1 // SPDX-License-Identifier: GPL-2.0-or-later 2 /* 3 * amd-pstate.c - AMD Processor P-state Frequency Driver 4 * 5 * Copyright (C) 2021 Advanced Micro Devices, Inc. All Rights Reserved. 6 * 7 * Author: Huang Rui <ray.huang@amd.com> 8 * 9 * AMD P-State introduces a new CPU performance scaling design for AMD 10 * processors using the ACPI Collaborative Performance and Power Control (CPPC) 11 * feature which works with the AMD SMU firmware providing a finer grained 12 * frequency control range. It is to replace the legacy ACPI P-States control, 13 * allows a flexible, low-latency interface for the Linux kernel to directly 14 * communicate the performance hints to hardware. 15 * 16 * AMD P-State is supported on recent AMD Zen base CPU series include some of 17 * Zen2 and Zen3 processors. _CPC needs to be present in the ACPI tables of AMD 18 * P-State supported system. And there are two types of hardware implementations 19 * for AMD P-State: 1) Full MSR Solution and 2) Shared Memory Solution. 20 * X86_FEATURE_CPPC CPU feature flag is used to distinguish the different types. 21 */ 22 23 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt 24 25 #include <linux/kernel.h> 26 #include <linux/module.h> 27 #include <linux/init.h> 28 #include <linux/smp.h> 29 #include <linux/sched.h> 30 #include <linux/cpufreq.h> 31 #include <linux/compiler.h> 32 #include <linux/dmi.h> 33 #include <linux/slab.h> 34 #include <linux/acpi.h> 35 #include <linux/io.h> 36 #include <linux/delay.h> 37 #include <linux/uaccess.h> 38 #include <linux/static_call.h> 39 #include <linux/amd-pstate.h> 40 #include <linux/topology.h> 41 42 #include <acpi/processor.h> 43 #include <acpi/cppc_acpi.h> 44 45 #include <asm/msr.h> 46 #include <asm/processor.h> 47 #include <asm/cpufeature.h> 48 #include <asm/cpu_device_id.h> 49 #include "amd-pstate-trace.h" 50 51 #define AMD_PSTATE_TRANSITION_LATENCY 20000 52 #define AMD_PSTATE_TRANSITION_DELAY 1000 53 #define CPPC_HIGHEST_PERF_PERFORMANCE 196 54 #define CPPC_HIGHEST_PERF_DEFAULT 166 55 56 /* 57 * TODO: We need more time to fine tune processors with shared memory solution 58 * with community together. 59 * 60 * There are some performance drops on the CPU benchmarks which reports from 61 * Suse. We are co-working with them to fine tune the shared memory solution. So 62 * we disable it by default to go acpi-cpufreq on these processors and add a 63 * module parameter to be able to enable it manually for debugging. 64 */ 65 static struct cpufreq_driver *current_pstate_driver; 66 static struct cpufreq_driver amd_pstate_driver; 67 static struct cpufreq_driver amd_pstate_epp_driver; 68 static int cppc_state = AMD_PSTATE_UNDEFINED; 69 static bool cppc_enabled; 70 static bool amd_pstate_prefcore = true; 71 72 /* 73 * AMD Energy Preference Performance (EPP) 74 * The EPP is used in the CCLK DPM controller to drive 75 * the frequency that a core is going to operate during 76 * short periods of activity. EPP values will be utilized for 77 * different OS profiles (balanced, performance, power savings) 78 * display strings corresponding to EPP index in the 79 * energy_perf_strings[] 80 * index String 81 *------------------------------------- 82 * 0 default 83 * 1 performance 84 * 2 balance_performance 85 * 3 balance_power 86 * 4 power 87 */ 88 enum energy_perf_value_index { 89 EPP_INDEX_DEFAULT = 0, 90 EPP_INDEX_PERFORMANCE, 