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  * @no_turbo:           User requested turbo state from intel_pstate sysfs
 191  * @turbo_disabled:     Platform turbo status either from msr
 192  *                      MSR_IA32_MISC_ENABLE or when maximum available pstate
 193  *                      matches the maximum turbo pstate
 194  * @max_perf_pct:       Effective maximum performance limit in percentage, this
 195  *                      is minimum of either limits enforced by cpufreq policy
 196  *                      or limits from user set limits via intel_pstate sysfs
 197  * @min_perf_pct:       Effective minimum performance limit in percentage, this
 198  *                      is maximum of either limits enforced by cpufreq policy
 199  *                      or limits from user set limits via intel_pstate sysfs
 200  * @max_perf:           This is a scaled value between 0 to 255 for max_perf_pct
 201  *                      This value is used to limit max pstate
 202  * @min_perf:           This is a scaled value between 0 to 255 for min_perf_pct
 203  *                      This value is used to limit min pstate
 204  * @max_policy_pct:     The maximum performance in percentage enforced by
 205  *                      cpufreq setpolicy interface
 206  * @max_sysfs_pct:      The maximum performance in percentage enforced by
 207  *                      intel pstate sysfs interface, unused when per cpu
 208  *                      controls are enforced
 209  * @min_policy_pct:     The minimum performance in percentage enforced by
 210  *                      cpufreq setpolicy interface
 211  * @min_sysfs_pct:      The minimum performance in percentage enforced by
 212  *                      intel pstate sysfs interface, unused when per cpu
 213  *                      controls are enforced
 214  *
 215  * Storage for user and policy defined limits.
 216  */
 217 struct perf_limits {
 218         int no_turbo;
 219         int turbo_disabled;
 220         int max_perf_pct;
 221         int min_perf_pct;
 222         int32_t max_perf;
 223         int32_t min_perf;
 224         int max_policy_pct;
 225         int max_sysfs_pct;
 226         int min_policy_pct;
 227         int min_sysfs_pct;
 228 };
 229
 230 /**
 231  * struct cpudata -     Per CPU instance data storage
 232  * @cpu:                CPU number for this instance data
 233  * @policy:             CPUFreq policy value
 234  * @update_util:        CPUFreq utility callback information
 235  * @update_util_set:    CPUFreq utility callback is set
 236  * @iowait_boost:       iowait-related boost fraction
 237  * @last_update:        Time of the last update.
 238  * @pstate:             Stores P state limits for this CPU
 239  * @vid:                Stores VID limits for this CPU
 240  * @pid:                Stores PID parameters for this CPU
 241  * @last_sample_time:   Last Sample time
 242  * @prev_aperf:         Last APERF value read from APERF MSR
 243  * @prev_mperf:         Last MPERF value read from MPERF MSR
 244  * @prev_tsc:           Last timestamp counter (TSC) value
 245  * @prev_cummulative_iowait: IO Wait time difference from last and
 246  *                      current sample
 247  * @sample:             Storage for storing last Sample data
 248  * @perf_limits:        Pointer to perf_limit unique to this CPU
 249  *                      Not all field in the structure are applicable
 250  *                      when per cpu controls are enforced
 251  * @acpi_perf_data:     Stores ACPI perf information read from _PSS
 252  * @valid_pss_table:    Set to true for valid ACPI _PSS entries found
 253  * @epp_powersave:      Last saved HWP energy performance preference
 254  *                      (EPP) or energy performance bias (EPB),
 255  *                      when policy switched to performance
 256  * @epp_policy:         Last saved policy used to set EPP/EPB
 257  * @epp_default:        Power on default HWP energy performance
 258  *                      preference/bias
 259  * @epp_saved:          Saved EPP/EPB during system suspend or CPU offline
 260  *                      operation
 261  *
 262  * This structure stores per CPU instance data for all CPUs.
 263  */
 264 struct cpudata {
 265         int cpu;
 266
 267         unsigned int policy;
 268         struct update_util_data update_util;
 269         bool   update_util_set;
 270
 271         struct pstate_data pstate;
 272         struct vid_data vid;
 273         struct _pid pid;
 274
 275         u64     last_update;
 276         u64     last_sample_time;
 277         u64     prev_aperf;
 278         u64     prev_mperf;
 279         u64     prev_tsc;
 280         u64     prev_cummulative_iowait;
 281         struct sample sample;
 282         struct perf_limits *perf_limits;
 283 #ifdef CONFIG_ACPI
 284         struct acpi_processor_performance acpi_perf_data;
 285         bool valid_pss_table;
 286 #endif
 287         unsigned int iowait_boost;
 288         s16 epp_powersave;
 289         s16 epp_policy;
 290         s16 epp_default;
 291         s16 epp_saved;
 292 };
 293
 294 static struct cpudata **all_cpu_data;
 295
 296 /**
 297  * struct pstate_adjust_policy - Stores static PID configuration data
 298  * @sample_rate_ms:     PID calculation sample rate in ms
 299  * @sample_rate_ns:     Sample rate calculation in ns
 300  * @deadband:           PID deadband
 301  * @setpoint:           PID Setpoint
 302  * @p_gain_pct:         PID proportional gain
 303  * @i_gain_pct:         PID integral gain
 304  * @d_gain_pct:         PID derivative gain
 305  *
 306  * Stores per CPU model static PID configuration data.
 307  */
 308 struct pstate_adjust_policy {
 309         int sample_rate_ms;
 310         s64 sample_rate_ns;
 311         int deadband;
 312         int setpoint;
 313         int p_gain_pct;
 314         int d_gain_pct;
 315         int i_gain_pct;
 316 };
 317
 318 /**
 319  * struct pstate_funcs - Per CPU model specific callbacks
 320  * @get_max:            Callback to get maximum non turbo effective P state
 321  * @get_max_physical:   Callback to get maximum non turbo physical P state
 322  * @get_min:            Callback to get minimum P state
 323  * @get_turbo:          Callback to get turbo P state
 324  * @get_scaling:        Callback to get frequency scaling factor
 325  * @get_val:            Callback to convert P state to actual MSR write value
 326  * @get_vid:            Callback to get VID data for Atom platforms
 327  * @get_target_pstate:  Callback to a function to calculate next P state to use
 328  *
 329  * Core and Atom CPU models have different way to get P State limits. This
 330  * structure is used to store those callbacks.
 331  */
 332 struct pstate_funcs {
 333         int (*get_max)(void);
 334         int (*get_max_physical)(void);
 335         int (*get_min)(void);
 336         int (*get_turbo)(void);
 337         int (*get_scaling)(void);
 338         u64 (*get_val)(struct cpudata*, int pstate);
 339         void (*get_vid)(struct cpudata *);
 340         int32_t (*get_target_pstate)(struct cpudata *);
 341 };
 342
 343 /**
 344  * struct cpu_defaults- Per CPU model default config data
 345  * @pid_policy: PID config data
 346  * @funcs:              Callback function data
 347  */
 348 struct cpu_defaults {
 349         struct pstate_adjust_policy pid_policy;
 350         struct pstate_funcs funcs;
 351 };
 352
 353 static inline int32_t get_target_pstate_use_performance(struct cpudata *cpu);
 354 static inline int32_t get_target_pstate_use_cpu_load(struct cpudata *cpu);
 355
 356 static struct pstate_adjust_policy pid_params __read_mostly;
 357 static struct pstate_funcs pstate_funcs __read_mostly;
 358 static int hwp_active __read_mostly;
 359 static bool per_cpu_limits __read_mostly;
 360
 361 static bool driver_registered __read_mostly;
 362
 363 #ifdef CONFIG_ACPI
 364 static bool acpi_ppc;
 365 #endif
 366
 367 static struct perf_limits global;
 368
 369 static void intel_pstate_init_limits(struct perf_limits *limits)
 370 {
 371         memset(limits, 0, sizeof(*limits));
 372         limits->max_perf_pct = 100;
 373         limits->max_perf = int_ext_tofp(1);
 374         limits->max_policy_pct = 100;
 375         limits->max_sysfs_pct = 100;
 376 }
 377
 378 static DEFINE_MUTEX(intel_pstate_driver_lock);
 379 static DEFINE_MUTEX(intel_pstate_limits_lock);
 380
 381 #ifdef CONFIG_ACPI
 382
 383 static bool intel_pstate_get_ppc_enable_status(void)
 384 {
 385         if (acpi_gbl_FADT.preferred_profile == PM_ENTERPRISE_SERVER ||
 386             acpi_gbl_FADT.preferred_profile == PM_PERFORMANCE_SERVER)
 387                 return true;
 388
 389         return acpi_ppc;
 390 }
 391
 392 #ifdef CONFIG_ACPI_CPPC_LIB
 393
 394 /* The work item is needed to avoid CPU hotplug locking issues */
 395 static void intel_pstste_sched_itmt_work_fn(struct work_struct *work)
 396 {
 397         sched_set_itmt_support();
 398 }
 399
 400 static DECLARE_WORK(sched_itmt_work, intel_pstste_sched_itmt_work_fn);
 401
 402 static void intel_pstate_set_itmt_prio(int cpu)
 403 {
 404         struct cppc_perf_caps cppc_perf;
 405         static u32 max_highest_perf = 0, min_highest_perf = U32_MAX;
 406         int ret;
 407
 408         ret = cppc_get_perf_caps(cpu, &cppc_perf);
 409         if (ret)
 410                 return;
 411
 412         /*
 413          * The priorities can be set regardless of whether or not
 414          * sched_set_itmt_support(true) has been called and it is valid to
 415          * update them at any time after it has been called.
 416          */
 417         sched_set_itmt_core_prio(cppc_perf.highest_perf, cpu);
 418
 419         if (max_highest_perf <= min_highest_perf) {
 420                 if (cppc_perf.highest_perf > max_highest_perf)
 421                         max_highest_perf = cppc_perf.highest_perf;
 422
 423                 if (cppc_perf.highest_perf < min_highest_perf)
 424                         min_highest_perf = cppc_perf.highest_perf;
 425
 426                 if (max_highest_perf > min_highest_perf) {
 427                         /*
 428                          * This code can be run during CPU online under the
 429                          * CPU hotplug locks, so sched_set_itmt_support()
 430                          * cannot be called from here.  Queue up a work item
 431                          * to invoke it.
 432                          */
 433                         schedule_work(&sched_itmt_work);
 434                 }
 435         }
 436 }
 437 #else
 438 static void intel_pstate_set_itmt_prio(int cpu)
 439 {
 440 }
 441 #endif
 442
 443 static void intel_pstate_init_acpi_perf_limits(struct cpufreq_policy *policy)
 444 {
 445         struct cpudata *cpu;
 446         int ret;
 447         int i;
 448
 449         if (hwp_active) {
 450                 intel_pstate_set_itmt_prio(policy->cpu);
 451                 return;
 452         }
 453
 454         if (!intel_pstate_get_ppc_enable_status())
 455                 return;
 456
 457         cpu = all_cpu_data[policy->cpu];
 458
 459         ret = acpi_processor_register_performance(&cpu->acpi_perf_data,
 460                                                   policy->cpu);
 461         if (ret)
 462                 return;
 463
 464         /*
 465          * Check if the control value in _PSS is for PERF_CTL MSR, which should
 466          * guarantee that the states returned by it map to the states in our
 467          * list directly.
 468          */
 469         if (cpu->acpi_perf_data.control_register.space_id !=
 470                                                 ACPI_ADR_SPACE_FIXED_HARDWARE)
 471                 goto err;
 472
 473         /*
 474          * If there is only one entry _PSS, simply ignore _PSS and continue as
 475          * usual without taking _PSS into account
 476          */
 477         if (cpu->acpi_perf_data.state_count < 2)
 478                 goto err;
 479
 480         pr_debug("CPU%u - ACPI _PSS perf data\n", policy->cpu);
 481         for (i = 0; i < cpu->acpi_perf_data.state_count; i++) {
 482                 pr_debug("     %cP%d: %u MHz, %u mW, 0x%x\n",
 483                          (i == cpu->acpi_perf_data.state ? '*' : ' '), i,
 484                          (u32) cpu->acpi_perf_data.states[i].core_frequency,
 485                          (u32) cpu->acpi_perf_data.states[i].power,
 486                          (u32) cpu->acpi_perf_data.states[i].control);
 487         }
 488
 489         /*
 490          * The _PSS table doesn't contain whole turbo frequency range.
