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; 140 141 struct perf_limits { 142 int no_turbo; 143 int turbo_disabled; 144 int max_perf_pct; 145 int min_perf_pct; 146 int32_t max_perf; 147 int32_t min_perf; 148 int max_policy_pct; 149 int max_sysfs_pct; 150 }; 151 152 static struct perf_limits limits = { 153 .no_turbo = 0, 154 .turbo_disabled = 0, 155 .max_perf_pct = 100, 156 .max_perf = int_tofp(1), 157 .min_perf_pct = 0, 158 .min_perf = 0, 159 .max_policy_pct = 100, 160 .max_sysfs_pct = 100, 161 }; 162 163 static inline void pid_reset(struct _pid *pid, int setpoint, int busy, 164 int deadband, int integral) { 165 pid->setpoint = setpoint; 166 pid->deadband = deadband; 167 pid->integral = int_tofp(integral); 168 pid->last_err = int_tofp(setpoint) - int_tofp(busy); 169 } 170 171 static inline void pid_p_gain_set(struct _pid *pid, int percent) 172 { 173 pid->p_gain = div_fp(int_tofp(percent), int_tofp(100)); 174 } 175 176 static inline void pid_i_gain_set(struct _pid *pid, int percent) 177 { 178 pid->i_gain = div_fp(int_tofp(percent), int_tofp(100)); 179 } 180 181 static inline void pid_d_gain_set(struct _pid *pid, int percent) 182 { 183 pid->d_gain = div_fp(int_tofp(percent), int_tofp(100)); 184 } 185 186 static signed int pid_calc(struct _pid *pid, int32_t busy) 187 { 188 signed int result; 189 int32_t pterm, dterm, fp_error; 190 int32_t integral_limit; 191 192 fp_error = int_tofp(pid->setpoint) - busy; 193 194 if (abs(fp_error) <= int_tofp(pid->deadband)) 195 return 0; 196 197 pterm = mul_fp(pid->p_gain, fp_error); 198 199 pid->integral += fp_error; 200 201 /* limit the integral term */ 202 integral_limit = int_tofp(30); 203 if (pid->integral > integral_limit) 204 pid->integral = integral_limit; 205 if (pid->integral < -integral_limit) 206 pid->integral = -integral_limit; 207 208 dterm = mul_fp(pid->d_gain, fp_error - pid->last_err); 209 pid->last_err = fp_error; 210 211 result = pterm + mul_fp(pid->integral, pid->i_gain) + dterm; 212 result = result + (1 << (FRAC_BITS-1)); 213 return (signed int)fp_toint(result); 214 } 215 216 static inline void intel_pstate_busy_pid_reset(struct cpudata *cpu) 217 { 218 pid_p_gain_set(&cpu->pid, pid_params.p_gain_pct); 219 pid_d_gain_set(&cpu->pid, pid_params.d_gain_pct); 220 pid_i_gain_set(&cpu->pid, pid_params.i_gain_pct); 221 222 pid_reset(&cpu->pid, pid_params.setpoint, 100, pid_params.deadband, 0); 223 } 224 225 static inline void intel_pstate_reset_all_pid(void) 226 { 227 unsigned int cpu; 228 229 for_each_online_cpu(cpu) { 230 if (all_cpu_data[cpu]) 231 intel_pstate_busy_pid_reset(all_cpu_data[cpu]); 232 } 233 } 234 235 static inline void update_turbo_state(void) 236 { 237 u64 misc_en; 238 struct cpudata *cpu; 239 240 cpu = all_cpu_data[0]; 241 rdmsrl(MSR_IA32_MISC_ENABLE, misc_en); 242 limits.turbo_disabled = 243 (misc_en & MSR_IA32_MISC_ENABLE_TURBO_DISABLE || 244 cpu->pstate.max_pstate == cpu->pstate.turbo_pstate); 245 } 246 247 /************************** debugfs begin ************************/ 248 static int