drivers/cpufreq/intel_pstate.c

   1 // SPDX-License-Identifier: GPL-2.0-only
   2 /*
   3  * intel_pstate.c: Native P state management for Intel processors
   4  *
   5  * (C) Copyright 2012 Intel Corporation
   6  * Author: Dirk Brandewie <dirk.j.brandewie@intel.com>
   7  */
   8
   9 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  10
  11 #include <linux/kernel.h>
  12 #include <linux/kernel_stat.h>
  13 #include <linux/module.h>
  14 #include <linux/ktime.h>
  15 #include <linux/hrtimer.h>
  16 #include <linux/tick.h>
  17 #include <linux/slab.h>
  18 #include <linux/sched/cpufreq.h>
  19 #include <linux/list.h>
  20 #include <linux/cpu.h>
  21 #include <linux/cpufreq.h>
  22 #include <linux/sysfs.h>
  23 #include <linux/types.h>
  24 #include <linux/fs.h>
  25 #include <linux/acpi.h>
  26 #include <linux/vmalloc.h>
  27 #include <linux/pm_qos.h>
  28 #include <trace/events/power.h>
  29
  30 #include <asm/div64.h>
  31 #include <asm/msr.h>
  32 #include <asm/cpu_device_id.h>
  33 #include <asm/cpufeature.h>
  34 #include <asm/intel-family.h>
  35
  36 #define INTEL_PSTATE_SAMPLING_INTERVAL  (10 * NSEC_PER_MSEC)
  37
  38 #define INTEL_CPUFREQ_TRANSITION_LATENCY        20000
  39 #define INTEL_CPUFREQ_TRANSITION_DELAY          500
  40
  41 #ifdef CONFIG_ACPI
  42 #include <acpi/processor.h>
  43 #include <acpi/cppc_acpi.h>
  44 #endif
  45
  46 #define FRAC_BITS 8
  47 #define int_tofp(X) ((int64_t)(X) << FRAC_BITS)
  48 #define fp_toint(X) ((X) >> FRAC_BITS)
  49
  50 #define ONE_EIGHTH_FP ((int64_t)1 << (FRAC_BITS - 3))
  51
  52 #define EXT_BITS 6
  53 #define EXT_FRAC_BITS (EXT_BITS + FRAC_BITS)
  54 #define fp_ext_toint(X) ((X) >> EXT_FRAC_BITS)
  55 #define int_ext_tofp(X) ((int64_t)(X) << EXT_FRAC_BITS)
  56
  57 static inline int32_t mul_fp(int32_t x, int32_t y)
  58 {
  59         return ((int64_t)x * (int64_t)y) >> FRAC_BITS;
  60 }
  61
  62 static inline int32_t div_fp(s64 x, s64 y)
  63 {
  64         return div64_s64((int64_t)x << FRAC_BITS, y);
  65 }
  66
  67 static inline int ceiling_fp(int32_t x)
  68 {
  69         int mask, ret;
  70
  71         ret = fp_toint(x);
  72         mask = (1 << FRAC_BITS) - 1;
  73         if (x & mask)
  74                 ret += 1;
  75         return ret;
  76 }
  77
  78 static inline int32_t percent_fp(int percent)
  79 {
  80         return div_fp(percent, 100);
  81 }
  82
  83 static inline u64 mul_ext_fp(u64 x, u64 y)
  84 {
  85         return (x * y) >> EXT_FRAC_BITS;
  86 }
  87
  88 static inline u64 div_ext_fp(u64 x, u64 y)
  89 {
  90         return div64_u64(x << EXT_FRAC_BITS, y);
  91 }
  92
  93 static inline int32_t percent_ext_fp(int percent)
  94 {
  95         return div_ext_fp(percent, 100);
  96 }
  97
  98 /**
  99  * struct sample -      Store performance sample
 100  * @core_avg_perf:      Ratio of APERF/MPERF which is the actual average
 101  *                      performance during last sample period
 102  * @busy_scaled:        Scaled busy value which is used to calculate next
 103  *                      P state. This can be different than core_avg_perf
 104  *                      to account for cpu idle period
 105  * @aperf:              Difference of actual performance frequency clock count
 106  *                      read from APERF MSR between last and current sample
 107  * @mperf:              Difference of maximum performance frequency clock count
 108  *                      read from MPERF MSR between last and current sample
 109  * @tsc:                Difference of time stamp counter between last and
 110  *                      current sample
 111  * @time:               Current time from scheduler
 112  *
 113  * This structure is used in the cpudata structure to store performance sample
 114  * data for choosing next P State.
 115  */
 116 struct sample {
 117         int32_t core_avg_perf;
 118         int32_t busy_scaled;
 119         u64 aperf;
 120         u64 mperf;
 121         u64 tsc;
 122         u64 time;
 123 };
 124
 125 /**
 126  * struct pstate_data - Store P state data
 127  * @current_pstate:     Current requested P state
 128  * @min_pstate:         Min P state possible for this platform
 129  * @max_pstate:         Max P state possible for this platform
 130  * @max_pstate_physical:This is physical Max P state for a processor
 131  *                      This can be higher than the max_pstate which can
 132  *                      be limited by platform thermal design power limits
 133  * @scaling:            Scaling factor to  convert frequency to cpufreq
 134  *                      frequency units
 135  * @turbo_pstate:       Max Turbo P state possible for this platform
 136  * @max_freq:           @max_pstate frequency in cpufreq units
 137  * @turbo_freq:         @turbo_pstate frequency in cpufreq units
 138  *
 139  * Stores the per cpu model P state limits and current P state.
 140  */
 141 struct pstate_data {
 142         int     current_pstate;
 143         int     min_pstate;
 144         int     max_pstate;
 145         int     max_pstate_physical;
 146         int     scaling;
 147         int     turbo_pstate;
 148         unsigned int max_freq;
 149         unsigned int turbo_freq;
 150 };
 151
 152 /**
 153  * struct vid_data -    Stores voltage information data
 154  * @min:                VID data for this platform corresponding to
 155  *                      the lowest P state
 156  * @max:                VID data corresponding to the highest P State.
 157  * @turbo:              VID data for turbo P state
 158  * @ratio:              Ratio of (vid max - vid min) /
 159  *                      (max P state - Min P State)
 160  *
 161  * Stores the voltage data for DVFS (Dynamic Voltage and Frequency Scaling)
 162  * This data is used in Atom platforms, where in addition to target P state,
 163  * the voltage data needs to be specified to select next P State.
 164  */
 165 struct vid_data {
 166         int min;
 167         int max;
 168         int turbo;
 169         int32_t ratio;
 170 };
 171
 172 /**
 173  * struct global_params - Global parameters, mostly tunable via sysfs.
 174  * @no_turbo:           Whether or not to use turbo P-states.
 175  * @turbo_disabled:     Whether or not turbo P-states are available at all,
 176  *                      based on the MSR_IA32_MISC_ENABLE value and whether or
 177  *                      not the maximum reported turbo P-state is different from
 178  *                      the maximum reported non-turbo one.
 179  * @turbo_disabled_mf:  The @turbo_disabled value reflected by cpuinfo.max_freq.
 180  * @min_perf_pct:       Minimum capacity limit in percent of the maximum turbo
 181  *                      P-state capacity.
 182  * @max_perf_pct:       Maximum capacity limit in percent of the maximum turbo
 183  *                      P-state capacity.
 184  */
 185 struct global_params {
 186         bool no_turbo;
 187         bool turbo_disabled;
 188         bool turbo_disabled_mf;
 189         int max_perf_pct;
 190         int min_perf_pct;
 191 };
 192
 193 /**
 194  * struct cpudata -     Per CPU instance data storage
 195  * @cpu:                CPU number for this instance data
 196  * @policy:             CPUFreq policy value
 197  * @update_util:        CPUFreq utility callback information
 198  * @update_util_set:    CPUFreq utility callback is set
 199  * @iowait_boost:       iowait-related boost fraction
 200  * @last_update:        Time of the last update.
 201  * @pstate:             Stores P state limits for this CPU
 202  * @vid:                Stores VID limits for this CPU
 203  * @last_sample_time:   Last Sample time
 204  * @aperf_mperf_shift:  Number of clock cycles after aperf, merf is incremented
 205  *                      This shift is a multiplier to mperf delta to
 206  *                      calculate CPU busy.
 207  * @prev_aperf:         Last APERF value read from APERF MSR
 208  * @prev_mperf:         Last MPERF value read from MPERF MSR
 209  * @prev_tsc:           Last timestamp counter (TSC) value
 210  * @prev_cummulative_iowait: IO Wait time difference from last and
 211  *                      current sample
 212  * @sample:             Storage for storing last Sample data
 213  * @min_perf_ratio:     Minimum capacity in terms of PERF or HWP ratios
 214  * @max_perf_ratio:     Maximum capacity in terms of PERF or HWP ratios
 215  * @acpi_perf_data:     Stores ACPI perf information read from _PSS
 216  * @valid_pss_table:    Set to true for valid ACPI _PSS entries found
 217  * @epp_powersave:      Last saved HWP energy performance preference
 218  *                      (EPP) or energy performance bias (EPB),
 219  *                      when policy switched to performance
 220  * @epp_policy:         Last saved policy used to set EPP/EPB
 221  * @epp_default:        Power on default HWP energy performance
 222  *                      preference/bias
 223  * @epp_saved:          Saved EPP/EPB during system suspend or CPU offline
 224  *                      operation
 225  * @hwp_req_cached:     Cached value of the last HWP Request MSR
 226  * @hwp_cap_cached:     Cached value of the last HWP Capabilities MSR
 227  * @last_io_update:     Last time when IO wake flag was set
 228  * @sched_flags:        Store scheduler flags for possible cross CPU update
 229  * @hwp_boost_min:      Last HWP boosted min performance
 230  *
 231  * This structure stores per CPU instance data for all CPUs.
 232  */
 233 struct cpudata {
 234         int cpu;
 235
 236         unsigned int policy;
 237         struct update_util_data update_util;
 238         bool   update_util_set;
 239
 240         struct pstate_data pstate;
 241         struct vid_data vid;
 242
 243         u64     last_update;
 244         u64     last_sample_time;
 245         u64     aperf_mperf_shift;
 246         u64     prev_aperf;
 247         u64     prev_mperf;
 248         u64     prev_tsc;
 249         u64     prev_cummulative_iowait;
 250         struct sample sample;
 251         int32_t min_perf_ratio;
 252         int32_t max_perf_ratio;
 253 #ifdef CONFIG_ACPI
 254         struct acpi_processor_performance acpi_perf_data;
 255         bool valid_pss_table;
 256 #endif
 257         unsigned int iowait_boost;
 258         s16 epp_powersave;
 259         s16 epp_policy;
 260         s16 epp_default;
 261         s16 epp_saved;
 262         u64 hwp_req_cached;
 263         u64 hwp_cap_cached;
 264         u64 last_io_update;
 265         unsigned int sched_flags;
 266         u32 hwp_boost_min;
 267 };
 268
 269 static struct cpudata **all_cpu_data;
 270
 271 /**
 272  * struct pstate_funcs - Per CPU model specific callbacks
 273  * @get_max:            Callback to get maximum non turbo effective P state
 274  * @get_max_physical:   Callback to get maximum non turbo physical P state
 275  * @get_min:            Callback to get minimum P state
 276  * @get_turbo:          Callback to get turbo P state
 277  * @get_scaling:        Callback to get frequency scaling factor
 278  * @get_val:            Callback to convert P state to actual MSR write value
 279  * @get_vid:            Callback to get VID data for Atom platforms
 280  *
 281  * Core and Atom CPU models have different way to get P State limits. This
 282  * structure is used to store those callbacks.
 283  */
 284 struct pstate_funcs {
 285         int (*get_max)(void);
 286         int (*get_max_physical)(void);
 287         int (*get_min)(void);
 288         int (*get_turbo)(void);
 289         int (*get_scaling)(void);
 290         int (*get_aperf_mperf_shift)(void);
 291         u64 (*get_val)(struct cpudata*, int pstate);
 292         void (*get_vid)(struct cpudata *);
 293 };
 294
 295 static struct pstate_funcs pstate_funcs __read_mostly;
 296
 297 static int hwp_active __read_mostly;
 298 static int hwp_mode_bdw __read_mostly;
 299 static bool per_cpu_limits __read_mostly;
 300 static bool hwp_boost __read_mostly;
 301
 302 static struct cpufreq_driver *intel_pstate_driver __read_mostly;
 303
 304 #ifdef CONFIG_ACPI
 305 static bool acpi_ppc;
 306 #endif
 307
 308 static struct global_params global;
 309
 310 static DEFINE_MUTEX(intel_pstate_driver_lock);
 311 static DEFINE_MUTEX(intel_pstate_limits_lock);
 312
 313 #ifdef CONFIG_ACPI
 314
 315 static bool intel_pstate_acpi_pm_profile_server(void)
 316 {
 317         if (acpi_gbl_FADT.preferred_profile == PM_ENTERPRISE_SERVER ||
 318             acpi_gbl_FADT.preferred_profile == PM_PERFORMANCE_SERVER)
 319                 return true;
 320
 321         return false;
 322 }
 323
 324 static bool intel_pstate_get_ppc_enable_status(void)
 325 {
 326         if (intel_pstate_acpi_pm_profile_server())
 327                 return true;
 328
 329         return acpi_ppc;
 330 }
 331
 332 #ifdef CONFIG_ACPI_CPPC_LIB
 333
 334 /* The work item is needed to avoid CPU hotplug locking issues */
 335 static void intel_pstste_sched_itmt_work_fn(struct work_struct *work)
 336 {
 337         sched_set_itmt_support();
 338 }
 339
 340 static DECLARE_WORK(sched_itmt_work, intel_pstste_sched_itmt_work_fn);
 341
 342 static void intel_pstate_set_itmt_prio(int cpu)
 343 {
 344         struct cppc_perf_caps cppc_perf;
 345         static u32 max_highest_perf = 0, min_highest_perf = U32_MAX;
 346         int ret;
 347
 348         ret = cppc_get_perf_caps(cpu, &cppc_perf);
 349         if (ret)
 350                 return;
 351
 352         /*
 353          * The priorities can be set regardless of whether or not
 354          * sched_set_itmt_support(true) has been called and it is valid to
 355          * update them at any time after it has been called.