 491          * This just contains +1 MHZ above the max non turbo frequency,
 492          * with control value corresponding to max turbo ratio. But
 493          * when cpufreq set policy is called, it will call with this
 494          * max frequency, which will cause a reduced performance as
 495          * this driver uses real max turbo frequency as the max
 496          * frequency. So correct this frequency in _PSS table to
 497          * correct max turbo frequency based on the turbo state.
 498          * Also need to convert to MHz as _PSS freq is in MHz.
 499          */
 500         if (!global.turbo_disabled)
 501                 cpu->acpi_perf_data.states[0].core_frequency =
 502                                         policy->cpuinfo.max_freq / 1000;
 503         cpu->valid_pss_table = true;
 504         pr_debug("_PPC limits will be enforced\n");
 505
 506         return;
 507
 508  err:
 509         cpu->valid_pss_table = false;
 510         acpi_processor_unregister_performance(policy->cpu);
 511 }
 512
 513 static void intel_pstate_exit_perf_limits(struct cpufreq_policy *policy)
 514 {
 515         struct cpudata *cpu;
 516
 517         cpu = all_cpu_data[policy->cpu];
 518         if (!cpu->valid_pss_table)
 519                 return;
 520
 521         acpi_processor_unregister_performance(policy->cpu);
 522 }
 523 #else
 524 static inline void intel_pstate_init_acpi_perf_limits(struct cpufreq_policy *policy)
 525 {
 526 }
 527
 528 static inline void intel_pstate_exit_perf_limits(struct cpufreq_policy *policy)
 529 {
 530 }
 531 #endif
 532
 533 static inline void pid_reset(struct _pid *pid, int setpoint, int busy,
 534                              int deadband, int integral) {
 535         pid->setpoint = int_tofp(setpoint);
 536         pid->deadband  = int_tofp(deadband);
 537         pid->integral  = int_tofp(integral);
 538         pid->last_err  = int_tofp(setpoint) - int_tofp(busy);
 539 }
 540
 541 static inline void pid_p_gain_set(struct _pid *pid, int percent)
 542 {
 543         pid->p_gain = div_fp(percent, 100);
 544 }
 545
 546 static inline void pid_i_gain_set(struct _pid *pid, int percent)
 547 {
 548         pid->i_gain = div_fp(percent, 100);
 549 }
 550
 551 static inline void pid_d_gain_set(struct _pid *pid, int percent)
 552 {
 553         pid->d_gain = div_fp(percent, 100);
 554 }
 555
 556 static signed int pid_calc(struct _pid *pid, int32_t busy)
 557 {
 558         signed int result;
 559         int32_t pterm, dterm, fp_error;
 560         int32_t integral_limit;
 561
 562         fp_error = pid->setpoint - busy;
 563
 564         if (abs(fp_error) <= pid->deadband)
 565                 return 0;
 566
 567         pterm = mul_fp(pid->p_gain, fp_error);
 568
 569         pid->integral += fp_error;
 570
 571         /*
 572          * We limit the integral here so that it will never
 573          * get higher than 30.  This prevents it from becoming
 574          * too large an input over long periods of time and allows
 575          * it to get factored out sooner.
 576          *
 577          * The value of 30 was chosen through experimentation.
 578          */
 579         integral_limit = int_tofp(30);
 580         if (pid->integral > integral_limit)
 581                 pid->integral = integral_limit;
 582         if (pid->integral < -integral_limit)
 583                 pid->integral = -integral_limit;
 584
 585         dterm = mul_fp(pid->d_gain, fp_error - pid->last_err);
 586         pid->last_err = fp_error;
 587
 588         result = pterm + mul_fp(pid->integral, pid->i_gain) + dterm;
 589         result = result + (1 << (FRAC_BITS-1));
 590         return (signed int)fp_toint(result);
 591 }
 592
 593 static inline void intel_pstate_busy_pid_reset(struct cpudata *cpu)
 594 {
 595         pid_p_gain_set(&cpu->pid, pid_params.p_gain_pct);
 596         pid_d_gain_set(&cpu->pid, pid_params.d_gain_pct);
 597         pid_i_gain_set(&cpu->pid, pid_params.i_gain_pct);
 598
 599         pid_reset(&cpu->pid, pid_params.setpoint, 100, pid_params.deadband, 0);
 600 }
 601
 602 static inline void intel_pstate_reset_all_pid(void)
 603 {
 604         unsigned int cpu;
 605
 606         for_each_online_cpu(cpu) {
 607                 if (all_cpu_data[cpu])
 608                         intel_pstate_busy_pid_reset(all_cpu_data[cpu]);
 609         }
 610 }
 611
 612 static inline void update_turbo_state(void)
 613 {
 614         u64 misc_en;
 615         struct cpudata *cpu;
 616
 617         cpu = all_cpu_data[0];
 618         rdmsrl(MSR_IA32_MISC_ENABLE, misc_en);
 619         global.turbo_disabled =
 620                 (misc_en & MSR_IA32_MISC_ENABLE_TURBO_DISABLE ||
 621                  cpu->pstate.max_pstate == cpu->pstate.turbo_pstate);
 622 }
 623
 624 static s16 intel_pstate_get_epb(struct cpudata *cpu_data)
 625 {
 626         u64 epb;
 627         int ret;
 628
 629         if (!static_cpu_has(X86_FEATURE_EPB))
 630                 return -ENXIO;
 631
 632         ret = rdmsrl_on_cpu(cpu_data->cpu, MSR_IA32_ENERGY_PERF_BIAS, &epb);
 633         if (ret)
 634                 return (s16)ret;
 635
 636         return (s16)(epb & 0x0f);
 637 }
 638
 639 static s16 intel_pstate_get_epp(struct cpudata *cpu_data, u64 hwp_req_data)
 640 {
 641         s16 epp;
 642
 643         if (static_cpu_has(X86_FEATURE_HWP_EPP)) {
 644                 /*
 645                  * When hwp_req_data is 0, means that caller didn't read
 646                  * MSR_HWP_REQUEST, so need to read and get EPP.
 647                  */
 648                 if (!hwp_req_data) {
 649                         epp = rdmsrl_on_cpu(cpu_data->cpu, MSR_HWP_REQUEST,
 650                                             &hwp_req_data);
 651                         if (epp)
 652                                 return epp;
 653                 }
 654                 epp = (hwp_req_data >> 24) & 0xff;
 655         } else {
 656                 /* When there is no EPP present, HWP uses EPB settings */
 657                 epp = intel_pstate_get_epb(cpu_data);
 658         }
 659
 660         return epp;
 661 }
 662
 663 static int intel_pstate_set_epb(int cpu, s16 pref)
 664 {
 665         u64 epb;
 666         int ret;
 667
 668         if (!static_cpu_has(X86_FEATURE_EPB))
 669                 return -ENXIO;
 670
 671         ret = rdmsrl_on_cpu(cpu, MSR_IA32_ENERGY_PERF_BIAS, &epb);
 672         if (ret)
 673                 return ret;
 674
 675         epb = (epb & ~0x0f) | pref;
 676         wrmsrl_on_cpu(cpu, MSR_IA32_ENERGY_PERF_BIAS, epb);
 677
 678         return 0;
 679 }
 680
 681 /*
 682  * EPP/EPB display strings corresponding to EPP index in the
 683  * energy_perf_strings[]
 684  *      index           String
 685  *-------------------------------------
 686  *      0               default
 687  *      1               performance
 688  *      2               balance_performance
 689  *      3               balance_power
 690  *      4               power
 691  */
 692 static const char * const energy_perf_strings[] = {
 693         "default",
 694         "performance",
 695         "balance_performance",
 696         "balance_power",
 697         "power",
 698         NULL
 699 };
 700
 701 static int intel_pstate_get_energy_pref_index(struct cpudata *cpu_data)
 702 {
 703         s16 epp;
 704         int index = -EINVAL;
 705
 706         epp = intel_pstate_get_epp(cpu_data, 0);
 707         if (epp < 0)
 708                 return epp;
 709
 710         if (static_cpu_has(X86_FEATURE_HWP_EPP)) {
 711                 /*
 712                  * Range:
 713                  *      0x00-0x3F       :       Performance
 714                  *      0x40-0x7F       :       Balance performance
 715                  *      0x80-0xBF       :       Balance power
 716                  *      0xC0-0xFF       :       Power
 717                  * The EPP is a 8 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 >> 6) + 1;
 722         } else if (static_cpu_has(X86_FEATURE_EPB)) {
 723                 /*
 724                  * Range:
 725                  *      0x00-0x03       :       Performance
 726                  *      0x04-0x07       :       Balance performance
 727                  *      0x08-0x0B       :       Balance power
 728                  *      0x0C-0x0F       :       Power
 729                  * The EPB is a 4 bit value, but our ranges restrict the
 730                  * value which can be set. Here only using top two bits
 731                  * effectively.
 732                  */
 733                 index = (epp >> 2) + 1;
 734         }
 735
 736         return index;
 737 }
 738
 739 static int intel_pstate_set_energy_pref_index(struct cpudata *cpu_data,
 740                                               int pref_index)
 741 {
 742         int epp = -EINVAL;
 743         int ret;
 744
 745         if (!pref_index)
 746                 epp = cpu_data->epp_default;
 747
 748         mutex_lock(&intel_pstate_limits_lock);
 749
 750         if (static_cpu_has(X86_FEATURE_HWP_EPP)) {
 751                 u64 value;
 752
 753                 ret = rdmsrl_on_cpu(cpu_data->cpu, MSR_HWP_REQUEST, &value);
 754                 if (ret)
 755                         goto return_pref;
 756
 757                 value &= ~GENMASK_ULL(31, 24);
 758
 759                 /*
 760                  * If epp is not default, convert from index into
 761                  * energy_perf_strings to epp value, by shifting 6
 762                  * bits left to use only top two bits in epp.
 763                  * The resultant epp need to shifted by 24 bits to
 764                  * epp position in MSR_HWP_REQUEST.