pid_param_set(void *data, u64 val) 249 { 250 *(u32 *)data = val; 251 intel_pstate_reset_all_pid(); 252 return 0; 253 } 254 255 static int pid_param_get(void *data, u64 *val) 256 { 257 *val = *(u32 *)data; 258 return 0; 259 } 260 DEFINE_SIMPLE_ATTRIBUTE(fops_pid_param, pid_param_get, pid_param_set, "%llu\n"); 261 262 struct pid_param { 263 char *name; 264 void *value; 265 }; 266 267 static struct pid_param pid_files[] = { 268 {"sample_rate_ms", &pid_params.sample_rate_ms}, 269 {"d_gain_pct", &pid_params.d_gain_pct}, 270 {"i_gain_pct", &pid_params.i_gain_pct}, 271 {"deadband", &pid_params.deadband}, 272 {"setpoint", &pid_params.setpoint}, 273 {"p_gain_pct", &pid_params.p_gain_pct}, 274 {NULL, NULL} 275 }; 276 277 static void __init intel_pstate_debug_expose_params(void) 278 { 279 struct dentry *debugfs_parent; 280 int i = 0; 281 282 debugfs_parent = debugfs_create_dir("pstate_snb", NULL); 283 if (IS_ERR_OR_NULL(debugfs_parent)) 284 return; 285 while (pid_files[i].name) { 286 debugfs_create_file(pid_files[i].name, 0660, 287 debugfs_parent, pid_files[i].value, 288 &fops_pid_param); 289 i++; 290 } 291 } 292 293 /************************** debugfs end ************************/ 294 295 /************************** sysfs begin ************************/ 296 #define show_one(file_name, object) \ 297 static ssize_t show_##file_name \ 298 (struct kobject *kobj, struct attribute *attr, char *buf) \ 299 { \ 300 return sprintf(buf, "%u\n", limits.object); \ 301 } 302 303 static ssize_t show_no_turbo(struct kobject *kobj, 304 struct attribute *attr, char *buf) 305 { 306 ssize_t ret; 307 308 update_turbo_state(); 309 if (limits.turbo_disabled) 310 ret = sprintf(buf, "%u\n", limits.turbo_disabled); 311 else 312 ret = sprintf(buf, "%u\n", limits.no_turbo); 313 314 return ret; 315 } 316 317 static ssize_t store_no_turbo(struct kobject *a, struct attribute *b, 318 const char *buf, size_t count) 319 { 320 unsigned int input; 321 int ret; 322 323 ret = sscanf(buf, "%u", &input); 324 if (ret != 1) 325 return -EINVAL; 326 327 update_turbo_state(); 328 if (limits.turbo_disabled) { 329 pr_warn("Turbo disabled by BIOS or unavailable on processor\n"); 330 return -EPERM; 331 } 332 limits.no_turbo = clamp_t(int, input, 0, 1); 333 334 return count; 335 } 336 337 static ssize_t store_max_perf_pct(struct kobject *a, struct attribute *b, 338 const char *buf, size_t count) 339 { 340 unsigned int input; 341 int ret; 342 343 ret = sscanf(buf, "%u", &input); 344 if (ret != 1) 345 return -EINVAL; 346 347 limits.max_sysfs_pct = clamp_t(int, input, 0 , 100); 348 limits.max_perf_pct = min(limits.max_policy_pct, limits.max_sysfs_pct); 349 limits.max_perf = div_fp(int_tofp(limits.max_perf_pct), int_tofp(100)); 350 351 return count; 352 } 353 354 static ssize_t store_min_perf_pct(struct kobject *a, struct attribute *b, 355 const char *buf, size_t count) 356 { 357 unsigned int input; 358 int ret; 359 360 ret = sscanf(buf, "%u", &input); 361 if (ret != 1) 362 return -EINVAL; 363 limits.min_perf_pct = clamp_t(int, input, 