 356          */
 357         sched_set_itmt_core_prio(cppc_perf.highest_perf, cpu);
 358
 359         if (max_highest_perf <= min_highest_perf) {
 360                 if (cppc_perf.highest_perf > max_highest_perf)
 361                         max_highest_perf = cppc_perf.highest_perf;
 362
 363                 if (cppc_perf.highest_perf < min_highest_perf)
 364                         min_highest_perf = cppc_perf.highest_perf;
 365
 366                 if (max_highest_perf > min_highest_perf) {
 367                         /*
 368                          * This code can be run during CPU online under the
 369                          * CPU hotplug locks, so sched_set_itmt_support()
 370                          * cannot be called from here.  Queue up a work item
 371                          * to invoke it.
 372                          */
 373                         schedule_work(&sched_itmt_work);
 374                 }
 375         }
 376 }
 377
 378 static int intel_pstate_get_cppc_guranteed(int cpu)
 379 {
 380         struct cppc_perf_caps cppc_perf;
 381         int ret;
 382
 383         ret = cppc_get_perf_caps(cpu, &cppc_perf);
 384         if (ret)
 385                 return ret;
 386
 387         if (cppc_perf.guaranteed_perf)
 388                 return cppc_perf.guaranteed_perf;
 389
 390         return cppc_perf.nominal_perf;
 391 }
 392
 393 #else /* CONFIG_ACPI_CPPC_LIB */
 394 static void intel_pstate_set_itmt_prio(int cpu)
 395 {
 396 }
 397 #endif /* CONFIG_ACPI_CPPC_LIB */
 398
 399 static void intel_pstate_init_acpi_perf_limits(struct cpufreq_policy *policy)
 400 {
 401         struct cpudata *cpu;
 402         int ret;
 403         int i;
 404
 405         if (hwp_active) {
 406                 intel_pstate_set_itmt_prio(policy->cpu);
 407                 return;
 408         }
 409
 410         if (!intel_pstate_get_ppc_enable_status())
 411                 return;
 412
 413         cpu = all_cpu_data[policy->cpu];
 414
 415         ret = acpi_processor_register_performance(&cpu->acpi_perf_data,
 416                                                   policy->cpu);
 417         if (ret)
 418                 return;
 419
 420         /*
 421          * Check if the control value in _PSS is for PERF_CTL MSR, which should
 422          * guarantee that the states returned by it map to the states in our
 423          * list directly.
 424          */
 425         if (cpu->acpi_perf_data.control_register.space_id !=
 426                                                 ACPI_ADR_SPACE_FIXED_HARDWARE)
 427                 goto err;
 428
 429         /*
 430          * If there is only one entry _PSS, simply ignore _PSS and continue as
 431          * usual without taking _PSS into account
 432          */
 433         if (cpu->acpi_perf_data.state_count < 2)
 434                 goto err;
 435
 436         pr_debug("CPU%u - ACPI _PSS perf data\n", policy->cpu);
 437         for (i = 0; i < cpu->acpi_perf_data.state_count; i++) {
 438                 pr_debug("     %cP%d: %u MHz, %u mW, 0x%x\n",
 439                          (i == cpu->acpi_perf_data.state ? '*' : ' '), i,
 440                          (u32) cpu->acpi_perf_data.states[i].core_frequency,
 441                          (u32) cpu->acpi_perf_data.states[i].power,
 442                          (u32) cpu->acpi_perf_data.states[i].control);
 443         }
 444
 445         /*
 446          * The _PSS table doesn't contain whole turbo frequency range.
 447          * This just contains +1 MHZ above the max non turbo frequency,
 448          * with control value corresponding to max turbo ratio. But
 449          * when cpufreq set policy is called, it will call with this
 450          * max frequency, which will cause a reduced performance as
 451          * this driver uses real max turbo frequency as the max
 452          * frequency. So correct this frequency in _PSS table to
 453          * correct max turbo frequency based on the turbo state.
 454          * Also need to convert to MHz as _PSS freq is in MHz.
 455          */
 456         if (!global.turbo_disabled)
 457                 cpu->acpi_perf_data.states[0].core_frequency =
 458                                         policy->cpuinfo.max_freq / 1000;
 459         cpu->valid_pss_table = true;
 460         pr_debug("_PPC limits will be enforced\n");
 461
 462         return;
 463
 464  err:
 465         cpu->valid_pss_table = false;
 466         acpi_processor_unregister_performance(policy->cpu);
 467 }
 468
 469 static void intel_pstate_exit_perf_limits(struct cpufreq_policy *policy)
 470 {
 471         struct cpudata *cpu;
 472
 473         cpu = all_cpu_data[policy->cpu];
 474         if (!cpu->valid_pss_table)
 475                 return;
 476
 477         acpi_processor_unregister_performance(policy->cpu);
 478 }
 479 #else /* CONFIG_ACPI */
 480 static inline void intel_pstate_init_acpi_perf_limits(struct cpufreq_policy *policy)
 481 {
 482 }
 483
 484 static inline void intel_pstate_exit_perf_limits(struct cpufreq_policy *policy)
 485 {
 486 }
 487
 488 static inline bool intel_pstate_acpi_pm_profile_server(void)
 489 {
 490         return false;
 491 }
 492 #endif /* CONFIG_ACPI */
 493
 494 #ifndef CONFIG_ACPI_CPPC_LIB
 495 static int intel_pstate_get_cppc_guranteed(int cpu)
 496 {
 497         return -ENOTSUPP;
 498 }
 499 #endif /* CONFIG_ACPI_CPPC_LIB */
 500
 501 static inline void update_turbo_state(void)
 502 {
 503         u64 misc_en;
 504         struct cpudata *cpu;
 505
 506         cpu = all_cpu_data[0];
 507         rdmsrl(MSR_IA32_MISC_ENABLE, misc_en);
 508         global.turbo_disabled =
 509                 (misc_en & MSR_IA32_MISC_ENABLE_TURBO_DISABLE ||
 510                  cpu->pstate.max_pstate == cpu->pstate.turbo_pstate);
 511 }
 512
 513 static int min_perf_pct_min(void)
 514 {
 515         struct cpudata *cpu = all_cpu_data[0];
 516         int turbo_pstate = cpu->pstate.turbo_pstate;
 517
 518         return turbo_pstate ?
 519                 (cpu->pstate.min_pstate * 100 / turbo_pstate) : 0;
 520 }
 521
 522 static s16 intel_pstate_get_epb(struct cpudata *cpu_data)
 523 {
 524         u64 epb;
 525         int ret;
 526
 527         if (!boot_cpu_has(X86_FEATURE_EPB))
 528                 return -ENXIO;
 529
 530         ret = rdmsrl_on_cpu(cpu_data->cpu, MSR_IA32_ENERGY_PERF_BIAS, &epb);
 531         if (ret)
 532                 return (s16)ret;
 533
 534         return (s16)(epb & 0x0f);
 535 }
 536
 537 static s16 intel_pstate_get_epp(struct cpudata *cpu_data, u64 hwp_req_data)
 538 {
 539         s16 epp;
 540
 541         if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
 542                 /*
 543                  * When hwp_req_data is 0, means that caller didn't read
 544                  * MSR_HWP_REQUEST, so need to read and get EPP.
 545                  */
 546                 if (!hwp_req_data) {
 547                         epp = rdmsrl_on_cpu(cpu_data->cpu, MSR_HWP_REQUEST,
 548                                             &hwp_req_data);
 549                         if (epp)
 550                                 return epp;
 551                 }
 552                 epp = (hwp_req_data >> 24) & 0xff;
 553         } else {
 554                 /* When there is no EPP present, HWP uses EPB settings */
 555                 epp = intel_pstate_get_epb(cpu_data);
 556         }
 557
 558         return epp;
 559 }
 560
 561 static int intel_pstate_set_epb(int cpu, s16 pref)
 562 {
 563         u64 epb;
 564         int ret;
 565
 566         if (!boot_cpu_has(X86_FEATURE_EPB))
 567                 return -ENXIO;
 568
 569         ret = rdmsrl_on_cpu(cpu, MSR_IA32_ENERGY_PERF_BIAS, &epb);
 570         if (ret)
 571                 return ret;
 572
 573         epb = (epb & ~0x0f) | pref;
 574         wrmsrl_on_cpu(cpu, MSR_IA32_ENERGY_PERF_BIAS, epb);
 575
 576         return 0;
 577 }
 578
 579 /*
 580  * EPP/EPB display strings corresponding to EPP index in the
 581  * energy_perf_strings[]
 582  *      index           String
 583  *-------------------------------------
 584  *      0               default
 585  *      1               performance
 586  *      2               balance_performance
 587  *      3               balance_power
 588  *      4               power
 589  */
 590 static const char * const energy_perf_strings[] = {
 591         "default",
 592         "performance",
 593         "balance_performance",
 594         "balance_power",
 595         "power",
 596         NULL
 597 };
 598 static const unsigned int epp_values[] = {
 599         HWP_EPP_PERFORMANCE,
 600         HWP_EPP_BALANCE_PERFORMANCE,
 601         HWP_EPP_BALANCE_POWERSAVE,
 602         HWP_EPP_POWERSAVE
 603 };
 604
 605 static int intel_pstate_get_energy_pref_index(struct cpudata *cpu_data)
 606 {
 607         s16 epp;
 608         int index = -EINVAL;
 609
 610         epp = intel_pstate_get_epp(cpu_data, 0);
 611         if (epp < 0)
 612                 return epp;
 613
 614         if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
 615                 if (epp == HWP_EPP_PERFORMANCE)
 616                         return 1;
 617                 if (epp <= HWP_EPP_BALANCE_PERFORMANCE)
 618                         return 2;
 619                 if (epp <= HWP_EPP_BALANCE_POWERSAVE)
 620                         return 3;
 621                 else
 622                         return 4;
 623         } else if (boot_cpu_has(X86_FEATURE_EPB)) {
 624                 /*
 625                  * Range:
 626                  *      0x00-0x03       :       Performance
 627                  *      0x04-0x07       :       Balance performance
 628                  *      0x08-0x0B       :       Balance power
 629                  *      0x0C-0x0F       :       Power
 630                  * The EPB is a 4 bit value, but our ranges restrict the
 631                  * value which can be set. Here only using top two bits
 632                  * effectively.
