drivers/cpufreq/intel_pstate.c

   1 /*
   2  * intel_pstate.c: Native P state management for Intel processors
   3  *
   4  * (C) Copyright 2012 Intel Corporation
   5  * Author: Dirk Brandewie <dirk.j.brandewie@intel.com>
   6  *
   7  * This program is free software; you can redistribute it and/or
   8  * modify it under the terms of the GNU General Public License
   9  * as published by the Free Software Foundation; version 2
  10  * of the License.
  11  */
  12
  13 #include <linux/kernel.h>
  14 #include <linux/kernel_stat.h>
  15 #include <linux/module.h>
  16 #include <linux/ktime.h>
  17 #include <linux/hrtimer.h>
  18 #include <linux/tick.h>
  19 #include <linux/slab.h>
  20 #include <linux/sched.h>
  21 #include <linux/list.h>
  22 #include <linux/cpu.h>
  23 #include <linux/cpufreq.h>
  24 #include <linux/sysfs.h>
  25 #include <linux/types.h>
  26 #include <linux/fs.h>
  27 #include <linux/debugfs.h>
  28 #include <linux/acpi.h>
  29 #include <linux/vmalloc.h>
  30 #include <trace/events/power.h>
  31
  32 #include <asm/div64.h>
  33 #include <asm/msr.h>
  34 #include <asm/cpu_device_id.h>
  35 #include <asm/cpufeature.h>
  36
  37 #define ATOM_RATIOS             0x66a
  38 #define ATOM_VIDS               0x66b
  39 #define ATOM_TURBO_RATIOS       0x66c
  40 #define ATOM_TURBO_VIDS         0x66d
  41
  42 #define FRAC_BITS 8
  43 #define int_tofp(X) ((int64_t)(X) << FRAC_BITS)
  44 #define fp_toint(X) ((X) >> FRAC_BITS)
  45
  46 static inline int32_t mul_fp(int32_t x, int32_t y)
  47 {
  48         return ((int64_t)x * (int64_t)y) >> FRAC_BITS;
  49 }
  50
  51 static inline int32_t div_fp(s64 x, s64 y)
  52 {
  53         return div64_s64((int64_t)x << FRAC_BITS, y);
  54 }
  55
  56 static inline int ceiling_fp(int32_t x)
  57 {
  58         int mask, ret;
  59
  60         ret = fp_toint(x);
  61         mask = (1 << FRAC_BITS) - 1;
  62         if (x & mask)
  63                 ret += 1;
  64         return ret;
  65 }
  66
  67 /**
  68  * struct sample -      Store performance sample
  69  * @core_pct_busy:      Ratio of APERF/MPERF in percent, which is actual
  70  *                      performance during last sample period
  71  * @busy_scaled:        Scaled busy value which is used to calculate next
  72  *                      P state. This can be different than core_pct_busy
  73  *                      to account for cpu idle period
  74  * @aperf:              Difference of actual performance frequency clock count
  75  *                      read from APERF MSR between last and current sample
  76  * @mperf:              Difference of maximum performance frequency clock count
  77  *                      read from MPERF MSR between last and current sample
  78  * @tsc:                Difference of time stamp counter between last and
  79  *                      current sample
  80  * @freq:               Effective frequency calculated from APERF/MPERF
  81  * @time:               Current time from scheduler
  82  *
  83  * This structure is used in the cpudata structure to store performance sample
  84  * data for choosing next P State.
  85  */
  86 struct sample {
  87         int32_t core_pct_busy;
  88         int32_t busy_scaled;
  89         u64 aperf;
  90         u64 mperf;
  91         u64 tsc;
  92         int freq;
  93         u64 time;
  94 };
  95
  96 /**
  97  * struct pstate_data - Store P state data
  98  * @current_pstate:     Current requested P state
  99  * @min_pstate:         Min P state possible for this platform
 100  * @max_pstate:         Max P state possible for this platform
 101  * @max_pstate_physical:This is physical Max P state for a processor
 102  *                      This can be higher than the max_pstate which can
 103  *                      be limited by platform thermal design power limits
 104  * @scaling:            Scaling factor to  convert frequency to cpufreq
 105  *                      frequency units
 106  * @turbo_pstate:       Max Turbo P state possible for this platform
 107  *
 108  * Stores the per cpu model P state limits and current P state.
 109  */
 110 struct pstate_data {
 111         int     current_pstate;
 112         int     min_pstate;
 113         int     max_pstate;
 114         int     max_pstate_physical;
 115         int     scaling;
 116         int     turbo_pstate;
 117 };
 118
 119 /**
 120  * struct vid_data -    Stores voltage information data
 121  * @min:                VID data for this platform corresponding to
 122  *                      the lowest P state
 123  * @max:                VID data corresponding to the highest P State.
 124  * @turbo:              VID data for turbo P state
 125  * @ratio:              Ratio of (vid max - vid min) /
 126  *                      (max P state - Min P State)
 127  *
 128  * Stores the voltage data for DVFS (Dynamic Voltage and Frequency Scaling)
 129  * This data is used in Atom platforms, where in addition to target P state,
 130  * the voltage data needs to be specified to select next P State.
 131  */
 132 struct vid_data {
 133         int min;
 134         int max;
 135         int turbo;
 136         int32_t ratio;
 137 };
 138
 139 /**
 140  * struct _pid -        Stores PID data
 141  * @setpoint:           Target set point for busyness or performance
 142  * @integral:           Storage for accumulated error values
 143  * @p_gain:             PID proportional gain
 144  * @i_gain:             PID integral gain
 145  * @d_gain:             PID derivative gain
 146  * @deadband:           PID deadband
 147  * @last_err:           Last error storage for integral part of PID calculation
 148  *
 149  * Stores PID coefficients and last error for PID controller.
 150  */
 151 struct _pid {
 152         int setpoint;
 153         int32_t integral;
 154         int32_t p_gain;
 155         int32_t i_gain;
 156         int32_t d_gain;
 157         int deadband;
 158         int32_t last_err;
 159 };
 160
 161 /**
 162  * struct cpudata -     Per CPU instance data storage
 163  * @cpu:                CPU number for this instance data
 164  * @update_util:        CPUFreq utility callback information
 165  * @pstate:             Stores P state limits for this CPU
 166  * @vid:                Stores VID limits for this CPU
 167  * @pid:                Stores PID parameters for this CPU
 168  * @last_sample_time:   Last Sample time
 169  * @prev_aperf:         Last APERF value read from APERF MSR
 170  * @prev_mperf:         Last MPERF value read from MPERF MSR
 171  * @prev_tsc:           Last timestamp counter (TSC) value
 172  * @prev_cummulative_iowait: IO Wait time difference from last and
 173  *                      current sample
 174  * @sample:             Storage for storing last Sample data
 175  *
 176  * This structure stores per CPU instance data for all CPUs.
 177  */
 178 struct cpudata {
 179         int cpu;
 180
 181         struct update_util_data update_util;
 182
 183         struct pstate_data pstate;
 184         struct vid_data vid;
 185         struct _pid pid;
 186
 187         u64     last_sample_time;
 188         u64     prev_aperf;
 189         u64     prev_mperf;
 190         u64     prev_tsc;
 191         u64     prev_cummulative_iowait;
 192         struct sample sample;
 193 };
 194
 195 static struct cpudata **all_cpu_data;
 196
 197 /**
 198  * struct pid_adjust_policy - Stores static PID configuration data
 199  * @sample_rate_ms:     PID calculation sample rate in ms
 200  * @sample_rate_ns:     Sample rate calculation in ns
 201  * @deadband:           PID deadband
 202  * @setpoint:           PID Setpoint
 203  * @p_gain_pct:         PID proportional gain
 204  * @i_gain_pct:         PID integral gain
 205  * @d_gain_pct:         PID derivative gain
 206  *
 207  * Stores per CPU model static PID configuration data.
