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 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt 14 15 #include <linux/kernel.h> 16 #include <linux/kernel_stat.h> 17 #include <linux/module.h> 18 #include <linux/ktime.h> 19 #include <linux/hrtimer.h> 20 #include <linux/tick.h> 21 #include <linux/slab.h> 22 #include <linux/sched/cpufreq.h> 23 #include <linux/list.h> 24 #include <linux/cpu.h> 25 #include <linux/cpufreq.h> 26 #include <linux/sysfs.h> 27 #include <linux/types.h> 28 #include <linux/fs.h> 29 #include <linux/acpi.h> 30 #include <linux/vmalloc.h> 31 #include <trace/events/power.h> 32 33 #include <asm/div64.h> 34 #include <asm/msr.h> 35 #include <asm/cpu_device_id.h> 36 #include <asm/cpufeature.h> 37 #include <asm/intel-family.h> 38 39 #define INTEL_PSTATE_SAMPLING_INTERVAL (10 * NSEC_PER_MSEC) 40 41 #define INTEL_CPUFREQ_TRANSITION_LATENCY 20000 42 #define INTEL_CPUFREQ_TRANSITION_DELAY 500 43 44 #ifdef CONFIG_ACPI 45 #include <acpi/processor.h> 46 #include <acpi/cppc_acpi.h> 47 #endif 48 49 #define FRAC_BITS 8 50 #define int_tofp(X) ((int64_t)(X) << FRAC_BITS) 51 #define fp_toint(X) ((X) >> FRAC_BITS) 52 53 #define EXT_BITS 6 54 #define EXT_FRAC_BITS (EXT_BITS + FRAC_BITS) 55 #define fp_ext_toint(X) ((X) >> EXT_FRAC_BITS) 56 #define int_ext_tofp(X) ((int64_t)(X) << EXT_FRAC_BITS) 57 58 static inline int32_t mul_fp(int32_t x, int32_t y) 59 { 60 return ((int64_t)x * (int64_t)y) >> FRAC_BITS; 61 } 62 63 static inline int32_t div_fp(s64 x, s64 y) 64 { 65 return div64_s64((int64_t)x << FRAC_BITS, y); 66 } 67 68 static inline int ceiling_fp(int32_t x) 69 { 70 int mask, ret; 71 72 ret = fp_toint(x); 73 mask = (1 << FRAC_BITS) - 1; 74 if (x & mask) 75 ret += 1; 76 return ret; 77 } 78 79 static inline int32_t percent_fp(int percent) 80 { 81 return div_fp(percent, 100); 82 } 83 84 static inline u64 mul_ext_fp(u64 x, u64 y) 85 { 86 return (x * y) >> EXT_FRAC_BITS; 87 } 88 89 static inline u64 div_ext_fp(u64 x, u64 y) 90 { 91 return div64_u64(x << EXT_FRAC_BITS, y); 92 } 93 94 static inline int32_t percent_ext_fp(int percent) 95 { 96 return div_ext_fp(percent, 100); 97 } 98 99 /** 100 * struct sample - Store performance sample 101 * @core_avg_perf: Ratio of APERF/MPERF which is the actual average 102 * performance during last sample period 103 * @busy_scaled: Scaled busy value which is used to calculate next 104 * P state. This can be different than core_avg_perf 105 * to account for cpu idle period 106 * @aperf: Difference of actual performance frequency clock count 107 * read from APERF MSR between last and current sample 108 * @mperf: Difference of maximum performance frequency clock count 109 * read from MPERF MSR between last and current sample 110 * @tsc: Difference of time stamp counter between last and 111 * current sample 112 * @time: Current time from scheduler 113 * 114 * This structure is used in the cpudata structure to store performance sample 115 * data for choosing next P State. 116 */ 117 struct sample { 118 int32_t core_avg_perf; 119 int32_t busy_scaled; 120 u64 aperf; 121 u64 mperf; 122 u64 tsc; 123 u64 time; 124 }; 125 126 /** 127 * struct pstate_data - Store P state data 128 * @current_pstate: Current requested P state 129 * @min_pstate: Min P state possible for this platform 130 * @max_pstate: Max P state possible for this platform 131 * @max_pstate_physical:This is physical Max P state for a processor 132 * This can be higher than the max_pstate which can 133 * be limited by platform thermal design power limits 134 * @scaling: Scaling factor to convert frequency to cpufreq 135 * frequency units 136 * @turbo_pstate: Max Turbo P state possible for this platform 137 * @max_freq: @max_pstate frequency in cpufreq units 138 * @turbo_freq: @turbo_pstate frequency in cpufreq units 139 * 140 * Stores the per cpu model P state limits and current P state. 141 */ 142 struct pstate_data { 143 int current_pstate; 144 int min_pstate; 145 int max_pstate; 146 int max_pstate_physical; 147 int scaling; 148 int turbo_pstate; 149 unsigned int max_freq; 150 unsigned int turbo_freq; 151 }; 152 153 /** 154 * struct vid_data - Stores voltage information data 155 * @min: VID data for this platform corresponding to 156 * the lowest P state 157 * @max: VID data corresponding to the highest P State. 158 * @turbo: VID data for turbo P state 159 * @ratio: Ratio of (vid max - vid min) / 160 * (max P state - Min P State) 161 * 162 * Stores the voltage data for DVFS (Dynamic Voltage and Frequency Scaling) 163 * This data is used in Atom platforms, where in addition to target P state, 164 * the voltage data needs to be specified to select next P State. 165 */ 166 struct vid_data { 167 int min; 168 int max; 169 int turbo; 170 int32_t ratio; 171 }; 172 173 /** 174 * struct global_params - Global parameters, mostly tunable via sysfs. 175 * @no_turbo: Whether or not to use turbo P-states. 176 * @turbo_disabled: Whethet or not turbo P-states are available at all, 177 * based on the MSR_IA32_MISC_ENABLE value and whether or 178 * not the maximum reported turbo P-state is different from 179 * the maximum reported non-turbo one. 180 * @min_perf_pct: Minimum capacity limit in percent of the maximum turbo 181 * P-state capacity. 182 * @max_perf_pct: Maximum capacity limit in percent of the maximum turbo 183 * P-state capacity. 184 */ 185 struct global_params { 186 bool no_turbo; 187 bool turbo_disabled; 188 int max_perf_pct; 189 int min_perf_pct; 190 }; 191 192 /** 193 * struct cpudata - Per CPU instance data storage 194 * @cpu: CPU number for this instance data 195 * @policy: CPUFreq policy value 196 * @update_util: CPUFreq utility callback information 197 * @update_util_set: CPUFreq utility callback is set 198 * @iowait_boost: iowait-related boost fraction 199 * @last_update: Time of the last update. 200 * @pstate: Stores P state limits for this CPU 201 * @vid: Stores VID limits for this CPU 202 * @last_sample_time: Last Sample time 203 * @aperf_mperf_shift: Number of clock cycles after aperf, merf is incremented 204 * This shift is a multiplier to mperf delta to 205 * calculate CPU busy. 206 * @prev_aperf: Last APERF value read from APERF MSR 207 * @prev_mperf: Last MPERF value read from MPERF MSR 208 * @prev_tsc: Last timestamp counter (TSC) value 209 * @prev_cummulative_iowait: IO Wait time difference from last and 210 * current sample 211 * @sample: Storage for storing last Sample data 212 * @min_perf_ratio: Minimum capacity in terms of PERF or HWP ratios 213 * @max_perf_ratio: Maximum capacity in terms of PERF or HWP ratios 214 * @acpi_perf_data: Stores ACPI perf information read from _PSS 215 * @valid_pss_table: Set to true for valid ACPI _PSS entries found 216 * @epp_powersave: Last saved HWP energy performance preference 217 * (EPP) or energy performance bias (EPB), 218 * when policy switched to performance 219 * @epp_policy: Last saved policy used to set EPP/EPB 220 * @epp_default: Power on default HWP energy performance 221 * preference/bias 222 * @epp_saved: Saved EPP/EPB during system suspend or CPU offline 223 * operation 224 * @hwp_req_cached: Cached value of the last HWP Request MSR 225 * @hwp_cap_cached: Cached value of the last HWP Capabilities MSR 226 * @last_io_update: Last time when IO wake flag was set 227 * @sched_flags: Store scheduler flags for possible cross CPU update 228 * @hwp_boost_min: Last HWP boosted min performance 229 * 230 * This structure stores per CPU instance data for all CPUs. 231 */ 232 struct cpudata { 233 int cpu; 234 235 unsigned int policy; 236 struct update_util_data update_util; 237 bool update_util_set; 238 239 struct pstate_data pstate; 240 struct vid_data vid; 241 242 u64 last_update; 243 u64 last_sample_time; 244 u64 aperf_mperf_shift; 245 u64 prev_aperf; 246 u64 prev_mperf; 247 u64 prev_tsc; 248 u64 prev_cummulative_iowait; 249 struct sample sample; 250 int32_t min_perf_ratio; 251 int32_t max_perf_ratio; 252 #ifdef CONFIG_ACPI 253 struct acpi_processor_performance acpi_perf_data; 254 bool valid_pss_table; 255 #endif 256 unsigned int iowait_boost; 257 s16 epp_powersave; 258 s16 epp_policy; 259 s16 epp_default; 260 s16 epp_saved; 261 u64 hwp_req_cached; 262 u64 hwp_cap_cached; 263 u64 last_io_update; 264 unsigned int sched_flags; 265 u32 hwp_boost_min; 266 }; 267 268 static struct cpudata **all_cpu_data; 269 270 /** 271 * struct pstate_funcs - Per CPU model specific callbacks 272 * @get_max: Callback to get maximum non turbo effective P state 273 * @get_max_physical: Callback to get maximum non turbo physical P state 274 * @get_min: Callback to get minimum P state 275 * @get_turbo: Callback to get turbo P state 276 * @get_scaling: Callback to get frequency scaling factor 277 * @get_val: Callback to convert P state to actual MSR write value 278 * @get_vid: Callback to get VID data for Atom platforms 279 * 280 * Core and Atom CPU models have different way to get P State limits. This 281 * structure is used to store those callbacks. 