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 struct sample { 56 int32_t core_pct_busy; 57 u64 aperf; 58 u64 mperf; 59 int freq; 60 ktime_t time; 61 }; 62 63 struct pstate_data { 64 int current_pstate; 65 int min_pstate; 66 int max_pstate; 67 int turbo_pstate; 68 }; 69 70 struct vid_data { 71 int min; 72 int max; 73 int turbo; 74 int32_t ratio; 75 }; 76 77 struct _pid { 78 int setpoint; 79 int32_t integral; 80 int32_t p_gain; 81 int32_t i_gain; 82 int32_t d_gain; 83 int deadband; 84 int32_t last_err; 85 }; 86 87 struct cpudata { 88 int cpu; 89 90 struct timer_list timer; 91 92 struct pstate_data pstate; 93 struct vid_data vid; 94 struct _pid pid; 95 96 ktime_t last_sample_time; 97 u64 prev_aperf; 98 u64 prev_mperf; 99 struct sample sample; 100 }; 101 102 static struct cpudata **all_cpu_data; 103 struct pstate_adjust_policy { 104 int sample_rate_ms; 105 int deadband; 106 int setpoint; 107 int p_gain_pct; 108 int d_gain_pct; 109 int i_gain_pct; 110 }; 111 112 struct pstate_funcs { 113 int (*get_max)(void); 114 int (*get_min)(void); 115 int (*get_turbo)(void); 116 void (*set)(struct cpudata*, int pstate); 117 void (*get_vid)(struct cpudata *); 118 }; 119 120 struct cpu_defaults { 121 struct pstate_adjust_policy pid_policy; 122 struct pstate_funcs funcs; 123 }; 124 125 static struct pstate_adjust_policy pid_params; 126 static struct pstate_funcs pstate_funcs; 127 128 struct perf_limits { 129 int no_turbo; 130 int turbo_disabled; 131 int max_perf_pct; 132 int min_perf_pct; 133 int32_t max_perf; 134 int32_t min_perf; 135 int max_policy_pct; 136 int max_sysfs_pct; 137 }; 138 139 static struct perf_limits limits = { 140 .no_turbo = 0, 141 .max_perf_pct = 100, 142 .max_perf = int_tofp(1), 143 .min_perf_pct = 0, 144 .min_perf = 0, 145 .max_policy_pct = 100, 146 .max_sysfs_pct = 100, 147 }; 148 149 static inline void pid_reset(struct _pid *pid, int setpoint, int busy, 150 int deadband, int integral) { 151 pid->setpoint = setpoint; 152 pid->deadband = deadband; 153 pid->integral = int_tofp(integral); 154 pid->last_err = int_tofp(setpoint) - int_tofp(busy); 155 } 156 157 static inline void pid_p_gain_set(struct _pid *pid, int percent) 158 { 159 pid->p_gain = div_fp(int_tofp(percent), int_tofp(100)); 160 } 161 162 static inline void pid_i_gain_set(struct _pid *pid, int percent) 163 { 164 pid->i_gain = div_fp(int_tofp(percent), int_tofp(100)); 165 } 166 167 static inline void pid_d_gain_set(struct _pid *pid, int percent) 168 { 169 pid->d_gain = div_fp(int_tofp(percent), int_tofp(100)); 170 } 171 172 static signed int pid_calc(struct _pid *pid, int32_t busy) 173 { 174 signed int result; 175 int32_t pterm, dterm, fp_error; 176 int32_t integral_limit; 177 178 fp_error = int_tofp(pid->setpoint) - busy; 179 180 if (abs(fp_error) <= int_tofp(pid->deadband)) 181 return 0; 182 183 pterm = mul_fp(pid->p_gain, fp_error); 184 185 pid->integral += fp_error; 186 187 /* limit the integral term */ 188 integral_limit = int_tofp(30); 189 if (pid->integral > integral_limit) 190 pid->integral = integral_limit; 191 if (pid->integral < -integral_limit) 192 pid->integral = -integral_limit; 193 194 dterm = mul_fp(pid->d_gain, fp_error - pid->last_err); 195 pid->last_err = fp_error; 196 197 result = pterm + mul_fp(pid->integral, pid->i_gain) + dterm; 198 result = result + (1 << (FRAC_BITS-1)); 199 return (signed int)fp_toint(result); 200 } 201 202 static inline void intel_pstate_busy_pid_reset(struct cpudata *cpu) 203 { 204 pid_p_gain_set(&cpu->pid, pid_params.p_gain_pct); 205 pid_d_gain_set(&cpu->pid, pid_params.d_gain_pct); 206 pid_i_gain_set(&cpu->pid, pid_params.i_gain_pct); 207 208 pid_reset(&cpu->pid, pid_params.setpoint, 100, pid_params.deadband, 0); 209 } 210 211 static inline void intel_pstate_reset_all_pid(void) 212 { 213 unsigned int cpu; 214 215 for_each_online_cpu(cpu) { 216 if (all_cpu_data[cpu]) 217 intel_pstate_busy_pid_reset(all_cpu_data[cpu]); 218 } 219 } 220 221 /************************** debugfs begin ************************/ 222 static int pid_param_set(void *data, u64 val) 223 { 224 *(u32 *)data = val; 225 intel_pstate_reset_all_pid(); 226 return 0; 227 } 228 229 static int pid_param_get(void *data, u64 *val) 230 { 231 *val = *(u32 *)data; 232 return 0; 233 } 234 DEFINE_SIMPLE_ATTRIBUTE(fops_pid_param, pid_param_get, pid_param_set, "%llu\n"); 235 236 struct pid_param { 237 char *name; 238 void *value; 239 }; 240 241 static struct pid_param pid_files[] = { 242 {"sample_rate_ms", &pid_params.sample_rate_ms}, 243 {"d_gain_pct", &pid_params.d_gain_pct}, 244 {"i_gain_pct", &pid_params.i_gain_pct}, 245 {"deadband", &pid_params.deadband}, 246 {"setpoint", &pid_params.setpoint}, 247 {"p_gain_pct", &pid_params.p_gain_pct}, 248 {NULL, NULL} 249 }; 250 251 static void __init intel_pstate_debug_expose_params(void) 252 { 253 struct dentry *debugfs_parent; 254 int i = 0; 255 256 debugfs_parent = debugfs_create_dir("pstate_snb", NULL); 257 if (IS_ERR_OR_NULL(debugfs_parent)) 258 return; 259 while (pid_files[i].name) { 260 debugfs_create_file(pid_files[i].name, 0660, 261 debugfs_parent, pid_files[i].value, 262 &fops_pid_param); 263 i++; 264 } 265 } 266 267 /************************** debugfs end ************************/ 268 269 /************************** sysfs begin ************************/ 270 #define show_one(file_name, object) \ 271 static ssize_t show_##file_name \ 272 (struct kobject *kobj, struct attribute *attr, char *buf) \ 273 { \ 274 return sprintf(buf, "%u\n", limits.object); \ 275 } 276 277 static ssize_t store_no_turbo(struct kobject *a, struct attribute *b, 278 const char *buf, size_t count) 279 { 280 unsigned int input; 281 int ret; 282 283 ret = sscanf(buf, "%u", &input); 284 if (ret != 1) 285 return -EINVAL; 286 limits.no_turbo = clamp_t(int, input, 0 , 1); 287 if (limits.turbo_disabled) { 288 pr_warn("Turbo disabled by BIOS or unavailable on processor\n"); 289 limits.no_turbo = limits.turbo_disabled; 290 } 291 return count; 292 } 293 294 static ssize_t store_max_perf_pct(struct kobject *a, struct attribute *b, 295 const char *buf, size_t count) 296 { 297 unsigned int input; 298 int ret; 299 300 ret = sscanf(buf, "%u", &input); 301 if (ret != 1) 302 return -EINVAL; 303 304 limits.max_sysfs_pct = clamp_t(int, input, 0 , 100); 305 limits.max_perf_pct = min(limits.max_policy_pct, limits.max_sysfs_pct); 306 limits.max_perf = div_fp(int_tofp(limits.max_perf_pct), int_tofp(100)); 307 308 return count; 309 } 310 311 static ssize_t store_min_perf_pct(struct kobject *a, struct attribute *b, 312 const char *buf, size_t count) 313 { 314 unsigned int input; 315 int ret; 316 317 ret = sscanf(buf, "%u", &input); 318 if (ret != 1) 319 return -EINVAL; 320 limits.min_perf_pct = clamp_t(int, input, 0 , 100); 321 limits.min_perf = div_fp(int_tofp(limits.min_perf_pct), int_tofp(100)); 322 323 return count; 324 } 325 326 show_one(no_turbo, no_turbo); 327 show_one(max_perf_pct, max_perf_pct); 328 show_one(min_perf_pct, min_perf_pct); 329 330 define_one_global_rw(no_turbo); 331 define_one_global_rw(max_perf_pct); 332 define_one_global_rw(min_perf_pct); 333 334 static struct attribute *intel_pstate_attributes[] = { 335 &no_turbo.attr, 336 &max_perf_pct.attr, 337 &min_perf_pct.attr, 338 NULL 339 }; 340 341 static struct attribute_group intel_pstate_attr_group = { 342 .attrs = intel_pstate_attributes, 343 }; 344 345 static void __init intel_pstate_sysfs_expose_params(void) 346 { 347 struct kobject *intel_pstate_kobject; 348 int rc; 349 350 intel_pstate_kobject = kobject_create_and_add("intel_pstate", 351 &cpu_subsys.dev_root->kobj); 352 BUG_ON(!intel_pstate_kobject); 353 rc = sysfs_create_group(intel_pstate_kobject, &intel_pstate_attr_group); 354 BUG_ON(rc); 355 } 356 357 /************************** sysfs end ************************/ 358 static int byt_get_min_pstate(void) 359 { 360 u64 value; 361 362 rdmsrl(BYT_RATIOS, value); 363 return (value >> 8) & 0x7F; 364 } 365 366 static int byt_get_max_pstate(void) 367 { 368 u64 value; 369 370 rdmsrl(BYT_RATIOS, value); 371 return (value >> 16) & 0x7F; 372 } 373 374 static int byt_get_turbo_pstate(void) 375 { 376 u64 value; 377 378 rdmsrl(BYT_TURBO_RATIOS, value); 379 return value & 0x7F; 380 } 381 382 static void byt_set_pstate(struct cpudata *cpudata, int pstate) 383 { 384 u64 val; 385 int32_t vid_fp; 386 u32 vid; 387 388 val = pstate << 8; 389 if (limits.no_turbo && !limits.turbo_disabled) 390 val |= (u64)1 << 32; 391 392 vid_fp = cpudata->vid.min + mul_fp( 393 int_tofp(pstate - cpudata->pstate.min_pstate), 394 cpudata->vid.ratio); 395 396 vid_fp = clamp_t(int32_t, vid_fp, cpudata->vid.min, cpudata->vid.max); 397 vid = fp_toint(vid_fp); 398 399 if (pstate > cpudata->pstate.max_pstate) 400 vid = cpudata->vid.turbo; 401 402 val |= vid; 403 404 wrmsrl(MSR_IA32_PERF_CTL, val); 405 } 406 407 static void byt_get_vid(struct cpudata *cpudata) 408 { 409 u64 value; 410 411 rdmsrl(BYT_VIDS, value); 412 cpudata->vid.min = int_tofp((value >> 8) & 0x7f); 413 cpudata->vid.max = int_tofp((value >> 16) & 0x7f); 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 static int