 208  */
 209 struct pstate_adjust_policy {
 210         int sample_rate_ms;
 211         s64 sample_rate_ns;
 212         int deadband;
 213         int setpoint;
 214         int p_gain_pct;
 215         int d_gain_pct;
 216         int i_gain_pct;
 217 };
 218
 219 /**
 220  * struct pstate_funcs - Per CPU model specific callbacks
 221  * @get_max:            Callback to get maximum non turbo effective P state
 222  * @get_max_physical:   Callback to get maximum non turbo physical P state
 223  * @get_min:            Callback to get minimum P state
 224  * @get_turbo:          Callback to get turbo P state
 225  * @get_scaling:        Callback to get frequency scaling factor
 226  * @get_val:            Callback to convert P state to actual MSR write value
 227  * @get_vid:            Callback to get VID data for Atom platforms
 228  * @get_target_pstate:  Callback to a function to calculate next P state to use
 229  *
 230  * Core and Atom CPU models have different way to get P State limits. This
 231  * structure is used to store those callbacks.
 232  */
 233 struct pstate_funcs {
 234         int (*get_max)(void);
 235         int (*get_max_physical)(void);
 236         int (*get_min)(void);
 237         int (*get_turbo)(void);
 238         int (*get_scaling)(void);
 239         u64 (*get_val)(struct cpudata*, int pstate);
 240         void (*get_vid)(struct cpudata *);
 241         int32_t (*get_target_pstate)(struct cpudata *);
 242 };
 243
 244 /**
 245  * struct cpu_defaults- Per CPU model default config data
 246  * @pid_policy: PID config data
 247  * @funcs:              Callback function data
 248  */
 249 struct cpu_defaults {
 250         struct pstate_adjust_policy pid_policy;
 251         struct pstate_funcs funcs;
 252 };
 253
 254 static inline int32_t get_target_pstate_use_performance(struct cpudata *cpu);
 255 static inline int32_t get_target_pstate_use_cpu_load(struct cpudata *cpu);
 256
 257 static struct pstate_adjust_policy pid_params;
 258 static struct pstate_funcs pstate_funcs;
 259 static int hwp_active;
 260
 261
 262 /**
 263  * struct perf_limits - Store user and policy limits
 264  * @no_turbo:           User requested turbo state from intel_pstate sysfs
 265  * @turbo_disabled:     Platform turbo status either from msr
 266  *                      MSR_IA32_MISC_ENABLE or when maximum available pstate
 267  *                      matches the maximum turbo pstate
 268  * @max_perf_pct:       Effective maximum performance limit in percentage, this
 269  *                      is minimum of either limits enforced by cpufreq policy
 270  *                      or limits from user set limits via intel_pstate sysfs
 271  * @min_perf_pct:       Effective minimum performance limit in percentage, this
 272  *                      is maximum of either limits enforced by cpufreq policy
 273  *                      or limits from user set limits via intel_pstate sysfs
 274  * @max_perf:           This is a scaled value between 0 to 255 for max_perf_pct
 275  *                      This value is used to limit max pstate
 276  * @min_perf:           This is a scaled value between 0 to 255 for min_perf_pct
 277  *                      This value is used to limit min pstate
 278  * @max_policy_pct:     The maximum performance in percentage enforced by
 279  *                      cpufreq setpolicy interface
 280  * @max_sysfs_pct:      The maximum performance in percentage enforced by
 281  *                      intel pstate sysfs interface
 282  * @min_policy_pct:     The minimum performance in percentage enforced by
 283  *                      cpufreq setpolicy interface
 284  * @min_sysfs_pct:      The minimum performance in percentage enforced by
 285  *                      intel pstate sysfs interface
 286  *
 287  * Storage for user and policy defined limits.
 288  */
 289 struct perf_limits {
 290         int no_turbo;
 291         int turbo_disabled;
 292         int max_perf_pct;
 293         int min_perf_pct;
 294         int32_t max_perf;
 295         int32_t min_perf;
 296         int max_policy_pct;
 297         int max_sysfs_pct;
 298         int min_policy_pct;
 299         int min_sysfs_pct;
 300 };
 301
 302 static struct perf_limits performance_limits = {
 303         .no_turbo = 0,
 304         .turbo_disabled = 0,
 305         .max_perf_pct = 100,
 306         .max_perf = int_tofp(1),
 307         .min_perf_pct = 100,
 308         .min_perf = int_tofp(1),
 309         .max_policy_pct = 100,
 310         .max_sysfs_pct = 100,
 311         .min_policy_pct = 0,
 312         .min_sysfs_pct = 0,
 313 };
 314
 315 static struct perf_limits powersave_limits = {
 316         .no_turbo = 0,
 317         .turbo_disabled = 0,
 318         .max_perf_pct = 100,
 319         .max_perf = int_tofp(1),
 320         .min_perf_pct = 0,
 321         .min_perf = 0,
 322         .max_policy_pct = 100,
 323         .max_sysfs_pct = 100,
 324         .min_policy_pct = 0,
 325         .min_sysfs_pct = 0,
 326 };
 327
 328 #ifdef CONFIG_CPU_FREQ_DEFAULT_GOV_PERFORMANCE
 329 static struct perf_limits *limits = &performance_limits;
 330 #else
 331 static struct perf_limits *limits = &powersave_limits;
 332 #endif
 333
 334 static inline void pid_reset(struct _pid *pid, int setpoint, int busy,
 335                              int deadband, int integral) {
 336         pid->setpoint = int_tofp(setpoint);
 337         pid->deadband  = int_tofp(deadband);
 338         pid->integral  = int_tofp(integral);
 339         pid->last_err  = int_tofp(setpoint) - int_tofp(busy);
 340 }
 341
 342 static inline void pid_p_gain_set(struct _pid *pid, int percent)
 343 {
 344         pid->p_gain = div_fp(int_tofp(percent), int_tofp(100));
 345 }
 346
 347 static inline void pid_i_gain_set(struct _pid *pid, int percent)
 348 {
 349         pid->i_gain = div_fp(int_tofp(percent), int_tofp(100));
 350 }
 351
 352 static inline void pid_d_gain_set(struct _pid *pid, int percent)
 353 {
 354         pid->d_gain = div_fp(int_tofp(percent), int_tofp(100));
 355 }
 356
 357 static signed int pid_calc(struct _pid *pid, int32_t busy)
 358 {
 359         signed int result;
 360         int32_t pterm, dterm, fp_error;
 361         int32_t integral_limit;
 362
 363         fp_error = pid->setpoint - busy;
 364
 365         if (abs(fp_error) <= pid->deadband)
 366                 return 0;
 367
 368         pterm = mul_fp(pid->p_gain, fp_error);
 369
 370         pid->integral += fp_error;
 371
 372         /*
 373          * We limit the integral here so that it will never
 374          * get higher than 30.  This prevents it from becoming
 375          * too large an input over long periods of time and allows
 376          * it to get factored out sooner.
 377          *
 378          * The value of 30 was chosen through experimentation.
