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
  35 #define BYT_RATIOS              0x66a
  36 #define BYT_VIDS                0x66b
  37 #define BYT_TURBO_RATIOS        0x66c
  38 #define BYT_TURBO_VIDS          0x66d
  39
  40 #define FRAC_BITS 8
  41 #define int_tofp(X) ((int64_t)(X) << FRAC_BITS)
  42 #define fp_toint(X) ((X) >> FRAC_BITS)
  43
  44
  45 static inline int32_t mul_fp(int32_t x, int32_t y)
  46 {
  47         return ((int64_t)x * (int64_t)y) >> FRAC_BITS;
  48 }
  49
  50 static inline int32_t div_fp(int32_t x, int32_t y)
  51 {
  52         return div_s64((int64_t)x << FRAC_BITS, y);
  53 }
  54
  55 static inline int ceiling_fp(int32_t x)
  56 {
  57         int mask, ret;
  58
  59         ret = fp_toint(x);
  60         mask = (1 << FRAC_BITS) - 1;
  61         if (x & mask)
  62                 ret += 1;
  63         return ret;
  64 }
  65
  66 struct sample {
  67         int32_t core_pct_busy;
  68         u64 aperf;
  69         u64 mperf;
  70         int freq;
  71         ktime_t time;
  72 };
  73
  74 struct pstate_data {
  75         int     current_pstate;
  76         int     min_pstate;
  77         int     max_pstate;
  78         int     scaling;
  79         int     turbo_pstate;
  80 };
  81
  82 struct vid_data {
  83         int min;
  84         int max;
  85         int turbo;
  86         int32_t ratio;
  87 };
  88
  89 struct _pid {
  90         int setpoint;
  91         int32_t integral;
  92         int32_t p_gain;
  93         int32_t i_gain;
  94         int32_t d_gain;
  95         int deadband;
  96         int32_t last_err;
  97 };
  98
  99 struct cpudata {
 100         int cpu;
 101
 102         struct timer_list timer;
 103
 104         struct pstate_data pstate;
 105         struct vid_data vid;
 106         struct _pid pid;
 107
 108         ktime_t last_sample_time;
 109         u64     prev_aperf;
 110         u64     prev_mperf;
 111         struct sample sample;
 112 };
 113
 114 static struct cpudata **all_cpu_data;
 115 struct pstate_adjust_policy {
 116         int sample_rate_ms;
 117         int deadband;
 118         int setpoint;
 119         int p_gain_pct;
 120         int d_gain_pct;
 121         int i_gain_pct;
 122 };
 123
 124 struct pstate_funcs {
 125         int (*get_max)(void);
 126         int (*get_min)(void);
 127         int (*get_turbo)(void);
 128         int (*get_scaling)(void);
 129         void (*set)(struct cpudata*, int pstate);
 130         void (*get_vid)(struct cpudata *);
 131 };
 132
 133 struct cpu_defaults {
 134         struct pstate_adjust_policy pid_policy;
 135         struct pstate_funcs funcs;
 136 };
 137
 138 static struct pstate_adjust_policy pid_params;
 139 static struct pstate_funcs pstate_funcs;
 140 static int hwp_active;
 141
 142 struct perf_limits {
 143         int no_turbo;
 144         int turbo_disabled;
 145         int max_perf_pct;
 146         int min_perf_pct;
 147         int32_t max_perf;
 148         int32_t min_perf;
 149         int max_policy_pct;
 150         int max_sysfs_pct;
 151 };
 152
 153 static struct perf_limits limits = {
 154         .no_turbo = 0,
 155         .turbo_disabled = 0,
 156         .max_perf_pct = 100,
 157         .max_perf = int_tofp(1),
 158         .min_perf_pct = 0,
 159         .min_perf = 0,
 160         .max_policy_pct = 100,
 161         .max_sysfs_pct = 100,
 162 };
 163
 164 static inline void pid_reset(struct _pid *pid, int setpoint, int busy,
 165                              int deadband, int integral) {
 166         pid->setpoint = setpoint;
 167         pid->deadband  = deadband;
 168         pid->integral  = int_tofp(integral);
 169         pid->last_err  = int_tofp(setpoint) - int_tofp(busy);
 170 }
 171
 172 static inline void pid_p_gain_set(struct _pid *pid, int percent)
 173 {
 174         pid->p_gain = div_fp(int_tofp(percent), int_tofp(100));
 175 }
 176
 177 static inline void pid_i_gain_set(struct _pid *pid, int percent)
 178 {
 179         pid->i_gain = div_fp(int_tofp(percent), int_tofp(100));
 180 }
 181
 182 static inline void pid_d_gain_set(struct _pid *pid, int percent)
 183 {
 184         pid->d_gain = div_fp(int_tofp(percent), int_tofp(100));
 185 }
 186
 187 static signed int pid_calc(struct _pid *pid, int32_t busy)
 188 {
 189         signed int result;
 190         int32_t pterm, dterm, fp_error;
 191         int32_t integral_limit;
 192
 193         fp_error = int_tofp(pid->setpoint) - busy;
 194
 195         if (abs(fp_error) <= int_tofp(pid->deadband))
 196                 return 0;
 197
 198         pterm = mul_fp(pid->p_gain, fp_error);
 199
 200         pid->integral += fp_error;
 201
 202         /*
 203          * We limit the integral here so that it will never
 204          * get higher than 30.  This prevents it from becoming
 205          * too large an input over long periods of time and allows
 206          * it to get factored out sooner.
 207          *
 208          * The value of 30 was chosen through experimentation.
