kernel/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 <trace/events/power.h>
  30
  31 #include <asm/div64.h>
  32 #include <asm/msr.h>
  33 #include <asm/cpu_device_id.h>
  34 #include <asm/cpufeature.h>
  35
  36 #define BYT_RATIOS              0x66a
  37 #define BYT_VIDS                0x66b
  38 #define BYT_TURBO_RATIOS        0x66c
  39 #define BYT_TURBO_VIDS          0x66d
  40
  41 #define FRAC_BITS 8
  42 #define int_tofp(X) ((int64_t)(X) << FRAC_BITS)
  43 #define fp_toint(X) ((X) >> FRAC_BITS)
  44
  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(int32_t x, int32_t y)
  52 {
  53         return div_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 struct sample {
  68         int32_t core_pct_busy;
  69         u64 aperf;
  70         u64 mperf;
  71         int freq;
  72         ktime_t time;
  73 };
  74
  75 struct pstate_data {
  76         int     current_pstate;
  77         int     min_pstate;
  78         int     max_pstate;
  79         int     scaling;
  80         int     turbo_pstate;
  81 };
  82
  83 struct vid_data {
  84         int min;
  85         int max;
  86         int turbo;
  87         int32_t ratio;
  88 };
  89
  90 struct _pid {
  91         int setpoint;
  92         int32_t integral;
  93         int32_t p_gain;
  94         int32_t i_gain;
  95         int32_t d_gain;
  96         int deadband;
  97         int32_t last_err;
  98 };
  99
 100 struct cpudata {
 101         int cpu;
 102
 103         struct timer_list timer;
 104
 105         struct pstate_data pstate;
 106         struct vid_data vid;
 107         struct _pid pid;
 108
 109         ktime_t last_sample_time;
 110         u64     prev_aperf;
 111         u64     prev_mperf;
 112         struct sample sample;
 113 };
 114
 115 static struct cpudata **all_cpu_data;
 116 struct pstate_adjust_policy {
 117         int sample_rate_ms;
 118         int deadband;
 119         int setpoint;
 120         int p_gain_pct;
 121         int d_gain_pct;
 122         int i_gain_pct;
 123 };
 124
 125 struct pstate_funcs {
 126         int (*get_max)(void);
 127         int (*get_min)(void);
 128         int (*get_turbo)(void);
 129         int (*get_scaling)(void);
 130         void (*set)(struct cpudata*, int pstate);
 131         void (*get_vid)(struct cpudata *);
 132 };
 133
 134 struct cpu_defaults {
 135         struct pstate_adjust_policy pid_policy;
 136         struct pstate_funcs funcs;
 137 };
 138
 139 static struct pstate_adjust_policy pid_params;
 140 static struct pstate_funcs pstate_funcs;
 141 static int hwp_active;
 142
 143 struct perf_limits {
 144         int no_turbo;
 145         int turbo_disabled;
 146         int max_perf_pct;
 147         int min_perf_pct;
 148         int32_t max_perf;
 149         int32_t min_perf;
 150         int max_policy_pct;
 151         int max_sysfs_pct;
 152         int min_policy_pct;
 153         int min_sysfs_pct;
 154 };
 155
 156 static struct perf_limits limits = {
 157         .no_turbo = 0,
 158         .turbo_disabled = 0,
 159         .max_perf_pct = 100,
 160         .max_perf = int_tofp(1),
 161         .min_perf_pct = 0,
 162         .min_perf = 0,
 163         .max_policy_pct = 100,
 164         .max_sysfs_pct = 100,
 165         .min_policy_pct = 0,
 166         .min_sysfs_pct = 0,
 167 };
 168
 169 static inline void pid_reset(struct _pid *pid, int setpoint, int busy,
 170                              int deadband, int integral) {
 171         pid->setpoint = setpoint;
 172         pid->deadband  = deadband;
 173         pid->integral  = int_tofp(integral);
 174         pid->last_err  = int_tofp(setpoint) - int_tofp(busy);
 175 }
 176
 177 static inline void pid_p_gain_set(struct _pid *pid, int percent)
 178 {
 179         pid->p_gain = div_fp(int_tofp(percent), int_tofp(100));
 180 }
 181
 182 static inline void pid_i_gain_set(struct _pid *pid, int percent)
 183 {
 184         pid->i_gain = div_fp(int_tofp(percent), int_tofp(100));
 185 }
 186
 187 static inline void pid_d_gain_set(struct _pid *pid, int percent)
 188 {
 189         pid->d_gain = div_fp(int_tofp(percent), int_tofp(100));
 190 }
 191
 192 static signed int pid_calc(struct _pid *pid, int32_t busy)
 193 {
 194         signed int result;
 195         int32_t pterm, dterm, fp_error;
 196         int32_t integral_limit;
 197
 198         fp_error = int_tofp(pid->setpoint) - busy;
 199
 200         if (abs(fp_error) <= int_tofp(pid->deadband))
 201                 return 0;
 202
 203         pterm = mul_fp(pid->p_gain, fp_error);
 204
 205         pid->integral += fp_error;
 206
 207         /*
 208          * We limit the integral here so that it will never
 209          * get higher than 30.  This prevents it from becoming
 210          * too large an input over long periods of time and allows
 211          * it to get factored out sooner.
 212          *
 213          * The value of 30 was chosen through experimentation.