0 , 100); 364 limits.min_perf = div_fp(int_tofp(limits.min_perf_pct), int_tofp(100)); 365 366 return count; 367 } 368 369 show_one(max_perf_pct, max_perf_pct); 370 show_one(min_perf_pct, min_perf_pct); 371 372 define_one_global_rw(no_turbo); 373 define_one_global_rw(max_perf_pct); 374 define_one_global_rw(min_perf_pct); 375 376 static struct attribute *intel_pstate_attributes[] = { 377 &no_turbo.attr, 378 &max_perf_pct.attr, 379 &min_perf_pct.attr, 380 NULL 381 }; 382 383 static struct attribute_group intel_pstate_attr_group = { 384 .attrs = intel_pstate_attributes, 385 }; 386 387 static void __init intel_pstate_sysfs_expose_params(void) 388 { 389 struct kobject *intel_pstate_kobject; 390 int rc; 391 392 intel_pstate_kobject = kobject_create_and_add("intel_pstate", 393 &cpu_subsys.dev_root->kobj); 394 BUG_ON(!intel_pstate_kobject); 395 rc = sysfs_create_group(intel_pstate_kobject, &intel_pstate_attr_group); 396 BUG_ON(rc); 397 } 398 399 /************************** sysfs end ************************/ 400 static int byt_get_min_pstate(void) 401 { 402 u64 value; 403 404 rdmsrl(BYT_RATIOS, value); 405 return (value >> 8) & 0x7F; 406 } 407 408 static int byt_get_max_pstate(void) 409 { 410 u64 value; 411 412 rdmsrl(BYT_RATIOS, value); 413 return (value >> 16) & 0x7F; 414 } 415 416 static int byt_get_turbo_pstate(void) 417 { 418 u64 value; 419 420 rdmsrl(BYT_TURBO_RATIOS, value); 421 return value & 0x7F; 422 } 423 424 static void byt_set_pstate(struct cpudata *cpudata, int pstate) 425 { 426 u64 val; 427 int32_t vid_fp; 428 u32 vid; 429 430 val = pstate << 8; 431 if (limits.no_turbo && !limits.turbo_disabled) 432 val |= (u64)1 << 32; 433 434 vid_fp = cpudata->vid.min + mul_fp( 435 int_tofp(pstate - cpudata->pstate.min_pstate), 436 cpudata->vid.ratio); 437 438 vid_fp = clamp_t(int32_t, vid_fp, cpudata->vid.min, cpudata->vid.max); 439 vid = ceiling_fp(vid_fp); 440 441 if (pstate > cpudata->pstate.max_pstate) 442 vid = cpudata->vid.turbo; 443 444 val |= vid; 445 446 wrmsrl(MSR_IA32_PERF_CTL, val); 447 } 448 449 #define BYT_BCLK_FREQS 5 450 static int byt_freq_table[BYT_BCLK_FREQS] = { 833, 1000, 1333, 1167, 800}; 451 452 static int byt_get_scaling(void) 453 { 454 u64 value; 455 int i; 456 457 rdmsrl(MSR_FSB_FREQ, value); 458 i = value & 0x3; 459 460 BUG_ON(i > BYT_BCLK_FREQS); 461 462 return byt_freq_table[i] * 100; 463 } 464 465 static void byt_get_vid(struct cpudata *cpudata) 466 { 467 u64 value; 468 469 rdmsrl(BYT_VIDS, value); 470 cpudata->vid.min = int_tofp((value >> 8) & 0x7f); 471 cpudata->vid.max = int_tofp((value >> 16) & 0x7f); 472 cpudata->vid.ratio = div_fp( 473 cpudata->vid.max - cpudata->vid.min, 474 int_tofp(cpudata->pstate.max_pstate - 475 cpudata->pstate.min_pstate)); 476 477 rdmsrl(BYT_TURBO_VIDS, value); 478 cpudata->vid.turbo = value & 0x7f; 479 } 480 481 static int core_get_min_pstate(void) 482 { 483 u64 value; 484 485 rdmsrl(MSR_PLATFORM_INFO, value); 486 return (value >> 40) & 0xFF; 487 } 488 489 static int core_get_max_pstate(void) 490 { 491 u64 value; 492 