 633                  */
 634                 index = (epp >> 2) + 1;
 635         }
 636
 637         return index;
 638 }
 639
 640 static int intel_pstate_set_energy_pref_index(struct cpudata *cpu_data,
 641                                               int pref_index)
 642 {
 643         int epp = -EINVAL;
 644         int ret;
 645
 646         if (!pref_index)
 647                 epp = cpu_data->epp_default;
 648
 649         mutex_lock(&intel_pstate_limits_lock);
 650
 651         if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
 652                 u64 value;
 653
 654                 ret = rdmsrl_on_cpu(cpu_data->cpu, MSR_HWP_REQUEST, &value);
 655                 if (ret)
 656                         goto return_pref;
 657
 658                 value &= ~GENMASK_ULL(31, 24);
 659
 660                 if (epp == -EINVAL)
 661                         epp = epp_values[pref_index - 1];
 662
 663                 value |= (u64)epp << 24;
 664                 ret = wrmsrl_on_cpu(cpu_data->cpu, MSR_HWP_REQUEST, value);
 665         } else {
 666                 if (epp == -EINVAL)
 667                         epp = (pref_index - 1) << 2;
 668                 ret = intel_pstate_set_epb(cpu_data->cpu, epp);
 669         }
 670 return_pref:
 671         mutex_unlock(&intel_pstate_limits_lock);
 672
 673         return ret;
 674 }
 675
 676 static ssize_t show_energy_performance_available_preferences(
 677                                 struct cpufreq_policy *policy, char *buf)
 678 {
 679         int i = 0;
 680         int ret = 0;
 681
 682         while (energy_perf_strings[i] != NULL)
 683                 ret += sprintf(&buf[ret], "%s ", energy_perf_strings[i++]);
 684
 685         ret += sprintf(&buf[ret], "\n");
 686
 687         return ret;
 688 }
 689
 690 cpufreq_freq_attr_ro(energy_performance_available_preferences);
 691
 692 static ssize_t store_energy_performance_preference(
 693                 struct cpufreq_policy *policy, const char *buf, size_t count)
 694 {
 695         struct cpudata *cpu_data = all_cpu_data[policy->cpu];
 696         char str_preference[21];
 697         int ret;
 698
 699         ret = sscanf(buf, "%20s", str_preference);
 700         if (ret != 1)
 701                 return -EINVAL;
 702
 703         ret = match_string(energy_perf_strings, -1, str_preference);
 704         if (ret < 0)
 705                 return ret;
 706
 707         intel_pstate_set_energy_pref_index(cpu_data, ret);
 708         return count;
 709 }
 710
 711 static ssize_t show_energy_performance_preference(
 712                                 struct cpufreq_policy *policy, char *buf)
 713 {
 714         struct cpudata *cpu_data = all_cpu_data[policy->cpu];
 715         int preference;
 716
 717         preference = intel_pstate_get_energy_pref_index(cpu_data);
 718         if (preference < 0)
 719                 return preference;
 720
 721         return  sprintf(buf, "%s\n", energy_perf_strings[preference]);
 722 }
 723
 724 cpufreq_freq_attr_rw(energy_performance_preference);
 725
 726 static ssize_t show_base_frequency(struct cpufreq_policy *policy, char *buf)
 727 {
 728         struct cpudata *cpu;
 729         u64 cap;
 730         int ratio;
 731
 732         ratio = intel_pstate_get_cppc_guranteed(policy->cpu);
 733         if (ratio <= 0) {
 734                 rdmsrl_on_cpu(policy->cpu, MSR_HWP_CAPABILITIES, &cap);
 735                 ratio = HWP_GUARANTEED_PERF(cap);
 736         }
 737
 738         cpu = all_cpu_data[policy->cpu];
 739
 740         return sprintf(buf, "%d\n", ratio * cpu->pstate.scaling);
 741 }
 742
 743 cpufreq_freq_attr_ro(base_frequency);
 744
 745 static struct freq_attr *hwp_cpufreq_attrs[] = {
 746         &energy_performance_preference,
 747         &energy_performance_available_preferences,
 748         &base_frequency,
 749         NULL,
 750 };
 751
 752 static void intel_pstate_get_hwp_max(unsigned int cpu, int *phy_max,
 753                                      int *current_max)
 754 {
 755         u64 cap;
 756
 757         rdmsrl_on_cpu(cpu, MSR_HWP_CAPABILITIES, &cap);
 758         WRITE_ONCE(all_cpu_data[cpu]->hwp_cap_cached, cap);
 759         if (global.no_turbo)
 760                 *current_max = HWP_GUARANTEED_PERF(cap);
 761         else
 762                 *current_max = HWP_HIGHEST_PERF(cap);
 763
 764         *phy_max = HWP_HIGHEST_PERF(cap);
 765 }
 766
 767 static void intel_pstate_hwp_set(unsigned int cpu)
 768 {
 769         struct cpudata *cpu_data = all_cpu_data[cpu];
 770         int max, min;
 771         u64 value;
 772         s16 epp;
 773
 774         max = cpu_data->max_perf_ratio;
 775         min = cpu_data->min_perf_ratio;
 776
 777         if (cpu_data->policy == CPUFREQ_POLICY_PERFORMANCE)
 778                 min = max;
 779
 780         rdmsrl_on_cpu(cpu, MSR_HWP_REQUEST, &value);
 781
 782         value &= ~HWP_MIN_PERF(~0L);
 783         value |= HWP_MIN_PERF(min);
 784
 785         value &= ~HWP_MAX_PERF(~0L);
 786         value |= HWP_MAX_PERF(max);
 787
 788         if (cpu_data->epp_policy == cpu_data->policy)
 789                 goto skip_epp;
 790
 791         cpu_data->epp_policy = cpu_data->policy;
 792
 793         if (cpu_data->epp_saved >= 0) {
 794                 epp = cpu_data->epp_saved;
 795                 cpu_data->epp_saved = -EINVAL;
 796                 goto update_epp;
 797         }
 798
 799         if (cpu_data->policy == CPUFREQ_POLICY_PERFORMANCE) {
 800                 epp = intel_pstate_get_epp(cpu_data, value);
 801                 cpu_data->epp_powersave = epp;
 802                 /* If EPP read was failed, then don't try to write */
 803                 if (epp < 0)
 804                         goto skip_epp;
 805
 806                 epp = 0;
 807         } else {
 808                 /* skip setting EPP, when saved value is invalid */
 809                 if (cpu_data->epp_powersave < 0)
 810                         goto skip_epp;
 811
 812                 /*
 813                  * No need to restore EPP when it is not zero. This
 814                  * means:
 815                  *  - Policy is not changed
 816                  *  - user has manually changed
 817                  *  - Error reading EPB
 818                  */
 819                 epp = intel_pstate_get_epp(cpu_data, value);
 820                 if (epp)
 821                         goto skip_epp;
 822
 823                 epp = cpu_data->epp_powersave;
 824         }
 825 update_epp:
 826         if (boot_cpu_has(X86_FEATURE_HWP_EPP)) {
 827                 value &= ~GENMASK_ULL(31, 24);
 828                 value |= (u64)epp << 24;
 829         } else {
 830                 intel_pstate_set_epb(cpu, epp);
 831         }
 832 skip_epp:
 833         WRITE_ONCE(cpu_data->hwp_req_cached, value);
 834         wrmsrl_on_cpu(cpu, MSR_HWP_REQUEST, value);
 835 }
 836
 837 static void intel_pstate_hwp_force_min_perf(int cpu)
 838 {
 839         u64 value;
 840         int min_perf;
 841
 842         value = all_cpu_data[cpu]->hwp_req_cached;
 843         value &= ~GENMASK_ULL(31, 0);
 844         min_perf = HWP_LOWEST_PERF(all_cpu_data[cpu]->hwp_cap_cached);
 845
 846         /* Set hwp_max = hwp_min */
 847         value |= HWP_MAX_PERF(min_perf);
 848         value |= HWP_MIN_PERF(min_perf);
 849
 850         /* Set EPP to min */
 851         if (boot_cpu_has(X86_FEATURE_HWP_EPP))
 852                 value |= HWP_ENERGY_PERF_PREFERENCE(HWP_EPP_POWERSAVE);
 853
 854         wrmsrl_on_cpu(cpu, MSR_HWP_REQUEST, value);
 855 }
 856
 857 static int intel_pstate_hwp_save_state(struct cpufreq_policy *policy)
 858 {
 859         struct cpudata *cpu_data = all_cpu_data[policy->cpu];
 860
 861         if (!hwp_active)
 862                 return 0;
 863
 864         cpu_data->epp_saved = intel_pstate_get_epp(cpu_data, 0);
 865
 866         return 0;
 867 }
 868
 869 static void intel_pstate_hwp_enable(struct cpudata *cpudata);
 870
 871 static int intel_pstate_resume(struct cpufreq_policy *policy)
 872 {
 873         if (!hwp_active)
 874                 return 0;
 875
 876         mutex_lock(&intel_pstate_limits_lock);
 877
 878         if (policy->cpu == 0)
 879                 intel_pstate_hwp_enable(all_cpu_data[policy->cpu]);
 880
 881         all_cpu_data[policy->cpu]->epp_policy = 0;
 882         intel_pstate_hwp_set(policy->cpu);
 883
 884         mutex_unlock(&intel_pstate_limits_lock);
 885
 886         return 0;
 887 }
 888
 889 static void intel_pstate_update_policies(void)
 890 {
 891         int cpu;
 892
 893         for_each_possible_cpu(cpu)
 894                 cpufreq_update_policy(cpu);
 895 }
 896
 897 static void intel_pstate_update_max_freq(unsigned int cpu)
 898 {
 899         struct cpufreq_policy *policy = cpufreq_cpu_acquire(cpu);
 900         struct cpudata *cpudata;
 901
 902         if (!policy)
 903                 return;
 904
 905         cpudata = all_cpu_data[cpu];
 906         policy->cpuinfo.max_freq = global.turbo_disabled_mf ?
 907                         cpudata->pstate.max_freq : cpudata->pstate.turbo_freq;
 908
 909         refresh_frequency_limits(policy);
 910
 911         cpufreq_cpu_release(policy);
 912 }
 913
 914 static void intel_pstate_update_limits(unsigned int cpu)
 915 {
 916         mutex_lock(&intel_pstate_driver_lock);
 917
 918         update_turbo_state();
 919         /*
 920          * If turbo has been turned on or off globally, policy limits for
 921          * all CPUs need to be updated to reflect that.
 922          */
 923         if (global.turbo_disabled_mf != global.turbo_disabled) {
 924                 global.turbo_disabled_mf = global.turbo_disabled;
 925                 for_each_possible_cpu(cpu)
 926                         intel_pstate_update_max_freq(cpu);
 927         } else {
 928                 cpufreq_update_policy(cpu);
 929         }
 930
 931         mutex_unlock(&intel_pstate_driver_lock);
 932 }
 933
 934 /************************** sysfs begin ************************/
 935 #define show_one(file_name, object)                                     \
 936         static ssize_t show_##file_name                                 \
 937         (struct kobject *kobj, struct kobj_attribute *attr, char *buf)  \
 938         {                                                               \
 939                 return sprintf(buf, "%u\n", global.object);             \
 940         }
 941
 942 static ssize_t intel_pstate_show_status(char *buf);
 943 static int intel_pstate_update_status(const char *buf, size_t size);
 944
 945 static ssize_t show_status(struct kobject *kobj,
 946                            struct kobj_attribute *attr, char *buf)
 947 {
 948         ssize_t ret;
 949
 950         mutex_lock(&intel_pstate_driver_lock);
 951         ret = intel_pstate_show_status(buf);
 952         mutex_unlock(&intel_pstate_driver_lock);
 953
 954         return ret;
 955 }
 956
 957 static ssize_t store_status(struct kobject *a, struct kobj_attribute *b,
 958                             const char *buf, size_t count)
 959 {
 960         char *p = memchr(buf, '\n', count);
 961         int ret;
 962
 963         mutex_lock(&intel_pstate_driver_lock);
 964         ret = intel_pstate_update_status(buf, p ? p - buf : count);
 965         mutex_unlock(&intel_pstate_driver_lock);
 966
 967         return ret < 0 ? ret : count;
 968 }
 969
 970 static ssize_t show_turbo_pct(struct kobject *kobj,
 971                                 struct kobj_attribute *attr, char *buf)
 972 {
 973         struct cpudata *cpu;
 974         int total, no_turbo, turbo_pct;
 975         uint32_t turbo_fp;
 976
 977         mutex_lock(&intel_pstate_driver_lock);
 978
 979         if (!intel_pstate_driver) {
 980                 mutex_unlock(&intel_pstate_driver_lock);
 981                 return -EAGAIN;
 982         }
 983
 984         cpu = all_cpu_data[0];
 985
 986         total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
 987         no_turbo = cpu->pstate.max_pstate - cpu->pstate.min_pstate + 1;
 988         turbo_fp = div_fp(no_turbo, total);
 989         turbo_pct = 100 - fp_toint(mul_fp(turbo_fp, int_tofp(100)));
 990
 991         mutex_unlock(&intel_pstate_driver_lock);
 992
 993         return sprintf(buf, "%u\n", turbo_pct);
 994 }
 995
 996 static ssize_t show_num_pstates(struct kobject *kobj,
 997                                 struct kobj_attribute *attr, char *buf)
 998 {
 999         struct cpudata *cpu;
1000         int total;
1001
1002         mutex_lock(&intel_pstate_driver_lock);
1003
1004         if (!intel_pstate_driver) {
1005                 mutex_unlock(&intel_pstate_driver_lock);
1006                 return -EAGAIN;
1007         }
1008
1009         cpu = all_cpu_data[0];
1010         total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
1011
1012         mutex_unlock(&intel_pstate_driver_lock);
1013
1014         return sprintf(buf, "%u\n", total);
1015 }
1016
1017 static ssize_t show_no_turbo(struct kobject *kobj,
1018                              struct kobj_attribute *attr, char *buf)
1019 {
1020         ssize_t ret;
1021
1022         mutex_lock(&intel_pstate_driver_lock);
1023
1024         if (!intel_pstate_driver) {
1025                 mutex_unlock(&intel_pstate_driver_lock);
1026                 return -EAGAIN;
1027         }
1028
1029         update_turbo_state();
1030         if (global.turbo_disabled)
1031                 ret = sprintf(buf, "%u\n", global.turbo_disabled);
1032         else
1033                 ret = sprintf(buf, "%u\n", global.no_turbo);
1034
1035         mutex_unlock(&intel_pstate_driver_lock);
1036
1037         return ret;
1038 }
1039
1040 static ssize_t store_no_turbo(struct kobject *a, struct kobj_attribute *b,
1041                               const char *buf, size_t count)
1042 {
1043         unsigned int input;
1044         int ret;
1045
1046         ret = sscanf(buf, "%u", &input);
1047         if (ret != 1)
1048                 return -EINVAL;
1049
1050         mutex_lock(&intel_pstate_driver_lock);
1051
1052         if (!intel_pstate_driver) {
1053                 mutex_unlock(&intel_pstate_driver_lock);
1054                 return -EAGAIN;
1055         }
1056
1057         mutex_lock(&intel_pstate_limits_lock);
1058
1059         update_turbo_state();
1060         if (global.turbo_disabled) {
1061                 pr_warn("Turbo disabled by BIOS or unavailable on processor\n");
1062                 mutex_unlock(&intel_pstate_limits_lock);
1063                 mutex_unlock(&intel_pstate_driver_lock);
1064                 return -EPERM;
1065         }
1066
1067         global.no_turbo = clamp_t(int, input, 0, 1);
1068
1069         if (global.no_turbo) {
1070                 struct cpudata *cpu = all_cpu_data[0];
1071                 int pct = cpu->pstate.max_pstate * 100 / cpu->pstate.turbo_pstate;
1072
1073                 /* Squash the global minimum into the permitted range. */
