intel_pstate.c revision d40a63c45b506b0681918d7c62a15cc9d48c8681
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 SAMPLE_COUNT 3 36 37#define BYT_RATIOS 0x66a 38#define BYT_VIDS 0x66b 39#define BYT_TURBO_RATIOS 0x66c 40#define BYT_TURBO_VIDS 0x66d 41 42 43#define FRAC_BITS 6 44#define int_tofp(X) ((int64_t)(X) << FRAC_BITS) 45#define fp_toint(X) ((X) >> FRAC_BITS) 46#define FP_ROUNDUP(X) ((X) += 1 << FRAC_BITS) 47 48static inline int32_t mul_fp(int32_t x, int32_t y) 49{ 50 return ((int64_t)x * (int64_t)y) >> FRAC_BITS; 51} 52 53static inline int32_t div_fp(int32_t x, int32_t y) 54{ 55 return div_s64((int64_t)x << FRAC_BITS, (int64_t)y); 56} 57 58struct sample { 59 int32_t core_pct_busy; 60 u64 aperf; 61 u64 mperf; 62 unsigned long long tsc; 63 int freq; 64}; 65 66struct pstate_data { 67 int current_pstate; 68 int min_pstate; 69 int max_pstate; 70 int turbo_pstate; 71}; 72 73struct vid_data { 74 int min; 75 int max; 76 int turbo; 77 int32_t ratio; 78}; 79 80struct _pid { 81 int setpoint; 82 int32_t integral; 83 int32_t p_gain; 84 int32_t i_gain; 85 int32_t d_gain; 86 int deadband; 87 int32_t last_err; 88}; 89 90struct cpudata { 91 int cpu; 92 93 char name[64]; 94 95 struct timer_list timer; 96 97 struct pstate_data pstate; 98 struct vid_data vid; 99 struct _pid pid; 100 101 u64 prev_aperf; 102 u64 prev_mperf; 103 unsigned long long prev_tsc; 104 struct sample sample; 105}; 106 107static struct cpudata **all_cpu_data; 108struct pstate_adjust_policy { 109 int sample_rate_ms; 110 int deadband; 111 int setpoint; 112 int p_gain_pct; 113 int d_gain_pct; 114 int i_gain_pct; 115}; 116 117struct pstate_funcs { 118 int (*get_max)(void); 119 int (*get_min)(void); 120 int (*get_turbo)(void); 121 void (*set)(struct cpudata*, int pstate); 122 void (*get_vid)(struct cpudata *); 123}; 124 125struct cpu_defaults { 126 struct pstate_adjust_policy pid_policy; 127 struct pstate_funcs funcs; 128}; 129 130static struct pstate_adjust_policy pid_params; 131static struct pstate_funcs pstate_funcs; 132 133struct perf_limits { 134 int no_turbo; 135 int max_perf_pct; 136 int min_perf_pct; 137 int32_t max_perf; 138 int32_t min_perf; 139 int max_policy_pct; 140 int max_sysfs_pct; 141}; 142 143static struct perf_limits limits = { 144 .no_turbo = 0, 145 .max_perf_pct = 100, 146 .max_perf = int_tofp(1), 147 .min_perf_pct = 0, 148 .min_perf = 0, 149 .max_policy_pct = 100, 150 .max_sysfs_pct = 100, 151}; 152 153static inline void pid_reset(struct _pid *pid, int setpoint, int busy, 154 int deadband, int integral) { 155 pid->setpoint = setpoint; 156 pid->deadband = deadband; 157 pid->integral = int_tofp(integral); 158 pid->last_err = int_tofp(setpoint) - int_tofp(busy); 159} 160 161static inline void pid_p_gain_set(struct _pid *pid, int percent) 162{ 163 pid->p_gain = div_fp(int_tofp(percent), int_tofp(100)); 164} 165 166static inline void pid_i_gain_set(struct _pid *pid, int percent) 167{ 168 pid->i_gain = div_fp(int_tofp(percent), int_tofp(100)); 169} 170 171static inline void pid_d_gain_set(struct _pid *pid, int percent) 172{ 173 174 pid->d_gain = div_fp(int_tofp(percent), int_tofp(100)); 