 209          */
 210         integral_limit = int_tofp(30);
 211         if (pid->integral > integral_limit)
 212                 pid->integral = integral_limit;
 213         if (pid->integral < -integral_limit)
 214                 pid->integral = -integral_limit;
 215
 216         dterm = mul_fp(pid->d_gain, fp_error - pid->last_err);
 217         pid->last_err = fp_error;
 218
 219         result = pterm + mul_fp(pid->integral, pid->i_gain) + dterm;
 220         result = result + (1 << (FRAC_BITS-1));
 221         return (signed int)fp_toint(result);
 222 }
 223
 224 static inline void intel_pstate_busy_pid_reset(struct cpudata *cpu)
 225 {
 226         pid_p_gain_set(&cpu->pid, pid_params.p_gain_pct);
 227         pid_d_gain_set(&cpu->pid, pid_params.d_gain_pct);
 228         pid_i_gain_set(&cpu->pid, pid_params.i_gain_pct);
 229
 230         pid_reset(&cpu->pid, pid_params.setpoint, 100, pid_params.deadband, 0);
 231 }
 232
 233 static inline void intel_pstate_reset_all_pid(void)
 234 {
 235         unsigned int cpu;
 236
 237         for_each_online_cpu(cpu) {
 238                 if (all_cpu_data[cpu])
 239                         intel_pstate_busy_pid_reset(all_cpu_data[cpu]);
 240         }
 241 }
 242
 243 static inline void update_turbo_state(void)
 244 {
 245         u64 misc_en;
 246         struct cpudata *cpu;
 247
 248         cpu = all_cpu_data[0];
 249         rdmsrl(MSR_IA32_MISC_ENABLE, misc_en);
 250         limits.turbo_disabled =
 251                 (misc_en & MSR_IA32_MISC_ENABLE_TURBO_DISABLE ||
 252                  cpu->pstate.max_pstate == cpu->pstate.turbo_pstate);
 253 }
 254
 255 #define PCT_TO_HWP(x) (x * 255 / 100)
 256 static void intel_pstate_hwp_set(void)
 257 {
 258         int min, max, cpu;
 259         u64 value, freq;
 260
 261         get_online_cpus();
 262
 263         for_each_online_cpu(cpu) {
 264                 rdmsrl_on_cpu(cpu, MSR_HWP_REQUEST, &value);
 265                 min = PCT_TO_HWP(limits.min_perf_pct);
 266                 value &= ~HWP_MIN_PERF(~0L);
 267                 value |= HWP_MIN_PERF(min);
 268
 269                 max = PCT_TO_HWP(limits.max_perf_pct);
 270                 if (limits.no_turbo) {
 271                         rdmsrl( MSR_HWP_CAPABILITIES, freq);
 272                         max = HWP_GUARANTEED_PERF(freq);
 273                 }
 274
 275                 value &= ~HWP_MAX_PERF(~0L);
 276                 value |= HWP_MAX_PERF(max);
 277                 wrmsrl_on_cpu(cpu, MSR_HWP_REQUEST, value);
 278         }
 279
 280         put_online_cpus();
 281 }
 282
 283 /************************** debugfs begin ************************/
 284 static int pid_param_set(void *data, u64 val)
 285 {
 286         *(u32 *)data = val;
 287         intel_pstate_reset_all_pid();
 288         return 0;
 289 }
 290
 291 static int pid_param_get(void *data, u64 *val)
 292 {
 293         *val = *(u32 *)data;
 294         return 0;
 295 }
 296 DEFINE_SIMPLE_ATTRIBUTE(fops_pid_param, pid_param_get, pid_param_set, "%llu\n");
 297
 298 struct pid_param {
 299         char *name;
 300         void *value;
 301 };
 302
 303 static struct pid_param pid_files[] = {
 304         {"sample_rate_ms", &pid_params.sample_rate_ms},
 305         {"d_gain_pct", &pid_params.d_gain_pct},
 306         {"i_gain_pct", &pid_params.i_gain_pct},
 307         {"deadband", &pid_params.deadband},
 308         {"setpoint", &pid_params.setpoint},
 309         {"p_gain_pct", &pid_params.p_gain_pct},
 310         {NULL, NULL}
 311 };
 312
 313 static void __init intel_pstate_debug_expose_params(void)
 314 {
 315         struct dentry *debugfs_parent;
 316         int i = 0;
 317
 318         if (hwp_active)
 319                 return;
 320         debugfs_parent = debugfs_create_dir("pstate_snb", NULL);
 321         if (IS_ERR_OR_NULL(debugfs_parent))
 322                 return;
 323         while (pid_files[i].name) {
 324                 debugfs_create_file(pid_files[i].name, 0660,
 325                                     debugfs_parent, pid_files[i].value,
 326                                     &fops_pid_param);
 327                 i++;
 328         }
 329 }
 330
 331 /************************** debugfs end ************************/
 332
 333 /************************** sysfs begin ************************/
 334 #define show_one(file_name, object)                                     \
 335         static ssize_t show_##file_name                                 \
 336         (struct kobject *kobj, struct attribute *attr, char *buf)       \