1074                 if (global.min_perf_pct > pct)
1075                         global.min_perf_pct = pct;
1076         }
1077
1078         mutex_unlock(&intel_pstate_limits_lock);
1079
1080         intel_pstate_update_policies();
1081
1082         mutex_unlock(&intel_pstate_driver_lock);
1083
1084         return count;
1085 }
1086
1087 static struct cpufreq_driver intel_pstate;
1088
1089 static void update_qos_request(enum freq_qos_req_type type)
1090 {
1091         int max_state, turbo_max, freq, i, perf_pct;
1092         struct freq_qos_request *req;
1093         struct cpufreq_policy *policy;
1094
1095         for_each_possible_cpu(i) {
1096                 struct cpudata *cpu = all_cpu_data[i];
1097
1098                 policy = cpufreq_cpu_get(i);
1099                 if (!policy)
1100                         continue;
1101
1102                 req = policy->driver_data;
1103                 cpufreq_cpu_put(policy);
1104
1105                 if (!req)
1106                         continue;
1107
1108                 if (hwp_active)
1109                         intel_pstate_get_hwp_max(i, &turbo_max, &max_state);
1110                 else
1111                         turbo_max = cpu->pstate.turbo_pstate;
1112
1113                 if (type == FREQ_QOS_MIN) {
1114                         perf_pct = global.min_perf_pct;
1115                 } else {
1116                         req++;
1117                         perf_pct = global.max_perf_pct;
1118                 }
1119
1120                 freq = DIV_ROUND_UP(turbo_max * perf_pct, 100);
1121                 freq *= cpu->pstate.scaling;
1122
1123                 if (freq_qos_update_request(req, freq) < 0)
1124                         pr_warn("Failed to update freq constraint: CPU%d\n", i);
1125         }
1126 }
1127
1128 static ssize_t store_max_perf_pct(struct kobject *a, struct kobj_attribute *b,
1129                                   const char *buf, size_t count)
1130 {
1131         unsigned int input;
1132         int ret;
1133
1134         ret = sscanf(buf, "%u", &input);
1135         if (ret != 1)
1136                 return -EINVAL;
1137
1138         mutex_lock(&intel_pstate_driver_lock);
1139
1140         if (!intel_pstate_driver) {
1141                 mutex_unlock(&intel_pstate_driver_lock);
1142                 return -EAGAIN;
1143         }
1144
1145         mutex_lock(&intel_pstate_limits_lock);
1146
1147         global.max_perf_pct = clamp_t(int, input, global.min_perf_pct, 100);
1148
1149         mutex_unlock(&intel_pstate_limits_lock);
1150
1151         if (intel_pstate_driver == &intel_pstate)
1152                 intel_pstate_update_policies();
1153         else
1154                 update_qos_request(FREQ_QOS_MAX);
1155
1156         mutex_unlock(&intel_pstate_driver_lock);
1157
1158         return count;
1159 }
1160
1161 static ssize_t store_min_perf_pct(struct kobject *a, struct kobj_attribute *b,
1162                                   const char *buf, size_t count)
1163 {
1164         unsigned int input;
1165         int ret;
1166
1167         ret = sscanf(buf, "%u", &input);
1168         if (ret != 1)
1169                 return -EINVAL;
1170
1171         mutex_lock(&intel_pstate_driver_lock);
1172
1173         if (!intel_pstate_driver) {
1174                 mutex_unlock(&intel_pstate_driver_lock);
1175                 return -EAGAIN;
1176         }
1177
1178         mutex_lock(&intel_pstate_limits_lock);
1179
1180         global.min_perf_pct = clamp_t(int, input,
1181                                       min_perf_pct_min(), global.max_perf_pct);
1182
1183         mutex_unlock(&intel_pstate_limits_lock);
1184
1185         if (intel_pstate_driver == &intel_pstate)
1186                 intel_pstate_update_policies();
1187         else
1188                 update_qos_request(FREQ_QOS_MIN);
1189
1190         mutex_unlock(&intel_pstate_driver_lock);
1191
1192         return count;
1193 }
1194
1195 static ssize_t show_hwp_dynamic_boost(struct kobject *kobj,
1196                                 struct kobj_attribute *attr, char *buf)
1197 {
1198         return sprintf(buf, "%u\n", hwp_boost);
1199 }
1200
1201 static ssize_t store_hwp_dynamic_boost(struct kobject *a,
1202                                        struct kobj_attribute *b,
1203                                        const char *buf, size_t count)
1204 {
1205         unsigned int input;
1206         int ret;
1207
1208         ret = kstrtouint(buf, 10, &input);
1209         if (ret)
1210                 return ret;
1211
1212         mutex_lock(&intel_pstate_driver_lock);
1213         hwp_boost = !!input;
1214         intel_pstate_update_policies();
1215         mutex_unlock(&intel_pstate_driver_lock);
1216
1217         return count;
1218 }
1219
1220 show_one(max_perf_pct, max_perf_pct);
1221 show_one(min_perf_pct, min_perf_pct);
1222
1223 define_one_global_rw(status);
1224 define_one_global_rw(no_turbo);
1225 define_one_global_rw(max_perf_pct);
1226 define_one_global_rw(min_perf_pct);
1227 define_one_global_ro(turbo_pct);
1228 define_one_global_ro(num_pstates);
1229 define_one_global_rw(hwp_dynamic_boost);
1230
1231 static struct attribute *intel_pstate_attributes[] = {
1232         &status.attr,
1233         &no_turbo.attr,
1234         &turbo_pct.attr,
1235         &num_pstates.attr,
1236         NULL
1237 };
1238
1239 static const struct attribute_group intel_pstate_attr_group = {
1240         .attrs = intel_pstate_attributes,
1241 };
1242
1243 static void __init intel_pstate_sysfs_expose_params(void)
1244 {
1245         struct kobject *intel_pstate_kobject;
1246         int rc;
1247
1248         intel_pstate_kobject = kobject_create_and_add("intel_pstate",
1249                                                 &cpu_subsys.dev_root->kobj);
1250         if (WARN_ON(!intel_pstate_kobject))
1251                 return;
1252
1253         rc = sysfs_create_group(intel_pstate_kobject, &intel_pstate_attr_group);
1254         if (WARN_ON(rc))
1255                 return;
1256
1257         /*
1258          * If per cpu limits are enforced there are no global limits, so
1259          * return without creating max/min_perf_pct attributes
1260          */
1261         if (per_cpu_limits)
1262                 return;
1263
1264         rc = sysfs_create_file(intel_pstate_kobject, &max_perf_pct.attr);
1265         WARN_ON(rc);
1266
1267         rc = sysfs_create_file(intel_pstate_kobject, &min_perf_pct.attr);
1268         WARN_ON(rc);
1269
1270         if (hwp_active) {
1271                 rc = sysfs_create_file(intel_pstate_kobject,
1272                                        &hwp_dynamic_boost.attr);
1273                 WARN_ON(rc);
1274         }
1275 }
1276 /************************** sysfs end ************************/
1277
1278 static void intel_pstate_hwp_enable(struct cpudata *cpudata)
1279 {
1280         /* First disable HWP notification interrupt as we don't process them */
1281         if (boot_cpu_has(X86_FEATURE_HWP_NOTIFY))
1282                 wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_INTERRUPT, 0x00);
1283
1284         wrmsrl_on_cpu(cpudata->cpu, MSR_PM_ENABLE, 0x1);
1285         cpudata->epp_policy = 0;
1286         if (cpudata->epp_default == -EINVAL)
1287                 cpudata->epp_default = intel_pstate_get_epp(cpudata, 0);
1288 }
1289
1290 #define MSR_IA32_POWER_CTL_BIT_EE       19
1291
1292 /* Disable energy efficiency optimization */
1293 static void intel_pstate_disable_ee(int cpu)
1294 {
1295         u64 power_ctl;
1296         int ret;
1297
1298         ret = rdmsrl_on_cpu(cpu, MSR_IA32_POWER_CTL, &power_ctl);
1299         if (ret)
1300                 return;
1301
1302         if (!(power_ctl & BIT(MSR_IA32_POWER_CTL_BIT_EE))) {
1303                 pr_info("Disabling energy efficiency optimization\n");
1304                 power_ctl |= BIT(MSR_IA32_POWER_CTL_BIT_EE);
1305                 wrmsrl_on_cpu(cpu, MSR_IA32_POWER_CTL, power_ctl);
1306         }
1307 }
1308
1309 static int atom_get_min_pstate(void)
1310 {
1311         u64 value;
1312
1313         rdmsrl(MSR_ATOM_CORE_RATIOS, value);
1314         return (value >> 8) & 0x7F;
1315 }
1316
1317 static int atom_get_max_pstate(void)
1318 {
1319         u64 value;
1320
1321         rdmsrl(MSR_ATOM_CORE_RATIOS, value);
1322         return (value >> 16) & 0x7F;
1323 }
1324
1325 static int atom_get_turbo_pstate(void)
1326 {
1327         u64 value;
1328
1329         rdmsrl(MSR_ATOM_CORE_TURBO_RATIOS, value);
1330         return value & 0x7F;
1331 }
1332
1333 static u64 atom_get_val(struct cpudata *cpudata, int pstate)
1334 {
1335         u64 val;
1336         int32_t vid_fp;
1337         u32 vid;
1338
1339         val = (u64)pstate << 8;
1340         if (global.no_turbo && !global.turbo_disabled)
1341                 val |= (u64)1 << 32;
1342
1343         vid_fp = cpudata->vid.min + mul_fp(
1344                 int_tofp(pstate - cpudata->pstate.min_pstate),
1345                 cpudata->vid.ratio);
1346
1347         vid_fp = clamp_t(int32_t, vid_fp, cpudata->vid.min, cpudata->vid.max);
1348         vid = ceiling_fp(vid_fp);
1349
1350         if (pstate > cpudata->pstate.max_pstate)
1351                 vid = cpudata->vid.turbo;
1352
1353         return val | vid;
1354 }
1355
1356 static int silvermont_get_scaling(void)
1357 {
1358         u64 value;
1359         int i;
1360         /* Defined in Table 35-6 from SDM (Sept 2015) */
1361         static int silvermont_freq_table[] = {
1362                 83300, 100000, 133300, 116700, 80000};
1363
1364         rdmsrl(MSR_FSB_FREQ, value);
1365         i = value & 0x7;
1366         WARN_ON(i > 4);
1367
1368         return silvermont_freq_table[i];
1369 }
1370
1371 static int airmont_get_scaling(void)
1372 {
1373         u64 value;
1374         int i;
1375         /* Defined in Table 35-10 from SDM (Sept 2015) */
1376         static int airmont_freq_table[] = {
1377                 83300, 100000, 133300, 116700, 80000,
1378                 93300, 90000, 88900, 87500};
1379
1380         rdmsrl(MSR_FSB_FREQ, value);
1381         i = value & 0xF;
1382         WARN_ON(i > 8);
1383
1384         return airmont_freq_table[i];
1385 }
1386
1387 static void atom_get_vid(struct cpudata *cpudata)
1388 {
1389         u64 value;
1390
1391         rdmsrl(MSR_ATOM_CORE_VIDS, value);
1392         cpudata->vid.min = int_tofp((value >> 8) & 0x7f);
1393         cpudata->vid.max = int_tofp((value >> 16) & 0x7f);
1394         cpudata->vid.ratio = div_fp(
1395                 cpudata->vid.max - cpudata->vid.min,
1396                 int_tofp(cpudata->pstate.max_pstate -
1397                         cpudata->pstate.min_pstate));
1398
1399         rdmsrl(MSR_ATOM_CORE_TURBO_VIDS, value);
1400         cpudata->vid.turbo = value & 0x7f;
1401 }
1402
1403 static int core_get_min_pstate(void)
1404 {
1405         u64 value;
1406
1407         rdmsrl(MSR_PLATFORM_INFO, value);
1408         return (value >> 40) & 0xFF;
1409 }
1410
1411 static int core_get_max_pstate_physical(void)
1412 {
1413         u64 value;
1414
1415         rdmsrl(MSR_PLATFORM_INFO, value);
1416         return (value >> 8) & 0xFF;
1417 }
1418
1419 static int core_get_tdp_ratio(u64 plat_info)
1420 {
1421         /* Check how many TDP levels present */
1422         if (plat_info & 0x600000000) {
1423                 u64 tdp_ctrl;
1424                 u64 tdp_ratio;
1425                 int tdp_msr;
1426                 int err;
1427
1428                 /* Get the TDP level (0, 1, 2) to get ratios */
1429                 err = rdmsrl_safe(MSR_CONFIG_TDP_CONTROL, &tdp_ctrl);
1430                 if (err)
1431                         return err;
1432
1433                 /* TDP MSR are continuous starting at 0x648 */
1434                 tdp_msr = MSR_CONFIG_TDP_NOMINAL + (tdp_ctrl & 0x03);
1435                 err = rdmsrl_safe(tdp_msr, &tdp_ratio);
1436                 if (err)
1437                         return err;
1438
1439                 /* For level 1 and 2, bits[23:16] contain the ratio */
1440                 if (tdp_ctrl & 0x03)
1441                         tdp_ratio >>= 16;
1442
1443                 tdp_ratio &= 0xff; /* ratios are only 8 bits long */
1444                 pr_debug("tdp_ratio %x\n", (int)tdp_ratio);
1445
1446                 return (int)tdp_ratio;
1447         }
1448
1449         return -ENXIO;
1450 }
1451
1452 static int core_get_max_pstate(void)
1453 {
1454         u64 tar;
1455         u64 plat_info;
1456         int max_pstate;
1457         int tdp_ratio;
1458         int err;
1459
1460         rdmsrl(MSR_PLATFORM_INFO, plat_info);
1461         max_pstate = (plat_info >> 8) & 0xFF;
1462
1463         tdp_ratio = core_get_tdp_ratio(plat_info);
1464         if (tdp_ratio <= 0)
1465                 return max_pstate;
1466
1467         if (hwp_active) {
1468                 /* Turbo activation ratio is not used on HWP platforms */
1469                 return tdp_ratio;
1470         }
1471
1472         err = rdmsrl_safe(MSR_TURBO_ACTIVATION_RATIO, &tar);
1473         if (!err) {
1474                 int tar_levels;
1475
1476                 /* Do some sanity checking for safety */
1477                 tar_levels = tar & 0xff;
1478                 if (tdp_ratio - 1 == tar_levels) {
1479                         max_pstate = tar_levels;
1480                         pr_debug("max_pstate=TAC %x\n", max_pstate);
1481                 }
1482         }
1483
1484         return max_pstate;
1485 }
1486
1487 static int core_get_turbo_pstate(void)
1488 {
1489         u64 value;
1490         int nont, ret;
1491
1492         rdmsrl(MSR_TURBO_RATIO_LIMIT, value);
1493         nont = core_get_max_pstate();
1494         ret = (value) & 255;
1495         if (ret <= nont)
1496                 ret = nont;
1497         return ret;
1498 }
1499
1500 static inline int core_get_scaling(void)
1501 {
1502         return 100000;
1503 }
1504
1505 static u64 core_get_val(struct cpudata *cpudata, int pstate)
1506 {
1507         u64 val;
1508
1509         val = (u64)pstate << 8;
1510         if (global.no_turbo && !global.turbo_disabled)
1511                 val |= (u64)1 << 32;
1512
1513         return val;
1514 }
1515
1516 static int knl_get_aperf_mperf_shift(void)
1517 {
1518         return 10;
1519 }
1520
1521 static int knl_get_turbo_pstate(void)
1522 {
1523         u64 value;
1524         int nont, ret;
1525
1526         rdmsrl(MSR_TURBO_RATIO_LIMIT, value);
1527         nont = core_get_max_pstate();
1528         ret = (((value) >> 8) & 0xFF);
1529         if (ret <= nont)
1530                 ret = nont;
1531         return ret;
1532 }
1533
1534 static void intel_pstate_set_pstate(struct cpudata *cpu, int pstate)
1535 {
1536         trace_cpu_frequency(pstate * cpu->pstate.scaling, cpu->cpu);
1537         cpu->pstate.current_pstate = pstate;
1538         /*
1539          * Generally, there is no guarantee that this code will always run on
1540          * the CPU being updated, so force the register update to run on the
1541          * right CPU.