493 rdmsrl(MSR_PLATFORM_INFO, value); 494 return (value >> 8) & 0xFF; 495 } 496 497 static int core_get_turbo_pstate(void) 498 { 499 u64 value; 500 int nont, ret; 501 502 rdmsrl(MSR_NHM_TURBO_RATIO_LIMIT, value); 503 nont = core_get_max_pstate(); 504 ret = (value) & 255; 505 if (ret <= nont) 506 ret = nont; 507 return ret; 508 } 509 510 static inline int core_get_scaling(void) 511 { 512 return 100000; 513 } 514 515 static void core_set_pstate(struct cpudata *cpudata, int pstate) 516 { 517 u64 val; 518 519 val = pstate << 8; 520 if (limits.no_turbo && !limits.turbo_disabled) 521 val |= (u64)1 << 32; 522 523 wrmsrl_on_cpu(cpudata->cpu, MSR_IA32_PERF_CTL, val); 524 } 525 526 static struct cpu_defaults core_params = { 527 .pid_policy = { 528 .sample_rate_ms = 10, 529 .deadband = 0, 530 .setpoint = 97, 531 .p_gain_pct = 20, 532 .d_gain_pct = 0, 533 .i_gain_pct = 0, 534 }, 535 .funcs = { 536 .get_max = core_get_max_pstate, 537 .get_min = core_get_min_pstate, 538 .get_turbo = core_get_turbo_pstate, 539 .get_scaling = core_get_scaling, 540 .set = core_set_pstate, 541 }, 542 }; 543 544 static struct cpu_defaults byt_params = { 545 .pid_policy = { 546 .sample_rate_ms = 10, 547 .deadband = 0, 548 .setpoint = 97, 549 .p_gain_pct = 14, 550 .d_gain_pct = 0, 551 .i_gain_pct = 4, 552 }, 553 .funcs = { 554 .get_max = byt_get_max_pstate, 555 .get_min = byt_get_min_pstate, 556 .get_turbo = byt_get_turbo_pstate, 557 .set = byt_set_pstate, 558 .get_scaling = byt_get_scaling, 559 .get_vid = byt_get_vid, 560 }, 561 }; 562 563 static void intel_pstate_get_min_max(struct cpudata *cpu, int *min, int *max) 564 { 565 int max_perf = cpu->pstate.turbo_pstate; 566 int max_perf_adj; 567 int min_perf; 568 569 if (limits.no_turbo || limits.turbo_disabled) 570 max_perf = cpu->pstate.max_pstate; 571 572 max_perf_adj = fp_toint(mul_fp(int_tofp(max_perf), limits.max_perf)); 573 *max = clamp_t(int, max_perf_adj, 574 cpu->pstate.min_pstate, cpu->pstate.turbo_pstate); 575 576 min_perf = fp_toint(mul_fp(int_tofp(max_perf), limits.min_perf)); 577 *min = clamp_t(int, min_perf, cpu->pstate.min_pstate, max_perf); 578 } 579 580 static void intel_pstate_set_pstate(struct cpudata *cpu, int pstate) 581 { 582 int max_perf, min_perf; 583 584 update_turbo_state(); 585 586 intel_pstate_get_min_max(cpu, &min_perf, &max_perf); 587 588 pstate = clamp_t(int, pstate, min_perf, max_perf); 589 590 if (pstate == cpu->pstate.current_pstate) 591 return; 592 593 trace_cpu_frequency(pstate * cpu->pstate.scaling, cpu->cpu); 594 595 cpu->pstate.current_pstate = pstate; 596 597 pstate_funcs.set(cpu, pstate); 598 } 599 600 static void intel_pstate_get_cpu_pstates(struct cpudata *cpu) 601 { 602 cpu->pstate.min_pstate = pstate_funcs.get_min(); 603 cpu->pstate.max_pstate = pstate_funcs.get_max(); 604 cpu->pstate.turbo_pstate = pstate_funcs.get_turbo(); 605 cpu->pstate.scaling = pstate_funcs.get_scaling(); 606 607 if (pstate_funcs.get_vid) 608 pstate_funcs.get_vid(cpu); 609 intel_pstate_set_pstate(cpu, cpu->pstate.min_pstate); 610 } 611 612 