_pstate_update_status(buf, p ? p - buf : count);
1228	mutex_unlock(&intel_pstate_driver_lock);
1229
1230	return ret < 0 ? ret : count;
1231}
1232
1233static ssize_t show_turbo_pct(struct kobject *kobj,
1234				struct kobj_attribute *attr, char *buf)
1235{
1236	struct cpudata *cpu;
1237	int total, no_turbo, turbo_pct;
1238	uint32_t turbo_fp;
1239
1240	mutex_lock(&intel_pstate_driver_lock);
1241
1242	if (!intel_pstate_driver) {
1243		mutex_unlock(&intel_pstate_driver_lock);
1244		return -EAGAIN;
1245	}
1246
1247	cpu = all_cpu_data[0];
1248
1249	total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
1250	no_turbo = cpu->pstate.max_pstate - cpu->pstate.min_pstate + 1;
1251	turbo_fp = div_fp(no_turbo, total);
1252	turbo_pct = 100 - fp_toint(mul_fp(turbo_fp, int_tofp(100)));
1253
1254	mutex_unlock(&intel_pstate_driver_lock);
1255
1256	return sprintf(buf, "%u\n", turbo_pct);
1257}
1258
1259static ssize_t show_num_pstates(struct kobject *kobj,
1260				struct kobj_attribute *attr, char *buf)
1261{
1262	struct cpudata *cpu;
1263	int total;
1264
1265	mutex_lock(&intel_pstate_driver_lock);
1266
1267	if (!intel_pstate_driver) {
1268		mutex_unlock(&intel_pstate_driver_lock);
1269		return -EAGAIN;
1270	}
1271
1272	cpu = all_cpu_data[0];
1273	total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
1274
1275	mutex_unlock(&intel_pstate_driver_lock);
1276
1277	return sprintf(buf, "%u\n", total);
1278}
1279
1280static ssize_t show_no_turbo(struct kobject *kobj,
1281			     struct kobj_attribute *attr, char *buf)
1282{
1283	ssize_t ret;
1284
1285	mutex_lock(&intel_pstate_driver_lock);
1286
1287	if (!intel_pstate_driver) {
1288		mutex_unlock(&intel_pstate_driver_lock);
1289		return -EAGAIN;
1290	}
1291
1292	update_turbo_state();
1293	if (global.turbo_disabled)
1294		ret = sprintf(buf, "%u\n", global.turbo_disabled);
1295	else
1296		ret = sprintf(buf, "%u\n", global.no_turbo);
1297
1298	mutex_unlock(&intel_pstate_driver_lock);
1299
1300	return ret;
1301}
1302
1303static ssize_t store_no_turbo(struct kobject *a, struct kobj_attribute *b,
1304			      const char *buf, size_t count)
1305{
1306	unsigned int input;
1307	int ret;
1308
1309	ret = sscanf(buf, "%u", &input);
1310	if (ret != 1)
1311		return -EINVAL;
1312
1313	mutex_lock(&intel_pstate_driver_lock);
1314
1315	if (!intel_pstate_driver) {
1316		mutex_unlock(&intel_pstate_driver_lock);
1317		return -EAGAIN;
1318	}
1319
1320	mutex_lock(&intel_pstate_limits_lock);
1321
1322	update_turbo_state();
1323	if (global.turbo_disabled) {
1324		pr_notice_once("Turbo disabled by BIOS or unavailable on processor\n");
1325		mutex_unlock(&intel_pstate_limits_lock);
1326		mutex_unlock(&intel_pstate_driver_lock);
1327		return -EPERM;
1328	}
1329
1330	global.no_turbo = clamp_t(int, input, 0, 1);
1331
1332	if (global.no_turbo) {
1333		struct cpudata *cpu = all_cpu_data[0];
1334		int pct = cpu->pstate.max_pstate * 100 / cpu->pstate.turbo_pstate;
1335
1336		/* Squash the global minimum into the permitted range. */
1337		if (global.min_perf_pct > pct)
1338			global.min_perf_pct = pct;
1339	}
1340
1341	mutex_unlock(&intel_pstate_limits_lock);
1342
1343	intel_pstate_update_policies();
1344	arch_set_max_freq_ratio(global.no_turbo);
1345
1346	mutex_unlock(&intel_pstate_driver_lock);
1347
1348	return count;
1349}
1350
1351static void update_qos_request(enum freq_qos_req_type type)
1352{
1353	struct freq_qos_request *req;
1354	struct cpufreq_policy *policy;
1355	int i;
1356
1357	for_each_possible_cpu(i) {
1358		struct cpudata *cpu = all_cpu_data[i];
1359		unsigned int freq, perf_pct;
1360
1361		policy = cpufreq_cpu_get(i);
1362		if (!policy)
1363			continue;
1364
1365		req = policy->driver_data;
1366		cpufreq_cpu_put(policy);
1367
1368		if (!req)
1369			continue;
1370
1371		if (hwp_active)
1372			intel_pstate_get_hwp_cap(cpu);
1373
1374		if (type == FREQ_QOS_MIN) {
1375			perf_pct = global.min_perf_pct;
1376		} else {
1377			req++;
1378			perf_pct = global.max_perf_pct;
1379		}
1380
1381		freq = DIV_ROUND_UP(cpu->pstate.turbo_freq * perf_pct, 100);
1382
1383		if (freq_qos_update_request(req, freq) < 0)
1384			pr_warn("Failed to update freq constraint: CPU%d\n", i);
1385	}
1386}
1387
1388static ssize_t store_max_perf_pct(struct kobject *a, struct kobj_attribute *b,
1389				  const char *buf, size_t count)
1390{
1391	unsigned int input;
1392	int ret;
1393
1394	ret = sscanf(buf, "%u", &input);
1395	if (ret != 1)
1396		return -EINVAL;
1397
1398	mutex_lock(&intel_pstate_driver_lock);
1399
1400	if (!intel_pstate_driver) {
1401		mutex_unlock(&intel_pstate_driver_lock);
1402		return -EAGAIN;
1403	}
1404
1405	mutex_lock(&intel_pstate_limits_lock);
1406
1407	global.max_perf_pct = clamp_t(int, input, global.min_perf_pct, 100);
1408
1409	mutex_unlock(&intel_pstate_limits_lock);
1410
1411	if (intel_pstate_driver == &intel_pstate)
1412		intel_pstate_update_policies();
1413	else
1414		update_qos_request(FREQ_QOS_MAX);
1415
1416	mutex_unlock(&intel_pstate_driver_lock);
1417
1418	return count;
1419}
1420
1421static ssize_t store_min_perf_pct(struct kobject *a, struct kobj_attribute *b,
1422				  const char *buf, size_t count)
1423{
1424	unsigned int input;
1425	int ret;
1426
1427	ret = sscanf(buf, "%u", &input);
1428	if (ret != 1)
1429		return -EINVAL;
1430
1431	mutex_lock(&intel_pstate_driver_lock);
1432
1433	if (!intel_pstate_driver) {
1434		mutex_unlock(&intel_pstate_driver_lock);
1435		return -EAGAIN;
1436	}
1437
1438	mutex_lock(&intel_pstate_limits_lock);
1439
1440	global.min_perf_pct = clamp_t(int, input,
1441				      min_perf_pct_min(), global.max_perf_pct);
1442
1443	mutex_unlock(&intel_pstate_limits_lock);
1444
1445	if (intel_pstate_driver == &intel_pstate)
1446		intel_pstate_update_policies();
1447	else
1448		update_qos_request(FREQ_QOS_MIN);
1449
1450	mutex_unlock(&intel_pstate_driver_lock);
1451
1452	return count;
1453}
1454
1455static ssize_t show_hwp_dynamic_boost(struct kobject *kobj,
1456				struct kobj_attribute *attr, char *buf)
1457{
1458	return sprintf(buf, "%u\n", hwp_boost);
1459}
1460
1461static ssize_t store_hwp_dynamic_boost(struct kobject *a,
1462				       struct kobj_attribute *b,
1463				       const char *buf, size_t count)
1464{
1465	unsigned int input;
1466	int ret;
1467
1468	ret = kstrtouint(buf, 10, &input);
1469	if (ret)
1470		return ret;
1471
1472	mutex_lock(&intel_pstate_driver_lock);
1473	hwp_boost = !!input;
1474	intel_pstate_update_policies();
1475	mutex_unlock(&intel_pstate_driver_lock);
1476
1477	return count;
1478}
1479
1480static ssize_t show_energy_efficiency(struct kobject *kobj, struct kobj_attribute *attr,
1481				      char *buf)
1482{
1483	u64 power_ctl;
1484	int enable;
1485
1486	rdmsrl(MSR_IA32_POWER_CTL, power_ctl);
1487	enable = !!(power_ctl & BIT(MSR_IA32_POWER_CTL_BIT_EE));
1488	return sprintf(buf, "%d\n", !enable);
1489}
1490
1491static ssize_t store_energy_efficiency(struct kobject *a, struct kobj_attribute *b,
1492				       const char *buf, size_t count)
1493{
1494	bool input;
1495	int ret;
1496
1497	ret = kstrtobool(buf, &input);
1498	if (ret)
1499		return ret;
1500
1501	set_power_ctl_ee_state(input);
1502
1503	return count;
1504}
1505
1506show_one(max_perf_pct, max_perf_pct);
1507show_one(min_perf_pct, min_perf_pct);
1508
1509define_one_global_rw(status);
1510define_one_global_rw(no_turbo);
1511define_one_global_rw(max_perf_pct);
1512define_one_global_rw(min_perf_pct);
1513define_one_global_ro(turbo_pct);
1514define_one_global_ro(num_pstates);
1515define_one_global_rw(hwp_dynamic_boost);
1516define_one_global_rw(energy_efficiency);
1517
1518static struct attribute *intel_pstate_attributes[] = {
1519	&status.attr,
1520	&no_turbo.attr,
1521	NULL
1522};
1523
1524static const struct attribute_group intel_pstate_attr_group = {
1525	.attrs = intel_pstate_attributes,
1526};
1527
1528static const struct x86_cpu_id intel_pstate_cpu_ee_disable_ids[];
1529
1530static struct kobject *intel_pstate_kobject;
1531
1532static void __init intel_pstate_sysfs_expose_params(void)
1533{
1534	struct device *dev_root = bus_get_dev_root(&cpu_subsys);
1535	int rc;
1536
1537	if (dev_root) {
1538		intel_pstate_kobject = kobject_create_and_add("intel_pstate", &dev_root->kobj);
1539		put_device(dev_root);
1540	}
1541	if (WARN_ON(!intel_pstate_kobject))
1542		return;
1543
1544	rc = sysfs_create_group(intel_pstate_kobject, &intel_pstate_attr_group);