 765                  */
 766                 if (epp == -EINVAL)
 767                         epp = (pref_index - 1) << 6;
 768
 769                 value |= (u64)epp << 24;
 770                 ret = wrmsrl_on_cpu(cpu_data->cpu, MSR_HWP_REQUEST, value);
 771         } else {
 772                 if (epp == -EINVAL)
 773                         epp = (pref_index - 1) << 2;
 774                 ret = intel_pstate_set_epb(cpu_data->cpu, epp);
 775         }
 776 return_pref:
 777         mutex_unlock(&intel_pstate_limits_lock);
 778
 779         return ret;
 780 }
 781
 782 static ssize_t show_energy_performance_available_preferences(
 783                                 struct cpufreq_policy *policy, char *buf)
 784 {
 785         int i = 0;
 786         int ret = 0;
 787
 788         while (energy_perf_strings[i] != NULL)
 789                 ret += sprintf(&buf[ret], "%s ", energy_perf_strings[i++]);
 790
 791         ret += sprintf(&buf[ret], "\n");
 792
 793         return ret;
 794 }
 795
 796 cpufreq_freq_attr_ro(energy_performance_available_preferences);
 797
 798 static ssize_t store_energy_performance_preference(
 799                 struct cpufreq_policy *policy, const char *buf, size_t count)
 800 {
 801         struct cpudata *cpu_data = all_cpu_data[policy->cpu];
 802         char str_preference[21];
 803         int ret, i = 0;
 804
 805         ret = sscanf(buf, "%20s", str_preference);
 806         if (ret != 1)
 807                 return -EINVAL;
 808
 809         while (energy_perf_strings[i] != NULL) {
 810                 if (!strcmp(str_preference, energy_perf_strings[i])) {
 811                         intel_pstate_set_energy_pref_index(cpu_data, i);
 812                         return count;
 813                 }
 814                 ++i;
 815         }
 816
 817         return -EINVAL;
 818 }
 819
 820 static ssize_t show_energy_performance_preference(
 821                                 struct cpufreq_policy *policy, char *buf)
 822 {
 823         struct cpudata *cpu_data = all_cpu_data[policy->cpu];
 824         int preference;
 825
 826         preference = intel_pstate_get_energy_pref_index(cpu_data);
 827         if (preference < 0)
 828                 return preference;
 829
 830         return  sprintf(buf, "%s\n", energy_perf_strings[preference]);
 831 }
 832
 833 cpufreq_freq_attr_rw(energy_performance_preference);
 834
 835 static struct freq_attr *hwp_cpufreq_attrs[] = {
 836         &energy_performance_preference,
 837         &energy_performance_available_preferences,
 838         NULL,
 839 };
 840
 841 static void intel_pstate_hwp_set(struct cpufreq_policy *policy)
 842 {
 843         int min, hw_min, max, hw_max, cpu;
 844         struct perf_limits *perf_limits = &global;
 845         u64 value, cap;
 846
 847         for_each_cpu(cpu, policy->cpus) {
 848                 struct cpudata *cpu_data = all_cpu_data[cpu];
 849                 s16 epp;
 850
 851                 if (per_cpu_limits)
 852                         perf_limits = all_cpu_data[cpu]->perf_limits;
 853
 854                 rdmsrl_on_cpu(cpu, MSR_HWP_CAPABILITIES, &cap);
 855                 hw_min = HWP_LOWEST_PERF(cap);
 856                 if (global.no_turbo)
 857                         hw_max = HWP_GUARANTEED_PERF(cap);
 858                 else
 859                         hw_max = HWP_HIGHEST_PERF(cap);
 860
 861                 max = fp_ext_toint(hw_max * perf_limits->max_perf);
 862                 if (cpu_data->policy == CPUFREQ_POLICY_PERFORMANCE)
 863                         min = max;
 864                 else
 865                         min = fp_ext_toint(hw_max * perf_limits->min_perf);
 866
 867                 rdmsrl_on_cpu(cpu, MSR_HWP_REQUEST, &value);
 868
 869                 value &= ~HWP_MIN_PERF(~0L);
 870                 value |= HWP_MIN_PERF(min);
 871
 872                 value &= ~HWP_MAX_PERF(~0L);
 873                 value |= HWP_MAX_PERF(max);
 874
 875                 if (cpu_data->epp_policy == cpu_data->policy)
 876                         goto skip_epp;
 877
 878                 cpu_data->epp_policy = cpu_data->policy;
 879
 880                 if (cpu_data->epp_saved >= 0) {
 881                         epp = cpu_data->epp_saved;
 882                         cpu_data->epp_saved = -EINVAL;
 883                         goto update_epp;
 884                 }
 885
 886                 if (cpu_data->policy == CPUFREQ_POLICY_PERFORMANCE) {
 887                         epp = intel_pstate_get_epp(cpu_data, value);
 888                         cpu_data->epp_powersave = epp;
 889                         /* If EPP read was failed, then don't try to write */
 890                         if (epp < 0)
 891                                 goto skip_epp;
 892
 893
 894                         epp = 0;
 895                 } else {
 896                         /* skip setting EPP, when saved value is invalid */
 897                         if (cpu_data->epp_powersave < 0)
 898                                 goto skip_epp;
 899
 900                         /*
 901                          * No need to restore EPP when it is not zero. This
 902                          * means:
 903                          *  - Policy is not changed
 904                          *  - user has manually changed
 905                          *  - Error reading EPB
 906                          */
 907                         epp = intel_pstate_get_epp(cpu_data, value);
 908                         if (epp)
 909                                 goto skip_epp;
 910
 911                         epp = cpu_data->epp_powersave;
 912                 }
 913 update_epp:
 914                 if (static_cpu_has(X86_FEATURE_HWP_EPP)) {
 915                         value &= ~GENMASK_ULL(31, 24);
 916                         value |= (u64)epp << 24;
 917                 } else {
 918                         intel_pstate_set_epb(cpu, epp);
 919                 }
 920 skip_epp:
 921                 wrmsrl_on_cpu(cpu, MSR_HWP_REQUEST, value);
 922         }
 923 }
 924
 925 static int intel_pstate_hwp_save_state(struct cpufreq_policy *policy)
 926 {
 927         struct cpudata *cpu_data = all_cpu_data[policy->cpu];
 928
 929         if (!hwp_active)
 930                 return 0;
 931
 932         cpu_data->epp_saved = intel_pstate_get_epp(cpu_data, 0);
 933
 934         return 0;
 935 }
 936
 937 static int intel_pstate_resume(struct cpufreq_policy *policy)
 938 {
 939         if (!hwp_active)
 940                 return 0;
 941
 942         mutex_lock(&intel_pstate_limits_lock);
 943
 944         all_cpu_data[policy->cpu]->epp_policy = 0;
 945         intel_pstate_hwp_set(policy);
 946
 947         mutex_unlock(&intel_pstate_limits_lock);
 948
 949         return 0;
 950 }
 951
 952 static void intel_pstate_update_policies(void)
 953 {
 954         int cpu;
 955
 956         for_each_possible_cpu(cpu)
 957                 cpufreq_update_policy(cpu);
 958 }
 959
 960 /************************** debugfs begin ************************/
 961 static int pid_param_set(void *data, u64 val)
 962 {
 963         *(u32 *)data = val;
 964         pid_params.sample_rate_ns = pid_params.sample_rate_ms * NSEC_PER_MSEC;
 965         intel_pstate_reset_all_pid();
 966         return 0;
 967 }
 968
 969 static int pid_param_get(void *data, u64 *val)
 970 {
 971         *val = *(u32 *)data;
 972         return 0;
 973 }
 974 DEFINE_SIMPLE_ATTRIBUTE(fops_pid_param, pid_param_get, pid_param_set, "%llu\n");
 975
 976 static struct dentry *debugfs_parent;
 977
 978 struct pid_param {
 979         char *name;
 980         void *value;
 981         struct dentry *dentry;
 982 };
 983
 984 static struct pid_param pid_files[] = {
 985         {"sample_rate_ms", &pid_params.sample_rate_ms, },
 986         {"d_gain_pct", &pid_params.d_gain_pct, },
 987         {"i_gain_pct", &pid_params.i_gain_pct, },
 988         {"deadband", &pid_params.deadband, },
 989         {"setpoint", &pid_params.setpoint, },
 990         {"p_gain_pct", &pid_params.p_gain_pct, },
 991         {NULL, NULL, }
 992 };
 993
 994 static void intel_pstate_debug_expose_params(void)
 995 {
 996         int i;
 997
 998         debugfs_parent = debugfs_create_dir("pstate_snb", NULL);
 999         if (IS_ERR_OR_NULL(debugfs_parent))
1000                 return;
1001
1002         for (i = 0; pid_files[i].name; i++) {
1003                 struct dentry *dentry;
1004
1005                 dentry = debugfs_create_file(pid_files[i].name, 0660,
1006                                              debugfs_parent, pid_files[i].value,
1007                                              &fops_pid_param);
1008                 if (!IS_ERR(dentry))
1009                         pid_files[i].dentry = dentry;
1010         }
1011 }
1012
1013 static void intel_pstate_debug_hide_params(void)
1014 {
1015         int i;
1016
1017         if (IS_ERR_OR_NULL(debugfs_parent))
1018                 return;
1019
1020         for (i = 0; pid_files[i].name; i++) {
1021                 debugfs_remove(pid_files[i].dentry);
1022                 pid_files[i].dentry = NULL;
1023         }
1024
1025         debugfs_remove(debugfs_parent);
1026         debugfs_parent = NULL;
1027 }
1028
1029 /************************** debugfs end ************************/
1030
1031 /************************** sysfs begin ************************/
1032 #define show_one(file_name, object)                                     \
1033         static ssize_t show_##file_name                                 \
1034         (struct kobject *kobj, struct attribute *attr, char *buf)       \
1035         {                                                               \
1036                 return sprintf(buf, "%u\n", global.object);             \
1037         }
1038
1039 static ssize_t intel_pstate_show_status(char *buf);
1040 static int intel_pstate_update_status(const char *buf, size_t size);
1041
1042 static ssize_t show_status(struct kobject *kobj,
1043                            struct attribute *attr, char *buf)
1044 {
1045         ssize_t ret;
1046
1047         mutex_lock(&intel_pstate_driver_lock);
1048         ret = intel_pstate_show_status(buf);
1049         mutex_unlock(&intel_pstate_driver_lock);
1050
1051         return ret;
1052 }
1053
1054 static ssize_t store_status(struct kobject *a, struct attribute *b,
1055                             const char *buf, size_t count)
1056 {
1057         char *p = memchr(buf, '\n', count);
1058         int ret;
1059
1060         mutex_lock(&intel_pstate_driver_lock);
1061         ret = intel_pstate_update_status(buf, p ? p - buf : count);
1062         mutex_unlock(&intel_pstate_driver_lock);
1063
1064         return ret < 0 ? ret : count;
1065 }
1066
1067 static ssize_t show_turbo_pct(struct kobject *kobj,
1068                                 struct attribute *attr, char *buf)
1069 {
1070         struct cpudata *cpu;
1071         int total, no_turbo, turbo_pct;
1072         uint32_t turbo_fp;
1073
1074         mutex_lock(&intel_pstate_driver_lock);
1075
1076         if (!driver_registered) {
1077                 mutex_unlock(&intel_pstate_driver_lock);
1078                 return -EAGAIN;
1079         }
1080
1081         cpu = all_cpu_data[0];
1082
1083         total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
1084         no_turbo = cpu->pstate.max_pstate - cpu->pstate.min_pstate + 1;
1085         turbo_fp = div_fp(no_turbo, total);
1086         turbo_pct = 100 - fp_toint(mul_fp(turbo_fp, int_tofp(100)));
1087
1088         mutex_unlock(&intel_pstate_driver_lock);
1089
1090         return sprintf(buf, "%u\n", turbo_pct);
1091 }
1092
1093 static ssize_t show_num_pstates(struct kobject *kobj,
1094                                 struct attribute *attr, char *buf)
1095 {
1096         struct cpudata *cpu;
1097         int total;
1098
1099         mutex_lock(&intel_pstate_driver_lock);
1100
1101         if (!driver_registered) {
1102                 mutex_unlock(&intel_pstate_driver_lock);
1103                 return -EAGAIN;
1104         }
1105
1106         cpu = all_cpu_data[0];
1107         total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
1108
1109         mutex_unlock(&intel_pstate_driver_lock);
1110
1111         return sprintf(buf, "%u\n", total);