static inline void intel_pstate_calc_busy(struct cpudata *cpu) 613 { 614 struct sample *sample = &cpu->sample; 615 int64_t core_pct; 616 617 core_pct = int_tofp(sample->aperf) * int_tofp(100); 618 core_pct = div64_u64(core_pct, int_tofp(sample->mperf)); 619 620 sample->freq = fp_toint( 621 mul_fp(int_tofp( 622 cpu->pstate.max_pstate * cpu->pstate.scaling / 100), 623 core_pct)); 624 625 sample->core_pct_busy = (int32_t)core_pct; 626 } 627 628 static inline void intel_pstate_sample(struct cpudata *cpu) 629 { 630 u64 aperf, mperf; 631 unsigned long flags; 632 633 local_irq_save(flags); 634 rdmsrl(MSR_IA32_APERF, aperf); 635 rdmsrl(MSR_IA32_MPERF, mperf); 636 local_irq_restore(flags); 637 638 cpu->last_sample_time = cpu->sample.time; 639 cpu->sample.time = ktime_get(); 640 cpu->sample.aperf = aperf; 641 cpu->sample.mperf = mperf; 642 cpu->sample.aperf -= cpu->prev_aperf; 643 cpu->sample.mperf -= cpu->prev_mperf; 644 645 intel_pstate_calc_busy(cpu); 646 647 cpu->prev_aperf = aperf; 648 cpu->prev_mperf = mperf; 649 } 650 651 static inline void intel_pstate_set_sample_time(struct cpudata *cpu) 652 { 653 int delay; 654 655 delay = msecs_to_jiffies(pid_params.sample_rate_ms); 656 mod_timer_pinned(&cpu->timer, jiffies + delay); 657 } 658 659 static inline int32_t intel_pstate_get_scaled_busy(struct cpudata *cpu) 660 { 661 int32_t core_busy, max_pstate, current_pstate, sample_ratio; 662 u32 duration_us; 663 u32 sample_time; 664 665 core_busy = cpu->sample.core_pct_busy; 666 max_pstate = int_tofp(cpu->pstate.max_pstate); 667 current_pstate = int_tofp(cpu->pstate.current_pstate); 668 core_busy = mul_fp(core_busy, div_fp(max_pstate, current_pstate)); 669 670 sample_time = pid_params.sample_rate_ms * USEC_PER_MSEC; 671 duration_us = (u32) ktime_us_delta(cpu->sample.time, 672 cpu->last_sample_time); 673 if (duration_us > sample_time * 3) { 674 sample_ratio = div_fp(int_tofp(sample_time), 675 int_tofp(duration_us)); 676 core_busy = mul_fp(core_busy, sample_ratio); 677 } 678 679 return core_busy; 680 } 681 682 static inline void intel_pstate_adjust_busy_pstate(struct cpudata *cpu) 683 { 684 int32_t busy_scaled; 685 struct _pid *pid; 686 signed int ctl; 687 688 pid = &cpu->pid; 689 busy_scaled = intel_pstate_get_scaled_busy(cpu); 690 691 ctl = pid_calc(pid, busy_scaled); 692 693 /* Negative values of ctl increase the pstate and vice versa */ 694 intel_pstate_set_pstate(cpu, cpu->pstate.current_pstate - ctl); 695 } 696 697 static void intel_pstate_timer_func(unsigned long __data) 698 { 699 struct cpudata *cpu = (struct cpudata *) __data; 700 struct sample *sample; 701 702 intel_pstate_sample(cpu); 703 704 sample = &cpu->sample; 705 706 intel_pstate_adjust_busy_pstate(cpu); 707 708 trace_pstate_sample(fp_toint(sample->core_pct_busy), 709 fp_toint(intel_pstate_get_scaled_busy(cpu)), 710 cpu->pstate.current_pstate, 711 sample->mperf, 712 sample->aperf, 713 sample->freq); 714 715 intel_pstate_set_sample_time(cpu); 716 } 717 718 #define ICPU(model, policy) \ 719 { X86_VENDOR_INTEL, 