 379          */
 380         integral_limit = int_tofp(30);
 381         if (pid->integral > integral_limit)
 382                 pid->integral = integral_limit;
 383         if (pid->integral < -integral_limit)
 384                 pid->integral = -integral_limit;
 385
 386         dterm = mul_fp(pid->d_gain, fp_error - pid->last_err);
 387         pid->last_err = fp_error;
 388
 389         result = pterm + mul_fp(pid->integral, pid->i_gain) + dterm;
 390         result = result + (1 << (FRAC_BITS-1));
 391         return (signed int)fp_toint(result);
 392 }
 393
 394 static inline void intel_pstate_busy_pid_reset(struct cpudata *cpu)
 395 {
 396         pid_p_gain_set(&cpu->pid, pid_params.p_gain_pct);
 397         pid_d_gain_set(&cpu->pid, pid_params.d_gain_pct);
 398         pid_i_gain_set(&cpu->pid, pid_params.i_gain_pct);
 399
 400         pid_reset(&cpu->pid, pid_params.setpoint, 100, pid_params.deadband, 0);
 401 }
 402
 403 static inline void intel_pstate_reset_all_pid(void)
 404 {
 405         unsigned int cpu;
 406
 407         for_each_online_cpu(cpu) {
 408                 if (all_cpu_data[cpu])
 409                         intel_pstate_busy_pid_reset(all_cpu_data[cpu]);
 410         }
 411 }
 412
 413 static inline void update_turbo_state(void)
 414 {
 415         u64 misc_en;
 416         struct cpudata *cpu;
 417
 418         cpu = all_cpu_data[0];
 419         rdmsrl(MSR_IA32_MISC_ENABLE, misc_en);
 420         limits->turbo_disabled =
 421                 (misc_en & MSR_IA32_MISC_ENABLE_TURBO_DISABLE ||
 422                  cpu->pstate.max_pstate == cpu->pstate.turbo_pstate);
 423 }
 424
 425 static void intel_pstate_hwp_set(const struct cpumask *cpumask)
 426 {
 427         int min, hw_min, max, hw_max, cpu, range, adj_range;
 428         u64 value, cap;
 429
 430         rdmsrl(MSR_HWP_CAPABILITIES, cap);
 431         hw_min = HWP_LOWEST_PERF(cap);
 432         hw_max = HWP_HIGHEST_PERF(cap);
 433         range = hw_max - hw_min;
 434
 435         for_each_cpu(cpu, cpumask) {
 436                 rdmsrl_on_cpu(cpu, MSR_HWP_REQUEST, &value);
 437                 adj_range = limits->min_perf_pct * range / 100;
 438                 min = hw_min + adj_range;
 439                 value &= ~HWP_MIN_PERF(~0L);
 440                 value |= HWP_MIN_PERF(min);
 441
 442                 adj_range = limits->max_perf_pct * range / 100;
 443                 max = hw_min + adj_range;
 444                 if (limits->no_turbo) {
 445                         hw_max = HWP_GUARANTEED_PERF(cap);
 446                         if (hw_max < max)
 447                                 max = hw_max;
 448                 }
 449
 450                 value &= ~HWP_MAX_PERF(~0L);
 451                 value |= HWP_MAX_PERF(max);
 452                 wrmsrl_on_cpu(cpu, MSR_HWP_REQUEST, value);
 453         }
 454 }
 455
 456 static int intel_pstate_hwp_set_policy(struct cpufreq_policy *policy)
 457 {
 458         if (hwp_active)
 459                 intel_pstate_hwp_set(policy->cpus);
 460
 461         return 0;
 462 }
 463
 464 static void intel_pstate_hwp_set_online_cpus(void)
 465 {
 466         get_online_cpus();
 467         intel_pstate_hwp_set(cpu_online_mask);
 468         put_online_cpus();
 469 }
 470
 471 /************************** debugfs begin ************************/
 472 static int pid_param_set(void *data, u64 val)
 473 {
 474         *(u32 *)data = val;
 475         intel_pstate_reset_all_pid();
 476         return 0;
 477 }
 478
 479 static int pid_param_get(void *data, u64 *val)
 480 {
 481         *val = *(u32 *)data;
 482         return 0;
 483 }
 484 DEFINE_SIMPLE_ATTRIBUTE(fops_pid_param, pid_param_get, pid_param_set, "%llu\n");
 485
 486 struct pid_param {
 487         char *name;
 488         void *value;
 489 };
 490
 491 static struct pid_param pid_files[] = {
 492         {"sample_rate_ms", &pid_params.sample_rate_ms},
 493         {"d_gain_pct", &pid_params.d_gain_pct},
 494         {"i_gain_pct", &pid_params.i_gain_pct},
 495         {"deadband", &pid_params.deadband},
 496         {"setpoint", &pid_params.setpoint},
 497         {"p_gain_pct", &pid_params.p_gain_pct},
 498         {NULL, NULL}
 499 };
 500
 501 static void __init intel_pstate_debug_expose_params(void)
 502 {
 503         struct dentry *debugfs_parent;
 504         int i = 0;
 505
 506         if (hwp_active)
 507                 return;
 508         debugfs_parent = debugfs_create_dir("pstate_snb", NULL);
 509         if (IS_ERR_OR_NULL(debugfs_parent))
 510                 return;
 511         while (pid_files[i].name) {
 512                 debugfs_create_file(pid_files[i].name, 0660,
 513                                     debugfs_parent, pid_files[i].value,
 514                                     &fops_pid_param);
 515                 i++;
 516         }
 517 }
 518
 519 /************************** debugfs end ************************/
 520
 521 /************************** sysfs begin ************************/
 522 #define show_one(file_name, object)                                     \
 523         static ssize_t show_##file_name                                 \
 524         (struct kobject *kobj, struct attribute *attr, char *buf)       \
 525         {                                                               \
 526                 return sprintf(buf, "%u\n", limits->object);            \
 527         }
 528
 529 static ssize_t show_turbo_pct(struct kobject *kobj,
 530                                 struct attribute *attr, char *buf)
 531 {
 532         struct cpudata *cpu;
 533         int total, no_turbo, turbo_pct;
 534         uint32_t turbo_fp;
 535
 536         cpu = all_cpu_data[0];
 537
 538         total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
 539         no_turbo = cpu->pstate.max_pstate - cpu->pstate.min_pstate + 1;
 540         turbo_fp = div_fp(int_tofp(no_turbo), int_tofp(total));
 541         turbo_pct = 100 - fp_toint(mul_fp(turbo_fp, int_tofp(100)));
 542         return sprintf(buf, "%u\n", turbo_pct);
 543 }
 544
 545 static ssize_t show_num_pstates(struct kobject *kobj,
 546                                 struct attribute *attr, char *buf)
 547 {
 548         struct cpudata *cpu;
 549         int total;
 550
 551         cpu = all_cpu_data[0];
 552         total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
 553         return sprintf(buf, "%u\n", total);
 554 }
 555
 556 static ssize_t show_no_turbo(struct kobject *kobj,
 557                              struct attribute *attr, char *buf)
 558 {
 559         ssize_t ret;
 560
 561         update_turbo_state();
 562         if (limits->turbo_disabled)
 563                 ret = sprintf(buf, "%u\n", limits->turbo_disabled);
 564         else
 565                 ret = sprintf(buf, "%u\n", limits->no_turbo);
 566
 567         return ret;
 568 }
 569
 570 static ssize_t store_no_turbo(struct kobject *a, struct attribute *b,
 571                               const char *buf, size_t count)
 572 {
 573         unsigned int input;
 574         int ret;
 575
 576         ret = sscanf(buf, "%u", &input);
 577         if (ret != 1)
 578                 return -EINVAL;
 579
 580         update_turbo_state();
 581         if (limits->turbo_disabled) {
 582                 pr_warn("intel_pstate: Turbo disabled by BIOS or unavailable on processor\n");
 583                 return -EPERM;
 584         }
 585
 586         limits->no_turbo = clamp_t(int, input, 0, 1);
 587
 588         if (hwp_active)
 589                 intel_pstate_hwp_set_online_cpus();
 590
 591         return count;
 592 }
 593
 594 static ssize_t store_max_perf_pct(struct kobject *a, struct attribute *b,
 595                                   const char *buf, size_t count)
 596 {
 597         unsigned int input;
 598         int ret;
 599
 600         ret = sscanf(buf, "%u", &input);
 601         if (ret != 1)
 602                 return -EINVAL;
 603
 604         limits->max_sysfs_pct = clamp_t(int, input, 0 , 100);
 605         limits->max_perf_pct = min(limits->max_policy_pct,
 606                                    limits->max_sysfs_pct);