91 EPP_INDEX_BALANCE_PERFORMANCE, 92 EPP_INDEX_BALANCE_POWERSAVE, 93 EPP_INDEX_POWERSAVE, 94 }; 95 96 static const char * const energy_perf_strings[] = { 97 [EPP_INDEX_DEFAULT] = "default", 98 [EPP_INDEX_PERFORMANCE] = "performance", 99 [EPP_INDEX_BALANCE_PERFORMANCE] = "balance_performance", 100 [EPP_INDEX_BALANCE_POWERSAVE] = "balance_power", 101 [EPP_INDEX_POWERSAVE] = "power", 102 NULL 103 }; 104 105 static unsigned int epp_values[] = { 106 [EPP_INDEX_DEFAULT] = 0, 107 [EPP_INDEX_PERFORMANCE] = AMD_CPPC_EPP_PERFORMANCE, 108 [EPP_INDEX_BALANCE_PERFORMANCE] = AMD_CPPC_EPP_BALANCE_PERFORMANCE, 109 [EPP_INDEX_BALANCE_POWERSAVE] = AMD_CPPC_EPP_BALANCE_POWERSAVE, 110 [EPP_INDEX_POWERSAVE] = AMD_CPPC_EPP_POWERSAVE, 111 }; 112 113 typedef int (*cppc_mode_transition_fn)(int); 114 115 static inline int get_mode_idx_from_str(const char *str, size_t size) 116 { 117 int i; 118 119 for (i=0; i < AMD_PSTATE_MAX; i++) { 120 if (!strncmp(str, amd_pstate_mode_string[i], size)) 121 return i; 122 } 123 return -EINVAL; 124 } 125 126 static DEFINE_MUTEX(amd_pstate_limits_lock); 127 static DEFINE_MUTEX(amd_pstate_driver_lock); 128 129 static s16 amd_pstate_get_epp(struct amd_cpudata *cpudata, u64 cppc_req_cached) 130 { 131 u64 epp; 132 int ret; 133 134 if (boot_cpu_has(X86_FEATURE_CPPC)) { 135 if (!cppc_req_cached) { 136 epp = rdmsrl_on_cpu(cpudata->cpu, MSR_AMD_CPPC_REQ, 137 &cppc_req_cached); 138 if (epp) 139 return epp; 140 } 141 epp = (cppc_req_cached >> 24) & 0xFF; 142 } else { 143 ret = cppc_get_epp_perf(cpudata->cpu, &epp); 144 if (ret < 0) { 145 pr_debug("Could not retrieve energy perf value (%d)\n", ret); 146 return -EIO; 147 } 148 } 149 150 return (s16)(epp & 0xff); 151 } 152 153 static int amd_pstate_get_energy_pref_index(struct amd_cpudata *cpudata) 154 { 155 s16 epp; 156 int index = -EINVAL; 157 158 epp = amd_pstate_get_epp(cpudata, 0); 159 if (epp < 0) 160 return epp; 161 162 switch (epp) { 163 case AMD_CPPC_EPP_PERFORMANCE: 164 index = EPP_INDEX_PERFORMANCE; 165 break; 166 case AMD_CPPC_EPP_BALANCE_PERFORMANCE: 167 index = EPP_INDEX_BALANCE_PERFORMANCE; 168 break; 169 case AMD_CPPC_EPP_BALANCE_POWERSAVE: 170 index = EPP_INDEX_BALANCE_POWERSAVE; 171 break; 172 case AMD_CPPC_EPP_POWERSAVE: 173 index = EPP_INDEX_POWERSAVE; 174 break; 175 default: 176 break; 177 } 178 179 return index; 180 } 181 182 static void pstate_update_perf(struct amd_cpudata *cpudata, u32 min_perf, 183 u32 des_perf, u32 max_perf, bool fast_switch) 184 { 185 if (fast_switch) 186 wrmsrl(MSR_AMD_CPPC_REQ, READ_ONCE(cpudata->cppc_req_cached)); 187 else 188 wrmsrl_on_cpu(cpudata->cpu, MSR_AMD_CPPC_REQ, 189 READ_ONCE(cpudata->cppc_req_cached)); 190 } 191 192 DEFINE_STATIC_CALL(amd_pstate_update_perf, pstate_update_perf); 193 194 static inline void amd_pstate_update_perf(struct amd_cpudata *cpudata, 195 u32 min_perf, u32 des_perf, 196 u32 max_perf, bool fast_switch) 197 { 198 static_call(amd_pstate_update_perf)(cpudata, min_perf, des_perf, 199 max_perf, fast_switch); 200 } 201 202 static int amd_pstate_set_epp(struct amd_cpudata *cpudata, u32 epp) 203 { 204 int ret; 205 struct cppc_perf_ctrls perf_ctrls; 206 207 if (boot_cpu_has(X86_FEATURE_CPPC)) { 208 u64 value = READ_ONCE(cpudata->cppc_req_cached); 209 210 value &= ~GENMASK_ULL(31, 24); 211 value |= (u64)epp << 24; 212 WRITE_ONCE(cpudata->cppc_req_cached, value); 213 214 ret = wrmsrl_on_cpu(cpudata->cpu, MSR_AMD_CPPC_REQ, value); 215 if (!ret) 216 cpudata->epp_cached = epp; 217 } else { 218 amd_pstate_update_perf(cpudata, cpudata->min_limit_perf, 0U, 219 cpudata->max_limit_perf, false); 220 221 perf_ctrls.energy_perf = epp; 222 ret = cppc_set_epp_perf(cpudata->cpu, &perf_ctrls, 1); 223 if (ret) { 224 pr_debug("failed to set energy perf value (%d)\n", ret); 225 return ret; 226 } 227 cpudata->epp_cached = epp; 228 } 229 230 return ret; 231 } 232 233 static int amd_pstate_set_energy_pref_index(struct amd_cpudata *cpudata, 234 int pref_index) 235 { 236 int epp = -EINVAL; 237 int ret; 238 239 if (!pref_index) { 240 pr_debug("EPP pref_index is invalid\n"); 241 return -EINVAL; 242 } 243 244 if (epp == -EINVAL) 245 epp = epp_values[pref_index]; 246 247 if (epp > 0 && cpudata->policy == CPUFREQ_POLICY_PERFORMANCE) { 248 pr_debug("EPP cannot be set under performance policy\n"); 249 return -EBUSY; 250 } 251 252 ret = amd_pstate_set_epp(cpudata, epp); 253 254 return ret; 255 } 256 257 static inline int pstate_enable(bool enable) 258 { 259 int ret, cpu; 260 unsigned long logical_proc_id_mask = 0; 261 262 if (enable == cppc_enabled) 263 return 0; 264 265 for_each_present_cpu(cpu) { 266 unsigned long logical_id = topology_logical_die_id(cpu); 267 268 if (test_bit(logical_id, &logical_proc_id_mask)) 269 continue; 270 271 set_bit(logical_id, &logical_proc_id_mask); 272 273 ret = wrmsrl_safe_on_cpu(cpu, MSR_AMD_CPPC_ENABLE, 274 enable); 275 if (ret) 276 return ret; 277 } 278 279 cppc_enabled = enable; 280 return 0; 281 } 282 283 static int cppc_enable(bool enable) 284 { 285 int cpu, ret = 0; 286 struct cppc_perf_ctrls perf_ctrls; 287 288 if (enable == cppc_enabled) 289 return 0; 290 291 for_each_present_cpu(cpu) { 292 ret = cppc_set_enable(cpu, enable); 293 if (ret) 294 return ret; 295 296 /* Enable autonomous mode for EPP */ 297 if (cppc_state == AMD_PSTATE_ACTIVE) { 298 /* Set desired perf as zero to allow EPP firmware control */ 299 perf_ctrls.desired_perf = 0; 300 ret = cppc_set_perf(cpu, &perf_ctrls); 301 if (ret) 302 return ret; 303 } 304 } 305 306 cppc_enabled = enable; 307 return ret; 308 } 309 310 DEFINE_STATIC_CALL(amd_pstate_enable, pstate_enable); 311 312 static inline int amd_pstate_enable(bool enable) 313 { 314 return static_call(amd_pstate_enable)(enable); 315 } 316 317 static u32 amd_pstate_highest_perf_set(struct amd_cpudata *cpudata) 318 { 319 struct cpuinfo_x86 *c = &cpu_data(0); 320 321 /* 322 * For AMD CPUs with Family ID 19H and Model ID range 0x70 to 0x7f, 323 * the highest performance level is set to 196. 324 * https://bugzilla.kernel.org/show_bug.cgi?id=218759 325 */ 326 if (c->x86 == 0x19 && (c->x86_model >= 0x70 && c->x86_model <= 0x7f)) 327 return CPPC_HIGHEST_PERF_PERFORMANCE; 328 329 return CPPC_HIGHEST_PERF_DEFAULT; 330 } 331 332 static int pstate_init_perf(struct amd_cpudata *cpudata) 333 { 334 u64 cap1; 335 u32 highest_perf; 336 337 int ret = rdmsrl_safe_on_cpu(cpudata->cpu, MSR_AMD_CPPC_CAP1, 338 &cap1); 339 if (ret) 340 return ret; 341 342 /* For platforms that do not support the preferred core feature, the 343 * highest_pef may be configured with 166 or 255, to avoid max frequency 344 * calculated wrongly. we take the AMD_CPPC_HIGHEST_PERF(cap1) value as 345 * the default max perf. 346 */ 347 if (cpudata->hw_prefcore) 348 highest_perf = amd_pstate_highest_perf_set(cpudata); 