1542          */
1543         wrmsrl_on_cpu(cpu->cpu, MSR_IA32_PERF_CTL,
1544                       pstate_funcs.get_val(cpu, pstate));
1545 }
1546
1547 static void intel_pstate_set_min_pstate(struct cpudata *cpu)
1548 {
1549         intel_pstate_set_pstate(cpu, cpu->pstate.min_pstate);
1550 }
1551
1552 static void intel_pstate_max_within_limits(struct cpudata *cpu)
1553 {
1554         int pstate = max(cpu->pstate.min_pstate, cpu->max_perf_ratio);
1555
1556         update_turbo_state();
1557         intel_pstate_set_pstate(cpu, pstate);
1558 }
1559
1560 static void intel_pstate_get_cpu_pstates(struct cpudata *cpu)
1561 {
1562         cpu->pstate.min_pstate = pstate_funcs.get_min();
1563         cpu->pstate.max_pstate = pstate_funcs.get_max();
1564         cpu->pstate.max_pstate_physical = pstate_funcs.get_max_physical();
1565         cpu->pstate.turbo_pstate = pstate_funcs.get_turbo();
1566         cpu->pstate.scaling = pstate_funcs.get_scaling();
1567         cpu->pstate.max_freq = cpu->pstate.max_pstate * cpu->pstate.scaling;
1568
1569         if (hwp_active && !hwp_mode_bdw) {
1570                 unsigned int phy_max, current_max;
1571
1572                 intel_pstate_get_hwp_max(cpu->cpu, &phy_max, &current_max);
1573                 cpu->pstate.turbo_freq = phy_max * cpu->pstate.scaling;
1574         } else {
1575                 cpu->pstate.turbo_freq = cpu->pstate.turbo_pstate * cpu->pstate.scaling;
1576         }
1577
1578         if (pstate_funcs.get_aperf_mperf_shift)
1579                 cpu->aperf_mperf_shift = pstate_funcs.get_aperf_mperf_shift();
1580
1581         if (pstate_funcs.get_vid)
1582                 pstate_funcs.get_vid(cpu);
1583
1584         intel_pstate_set_min_pstate(cpu);
1585 }
1586
1587 /*
1588  * Long hold time will keep high perf limits for long time,
1589  * which negatively impacts perf/watt for some workloads,
1590  * like specpower. 3ms is based on experiements on some
1591  * workoads.
1592  */
1593 static int hwp_boost_hold_time_ns = 3 * NSEC_PER_MSEC;
1594
1595 static inline void intel_pstate_hwp_boost_up(struct cpudata *cpu)
1596 {
1597         u64 hwp_req = READ_ONCE(cpu->hwp_req_cached);
1598         u32 max_limit = (hwp_req & 0xff00) >> 8;
1599         u32 min_limit = (hwp_req & 0xff);
1600         u32 boost_level1;
1601
1602         /*
1603          * Cases to consider (User changes via sysfs or boot time):
1604          * If, P0 (Turbo max) = P1 (Guaranteed max) = min:
1605          *      No boost, return.
1606          * If, P0 (Turbo max) > P1 (Guaranteed max) = min:
1607          *     Should result in one level boost only for P0.
1608          * If, P0 (Turbo max) = P1 (Guaranteed max) > min:
1609          *     Should result in two level boost:
1610          *         (min + p1)/2 and P1.
1611          * If, P0 (Turbo max) > P1 (Guaranteed max) > min:
1612          *     Should result in three level boost:
1613          *        (min + p1)/2, P1 and P0.
1614          */
1615
1616         /* If max and min are equal or already at max, nothing to boost */
1617         if (max_limit == min_limit || cpu->hwp_boost_min >= max_limit)
1618                 return;
1619
1620         if (!cpu->hwp_boost_min)
1621                 cpu->hwp_boost_min = min_limit;
1622
1623         /* level at half way mark between min and guranteed */
1624         boost_level1 = (HWP_GUARANTEED_PERF(cpu->hwp_cap_cached) + min_limit) >> 1;
1625
1626         if (cpu->hwp_boost_min < boost_level1)
1627                 cpu->hwp_boost_min = boost_level1;
1628         else if (cpu->hwp_boost_min < HWP_GUARANTEED_PERF(cpu->hwp_cap_cached))
1629                 cpu->hwp_boost_min = HWP_GUARANTEED_PERF(cpu->hwp_cap_cached);
1630         else if (cpu->hwp_boost_min == HWP_GUARANTEED_PERF(cpu->hwp_cap_cached) &&
1631                  max_limit != HWP_GUARANTEED_PERF(cpu->hwp_cap_cached))
1632                 cpu->hwp_boost_min = max_limit;
1633         else
1634                 return;
1635
1636         hwp_req = (hwp_req & ~GENMASK_ULL(7, 0)) | cpu->hwp_boost_min;
1637         wrmsrl(MSR_HWP_REQUEST, hwp_req);
1638         cpu->last_update = cpu->sample.time;
1639 }
1640
1641 static inline void intel_pstate_hwp_boost_down(struct cpudata *cpu)
1642 {
1643         if (cpu->hwp_boost_min) {
1644                 bool expired;
1645
1646                 /* Check if we are idle for hold time to boost down */
1647                 expired = time_after64(cpu->sample.time, cpu->last_update +
1648                                        hwp_boost_hold_time_ns);
1649                 if (expired) {
1650                         wrmsrl(MSR_HWP_REQUEST, cpu->hwp_req_cached);
1651                         cpu->hwp_boost_min = 0;
1652                 }
1653         }
1654         cpu->last_update = cpu->sample.time;
1655 }
1656
1657 static inline void intel_pstate_update_util_hwp_local(struct cpudata *cpu,
1658                                                       u64 time)
1659 {
1660         cpu->sample.time = time;
1661
1662         if (cpu->sched_flags & SCHED_CPUFREQ_IOWAIT) {
1663                 bool do_io = false;
1664
1665                 cpu->sched_flags = 0;
1666                 /*
1667                  * Set iowait_boost flag and update time. Since IO WAIT flag
1668                  * is set all the time, we can't just conclude that there is
1669                  * some IO bound activity is scheduled on this CPU with just
1670                  * one occurrence. If we receive at least two in two
1671                  * consecutive ticks, then we treat as boost candidate.
1672                  */
1673                 if (time_before64(time, cpu->last_io_update + 2 * TICK_NSEC))
1674                         do_io = true;
1675
1676                 cpu->last_io_update = time;
1677
1678                 if (do_io)
1679                         intel_pstate_hwp_boost_up(cpu);
1680
1681         } else {
1682                 intel_pstate_hwp_boost_down(cpu);
1683         }
1684 }
1685
1686 static inline void intel_pstate_update_util_hwp(struct update_util_data *data,
1687                                                 u64 time, unsigned int flags)
1688 {
1689         struct cpudata *cpu = container_of(data, struct cpudata, update_util);
1690
1691         cpu->sched_flags |= flags;
1692
1693         if (smp_processor_id() == cpu->cpu)
1694                 intel_pstate_update_util_hwp_local(cpu, time);
1695 }
1696
1697 static inline void intel_pstate_calc_avg_perf(struct cpudata *cpu)
1698 {
1699         struct sample *sample = &cpu->sample;
1700
1701         sample->core_avg_perf = div_ext_fp(sample->aperf, sample->mperf);
1702 }
1703
1704 static inline bool intel_pstate_sample(struct cpudata *cpu, u64 time)
1705 {
1706         u64 aperf, mperf;
1707         unsigned long flags;
1708         u64 tsc;
1709
1710         local_irq_save(flags);
1711         rdmsrl(MSR_IA32_APERF, aperf);
1712         rdmsrl(MSR_IA32_MPERF, mperf);
1713         tsc = rdtsc();
1714         if (cpu->prev_mperf == mperf || cpu->prev_tsc == tsc) {
1715                 local_irq_restore(flags);
1716                 return false;
1717         }
1718         local_irq_restore(flags);
1719
1720         cpu->last_sample_time = cpu->sample.time;
1721         cpu->sample.time = time;
1722         cpu->sample.aperf = aperf;
1723         cpu->sample.mperf = mperf;
1724         cpu->sample.tsc =  tsc;
1725         cpu->sample.aperf -= cpu->prev_aperf;
1726         cpu->sample.mperf -= cpu->prev_mperf;
1727         cpu->sample.tsc -= cpu->prev_tsc;
1728
1729         cpu->prev_aperf = aperf;
1730         cpu->prev_mperf = mperf;
1731         cpu->prev_tsc = tsc;
1732         /*
1733          * First time this function is invoked in a given cycle, all of the
1734          * previous sample data fields are equal to zero or stale and they must
1735          * be populated with meaningful numbers for things to work, so assume
1736          * that sample.time will always be reset before setting the utilization
1737          * update hook and make the caller skip the sample then.
1738          */
1739         if (cpu->last_sample_time) {
1740                 intel_pstate_calc_avg_perf(cpu);
1741                 return true;
1742         }
1743         return false;
1744 }
1745
1746 static inline int32_t get_avg_frequency(struct cpudata *cpu)
1747 {
1748         return mul_ext_fp(cpu->sample.core_avg_perf, cpu_khz);
1749 }
1750
1751 static inline int32_t get_avg_pstate(struct cpudata *cpu)
1752 {
1753         return mul_ext_fp(cpu->pstate.max_pstate_physical,
1754                           cpu->sample.core_avg_perf);
1755 }
1756
1757 static inline int32_t get_target_pstate(struct cpudata *cpu)
1758 {
1759         struct sample *sample = &cpu->sample;
1760         int32_t busy_frac;
1761         int target, avg_pstate;
1762
1763         busy_frac = div_fp(sample->mperf << cpu->aperf_mperf_shift,
1764                            sample->tsc);
1765
1766         if (busy_frac < cpu->iowait_boost)
1767                 busy_frac = cpu->iowait_boost;
1768
1769         sample->busy_scaled = busy_frac * 100;
1770
1771         target = global.no_turbo || global.turbo_disabled ?
1772                         cpu->pstate.max_pstate : cpu->pstate.turbo_pstate;
1773         target += target >> 2;
1774         target = mul_fp(target, busy_frac);
1775         if (target < cpu->pstate.min_pstate)
1776                 target = cpu->pstate.min_pstate;
1777
1778         /*
1779          * If the average P-state during the previous cycle was higher than the
1780          * current target, add 50% of the difference to the target to reduce
1781          * possible performance oscillations and offset possible performance
1782          * loss related to moving the workload from one CPU to another within
1783          * a package/module.