6, model, X86_FEATURE_APERFMPERF,\ 720 (unsigned long)&policy } 721 722 static const struct x86_cpu_id intel_pstate_cpu_ids[] = { 723 ICPU(0x2a, core_params), 724 ICPU(0x2d, core_params), 725 ICPU(0x37, byt_params), 726 ICPU(0x3a, core_params), 727 ICPU(0x3c, core_params), 728 ICPU(0x3d, core_params), 729 ICPU(0x3e, core_params), 730 ICPU(0x3f, core_params), 731 ICPU(0x45, core_params), 732 ICPU(0x46, core_params), 733 ICPU(0x4c, byt_params), 734 ICPU(0x4f, core_params), 735 ICPU(0x56, core_params), 736 {} 737 }; 738 MODULE_DEVICE_TABLE(x86cpu, intel_pstate_cpu_ids); 739 740 static int intel_pstate_init_cpu(unsigned int cpunum) 741 { 742 struct cpudata *cpu; 743 744 if (!all_cpu_data[cpunum]) 745 all_cpu_data[cpunum] = kzalloc(sizeof(struct cpudata), 746 GFP_KERNEL); 747 if (!all_cpu_data[cpunum]) 748 return -ENOMEM; 749 750 cpu = all_cpu_data[cpunum]; 751 752 cpu->cpu = cpunum; 753 intel_pstate_get_cpu_pstates(cpu); 754 755 init_timer_deferrable(&cpu->timer); 756 cpu->timer.function = intel_pstate_timer_func; 757 cpu->timer.data = (unsigned long)cpu; 758 cpu->timer.expires = jiffies + HZ/100; 759 intel_pstate_busy_pid_reset(cpu); 760 intel_pstate_sample(cpu); 761 762 add_timer_on(&cpu->timer, cpunum); 763 764 pr_debug("Intel pstate controlling: cpu %d\n", cpunum); 765 766 return 0; 767 } 768 769 static unsigned int intel_pstate_get(unsigned int cpu_num) 770 { 771 struct sample *sample; 772 struct cpudata *cpu; 773 774 cpu = all_cpu_data[cpu_num]; 775 if (!cpu) 776 return 0; 777 sample = &cpu->sample; 778 return sample->freq; 779 } 780 781 static int intel_pstate_set_policy(struct cpufreq_policy *policy) 782 { 783 if (!policy->cpuinfo.max_freq) 784 return -ENODEV; 785 786 if (policy->policy == CPUFREQ_POLICY_PERFORMANCE) { 787 limits.min_perf_pct = 100; 788 limits.min_perf = int_tofp(1); 789 limits.max_policy_pct = 100; 790 limits.max_perf_pct = 100; 791 limits.max_perf = int_tofp(1); 792 limits.no_turbo = 0; 793 return 0; 794 } 795 limits.min_perf_pct = (policy->min * 100) / policy->cpuinfo.max_freq; 796 limits.min_perf_pct = clamp_t(int, limits.min_perf_pct, 0 , 100); 797 limits.min_perf = div_fp(int_tofp(limits.min_perf_pct), int_tofp(100)); 798 799 limits.max_policy_pct = (policy->max * 100) / policy->cpuinfo.max_freq; 800 limits.max_policy_pct = clamp_t(int, limits.max_policy_pct, 0 , 100); 801 limits.max_perf_pct = min(limits.max_policy_pct, limits.max_sysfs_pct); 802 limits.max_perf = div_fp(int_tofp(limits.max_perf_pct), int_tofp(100)); 803 804 return 0; 805 } 806 807 static int intel_pstate_verify_policy(struct cpufreq_policy *policy) 808 { 809 cpufreq_verify_within_cpu_limits(policy); 810 811 if (policy->policy != CPUFREQ_POLICY_POWERSAVE && 812 policy->policy != CPUFREQ_POLICY_PERFORMANCE) 813 return -EINVAL; 814 815 return 0; 816 } 817 818 static void intel_pstate_stop_cpu(struct cpufreq_policy *policy) 819 { 820 int cpu_num = policy->cpu; 821 struct cpudata *cpu = all_cpu_data[cpu_num]; 822 823 pr_info("intel_pstate CPU %d exiting\n", cpu_num); 824 825 