drivers/cpufreq/intel_pstate.c

   1 /*
   2  * intel_pstate.c: Native P state management for Intel processors
   3  *
   4  * (C) Copyright 2012 Intel Corporation
   5  * Author: Dirk Brandewie <dirk.j.brandewie@intel.com>
   6  *
   7  * This program is free software; you can redistribute it and/or
   8  * modify it under the terms of the GNU General Public License
   9  * as published by the Free Software Foundation; version 2
  10  * of the License.
  11  */
  12
  13 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  14
  15 #include <linux/kernel.h>
  16 #include <linux/kernel_stat.h>
  17 #include <linux/module.h>
  18 #include <linux/ktime.h>
  19 #include <linux/hrtimer.h>
  20 #include <linux/tick.h>
  21 #include <linux/slab.h>
  22 #include <linux/sched/cpufreq.h>
  23 #include <linux/list.h>
  24 #include <linux/cpu.h>
  25 #include <linux/cpufreq.h>
  26 #include <linux/sysfs.h>
  27 #include <linux/types.h>
  28 #include <linux/fs.h>
  29 #include <linux/debugfs.h>
  30 #include <linux/acpi.h>
  31 #include <linux/vmalloc.h>
  32 #include <trace/events/power.h>
  33
  34 #include <asm/div64.h>
  35 #include <asm/msr.h>
  36 #include <asm/cpu_device_id.h>
  37 #include <asm/cpufeature.h>
  38 #include <asm/intel-family.h>
  39
  40 #define INTEL_CPUFREQ_TRANSITION_LATENCY        20000
  41
  42 #ifdef CONFIG_ACPI
  43 #include <acpi/processor.h>
  44 #include <acpi/cppc_acpi.h>
  45 #endif
  46
  47 #define FRAC_BITS 8
  48 #define int_tofp(X) ((int64_t)(X) << FRAC_BITS)
  49 #define fp_toint(X) ((X) >> FRAC_BITS)
  50
  51 #define EXT_BITS 6
  52 #define EXT_FRAC_BITS (EXT_BITS + FRAC_BITS)
  53 #define fp_ext_toint(X) ((X) >> EXT_FRAC_BITS)
  54 #define int_ext_tofp(X) ((int64_t)(X) << EXT_FRAC_BITS)
  55
  56 static inline int32_t mul_fp(int32_t x, int32_t y)
  57 {
  58         return ((int64_t)x * (int64_t)y) >> FRAC_BITS;
  59 }
  60
  61 static inline int32_t div_fp(s64 x, s64 y)
  62 {
  63         return div64_s64((int64_t)x << FRAC_BITS, y);
  64 }
  65
  66 static inline int ceiling_fp(int32_t x)
  67 {
  68         int mask, ret;
  69
  70         ret = fp_toint(x);
  71         mask = (1 << FRAC_BITS) - 1;
  72         if (x & mask)
  73                 ret += 1;
  74         return ret;
  75 }
  76
  77 static inline u64 mul_ext_fp(u64 x, u64 y)
  78 {
  79         return (x * y) >> EXT_FRAC_BITS;
  80 }
  81
  82 static inline u64 div_ext_fp(u64 x, u64 y)
  83 {
  84         return div64_u64(x << EXT_FRAC_BITS, y);
  85 }
  86
  87 static inline int32_t percent_ext_fp(int percent)
  88 {
  89         return div_ext_fp(percent, 100);
  90 }
  91
  92 /**
  93  * struct sample -      Store performance sample
  94  * @core_avg_perf:      Ratio of APERF/MPERF which is the actual average
  95  *                      performance during last sample period
  96  * @busy_scaled:        Scaled busy value which is used to calculate next
  97  *                      P state. This can be different than core_avg_perf
  98  *                      to account for cpu idle period
  99  * @aperf:              Difference of actual performance frequency clock count
 100  *                      read from APERF MSR between last and current sample
 101  * @mperf:              Difference of maximum performance frequency clock count
 102  *                      read from MPERF MSR between last and current sample
 103  * @tsc:                Difference of time stamp counter between last and
 104  *                      current sample
 105  * @time:               Current time from scheduler
 106  *
 107  * This structure is used in the cpudata structure to store performance sample
 108  * data for choosing next P State.
 109  */
 110 struct sample {
 111         int32_t core_avg_perf;
 112         int32_t busy_scaled;
 113         u64 aperf;
 114         u64 mperf;
 115         u64 tsc;
 116         u64 time;
 117 };
 118
 119 /**
 120  * struct pstate_data - Store P state data
 121  * @current_pstate:     Current requested P state
 122  * @min_pstate:         Min P state possible for this platform
 123  * @max_pstate:         Max P state possible for this platform
 124  * @max_pstate_physical:This is physical Max P state for a processor
 125  *                      This can be higher than the max_pstate which can
 126  *                      be limited by platform thermal design power limits
 127  * @scaling:            Scaling factor to  convert frequency to cpufreq
 128  *                      frequency units
 129  * @turbo_pstate:       Max Turbo P state possible for this platform
 130  * @max_freq:           @max_pstate frequency in cpufreq units
 131  * @turbo_freq:         @turbo_pstate frequency in cpufreq units
 132  *
 133  * Stores the per cpu model P state limits and current P state.
 134  */
 135 struct pstate_data {
 136         int     current_pstate;
 137         int     min_pstate;
 138         int     max_pstate;
 139         int     max_pstate_physical;
 140         int     scaling;
 141         int     turbo_pstate;
 142         unsigned int max_freq;
 143         unsigned int turbo_freq;
 144 };
 145
 146 /**
 147  * struct vid_data -    Stores voltage information data
 148  * @min:                VID data for this platform corresponding to
 149  *                      the lowest P state
 150  * @max:                VID data corresponding to the highest P State.
 151  * @turbo:              VID data for turbo P state
 152  * @ratio:              Ratio of (vid max - vid min) /
 153  *                      (max P state - Min P State)
 154  *
 155  * Stores the voltage data for DVFS (Dynamic Voltage and Frequency Scaling)
 156  * This data is used in Atom platforms, where in addition to target P state,
 157  * the voltage data needs to be specified to select next P State.
 158  */
 159 struct vid_data {
 160         int min;
 161         int max;
 162         int turbo;
 163         int32_t ratio;
 164 };
 165
 166 /**
 167  * struct _pid -        Stores PID data
 168  * @setpoint:           Target set point for busyness or performance
 169  * @integral:           Storage for accumulated error values
 170  * @p_gain:             PID proportional gain
 171  * @i_gain:             PID integral gain
 172  * @d_gain:             PID derivative gain
 173  * @deadband:           PID deadband
 174  * @last_err:           Last error storage for integral part of PID calculation
 175  *
 176  * Stores PID coefficients and last error for PID controller.
 177  */
 178 struct _pid {
 179         int setpoint;
 180         int32_t integral;
 181         int32_t p_gain;
 182         int32_t i_gain;
 183         int32_t d_gain;
 184         int deadband;
 185         int32_t last_err;
 186 };
 187
 188 /**
 189  * struct perf_limits - Store user and policy limits
 190  * @max_perf:           This is a scaled value between 0 to 255 for max_perf_pct
 191  *                      This value is used to limit max pstate
 192  * @min_perf:           This is a scaled value between 0 to 255 for min_perf_pct
 193  *                      This value is used to limit min pstate
 194  *
 195  * Storage for policy defined limits.
 196  */
 197 struct perf_limits {
 198         int32_t max_perf;
 199         int32_t min_perf;
 200 };
 201
 202 /**
 203  * struct global_params - Global parameters, mostly tunable via sysfs.
 204  * @no_turbo:           Whether or not to use turbo P-states.
 205  * @turbo_disabled:     Whethet or not turbo P-states are available at all,
 206  *                      based on the MSR_IA32_MISC_ENABLE value and whether or
 207  *                      not the maximum reported turbo P-state is different from
 208  *                      the maximum reported non-turbo one.
 209  * @min_perf_pct:       Minimum capacity limit in percent of the maximum turbo
 210  *                      P-state capacity.
 211  * @max_perf_pct:       Maximum capacity limit in percent of the maximum turbo
 212  *                      P-state capacity.
 213  */
 214 struct global_params {
 215         bool no_turbo;
 216         bool turbo_disabled;
 217         int max_perf_pct;
 218         int min_perf_pct;
 219 };
 220
 221 /**
 222  * struct cpudata -     Per CPU instance data storage
 223  * @cpu:                CPU number for this instance data
 224  * @policy:             CPUFreq policy value
 225  * @update_util:        CPUFreq utility callback information
 226  * @update_util_set:    CPUFreq utility callback is set
 227  * @iowait_boost:       iowait-related boost fraction
 228  * @last_update:        Time of the last update.
 229  * @pstate:             Stores P state limits for this CPU
 230  * @vid:                Stores VID limits for this CPU
 231  * @pid:                Stores PID parameters for this CPU
 232  * @last_sample_time:   Last Sample time
 233  * @prev_aperf:         Last APERF value read from APERF MSR
 234  * @prev_mperf:         Last MPERF value read from MPERF MSR
 235  * @prev_tsc:           Last timestamp counter (TSC) value
 236  * @prev_cummulative_iowait: IO Wait time difference from last and
 237  *                      current sample
 238  * @sample:             Storage for storing last Sample data
 239  * @perf_limits:        Capacity limits unique to this CPU
 240  * @acpi_perf_data:     Stores ACPI perf information read from _PSS
 241  * @valid_pss_table:    Set to true for valid ACPI _PSS entries found
 242  * @epp_powersave:      Last saved HWP energy performance preference
 243  *                      (EPP) or energy performance bias (EPB),
 244  *                      when policy switched to performance
 245  * @epp_policy:         Last saved policy used to set EPP/EPB
 246  * @epp_default:        Power on default HWP energy performance
 247  *                      preference/bias
 248  * @epp_saved:          Saved EPP/EPB during system suspend or CPU offline
 249  *                      operation
 250  *
 251  * This structure stores per CPU instance data for all CPUs.
 252  */
 253 struct cpudata {
 254         int cpu;
 255
 256         unsigned int policy;
 257         struct update_util_data update_util;
 258         bool   update_util_set;
 259
 260         struct pstate_data pstate;
 261         struct vid_data vid;
 262         struct _pid pid;
 263
 264         u64     last_update;
 265         u64     last_sample_time;
 266         u64     prev_aperf;
 267         u64     prev_mperf;
 268         u64     prev_tsc;
 269         u64     prev_cummulative_iowait;
 270         struct sample sample;
 271         struct perf_limits perf_limits;
 272 #ifdef CONFIG_ACPI
 273         struct acpi_processor_performance acpi_perf_data;
 274         bool valid_pss_table;
 275 #endif
 276         unsigned int iowait_boost;
 277         s16 epp_powersave;
 278         s16 epp_policy;
 279         s16 epp_default;
 280         s16 epp_saved;
 281 };
 282
 283 static struct cpudata **all_cpu_data;
 284
 285 /**
 286  * struct pstate_adjust_policy - Stores static PID configuration data
 287  * @sample_rate_ms:     PID calculation sample rate in ms
 288  * @sample_rate_ns:     Sample rate calculation in ns
 289  * @deadband:           PID deadband
 290  * @setpoint:           PID Setpoint
 291  * @p_gain_pct:         PID proportional gain
 292  * @i_gain_pct:         PID integral gain
 293  * @d_gain_pct:         PID derivative gain
 294  *
 295  * Stores per CPU model static PID configuration data.
 296  */
 297 struct pstate_adjust_policy {
 298         int sample_rate_ms;
 299         s64 sample_rate_ns;
 300         int deadband;
 301         int setpoint;
 302         int p_gain_pct;
 303         int d_gain_pct;
 304         int i_gain_pct;
 305 };
 306
 307 /**
 308  * struct pstate_funcs - Per CPU model specific callbacks
 309  * @get_max:            Callback to get maximum non turbo effective P state
 310  * @get_max_physical:   Callback to get maximum non turbo physical P state
 311  * @get_min:            Callback to get minimum P state
 312  * @get_turbo:          Callback to get turbo P state
 313  * @get_scaling:        Callback to get frequency scaling factor
 314  * @get_val:            Callback to convert P state to actual MSR write value
 315  * @get_vid:            Callback to get VID data for Atom platforms
 316  * @get_target_pstate:  Callback to a function to calculate next P state to use
 317  *
 318  * Core and Atom CPU models have different way to get P State limits. This
 319  * structure is used to store those callbacks.