 337         {                                                               \
 338                 return sprintf(buf, "%u\n", limits.object);             \
 339         }
 340
 341 static ssize_t show_no_turbo(struct kobject *kobj,
 342                              struct attribute *attr, char *buf)
 343 {
 344         ssize_t ret;
 345
 346         update_turbo_state();
 347         if (limits.turbo_disabled)
 348                 ret = sprintf(buf, "%u\n", limits.turbo_disabled);
 349         else
 350                 ret = sprintf(buf, "%u\n", limits.no_turbo);
 351
 352         return ret;
 353 }
 354
 355 static ssize_t store_no_turbo(struct kobject *a, struct attribute *b,
 356                               const char *buf, size_t count)
 357 {
 358         unsigned int input;
 359         int ret;
 360
 361         ret = sscanf(buf, "%u", &input);
 362         if (ret != 1)
 363                 return -EINVAL;
 364
 365         update_turbo_state();
 366         if (limits.turbo_disabled) {
 367                 pr_warn("Turbo disabled by BIOS or unavailable on processor\n");
 368                 return -EPERM;
 369         }
 370
 371         limits.no_turbo = clamp_t(int, input, 0, 1);
 372
 373         if (hwp_active)
 374                 intel_pstate_hwp_set();
 375
 376         return count;
 377 }
 378
 379 static ssize_t store_max_perf_pct(struct kobject *a, struct attribute *b,
 380                                   const char *buf, size_t count)
 381 {
 382         unsigned int input;
 383         int ret;
 384
 385         ret = sscanf(buf, "%u", &input);
 386         if (ret != 1)
 387                 return -EINVAL;
 388
 389         limits.max_sysfs_pct = clamp_t(int, input, 0 , 100);
 390         limits.max_perf_pct = min(limits.max_policy_pct, limits.max_sysfs_pct);
 391         limits.max_perf = div_fp(int_tofp(limits.max_perf_pct), int_tofp(100));
 392
 393         if (hwp_active)
 394                 intel_pstate_hwp_set();
 395         return count;
 396 }
 397
 398 static ssize_t store_min_perf_pct(struct kobject *a, struct attribute *b,
 399                                   const char *buf, size_t count)
 400 {
 401         unsigned int input;
 402         int ret;
 403
 404         ret = sscanf(buf, "%u", &input);
 405         if (ret != 1)
 406                 return -EINVAL;
 407         limits.min_perf_pct = clamp_t(int, input, 0 , 100);
 408         limits.min_perf = div_fp(int_tofp(limits.min_perf_pct), int_tofp(100));
 409
 410         if (hwp_active)
 411                 intel_pstate_hwp_set();
 412         return count;
 413 }
 414
 415 show_one(max_perf_pct, max_perf_pct);
 416 show_one(min_perf_pct, min_perf_pct);
 417
 418 define_one_global_rw(no_turbo);
 419 define_one_global_rw(max_perf_pct);
 420 define_one_global_rw(min_perf_pct);
 421
 422 static struct attribute *intel_pstate_attributes[] = {
 423         &no_turbo.attr,
 424         &max_perf_pct.attr,
 425         &min_perf_pct.attr,
 426         NULL
 427 };
 428
 429 static struct attribute_group intel_pstate_attr_group = {
 430         .attrs = intel_pstate_attributes,
 431 };
 432
 433 static void __init intel_pstate_sysfs_expose_params(void)
 434 {
 435         struct kobject *intel_pstate_kobject;
 436         int rc;
 437
 438         intel_pstate_kobject = kobject_create_and_add("intel_pstate",
 439                                                 &cpu_subsys.dev_root->kobj);
 440         BUG_ON(!intel_pstate_kobject);
 441         rc = sysfs_create_group(intel_pstate_kobject, &intel_pstate_attr_group);
 442         BUG_ON(rc);
 443 }
 444 /************************** sysfs end ************************/
 445
 446 static void intel_pstate_hwp_enable(void)
 447 {
 448         hwp_active++;
 449         pr_info("intel_pstate HWP enabled\n");
 450
 451         wrmsrl( MSR_PM_ENABLE, 0x1);
 452 }
 453
 454 static int byt_get_min_pstate(void)
 455 {
 456         u64 value;
 457
 458         rdmsrl(BYT_RATIOS, value);
 459         return (value >> 8) & 0x7F;
 460 }
 461
 462 static int byt_get_max_pstate(void)
 463 {
 464         u64 value;
 465
 466         rdmsrl(BYT_RATIOS, value);
 467         return (value >> 16) & 0x7F;
 468 }
 469
 470 static int byt_get_turbo_pstate(void)
 471 {
 472         u64 value;
 473
 474         rdmsrl(BYT_TURBO_RATIOS, value);
 475         return value & 0x7F;
 476 }
 477
 478 static void byt_set_pstate(struct cpudata *cpudata, int pstate)
 479 {
 480         u64 val;
 481         int32_t vid_fp;
 482         u32 vid;
 483