1112 }
1113
1114 static ssize_t show_no_turbo(struct kobject *kobj,
1115                              struct attribute *attr, char *buf)
1116 {
1117         ssize_t ret;
1118
1119         mutex_lock(&intel_pstate_driver_lock);
1120
1121         if (!driver_registered) {
1122                 mutex_unlock(&intel_pstate_driver_lock);
1123                 return -EAGAIN;
1124         }
1125
1126         update_turbo_state();
1127         if (global.turbo_disabled)
1128                 ret = sprintf(buf, "%u\n", global.turbo_disabled);
1129         else
1130                 ret = sprintf(buf, "%u\n", global.no_turbo);
1131
1132         mutex_unlock(&intel_pstate_driver_lock);
1133
1134         return ret;
1135 }
1136
1137 static ssize_t store_no_turbo(struct kobject *a, struct attribute *b,
1138                               const char *buf, size_t count)
1139 {
1140         unsigned int input;
1141         int ret;
1142
1143         ret = sscanf(buf, "%u", &input);
1144         if (ret != 1)
1145                 return -EINVAL;
1146
1147         mutex_lock(&intel_pstate_driver_lock);
1148
1149         if (!driver_registered) {
1150                 mutex_unlock(&intel_pstate_driver_lock);
1151                 return -EAGAIN;
1152         }
1153
1154         mutex_lock(&intel_pstate_limits_lock);
1155
1156         update_turbo_state();
1157         if (global.turbo_disabled) {
1158                 pr_warn("Turbo disabled by BIOS or unavailable on processor\n");
1159                 mutex_unlock(&intel_pstate_limits_lock);
1160                 mutex_unlock(&intel_pstate_driver_lock);
1161                 return -EPERM;
1162         }
1163
1164         global.no_turbo = clamp_t(int, input, 0, 1);
1165
1166         mutex_unlock(&intel_pstate_limits_lock);
1167
1168         intel_pstate_update_policies();
1169
1170         mutex_unlock(&intel_pstate_driver_lock);
1171
1172         return count;
1173 }
1174
1175 static ssize_t store_max_perf_pct(struct kobject *a, struct attribute *b,
1176                                   const char *buf, size_t count)
1177 {
1178         unsigned int input;
1179         int ret;
1180
1181         ret = sscanf(buf, "%u", &input);
1182         if (ret != 1)
1183                 return -EINVAL;
1184
1185         mutex_lock(&intel_pstate_driver_lock);
1186
1187         if (!driver_registered) {
1188                 mutex_unlock(&intel_pstate_driver_lock);
1189                 return -EAGAIN;
1190         }
1191
1192         mutex_lock(&intel_pstate_limits_lock);
1193
1194         global.max_sysfs_pct = clamp_t(int, input, 0 , 100);
1195         global.max_perf_pct = min(global.max_policy_pct, global.max_sysfs_pct);
1196         global.max_perf_pct = max(global.min_policy_pct, global.max_perf_pct);
1197         global.max_perf_pct = max(global.min_perf_pct, global.max_perf_pct);
1198         global.max_perf = percent_ext_fp(global.max_perf_pct);
1199
1200         mutex_unlock(&intel_pstate_limits_lock);
1201
1202         intel_pstate_update_policies();
1203
1204         mutex_unlock(&intel_pstate_driver_lock);
1205
1206         return count;
1207 }
1208
1209 static ssize_t store_min_perf_pct(struct kobject *a, struct attribute *b,
1210                                   const char *buf, size_t count)
1211 {
1212         unsigned int input;
1213         int ret;
1214
1215         ret = sscanf(buf, "%u", &input);
1216         if (ret != 1)
1217                 return -EINVAL;
1218
1219         mutex_lock(&intel_pstate_driver_lock);
1220
1221         if (!driver_registered) {
1222                 mutex_unlock(&intel_pstate_driver_lock);
1223                 return -EAGAIN;
1224         }
1225
1226         mutex_lock(&intel_pstate_limits_lock);
1227
1228         global.min_sysfs_pct = clamp_t(int, input, 0 , 100);
1229         global.min_perf_pct = max(global.min_policy_pct, global.min_sysfs_pct);
1230         global.min_perf_pct = min(global.max_policy_pct, global.min_perf_pct);
1231         global.min_perf_pct = min(global.max_perf_pct, global.min_perf_pct);
1232         global.min_perf = percent_ext_fp(global.min_perf_pct);
1233
1234         mutex_unlock(&intel_pstate_limits_lock);
1235
1236         intel_pstate_update_policies();
1237
1238         mutex_unlock(&intel_pstate_driver_lock);
1239
1240         return count;
1241 }
1242
1243 show_one(max_perf_pct, max_perf_pct);
1244 show_one(min_perf_pct, min_perf_pct);
1245
1246 define_one_global_rw(status);
1247 define_one_global_rw(no_turbo);
1248 define_one_global_rw(max_perf_pct);
1249 define_one_global_rw(min_perf_pct);
1250 define_one_global_ro(turbo_pct);
1251 define_one_global_ro(num_pstates);
1252
1253 static struct attribute *intel_pstate_attributes[] = {
1254         &status.attr,
1255         &no_turbo.attr,
1256         &turbo_pct.attr,
1257         &num_pstates.attr,
1258         NULL
1259 };
1260
1261 static struct attribute_group intel_pstate_attr_group = {
1262         .attrs = intel_pstate_attributes,
1263 };
1264
1265 static void __init intel_pstate_sysfs_expose_params(void)
1266 {
1267         struct kobject *intel_pstate_kobject;
1268         int rc;
1269
1270         intel_pstate_kobject = kobject_create_and_add("intel_pstate",
1271                                                 &cpu_subsys.dev_root->kobj);
1272         if (WARN_ON(!intel_pstate_kobject))
1273                 return;
1274
1275         rc = sysfs_create_group(intel_pstate_kobject, &intel_pstate_attr_group);
1276         if (WARN_ON(rc))
1277                 return;
1278
1279         /*
1280          * If per cpu limits are enforced there are no global limits, so
1281          * return without creating max/min_perf_pct attributes
1282          */
1283         if (per_cpu_limits)
1284                 return;
1285
1286         rc = sysfs_create_file(intel_pstate_kobject, &max_perf_pct.attr);
1287         WARN_ON(rc);
1288
1289         rc = sysfs_create_file(intel_pstate_kobject, &min_perf_pct.attr);
1290         WARN_ON(rc);
1291
1292 }
1293 /************************** sysfs end ************************/
1294
1295 static void intel_pstate_hwp_enable(struct cpudata *cpudata)
1296 {
1297         /* First disable HWP notification interrupt as we don't process them */
1298         if (static_cpu_has(X86_FEATURE_HWP_NOTIFY))
1299                 wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_INTERRUPT, 0x00);
1300
1301         wrmsrl_on_cpu(cpudata->cpu, MSR_PM_ENABLE, 0x1);
1302         cpudata->epp_policy = 0;
1303         if (cpudata->epp_default == -EINVAL)
1304                 cpudata->epp_default = intel_pstate_get_epp(cpudata, 0);
1305 }
1306
1307 #define MSR_IA32_POWER_CTL_BIT_EE       19
1308
1309 /* Disable energy efficiency optimization */
1310 static void intel_pstate_disable_ee(int cpu)
1311 {
1312         u64 power_ctl;
1313         int ret;
1314
1315         ret = rdmsrl_on_cpu(cpu, MSR_IA32_POWER_CTL, &power_ctl);
1316         if (ret)
1317                 return;
1318
1319         if (!(power_ctl & BIT(MSR_IA32_POWER_CTL_BIT_EE))) {
1320                 pr_info("Disabling energy efficiency optimization\n");
1321                 power_ctl |= BIT(MSR_IA32_POWER_CTL_BIT_EE);
1322                 wrmsrl_on_cpu(cpu, MSR_IA32_POWER_CTL, power_ctl);
1323         }
1324 }
1325
1326 static int atom_get_min_pstate(void)
1327 {
1328         u64 value;
1329
1330         rdmsrl(MSR_ATOM_CORE_RATIOS, value);
1331         return (value >> 8) & 0x7F;
1332 }
1333
1334 static int atom_get_max_pstate(void)
1335 {
1336         u64 value;
1337
1338         rdmsrl(MSR_ATOM_CORE_RATIOS, value);
1339         return (value >> 16) & 0x7F;
1340 }
1341
1342 static int atom_get_turbo_pstate(void)
1343 {
1344         u64 value;
1345
1346         rdmsrl(MSR_ATOM_CORE_TURBO_RATIOS, value);
1347         return value & 0x7F;
1348 }
1349
1350 static u64 atom_get_val(struct cpudata *cpudata, int pstate)
1351 {
1352         u64 val;
1353         int32_t vid_fp;
1354         u32 vid;
1355
1356         val = (u64)pstate << 8;
1357         if (global.no_turbo && !global.turbo_disabled)
1358                 val |= (u64)1 << 32;
1359
1360         vid_fp = cpudata->vid.min + mul_fp(
1361                 int_tofp(pstate - cpudata->pstate.min_pstate),
1362                 cpudata->vid.ratio);
1363
1364         vid_fp = clamp_t(int32_t, vid_fp, cpudata->vid.min, cpudata->vid.max);
1365         vid = ceiling_fp(vid_fp);
1366
1367         if (pstate > cpudata->pstate.max_pstate)
1368                 vid = cpudata->vid.turbo;
1369
1370         return val | vid;
1371 }
1372
1373 static int silvermont_get_scaling(void)
1374 {
1375         u64 value;
1376         int i;
1377         /* Defined in Table 35-6 from SDM (Sept 2015) */
1378         static int silvermont_freq_table[] = {
1379                 83300, 100000, 133300, 116700, 80000};
1380
1381         rdmsrl(MSR_FSB_FREQ, value);
1382         i = value & 0x7;
1383         WARN_ON(i > 4);
1384
1385         return silvermont_freq_table[i];
1386 }
1387
1388 static int airmont_get_scaling(void)
1389 {
1390         u64 value;
1391         int i;
1392         /* Defined in Table 35-10 from SDM (Sept 2015) */
1393         static int airmont_freq_table[] = {
1394                 83300, 100000, 133300, 116700, 80000,
1395                 93300, 90000, 88900, 87500};
1396
1397         rdmsrl(MSR_FSB_FREQ, value);
1398         i = value & 0xF;
1399         WARN_ON(i > 8);
1400
1401         return airmont_freq_table[i];
1402 }
1403
1404 static void atom_get_vid(struct cpudata *cpudata)
1405 {
1406         u64 value;
1407
1408         rdmsrl(MSR_ATOM_CORE_VIDS, value);
1409         cpudata->vid.min = int_tofp((value >> 8) & 0x7f);
1410         cpudata->vid.max = int_tofp((value >> 16) & 0x7f);
1411         cpudata->vid.ratio = div_fp(
1412                 cpudata->vid.max - cpudata->vid.min,
1413                 int_tofp(cpudata->pstate.max_pstate -
1414                         cpudata->pstate.min_pstate));
1415
1416         rdmsrl(MSR_ATOM_CORE_TURBO_VIDS, value);
1417         cpudata->vid.turbo = value & 0x7f;
1418 }
1419
1420 static int core_get_min_pstate(void)
1421 {
1422         u64 value;
1423
1424         rdmsrl(MSR_PLATFORM_INFO, value);
1425         return (value >> 40) & 0xFF;
1426 }
1427
1428 static int core_get_max_pstate_physical(void)
1429 {
1430         u64 value;
1431
1432         rdmsrl(MSR_PLATFORM_INFO, value);
1433         return (value >> 8) & 0xFF;
1434 }
1435
1436 static int core_get_tdp_ratio(u64 plat_info)
1437 {
1438         /* Check how many TDP levels present */
1439         if (plat_info & 0x600000000) {
1440                 u64 tdp_ctrl;
1441                 u64 tdp_ratio;
1442                 int tdp_msr;
1443                 int err;
1444
1445                 /* Get the TDP level (0, 1, 2) to get ratios */
1446                 err = rdmsrl_safe(MSR_CONFIG_TDP_CONTROL, &tdp_ctrl);
1447                 if (err)
1448                         return err;
1449
1450                 /* TDP MSR are continuous starting at 0x648 */
1451                 tdp_msr = MSR_CONFIG_TDP_NOMINAL + (tdp_ctrl & 0x03);
1452                 err = rdmsrl_safe(tdp_msr, &tdp_ratio);
1453                 if (err)
1454                         return err;
1455
1456                 /* For level 1 and 2, bits[23:16] contain the ratio */
1457                 if (tdp_ctrl & 0x03)
1458                         tdp_ratio >>= 16;
1459
1460                 tdp_ratio &= 0xff; /* ratios are only 8 bits long */
1461                 pr_debug("tdp_ratio %x\n", (int)tdp_ratio);
1462
1463                 return (int)tdp_ratio;
1464         }
1465
1466         return -ENXIO;
1467 }
1468
1469 static int core_get_max_pstate(void)
1470 {
1471         u64 tar;
1472         u64 plat_info;
1473         int max_pstate;
1474         int tdp_ratio;
1475         int err;
1476
1477         rdmsrl(MSR_PLATFORM_INFO, plat_info);
1478         max_pstate = (plat_info >> 8) & 0xFF;
1479
1480         tdp_ratio = core_get_tdp_ratio(plat_info);
1481         if (tdp_ratio <= 0)
1482                 return max_pstate;
1483
1484         if (hwp_active) {
1485                 /* Turbo activation ratio is not used on HWP platforms */
1486                 return tdp_ratio;
1487         }
1488
1489         err = rdmsrl_safe(MSR_TURBO_ACTIVATION_RATIO, &tar);