 320  */
 321 struct pstate_funcs {
 322         int (*get_max)(void);
 323         int (*get_max_physical)(void);
 324         int (*get_min)(void);
 325         int (*get_turbo)(void);
 326         int (*get_scaling)(void);
 327         u64 (*get_val)(struct cpudata*, int pstate);
 328         void (*get_vid)(struct cpudata *);
 329         int32_t (*get_target_pstate)(struct cpudata *);
 330 };
 331
 332 /**
 333  * struct cpu_defaults- Per CPU model default config data
 334  * @pid_policy: PID config data
 335  * @funcs:              Callback function data
 336  */
 337 struct cpu_defaults {
 338         struct pstate_adjust_policy pid_policy;
 339         struct pstate_funcs funcs;
 340 };
 341
 342 static inline int32_t get_target_pstate_use_performance(struct cpudata *cpu);
 343 static inline int32_t get_target_pstate_use_cpu_load(struct cpudata *cpu);
 344
 345 static struct pstate_adjust_policy pid_params __read_mostly;
 346 static struct pstate_funcs pstate_funcs __read_mostly;
 347 static int hwp_active __read_mostly;
 348 static bool per_cpu_limits __read_mostly;
 349
 350 static bool driver_registered __read_mostly;
 351
 352 #ifdef CONFIG_ACPI
 353 static bool acpi_ppc;
 354 #endif
 355
 356 static struct global_params global;
 357
 358 static DEFINE_MUTEX(intel_pstate_driver_lock);
 359 static DEFINE_MUTEX(intel_pstate_limits_lock);
 360
 361 #ifdef CONFIG_ACPI
 362
 363 static bool intel_pstate_get_ppc_enable_status(void)
 364 {
 365         if (acpi_gbl_FADT.preferred_profile == PM_ENTERPRISE_SERVER ||
 366             acpi_gbl_FADT.preferred_profile == PM_PERFORMANCE_SERVER)
 367                 return true;
 368
 369         return acpi_ppc;
 370 }
 371
 372 #ifdef CONFIG_ACPI_CPPC_LIB
 373
 374 /* The work item is needed to avoid CPU hotplug locking issues */
 375 static void intel_pstste_sched_itmt_work_fn(struct work_struct *work)
 376 {
 377         sched_set_itmt_support();
 378 }
 379
 380 static DECLARE_WORK(sched_itmt_work, intel_pstste_sched_itmt_work_fn);
 381
 382 static void intel_pstate_set_itmt_prio(int cpu)
 383 {
 384         struct cppc_perf_caps cppc_perf;
 385         static u32 max_highest_perf = 0, min_highest_perf = U32_MAX;
 386         int ret;
 387
 388         ret = cppc_get_perf_caps(cpu, &cppc_perf);
 389         if (ret)
 390                 return;
 391
 392         /*
 393          * The priorities can be set regardless of whether or not
 394          * sched_set_itmt_support(true) has been called and it is valid to
 395          * update them at any time after it has been called.
 396          */
 397         sched_set_itmt_core_prio(cppc_perf.highest_perf, cpu);
 398
 399         if (max_highest_perf <= min_highest_perf) {
 400                 if (cppc_perf.highest_perf > max_highest_perf)
 401                         max_highest_perf = cppc_perf.highest_perf;
 402
 403                 if (cppc_perf.highest_perf < min_highest_perf)
 404                         min_highest_perf = cppc_perf.highest_perf;
 405
 406                 if (max_highest_perf > min_highest_perf) {
 407                         /*
 408                          * This code can be run during CPU online under the
 409                          * CPU hotplug locks, so sched_set_itmt_support()
 410                          * cannot be called from here.  Queue up a work item
 411                          * to invoke it.
 412                          */
 413                         schedule_work(&sched_itmt_work);
 414                 }
 415         }
 416 }
 417 #else
 418 static void intel_pstate_set_itmt_prio(int cpu)
 419 {
 420 }
 421 #endif
 422
 423 static void intel_pstate_init_acpi_perf_limits(struct cpufreq_policy *policy)
 424 {
 425         struct cpudata *cpu;
 426         int ret;
 427         int i;
 428
 429         if (hwp_active) {
 430                 intel_pstate_set_itmt_prio(policy->cpu);
 431                 return;
 432         }
 433
 434         if (!intel_pstate_get_ppc_enable_status())
 435                 return;
 436
 437         cpu = all_cpu_data[policy->cpu];
 438
 439         ret = acpi_processor_register_performance(&cpu->acpi_perf_data,
 440                                                   policy->cpu);
 441         if (ret)
 442                 return;
 443
 444         /*
 445          * Check if the control value in _PSS is for PERF_CTL MSR, which should
 446          * guarantee that the states returned by it map to the states in our
 447          * list directly.
 448          */
 449         if (cpu->acpi_perf_data.control_register.space_id !=
 450                                                 ACPI_ADR_SPACE_FIXED_HARDWARE)
 451                 goto err;
 452
 453         /*
 454          * If there is only one entry _PSS, simply ignore _PSS and continue as
 455          * usual without taking _PSS into account
 456          */
 457         if (cpu->acpi_perf_data.state_count < 2)
 458                 goto err;
 459
 460         pr_debug("CPU%u - ACPI _PSS perf data\n", policy->cpu);
 461         for (i = 0; i < cpu->acpi_perf_data.state_count; i++) {
 462                 pr_debug("     %cP%d: %u MHz, %u mW, 0x%x\n",
 463                          (i == cpu->acpi_perf_data.state ? '*' : ' '), i,
 464                          (u32) cpu->acpi_perf_data.states[i].core_frequency,
 465                          (u32) cpu->acpi_perf_data.states[i].power,
 466                          (u32) cpu->acpi_perf_data.states[i].control);
 467         }
 468
 469         /*
 470          * The _PSS table doesn't contain whole turbo frequency range.
 471          * This just contains +1 MHZ above the max non turbo frequency,
 472          * with control value corresponding to max turbo ratio. But
 473          * when cpufreq set policy is called, it will call with this
 474          * max frequency, which will cause a reduced performance as
 475          * this driver uses real max turbo frequency as the max
 476          * frequency. So correct this frequency in _PSS table to
 477          * correct max turbo frequency based on the turbo state.
 478          * Also need to convert to MHz as _PSS freq is in MHz.
 479          */
 480         if (!global.turbo_disabled)
 481                 cpu->acpi_perf_data.states[0].core_frequency =
 482                                         policy->cpuinfo.max_freq / 1000;
 483         cpu->valid_pss_table = true;
 484         pr_debug("_PPC limits will be enforced\n");
 485
 486         return;
 487
 488  err:
 489         cpu->valid_pss_table = false;
 490         acpi_processor_unregister_performance(policy->cpu);
 491 }
 492
 493 static void intel_pstate_exit_perf_limits(struct cpufreq_policy *policy)
 494 {
 495         struct cpudata *cpu;
 496
 497         cpu = all_cpu_data[policy->cpu];
 498         if (!cpu->valid_pss_table)
 499                 return;
 500
 501         acpi_processor_unregister_performance(policy->cpu);
 502 }
 503 #else
 504 static inline void intel_pstate_init_acpi_perf_limits(struct cpufreq_policy *policy)
 505 {
 506 }
 507
 508 static inline void intel_pstate_exit_perf_limits(struct cpufreq_policy *policy)
 509 {
 510 }
 511 #endif
 512
 513 static inline void pid_reset(struct _pid *pid, int setpoint, int busy,
 514                              int deadband, int integral) {
 515         pid->setpoint = int_tofp(setpoint);
 516         pid->deadband  = int_tofp(deadband);
 517         pid->integral  = int_tofp(integral);
 518         pid->last_err  = int_tofp(setpoint) - int_tofp(busy);
 519 }
 520
 521 static inline void pid_p_gain_set(struct _pid *pid, int percent)
 522 {
 523         pid->p_gain = div_fp(percent, 100);
 524 }
 525
 526 static inline void pid_i_gain_set(struct _pid *pid, int percent)
 527 {
 528         pid->i_gain = div_fp(percent, 100);
 529 }
 530
 531 static inline void pid_d_gain_set(struct _pid *pid, int percent)
 532 {
 533         pid->d_gain = div_fp(percent, 100);
 534 }
 535
 536 static signed int pid_calc(struct _pid *pid, int32_t busy)
 537 {
 538         signed int result;
 539         int32_t pterm, dterm, fp_error;
 540         int32_t integral_limit;
 541
 542         fp_error = pid->setpoint - busy;
 543
 544         if (abs(fp_error) <= pid->deadband)
 545                 return 0;
 546
 547         pterm = mul_fp(pid->p_gain, fp_error);
 548
 549         pid->integral += fp_error;
 550
 551         /*
 552          * We limit the integral here so that it will never
 553          * get higher than 30.  This prevents it from becoming
 554          * too large an input over long periods of time and allows
 555          * it to get factored out sooner.
 556          *
 557          * The value of 30 was chosen through experimentation.
 558          */
 559         integral_limit = int_tofp(30);
 560         if (pid->integral > integral_limit)
 561                 pid->integral = integral_limit;
 562         if (pid->integral < -integral_limit)
 563                 pid->integral = -integral_limit;
 564
 565         dterm = mul_fp(pid->d_gain, fp_error - pid->last_err);
 566         pid->last_err = fp_error;
 567
 568         result = pterm + mul_fp(pid->integral, pid->i_gain) + dterm;
 569         result = result + (1 << (FRAC_BITS-1));
 570         return (signed int)fp_toint(result);
 571 }
 572
 573 static inline void intel_pstate_busy_pid_reset(struct cpudata *cpu)
 574 {
 575         pid_p_gain_set(&cpu->pid, pid_params.p_gain_pct);
 576         pid_d_gain_set(&cpu->pid, pid_params.d_gain_pct);
 577         pid_i_gain_set(&cpu->pid, pid_params.i_gain_pct);
 578
 579         pid_reset(&cpu->pid, pid_params.setpoint, 100, pid_params.deadband, 0);
 580 }
 581
 582 static inline void intel_pstate_reset_all_pid(void)
 583 {
 584         unsigned int cpu;
 585
 586         for_each_online_cpu(cpu) {
 587                 if (all_cpu_data[cpu])
 588                         intel_pstate_busy_pid_reset(all_cpu_data[cpu]);
 589         }
 590 }
 591
 592 static inline void update_turbo_state(void)
 593 {
 594         u64 misc_en;
 595         struct cpudata *cpu;
 596
 597         cpu = all_cpu_data[0];
 598         rdmsrl(MSR_IA32_MISC_ENABLE, misc_en);
 599         global.turbo_disabled =
 600                 (misc_en & MSR_IA32_MISC_ENABLE_TURBO_DISABLE ||
 601                  cpu->pstate.max_pstate == cpu->pstate.turbo_pstate);
 602 }
 603
 604 static int min_perf_pct_min(void)
 605 {
 606         struct cpudata *cpu = all_cpu_data[0];
 607
 608         return DIV_ROUND_UP(cpu->pstate.min_pstate * 100,
 609                             cpu->pstate.turbo_pstate);
 610 }
 611
 612 static s16 intel_pstate_get_epb(struct cpudata *cpu_data)
 613 {
 614         u64 epb;
 615         int ret;
 616
 617         if (!static_cpu_has(X86_FEATURE_EPB))
 618                 return -ENXIO;
 619
 620         ret = rdmsrl_on_cpu(cpu_data->cpu, MSR_IA32_ENERGY_PERF_BIAS, &epb);
 621         if (ret)
 622                 return (s16)ret;
 623
 624         return (s16)(epb & 0x0f);
 625 }
 626
 627 static s16 intel_pstate_get_epp(struct cpudata *cpu_data, u64 hwp_req_data)
 628 {
 629         s16 epp;
 630
 631         if (static_cpu_has(X86_FEATURE_HWP_EPP)) {
 632                 /*
 633                  * When hwp_req_data is 0, means that caller didn't read
 634                  * MSR_HWP_REQUEST, so need to read and get EPP.
 635                  */
 636                 if (!hwp_req_data) {
 637                         epp = rdmsrl_on_cpu(cpu_data->cpu, MSR_HWP_REQUEST,
 638                                             &hwp_req_data);
 639                         if (epp)
 640                                 return epp;
 641                 }
 642                 epp = (hwp_req_data >> 24) & 0xff;
 643         } else {
 644                 /* When there is no EPP present, HWP uses EPB settings */
 645                 epp = intel_pstate_get_epb(cpu_data);
 646         }
 647
 648         return epp;
 649 }
 650
 651 static int intel_pstate_set_epb(int cpu, s16 pref)
 652 {
 653         u64 epb;
 654         int ret;
 655
 656         if (!static_cpu_has(X86_FEATURE_EPB))
 657                 return -ENXIO;
 658
 659         ret = rdmsrl_on_cpu(cpu, MSR_IA32_ENERGY_PERF_BIAS, &epb);
 660         if (ret)
 661                 return ret;
 662
 663         epb = (epb & ~0x0f) | pref;
 664         wrmsrl_on_cpu(cpu, MSR_IA32_ENERGY_PERF_BIAS, epb);
 665
 666         return 0;
 667 }
 668
 669 /*
 670  * EPP/EPB display strings corresponding to EPP index in the
 671  * energy_perf_strings[]
 672  *      index           String
 673  *-------------------------------------
 674  *      0               default
 675  *      1               performance
 676  *      2               balance_performance
 677  *      3               balance_power
 678  *      4               power
 679  */
 680 static const char * const energy_perf_strings[] = {
 681         "default",
 682         "performance",
 683         "balance_performance",
 684         "balance_power",
 685         "power",
 686         NULL
 687 };
 688
 689 static int intel_pstate_get_energy_pref_index(struct cpudata *cpu_data)
 690 {
 691         s16 epp;
 692         int index = -EINVAL;
 693
 694         epp = intel_pstate_get_epp(cpu_data, 0);
 695         if (epp < 0)
 696                 return epp;
 697
 698         if (static_cpu_has(X86_FEATURE_HWP_EPP)) {
 699                 /*
 700                  * Range:
 701                  *      0x00-0x3F       :       Performance
 702                  *      0x40-0x7F       :       Balance performance
 703                  *      0x80-0xBF       :       Balance power
 704                  *      0xC0-0xFF       :       Power
 705                  * The EPP is a 8 bit value, but our ranges restrict the
 706                  * value which can be set. Here only using top two bits
 707                  * effectively.