 484         val = pstate << 8;
 485         if (limits.no_turbo && !limits.turbo_disabled)
 486                 val |= (u64)1 << 32;
 487
 488         vid_fp = cpudata->vid.min + mul_fp(
 489                 int_tofp(pstate - cpudata->pstate.min_pstate),
 490                 cpudata->vid.ratio);
 491
 492         vid_fp = clamp_t(int32_t, vid_fp, cpudata->vid.min, cpudata->vid.max);
 493         vid = ceiling_fp(vid_fp);
 494
 495         if (pstate > cpudata->pstate.max_pstate)
 496                 vid = cpudata->vid.turbo;
 497
 498         val |= vid;
 499
 500         wrmsrl(MSR_IA32_PERF_CTL, val);
 501 }
 502
 503 #define BYT_BCLK_FREQS 5
 504 static int byt_freq_table[BYT_BCLK_FREQS] = { 833, 1000, 1333, 1167, 800};
 505
 506 static int byt_get_scaling(void)
 507 {
 508         u64 value;
 509         int i;
 510
 511         rdmsrl(MSR_FSB_FREQ, value);
 512         i = value & 0x3;
 513
 514         BUG_ON(i > BYT_BCLK_FREQS);
 515
 516         return byt_freq_table[i] * 100;
 517 }
 518
 519 static void byt_get_vid(struct cpudata *cpudata)
 520 {
 521         u64 value;
 522
 523         rdmsrl(BYT_VIDS, value);
 524         cpudata->vid.min = int_tofp((value >> 8) & 0x7f);
 525         cpudata->vid.max = int_tofp((value >> 16) & 0x7f);
 526         cpudata->vid.ratio = div_fp(
 527                 cpudata->vid.max - cpudata->vid.min,
 528                 int_tofp(cpudata->pstate.max_pstate -
 529                         cpudata->pstate.min_pstate));
 530
 531         rdmsrl(BYT_TURBO_VIDS, value);
 532         cpudata->vid.turbo = value & 0x7f;
 533 }
 534
 535 static int core_get_min_pstate(void)
 536 {
 537         u64 value;
 538
 539         rdmsrl(MSR_PLATFORM_INFO, value);
 540         return (value >> 40) & 0xFF;
 541 }
 542
 543 static int core_get_max_pstate(void)
 544 {
 545         u64 value;
 546
 547         rdmsrl(MSR_PLATFORM_INFO, value);
 548         return (value >> 8) & 0xFF;
 549 }
 550
 551 static int core_get_turbo_pstate(void)
 552 {
 553         u64 value;
 554         int nont, ret;
 555
 556         rdmsrl(MSR_NHM_TURBO_RATIO_LIMIT, value);
 557         nont = core_get_max_pstate();
 558         ret = (value) & 255;
 559         if (ret <= nont)
 560                 ret = nont;
 561         return ret;
 562 }
 563
 564 static inline int core_get_scaling(void)
 565 {
 566         return 100000;
 567 }
 568
 569 static void core_set_pstate(struct cpudata *cpudata, int pstate)
 570 {
 571         u64 val;
 572
 573         val = pstate << 8;
 574         if (limits.no_turbo && !limits.turbo_disabled)
 575                 val |= (u64)1 << 32;
 576
 577         wrmsrl_on_cpu(cpudata->cpu, MSR_IA32_PERF_CTL, val);
 578 }
 579
 580 static struct cpu_defaults core_params = {
 581         .pid_policy = {
 582                 .sample_rate_ms = 10,
 583                 .deadband = 0,
 584                 .setpoint = 97,
 585                 .p_gain_pct = 20,
 586                 .d_gain_pct = 0,
 587                 .i_gain_pct = 0,
 588         },
 589         .funcs = {
 590                 .get_max = core_get_max_pstate,
 591                 .get_min = core_get_min_pstate,
 592                 .get_turbo = core_get_turbo_pstate,
 593                 .get_scaling = core_get_scaling,
 594                 .set = core_set_pstate,
 595         },
 596 };
 597
 598 static struct cpu_defaults byt_params = {
 599         .pid_policy = {
 600                 .sample_rate_ms = 10,
 601                 .deadband = 0,
 602                 .setpoint = 97,
 603                 .p_gain_pct = 14,
 604                 .d_gain_pct = 0,
 605                 .i_gain_pct = 4,
 606         },
 607         .funcs = {
 608                 .get_max = byt_get_max_pstate,
 609                 .get_min = byt_get_min_pstate,
 610                 .get_turbo = byt_get_turbo_pstate,
 611                 .set = byt_set_pstate,
 612                 .get_scaling = byt_get_scaling,
 613                 .get_vid = byt_get_vid,
 614         },
 615 };
 616
 617 static void intel_pstate_get_min_max(struct cpudata *cpu, int *min, int *max)
 618 {
 619         int max_perf = cpu->pstate.turbo_pstate;
 620         int max_perf_adj;
 621         int min_perf;
 622
 623         if (limits.no_turbo || limits.turbo_disabled)
 624                 max_perf = cpu->pstate.max_pstate;
 625
 626         /*
 627          * performance can be limited by user through sysfs, by cpufreq
 628          * policy, or by cpu specific default values determined through
 629          * experimentation.