 607         limits->max_perf_pct = max(limits->min_policy_pct,
 608                                    limits->max_perf_pct);
 609         limits->max_perf_pct = max(limits->min_perf_pct,
 610                                    limits->max_perf_pct);
 611         limits->max_perf = div_fp(int_tofp(limits->max_perf_pct),
 612                                   int_tofp(100));
 613
 614         if (hwp_active)
 615                 intel_pstate_hwp_set_online_cpus();
 616         return count;
 617 }
 618
 619 static ssize_t store_min_perf_pct(struct kobject *a, struct attribute *b,
 620                                   const char *buf, size_t count)
 621 {
 622         unsigned int input;
 623         int ret;
 624
 625         ret = sscanf(buf, "%u", &input);
 626         if (ret != 1)
 627                 return -EINVAL;
 628
 629         limits->min_sysfs_pct = clamp_t(int, input, 0 , 100);
 630         limits->min_perf_pct = max(limits->min_policy_pct,
 631                                    limits->min_sysfs_pct);
 632         limits->min_perf_pct = min(limits->max_policy_pct,
 633                                    limits->min_perf_pct);
 634         limits->min_perf_pct = min(limits->max_perf_pct,
 635                                    limits->min_perf_pct);
 636         limits->min_perf = div_fp(int_tofp(limits->min_perf_pct),
 637                                   int_tofp(100));
 638
 639         if (hwp_active)
 640                 intel_pstate_hwp_set_online_cpus();
 641         return count;
 642 }
 643
 644 show_one(max_perf_pct, max_perf_pct);
 645 show_one(min_perf_pct, min_perf_pct);
 646
 647 define_one_global_rw(no_turbo);
 648 define_one_global_rw(max_perf_pct);
 649 define_one_global_rw(min_perf_pct);
 650 define_one_global_ro(turbo_pct);
 651 define_one_global_ro(num_pstates);
 652
 653 static struct attribute *intel_pstate_attributes[] = {
 654         &no_turbo.attr,
 655         &max_perf_pct.attr,
 656         &min_perf_pct.attr,
 657         &turbo_pct.attr,
 658         &num_pstates.attr,
 659         NULL
 660 };
 661
 662 static struct attribute_group intel_pstate_attr_group = {
 663         .attrs = intel_pstate_attributes,
 664 };
 665
 666 static void __init intel_pstate_sysfs_expose_params(void)
 667 {
 668         struct kobject *intel_pstate_kobject;
 669         int rc;
 670
 671         intel_pstate_kobject = kobject_create_and_add("intel_pstate",
 672                                                 &cpu_subsys.dev_root->kobj);
 673         BUG_ON(!intel_pstate_kobject);
 674         rc = sysfs_create_group(intel_pstate_kobject, &intel_pstate_attr_group);
 675         BUG_ON(rc);
 676 }
 677 /************************** sysfs end ************************/
 678
 679 static void intel_pstate_hwp_enable(struct cpudata *cpudata)
 680 {
 681         /* First disable HWP notification interrupt as we don't process them */
 682         wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_INTERRUPT, 0x00);
 683
 684         wrmsrl_on_cpu(cpudata->cpu, MSR_PM_ENABLE, 0x1);
 685 }
 686
 687 static int atom_get_min_pstate(void)
 688 {
 689         u64 value;
 690
 691         rdmsrl(ATOM_RATIOS, value);
 692         return (value >> 8) & 0x7F;
 693 }
 694
 695 static int atom_get_max_pstate(void)
 696 {
 697         u64 value;
 698
 699         rdmsrl(ATOM_RATIOS, value);
 700         return (value >> 16) & 0x7F;
 701 }
 702
 703 static int atom_get_turbo_pstate(void)
 704 {
 705         u64 value;
 706
 707         rdmsrl(ATOM_TURBO_RATIOS, value);
 708         return value & 0x7F;
 709 }
 710
 711 static u64 atom_get_val(struct cpudata *cpudata, int pstate)
 712 {
 713         u64 val;
 714         int32_t vid_fp;
 715         u32 vid;
 716
 717         val = (u64)pstate << 8;
 718         if (limits->no_turbo && !limits->turbo_disabled)
 719                 val |= (u64)1 << 32;
 720
 721         vid_fp = cpudata->vid.min + mul_fp(
 722                 int_tofp(pstate - cpudata->pstate.min_pstate),
 723                 cpudata->vid.ratio);
 724
 725         vid_fp = clamp_t(int32_t, vid_fp, cpudata->vid.min, cpudata->vid.max);
 726         vid = ceiling_fp(vid_fp);
 727
 728         if (pstate > cpudata->pstate.max_pstate)
 729                 vid = cpudata->vid.turbo;
 730
 731         return val | vid;
 732 }
 733
 734 static int silvermont_get_scaling(void)
 735 {
 736         u64 value;
 737         int i;
 738         /* Defined in Table 35-6 from SDM (Sept 2015) */
 739         static int silvermont_freq_table[] = {
 740                 83300, 100000, 133300, 116700, 80000};
 741
 742         rdmsrl(MSR_FSB_FREQ, value);
 743         i = value & 0x7;
 744         WARN_ON(i > 4);
 745
 746         return silvermont_freq_table[i];
 747 }
 748
 749 static int airmont_get_scaling(void)
 750 {
 751         u64 value;
 752         int i;
 753         /* Defined in Table 35-10 from SDM (Sept 2015) */
 754         static int airmont_freq_table[] = {
 755                 83300, 100000, 133300, 116700, 80000,
 756                 93300, 90000, 88900, 87500};
 757
 758         rdmsrl(MSR_FSB_FREQ, value);
 759         i = value & 0xF;
 760         WARN_ON(i > 8);
 761
 762         return airmont_freq_table[i];
 763 }
 764
 765 static void atom_get_vid(struct cpudata *cpudata)
 766 {
 767         u64 value;
 768
 769         rdmsrl(ATOM_VIDS, value);
 770         cpudata->vid.min = int_tofp((value >> 8) & 0x7f);
 771         cpudata->vid.max = int_tofp((value >> 16) & 0x7f);
 772         cpudata->vid.ratio = div_fp(
 773                 cpudata->vid.max - cpudata->vid.min,
 774                 int_tofp(cpudata->pstate.max_pstate -
 775                         cpudata->pstate.min_pstate));
 776
 777         rdmsrl(ATOM_TURBO_VIDS, value);
 778         cpudata->vid.turbo = value & 0x7f;
 779 }
 780
 781 static int core_get_min_pstate(void)
 782 {
 783         u64 value;
 784
 785         rdmsrl(MSR_PLATFORM_INFO, value);
 786         return (value >> 40) & 0xFF;
 787 }
 788
 789 static int core_get_max_pstate_physical(void)
 790 {
 791         u64 value;
 792
 793         rdmsrl(MSR_PLATFORM_INFO, value);
 794         return (value >> 8) & 0xFF;
 795 }
 796
 797 static int core_get_max_pstate(void)
 798 {
 799         u64 tar;
 800         u64 plat_info;
 801         int max_pstate;
 802         int err;
 803
 804         rdmsrl(MSR_PLATFORM_INFO, plat_info);
 805         max_pstate = (plat_info >> 8) & 0xFF;
 806
 807         err = rdmsrl_safe(MSR_TURBO_ACTIVATION_RATIO, &tar);
 808         if (!err) {
 809                 /* Do some sanity checking for safety */
 810                 if (plat_info & 0x600000000) {
 811                         u64 tdp_ctrl;
 812                         u64 tdp_ratio;
 813                         int tdp_msr;
 814
 815                         err = rdmsrl_safe(MSR_CONFIG_TDP_CONTROL, &tdp_ctrl);
 816                         if (err)
 817                                 goto skip_tar;
 818
 819                         tdp_msr = MSR_CONFIG_TDP_NOMINAL + tdp_ctrl;
 820                         err = rdmsrl_safe(tdp_msr, &tdp_ratio);
 821                         if (err)
 822                                 goto skip_tar;
 823
 824                         /* For level 1 and 2, bits[23:16] contain the ratio */
 825                         if (tdp_ctrl)
 826                                 tdp_ratio >>= 16;
 827
 828                         tdp_ratio &= 0xff; /* ratios are only 8 bits long */
 829                         if (tdp_ratio - 1 == tar) {
 830                                 max_pstate = tar;
 831                                 pr_debug("max_pstate=TAC %x\n", max_pstate);
 832                         } else {
 833                                 goto skip_tar;
 834                         }
 835                 }
 836         }
 837
 838 skip_tar:
 839         return max_pstate;
 840 }
 841
 842 static int core_get_turbo_pstate(void)
 843 {
 844         u64 value;
 845         int nont, ret;
 846
 847         rdmsrl(MSR_NHM_TURBO_RATIO_LIMIT, value);
 848         nont = core_get_max_pstate();