1545	if (WARN_ON(rc))
1546		return;
1547
1548	if (!boot_cpu_has(X86_FEATURE_HYBRID_CPU)) {
1549		rc = sysfs_create_file(intel_pstate_kobject, &turbo_pct.attr);
1550		WARN_ON(rc);
1551
1552		rc = sysfs_create_file(intel_pstate_kobject, &num_pstates.attr);
1553		WARN_ON(rc);
1554	}
1555
1556	/*
1557	 * If per cpu limits are enforced there are no global limits, so
1558	 * return without creating max/min_perf_pct attributes
1559	 */
1560	if (per_cpu_limits)
1561		return;
1562
1563	rc = sysfs_create_file(intel_pstate_kobject, &max_perf_pct.attr);
1564	WARN_ON(rc);
1565
1566	rc = sysfs_create_file(intel_pstate_kobject, &min_perf_pct.attr);
1567	WARN_ON(rc);
1568
1569	if (x86_match_cpu(intel_pstate_cpu_ee_disable_ids)) {
1570		rc = sysfs_create_file(intel_pstate_kobject, &energy_efficiency.attr);
1571		WARN_ON(rc);
1572	}
1573}
1574
1575static void __init intel_pstate_sysfs_remove(void)
1576{
1577	if (!intel_pstate_kobject)
1578		return;
1579
1580	sysfs_remove_group(intel_pstate_kobject, &intel_pstate_attr_group);
1581
1582	if (!boot_cpu_has(X86_FEATURE_HYBRID_CPU)) {
1583		sysfs_remove_file(intel_pstate_kobject, &num_pstates.attr);
1584		sysfs_remove_file(intel_pstate_kobject, &turbo_pct.attr);
1585	}
1586
1587	if (!per_cpu_limits) {
1588		sysfs_remove_file(intel_pstate_kobject, &max_perf_pct.attr);
1589		sysfs_remove_file(intel_pstate_kobject, &min_perf_pct.attr);
1590
1591		if (x86_match_cpu(intel_pstate_cpu_ee_disable_ids))
1592			sysfs_remove_file(intel_pstate_kobject, &energy_efficiency.attr);
1593	}
1594
1595	kobject_put(intel_pstate_kobject);
1596}
1597
1598static void intel_pstate_sysfs_expose_hwp_dynamic_boost(void)
1599{
1600	int rc;
1601
1602	if (!hwp_active)
1603		return;
1604
1605	rc = sysfs_create_file(intel_pstate_kobject, &hwp_dynamic_boost.attr);
1606	WARN_ON_ONCE(rc);
1607}
1608
1609static void intel_pstate_sysfs_hide_hwp_dynamic_boost(void)
1610{
1611	if (!hwp_active)
1612		return;
1613
1614	sysfs_remove_file(intel_pstate_kobject, &hwp_dynamic_boost.attr);
1615}
1616
1617/************************** sysfs end ************************/
1618
1619static void intel_pstate_notify_work(struct work_struct *work)
1620{
1621	struct cpudata *cpudata =
1622		container_of(to_delayed_work(work), struct cpudata, hwp_notify_work);
1623	struct cpufreq_policy *policy = cpufreq_cpu_acquire(cpudata->cpu);
1624
1625	if (policy) {
1626		intel_pstate_get_hwp_cap(cpudata);
1627		__intel_pstate_update_max_freq(cpudata, policy);
1628
1629		cpufreq_cpu_release(policy);
1630	}
1631
1632	wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_STATUS, 0);
1633}
1634
1635static DEFINE_SPINLOCK(hwp_notify_lock);
1636static cpumask_t hwp_intr_enable_mask;
1637
1638void notify_hwp_interrupt(void)
1639{
1640	unsigned int this_cpu = smp_processor_id();
1641	struct cpudata *cpudata;
1642	unsigned long flags;
1643	u64 value;
1644
1645	if (!READ_ONCE(hwp_active) || !boot_cpu_has(X86_FEATURE_HWP_NOTIFY))
1646		return;
1647
1648	rdmsrl_safe(MSR_HWP_STATUS, &value);
1649	if (!(value & 0x01))
1650		return;
1651
1652	spin_lock_irqsave(&hwp_notify_lock, flags);
1653
1654	if (!cpumask_test_cpu(this_cpu, &hwp_intr_enable_mask))
1655		goto ack_intr;
1656
1657	/*
1658	 * Currently we never free all_cpu_data. And we can't reach here
1659	 * without this allocated. But for safety for future changes, added
1660	 * check.
1661	 */
1662	if (unlikely(!READ_ONCE(all_cpu_data)))
1663		goto ack_intr;
1664
1665	/*
1666	 * The free is done during cleanup, when cpufreq registry is failed.
1667	 * We wouldn't be here if it fails on init or switch status. But for
1668	 * future changes, added check.
1669	 */
1670	cpudata = READ_ONCE(all_cpu_data[this_cpu]);
1671	if (unlikely(!cpudata))
1672		goto ack_intr;
1673
1674	schedule_delayed_work(&cpudata->hwp_notify_work, msecs_to_jiffies(10));
1675
1676	spin_unlock_irqrestore(&hwp_notify_lock, flags);
1677
1678	return;
1679
1680ack_intr:
1681	wrmsrl_safe(MSR_HWP_STATUS, 0);
1682	spin_unlock_irqrestore(&hwp_notify_lock, flags);
1683}
1684
1685static void intel_pstate_disable_hwp_interrupt(struct cpudata *cpudata)
1686{
1687	unsigned long flags;
1688
1689	if (!boot_cpu_has(X86_FEATURE_HWP_NOTIFY))
1690		return;
1691
1692	/* wrmsrl_on_cpu has to be outside spinlock as this can result in IPC */
1693	wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_INTERRUPT, 0x00);
1694
1695	spin_lock_irqsave(&hwp_notify_lock, flags);
1696	if (cpumask_test_and_clear_cpu(cpudata->cpu, &hwp_intr_enable_mask))
1697		cancel_delayed_work(&cpudata->hwp_notify_work);
1698	spin_unlock_irqrestore(&hwp_notify_lock, flags);
1699}
1700
1701static void intel_pstate_enable_hwp_interrupt(struct cpudata *cpudata)
1702{
1703	/* Enable HWP notification interrupt for guaranteed performance change */
1704	if (boot_cpu_has(X86_FEATURE_HWP_NOTIFY)) {
1705		unsigned long flags;
1706
1707		spin_lock_irqsave(&hwp_notify_lock, flags);
1708		INIT_DELAYED_WORK(&cpudata->hwp_notify_work, intel_pstate_notify_work);
1709		cpumask_set_cpu(cpudata->cpu, &hwp_intr_enable_mask);
1710		spin_unlock_irqrestore(&hwp_notify_lock, flags);
1711
1712		/* wrmsrl_on_cpu has to be outside spinlock as this can result in IPC */
1713		wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_INTERRUPT, 0x01);
1714		wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_STATUS, 0);
1715	}
1716}
1717
1718static void intel_pstate_update_epp_defaults(struct cpudata *cpudata)
1719{
1720	cpudata->epp_default = intel_pstate_get_epp(cpudata, 0);
1721
1722	/*
1723	 * If the EPP is set by firmware, which means that firmware enabled HWP
1724	 * - Is equal or less than 0x80 (default balance_perf EPP)
1725	 * - But less performance oriented than performance EPP
1726	 *   then use this as new balance_perf EPP.
1727	 */
1728	if (hwp_forced && cpudata->epp_default <= HWP_EPP_BALANCE_PERFORMANCE &&
1729	    cpudata->epp_default > HWP_EPP_PERFORMANCE) {
1730		epp_values[EPP_INDEX_BALANCE_PERFORMANCE] = cpudata->epp_default;
1731		return;
1732	}
1733
1734	/*
1735	 * If this CPU gen doesn't call for change in balance_perf
1736	 * EPP return.
1737	 */
1738	if (epp_values[EPP_INDEX_BALANCE_PERFORMANCE] == HWP_EPP_BALANCE_PERFORMANCE)
1739		return;
1740
1741	/*
1742	 * Use hard coded value per gen to update the balance_perf
1743	 * and default EPP.
1744	 */
1745	cpudata->epp_default = epp_values[EPP_INDEX_BALANCE_PERFORMANCE];
1746	intel_pstate_set_epp(cpudata, cpudata->epp_default);
1747}
1748
1749static void intel_pstate_hwp_enable(struct cpudata *cpudata)
1750{
1751	/* First disable HWP notification interrupt till we activate again */
1752	if (boot_cpu_has(X86_FEATURE_HWP_NOTIFY))
1753		wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_INTERRUPT, 0x00);
1754
1755	wrmsrl_on_cpu(cpudata->cpu, MSR_PM_ENABLE, 0x1);
1756
1757	intel_pstate_enable_hwp_interrupt(cpudata);
1758
1759	if (cpudata->epp_default >= 0)
1760		return;
1761
1762	intel_pstate_update_epp_defaults(cpudata);
1763}
1764
1765static int atom_get_min_pstate(int not_used)
1766{
1767	u64 value;
1768
1769	rdmsrl(MSR_ATOM_CORE_RATIOS, value);
1770	return (value >> 8) & 0x7F;
1771}
1772
1773static int atom_get_max_pstate(int not_used)
1774{
1775	u64 value;
1776
1777	rdmsrl(MSR_ATOM_CORE_RATIOS, value);
1778	return (value >> 16) & 0x7F;
1779}
1780
1781static int atom_get_turbo_pstate(int not_used)
1782{
1783	u64 value;
1784
1785	rdmsrl(MSR_ATOM_CORE_TURBO_RATIOS, value);
1786	return value & 0x7F;
1787}
1788
1789static u64 atom_get_val(struct cpudata *cpudata, int pstate)
1790{
1791	u64 val;
1792	int32_t vid_fp;
1793	u32 vid;
1794
1795	val = (u64)pstate << 8;
1796	if (global.no_turbo && !global.turbo_disabled)
1797		val |= (u64)1 << 32;
1798
1799	vid_fp = cpudata->vid.min + mul_fp(
1800		int_tofp(pstate - cpudata->pstate.min_pstate),
1801		cpudata->vid.ratio);
1802
1803	vid_fp = clamp_t(int32_t, vid_fp, cpudata->vid.min, cpudata->vid.max);
1804	vid = ceiling_fp(vid_fp);
1805
1806	if (pstate > cpudata->pstate.max_pstate)
1807		vid = cpudata->vid.turbo;
1808
1809	return val | vid;
1810}
1811
1812static int silvermont_get_scaling(void)
1813{
1814	u64 value;
1815	int i;
1816	/* Defined in Table 35-6 from SDM (Sept 2015) */
1817	static int silvermont_freq_table[] = {
1818		83300, 100000, 133300, 116700, 80000};
1819
1820	rdmsrl(MSR_FSB_FREQ, value);
1821	i = value & 0x7;
1822	WARN_ON(i > 4);