 214          */
 215         integral_limit = int_tofp(30);
 216         if (pid->integral > integral_limit)
 217                 pid->integral = integral_limit;
 218         if (pid->integral < -integral_limit)
 219                 pid->integral = -integral_limit;
 220
 221         dterm = mul_fp(pid->d_gain, fp_error - pid->last_err);
 222         pid->last_err = fp_error;
 223
 224         result = pterm + mul_fp(pid->integral, pid->i_gain) + dterm;
 225         result = result + (1 << (FRAC_BITS-1));
 226         return (signed int)fp_toint(result);
 227 }
 228
 229 static inline void intel_pstate_busy_pid_reset(struct cpudata *cpu)
 230 {
 231         pid_p_gain_set(&cpu->pid, pid_params.p_gain_pct);
 232         pid_d_gain_set(&cpu->pid, pid_params.d_gain_pct);
 233         pid_i_gain_set(&cpu->pid, pid_params.i_gain_pct);
 234
 235         pid_reset(&cpu->pid, pid_params.setpoint, 100, pid_params.deadband, 0);
 236 }
 237
 238 static inline void intel_pstate_reset_all_pid(void)
 239 {
 240         unsigned int cpu;
 241
 242         for_each_online_cpu(cpu) {
 243                 if (all_cpu_data[cpu])
 244                         intel_pstate_busy_pid_reset(all_cpu_data[cpu]);
 245         }
 246 }
 247
 248 static inline void update_turbo_state(void)
 249 {
 250         u64 misc_en;
 251         struct cpudata *cpu;
 252
 253         cpu = all_cpu_data[0];
 254         rdmsrl(MSR_IA32_MISC_ENABLE, misc_en);
 255         limits.turbo_disabled =
 256                 (misc_en & MSR_IA32_MISC_ENABLE_TURBO_DISABLE ||
 257                  cpu->pstate.max_pstate == cpu->pstate.turbo_pstate);
 258 }
 259
 260 #define PCT_TO_HWP(x) (x * 255 / 100)
 261 static void intel_pstate_hwp_set(void)
 262 {
 263         int min, max, cpu;
 264         u64 value, freq;
 265
 266         get_online_cpus();
 267
 268         for_each_online_cpu(cpu) {
 269                 rdmsrl_on_cpu(cpu, MSR_HWP_REQUEST, &value);
 270                 min = PCT_TO_HWP(limits.min_perf_pct);
 271                 value &= ~HWP_MIN_PERF(~0L);
 272                 value |= HWP_MIN_PERF(min);
 273
 274                 max = PCT_TO_HWP(limits.max_perf_pct);
 275                 if (limits.no_turbo) {
 276                         rdmsrl( MSR_HWP_CAPABILITIES, freq);
 277                         max = HWP_GUARANTEED_PERF(freq);
 278                 }
 279
 280                 value &= ~HWP_MAX_PERF(~0L);
 281                 value |= HWP_MAX_PERF(max);
 282                 wrmsrl_on_cpu(cpu, MSR_HWP_REQUEST, value);
 283         }
 284
 285         put_online_cpus();
 286 }
 287
 288 /************************** debugfs begin ************************/
 289 static int pid_param_set(void *data, u64 val)
 290 {
 291         *(u32 *)data = val;
 292         intel_pstate_reset_all_pid();
 293         return 0;
 294 }
 295
 296 static int pid_param_get(void *data, u64 *val)
 297 {
 298         *val = *(u32 *)data;
 299         return 0;
 300 }
 301 DEFINE_SIMPLE_ATTRIBUTE(fops_pid_param, pid_param_get, pid_param_set, "%llu\n");
 302
 303 struct pid_param {
 304         char *name;
 305         void *value;
 306 };
 307
 308 static struct pid_param pid_files[] = {
 309         {"sample_rate_ms", &pid_params.sample_rate_ms},
 310         {"d_gain_pct", &pid_params.d_gain_pct},
 311         {"i_gain_pct", &pid_params.i_gain_pct},
 312         {"deadband", &pid_params.deadband},
 313         {"setpoint", &pid_params.setpoint},
 314         {"p_gain_pct", &pid_params.p_gain_pct},
 315         {NULL, NULL}
 316 };
 317
 318 static void __init intel_pstate_debug_expose_params(void)
 319 {
 320         struct dentry *debugfs_parent;
 321         int i = 0;
 322
 323         if (hwp_active)
 324                 return;
 325         debugfs_parent = debugfs_create_dir("pstate_snb", NULL);
 326         if (IS_ERR_OR_NULL(debugfs_parent))
 327                 return;
 328         while (pid_files[i].name) {
 329                 debugfs_create_file(pid_files[i].name, 0660,
 330                                     debugfs_parent, pid_files[i].value,
 331                                     &fops_pid_param);
 332                 i++;
 333         }
 334 }
 335
 336 /************************** debugfs end ************************/
 337
 338 /************************** sysfs begin ************************/
 339 #define show_one(file_name, object)                                     \
 340         static ssize_t show_##file_name                                 \
 341         (struct kobject *kobj, struct attribute *attr, char *buf)       \
 342         {                                                               \
 343                 return sprintf(buf, "%u\n", limits.object);             \
 344         }
 345
 346 static ssize_t show_turbo_pct(struct kobject *kobj,
 347                                 struct attribute *attr, char *buf)
 348 {
 349         struct cpudata *cpu;
 350         int total, no_turbo, turbo_pct;
 351         uint32_t turbo_fp;
 352
 353         cpu = all_cpu_data[0];
 354
 355         total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
 356         no_turbo = cpu->pstate.max_pstate - cpu->pstate.min_pstate + 1;
 357         turbo_fp = div_fp(int_tofp(no_turbo), int_tofp(total));
 358         turbo_pct = 100 - fp_toint(mul_fp(turbo_fp, int_tofp(100)));