282 */ 283 struct pstate_funcs { 284 int (*get_max)(void); 285 int (*get_max_physical)(void); 286 int (*get_min)(void); 287 int (*get_turbo)(void); 288 int (*get_scaling)(void); 289 int (*get_aperf_mperf_shift)(void); 290 u64 (*get_val)(struct cpudata*, int pstate); 291 void (*get_vid)(struct cpudata *); 292 }; 293 294 static struct pstate_funcs pstate_funcs __read_mostly; 295 296 static int hwp_active __read_mostly; 297 static int hwp_mode_bdw __read_mostly; 298 static bool per_cpu_limits __read_mostly; 299 static bool hwp_boost __read_mostly; 300 301 static struct cpufreq_driver *intel_pstate_driver __read_mostly; 302 303 #ifdef CONFIG_ACPI 304 static bool acpi_ppc; 305 #endif 306 307 static struct global_params global; 308 309 static DEFINE_MUTEX(intel_pstate_driver_lock); 310 static DEFINE_MUTEX(intel_pstate_limits_lock); 311 312 #ifdef CONFIG_ACPI 313 314 static bool intel_pstate_get_ppc_enable_status(void) 315 { 316 if (acpi_gbl_FADT.preferred_profile == PM_ENTERPRISE_SERVER || 317 acpi_gbl_FADT.preferred_profile == PM_PERFORMANCE_SERVER) 318 return true; 319 320 return acpi_ppc; 321 } 322 323 #ifdef CONFIG_ACPI_CPPC_LIB 324 325 /* The work item is needed to avoid CPU hotplug locking issues */ 326 static void intel_pstste_sched_itmt_work_fn(struct work_struct *work) 327 { 328 sched_set_itmt_support(); 329 } 330 331 static DECLARE_WORK(sched_itmt_work, intel_pstste_sched_itmt_work_fn); 332 333 static void intel_pstate_set_itmt_prio(int cpu) 334 { 335 struct cppc_perf_caps cppc_perf; 336 static u32 max_highest_perf = 0, min_highest_perf = U32_MAX; 337 int ret; 338 339 ret = cppc_get_perf_caps(cpu, &cppc_perf); 340 if (ret) 341 return; 342 343 /* 344 * The priorities can be set regardless of whether or not 345 * sched_set_itmt_support(true) has been called and it is valid to 346 * update them at any time after it has been called. 347 */ 348 sched_set_itmt_core_prio(cppc_perf.highest_perf, cpu); 349 350 if (max_highest_perf <= min_highest_perf) { 351 if (cppc_perf.highest_perf > max_highest_perf) 352 max_highest_perf = cppc_perf.highest_perf; 353 354 if (cppc_perf.highest_perf < min_highest_perf) 355 min_highest_perf = cppc_perf.highest_perf; 356 357 if (max_highest_perf > min_highest_perf) { 358 /* 359 * This code can be run during CPU online under the 360 * CPU hotplug locks, so sched_set_itmt_support() 361 * cannot be called from here. Queue up a work item 362 * to invoke it. 363 */ 364 schedule_work(&sched_itmt_work); 365 } 366 } 367 } 368 #else 369 static void intel_pstate_set_itmt_prio(int cpu) 370 { 371 } 372 #endif 373 374 static void intel_pstate_init_acpi_perf_limits(struct cpufreq_policy *policy) 375 { 376 struct cpudata *cpu; 377 int ret; 378 int i; 379 380 if (hwp_active) { 381 intel_pstate_set_itmt_prio(policy->cpu); 382 return; 383 } 384 385 if (!intel_pstate_get_ppc_enable_status()) 386 return; 387 388 cpu = all_cpu_data[policy->cpu]; 389 390 ret = acpi_processor_register_performance(&cpu->acpi_perf_data, 391 policy->cpu); 392 if (ret) 393 return; 394 395 /* 396 * Check if the control value in _PSS is for PERF_CTL MSR, which should 397 * guarantee that the states returned by it map to the states in our 398 * list directly. 399 */ 400 if (cpu->acpi_perf_data.control_register.space_id != 401 ACPI_ADR_SPACE_FIXED_HARDWARE) 402 goto err; 403 404 /* 405 * If there is only one entry _PSS, simply ignore _PSS and continue as 406 * usual without taking _PSS into account 407 */ 408 if (cpu->acpi_perf_data.state_count < 2) 409 goto err; 410 411 pr_debug("CPU%u - ACPI _PSS perf data\n", policy->cpu); 412 for (i = 0; i < cpu->acpi_perf_data.state_count; i++) { 413 pr_debug(" %cP%d: %u MHz, %u mW, 0x%x\n", 414 (i == cpu->acpi_perf_data.state ? '*' : ' '), i, 415 (u32) cpu->acpi_perf_data.states[i].core_frequency, 416 (u32) cpu->acpi_perf_data.states[i].power, 417 (u32) cpu->acpi_perf_data.states[i].control); 418 } 419 420 /* 421 * The _PSS table doesn't contain whole turbo frequency range. 422 * This just contains +1 MHZ above the max non turbo frequency, 423 * with control value corresponding to max turbo ratio. But 424 * when cpufreq set policy is called, it will call with this 425 * max frequency, which will cause a reduced performance as 426 * this driver uses real max turbo frequency as the max 427 * frequency. So correct this frequency in _PSS table to 428 * correct max turbo frequency based on the turbo state. 429 * Also need to convert to MHz as _PSS freq is in MHz. 430 */ 431 if (!global.turbo_disabled) 432 cpu->acpi_perf_data.states[0].core_frequency = 433 policy->cpuinfo.max_freq / 1000; 434 cpu->valid_pss_table = true; 435 pr_debug("_PPC limits will be enforced\n"); 436 437 return; 438 439 err: 440 cpu->valid_pss_table = false; 441 acpi_processor_unregister_performance(policy->cpu); 442 } 443 444 static void intel_pstate_exit_perf_limits(struct cpufreq_policy *policy) 445 { 446 struct cpudata *cpu; 447 448 cpu = all_cpu_data[policy->cpu]; 449 if (!cpu->valid_pss_table) 450 return; 451 452 acpi_processor_unregister_performance(policy->cpu); 453 } 454 #else 455 static inline void intel_pstate_init_acpi_perf_limits(struct cpufreq_policy *policy) 456 { 457 } 458 459 static inline void intel_pstate_exit_perf_limits(struct cpufreq_policy *policy) 460 { 461 } 462 #endif 463 464 static inline void update_turbo_state(void) 465 { 466 u64 misc_en; 467 struct cpudata *cpu; 468 469 cpu = all_cpu_data[0]; 470 rdmsrl(MSR_IA32_MISC_ENABLE, misc_en); 471 global.turbo_disabled = 472 (misc_en & MSR_IA32_MISC_ENABLE_TURBO_DISABLE || 473 cpu->pstate.max_pstate == cpu->pstate.turbo_pstate); 474 } 475 476 static int min_perf_pct_min(void) 477 { 478 struct cpudata *cpu = all_cpu_data[0]; 479 int turbo_pstate = cpu->pstate.turbo_pstate; 480 481 return turbo_pstate ? 482 (cpu->pstate.min_pstate * 100 / turbo_pstate) : 0; 483 } 484 485 static s16 intel_pstate_get_epb(struct cpudata *cpu_data) 486 { 487 u64 epb; 488 int ret; 489 490 if (!static_cpu_has(X86_FEATURE_EPB)) 491 return -ENXIO; 492 493 ret = rdmsrl_on_cpu(cpu_data->cpu, MSR_IA32_ENERGY_PERF_BIAS, &epb); 494 if (ret) 495 return (s16)ret; 496 497 return (s16)(epb & 0x0f); 498 } 499 500 static s16 intel_pstate_get_epp(struct cpudata *cpu_data, u64 hwp_req_data) 501 { 502 s16 epp; 503 504 if (static_cpu_has(X86_FEATURE_HWP_EPP)) { 505 /* 506 * When hwp_req_data is 0, means that caller didn't read 507 * MSR_HWP_REQUEST, so need to read and get EPP. 508 */ 509 if (!hwp_req_data) { 510 epp = rdmsrl_on_cpu(cpu_data->cpu, MSR_HWP_REQUEST, 511 &hwp_req_data); 512 if (epp) 513 return epp; 514 } 515 epp = (hwp_req_data >> 24) & 0xff; 516 } else { 517 /* When there is no EPP present, HWP uses EPB settings */ 518 epp = intel_pstate_get_epb(cpu_data); 519 } 520 521 return epp; 522 } 523 524 static int intel_pstate_set_epb(int cpu, s16 pref) 525 { 526 u64 epb; 527 int ret; 528 529 if (!static_cpu_has(X86_FEATURE_EPB)) 530 return -ENXIO; 531 532 ret = rdmsrl_on_cpu(cpu, MSR_IA32_ENERGY_PERF_BIAS, &epb); 533 if (ret) 534 return ret; 535 536 epb = (epb & ~0x0f) | pref; 537 wrmsrl_on_cpu(cpu, MSR_IA32_ENERGY_PERF_BIAS, epb); 538 539 return 0; 540 } 541 542 /* 543 * EPP/EPB display strings corresponding to EPP index in the 544 * energy_perf_strings[] 545 * index String 546 *------------------------------------- 547 * 0 default 548 * 1 performance 549 * 2 balance_performance 550 * 3 balance_power 551 * 4 power 552 */ 553 static const char * const energy_perf_strings[] = { 554 "default", 555 "performance", 556 "balance_performance", 557 "balance_power", 558 "power", 559 NULL 560 }; 561 static const unsigned int epp_values[] = { 562 HWP_EPP_PERFORMANCE, 563 HWP_EPP_BALANCE_PERFORMANCE, 564 HWP_EPP_BALANCE_POWERSAVE, 565 HWP_EPP_POWERSAVE 566 }; 567 568 static int intel_pstate_get_energy_pref_index(struct cpudata *cpu_data) 569 { 570 s16 epp; 571 int index = -EINVAL; 572 573 epp = intel_pstate_get_epp(cpu_data, 0); 574 if (epp < 0) 575 return epp; 576 577 if (static_cpu_has(X86_FEATURE_HWP_EPP)) { 578 if (epp == HWP_EPP_PERFORMANCE) 579 return 1; 580 if (epp <= HWP_EPP_BALANCE_PERFORMANCE) 581 return 2; 582 if (epp <= HWP_EPP_BALANCE_POWERSAVE) 583 return 3; 584 else 585 return 4; 586 } else if (static_cpu_has(X86_FEATURE_EPB)) { 587 /* 588 * Range: 589 * 0x00-0x03 : Performance 590 * 0x04-0x07 : Balance performance 591 * 0x08-0x0B : Balance power 592 * 0x0C-0x0F : Power 593 * The EPB is a 4 bit value, but our ranges restrict the 594 * value which can be set. Here only using top two bits 595 * effectively. 