1823
1824	return silvermont_freq_table[i];
1825}
1826
1827static int airmont_get_scaling(void)
1828{
1829	u64 value;
1830	int i;
1831	/* Defined in Table 35-10 from SDM (Sept 2015) */
1832	static int airmont_freq_table[] = {
1833		83300, 100000, 133300, 116700, 80000,
1834		93300, 90000, 88900, 87500};
1835
1836	rdmsrl(MSR_FSB_FREQ, value);
1837	i = value & 0xF;
1838	WARN_ON(i > 8);
1839
1840	return airmont_freq_table[i];
1841}
1842
1843static void atom_get_vid(struct cpudata *cpudata)
1844{
1845	u64 value;
1846
1847	rdmsrl(MSR_ATOM_CORE_VIDS, value);
1848	cpudata->vid.min = int_tofp((value >> 8) & 0x7f);
1849	cpudata->vid.max = int_tofp((value >> 16) & 0x7f);
1850	cpudata->vid.ratio = div_fp(
1851		cpudata->vid.max - cpudata->vid.min,
1852		int_tofp(cpudata->pstate.max_pstate -
1853			cpudata->pstate.min_pstate));
1854
1855	rdmsrl(MSR_ATOM_CORE_TURBO_VIDS, value);
1856	cpudata->vid.turbo = value & 0x7f;
1857}
1858
1859static int core_get_min_pstate(int cpu)
1860{
1861	u64 value;
1862
1863	rdmsrl_on_cpu(cpu, MSR_PLATFORM_INFO, &value);
1864	return (value >> 40) & 0xFF;
1865}
1866
1867static int core_get_max_pstate_physical(int cpu)
1868{
1869	u64 value;
1870
1871	rdmsrl_on_cpu(cpu, MSR_PLATFORM_INFO, &value);
1872	return (value >> 8) & 0xFF;
1873}
1874
1875static int core_get_tdp_ratio(int cpu, u64 plat_info)
1876{
1877	/* Check how many TDP levels present */
1878	if (plat_info & 0x600000000) {
1879		u64 tdp_ctrl;
1880		u64 tdp_ratio;
1881		int tdp_msr;
1882		int err;
1883
1884		/* Get the TDP level (0, 1, 2) to get ratios */
1885		err = rdmsrl_safe_on_cpu(cpu, MSR_CONFIG_TDP_CONTROL, &tdp_ctrl);
1886		if (err)
1887			return err;
1888
1889		/* TDP MSR are continuous starting at 0x648 */
1890		tdp_msr = MSR_CONFIG_TDP_NOMINAL + (tdp_ctrl & 0x03);
1891		err = rdmsrl_safe_on_cpu(cpu, tdp_msr, &tdp_ratio);
1892		if (err)
1893			return err;
1894
1895		/* For level 1 and 2, bits[23:16] contain the ratio */
1896		if (tdp_ctrl & 0x03)
1897			tdp_ratio >>= 16;
1898
1899		tdp_ratio &= 0xff; /* ratios are only 8 bits long */
1900		pr_debug("tdp_ratio %x\n", (int)tdp_ratio);
1901
1902		return (int)tdp_ratio;
1903	}
1904
1905	return -ENXIO;
1906}
1907
1908static int core_get_max_pstate(int cpu)
1909{
1910	u64 tar;
1911	u64 plat_info;
1912	int max_pstate;
1913	int tdp_ratio;
1914	int err;
1915
1916	rdmsrl_on_cpu(cpu, MSR_PLATFORM_INFO, &plat_info);
1917	max_pstate = (plat_info >> 8) & 0xFF;
1918
1919	tdp_ratio = core_get_tdp_ratio(cpu, plat_info);
1920	if (tdp_ratio <= 0)
1921		return max_pstate;
1922
1923	if (hwp_active) {
1924		/* Turbo activation ratio is not used on HWP platforms */
1925		return tdp_ratio;
1926	}
1927
1928	err = rdmsrl_safe_on_cpu(cpu, MSR_TURBO_ACTIVATION_RATIO, &tar);
1929	if (!err) {
1930		int tar_levels;
1931
1932		/* Do some sanity checking for safety */
1933		tar_levels = tar & 0xff;
1934		if (tdp_ratio - 1 == tar_levels) {
1935			max_pstate = tar_levels;
1936			pr_debug("max_pstate=TAC %x\n", max_pstate);
1937		}
1938	}
1939
1940	return max_pstate;
1941}
1942
1943static int core_get_turbo_pstate(int cpu)
1944{
1945	u64 value;
1946	int nont, ret;
1947
1948	rdmsrl_on_cpu(cpu, MSR_TURBO_RATIO_LIMIT, &value);
1949	nont = core_get_max_pstate(cpu);
1950	ret = (value) & 255;
1951	if (ret <= nont)
1952		ret = nont;
1953	return ret;
1954}
1955
1956static u64 core_get_val(struct cpudata *cpudata, int pstate)
1957{
1958	u64 val;
1959
1960	val = (u64)pstate << 8;
1961	if (global.no_turbo && !global.turbo_disabled)
1962		val |= (u64)1 << 32;
1963
1964	return val;
1965}
1966
1967static int knl_get_aperf_mperf_shift(void)
1968{
1969	return 10;
1970}
1971
1972static int knl_get_turbo_pstate(int cpu)
1973{
1974	u64 value;
1975	int nont, ret;
1976
1977	rdmsrl_on_cpu(cpu, MSR_TURBO_RATIO_LIMIT, &value);
1978	nont = core_get_max_pstate(cpu);
1979	ret = (((value) >> 8) & 0xFF);
1980	if (ret <= nont)
1981		ret = nont;
1982	return ret;
1983}
1984
1985static void hybrid_get_type(void *data)
1986{
1987	u8 *cpu_type = data;
1988
1989	*cpu_type = get_this_hybrid_cpu_type();
1990}
1991
1992static int hwp_get_cpu_scaling(int cpu)
1993{
1994	u8 cpu_type = 0;
1995
1996	smp_call_function_single(cpu, hybrid_get_type, &cpu_type, 1);
1997	/* P-cores have a smaller perf level-to-freqency scaling factor. */
1998	if (cpu_type == 0x40)
1999		return hybrid_scaling_factor;
2000
2001	/* Use default core scaling for E-cores */
2002	if (cpu_type == 0x20)
2003		return core_get_scaling();
2004
2005	/*
2006	 * If reached here, this system is either non-hybrid (like Tiger
2007	 * Lake) or hybrid-capable (like Alder Lake or Raptor Lake) with
2008	 * no E cores (in which case CPUID for hybrid support is 0).
2009	 *
2010	 * The CPPC nominal_frequency field is 0 for non-hybrid systems,
2011	 * so the default core scaling will be used for them.
2012	 */
2013	return intel_pstate_cppc_get_scaling(cpu);
2014}
2015
2016static void intel_pstate_set_pstate(struct cpudata *cpu, int pstate)
2017{
2018	trace_cpu_frequency(pstate * cpu->pstate.scaling, cpu->cpu);
2019	cpu->pstate.current_pstate = pstate;
2020	/*
2021	 * Generally, there is no guarantee that this code will always run on
2022	 * the CPU being updated, so force the register update to run on the
2023	 * right CPU.
2024	 */
2025	wrmsrl_on_cpu(cpu->cpu, MSR_IA32_PERF_CTL,
2026		      pstate_funcs.get_val(cpu, pstate));
2027}
2028
2029static void intel_pstate_set_min_pstate(struct cpudata *cpu)
2030{
2031	intel_pstate_set_pstate(cpu, cpu->pstate.min_pstate);
2032}
2033
2034static void intel_pstate_max_within_limits(struct cpudata *cpu)
2035{
2036	int pstate = max(cpu->pstate.min_pstate, cpu->max_perf_ratio);
2037
2038	update_turbo_state();
2039	intel_pstate_set_pstate(cpu, pstate);
2040}
2041
2042static void intel_pstate_get_cpu_pstates(struct cpudata *cpu)
2043{
2044	int perf_ctl_max_phys = pstate_funcs.get_max_physical(cpu->cpu);
2045	int perf_ctl_scaling = pstate_funcs.get_scaling();
2046
2047	cpu->pstate.min_pstate = pstate_funcs.get_min(cpu->cpu);
2048	cpu->pstate.max_pstate_physical = perf_ctl_max_phys;
2049	cpu->pstate.perf_ctl_scaling = perf_ctl_scaling;
2050
2051	if (hwp_active && !hwp_mode_bdw) {
2052		__intel_pstate_get_hwp_cap(cpu);
2053
2054		if (pstate_funcs.get_cpu_scaling) {
2055			cpu->pstate.scaling = pstate_funcs.get_cpu_scaling(cpu->cpu);
2056			if (cpu->pstate.scaling != perf_ctl_scaling)
2057				intel_pstate_hybrid_hwp_adjust(cpu);
2058		} else {
2059			cpu->pstate.scaling = perf_ctl_scaling;
2060		}
2061	} else {
2062		cpu->pstate.scaling = perf_ctl_scaling;
2063		cpu->pstate.max_pstate = pstate_funcs.get_max(cpu->cpu);
2064		cpu->pstate.turbo_pstate = pstate_funcs.get_turbo(cpu->cpu);
2065	}
2066
2067	if (cpu->pstate.scaling == perf_ctl_scaling) {
2068		cpu->pstate.min_freq = cpu->pstate.min_pstate * perf_ctl_scaling;
2069		cpu->pstate.max_freq = cpu->pstate.max_pstate * perf_ctl_scaling;
2070		cpu->pstate.turbo_freq = cpu->pstate.turbo_pstate * perf_ctl_scaling;
2071	}
2072
2073	if (pstate_funcs.get_aperf_mperf_shift)
2074		cpu->aperf_mperf_shift = pstate_funcs.get_aperf_mperf_shift();
2075
2076	if (pstate_funcs.get_vid)
2077		pstate_funcs.get_vid(cpu);
2078
2079	intel_pstate_set_min_pstate(cpu);
2080}
2081
2082/*
2083 * Long hold time will keep high perf limits for long time,
2084 * which negatively impacts perf/watt for some workloads,
2085 * like specpower. 3ms is based on experiements on some
2086 * workoads.
2087 */
2088static int hwp_boost_hold_time_ns = 3 * NSEC_PER_MSEC;
2089
2090static inline void intel_pstate_hwp_boost_up(struct cpudata *cpu)
2091{
2092	u64 hwp_req = READ_ONCE(cpu->hwp_req_cached);
2093	u64 hwp_cap = READ_ONCE(cpu->hwp_cap_cached);
2094	u32 max_limit = (hwp_req & 0xff00) >> 8;
2095	u32 min_limit = (hwp_req & 0xff);
2096	u32 boost_level1;
2097
2098	/*
2099	 * Cases to consider (User changes via sysfs or boot time):
2100	 * If, P0 (Turbo max) = P1 (Guaranteed max) = min:
2101	 *	No boost, return.
2102	 * If, P0 (Turbo max) > P1 (Guaranteed max) = min:
2103	 *     Should result in one level boost only for P0.
2104	 * If, P0 (Turbo max) = P1 (Guaranteed max) > min:
2105	 *     Should result in two level boost:
2106	 *         (min + p1)/2 and P1.
2107	 * If, P0 (Turbo max) > P1 (Guaranteed max) > min:
2108	 *     Should result in three level boost:
2109	 *        (min + p1)/2, P1 and P0.