 630          */
 631         max_perf_adj = fp_toint(mul_fp(int_tofp(max_perf), limits.max_perf));
 632         *max = clamp_t(int, max_perf_adj,
 633                         cpu->pstate.min_pstate, cpu->pstate.turbo_pstate);
 634
 635         min_perf = fp_toint(mul_fp(int_tofp(max_perf), limits.min_perf));
 636         *min = clamp_t(int, min_perf, cpu->pstate.min_pstate, max_perf);
 637 }
 638
 639 static void intel_pstate_set_pstate(struct cpudata *cpu, int pstate)
 640 {
 641         int max_perf, min_perf;
 642
 643         update_turbo_state();
 644
 645         intel_pstate_get_min_max(cpu, &min_perf, &max_perf);
 646
 647         pstate = clamp_t(int, pstate, min_perf, max_perf);
 648
 649         if (pstate == cpu->pstate.current_pstate)
 650                 return;
 651
 652         trace_cpu_frequency(pstate * cpu->pstate.scaling, cpu->cpu);
 653
 654         cpu->pstate.current_pstate = pstate;
 655
 656         pstate_funcs.set(cpu, pstate);
 657 }
 658
 659 static void intel_pstate_get_cpu_pstates(struct cpudata *cpu)
 660 {
 661         cpu->pstate.min_pstate = pstate_funcs.get_min();
 662         cpu->pstate.max_pstate = pstate_funcs.get_max();
 663         cpu->pstate.turbo_pstate = pstate_funcs.get_turbo();
 664         cpu->pstate.scaling = pstate_funcs.get_scaling();
 665
 666         if (pstate_funcs.get_vid)
 667                 pstate_funcs.get_vid(cpu);
 668         intel_pstate_set_pstate(cpu, cpu->pstate.min_pstate);
 669 }
 670
 671 static inline void intel_pstate_calc_busy(struct cpudata *cpu)
 672 {
 673         struct sample *sample = &cpu->sample;
 674         int64_t core_pct;
 675
 676         core_pct = int_tofp(sample->aperf) * int_tofp(100);
 677         core_pct = div64_u64(core_pct, int_tofp(sample->mperf));
 678
 679         sample->freq = fp_toint(
 680                 mul_fp(int_tofp(
 681                         cpu->pstate.max_pstate * cpu->pstate.scaling / 100),
 682                         core_pct));
 683
 684         sample->core_pct_busy = (int32_t)core_pct;
 685 }
 686
 687 static inline void intel_pstate_sample(struct cpudata *cpu)
 688 {
 689         u64 aperf, mperf;
 690         unsigned long flags;
 691
 692         local_irq_save(flags);
 693         rdmsrl(MSR_IA32_APERF, aperf);
 694         rdmsrl(MSR_IA32_MPERF, mperf);
 695         local_irq_restore(flags);
 696
 697         cpu->last_sample_time = cpu->sample.time;
 698         cpu->sample.time = ktime_get();
 699         cpu->sample.aperf = aperf;
 700         cpu->sample.mperf = mperf;
 701         cpu->sample.aperf -= cpu->prev_aperf;
 702         cpu->sample.mperf -= cpu->prev_mperf;
 703
 704         intel_pstate_calc_busy(cpu);
 705
 706         cpu->prev_aperf = aperf;
 707         cpu->prev_mperf = mperf;
 708 }
 709
 710 static inline void intel_hwp_set_sample_time(struct cpudata *cpu)
 711 {
 712         int delay;
 713
 714         delay = msecs_to_jiffies(50);
 715         mod_timer_pinned(&cpu->timer, jiffies + delay);
 716 }
 717
 718 static inline void intel_pstate_set_sample_time(struct cpudata *cpu)
 719 {
 720         int delay;
 721
 722         delay = msecs_to_jiffies(pid_params.sample_rate_ms);
 723         mod_timer_pinned(&cpu->timer, jiffies + delay);
 724 }
 725
 726 static inline int32_t intel_pstate_get_scaled_busy(struct cpudata *cpu)
 727 {
 728         int32_t core_busy, max_pstate, current_pstate, sample_ratio;
 729         u32 duration_us;
 730         u32 sample_time;
 731
 732         /*
 733          * core_busy is the ratio of actual performance to max
 734          * max_pstate is the max non turbo pstate available
 735          * current_pstate was the pstate that was requested during
 736          *      the last sample period.
 737          *
 738          * We normalize core_busy, which was our actual percent
 739          * performance to what we requested during the last sample
 740          * period. The result will be a percentage of busy at a
 741          * specified pstate.
 742          */
 743         core_busy = cpu->sample.core_pct_busy;
 744         max_pstate = int_tofp(cpu->pstate.max_pstate);
 745         current_pstate = int_tofp(cpu->pstate.current_pstate);
 746         core_busy = mul_fp(core_busy, div_fp(max_pstate, current_pstate));
 747
 748         /*
 749          * Since we have a deferred timer, it will not fire unless
 750          * we are in C0.  So, determine if the actual elapsed time
 751          * is significantly greater (3x) than our sample interval.  If it
 752          * is, then we were idle for a long enough period of time
 753          * to adjust our busyness.
 754          */
 755         sample_time = pid_params.sample_rate_ms  * USEC_PER_MSEC;
 756         duration_us = (u32) ktime_us_delta(cpu->sample.time,
 757                                            cpu->last_sample_time);
 758         if (duration_us > sample_time * 3) {
 759                 sample_ratio = div_fp(int_tofp(sample_time),
 760                                       int_tofp(duration_us));
 761                 core_busy = mul_fp(core_busy, sample_ratio);
 762         }
 763
 764         return core_busy;
 765 }
 766