 849         ret = (value) & 255;
 850         if (ret <= nont)
 851                 ret = nont;
 852         return ret;
 853 }
 854
 855 static inline int core_get_scaling(void)
 856 {
 857         return 100000;
 858 }
 859
 860 static u64 core_get_val(struct cpudata *cpudata, int pstate)
 861 {
 862         u64 val;
 863
 864         val = (u64)pstate << 8;
 865         if (limits->no_turbo && !limits->turbo_disabled)
 866                 val |= (u64)1 << 32;
 867
 868         return val;
 869 }
 870
 871 static int knl_get_turbo_pstate(void)
 872 {
 873         u64 value;
 874         int nont, ret;
 875
 876         rdmsrl(MSR_NHM_TURBO_RATIO_LIMIT, value);
 877         nont = core_get_max_pstate();
 878         ret = (((value) >> 8) & 0xFF);
 879         if (ret <= nont)
 880                 ret = nont;
 881         return ret;
 882 }
 883
 884 static struct cpu_defaults core_params = {
 885         .pid_policy = {
 886                 .sample_rate_ms = 10,
 887                 .deadband = 0,
 888                 .setpoint = 97,
 889                 .p_gain_pct = 20,
 890                 .d_gain_pct = 0,
 891                 .i_gain_pct = 0,
 892         },
 893         .funcs = {
 894                 .get_max = core_get_max_pstate,
 895                 .get_max_physical = core_get_max_pstate_physical,
 896                 .get_min = core_get_min_pstate,
 897                 .get_turbo = core_get_turbo_pstate,
 898                 .get_scaling = core_get_scaling,
 899                 .get_val = core_get_val,
 900                 .get_target_pstate = get_target_pstate_use_performance,
 901         },
 902 };
 903
 904 static struct cpu_defaults silvermont_params = {
 905         .pid_policy = {
 906                 .sample_rate_ms = 10,
 907                 .deadband = 0,
 908                 .setpoint = 60,
 909                 .p_gain_pct = 14,
 910                 .d_gain_pct = 0,
 911                 .i_gain_pct = 4,
 912         },
 913         .funcs = {
 914                 .get_max = atom_get_max_pstate,
 915                 .get_max_physical = atom_get_max_pstate,
 916                 .get_min = atom_get_min_pstate,
 917                 .get_turbo = atom_get_turbo_pstate,
 918                 .get_val = atom_get_val,
 919                 .get_scaling = silvermont_get_scaling,
 920                 .get_vid = atom_get_vid,
 921                 .get_target_pstate = get_target_pstate_use_cpu_load,
 922         },
 923 };
 924
 925 static struct cpu_defaults airmont_params = {
 926         .pid_policy = {
 927                 .sample_rate_ms = 10,
 928                 .deadband = 0,
 929                 .setpoint = 60,
 930                 .p_gain_pct = 14,
 931                 .d_gain_pct = 0,
 932                 .i_gain_pct = 4,
 933         },
 934         .funcs = {
 935                 .get_max = atom_get_max_pstate,
 936                 .get_max_physical = atom_get_max_pstate,
 937                 .get_min = atom_get_min_pstate,
 938                 .get_turbo = atom_get_turbo_pstate,
 939                 .get_val = atom_get_val,
 940                 .get_scaling = airmont_get_scaling,
 941                 .get_vid = atom_get_vid,
 942                 .get_target_pstate = get_target_pstate_use_cpu_load,
 943         },
 944 };
 945
 946 static struct cpu_defaults knl_params = {
 947         .pid_policy = {
 948                 .sample_rate_ms = 10,
 949                 .deadband = 0,
 950                 .setpoint = 97,
 951                 .p_gain_pct = 20,
 952                 .d_gain_pct = 0,
 953                 .i_gain_pct = 0,
 954         },
 955         .funcs = {
 956                 .get_max = core_get_max_pstate,
 957                 .get_max_physical = core_get_max_pstate_physical,
 958                 .get_min = core_get_min_pstate,
 959                 .get_turbo = knl_get_turbo_pstate,
 960                 .get_scaling = core_get_scaling,
 961                 .get_val = core_get_val,
 962                 .get_target_pstate = get_target_pstate_use_performance,
 963         },
 964 };
 965
 966 static void intel_pstate_get_min_max(struct cpudata *cpu, int *min, int *max)
 967 {
 968         int max_perf = cpu->pstate.turbo_pstate;
 969         int max_perf_adj;
 970         int min_perf;
 971
 972         if (limits->no_turbo || limits->turbo_disabled)
 973                 max_perf = cpu->pstate.max_pstate;
 974
 975         /*
 976          * performance can be limited by user through sysfs, by cpufreq
 977          * policy, or by cpu specific default values determined through
 978          * experimentation.
 979          */
 980         max_perf_adj = fp_toint(max_perf * limits->max_perf);
 981         *max = clamp_t(int, max_perf_adj,
 982                         cpu->pstate.min_pstate, cpu->pstate.turbo_pstate);
 983
 984         min_perf = fp_toint(max_perf * limits->min_perf);
 985         *min = clamp_t(int, min_perf, cpu->pstate.min_pstate, max_perf);
 986 }
 987
 988 static inline void intel_pstate_record_pstate(struct cpudata *cpu, int pstate)
 989 {
 990         trace_cpu_frequency(pstate * cpu->pstate.scaling, cpu->cpu);
 991         cpu->pstate.current_pstate = pstate;
 992 }
 993
 994 static void intel_pstate_set_min_pstate(struct cpudata *cpu)
 995 {
 996         int pstate = cpu->pstate.min_pstate;
 997
 998         intel_pstate_record_pstate(cpu, pstate);
 999         /*
1000          * Generally, there is no guarantee that this code will always run on
1001          * the CPU being updated, so force the register update to run on the
1002          * right CPU.
1003          */
1004         wrmsrl_on_cpu(cpu->cpu, MSR_IA32_PERF_CTL,
1005                       pstate_funcs.get_val(cpu, pstate));
1006 }
1007
1008 static void intel_pstate_get_cpu_pstates(struct cpudata *cpu)
1009 {
1010         cpu->pstate.min_pstate = pstate_funcs.get_min();
1011         cpu->pstate.max_pstate = pstate_funcs.get_max();
1012         cpu->pstate.max_pstate_physical = pstate_funcs.get_max_physical();
1013         cpu->pstate.turbo_pstate = pstate_funcs.get_turbo();
1014         cpu->pstate.scaling = pstate_funcs.get_scaling();
1015
1016         if (pstate_funcs.get_vid)
1017                 pstate_funcs.get_vid(cpu);
1018
1019         intel_pstate_set_min_pstate(cpu);
1020 }
1021
1022 static inline void intel_pstate_calc_busy(struct cpudata *cpu)
1023 {
1024         struct sample *sample = &cpu->sample;
1025         int64_t core_pct;
1026
1027         core_pct = int_tofp(sample->aperf) * int_tofp(100);
1028         core_pct = div64_u64(core_pct, int_tofp(sample->mperf));
1029
1030         sample->core_pct_busy = (int32_t)core_pct;
1031 }
1032
1033 static inline bool intel_pstate_sample(struct cpudata *cpu, u64 time)
1034 {
1035         u64 aperf, mperf;
1036         unsigned long flags;
1037         u64 tsc;
1038
1039         local_irq_save(flags);
1040         rdmsrl(MSR_IA32_APERF, aperf);
1041         rdmsrl(MSR_IA32_MPERF, mperf);
1042         tsc = rdtsc();
1043         if (cpu->prev_mperf == mperf || cpu->prev_tsc == tsc) {
1044                 local_irq_restore(flags);
1045                 return false;
1046         }
1047         local_irq_restore(flags);
1048
1049         cpu->last_sample_time = cpu->sample.time;
1050         cpu->sample.time = time;
1051         cpu->sample.aperf = aperf;
1052         cpu->sample.mperf = mperf;
1053         cpu->sample.tsc =  tsc;
1054         cpu->sample.aperf -= cpu->prev_aperf;
1055         cpu->sample.mperf -= cpu->prev_mperf;
1056         cpu->sample.tsc -= cpu->prev_tsc;
1057
1058         cpu->prev_aperf = aperf;
1059         cpu->prev_mperf = mperf;
1060         cpu->prev_tsc = tsc;
1061         /*
1062          * First time this function is invoked in a given cycle, all of the
1063          * previous sample data fields are equal to zero or stale and they must
1064          * be populated with meaningful numbers for things to work, so assume
1065          * that sample.time will always be reset before setting the utilization
1066          * update hook and make the caller skip the sample then.