1490         if (!err) {
1491                 int tar_levels;
1492
1493                 /* Do some sanity checking for safety */
1494                 tar_levels = tar & 0xff;
1495                 if (tdp_ratio - 1 == tar_levels) {
1496                         max_pstate = tar_levels;
1497                         pr_debug("max_pstate=TAC %x\n", max_pstate);
1498                 }
1499         }
1500
1501         return max_pstate;
1502 }
1503
1504 static int core_get_turbo_pstate(void)
1505 {
1506         u64 value;
1507         int nont, ret;
1508
1509         rdmsrl(MSR_TURBO_RATIO_LIMIT, value);
1510         nont = core_get_max_pstate();
1511         ret = (value) & 255;
1512         if (ret <= nont)
1513                 ret = nont;
1514         return ret;
1515 }
1516
1517 static inline int core_get_scaling(void)
1518 {
1519         return 100000;
1520 }
1521
1522 static u64 core_get_val(struct cpudata *cpudata, int pstate)
1523 {
1524         u64 val;
1525
1526         val = (u64)pstate << 8;
1527         if (global.no_turbo && !global.turbo_disabled)
1528                 val |= (u64)1 << 32;
1529
1530         return val;
1531 }
1532
1533 static int knl_get_turbo_pstate(void)
1534 {
1535         u64 value;
1536         int nont, ret;
1537
1538         rdmsrl(MSR_TURBO_RATIO_LIMIT, value);
1539         nont = core_get_max_pstate();
1540         ret = (((value) >> 8) & 0xFF);
1541         if (ret <= nont)
1542                 ret = nont;
1543         return ret;
1544 }
1545
1546 static struct cpu_defaults core_params = {
1547         .pid_policy = {
1548                 .sample_rate_ms = 10,
1549                 .deadband = 0,
1550                 .setpoint = 97,
1551                 .p_gain_pct = 20,
1552                 .d_gain_pct = 0,
1553                 .i_gain_pct = 0,
1554         },
1555         .funcs = {
1556                 .get_max = core_get_max_pstate,
1557                 .get_max_physical = core_get_max_pstate_physical,
1558                 .get_min = core_get_min_pstate,
1559                 .get_turbo = core_get_turbo_pstate,
1560                 .get_scaling = core_get_scaling,
1561                 .get_val = core_get_val,
1562                 .get_target_pstate = get_target_pstate_use_performance,
1563         },
1564 };
1565
1566 static const struct cpu_defaults silvermont_params = {
1567         .pid_policy = {
1568                 .sample_rate_ms = 10,
1569                 .deadband = 0,
1570                 .setpoint = 60,
1571                 .p_gain_pct = 14,
1572                 .d_gain_pct = 0,
1573                 .i_gain_pct = 4,
1574         },
1575         .funcs = {
1576                 .get_max = atom_get_max_pstate,
1577                 .get_max_physical = atom_get_max_pstate,
1578                 .get_min = atom_get_min_pstate,
1579                 .get_turbo = atom_get_turbo_pstate,
1580                 .get_val = atom_get_val,
1581                 .get_scaling = silvermont_get_scaling,
1582                 .get_vid = atom_get_vid,
1583                 .get_target_pstate = get_target_pstate_use_cpu_load,
1584         },
1585 };
1586
1587 static const struct cpu_defaults airmont_params = {
1588         .pid_policy = {
1589                 .sample_rate_ms = 10,
1590                 .deadband = 0,
1591                 .setpoint = 60,
1592                 .p_gain_pct = 14,
1593                 .d_gain_pct = 0,
1594                 .i_gain_pct = 4,
1595         },
1596         .funcs = {
1597                 .get_max = atom_get_max_pstate,
1598                 .get_max_physical = atom_get_max_pstate,
1599                 .get_min = atom_get_min_pstate,
1600                 .get_turbo = atom_get_turbo_pstate,
1601                 .get_val = atom_get_val,
1602                 .get_scaling = airmont_get_scaling,
1603                 .get_vid = atom_get_vid,
1604                 .get_target_pstate = get_target_pstate_use_cpu_load,
1605         },
1606 };
1607
1608 static const struct cpu_defaults knl_params = {
1609         .pid_policy = {
1610                 .sample_rate_ms = 10,
1611                 .deadband = 0,
1612                 .setpoint = 97,
1613                 .p_gain_pct = 20,
1614                 .d_gain_pct = 0,
1615                 .i_gain_pct = 0,
1616         },
1617         .funcs = {
1618                 .get_max = core_get_max_pstate,
1619                 .get_max_physical = core_get_max_pstate_physical,
1620                 .get_min = core_get_min_pstate,
1621                 .get_turbo = knl_get_turbo_pstate,
1622                 .get_scaling = core_get_scaling,
1623                 .get_val = core_get_val,
1624                 .get_target_pstate = get_target_pstate_use_performance,
1625         },
1626 };
1627
1628 static const struct cpu_defaults bxt_params = {
1629         .pid_policy = {
1630                 .sample_rate_ms = 10,
1631                 .deadband = 0,
1632                 .setpoint = 60,
1633                 .p_gain_pct = 14,
1634                 .d_gain_pct = 0,
1635                 .i_gain_pct = 4,
1636         },
1637         .funcs = {
1638                 .get_max = core_get_max_pstate,
1639                 .get_max_physical = core_get_max_pstate_physical,
1640                 .get_min = core_get_min_pstate,
1641                 .get_turbo = core_get_turbo_pstate,
1642                 .get_scaling = core_get_scaling,
1643                 .get_val = core_get_val,
1644                 .get_target_pstate = get_target_pstate_use_cpu_load,
1645         },
1646 };
1647
1648 static void intel_pstate_get_min_max(struct cpudata *cpu, int *min, int *max)
1649 {
1650         int max_perf = cpu->pstate.turbo_pstate;
1651         int max_perf_adj;
1652         int min_perf;
1653         struct perf_limits *perf_limits = &global;
1654
1655         if (global.no_turbo || global.turbo_disabled)
1656                 max_perf = cpu->pstate.max_pstate;
1657
1658         if (per_cpu_limits)
1659                 perf_limits = cpu->perf_limits;
1660
1661         /*
1662          * performance can be limited by user through sysfs, by cpufreq
1663          * policy, or by cpu specific default values determined through
1664          * experimentation.
1665          */
1666         max_perf_adj = fp_ext_toint(max_perf * perf_limits->max_perf);
1667         *max = clamp_t(int, max_perf_adj,
1668                         cpu->pstate.min_pstate, cpu->pstate.turbo_pstate);
1669
1670         min_perf = fp_ext_toint(max_perf * perf_limits->min_perf);
1671         *min = clamp_t(int, min_perf, cpu->pstate.min_pstate, max_perf);
1672 }
1673
1674 static void intel_pstate_set_pstate(struct cpudata *cpu, int pstate)
1675 {
1676         trace_cpu_frequency(pstate * cpu->pstate.scaling, cpu->cpu);
1677         cpu->pstate.current_pstate = pstate;
1678         /*
1679          * Generally, there is no guarantee that this code will always run on
1680          * the CPU being updated, so force the register update to run on the
1681          * right CPU.
1682          */
1683         wrmsrl_on_cpu(cpu->cpu, MSR_IA32_PERF_CTL,
1684                       pstate_funcs.get_val(cpu, pstate));
1685 }
1686
1687 static void intel_pstate_set_min_pstate(struct cpudata *cpu)
1688 {
1689         intel_pstate_set_pstate(cpu, cpu->pstate.min_pstate);
1690 }
1691
1692 static void intel_pstate_max_within_limits(struct cpudata *cpu)
1693 {
1694         int min_pstate, max_pstate;
1695
1696         update_turbo_state();
1697         intel_pstate_get_min_max(cpu, &min_pstate, &max_pstate);
1698         intel_pstate_set_pstate(cpu, max_pstate);
1699 }
1700
1701 static void intel_pstate_get_cpu_pstates(struct cpudata *cpu)
1702 {
1703         cpu->pstate.min_pstate = pstate_funcs.get_min();
1704         cpu->pstate.max_pstate = pstate_funcs.get_max();
1705         cpu->pstate.max_pstate_physical = pstate_funcs.get_max_physical();
1706         cpu->pstate.turbo_pstate = pstate_funcs.get_turbo();
1707         cpu->pstate.scaling = pstate_funcs.get_scaling();
1708         cpu->pstate.max_freq = cpu->pstate.max_pstate * cpu->pstate.scaling;
1709         cpu->pstate.turbo_freq = cpu->pstate.turbo_pstate * cpu->pstate.scaling;
1710
1711         if (pstate_funcs.get_vid)
1712                 pstate_funcs.get_vid(cpu);
1713
1714         intel_pstate_set_min_pstate(cpu);
1715 }
1716
1717 static inline void intel_pstate_calc_avg_perf(struct cpudata *cpu)
1718 {
1719         struct sample *sample = &cpu->sample;
1720
1721         sample->core_avg_perf = div_ext_fp(sample->aperf, sample->mperf);
1722 }
1723
1724 static inline bool intel_pstate_sample(struct cpudata *cpu, u64 time)
1725 {
1726         u64 aperf, mperf;
1727         unsigned long flags;
1728         u64 tsc;
1729
1730         local_irq_save(flags);
1731         rdmsrl(MSR_IA32_APERF, aperf);
1732         rdmsrl(MSR_IA32_MPERF, mperf);
1733         tsc = rdtsc();
1734         if (cpu->prev_mperf == mperf || cpu->prev_tsc == tsc) {
1735                 local_irq_restore(flags);
1736                 return false;
1737         }
1738         local_irq_restore(flags);
1739
1740         cpu->last_sample_time = cpu->sample.time;
1741         cpu->sample.time = time;
1742         cpu->sample.aperf = aperf;
1743         cpu->sample.mperf = mperf;
1744         cpu->sample.tsc =  tsc;
1745         cpu->sample.aperf -= cpu->prev_aperf;
1746         cpu->sample.mperf -= cpu->prev_mperf;
1747         cpu->sample.tsc -= cpu->prev_tsc;
1748
1749         cpu->prev_aperf = aperf;
1750         cpu->prev_mperf = mperf;
1751         cpu->prev_tsc = tsc;
1752         /*
1753          * First time this function is invoked in a given cycle, all of the
1754          * previous sample data fields are equal to zero or stale and they must
1755          * be populated with meaningful numbers for things to work, so assume
1756          * that sample.time will always be reset before setting the utilization
1757          * update hook and make the caller skip the sample then.
1758          */
1759         return !!cpu->last_sample_time;
1760 }
1761
1762 static inline int32_t get_avg_frequency(struct cpudata *cpu)
1763 {
1764         return mul_ext_fp(cpu->sample.core_avg_perf,
1765                           cpu->pstate.max_pstate_physical * cpu->pstate.scaling);
1766 }
1767
1768 static inline int32_t get_avg_pstate(struct cpudata *cpu)
1769 {
1770         return mul_ext_fp(cpu->pstate.max_pstate_physical,
1771                           cpu->sample.core_avg_perf);
1772 }
1773
1774 static inline int32_t get_target_pstate_use_cpu_load(struct cpudata *cpu)
1775 {
1776         struct sample *sample = &cpu->sample;
1777         int32_t busy_frac, boost;
1778         int target, avg_pstate;
1779
1780         busy_frac = div_fp(sample->mperf, sample->tsc);
1781
1782         boost = cpu->iowait_boost;
1783         cpu->iowait_boost >>= 1;
1784
1785         if (busy_frac < boost)
1786                 busy_frac = boost;
1787
1788         sample->busy_scaled = busy_frac * 100;
1789
1790         target = global.no_turbo || global.turbo_disabled ?
1791                         cpu->pstate.max_pstate : cpu->pstate.turbo_pstate;
1792         target += target >> 2;
1793         target = mul_fp(target, busy_frac);
1794         if (target < cpu->pstate.min_pstate)
1795                 target = cpu->pstate.min_pstate;
1796
1797         /*
1798          * If the average P-state during the previous cycle was higher than the
1799          * current target, add 50% of the difference to the target to reduce
1800          * possible performance oscillations and offset possible performance
1801          * loss related to moving the workload from one CPU to another within
1802          * a package/module.
1803          */
1804         avg_pstate = get_avg_pstate(cpu);
1805         if (avg_pstate > target)
1806                 target += (avg_pstate - target) >> 1;
1807
1808         return target;
1809 }
1810
1811 static inline int32_t get_target_pstate_use_performance(struct cpudata *cpu)
1812 {
1813         int32_t perf_scaled, max_pstate, current_pstate, sample_ratio;
1814         u64 duration_ns;
1815
1816         /*
1817          * perf_scaled is the ratio of the average P-state during the last
1818          * sampling period to the P-state requested last time (in percent).