 767 static inline void intel_pstate_adjust_busy_pstate(struct cpudata *cpu)
 768 {
 769         int32_t busy_scaled;
 770         struct _pid *pid;
 771         signed int ctl;
 772
 773         pid = &cpu->pid;
 774         busy_scaled = intel_pstate_get_scaled_busy(cpu);
 775
 776         ctl = pid_calc(pid, busy_scaled);
 777
 778         /* Negative values of ctl increase the pstate and vice versa */
 779         intel_pstate_set_pstate(cpu, cpu->pstate.current_pstate - ctl);
 780 }
 781
 782 static void intel_hwp_timer_func(unsigned long __data)
 783 {
 784         struct cpudata *cpu = (struct cpudata *) __data;
 785
 786         intel_pstate_sample(cpu);
 787         intel_hwp_set_sample_time(cpu);
 788 }
 789
 790 static void intel_pstate_timer_func(unsigned long __data)
 791 {
 792         struct cpudata *cpu = (struct cpudata *) __data;
 793         struct sample *sample;
 794
 795         intel_pstate_sample(cpu);
 796
 797         sample = &cpu->sample;
 798
 799         intel_pstate_adjust_busy_pstate(cpu);
 800
 801         trace_pstate_sample(fp_toint(sample->core_pct_busy),
 802                         fp_toint(intel_pstate_get_scaled_busy(cpu)),
 803                         cpu->pstate.current_pstate,
 804                         sample->mperf,
 805                         sample->aperf,
 806                         sample->freq);
 807
 808         intel_pstate_set_sample_time(cpu);
 809 }
 810
 811 #define ICPU(model, policy) \
 812         { X86_VENDOR_INTEL, 6, model, X86_FEATURE_APERFMPERF,\
 813                         (unsigned long)&policy }
 814
 815 static const struct x86_cpu_id intel_pstate_cpu_ids[] = {
 816         ICPU(0x2a, core_params),
 817         ICPU(0x2d, core_params),
 818         ICPU(0x37, byt_params),
 819         ICPU(0x3a, core_params),
 820         ICPU(0x3c, core_params),
 821         ICPU(0x3d, core_params),
 822         ICPU(0x3e, core_params),
 823         ICPU(0x3f, core_params),
 824         ICPU(0x45, core_params),
 825         ICPU(0x46, core_params),
 826         ICPU(0x47, core_params),
 827         ICPU(0x4c, byt_params),
 828         ICPU(0x4f, core_params),
 829         ICPU(0x56, core_params),
 830         {}
 831 };
 832 MODULE_DEVICE_TABLE(x86cpu, intel_pstate_cpu_ids);
 833
 834 static const struct x86_cpu_id intel_pstate_cpu_oob_ids[] = {
 835         ICPU(0x56, core_params),
 836         {}
 837 };
 838
 839 static int intel_pstate_init_cpu(unsigned int cpunum)
 840 {
 841         struct cpudata *cpu;
 842
 843         if (!all_cpu_data[cpunum])
 844                 all_cpu_data[cpunum] = kzalloc(sizeof(struct cpudata),
 845                                                GFP_KERNEL);
 846         if (!all_cpu_data[cpunum])
 847                 return -ENOMEM;
 848
 849         cpu = all_cpu_data[cpunum];
 850
 851         cpu->cpu = cpunum;
 852         intel_pstate_get_cpu_pstates(cpu);
 853
 854         init_timer_deferrable(&cpu->timer);
 855         cpu->timer.data = (unsigned long)cpu;
 856         cpu->timer.expires = jiffies + HZ/100;
 857
 858         if (!hwp_active)
 859                 cpu->timer.function = intel_pstate_timer_func;
 860         else
 861                 cpu->timer.function = intel_hwp_timer_func;
 862
 863         intel_pstate_busy_pid_reset(cpu);
 864         intel_pstate_sample(cpu);
 865
 866         add_timer_on(&cpu->timer, cpunum);
 867
 868         pr_debug("Intel pstate controlling: cpu %d\n", cpunum);
 869
 870         return 0;
 871 }
 872
 873 static unsigned int intel_pstate_get(unsigned int cpu_num)
 874 {
 875         struct sample *sample;
 876         struct cpudata *cpu;
 877
 878         cpu = all_cpu_data[cpu_num];
 879         if (!cpu)
 880                 return 0;
 881         sample = &cpu->sample;
 882         return sample->freq;
 883 }
 884
 885 static int intel_pstate_set_policy(struct cpufreq_policy *policy)
 886 {
 887         if (!policy->cpuinfo.max_freq)
 888                 return -ENODEV;
 889
 890         if (policy->policy == CPUFREQ_POLICY_PERFORMANCE) {
 891                 limits.min_perf_pct = 100;
 892                 limits.min_perf = int_tofp(1);
 893                 limits.max_policy_pct = 100;
 894                 limits.max_perf_pct = 100;
 895                 limits.max_perf = int_tofp(1);
 896                 limits.no_turbo = 0;
 897                 return 0;
 898         }
 899
 900         limits.min_perf_pct = (policy->min * 100) / policy->cpuinfo.max_freq;
 901         limits.min_perf_pct = clamp_t(int, limits.min_perf_pct, 0 , 100);
 902         limits.min_perf = div_fp(int_tofp(limits.min_perf_pct), int_tofp(100));
 903
 904         limits.max_policy_pct = (policy->max * 100) / policy->cpuinfo.max_freq;
 905         limits.max_policy_pct = clamp_t(int, limits.max_policy_pct, 0 , 100);
 906         limits.max_perf_pct = min(limits.max_policy_pct, limits.max_sysfs_pct);
 907         limits.max_perf = div_fp(int_tofp(limits.max_perf_pct), int_tofp(100));
 908
 909         if (hwp_active)
 910                 intel_pstate_hwp_set();
 911
 912         return 0;
 913 }
 914
 915 static int intel_pstate_verify_policy(struct cpufreq_policy *policy)
 916 {
 917         cpufreq_verify_within_cpu_limits(policy);