 708                  */
 709                 index = (epp >> 6) + 1;
 710         } else if (static_cpu_has(X86_FEATURE_EPB)) {
 711                 /*
 712                  * Range:
 713                  *      0x00-0x03       :       Performance
 714                  *      0x04-0x07       :       Balance performance
 715                  *      0x08-0x0B       :       Balance power
 716                  *      0x0C-0x0F       :       Power
 717                  * The EPB is a 4 bit value, but our ranges restrict the
 718                  * value which can be set. Here only using top two bits
 719                  * effectively.
 720                  */
 721                 index = (epp >> 2) + 1;
 722         }
 723
 724         return index;
 725 }
 726
 727 static int intel_pstate_set_energy_pref_index(struct cpudata *cpu_data,
 728                                               int pref_index)
 729 {
 730         int epp = -EINVAL;
 731         int ret;
 732
 733         if (!pref_index)
 734                 epp = cpu_data->epp_default;
 735
 736         mutex_lock(&intel_pstate_limits_lock);
 737
 738         if (static_cpu_has(X86_FEATURE_HWP_EPP)) {
 739                 u64 value;
 740
 741                 ret = rdmsrl_on_cpu(cpu_data->cpu, MSR_HWP_REQUEST, &value);
 742                 if (ret)
 743                         goto return_pref;
 744
 745                 value &= ~GENMASK_ULL(31, 24);
 746
 747                 /*
 748                  * If epp is not default, convert from index into
 749                  * energy_perf_strings to epp value, by shifting 6
 750                  * bits left to use only top two bits in epp.
 751                  * The resultant epp need to shifted by 24 bits to
 752                  * epp position in MSR_HWP_REQUEST.
 753                  */
 754                 if (epp == -EINVAL)
 755                         epp = (pref_index - 1) << 6;
 756
 757                 value |= (u64)epp << 24;
 758                 ret = wrmsrl_on_cpu(cpu_data->cpu, MSR_HWP_REQUEST, value);
 759         } else {
 760                 if (epp == -EINVAL)
 761                         epp = (pref_index - 1) << 2;
 762                 ret = intel_pstate_set_epb(cpu_data->cpu, epp);
 763         }
 764 return_pref:
 765         mutex_unlock(&intel_pstate_limits_lock);
 766
 767         return ret;
 768 }
 769
 770 static ssize_t show_energy_performance_available_preferences(
 771                                 struct cpufreq_policy *policy, char *buf)
 772 {
 773         int i = 0;
 774         int ret = 0;
 775
 776         while (energy_perf_strings[i] != NULL)
 777                 ret += sprintf(&buf[ret], "%s ", energy_perf_strings[i++]);
 778
 779         ret += sprintf(&buf[ret], "\n");
 780
 781         return ret;
 782 }
 783
 784 cpufreq_freq_attr_ro(energy_performance_available_preferences);
 785
 786 static ssize_t store_energy_performance_preference(
 787                 struct cpufreq_policy *policy, const char *buf, size_t count)
 788 {
 789         struct cpudata *cpu_data = all_cpu_data[policy->cpu];
 790         char str_preference[21];
 791         int ret, i = 0;
 792
 793         ret = sscanf(buf, "%20s", str_preference);
 794         if (ret != 1)
 795                 return -EINVAL;
 796
 797         while (energy_perf_strings[i] != NULL) {
 798                 if (!strcmp(str_preference, energy_perf_strings[i])) {
 799                         intel_pstate_set_energy_pref_index(cpu_data, i);
 800                         return count;
 801                 }
 802                 ++i;
 803         }
 804
 805         return -EINVAL;
 806 }
 807
 808 static ssize_t show_energy_performance_preference(
 809                                 struct cpufreq_policy *policy, char *buf)
 810 {
 811         struct cpudata *cpu_data = all_cpu_data[policy->cpu];
 812         int preference;
 813
 814         preference = intel_pstate_get_energy_pref_index(cpu_data);
 815         if (preference < 0)
 816                 return preference;
 817
 818         return  sprintf(buf, "%s\n", energy_perf_strings[preference]);
 819 }
 820
 821 cpufreq_freq_attr_rw(energy_performance_preference);
 822
 823 static struct freq_attr *hwp_cpufreq_attrs[] = {
 824         &energy_performance_preference,
 825         &energy_performance_available_preferences,
 826         NULL,
 827 };
 828
 829 static void intel_pstate_hwp_set(struct cpufreq_policy *policy)
 830 {
 831         int min, hw_min, max, hw_max, cpu;
 832         u64 value, cap;
 833
 834         for_each_cpu(cpu, policy->cpus) {
 835                 struct cpudata *cpu_data = all_cpu_data[cpu];
 836                 struct perf_limits *perf_limits = &cpu_data->perf_limits;
 837                 s16 epp;
 838
 839                 rdmsrl_on_cpu(cpu, MSR_HWP_CAPABILITIES, &cap);
 840                 hw_min = HWP_LOWEST_PERF(cap);
 841                 if (global.no_turbo)
 842                         hw_max = HWP_GUARANTEED_PERF(cap);
 843                 else
 844                         hw_max = HWP_HIGHEST_PERF(cap);
 845
 846                 max = fp_ext_toint(hw_max * perf_limits->max_perf);
 847                 if (cpu_data->policy == CPUFREQ_POLICY_PERFORMANCE)
 848                         min = max;
 849                 else
 850                         min = fp_ext_toint(hw_max * perf_limits->min_perf);
 851
 852                 rdmsrl_on_cpu(cpu, MSR_HWP_REQUEST, &value);
 853
 854                 value &= ~HWP_MIN_PERF(~0L);
 855                 value |= HWP_MIN_PERF(min);
 856
 857                 value &= ~HWP_MAX_PERF(~0L);
 858                 value |= HWP_MAX_PERF(max);
 859
 860                 if (cpu_data->epp_policy == cpu_data->policy)
 861                         goto skip_epp;
 862
 863                 cpu_data->epp_policy = cpu_data->policy;
 864
 865                 if (cpu_data->epp_saved >= 0) {
 866                         epp = cpu_data->epp_saved;
 867                         cpu_data->epp_saved = -EINVAL;
 868                         goto update_epp;
 869                 }
 870
 871                 if (cpu_data->policy == CPUFREQ_POLICY_PERFORMANCE) {
 872                         epp = intel_pstate_get_epp(cpu_data, value);
 873                         cpu_data->epp_powersave = epp;
 874                         /* If EPP read was failed, then don't try to write */
 875                         if (epp < 0)
 876                                 goto skip_epp;
 877
 878
 879                         epp = 0;
 880                 } else {
 881                         /* skip setting EPP, when saved value is invalid */
 882                         if (cpu_data->epp_powersave < 0)
 883                                 goto skip_epp;
 884
 885                         /*
 886                          * No need to restore EPP when it is not zero. This
 887                          * means:
 888                          *  - Policy is not changed
 889                          *  - user has manually changed
 890                          *  - Error reading EPB
 891                          */
 892                         epp = intel_pstate_get_epp(cpu_data, value);
 893                         if (epp)
 894                                 goto skip_epp;
 895
 896                         epp = cpu_data->epp_powersave;
 897                 }
 898 update_epp:
 899                 if (static_cpu_has(X86_FEATURE_HWP_EPP)) {
 900                         value &= ~GENMASK_ULL(31, 24);
 901                         value |= (u64)epp << 24;
 902                 } else {
 903                         intel_pstate_set_epb(cpu, epp);
 904                 }
 905 skip_epp:
 906                 wrmsrl_on_cpu(cpu, MSR_HWP_REQUEST, value);
 907         }
 908 }
 909
 910 static int intel_pstate_hwp_save_state(struct cpufreq_policy *policy)
 911 {
 912         struct cpudata *cpu_data = all_cpu_data[policy->cpu];
 913
 914         if (!hwp_active)
 915                 return 0;
 916
 917         cpu_data->epp_saved = intel_pstate_get_epp(cpu_data, 0);
 918
 919         return 0;
 920 }
 921
 922 static int intel_pstate_resume(struct cpufreq_policy *policy)
 923 {
 924         if (!hwp_active)
 925                 return 0;
 926
 927         mutex_lock(&intel_pstate_limits_lock);
 928
 929         all_cpu_data[policy->cpu]->epp_policy = 0;
 930         intel_pstate_hwp_set(policy);
 931
 932         mutex_unlock(&intel_pstate_limits_lock);
 933
 934         return 0;
 935 }
 936
 937 static void intel_pstate_update_policies(void)
 938 {
 939         int cpu;
 940
 941         for_each_possible_cpu(cpu)
 942                 cpufreq_update_policy(cpu);
 943 }
 944
 945 /************************** debugfs begin ************************/
 946 static int pid_param_set(void *data, u64 val)
 947 {
 948         *(u32 *)data = val;
 949         pid_params.sample_rate_ns = pid_params.sample_rate_ms * NSEC_PER_MSEC;
 950         intel_pstate_reset_all_pid();
 951         return 0;
 952 }
 953
 954 static int pid_param_get(void *data, u64 *val)
 955 {
 956         *val = *(u32 *)data;
 957         return 0;
 958 }
 959 DEFINE_SIMPLE_ATTRIBUTE(fops_pid_param, pid_param_get, pid_param_set, "%llu\n");
 960
 961 static struct dentry *debugfs_parent;
 962
 963 struct pid_param {
 964         char *name;
 965         void *value;
 966         struct dentry *dentry;
 967 };
 968
 969 static struct pid_param pid_files[] = {
 970         {"sample_rate_ms", &pid_params.sample_rate_ms, },
 971         {"d_gain_pct", &pid_params.d_gain_pct, },
 972         {"i_gain_pct", &pid_params.i_gain_pct, },
 973         {"deadband", &pid_params.deadband, },
 974         {"setpoint", &pid_params.setpoint, },
 975         {"p_gain_pct", &pid_params.p_gain_pct, },
 976         {NULL, NULL, }
 977 };
 978
 979 static void intel_pstate_debug_expose_params(void)
 980 {
 981         int i;
 982
 983         debugfs_parent = debugfs_create_dir("pstate_snb", NULL);
 984         if (IS_ERR_OR_NULL(debugfs_parent))
 985                 return;
 986
 987         for (i = 0; pid_files[i].name; i++) {
 988                 struct dentry *dentry;
 989
 990                 dentry = debugfs_create_file(pid_files[i].name, 0660,
 991                                              debugfs_parent, pid_files[i].value,
 992                                              &fops_pid_param);
 993                 if (!IS_ERR(dentry))
 994                         pid_files[i].dentry = dentry;
 995         }
 996 }
 997
 998 static void intel_pstate_debug_hide_params(void)
 999 {
1000         int i;
1001
1002         if (IS_ERR_OR_NULL(debugfs_parent))
1003                 return;
1004
1005         for (i = 0; pid_files[i].name; i++) {
1006                 debugfs_remove(pid_files[i].dentry);
1007                 pid_files[i].dentry = NULL;
1008         }
1009
1010         debugfs_remove(debugfs_parent);
1011         debugfs_parent = NULL;
1012 }
1013
1014 /************************** debugfs end ************************/
1015
1016 /************************** sysfs begin ************************/
1017 #define show_one(file_name, object)                                     \
1018         static ssize_t show_##file_name                                 \
1019         (struct kobject *kobj, struct attribute *attr, char *buf)       \
1020         {                                                               \
1021                 return sprintf(buf, "%u\n", global.object);             \
1022         }
1023
1024 static ssize_t intel_pstate_show_status(char *buf);
1025 static int intel_pstate_update_status(const char *buf, size_t size);
1026
1027 static ssize_t show_status(struct kobject *kobj,
1028                            struct attribute *attr, char *buf)
1029 {
1030         ssize_t ret;
1031
1032         mutex_lock(&intel_pstate_driver_lock);
1033         ret = intel_pstate_show_status(buf);
1034         mutex_unlock(&intel_pstate_driver_lock);
1035
1036         return ret;
1037 }
1038
