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 <trace/events/power.h>
  29
  30 #include <asm/div64.h>
  31 #include <asm/msr.h>
  32 #include <asm/cpu_device_id.h>
  33
  34 #define SAMPLE_COUNT            3
  35
  36 #define FRAC_BITS 8
  37 #define int_tofp(X) ((int64_t)(X) << FRAC_BITS)
  38 #define fp_toint(X) ((X) >> FRAC_BITS)
  39
  40 static inline int32_t mul_fp(int32_t x, int32_t y)
  41 {
  42         return ((int64_t)x * (int64_t)y) >> FRAC_BITS;
  43 }
  44
  45 static inline int32_t div_fp(int32_t x, int32_t y)
  46 {
  47         return div_s64((int64_t)x << FRAC_BITS, (int64_t)y);
  48 }
  49
  50 struct sample {
  51         int32_t core_pct_busy;
  52         u64 aperf;
  53         u64 mperf;
  54         int freq;
  55 };
  56
  57 struct pstate_data {
  58         int     current_pstate;
  59         int     min_pstate;
  60         int     max_pstate;
  61         int     turbo_pstate;
  62 };
  63
  64 struct _pid {
  65         int setpoint;
  66         int32_t integral;
  67         int32_t p_gain;
  68         int32_t i_gain;
  69         int32_t d_gain;
  70         int deadband;
  71         int32_t last_err;
  72 };
  73
  74 struct cpudata {
  75         int cpu;
  76
  77         char name[64];
  78
  79         struct timer_list timer;
  80
  81         struct pstate_adjust_policy *pstate_policy;
  82         struct pstate_data pstate;
  83         struct _pid pid;
  84
  85         int min_pstate_count;
  86
  87         u64     prev_aperf;
  88         u64     prev_mperf;
  89         int     sample_ptr;
  90         struct sample samples[SAMPLE_COUNT];
  91 };
  92
  93 static struct cpudata **all_cpu_data;
  94 struct pstate_adjust_policy {
  95         int sample_rate_ms;
  96         int deadband;
  97         int setpoint;
  98         int p_gain_pct;
  99         int d_gain_pct;
 100         int i_gain_pct;
 101 };
 102
 103 static struct pstate_adjust_policy default_policy = {
 104         .sample_rate_ms = 10,
 105         .deadband = 0,
 106         .setpoint = 97,
 107         .p_gain_pct = 20,
 108         .d_gain_pct = 0,
 109         .i_gain_pct = 0,
 110 };
 111
 112 struct perf_limits {
 113         int no_turbo;
 114         int max_perf_pct;
 115         int min_perf_pct;
 116         int32_t max_perf;
 117         int32_t min_perf;
 118         int max_policy_pct;
 119         int max_sysfs_pct;
 120 };
 121
 122 static struct perf_limits limits = {
 123         .no_turbo = 0,
 124         .max_perf_pct = 100,
 125         .max_perf = int_tofp(1),
 126         .min_perf_pct = 0,
 127         .min_perf = 0,
 128         .max_policy_pct = 100,
 129         .max_sysfs_pct = 100,
 130 };
 131
 132 static inline void pid_reset(struct _pid *pid, int setpoint, int busy,
 133                         int deadband, int integral) {
 134         pid->setpoint = setpoint;
 135         pid->deadband  = deadband;
 136         pid->integral  = int_tofp(integral);
 137         pid->last_err  = setpoint - busy;
 138 }
 139
 140 static inline void pid_p_gain_set(struct _pid *pid, int percent)
 141 {
 142         pid->p_gain = div_fp(int_tofp(percent), int_tofp(100));
 143 }
 144
 145 static inline void pid_i_gain_set(struct _pid *pid, int percent)
 146 {
 147         pid->i_gain = div_fp(int_tofp(percent), int_tofp(100));
 148 }
 149
 150 static inline void pid_d_gain_set(struct _pid *pid, int percent)
 151 {
 152
 153         pid->d_gain = div_fp(int_tofp(percent), int_tofp(100));
 154 }
 155
 156 static signed int pid_calc(struct _pid *pid, int32_t busy)