core_get_min_pstate(void) 424 { 425 u64 value; 426 427 rdmsrl(MSR_PLATFORM_INFO, value); 428 return (value >> 40) & 0xFF; 429 } 430 431 static int core_get_max_pstate(void) 432 { 433 u64 value; 434 435 rdmsrl(MSR_PLATFORM_INFO, value); 436 return (value >> 8) & 0xFF; 437 } 438 439 static int core_get_turbo_pstate(void) 440 { 441 u64 value; 442 int nont, ret; 443 444 rdmsrl(MSR_NHM_TURBO_RATIO_LIMIT, value); 445 nont = core_get_max_pstate(); 446 ret = (value) & 255; 447 if (ret <= nont) 448 ret = nont; 449 return ret; 450 } 451 452 static void core_set_pstate(struct cpudata *cpudata, int pstate) 453 { 454 u64 val; 455 456 val = pstate << 8; 457 if (limits.no_turbo && !limits.turbo_disabled) 458 val |= (u64)1 << 32; 459 460 wrmsrl_on_cpu(cpudata->cpu, MSR_IA32_PERF_CTL, val); 461 } 462 463 static struct cpu_defaults core_params = { 464 .pid_policy = { 465 .sample_rate_ms = 10, 466 .deadband = 0, 467 .setpoint = 97, 468 .p_gain_pct = 20, 469 .d_gain_pct = 0, 470 .i_gain_pct = 0, 471 }, 472 .funcs = { 473 .get_max = core_get_max_pstate, 474 .get_min = core_get_min_pstate, 475 .get_turbo = core_get_turbo_pstate, 476 .set = core_set_pstate, 477 }, 478 }; 479 480 static struct cpu_defaults byt_params = { 481 .pid_policy = { 482 .sample_rate_ms = 10, 483 .deadband = 0, 484 .setpoint = 97, 485 .p_gain_pct = 14, 486 .d_gain_pct = 0, 487 .i_gain_pct = 4, 488 }, 489 .funcs = { 490 .get_max = byt_get_max_pstate, 491 .get_min = byt_get_min_pstate, 492 .get_turbo = byt_get_turbo_pstate, 493 .set = byt_set_pstate, 494 .get_vid = byt_get_vid, 495 }, 496 }; 497 498 static void intel_pstate_get_min_max(struct cpudata *cpu, int *min, int *max) 499 { 500 int max_perf = cpu->pstate.turbo_pstate; 501 int max_perf_adj; 502 int min_perf; 503 504 if (limits.no_turbo) 505 max_perf = cpu->pstate.max_pstate; 506 507 max_perf_adj = fp_toint(mul_fp(int_tofp(max_perf), limits.max_perf)); 508 *max = clamp_t(int, max_perf_adj, 509 cpu->pstate.min_pstate, cpu->pstate.turbo_pstate); 510 511 min_perf = fp_toint(mul_fp(int_tofp(max_perf), limits.min_perf)); 512 *min = clamp_t(int, min_perf, cpu->pstate.min_pstate, max_perf); 513 } 514 515 static void intel_pstate_set_pstate(struct cpudata *cpu, int pstate) 516 { 517 int max_perf, min_perf; 518 519 intel_pstate_get_min_max(cpu, &min_perf, &max_perf); 520 521 pstate = clamp_t(int, pstate, min_perf, max_perf); 522 523 if (pstate == cpu->pstate.current_pstate) 524 return; 525 526 trace_cpu_frequency(pstate * 100000, cpu->cpu); 527 528 cpu->pstate.current_pstate = pstate; 529 530 pstate_funcs.set(cpu, pstate); 531 } 532 533 static void intel_pstate_get_cpu_pstates(struct cpudata *cpu) 534 { 535 cpu->pstate.min_pstate = pstate_funcs.get_min(); 536 cpu->pstate.max_pstate = pstate_funcs.get_max(); 537 cpu->pstate.turbo_pstate = pstate_funcs.get_turbo(); 538 539 if (pstate_funcs.get_vid) 540 pstate_funcs.get_vid(cpu); 541 intel_pstate_set_pstate(cpu, cpu->pstate.min_pstate); 542 } 543 544 static inline void intel_pstate_calc_busy(struct cpudata *cpu) 545 { 546 struct