2110	 */
2111
2112	/* If max and min are equal or already at max, nothing to boost */
2113	if (max_limit == min_limit || cpu->hwp_boost_min >= max_limit)
2114		return;
2115
2116	if (!cpu->hwp_boost_min)
2117		cpu->hwp_boost_min = min_limit;
2118
2119	/* level at half way mark between min and guranteed */
2120	boost_level1 = (HWP_GUARANTEED_PERF(hwp_cap) + min_limit) >> 1;
2121
2122	if (cpu->hwp_boost_min < boost_level1)
2123		cpu->hwp_boost_min = boost_level1;
2124	else if (cpu->hwp_boost_min < HWP_GUARANTEED_PERF(hwp_cap))
2125		cpu->hwp_boost_min = HWP_GUARANTEED_PERF(hwp_cap);
2126	else if (cpu->hwp_boost_min == HWP_GUARANTEED_PERF(hwp_cap) &&
2127		 max_limit != HWP_GUARANTEED_PERF(hwp_cap))
2128		cpu->hwp_boost_min = max_limit;
2129	else
2130		return;
2131
2132	hwp_req = (hwp_req & ~GENMASK_ULL(7, 0)) | cpu->hwp_boost_min;
2133	wrmsrl(MSR_HWP_REQUEST, hwp_req);
2134	cpu->last_update = cpu->sample.time;
2135}
2136
2137static inline void intel_pstate_hwp_boost_down(struct cpudata *cpu)
2138{
2139	if (cpu->hwp_boost_min) {
2140		bool expired;
2141
2142		/* Check if we are idle for hold time to boost down */
2143		expired = time_after64(cpu->sample.time, cpu->last_update +
2144				       hwp_boost_hold_time_ns);
2145		if (expired) {
2146			wrmsrl(MSR_HWP_REQUEST, cpu->hwp_req_cached);
2147			cpu->hwp_boost_min = 0;
2148		}
2149	}
2150	cpu->last_update = cpu->sample.time;
2151}
2152
2153static inline void intel_pstate_update_util_hwp_local(struct cpudata *cpu,
2154						      u64 time)
2155{
2156	cpu->sample.time = time;
2157
2158	if (cpu->sched_flags & SCHED_CPUFREQ_IOWAIT) {
2159		bool do_io = false;
2160
2161		cpu->sched_flags = 0;
2162		/*
2163		 * Set iowait_boost flag and update time. Since IO WAIT flag
2164		 * is set all the time, we can't just conclude that there is
2165		 * some IO bound activity is scheduled on this CPU with just
2166		 * one occurrence. If we receive at least two in two
2167		 * consecutive ticks, then we treat as boost candidate.
2168		 */
2169		if (time_before64(time, cpu->last_io_update + 2 * TICK_NSEC))
2170			do_io = true;
2171
2172		cpu->last_io_update = time;
2173
2174		if (do_io)
2175			intel_pstate_hwp_boost_up(cpu);
2176
2177	} else {
2178		intel_pstate_hwp_boost_down(cpu);
2179	}
2180}
2181
2182static inline void intel_pstate_update_util_hwp(struct update_util_data *data,
2183						u64 time, unsigned int flags)
2184{
2185	struct cpudata *cpu = container_of(data, struct cpudata, update_util);
2186
2187	cpu->sched_flags |= flags;
2188
2189	if (smp_processor_id() == cpu->cpu)
2190		intel_pstate_update_util_hwp_local(cpu, time);
2191}
2192
2193static inline void intel_pstate_calc_avg_perf(struct cpudata *cpu)
2194{
2195	struct sample *sample = &cpu->sample;
2196
2197	sample->core_avg_perf = div_ext_fp(sample->aperf, sample->mperf);
2198}
2199
2200static inline bool intel_pstate_sample(struct cpudata *cpu, u64 time)
2201{
2202	u64 aperf, mperf;
2203	unsigned long flags;
2204	u64 tsc;
2205
2206	local_irq_save(flags);
2207	rdmsrl(MSR_IA32_APERF, aperf);
2208	rdmsrl(MSR_IA32_MPERF, mperf);
2209	tsc = rdtsc();
2210	if (cpu->prev_mperf == mperf || cpu->prev_tsc == tsc) {
2211		local_irq_restore(flags);
2212		return false;
2213	}
2214	local_irq_restore(flags);
2215
2216	cpu->last_sample_time = cpu->sample.time;
2217	cpu->sample.time = time;
2218	cpu->sample.aperf = aperf;
2219	cpu->sample.mperf = mperf;
2220	cpu->sample.tsc =  tsc;
2221	cpu->sample.aperf -= cpu->prev_aperf;
2222	cpu->sample.mperf -= cpu->prev_mperf;
2223	cpu->sample.tsc -= cpu->prev_tsc;
2224
2225	cpu->prev_aperf = aperf;
2226	cpu->prev_mperf = mperf;
2227	cpu->prev_tsc = tsc;
2228	/*
2229	 * First time this function is invoked in a given cycle, all of the
2230	 * previous sample data fields are equal to zero or stale and they must
2231	 * be populated with meaningful numbers for things to work, so assume
2232	 * that sample.time will always be reset before setting the utilization
2233	 * update hook and make the caller skip the sample then.
2234	 */
2235	if (cpu->last_sample_time) {
2236		intel_pstate_calc_avg_perf(cpu);
2237		return true;
2238	}
2239	return false;
2240}
2241
2242static inline int32_t get_avg_frequency(struct cpudata *cpu)
2243{
2244	return mul_ext_fp(cpu->sample.core_avg_perf, cpu_khz);
2245}
2246
2247static inline int32_t get_avg_pstate(struct cpudata *cpu)
2248{
2249	return mul_ext_fp(cpu->pstate.max_pstate_physical,
2250			  cpu->sample.core_avg_perf);
2251}
2252
2253static inline int32_t get_target_pstate(struct cpudata *cpu)
2254{
2255	struct sample *sample = &cpu->sample;
2256	int32_t busy_frac;
2257	int target, avg_pstate;
2258
2259	busy_frac = div_fp(sample->mperf << cpu->aperf_mperf_shift,
2260			   sample->tsc);
2261
2262	if (busy_frac < cpu->iowait_boost)
2263		busy_frac = cpu->iowait_boost;
2264
2265	sample->busy_scaled = busy_frac * 100;
2266
2267	target = global.no_turbo || global.turbo_disabled ?
2268			cpu->pstate.max_pstate : cpu->pstate.turbo_pstate;
2269	target += target >> 2;
2270	target = mul_fp(target, busy_frac);
2271	if (target < cpu->pstate.min_pstate)
2272		target = cpu->pstate.min_pstate;
2273
2274	/*
2275	 * If the average P-state during the previous cycle was higher than the
2276	 * current target, add 50% of the difference to the target to reduce
2277	 * possible performance oscillations and offset possible performance
2278	 * loss related to moving the workload from one CPU to another within
2279	 * a package/module.
2280	 */
2281	avg_pstate = get_avg_pstate(cpu);
2282	if (avg_pstate > target)
2283		target += (avg_pstate - target) >> 1;
2284
2285	return target;
2286}
2287
2288static int intel_pstate_prepare_request(struct cpudata *cpu, int pstate)
2289{
2290	int min_pstate = max(cpu->pstate.min_pstate, cpu->min_perf_ratio);
2291	int max_pstate = max(min_pstate, cpu->max_perf_ratio);
2292
2293	return clamp_t(int, pstate, min_pstate, max_pstate);
2294}
2295
2296static void intel_pstate_update_pstate(struct cpudata *cpu, int pstate)
2297{
2298	if (pstate == cpu->pstate.current_pstate)
2299		return;
2300
2301	cpu->pstate.current_pstate = pstate;
2302	wrmsrl(MSR_IA32_PERF_CTL, pstate_funcs.get_val(cpu, pstate));
2303}
2304
2305static void intel_pstate_adjust_pstate(struct cpudata *cpu)
2306{
2307	int from = cpu->pstate.current_pstate;
2308	struct sample *sample;
2309	int target_pstate;
2310
2311	update_turbo_state();
2312
2313	target_pstate = get_target_pstate(cpu);
2314	target_pstate = intel_pstate_prepare_request(cpu, target_pstate);
2315	trace_cpu_frequency(target_pstate * cpu->pstate.scaling, cpu->cpu);
2316	intel_pstate_update_pstate(cpu, target_pstate);
2317
2318	sample = &cpu->sample;
2319	trace_pstate_sample(mul_ext_fp(100, sample->core_avg_perf),
2320		fp_toint(sample->busy_scaled),
2321		from,
2322		cpu->pstate.current_pstate,
2323		sample->mperf,
2324		sample->aperf,
2325		sample->tsc,
2326		get_avg_frequency(cpu),
2327		fp_toint(cpu->iowait_boost * 100));
2328}
2329
2330static void intel_pstate_update_util(struct update_util_data *data, u64 time,
2331				     unsigned int flags)
2332{
2333	struct cpudata *cpu = container_of(data, struct cpudata, update_util);
2334	u64 delta_ns;
2335
2336	/* Don't allow remote callbacks */
2337	if (smp_processor_id() != cpu->cpu)
2338		return;
2339
2340	delta_ns = time - cpu->last_update;
2341	if (flags & SCHED_CPUFREQ_IOWAIT) {
2342		/* Start over if the CPU may have been idle. */
2343		if (delta_ns > TICK_NSEC) {
2344			cpu->iowait_boost = ONE_EIGHTH_FP;
2345		} else if (cpu->iowait_boost >= ONE_EIGHTH_FP) {
2346			cpu->iowait_boost <<= 1;
2347			if (cpu->iowait_boost > int_tofp(1))
2348				cpu->iowait_boost = int_tofp(1);
2349		} else {
2350			cpu->iowait_boost = ONE_EIGHTH_FP;
2351		}
2352	} else if (cpu->iowait_boost) {
2353		/* Clear iowait_boost if the CPU may have been idle. */
2354		if (delta_ns > TICK_NSEC)
2355			cpu->iowait_boost = 0;
2356		else
2357			cpu->iowait_boost >>= 1;
2358	}
2359	cpu->last_update = time;
2360	delta_ns = time - cpu->sample.time;
2361	if ((s64)delta_ns < INTEL_PSTATE_SAMPLING_INTERVAL)
2362		return;
2363
2364	if (intel_pstate_sample(cpu, time))
2365		intel_pstate_adjust_pstate(cpu);
2366}
2367
2368static struct pstate_funcs core_funcs = {
2369	.get_max = core_get_max_pstate,
2370	.get_max_physical = core_get_max_pstate_physical,
2371	.get_min = core_get_min_pstate,
2372	.get_turbo = core_get_turbo_pstate,
2373	.get_scaling = core_get_scaling,
2374	.get_val = core_get_val,
2375};
2376
2377static const struct pstate_funcs silvermont_funcs = {
2378	.get_max = atom_get_max_pstate,
2379	.get_max_physical = atom_get_max_pstate,
2380	.get_min = atom_get_min_pstate,
2381	.get_turbo = atom_get_turbo_pstate,
2382	.get_val = atom_get_val,
2383	.get_scaling = silvermont_get_scaling,
2384	.get_vid = atom_get_vid,
2385};
2386
2387static const struct pstate_funcs airmont_funcs = {
2388	.get_max = atom_get_max_pstate,
2389	.get_max_physical = atom_get_max_pstate,