1039 static ssize_t store_status(struct kobject *a, struct attribute *b,
1040                             const char *buf, size_t count)
1041 {
1042         char *p = memchr(buf, '\n', count);
1043         int ret;
1044
1045         mutex_lock(&intel_pstate_driver_lock);
1046         ret = intel_pstate_update_status(buf, p ? p - buf : count);
1047         mutex_unlock(&intel_pstate_driver_lock);
1048
1049         return ret < 0 ? ret : count;
1050 }
1051
1052 static ssize_t show_turbo_pct(struct kobject *kobj,
1053                                 struct attribute *attr, char *buf)
1054 {
1055         struct cpudata *cpu;
1056         int total, no_turbo, turbo_pct;
1057         uint32_t turbo_fp;
1058
1059         mutex_lock(&intel_pstate_driver_lock);
1060
1061         if (!driver_registered) {
1062                 mutex_unlock(&intel_pstate_driver_lock);
1063                 return -EAGAIN;
1064         }
1065
1066         cpu = all_cpu_data[0];
1067
1068         total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
1069         no_turbo = cpu->pstate.max_pstate - cpu->pstate.min_pstate + 1;
1070         turbo_fp = div_fp(no_turbo, total);
1071         turbo_pct = 100 - fp_toint(mul_fp(turbo_fp, int_tofp(100)));
1072
1073         mutex_unlock(&intel_pstate_driver_lock);
1074
1075         return sprintf(buf, "%u\n", turbo_pct);
1076 }
1077
1078 static ssize_t show_num_pstates(struct kobject *kobj,
1079                                 struct attribute *attr, char *buf)
1080 {
1081         struct cpudata *cpu;
1082         int total;
1083
1084         mutex_lock(&intel_pstate_driver_lock);
1085
1086         if (!driver_registered) {
1087                 mutex_unlock(&intel_pstate_driver_lock);
1088                 return -EAGAIN;
1089         }
1090
1091         cpu = all_cpu_data[0];
1092         total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
1093
1094         mutex_unlock(&intel_pstate_driver_lock);
1095
1096         return sprintf(buf, "%u\n", total);
1097 }
1098
1099 static ssize_t show_no_turbo(struct kobject *kobj,
1100                              struct attribute *attr, char *buf)
1101 {
1102         ssize_t ret;
1103
1104         mutex_lock(&intel_pstate_driver_lock);
1105
1106         if (!driver_registered) {
1107                 mutex_unlock(&intel_pstate_driver_lock);
1108                 return -EAGAIN;
1109         }
1110
1111         update_turbo_state();
1112         if (global.turbo_disabled)
1113                 ret = sprintf(buf, "%u\n", global.turbo_disabled);
1114         else
1115                 ret = sprintf(buf, "%u\n", global.no_turbo);
1116
1117         mutex_unlock(&intel_pstate_driver_lock);
1118
1119         return ret;
1120 }
1121
1122 static ssize_t store_no_turbo(struct kobject *a, struct attribute *b,
1123                               const char *buf, size_t count)
1124 {
1125         unsigned int input;
1126         int ret;
1127
1128         ret = sscanf(buf, "%u", &input);
1129         if (ret != 1)
1130                 return -EINVAL;
1131
1132         mutex_lock(&intel_pstate_driver_lock);
1133
1134         if (!driver_registered) {
1135                 mutex_unlock(&intel_pstate_driver_lock);
1136                 return -EAGAIN;
1137         }
1138
1139         mutex_lock(&intel_pstate_limits_lock);
1140
1141         update_turbo_state();
1142         if (global.turbo_disabled) {
1143                 pr_warn("Turbo disabled by BIOS or unavailable on processor\n");
1144                 mutex_unlock(&intel_pstate_limits_lock);
1145                 mutex_unlock(&intel_pstate_driver_lock);
1146                 return -EPERM;
1147         }
1148
1149         global.no_turbo = clamp_t(int, input, 0, 1);
1150
1151         if (global.no_turbo) {
1152                 struct cpudata *cpu = all_cpu_data[0];
1153                 int pct = cpu->pstate.max_pstate * 100 / cpu->pstate.turbo_pstate;
1154
1155                 /* Squash the global minimum into the permitted range. */
1156                 if (global.min_perf_pct > pct)
1157                         global.min_perf_pct = pct;
1158         }
1159
1160         mutex_unlock(&intel_pstate_limits_lock);
1161
1162         intel_pstate_update_policies();
1163
1164         mutex_unlock(&intel_pstate_driver_lock);
1165
1166         return count;
1167 }
1168
1169 static ssize_t store_max_perf_pct(struct kobject *a, struct attribute *b,
1170                                   const char *buf, size_t count)
1171 {
1172         unsigned int input;
1173         int ret;
1174
1175         ret = sscanf(buf, "%u", &input);
1176         if (ret != 1)
1177                 return -EINVAL;
1178
1179         mutex_lock(&intel_pstate_driver_lock);
1180
1181         if (!driver_registered) {
1182                 mutex_unlock(&intel_pstate_driver_lock);
1183                 return -EAGAIN;
1184         }
1185
1186         mutex_lock(&intel_pstate_limits_lock);
1187
1188         global.max_perf_pct = clamp_t(int, input, global.min_perf_pct, 100);
1189
1190         mutex_unlock(&intel_pstate_limits_lock);
1191
1192         intel_pstate_update_policies();
1193
1194         mutex_unlock(&intel_pstate_driver_lock);
1195
1196         return count;
1197 }
1198
1199 static ssize_t store_min_perf_pct(struct kobject *a, struct attribute *b,
1200                                   const char *buf, size_t count)
1201 {
1202         unsigned int input;
1203         int ret;
1204
1205         ret = sscanf(buf, "%u", &input);
1206         if (ret != 1)
1207                 return -EINVAL;
1208
1209         mutex_lock(&intel_pstate_driver_lock);
1210
1211         if (!driver_registered) {
1212                 mutex_unlock(&intel_pstate_driver_lock);
1213                 return -EAGAIN;
1214         }
1215
1216         mutex_lock(&intel_pstate_limits_lock);
1217
1218         global.min_perf_pct = clamp_t(int, input,
1219                                       min_perf_pct_min(), global.max_perf_pct);
1220
1221         mutex_unlock(&intel_pstate_limits_lock);
1222
1223         intel_pstate_update_policies();
1224
1225         mutex_unlock(&intel_pstate_driver_lock);
1226
1227         return count;
1228 }
1229
1230 show_one(max_perf_pct, max_perf_pct);
1231 show_one(min_perf_pct, min_perf_pct);
1232
1233 define_one_global_rw(status);
1234 define_one_global_rw(no_turbo);
1235 define_one_global_rw(max_perf_pct);
1236 define_one_global_rw(min_perf_pct);
1237 define_one_global_ro(turbo_pct);
1238 define_one_global_ro(num_pstates);
1239
1240 static struct attribute *intel_pstate_attributes[] = {
1241         &status.attr,
1242         &no_turbo.attr,
1243         &turbo_pct.attr,
1244         &num_pstates.attr,
1245         NULL
1246 };
1247
1248 static struct attribute_group intel_pstate_attr_group = {
1249         .attrs = intel_pstate_attributes,
1250 };
1251
1252 static void __init intel_pstate_sysfs_expose_params(void)
1253 {
1254         struct kobject *intel_pstate_kobject;
1255         int rc;
1256
1257         intel_pstate_kobject = kobject_create_and_add("intel_pstate",
1258                                                 &cpu_subsys.dev_root->kobj);
1259         if (WARN_ON(!intel_pstate_kobject))
1260                 return;
1261
1262         rc = sysfs_create_group(intel_pstate_kobject, &intel_pstate_attr_group);
1263         if (WARN_ON(rc))
1264                 return;
1265
1266         /*
1267          * If per cpu limits are enforced there are no global limits, so
1268          * return without creating max/min_perf_pct attributes
1269          */
1270         if (per_cpu_limits)
1271                 return;
1272
1273         rc = sysfs_create_file(intel_pstate_kobject, &max_perf_pct.attr);
1274         WARN_ON(rc);
1275
1276         rc = sysfs_create_file(intel_pstate_kobject, &min_perf_pct.attr);
1277         WARN_ON(rc);
1278
1279 }
1280 /************************** sysfs end ************************/
1281
1282 static void intel_pstate_hwp_enable(struct cpudata *cpudata)
1283 {
1284         /* First disable HWP notification interrupt as we don't process them */
1285         if (static_cpu_has(X86_FEATURE_HWP_NOTIFY))
1286                 wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_INTERRUPT, 0x00);
1287
1288         wrmsrl_on_cpu(cpudata->cpu, MSR_PM_ENABLE, 0x1);
1289         cpudata->epp_policy = 0;
1290         if (cpudata->epp_default == -EINVAL)
1291                 cpudata->epp_default = intel_pstate_get_epp(cpudata, 0);
1292 }
1293
1294 #define MSR_IA32_POWER_CTL_BIT_EE       19
1295
1296 /* Disable energy efficiency optimization */
1297 static void intel_pstate_disable_ee(int cpu)
1298 {
1299         u64 power_ctl;
1300         int ret;
1301
1302         ret = rdmsrl_on_cpu(cpu, MSR_IA32_POWER_CTL, &power_ctl);
1303         if (ret)
1304                 return;
1305
1306         if (!(power_ctl & BIT(MSR_IA32_POWER_CTL_BIT_EE))) {
1307                 pr_info("Disabling energy efficiency optimization\n");
1308                 power_ctl |= BIT(MSR_IA32_POWER_CTL_BIT_EE);
1309                 wrmsrl_on_cpu(cpu, MSR_IA32_POWER_CTL, power_ctl);
1310         }
1311 }
1312
1313 static int atom_get_min_pstate(void)
1314 {
1315         u64 value;
1316
1317         rdmsrl(MSR_ATOM_CORE_RATIOS, value);
1318         return (value >> 8) & 0x7F;
1319 }
1320
1321 static int atom_get_max_pstate(void)
1322 {
1323         u64 value;
1324
1325         rdmsrl(MSR_ATOM_CORE_RATIOS, value);
1326         return (value >> 16) & 0x7F;
1327 }
1328
1329 static int atom_get_turbo_pstate(void)
1330 {
1331         u64 value;
1332
1333         rdmsrl(MSR_ATOM_CORE_TURBO_RATIOS, value);
1334         return value & 0x7F;
1335 }
1336
1337 static u64 atom_get_val(struct cpudata *cpudata, int pstate)
1338 {
1339         u64 val;
1340         int32_t vid_fp;
1341         u32 vid;
1342
1343         val = (u64)pstate << 8;
1344         if (global.no_turbo && !global.turbo_disabled)
1345                 val |= (u64)1 << 32;
1346
1347         vid_fp = cpudata->vid.min + mul_fp(
1348                 int_tofp(pstate - cpudata->pstate.min_pstate),
1349                 cpudata->vid.ratio);
1350
1351         vid_fp = clamp_t(int32_t, vid_fp, cpudata->vid.min, cpudata->vid.max);
1352         vid = ceiling_fp(vid_fp);
1353
1354         if (pstate > cpudata->pstate.max_pstate)
1355                 vid = cpudata->vid.turbo;
1356
1357         return val | vid;
1358 }
1359
1360 static int silvermont_get_scaling(void)
1361 {
1362         u64 value;
1363         int i;
1364         /* Defined in Table 35-6 from SDM (Sept 2015) */
1365         static int silvermont_freq_table[] = {
1366                 83300, 100000, 133300, 116700, 80000};
1367
1368         rdmsrl(MSR_FSB_FREQ, value);
1369         i = value & 0x7;
1370         WARN_ON(i > 4);
1371
1372         return silvermont_freq_table[i];
1373 }
1374
1375 static int airmont_get_scaling(void)
1376 {
1377         u64 value;
1378         int i;
1379         /* Defined in Table 35-10 from SDM (Sept 2015) */
1380         static int airmont_freq_table[] = {
1381                 83300, 100000, 133300, 116700, 80000,
1382                 93300, 90000, 88900, 87500};
1383
1384         rdmsrl(MSR_FSB_FREQ, value);
1385         i = value & 0xF;
1386         WARN_ON(i > 8);
1387
1388         return airmont_freq_table[i];
1389 }
1390
1391 static void atom_get_vid(struct cpudata *cpudata)
1392 {
1393         u64 value;
1394
1395         rdmsrl(MSR_ATOM_CORE_VIDS, value);
1396         cpudata->vid.min = int_tofp((value >> 8) & 0x7f);
1397         cpudata->vid.max = int_tofp((value >> 16) & 0x7f);
1398         cpudata->vid.ratio = div_fp(
1399                 cpudata->vid.max - cpudata->vid.min,
1400                 int_tofp(cpudata->pstate.max_pstate -
1401                         cpudata->pstate.min_pstate));
1402
1403         rdmsrl(MSR_ATOM_CORE_TURBO_VIDS, value);
1404         cpudata->vid.turbo = value & 0x7f;
1405 }
1406
1407 static int core_get_min_pstate(void)
1408 {
1409         u64 value;
1410
1411         rdmsrl(MSR_PLATFORM_INFO, value);