1784          */
1785         avg_pstate = get_avg_pstate(cpu);
1786         if (avg_pstate > target)
1787                 target += (avg_pstate - target) >> 1;
1788
1789         return target;
1790 }
1791
1792 static int intel_pstate_prepare_request(struct cpudata *cpu, int pstate)
1793 {
1794         int min_pstate = max(cpu->pstate.min_pstate, cpu->min_perf_ratio);
1795         int max_pstate = max(min_pstate, cpu->max_perf_ratio);
1796
1797         return clamp_t(int, pstate, min_pstate, max_pstate);
1798 }
1799
1800 static void intel_pstate_update_pstate(struct cpudata *cpu, int pstate)
1801 {
1802         if (pstate == cpu->pstate.current_pstate)
1803                 return;
1804
1805         cpu->pstate.current_pstate = pstate;
1806         wrmsrl(MSR_IA32_PERF_CTL, pstate_funcs.get_val(cpu, pstate));
1807 }
1808
1809 static void intel_pstate_adjust_pstate(struct cpudata *cpu)
1810 {
1811         int from = cpu->pstate.current_pstate;
1812         struct sample *sample;
1813         int target_pstate;
1814
1815         update_turbo_state();
1816
1817         target_pstate = get_target_pstate(cpu);
1818         target_pstate = intel_pstate_prepare_request(cpu, target_pstate);
1819         trace_cpu_frequency(target_pstate * cpu->pstate.scaling, cpu->cpu);
1820         intel_pstate_update_pstate(cpu, target_pstate);
1821
1822         sample = &cpu->sample;
1823         trace_pstate_sample(mul_ext_fp(100, sample->core_avg_perf),
1824                 fp_toint(sample->busy_scaled),
1825                 from,
1826                 cpu->pstate.current_pstate,
1827                 sample->mperf,
1828                 sample->aperf,
1829                 sample->tsc,
1830                 get_avg_frequency(cpu),
1831                 fp_toint(cpu->iowait_boost * 100));
1832 }
1833
1834 static void intel_pstate_update_util(struct update_util_data *data, u64 time,
1835                                      unsigned int flags)
1836 {
1837         struct cpudata *cpu = container_of(data, struct cpudata, update_util);
1838         u64 delta_ns;
1839
1840         /* Don't allow remote callbacks */
1841         if (smp_processor_id() != cpu->cpu)
1842                 return;
1843
1844         delta_ns = time - cpu->last_update;
1845         if (flags & SCHED_CPUFREQ_IOWAIT) {
1846                 /* Start over if the CPU may have been idle. */
1847                 if (delta_ns > TICK_NSEC) {
1848                         cpu->iowait_boost = ONE_EIGHTH_FP;
1849                 } else if (cpu->iowait_boost >= ONE_EIGHTH_FP) {
1850                         cpu->iowait_boost <<= 1;
1851                         if (cpu->iowait_boost > int_tofp(1))
1852                                 cpu->iowait_boost = int_tofp(1);
1853                 } else {
1854                         cpu->iowait_boost = ONE_EIGHTH_FP;
1855                 }
1856         } else if (cpu->iowait_boost) {
1857                 /* Clear iowait_boost if the CPU may have been idle. */
1858                 if (delta_ns > TICK_NSEC)
1859                         cpu->iowait_boost = 0;
1860                 else
1861                         cpu->iowait_boost >>= 1;
1862         }
1863         cpu->last_update = time;
1864         delta_ns = time - cpu->sample.time;
1865         if ((s64)delta_ns < INTEL_PSTATE_SAMPLING_INTERVAL)
1866                 return;
1867
1868         if (intel_pstate_sample(cpu, time))
1869                 intel_pstate_adjust_pstate(cpu);
1870 }
1871
1872 static struct pstate_funcs core_funcs = {
1873         .get_max = core_get_max_pstate,
1874         .get_max_physical = core_get_max_pstate_physical,
1875         .get_min = core_get_min_pstate,
1876         .get_turbo = core_get_turbo_pstate,
1877         .get_scaling = core_get_scaling,
1878         .get_val = core_get_val,
1879 };
1880
1881 static const struct pstate_funcs silvermont_funcs = {
1882         .get_max = atom_get_max_pstate,
1883         .get_max_physical = atom_get_max_pstate,
1884         .get_min = atom_get_min_pstate,
1885         .get_turbo = atom_get_turbo_pstate,
1886         .get_val = atom_get_val,
1887         .get_scaling = silvermont_get_scaling,
1888         .get_vid = atom_get_vid,
1889 };
1890
1891 static const struct pstate_funcs airmont_funcs = {
1892         .get_max = atom_get_max_pstate,
1893         .get_max_physical = atom_get_max_pstate,
1894         .get_min = atom_get_min_pstate,
1895         .get_turbo = atom_get_turbo_pstate,
1896         .get_val = atom_get_val,
1897         .get_scaling = airmont_get_scaling,
1898         .get_vid = atom_get_vid,
1899 };
1900
1901 static const struct pstate_funcs knl_funcs = {
1902         .get_max = core_get_max_pstate,
1903         .get_max_physical = core_get_max_pstate_physical,
1904         .get_min = core_get_min_pstate,
1905         .get_turbo = knl_get_turbo_pstate,
1906         .get_aperf_mperf_shift = knl_get_aperf_mperf_shift,
1907         .get_scaling = core_get_scaling,
1908         .get_val = core_get_val,
1909 };
1910
1911 #define ICPU(model, policy) \
1912         { X86_VENDOR_INTEL, 6, model, X86_FEATURE_APERFMPERF,\
1913                         (unsigned long)&policy }
1914
1915 static const struct x86_cpu_id intel_pstate_cpu_ids[] = {
1916         ICPU(INTEL_FAM6_SANDYBRIDGE,            core_funcs),
1917         ICPU(INTEL_FAM6_SANDYBRIDGE_X,          core_funcs),
1918         ICPU(INTEL_FAM6_ATOM_SILVERMONT,        silvermont_funcs),
1919         ICPU(INTEL_FAM6_IVYBRIDGE,              core_funcs),
1920         ICPU(INTEL_FAM6_HASWELL,                core_funcs),
1921         ICPU(INTEL_FAM6_BROADWELL,              core_funcs),
1922         ICPU(INTEL_FAM6_IVYBRIDGE_X,            core_funcs),
1923         ICPU(INTEL_FAM6_HASWELL_X,              core_funcs),
1924         ICPU(INTEL_FAM6_HASWELL_L,              core_funcs),
1925         ICPU(INTEL_FAM6_HASWELL_G,              core_funcs),
1926         ICPU(INTEL_FAM6_BROADWELL_G,            core_funcs),
1927         ICPU(INTEL_FAM6_ATOM_AIRMONT,           airmont_funcs),
1928         ICPU(INTEL_FAM6_SKYLAKE_L,              core_funcs),
1929         ICPU(INTEL_FAM6_BROADWELL_X,            core_funcs),
1930         ICPU(INTEL_FAM6_SKYLAKE,                core_funcs),
1931         ICPU(INTEL_FAM6_BROADWELL_D,            core_funcs),
1932         ICPU(INTEL_FAM6_XEON_PHI_KNL,           knl_funcs),
1933         ICPU(INTEL_FAM6_XEON_PHI_KNM,           knl_funcs),
1934         ICPU(INTEL_FAM6_ATOM_GOLDMONT,          core_funcs),
1935         ICPU(INTEL_FAM6_ATOM_GOLDMONT_PLUS,     core_funcs),
1936         ICPU(INTEL_FAM6_SKYLAKE_X,              core_funcs),
1937         {}
1938 };
1939 MODULE_DEVICE_TABLE(x86cpu, intel_pstate_cpu_ids);
1940
1941 static const struct x86_cpu_id intel_pstate_cpu_oob_ids[] __initconst = {
1942         ICPU(INTEL_FAM6_BROADWELL_D, core_funcs),
1943         ICPU(INTEL_FAM6_BROADWELL_X, core_funcs),
1944         ICPU(INTEL_FAM6_SKYLAKE_X, core_funcs),
1945         {}
1946 };
1947
1948 static const struct x86_cpu_id intel_pstate_cpu_ee_disable_ids[] = {
1949         ICPU(INTEL_FAM6_KABYLAKE, core_funcs),
1950         {}
1951 };
1952
1953 static const struct x86_cpu_id intel_pstate_hwp_boost_ids[] = {
1954         ICPU(INTEL_FAM6_SKYLAKE_X, core_funcs),
1955         ICPU(INTEL_FAM6_SKYLAKE, core_funcs),
1956         {}
1957 };
1958
1959 static int intel_pstate_init_cpu(unsigned int cpunum)
1960 {
1961         struct cpudata *cpu;
1962
1963         cpu = all_cpu_data[cpunum];
1964
1965         if (!cpu) {
1966                 cpu = kzalloc(sizeof(*cpu), GFP_KERNEL);
1967                 if (!cpu)
1968                         return -ENOMEM;
1969
1970                 all_cpu_data[cpunum] = cpu;
1971
1972                 cpu->epp_default = -EINVAL;
1973                 cpu->epp_powersave = -EINVAL;
1974                 cpu->epp_saved = -EINVAL;
1975         }
1976
1977         cpu = all_cpu_data[cpunum];
1978
1979         cpu->cpu = cpunum;
1980
1981         if (hwp_active) {
1982                 const struct x86_cpu_id *id;
1983
1984                 id = x86_match_cpu(intel_pstate_cpu_ee_disable_ids);
1985                 if (id)
1986                         intel_pstate_disable_ee(cpunum);
1987
1988                 intel_pstate_hwp_enable(cpu);
1989
1990                 id = x86_match_cpu(intel_pstate_hwp_boost_ids);
1991                 if (id && intel_pstate_acpi_pm_profile_server())
1992                         hwp_boost = true;
1993         }
1994
1995         intel_pstate_get_cpu_pstates(cpu);
1996
1997         pr_debug("controlling: cpu %d\n", cpunum);
1998
1999         return 0;
2000 }
2001
2002 static void intel_pstate_set_update_util_hook(unsigned int cpu_num)
2003 {
2004         struct cpudata *cpu = all_cpu_data[cpu_num];
2005
2006         if (hwp_active && !hwp_boost)
2007                 return;
2008
2009         if (cpu->update_util_set)
2010                 return;
2011
2012         /* Prevent intel_pstate_update_util() from using stale data. */
2013         cpu->sample.time = 0;
2014         cpufreq_add_update_util_hook(cpu_num, &cpu->update_util,
2015                                      (hwp_active ?
2016                                       intel_pstate_update_util_hwp :
2017                                       intel_pstate_update_util));
2018         cpu->update_util_set = true;
2019 }
2020
2021 static void intel_pstate_clear_update_util_hook(unsigned int cpu)
2022 {
2023         struct cpudata *cpu_data = all_cpu_data[cpu];
2024
2025         if (!cpu_data->update_util_set)
2026                 return;
2027
2028         cpufreq_remove_update_util_hook(cpu);
2029         cpu_data->update_util_set = false;
2030         synchronize_rcu();
2031 }
2032
2033 static int intel_pstate_get_max_freq(struct cpudata *cpu)
2034 {
2035         return global.turbo_disabled || global.no_turbo ?
2036                         cpu->pstate.max_freq : cpu->pstate.turbo_freq;
2037 }
2038
2039 static void intel_pstate_update_perf_limits(struct cpudata *cpu,
2040                                             unsigned int policy_min,
2041                                             unsigned int policy_max)
2042 {
2043         int max_freq = intel_pstate_get_max_freq(cpu);
2044         int32_t max_policy_perf, min_policy_perf;
2045         int max_state, turbo_max;
2046
2047         /*
2048          * HWP needs some special consideration, because on BDX the
2049          * HWP_REQUEST uses abstract value to represent performance
2050          * rather than pure ratios.
2051          */
2052         if (hwp_active) {
2053                 intel_pstate_get_hwp_max(cpu->cpu, &turbo_max, &max_state);
2054         } else {
2055                 max_state = global.no_turbo || global.turbo_disabled ?
2056                         cpu->pstate.max_pstate : cpu->pstate.turbo_pstate;
2057                 turbo_max = cpu->pstate.turbo_pstate;
2058         }
2059
2060         max_policy_perf = max_state * policy_max / max_freq;
2061         if (policy_max == policy_min) {
2062                 min_policy_perf = max_policy_perf;
2063         } else {
2064                 min_policy_perf = max_state * policy_min / max_freq;
2065                 min_policy_perf = clamp_t(int32_t, min_policy_perf,
2066                                           0, max_policy_perf);
2067         }
2068
2069         pr_debug("cpu:%d max_state %d min_policy_perf:%d max_policy_perf:%d\n",
2070                  cpu->cpu, max_state, min_policy_perf, max_policy_perf);
2071
2072         /* Normalize user input to [min_perf, max_perf] */
2073         if (per_cpu_limits) {
2074                 cpu->min_perf_ratio = min_policy_perf;
2075                 cpu->max_perf_ratio = max_policy_perf;
2076         } else {
2077                 int32_t global_min, global_max;
2078
2079                 /* Global limits are in percent of the maximum turbo P-state. */
2080                 global_max = DIV_ROUND_UP(turbo_max * global.max_perf_pct, 100);
2081                 global_min = DIV_ROUND_UP(turbo_max * global.min_perf_pct, 100);
2082                 global_min = clamp_t(int32_t, global_min, 0, global_max);
2083
2084                 pr_debug("cpu:%d global_min:%d global_max:%d\n", cpu->cpu,
2085                          global_min, global_max);
2086
2087                 cpu->min_perf_ratio = max(min_policy_perf, global_min);
2088                 cpu->min_perf_ratio = min(cpu->min_perf_ratio, max_policy_perf);
2089                 cpu->max_perf_ratio = min(max_policy_perf, global_max);
2090                 cpu->max_perf_ratio = max(min_policy_perf, cpu->max_perf_ratio);
2091
2092                 /* Make sure min_perf <= max_perf */
2093                 cpu->min_perf_ratio = min(cpu->min_perf_ratio,
2094                                           cpu->max_perf_ratio);
2095
2096         }
2097         pr_debug("cpu:%d max_perf_ratio:%d min_perf_ratio:%d\n", cpu->cpu,
2098                  cpu->max_perf_ratio,
2099                  cpu->min_perf_ratio);
2100 }
2101
2102 static int intel_pstate_set_policy(struct cpufreq_policy *policy)
2103 {
2104         struct cpudata *cpu;
2105
2106         if (!policy->cpuinfo.max_freq)
2107                 return -ENODEV;
2108
2109         pr_debug("set_policy cpuinfo.max %u policy->max %u\n",
2110                  policy->cpuinfo.max_freq, policy->max);
2111
2112         cpu = all_cpu_data[policy->cpu];
2113         cpu->policy = policy->policy;
2114
2115         mutex_lock(&intel_pstate_limits_lock);
2116
2117         intel_pstate_update_perf_limits(cpu, policy->min, policy->max);
2118
2119         if (cpu->policy == CPUFREQ_POLICY_PERFORMANCE) {
2120                 /*
2121                  * NOHZ_FULL CPUs need this as the governor callback may not
2122                  * be invoked on them.