 157 {
 158         signed int result;
 159         int32_t pterm, dterm, fp_error;
 160         int32_t integral_limit;
 161
 162         fp_error = int_tofp(pid->setpoint) - busy;
 163
 164         if (abs(fp_error) <= int_tofp(pid->deadband))
 165                 return 0;
 166
 167         pterm = mul_fp(pid->p_gain, fp_error);
 168
 169         pid->integral += fp_error;
 170
 171         /* limit the integral term */
 172         integral_limit = int_tofp(30);
 173         if (pid->integral > integral_limit)
 174                 pid->integral = integral_limit;
 175         if (pid->integral < -integral_limit)
 176                 pid->integral = -integral_limit;
 177
 178         dterm = mul_fp(pid->d_gain, fp_error - pid->last_err);
 179         pid->last_err = fp_error;
 180
 181         result = pterm + mul_fp(pid->integral, pid->i_gain) + dterm;
 182
 183         return (signed int)fp_toint(result);
 184 }
 185
 186 static inline void intel_pstate_busy_pid_reset(struct cpudata *cpu)
 187 {
 188         pid_p_gain_set(&cpu->pid, cpu->pstate_policy->p_gain_pct);
 189         pid_d_gain_set(&cpu->pid, cpu->pstate_policy->d_gain_pct);
 190         pid_i_gain_set(&cpu->pid, cpu->pstate_policy->i_gain_pct);
 191
 192         pid_reset(&cpu->pid,
 193                 cpu->pstate_policy->setpoint,
 194                 100,
 195                 cpu->pstate_policy->deadband,
 196                 0);
 197 }
 198
 199 static inline void intel_pstate_reset_all_pid(void)
 200 {
 201         unsigned int cpu;
 202         for_each_online_cpu(cpu) {
 203                 if (all_cpu_data[cpu])
 204                         intel_pstate_busy_pid_reset(all_cpu_data[cpu]);
 205         }
 206 }
 207
 208 /************************** debugfs begin ************************/
 209 static int pid_param_set(void *data, u64 val)
 210 {
 211         *(u32 *)data = val;
 212         intel_pstate_reset_all_pid();
 213         return 0;
 214 }
 215 static int pid_param_get(void *data, u64 *val)
 216 {
 217         *val = *(u32 *)data;
 218         return 0;
 219 }
 220 DEFINE_SIMPLE_ATTRIBUTE(fops_pid_param, pid_param_get,
 221                         pid_param_set, "%llu\n");
 222
 223 struct pid_param {
 224         char *name;
 225         void *value;
 226 };
 227
 228 static struct pid_param pid_files[] = {
 229         {"sample_rate_ms", &default_policy.sample_rate_ms},
 230         {"d_gain_pct", &default_policy.d_gain_pct},
 231         {"i_gain_pct", &default_policy.i_gain_pct},
 232         {"deadband", &default_policy.deadband},
 233         {"setpoint", &default_policy.setpoint},
 234         {"p_gain_pct", &default_policy.p_gain_pct},
 235         {NULL, NULL}
 236 };
 237
 238 static struct dentry *debugfs_parent;
 239 static void intel_pstate_debug_expose_params(void)
 240 {
 241         int i = 0;
 242
 243         debugfs_parent = debugfs_create_dir("pstate_snb", NULL);
 244         if (IS_ERR_OR_NULL(debugfs_parent))
 245                 return;
 246         while (pid_files[i].name) {
 247                 debugfs_create_file(pid_files[i].name, 0660,
 248                                 debugfs_parent, pid_files[i].value,
 249                                 &fops_pid_param);
 250                 i++;
 251         }
 252 }
 253
 254 /************************** debugfs end ************************/
 255
 256 /************************** sysfs begin ************************/
 257 #define show_one(file_name, object)                                     \