2390	.get_min = atom_get_min_pstate,
2391	.get_turbo = atom_get_turbo_pstate,
2392	.get_val = atom_get_val,
2393	.get_scaling = airmont_get_scaling,
2394	.get_vid = atom_get_vid,
2395};
2396
2397static const struct pstate_funcs knl_funcs = {
2398	.get_max = core_get_max_pstate,
2399	.get_max_physical = core_get_max_pstate_physical,
2400	.get_min = core_get_min_pstate,
2401	.get_turbo = knl_get_turbo_pstate,
2402	.get_aperf_mperf_shift = knl_get_aperf_mperf_shift,
2403	.get_scaling = core_get_scaling,
2404	.get_val = core_get_val,
2405};
2406
2407#define X86_MATCH(model, policy)					 \
2408	X86_MATCH_VENDOR_FAM_MODEL_FEATURE(INTEL, 6, INTEL_FAM6_##model, \
2409					   X86_FEATURE_APERFMPERF, &policy)
2410
2411static const struct x86_cpu_id intel_pstate_cpu_ids[] = {
2412	X86_MATCH(SANDYBRIDGE,		core_funcs),
2413	X86_MATCH(SANDYBRIDGE_X,	core_funcs),
2414	X86_MATCH(ATOM_SILVERMONT,	silvermont_funcs),
2415	X86_MATCH(IVYBRIDGE,		core_funcs),
2416	X86_MATCH(HASWELL,		core_funcs),
2417	X86_MATCH(BROADWELL,		core_funcs),
2418	X86_MATCH(IVYBRIDGE_X,		core_funcs),
2419	X86_MATCH(HASWELL_X,		core_funcs),
2420	X86_MATCH(HASWELL_L,		core_funcs),
2421	X86_MATCH(HASWELL_G,		core_funcs),
2422	X86_MATCH(BROADWELL_G,		core_funcs),
2423	X86_MATCH(ATOM_AIRMONT,		airmont_funcs),
2424	X86_MATCH(SKYLAKE_L,		core_funcs),
2425	X86_MATCH(BROADWELL_X,		core_funcs),
2426	X86_MATCH(SKYLAKE,		core_funcs),
2427	X86_MATCH(BROADWELL_D,		core_funcs),
2428	X86_MATCH(XEON_PHI_KNL,		knl_funcs),
2429	X86_MATCH(XEON_PHI_KNM,		knl_funcs),
2430	X86_MATCH(ATOM_GOLDMONT,	core_funcs),
2431	X86_MATCH(ATOM_GOLDMONT_PLUS,	core_funcs),
2432	X86_MATCH(SKYLAKE_X,		core_funcs),
2433	X86_MATCH(COMETLAKE,		core_funcs),
2434	X86_MATCH(ICELAKE_X,		core_funcs),
2435	X86_MATCH(TIGERLAKE,		core_funcs),
2436	X86_MATCH(SAPPHIRERAPIDS_X,	core_funcs),
2437	X86_MATCH(EMERALDRAPIDS_X,      core_funcs),
2438	{}
2439};
2440MODULE_DEVICE_TABLE(x86cpu, intel_pstate_cpu_ids);
2441
2442static const struct x86_cpu_id intel_pstate_cpu_oob_ids[] __initconst = {
2443	X86_MATCH(BROADWELL_D,		core_funcs),
2444	X86_MATCH(BROADWELL_X,		core_funcs),
2445	X86_MATCH(SKYLAKE_X,		core_funcs),
2446	X86_MATCH(ICELAKE_X,		core_funcs),
2447	X86_MATCH(SAPPHIRERAPIDS_X,	core_funcs),
2448	{}
2449};
2450
2451static const struct x86_cpu_id intel_pstate_cpu_ee_disable_ids[] = {
2452	X86_MATCH(KABYLAKE,		core_funcs),
2453	{}
2454};
2455
2456static int intel_pstate_init_cpu(unsigned int cpunum)
2457{
2458	struct cpudata *cpu;
2459
2460	cpu = all_cpu_data[cpunum];
2461
2462	if (!cpu) {
2463		cpu = kzalloc(sizeof(*cpu), GFP_KERNEL);
2464		if (!cpu)
2465			return -ENOMEM;
2466
2467		WRITE_ONCE(all_cpu_data[cpunum], cpu);
2468
2469		cpu->cpu = cpunum;
2470
2471		cpu->epp_default = -EINVAL;
2472
2473		if (hwp_active) {
2474			intel_pstate_hwp_enable(cpu);
2475
2476			if (intel_pstate_acpi_pm_profile_server())
2477				hwp_boost = true;
2478		}
2479	} else if (hwp_active) {
2480		/*
2481		 * Re-enable HWP in case this happens after a resume from ACPI
2482		 * S3 if the CPU was offline during the whole system/resume
2483		 * cycle.
2484		 */
2485		intel_pstate_hwp_reenable(cpu);
2486	}
2487
2488	cpu->epp_powersave = -EINVAL;
2489	cpu->epp_policy = 0;
2490
2491	intel_pstate_get_cpu_pstates(cpu);
2492
2493	pr_debug("controlling: cpu %d\n", cpunum);
2494
2495	return 0;
2496}
2497
2498static void intel_pstate_set_update_util_hook(unsigned int cpu_num)
2499{
2500	struct cpudata *cpu = all_cpu_data[cpu_num];
2501
2502	if (hwp_active && !hwp_boost)
2503		return;
2504
2505	if (cpu->update_util_set)
2506		return;
2507
2508	/* Prevent intel_pstate_update_util() from using stale data. */
2509	cpu->sample.time = 0;
2510	cpufreq_add_update_util_hook(cpu_num, &cpu->update_util,
2511				     (hwp_active ?
2512				      intel_pstate_update_util_hwp :
2513				      intel_pstate_update_util));
2514	cpu->update_util_set = true;
2515}
2516
2517static void intel_pstate_clear_update_util_hook(unsigned int cpu)
2518{
2519	struct cpudata *cpu_data = all_cpu_data[cpu];
2520
2521	if (!cpu_data->update_util_set)
2522		return;
2523
2524	cpufreq_remove_update_util_hook(cpu);
2525	cpu_data->update_util_set = false;
2526	synchronize_rcu();
2527}
2528
2529static int intel_pstate_get_max_freq(struct cpudata *cpu)
2530{
2531	return global.turbo_disabled || global.no_turbo ?
2532			cpu->pstate.max_freq : cpu->pstate.turbo_freq;
2533}
2534
2535static void intel_pstate_update_perf_limits(struct cpudata *cpu,
2536					    unsigned int policy_min,
2537					    unsigned int policy_max)
2538{
2539	int perf_ctl_scaling = cpu->pstate.perf_ctl_scaling;
2540	int32_t max_policy_perf, min_policy_perf;
2541
2542	max_policy_perf = policy_max / perf_ctl_scaling;
2543	if (policy_max == policy_min) {
2544		min_policy_perf = max_policy_perf;
2545	} else {
2546		min_policy_perf = policy_min / perf_ctl_scaling;
2547		min_policy_perf = clamp_t(int32_t, min_policy_perf,
2548					  0, max_policy_perf);
2549	}
2550
2551	/*
2552	 * HWP needs some special consideration, because HWP_REQUEST uses
2553	 * abstract values to represent performance rather than pure ratios.
2554	 */
2555	if (hwp_active && cpu->pstate.scaling != perf_ctl_scaling) {
2556		int freq;
2557
2558		freq = max_policy_perf * perf_ctl_scaling;
2559		max_policy_perf = intel_pstate_freq_to_hwp(cpu, freq);
2560		freq = min_policy_perf * perf_ctl_scaling;
2561		min_policy_perf = intel_pstate_freq_to_hwp(cpu, freq);
2562	}
2563
2564	pr_debug("cpu:%d min_policy_perf:%d max_policy_perf:%d\n",
2565		 cpu->cpu, min_policy_perf, max_policy_perf);
2566
2567	/* Normalize user input to [min_perf, max_perf] */
2568	if (per_cpu_limits) {
2569		cpu->min_perf_ratio = min_policy_perf;
2570		cpu->max_perf_ratio = max_policy_perf;
2571	} else {
2572		int turbo_max = cpu->pstate.turbo_pstate;
2573		int32_t global_min, global_max;
2574
2575		/* Global limits are in percent of the maximum turbo P-state. */
2576		global_max = DIV_ROUND_UP(turbo_max * global.max_perf_pct, 100);
2577		global_min = DIV_ROUND_UP(turbo_max * global.min_perf_pct, 100);
2578		global_min = clamp_t(int32_t, global_min, 0, global_max);
2579
2580		pr_debug("cpu:%d global_min:%d global_max:%d\n", cpu->cpu,
2581			 global_min, global_max);
2582
2583		cpu->min_perf_ratio = max(min_policy_perf, global_min);
2584		cpu->min_perf_ratio = min(cpu->min_perf_ratio, max_policy_perf);
2585		cpu->max_perf_ratio = min(max_policy_perf, global_max);
2586		cpu->max_perf_ratio = max(min_policy_perf, cpu->max_perf_ratio);
2587
2588		/* Make sure min_perf <= max_perf */
2589		cpu->min_perf_ratio = min(cpu->min_perf_ratio,
2590					  cpu->max_perf_ratio);
2591
2592	}
2593	pr_debug("cpu:%d max_perf_ratio:%d min_perf_ratio:%d\n", cpu->cpu,
2594		 cpu->max_perf_ratio,
2595		 cpu->min_perf_ratio);
2596}
2597
2598static int intel_pstate_set_policy(struct cpufreq_policy *policy)
2599{
2600	struct cpudata *cpu;
2601
2602	if (!policy->cpuinfo.max_freq)
2603		return -ENODEV;
2604
2605	pr_debug("set_policy cpuinfo.max %u policy->max %u\n",
2606		 policy->cpuinfo.max_freq, policy->max);
2607
2608	cpu = all_cpu_data[policy->cpu];
2609	cpu->policy = policy->policy;
2610
2611	mutex_lock(&intel_pstate_limits_lock);
2612
2613	intel_pstate_update_perf_limits(cpu, policy->min, policy->max);
2614
2615	if (cpu->policy == CPUFREQ_POLICY_PERFORMANCE) {
2616		/*
2617		 * NOHZ_FULL CPUs need this as the governor callback may not
2618		 * be invoked on them.
2619		 */
2620		intel_pstate_clear_update_util_hook(policy->cpu);
2621		intel_pstate_max_within_limits(cpu);
2622	} else {
2623		intel_pstate_set_update_util_hook(policy->cpu);
2624	}
2625
2626	if (hwp_active) {
2627		/*
2628		 * When hwp_boost was active before and dynamically it
2629		 * was turned off, in that case we need to clear the
2630		 * update util hook.
2631		 */
2632		if (!hwp_boost)
2633			intel_pstate_clear_update_util_hook(policy->cpu);
2634		intel_pstate_hwp_set(policy->cpu);
2635	}
2636	/*
2637	 * policy->cur is never updated with the intel_pstate driver, but it
2638	 * is used as a stale frequency value. So, keep it within limits.
2639	 */
2640	policy->cur = policy->min;
2641
2642	mutex_unlock(&intel_pstate_limits_lock);
2643
2644	return 0;
2645}
2646
2647static void intel_pstate_adjust_policy_max(struct cpudata *cpu,
2648					   struct cpufreq_policy_data *policy)
2649{
2650	if (!hwp_active &&
2651	    cpu->pstate.max_pstate_physical > cpu->pstate.max_pstate &&
2652	    policy->max < policy->cpuinfo.max_freq &&
2653	    policy->max > cpu->pstate.max_freq) {
2654		pr_debug("policy->max > max non turbo frequency\n");
2655		policy->max = policy->cpuinfo.max_freq;
2656	}
2657}
2658
2659static void intel_pstate_verify_cpu_policy(struct cpudata *cpu,
2660					   struct cpufreq_policy_data *policy)
2661{
2662	int max_freq;
2663
2664	update_turbo_state();
2665	if (hwp_active) {
2666		intel_pstate_get_hwp_cap(cpu);
2667		max_freq = global.no_turbo || global.turbo_disabled ?