596 */ 597 index = (epp >> 2) + 1; 598 } 599 600 return index; 601 } 602 603 static int intel_pstate_set_energy_pref_index(struct cpudata *cpu_data, 604 int pref_index) 605 { 606 int epp = -EINVAL; 607 int ret; 608 609 if (!pref_index) 610 epp = cpu_data->epp_default; 611 612 mutex_lock(&intel_pstate_limits_lock); 613 614 if (static_cpu_has(X86_FEATURE_HWP_EPP)) { 615 u64 value; 616 617 ret = rdmsrl_on_cpu(cpu_data->cpu, MSR_HWP_REQUEST, &value); 618 if (ret) 619 goto return_pref; 620 621 value &= ~GENMASK_ULL(31, 24); 622 623 if (epp == -EINVAL) 624 epp = epp_values[pref_index - 1]; 625 626 value |= (u64)epp << 24; 627 ret = wrmsrl_on_cpu(cpu_data->cpu, MSR_HWP_REQUEST, value); 628 } else { 629 if (epp == -EINVAL) 630 epp = (pref_index - 1) << 2; 631 ret = intel_pstate_set_epb(cpu_data->cpu, epp); 632 } 633 return_pref: 634 mutex_unlock(&intel_pstate_limits_lock); 635 636 return ret; 637 } 638 639 static ssize_t show_energy_performance_available_preferences( 640 struct cpufreq_policy *policy, char *buf) 641 { 642 int i = 0; 643 int ret = 0; 644 645 while (energy_perf_strings[i] != NULL) 646 ret += sprintf(&buf[ret], "%s ", energy_perf_strings[i++]); 647 648 ret += sprintf(&buf[ret], "\n"); 649 650 return ret; 651 } 652 653 cpufreq_freq_attr_ro(energy_performance_available_preferences); 654 655 static ssize_t store_energy_performance_preference( 656 struct cpufreq_policy *policy, const char *buf, size_t count) 657 { 658 struct cpudata *cpu_data = all_cpu_data[policy->cpu]; 659 char str_preference[21]; 660 int ret, i = 0; 661 662 ret = sscanf(buf, "%20s", str_preference); 663 if (ret != 1) 664 return -EINVAL; 665 666 while (energy_perf_strings[i] != NULL) { 667 if (!strcmp(str_preference, energy_perf_strings[i])) { 668 intel_pstate_set_energy_pref_index(cpu_data, i); 669 return count; 670 } 671 ++i; 672 } 673 674 return -EINVAL; 675 } 676 677 static ssize_t show_energy_performance_preference( 678 struct cpufreq_policy *policy, char *buf) 679 { 680 struct cpudata *cpu_data = all_cpu_data[policy->cpu]; 681 int preference; 682 683 preference = intel_pstate_get_energy_pref_index(cpu_data); 684 if (preference < 0) 685 return preference; 686 687 return sprintf(buf, "%s\n", energy_perf_strings[preference]); 688 } 689 690 cpufreq_freq_attr_rw(energy_performance_preference); 691 692 static struct freq_attr *hwp_cpufreq_attrs[] = { 693 &energy_performance_preference, 694 &energy_performance_available_preferences, 695 NULL, 696 }; 697 698 static void intel_pstate_get_hwp_max(unsigned int cpu, int *phy_max, 699 int *current_max) 700 { 701 u64 cap; 702 703 rdmsrl_on_cpu(cpu, MSR_HWP_CAPABILITIES, &cap); 704 WRITE_ONCE(all_cpu_data[cpu]->hwp_cap_cached, cap); 705 if (global.no_turbo) 706 *current_max = HWP_GUARANTEED_PERF(cap); 707 else 708 *current_max = HWP_HIGHEST_PERF(cap); 709 710 *phy_max = HWP_HIGHEST_PERF(cap); 711 } 712 713 static void intel_pstate_hwp_set(unsigned int cpu) 714 { 715 struct cpudata *cpu_data = all_cpu_data[cpu]; 716 int max, min; 717 u64 value; 718 s16 epp; 719 720 max = cpu_data->max_perf_ratio; 721 min = cpu_data->min_perf_ratio; 722 723 if (cpu_data->policy == CPUFREQ_POLICY_PERFORMANCE) 724 min = max; 725 726 rdmsrl_on_cpu(cpu, MSR_HWP_REQUEST, &value); 727 728 value &= ~HWP_MIN_PERF(~0L); 729 value |= HWP_MIN_PERF(min); 730 731 value &= ~HWP_MAX_PERF(~0L); 732 value |= HWP_MAX_PERF(max); 733 734 if (cpu_data->epp_policy == cpu_data->policy) 735 goto skip_epp; 736 737 cpu_data->epp_policy = cpu_data->policy; 738 739 if (cpu_data->epp_saved >= 0) { 740 epp = cpu_data->epp_saved; 741 cpu_data->epp_saved = -EINVAL; 742 goto update_epp; 743 } 744 745 if (cpu_data->policy == CPUFREQ_POLICY_PERFORMANCE) { 746 epp = intel_pstate_get_epp(cpu_data, value); 747 cpu_data->epp_powersave = epp; 748 /* If EPP read was failed, then don't try to write */ 749 if (epp < 0) 750 goto skip_epp; 751 752 epp = 0; 753 } else { 754 /* skip setting EPP, when saved value is invalid */ 755 if (cpu_data->epp_powersave < 0) 756 goto skip_epp; 757 758 /* 759 * No need to restore EPP when it is not zero. This 760 * means: 761 * - Policy is not changed 762 * - user has manually changed 763 * - Error reading EPB 764 */ 765 epp = intel_pstate_get_epp(cpu_data, value); 766 if (epp) 767 goto skip_epp; 768 769 epp = cpu_data->epp_powersave; 770 } 771 update_epp: 772 if (static_cpu_has(X86_FEATURE_HWP_EPP)) { 773 value &= ~GENMASK_ULL(31, 24); 774 value |= (u64)epp << 24; 775 } else { 776 intel_pstate_set_epb(cpu, epp); 777 } 778 skip_epp: 779 WRITE_ONCE(cpu_data->hwp_req_cached, value); 780 wrmsrl_on_cpu(cpu, MSR_HWP_REQUEST, value); 781 } 782 783 static int intel_pstate_hwp_save_state(struct cpufreq_policy *policy) 784 { 785 struct cpudata *cpu_data = all_cpu_data[policy->cpu]; 786 787 if (!hwp_active) 788 return 0; 789 790 cpu_data->epp_saved = intel_pstate_get_epp(cpu_data, 0); 791 792 return 0; 793 } 794 795 static void intel_pstate_hwp_enable(struct cpudata *cpudata); 796 797 static int intel_pstate_resume(struct cpufreq_policy *policy) 798 { 799 if (!hwp_active) 800 return 0; 801 802 mutex_lock(&intel_pstate_limits_lock); 803 804 if (policy->cpu == 0) 805 intel_pstate_hwp_enable(all_cpu_data[policy->cpu]); 806 807 all_cpu_data[policy->cpu]->epp_policy = 0; 808 intel_pstate_hwp_set(policy->cpu); 809 810 mutex_unlock(&intel_pstate_limits_lock); 811 812 return 0; 813 } 814 815 static void intel_pstate_update_policies(void) 816 { 817 int cpu; 818 819 for_each_possible_cpu(cpu) 820 cpufreq_update_policy(cpu); 821 } 822 823 /************************** sysfs begin ************************/ 824 #define show_one(file_name, object) \ 825 static ssize_t show_##file_name \ 826 (struct kobject *kobj, struct attribute *attr, char *buf) \ 827 { \ 828 return sprintf(buf, "%u\n", global.object); \ 829 } 830 831 static ssize_t intel_pstate_show_status(char *buf); 832 static int intel_pstate_update_status(const char *buf, size_t size); 833 834 static ssize_t show_status(struct kobject *kobj, 835 struct attribute *attr, char *buf) 836 { 837 ssize_t ret; 838 839 mutex_lock(&intel_pstate_driver_lock); 840 ret = intel_pstate_show_status(buf); 841 mutex_unlock(&intel_pstate_driver_lock); 842 843 return ret; 844 } 845 846 static ssize_t store_status(struct kobject *a, struct attribute *b, 847 const char *buf, size_t count) 848 { 849 char *p = memchr(buf, '\n', count); 850 int ret; 851 852 mutex_lock(&intel_pstate_driver_lock); 853 ret = intel_pstate_update_status(buf, p ? p - buf : count); 854 mutex_unlock(&intel_pstate_driver_lock); 855 856 return ret < 0 ? ret : count; 857 } 858 859 static ssize_t show_turbo_pct(struct kobject *kobj, 860 struct attribute *attr, char *buf) 861 { 862 struct cpudata *cpu; 863 int total, no_turbo, turbo_pct; 864 uint32_t turbo_fp; 865 866 mutex_lock(&intel_pstate_driver_lock); 867 868 if (!intel_pstate_driver) { 869 mutex_unlock(&intel_pstate_driver_lock); 870 return -EAGAIN; 871 } 872 873 cpu = all_cpu_data[0]; 874 875 total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1; 876 no_turbo = cpu->pstate.max_pstate - cpu->pstate.min_pstate + 1; 877 turbo_fp = div_fp(no_turbo, total); 878 turbo_pct = 100 - fp_toint(mul_fp(turbo_fp, int_tofp(100))); 879 880 mutex_unlock(&intel_pstate_driver_lock); 881 882 return sprintf(buf, "%u\n", turbo_pct); 883 } 884 885 static ssize_t show_num_pstates(struct kobject *kobj, 886 struct attribute *attr, char *buf) 887 { 888 struct cpudata *cpu; 889 int total; 890 891 mutex_lock(&intel_pstate_driver_lock); 892 893 if (!intel_pstate_driver) { 894 mutex_unlock(&intel_pstate_driver_lock); 895 return -EAGAIN; 896 } 897 898 cpu = all_cpu_data[0]; 899 total = cpu->pstate.turbo_pstate - cpu->pstate.min_pstate + 1; 900 901 mutex_unlock(&intel_pstate_driver_lock); 902 903 return sprintf(buf, "%u\n", total); 904 } 905 906 static ssize_t show_no_turbo(struct kobject *kobj, 907 struct attribute *attr, char *buf) 908 { 909 ssize_t ret; 910 911 mutex_lock(&intel_pstate_driver_lock); 912 913 if (!intel_pstate_driver) { 914 mutex_unlock(&intel_pstate_driver_lock); 915 return -EAGAIN; 916 } 917 918 update_turbo_state(); 919 if (global.turbo_disabled) 920 ret = sprintf(buf, "%u\n", global.turbo_disabled); 921 else 922 ret = sprintf(buf, "%u\n", global.no_turbo); 923 924 mutex_unlock(&intel_pstate_driver_lock); 925 926 return ret; 927 } 928 929 static ssize_t store_no_turbo(struct kobject *a, struct attribute *b, 930 const char *buf, size_t count) 931 { 932 unsigned int input; 933 int ret; 934 935 ret = sscanf(buf, "%u", &input); 936 if (ret != 1) 937 return -EINVAL; 938 939 mutex_lock(&intel_pstate_driver_lock); 940 941 if (!intel_pstate_driver) { 942 mutex_unlock(&intel_pstate_driver_lock); 943 return -EAGAIN; 944 } 945 946 mutex_lock(&intel_pstate_limits_lock); 947 948 update_turbo_state(); 949 if (global.turbo_disabled) { 950 pr_warn("Turbo disabled by BIOS or unavailable on processor\n"); 951 mutex_unlock(&intel_pstate_limits_lock); 952 mutex_unlock(&intel_pstate_driver_lock); 953 return -EPERM; 954 } 955 956 global.no_turbo = clamp_t(int, input, 0, 1); 957 958 if (global.no_turbo) { 959 struct cpudata *cpu = all_cpu_data[0]; 960 int pct = cpu->pstate.max_pstate * 100 / cpu->pstate.turbo_pstate; 961 962 /* Squash the global minimum into the permitted range. */ 963 if (global.min_perf_pct > pct) 964 global.min_perf_pct = pct; 965 } 966 967 mutex_unlock(&intel_pstate_limits_lock); 968 969 intel_pstate_update_policies(); 970 971 mutex_unlock(&intel_pstate_driver_lock); 972 973 return count; 974 } 975 976 static ssize_t store_max_perf_pct(struct kobject *a, struct attribute *b, 977 const char *buf, size_t count) 978 { 979 unsigned int input; 980 int ret; 981 982 ret = sscanf(buf, "%u", &input); 983 if (ret != 1) 984 return -EINVAL; 985 986 mutex_lock(&intel_pstate_driver_lock); 987 988 if (!intel_pstate_driver) { 989 mutex_unlock(&intel_pstate_driver_lock); 990 return -EAGAIN; 991 } 992 993 mutex_lock(&intel_pstate_limits_lock); 994 995 global.max_perf_pct = clamp_t(int, input, global.min_perf_pct, 100); 996 997 mutex_unlock(&intel_pstate_limits_lock); 998 999 intel_pstate_update_policies(); 1000 1001 mutex_unlock(&intel_pstate_driver_lock); 1002 1003 return count; 1004 } 1005 1006 static ssize_t store_min_perf_pct(struct kobject *a, struct attribute *b, 1007 const char *buf, size_t count) 1008 { 1009 unsigned int input; 1010 int ret; 1011 1012 ret = sscanf(buf, "%u", &input); 1013 if (ret != 1) 1014 return -EINVAL; 1015 1016 mutex_lock(&intel_pstate_driver_lock); 1017 1018 if (!intel_pstate_driver) { 1019 mutex_unlock(&intel_pstate_driver_lock); 1020 return -EAGAIN; 1021 } 1022 1023 mutex_lock(&intel_pstate_limits_lock); 1024 1025 global.min_perf_pct = clamp_t(int, input, 1026 min_perf_pct_min(), global.max_perf_pct); 1027 1028 mutex_unlock(&intel_pstate_limits_lock); 1029 1030 intel_pstate_update_policies(); 1031 1032 mutex_unlock(&intel_pstate_driver_lock); 1033 1034 return count; 1035 } 1036 1037 static ssize_t show_hwp_dynamic_boost(struct kobject *kobj, 1038 struct attribute *attr, char *buf) 1039 { 1040 return sprintf(buf, "%u\n", hwp_boost); 1041 } 1042 1043 static ssize_t store_hwp_dynamic_boost(struct kobject *a, struct attribute *b, 1044 const char *buf, size_t count) 1045 { 1046 unsigned int input; 1047 int ret; 1048 1049 ret = kstrtouint(buf, 10, &input); 1050 if (ret) 1051 return ret; 1052 1053 mutex_lock(&intel_pstate_driver_lock); 1054 hwp_boost = !!input; 1055 intel_pstate_update_policies(); 1056 mutex_unlock(&intel_pstate_driver_lock); 1057 1058 return count; 1059 } 1060 1061 show_one(max_perf_pct, max_perf_pct); 1062 show_one(min_perf_pct, min_perf_pct); 1063 1064 define_one_global_rw(status); 1065 define_one_global_rw(no_turbo); 1066 define_one_global_rw(max_perf_pct); 1067 define_one_global_rw(min_perf_pct); 1068 define_one_global_ro(turbo_pct); 1069 define_one_global_ro(num_pstates); 1070 define_one_global_rw(hwp_dynamic_boost); 1071 1072 static struct attribute *intel_pstate_attributes[] = { 1073 &status.attr, 1074 &no_turbo.attr, 1075 &turbo_pct.attr, 1076 &num_pstates.attr, 1077 NULL 1078 }; 1079 1080 static const struct attribute_group intel_pstate_attr_group = { 1081 .attrs = intel_pstate_attributes, 1082 }; 1083 1084 static void __init intel_pstate_sysfs_expose_params(void) 1085 { 1086 struct kobject *intel_pstate_kobject; 1087 int rc; 1088 1089 intel_pstate_kobject = kobject_create_and_add("intel_pstate", 1090 &cpu_subsys.dev_root->kobj); 1091 if (WARN_ON(!intel_pstate_kobject)) 1092 return; 1093 1094 rc = sysfs_create_group(intel_pstate_kobject, &intel_pstate_attr_group); 1095 if (WARN_ON(rc)) 1096 return; 1097 1098 /* 1099 * If per cpu limits are enforced there are no global limits, so 1100 * return without creating max/min_perf_pct attributes 1101 */ 1102 if (per_cpu_limits) 1103 return; 1104 1105 rc = sysfs_create_file(intel_pstate_kobject, &max_perf_pct.attr); 1106 WARN_ON(rc); 1107 1108 rc = sysfs_create_file(intel_pstate_kobject, &min_perf_pct.attr); 1109 WARN_ON(rc); 1110 1111 if (hwp_active) { 1112 rc = sysfs_create_file(intel_pstate_kobject, 1113 &hwp_dynamic_boost.attr); 1114 WARN_ON(rc); 1115 } 1116 } 1117 /************************** sysfs end ************************/ 1118 1119 static void intel_pstate_hwp_enable(struct cpudata *cpudata) 1120 { 1121 /* First disable HWP notification interrupt as we don't process them */ 1122 if (static_cpu_has(X86_FEATURE_HWP_NOTIFY)) 1123 wrmsrl_on_cpu(cpudata->cpu, MSR_HWP_INTERRUPT, 0x00); 1124 1125 wrmsrl_on_cpu(cpudata->cpu, MSR_PM_ENABLE, 0x1); 1126 cpudata->epp_policy = 0; 1127 if (cpudata->epp_default == -EINVAL) 1128 cpudata->epp_default = intel_pstate_get_epp(cpudata, 0); 1129 } 1130 1131 #define MSR_IA32_POWER_CTL_BIT_EE 19 1132 1133 /* Disable energy efficiency optimization */ 1134 static void intel_pstate_disable_ee(int cpu) 1135 { 1136 u64 power_ctl; 1137 int ret; 1138 1139 ret = rdmsrl_on_cpu(cpu, MSR_IA32_POWER_CTL, &power_ctl); 1140 if (ret) 1141 return; 1142 1143 if (!(power_ctl & BIT(MSR_IA32_POWER_CTL_BIT_EE))) { 1144 pr_info("Disabling energy efficiency optimization\n"); 1145 power_ctl |= BIT(MSR_IA32_POWER_CTL_BIT_EE); 1146 wrmsrl_on_cpu(cpu, MSR_IA32_POWER_CTL, power_ctl); 1147 } 1148 } 1149 1150 static int atom_get_min_pstate(void) 1151 { 1152 u64 value; 1153 1154 rdmsrl(MSR_ATOM_CORE_RATIOS, value); 1155 return (value >> 8) & 0x7F; 1156 } 1157 1158 static int atom_get_max_pstate(void) 1159 { 1160 u64 value; 1161 1162 rdmsrl(MSR_ATOM_CORE_RATIOS, value); 1163 return (value >> 16) & 0x7F; 1164 } 1165 1166 static int atom_get_turbo_pstate(void) 1167 { 1168 u64 value; 1169 1170 rdmsrl(MSR_ATOM_CORE_TURBO_RATIOS, value); 1171 return value & 0x7F; 1172 } 1173 1174 static u64 atom_get_val(struct cpudata *cpudata, int pstate) 1175 { 1176 u64 val; 1177 int32_t vid_fp; 1178 u32 vid; 1179 1180 val = (u64)pstate << 8; 1181 if (global.no_turbo && !global.turbo_disabled) 1182 val |= (u64)1 << 32; 1183 1184 vid_fp = cpudata->vid.min + mul_fp( 1185 int_tofp(pstate - cpudata->pstate.min_pstate), 1186 cpudata->vid.ratio); 1187 1188 vid_fp = clamp_t(int32_t, vid_fp, cpudata->vid.min, cpudata->vid.max); 1189 vid = ceiling_fp(vid_fp); 1190 1191 if (pstate > cpudata->pstate.max_pstate) 1192 vid = cpudata->vid.turbo; 1193 1194 return val | vid; 1195 } 1196 1197 static int silvermont_get_scaling(void) 1198 { 1199 u64 value; 1200 int i; 1201 /* Defined in Table 35-6 from SDM (Sept 2015) */ 1202 static int silvermont_freq_table[] = { 1203 83300, 100000, 133300, 116700, 80000}; 1204 1205 rdmsrl(MSR_FSB_FREQ, value); 1206 i = value & 0x7; 1207 WARN_ON(i > 4); 1208 1209 return silvermont_freq_table[i]; 1210 } 1211 1212 static int airmont_get_scaling(void) 1213 { 1214 u64 value; 1215 int i; 1216 /* Defined in Table 35-10 from SDM (Sept 2015) */ 1217 static int airmont_freq_table[] = { 1218 83300, 100000, 133300, 116700, 80000, 1219 93300, 90000, 88900, 87500}; 1220 1221 rdmsrl(MSR_FSB_FREQ, value); 1222 i = value & 0xF; 1223 WARN_ON(i > 8); 1224 1225 return airmont_freq_table[i]; 1226 } 1227 1228 static void atom_get_vid(struct cpudata *cpudata) 1229 { 1230 u64 value; 1231 1232 rdmsrl(MSR_ATOM_CORE_VIDS, value); 1233 cpudata->vid.min = int_tofp((value >> 8) & 0x7f); 1234 cpudata->vid.max = int_tofp((value >> 16) & 0x7f); 1235 cpudata->vid.ratio = div_fp( 1236 cpudata->vid.max - cpudata->vid.min, 1237 int_tofp(cpudata->pstate.max_pstate - 1238 cpudata->pstate.min_pstate)); 1239 1240 rdmsrl(MSR_ATOM_CORE_TURBO_VIDS, value); 1241 cpudata->vid.turbo = value & 0x7f; 1242 } 1243 1244 static int core_get_min_pstate(void) 1245 { 1246 u64 value; 1247 1248 rdmsrl(MSR_PLATFORM_INFO, value); 1249 return (value >> 40) & 0xFF; 1250 } 1251 1252 static int core_get_max_pstate_physical(void) 1253 { 1254 u64 value; 1255 1256 rdmsrl(MSR_PLATFORM_INFO, value); 1257 return (value >> 8) & 0xFF; 1258 } 1259 1260 static int core_get_tdp_ratio(u64 plat_info) 1261 { 1262 /* Check how many TDP levels present */ 1263 if (plat_info & 0x600000000) { 1264 u64 tdp_ctrl; 1265 u64 tdp_ratio; 1266 int tdp_msr; 1267 int err; 1268 1269 /* Get the TDP level (0, 1, 2) to get ratios */ 1270 err = rdmsrl_safe(MSR_CONFIG_TDP_CONTROL, &tdp_ctrl); 1271 if (err) 1272 return err; 1273 1274 /* TDP MSR are continuous starting at 0x648 */ 1275 tdp_msr = MSR_CONFIG_TDP_NOMINAL + (tdp_ctrl & 0x03); 1276 err = rdmsrl_safe(tdp_msr, &tdp_ratio); 1277 if (err) 1278 return err; 1279 1280 /* For level 1 and 2, bits[23:16] contain the ratio */ 1281 if (tdp_ctrl & 0x03) 1282 tdp_ratio >>= 16; 1283 1284 tdp_ratio &= 0xff; /* ratios are only 8 bits long */ 1285 pr_debug("tdp_ratio %x\n", (int)tdp_ratio); 1286 1287 return (int)tdp_ratio; 1288 } 1289 1290 return -ENXIO; 1291 } 1292 1293 static int core_get_max_pstate(void) 1294 { 1295 u64 tar; 1296 u64 plat_info; 1297 int max_pstate; 1298 int tdp_ratio; 1299 int err; 1300 1301 rdmsrl(MSR_PLATFORM_INFO, plat_info); 1302 max_pstate = (plat_info >> 8) & 0xFF; 1303 1304 tdp_ratio = core_get_tdp_ratio(plat_info); 1305 if (tdp_ratio <= 0) 1306 return max_pstate; 1307 1308 if (hwp_active) { 1309 /* Turbo activation ratio is not used on HWP platforms */ 1310 return tdp_ratio; 1311 } 1312 1313 err = rdmsrl_safe(MSR_TURBO_ACTIVATION_RATIO, &tar); 1314 if (!err) { 1315 int tar_levels; 1316 1317 /* Do some sanity checking for safety */ 1318 tar_levels = tar & 0xff; 1319 if (tdp_ratio - 1 == tar_levels) { 1320 max_pstate = tar_levels; 1321 pr_debug("max_pstate=TAC %x\n", max_pstate); 1322 } 1323 } 1324 1325 return max_pstate; 1326 } 1327 1328 static int core_get_turbo_pstate(void) 1329 { 1330 u64 value; 1331 int nont, ret; 1332 1333 rdmsrl(MSR_TURBO_RATIO_LIMIT, value); 1334 nont = core_get_max_pstate(); 1335 ret = (value) & 255; 1336 if (ret <= nont) 1337 ret = nont; 1338 return ret; 1339 } 1340 1341 static inline int core_get_scaling(void) 1342 { 1343 return 100000; 1344 } 1345 1346 static u64 core_get_val(struct cpudata *cpudata, int pstate) 1347 { 1348 u64 val; 1349 1350 val = (u64)pstate << 8; 1351 if (global.no_turbo && !global.turbo_disabled) 1352 val |= (u64)1 << 32; 1353 1354 return val; 1355 } 1356 1357 static int knl_get_aperf_mperf_shift(void) 1358 { 1359 return 10; 1360 } 1361 1362 static int knl_get_turbo_pstate(void) 1363 { 1364 u64 value; 1365 int nont, ret; 1366 1367 rdmsrl(MSR_TURBO_RATIO_LIMIT, value); 1368 nont = core_get_max_pstate(); 1369 ret = (((value) >> 8) & 0xFF); 1370 if (ret <= nont) 1371 ret = nont; 1372 return ret; 1373 } 1374 1375 static int intel_pstate_get_base_pstate(struct cpudata *cpu) 1376 { 1377 return global.no_turbo || global.turbo_disabled ? 