 359         return sprintf(buf, "%u\n", turbo_pct);
 360 }
 361
 362 static ssize_t show_num_pstates(struct kobject *kobj,
 363                                 struct attribute *attr, char *buf)
 364 {
 365         struct cpudata *cpu;
 366         int total;
 367
 368         cpu = all_cpu_data[0];
 369         total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1;
 370         return sprintf(buf, "%u\n", total);
 371 }
 372
 373 static ssize_t show_no_turbo(struct kobject *kobj,
 374                              struct attribute *attr, char *buf)
 375 {
 376         ssize_t ret;
 377
 378         update_turbo_state();
 379         if (limits.turbo_disabled)
 380                 ret = sprintf(buf, "%u\n", limits.turbo_disabled);
 381         else
 382                 ret = sprintf(buf, "%u\n", limits.no_turbo);
 383
 384         return ret;
 385 }
 386
 387 static ssize_t store_no_turbo(struct kobject *a, struct attribute *b,
 388                               const char *buf, size_t count)
 389 {
 390         unsigned int input;
 391         int ret;
 392
 393         ret = sscanf(buf, "%u", &input);
 394         if (ret != 1)
 395                 return -EINVAL;
 396
 397         update_turbo_state();
 398         if (limits.turbo_disabled) {
 399                 pr_warn("Turbo disabled by BIOS or unavailable on processor\n");
 400                 return -EPERM;
 401         }
 402
 403         limits.no_turbo = clamp_t(int, input, 0, 1);
 404
 405         if (hwp_active)
 406                 intel_pstate_hwp_set();
 407
 408         return count;
 409 }
 410
 411 static ssize_t store_max_perf_pct(struct kobject *a, struct attribute *b,
 412                                   const char *buf, size_t count)
 413 {
 414         unsigned int input;
 415         int ret;
 416
 417         ret = sscanf(buf, "%u", &input);
 418         if (ret != 1)
 419                 return -EINVAL;
 420
 421         limits.max_sysfs_pct = clamp_t(int, input, 0 , 100);
 422         limits.max_perf_pct = min(limits.max_policy_pct, limits.max_sysfs_pct);
 423         limits.max_perf = div_fp(int_tofp(limits.max_perf_pct), int_tofp(100));
 424
 425         if (hwp_active)
 426                 intel_pstate_hwp_set();
 427         return count;
 428 }
 429
 430 static ssize_t store_min_perf_pct(struct kobject *a, struct attribute *b,
 431                                   const char *buf, size_t count)
 432 {
 433         unsigned int input;
 434         int ret;
 435
 436         ret = sscanf(buf, "%u", &input);
 437         if (ret != 1)
 438                 return -EINVAL;
 439
 440         limits.min_sysfs_pct = clamp_t(int, input, 0 , 100);
 441         limits.min_perf_pct = max(limits.min_policy_pct, limits.min_sysfs_pct);
 442         limits.min_perf = div_fp(int_tofp(limits.min_perf_pct), int_tofp(100));
 443
 444         if (hwp_active)
 445                 intel_pstate_hwp_set();
 446         return count;
 447 }
 448
 449 show_one(max_perf_pct, max_perf_pct);
 450 show_one(min_perf_pct, min_perf_pct);
 451
 452 define_one_global_rw(no_turbo);
 453 define_one_global_rw(max_perf_pct);
 454 define_one_global_rw(min_perf_pct);
 455 define_one_global_ro(turbo_pct);
 456 define_one_global_ro(num_pstates);
 457
 458 static struct attribute *intel_pstate_attributes[] = {
 459         &no_turbo.attr,
 460         &max_perf_pct.attr,
 461         &min_perf_pct.attr,
 462         &turbo_pct.attr,
 463         &num_pstates.attr,
 464         NULL
 465 };
 466
 467 static struct attribute_group intel_pstate_attr_group = {
 468         .attrs = intel_pstate_attributes,
 469 };
 470
 471 static void __init intel_pstate_sysfs_expose_params(void)
 472 {
 473         struct kobject *intel_pstate_kobject;
 474         int rc;
 475
 476         intel_pstate_kobject = kobject_create_and_add("intel_pstate",
 477                                                 &cpu_subsys.dev_root->kobj);
 478         BUG_ON(!intel_pstate_kobject);
 479         rc = sysfs_create_group(intel_pstate_kobject, &intel_pstate_attr_group);
 480         BUG_ON(rc);
 481 }
 482 /************************** sysfs end ************************/
 483
 484 static void intel_pstate_hwp_enable(void)
 485 {
 486         hwp_active++;
 487         pr_info("intel_pstate HWP enabled\n");
 488
 489         wrmsrl( MSR_PM_ENABLE, 0x1);
 490 }
 491
 492 static int byt_get_min_pstate(void)
 493 {
 494         u64 value;
 495
 496         rdmsrl(BYT_RATIOS, value);
 497         return (value >> 8) & 0x7F;
 498 }
 499
 500 static int byt_get_max_pstate(void)
 501 {
 502         u64 value;
 503
 504         rdmsrl(BYT_RATIOS, value);
 505         return (value >> 16) & 0x7F;
 506 }
 507
 508 static int byt_get_turbo_pstate(void)
 509 {
 510         u64 value;
 511
 512         rdmsrl(BYT_TURBO_RATIOS, value);
 513         return value & 0x7F;
 514 }
 515
 516 static void byt_set_pstate(struct cpudata *cpudata, int pstate)
 517 {
 518         u64 val;
 519         int32_t vid_fp;
 520         u32 vid;
 521
 522         val = pstate << 8;
 523         if (limits.no_turbo && !limits.turbo_disabled)
 524                 val |= (u64)1 << 32;
 525
 526         vid_fp = cpudata->vid.min + mul_fp(
 527                 int_tofp(pstate - cpudata->pstate.min_pstate),