175} 176 177static signed int pid_calc(struct _pid *pid, int32_t busy) 178{ 179 signed int result; 180 int32_t pterm, dterm, fp_error; 181 int32_t integral_limit; 182 183 fp_error = int_tofp(pid->setpoint) - busy; 184 185 if (abs(fp_error) <= int_tofp(pid->deadband)) 186 return 0; 187 188 pterm = mul_fp(pid->p_gain, fp_error); 189 190 pid->integral += fp_error; 191 192 /* limit the integral term */ 193 integral_limit = int_tofp(30); 194 if (pid->integral > integral_limit) 195 pid->integral = integral_limit; 196 if (pid->integral < -integral_limit) 197 pid->integral = -integral_limit; 198 199 dterm = mul_fp(pid->d_gain, fp_error - pid->last_err); 200 pid->last_err = fp_error; 201 202 result = pterm + mul_fp(pid->integral, pid->i_gain) + dterm; 203 204 return (signed int)fp_toint(result); 205} 206 207static inline void intel_pstate_busy_pid_reset(struct cpudata *cpu) 208{ 209 pid_p_gain_set(&cpu->pid, pid_params.p_gain_pct); 210 pid_d_gain_set(&cpu->pid, pid_params.d_gain_pct); 211 pid_i_gain_set(&cpu->pid, pid_params.i_gain_pct); 212 213 pid_reset(&cpu->pid, 214 pid_params.setpoint, 215 100, 216 pid_params.deadband, 217 0); 218} 219 220static inline void intel_pstate_reset_all_pid(void) 221{ 222 unsigned int cpu; 223 for_each_online_cpu(cpu) { 224 if (all_cpu_data[cpu]) 225 intel_pstate_busy_pid_reset(all_cpu_data[cpu]); 226 } 227} 228 229/************************** debugfs begin ************************/ 230static int pid_param_set(void *data, u64 val) 231{ 232 *(u32 *)data = val; 233 intel_pstate_reset_all_pid(); 234 return 0; 235} 236static int pid_param_get(void *data, u64 *val) 237{ 238 *val = *(u32 *)data; 239 return 0; 240} 241DEFINE_SIMPLE_ATTRIBUTE(fops_pid_param, pid_param_get, 242 pid_param_set, "%llu\n"); 243 244struct pid_param { 245 char *name; 246 void *value; 247}; 248 249static struct pid_param pid_files[] = { 250 {"sample_rate_ms", &pid_params.sample_rate_ms}, 251 {"d_gain_pct", &pid_params.d_gain_pct}, 252 {"i_gain_pct", &pid_params.i_gain_pct}, 253 {"deadband", &pid_params.deadband}, 254 {"setpoint", &pid_params.setpoint}, 255 {"p_gain_pct", &pid_params.p_gain_pct}, 256 {NULL, NULL} 257}; 258 259static struct dentry *debugfs_parent; 260static void intel_pstate_debug_expose_params(void) 261{ 262 int i = 0; 263 264 debugfs_parent = debugfs_create_dir("pstate_snb", NULL); 265 if (IS_ERR_OR_NULL(debugfs_parent)) 266 return; 267 while (pid_files[i].name) { 268 debugfs_create_file(pid_files[i].name, 0660, 269 debugfs_parent, pid_files[i].value, 270 &fops_pid_param); 271 i++; 272 } 273} 274 275/************************** debugfs end ************************/ 276 277/************************** sysfs begin ************************/ 278#define show_one(file_name, object) \ 279 static ssize_t show_##file_name \ 280 (struct kobject *kobj, struct attribute *attr, char *buf) \ 281 { \ 282 return sprintf(buf, "%u\n", limits.object); \ 283 } 284 285static ssize_t store_no_turbo(struct kobject *a, struct attribute *b, 286 const char *buf, size_t count) 287{ 288 unsigned int input; 289 int ret; 290 ret = sscanf(buf, "%u", &input); 291 if (ret != 1) 292 return -EINVAL; 293 limits.no_turbo = clamp_t(int, input, 0 , 1); 294 295 return count; 296} 297 298static ssize_t store_max_perf_pct(struct kobject *a, struct attribute *b, 299 const char *buf, size_t count) 300{ 301 unsigned int input; 302 int ret; 303 ret = sscanf(buf, "%u", &input); 304 if (ret != 1) 305 return -EINVAL; 306 307 limits.max_sysfs_pct = clamp_t(int, input, 0 , 100); 308 limits.max_perf_pct = min(limits.max_policy_pct, limits.max_sysfs_pct); 309 limits.max_perf = div_fp(int_tofp(limits.max_perf_pct), int_tofp(100)); 310 return count; 311} 312 313static ssize_t store_min_perf_pct(struct kobject *a, struct attribute *b, 314 const char *buf, size_t count) 315{ 316 unsigned int input; 317 int ret; 318 ret = sscanf(buf, "%u", &input); 319 if (ret != 1) 320 return -EINVAL; 321 limits.min_perf_pct = clamp_t(int, input, 0 , 100); 322 limits.min_perf = div_fp(int_tofp(limits.min_perf_pct), int_tofp(100)); 323 324 return count; 325} 326 327show_one(no_turbo, no_turbo); 328show_one(max_perf_pct, max_perf_pct); 329show_one(min_perf_pct, min_perf_pct); 330 331define_one_global_rw(no_turbo); 332define_one_global_rw(max_perf_pct); 333define_one_global_rw(min_perf_pct); 334 335static struct attribute *intel_pstate_attributes[] = { 336 &no_turbo.attr, 337 &max_perf_pct.attr, 338 &min_perf_pct.attr, 339 NULL 340}; 341 342static struct attribute_group intel_pstate_attr_group = { 343 .attrs = intel_pstate_attributes, 344}; 345static struct kobject *intel_pstate_kobject; 346 347static void intel_pstate_sysfs_expose_params(void) 348{ 349 int rc; 350 351 intel_pstate_kobject = kobject_create_and_add("intel_pstate", 352 &cpu_subsys.dev_root->kobj); 353 BUG_ON(!intel_pstate_kobject); 354 rc = sysfs_create_group(intel_pstate_kobject, 355 &intel_pstate_attr_group); 356 BUG_ON(rc); 357} 358 359/************************** sysfs end ************************/ 360static int byt_get_min_pstate(void) 361{ 362 u64 value; 363 rdmsrl(BYT_RATIOS, value); 364 return (value >> 8) & 0x3F; 365} 366 367static int byt_get_max_pstate(void) 368{ 369 u64 value; 370 rdmsrl(BYT_RATIOS, value); 371 return (value >> 16) & 0x3F; 372} 373 374static int byt_get_turbo_pstate(void) 375{ 376 u64 value; 377 rdmsrl(BYT_TURBO_RATIOS, value); 378 return value & 0x3F; 379} 380 381static void byt_set_pstate(struct cpudata *cpudata, int pstate) 382{ 383 u64 val; 384 int32_t vid_fp; 385 u32 vid; 386 387 val = pstate << 8; 388 if (limits.no_turbo) 389 val |= (u64)1 << 32; 390 391 vid_fp = cpudata->vid.min + mul_fp( 392 int_tofp(pstate - cpudata->pstate.min_pstate), 393 cpudata->vid.ratio); 394 395 vid_fp = clamp_t(int32_t, vid_fp, cpudata->vid.min, cpudata->vid.max); 396 vid = fp_toint(vid_fp); 397 398 if (pstate > cpudata->pstate.max_pstate) 399 vid = cpudata->vid.turbo; 400 401 val |= vid; 402 403 wrmsrl(MSR_IA32_PERF_CTL, val); 404} 405 406static void byt_get_vid(struct cpudata *cpudata) 407{ 408 u64 value; 409 410 411 rdmsrl(BYT_VIDS, value); 412 cpudata->vid.min = int_tofp((value >> 8) & 0x3f); 413 cpudata->vid.max = int_tofp((value >> 16) & 