 918
 919         if (policy->policy != CPUFREQ_POLICY_POWERSAVE &&
 920             policy->policy != CPUFREQ_POLICY_PERFORMANCE)
 921                 return -EINVAL;
 922
 923         return 0;
 924 }
 925
 926 static void intel_pstate_stop_cpu(struct cpufreq_policy *policy)
 927 {
 928         int cpu_num = policy->cpu;
 929         struct cpudata *cpu = all_cpu_data[cpu_num];
 930
 931         pr_info("intel_pstate CPU %d exiting\n", cpu_num);
 932
 933         del_timer_sync(&all_cpu_data[cpu_num]->timer);
 934         if (hwp_active)
 935                 return;
 936
 937         intel_pstate_set_pstate(cpu, cpu->pstate.min_pstate);
 938 }
 939
 940 static int intel_pstate_cpu_init(struct cpufreq_policy *policy)
 941 {
 942         struct cpudata *cpu;
 943         int rc;
 944
 945         rc = intel_pstate_init_cpu(policy->cpu);
 946         if (rc)
 947                 return rc;
 948
 949         cpu = all_cpu_data[policy->cpu];
 950
 951         if (limits.min_perf_pct == 100 && limits.max_perf_pct == 100)
 952                 policy->policy = CPUFREQ_POLICY_PERFORMANCE;
 953         else
 954                 policy->policy = CPUFREQ_POLICY_POWERSAVE;
 955
 956         policy->min = cpu->pstate.min_pstate * cpu->pstate.scaling;
 957         policy->max = cpu->pstate.turbo_pstate * cpu->pstate.scaling;
 958
 959         /* cpuinfo and default policy values */
 960         policy->cpuinfo.min_freq = cpu->pstate.min_pstate * cpu->pstate.scaling;
 961         policy->cpuinfo.max_freq =
 962                 cpu->pstate.turbo_pstate * cpu->pstate.scaling;
 963         policy->cpuinfo.transition_latency = CPUFREQ_ETERNAL;
 964         cpumask_set_cpu(policy->cpu, policy->cpus);
 965
 966         return 0;
 967 }
 968
 969 static struct cpufreq_driver intel_pstate_driver = {
 970         .flags          = CPUFREQ_CONST_LOOPS,
 971         .verify         = intel_pstate_verify_policy,
 972         .setpolicy      = intel_pstate_set_policy,
 973         .get            = intel_pstate_get,
 974         .init           = intel_pstate_cpu_init,
 975         .stop_cpu       = intel_pstate_stop_cpu,
 976         .name           = "intel_pstate",
 977 };
 978
 979 static int __initdata no_load;
 980 static int __initdata no_hwp;
 981 static unsigned int force_load;
 982
 983 static int intel_pstate_msrs_not_valid(void)
 984 {
 985         /* Check that all the msr's we are using are valid. */
 986         u64 aperf, mperf, tmp;
 987
 988         rdmsrl(MSR_IA32_APERF, aperf);
 989         rdmsrl(MSR_IA32_MPERF, mperf);
 990
 991         if (!pstate_funcs.get_max() ||
 992             !pstate_funcs.get_min() ||
 993             !pstate_funcs.get_turbo())
 994                 return -ENODEV;
 995
 996         rdmsrl(MSR_IA32_APERF, tmp);
 997         if (!(tmp - aperf))
 998                 return -ENODEV;
 999
1000         rdmsrl(MSR_IA32_MPERF, tmp);
1001         if (!(tmp - mperf))
1002                 return -ENODEV;
1003
1004         return 0;
1005 }
1006
1007 static void copy_pid_params(struct pstate_adjust_policy *policy)
1008 {
1009         pid_params.sample_rate_ms = policy->sample_rate_ms;
1010         pid_params.p_gain_pct = policy->p_gain_pct;
1011         pid_params.i_gain_pct = policy->i_gain_pct;
1012         pid_params.d_gain_pct = policy->d_gain_pct;
1013         pid_params.deadband = policy->deadband;
1014         pid_params.setpoint = policy->setpoint;
1015 }
1016
1017 static void copy_cpu_funcs(struct pstate_funcs *funcs)
1018 {
1019         pstate_funcs.get_max   = funcs->get_max;
1020         pstate_funcs.get_min   = funcs->get_min;
1021         pstate_funcs.get_turbo = funcs->get_turbo;
1022         pstate_funcs.get_scaling = funcs->get_scaling;
1023         pstate_funcs.set       = funcs->set;
1024         pstate_funcs.get_vid   = funcs->get_vid;
1025 }
1026
1027 #if IS_ENABLED(CONFIG_ACPI)
1028 #include <acpi/processor.h>
1029
1030 static bool intel_pstate_no_acpi_pss(void)
1031 {
1032         int i;
1033
1034         for_each_possible_cpu(i) {
1035                 acpi_status status;
1036                 union acpi_object *pss;
1037                 struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
1038                 struct acpi_processor *pr = per_cpu(processors, i);
1039
1040                 if (!pr)
1041                         continue;
1042
1043                 status = acpi_evaluate_object(pr->handle, "_PSS", NULL, &buffer);
1044                 if (ACPI_FAILURE(status))
1045                         continue;
1046
1047                 pss = buffer.pointer;
1048                 if (pss && pss->type == ACPI_TYPE_PACKAGE) {
1049                         kfree(pss);
1050                         return false;
1051                 }
1052
1053                 kfree(pss);
1054         }
1055
1056         return true;
1057 }
1058
1059 static bool intel_pstate_has_acpi_ppc(void)
1060 {
1061         int i;
1062
1063         for_each_possible_cpu(i) {
1064                 struct acpi_processor *pr = per_cpu(processors, i);
1065
1066                 if (!pr)
1067                         continue;
1068                 if (acpi_has_method(pr->handle, "_PPC"))