sample *sample = &cpu->sample; 547 int64_t core_pct; 548 549 core_pct = int_tofp(sample->aperf) * int_tofp(100); 550 core_pct = div64_u64(core_pct, int_tofp(sample->mperf)); 551 552 sample->freq = fp_toint( 553 mul_fp(int_tofp(cpu->pstate.max_pstate * 1000), core_pct)); 554 555 sample->core_pct_busy = (int32_t)core_pct; 556 } 557 558 static inline void intel_pstate_sample(struct cpudata *cpu) 559 { 560 u64 aperf, mperf; 561 unsigned long flags; 562 563 local_irq_save(flags); 564 rdmsrl(MSR_IA32_APERF, aperf); 565 rdmsrl(MSR_IA32_MPERF, mperf); 566 local_irq_restore(flags); 567 568 cpu->last_sample_time = cpu->sample.time; 569 cpu->sample.time = ktime_get(); 570 cpu->sample.aperf = aperf; 571 cpu->sample.mperf = mperf; 572 cpu->sample.aperf -= cpu->prev_aperf; 573 cpu->sample.mperf -= cpu->prev_mperf; 574 575 intel_pstate_calc_busy(cpu); 576 577 cpu->prev_aperf = aperf; 578 cpu->prev_mperf = mperf; 579 } 580 581 static inline void intel_pstate_set_sample_time(struct cpudata *cpu) 582 { 583 int delay; 584 585 delay = msecs_to_jiffies(pid_params.sample_rate_ms); 586 mod_timer_pinned(&cpu->timer, jiffies + delay); 587 } 588 589 static inline int32_t intel_pstate_get_scaled_busy(struct cpudata *cpu) 590 { 591 int32_t core_busy, max_pstate, current_pstate, sample_ratio; 592 u32 duration_us; 593 u32 sample_time; 594 595 core_busy = cpu->sample.core_pct_busy; 596 max_pstate = int_tofp(cpu->pstate.max_pstate); 597 current_pstate = int_tofp(cpu->pstate.current_pstate); 598 core_busy = mul_fp(core_busy, div_fp(max_pstate, current_pstate)); 599 600 sample_time = pid_params.sample_rate_ms * USEC_PER_MSEC; 601 duration_us = (u32) ktime_us_delta(cpu->sample.time, 602 cpu->last_sample_time); 603 if (duration_us > sample_time * 3) { 604 sample_ratio = div_fp(int_tofp(sample_time), 605 int_tofp(duration_us)); 606 core_busy = mul_fp(core_busy, sample_ratio); 607 } 608 609 return core_busy; 610 } 611 612 static inline void intel_pstate_adjust_busy_pstate(struct cpudata *cpu) 613 { 614 int32_t busy_scaled; 615 struct _pid *pid; 616 signed int ctl; 617 618 pid = &cpu->pid; 619 busy_scaled = intel_pstate_get_scaled_busy(cpu); 620 621 ctl = pid_calc(pid, busy_scaled); 622 623 /* Negative values of ctl increase the pstate and vice versa */ 624 intel_pstate_set_pstate(cpu, cpu->pstate.current_pstate - ctl); 625 } 626 627 static void intel_pstate_timer_func(unsigned long __data) 628 { 629 struct cpudata *cpu = (struct cpudata *) __data; 630 struct sample *sample; 631 632 intel_pstate_sample(cpu); 633 634 sample = &cpu->sample; 635 636 intel_pstate_adjust_busy_pstate(cpu); 637 638 trace_pstate_sample(fp_toint(sample->core_pct_busy), 639 fp_toint(intel_pstate_get_scaled_busy(cpu)), 640 cpu->pstate.current_pstate, 641 sample->mperf, 642 sample->aperf, 643 sample->freq); 644 645 intel_pstate_set_sample_time(cpu); 646 } 647 648 #define ICPU(model, policy) \ 649 { X86_VENDOR_INTEL, 6, model, X86_FEATURE_APERFMPERF,\ 650 (unsigned long)&policy } 651 652 static const struct x86_cpu_id intel_pstate_cpu_ids[] = { 