 528                 cpudata->vid.ratio);
 529
 530         vid_fp = clamp_t(int32_t, vid_fp, cpudata->vid.min, cpudata->vid.max);
 531         vid = ceiling_fp(vid_fp);
 532
 533         if (pstate > cpudata->pstate.max_pstate)
 534                 vid = cpudata->vid.turbo;
 535
 536         val |= vid;
 537
 538         wrmsrl_on_cpu(cpudata->cpu, MSR_IA32_PERF_CTL, val);
 539 }
 540
 541 #define BYT_BCLK_FREQS 5
 542 static int byt_freq_table[BYT_BCLK_FREQS] = { 833, 1000, 1333, 1167, 800};
 543
 544 static int byt_get_scaling(void)
 545 {
 546         u64 value;
 547         int i;
 548
 549         rdmsrl(MSR_FSB_FREQ, value);
 550         i = value & 0x3;
 551
 552         BUG_ON(i > BYT_BCLK_FREQS);
 553
 554         return byt_freq_table[i] * 100;
 555 }
 556
 557 static void byt_get_vid(struct cpudata *cpudata)
 558 {
 559         u64 value;
 560
 561         rdmsrl(BYT_VIDS, value);
 562         cpudata->vid.min = int_tofp((value >> 8) & 0x7f);
 563         cpudata->vid.max = int_tofp((value >> 16) & 0x7f);
 564         cpudata->vid.ratio = div_fp(
 565                 cpudata->vid.max - cpudata->vid.min,
 566                 int_tofp(cpudata->pstate.max_pstate -
 567                         cpudata->pstate.min_pstate));
 568
 569         rdmsrl(BYT_TURBO_VIDS, value);
 570         cpudata->vid.turbo = value & 0x7f;
 571 }
 572
 573 static int core_get_min_pstate(void)
 574 {
 575         u64 value;
 576
 577         rdmsrl(MSR_PLATFORM_INFO, value);
 578         return (value >> 40) & 0xFF;
 579 }
 580
 581 static int core_get_max_pstate(void)
 582 {
 583         u64 value;
 584
 585         rdmsrl(MSR_PLATFORM_INFO, value);
 586         return (value >> 8) & 0xFF;
 587 }
 588
 589 static int core_get_turbo_pstate(void)
 590 {
 591         u64 value;
 592         int nont, ret;
 593
 594         rdmsrl(MSR_NHM_TURBO_RATIO_LIMIT, value);
 595         nont = core_get_max_pstate();
 596         ret = (value) & 255;
 597         if (ret <= nont)
 598                 ret = nont;
 599         return ret;
 600 }
 601
 602 static inline int core_get_scaling(void)
 603 {
 604         return 100000;
 605 }
 606
 607 static void core_set_pstate(struct cpudata *cpudata, int pstate)
 608 {
 609         u64 val;
 610
 611         val = pstate << 8;
 612         if (limits.no_turbo && !limits.turbo_disabled)
 613                 val |= (u64)1 << 32;
 614
 615         wrmsrl_on_cpu(cpudata->cpu, MSR_IA32_PERF_CTL, val);
 616 }
 617
 618 static int knl_get_turbo_pstate(void)
 619 {
 620         u64 value;
 621         int nont, ret;
 622
 623         rdmsrl(MSR_NHM_TURBO_RATIO_LIMIT, value);
 624         nont = core_get_max_pstate();
 625         ret = (((value) >> 8) & 0xFF);
 626         if (ret <= nont)
 627                 ret = nont;
 628         return ret;
 629 }
 630
 631 static struct cpu_defaults core_params = {
 632         .pid_policy = {
 633                 .sample_rate_ms = 10,
 634                 .deadband = 0,
 635                 .setpoint = 97,
 636                 .p_gain_pct = 20,
 637                 .d_gain_pct = 0,
 638                 .i_gain_pct = 0,
 639         },
 640         .funcs = {
 641                 .get_max = core_get_max_pstate,
 642                 .get_min = core_get_min_pstate,
 643                 .get_turbo = core_get_turbo_pstate,
 644                 .get_scaling = core_get_scaling,
 645                 .set = core_set_pstate,
 646         },
 647 };
 648
 649 static struct cpu_defaults byt_params = {
 650         .pid_policy = {
 651                 .sample_rate_ms = 10,
 652                 .deadband = 0,
 653                 .setpoint = 60,
 654                 .p_gain_pct = 14,
 655                 .d_gain_pct = 0,
 656                 .i_gain_pct = 4,
 657         },
 658         .funcs = {
 659                 .get_max = byt_get_max_pstate,
 660                 .get_min = byt_get_min_pstate,
 661                 .get_turbo = byt_get_turbo_pstate,
 662                 .set = byt_set_pstate,
 663                 .get_scaling = byt_get_scaling,
 664                 .get_vid = byt_get_vid,
 665         },
 666 };
 667
 668 static struct cpu_defaults knl_params = {
 669         .pid_policy = {
 670                 .sample_rate_ms = 10,
 671                 .deadband = 0,
 672                 .setpoint = 97,
 673                 .p_gain_pct = 20,
 674                 .d_gain_pct = 0,
 675                 .i_gain_pct = 0,
 676         },
 677         .funcs = {
 678                 .get_max = core_get_max_pstate,
 679                 .get_min = core_get_min_pstate,
 680                 .get_turbo = knl_get_turbo_pstate,
 681                 .get_scaling = core_get_scaling,
 682                 .set = core_set_pstate,
 683         },
 684 };
 685
 686 static void intel_pstate_get_min_max(struct cpudata *cpu, int *min, int *max)
 687 {
 688         int max_perf = cpu->pstate.turbo_pstate;
 689         int max_perf_adj;
 690         int min_perf;
 691
 692         if (limits.no_turbo || limits.turbo_disabled)
 693                 max_perf = cpu->pstate.max_pstate;
 694
 695         /*
 696          * performance can be limited by user through sysfs, by cpufreq
 697          * policy, or by cpu specific default values determined through
 698          * experimentation.