2123                  */
2124                 intel_pstate_clear_update_util_hook(policy->cpu);
2125                 intel_pstate_max_within_limits(cpu);
2126         } else {
2127                 intel_pstate_set_update_util_hook(policy->cpu);
2128         }
2129
2130         if (hwp_active) {
2131                 /*
2132                  * When hwp_boost was active before and dynamically it
2133                  * was turned off, in that case we need to clear the
2134                  * update util hook.
2135                  */
2136                 if (!hwp_boost)
2137                         intel_pstate_clear_update_util_hook(policy->cpu);
2138                 intel_pstate_hwp_set(policy->cpu);
2139         }
2140
2141         mutex_unlock(&intel_pstate_limits_lock);
2142
2143         return 0;
2144 }
2145
2146 static void intel_pstate_adjust_policy_max(struct cpudata *cpu,
2147                                            struct cpufreq_policy_data *policy)
2148 {
2149         if (!hwp_active &&
2150             cpu->pstate.max_pstate_physical > cpu->pstate.max_pstate &&
2151             policy->max < policy->cpuinfo.max_freq &&
2152             policy->max > cpu->pstate.max_freq) {
2153                 pr_debug("policy->max > max non turbo frequency\n");
2154                 policy->max = policy->cpuinfo.max_freq;
2155         }
2156 }
2157
2158 static int intel_pstate_verify_policy(struct cpufreq_policy_data *policy)
2159 {
2160         struct cpudata *cpu = all_cpu_data[policy->cpu];
2161
2162         update_turbo_state();
2163         cpufreq_verify_within_limits(policy, policy->cpuinfo.min_freq,
2164                                      intel_pstate_get_max_freq(cpu));
2165
2166         intel_pstate_adjust_policy_max(cpu, policy);
2167
2168         return 0;
2169 }
2170
2171 static void intel_cpufreq_stop_cpu(struct cpufreq_policy *policy)
2172 {
2173         intel_pstate_set_min_pstate(all_cpu_data[policy->cpu]);
2174 }
2175
2176 static void intel_pstate_stop_cpu(struct cpufreq_policy *policy)
2177 {
2178         pr_debug("CPU %d exiting\n", policy->cpu);
2179
2180         intel_pstate_clear_update_util_hook(policy->cpu);
2181         if (hwp_active) {
2182                 intel_pstate_hwp_save_state(policy);
2183                 intel_pstate_hwp_force_min_perf(policy->cpu);
2184         } else {
2185                 intel_cpufreq_stop_cpu(policy);
2186         }
2187 }
2188
2189 static int intel_pstate_cpu_exit(struct cpufreq_policy *policy)
2190 {
2191         intel_pstate_exit_perf_limits(policy);
2192
2193         policy->fast_switch_possible = false;
2194
2195         return 0;
2196 }
2197
2198 static int __intel_pstate_cpu_init(struct cpufreq_policy *policy)
2199 {
2200         struct cpudata *cpu;
2201         int rc;
2202
2203         rc = intel_pstate_init_cpu(policy->cpu);
2204         if (rc)
2205                 return rc;
2206
2207         cpu = all_cpu_data[policy->cpu];
2208
2209         cpu->max_perf_ratio = 0xFF;
2210         cpu->min_perf_ratio = 0;
2211
2212         policy->min = cpu->pstate.min_pstate * cpu->pstate.scaling;
2213         policy->max = cpu->pstate.turbo_pstate * cpu->pstate.scaling;
2214
2215         /* cpuinfo and default policy values */
2216         policy->cpuinfo.min_freq = cpu->pstate.min_pstate * cpu->pstate.scaling;
2217         update_turbo_state();
2218         global.turbo_disabled_mf = global.turbo_disabled;
2219         policy->cpuinfo.max_freq = global.turbo_disabled ?
2220                         cpu->pstate.max_pstate : cpu->pstate.turbo_pstate;
2221         policy->cpuinfo.max_freq *= cpu->pstate.scaling;
2222
2223         if (hwp_active) {
2224                 unsigned int max_freq;
2225
2226                 max_freq = global.turbo_disabled ?
2227                         cpu->pstate.max_freq : cpu->pstate.turbo_freq;
2228                 if (max_freq < policy->cpuinfo.max_freq)
2229                         policy->cpuinfo.max_freq = max_freq;
2230         }
2231
2232         intel_pstate_init_acpi_perf_limits(policy);
2233
2234         policy->fast_switch_possible = true;
2235
2236         return 0;
2237 }
2238
2239 static int intel_pstate_cpu_init(struct cpufreq_policy *policy)
2240 {
2241         int ret = __intel_pstate_cpu_init(policy);
2242
2243         if (ret)
2244                 return ret;
2245
2246         if (IS_ENABLED(CONFIG_CPU_FREQ_DEFAULT_GOV_PERFORMANCE))
2247                 policy->policy = CPUFREQ_POLICY_PERFORMANCE;
2248         else
2249                 policy->policy = CPUFREQ_POLICY_POWERSAVE;
2250
2251         return 0;
2252 }
2253
2254 static struct cpufreq_driver intel_pstate = {
2255         .flags          = CPUFREQ_CONST_LOOPS,
2256         .verify         = intel_pstate_verify_policy,
2257         .setpolicy      = intel_pstate_set_policy,
2258         .suspend        = intel_pstate_hwp_save_state,
2259         .resume         = intel_pstate_resume,
2260         .init           = intel_pstate_cpu_init,
2261         .exit           = intel_pstate_cpu_exit,
2262         .stop_cpu       = intel_pstate_stop_cpu,
2263         .update_limits  = intel_pstate_update_limits,
2264         .name           = "intel_pstate",
2265 };
2266
2267 static int intel_cpufreq_verify_policy(struct cpufreq_policy_data *policy)
2268 {
2269         struct cpudata *cpu = all_cpu_data[policy->cpu];
2270
2271         update_turbo_state();
2272         cpufreq_verify_within_limits(policy, policy->cpuinfo.min_freq,
2273                                      intel_pstate_get_max_freq(cpu));
2274
2275         intel_pstate_adjust_policy_max(cpu, policy);
2276
2277         intel_pstate_update_perf_limits(cpu, policy->min, policy->max);
2278
2279         return 0;
2280 }
2281
2282 /* Use of trace in passive mode:
2283  *
2284  * In passive mode the trace core_busy field (also known as the
2285  * performance field, and lablelled as such on the graphs; also known as
2286  * core_avg_perf) is not needed and so is re-assigned to indicate if the
2287  * driver call was via the normal or fast switch path. Various graphs
2288  * output from the intel_pstate_tracer.py utility that include core_busy
2289  * (or performance or core_avg_perf) have a fixed y-axis from 0 to 100%,
2290  * so we use 10 to indicate the the normal path through the driver, and
2291  * 90 to indicate the fast switch path through the driver.
2292  * The scaled_busy field is not used, and is set to 0.
2293  */
2294
2295 #define INTEL_PSTATE_TRACE_TARGET 10
2296 #define INTEL_PSTATE_TRACE_FAST_SWITCH 90
2297
2298 static void intel_cpufreq_trace(struct cpudata *cpu, unsigned int trace_type, int old_pstate)
2299 {
2300         struct sample *sample;
2301
2302         if (!trace_pstate_sample_enabled())
2303                 return;
2304
2305         if (!intel_pstate_sample(cpu, ktime_get()))
2306                 return;
2307
2308         sample = &cpu->sample;
2309         trace_pstate_sample(trace_type,
2310                 0,
2311                 old_pstate,
2312                 cpu->pstate.current_pstate,
2313                 sample->mperf,
2314                 sample->aperf,
2315                 sample->tsc,
2316                 get_avg_frequency(cpu),
2317                 fp_toint(cpu->iowait_boost * 100));
2318 }
2319
2320 static int intel_cpufreq_target(struct cpufreq_policy *policy,
2321                                 unsigned int target_freq,
2322                                 unsigned int relation)
2323 {
2324         struct cpudata *cpu = all_cpu_data[policy->cpu];
2325         struct cpufreq_freqs freqs;
2326         int target_pstate, old_pstate;
2327
2328         update_turbo_state();
2329
2330         freqs.old = policy->cur;
2331         freqs.new = target_freq;
2332
2333         cpufreq_freq_transition_begin(policy, &freqs);
2334         switch (relation) {
2335         case CPUFREQ_RELATION_L:
2336                 target_pstate = DIV_ROUND_UP(freqs.new, cpu->pstate.scaling);
2337                 break;
2338         case CPUFREQ_RELATION_H:
2339                 target_pstate = freqs.new / cpu->pstate.scaling;
2340                 break;
2341         default:
2342                 target_pstate = DIV_ROUND_CLOSEST(freqs.new, cpu->pstate.scaling);
2343                 break;
2344         }
2345         target_pstate = intel_pstate_prepare_request(cpu, target_pstate);
2346         old_pstate = cpu->pstate.current_pstate;
2347         if (target_pstate != cpu->pstate.current_pstate) {
2348                 cpu->pstate.current_pstate = target_pstate;
2349                 wrmsrl_on_cpu(policy->cpu, MSR_IA32_PERF_CTL,
2350                               pstate_funcs.get_val(cpu, target_pstate));
2351         }
2352         freqs.new = target_pstate * cpu->pstate.scaling;
2353         intel_cpufreq_trace(cpu, INTEL_PSTATE_TRACE_TARGET, old_pstate);
2354         cpufreq_freq_transition_end(policy, &freqs, false);
2355
2356         return 0;
2357 }
2358
2359 static unsigned int intel_cpufreq_fast_switch(struct cpufreq_policy *policy,
2360                                               unsigned int target_freq)
2361 {
2362         struct cpudata *cpu = all_cpu_data[policy->cpu];
2363         int target_pstate, old_pstate;
2364
2365         update_turbo_state();
2366
2367         target_pstate = DIV_ROUND_UP(target_freq, cpu->pstate.scaling);
2368         target_pstate = intel_pstate_prepare_request(cpu, target_pstate);
2369         old_pstate = cpu->pstate.current_pstate;
2370         intel_pstate_update_pstate(cpu, target_pstate);
2371         intel_cpufreq_trace(cpu, INTEL_PSTATE_TRACE_FAST_SWITCH, old_pstate);
2372         return target_pstate * cpu->pstate.scaling;
2373 }
2374
2375 static int intel_cpufreq_cpu_init(struct cpufreq_policy *policy)
2376 {
2377         int max_state, turbo_max, min_freq, max_freq, ret;
2378         struct freq_qos_request *req;
2379         struct cpudata *cpu;
2380         struct device *dev;
2381
2382         dev = get_cpu_device(policy->cpu);
2383         if (!dev)
2384                 return -ENODEV;
2385
2386         ret = __intel_pstate_cpu_init(policy);
2387         if (ret)
2388                 return ret;
2389
2390         policy->cpuinfo.transition_latency = INTEL_CPUFREQ_TRANSITION_LATENCY;
2391         policy->transition_delay_us = INTEL_CPUFREQ_TRANSITION_DELAY;
2392         /* This reflects the intel_pstate_get_cpu_pstates() setting. */
2393         policy->cur = policy->cpuinfo.min_freq;
2394
2395         req = kcalloc(2, sizeof(*req), GFP_KERNEL);
2396         if (!req) {
2397                 ret = -ENOMEM;
2398                 goto pstate_exit;
2399         }
2400
2401         cpu = all_cpu_data[policy->cpu];
2402
2403         if (hwp_active)
2404                 intel_pstate_get_hwp_max(policy->cpu, &turbo_max, &max_state);
2405         else
2406                 turbo_max = cpu->pstate.turbo_pstate;
2407
2408         min_freq = DIV_ROUND_UP(turbo_max * global.min_perf_pct, 100);
2409         min_freq *= cpu->pstate.scaling;
2410         max_freq = DIV_ROUND_UP(turbo_max * global.max_perf_pct, 100);
2411         max_freq *= cpu->pstate.scaling;
2412
2413         ret = freq_qos_add_request(&policy->constraints, req, FREQ_QOS_MIN,
2414                                    min_freq);
2415         if (ret < 0) {
2416                 dev_err(dev, "Failed to add min-freq constraint (%d)\n", ret);
2417                 goto free_req;
2418         }
2419
2420         ret = freq_qos_add_request(&policy->constraints, req + 1, FREQ_QOS_MAX,
2421                                    max_freq);
2422         if (ret < 0) {
2423                 dev_err(dev, "Failed to add max-freq constraint (%d)\n", ret);
2424                 goto remove_min_req;
2425         }
2426
2427         policy->driver_data = req;
2428
2429         return 0;
2430
2431 remove_min_req:
2432         freq_qos_remove_request(req);
2433 free_req:
2434         kfree(req);
2435 pstate_exit:
2436         intel_pstate_exit_perf_limits(policy);
2437
2438         return ret;
2439 }
2440
2441 static int intel_cpufreq_cpu_exit(struct cpufreq_policy *policy)
2442 {
2443         struct freq_qos_request *req;
2444
2445         req = policy->driver_data;
2446
2447         freq_qos_remove_request(req + 1);
2448         freq_qos_remove_request(req);
2449         kfree(req);
2450
2451         return intel_pstate_cpu_exit(policy);
2452 }
2453
2454 static struct cpufreq_driver intel_cpufreq = {
2455         .flags          = CPUFREQ_CONST_LOOPS,
2456         .verify         = intel_cpufreq_verify_policy,
2457         .target         = intel_cpufreq_target,
2458         .fast_switch    = intel_cpufreq_fast_switch,
2459         .init           = intel_cpufreq_cpu_init,
2460         .exit           = intel_cpufreq_cpu_exit,
2461         .stop_cpu       = intel_cpufreq_stop_cpu,
2462         .update_limits  = intel_pstate_update_limits,
2463         .name           = "intel_cpufreq",
2464 };
2465
2466 static struct cpufreq_driver *default_driver = &intel_pstate;
2467
2468 static void intel_pstate_driver_cleanup(void)
2469 {
2470         unsigned int cpu;
2471
2472         get_online_cpus();
2473         for_each_online_cpu(cpu) {
2474                 if (all_cpu_data[cpu]) {
2475                         if (intel_pstate_driver == &intel_pstate)
2476                                 intel_pstate_clear_update_util_hook(cpu);
2477
2478                         kfree(all_cpu_data[cpu]);
2479                         all_cpu_data[cpu] = NULL;
2480                 }
2481         }
2482         put_online_cpus();
2483         intel_pstate_driver = NULL;
2484 }
2485
2486 static int intel_pstate_register_driver(struct cpufreq_driver *driver)
2487 {
2488         int ret;
2489
2490         memset(&global, 0, sizeof(global));
2491         global.max_perf_pct = 100;
2492
2493         intel_pstate_driver = driver;
2494         ret = cpufreq_register_driver(intel_pstate_driver);
2495         if (ret) {
2496                 intel_pstate_driver_cleanup();
2497                 return ret;
2498         }
2499
2500         global.min_perf_pct = min_perf_pct_min();
2501
2502         return 0;
2503 }
2504
2505 static int intel_pstate_unregister_driver(void)
2506 {
2507         if (hwp_active)
2508                 return -EBUSY;
2509
2510         cpufreq_unregister_driver(intel_pstate_driver);
2511         intel_pstate_driver_cleanup();
2512
2513         return 0;
2514 }
2515
2516 static ssize_t intel_pstate_show_status(char *buf)
2517 {
2518         if (!intel_pstate_driver)
2519                 return sprintf(buf, "off\n");
2520
2521         return sprintf(buf, "%s\n", intel_pstate_driver == &intel_pstate ?