1067          */
1068         return !!cpu->last_sample_time;
1069 }
1070
1071 static inline int32_t get_avg_frequency(struct cpudata *cpu)
1072 {
1073         return fp_toint(mul_fp(cpu->sample.core_pct_busy,
1074                                int_tofp(cpu->pstate.max_pstate_physical *
1075                                                 cpu->pstate.scaling / 100)));
1076 }
1077
1078 static inline int32_t get_target_pstate_use_cpu_load(struct cpudata *cpu)
1079 {
1080         struct sample *sample = &cpu->sample;
1081         u64 cummulative_iowait, delta_iowait_us;
1082         u64 delta_iowait_mperf;
1083         u64 mperf, now;
1084         int32_t cpu_load;
1085
1086         cummulative_iowait = get_cpu_iowait_time_us(cpu->cpu, &now);
1087
1088         /*
1089          * Convert iowait time into number of IO cycles spent at max_freq.
1090          * IO is considered as busy only for the cpu_load algorithm. For
1091          * performance this is not needed since we always try to reach the
1092          * maximum P-State, so we are already boosting the IOs.
1093          */
1094         delta_iowait_us = cummulative_iowait - cpu->prev_cummulative_iowait;
1095         delta_iowait_mperf = div64_u64(delta_iowait_us * cpu->pstate.scaling *
1096                 cpu->pstate.max_pstate, MSEC_PER_SEC);
1097
1098         mperf = cpu->sample.mperf + delta_iowait_mperf;
1099         cpu->prev_cummulative_iowait = cummulative_iowait;
1100
1101         /*
1102          * The load can be estimated as the ratio of the mperf counter
1103          * running at a constant frequency during active periods
1104          * (C0) and the time stamp counter running at the same frequency
1105          * also during C-states.
1106          */
1107         cpu_load = div64_u64(int_tofp(100) * mperf, sample->tsc);
1108         cpu->sample.busy_scaled = cpu_load;
1109
1110         return cpu->pstate.current_pstate - pid_calc(&cpu->pid, cpu_load);
1111 }
1112
1113 static inline int32_t get_target_pstate_use_performance(struct cpudata *cpu)
1114 {
1115         int32_t core_busy, max_pstate, current_pstate, sample_ratio;
1116         u64 duration_ns;
1117
1118         /*
1119          * core_busy is the ratio of actual performance to max
1120          * max_pstate is the max non turbo pstate available
1121          * current_pstate was the pstate that was requested during
1122          *      the last sample period.
1123          *
1124          * We normalize core_busy, which was our actual percent
1125          * performance to what we requested during the last sample
1126          * period. The result will be a percentage of busy at a
1127          * specified pstate.
1128          */
1129         core_busy = cpu->sample.core_pct_busy;
1130         max_pstate = int_tofp(cpu->pstate.max_pstate_physical);
1131         current_pstate = int_tofp(cpu->pstate.current_pstate);
1132         core_busy = mul_fp(core_busy, div_fp(max_pstate, current_pstate));
1133
1134         /*
1135          * Since our utilization update callback will not run unless we are
1136          * in C0, check if the actual elapsed time is significantly greater (3x)
1137          * than our sample interval.  If it is, then we were idle for a long
1138          * enough period of time to adjust our busyness.
1139          */
1140         duration_ns = cpu->sample.time - cpu->last_sample_time;
1141         if ((s64)duration_ns > pid_params.sample_rate_ns * 3) {
1142                 sample_ratio = div_fp(int_tofp(pid_params.sample_rate_ns),
1143                                       int_tofp(duration_ns));
1144                 core_busy = mul_fp(core_busy, sample_ratio);
1145         } else {
1146                 sample_ratio = div_fp(100 * cpu->sample.mperf, cpu->sample.tsc);
1147                 if (sample_ratio < int_tofp(1))
1148                         core_busy = 0;
1149         }
1150
1151         cpu->sample.busy_scaled = core_busy;
1152         return cpu->pstate.current_pstate - pid_calc(&cpu->pid, core_busy);
1153 }
1154
1155 static inline void intel_pstate_update_pstate(struct cpudata *cpu, int pstate)
1156 {
1157         int max_perf, min_perf;
1158
1159         update_turbo_state();
1160
1161         intel_pstate_get_min_max(cpu, &min_perf, &max_perf);
1162         pstate = clamp_t(int, pstate, min_perf, max_perf);
1163         if (pstate == cpu->pstate.current_pstate)
1164                 return;
1165
1166         intel_pstate_record_pstate(cpu, pstate);
1167         wrmsrl(MSR_IA32_PERF_CTL, pstate_funcs.get_val(cpu, pstate));
1168 }
1169
1170 static inline void intel_pstate_adjust_busy_pstate(struct cpudata *cpu)
1171 {
1172         int from, target_pstate;
1173         struct sample *sample;
1174
1175         from = cpu->pstate.current_pstate;
1176
1177         target_pstate = pstate_funcs.get_target_pstate(cpu);
1178
1179         intel_pstate_update_pstate(cpu, target_pstate);
1180
1181         sample = &cpu->sample;
1182         trace_pstate_sample(fp_toint(sample->core_pct_busy),
1183                 fp_toint(sample->busy_scaled),
1184                 from,
1185                 cpu->pstate.current_pstate,
1186                 sample->mperf,
1187                 sample->aperf,
1188                 sample->tsc,
1189                 get_avg_frequency(cpu));
1190 }
1191
1192 static void intel_pstate_update_util(struct update_util_data *data, u64 time,
1193                                      unsigned long util, unsigned long max)
1194 {
1195         struct cpudata *cpu = container_of(data, struct cpudata, update_util);
1196         u64 delta_ns = time - cpu->sample.time;
1197
1198         if ((s64)delta_ns >= pid_params.sample_rate_ns) {
1199                 bool sample_taken = intel_pstate_sample(cpu, time);
1200
1201                 if (sample_taken) {
1202                         intel_pstate_calc_busy(cpu);
1203                         if (!hwp_active)
1204                                 intel_pstate_adjust_busy_pstate(cpu);
1205                 }
1206         }
1207 }
1208
1209 #define ICPU(model, policy) \
1210         { X86_VENDOR_INTEL, 6, model, X86_FEATURE_APERFMPERF,\
1211                         (unsigned long)&policy }
1212
1213 static const struct x86_cpu_id intel_pstate_cpu_ids[] = {
1214         ICPU(0x2a, core_params),
1215         ICPU(0x2d, core_params),
1216         ICPU(0x37, silvermont_params),
1217         ICPU(0x3a, core_params),
1218         ICPU(0x3c, core_params),
1219         ICPU(0x3d, core_params),
1220         ICPU(0x3e, core_params),
1221         ICPU(0x3f, core_params),
1222         ICPU(0x45, core_params),
1223         ICPU(0x46, core_params),
1224         ICPU(0x47, core_params),
1225         ICPU(0x4c, airmont_params),
1226         ICPU(0x4e, core_params),
1227         ICPU(0x4f, core_params),
1228         ICPU(0x5e, core_params),
1229         ICPU(0x56, core_params),
1230         ICPU(0x57, knl_params),
1231         {}
1232 };
1233 MODULE_DEVICE_TABLE(x86cpu, intel_pstate_cpu_ids);
1234
1235 static const struct x86_cpu_id intel_pstate_cpu_oob_ids[] = {
1236         ICPU(0x56, core_params),
1237         {}
1238 };
1239
1240 static int intel_pstate_init_cpu(unsigned int cpunum)
1241 {
1242         struct cpudata *cpu;
1243
1244         if (!all_cpu_data[cpunum])
1245                 all_cpu_data[cpunum] = kzalloc(sizeof(struct cpudata),
1246                                                GFP_KERNEL);
1247         if (!all_cpu_data[cpunum])
1248                 return -ENOMEM;
1249
1250         cpu = all_cpu_data[cpunum];
1251
1252         cpu->cpu = cpunum;
1253
1254         if (hwp_active) {
1255                 intel_pstate_hwp_enable(cpu);