1412         return (value >> 40) & 0xFF;
1413 }
1414
1415 static int core_get_max_pstate_physical(void)
1416 {
1417         u64 value;
1418
1419         rdmsrl(MSR_PLATFORM_INFO, value);
1420         return (value >> 8) & 0xFF;
1421 }
1422
1423 static int core_get_tdp_ratio(u64 plat_info)
1424 {
1425         /* Check how many TDP levels present */
1426         if (plat_info & 0x600000000) {
1427                 u64 tdp_ctrl;
1428                 u64 tdp_ratio;
1429                 int tdp_msr;
1430                 int err;
1431
1432                 /* Get the TDP level (0, 1, 2) to get ratios */
1433                 err = rdmsrl_safe(MSR_CONFIG_TDP_CONTROL, &tdp_ctrl);
1434                 if (err)
1435                         return err;
1436
1437                 /* TDP MSR are continuous starting at 0x648 */
1438                 tdp_msr = MSR_CONFIG_TDP_NOMINAL + (tdp_ctrl & 0x03);
1439                 err = rdmsrl_safe(tdp_msr, &tdp_ratio);
1440                 if (err)
1441                         return err;
1442
1443                 /* For level 1 and 2, bits[23:16] contain the ratio */
1444                 if (tdp_ctrl & 0x03)
1445                         tdp_ratio >>= 16;
1446
1447                 tdp_ratio &= 0xff; /* ratios are only 8 bits long */
1448                 pr_debug("tdp_ratio %x\n", (int)tdp_ratio);
1449
1450                 return (int)tdp_ratio;
1451         }
1452
1453         return -ENXIO;
1454 }
1455
1456 static int core_get_max_pstate(void)
1457 {
1458         u64 tar;
1459         u64 plat_info;
1460         int max_pstate;
1461         int tdp_ratio;
1462         int err;
1463
1464         rdmsrl(MSR_PLATFORM_INFO, plat_info);
1465         max_pstate = (plat_info >> 8) & 0xFF;
1466
1467         tdp_ratio = core_get_tdp_ratio(plat_info);
1468         if (tdp_ratio <= 0)
1469                 return max_pstate;
1470
1471         if (hwp_active) {
1472                 /* Turbo activation ratio is not used on HWP platforms */
1473                 return tdp_ratio;
1474         }
1475
1476         err = rdmsrl_safe(MSR_TURBO_ACTIVATION_RATIO, &tar);
1477         if (!err) {
1478                 int tar_levels;
1479
1480                 /* Do some sanity checking for safety */
1481                 tar_levels = tar & 0xff;
1482                 if (tdp_ratio - 1 == tar_levels) {
1483                         max_pstate = tar_levels;
1484                         pr_debug("max_pstate=TAC %x\n", max_pstate);
1485                 }
1486         }
1487
1488         return max_pstate;
1489 }
1490
1491 static int core_get_turbo_pstate(void)
1492 {
1493         u64 value;
1494         int nont, ret;
1495
1496         rdmsrl(MSR_TURBO_RATIO_LIMIT, value);
1497         nont = core_get_max_pstate();
1498         ret = (value) & 255;
1499         if (ret <= nont)
1500                 ret = nont;
1501         return ret;
1502 }
1503
1504 static inline int core_get_scaling(void)
1505 {
1506         return 100000;
1507 }
1508
1509 static u64 core_get_val(struct cpudata *cpudata, int pstate)
1510 {
1511         u64 val;
1512
1513         val = (u64)pstate << 8;
1514         if (global.no_turbo && !global.turbo_disabled)
1515                 val |= (u64)1 << 32;
1516
1517         return val;
1518 }
1519
1520 static int knl_get_turbo_pstate(void)
1521 {
1522         u64 value;
1523         int nont, ret;
1524
1525         rdmsrl(MSR_TURBO_RATIO_LIMIT, value);
1526         nont = core_get_max_pstate();
1527         ret = (((value) >> 8) & 0xFF);
1528         if (ret <= nont)
1529                 ret = nont;
1530         return ret;
1531 }
1532
1533 static struct cpu_defaults core_params = {
1534         .pid_policy = {
1535                 .sample_rate_ms = 10,
1536                 .deadband = 0,
1537                 .setpoint = 97,
1538                 .p_gain_pct = 20,
1539                 .d_gain_pct = 0,
1540                 .i_gain_pct = 0,
1541         },
1542         .funcs = {
1543                 .get_max = core_get_max_pstate,
1544                 .get_max_physical = core_get_max_pstate_physical,
1545                 .get_min = core_get_min_pstate,
1546                 .get_turbo = core_get_turbo_pstate,
1547                 .get_scaling = core_get_scaling,
1548                 .get_val = core_get_val,
1549                 .get_target_pstate = get_target_pstate_use_performance,
1550         },
1551 };
1552
1553 static const struct cpu_defaults silvermont_params = {
1554         .pid_policy = {
1555                 .sample_rate_ms = 10,
1556                 .deadband = 0,
1557                 .setpoint = 60,
1558                 .p_gain_pct = 14,
1559                 .d_gain_pct = 0,
1560                 .i_gain_pct = 4,
1561         },
1562         .funcs = {
1563                 .get_max = atom_get_max_pstate,
1564                 .get_max_physical = atom_get_max_pstate,
1565                 .get_min = atom_get_min_pstate,
1566                 .get_turbo = atom_get_turbo_pstate,
1567                 .get_val = atom_get_val,
1568                 .get_scaling = silvermont_get_scaling,
1569                 .get_vid = atom_get_vid,
1570                 .get_target_pstate = get_target_pstate_use_cpu_load,
1571         },
1572 };
1573
1574 static const struct cpu_defaults airmont_params = {
1575         .pid_policy = {
1576                 .sample_rate_ms = 10,
1577                 .deadband = 0,
1578                 .setpoint = 60,
1579                 .p_gain_pct = 14,
1580                 .d_gain_pct = 0,
1581                 .i_gain_pct = 4,
1582         },
1583         .funcs = {
1584                 .get_max = atom_get_max_pstate,
1585                 .get_max_physical = atom_get_max_pstate,
1586                 .get_min = atom_get_min_pstate,
1587                 .get_turbo = atom_get_turbo_pstate,
1588                 .get_val = atom_get_val,
1589                 .get_scaling = airmont_get_scaling,
1590                 .get_vid = atom_get_vid,
1591                 .get_target_pstate = get_target_pstate_use_cpu_load,
1592         },
1593 };
1594
1595 static const struct cpu_defaults knl_params = {
1596         .pid_policy = {
1597                 .sample_rate_ms = 10,
1598                 .deadband = 0,
1599                 .setpoint = 97,
1600                 .p_gain_pct = 20,
1601                 .d_gain_pct = 0,
1602                 .i_gain_pct = 0,
1603         },
1604         .funcs = {
1605                 .get_max = core_get_max_pstate,
1606                 .get_max_physical = core_get_max_pstate_physical,
1607                 .get_min = core_get_min_pstate,
1608                 .get_turbo = knl_get_turbo_pstate,
1609                 .get_scaling = core_get_scaling,
1610                 .get_val = core_get_val,
1611                 .get_target_pstate = get_target_pstate_use_performance,
1612         },
1613 };
1614
1615 static const struct cpu_defaults bxt_params = {
1616         .pid_policy = {
1617                 .sample_rate_ms = 10,
1618                 .deadband = 0,
1619                 .setpoint = 60,
1620                 .p_gain_pct = 14,
1621                 .d_gain_pct = 0,
1622                 .i_gain_pct = 4,
1623         },
1624         .funcs = {
1625                 .get_max = core_get_max_pstate,
1626                 .get_max_physical = core_get_max_pstate_physical,
1627                 .get_min = core_get_min_pstate,
1628                 .get_turbo = core_get_turbo_pstate,
1629                 .get_scaling = core_get_scaling,
1630                 .get_val = core_get_val,
1631                 .get_target_pstate = get_target_pstate_use_cpu_load,
1632         },
1633 };
1634
1635 static void intel_pstate_get_min_max(struct cpudata *cpu, int *min, int *max)
1636 {
1637         int max_perf = cpu->pstate.turbo_pstate;
1638         int max_perf_adj;
1639         int min_perf;
1640         struct perf_limits *perf_limits = &cpu->perf_limits;
1641
1642         if (global.no_turbo || global.turbo_disabled)
1643                 max_perf = cpu->pstate.max_pstate;
1644
1645         /*
1646          * performance can be limited by user through sysfs, by cpufreq
1647          * policy, or by cpu specific default values determined through
1648          * experimentation.
1649          */
1650         max_perf_adj = fp_ext_toint(max_perf * perf_limits->max_perf);
1651         *max = clamp_t(int, max_perf_adj,
1652                         cpu->pstate.min_pstate, cpu->pstate.turbo_pstate);
1653
1654         min_perf = fp_ext_toint(max_perf * perf_limits->min_perf);
1655         *min = clamp_t(int, min_perf, cpu->pstate.min_pstate, max_perf);
1656 }
1657
1658 static void intel_pstate_set_pstate(struct cpudata *cpu, int pstate)
1659 {
1660         trace_cpu_frequency(pstate * cpu->pstate.scaling, cpu->cpu);
1661         cpu->pstate.current_pstate = pstate;
1662         /*
1663          * Generally, there is no guarantee that this code will always run on
1664          * the CPU being updated, so force the register update to run on the
1665          * right CPU.
1666          */
1667         wrmsrl_on_cpu(cpu->cpu, MSR_IA32_PERF_CTL,
1668                       pstate_funcs.get_val(cpu, pstate));
1669 }
1670
1671 static void intel_pstate_set_min_pstate(struct cpudata *cpu)
1672 {
1673         intel_pstate_set_pstate(cpu, cpu->pstate.min_pstate);
1674 }
1675
1676 static void intel_pstate_max_within_limits(struct cpudata *cpu)
1677 {
1678         int min_pstate, max_pstate;
1679
1680         update_turbo_state();
1681         intel_pstate_get_min_max(cpu, &min_pstate, &max_pstate);
1682         intel_pstate_set_pstate(cpu, max_pstate);
1683 }
1684
1685 static void intel_pstate_get_cpu_pstates(struct cpudata *cpu)
1686 {
1687         cpu->pstate.min_pstate = pstate_funcs.get_min();
1688         cpu->pstate.max_pstate = pstate_funcs.get_max();
1689         cpu->pstate.max_pstate_physical = pstate_funcs.get_max_physical();
1690         cpu->pstate.turbo_pstate = pstate_funcs.get_turbo();
1691         cpu->pstate.scaling = pstate_funcs.get_scaling();
1692         cpu->pstate.max_freq = cpu->pstate.max_pstate * cpu->pstate.scaling;
1693         cpu->pstate.turbo_freq = cpu->pstate.turbo_pstate * cpu->pstate.scaling;
1694
1695         if (pstate_funcs.get_vid)
1696                 pstate_funcs.get_vid(cpu);
1697
1698         intel_pstate_set_min_pstate(cpu);
1699 }
1700
1701 static inline void intel_pstate_calc_avg_perf(struct cpudata *cpu)
1702 {
1703         struct sample *sample = &cpu->sample;
1704
1705         sample->core_avg_perf = div_ext_fp(sample->aperf, sample->mperf);
1706 }
1707
1708 static inline bool intel_pstate_sample(struct cpudata *cpu, u64 time)
1709 {
1710         u64 aperf, mperf;
1711         unsigned long flags;
1712         u64 tsc;
1713
1714         local_irq_save(flags);
1715         rdmsrl(MSR_IA32_APERF, aperf);
1716         rdmsrl(MSR_IA32_MPERF, mperf);
1717         tsc = rdtsc();
1718         if (cpu->prev_mperf == mperf || cpu->prev_tsc == tsc) {
1719                 local_irq_restore(flags);
1720                 return false;
1721         }
1722         local_irq_restore(flags);
1723
1724         cpu->last_sample_time = cpu->sample.time;
1725         cpu->sample.time = time;
1726         cpu->sample.aperf = aperf;
1727         cpu->sample.mperf = mperf;
1728         cpu->sample.tsc =  tsc;
1729         cpu->sample.aperf -= cpu->prev_aperf;
1730         cpu->sample.mperf -= cpu->prev_mperf;
1731         cpu->sample.tsc -= cpu->prev_tsc;
1732
1733         cpu->prev_aperf = aperf;
1734         cpu->prev_mperf = mperf;
1735         cpu->prev_tsc = tsc;
1736         /*
1737          * First time this function is invoked in a given cycle, all of the
1738          * previous sample data fields are equal to zero or stale and they must
1739          * be populated with meaningful numbers for things to work, so assume
1740          * that sample.time will always be reset before setting the utilization
1741          * update hook and make the caller skip the sample then.