1819          *
1820          * That measures the system's response to the previous P-state
1821          * selection.
1822          */
1823         max_pstate = cpu->pstate.max_pstate_physical;
1824         current_pstate = cpu->pstate.current_pstate;
1825         perf_scaled = mul_ext_fp(cpu->sample.core_avg_perf,
1826                                div_fp(100 * max_pstate, current_pstate));
1827
1828         /*
1829          * Since our utilization update callback will not run unless we are
1830          * in C0, check if the actual elapsed time is significantly greater (3x)
1831          * than our sample interval.  If it is, then we were idle for a long
1832          * enough period of time to adjust our performance metric.
1833          */
1834         duration_ns = cpu->sample.time - cpu->last_sample_time;
1835         if ((s64)duration_ns > pid_params.sample_rate_ns * 3) {
1836                 sample_ratio = div_fp(pid_params.sample_rate_ns, duration_ns);
1837                 perf_scaled = mul_fp(perf_scaled, sample_ratio);
1838         } else {
1839                 sample_ratio = div_fp(100 * cpu->sample.mperf, cpu->sample.tsc);
1840                 if (sample_ratio < int_tofp(1))
1841                         perf_scaled = 0;
1842         }
1843
1844         cpu->sample.busy_scaled = perf_scaled;
1845         return cpu->pstate.current_pstate - pid_calc(&cpu->pid, perf_scaled);
1846 }
1847
1848 static int intel_pstate_prepare_request(struct cpudata *cpu, int pstate)
1849 {
1850         int max_perf, min_perf;
1851
1852         intel_pstate_get_min_max(cpu, &min_perf, &max_perf);
1853         pstate = clamp_t(int, pstate, min_perf, max_perf);
1854         return pstate;
1855 }
1856
1857 static void intel_pstate_update_pstate(struct cpudata *cpu, int pstate)
1858 {
1859         if (pstate == cpu->pstate.current_pstate)
1860                 return;
1861
1862         cpu->pstate.current_pstate = pstate;
1863         wrmsrl(MSR_IA32_PERF_CTL, pstate_funcs.get_val(cpu, pstate));
1864 }
1865
1866 static inline void intel_pstate_adjust_busy_pstate(struct cpudata *cpu)
1867 {
1868         int from, target_pstate;
1869         struct sample *sample;
1870
1871         from = cpu->pstate.current_pstate;
1872
1873         target_pstate = cpu->policy == CPUFREQ_POLICY_PERFORMANCE ?
1874                 cpu->pstate.turbo_pstate : pstate_funcs.get_target_pstate(cpu);
1875
1876         update_turbo_state();
1877
1878         target_pstate = intel_pstate_prepare_request(cpu, target_pstate);
1879         trace_cpu_frequency(target_pstate * cpu->pstate.scaling, cpu->cpu);
1880         intel_pstate_update_pstate(cpu, target_pstate);
1881
1882         sample = &cpu->sample;
1883         trace_pstate_sample(mul_ext_fp(100, sample->core_avg_perf),
1884                 fp_toint(sample->busy_scaled),
1885                 from,
1886                 cpu->pstate.current_pstate,
1887                 sample->mperf,
1888                 sample->aperf,
1889                 sample->tsc,
1890                 get_avg_frequency(cpu),
1891                 fp_toint(cpu->iowait_boost * 100));
1892 }
1893
1894 static void intel_pstate_update_util(struct update_util_data *data, u64 time,
1895                                      unsigned int flags)
1896 {
1897         struct cpudata *cpu = container_of(data, struct cpudata, update_util);
1898         u64 delta_ns;
1899
1900         if (pstate_funcs.get_target_pstate == get_target_pstate_use_cpu_load) {
1901                 if (flags & SCHED_CPUFREQ_IOWAIT) {
1902                         cpu->iowait_boost = int_tofp(1);
1903                 } else if (cpu->iowait_boost) {
1904                         /* Clear iowait_boost if the CPU may have been idle. */
1905                         delta_ns = time - cpu->last_update;
1906                         if (delta_ns > TICK_NSEC)
1907                                 cpu->iowait_boost = 0;
1908                 }
1909                 cpu->last_update = time;
1910         }
1911
1912         delta_ns = time - cpu->sample.time;
1913         if ((s64)delta_ns >= pid_params.sample_rate_ns) {
1914                 bool sample_taken = intel_pstate_sample(cpu, time);
1915
1916                 if (sample_taken) {
1917                         intel_pstate_calc_avg_perf(cpu);
1918                         if (!hwp_active)
1919                                 intel_pstate_adjust_busy_pstate(cpu);
1920                 }
1921         }
1922 }
1923
1924 #define ICPU(model, policy) \
1925         { X86_VENDOR_INTEL, 6, model, X86_FEATURE_APERFMPERF,\
1926                         (unsigned long)&policy }
1927
1928 static const struct x86_cpu_id intel_pstate_cpu_ids[] = {
1929         ICPU(INTEL_FAM6_SANDYBRIDGE,            core_params),
1930         ICPU(INTEL_FAM6_SANDYBRIDGE_X,          core_params),
1931         ICPU(INTEL_FAM6_ATOM_SILVERMONT1,       silvermont_params),
1932         ICPU(INTEL_FAM6_IVYBRIDGE,              core_params),
1933         ICPU(INTEL_FAM6_HASWELL_CORE,           core_params),
1934         ICPU(INTEL_FAM6_BROADWELL_CORE,         core_params),
1935         ICPU(INTEL_FAM6_IVYBRIDGE_X,            core_params),
1936         ICPU(INTEL_FAM6_HASWELL_X,              core_params),
1937         ICPU(INTEL_FAM6_HASWELL_ULT,            core_params),
1938         ICPU(INTEL_FAM6_HASWELL_GT3E,           core_params),
1939         ICPU(INTEL_FAM6_BROADWELL_GT3E,         core_params),
1940         ICPU(INTEL_FAM6_ATOM_AIRMONT,           airmont_params),
1941         ICPU(INTEL_FAM6_SKYLAKE_MOBILE,         core_params),
1942         ICPU(INTEL_FAM6_BROADWELL_X,            core_params),
1943         ICPU(INTEL_FAM6_SKYLAKE_DESKTOP,        core_params),
1944         ICPU(INTEL_FAM6_BROADWELL_XEON_D,       core_params),
1945         ICPU(INTEL_FAM6_XEON_PHI_KNL,           knl_params),
1946         ICPU(INTEL_FAM6_XEON_PHI_KNM,           knl_params),
1947         ICPU(INTEL_FAM6_ATOM_GOLDMONT,          bxt_params),
1948         {}
1949 };
1950 MODULE_DEVICE_TABLE(x86cpu, intel_pstate_cpu_ids);
1951
1952 static const struct x86_cpu_id intel_pstate_cpu_oob_ids[] __initconst = {
1953         ICPU(INTEL_FAM6_BROADWELL_XEON_D, core_params),
1954         ICPU(INTEL_FAM6_BROADWELL_X, core_params),
1955         ICPU(INTEL_FAM6_SKYLAKE_X, core_params),
1956         {}
1957 };
1958
1959 static const struct x86_cpu_id intel_pstate_cpu_ee_disable_ids[] = {
1960         ICPU(INTEL_FAM6_KABYLAKE_DESKTOP, core_params),
1961         {}
1962 };
1963
1964 static int intel_pstate_init_cpu(unsigned int cpunum)
1965 {
1966         struct cpudata *cpu;
1967
1968         cpu = all_cpu_data[cpunum];
1969
1970         if (!cpu) {
1971                 unsigned int size = sizeof(struct cpudata);
1972
1973                 if (per_cpu_limits)
1974                         size += sizeof(struct perf_limits);
1975
1976                 cpu = kzalloc(size, GFP_KERNEL);
1977                 if (!cpu)
1978                         return -ENOMEM;
1979
1980                 all_cpu_data[cpunum] = cpu;
1981                 if (per_cpu_limits)
1982                         cpu->perf_limits = (struct perf_limits *)(cpu + 1);
1983
1984                 cpu->epp_default = -EINVAL;
1985                 cpu->epp_powersave = -EINVAL;
1986                 cpu->epp_saved = -EINVAL;
1987         }
1988
1989         cpu = all_cpu_data[cpunum];
1990
1991         cpu->cpu = cpunum;
1992
1993         if (hwp_active) {
1994                 const struct x86_cpu_id *id;
1995
1996                 id = x86_match_cpu(intel_pstate_cpu_ee_disable_ids);
1997                 if (id)
1998                         intel_pstate_disable_ee(cpunum);
1999
2000                 intel_pstate_hwp_enable(cpu);
2001                 pid_params.sample_rate_ms = 50;
2002                 pid_params.sample_rate_ns = 50 * NSEC_PER_MSEC;
2003         }
2004
2005         intel_pstate_get_cpu_pstates(cpu);
2006
2007         intel_pstate_busy_pid_reset(cpu);
2008
2009         pr_debug("controlling: cpu %d\n", cpunum);
2010
2011         return 0;
2012 }
2013
2014 static unsigned int intel_pstate_get(unsigned int cpu_num)
2015 {
2016         struct cpudata *cpu = all_cpu_data[cpu_num];
2017
2018         return cpu ? get_avg_frequency(cpu) : 0;
2019 }
2020
2021 static void intel_pstate_set_update_util_hook(unsigned int cpu_num)
2022 {
2023         struct cpudata *cpu = all_cpu_data[cpu_num];
2024
2025         if (cpu->update_util_set)
2026                 return;
2027
2028         /* Prevent intel_pstate_update_util() from using stale data. */
2029         cpu->sample.time = 0;
2030         cpufreq_add_update_util_hook(cpu_num, &cpu->update_util,
2031                                      intel_pstate_update_util);
2032         cpu->update_util_set = true;
2033 }
2034
2035 static void intel_pstate_clear_update_util_hook(unsigned int cpu)
2036 {
2037         struct cpudata *cpu_data = all_cpu_data[cpu];
2038
2039         if (!cpu_data->update_util_set)
2040                 return;
2041
2042         cpufreq_remove_update_util_hook(cpu);
2043         cpu_data->update_util_set = false;
2044         synchronize_sched();
2045 }
2046
2047 static void intel_pstate_update_perf_limits(struct cpufreq_policy *policy,
2048                                             struct perf_limits *limits)
2049 {
2050         int32_t max_policy_perf, min_policy_perf;
2051
2052         max_policy_perf = div_ext_fp(policy->max, policy->cpuinfo.max_freq);
2053         max_policy_perf = clamp_t(int32_t, max_policy_perf, 0, int_ext_tofp(1));
2054         if (policy->max == policy->min) {
2055                 min_policy_perf = max_policy_perf;
2056         } else {
2057                 min_policy_perf = div_ext_fp(policy->min,
2058                                              policy->cpuinfo.max_freq);
2059                 min_policy_perf = clamp_t(int32_t, min_policy_perf,
2060                                           0, max_policy_perf);
2061         }
2062
2063         /* Normalize user input to [min_perf, max_perf] */
2064         limits->min_perf = max(min_policy_perf,
2065                                percent_ext_fp(limits->min_sysfs_pct));
2066         limits->min_perf = min(limits->min_perf, max_policy_perf);
2067         limits->max_perf = min(max_policy_perf,
2068                                percent_ext_fp(limits->max_sysfs_pct));
2069         limits->max_perf = max(min_policy_perf, limits->max_perf);
2070
2071         /* Make sure min_perf <= max_perf */
2072         limits->min_perf = min(limits->min_perf, limits->max_perf);
2073
2074         limits->max_perf = round_up(limits->max_perf, EXT_FRAC_BITS);
2075         limits->min_perf = round_up(limits->min_perf, EXT_FRAC_BITS);
2076         limits->max_perf_pct = fp_ext_toint(limits->max_perf * 100);
2077         limits->min_perf_pct = fp_ext_toint(limits->min_perf * 100);
2078
2079         pr_debug("cpu:%d max_perf_pct:%d min_perf_pct:%d\n", policy->cpu,
2080                  limits->max_perf_pct, limits->min_perf_pct);
2081 }
2082
2083 static int intel_pstate_set_policy(struct cpufreq_policy *policy)
2084 {
2085         struct cpudata *cpu;
2086         struct perf_limits *perf_limits = &global;
2087
2088         if (!policy->cpuinfo.max_freq)
2089                 return -ENODEV;
2090
2091         pr_debug("set_policy cpuinfo.max %u policy->max %u\n",
2092                  policy->cpuinfo.max_freq, policy->max);
2093
2094         cpu = all_cpu_data[policy->cpu];
2095         cpu->policy = policy->policy;
2096
2097         if (cpu->pstate.max_pstate_physical > cpu->pstate.max_pstate &&
2098             policy->max < policy->cpuinfo.max_freq &&
2099             policy->max > cpu->pstate.max_pstate * cpu->pstate.scaling) {
2100                 pr_debug("policy->max > max non turbo frequency\n");
2101                 policy->max = policy->cpuinfo.max_freq;
2102         }
2103
2104         if (per_cpu_limits)
2105                 perf_limits = cpu->perf_limits;
2106
2107         mutex_lock(&intel_pstate_limits_lock);
2108
2109         intel_pstate_update_perf_limits(policy, perf_limits);
2110
2111         if (cpu->policy == CPUFREQ_POLICY_PERFORMANCE) {
2112                 /*
2113                  * NOHZ_FULL CPUs need this as the governor callback may not
2114                  * be invoked on them.