 258         static ssize_t show_##file_name                                 \
 259         (struct kobject *kobj, struct attribute *attr, char *buf)       \
 260         {                                                               \
 261                 return sprintf(buf, "%u\n", limits.object);             \
 262         }
 263
 264 static ssize_t store_no_turbo(struct kobject *a, struct attribute *b,
 265                                 const char *buf, size_t count)
 266 {
 267         unsigned int input;
 268         int ret;
 269         ret = sscanf(buf, "%u", &input);
 270         if (ret != 1)
 271                 return -EINVAL;
 272         limits.no_turbo = clamp_t(int, input, 0 , 1);
 273
 274         return count;
 275 }
 276
 277 static ssize_t store_max_perf_pct(struct kobject *a, struct attribute *b,
 278                                 const char *buf, size_t count)
 279 {
 280         unsigned int input;
 281         int ret;
 282         ret = sscanf(buf, "%u", &input);
 283         if (ret != 1)
 284                 return -EINVAL;
 285
 286         limits.max_sysfs_pct = clamp_t(int, input, 0 , 100);
 287         limits.max_perf_pct = min(limits.max_policy_pct, limits.max_sysfs_pct);
 288         limits.max_perf = div_fp(int_tofp(limits.max_perf_pct), int_tofp(100));
 289         return count;
 290 }
 291
 292 static ssize_t store_min_perf_pct(struct kobject *a, struct attribute *b,
 293                                 const char *buf, size_t count)
 294 {
 295         unsigned int input;
 296         int ret;
 297         ret = sscanf(buf, "%u", &input);
 298         if (ret != 1)
 299                 return -EINVAL;
 300         limits.min_perf_pct = clamp_t(int, input, 0 , 100);
 301         limits.min_perf = div_fp(int_tofp(limits.min_perf_pct), int_tofp(100));
 302
 303         return count;
 304 }
 305
 306 show_one(no_turbo, no_turbo);
 307 show_one(max_perf_pct, max_perf_pct);
 308 show_one(min_perf_pct, min_perf_pct);
 309
 310 define_one_global_rw(no_turbo);
 311 define_one_global_rw(max_perf_pct);
 312 define_one_global_rw(min_perf_pct);
 313
 314 static struct attribute *intel_pstate_attributes[] = {
 315         &no_turbo.attr,
 316         &max_perf_pct.attr,
 317         &min_perf_pct.attr,
 318         NULL
 319 };
 320
 321 static struct attribute_group intel_pstate_attr_group = {
 322         .attrs = intel_pstate_attributes,
 323 };
 324 static struct kobject *intel_pstate_kobject;
 325
 326 static void intel_pstate_sysfs_expose_params(void)
 327 {
 328         int rc;
 329
 330         intel_pstate_kobject = kobject_create_and_add("intel_pstate",
 331                                                 &cpu_subsys.dev_root->kobj);
 332         BUG_ON(!intel_pstate_kobject);
 333         rc = sysfs_create_group(intel_pstate_kobject,
 334                                 &intel_pstate_attr_group);
 335         BUG_ON(rc);
 336 }
 337
 338 /************************** sysfs end ************************/
 339
 340 static int intel_pstate_min_pstate(void)
 341 {
 342         u64 value;
 343         rdmsrl(MSR_PLATFORM_INFO, value);
 344         return (value >> 40) & 0xFF;
 345 }
 346
 347 static int intel_pstate_max_pstate(void)
 348 {
 349         u64 value;
 350         rdmsrl(MSR_PLATFORM_INFO, value);
 351         return (value >> 8) & 0xFF;
 352 }
 353
 354 static int intel_pstate_turbo_pstate(void)
 355 {
 356         u64 value;
 357         int nont, ret;
 358         rdmsrl(MSR_NHM_TURBO_RATIO_LIMIT, value);
 359         nont = intel_pstate_max_pstate();
 360         ret = ((value) & 255);