1742          */
1743         return !!cpu->last_sample_time;
1744 }
1745
1746 static inline int32_t get_avg_frequency(struct cpudata *cpu)
1747 {
1748         return mul_ext_fp(cpu->sample.core_avg_perf,
1749                           cpu->pstate.max_pstate_physical * cpu->pstate.scaling);
1750 }
1751
1752 static inline int32_t get_avg_pstate(struct cpudata *cpu)
1753 {
1754         return mul_ext_fp(cpu->pstate.max_pstate_physical,
1755                           cpu->sample.core_avg_perf);
1756 }
1757
1758 static inline int32_t get_target_pstate_use_cpu_load(struct cpudata *cpu)
1759 {
1760         struct sample *sample = &cpu->sample;
1761         int32_t busy_frac, boost;
1762         int target, avg_pstate;
1763
1764         busy_frac = div_fp(sample->mperf, sample->tsc);
1765
1766         boost = cpu->iowait_boost;
1767         cpu->iowait_boost >>= 1;
1768
1769         if (busy_frac < boost)
1770                 busy_frac = boost;
1771
1772         sample->busy_scaled = busy_frac * 100;
1773
1774         target = global.no_turbo || global.turbo_disabled ?
1775                         cpu->pstate.max_pstate : cpu->pstate.turbo_pstate;
1776         target += target >> 2;
1777         target = mul_fp(target, busy_frac);
1778         if (target < cpu->pstate.min_pstate)
1779                 target = cpu->pstate.min_pstate;
1780
1781         /*
1782          * If the average P-state during the previous cycle was higher than the
1783          * current target, add 50% of the difference to the target to reduce
1784          * possible performance oscillations and offset possible performance
1785          * loss related to moving the workload from one CPU to another within
1786          * a package/module.
1787          */
1788         avg_pstate = get_avg_pstate(cpu);
1789         if (avg_pstate > target)
1790                 target += (avg_pstate - target) >> 1;
1791
1792         return target;
1793 }
1794
1795 static inline int32_t get_target_pstate_use_performance(struct cpudata *cpu)
1796 {
1797         int32_t perf_scaled, max_pstate, current_pstate, sample_ratio;
1798         u64 duration_ns;
1799
1800         /*
1801          * perf_scaled is the ratio of the average P-state during the last
1802          * sampling period to the P-state requested last time (in percent).
1803          *
1804          * That measures the system's response to the previous P-state
1805          * selection.
1806          */
1807         max_pstate = cpu->pstate.max_pstate_physical;
1808         current_pstate = cpu->pstate.current_pstate;
1809         perf_scaled = mul_ext_fp(cpu->sample.core_avg_perf,
1810                                div_fp(100 * max_pstate, current_pstate));
1811
1812         /*
1813          * Since our utilization update callback will not run unless we are
1814          * in C0, check if the actual elapsed time is significantly greater (3x)
1815          * than our sample interval.  If it is, then we were idle for a long
1816          * enough period of time to adjust our performance metric.
1817          */
1818         duration_ns = cpu->sample.time - cpu->last_sample_time;
1819         if ((s64)duration_ns > pid_params.sample_rate_ns * 3) {
1820                 sample_ratio = div_fp(pid_params.sample_rate_ns, duration_ns);
1821                 perf_scaled = mul_fp(perf_scaled, sample_ratio);
1822         } else {
1823                 sample_ratio = div_fp(100 * cpu->sample.mperf, cpu->sample.tsc);
1824                 if (sample_ratio < int_tofp(1))
1825                         perf_scaled = 0;
1826         }
1827
1828         cpu->sample.busy_scaled = perf_scaled;
1829         return cpu->pstate.current_pstate - pid_calc(&cpu->pid, perf_scaled);
1830 }
1831
1832 static int intel_pstate_prepare_request(struct cpudata *cpu, int pstate)
1833 {
1834         int max_perf, min_perf;
1835
1836         intel_pstate_get_min_max(cpu, &min_perf, &max_perf);
1837         pstate = clamp_t(int, pstate, min_perf, max_perf);
1838         return pstate;
1839 }
1840
1841 static void intel_pstate_update_pstate(struct cpudata *cpu, int pstate)
1842 {
1843         if (pstate == cpu->pstate.current_pstate)
1844                 return;
1845
1846         cpu->pstate.current_pstate = pstate;
1847         wrmsrl(MSR_IA32_PERF_CTL, pstate_funcs.get_val(cpu, pstate));
1848 }
1849
1850 static inline void intel_pstate_adjust_busy_pstate(struct cpudata *cpu)
1851 {
1852         int from, target_pstate;
1853         struct sample *sample;
1854
1855         from = cpu->pstate.current_pstate;
1856
1857         target_pstate = cpu->policy == CPUFREQ_POLICY_PERFORMANCE ?
1858                 cpu->pstate.turbo_pstate : pstate_funcs.get_target_pstate(cpu);
1859
1860         update_turbo_state();
1861
1862         target_pstate = intel_pstate_prepare_request(cpu, target_pstate);
1863         trace_cpu_frequency(target_pstate * cpu->pstate.scaling, cpu->cpu);
1864         intel_pstate_update_pstate(cpu, target_pstate);
1865
1866         sample = &cpu->sample;
1867         trace_pstate_sample(mul_ext_fp(100, sample->core_avg_perf),
1868                 fp_toint(sample->busy_scaled),
1869                 from,
1870                 cpu->pstate.current_pstate,
1871                 sample->mperf,
1872                 sample->aperf,
1873                 sample->tsc,
1874                 get_avg_frequency(cpu),
1875                 fp_toint(cpu->iowait_boost * 100));
1876 }
1877
1878 static void intel_pstate_update_util(struct update_util_data *data, u64 time,
1879                                      unsigned int flags)
1880 {
1881         struct cpudata *cpu = container_of(data, struct cpudata, update_util);
1882         u64 delta_ns;
1883
1884         if (pstate_funcs.get_target_pstate == get_target_pstate_use_cpu_load) {
1885                 if (flags & SCHED_CPUFREQ_IOWAIT) {
1886                         cpu->iowait_boost = int_tofp(1);
1887                 } else if (cpu->iowait_boost) {
1888                         /* Clear iowait_boost if the CPU may have been idle. */
1889                         delta_ns = time - cpu->last_update;
1890                         if (delta_ns > TICK_NSEC)
1891                                 cpu->iowait_boost = 0;
1892                 }
1893                 cpu->last_update = time;
1894         }
1895
1896         delta_ns = time - cpu->sample.time;
1897         if ((s64)delta_ns >= pid_params.sample_rate_ns) {
1898                 bool sample_taken = intel_pstate_sample(cpu, time);
1899
1900                 if (sample_taken) {
1901                         intel_pstate_calc_avg_perf(cpu);
1902                         if (!hwp_active)
1903                                 intel_pstate_adjust_busy_pstate(cpu);
1904                 }
1905         }
1906 }
1907
1908 #define ICPU(model, policy) \
1909         { X86_VENDOR_INTEL, 6, model, X86_FEATURE_APERFMPERF,\
1910                         (unsigned long)&policy }
1911
1912 static const struct x86_cpu_id intel_pstate_cpu_ids[] = {
1913         ICPU(INTEL_FAM6_SANDYBRIDGE,            core_params),
1914         ICPU(INTEL_FAM6_SANDYBRIDGE_X,          core_params),
1915         ICPU(INTEL_FAM6_ATOM_SILVERMONT1,       silvermont_params),
1916         ICPU(INTEL_FAM6_IVYBRIDGE,              core_params),
1917         ICPU(INTEL_FAM6_HASWELL_CORE,           core_params),
1918         ICPU(INTEL_FAM6_BROADWELL_CORE,         core_params),
1919         ICPU(INTEL_FAM6_IVYBRIDGE_X,            core_params),
1920         ICPU(INTEL_FAM6_HASWELL_X,              core_params),
1921         ICPU(INTEL_FAM6_HASWELL_ULT,            core_params),
1922         ICPU(INTEL_FAM6_HASWELL_GT3E,           core_params),
1923         ICPU(INTEL_FAM6_BROADWELL_GT3E,         core_params),
1924         ICPU(INTEL_FAM6_ATOM_AIRMONT,           airmont_params),
1925         ICPU(INTEL_FAM6_SKYLAKE_MOBILE,         core_params),
1926         ICPU(INTEL_FAM6_BROADWELL_X,            core_params),
1927         ICPU(INTEL_FAM6_SKYLAKE_DESKTOP,        core_params),
1928         ICPU(INTEL_FAM6_BROADWELL_XEON_D,       core_params),
1929         ICPU(INTEL_FAM6_XEON_PHI_KNL,           knl_params),
1930         ICPU(INTEL_FAM6_XEON_PHI_KNM,           knl_params),
1931         ICPU(INTEL_FAM6_ATOM_GOLDMONT,          bxt_params),
1932         {}
1933 };
1934 MODULE_DEVICE_TABLE(x86cpu, intel_pstate_cpu_ids);
1935
1936 static const struct x86_cpu_id intel_pstate_cpu_oob_ids[] __initconst = {
1937         ICPU(INTEL_FAM6_BROADWELL_XEON_D, core_params),
1938         ICPU(INTEL_FAM6_BROADWELL_X, core_params),
1939         ICPU(INTEL_FAM6_SKYLAKE_X, core_params),
1940         {}
1941 };
1942
1943 static const struct x86_cpu_id intel_pstate_cpu_ee_disable_ids[] = {
1944         ICPU(INTEL_FAM6_KABYLAKE_DESKTOP, core_params),
1945         {}
1946 };
1947
1948 static int intel_pstate_init_cpu(unsigned int cpunum)
1949 {
1950         struct cpudata *cpu;
1951
1952         cpu = all_cpu_data[cpunum];
1953
1954         if (!cpu) {
1955                 cpu = kzalloc(sizeof(*cpu), GFP_KERNEL);
1956                 if (!cpu)
1957                         return -ENOMEM;
1958
1959                 all_cpu_data[cpunum] = cpu;
1960
1961                 cpu->epp_default = -EINVAL;
1962                 cpu->epp_powersave = -EINVAL;
1963                 cpu->epp_saved = -EINVAL;
1964         }
1965
1966         cpu = all_cpu_data[cpunum];
1967
1968         cpu->cpu = cpunum;
1969
1970         if (hwp_active) {
1971                 const struct x86_cpu_id *id;
1972
1973                 id = x86_match_cpu(intel_pstate_cpu_ee_disable_ids);
1974                 if (id)
1975                         intel_pstate_disable_ee(cpunum);
1976
1977                 intel_pstate_hwp_enable(cpu);
1978                 pid_params.sample_rate_ms = 50;
1979                 pid_params.sample_rate_ns = 50 * NSEC_PER_MSEC;
1980         }
1981
1982         intel_pstate_get_cpu_pstates(cpu);
1983
1984         intel_pstate_busy_pid_reset(cpu);
1985
1986         pr_debug("controlling: cpu %d\n", cpunum);
1987
1988         return 0;
1989 }
1990
1991 static unsigned int intel_pstate_get(unsigned int cpu_num)
1992 {
1993         struct cpudata *cpu = all_cpu_data[cpu_num];
1994
1995         return cpu ? get_avg_frequency(cpu) : 0;
1996 }
1997
1998 static void intel_pstate_set_update_util_hook(unsigned int cpu_num)
1999 {
2000         struct cpudata *cpu = all_cpu_data[cpu_num];
2001
2002         if (cpu->update_util_set)
2003                 return;
2004
2005         /* Prevent intel_pstate_update_util() from using stale data. */
2006         cpu->sample.time = 0;
2007         cpufreq_add_update_util_hook(cpu_num, &cpu->update_util,
2008                                      intel_pstate_update_util);
2009         cpu->update_util_set = true;
2010 }
2011
2012 static void intel_pstate_clear_update_util_hook(unsigned int cpu)
2013 {
2014         struct cpudata *cpu_data = all_cpu_data[cpu];
2015
2016         if (!cpu_data->update_util_set)
2017                 return;
2018
2019         cpufreq_remove_update_util_hook(cpu);
2020         cpu_data->update_util_set = false;
2021         synchronize_sched();
2022 }
2023
2024 static void intel_pstate_update_perf_limits(struct cpufreq_policy *policy,
2025                                             struct cpudata *cpu)
2026 {
2027         struct perf_limits *limits = &cpu->perf_limits;
2028         int32_t max_policy_perf, min_policy_perf;
2029
2030         max_policy_perf = div_ext_fp(policy->max, policy->cpuinfo.max_freq);
2031         max_policy_perf = clamp_t(int32_t, max_policy_perf, 0, int_ext_tofp(1));
2032         if (policy->max == policy->min) {
2033                 min_policy_perf = max_policy_perf;
2034         } else {
2035                 min_policy_perf = div_ext_fp(policy->min,
2036                                              policy->cpuinfo.max_freq);
2037                 min_policy_perf = clamp_t(int32_t, min_policy_perf,
2038                                           0, max_policy_perf);
2039         }
2040
2041         /* Normalize user input to [min_perf, max_perf] */
2042         if (per_cpu_limits) {
2043                 limits->min_perf = min_policy_perf;
2044                 limits->max_perf = max_policy_perf;
2045         } else {
2046                 int32_t global_min, global_max;
2047
2048                 /* Global limits are in percent of the maximum turbo P-state. */
2049                 global_max = percent_ext_fp(global.max_perf_pct);
2050                 global_min = percent_ext_fp(global.min_perf_pct);
2051                 if (policy->cpuinfo.max_freq != cpu->pstate.turbo_freq) {
2052                         int32_t turbo_factor;
2053
2054                         turbo_factor = div_ext_fp(cpu->pstate.turbo_pstate,
2055                                                   cpu->pstate.max_pstate);
2056                         global_min = mul_ext_fp(global_min, turbo_factor);
2057                         global_max = mul_ext_fp(global_max, turbo_factor);
2058                 }
2059                 global_min = clamp_t(int32_t, global_min, 0, global_max);
2060
2061                 limits->min_perf = max(min_policy_perf, global_min);
2062                 limits->min_perf = min(limits->min_perf, max_policy_perf);
2063                 limits->max_perf = min(max_policy_perf, global_max);
2064                 limits->max_perf = max(min_policy_perf, limits->max_perf);
2065
2066                 /* Make sure min_perf <= max_perf */
2067                 limits->min_perf = min(limits->min_perf, limits->max_perf);
2068         }
2069
2070         limits->max_perf = round_up(limits->max_perf, EXT_FRAC_BITS);
2071         limits->min_perf = round_up(limits->min_perf, EXT_FRAC_BITS);
2072
2073         pr_debug("cpu:%d max_perf_pct:%d min_perf_pct:%d\n", policy->cpu,
2074                  fp_ext_toint(limits->max_perf * 100),
2075                  fp_ext_toint(limits->min_perf * 100));
2076 }
2077
2078 static int intel_pstate_set_policy(struct cpufreq_policy *policy)
2079 {
2080         struct cpudata *cpu;
2081
2082         if (!policy->cpuinfo.max_freq)
2083                 return -ENODEV;
2084
2085         pr_debug("set_policy cpuinfo.max %u policy->max %u\n",
2086                  policy->cpuinfo.max_freq, policy->max);
2087
2088         cpu = all_cpu_data[policy->cpu];
2089         cpu->policy = policy->policy;
2090
2091         if (cpu->pstate.max_pstate_physical > cpu->pstate.max_pstate &&
2092             policy->max < policy->cpuinfo.max_freq &&
2093             policy->max > cpu->pstate.max_pstate * cpu->pstate.scaling) {
2094                 pr_debug("policy->max > max non turbo frequency\n");
2095                 policy->max = policy->cpuinfo.max_freq;
2096         }
2097
2098         mutex_lock(&intel_pstate_limits_lock);
2099
2100         intel_pstate_update_perf_limits(policy, cpu);
2101
2102         if (cpu->policy == CPUFREQ_POLICY_PERFORMANCE) {
2103                 /*
2104                  * NOHZ_FULL CPUs need this as the governor callback may not
2105                  * be invoked on them.