653 ICPU(0x2a, core_params), 654 ICPU(0x2d, core_params), 655 ICPU(0x37, byt_params), 656 ICPU(0x3a, core_params), 657 ICPU(0x3c, core_params), 658 ICPU(0x3d, core_params), 659 ICPU(0x3e, core_params), 660 ICPU(0x3f, core_params), 661 ICPU(0x45, core_params), 662 ICPU(0x46, core_params), 663 ICPU(0x4c, byt_params), 664 ICPU(0x4f, core_params), 665 ICPU(0x56, core_params), 666 {} 667 }; 668 MODULE_DEVICE_TABLE(x86cpu, intel_pstate_cpu_ids); 669 670 static int intel_pstate_init_cpu(unsigned int cpunum) 671 { 672 struct cpudata *cpu; 673 674 all_cpu_data[cpunum] = kzalloc(sizeof(struct cpudata), GFP_KERNEL); 675 if (!all_cpu_data[cpunum]) 676 return -ENOMEM; 677 678 cpu = all_cpu_data[cpunum]; 679 680 cpu->cpu = cpunum; 681 intel_pstate_get_cpu_pstates(cpu); 682 683 init_timer_deferrable(&cpu->timer); 684 cpu->timer.function = intel_pstate_timer_func; 685 cpu->timer.data = (unsigned long)cpu; 686 cpu->timer.expires = jiffies + HZ/100; 687 intel_pstate_busy_pid_reset(cpu); 688 intel_pstate_sample(cpu); 689 690 add_timer_on(&cpu->timer, cpunum); 691 692 pr_debug("Intel pstate controlling: cpu %d\n", cpunum); 693 694 return 0; 695 } 696 697 static unsigned int intel_pstate_get(unsigned int cpu_num) 698 { 699 struct sample *sample; 700 struct cpudata *cpu; 701 702 cpu = all_cpu_data[cpu_num]; 703 if (!cpu) 704 return 0; 705 sample = &cpu->sample; 706 return sample->freq; 707 } 708 709 static int intel_pstate_set_policy(struct cpufreq_policy *policy) 710 { 711 if (!policy->cpuinfo.max_freq) 712 return -ENODEV; 713 714 if (policy->policy == CPUFREQ_POLICY_PERFORMANCE) { 715 limits.min_perf_pct = 100; 716 limits.min_perf = int_tofp(1); 717 limits.max_perf_pct = 100; 718 limits.max_perf = int_tofp(1); 719 limits.no_turbo = limits.turbo_disabled; 720 return 0; 721 } 722 limits.min_perf_pct = (policy->min * 100) / policy->cpuinfo.max_freq; 723 limits.min_perf_pct = clamp_t(int, limits.min_perf_pct, 0 , 100); 724 limits.min_perf = div_fp(int_tofp(limits.min_perf_pct), int_tofp(100)); 725 726 limits.max_policy_pct = (policy->max * 100) / policy->cpuinfo.max_freq; 727 limits.max_policy_pct = clamp_t(int, limits.max_policy_pct, 0 , 100); 728 limits.max_perf_pct = min(limits.max_policy_pct, limits.max_sysfs_pct); 729 limits.max_perf = div_fp(int_tofp(limits.max_perf_pct), int_tofp(100)); 730 731 return 0; 732 } 733 734 static int intel_pstate_verify_policy(struct cpufreq_policy *policy) 735 { 736 cpufreq_verify_within_cpu_limits(policy); 737 738 if (policy->policy != CPUFREQ_POLICY_POWERSAVE && 739 policy->policy != CPUFREQ_POLICY_PERFORMANCE) 740 return -EINVAL; 741 742 return 0; 743 } 744 745 static void intel_pstate_stop_cpu(struct cpufreq_policy *policy) 746 { 747 int cpu_num = policy->cpu; 748 struct cpudata *cpu = all_cpu_data[cpu_num]; 749 750 pr_info("intel_pstate CPU %d exiting\n", cpu_num); 751 752 del_timer_sync(&all_cpu_data[cpu_num]->timer); 753 intel_pstate_set_pstate(cpu, cpu->pstate.min_pstate); 754 kfree(all_cpu_data[cpu_num]); 755 all_cpu_data[cpu_num] = NULL; 