 699          */
 700         max_perf_adj = fp_toint(mul_fp(int_tofp(max_perf), limits.max_perf));
 701         *max = clamp_t(int, max_perf_adj,
 702                         cpu->pstate.min_pstate, cpu->pstate.turbo_pstate);
 703
 704         min_perf = fp_toint(mul_fp(int_tofp(max_perf), limits.min_perf));
 705         *min = clamp_t(int, min_perf, cpu->pstate.min_pstate, max_perf);
 706 }
 707
 708 static void intel_pstate_set_pstate(struct cpudata *cpu, int pstate)
 709 {
 710         int max_perf, min_perf;
 711
 712         update_turbo_state();
 713
 714         intel_pstate_get_min_max(cpu, &min_perf, &max_perf);
 715
 716         pstate = clamp_t(int, pstate, min_perf, max_perf);
 717
 718         if (pstate == cpu->pstate.current_pstate)
 719                 return;
 720
 721         trace_cpu_frequency(pstate * cpu->pstate.scaling, cpu->cpu);
 722
 723         cpu->pstate.current_pstate = pstate;
 724
 725         pstate_funcs.set(cpu, pstate);
 726 }
 727
 728 static void intel_pstate_get_cpu_pstates(struct cpudata *cpu)
 729 {
 730         cpu->pstate.min_pstate = pstate_funcs.get_min();
 731         cpu->pstate.max_pstate = pstate_funcs.get_max();
 732         cpu->pstate.turbo_pstate = pstate_funcs.get_turbo();
 733         cpu->pstate.scaling = pstate_funcs.get_scaling();
 734
 735         if (pstate_funcs.get_vid)
 736                 pstate_funcs.get_vid(cpu);
 737         intel_pstate_set_pstate(cpu, cpu->pstate.min_pstate);
 738 }
 739
 740 static inline void intel_pstate_calc_busy(struct cpudata *cpu)
 741 {
 742         struct sample *sample = &cpu->sample;
 743         int64_t core_pct;
 744
 745         core_pct = int_tofp(sample->aperf) * int_tofp(100);
 746         core_pct = div64_u64(core_pct, int_tofp(sample->mperf));
 747
 748         sample->freq = fp_toint(
 749                 mul_fp(int_tofp(
 750                         cpu->pstate.max_pstate * cpu->pstate.scaling / 100),
 751                         core_pct));
 752
 753         sample->core_pct_busy = (int32_t)core_pct;
 754 }
 755
 756 static inline void intel_pstate_sample(struct cpudata *cpu)
 757 {
 758         u64 aperf, mperf;
 759         unsigned long flags;
 760
 761         local_irq_save(flags);
 762         rdmsrl(MSR_IA32_APERF, aperf);
 763         rdmsrl(MSR_IA32_MPERF, mperf);
 764         local_irq_restore(flags);
 765
 766         cpu->last_sample_time = cpu->sample.time;
 767         cpu->sample.time = ktime_get();
 768         cpu->sample.aperf = aperf;
 769         cpu->sample.mperf = mperf;
 770         cpu->sample.aperf -= cpu->prev_aperf;
 771         cpu->sample.mperf -= cpu->prev_mperf;
 772
 773         intel_pstate_calc_busy(cpu);
 774
 775         cpu->prev_aperf = aperf;
 776         cpu->prev_mperf = mperf;
 777 }
 778
 779 static inline void intel_hwp_set_sample_time(struct cpudata *cpu)
 780 {
 781         int delay;
 782
 783         delay = msecs_to_jiffies(50);
 784         mod_timer_pinned(&cpu->timer, jiffies + delay);
 785 }
 786
 787 static inline void intel_pstate_set_sample_time(struct cpudata *cpu)
 788 {
 789         int delay;
 790
 791         delay = msecs_to_jiffies(pid_params.sample_rate_ms);
 792         mod_timer_pinned(&cpu->timer, jiffies + delay);
 793 }
 794
 795 static inline int32_t intel_pstate_get_scaled_busy(struct cpudata *cpu)
 796 {
 797         int32_t core_busy, max_pstate, current_pstate, sample_ratio;
 798         u32 duration_us;
 799         u32 sample_time;
 800
 801         /*
 802          * core_busy is the ratio of actual performance to max
 803          * max_pstate is the max non turbo pstate available
 804          * current_pstate was the pstate that was requested during
 805          *      the last sample period.
 806          *
 807          * We normalize core_busy, which was our actual percent
 808          * performance to what we requested during the last sample
 809          * period. The result will be a percentage of busy at a
 810          * specified pstate.
 811          */
 812         core_busy = cpu->sample.core_pct_busy;
 813         max_pstate = int_tofp(cpu->pstate.max_pstate);
 814         current_pstate = int_tofp(cpu->pstate.current_pstate);
 815         core_busy = mul_fp(core_busy, div_fp(max_pstate, current_pstate));
 816
 817         /*
 818          * Since we have a deferred timer, it will not fire unless
 819          * we are in C0.  So, determine if the actual elapsed time
 820          * is significantly greater (3x) than our sample interval.  If it
 821          * is, then we were idle for a long enough period of time
 822          * to adjust our busyness.