 361         if (ret <= nont)
 362                 ret = nont;
 363         return ret;
 364 }
 365
 366 static void intel_pstate_get_min_max(struct cpudata *cpu, int *min, int *max)
 367 {
 368         int max_perf = cpu->pstate.turbo_pstate;
 369         int max_perf_adj;
 370         int min_perf;
 371         if (limits.no_turbo)
 372                 max_perf = cpu->pstate.max_pstate;
 373
 374         max_perf_adj = fp_toint(mul_fp(int_tofp(max_perf), limits.max_perf));
 375         *max = clamp_t(int, max_perf_adj,
 376                         cpu->pstate.min_pstate, cpu->pstate.turbo_pstate);
 377
 378         min_perf = fp_toint(mul_fp(int_tofp(max_perf), limits.min_perf));
 379         *min = clamp_t(int, min_perf,
 380                         cpu->pstate.min_pstate, max_perf);
 381 }
 382
 383 static void intel_pstate_set_pstate(struct cpudata *cpu, int pstate)
 384 {
 385         int max_perf, min_perf;
 386
 387         intel_pstate_get_min_max(cpu, &min_perf, &max_perf);
 388
 389         pstate = clamp_t(int, pstate, min_perf, max_perf);
 390
 391         if (pstate == cpu->pstate.current_pstate)
 392                 return;
 393
 394         trace_cpu_frequency(pstate * 100000, cpu->cpu);
 395
 396         cpu->pstate.current_pstate = pstate;
 397         if (limits.no_turbo)
 398                 wrmsrl(MSR_IA32_PERF_CTL, BIT(32) | (pstate << 8));
 399         else
 400                 wrmsrl(MSR_IA32_PERF_CTL, pstate << 8);
 401
 402 }
 403
 404 static inline void intel_pstate_pstate_increase(struct cpudata *cpu, int steps)
 405 {
 406         int target;
 407         target = cpu->pstate.current_pstate + steps;
 408
 409         intel_pstate_set_pstate(cpu, target);
 410 }
 411
 412 static inline void intel_pstate_pstate_decrease(struct cpudata *cpu, int steps)
 413 {
 414         int target;
 415         target = cpu->pstate.current_pstate - steps;
 416         intel_pstate_set_pstate(cpu, target);
 417 }
 418
 419 static void intel_pstate_get_cpu_pstates(struct cpudata *cpu)
 420 {
 421         sprintf(cpu->name, "Intel 2nd generation core");
 422
 423         cpu->pstate.min_pstate = intel_pstate_min_pstate();
 424         cpu->pstate.max_pstate = intel_pstate_max_pstate();
 425         cpu->pstate.turbo_pstate = intel_pstate_turbo_pstate();
 426
 427         /*
 428          * goto max pstate so we don't slow up boot if we are built-in if we are
 429          * a module we will take care of it during normal operation
 430          */
 431         intel_pstate_set_pstate(cpu, cpu->pstate.max_pstate);
 432 }
 433
 434 static inline void intel_pstate_calc_busy(struct cpudata *cpu,
 435                                         struct sample *sample)
 436 {
 437         u64 core_pct;
 438         core_pct = div64_u64(int_tofp(sample->aperf * 100),
 439                              sample->mperf);
 440         sample->freq = fp_toint(cpu->pstate.max_pstate * core_pct * 1000);
 441
 442         sample->core_pct_busy = core_pct;
 443 }
 444
 445 static inline void intel_pstate_sample(struct cpudata *cpu)
 446 {
 447         u64 aperf, mperf;
 448
 449         rdmsrl(MSR_IA32_APERF, aperf);
 450         rdmsrl(MSR_IA32_MPERF, mperf);
 451         cpu->sample_ptr = (cpu->sample_ptr + 1) % SAMPLE_COUNT;
 452         cpu->samples[cpu->sample_ptr].aperf = aperf;
 453         cpu->samples[cpu->sample_ptr].mperf = mperf;
 454         cpu->samples[cpu->sample_ptr].aperf -= cpu->prev_aperf;
 455         cpu->samples[cpu->sample_ptr].mperf -= cpu->prev_mperf;
 456