0x3f); 414 cpudata->vid.ratio = div_fp( 415 cpudata->vid.max - cpudata->vid.min, 416 int_tofp(cpudata->pstate.max_pstate - 417 cpudata->pstate.min_pstate)); 418 419 rdmsrl(BYT_TURBO_VIDS, value); 420 cpudata->vid.turbo = value & 0x7f; 421} 422 423 424static int core_get_min_pstate(void) 425{ 426 u64 value; 427 rdmsrl(MSR_PLATFORM_INFO, value); 428 return (value >> 40) & 0xFF; 429} 430 431static int core_get_max_pstate(void) 432{ 433 u64 value; 434 rdmsrl(MSR_PLATFORM_INFO, value); 435 return (value >> 8) & 0xFF; 436} 437 438static int core_get_turbo_pstate(void) 439{ 440 u64 value; 441 int nont, ret; 442 rdmsrl(MSR_NHM_TURBO_RATIO_LIMIT, value); 443 nont = core_get_max_pstate(); 444 ret = ((value) & 255); 445 if (ret <= nont) 446 ret = nont; 447 return ret; 448} 449 450static void core_set_pstate(struct cpudata *cpudata, int pstate) 451{ 452 u64 val; 453 454 val = pstate << 8; 455 if (limits.no_turbo) 456 val |= (u64)1 << 32; 457 458 wrmsrl_on_cpu(cpudata->cpu, MSR_IA32_PERF_CTL, val); 459} 460 461static struct cpu_defaults core_params = { 462 .pid_policy = { 463 .sample_rate_ms = 10, 464 .deadband = 0, 465 .setpoint = 97, 466 .p_gain_pct = 20, 467 .d_gain_pct = 0, 468 .i_gain_pct = 0, 469 }, 470 .funcs = { 471 .get_max = core_get_max_pstate, 472 .get_min = core_get_min_pstate, 473 .get_turbo = core_get_turbo_pstate, 474 .set = core_set_pstate, 475 }, 476}; 477 478static struct cpu_defaults byt_params = { 479 .pid_policy = { 480 .sample_rate_ms = 10, 481 .deadband = 0, 482 .setpoint = 97, 483 .p_gain_pct = 14, 484 .d_gain_pct = 0, 485 .i_gain_pct = 4, 486 }, 487 .funcs = { 488 .get_max = byt_get_max_pstate, 489 .get_min = byt_get_min_pstate, 490 .get_turbo = byt_get_turbo_pstate, 491 .set = byt_set_pstate, 492 .get_vid = byt_get_vid, 493 }, 494}; 495 496 497static void intel_pstate_get_min_max(struct cpudata *cpu, int *min, int *max) 498{ 499 int max_perf = cpu->pstate.turbo_pstate; 500 int max_perf_adj; 501 int min_perf; 502 if (limits.no_turbo) 503 max_perf = cpu->pstate.max_pstate; 504 505 max_perf_adj = fp_toint(mul_fp(int_tofp(max_perf), limits.max_perf)); 506 *max = clamp_t(int, max_perf_adj, 507 cpu->pstate.min_pstate, cpu->pstate.turbo_pstate); 508 509 min_perf = fp_toint(mul_fp(int_tofp(max_perf), limits.min_perf)); 510 *min = clamp_t(int, min_perf, 511 cpu->pstate.min_pstate, max_perf); 512} 513 514static void intel_pstate_set_pstate(struct cpudata *cpu, int pstate) 515{ 516 int max_perf, min_perf; 517 518 intel_pstate_get_min_max(cpu, &min_perf, &max_perf); 519 520 pstate = clamp_t(int, pstate, min_perf, max_perf); 521 522 if (pstate == cpu->pstate.current_pstate) 523 return; 524 525 trace_cpu_frequency(pstate * 100000, cpu->cpu); 526 527 cpu->pstate.current_pstate = pstate; 528 529 pstate_funcs.set(cpu, pstate); 530} 531 532static inline void intel_pstate_pstate_increase(struct cpudata *cpu, int steps) 533{ 534 int target; 535 target = cpu->pstate.current_pstate + steps; 536 537 intel_pstate_set_pstate(cpu, target); 538} 539 540static inline void intel_pstate_pstate_decrease(struct cpudata *cpu, int steps) 541{ 542 int target; 