2115                  */
2116                 intel_pstate_clear_update_util_hook(policy->cpu);
2117                 intel_pstate_max_within_limits(cpu);
2118         }
2119
2120         intel_pstate_set_update_util_hook(policy->cpu);
2121
2122         if (hwp_active)
2123                 intel_pstate_hwp_set(policy);
2124
2125         mutex_unlock(&intel_pstate_limits_lock);
2126
2127         return 0;
2128 }
2129
2130 static int intel_pstate_verify_policy(struct cpufreq_policy *policy)
2131 {
2132         struct cpudata *cpu = all_cpu_data[policy->cpu];
2133
2134         update_turbo_state();
2135         policy->cpuinfo.max_freq = global.turbo_disabled || global.no_turbo ?
2136                                         cpu->pstate.max_freq :
2137                                         cpu->pstate.turbo_freq;
2138
2139         cpufreq_verify_within_cpu_limits(policy);
2140
2141         if (policy->policy != CPUFREQ_POLICY_POWERSAVE &&
2142             policy->policy != CPUFREQ_POLICY_PERFORMANCE)
2143                 return -EINVAL;
2144
2145         /* When per-CPU limits are used, sysfs limits are not used */
2146         if (!per_cpu_limits) {
2147                 unsigned int max_freq, min_freq;
2148
2149                 max_freq = policy->cpuinfo.max_freq *
2150                                         global.max_sysfs_pct / 100;
2151                 min_freq = policy->cpuinfo.max_freq *
2152                                         global.min_sysfs_pct / 100;
2153                 cpufreq_verify_within_limits(policy, min_freq, max_freq);
2154         }
2155
2156         return 0;
2157 }
2158
2159 static void intel_cpufreq_stop_cpu(struct cpufreq_policy *policy)
2160 {
2161         intel_pstate_set_min_pstate(all_cpu_data[policy->cpu]);
2162 }
2163
2164 static void intel_pstate_stop_cpu(struct cpufreq_policy *policy)
2165 {
2166         pr_debug("CPU %d exiting\n", policy->cpu);
2167
2168         intel_pstate_clear_update_util_hook(policy->cpu);
2169         if (hwp_active)
2170                 intel_pstate_hwp_save_state(policy);
2171         else
2172                 intel_cpufreq_stop_cpu(policy);
2173 }
2174
2175 static int intel_pstate_cpu_exit(struct cpufreq_policy *policy)
2176 {
2177         intel_pstate_exit_perf_limits(policy);
2178
2179         policy->fast_switch_possible = false;
2180
2181         return 0;
2182 }
2183
2184 static int __intel_pstate_cpu_init(struct cpufreq_policy *policy)
2185 {
2186         struct cpudata *cpu;
2187         int rc;
2188
2189         rc = intel_pstate_init_cpu(policy->cpu);
2190         if (rc)
2191                 return rc;
2192
2193         cpu = all_cpu_data[policy->cpu];
2194
2195         if (per_cpu_limits)
2196                 intel_pstate_init_limits(cpu->perf_limits);
2197
2198         policy->min = cpu->pstate.min_pstate * cpu->pstate.scaling;
2199         policy->max = cpu->pstate.turbo_pstate * cpu->pstate.scaling;
2200
2201         /* cpuinfo and default policy values */
2202         policy->cpuinfo.min_freq = cpu->pstate.min_pstate * cpu->pstate.scaling;
2203         update_turbo_state();
2204         policy->cpuinfo.max_freq = global.turbo_disabled ?
2205                         cpu->pstate.max_pstate : cpu->pstate.turbo_pstate;
2206         policy->cpuinfo.max_freq *= cpu->pstate.scaling;
2207
2208         intel_pstate_init_acpi_perf_limits(policy);
2209         cpumask_set_cpu(policy->cpu, policy->cpus);
2210
2211         policy->fast_switch_possible = true;
2212
2213         return 0;
2214 }
2215
2216 static int intel_pstate_cpu_init(struct cpufreq_policy *policy)
2217 {
2218         int ret = __intel_pstate_cpu_init(policy);
2219
2220         if (ret)
2221                 return ret;
2222
2223         policy->cpuinfo.transition_latency = CPUFREQ_ETERNAL;
2224         if (IS_ENABLED(CONFIG_CPU_FREQ_DEFAULT_GOV_PERFORMANCE))
2225                 policy->policy = CPUFREQ_POLICY_PERFORMANCE;
2226         else
2227                 policy->policy = CPUFREQ_POLICY_POWERSAVE;
2228
2229         return 0;
2230 }
2231
2232 static struct cpufreq_driver intel_pstate = {
2233         .flags          = CPUFREQ_CONST_LOOPS,
2234         .verify         = intel_pstate_verify_policy,
2235         .setpolicy      = intel_pstate_set_policy,
2236         .suspend        = intel_pstate_hwp_save_state,
2237         .resume         = intel_pstate_resume,
2238         .get            = intel_pstate_get,
2239         .init           = intel_pstate_cpu_init,
2240         .exit           = intel_pstate_cpu_exit,
2241         .stop_cpu       = intel_pstate_stop_cpu,
2242         .name           = "intel_pstate",
2243 };
2244
2245 static int intel_cpufreq_verify_policy(struct cpufreq_policy *policy)
2246 {
2247         struct cpudata *cpu = all_cpu_data[policy->cpu];
2248
2249         update_turbo_state();
2250         policy->cpuinfo.max_freq = global.no_turbo || global.turbo_disabled ?
2251                         cpu->pstate.max_freq : cpu->pstate.turbo_freq;
2252
2253         cpufreq_verify_within_cpu_limits(policy);
2254
2255         return 0;
2256 }
2257
2258 static int intel_cpufreq_target(struct cpufreq_policy *policy,
2259                                 unsigned int target_freq,
2260                                 unsigned int relation)
2261 {
2262         struct cpudata *cpu = all_cpu_data[policy->cpu];
2263         struct cpufreq_freqs freqs;
2264         int target_pstate;
2265
2266         update_turbo_state();
2267
2268         freqs.old = policy->cur;
2269         freqs.new = target_freq;
2270
2271         cpufreq_freq_transition_begin(policy, &freqs);
2272         switch (relation) {
2273         case CPUFREQ_RELATION_L:
2274                 target_pstate = DIV_ROUND_UP(freqs.new, cpu->pstate.scaling);
2275                 break;
2276         case CPUFREQ_RELATION_H:
2277                 target_pstate = freqs.new / cpu->pstate.scaling;
2278                 break;
2279         default:
2280                 target_pstate = DIV_ROUND_CLOSEST(freqs.new, cpu->pstate.scaling);
2281                 break;
2282         }
2283         target_pstate = intel_pstate_prepare_request(cpu, target_pstate);
2284         if (target_pstate != cpu->pstate.current_pstate) {
2285                 cpu->pstate.current_pstate = target_pstate;
2286                 wrmsrl_on_cpu(policy->cpu, MSR_IA32_PERF_CTL,
2287                               pstate_funcs.get_val(cpu, target_pstate));
2288         }
2289         freqs.new = target_pstate * cpu->pstate.scaling;
2290         cpufreq_freq_transition_end(policy, &freqs, false);
2291
2292         return 0;
2293 }
2294
2295 static unsigned int intel_cpufreq_fast_switch(struct cpufreq_policy *policy,
2296                                               unsigned int target_freq)
2297 {
2298         struct cpudata *cpu = all_cpu_data[policy->cpu];
2299         int target_pstate;
2300
2301         update_turbo_state();
2302
2303         target_pstate = DIV_ROUND_UP(target_freq, cpu->pstate.scaling);
2304         target_pstate = intel_pstate_prepare_request(cpu, target_pstate);
2305         intel_pstate_update_pstate(cpu, target_pstate);
2306         return target_pstate * cpu->pstate.scaling;
2307 }
2308
2309 static int intel_cpufreq_cpu_init(struct cpufreq_policy *policy)
2310 {
2311         int ret = __intel_pstate_cpu_init(policy);
2312
2313         if (ret)
2314                 return ret;
2315
2316         policy->cpuinfo.transition_latency = INTEL_CPUFREQ_TRANSITION_LATENCY;
2317         /* This reflects the intel_pstate_get_cpu_pstates() setting. */
2318         policy->cur = policy->cpuinfo.min_freq;
2319
2320         return 0;
2321 }
2322
2323 static struct cpufreq_driver intel_cpufreq = {
2324         .flags          = CPUFREQ_CONST_LOOPS,
2325         .verify         = intel_cpufreq_verify_policy,
2326         .target         = intel_cpufreq_target,
2327         .fast_switch    = intel_cpufreq_fast_switch,
2328         .init           = intel_cpufreq_cpu_init,
2329         .exit           = intel_pstate_cpu_exit,
2330         .stop_cpu       = intel_cpufreq_stop_cpu,
2331         .name           = "intel_cpufreq",
2332 };
2333
2334 static struct cpufreq_driver *intel_pstate_driver = &intel_pstate;
2335
2336 static void intel_pstate_driver_cleanup(void)
2337 {
2338         unsigned int cpu;
2339
2340         get_online_cpus();
2341         for_each_online_cpu(cpu) {
2342                 if (all_cpu_data[cpu]) {
2343                         if (intel_pstate_driver == &intel_pstate)
2344                                 intel_pstate_clear_update_util_hook(cpu);
2345
2346                         kfree(all_cpu_data[cpu]);
2347                         all_cpu_data[cpu] = NULL;
2348                 }
2349         }
2350         put_online_cpus();
2351 }
2352
2353 static int intel_pstate_register_driver(void)
2354 {
2355         int ret;
2356
2357         intel_pstate_init_limits(&global);
2358
2359         ret = cpufreq_register_driver(intel_pstate_driver);
2360         if (ret) {
2361                 intel_pstate_driver_cleanup();
2362                 return ret;
2363         }
2364
2365         mutex_lock(&intel_pstate_limits_lock);
2366         driver_registered = true;
2367         mutex_unlock(&intel_pstate_limits_lock);
2368
2369         if (intel_pstate_driver == &intel_pstate && !hwp_active &&
2370             pstate_funcs.get_target_pstate != get_target_pstate_use_cpu_load)
2371                 intel_pstate_debug_expose_params();
2372
2373         return 0;
2374 }
2375
2376 static int intel_pstate_unregister_driver(void)
2377 {
2378         if (hwp_active)
2379                 return -EBUSY;
2380
2381         if (intel_pstate_driver == &intel_pstate && !hwp_active &&
2382             pstate_funcs.get_target_pstate != get_target_pstate_use_cpu_load)
2383                 intel_pstate_debug_hide_params();
2384
2385         mutex_lock(&intel_pstate_limits_lock);
2386         driver_registered = false;
2387         mutex_unlock(&intel_pstate_limits_lock);
2388
2389         cpufreq_unregister_driver(intel_pstate_driver);
2390         intel_pstate_driver_cleanup();
2391
2392         return 0;
2393 }
2394
2395 static ssize_t intel_pstate_show_status(char *buf)
2396 {
2397         if (!driver_registered)
2398                 return sprintf(buf, "off\n");
2399
2400         return sprintf(buf, "%s\n", intel_pstate_driver == &intel_pstate ?