2668				cpu->pstate.max_freq : cpu->pstate.turbo_freq;
2669	} else {
2670		max_freq = intel_pstate_get_max_freq(cpu);
2671	}
2672	cpufreq_verify_within_limits(policy, policy->cpuinfo.min_freq, max_freq);
2673
2674	intel_pstate_adjust_policy_max(cpu, policy);
2675}
2676
2677static int intel_pstate_verify_policy(struct cpufreq_policy_data *policy)
2678{
2679	intel_pstate_verify_cpu_policy(all_cpu_data[policy->cpu], policy);
2680
2681	return 0;
2682}
2683
2684static int intel_cpufreq_cpu_offline(struct cpufreq_policy *policy)
2685{
2686	struct cpudata *cpu = all_cpu_data[policy->cpu];
2687
2688	pr_debug("CPU %d going offline\n", cpu->cpu);
2689
2690	if (cpu->suspended)
2691		return 0;
2692
2693	/*
2694	 * If the CPU is an SMT thread and it goes offline with the performance
2695	 * settings different from the minimum, it will prevent its sibling
2696	 * from getting to lower performance levels, so force the minimum
2697	 * performance on CPU offline to prevent that from happening.
2698	 */
2699	if (hwp_active)
2700		intel_pstate_hwp_offline(cpu);
2701	else
2702		intel_pstate_set_min_pstate(cpu);
2703
2704	intel_pstate_exit_perf_limits(policy);
2705
2706	return 0;
2707}
2708
2709static int intel_pstate_cpu_online(struct cpufreq_policy *policy)
2710{
2711	struct cpudata *cpu = all_cpu_data[policy->cpu];
2712
2713	pr_debug("CPU %d going online\n", cpu->cpu);
2714
2715	intel_pstate_init_acpi_perf_limits(policy);
2716
2717	if (hwp_active) {
2718		/*
2719		 * Re-enable HWP and clear the "suspended" flag to let "resume"
2720		 * know that it need not do that.
2721		 */
2722		intel_pstate_hwp_reenable(cpu);
2723		cpu->suspended = false;
2724	}
2725
2726	return 0;
2727}
2728
2729static int intel_pstate_cpu_offline(struct cpufreq_policy *policy)
2730{
2731	intel_pstate_clear_update_util_hook(policy->cpu);
2732
2733	return intel_cpufreq_cpu_offline(policy);
2734}
2735
2736static int intel_pstate_cpu_exit(struct cpufreq_policy *policy)
2737{
2738	pr_debug("CPU %d exiting\n", policy->cpu);
2739
2740	policy->fast_switch_possible = false;
2741
2742	return 0;
2743}
2744
2745static int __intel_pstate_cpu_init(struct cpufreq_policy *policy)
2746{
2747	struct cpudata *cpu;
2748	int rc;
2749
2750	rc = intel_pstate_init_cpu(policy->cpu);
2751	if (rc)
2752		return rc;
2753
2754	cpu = all_cpu_data[policy->cpu];
2755
2756	cpu->max_perf_ratio = 0xFF;
2757	cpu->min_perf_ratio = 0;
2758
2759	/* cpuinfo and default policy values */
2760	policy->cpuinfo.min_freq = cpu->pstate.min_freq;
2761	update_turbo_state();
2762	global.turbo_disabled_mf = global.turbo_disabled;
2763	policy->cpuinfo.max_freq = global.turbo_disabled ?
2764			cpu->pstate.max_freq : cpu->pstate.turbo_freq;
2765
2766	policy->min = policy->cpuinfo.min_freq;
2767	policy->max = policy->cpuinfo.max_freq;
2768
2769	intel_pstate_init_acpi_perf_limits(policy);
2770
2771	policy->fast_switch_possible = true;
2772
2773	return 0;
2774}
2775
2776static int intel_pstate_cpu_init(struct cpufreq_policy *policy)
2777{
2778	int ret = __intel_pstate_cpu_init(policy);
2779
2780	if (ret)
2781		return ret;
2782
2783	/*
2784	 * Set the policy to powersave to provide a valid fallback value in case
2785	 * the default cpufreq governor is neither powersave nor performance.
2786	 */
2787	policy->policy = CPUFREQ_POLICY_POWERSAVE;
2788
2789	if (hwp_active) {
2790		struct cpudata *cpu = all_cpu_data[policy->cpu];
2791
2792		cpu->epp_cached = intel_pstate_get_epp(cpu, 0);
2793	}
2794
2795	return 0;
2796}
2797
2798static struct cpufreq_driver intel_pstate = {
2799	.flags		= CPUFREQ_CONST_LOOPS,
2800	.verify		= intel_pstate_verify_policy,
2801	.setpolicy	= intel_pstate_set_policy,
2802	.suspend	= intel_pstate_suspend,
2803	.resume		= intel_pstate_resume,
2804	.init		= intel_pstate_cpu_init,
2805	.exit		= intel_pstate_cpu_exit,
2806	.offline	= intel_pstate_cpu_offline,
2807	.online		= intel_pstate_cpu_online,
2808	.update_limits	= intel_pstate_update_limits,
2809	.name		= "intel_pstate",
2810};
2811
2812static int intel_cpufreq_verify_policy(struct cpufreq_policy_data *policy)
2813{
2814	struct cpudata *cpu = all_cpu_data[policy->cpu];
2815
2816	intel_pstate_verify_cpu_policy(cpu, policy);
2817	intel_pstate_update_perf_limits(cpu, policy->min, policy->max);
2818
2819	return 0;
2820}
2821
2822/* Use of trace in passive mode:
2823 *
2824 * In passive mode the trace core_busy field (also known as the
2825 * performance field, and lablelled as such on the graphs; also known as
2826 * core_avg_perf) is not needed and so is re-assigned to indicate if the
2827 * driver call was via the normal or fast switch path. Various graphs
2828 * output from the intel_pstate_tracer.py utility that include core_busy
2829 * (or performance or core_avg_perf) have a fixed y-axis from 0 to 100%,
2830 * so we use 10 to indicate the normal path through the driver, and
2831 * 90 to indicate the fast switch path through the driver.
2832 * The scaled_busy field is not used, and is set to 0.
2833 */
2834
2835#define	INTEL_PSTATE_TRACE_TARGET 10
2836#define	INTEL_PSTATE_TRACE_FAST_SWITCH 90
2837
2838static void intel_cpufreq_trace(struct cpudata *cpu, unsigned int trace_type, int old_pstate)
2839{
2840	struct sample *sample;
2841
2842	if (!trace_pstate_sample_enabled())
2843		return;
2844
2845	if (!intel_pstate_sample(cpu, ktime_get()))
2846		return;
2847
2848	sample = &cpu->sample;
2849	trace_pstate_sample(trace_type,
2850		0,
2851		old_pstate,
2852		cpu->pstate.current_pstate,
2853		sample->mperf,
2854		sample->aperf,
2855		sample->tsc,
2856		get_avg_frequency(cpu),
2857		fp_toint(cpu->iowait_boost * 100));
2858}
2859
2860static void intel_cpufreq_hwp_update(struct cpudata *cpu, u32 min, u32 max,
2861				     u32 desired, bool fast_switch)
2862{
2863	u64 prev = READ_ONCE(cpu->hwp_req_cached), value = prev;
2864
2865	value &= ~HWP_MIN_PERF(~0L);
2866	value |= HWP_MIN_PERF(min);
2867
2868	value &= ~HWP_MAX_PERF(~0L);
2869	value |= HWP_MAX_PERF(max);
2870
2871	value &= ~HWP_DESIRED_PERF(~0L);
2872	value |= HWP_DESIRED_PERF(desired);
2873
2874	if (value == prev)
2875		return;
2876
2877	WRITE_ONCE(cpu->hwp_req_cached, value);
2878	if (fast_switch)
2879		wrmsrl(MSR_HWP_REQUEST, value);
2880	else
2881		wrmsrl_on_cpu(cpu->cpu, MSR_HWP_REQUEST, value);
2882}
2883
2884static void intel_cpufreq_perf_ctl_update(struct cpudata *cpu,
2885					  u32 target_pstate, bool fast_switch)
2886{
2887	if (fast_switch)
2888		wrmsrl(MSR_IA32_PERF_CTL,
2889		       pstate_funcs.get_val(cpu, target_pstate));
2890	else
2891		wrmsrl_on_cpu(cpu->cpu, MSR_IA32_PERF_CTL,
2892			      pstate_funcs.get_val(cpu, target_pstate));
2893}
2894
2895static int intel_cpufreq_update_pstate(struct cpufreq_policy *policy,
2896				       int target_pstate, bool fast_switch)
2897{
2898	struct cpudata *cpu = all_cpu_data[policy->cpu];
2899	int old_pstate = cpu->pstate.current_pstate;
2900
2901	target_pstate = intel_pstate_prepare_request(cpu, target_pstate);
2902	if (hwp_active) {
2903		int max_pstate = policy->strict_target ?