756 } 757 758 static int intel_pstate_cpu_init(struct cpufreq_policy *policy) 759 { 760 struct cpudata *cpu; 761 int rc; 762 u64 misc_en; 763 764 rc = intel_pstate_init_cpu(policy->cpu); 765 if (rc) 766 return rc; 767 768 cpu = all_cpu_data[policy->cpu]; 769 770 rdmsrl(MSR_IA32_MISC_ENABLE, misc_en); 771 if (misc_en & MSR_IA32_MISC_ENABLE_TURBO_DISABLE || 772 cpu->pstate.max_pstate == cpu->pstate.turbo_pstate) { 773 limits.turbo_disabled = 1; 774 limits.no_turbo = 1; 775 } 776 if (limits.min_perf_pct == 100 && limits.max_perf_pct == 100) 777 policy->policy = CPUFREQ_POLICY_PERFORMANCE; 778 else 779 policy->policy = CPUFREQ_POLICY_POWERSAVE; 780 781 policy->min = cpu->pstate.min_pstate * 100000; 782 policy->max = cpu->pstate.turbo_pstate * 100000; 783 784 /* cpuinfo and default policy values */ 785 policy->cpuinfo.min_freq = cpu->pstate.min_pstate * 100000; 786 policy->cpuinfo.max_freq = cpu->pstate.turbo_pstate * 100000; 787 policy->cpuinfo.transition_latency = CPUFREQ_ETERNAL; 788 cpumask_set_cpu(policy->cpu, policy->cpus); 789 790 return 0; 791 } 792 793 static struct cpufreq_driver intel_pstate_driver = { 794 .flags = CPUFREQ_CONST_LOOPS, 795 .verify = intel_pstate_verify_policy, 796 .setpolicy = intel_pstate_set_policy, 797 .get = intel_pstate_get, 798 .init = intel_pstate_cpu_init, 799 .stop_cpu = intel_pstate_stop_cpu, 800 .name = "intel_pstate", 801 }; 802 803 static int __initdata no_load; 804 805 static int intel_pstate_msrs_not_valid(void) 806 { 807 /* Check that all the msr's we are using are valid. */ 808 u64 aperf, mperf, tmp; 809 810 rdmsrl(MSR_IA32_APERF, aperf); 811 rdmsrl(MSR_IA32_MPERF, mperf); 812 813 if (!pstate_funcs.get_max() || 814 !pstate_funcs.get_min() || 815 !pstate_funcs.get_turbo()) 816 return -ENODEV; 817 818 rdmsrl(MSR_IA32_APERF, tmp); 819 if (!(tmp - aperf)) 820 return -ENODEV; 821 822 rdmsrl(MSR_IA32_MPERF, tmp); 823 if (!(tmp - mperf)) 824 return -ENODEV; 825 826 return 0; 827 } 828 829 static void copy_pid_params(struct pstate_adjust_policy *policy) 830 { 831 pid_params.sample_rate_ms = policy->sample_rate_ms; 832 pid_params.p_gain_pct = policy->p_gain_pct; 833 pid_params.i_gain_pct = policy->i_gain_pct; 834 pid_params.d_gain_pct = policy->d_gain_pct; 835 pid_params.deadband = policy->deadband; 836 pid_params.setpoint = policy->setpoint; 837 } 838 839 static void copy_cpu_funcs(struct pstate_funcs *funcs) 840 { 841 pstate_funcs.get_max = funcs->get_max; 842 pstate_funcs.get_min = funcs->get_min; 843 pstate_funcs.get_turbo = funcs->get_turbo; 844 pstate_funcs.set = funcs->set; 845 pstate_funcs.get_vid = funcs->get_vid; 846 } 847 848 #if IS_ENABLED(CONFIG_ACPI) 849 #include <acpi/processor.h> 850 851 static bool intel_pstate_no_acpi_pss(void) 852 { 853 int i; 854 855 for_each_possible_cpu(i) { 856 acpi_status status; 857 union acpi_object *pss; 858 struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL }; 859 struct acpi_processor *pr = per_cpu(processors, i); 860 861 if (!pr) 862 continue; 863 864 status = acpi_evaluate_object(pr->handle, "_PSS", NULL, &buffer); 865 if (ACPI_FAILURE(status)) 866 continue; 867 868 pss = buffer.pointer; 869 if (pss && pss->type == ACPI_TYPE_PACKAGE) { 870 kfree(pss); 871 return false; 872 } 873 874 kfree(pss); 875 } 876 877 return true; 878 } 879 880 struct hw_vendor_info { 881 u16 valid; 882 char oem_id[ACPI_OEM_ID_SIZE]; 883 char oem_table_id[ACPI_OEM_TABLE_ID_SIZE]; 884 }; 885 886 /* Hardware vendor-specific info that has its own power management modes */ 887 static struct hw_vendor_info vendor_info[] = { 888 {1, "HP ", "ProLiant"}, 889 {0, "", ""}, 890 }; 891 892 static bool intel_pstate_platform_pwr_mgmt_exists(void) 893 { 894 struct acpi_table_header hdr; 895 struct hw_vendor_info *v_info; 896 897 if (acpi_disabled || 898 ACPI_FAILURE(acpi_get_table_header(ACPI_SIG_FADT, 0, &hdr))) 899 return false; 900 901 for (v_info = vendor_info; v_info->valid; v_info++) { 902 if (!strncmp(hdr.oem_id, v_info->oem_id, ACPI_OEM_ID_SIZE) && 903 !strncmp(hdr.oem_table_id, v_info->oem_table_id, ACPI_OEM_TABLE_ID_SIZE) && 904 intel_pstate_no_acpi_pss()) 905 return true; 906 } 907 908 return false; 909 } 910 #else /* CONFIG_ACPI not enabled */ 911 static inline bool intel_pstate_platform_pwr_mgmt_exists(void) { return false; } 912 #endif /* CONFIG_ACPI */ 913 914 static int __init intel_pstate_init(void) 915 { 916 int cpu, rc = 0; 917 const struct x86_cpu_id *id; 918 struct cpu_defaults *cpu_info; 919 920 if (no_load) 921 return -ENODEV; 922 923 id = x86_match_cpu(intel_pstate_cpu_ids); 924 if (!id) 925 return -ENODEV; 926 927 /* 928 * The Intel pstate driver will be ignored if the platform 929 * firmware has its own power management modes. 930 */ 931 if (intel_pstate_platform_pwr_mgmt_exists()) 932 return -ENODEV; 933 934 cpu_info = (struct cpu_defaults *)id->driver_data; 935 936 copy_pid_params(&cpu_info->pid_policy); 937 copy_cpu_funcs(&cpu_info->funcs); 938 939 if (intel_pstate_msrs_not_valid()) 940 return -ENODEV; 941 942 pr_info("Intel P-state driver initializing.\n"); 943 944 all_cpu_data = vzalloc(sizeof(void *) * num_possible_cpus()); 945 if (!all_cpu_data) 946 return -ENOMEM; 947 948 rc = cpufreq_register_driver(&intel_pstate_driver); 949 if (rc) 950 goto out; 951 952 intel_pstate_debug_expose_params(); 953 intel_pstate_sysfs_expose_params(); 954 955 return rc; 956 out: 957 get_online_cpus(); 958 for_each_online_cpu(cpu) { 959 if (all_cpu_data[cpu]) { 960 del_timer_sync(&all_cpu_data[cpu]->timer); 961 kfree(all_cpu_data[cpu]); 962 } 963 } 964 965 put_online_cpus(); 966 vfree(all_cpu_data); 967 return -ENODEV; 968 } 969 device_initcall(intel_pstate_init); 970 971 static int __init intel_pstate_setup(char *str) 972 { 973 if (!str) 974 return -EINVAL; 975 976 if (!strcmp(str, "disable")) 977 no_load = 1; 978 return 0; 979 } 980 early_param("intel_pstate", intel_pstate_setup); 981 982 MODULE_AUTHOR("Dirk Brandewie <dirk.j.brandewie@intel.com>"); 983 MODULE_DESCRIPTION("'intel_pstate' - P state driver Intel Core processors"); 984 MODULE_LICENSE("GPL"); 985