del_timer_sync(&all_cpu_data[cpu_num]->timer); 826 intel_pstate_set_pstate(cpu, cpu->pstate.min_pstate); 827 } 828 829 static int intel_pstate_cpu_init(struct cpufreq_policy *policy) 830 { 831 struct cpudata *cpu; 832 int rc; 833 834 rc = intel_pstate_init_cpu(policy->cpu); 835 if (rc) 836 return rc; 837 838 cpu = all_cpu_data[policy->cpu]; 839 840 if (limits.min_perf_pct == 100 && limits.max_perf_pct == 100) 841 policy->policy = CPUFREQ_POLICY_PERFORMANCE; 842 else 843 policy->policy = CPUFREQ_POLICY_POWERSAVE; 844 845 policy->min = cpu->pstate.min_pstate * cpu->pstate.scaling; 846 policy->max = cpu->pstate.turbo_pstate * cpu->pstate.scaling; 847 848 /* cpuinfo and default policy values */ 849 policy->cpuinfo.min_freq = cpu->pstate.min_pstate * cpu->pstate.scaling; 850 policy->cpuinfo.max_freq = 851 cpu->pstate.turbo_pstate * cpu->pstate.scaling; 852 policy->cpuinfo.transition_latency = CPUFREQ_ETERNAL; 853 cpumask_set_cpu(policy->cpu, policy->cpus); 854 855 return 0; 856 } 857 858 static struct cpufreq_driver intel_pstate_driver = { 859 .flags = CPUFREQ_CONST_LOOPS, 860 .verify = intel_pstate_verify_policy, 861 .setpolicy = intel_pstate_set_policy, 862 .get = intel_pstate_get, 863 .init = intel_pstate_cpu_init, 864 .stop_cpu = intel_pstate_stop_cpu, 865 .name = "intel_pstate", 866 }; 867 868 static int __initdata no_load; 869 870 static int intel_pstate_msrs_not_valid(void) 871 { 872 /* Check that all the msr's we are using are valid. */ 873 u64 aperf, mperf, tmp; 874 875 rdmsrl(MSR_IA32_APERF, aperf); 876 rdmsrl(MSR_IA32_MPERF, mperf); 877 878 if (!pstate_funcs.get_max() || 879 !pstate_funcs.get_min() || 880 !pstate_funcs.get_turbo()) 881 return -ENODEV; 882 883 rdmsrl(MSR_IA32_APERF, tmp); 884 if (!(tmp - aperf)) 885 return -ENODEV; 886 887 rdmsrl(MSR_IA32_MPERF, tmp); 888 if (!(tmp - mperf)) 889 return -ENODEV; 890 891 return 0; 892 } 893 894 static void copy_pid_params(struct pstate_adjust_policy *policy) 895 { 896 pid_params.sample_rate_ms = policy->sample_rate_ms; 897 pid_params.p_gain_pct = policy->p_gain_pct; 898 pid_params.i_gain_pct = policy->i_gain_pct; 899 pid_params.d_gain_pct = policy->d_gain_pct; 900 pid_params.deadband = policy->deadband; 901 pid_params.setpoint = policy->setpoint; 902 } 903 904 static void copy_cpu_funcs(struct pstate_funcs *funcs) 905 { 906 pstate_funcs.get_max = funcs->get_max; 907 pstate_funcs.get_min = funcs->get_min; 908 pstate_funcs.get_turbo = funcs->get_turbo; 909 pstate_funcs.get_scaling = funcs->get_scaling; 910 pstate_funcs.set = funcs->set; 911 pstate_funcs.get_vid = funcs->get_vid; 912 } 913 914 #if IS_ENABLED(CONFIG_ACPI) 915 #include <acpi/processor.h> 916 917 static bool intel_pstate_no_acpi_pss(void) 918 { 919 int i; 920 921 for_each_possible_cpu(i) { 922 acpi_status status; 923 union acpi_object *pss; 924 struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL }; 925 struct acpi_processor *pr = per_cpu(processors, i); 926 927 if (!pr) 928 continue; 929 930 status = acpi_evaluate_object(pr->handle, "_PSS", NULL, &buffer); 931 