543 target = cpu->pstate.current_pstate - steps; 544 intel_pstate_set_pstate(cpu, target); 545} 546 547static void intel_pstate_get_cpu_pstates(struct cpudata *cpu) 548{ 549 sprintf(cpu->name, "Intel 2nd generation core"); 550 551 cpu->pstate.min_pstate = pstate_funcs.get_min(); 552 cpu->pstate.max_pstate = pstate_funcs.get_max(); 553 cpu->pstate.turbo_pstate = pstate_funcs.get_turbo(); 554 555 if (pstate_funcs.get_vid) 556 pstate_funcs.get_vid(cpu); 557 intel_pstate_set_pstate(cpu, cpu->pstate.min_pstate); 558} 559 560static inline void intel_pstate_calc_busy(struct cpudata *cpu, 561 struct sample *sample) 562{ 563 int32_t core_pct; 564 int32_t c0_pct; 565 566 core_pct = div_fp(int_tofp((sample->aperf)), 567 int_tofp((sample->mperf))); 568 core_pct = mul_fp(core_pct, int_tofp(100)); 569 FP_ROUNDUP(core_pct); 570 571 c0_pct = div_fp(int_tofp(sample->mperf), int_tofp(sample->tsc)); 572 573 sample->freq = fp_toint( 574 mul_fp(int_tofp(cpu->pstate.max_pstate * 1000), core_pct)); 575 576 sample->core_pct_busy = mul_fp(core_pct, c0_pct); 577} 578 579static inline void intel_pstate_sample(struct cpudata *cpu) 580{ 581 u64 aperf, mperf; 582 unsigned long long tsc; 583 584 rdmsrl(MSR_IA32_APERF, aperf); 585 rdmsrl(MSR_IA32_MPERF, mperf); 586 tsc = native_read_tsc(); 587 588 aperf = aperf >> FRAC_BITS; 589 mperf = mperf >> FRAC_BITS; 590 tsc = tsc >> FRAC_BITS; 591 592 cpu->sample.aperf = aperf; 593 cpu->sample.mperf = mperf; 594 cpu->sample.tsc = tsc; 595 cpu->sample.aperf -= cpu->prev_aperf; 596 cpu->sample.mperf -= cpu->prev_mperf; 597 cpu->sample.tsc -= cpu->prev_tsc; 598 599 intel_pstate_calc_busy(cpu, &cpu->sample); 600 601 cpu->prev_aperf = aperf; 602 cpu->prev_mperf = mperf; 603 cpu->prev_tsc = tsc; 604} 605 606static inline void intel_pstate_set_sample_time(struct cpudata *cpu) 607{ 608 int sample_time, delay; 609 610 sample_time = pid_params.sample_rate_ms; 611 delay = msecs_to_jiffies(sample_time); 612 mod_timer_pinned(&cpu->timer, jiffies + delay); 613} 614 615static inline int32_t intel_pstate_get_scaled_busy(struct cpudata *cpu) 616{ 617 int32_t core_busy, max_pstate, current_pstate; 618 619 core_busy = cpu->sample.core_pct_busy; 620 max_pstate = int_tofp(cpu->pstate.max_pstate); 621 current_pstate = int_tofp(cpu->pstate.current_pstate); 622 core_busy = mul_fp(core_busy, div_fp(max_pstate, current_pstate)); 623 return FP_ROUNDUP(core_busy); 624} 625 626static inline void intel_pstate_adjust_busy_pstate(struct cpudata *cpu) 627{ 628 int32_t busy_scaled; 629 struct _pid *pid; 630 signed int ctl = 0; 631 int steps; 632 633 pid = &cpu->pid; 634 busy_scaled = intel_pstate_get_scaled_busy(cpu); 635 636 ctl = pid_calc(pid, busy_scaled); 637 638 steps = abs(ctl); 639 640 if (ctl < 0) 641 intel_pstate_pstate_increase(cpu, steps); 642 else 643 intel_pstate_pstate_decrease(cpu, steps); 644} 645 646static void intel_pstate_timer_func(unsigned long __data) 647{ 648 struct cpudata *cpu = (struct cpudata *) __data; 649 struct sample *sample; 650 651 intel_pstate_sample(cpu); 652 653 sample = &cpu->sample; 654 655 intel_pstate_adjust_busy_pstate(cpu); 