1378 cpu->pstate.max_pstate : cpu->pstate.turbo_pstate; 1379 } 1380 1381 static void intel_pstate_set_pstate(struct cpudata *cpu, int pstate) 1382 { 1383 trace_cpu_frequency(pstate * cpu->pstate.scaling, cpu->cpu); 1384 cpu->pstate.current_pstate = pstate; 1385 /* 1386 * Generally, there is no guarantee that this code will always run on 1387 * the CPU being updated, so force the register update to run on the 1388 * right CPU. 1389 */ 1390 wrmsrl_on_cpu(cpu->cpu, MSR_IA32_PERF_CTL, 1391 pstate_funcs.get_val(cpu, pstate)); 1392 } 1393 1394 static void intel_pstate_set_min_pstate(struct cpudata *cpu) 1395 { 1396 intel_pstate_set_pstate(cpu, cpu->pstate.min_pstate); 1397 } 1398 1399 static void intel_pstate_max_within_limits(struct cpudata *cpu) 1400 { 1401 int pstate; 1402 1403 update_turbo_state(); 1404 pstate = intel_pstate_get_base_pstate(cpu); 1405 pstate = max(cpu->pstate.min_pstate, cpu->max_perf_ratio); 1406 intel_pstate_set_pstate(cpu, pstate); 1407 } 1408 1409 static void intel_pstate_get_cpu_pstates(struct cpudata *cpu) 1410 { 1411 cpu->pstate.min_pstate = pstate_funcs.get_min(); 1412 cpu->pstate.max_pstate = pstate_funcs.get_max(); 1413 cpu->pstate.max_pstate_physical = pstate_funcs.get_max_physical(); 1414 cpu->pstate.turbo_pstate = pstate_funcs.get_turbo(); 1415 cpu->pstate.scaling = pstate_funcs.get_scaling(); 1416 cpu->pstate.max_freq = cpu->pstate.max_pstate * cpu->pstate.scaling; 1417 1418 if (hwp_active && !hwp_mode_bdw) { 1419 unsigned int phy_max, current_max; 1420 1421 intel_pstate_get_hwp_max(cpu->cpu, &phy_max, ¤t_max); 1422 cpu->pstate.turbo_freq = phy_max * cpu->pstate.scaling; 1423 } else { 1424 cpu->pstate.turbo_freq = cpu->pstate.turbo_pstate * cpu->pstate.scaling; 1425 } 1426 1427 if (pstate_funcs.get_aperf_mperf_shift) 1428 cpu->aperf_mperf_shift = pstate_funcs.get_aperf_mperf_shift(); 1429 1430 if (pstate_funcs.get_vid) 1431 pstate_funcs.get_vid(cpu); 1432 1433 intel_pstate_set_min_pstate(cpu); 1434 } 1435 1436 /* 1437 * Long hold time will keep high perf limits for long time, 1438 * which negatively impacts perf/watt for some workloads, 1439 * like specpower. 3ms is based on experiements on some 1440 * workoads. 1441 */ 1442 static int hwp_boost_hold_time_ns = 3 * NSEC_PER_MSEC; 1443 1444 static inline void intel_pstate_hwp_boost_up(struct cpudata *cpu) 1445 { 1446 u64 hwp_req = READ_ONCE(cpu->hwp_req_cached); 1447 u32 max_limit = (hwp_req & 0xff00) >> 8; 1448 u32 min_limit = (hwp_req & 0xff); 1449 u32 boost_level1; 1450 1451 /* 1452 * Cases to consider (User changes via sysfs or boot time): 1453 * If, P0 (Turbo max) = P1 (Guaranteed max) = min: 1454 * No boost, return. 1455 * If, P0 (Turbo max) > P1 (Guaranteed max) = min: 1456 * Should result in one level boost only for P0. 1457 * If, P0 (Turbo max) = P1 (Guaranteed max) > min: 1458 * Should result in two level boost: 1459 * (min + p1)/2 and P1. 1460 * If, P0 (Turbo max) > P1 (Guaranteed max) > min: 1461 * Should result in three level boost: 1462 * (min + p1)/2, P1 and P0. 1463 */ 1464 1465 /* If max and min are equal or already at max, nothing to boost */ 1466 if (max_limit == min_limit || cpu->hwp_boost_min >= max_limit) 1467 return; 1468 1469 if (!cpu->hwp_boost_min) 1470 cpu->hwp_boost_min = min_limit; 1471 1472 /* level at half way mark between min and guranteed */ 1473 boost_level1 = (HWP_GUARANTEED_PERF(cpu->hwp_cap_cached) + min_limit) >> 1; 1474 1475 if (cpu->hwp_boost_min < boost_level1) 1476 cpu->hwp_boost_min = boost_level1; 1477 else if (cpu->hwp_boost_min < HWP_GUARANTEED_PERF(cpu->hwp_cap_cached)) 1478 cpu->hwp_boost_min = HWP_GUARANTEED_PERF(cpu->hwp_cap_cached); 1479 else if (cpu->hwp_boost_min == HWP_GUARANTEED_PERF(cpu->hwp_cap_cached) && 1480 max_limit != HWP_GUARANTEED_PERF(cpu->hwp_cap_cached)) 1481 cpu->hwp_boost_min = max_limit; 1482 else 1483 return; 1484 1485 hwp_req = (hwp_req & ~GENMASK_ULL(7, 0)) | cpu->hwp_boost_min; 1486 wrmsrl(MSR_HWP_REQUEST, hwp_req); 1487 cpu->last_update = cpu->sample.time; 1488 } 1489 1490 static inline void intel_pstate_hwp_boost_down(struct cpudata *cpu) 1491 { 1492 if (cpu->hwp_boost_min) { 1493 bool expired; 1494 1495 /* Check if we are idle for hold time to boost down */ 1496 expired = time_after64(cpu->sample.time, cpu->last_update + 1497 hwp_boost_hold_time_ns); 1498 if (expired) { 1499 wrmsrl(MSR_HWP_REQUEST, cpu->hwp_req_cached); 1500 cpu->hwp_boost_min = 0; 1501 } 1502 } 1503 cpu->last_update = cpu->sample.time; 1504 } 1505 1506 static inline void intel_pstate_update_util_hwp_local(struct cpudata *cpu, 1507 u64 time) 1508 { 1509 cpu->sample.time = time; 1510 1511 if (cpu->sched_flags & SCHED_CPUFREQ_IOWAIT) { 1512 bool do_io = false; 1513 1514 cpu->sched_flags = 0; 1515 /* 1516 * Set iowait_boost flag and update time. Since IO WAIT flag 1517 * is set all the time, we can't just conclude that there is 1518 * some IO bound activity is scheduled on this CPU with just 1519 * one occurrence. If we receive at least two in two 1520 * consecutive ticks, then we treat as boost candidate. 1521 */ 1522 if (time_before64(time, cpu->last_io_update + 2 * TICK_NSEC)) 1523 do_io = true; 1524 1525 cpu->last_io_update = time; 1526 1527 if (do_io) 1528 intel_pstate_hwp_boost_up(cpu); 1529 1530 } else { 1531 intel_pstate_hwp_boost_down(cpu); 1532 } 1533 } 1534 1535 static inline void intel_pstate_update_util_hwp(struct update_util_data *data, 1536 u64 time, unsigned int flags) 1537 { 1538 struct cpudata *cpu = container_of(data, struct cpudata, update_util); 1539 1540 cpu->sched_flags |= flags; 1541 1542 if (smp_processor_id() == cpu->cpu) 1543 intel_pstate_update_util_hwp_local(cpu, time); 1544 } 1545 1546 static inline void intel_pstate_calc_avg_perf(struct cpudata *cpu) 1547 { 1548 struct sample *sample = &cpu->sample; 1549 1550 sample->core_avg_perf = div_ext_fp(sample->aperf, sample->mperf); 1551 } 1552 1553 static inline bool intel_pstate_sample(struct cpudata *cpu, u64 time) 1554 { 1555 u64 aperf, mperf; 1556 unsigned long flags; 1557 u64 tsc; 1558 1559 local_irq_save(flags); 1560 rdmsrl(MSR_IA32_APERF, aperf); 1561 rdmsrl(MSR_IA32_MPERF, mperf); 1562 tsc = rdtsc(); 1563 if (cpu->prev_mperf == mperf || cpu->prev_tsc == tsc) { 1564 local_irq_restore(flags); 1565 return false; 1566 } 1567 local_irq_restore(flags); 1568 1569 cpu->last_sample_time = cpu->sample.time; 1570 cpu->sample.time = time; 1571 cpu->sample.aperf = aperf; 1572 cpu->sample.mperf = mperf; 1573 cpu->sample.tsc = tsc; 1574 cpu->sample.aperf -= cpu->prev_aperf; 1575 cpu->sample.mperf -= cpu->prev_mperf; 1576 cpu->sample.tsc -= cpu->prev_tsc; 1577 1578 cpu->prev_aperf = aperf; 1579 cpu->prev_mperf = mperf; 1580 cpu->prev_tsc = tsc; 1581 /* 1582 * First time this function is invoked in a given cycle, all of the 1583 * previous sample data fields are equal to zero or stale and they must 1584 * be populated with meaningful numbers for things to work, so assume 1585 * that sample.time will always be reset before setting the utilization 1586 * update hook and make the caller skip the sample then. 1587 */ 1588 if (cpu->last_sample_time) { 1589 intel_pstate_calc_avg_perf(cpu); 1590 return true; 1591 } 1592 return false; 1593 } 1594 1595 static inline int32_t get_avg_frequency(struct cpudata *cpu) 1596 { 1597 return mul_ext_fp(cpu->sample.core_avg_perf, cpu_khz); 1598 } 1599 1600 static inline int32_t get_avg_pstate(struct cpudata *cpu) 1601 { 1602 return mul_ext_fp(cpu->pstate.max_pstate_physical, 1603 cpu->sample.core_avg_perf); 1604 } 1605 1606 static inline int32_t get_target_pstate(struct cpudata *cpu) 1607 { 1608 struct sample *sample = &cpu->sample; 1609 int32_t busy_frac, boost; 1610 int target, avg_pstate; 1611 1612 busy_frac = div_fp(sample->mperf << cpu->aperf_mperf_shift, 1613 sample->tsc); 1614 1615 boost = cpu->iowait_boost; 1616 cpu->iowait_boost >>= 1; 1617 1618 if (busy_frac < boost) 1619 busy_frac = boost; 1620 1621 sample->busy_scaled = busy_frac * 100; 1622 1623 target = global.no_turbo || global.turbo_disabled ? 1624 cpu->pstate.max_pstate : cpu->pstate.turbo_pstate; 1625 target += target >> 2; 1626 target = mul_fp(target, busy_frac); 1627 if (target < cpu->pstate.min_pstate) 1628 target = cpu->pstate.min_pstate; 1629 1630 /* 1631 * If the average P-state during the previous cycle was higher than the 1632 * current target, add 50% of the difference to the target to reduce 1633 * possible performance oscillations and offset possible performance 1634 * loss related to moving the workload from one CPU to another within 1635 * a package/module. 