2106                  */
2107                 intel_pstate_clear_update_util_hook(policy->cpu);
2108                 intel_pstate_max_within_limits(cpu);
2109         }
2110
2111         intel_pstate_set_update_util_hook(policy->cpu);
2112
2113         if (hwp_active)
2114                 intel_pstate_hwp_set(policy);
2115
2116         mutex_unlock(&intel_pstate_limits_lock);
2117
2118         return 0;
2119 }
2120
2121 static int intel_pstate_verify_policy(struct cpufreq_policy *policy)
2122 {
2123         struct cpudata *cpu = all_cpu_data[policy->cpu];
2124
2125         update_turbo_state();
2126         policy->cpuinfo.max_freq = global.turbo_disabled || global.no_turbo ?
2127                                         cpu->pstate.max_freq :
2128                                         cpu->pstate.turbo_freq;
2129
2130         cpufreq_verify_within_cpu_limits(policy);
2131
2132         if (policy->policy != CPUFREQ_POLICY_POWERSAVE &&
2133             policy->policy != CPUFREQ_POLICY_PERFORMANCE)
2134                 return -EINVAL;
2135
2136         return 0;
2137 }
2138
2139 static void intel_cpufreq_stop_cpu(struct cpufreq_policy *policy)
2140 {
2141         intel_pstate_set_min_pstate(all_cpu_data[policy->cpu]);
2142 }
2143
2144 static void intel_pstate_stop_cpu(struct cpufreq_policy *policy)
2145 {
2146         pr_debug("CPU %d exiting\n", policy->cpu);
2147
2148         intel_pstate_clear_update_util_hook(policy->cpu);
2149         if (hwp_active)
2150                 intel_pstate_hwp_save_state(policy);
2151         else
2152                 intel_cpufreq_stop_cpu(policy);
2153 }
2154
2155 static int intel_pstate_cpu_exit(struct cpufreq_policy *policy)
2156 {
2157         intel_pstate_exit_perf_limits(policy);
2158
2159         policy->fast_switch_possible = false;
2160
2161         return 0;
2162 }
2163
2164 static int __intel_pstate_cpu_init(struct cpufreq_policy *policy)
2165 {
2166         struct cpudata *cpu;
2167         int rc;
2168
2169         rc = intel_pstate_init_cpu(policy->cpu);
2170         if (rc)
2171                 return rc;
2172
2173         cpu = all_cpu_data[policy->cpu];
2174
2175         cpu->perf_limits.max_perf = int_ext_tofp(1);
2176         cpu->perf_limits.min_perf = 0;
2177
2178         policy->min = cpu->pstate.min_pstate * cpu->pstate.scaling;
2179         policy->max = cpu->pstate.turbo_pstate * cpu->pstate.scaling;
2180
2181         /* cpuinfo and default policy values */
2182         policy->cpuinfo.min_freq = cpu->pstate.min_pstate * cpu->pstate.scaling;
2183         update_turbo_state();
2184         policy->cpuinfo.max_freq = global.turbo_disabled ?
2185                         cpu->pstate.max_pstate : cpu->pstate.turbo_pstate;
2186         policy->cpuinfo.max_freq *= cpu->pstate.scaling;
2187
2188         intel_pstate_init_acpi_perf_limits(policy);
2189         cpumask_set_cpu(policy->cpu, policy->cpus);
2190
2191         policy->fast_switch_possible = true;
2192
2193         return 0;
2194 }
2195
2196 static int intel_pstate_cpu_init(struct cpufreq_policy *policy)
2197 {
2198         int ret = __intel_pstate_cpu_init(policy);
2199
2200         if (ret)
2201                 return ret;
2202
2203         policy->cpuinfo.transition_latency = CPUFREQ_ETERNAL;
2204         if (IS_ENABLED(CONFIG_CPU_FREQ_DEFAULT_GOV_PERFORMANCE))
2205                 policy->policy = CPUFREQ_POLICY_PERFORMANCE;
2206         else
2207                 policy->policy = CPUFREQ_POLICY_POWERSAVE;
2208
2209         return 0;
2210 }
2211
2212 static struct cpufreq_driver intel_pstate = {
2213         .flags          = CPUFREQ_CONST_LOOPS,
2214         .verify         = intel_pstate_verify_policy,
2215         .setpolicy      = intel_pstate_set_policy,
2216         .suspend        = intel_pstate_hwp_save_state,
2217         .resume         = intel_pstate_resume,
2218         .get            = intel_pstate_get,
2219         .init           = intel_pstate_cpu_init,
2220         .exit           = intel_pstate_cpu_exit,
2221         .stop_cpu       = intel_pstate_stop_cpu,
2222         .name           = "intel_pstate",
2223 };
2224
2225 static int intel_cpufreq_verify_policy(struct cpufreq_policy *policy)
2226 {
2227         struct cpudata *cpu = all_cpu_data[policy->cpu];
2228
2229         update_turbo_state();
2230         policy->cpuinfo.max_freq = global.no_turbo || global.turbo_disabled ?
2231                         cpu->pstate.max_freq : cpu->pstate.turbo_freq;
2232
2233         cpufreq_verify_within_cpu_limits(policy);
2234
2235         intel_pstate_update_perf_limits(policy, cpu);
2236
2237         return 0;
2238 }
2239
2240 static int intel_cpufreq_target(struct cpufreq_policy *policy,
2241                                 unsigned int target_freq,
2242                                 unsigned int relation)
2243 {
2244         struct cpudata *cpu = all_cpu_data[policy->cpu];
2245         struct cpufreq_freqs freqs;
2246         int target_pstate;
2247
2248         update_turbo_state();
2249
2250         freqs.old = policy->cur;
2251         freqs.new = target_freq;
2252
2253         cpufreq_freq_transition_begin(policy, &freqs);
2254         switch (relation) {
2255         case CPUFREQ_RELATION_L:
2256                 target_pstate = DIV_ROUND_UP(freqs.new, cpu->pstate.scaling);
2257                 break;
2258         case CPUFREQ_RELATION_H:
2259                 target_pstate = freqs.new / cpu->pstate.scaling;
2260                 break;
2261         default:
2262                 target_pstate = DIV_ROUND_CLOSEST(freqs.new, cpu->pstate.scaling);
2263                 break;
2264         }
2265         target_pstate = intel_pstate_prepare_request(cpu, target_pstate);
2266         if (target_pstate != cpu->pstate.current_pstate) {
2267                 cpu->pstate.current_pstate = target_pstate;
2268                 wrmsrl_on_cpu(policy->cpu, MSR_IA32_PERF_CTL,
2269                               pstate_funcs.get_val(cpu, target_pstate));
2270         }
2271         freqs.new = target_pstate * cpu->pstate.scaling;
2272         cpufreq_freq_transition_end(policy, &freqs, false);
2273
2274         return 0;
2275 }
2276
2277 static unsigned int intel_cpufreq_fast_switch(struct cpufreq_policy *policy,
2278                                               unsigned int target_freq)
2279 {
2280         struct cpudata *cpu = all_cpu_data[policy->cpu];
2281         int target_pstate;
2282
2283         update_turbo_state();
2284
2285         target_pstate = DIV_ROUND_UP(target_freq, cpu->pstate.scaling);
2286         target_pstate = intel_pstate_prepare_request(cpu, target_pstate);
2287         intel_pstate_update_pstate(cpu, target_pstate);
2288         return target_pstate * cpu->pstate.scaling;
2289 }
2290
2291 static int intel_cpufreq_cpu_init(struct cpufreq_policy *policy)
2292 {
2293         int ret = __intel_pstate_cpu_init(policy);
2294
2295         if (ret)
2296                 return ret;
2297
2298         policy->cpuinfo.transition_latency = INTEL_CPUFREQ_TRANSITION_LATENCY;
2299         /* This reflects the intel_pstate_get_cpu_pstates() setting. */
2300         policy->cur = policy->cpuinfo.min_freq;
2301
2302         return 0;
2303 }
2304
2305 static struct cpufreq_driver intel_cpufreq = {
2306         .flags          = CPUFREQ_CONST_LOOPS,
2307         .verify         = intel_cpufreq_verify_policy,
2308         .target         = intel_cpufreq_target,
2309         .fast_switch    = intel_cpufreq_fast_switch,
2310         .init           = intel_cpufreq_cpu_init,
2311         .exit           = intel_pstate_cpu_exit,
2312         .stop_cpu       = intel_cpufreq_stop_cpu,
2313         .name           = "intel_cpufreq",
2314 };
2315
2316 static struct cpufreq_driver *intel_pstate_driver = &intel_pstate;
2317
2318 static void intel_pstate_driver_cleanup(void)
2319 {
2320         unsigned int cpu;
2321
2322         get_online_cpus();
2323         for_each_online_cpu(cpu) {
2324                 if (all_cpu_data[cpu]) {
2325                         if (intel_pstate_driver == &intel_pstate)
2326                                 intel_pstate_clear_update_util_hook(cpu);
2327
2328                         kfree(all_cpu_data[cpu]);
2329                         all_cpu_data[cpu] = NULL;
2330                 }
2331         }
2332         put_online_cpus();
2333 }
2334
2335 static int intel_pstate_register_driver(void)
2336 {
2337         int ret;
2338
2339         memset(&global, 0, sizeof(global));
2340         global.max_perf_pct = 100;
2341
2342         ret = cpufreq_register_driver(intel_pstate_driver);
2343         if (ret) {
2344                 intel_pstate_driver_cleanup();
2345                 return ret;
2346         }
2347
2348         global.min_perf_pct = min_perf_pct_min();
2349
2350         mutex_lock(&intel_pstate_limits_lock);
2351         driver_registered = true;
2352         mutex_unlock(&intel_pstate_limits_lock);
2353
2354         if (intel_pstate_driver == &intel_pstate && !hwp_active &&
2355             pstate_funcs.get_target_pstate != get_target_pstate_use_cpu_load)
2356                 intel_pstate_debug_expose_params();
2357
2358         return 0;
2359 }
2360
2361 static int intel_pstate_unregister_driver(void)
2362 {
2363         if (hwp_active)
2364                 return -EBUSY;
2365
2366         if (intel_pstate_driver == &intel_pstate && !hwp_active &&
2367             pstate_funcs.get_target_pstate != get_target_pstate_use_cpu_load)
2368                 intel_pstate_debug_hide_params();
2369
2370         mutex_lock(&intel_pstate_limits_lock);
2371         driver_registered = false;
2372         mutex_unlock(&intel_pstate_limits_lock);
2373
2374         cpufreq_unregister_driver(intel_pstate_driver);
2375         intel_pstate_driver_cleanup();
2376
2377         return 0;
2378 }
2379
2380 static ssize_t intel_pstate_show_status(char *buf)
2381 {
2382         if (!driver_registered)
2383                 return sprintf(buf, "off\n");
2384
2385         return sprintf(buf, "%s\n", intel_pstate_driver == &intel_pstate ?