656 657 trace_pstate_sample(fp_toint(sample->core_pct_busy), 658 fp_toint(intel_pstate_get_scaled_busy(cpu)), 659 cpu->pstate.current_pstate, 660 sample->mperf, 661 sample->aperf, 662 sample->freq); 663 664 intel_pstate_set_sample_time(cpu); 665} 666 667#define ICPU(model, policy) \ 668 { X86_VENDOR_INTEL, 6, model, X86_FEATURE_APERFMPERF,\ 669 (unsigned long)&policy } 670 671static const struct x86_cpu_id intel_pstate_cpu_ids[] = { 672 ICPU(0x2a, core_params), 673 ICPU(0x2d, core_params), 674 ICPU(0x37, byt_params), 675 ICPU(0x3a, core_params), 676 ICPU(0x3c, core_params), 677 ICPU(0x3e, core_params), 678 ICPU(0x3f, core_params), 679 ICPU(0x45, core_params), 680 ICPU(0x46, core_params), 681 {} 682}; 683MODULE_DEVICE_TABLE(x86cpu, intel_pstate_cpu_ids); 684 685static int intel_pstate_init_cpu(unsigned int cpunum) 686{ 687 688 const struct x86_cpu_id *id; 689 struct cpudata *cpu; 690 691 id = x86_match_cpu(intel_pstate_cpu_ids); 692 if (!id) 693 return -ENODEV; 694 695 all_cpu_data[cpunum] = kzalloc(sizeof(struct cpudata), GFP_KERNEL); 696 if (!all_cpu_data[cpunum]) 697 return -ENOMEM; 698 699 cpu = all_cpu_data[cpunum]; 700 701 intel_pstate_get_cpu_pstates(cpu); 702 703 cpu->cpu = cpunum; 704 705 init_timer_deferrable(&cpu->timer); 706 cpu->timer.function = intel_pstate_timer_func; 707 cpu->timer.data = 708 (unsigned long)cpu; 709 cpu->timer.expires = jiffies + HZ/100; 710 intel_pstate_busy_pid_reset(cpu); 711 intel_pstate_sample(cpu); 712 713 add_timer_on(&cpu->timer, cpunum); 714 715 pr_info("Intel pstate controlling: cpu %d\n", cpunum); 716 717 return 0; 718} 719 720static unsigned int intel_pstate_get(unsigned int cpu_num) 721{ 722 struct sample *sample; 723 struct cpudata *cpu; 724 725 cpu = all_cpu_data[cpu_num]; 726 if (!cpu) 727 return 0; 728 sample = &cpu->sample; 729 return sample->freq; 730} 731 732static int intel_pstate_set_policy(struct cpufreq_policy *policy) 733{ 734 struct cpudata *cpu; 735 736 cpu = all_cpu_data[policy->cpu]; 737 738 if (!policy->cpuinfo.max_freq) 739 return -ENODEV; 740 741 if (policy->policy == CPUFREQ_POLICY_PERFORMANCE) { 742 limits.min_perf_pct = 100; 743 limits.min_perf = int_tofp(1); 744 limits.max_perf_pct = 100; 745 limits.max_perf = int_tofp(1); 746 limits.no_turbo = 0; 747 return 0; 748 } 749 limits.min_perf_pct = (policy->min * 100) / policy->cpuinfo.max_freq; 750 limits.min_perf_pct = clamp_t(int, limits.min_perf_pct, 0 , 100); 751 limits.min_perf = div_fp(int_tofp(limits.min_perf_pct), int_tofp(100)); 752 753 limits.max_policy_pct = policy->max * 100 / policy->cpuinfo.max_freq; 754 limits.max_policy_pct = clamp_t(int, limits.max_policy_pct, 0 , 100); 755 limits.max_perf_pct = min(limits.max_policy_pct, limits.max_sysfs_pct); 756 limits.max_perf = div_fp(int_tofp(limits.max_perf_pct), int_tofp(100)); 757 758 return 0; 759} 760 761static int intel_pstate_verify_policy(struct cpufreq_policy *policy) 762{ 763 cpufreq_verify_within_cpu_limits(policy); 764 765 if ((policy->policy != CPUFREQ_POLICY_POWERSAVE) && 766 (policy->policy != CPUFREQ_POLICY_PERFORMANCE)) 767 return -EINVAL; 768 769 return 