1069                         return true;
1070         }
1071         return false;
1072 }
1073
1074 enum {
1075         PSS,
1076         PPC,
1077 };
1078
1079 struct hw_vendor_info {
1080         u16  valid;
1081         char oem_id[ACPI_OEM_ID_SIZE];
1082         char oem_table_id[ACPI_OEM_TABLE_ID_SIZE];
1083         int  oem_pwr_table;
1084 };
1085
1086 /* Hardware vendor-specific info that has its own power management modes */
1087 static struct hw_vendor_info vendor_info[] = {
1088         {1, "HP    ", "ProLiant", PSS},
1089         {1, "ORACLE", "X4-2    ", PPC},
1090         {1, "ORACLE", "X4-2L   ", PPC},
1091         {1, "ORACLE", "X4-2B   ", PPC},
1092         {1, "ORACLE", "X3-2    ", PPC},
1093         {1, "ORACLE", "X3-2L   ", PPC},
1094         {1, "ORACLE", "X3-2B   ", PPC},
1095         {1, "ORACLE", "X4470M2 ", PPC},
1096         {1, "ORACLE", "X4270M3 ", PPC},
1097         {1, "ORACLE", "X4270M2 ", PPC},
1098         {1, "ORACLE", "X4170M2 ", PPC},
1099         {0, "", ""},
1100 };
1101
1102 static bool intel_pstate_platform_pwr_mgmt_exists(void)
1103 {
1104         struct acpi_table_header hdr;
1105         struct hw_vendor_info *v_info;
1106         const struct x86_cpu_id *id;
1107         u64 misc_pwr;
1108
1109         id = x86_match_cpu(intel_pstate_cpu_oob_ids);
1110         if (id) {
1111                 rdmsrl(MSR_MISC_PWR_MGMT, misc_pwr);
1112                 if ( misc_pwr & (1 << 8))
1113                         return true;
1114         }
1115
1116         if (acpi_disabled ||
1117             ACPI_FAILURE(acpi_get_table_header(ACPI_SIG_FADT, 0, &hdr)))
1118                 return false;
1119
1120         for (v_info = vendor_info; v_info->valid; v_info++) {
1121                 if (!strncmp(hdr.oem_id, v_info->oem_id, ACPI_OEM_ID_SIZE) &&
1122                         !strncmp(hdr.oem_table_id, v_info->oem_table_id,
1123                                                 ACPI_OEM_TABLE_ID_SIZE))
1124                         switch (v_info->oem_pwr_table) {
1125                         case PSS:
1126                                 return intel_pstate_no_acpi_pss();
1127                         case PPC:
1128                                 return intel_pstate_has_acpi_ppc() &&
1129                                         (!force_load);
1130                         }
1131         }
1132
1133         return false;
1134 }
1135 #else /* CONFIG_ACPI not enabled */
1136 static inline bool intel_pstate_platform_pwr_mgmt_exists(void) { return false; }
1137 static inline bool intel_pstate_has_acpi_ppc(void) { return false; }
1138 #endif /* CONFIG_ACPI */
1139
1140 static int __init intel_pstate_init(void)
1141 {
1142         int cpu, rc = 0;
1143         const struct x86_cpu_id *id;
1144         struct cpu_defaults *cpu_info;
1145         struct cpuinfo_x86 *c = &boot_cpu_data;
1146
1147         if (no_load)
1148                 return -ENODEV;
1149
1150         id = x86_match_cpu(intel_pstate_cpu_ids);
1151         if (!id)
1152                 return -ENODEV;
1153
1154         /*
1155          * The Intel pstate driver will be ignored if the platform
1156          * firmware has its own power management modes.
1157          */
1158         if (intel_pstate_platform_pwr_mgmt_exists())
1159                 return -ENODEV;
1160
1161         cpu_info = (struct cpu_defaults *)id->driver_data;
1162
1163         copy_pid_params(&cpu_info->pid_policy);
1164         copy_cpu_funcs(&cpu_info->funcs);
1165
1166         if (intel_pstate_msrs_not_valid())
1167                 return -ENODEV;
1168
1169         pr_info("Intel P-state driver initializing.\n");
1170
1171         all_cpu_data = vzalloc(sizeof(void *) * num_possible_cpus());
1172         if (!all_cpu_data)
1173                 return -ENOMEM;
1174
1175         if (cpu_has(c,X86_FEATURE_HWP) && !no_hwp)
1176                 intel_pstate_hwp_enable();
1177
1178         rc = cpufreq_register_driver(&intel_pstate_driver);
1179         if (rc)
1180                 goto out;
1181
1182         intel_pstate_debug_expose_params();
1183         intel_pstate_sysfs_expose_params();
1184
1185         return rc;
1186 out:
1187         get_online_cpus();
1188         for_each_online_cpu(cpu) {
1189                 if (all_cpu_data[cpu]) {
1190                         del_timer_sync(&all_cpu_data[cpu]->timer);
1191                         kfree(all_cpu_data[cpu]);
1192                 }
1193         }
1194
1195         put_online_cpus();
1196         vfree(all_cpu_data);
1197         return -ENODEV;
1198 }
1199 device_initcall(intel_pstate_init);
1200
1201 static int __init intel_pstate_setup(char *str)
1202 {
1203         if (!str)
1204                 return -EINVAL;
1205
1206         if (!strcmp(str, "disable"))
1207                 no_load = 1;
1208         if (!strcmp(str, "no_hwp"))
1209                 no_hwp = 1;
1210         if (!strcmp(str, "force"))
1211                 force_load = 1;
1212         return 0;
1213 }
1214 early_param("intel_pstate", intel_pstate_setup);
1215
1216 MODULE_AUTHOR("Dirk Brandewie <dirk.j.brandewie@intel.com>");
1217 MODULE_DESCRIPTION("'intel_pstate' - P state driver Intel Core processors");
1218 MODULE_LICENSE("GPL");