2386                                         "active" : "passive");
2387 }
2388
2389 static int intel_pstate_update_status(const char *buf, size_t size)
2390 {
2391         int ret;
2392
2393         if (size == 3 && !strncmp(buf, "off", size))
2394                 return driver_registered ?
2395                         intel_pstate_unregister_driver() : -EINVAL;
2396
2397         if (size == 6 && !strncmp(buf, "active", size)) {
2398                 if (driver_registered) {
2399                         if (intel_pstate_driver == &intel_pstate)
2400                                 return 0;
2401
2402                         ret = intel_pstate_unregister_driver();
2403                         if (ret)
2404                                 return ret;
2405                 }
2406
2407                 intel_pstate_driver = &intel_pstate;
2408                 return intel_pstate_register_driver();
2409         }
2410
2411         if (size == 7 && !strncmp(buf, "passive", size)) {
2412                 if (driver_registered) {
2413                         if (intel_pstate_driver != &intel_pstate)
2414                                 return 0;
2415
2416                         ret = intel_pstate_unregister_driver();
2417                         if (ret)
2418                                 return ret;
2419                 }
2420
2421                 intel_pstate_driver = &intel_cpufreq;
2422                 return intel_pstate_register_driver();
2423         }
2424
2425         return -EINVAL;
2426 }
2427
2428 static int no_load __initdata;
2429 static int no_hwp __initdata;
2430 static int hwp_only __initdata;
2431 static unsigned int force_load __initdata;
2432
2433 static int __init intel_pstate_msrs_not_valid(void)
2434 {
2435         if (!pstate_funcs.get_max() ||
2436             !pstate_funcs.get_min() ||
2437             !pstate_funcs.get_turbo())
2438                 return -ENODEV;
2439
2440         return 0;
2441 }
2442
2443 static void __init copy_pid_params(struct pstate_adjust_policy *policy)
2444 {
2445         pid_params.sample_rate_ms = policy->sample_rate_ms;
2446         pid_params.sample_rate_ns = pid_params.sample_rate_ms * NSEC_PER_MSEC;
2447         pid_params.p_gain_pct = policy->p_gain_pct;
2448         pid_params.i_gain_pct = policy->i_gain_pct;
2449         pid_params.d_gain_pct = policy->d_gain_pct;
2450         pid_params.deadband = policy->deadband;
2451         pid_params.setpoint = policy->setpoint;
2452 }
2453
2454 #ifdef CONFIG_ACPI
2455 static void intel_pstate_use_acpi_profile(void)
2456 {
2457         switch (acpi_gbl_FADT.preferred_profile) {
2458         case PM_MOBILE:
2459         case PM_TABLET:
2460         case PM_APPLIANCE_PC:
2461         case PM_DESKTOP:
2462         case PM_WORKSTATION:
2463                 pstate_funcs.get_target_pstate =
2464                                 get_target_pstate_use_cpu_load;
2465         }
2466 }
2467 #else
2468 static void intel_pstate_use_acpi_profile(void)
2469 {
2470 }
2471 #endif
2472
2473 static void __init copy_cpu_funcs(struct pstate_funcs *funcs)
2474 {
2475         pstate_funcs.get_max   = funcs->get_max;
2476         pstate_funcs.get_max_physical = funcs->get_max_physical;
2477         pstate_funcs.get_min   = funcs->get_min;
2478         pstate_funcs.get_turbo = funcs->get_turbo;
2479         pstate_funcs.get_scaling = funcs->get_scaling;
2480         pstate_funcs.get_val   = funcs->get_val;
2481         pstate_funcs.get_vid   = funcs->get_vid;
2482         pstate_funcs.get_target_pstate = funcs->get_target_pstate;
2483
2484         intel_pstate_use_acpi_profile();
2485 }
2486
2487 #ifdef CONFIG_ACPI
2488
2489 static bool __init intel_pstate_no_acpi_pss(void)
2490 {
2491         int i;
2492
2493         for_each_possible_cpu(i) {
2494                 acpi_status status;
2495                 union acpi_object *pss;
2496                 struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
2497                 struct acpi_processor *pr = per_cpu(processors, i);
2498
2499                 if (!pr)
2500                         continue;
2501
2502                 status = acpi_evaluate_object(pr->handle, "_PSS", NULL, &buffer);
2503                 if (ACPI_FAILURE(status))
2504                         continue;
2505
2506                 pss = buffer.pointer;
2507                 if (pss && pss->type == ACPI_TYPE_PACKAGE) {
2508                         kfree(pss);
2509                         return false;
2510                 }
2511
2512                 kfree(pss);
2513         }
2514
2515         return true;
2516 }
2517
2518 static bool __init intel_pstate_has_acpi_ppc(void)
2519 {
2520         int i;
2521
2522         for_each_possible_cpu(i) {
2523                 struct acpi_processor *pr = per_cpu(processors, i);
2524
2525                 if (!pr)
2526                         continue;
2527                 if (acpi_has_method(pr->handle, "_PPC"))
2528                         return true;
2529         }
2530         return false;
2531 }
2532
2533 enum {
2534         PSS,
2535         PPC,
2536 };
2537
2538 struct hw_vendor_info {
2539         u16  valid;
2540         char oem_id[ACPI_OEM_ID_SIZE];
2541         char oem_table_id[ACPI_OEM_TABLE_ID_SIZE];
2542         int  oem_pwr_table;
2543 };
2544
2545 /* Hardware vendor-specific info that has its own power management modes */
2546 static struct hw_vendor_info vendor_info[] __initdata = {
2547         {1, "HP    ", "ProLiant", PSS},
2548         {1, "ORACLE", "X4-2    ", PPC},
2549         {1, "ORACLE", "X4-2L   ", PPC},
2550         {1, "ORACLE", "X4-2B   ", PPC},
2551         {1, "ORACLE", "X3-2    ", PPC},
2552         {1, "ORACLE", "X3-2L   ", PPC},
2553         {1, "ORACLE", "X3-2B   ", PPC},
2554         {1, "ORACLE", "X4470M2 ", PPC},
2555         {1, "ORACLE", "X4270M3 ", PPC},
2556         {1, "ORACLE", "X4270M2 ", PPC},
2557         {1, "ORACLE", "X4170M2 ", PPC},
2558         {1, "ORACLE", "X4170 M3", PPC},
2559         {1, "ORACLE", "X4275 M3", PPC},
2560         {1, "ORACLE", "X6-2    ", PPC},
2561         {1, "ORACLE", "Sudbury ", PPC},
2562         {0, "", ""},
2563 };
2564
2565 static bool __init intel_pstate_platform_pwr_mgmt_exists(void)
2566 {
2567         struct acpi_table_header hdr;
2568         struct hw_vendor_info *v_info;
2569         const struct x86_cpu_id *id;
2570         u64 misc_pwr;
2571
2572         id = x86_match_cpu(intel_pstate_cpu_oob_ids);
2573         if (id) {
2574                 rdmsrl(MSR_MISC_PWR_MGMT, misc_pwr);
2575                 if ( misc_pwr & (1 << 8))
2576                         return true;
2577         }
2578
2579         if (acpi_disabled ||
2580             ACPI_FAILURE(acpi_get_table_header(ACPI_SIG_FADT, 0, &hdr)))
2581                 return false;
2582
2583         for (v_info = vendor_info; v_info->valid; v_info++) {
2584                 if (!strncmp(hdr.oem_id, v_info->oem_id, ACPI_OEM_ID_SIZE) &&
2585                         !strncmp(hdr.oem_table_id, v_info->oem_table_id,
2586                                                 ACPI_OEM_TABLE_ID_SIZE))
2587                         switch (v_info->oem_pwr_table) {
2588                         case PSS:
2589                                 return intel_pstate_no_acpi_pss();
2590                         case PPC:
2591                                 return intel_pstate_has_acpi_ppc() &&
2592                                         (!force_load);
2593                         }
2594         }
2595
2596         return false;
2597 }
2598
2599 static void intel_pstate_request_control_from_smm(void)
2600 {
2601         /*
2602          * It may be unsafe to request P-states control from SMM if _PPC support
2603          * has not been enabled.
2604          */
2605         if (acpi_ppc)
2606                 acpi_processor_pstate_control();
2607 }
2608 #else /* CONFIG_ACPI not enabled */
2609 static inline bool intel_pstate_platform_pwr_mgmt_exists(void) { return false; }
2610 static inline bool intel_pstate_has_acpi_ppc(void) { return false; }
2611 static inline void intel_pstate_request_control_from_smm(void) {}
2612 #endif /* CONFIG_ACPI */
2613
2614 static const struct x86_cpu_id hwp_support_ids[] __initconst = {
2615         { X86_VENDOR_INTEL, 6, X86_MODEL_ANY, X86_FEATURE_HWP },
2616         {}
2617 };
2618
2619 static int __init intel_pstate_init(void)
2620 {
2621         int rc;
2622
2623         if (no_load)
2624                 return -ENODEV;
2625
2626         if (x86_match_cpu(hwp_support_ids)) {
2627                 copy_cpu_funcs(&core_params.funcs);
2628                 if (no_hwp) {
2629                         pstate_funcs.get_target_pstate = get_target_pstate_use_cpu_load;
2630                 } else {
2631                         hwp_active++;
2632                         intel_pstate.attr = hwp_cpufreq_attrs;
2633                         goto hwp_cpu_matched;
2634                 }
2635         } else {
2636                 const struct x86_cpu_id *id;
2637                 struct cpu_defaults *cpu_def;
2638
2639                 id = x86_match_cpu(intel_pstate_cpu_ids);
2640                 if (!id)
2641                         return -ENODEV;
2642
2643                 cpu_def = (struct cpu_defaults *)id->driver_data;
2644
2645                 copy_pid_params(&cpu_def->pid_policy);
2646                 copy_cpu_funcs(&cpu_def->funcs);
2647         }
2648
2649         if (intel_pstate_msrs_not_valid())
2650                 return -ENODEV;
2651
2652 hwp_cpu_matched:
2653         /*
2654          * The Intel pstate driver will be ignored if the platform
2655          * firmware has its own power management modes.
2656          */
2657         if (intel_pstate_platform_pwr_mgmt_exists())
2658                 return -ENODEV;
2659
2660         if (!hwp_active && hwp_only)
2661                 return -ENOTSUPP;
2662
2663         pr_info("Intel P-state driver initializing\n");
2664
2665         all_cpu_data = vzalloc(sizeof(void *) * num_possible_cpus());
2666         if (!all_cpu_data)
2667                 return -ENOMEM;
2668
2669         intel_pstate_request_control_from_smm();
2670
2671         intel_pstate_sysfs_expose_params();
2672
2673         mutex_lock(&intel_pstate_driver_lock);
2674         rc = intel_pstate_register_driver();
2675         mutex_unlock(&intel_pstate_driver_lock);
2676         if (rc)
2677                 return rc;
2678
2679         if (hwp_active)
2680                 pr_info("HWP enabled\n");
2681
2682         return 0;
2683 }
2684 device_initcall(intel_pstate_init);
2685
2686 static int __init intel_pstate_setup(char *str)
2687 {
2688         if (!str)
2689                 return -EINVAL;
2690
2691         if (!strcmp(str, "disable")) {
2692                 no_load = 1;
2693         } else if (!strcmp(str, "passive")) {
2694                 pr_info("Passive mode enabled\n");
2695                 intel_pstate_driver = &intel_cpufreq;
2696                 no_hwp = 1;
2697         }
2698         if (!strcmp(str, "no_hwp")) {
2699                 pr_info("HWP disabled\n");
2700                 no_hwp = 1;
2701         }
2702         if (!strcmp(str, "force"))
2703                 force_load = 1;
2704         if (!strcmp(str, "hwp_only"))
2705                 hwp_only = 1;
2706         if (!strcmp(str, "per_cpu_perf_limits"))
2707                 per_cpu_limits = true;
2708
2709 #ifdef CONFIG_ACPI
2710         if (!strcmp(str, "support_acpi_ppc"))
2711                 acpi_ppc = true;
2712 #endif
2713
2714         return 0;
2715 }
2716 early_param("intel_pstate", intel_pstate_setup);
2717
2718 MODULE_AUTHOR("Dirk Brandewie <dirk.j.brandewie@intel.com>");
2719 MODULE_DESCRIPTION("'intel_pstate' - P state driver Intel Core processors");
2720 MODULE_LICENSE("GPL");