0; 770} 771 772static void intel_pstate_stop_cpu(struct cpufreq_policy *policy) 773{ 774 int cpu_num = policy->cpu; 775 struct cpudata *cpu = all_cpu_data[cpu_num]; 776 777 pr_info("intel_pstate CPU %d exiting\n", cpu_num); 778 779 del_timer_sync(&all_cpu_data[cpu_num]->timer); 780 intel_pstate_set_pstate(cpu, cpu->pstate.min_pstate); 781 kfree(all_cpu_data[cpu_num]); 782 all_cpu_data[cpu_num] = NULL; 783} 784 785static int intel_pstate_cpu_init(struct cpufreq_policy *policy) 786{ 787 struct cpudata *cpu; 788 int rc; 789 790 rc = intel_pstate_init_cpu(policy->cpu); 791 if (rc) 792 return rc; 793 794 cpu = all_cpu_data[policy->cpu]; 795 796 if (!limits.no_turbo && 797 limits.min_perf_pct == 100 && limits.max_perf_pct == 100) 798 policy->policy = CPUFREQ_POLICY_PERFORMANCE; 799 else 800 policy->policy = CPUFREQ_POLICY_POWERSAVE; 801 802 policy->min = cpu->pstate.min_pstate * 100000; 803 policy->max = cpu->pstate.turbo_pstate * 100000; 804 805 /* cpuinfo and default policy values */ 806 policy->cpuinfo.min_freq = cpu->pstate.min_pstate * 100000; 807 policy->cpuinfo.max_freq = cpu->pstate.turbo_pstate * 100000; 808 policy->cpuinfo.transition_latency = CPUFREQ_ETERNAL; 809 cpumask_set_cpu(policy->cpu, policy->cpus); 810 811 return 0; 812} 813 814static struct cpufreq_driver intel_pstate_driver = { 815 .flags = CPUFREQ_CONST_LOOPS, 816 .verify = intel_pstate_verify_policy, 817 .setpolicy = intel_pstate_set_policy, 818 .get = intel_pstate_get, 819 .init = intel_pstate_cpu_init, 820 .stop_cpu = intel_pstate_stop_cpu, 821 .name = "intel_pstate", 822}; 823 824static int __initdata no_load; 825 826static int intel_pstate_msrs_not_valid(void) 827{ 828 /* Check that all the msr's we are using are valid. */ 829 u64 aperf, mperf, tmp; 830 831 rdmsrl(MSR_IA32_APERF, aperf); 832 rdmsrl(MSR_IA32_MPERF, mperf); 833 834 if (!pstate_funcs.get_max() || 835 !pstate_funcs.get_min() || 836 !pstate_funcs.get_turbo()) 837 return -ENODEV; 838 839 rdmsrl(MSR_IA32_APERF, tmp); 840 if (!(tmp - aperf)) 841 return -ENODEV; 842 843 rdmsrl(MSR_IA32_MPERF, tmp); 844 if (!(tmp - mperf)) 845 return -ENODEV; 846 847 return 0; 848} 849 850static void copy_pid_params(struct pstate_adjust_policy *policy) 851{ 852 pid_params.sample_rate_ms = policy->sample_rate_ms; 853 pid_params.p_gain_pct = policy->p_gain_pct; 854 pid_params.i_gain_pct = policy->i_gain_pct; 855 pid_params.d_gain_pct = policy->d_gain_pct; 856 pid_params.deadband = policy->deadband; 857 pid_params.setpoint = policy->setpoint; 858} 859 860static void copy_cpu_funcs(struct pstate_funcs *funcs) 861{ 862 pstate_funcs.get_max = funcs->get_max; 863 pstate_funcs.get_min = funcs->get_min; 864 pstate_funcs.get_turbo = funcs->get_turbo; 865 pstate_funcs.set = funcs->set; 866 pstate_funcs.get_vid = funcs->get_vid; 867} 868 869#if IS_ENABLED(CONFIG_ACPI) 870#include <acpi/processor.h> 871 872static bool intel_pstate_no_acpi_pss(void) 873{ 874 int i; 875 876 for_each_possible_cpu(i) { 877 acpi_status status; 878 union acpi_object *pss; 879 struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL }; 880 struct acpi_processor *pr = per_cpu(processors, i); 881 882 if (!pr) 883 continue; 884 885 status = acpi_evaluate_object(pr->handle, "_PSS", NULL, &buffer); 886 if (ACPI_FAILURE(status)) 887 continue; 888 889 pss = buffer.pointer; 890 if (pss && pss->type == ACPI_TYPE_PACKAGE) { 891 kfree(pss); 892 return false; 893 } 894 895 kfree(pss); 896 } 897 898 return true; 899} 900 901struct hw_vendor_info { 902 u16 valid; 903 char oem_id[ACPI_OEM_ID_SIZE]; 904 char oem_table_id[ACPI_OEM_TABLE_ID_SIZE]; 905}; 906 907/* Hardware vendor-specific info that has its own power management modes */ 908static struct hw_vendor_info vendor_info[] = { 909 {1, "HP ", "ProLiant"}, 910 {0, "", ""}, 911}; 912 913static bool intel_pstate_platform_pwr_mgmt_exists(void) 914{ 915 struct acpi_table_header hdr; 916 struct hw_vendor_info *v_info; 917 918 if (acpi_disabled 919 || ACPI_FAILURE(acpi_get_table_header(ACPI_SIG_FADT, 0, &hdr))) 920 return false; 921 922 for (v_info = vendor_info; v_info->valid; v_info++) { 923 if (!strncmp(hdr.oem_id, v_info->oem_id, ACPI_OEM_ID_SIZE) 924 && !strncmp(hdr.oem_table_id, v_info->oem_table_id, ACPI_OEM_TABLE_ID_SIZE) 925 && intel_pstate_no_acpi_pss()) 926 return true; 927 } 928 929 return false; 930} 931#else /* CONFIG_ACPI not enabled */ 932static inline bool intel_pstate_platform_pwr_mgmt_exists(void) { return false; } 933#endif /* CONFIG_ACPI */ 934 935static int __init intel_pstate_init(void) 936{ 937 int cpu, rc = 0; 938 const struct x86_cpu_id *id; 939 struct cpu_defaults *cpu_info; 940 941 if (no_load) 942 return -ENODEV; 943 944 id = x86_match_cpu(intel_pstate_cpu_ids); 945 if (!id) 946 return -ENODEV; 947 948 /* 949 * The Intel pstate driver will be ignored if the platform 950 * firmware has its own power management modes. 951 */ 952 if (intel_pstate_platform_pwr_mgmt_exists()) 953 return -ENODEV; 954 955 cpu_info = (struct cpu_defaults *)id->driver_data; 956 957 copy_pid_params(&cpu_info->pid_policy); 958 copy_cpu_funcs(&cpu_info->funcs); 959 960 if (intel_pstate_msrs_not_valid()) 961 return -ENODEV; 962 963 pr_info("Intel P-state driver initializing.\n"); 964 965 all_cpu_data = vzalloc(sizeof(void *) * num_possible_cpus()); 966 if (!all_cpu_data) 967 return -ENOMEM; 968 969 rc = cpufreq_register_driver(&intel_pstate_driver); 970 if (rc) 971 goto out; 972 973 intel_pstate_debug_expose_params(); 974 intel_pstate_sysfs_expose_params(); 975 976 return rc; 977out: 978 get_online_cpus(); 979 for_each_online_cpu(cpu) { 980 if (all_cpu_data[cpu]) { 981 del_timer_sync(&all_cpu_data[cpu]->timer); 982 kfree(all_cpu_data[cpu]); 983 } 984 } 985 986 put_online_cpus(); 987 vfree(all_cpu_data); 988 return -ENODEV; 989} 990device_initcall(intel_pstate_init); 991 992static int __init intel_pstate_setup(char *str) 993{ 994 if (!str) 995 return -EINVAL; 996 997 if (!strcmp(str, "disable")) 998 no_load = 1; 999 return 0; 1000} 1001early_param("intel_pstate", intel_pstate_setup); 1002 1003MODULE_AUTHOR("Dirk Brandewie <dirk.j.brandewie@intel.com>"); 1004MODULE_DESCRIPTION("'intel_pstate' - P state driver Intel Core processors"); 1005MODULE_LICENSE("GPL"); 1006