 457         intel_pstate_calc_busy(cpu, &cpu->samples[cpu->sample_ptr]);
 458
 459         cpu->prev_aperf = aperf;
 460         cpu->prev_mperf = mperf;
 461 }
 462
 463 static inline void intel_pstate_set_sample_time(struct cpudata *cpu)
 464 {
 465         int sample_time, delay;
 466
 467         sample_time = cpu->pstate_policy->sample_rate_ms;
 468         delay = msecs_to_jiffies(sample_time);
 469         mod_timer_pinned(&cpu->timer, jiffies + delay);
 470 }
 471
 472 static inline int32_t intel_pstate_get_scaled_busy(struct cpudata *cpu)
 473 {
 474         int32_t core_busy, max_pstate, current_pstate;
 475
 476         core_busy = cpu->samples[cpu->sample_ptr].core_pct_busy;
 477         max_pstate = int_tofp(cpu->pstate.max_pstate);
 478         current_pstate = int_tofp(cpu->pstate.current_pstate);
 479         return mul_fp(core_busy, div_fp(max_pstate, current_pstate));
 480 }
 481
 482 static inline void intel_pstate_adjust_busy_pstate(struct cpudata *cpu)
 483 {
 484         int32_t busy_scaled;
 485         struct _pid *pid;
 486         signed int ctl = 0;
 487         int steps;
 488
 489         pid = &cpu->pid;
 490         busy_scaled = intel_pstate_get_scaled_busy(cpu);
 491
 492         ctl = pid_calc(pid, busy_scaled);
 493
 494         steps = abs(ctl);
 495         if (ctl < 0)
 496                 intel_pstate_pstate_increase(cpu, steps);
 497         else
 498                 intel_pstate_pstate_decrease(cpu, steps);
 499 }
 500
 501 static void intel_pstate_timer_func(unsigned long __data)
 502 {
 503         struct cpudata *cpu = (struct cpudata *) __data;
 504
 505         intel_pstate_sample(cpu);
 506         intel_pstate_adjust_busy_pstate(cpu);
 507
 508         if (cpu->pstate.current_pstate == cpu->pstate.min_pstate) {
 509                 cpu->min_pstate_count++;
 510                 if (!(cpu->min_pstate_count % 5)) {
 511                         intel_pstate_set_pstate(cpu, cpu->pstate.max_pstate);
 512                 }
 513         } else
 514                 cpu->min_pstate_count = 0;
 515
 516         intel_pstate_set_sample_time(cpu);
 517 }
 518
 519 #define ICPU(model, policy) \
 520         { X86_VENDOR_INTEL, 6, model, X86_FEATURE_APERFMPERF,\
 521                         (unsigned long)&policy }
 522
 523 static const struct x86_cpu_id intel_pstate_cpu_ids[] = {
 524         ICPU(0x2a, default_policy),
 525         ICPU(0x2d, default_policy),
 526         ICPU(0x3a, default_policy),
 527         ICPU(0x3c, default_policy),
 528         ICPU(0x3e, default_policy),
 529         ICPU(0x3f, default_policy),
 530         ICPU(0x45, default_policy),
 531         ICPU(0x46, default_policy),
 532         {}
 533 };
 534 MODULE_DEVICE_TABLE(x86cpu, intel_pstate_cpu_ids);
 535
 536 static int intel_pstate_init_cpu(unsigned int cpunum)
 537 {
 538
 539         const struct x86_cpu_id *id;
 540         struct cpudata *cpu;
 541
 542         id = x86_match_cpu(intel_pstate_cpu_ids);
 543         if (!id)
 544                 return -ENODEV;
 545
 546         all_cpu_data[cpunum] = kzalloc(sizeof(struct cpudata), GFP_KERNEL);
 547         if (!all_cpu_data[cpunum])
 548                 return -ENOMEM;
 549
 550         cpu = all_cpu_data[cpunum];
 551
 552         intel_pstate_get_cpu_pstates(cpu);
 553         if (!cpu->pstate.current_pstate) {
 554                 all_cpu_data[cpunum] = NULL;
 555                 kfree(cpu);
 556                 return -ENODATA;
 557         }
 558
 559         cpu->cpu = cpunum;
 560         cpu->pstate_policy =