1256                 pid_params.sample_rate_ms = 50;
1257                 pid_params.sample_rate_ns = 50 * NSEC_PER_MSEC;
1258         }
1259
1260         intel_pstate_get_cpu_pstates(cpu);
1261
1262         intel_pstate_busy_pid_reset(cpu);
1263
1264         cpu->update_util.func = intel_pstate_update_util;
1265
1266         pr_debug("intel_pstate: controlling: cpu %d\n", cpunum);
1267
1268         return 0;
1269 }
1270
1271 static unsigned int intel_pstate_get(unsigned int cpu_num)
1272 {
1273         struct sample *sample;
1274         struct cpudata *cpu;
1275
1276         cpu = all_cpu_data[cpu_num];
1277         if (!cpu)
1278                 return 0;
1279         sample = &cpu->sample;
1280         return get_avg_frequency(cpu);
1281 }
1282
1283 static void intel_pstate_set_update_util_hook(unsigned int cpu_num)
1284 {
1285         struct cpudata *cpu = all_cpu_data[cpu_num];
1286
1287         /* Prevent intel_pstate_update_util() from using stale data. */
1288         cpu->sample.time = 0;
1289         cpufreq_set_update_util_data(cpu_num, &cpu->update_util);
1290 }
1291
1292 static void intel_pstate_clear_update_util_hook(unsigned int cpu)
1293 {
1294         cpufreq_set_update_util_data(cpu, NULL);
1295         synchronize_sched();
1296 }
1297
1298 static void intel_pstate_set_performance_limits(struct perf_limits *limits)
1299 {
1300         limits->no_turbo = 0;
1301         limits->turbo_disabled = 0;
1302         limits->max_perf_pct = 100;
1303         limits->max_perf = int_tofp(1);
1304         limits->min_perf_pct = 100;
1305         limits->min_perf = int_tofp(1);
1306         limits->max_policy_pct = 100;
1307         limits->max_sysfs_pct = 100;
1308         limits->min_policy_pct = 0;
1309         limits->min_sysfs_pct = 0;
1310 }
1311
1312 static int intel_pstate_set_policy(struct cpufreq_policy *policy)
1313 {
1314         if (!policy->cpuinfo.max_freq)
1315                 return -ENODEV;
1316
1317         intel_pstate_clear_update_util_hook(policy->cpu);
1318
1319         if (policy->policy == CPUFREQ_POLICY_PERFORMANCE) {
1320                 limits = &performance_limits;
1321                 if (policy->max >= policy->cpuinfo.max_freq) {
1322                         pr_debug("intel_pstate: set performance\n");
1323                         intel_pstate_set_performance_limits(limits);
1324                         goto out;
1325                 }
1326         } else {
1327                 pr_debug("intel_pstate: set powersave\n");
1328                 limits = &powersave_limits;
1329         }
1330
1331         limits->min_policy_pct = (policy->min * 100) / policy->cpuinfo.max_freq;
1332         limits->min_policy_pct = clamp_t(int, limits->min_policy_pct, 0 , 100);
1333         limits->max_policy_pct = DIV_ROUND_UP(policy->max * 100,
1334                                               policy->cpuinfo.max_freq);
1335         limits->max_policy_pct = clamp_t(int, limits->max_policy_pct, 0 , 100);
1336
1337         /* Normalize user input to [min_policy_pct, max_policy_pct] */
1338         limits->min_perf_pct = max(limits->min_policy_pct,
1339                                    limits->min_sysfs_pct);
1340         limits->min_perf_pct = min(limits->max_policy_pct,
1341                                    limits->min_perf_pct);
1342         limits->max_perf_pct = min(limits->max_policy_pct,
1343                                    limits->max_sysfs_pct);
1344         limits->max_perf_pct = max(limits->min_policy_pct,
1345                                    limits->max_perf_pct);
1346         limits->max_perf = round_up(limits->max_perf, FRAC_BITS);
1347
1348         /* Make sure min_perf_pct <= max_perf_pct */
1349         limits->min_perf_pct = min(limits->max_perf_pct, limits->min_perf_pct);
1350
1351         limits->min_perf = div_fp(int_tofp(limits->min_perf_pct),
1352                                   int_tofp(100));
1353         limits->max_perf = div_fp(int_tofp(limits->max_perf_pct),
1354                                   int_tofp(100));
1355
1356  out:
1357         intel_pstate_set_update_util_hook(policy->cpu);
1358
1359         intel_pstate_hwp_set_policy(policy);
1360
1361         return 0;
1362 }
1363
1364 static int intel_pstate_verify_policy(struct cpufreq_policy *policy)
1365 {
1366         cpufreq_verify_within_cpu_limits(policy);
1367
1368         if (policy->policy != CPUFREQ_POLICY_POWERSAVE &&
1369             policy->policy != CPUFREQ_POLICY_PERFORMANCE)
1370                 return -EINVAL;
1371
1372         return 0;
1373 }
1374
1375 static void intel_pstate_stop_cpu(struct cpufreq_policy *policy)
1376 {
1377         int cpu_num = policy->cpu;
1378         struct cpudata *cpu = all_cpu_data[cpu_num];
1379
1380         pr_debug("intel_pstate: CPU %d exiting\n", cpu_num);
1381
1382         intel_pstate_clear_update_util_hook(cpu_num);
1383
1384         if (hwp_active)
1385                 return;
1386
1387         intel_pstate_set_min_pstate(cpu);
1388 }
1389
1390 static int intel_pstate_cpu_init(struct cpufreq_policy *policy)
1391 {
1392         struct cpudata *cpu;
1393         int rc;
1394
1395         rc = intel_pstate_init_cpu(policy->cpu);
1396         if (rc)
1397                 return rc;
1398
1399         cpu = all_cpu_data[policy->cpu];
1400
1401         if (limits->min_perf_pct == 100 && limits->max_perf_pct == 100)
1402                 policy->policy = CPUFREQ_POLICY_PERFORMANCE;
1403         else
1404                 policy->policy = CPUFREQ_POLICY_POWERSAVE;
1405
1406         policy->min = cpu->pstate.min_pstate * cpu->pstate.scaling;
1407         policy->max = cpu->pstate.turbo_pstate * cpu->pstate.scaling;
1408
1409         /* cpuinfo and default policy values */
1410         policy->cpuinfo.min_freq = cpu->pstate.min_pstate * cpu->pstate.scaling;
1411         policy->cpuinfo.max_freq =
1412                 cpu->pstate.turbo_pstate * cpu->pstate.scaling;
1413         policy->cpuinfo.transition_latency = CPUFREQ_ETERNAL;
1414         cpumask_set_cpu(policy->cpu, policy->cpus);
1415
1416         return 0;
1417 }
1418
1419 static struct cpufreq_driver intel_pstate_driver = {
1420         .flags          = CPUFREQ_CONST_LOOPS,
1421         .verify         = intel_pstate_verify_policy,
1422         .setpolicy      = intel_pstate_set_policy,
1423         .resume         = intel_pstate_hwp_set_policy,
1424         .get            = intel_pstate_get,
1425         .init           = intel_pstate_cpu_init,
1426         .stop_cpu       = intel_pstate_stop_cpu,
1427         .name           = "intel_pstate",
1428 };
1429
1430 static int __initdata no_load;
1431 static int __initdata no_hwp;
1432 static int __initdata hwp_only;
1433 static unsigned int force_load;
1434
1435 static int intel_pstate_msrs_not_valid(void)
1436 {
1437         if (!pstate_funcs.get_max() ||
1438             !pstate_funcs.get_min() ||
1439             !pstate_funcs.get_turbo())
1440                 return -ENODEV;
1441
1442         return 0;
1443 }
1444
1445 static void copy_pid_params(struct pstate_adjust_policy *policy)
1446 {
1447         pid_params.sample_rate_ms = policy->sample_rate_ms;
1448         pid_params.sample_rate_ns = pid_params.sample_rate_ms * NSEC_PER_MSEC;
1449         pid_params.p_gain_pct = policy->p_gain_pct;
1450         pid_params.i_gain_pct = policy->i_gain_pct;
1451         pid_params.d_gain_pct = policy->d_gain_pct;
1452         pid_params.deadband = policy->deadband;
1453         pid_params.setpoint = policy->setpoint;
1454 }
1455
1456 static void copy_cpu_funcs(struct pstate_funcs *funcs)
1457 {
1458         pstate_funcs.get_max   = funcs->get_max;
1459         pstate_funcs.get_max_physical = funcs->get_max_physical;