1636 */ 1637 avg_pstate = get_avg_pstate(cpu); 1638 if (avg_pstate > target) 1639 target += (avg_pstate - target) >> 1; 1640 1641 return target; 1642 } 1643 1644 static int intel_pstate_prepare_request(struct cpudata *cpu, int pstate) 1645 { 1646 int max_pstate = intel_pstate_get_base_pstate(cpu); 1647 int min_pstate; 1648 1649 min_pstate = max(cpu->pstate.min_pstate, cpu->min_perf_ratio); 1650 max_pstate = max(min_pstate, cpu->max_perf_ratio); 1651 return clamp_t(int, pstate, min_pstate, max_pstate); 1652 } 1653 1654 static void intel_pstate_update_pstate(struct cpudata *cpu, int pstate) 1655 { 1656 if (pstate == cpu->pstate.current_pstate) 1657 return; 1658 1659 cpu->pstate.current_pstate = pstate; 1660 wrmsrl(MSR_IA32_PERF_CTL, pstate_funcs.get_val(cpu, pstate)); 1661 } 1662 1663 static void intel_pstate_adjust_pstate(struct cpudata *cpu) 1664 { 1665 int from = cpu->pstate.current_pstate; 1666 struct sample *sample; 1667 int target_pstate; 1668 1669 update_turbo_state(); 1670 1671 target_pstate = get_target_pstate(cpu); 1672 target_pstate = intel_pstate_prepare_request(cpu, target_pstate); 1673 trace_cpu_frequency(target_pstate * cpu->pstate.scaling, cpu->cpu); 1674 intel_pstate_update_pstate(cpu, target_pstate); 1675 1676 sample = &cpu->sample; 1677 trace_pstate_sample(mul_ext_fp(100, sample->core_avg_perf), 1678 fp_toint(sample->busy_scaled), 1679 from, 1680 cpu->pstate.current_pstate, 1681 sample->mperf, 1682 sample->aperf, 1683 sample->tsc, 1684 get_avg_frequency(cpu), 1685 fp_toint(cpu->iowait_boost * 100)); 1686 } 1687 1688 static void intel_pstate_update_util(struct update_util_data *data, u64 time, 1689 unsigned int flags) 1690 { 1691 struct cpudata *cpu = container_of(data, struct cpudata, update_util); 1692 u64 delta_ns; 1693 1694 /* Don't allow remote callbacks */ 1695 if (smp_processor_id() != cpu->cpu) 1696 return; 1697 1698 if (flags & SCHED_CPUFREQ_IOWAIT) { 1699 cpu->iowait_boost = int_tofp(1); 1700 cpu->last_update = time; 1701 /* 1702 * The last time the busy was 100% so P-state was max anyway 1703 * so avoid overhead of computation. 1704 */ 1705 if (fp_toint(cpu->sample.busy_scaled) == 100) 1706 return; 1707 1708 goto set_pstate; 1709 } else if (cpu->iowait_boost) { 1710 /* Clear iowait_boost if the CPU may have been idle. */ 1711 delta_ns = time - cpu->last_update; 1712 if (delta_ns > TICK_NSEC) 1713 cpu->iowait_boost = 0; 1714 } 1715 cpu->last_update = time; 1716 delta_ns = time - cpu->sample.time; 1717 if ((s64)delta_ns < INTEL_PSTATE_SAMPLING_INTERVAL) 1718 return; 1719 1720 set_pstate: 1721 if (intel_pstate_sample(cpu, time)) 1722 intel_pstate_adjust_pstate(cpu); 1723 } 1724 1725 static struct pstate_funcs core_funcs = { 1726 .get_max = core_get_max_pstate, 1727 .get_max_physical = core_get_max_pstate_physical, 1728 .get_min = core_get_min_pstate, 1729 .get_turbo = core_get_turbo_pstate, 1730 .get_scaling = core_get_scaling, 1731 .get_val = core_get_val, 1732 }; 1733 1734 static const struct pstate_funcs silvermont_funcs = { 1735 .get_max = atom_get_max_pstate, 1736 .get_max_physical = atom_get_max_pstate, 1737 .get_min = atom_get_min_pstate, 1738 .get_turbo = atom_get_turbo_pstate, 1739 .get_val = atom_get_val, 1740 .get_scaling = silvermont_get_scaling, 1741 .get_vid = atom_get_vid, 1742 }; 1743 1744 static const struct pstate_funcs airmont_funcs = { 1745 .get_max = atom_get_max_pstate, 1746 .get_max_physical = atom_get_max_pstate, 1747 .get_min = atom_get_min_pstate, 1748 .get_turbo = atom_get_turbo_pstate, 1749 .get_val = atom_get_val, 1750 .get_scaling = airmont_get_scaling, 1751 .get_vid = atom_get_vid, 1752 }; 1753 1754 static const struct pstate_funcs knl_funcs = { 1755 .get_max = core_get_max_pstate, 1756 .get_max_physical = core_get_max_pstate_physical, 1757 .get_min = core_get_min_pstate, 1758 .get_turbo = knl_get_turbo_pstate, 1759 .get_aperf_mperf_shift = knl_get_aperf_mperf_shift, 1760 .get_scaling = core_get_scaling, 1761 .get_val = core_get_val, 1762 }; 1763 1764 #define ICPU(model, policy) \ 1765 { X86_VENDOR_INTEL, 6, model, X86_FEATURE_APERFMPERF,\ 1766 (unsigned long)&policy } 1767 1768 static const struct x86_cpu_id intel_pstate_cpu_ids[] = { 1769 ICPU(INTEL_FAM6_SANDYBRIDGE, core_funcs), 1770 ICPU(INTEL_FAM6_SANDYBRIDGE_X, core_funcs), 1771 ICPU(INTEL_FAM6_ATOM_SILVERMONT1, silvermont_funcs), 1772 ICPU(INTEL_FAM6_IVYBRIDGE, core_funcs), 1773 ICPU(INTEL_FAM6_HASWELL_CORE, core_funcs), 1774 ICPU(INTEL_FAM6_BROADWELL_CORE, core_funcs), 1775 ICPU(INTEL_FAM6_IVYBRIDGE_X, core_funcs), 1776 ICPU(INTEL_FAM6_HASWELL_X, core_funcs), 1777 ICPU(INTEL_FAM6_HASWELL_ULT, core_funcs), 1778 ICPU(INTEL_FAM6_HASWELL_GT3E, core_funcs), 1779 ICPU(INTEL_FAM6_BROADWELL_GT3E, core_funcs), 1780 ICPU(INTEL_FAM6_ATOM_AIRMONT, airmont_funcs), 1781 ICPU(INTEL_FAM6_SKYLAKE_MOBILE, core_funcs), 1782 ICPU(INTEL_FAM6_BROADWELL_X, core_funcs), 1783 ICPU(INTEL_FAM6_SKYLAKE_DESKTOP, core_funcs), 1784 ICPU(INTEL_FAM6_BROADWELL_XEON_D, core_funcs), 1785 ICPU(INTEL_FAM6_XEON_PHI_KNL, knl_funcs), 1786 ICPU(INTEL_FAM6_XEON_PHI_KNM, knl_funcs), 1787 ICPU(INTEL_FAM6_ATOM_GOLDMONT, core_funcs), 1788 ICPU(INTEL_FAM6_ATOM_GEMINI_LAKE, core_funcs), 1789 ICPU(INTEL_FAM6_SKYLAKE_X, core_funcs), 1790 {} 1791 }; 1792 MODULE_DEVICE_TABLE(x86cpu, intel_pstate_cpu_ids); 1793 1794 static const struct x86_cpu_id intel_pstate_cpu_oob_ids[] __initconst = { 1795 ICPU(INTEL_FAM6_BROADWELL_XEON_D, core_funcs), 1796 ICPU(INTEL_FAM6_BROADWELL_X, core_funcs), 1797 ICPU(INTEL_FAM6_SKYLAKE_X, core_funcs), 1798 {} 1799 }; 1800 1801 static const struct x86_cpu_id intel_pstate_cpu_ee_disable_ids[] = { 1802 ICPU(INTEL_FAM6_KABYLAKE_DESKTOP, core_funcs), 1803 {} 1804 }; 1805 1806 static const struct x86_cpu_id intel_pstate_hwp_boost_ids[] = { 1807 ICPU(INTEL_FAM6_SKYLAKE_X, core_funcs), 1808 ICPU(INTEL_FAM6_SKYLAKE_DESKTOP, core_funcs), 1809 {} 1810 }; 1811 1812 static int intel_pstate_init_cpu(unsigned int cpunum) 1813 { 1814 struct cpudata *cpu; 1815 1816 cpu = all_cpu_data[cpunum]; 1817 1818 if (!cpu) { 1819 cpu = kzalloc(sizeof(*cpu), GFP_KERNEL); 1820 if (!cpu) 1821 return -ENOMEM; 1822 1823 all_cpu_data[cpunum] = cpu; 1824 1825 cpu->epp_default = -EINVAL; 1826 cpu->epp_powersave = -EINVAL; 1827 cpu->epp_saved = -EINVAL; 1828 } 1829 1830 cpu = all_cpu_data[cpunum]; 1831 1832 cpu->cpu = cpunum; 1833 1834 if (hwp_active) { 1835 const struct x86_cpu_id *id; 1836 1837 id = x86_match_cpu(intel_pstate_cpu_ee_disable_ids); 1838 if (id) 1839 intel_pstate_disable_ee(cpunum); 1840 1841 intel_pstate_hwp_enable(cpu); 1842 1843 id = x86_match_cpu(intel_pstate_hwp_boost_ids); 1844 if (id) 1845 hwp_boost = true; 1846 } 1847 1848 intel_pstate_get_cpu_pstates(cpu); 1849 1850 pr_debug("controlling: cpu %d\n", cpunum); 1851 1852 return 0; 1853 } 1854 1855 static void intel_pstate_set_update_util_hook(unsigned int cpu_num) 1856 { 1857 struct cpudata *cpu = all_cpu_data[cpu_num]; 1858 1859 if (hwp_active && !hwp_boost) 1860 return; 1861 1862 if (cpu->update_util_set) 1863 return; 1864 1865 /* Prevent intel_pstate_update_util() from using stale data. */ 1866 cpu->sample.time = 0; 1867 cpufreq_add_update_util_hook(cpu_num, &cpu->update_util, 1868 (hwp_active ? 1869 intel_pstate_update_util_hwp : 1870 intel_pstate_update_util)); 1871 cpu->update_util_set = true; 1872 } 1873 1874 static void intel_pstate_clear_update_util_hook(unsigned int cpu) 1875 { 1876 struct cpudata *cpu_data = all_cpu_data[cpu]; 1877 1878 if (!cpu_data->update_util_set) 1879 return; 1880 1881 cpufreq_remove_update_util_hook(cpu); 1882 cpu_data->update_util_set = false; 1883 synchronize_sched(); 1884 } 1885 1886 static int intel_pstate_get_max_freq(struct cpudata *cpu) 1887 { 1888 return global.turbo_disabled || global.no_turbo ? 1889 cpu->pstate.max_freq : cpu->pstate.turbo_freq; 1890 } 1891 1892 static void intel_pstate_update_perf_limits(struct cpufreq_policy *policy, 1893 struct cpudata *cpu) 1894 { 1895 int max_freq = intel_pstate_get_max_freq(cpu); 1896 int32_t max_policy_perf, min_policy_perf; 1897 int max_state, turbo_max; 1898 1899 /* 1900 * HWP needs some special consideration, because on BDX the 1901 * HWP_REQUEST uses abstract value to represent performance 1902 * rather than pure ratios. 1903 */ 1904 if (hwp_active) { 1905 intel_pstate_get_hwp_max(cpu->cpu, &turbo_max, &max_state); 1906 } else { 1907 max_state = intel_pstate_get_base_pstate(cpu); 1908 turbo_max = cpu->pstate.turbo_pstate; 1909 } 1910 1911 max_policy_perf = max_state * policy->max / max_freq; 1912 if (policy->max == policy->min) { 1913 min_policy_perf = max_policy_perf; 1914 } else { 1915 min_policy_perf = max_state * policy->min / max_freq; 1916 min_policy_perf = clamp_t(int32_t, min_policy_perf, 1917 0, max_policy_perf); 1918 } 1919 1920 pr_debug("cpu:%d max_state %d min_policy_perf:%d max_policy_perf:%d\n", 1921 policy->cpu, max_state, 1922 min_policy_perf, max_policy_perf); 1923 1924 /* Normalize user input to [min_perf, max_perf] */ 1925 if (per_cpu_limits) { 1926 cpu->min_perf_ratio = min_policy_perf; 1927 cpu->max_perf_ratio = max_policy_perf; 1928 } else { 1929 int32_t global_min, global_max; 1930 1931 /* Global limits are in percent of the maximum turbo P-state. */ 1932 global_max = DIV_ROUND_UP(turbo_max * global.max_perf_pct, 100); 1933 global_min = DIV_ROUND_UP(turbo_max * global.min_perf_pct, 100); 1934 global_min = clamp_t(int32_t, global_min, 0, global_max); 1935 1936 pr_debug("cpu:%d global_min:%d global_max:%d\n", policy->cpu, 1937 global_min, global_max); 1938 1939 cpu->min_perf_ratio = max(min_policy_perf, global_min); 1940 cpu->min_perf_ratio = min(cpu->min_perf_ratio, max_policy_perf); 1941 cpu->max_perf_ratio = min(max_policy_perf, global_max); 1942 cpu->max_perf_ratio = max(min_policy_perf, cpu->max_perf_ratio); 1943 1944 /* Make sure min_perf <= max_perf */ 1945 cpu->min_perf_ratio = min(cpu->min_perf_ratio, 1946 cpu->max_perf_ratio); 1947 1948 } 1949 pr_debug("cpu:%d max_perf_ratio:%d min_perf_ratio:%d\n", policy->cpu, 1950 cpu->max_perf_ratio, 1951 cpu->min_perf_ratio); 1952 } 1953 1954 static int intel_pstate_set_policy(struct cpufreq_policy *policy) 1955 { 1956 struct cpudata *cpu; 1957 1958 if (!policy->cpuinfo.max_freq) 1959 return -ENODEV; 1960 1961 pr_debug("set_policy cpuinfo.max %u policy->max %u\n", 1962 policy->cpuinfo.max_freq, policy->max); 1963 1964 cpu = all_cpu_data[policy->cpu]; 1965 cpu->policy = policy->policy; 1966 1967 mutex_lock(&intel_pstate_limits_lock); 1968 1969 intel_pstate_update_perf_limits(policy, cpu); 1970 1971 if (cpu->policy == CPUFREQ_POLICY_PERFORMANCE) { 1972 /* 1973 * NOHZ_FULL CPUs need this as the governor callback may not 1974 * be invoked on them. 1975 */ 1976 intel_pstate_clear_update_util_hook(policy->cpu); 1977 intel_pstate_max_within_limits(cpu); 1978 } else { 1979 intel_pstate_set_update_util_hook(policy->cpu); 1980 } 1981 1982 if (hwp_active) { 1983 /* 1984 * When hwp_boost was active before and dynamically it 1985 * was turned off, in that case we need to clear the 1986 * update util hook. 1987 */ 1988 if (!hwp_boost) 1989 intel_pstate_clear_update_util_hook(policy->cpu); 1990 intel_pstate_hwp_set(policy->cpu); 1991 } 1992 1993 mutex_unlock(&intel_pstate_limits_lock); 1994 1995 return 0; 1996 } 1997 1998 static void intel_pstate_adjust_policy_max(struct cpufreq_policy *policy, 1999 struct cpudata *cpu) 2000 { 2001 if (cpu->pstate.max_pstate_physical > cpu->pstate.max_pstate && 2002 policy->max < policy->cpuinfo.max_freq && 2003 policy->max > cpu->pstate.max_freq) { 2004 pr_debug("policy->max > max non turbo frequency\n"); 2005 policy->max = policy->cpuinfo.max_freq; 2006 } 2007 } 2008 2009 static int intel_pstate_verify_policy(struct cpufreq_policy *policy) 2010 { 2011 struct cpudata *cpu = all_cpu_data[policy->cpu]; 2012 2013 update_turbo_state(); 2014 cpufreq_verify_within_limits(policy, policy->cpuinfo.min_freq, 2015 intel_pstate_get_max_freq(cpu)); 2016 2017 if (policy->policy != CPUFREQ_POLICY_POWERSAVE && 2018 policy->policy != CPUFREQ_POLICY_PERFORMANCE) 2019 return -EINVAL; 2020 2021 intel_pstate_adjust_policy_max(policy, cpu); 2022 2023 return 0; 2024 } 2025 2026 static void intel_cpufreq_stop_cpu(struct cpufreq_policy *policy) 2027 { 2028 intel_pstate_set_min_pstate(all_cpu_data[policy->cpu]); 2029 } 2030 2031 static void intel_pstate_stop_cpu(struct cpufreq_policy *policy) 2032 { 2033 pr_debug("CPU %d exiting\n", policy->cpu); 2034 2035 intel_pstate_clear_update_util_hook(policy->cpu); 2036 if (hwp_active) 2037 intel_pstate_hwp_save_state(policy); 2038 else 2039 intel_cpufreq_stop_cpu(policy); 2040 } 2041 2042 static int intel_pstate_cpu_exit(struct cpufreq_policy *policy) 2043 { 2044 intel_pstate_exit_perf_limits(policy); 2045 2046 policy->fast_switch_possible = false; 2047 2048 return 0; 2049 } 2050 2051 static int __intel_pstate_cpu_init(struct cpufreq_policy *policy) 2052 { 2053 struct cpudata *cpu; 2054 int rc; 2055 2056 rc = intel_pstate_init_cpu(policy->cpu); 2057 if (rc) 2058 return rc; 2059 2060 cpu = all_cpu_data[policy->cpu]; 2061 2062 cpu->max_perf_ratio = 0xFF; 2063 cpu->min_perf_ratio = 0; 2064 2065 policy->min = cpu->pstate.min_pstate * cpu->pstate.scaling; 2066 policy->max = cpu->pstate.turbo_pstate * cpu->pstate.scaling; 2067 2068 /* cpuinfo and default policy values */ 2069 policy->cpuinfo.min_freq = cpu->pstate.min_pstate * cpu->pstate.scaling; 2070 update_turbo_state(); 2071 policy->cpuinfo.max_freq = global.turbo_disabled ? 2072 cpu->pstate.max_pstate : cpu->pstate.turbo_pstate; 2073 policy->cpuinfo.max_freq *= cpu->pstate.scaling; 2074 2075 intel_pstate_init_acpi_perf_limits(policy); 2076 2077 policy->fast_switch_possible = true; 2078 2079 return 0; 2080 } 2081 2082 static int intel_pstate_cpu_init(struct cpufreq_policy *policy) 2083 { 2084 int ret = __intel_pstate_cpu_init(policy); 2085 2086 if (ret) 2087 return ret; 2088 2089 if (IS_ENABLED(CONFIG_CPU_FREQ_DEFAULT_GOV_PERFORMANCE)) 2090 policy->policy = CPUFREQ_POLICY_PERFORMANCE; 2091 else 2092 policy->policy = CPUFREQ_POLICY_POWERSAVE; 2093 2094 return 0; 2095 } 2096 2097 static struct cpufreq_driver intel_pstate = { 2098 .flags = CPUFREQ_CONST_LOOPS, 2099 .verify = intel_pstate_verify_policy, 2100 .setpolicy = intel_pstate_set_policy, 2101 .suspend = intel_pstate_hwp_save_state, 2102 .resume = intel_pstate_resume, 2103 .init = intel_pstate_cpu_init, 2104 .exit = intel_pstate_cpu_exit, 2105 .stop_cpu = intel_pstate_stop_cpu, 2106 .name = "intel_pstate", 2107 }; 2108 2109 static int intel_cpufreq_verify_policy(struct cpufreq_policy *policy) 2110 { 2111 struct cpudata *cpu = all_cpu_data[policy->cpu]; 2112 2113 update_turbo_state(); 2114 cpufreq_verify_within_limits(policy, policy->cpuinfo.min_freq, 2115 intel_pstate_get_max_freq(cpu)); 2116 2117 intel_pstate_adjust_policy_max(policy, cpu); 2118 2119 intel_pstate_update_perf_limits(policy, cpu); 2120 2121 return 0; 2122 } 2123 2124 /* Use of trace in passive mode: 2125 * 2126 * In passive mode the trace core_busy field (also known as the 2127 * performance field, and lablelled as such on the graphs; also known as 2128 * core_avg_perf) is not needed and so is re-assigned to indicate if the 2129 * driver call was via the normal or fast switch path. Various graphs 2130 * output from the intel_pstate_tracer.py utility that include core_busy 2131 * (or performance or core_avg_perf) have a fixed y-axis from 0 to 100%, 2132 * so we use 10 to indicate the the normal path through the driver, and 2133 * 90 to indicate the fast switch path through the driver. 2134 * The scaled_busy field is not used, and is set to 0. 2135 */ 2136 2137 #define INTEL_PSTATE_TRACE_TARGET 10 2138 #define INTEL_PSTATE_TRACE_FAST_SWITCH 90 2139 2140 static void intel_cpufreq_trace(struct cpudata *cpu, unsigned int trace_type, int old_pstate) 2141 { 2142 struct sample *sample; 2143 2144 if (!trace_pstate_sample_enabled()) 2145 return; 2146 2147 if (!intel_pstate_sample(cpu, ktime_get())) 2148 return; 2149 2150 sample = &cpu->sample; 2151 trace_pstate_sample(trace_type, 2152 0, 2153 old_pstate, 2154 cpu->pstate.current_pstate, 2155 sample->mperf, 2156 sample->aperf, 2157 sample->tsc, 2158 get_avg_frequency(cpu), 2159 fp_toint(cpu->iowait_boost * 100)); 2160 } 2161 2162 static int intel_cpufreq_target(struct cpufreq_policy *policy, 2163 unsigned int target_freq, 2164 unsigned int relation) 2165 { 2166 struct cpudata *cpu = all_cpu_data[policy->cpu]; 2167 struct cpufreq_freqs freqs; 2168 int target_pstate, old_pstate; 2169 2170 update_turbo_state(); 2171 2172 freqs.old = policy->cur; 2173 freqs.new = target_freq; 2174 2175 cpufreq_freq_transition_begin(policy, &freqs); 2176 switch (relation) { 2177 case CPUFREQ_RELATION_L: 2178 target_pstate = DIV_ROUND_UP(freqs.new, cpu->pstate.scaling); 2179 break; 2180 case CPUFREQ_RELATION_H: 2181 target_pstate = freqs.new / cpu->pstate.scaling; 2182 break; 2183 default: 2184 target_pstate = DIV_ROUND_CLOSEST(freqs.new, cpu->pstate.scaling); 2185 break; 2186 } 2187 target_pstate = intel_pstate_prepare_request(cpu, target_pstate); 2188 old_pstate = cpu->pstate.current_pstate; 2189 if (target_pstate != cpu->pstate.current_pstate) { 2190 cpu->pstate.current_pstate = target_pstate; 2191 wrmsrl_on_cpu(policy->cpu, MSR_IA32_PERF_CTL, 2192 pstate_funcs.get_val(cpu, target_pstate)); 2193 } 2194 freqs.new = target_pstate * cpu->pstate.scaling; 2195 intel_cpufreq_trace(cpu, INTEL_PSTATE_TRACE_TARGET, old_pstate); 2196 cpufreq_freq_transition_end(policy, &freqs, false); 2197 2198 return 0; 2199 } 2200 2201 static unsigned int intel_cpufreq_fast_switch(struct cpufreq_policy *policy, 2202 unsigned int target_freq) 2203 { 2204 struct cpudata *cpu = all_cpu_data[policy->cpu]; 2205 int target_pstate, old_pstate; 2206 2207 update_turbo_state(); 2208 2209 target_pstate = DIV_ROUND_UP(target_freq, cpu->pstate.scaling); 2210 target_pstate = intel_pstate_prepare_request(cpu, target_pstate); 2211 old_pstate = cpu->pstate.current_pstate; 2212 intel_pstate_update_pstate(cpu, target_pstate); 2213 intel_cpufreq_trace(cpu, INTEL_PSTATE_TRACE_FAST_SWITCH, old_pstate); 2214 return target_pstate * cpu->pstate.scaling; 2215 } 2216 2217 static int intel_cpufreq_cpu_init(struct cpufreq_policy *policy) 2218 { 2219 int ret = __intel_pstate_cpu_init(policy); 2220 2221 if (ret) 2222 return ret; 2223 2224 policy->cpuinfo.transition_latency = INTEL_CPUFREQ_TRANSITION_LATENCY; 2225 policy->transition_delay_us = INTEL_CPUFREQ_TRANSITION_DELAY; 2226 /* This reflects the intel_pstate_get_cpu_pstates() setting. */ 2227 policy->cur = policy->cpuinfo.min_freq; 2228 2229 return 0; 2230 } 2231 2232 static struct cpufreq_driver intel_cpufreq = { 2233 .flags = CPUFREQ_CONST_LOOPS, 2234 .verify = intel_cpufreq_verify_policy, 2235 .target = intel_cpufreq_target, 2236 .fast_switch = intel_cpufreq_fast_switch, 2237 .init = intel_cpufreq_cpu_init, 2238 .exit = intel_pstate_cpu_exit, 2239 .stop_cpu = intel_cpufreq_stop_cpu, 2240 .name = "intel_cpufreq", 2241 }; 2242 2243 static struct cpufreq_driver *default_driver = &intel_pstate; 2244 2245 static void intel_pstate_driver_cleanup(void) 2246 { 2247 unsigned int cpu; 2248 2249 get_online_cpus(); 2250 for_each_online_cpu(cpu) { 2251 if (all_cpu_data[cpu]) { 2252 if (intel_pstate_driver == &intel_pstate) 2253 intel_pstate_clear_update_util_hook(cpu); 2254 2255 kfree(all_cpu_data[cpu]); 2256 all_cpu_data[cpu] = NULL; 2257 } 2258 } 2259 put_online_cpus(); 2260 intel_pstate_driver = NULL; 2261 } 2262 2263 static int intel_pstate_register_driver(struct cpufreq_driver *driver) 2264 { 2265 int ret; 2266 2267 memset(&global, 0, sizeof(global)); 2268 global.max_perf_pct = 100; 2269 2270 intel_pstate_driver = driver; 2271 ret = cpufreq_register_driver(intel_pstate_driver); 2272 if (ret) { 2273 intel_pstate_driver_cleanup(); 2274 return ret; 2275 } 2276 2277 global.min_perf_pct = min_perf_pct_min(); 2278 2279 return 0; 2280 } 2281 2282 static int intel_pstate_unregister_driver(void) 2283 { 2284 if (hwp_active) 2285 return -EBUSY; 2286 2287 cpufreq_unregister_driver(intel_pstate_driver); 2288 intel_pstate_driver_cleanup(); 2289 2290 return 0; 2291 } 2292 2293 static ssize_t intel_pstate_show_status(char *buf) 2294 { 2295 if (!intel_pstate_driver) 2296 return sprintf(buf, "off\n"); 2297 2298 return sprintf(buf, "%s\n", intel_pstate_driver == &intel_pstate ? 