 561                 (struct pstate_adjust_policy *)id->driver_data;
 562         init_timer_deferrable(&cpu->timer);
 563         cpu->timer.function = intel_pstate_timer_func;
 564         cpu->timer.data =
 565                 (unsigned long)cpu;
 566         cpu->timer.expires = jiffies + HZ/100;
 567         intel_pstate_busy_pid_reset(cpu);
 568         intel_pstate_sample(cpu);
 569         intel_pstate_set_pstate(cpu, cpu->pstate.max_pstate);
 570
 571         add_timer_on(&cpu->timer, cpunum);
 572
 573         pr_info("Intel pstate controlling: cpu %d\n", cpunum);
 574
 575         return 0;
 576 }
 577
 578 static unsigned int intel_pstate_get(unsigned int cpu_num)
 579 {
 580         struct sample *sample;
 581         struct cpudata *cpu;
 582
 583         cpu = all_cpu_data[cpu_num];
 584         if (!cpu)
 585                 return 0;
 586         sample = &cpu->samples[cpu->sample_ptr];
 587         return sample->freq;
 588 }
 589
 590 static int intel_pstate_set_policy(struct cpufreq_policy *policy)
 591 {
 592         struct cpudata *cpu;
 593
 594         cpu = all_cpu_data[policy->cpu];
 595
 596         if (!policy->cpuinfo.max_freq)
 597                 return -ENODEV;
 598
 599         if (policy->policy == CPUFREQ_POLICY_PERFORMANCE) {
 600                 limits.min_perf_pct = 100;
 601                 limits.min_perf = int_tofp(1);
 602                 limits.max_policy_pct = 100;
 603                 limits.max_perf_pct = 100;
 604                 limits.max_perf = int_tofp(1);
 605                 limits.no_turbo = 0;
 606                 return 0;
 607         }
 608         limits.min_perf_pct = (policy->min * 100) / policy->cpuinfo.max_freq;
 609         limits.min_perf_pct = clamp_t(int, limits.min_perf_pct, 0 , 100);
 610         limits.min_perf = div_fp(int_tofp(limits.min_perf_pct), int_tofp(100));
 611
 612         limits.max_policy_pct = policy->max * 100 / policy->cpuinfo.max_freq;
 613         limits.max_policy_pct = clamp_t(int, limits.max_policy_pct, 0 , 100);
 614         limits.max_perf_pct = min(limits.max_policy_pct, limits.max_sysfs_pct);
 615         limits.max_perf = div_fp(int_tofp(limits.max_perf_pct), int_tofp(100));
 616
 617         return 0;
 618 }
 619
 620 static int intel_pstate_verify_policy(struct cpufreq_policy *policy)
 621 {
 622         cpufreq_verify_within_limits(policy,
 623                                 policy->cpuinfo.min_freq,
 624                                 policy->cpuinfo.max_freq);
 625
 626         if ((policy->policy != CPUFREQ_POLICY_POWERSAVE) &&
 627                 (policy->policy != CPUFREQ_POLICY_PERFORMANCE))
 628                 return -EINVAL;
 629
 630         return 0;
 631 }
 632
 633 static int __cpuinit intel_pstate_cpu_exit(struct cpufreq_policy *policy)
 634 {
 635         int cpu = policy->cpu;
 636
 637         del_timer(&all_cpu_data[cpu]->timer);
 638         kfree(all_cpu_data[cpu]);
 639         all_cpu_data[cpu] = NULL;
 640         return 0;
 641 }
 642
 643 static int __cpuinit intel_pstate_cpu_init(struct cpufreq_policy *policy)
 644 {
 645         struct cpudata *cpu;
 646         int rc;
 647
 648         rc = intel_pstate_init_cpu(policy->cpu);
 649         if (rc)
 650                 return rc;
 651
 652         cpu = all_cpu_data[policy->cpu];
 653
 654         if (!limits.no_turbo &&
 655                 limits.min_perf_pct == 100 && limits.max_perf_pct == 100)
 656                 policy->policy = CPUFREQ_POLICY_PERFORMANCE;
 657         else