2904					target_pstate : cpu->max_perf_ratio;
2905
2906		intel_cpufreq_hwp_update(cpu, target_pstate, max_pstate, 0,
2907					 fast_switch);
2908	} else if (target_pstate != old_pstate) {
2909		intel_cpufreq_perf_ctl_update(cpu, target_pstate, fast_switch);
2910	}
2911
2912	cpu->pstate.current_pstate = target_pstate;
2913
2914	intel_cpufreq_trace(cpu, fast_switch ? INTEL_PSTATE_TRACE_FAST_SWITCH :
2915			    INTEL_PSTATE_TRACE_TARGET, old_pstate);
2916
2917	return target_pstate;
2918}
2919
2920static int intel_cpufreq_target(struct cpufreq_policy *policy,
2921				unsigned int target_freq,
2922				unsigned int relation)
2923{
2924	struct cpudata *cpu = all_cpu_data[policy->cpu];
2925	struct cpufreq_freqs freqs;
2926	int target_pstate;
2927
2928	update_turbo_state();
2929
2930	freqs.old = policy->cur;
2931	freqs.new = target_freq;
2932
2933	cpufreq_freq_transition_begin(policy, &freqs);
2934
2935	target_pstate = intel_pstate_freq_to_hwp_rel(cpu, freqs.new, relation);
2936	target_pstate = intel_cpufreq_update_pstate(policy, target_pstate, false);
2937
2938	freqs.new = target_pstate * cpu->pstate.scaling;
2939
2940	cpufreq_freq_transition_end(policy, &freqs, false);
2941
2942	return 0;
2943}
2944
2945static unsigned int intel_cpufreq_fast_switch(struct cpufreq_policy *policy,
2946					      unsigned int target_freq)
2947{
2948	struct cpudata *cpu = all_cpu_data[policy->cpu];
2949	int target_pstate;
2950
2951	update_turbo_state();
2952
2953	target_pstate = intel_pstate_freq_to_hwp(cpu, target_freq);
2954
2955	target_pstate = intel_cpufreq_update_pstate(policy, target_pstate, true);
2956
2957	return target_pstate * cpu->pstate.scaling;
2958}
2959
2960static void intel_cpufreq_adjust_perf(unsigned int cpunum,
2961				      unsigned long min_perf,
2962				      unsigned long target_perf,
2963				      unsigned long capacity)
2964{
2965	struct cpudata *cpu = all_cpu_data[cpunum];
2966	u64 hwp_cap = READ_ONCE(cpu->hwp_cap_cached);
2967	int old_pstate = cpu->pstate.current_pstate;
2968	int cap_pstate, min_pstate, max_pstate, target_pstate;
2969
2970	update_turbo_state();
2971	cap_pstate = global.turbo_disabled ? HWP_GUARANTEED_PERF(hwp_cap) :
2972					     HWP_HIGHEST_PERF(hwp_cap);
2973
2974	/* Optimization: Avoid unnecessary divisions. */
2975
2976	target_pstate = cap_pstate;
2977	if (target_perf < capacity)
2978		target_pstate = DIV_ROUND_UP(cap_pstate * target_perf, capacity);
2979
2980	min_pstate = cap_pstate;
2981	if (min_perf < capacity)
2982		min_pstate = DIV_ROUND_UP(cap_pstate * min_perf, capacity);
2983
2984	if (min_pstate < cpu->pstate.min_pstate)
2985		min_pstate = cpu->pstate.min_pstate;
2986
2987	if (min_pstate < cpu->min_perf_ratio)
2988		min_pstate = cpu->min_perf_ratio;
2989
2990	if (min_pstate > cpu->max_perf_ratio)
2991		min_pstate = cpu->max_perf_ratio;
2992
2993	max_pstate = min(cap_pstate, cpu->max_perf_ratio);
2994	if (max_pstate < min_pstate)
2995		max_pstate = min_pstate;
2996
2997	target_pstate = clamp_t(int, target_pstate, min_pstate, max_pstate);
2998
2999	intel_cpufreq_hwp_update(cpu, min_pstate, max_pstate, target_pstate, true);
3000
3001	cpu->pstate.current_pstate = target_pstate;
3002	intel_cpufreq_trace(cpu, INTEL_PSTATE_TRACE_FAST_SWITCH, old_pstate);
3003}
3004
3005static int intel_cpufreq_cpu_init(struct cpufreq_policy *policy)
3006{
3007	struct freq_qos_request *req;
3008	struct cpudata *cpu;
3009	struct device *dev;
3010	int ret, freq;
3011
3012	dev = get_cpu_device(policy->cpu);
3013	if (!dev)
3014		return -ENODEV;
3015
3016	ret = __intel_pstate_cpu_init(policy);
3017	if (ret)
3018		return ret;
3019
3020	policy->cpuinfo.transition_latency = INTEL_CPUFREQ_TRANSITION_LATENCY;
3021	/* This reflects the intel_pstate_get_cpu_pstates() setting. */
3022	policy->cur = policy->cpuinfo.min_freq;
3023
3024	req = kcalloc(2, sizeof(*req), GFP_KERNEL);
3025	if (!req) {
3026		ret = -ENOMEM;
3027		goto pstate_exit;
3028	}
3029
3030	cpu = all_cpu_data[policy->cpu];
3031
3032	if (hwp_active) {
3033		u64 value;
3034
3035		policy->transition_delay_us = INTEL_CPUFREQ_TRANSITION_DELAY_HWP;
3036
3037		intel_pstate_get_hwp_cap(cpu);
3038
3039		rdmsrl_on_cpu(cpu->cpu, MSR_HWP_REQUEST, &value);
3040		WRITE_ONCE(cpu->hwp_req_cached, value);
3041
3042		cpu->epp_cached = intel_pstate_get_epp(cpu, value);
3043	} else {
3044		policy->transition_delay_us = INTEL_CPUFREQ_TRANSITION_DELAY;
3045	}
3046
3047	freq = DIV_ROUND_UP(cpu->pstate.turbo_freq * global.min_perf_pct, 100);
3048
3049	ret = freq_qos_add_request(&policy->constraints, req, FREQ_QOS_MIN,
3050				   freq);
3051	if (ret < 0) {
3052		dev_err(dev, "Failed to add min-freq constraint (%d)\n", ret);
3053		goto free_req;
3054	}
3055
3056	freq = DIV_ROUND_UP(cpu->pstate.turbo_freq * global.max_perf_pct, 100);
3057
3058	ret = freq_qos_add_request(&policy->constraints, req + 1, FREQ_QOS_MAX,
3059				   freq);
3060	if (ret < 0) {
3061		dev_err(dev, "Failed to add max-freq constraint (%d)\n", ret);
3062		goto remove_min_req;
3063	}
3064
3065	policy->driver_data = req;
3066
3067	return 0;
3068
3069remove_min_req:
3070	freq_qos_remove_request(req);
3071free_req:
3072	kfree(req);
3073pstate_exit:
3074	intel_pstate_exit_perf_limits(policy);
3075
3076	return ret;
3077}
3078
3079static int intel_cpufreq_cpu_exit(struct cpufreq_policy *policy)
3080{
3081	struct freq_qos_request *req;
3082
3083	req = policy->driver_data;
3084
3085	freq_qos_remove_request(req + 1);
3086	freq_qos_remove_request(req);
3087	kfree(req);
3088
3089	return intel_pstate_cpu_exit(policy);
3090}
3091
3092static int intel_cpufreq_suspend(struct cpufreq_policy *policy)
3093{
3094	intel_pstate_suspend(policy);
3095
3096	if (hwp_active) {
3097		struct cpudata *cpu = all_cpu_data[policy->cpu];
3098		u64 value = READ_ONCE(cpu->hwp_req_cached);
3099
3100		/*
3101		 * Clear the desired perf field in MSR_HWP_REQUEST in case
3102		 * intel_cpufreq_adjust_perf() is in use and the last value
3103		 * written by it may not be suitable.
3104		 */
3105		value &= ~HWP_DESIRED_PERF(~0L);
3106		wrmsrl_on_cpu(cpu->cpu, MSR_HWP_REQUEST, value);
3107		WRITE_ONCE(cpu->hwp_req_cached, value);
3108	}
3109
3110	return 0;
3111}
3112
3113static struct cpufreq_driver intel_cpufreq = {
3114	.flags		= CPUFREQ_CONST_LOOPS,
3115	.verify		= intel_cpufreq_verify_policy,
3116	.target		= intel_cpufreq_target,
3117	.fast_switch	= intel_cpufreq_fast_switch,
3118	.init		= intel_cpufreq_cpu_init,
3119	.exit		= intel_cpufreq_cpu_exit,
3120	.offline	= intel_cpufreq_cpu_offline,
3121	.online		= intel_pstate_cpu_online,
3122	.suspend	= intel_cpufreq_suspend,
3123	.resume		= intel_pstate_resume,
3124	.update_limits	= intel_pstate_update_limits,
3125	.name		= "intel_cpufreq",
3126};
3127
3128static struct cpufreq_driver *default_driver;
3129
3130static void intel_pstate_driver_cleanup(void)
3131{
3132	unsigned int cpu;
3133
3134	cpus_read_lock();
3135	for_each_online_cpu(cpu) {
3136		if (all_cpu_data[cpu]) {
3137			if (intel_pstate_driver == &intel_pstate)
3138				intel_pstate_clear_update_util_hook(cpu);
3139
3140			spin_lock(&hwp_notify_lock);
3141			kfree(all_cpu_data[cpu]);
3142			WRITE_ONCE(all_cpu_data[cpu], NULL);
3143			spin_unlock(&hwp_notify_lock);
3144		}
3145	}
3146	cpus_read_unlock();
3147
3148	intel_pstate_driver = NULL;
3149}
3150
3151static int intel_pstate_register_driver(struct cpufreq_driver *driver)
3152{
3153	int ret;
3154
3155	if (driver == &intel_pstate)
3156		intel_pstate_sysfs_expose_hwp_dynamic_boost();
3157
3158	memset(&global, 0, sizeof(global));
3159	global.max_perf_pct = 100;
3160
3161	intel_pstate_driver = driver;
3162	ret = cpufreq_register_driver(intel_pstate_driver);
3163	if (ret) {
3164		intel_pstate_driver_cleanup();
3165		return ret;
3166	}
3167
3168	global.min_perf_pct = min_perf_pct_min();
3169
3170	return 0;
3171}
3172
3173static ssize_t intel_pstate_show_status(char *buf)
3174{
3175	if (!intel_pstate_driver)
3176		return sprintf(buf, "off\n");
3177
3178	return sprintf(buf, "%s\n", intel_pstate_driver == &intel_pstate ?
3179					"active" : "passive");
3180}
3181
3182static int intel_pstate_update_status(const char *buf, size_t size)
3183{
3184	if (size == 3 && !strncmp(buf, "off", size)) {
3185		if (!intel_pstate_driver)
3186			return -EINVAL;
3187
3188		if (hwp_active)
3189			return -EBUSY;
3190
3191		cpufreq_unregister_driver(intel_pstate_driver);
3192		intel_pstate_driver_cleanup();
3193		return 0;
3194	}
3195
3196	if (size == 6 && !strncmp(buf, "active", size)) {
3197		if (intel_pstate_driver) {
3198			if (intel_pstate_driver == &intel_pstate)
3199				return 0;
3200
3201			cpufreq_unregister_driver(intel_pstate_driver);
3202		}
3203
3204		return intel_pstate_register_driver(&intel_pstate);
3205	}
3206
3207	if (size == 7 && !strncmp(buf, "passive", size)) {
3208		if (intel_pstate_driver) {
3209			if (intel_pstate_driver == &intel_cpufreq)
3210				return 0;
3211
3212			cpufreq_unregister_driver(intel_pstate_driver);
3213			intel_pstate_sysfs_hide_hwp_dynamic_boost();
3214		}
3215
3216		return intel_pstate_register_driver(&intel_cpufreq);
3217	}
3218
3219	return -EINVAL;
3220}
3221
3222static int no_load __initdata;
3223static int no_hwp __initdata;
3224static int hwp_only __initdata;
3225static unsigned int force_load __initdata;
3226
3227static int __init intel_pstate_msrs_not_valid(void)
3228{
3229	if (!pstate_funcs.get_max(0) ||
3230	    !pstate_funcs.get_min(0) ||
3231	    !pstate_funcs.get_turbo(0))
3232		return -ENODEV;
3233
3234	return 0;
3235}
3236
3237static void __init copy_cpu_funcs(struct pstate_funcs *funcs)
3238{
3239	pstate_funcs.get_max   = funcs->get_max;
3240	pstate_funcs.get_max_physical = funcs->get_max_physical;
3241	pstate_funcs.get_min   = funcs->get_min;
3242	pstate_funcs.get_turbo = funcs->get_turbo;
3243	pstate_funcs.get_scaling = funcs->get_scaling;
3244	pstate_funcs.get_val   = funcs->get_val;
3245	pstate_funcs.get_vid   = funcs->get_vid;
3246	pstate_funcs.get_aperf_mperf_shift = funcs->get_aperf_mperf_shift;
3247}
3248
3249#ifdef CONFIG_ACPI
3250
3251static bool __init intel_pstate_no_acpi_pss(void)
3252{
3253	int i;
3254
3255	for_each_possible_cpu(i) {
3256		acpi_status status;
3257		union acpi_object *pss;
3258		struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
3259		struct acpi_processor *pr = per_cpu(processors, i);
3260
3261		if (!pr)
3262			continue;
3263
3264		status = acpi_evaluate_object(pr->handle, "_PSS", NULL, &buffer);
3265		if (ACPI_FAILURE(status))
3266			continue;
3267
3268		pss = buffer.pointer;
3269		if (pss && pss->type == ACPI_TYPE_PACKAGE) {
3270			kfree(pss);
3271			return false;
3272		}
3273
3274		kfree(pss);
3275	}
3276
3277	pr_debug("ACPI _PSS not found\n");
3278	return true;
3279}
3280
3281static bool __init intel_pstate_no_acpi_pcch(void)
3282{
3283	acpi_status status;
3284	acpi_handle handle;
3285
3286	status = acpi_get_handle(NULL, "\\_SB", &handle);
3287	if (ACPI_FAILURE(status))
3288		goto not_found;
3289
3290	if (acpi_has_method(handle, "PCCH"))
3291		return false;
3292
3293not_found:
3294	pr_debug("ACPI PCCH not found\n");
3295	return true;
3296}
3297
3298static bool __init intel_pstate_has_acpi_ppc(void)
3299{
3300	int i;
3301
3302	for_each_possible_cpu(i) {
3303		struct acpi_processor *pr = per_cpu(processors, i);
3304
3305		if (!pr)
3306			continue;
3307		if (acpi_has_method(pr->handle, "_PPC"))
3308			return true;
3309	}
3310	pr_debug("ACPI _PPC not found\n");
3311	return false;
3312}
3313
3314enum {
3315	PSS,
3316	PPC,
3317};
3318
3319/* Hardware vendor-specific info that has its own power management modes */
3320static struct acpi_platform_list plat_info[] __initdata = {
3321	{"HP    ", "ProLiant", 0, ACPI_SIG_FADT, all_versions, NULL, PSS},
3322	{"ORACLE", "X4-2    ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3323	{"ORACLE", "X4-2L   ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3324	{"ORACLE", "X4-2B   ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3325	{"ORACLE", "X3-2    ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3326	{"ORACLE", "X3-2L   ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3327	{"ORACLE", "X3-2B   ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3328	{"ORACLE", "X4470M2 ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3329	{"ORACLE", "X4270M3 ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3330	{"ORACLE", "X4270M2 ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3331	{"ORACLE", "X4170M2 ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3332	{"ORACLE", "X4170 M3", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3333	{"ORACLE", "X4275 M3", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3334	{"ORACLE", "X6-2    ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3335	{"ORACLE", "Sudbury ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
3336	{ } /* End */
3337};
3338
3339#define BITMASK_OOB	(BIT(8) | BIT(18))
3340
3341static bool __init intel_pstate_platform_pwr_mgmt_exists(void)
3342{
3343	const struct x86_cpu_id *id;
3344	u64 misc_pwr;
3345	int idx;
3346
3347	id = x86_match_cpu(intel_pstate_cpu_oob_ids);
3348	if (id) {
3349		rdmsrl(MSR_MISC_PWR_MGMT, misc_pwr);
3350		if (misc_pwr & BITMASK_OOB) {
3351			pr_debug("Bit 8 or 18 in the MISC_PWR_MGMT MSR set\n");
3352			pr_debug("P states are controlled in Out of Band mode by the firmware/hardware\n");
3353			return true;
3354		}
3355	}
3356
3357	idx = acpi_match_platform_list(plat_info);
3358	if (idx < 0)
3359		return false;
3360
3361	switch (plat_info[idx].data) {
3362	case PSS:
3363		if (!intel_pstate_no_acpi_pss())
3364			return false;
3365
3366		return intel_pstate_no_acpi_pcch();
3367	case PPC:
3368		return intel_pstate_has_acpi_ppc() && !force_load;
3369	}
3370
3371	return false;
3372}
3373
3374static void intel_pstate_request_control_from_smm(void)
3375{
3376	/*
3377	 * It may be unsafe to request P-states control from SMM if _PPC support
3378	 * has not been enabled.