 823          */
 824         sample_time = pid_params.sample_rate_ms  * USEC_PER_MSEC;
 825         duration_us = (u32) ktime_us_delta(cpu->sample.time,
 826                                            cpu->last_sample_time);
 827         if (duration_us > sample_time * 3) {
 828                 sample_ratio = div_fp(int_tofp(sample_time),
 829                                       int_tofp(duration_us));
 830                 core_busy = mul_fp(core_busy, sample_ratio);
 831         }
 832
 833         return core_busy;
 834 }
 835
 836 static inline void intel_pstate_adjust_busy_pstate(struct cpudata *cpu)
 837 {
 838         int32_t busy_scaled;
 839         struct _pid *pid;
 840         signed int ctl;
 841
 842         pid = &cpu->pid;
 843         busy_scaled = intel_pstate_get_scaled_busy(cpu);
 844
 845         ctl = pid_calc(pid, busy_scaled);
 846
 847         /* Negative values of ctl increase the pstate and vice versa */
 848         intel_pstate_set_pstate(cpu, cpu->pstate.current_pstate - ctl);
 849 }
 850
 851 static void intel_hwp_timer_func(unsigned long __data)
 852 {
 853         struct cpudata *cpu = (struct cpudata *) __data;
 854
 855         intel_pstate_sample(cpu);
 856         intel_hwp_set_sample_time(cpu);
 857 }
 858
 859 static void intel_pstate_timer_func(unsigned long __data)
 860 {
 861         struct cpudata *cpu = (struct cpudata *) __data;
 862         struct sample *sample;
 863
 864         intel_pstate_sample(cpu);
 865
 866         sample = &cpu->sample;
 867
 868         intel_pstate_adjust_busy_pstate(cpu);
 869
 870         trace_pstate_sample(fp_toint(sample->core_pct_busy),
 871                         fp_toint(intel_pstate_get_scaled_busy(cpu)),
 872                         cpu->pstate.current_pstate,
 873                         sample->mperf,
 874                         sample->aperf,
 875                         sample->freq);
 876
 877         intel_pstate_set_sample_time(cpu);
 878 }
 879
 880 #define ICPU(model, policy) \
 881         { X86_VENDOR_INTEL, 6, model, X86_FEATURE_APERFMPERF,\
 882                         (unsigned long)&policy }
 883
 884 static const struct x86_cpu_id intel_pstate_cpu_ids[] = {
 885         ICPU(0x2a, core_params),
 886         ICPU(0x2d, core_params),
 887         ICPU(0x37, byt_params),
 888         ICPU(0x3a, core_params),
 889         ICPU(0x3c, core_params),
 890         ICPU(0x3d, core_params),
 891         ICPU(0x3e, core_params),
 892         ICPU(0x3f, core_params),
 893         ICPU(0x45, core_params),
 894         ICPU(0x46, core_params),
 895         ICPU(0x47, core_params),
 896         ICPU(0x4c, byt_params),
 897         ICPU(0x4e, core_params),
 898         ICPU(0x4f, core_params),
 899         ICPU(0x56, core_params),
 900         ICPU(0x57, knl_params),
 901         {}
 902 };
 903 MODULE_DEVICE_TABLE(x86cpu, intel_pstate_cpu_ids);
 904
 905 static const struct x86_cpu_id intel_pstate_cpu_oob_ids[] = {
 906         ICPU(0x56, core_params),
 907         {}
 908 };
 909
 910 static int intel_pstate_init_cpu(unsigned int cpunum)
 911 {
 912         struct cpudata *cpu;
 913
 914         if (!all_cpu_data[cpunum])
 915                 all_cpu_data[cpunum] = kzalloc(sizeof(struct cpudata),
 916                                                GFP_KERNEL);
 917         if (!all_cpu_data[cpunum])
 918                 return -ENOMEM;
 919
 920         cpu = all_cpu_data[cpunum];
 921
 922         cpu->cpu = cpunum;
 923         intel_pstate_get_cpu_pstates(cpu);
 924
 925         init_timer_deferrable(&cpu->timer);
 926         cpu->timer.data = (unsigned long)cpu;
 927         cpu->timer.expires = jiffies + HZ/100;
 928
 929         if (!hwp_active)
 930                 cpu->timer.function = intel_pstate_timer_func;
 931         else
 932                 cpu->timer.function = intel_hwp_timer_func;
 933
 934         intel_pstate_busy_pid_reset(cpu);
 935         intel_pstate_sample(cpu);
 936
 937         add_timer_on(&cpu->timer, cpunum);
 938
 939         pr_debug("Intel pstate controlling: cpu %d\n", cpunum);
 940
 941         return 0;
 942 }
 943
 944 static unsigned int intel_pstate_get(unsigned int cpu_num)
 945 {
 946         struct sample *sample;
 947         struct cpudata *cpu;
 948
 949         cpu = all_cpu_data[cpu_num];
 950         if (!cpu)
 951                 return 0;
 952         sample = &cpu->sample;
 953         return sample->freq;
 954 }
 955
 956 static int intel_pstate_set_policy(struct cpufreq_policy *policy)
 957 {
 958         if (!policy->cpuinfo.max_freq)
 959                 return -ENODEV;
 960
 961         if (policy->policy == CPUFREQ_POLICY_PERFORMANCE &&
 962             policy->max >= policy->cpuinfo.max_freq) {
 963                 limits.min_policy_pct = 100;
 964                 limits.min_perf_pct = 100;
 965                 limits.min_perf = int_tofp(1);
 966                 limits.max_policy_pct = 100;
 967                 limits.max_perf_pct = 100;
 968                 limits.max_perf = int_tofp(1);
 969                 limits.no_turbo = 0;
 970                 return 0;
 971         }
 972
 973         limits.min_policy_pct = (policy->min * 100) / policy->cpuinfo.max_freq;
 974         limits.min_policy_pct = clamp_t(int, limits.min_policy_pct, 0 , 100);
 975         limits.min_perf_pct = max(limits.min_policy_pct, limits.min_sysfs_pct);