 658                 policy->policy = CPUFREQ_POLICY_POWERSAVE;
 659
 660         policy->min = cpu->pstate.min_pstate * 100000;
 661         policy->max = cpu->pstate.turbo_pstate * 100000;
 662
 663         /* cpuinfo and default policy values */
 664         policy->cpuinfo.min_freq = cpu->pstate.min_pstate * 100000;
 665         policy->cpuinfo.max_freq = cpu->pstate.turbo_pstate * 100000;
 666         policy->cpuinfo.transition_latency = CPUFREQ_ETERNAL;
 667         cpumask_set_cpu(policy->cpu, policy->cpus);
 668
 669         return 0;
 670 }
 671
 672 static struct cpufreq_driver intel_pstate_driver = {
 673         .flags          = CPUFREQ_CONST_LOOPS,
 674         .verify         = intel_pstate_verify_policy,
 675         .setpolicy      = intel_pstate_set_policy,
 676         .get            = intel_pstate_get,
 677         .init           = intel_pstate_cpu_init,
 678         .exit           = intel_pstate_cpu_exit,
 679         .name           = "intel_pstate",
 680         .owner          = THIS_MODULE,
 681 };
 682
 683 static int __initdata no_load;
 684
 685 static int intel_pstate_msrs_not_valid(void)
 686 {
 687         /* Check that all the msr's we are using are valid. */
 688         u64 aperf, mperf, tmp;
 689
 690         rdmsrl(MSR_IA32_APERF, aperf);
 691         rdmsrl(MSR_IA32_MPERF, mperf);
 692
 693         if (!intel_pstate_min_pstate() ||
 694                 !intel_pstate_max_pstate() ||
 695                 !intel_pstate_turbo_pstate())
 696                 return -ENODEV;
 697
 698         rdmsrl(MSR_IA32_APERF, tmp);
 699         if (!(tmp - aperf))
 700                 return -ENODEV;
 701
 702         rdmsrl(MSR_IA32_MPERF, tmp);
 703         if (!(tmp - mperf))
 704                 return -ENODEV;
 705
 706         return 0;
 707 }
 708 static int __init intel_pstate_init(void)
 709 {
 710         int cpu, rc = 0;
 711         const struct x86_cpu_id *id;
 712
 713         if (no_load)
 714                 return -ENODEV;
 715
 716         id = x86_match_cpu(intel_pstate_cpu_ids);
 717         if (!id)
 718                 return -ENODEV;
 719
 720         if (intel_pstate_msrs_not_valid())
 721                 return -ENODEV;
 722
 723         pr_info("Intel P-state driver initializing.\n");
 724
 725         all_cpu_data = vzalloc(sizeof(void *) * num_possible_cpus());
 726         if (!all_cpu_data)
 727                 return -ENOMEM;
 728
 729         rc = cpufreq_register_driver(&intel_pstate_driver);
 730         if (rc)
 731                 goto out;
 732
 733         intel_pstate_debug_expose_params();
 734         intel_pstate_sysfs_expose_params();
 735         return rc;
 736 out:
 737         get_online_cpus();
 738         for_each_online_cpu(cpu) {
 739                 if (all_cpu_data[cpu]) {
 740                         del_timer_sync(&all_cpu_data[cpu]->timer);
 741                         kfree(all_cpu_data[cpu]);
 742                 }
 743         }
 744
 745         put_online_cpus();
 746         vfree(all_cpu_data);
 747         return -ENODEV;
 748 }
 749 device_initcall(intel_pstate_init);
 750
 751 static int __init intel_pstate_setup(char *str)
 752 {
 753         if (!str)
 754                 return -EINVAL;
 755
 756         if (!strcmp(str, "disable"))
 757                 no_load = 1;
 758         return 0;
 759 }
 760 early_param("intel_pstate", intel_pstate_setup);
 761
 762 MODULE_AUTHOR("Dirk Brandewie <dirk.j.brandewie@intel.com>");
 763 MODULE_DESCRIPTION("'intel_pstate' - P state driver Intel Core processors");
 764 MODULE_LICENSE("GPL");