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