 976         limits.min_perf = div_fp(int_tofp(limits.min_perf_pct), int_tofp(100));
 977
 978         limits.max_policy_pct = (policy->max * 100) / policy->cpuinfo.max_freq;
 979         limits.max_policy_pct = clamp_t(int, limits.max_policy_pct, 0 , 100);
 980         limits.max_perf_pct = min(limits.max_policy_pct, limits.max_sysfs_pct);
 981         limits.max_perf = div_fp(int_tofp(limits.max_perf_pct), int_tofp(100));
 982
 983         if (hwp_active)
 984                 intel_pstate_hwp_set();
 985
 986         return 0;
 987 }
 988
 989 static int intel_pstate_verify_policy(struct cpufreq_policy *policy)
 990 {
 991         cpufreq_verify_within_cpu_limits(policy);
 992
 993         if (policy->policy != CPUFREQ_POLICY_POWERSAVE &&
 994             policy->policy != CPUFREQ_POLICY_PERFORMANCE)
 995                 return -EINVAL;
 996
 997         return 0;
 998 }
 999
1000 static void intel_pstate_stop_cpu(struct cpufreq_policy *policy)
1001 {
1002         int cpu_num = policy->cpu;
1003         struct cpudata *cpu = all_cpu_data[cpu_num];
1004
1005         pr_info("intel_pstate CPU %d exiting\n", cpu_num);
1006
1007         del_timer_sync(&all_cpu_data[cpu_num]->timer);
1008         if (hwp_active)
1009                 return;
1010
1011         intel_pstate_set_pstate(cpu, cpu->pstate.min_pstate);
1012 }
1013
1014 static int intel_pstate_cpu_init(struct cpufreq_policy *policy)
1015 {
1016         struct cpudata *cpu;
1017         int rc;
1018
1019         rc = intel_pstate_init_cpu(policy->cpu);
1020         if (rc)
1021                 return rc;
1022
1023         cpu = all_cpu_data[policy->cpu];
1024
1025         if (limits.min_perf_pct == 100 && limits.max_perf_pct == 100)
1026                 policy->policy = CPUFREQ_POLICY_PERFORMANCE;
1027         else
1028                 policy->policy = CPUFREQ_POLICY_POWERSAVE;
1029
1030         policy->min = cpu->pstate.min_pstate * cpu->pstate.scaling;
1031         policy->max = cpu->pstate.turbo_pstate * cpu->pstate.scaling;
1032
1033         /* cpuinfo and default policy values */
1034         policy->cpuinfo.min_freq = cpu->pstate.min_pstate * cpu->pstate.scaling;
1035         policy->cpuinfo.max_freq =
1036                 cpu->pstate.turbo_pstate * cpu->pstate.scaling;
1037         policy->cpuinfo.transition_latency = CPUFREQ_ETERNAL;
1038         cpumask_set_cpu(policy->cpu, policy->cpus);
1039
1040         return 0;
1041 }
1042
1043 static struct cpufreq_driver intel_pstate_driver = {
1044         .flags          = CPUFREQ_CONST_LOOPS,
1045         .verify         = intel_pstate_verify_policy,
1046         .setpolicy      = intel_pstate_set_policy,
1047         .get            = intel_pstate_get,
1048         .init           = intel_pstate_cpu_init,
1049         .stop_cpu       = intel_pstate_stop_cpu,
1050         .name           = "intel_pstate",
1051 };
1052
1053 static int __initdata no_load;
1054 static int __initdata no_hwp;
1055 static int __initdata hwp_only;
1056 static unsigned int force_load;
1057
1058 static int intel_pstate_msrs_not_valid(void)
1059 {
1060         if (!pstate_funcs.get_max() ||
1061             !pstate_funcs.get_min() ||
1062             !pstate_funcs.get_turbo())
1063                 return -ENODEV;
1064
1065         return 0;
1066 }
1067
1068 static void copy_pid_params(struct pstate_adjust_policy *policy)
1069 {
1070         pid_params.sample_rate_ms = policy->sample_rate_ms;
1071         pid_params.p_gain_pct = policy->p_gain_pct;
1072         pid_params.i_gain_pct = policy->i_gain_pct;
1073         pid_params.d_gain_pct = policy->d_gain_pct;
1074         pid_params.deadband = policy->deadband;
1075         pid_params.setpoint = policy->setpoint;
1076 }
1077
1078 static void copy_cpu_funcs(struct pstate_funcs *funcs)
1079 {
1080         pstate_funcs.get_max   = funcs->get_max;
1081         pstate_funcs.get_min   = funcs->get_min;
1082         pstate_funcs.get_turbo = funcs->get_turbo;
1083         pstate_funcs.get_scaling = funcs->get_scaling;
1084         pstate_funcs.set       = funcs->set;
1085         pstate_funcs.get_vid   = funcs->get_vid;
1086 }
1087
1088 #if IS_ENABLED(CONFIG_ACPI)
1089 #include <acpi/processor.h>
1090
1091 static bool intel_pstate_no_acpi_pss(void)
1092 {
1093         int i;
1094
1095         for_each_possible_cpu(i) {
1096                 acpi_status status;
1097                 union acpi_object *pss;
1098                 struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
1099                 struct acpi_processor *pr = per_cpu(processors, i);
1100
1101                 if (!pr)
1102                         continue;
1103
1104                 status = acpi_evaluate_object(pr->handle, "_PSS", NULL, &buffer);
1105                 if (ACPI_FAILURE(status))
1106                         continue;
1107
1108                 pss = buffer.pointer;
1109                 if (pss && pss->type == ACPI_TYPE_PACKAGE) {
1110                         kfree(pss);
1111                         return false;
1112                 }
1113
1114                 kfree(pss);
1115         }
1116
1117         return true;
1118 }
1119
1120 static bool intel_pstate_has_acpi_ppc(void)
1121 {
1122         int i;
1123
1124         for_each_possible_cpu(i) {
1125                 struct acpi_processor *pr = per_cpu(processors, i);
1126
1127                 if (!pr)
1128                         continue;