2299 "active" : "passive"); 2300 } 2301 2302 static int intel_pstate_update_status(const char *buf, size_t size) 2303 { 2304 int ret; 2305 2306 if (size == 3 && !strncmp(buf, "off", size)) 2307 return intel_pstate_driver ? 2308 intel_pstate_unregister_driver() : -EINVAL; 2309 2310 if (size == 6 && !strncmp(buf, "active", size)) { 2311 if (intel_pstate_driver) { 2312 if (intel_pstate_driver == &intel_pstate) 2313 return 0; 2314 2315 ret = intel_pstate_unregister_driver(); 2316 if (ret) 2317 return ret; 2318 } 2319 2320 return intel_pstate_register_driver(&intel_pstate); 2321 } 2322 2323 if (size == 7 && !strncmp(buf, "passive", size)) { 2324 if (intel_pstate_driver) { 2325 if (intel_pstate_driver == &intel_cpufreq) 2326 return 0; 2327 2328 ret = intel_pstate_unregister_driver(); 2329 if (ret) 2330 return ret; 2331 } 2332 2333 return intel_pstate_register_driver(&intel_cpufreq); 2334 } 2335 2336 return -EINVAL; 2337 } 2338 2339 static int no_load __initdata; 2340 static int no_hwp __initdata; 2341 static int hwp_only __initdata; 2342 static unsigned int force_load __initdata; 2343 2344 static int __init intel_pstate_msrs_not_valid(void) 2345 { 2346 if (!pstate_funcs.get_max() || 2347 !pstate_funcs.get_min() || 2348 !pstate_funcs.get_turbo()) 2349 return -ENODEV; 2350 2351 return 0; 2352 } 2353 2354 static void __init copy_cpu_funcs(struct pstate_funcs *funcs) 2355 { 2356 pstate_funcs.get_max = funcs->get_max; 2357 pstate_funcs.get_max_physical = funcs->get_max_physical; 2358 pstate_funcs.get_min = funcs->get_min; 2359 pstate_funcs.get_turbo = funcs->get_turbo; 2360 pstate_funcs.get_scaling = funcs->get_scaling; 2361 pstate_funcs.get_val = funcs->get_val; 2362 pstate_funcs.get_vid = funcs->get_vid; 2363 pstate_funcs.get_aperf_mperf_shift = funcs->get_aperf_mperf_shift; 2364 } 2365 2366 #ifdef CONFIG_ACPI 2367 2368 static bool __init intel_pstate_no_acpi_pss(void) 2369 { 2370 int i; 2371 2372 for_each_possible_cpu(i) { 2373 acpi_status status; 2374 union acpi_object *pss; 2375 struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL }; 2376 struct acpi_processor *pr = per_cpu(processors, i); 2377 2378 if (!pr) 2379 continue; 2380 2381 status = acpi_evaluate_object(pr->handle, "_PSS", NULL, &buffer); 2382 if (ACPI_FAILURE(status)) 2383 continue; 2384 2385 pss = buffer.pointer; 2386 if (pss && pss->type == ACPI_TYPE_PACKAGE) { 2387 kfree(pss); 2388 return false; 2389 } 2390 2391 kfree(pss); 2392 } 2393 2394 return true; 2395 } 2396 2397 static bool __init intel_pstate_no_acpi_pcch(void) 2398 { 2399 acpi_status status; 2400 acpi_handle handle; 2401 2402 status = acpi_get_handle(NULL, "\\_SB", &handle); 2403 if (ACPI_FAILURE(status)) 2404 return true; 2405 2406 return !acpi_has_method(handle, "PCCH"); 2407 } 2408 2409 static bool __init intel_pstate_has_acpi_ppc(void) 2410 { 2411 int i; 2412 2413 for_each_possible_cpu(i) { 2414 struct acpi_processor *pr = per_cpu(processors, i); 2415 2416 if (!pr) 2417 continue; 2418 if (acpi_has_method(pr->handle, "_PPC")) 2419 return true; 2420 } 2421 return false; 2422 } 2423 2424 enum { 2425 PSS, 2426 PPC, 2427 }; 2428 2429 /* Hardware vendor-specific info that has its own power management modes */ 2430 static struct acpi_platform_list plat_info[] __initdata = { 2431 {"HP ", "ProLiant", 0, ACPI_SIG_FADT, all_versions, 0, PSS}, 2432 {"ORACLE", "X4-2 ", 0, ACPI_SIG_FADT, all_versions, 0, PPC}, 2433 {"ORACLE", "X4-2L ", 0, ACPI_SIG_FADT, all_versions, 0, PPC}, 2434 {"ORACLE", "X4-2B ", 0, ACPI_SIG_FADT, all_versions, 0, PPC}, 2435 {"ORACLE", "X3-2 ", 0, ACPI_SIG_FADT, all_versions, 0, PPC}, 2436 {"ORACLE", "X3-2L ", 0, ACPI_SIG_FADT, all_versions, 0, PPC}, 2437 {"ORACLE", "X3-2B ", 0, ACPI_SIG_FADT, all_versions, 0, PPC}, 2438 {"ORACLE", "X4470M2 ", 0, ACPI_SIG_FADT, all_versions, 0, PPC}, 2439 {"ORACLE", "X4270M3 ", 0, ACPI_SIG_FADT, all_versions, 0, PPC}, 2440 {"ORACLE", "X4270M2 ", 0, ACPI_SIG_FADT, all_versions, 0, PPC}, 2441 {"ORACLE", "X4170M2 ", 0, ACPI_SIG_FADT, all_versions, 0, PPC}, 2442 {"ORACLE", "X4170 M3", 0, ACPI_SIG_FADT, all_versions, 0, PPC}, 2443 {"ORACLE", "X4275 M3", 0, ACPI_SIG_FADT, all_versions, 0, PPC}, 2444 {"ORACLE", "X6-2 ", 0, ACPI_SIG_FADT, all_versions, 0, PPC}, 2445 {"ORACLE", "Sudbury ", 0, ACPI_SIG_FADT, all_versions, 0, PPC}, 2446 { } /* End */ 2447 }; 2448 2449 static bool __init intel_pstate_platform_pwr_mgmt_exists(void) 2450 { 2451 const struct x86_cpu_id *id; 2452 u64 misc_pwr; 2453 int idx; 2454 2455 id = x86_match_cpu(intel_pstate_cpu_oob_ids); 2456 if (id) { 2457 rdmsrl(MSR_MISC_PWR_MGMT, misc_pwr); 2458 if ( misc_pwr & (1 << 8)) 2459 return true; 2460 } 2461 2462 idx = acpi_match_platform_list(plat_info); 2463 if (idx < 0) 2464 return false; 2465 2466 switch (plat_info[idx].data) { 2467 case PSS: 2468 if (!intel_pstate_no_acpi_pss()) 2469 return false; 2470 2471 return intel_pstate_no_acpi_pcch(); 2472 case PPC: 2473 return intel_pstate_has_acpi_ppc() && !force_load; 2474 } 2475 2476 return false; 2477 } 2478 2479 static void intel_pstate_request_control_from_smm(void) 2480 { 2481 /* 2482 * It may be unsafe to request P-states control from SMM if _PPC support 2483 * has not been enabled. 2484 */ 2485 if (acpi_ppc) 2486 acpi_processor_pstate_control(); 2487 } 2488 #else /* CONFIG_ACPI not enabled */ 2489 static inline bool intel_pstate_platform_pwr_mgmt_exists(void) { return false; } 2490 static inline bool intel_pstate_has_acpi_ppc(void) { return false; } 2491 static inline void intel_pstate_request_control_from_smm(void) {} 2492 #endif /* CONFIG_ACPI */ 2493 2494 #define INTEL_PSTATE_HWP_BROADWELL 0x01 2495 2496 #define ICPU_HWP(model, hwp_mode) \ 2497 { X86_VENDOR_INTEL, 6, model, X86_FEATURE_HWP, hwp_mode } 2498 2499 static const struct x86_cpu_id hwp_support_ids[] __initconst = { 2500 ICPU_HWP(INTEL_FAM6_BROADWELL_X, INTEL_PSTATE_HWP_BROADWELL), 2501 ICPU_HWP(INTEL_FAM6_BROADWELL_XEON_D, INTEL_PSTATE_HWP_BROADWELL), 2502 ICPU_HWP(X86_MODEL_ANY, 0), 2503 {} 2504 }; 2505 2506 static int __init intel_pstate_init(void) 2507 { 2508 const struct x86_cpu_id *id; 2509 int rc; 2510 2511 if (no_load) 2512 return -ENODEV; 2513 2514 id = x86_match_cpu(hwp_support_ids); 2515 if (id) { 2516 copy_cpu_funcs(&core_funcs); 2517 if (!no_hwp) { 2518 hwp_active++; 2519 hwp_mode_bdw = id->driver_data; 2520 intel_pstate.attr = hwp_cpufreq_attrs; 2521 goto hwp_cpu_matched; 2522 } 2523 } else { 2524 id = x86_match_cpu(intel_pstate_cpu_ids); 2525 if (!id) 2526 return -ENODEV; 2527 2528 copy_cpu_funcs((struct pstate_funcs *)id->driver_data); 2529 } 2530 2531 if (intel_pstate_msrs_not_valid()) 2532 return -ENODEV; 2533 2534 hwp_cpu_matched: 2535 /* 2536 * The Intel pstate driver will be ignored if the platform 2537 * firmware has its own power management modes. 2538 */ 2539 if (intel_pstate_platform_pwr_mgmt_exists()) 2540 return -ENODEV; 2541 2542 if (!hwp_active && hwp_only) 2543 return -ENOTSUPP; 2544 2545 pr_info("Intel P-state driver initializing\n"); 2546 2547 all_cpu_data = vzalloc(array_size(sizeof(void *), num_possible_cpus())); 2548 if (!all_cpu_data) 2549 return -ENOMEM; 2550 2551 intel_pstate_request_control_from_smm(); 2552 2553 intel_pstate_sysfs_expose_params(); 2554 2555 mutex_lock(&intel_pstate_driver_lock); 2556 rc = intel_pstate_register_driver(default_driver); 2557 mutex_unlock(&intel_pstate_driver_lock); 2558 if (rc) 2559 return rc; 2560 2561 if (hwp_active) 2562 pr_info("HWP enabled\n"); 2563 2564 return 0; 2565 } 2566 device_initcall(intel_pstate_init); 2567 2568 static int __init intel_pstate_setup(char *str) 2569 { 2570 if (!str) 2571 return -EINVAL; 2572 2573 if (!strcmp(str, "disable")) { 2574 no_load = 1; 2575 } else if (!strcmp(str, "passive")) { 2576 pr_info("Passive mode enabled\n"); 2577 default_driver = &intel_cpufreq; 2578 no_hwp = 1; 2579 } 2580 if (!strcmp(str, "no_hwp")) { 2581 pr_info("HWP disabled\n"); 2582 no_hwp = 1; 2583 } 2584 if (!strcmp(str, "force")) 2585 force_load = 1; 2586 if (!strcmp(str, "hwp_only")) 2587 hwp_only = 1; 2588 if (!strcmp(str, "per_cpu_perf_limits")) 2589 per_cpu_limits = true; 2590 2591 #ifdef CONFIG_ACPI 2592 if (!strcmp(str, "support_acpi_ppc")) 2593 acpi_ppc = true; 2594 #endif 2595 2596 return 0; 2597 } 2598 early_param("intel_pstate", intel_pstate_setup); 2599 2600 MODULE_AUTHOR("Dirk Brandewie <dirk.j.brandewie@intel.com>"); 2601 MODULE_DESCRIPTION("'intel_pstate' - P state driver Intel Core processors"); 2602 MODULE_LICENSE("GPL"); 2603