2401                                         "active" : "passive");
2402 }
2403
2404 static int intel_pstate_update_status(const char *buf, size_t size)
2405 {
2406         int ret;
2407
2408         if (size == 3 && !strncmp(buf, "off", size))
2409                 return driver_registered ?
2410                         intel_pstate_unregister_driver() : -EINVAL;
2411
2412         if (size == 6 && !strncmp(buf, "active", size)) {
2413                 if (driver_registered) {
2414                         if (intel_pstate_driver == &intel_pstate)
2415                                 return 0;
2416
2417                         ret = intel_pstate_unregister_driver();
2418                         if (ret)
2419                                 return ret;
2420                 }
2421
2422                 intel_pstate_driver = &intel_pstate;
2423                 return intel_pstate_register_driver();
2424         }
2425
2426         if (size == 7 && !strncmp(buf, "passive", size)) {
2427                 if (driver_registered) {
2428                         if (intel_pstate_driver != &intel_pstate)
2429                                 return 0;
2430
2431                         ret = intel_pstate_unregister_driver();
2432                         if (ret)
2433                                 return ret;
2434                 }
2435
2436                 intel_pstate_driver = &intel_cpufreq;
2437                 return intel_pstate_register_driver();
2438         }
2439
2440         return -EINVAL;
2441 }
2442
2443 static int no_load __initdata;
2444 static int no_hwp __initdata;
2445 static int hwp_only __initdata;
2446 static unsigned int force_load __initdata;
2447
2448 static int __init intel_pstate_msrs_not_valid(void)
2449 {
2450         if (!pstate_funcs.get_max() ||
2451             !pstate_funcs.get_min() ||
2452             !pstate_funcs.get_turbo())
2453                 return -ENODEV;
2454
2455         return 0;
2456 }
2457
2458 static void __init copy_pid_params(struct pstate_adjust_policy *policy)
2459 {
2460         pid_params.sample_rate_ms = policy->sample_rate_ms;
2461         pid_params.sample_rate_ns = pid_params.sample_rate_ms * NSEC_PER_MSEC;
2462         pid_params.p_gain_pct = policy->p_gain_pct;
2463         pid_params.i_gain_pct = policy->i_gain_pct;
2464         pid_params.d_gain_pct = policy->d_gain_pct;
2465         pid_params.deadband = policy->deadband;
2466         pid_params.setpoint = policy->setpoint;
2467 }
2468
2469 #ifdef CONFIG_ACPI
2470 static void intel_pstate_use_acpi_profile(void)
2471 {
2472         if (acpi_gbl_FADT.preferred_profile == PM_MOBILE)
2473                 pstate_funcs.get_target_pstate =
2474                                 get_target_pstate_use_cpu_load;
2475 }
2476 #else
2477 static void intel_pstate_use_acpi_profile(void)
2478 {
2479 }
2480 #endif
2481
2482 static void __init copy_cpu_funcs(struct pstate_funcs *funcs)
2483 {
2484         pstate_funcs.get_max   = funcs->get_max;
2485         pstate_funcs.get_max_physical = funcs->get_max_physical;
2486         pstate_funcs.get_min   = funcs->get_min;
2487         pstate_funcs.get_turbo = funcs->get_turbo;
2488         pstate_funcs.get_scaling = funcs->get_scaling;
2489         pstate_funcs.get_val   = funcs->get_val;
2490         pstate_funcs.get_vid   = funcs->get_vid;
2491         pstate_funcs.get_target_pstate = funcs->get_target_pstate;
2492
2493         intel_pstate_use_acpi_profile();
2494 }
2495
2496 #ifdef CONFIG_ACPI
2497
2498 static bool __init intel_pstate_no_acpi_pss(void)
2499 {
2500         int i;
2501
2502         for_each_possible_cpu(i) {
2503                 acpi_status status;
2504                 union acpi_object *pss;
2505                 struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
2506                 struct acpi_processor *pr = per_cpu(processors, i);
2507
2508                 if (!pr)
2509                         continue;
2510
2511                 status = acpi_evaluate_object(pr->handle, "_PSS", NULL, &buffer);
2512                 if (ACPI_FAILURE(status))
2513                         continue;
2514
2515                 pss = buffer.pointer;
2516                 if (pss && pss->type == ACPI_TYPE_PACKAGE) {
2517                         kfree(pss);
2518                         return false;
2519                 }
2520
2521                 kfree(pss);
2522         }
2523
2524         return true;
2525 }
2526
2527 static bool __init intel_pstate_has_acpi_ppc(void)
2528 {
2529         int i;
2530
2531         for_each_possible_cpu(i) {
2532                 struct acpi_processor *pr = per_cpu(processors, i);
2533
2534                 if (!pr)
2535                         continue;
2536                 if (acpi_has_method(pr->handle, "_PPC"))
2537                         return true;
2538         }
2539         return false;
2540 }
2541
2542 enum {
2543         PSS,
2544         PPC,
2545 };
2546
2547 struct hw_vendor_info {
2548         u16  valid;
2549         char oem_id[ACPI_OEM_ID_SIZE];
2550         char oem_table_id[ACPI_OEM_TABLE_ID_SIZE];
2551         int  oem_pwr_table;
2552 };
2553
2554 /* Hardware vendor-specific info that has its own power management modes */
2555 static struct hw_vendor_info vendor_info[] __initdata = {
2556         {1, "HP    ", "ProLiant", PSS},
2557         {1, "ORACLE", "X4-2    ", PPC},
2558         {1, "ORACLE", "X4-2L   ", PPC},
2559         {1, "ORACLE", "X4-2B   ", PPC},
2560         {1, "ORACLE", "X3-2    ", PPC},
2561         {1, "ORACLE", "X3-2L   ", PPC},
2562         {1, "ORACLE", "X3-2B   ", PPC},
2563         {1, "ORACLE", "X4470M2 ", PPC},
2564         {1, "ORACLE", "X4270M3 ", PPC},
2565         {1, "ORACLE", "X4270M2 ", PPC},
2566         {1, "ORACLE", "X4170M2 ", PPC},
2567         {1, "ORACLE", "X4170 M3", PPC},
2568         {1, "ORACLE", "X4275 M3", PPC},
2569         {1, "ORACLE", "X6-2    ", PPC},
2570         {1, "ORACLE", "Sudbury ", PPC},
2571         {0, "", ""},
2572 };
2573
2574 static bool __init intel_pstate_platform_pwr_mgmt_exists(void)
2575 {
2576         struct acpi_table_header hdr;
2577         struct hw_vendor_info *v_info;
2578         const struct x86_cpu_id *id;
2579         u64 misc_pwr;
2580
2581         id = x86_match_cpu(intel_pstate_cpu_oob_ids);
2582         if (id) {
2583                 rdmsrl(MSR_MISC_PWR_MGMT, misc_pwr);
2584                 if ( misc_pwr & (1 << 8))
2585                         return true;
2586         }
2587
2588         if (acpi_disabled ||
2589             ACPI_FAILURE(acpi_get_table_header(ACPI_SIG_FADT, 0, &hdr)))
2590                 return false;
2591
2592         for (v_info = vendor_info; v_info->valid; v_info++) {
2593                 if (!strncmp(hdr.oem_id, v_info->oem_id, ACPI_OEM_ID_SIZE) &&
2594                         !strncmp(hdr.oem_table_id, v_info->oem_table_id,
2595                                                 ACPI_OEM_TABLE_ID_SIZE))
2596                         switch (v_info->oem_pwr_table) {
2597                         case PSS:
2598                                 return intel_pstate_no_acpi_pss();
2599                         case PPC:
2600                                 return intel_pstate_has_acpi_ppc() &&
2601                                         (!force_load);
2602                         }
2603         }
2604
2605         return false;
2606 }
2607
2608 static void intel_pstate_request_control_from_smm(void)
2609 {
2610         /*
2611          * It may be unsafe to request P-states control from SMM if _PPC support
2612          * has not been enabled.
2613          */
2614         if (acpi_ppc)
2615                 acpi_processor_pstate_control();
2616 }
2617 #else /* CONFIG_ACPI not enabled */
2618 static inline bool intel_pstate_platform_pwr_mgmt_exists(void) { return false; }
2619 static inline bool intel_pstate_has_acpi_ppc(void) { return false; }
2620 static inline void intel_pstate_request_control_from_smm(void) {}
2621 #endif /* CONFIG_ACPI */
2622
2623 static const struct x86_cpu_id hwp_support_ids[] __initconst = {
2624         { X86_VENDOR_INTEL, 6, X86_MODEL_ANY, X86_FEATURE_HWP },
2625         {}
2626 };
2627
2628 static int __init intel_pstate_init(void)
2629 {
2630         const struct x86_cpu_id *id;
2631         struct cpu_defaults *cpu_def;
2632         int rc = 0;
2633
2634         if (no_load)
2635                 return -ENODEV;
2636
2637         if (x86_match_cpu(hwp_support_ids) && !no_hwp) {
2638                 copy_cpu_funcs(&core_params.funcs);
2639                 hwp_active++;
2640                 intel_pstate.attr = hwp_cpufreq_attrs;
2641                 goto hwp_cpu_matched;
2642         }
2643
2644         id = x86_match_cpu(intel_pstate_cpu_ids);
2645         if (!id)
2646                 return -ENODEV;
2647
2648         cpu_def = (struct cpu_defaults *)id->driver_data;
2649
2650         copy_pid_params(&cpu_def->pid_policy);
2651         copy_cpu_funcs(&cpu_def->funcs);
2652
2653         if (intel_pstate_msrs_not_valid())
2654                 return -ENODEV;
2655
2656 hwp_cpu_matched:
2657         /*
2658          * The Intel pstate driver will be ignored if the platform
2659          * firmware has its own power management modes.
2660          */
2661         if (intel_pstate_platform_pwr_mgmt_exists())
2662                 return -ENODEV;
2663
2664         if (!hwp_active && hwp_only)
2665                 return -ENOTSUPP;
2666
2667         pr_info("Intel P-state driver initializing\n");
2668
2669         all_cpu_data = vzalloc(sizeof(void *) * num_possible_cpus());
2670         if (!all_cpu_data)
2671                 return -ENOMEM;
2672
2673         intel_pstate_request_control_from_smm();
2674
2675         intel_pstate_sysfs_expose_params();
2676
2677         mutex_lock(&intel_pstate_driver_lock);
2678         rc = intel_pstate_register_driver();
2679         mutex_unlock(&intel_pstate_driver_lock);
2680         if (rc)
2681                 return rc;
2682
2683         if (hwp_active)
2684                 pr_info("HWP enabled\n");
2685
2686         return 0;
2687 }
2688 device_initcall(intel_pstate_init);
2689
2690 static int __init intel_pstate_setup(char *str)
2691 {
2692         if (!str)
2693                 return -EINVAL;
2694
2695         if (!strcmp(str, "disable")) {
2696                 no_load = 1;
2697         } else if (!strcmp(str, "passive")) {
2698                 pr_info("Passive mode enabled\n");
2699                 intel_pstate_driver = &intel_cpufreq;
2700                 no_hwp = 1;
2701         }
2702         if (!strcmp(str, "no_hwp")) {
2703                 pr_info("HWP disabled\n");
2704                 no_hwp = 1;
2705         }
2706         if (!strcmp(str, "force"))
2707                 force_load = 1;
2708         if (!strcmp(str, "hwp_only"))
2709                 hwp_only = 1;
2710         if (!strcmp(str, "per_cpu_perf_limits"))
2711                 per_cpu_limits = true;
2712
2713 #ifdef CONFIG_ACPI
2714         if (!strcmp(str, "support_acpi_ppc"))
2715                 acpi_ppc = true;
2716 #endif
2717
2718         return 0;
2719 }
2720 early_param("intel_pstate", intel_pstate_setup);
2721
2722 MODULE_AUTHOR("Dirk Brandewie <dirk.j.brandewie@intel.com>");
2723 MODULE_DESCRIPTION("'intel_pstate' - P state driver Intel Core processors");
2724 MODULE_LICENSE("GPL");