3379	 */
3380	if (acpi_ppc)
3381		acpi_processor_pstate_control();
3382}
3383#else /* CONFIG_ACPI not enabled */
3384static inline bool intel_pstate_platform_pwr_mgmt_exists(void) { return false; }
3385static inline bool intel_pstate_has_acpi_ppc(void) { return false; }
3386static inline void intel_pstate_request_control_from_smm(void) {}
3387#endif /* CONFIG_ACPI */
3388
3389#define INTEL_PSTATE_HWP_BROADWELL	0x01
3390
3391#define X86_MATCH_HWP(model, hwp_mode)					\
3392	X86_MATCH_VENDOR_FAM_MODEL_FEATURE(INTEL, 6, INTEL_FAM6_##model, \
3393					   X86_FEATURE_HWP, hwp_mode)
3394
3395static const struct x86_cpu_id hwp_support_ids[] __initconst = {
3396	X86_MATCH_HWP(BROADWELL_X,	INTEL_PSTATE_HWP_BROADWELL),
3397	X86_MATCH_HWP(BROADWELL_D,	INTEL_PSTATE_HWP_BROADWELL),
3398	X86_MATCH_HWP(ANY,		0),
3399	{}
3400};
3401
3402static bool intel_pstate_hwp_is_enabled(void)
3403{
3404	u64 value;
3405
3406	rdmsrl(MSR_PM_ENABLE, value);
3407	return !!(value & 0x1);
3408}
3409
3410static const struct x86_cpu_id intel_epp_balance_perf[] = {
3411	/*
3412	 * Set EPP value as 102, this is the max suggested EPP
3413	 * which can result in one core turbo frequency for
3414	 * AlderLake Mobile CPUs.
3415	 */
3416	X86_MATCH_INTEL_FAM6_MODEL(ALDERLAKE_L, 102),
3417	X86_MATCH_INTEL_FAM6_MODEL(SAPPHIRERAPIDS_X, 32),
3418	{}
3419};
3420
3421static const struct x86_cpu_id intel_hybrid_scaling_factor[] = {
3422	X86_MATCH_INTEL_FAM6_MODEL(METEORLAKE_L, HYBRID_SCALING_FACTOR_MTL),
3423	{}
3424};
3425
3426static int __init intel_pstate_init(void)
3427{
3428	static struct cpudata **_all_cpu_data;
3429	const struct x86_cpu_id *id;
3430	int rc;
3431
3432	if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL)
3433		return -ENODEV;
3434
3435	id = x86_match_cpu(hwp_support_ids);
3436	if (id) {
3437		hwp_forced = intel_pstate_hwp_is_enabled();
3438
3439		if (hwp_forced)
3440			pr_info("HWP enabled by BIOS\n");
3441		else if (no_load)
3442			return -ENODEV;
3443
3444		copy_cpu_funcs(&core_funcs);
3445		/*
3446		 * Avoid enabling HWP for processors without EPP support,
3447		 * because that means incomplete HWP implementation which is a
3448		 * corner case and supporting it is generally problematic.
3449		 *
3450		 * If HWP is enabled already, though, there is no choice but to
3451		 * deal with it.
3452		 */
3453		if ((!no_hwp && boot_cpu_has(X86_FEATURE_HWP_EPP)) || hwp_forced) {
3454			WRITE_ONCE(hwp_active, 1);
3455			hwp_mode_bdw = id->driver_data;
3456			intel_pstate.attr = hwp_cpufreq_attrs;
3457			intel_cpufreq.attr = hwp_cpufreq_attrs;
3458			intel_cpufreq.flags |= CPUFREQ_NEED_UPDATE_LIMITS;
3459			intel_cpufreq.adjust_perf = intel_cpufreq_adjust_perf;
3460			if (!default_driver)
3461				default_driver = &intel_pstate;
3462
3463			pstate_funcs.get_cpu_scaling = hwp_get_cpu_scaling;
3464
3465			goto hwp_cpu_matched;
3466		}
3467		pr_info("HWP not enabled\n");
3468	} else {
3469		if (no_load)
3470			return -ENODEV;
3471
3472		id = x86_match_cpu(intel_pstate_cpu_ids);
3473		if (!id) {
3474			pr_info("CPU model not supported\n");
3475			return -ENODEV;
3476		}
3477
3478		copy_cpu_funcs((struct pstate_funcs *)id->driver_data);
3479	}
3480
3481	if (intel_pstate_msrs_not_valid()) {
3482		pr_info("Invalid MSRs\n");
3483		return -ENODEV;
3484	}
3485	/* Without HWP start in the passive mode. */
3486	if (!default_driver)
3487		default_driver = &intel_cpufreq;
3488
3489hwp_cpu_matched:
3490	/*
3491	 * The Intel pstate driver will be ignored if the platform
3492	 * firmware has its own power management modes.
3493	 */
3494	if (intel_pstate_platform_pwr_mgmt_exists()) {
3495		pr_info("P-states controlled by the platform\n");
3496		return -ENODEV;
3497	}
3498
3499	if (!hwp_active && hwp_only)
3500		return -ENOTSUPP;
3501
3502	pr_info("Intel P-state driver initializing\n");
3503
3504	_all_cpu_data = vzalloc(array_size(sizeof(void *), num_possible_cpus()));
3505	if (!_all_cpu_data)
3506		return -ENOMEM;
3507
3508	WRITE_ONCE(all_cpu_data, _all_cpu_data);
3509
3510	intel_pstate_request_control_from_smm();
3511
3512	intel_pstate_sysfs_expose_params();
3513
3514	if (hwp_active) {
3515		const struct x86_cpu_id *id = x86_match_cpu(intel_epp_balance_perf);
3516		const struct x86_cpu_id *hybrid_id = x86_match_cpu(intel_hybrid_scaling_factor);
3517
3518		if (id)
3519			epp_values[EPP_INDEX_BALANCE_PERFORMANCE] = id->driver_data;
3520
3521		if (hybrid_id) {
3522			hybrid_scaling_factor = hybrid_id->driver_data;
3523			pr_debug("hybrid scaling factor: %d\n", hybrid_scaling_factor);
3524		}
3525
3526	}
3527
3528	mutex_lock(&intel_pstate_driver_lock);
3529	rc = intel_pstate_register_driver(default_driver);
3530	mutex_unlock(&intel_pstate_driver_lock);
3531	if (rc) {
3532		intel_pstate_sysfs_remove();
3533		return rc;
3534	}
3535
3536	if (hwp_active) {
3537		const struct x86_cpu_id *id;
3538
3539		id = x86_match_cpu(intel_pstate_cpu_ee_disable_ids);
3540		if (id) {
3541			set_power_ctl_ee_state(false);
3542			pr_info("Disabling energy efficiency optimization\n");
3543		}
3544
3545		pr_info("HWP enabled\n");
3546	} else if (boot_cpu_has(X86_FEATURE_HYBRID_CPU)) {
3547		pr_warn("Problematic setup: Hybrid processor with disabled HWP\n");
3548	}
3549
3550	return 0;
3551}
3552device_initcall(intel_pstate_init);
3553
3554static int __init intel_pstate_setup(char *str)
3555{
3556	if (!str)
3557		return -EINVAL;
3558
3559	if (!strcmp(str, "disable"))
3560		no_load = 1;
3561	else if (!strcmp(str, "active"))
3562		default_driver = &intel_pstate;
3563	else if (!strcmp(str, "passive"))
3564		default_driver = &intel_cpufreq;
3565
3566	if (!strcmp(str, "no_hwp"))
3567		no_hwp = 1;
3568
3569	if (!strcmp(str, "force"))
3570		force_load = 1;
3571	if (!strcmp(str, "hwp_only"))
3572		hwp_only = 1;
3573	if (!strcmp(str, "per_cpu_perf_limits"))
3574		per_cpu_limits = true;
3575
3576#ifdef CONFIG_ACPI
3577	if (!strcmp(str, "support_acpi_ppc"))
3578		acpi_ppc = true;
3579#endif
3580
3581	return 0;
3582}
3583early_param("intel_pstate", intel_pstate_setup);
3584
3585MODULE_AUTHOR("Dirk Brandewie <dirk.j.brandewie@intel.com>");
3586MODULE_DESCRIPTION("'intel_pstate' - P state driver Intel Core processors");