1129                 if (acpi_has_method(pr->handle, "_PPC"))
1130                         return true;
1131         }
1132         return false;
1133 }
1134
1135 enum {
1136         PSS,
1137         PPC,
1138 };
1139
1140 struct hw_vendor_info {
1141         u16  valid;
1142         char oem_id[ACPI_OEM_ID_SIZE];
1143         char oem_table_id[ACPI_OEM_TABLE_ID_SIZE];
1144         int  oem_pwr_table;
1145 };
1146
1147 /* Hardware vendor-specific info that has its own power management modes */
1148 static struct hw_vendor_info vendor_info[] = {
1149         {1, "HP    ", "ProLiant", PSS},
1150         {1, "ORACLE", "X4-2    ", PPC},
1151         {1, "ORACLE", "X4-2L   ", PPC},
1152         {1, "ORACLE", "X4-2B   ", PPC},
1153         {1, "ORACLE", "X3-2    ", PPC},
1154         {1, "ORACLE", "X3-2L   ", PPC},
1155         {1, "ORACLE", "X3-2B   ", PPC},
1156         {1, "ORACLE", "X4470M2 ", PPC},
1157         {1, "ORACLE", "X4270M3 ", PPC},
1158         {1, "ORACLE", "X4270M2 ", PPC},
1159         {1, "ORACLE", "X4170M2 ", PPC},
1160         {0, "", ""},
1161 };
1162
1163 static bool intel_pstate_platform_pwr_mgmt_exists(void)
1164 {
1165         struct acpi_table_header hdr;
1166         struct hw_vendor_info *v_info;
1167         const struct x86_cpu_id *id;
1168         u64 misc_pwr;
1169
1170         id = x86_match_cpu(intel_pstate_cpu_oob_ids);
1171         if (id) {
1172                 rdmsrl(MSR_MISC_PWR_MGMT, misc_pwr);
1173                 if ( misc_pwr & (1 << 8))
1174                         return true;
1175         }
1176
1177         if (acpi_disabled ||
1178             ACPI_FAILURE(acpi_get_table_header(ACPI_SIG_FADT, 0, &hdr)))
1179                 return false;
1180
1181         for (v_info = vendor_info; v_info->valid; v_info++) {
1182                 if (!strncmp(hdr.oem_id, v_info->oem_id, ACPI_OEM_ID_SIZE) &&
1183                         !strncmp(hdr.oem_table_id, v_info->oem_table_id,
1184                                                 ACPI_OEM_TABLE_ID_SIZE))
1185                         switch (v_info->oem_pwr_table) {
1186                         case PSS:
1187                                 return intel_pstate_no_acpi_pss();
1188                         case PPC:
1189                                 return intel_pstate_has_acpi_ppc() &&
1190                                         (!force_load);
1191                         }
1192         }
1193
1194         return false;
1195 }
1196 #else /* CONFIG_ACPI not enabled */
1197 static inline bool intel_pstate_platform_pwr_mgmt_exists(void) { return false; }
1198 static inline bool intel_pstate_has_acpi_ppc(void) { return false; }
1199 #endif /* CONFIG_ACPI */
1200
1201 static int __init intel_pstate_init(void)
1202 {
1203         int cpu, rc = 0;
1204         const struct x86_cpu_id *id;
1205         struct cpu_defaults *cpu_def;
1206
1207         if (no_load)
1208                 return -ENODEV;
1209
1210         id = x86_match_cpu(intel_pstate_cpu_ids);
1211         if (!id)
1212                 return -ENODEV;
1213
1214         /*
1215          * The Intel pstate driver will be ignored if the platform
1216          * firmware has its own power management modes.
1217          */
1218         if (intel_pstate_platform_pwr_mgmt_exists())
1219                 return -ENODEV;
1220
1221         cpu_def = (struct cpu_defaults *)id->driver_data;
1222
1223         copy_pid_params(&cpu_def->pid_policy);
1224         copy_cpu_funcs(&cpu_def->funcs);
1225
1226         if (intel_pstate_msrs_not_valid())
1227                 return -ENODEV;
1228
1229         pr_info("Intel P-state driver initializing.\n");
1230
1231         all_cpu_data = vzalloc(sizeof(void *) * num_possible_cpus());
1232         if (!all_cpu_data)
1233                 return -ENOMEM;
1234
1235         if (static_cpu_has_safe(X86_FEATURE_HWP) && !no_hwp)
1236                 intel_pstate_hwp_enable();
1237
1238         if (!hwp_active && hwp_only)
1239                 goto out;
1240
1241         rc = cpufreq_register_driver(&intel_pstate_driver);
1242         if (rc)
1243                 goto out;
1244
1245         intel_pstate_debug_expose_params();
1246         intel_pstate_sysfs_expose_params();
1247
1248         return rc;
1249 out:
1250         get_online_cpus();
1251         for_each_online_cpu(cpu) {
1252                 if (all_cpu_data[cpu]) {
1253                         del_timer_sync(&all_cpu_data[cpu]->timer);
1254                         kfree(all_cpu_data[cpu]);
1255                 }
1256         }
1257
1258         put_online_cpus();
1259         vfree(all_cpu_data);
1260         return -ENODEV;
1261 }
1262 device_initcall(intel_pstate_init);
1263
1264 static int __init intel_pstate_setup(char *str)
1265 {
1266         if (!str)
1267                 return -EINVAL;
1268
1269         if (!strcmp(str, "disable"))
1270                 no_load = 1;
1271         if (!strcmp(str, "no_hwp"))
1272                 no_hwp = 1;
1273         if (!strcmp(str, "force"))
1274                 force_load = 1;
1275         if (!strcmp(str, "hwp_only"))
1276                 hwp_only = 1;
1277         return 0;
1278 }
1279 early_param("intel_pstate", intel_pstate_setup);
1280
1281 MODULE_AUTHOR("Dirk Brandewie <dirk.j.brandewie@intel.com>");
1282 MODULE_DESCRIPTION("'intel_pstate' - P state driver Intel Core processors");
1283 MODULE_LICENSE("GPL");