1460         pstate_funcs.get_min   = funcs->get_min;
1461         pstate_funcs.get_turbo = funcs->get_turbo;
1462         pstate_funcs.get_scaling = funcs->get_scaling;
1463         pstate_funcs.get_val   = funcs->get_val;
1464         pstate_funcs.get_vid   = funcs->get_vid;
1465         pstate_funcs.get_target_pstate = funcs->get_target_pstate;
1466
1467 }
1468
1469 #if IS_ENABLED(CONFIG_ACPI)
1470 #include <acpi/processor.h>
1471
1472 static bool intel_pstate_no_acpi_pss(void)
1473 {
1474         int i;
1475
1476         for_each_possible_cpu(i) {
1477                 acpi_status status;
1478                 union acpi_object *pss;
1479                 struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
1480                 struct acpi_processor *pr = per_cpu(processors, i);
1481
1482                 if (!pr)
1483                         continue;
1484
1485                 status = acpi_evaluate_object(pr->handle, "_PSS", NULL, &buffer);
1486                 if (ACPI_FAILURE(status))
1487                         continue;
1488
1489                 pss = buffer.pointer;
1490                 if (pss && pss->type == ACPI_TYPE_PACKAGE) {
1491                         kfree(pss);
1492                         return false;
1493                 }
1494
1495                 kfree(pss);
1496         }
1497
1498         return true;
1499 }
1500
1501 static bool intel_pstate_has_acpi_ppc(void)
1502 {
1503         int i;
1504
1505         for_each_possible_cpu(i) {
1506                 struct acpi_processor *pr = per_cpu(processors, i);
1507
1508                 if (!pr)
1509                         continue;
1510                 if (acpi_has_method(pr->handle, "_PPC"))
1511                         return true;
1512         }
1513         return false;
1514 }
1515
1516 enum {
1517         PSS,
1518         PPC,
1519 };
1520
1521 struct hw_vendor_info {
1522         u16  valid;
1523         char oem_id[ACPI_OEM_ID_SIZE];
1524         char oem_table_id[ACPI_OEM_TABLE_ID_SIZE];
1525         int  oem_pwr_table;
1526 };
1527
1528 /* Hardware vendor-specific info that has its own power management modes */
1529 static struct hw_vendor_info vendor_info[] = {
1530         {1, "HP    ", "ProLiant", PSS},
1531         {1, "ORACLE", "X4-2    ", PPC},
1532         {1, "ORACLE", "X4-2L   ", PPC},
1533         {1, "ORACLE", "X4-2B   ", PPC},
1534         {1, "ORACLE", "X3-2    ", PPC},
1535         {1, "ORACLE", "X3-2L   ", PPC},
1536         {1, "ORACLE", "X3-2B   ", PPC},
1537         {1, "ORACLE", "X4470M2 ", PPC},
1538         {1, "ORACLE", "X4270M3 ", PPC},
1539         {1, "ORACLE", "X4270M2 ", PPC},
1540         {1, "ORACLE", "X4170M2 ", PPC},
1541         {1, "ORACLE", "X4170 M3", PPC},
1542         {1, "ORACLE", "X4275 M3", PPC},
1543         {1, "ORACLE", "X6-2    ", PPC},
1544         {1, "ORACLE", "Sudbury ", PPC},
1545         {0, "", ""},
1546 };
1547
1548 static bool intel_pstate_platform_pwr_mgmt_exists(void)
1549 {
1550         struct acpi_table_header hdr;
1551         struct hw_vendor_info *v_info;
1552         const struct x86_cpu_id *id;
1553         u64 misc_pwr;
1554
1555         id = x86_match_cpu(intel_pstate_cpu_oob_ids);
1556         if (id) {
1557                 rdmsrl(MSR_MISC_PWR_MGMT, misc_pwr);
1558                 if ( misc_pwr & (1 << 8))
1559                         return true;
1560         }
1561
1562         if (acpi_disabled ||
1563             ACPI_FAILURE(acpi_get_table_header(ACPI_SIG_FADT, 0, &hdr)))
1564                 return false;
1565
1566         for (v_info = vendor_info; v_info->valid; v_info++) {
1567                 if (!strncmp(hdr.oem_id, v_info->oem_id, ACPI_OEM_ID_SIZE) &&
1568                         !strncmp(hdr.oem_table_id, v_info->oem_table_id,
1569                                                 ACPI_OEM_TABLE_ID_SIZE))
1570                         switch (v_info->oem_pwr_table) {
1571                         case PSS:
1572                                 return intel_pstate_no_acpi_pss();
1573                         case PPC:
1574                                 return intel_pstate_has_acpi_ppc() &&
1575                                         (!force_load);
1576                         }
1577         }
1578
1579         return false;
1580 }
1581 #else /* CONFIG_ACPI not enabled */
1582 static inline bool intel_pstate_platform_pwr_mgmt_exists(void) { return false; }
1583 static inline bool intel_pstate_has_acpi_ppc(void) { return false; }
1584 #endif /* CONFIG_ACPI */
1585
1586 static const struct x86_cpu_id hwp_support_ids[] __initconst = {
1587         { X86_VENDOR_INTEL, 6, X86_MODEL_ANY, X86_FEATURE_HWP },
1588         {}
1589 };
1590
1591 static int __init intel_pstate_init(void)
1592 {
1593         int cpu, rc = 0;
1594         const struct x86_cpu_id *id;
1595         struct cpu_defaults *cpu_def;
1596
1597         if (no_load)
1598                 return -ENODEV;
1599
1600         if (x86_match_cpu(hwp_support_ids) && !no_hwp) {
1601                 copy_cpu_funcs(&core_params.funcs);
1602                 hwp_active++;
1603                 goto hwp_cpu_matched;
1604         }
1605
1606         id = x86_match_cpu(intel_pstate_cpu_ids);
1607         if (!id)
1608                 return -ENODEV;
1609
1610         cpu_def = (struct cpu_defaults *)id->driver_data;
1611
1612         copy_pid_params(&cpu_def->pid_policy);
1613         copy_cpu_funcs(&cpu_def->funcs);
1614
1615         if (intel_pstate_msrs_not_valid())
1616                 return -ENODEV;
1617
1618 hwp_cpu_matched:
1619         /*
1620          * The Intel pstate driver will be ignored if the platform
1621          * firmware has its own power management modes.
1622          */
1623         if (intel_pstate_platform_pwr_mgmt_exists())
1624                 return -ENODEV;
1625
1626         pr_info("Intel P-state driver initializing.\n");
1627
1628         all_cpu_data = vzalloc(sizeof(void *) * num_possible_cpus());
1629         if (!all_cpu_data)
1630                 return -ENOMEM;
1631
1632         if (!hwp_active && hwp_only)
1633                 goto out;
1634
1635         rc = cpufreq_register_driver(&intel_pstate_driver);
1636         if (rc)
1637                 goto out;
1638
1639         intel_pstate_debug_expose_params();
1640         intel_pstate_sysfs_expose_params();
1641
1642         if (hwp_active)
1643                 pr_info("intel_pstate: HWP enabled\n");
1644
1645         return rc;
1646 out:
1647         get_online_cpus();
1648         for_each_online_cpu(cpu) {
1649                 if (all_cpu_data[cpu]) {
1650                         intel_pstate_clear_update_util_hook(cpu);
1651                         kfree(all_cpu_data[cpu]);
1652                 }
1653         }
1654
1655         put_online_cpus();
1656         vfree(all_cpu_data);
1657         return -ENODEV;
1658 }
1659 device_initcall(intel_pstate_init);
1660
1661 static int __init intel_pstate_setup(char *str)
1662 {
1663         if (!str)
1664                 return -EINVAL;
1665
1666         if (!strcmp(str, "disable"))
1667                 no_load = 1;
1668         if (!strcmp(str, "no_hwp")) {
1669                 pr_info("intel_pstate: HWP disabled\n");
1670                 no_hwp = 1;
1671         }
1672         if (!strcmp(str, "force"))
1673                 force_load = 1;
1674         if (!strcmp(str, "hwp_only"))
1675                 hwp_only = 1;
1676         return 0;
1677 }
1678 early_param("intel_pstate", intel_pstate_setup);
1679
1680 MODULE_AUTHOR("Dirk Brandewie <dirk.j.brandewie@intel.com>");
1681 MODULE_DESCRIPTION("'intel_pstate' - P state driver Intel Core processors");
1682 MODULE_LICENSE("GPL");