if (ACPI_FAILURE(status)) 932 continue; 933 934 pss = buffer.pointer; 935 if (pss && pss->type == ACPI_TYPE_PACKAGE) { 936 kfree(pss); 937 return false; 938 } 939 940 kfree(pss); 941 } 942 943 return true; 944 } 945 946 struct hw_vendor_info { 947 u16 valid; 948 char oem_id[ACPI_OEM_ID_SIZE]; 949 char oem_table_id[ACPI_OEM_TABLE_ID_SIZE]; 950 }; 951 952 /* Hardware vendor-specific info that has its own power management modes */ 953 static struct hw_vendor_info vendor_info[] = { 954 {1, "HP ", "ProLiant"}, 955 {0, "", ""}, 956 }; 957 958 static bool intel_pstate_platform_pwr_mgmt_exists(void) 959 { 960 struct acpi_table_header hdr; 961 struct hw_vendor_info *v_info; 962 963 if (acpi_disabled || 964 ACPI_FAILURE(acpi_get_table_header(ACPI_SIG_FADT, 0, &hdr))) 965 return false; 966 967 for (v_info = vendor_info; v_info->valid; v_info++) { 968 if (!strncmp(hdr.oem_id, v_info->oem_id, ACPI_OEM_ID_SIZE) && 969 !strncmp(hdr.oem_table_id, v_info->oem_table_id, ACPI_OEM_TABLE_ID_SIZE) && 970 intel_pstate_no_acpi_pss()) 971 return true; 972 } 973 974 return false; 975 } 976 #else /* CONFIG_ACPI not enabled */ 977 static inline bool intel_pstate_platform_pwr_mgmt_exists(void) { return false; } 978 #endif /* CONFIG_ACPI */ 979 980 static int __init intel_pstate_init(void) 981 { 982 int cpu, rc = 0; 983 const struct x86_cpu_id *id; 984 struct cpu_defaults *cpu_info; 985 986 if (no_load) 987 return -ENODEV; 988 989 id = x86_match_cpu(intel_pstate_cpu_ids); 990 if (!id) 991 return -ENODEV; 992 993 /* 994 * The Intel pstate driver will be ignored if the platform 995 * firmware has its own power management modes. 996 */ 997 if (intel_pstate_platform_pwr_mgmt_exists()) 998 return -ENODEV; 999 1000 cpu_info = (struct cpu_defaults *)id->driver_data; 1001 1002 copy_pid_params(&cpu_info->pid_policy); 1003 copy_cpu_funcs(&cpu_info->funcs); 1004 1005 if (intel_pstate_msrs_not_valid()) 1006 return -ENODEV; 1007 1008 pr_info("Intel P-state driver initializing.\n"); 1009 1010 all_cpu_data = vzalloc(sizeof(void *) * num_possible_cpus()); 1011 if (!all_cpu_data) 1012 return -ENOMEM; 1013 1014 rc = cpufreq_register_driver(&intel_pstate_driver); 1015 if (rc) 1016 goto out; 1017 1018 intel_pstate_debug_expose_params(); 1019 intel_pstate_sysfs_expose_params(); 1020 1021 return rc; 1022 out: 1023 get_online_cpus(); 1024 for_each_online_cpu(cpu) { 1025 if (all_cpu_data[cpu]) { 1026 del_timer_sync(&all_cpu_data[cpu]->timer); 1027 kfree(all_cpu_data[cpu]); 1028 } 1029 } 1030 1031 put_online_cpus(); 1032 vfree(all_cpu_data); 1033 return -ENODEV; 1034 } 1035 device_initcall(intel_pstate_init); 1036 1037 static int __init intel_pstate_setup(char *str) 1038 { 1039 if (!str) 1040 return -EINVAL; 1041 1042 if (!strcmp(str, "disable")) 1043 no_load = 1; 1044 return 0; 1045 } 1046 early_param("intel_pstate", intel_pstate_setup); 1047 1048 MODULE_AUTHOR("Dirk Brandewie <dirk.j.brandewie@intel.com>"); 1049 MODULE_DESCRIPTION("'intel_pstate' - P state driver Intel Core processors"); 1050 MODULE_LICENSE("GPL"); 1051