2522                                         "active" : "passive");
2523 }
2524
2525 static int intel_pstate_update_status(const char *buf, size_t size)
2526 {
2527         int ret;
2528
2529         if (size == 3 && !strncmp(buf, "off", size))
2530                 return intel_pstate_driver ?
2531                         intel_pstate_unregister_driver() : -EINVAL;
2532
2533         if (size == 6 && !strncmp(buf, "active", size)) {
2534                 if (intel_pstate_driver) {
2535                         if (intel_pstate_driver == &intel_pstate)
2536                                 return 0;
2537
2538                         ret = intel_pstate_unregister_driver();
2539                         if (ret)
2540                                 return ret;
2541                 }
2542
2543                 return intel_pstate_register_driver(&intel_pstate);
2544         }
2545
2546         if (size == 7 && !strncmp(buf, "passive", size)) {
2547                 if (intel_pstate_driver) {
2548                         if (intel_pstate_driver == &intel_cpufreq)
2549                                 return 0;
2550
2551                         ret = intel_pstate_unregister_driver();
2552                         if (ret)
2553                                 return ret;
2554                 }
2555
2556                 return intel_pstate_register_driver(&intel_cpufreq);
2557         }
2558
2559         return -EINVAL;
2560 }
2561
2562 static int no_load __initdata;
2563 static int no_hwp __initdata;
2564 static int hwp_only __initdata;
2565 static unsigned int force_load __initdata;
2566
2567 static int __init intel_pstate_msrs_not_valid(void)
2568 {
2569         if (!pstate_funcs.get_max() ||
2570             !pstate_funcs.get_min() ||
2571             !pstate_funcs.get_turbo())
2572                 return -ENODEV;
2573
2574         return 0;
2575 }
2576
2577 static void __init copy_cpu_funcs(struct pstate_funcs *funcs)
2578 {
2579         pstate_funcs.get_max   = funcs->get_max;
2580         pstate_funcs.get_max_physical = funcs->get_max_physical;
2581         pstate_funcs.get_min   = funcs->get_min;
2582         pstate_funcs.get_turbo = funcs->get_turbo;
2583         pstate_funcs.get_scaling = funcs->get_scaling;
2584         pstate_funcs.get_val   = funcs->get_val;
2585         pstate_funcs.get_vid   = funcs->get_vid;
2586         pstate_funcs.get_aperf_mperf_shift = funcs->get_aperf_mperf_shift;
2587 }
2588
2589 #ifdef CONFIG_ACPI
2590
2591 static bool __init intel_pstate_no_acpi_pss(void)
2592 {
2593         int i;
2594
2595         for_each_possible_cpu(i) {
2596                 acpi_status status;
2597                 union acpi_object *pss;
2598                 struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
2599                 struct acpi_processor *pr = per_cpu(processors, i);
2600
2601                 if (!pr)
2602                         continue;
2603
2604                 status = acpi_evaluate_object(pr->handle, "_PSS", NULL, &buffer);
2605                 if (ACPI_FAILURE(status))
2606                         continue;
2607
2608                 pss = buffer.pointer;
2609                 if (pss && pss->type == ACPI_TYPE_PACKAGE) {
2610                         kfree(pss);
2611                         return false;
2612                 }
2613
2614                 kfree(pss);
2615         }
2616
2617         pr_debug("ACPI _PSS not found\n");
2618         return true;
2619 }
2620
2621 static bool __init intel_pstate_no_acpi_pcch(void)
2622 {
2623         acpi_status status;
2624         acpi_handle handle;
2625
2626         status = acpi_get_handle(NULL, "\\_SB", &handle);
2627         if (ACPI_FAILURE(status))
2628                 goto not_found;
2629
2630         if (acpi_has_method(handle, "PCCH"))
2631                 return false;
2632
2633 not_found:
2634         pr_debug("ACPI PCCH not found\n");
2635         return true;
2636 }
2637
2638 static bool __init intel_pstate_has_acpi_ppc(void)
2639 {
2640         int i;
2641
2642         for_each_possible_cpu(i) {
2643                 struct acpi_processor *pr = per_cpu(processors, i);
2644
2645                 if (!pr)
2646                         continue;
2647                 if (acpi_has_method(pr->handle, "_PPC"))
2648                         return true;
2649         }
2650         pr_debug("ACPI _PPC not found\n");
2651         return false;
2652 }
2653
2654 enum {
2655         PSS,
2656         PPC,
2657 };
2658
2659 /* Hardware vendor-specific info that has its own power management modes */
2660 static struct acpi_platform_list plat_info[] __initdata = {
2661         {"HP    ", "ProLiant", 0, ACPI_SIG_FADT, all_versions, NULL, PSS},
2662         {"ORACLE", "X4-2    ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
2663         {"ORACLE", "X4-2L   ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
2664         {"ORACLE", "X4-2B   ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
2665         {"ORACLE", "X3-2    ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
2666         {"ORACLE", "X3-2L   ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
2667         {"ORACLE", "X3-2B   ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
2668         {"ORACLE", "X4470M2 ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
2669         {"ORACLE", "X4270M3 ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
2670         {"ORACLE", "X4270M2 ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
2671         {"ORACLE", "X4170M2 ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
2672         {"ORACLE", "X4170 M3", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
2673         {"ORACLE", "X4275 M3", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
2674         {"ORACLE", "X6-2    ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
2675         {"ORACLE", "Sudbury ", 0, ACPI_SIG_FADT, all_versions, NULL, PPC},
2676         { } /* End */
2677 };
2678
2679 static bool __init intel_pstate_platform_pwr_mgmt_exists(void)
2680 {
2681         const struct x86_cpu_id *id;
2682         u64 misc_pwr;
2683         int idx;
2684
2685         id = x86_match_cpu(intel_pstate_cpu_oob_ids);
2686         if (id) {
2687                 rdmsrl(MSR_MISC_PWR_MGMT, misc_pwr);
2688                 if (misc_pwr & (1 << 8)) {
2689                         pr_debug("Bit 8 in the MISC_PWR_MGMT MSR set\n");
2690                         return true;
2691                 }
2692         }
2693
2694         idx = acpi_match_platform_list(plat_info);
2695         if (idx < 0)
2696                 return false;
2697
2698         switch (plat_info[idx].data) {
2699         case PSS:
2700                 if (!intel_pstate_no_acpi_pss())
2701                         return false;
2702
2703                 return intel_pstate_no_acpi_pcch();
2704         case PPC:
2705                 return intel_pstate_has_acpi_ppc() && !force_load;
2706         }
2707
2708         return false;
2709 }
2710
2711 static void intel_pstate_request_control_from_smm(void)
2712 {
2713         /*
2714          * It may be unsafe to request P-states control from SMM if _PPC support
2715          * has not been enabled.
2716          */
2717         if (acpi_ppc)
2718                 acpi_processor_pstate_control();
2719 }
2720 #else /* CONFIG_ACPI not enabled */
2721 static inline bool intel_pstate_platform_pwr_mgmt_exists(void) { return false; }
2722 static inline bool intel_pstate_has_acpi_ppc(void) { return false; }
2723 static inline void intel_pstate_request_control_from_smm(void) {}
2724 #endif /* CONFIG_ACPI */
2725
2726 #define INTEL_PSTATE_HWP_BROADWELL      0x01
2727
2728 #define ICPU_HWP(model, hwp_mode) \
2729         { X86_VENDOR_INTEL, 6, model, X86_FEATURE_HWP, hwp_mode }
2730
2731 static const struct x86_cpu_id hwp_support_ids[] __initconst = {
2732         ICPU_HWP(INTEL_FAM6_BROADWELL_X, INTEL_PSTATE_HWP_BROADWELL),
2733         ICPU_HWP(INTEL_FAM6_BROADWELL_D, INTEL_PSTATE_HWP_BROADWELL),
2734         ICPU_HWP(X86_MODEL_ANY, 0),
2735         {}
2736 };
2737
2738 static int __init intel_pstate_init(void)
2739 {
2740         const struct x86_cpu_id *id;
2741         int rc;
2742
2743         if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL)
2744                 return -ENODEV;
2745
2746         if (no_load)
2747                 return -ENODEV;
2748
2749         id = x86_match_cpu(hwp_support_ids);
2750         if (id) {
2751                 copy_cpu_funcs(&core_funcs);
2752                 if (!no_hwp) {
2753                         hwp_active++;
2754                         hwp_mode_bdw = id->driver_data;
2755                         intel_pstate.attr = hwp_cpufreq_attrs;
2756                         goto hwp_cpu_matched;
2757                 }
2758         } else {
2759                 id = x86_match_cpu(intel_pstate_cpu_ids);
2760                 if (!id) {
2761                         pr_info("CPU model not supported\n");
2762                         return -ENODEV;
2763                 }
2764
2765                 copy_cpu_funcs((struct pstate_funcs *)id->driver_data);
2766         }
2767
2768         if (intel_pstate_msrs_not_valid()) {
2769                 pr_info("Invalid MSRs\n");
2770                 return -ENODEV;
2771         }
2772
2773 hwp_cpu_matched:
2774         /*
2775          * The Intel pstate driver will be ignored if the platform
2776          * firmware has its own power management modes.
2777          */
2778         if (intel_pstate_platform_pwr_mgmt_exists()) {
2779                 pr_info("P-states controlled by the platform\n");
2780                 return -ENODEV;
2781         }
2782
2783         if (!hwp_active && hwp_only)
2784                 return -ENOTSUPP;
2785
2786         pr_info("Intel P-state driver initializing\n");
2787
2788         all_cpu_data = vzalloc(array_size(sizeof(void *), num_possible_cpus()));
2789         if (!all_cpu_data)
2790                 return -ENOMEM;
2791
2792         intel_pstate_request_control_from_smm();
2793
2794         intel_pstate_sysfs_expose_params();
2795
2796         mutex_lock(&intel_pstate_driver_lock);
2797         rc = intel_pstate_register_driver(default_driver);
2798         mutex_unlock(&intel_pstate_driver_lock);
2799         if (rc)
2800                 return rc;
2801
2802         if (hwp_active)
2803                 pr_info("HWP enabled\n");
2804
2805         return 0;
2806 }
2807 device_initcall(intel_pstate_init);
2808
2809 static int __init intel_pstate_setup(char *str)
2810 {
2811         if (!str)
2812                 return -EINVAL;
2813
2814         if (!strcmp(str, "disable")) {
2815                 no_load = 1;
2816         } else if (!strcmp(str, "passive")) {
2817                 pr_info("Passive mode enabled\n");
2818                 default_driver = &intel_cpufreq;
2819                 no_hwp = 1;
2820         }
2821         if (!strcmp(str, "no_hwp")) {
2822                 pr_info("HWP disabled\n");
2823                 no_hwp = 1;
2824         }
2825         if (!strcmp(str, "force"))
2826                 force_load = 1;
2827         if (!strcmp(str, "hwp_only"))
2828                 hwp_only = 1;
2829         if (!strcmp(str, "per_cpu_perf_limits"))
2830                 per_cpu_limits = true;
2831
2832 #ifdef CONFIG_ACPI
2833         if (!strcmp(str, "support_acpi_ppc"))
2834                 acpi_ppc = true;
2835 #endif
2836
2837         return 0;
2838 }
2839 early_param("intel_pstate", intel_pstate_setup);
2840
2841 MODULE_AUTHOR("Dirk Brandewie <dirk.j.brandewie@intel.com>");
2842 MODULE_DESCRIPTION("'intel_pstate' - P state driver Intel Core processors");
2843 MODULE_LICENSE("GPL");