349 else 350 highest_perf = AMD_CPPC_HIGHEST_PERF(cap1); 351 352 WRITE_ONCE(cpudata->highest_perf, highest_perf); 353 WRITE_ONCE(cpudata->max_limit_perf, highest_perf); 354 WRITE_ONCE(cpudata->nominal_perf, AMD_CPPC_NOMINAL_PERF(cap1)); 355 WRITE_ONCE(cpudata->lowest_nonlinear_perf, AMD_CPPC_LOWNONLIN_PERF(cap1)); 356 WRITE_ONCE(cpudata->lowest_perf, AMD_CPPC_LOWEST_PERF(cap1)); 357 WRITE_ONCE(cpudata->min_limit_perf, AMD_CPPC_LOWEST_PERF(cap1)); 358 return 0; 359 } 360 361 static int cppc_init_perf(struct amd_cpudata *cpudata) 362 { 363 struct cppc_perf_caps cppc_perf; 364 u32 highest_perf; 365 366 int ret = cppc_get_perf_caps(cpudata->cpu, &cppc_perf); 367 if (ret) 368 return ret; 369 370 if (cpudata->hw_prefcore) 371 highest_perf = amd_pstate_highest_perf_set(cpudata); 372 else 373 highest_perf = cppc_perf.highest_perf; 374 375 WRITE_ONCE(cpudata->highest_perf, highest_perf); 376 WRITE_ONCE(cpudata->max_limit_perf, highest_perf); 377 WRITE_ONCE(cpudata->nominal_perf, cppc_perf.nominal_perf); 378 WRITE_ONCE(cpudata->lowest_nonlinear_perf, 379 cppc_perf.lowest_nonlinear_perf); 380 WRITE_ONCE(cpudata->lowest_perf, cppc_perf.lowest_perf); 381 WRITE_ONCE(cpudata->min_limit_perf, cppc_perf.lowest_perf); 382 383 if (cppc_state == AMD_PSTATE_ACTIVE) 384 return 0; 385 386 ret = cppc_get_auto_sel_caps(cpudata->cpu, &cppc_perf); 387 if (ret) { 388 pr_warn("failed to get auto_sel, ret: %d\n", ret); 389 return 0; 390 } 391 392 ret = cppc_set_auto_sel(cpudata->cpu, 393 (cppc_state == AMD_PSTATE_PASSIVE) ? 0 : 1); 394 395 if (ret) 396 pr_warn("failed to set auto_sel, ret: %d\n", ret); 397 398 return ret; 399 } 400 401 DEFINE_STATIC_CALL(amd_pstate_init_perf, pstate_init_perf); 402 403 static inline int amd_pstate_init_perf(struct amd_cpudata *cpudata) 404 { 405 return static_call(amd_pstate_init_perf)(cpudata); 406 } 407 408 static void cppc_update_perf(struct amd_cpudata *cpudata, 409 u32 min_perf, u32 des_perf, 410 u32 max_perf, bool fast_switch) 411 { 412 struct cppc_perf_ctrls perf_ctrls; 413 414 perf_ctrls.max_perf = max_perf; 415 perf_ctrls.min_perf = min_perf; 416 perf_ctrls.desired_perf = des_perf; 417 418 cppc_set_perf(cpudata->cpu, &perf_ctrls); 419 } 420 421 static inline bool amd_pstate_sample(struct amd_cpudata *cpudata) 422 { 423 u64 aperf, mperf, tsc; 424 unsigned long flags; 425 426 local_irq_save(flags); 427 rdmsrl(MSR_IA32_APERF, aperf); 428 rdmsrl(MSR_IA32_MPERF, mperf); 429 tsc = rdtsc(); 430 431 if (cpudata->prev.mperf == mperf || cpudata->prev.tsc == tsc) { 432 local_irq_restore(flags); 433 return false; 434 } 435 436 local_irq_restore(flags); 437 438 cpudata->cur.aperf = aperf; 439 cpudata->cur.mperf = mperf; 440 cpudata->cur.tsc = tsc; 441 cpudata->cur.aperf -= cpudata->prev.aperf; 442 cpudata->cur.mperf -= cpudata->prev.mperf; 443 cpudata->cur.tsc -= cpudata->prev.tsc; 444 445 cpudata->prev.aperf = aperf; 446 cpudata->prev.mperf = mperf; 447 cpudata->prev.tsc = tsc; 448 449 cpudata->freq = div64_u64((cpudata->cur.aperf * cpu_khz), cpudata->cur.mperf); 450 451 return true; 452 } 453 454 static void amd_pstate_update(struct amd_cpudata *cpudata, u32 min_perf, 455 u32 des_perf, u32 max_perf, bool fast_switch, int gov_flags) 456 { 457 u64 prev = READ_ONCE(cpudata->cppc_req_cached); 458 u64 value = prev; 459 460 min_perf = clamp_t(unsigned long, min_perf, cpudata->min_limit_perf, 461 cpudata->max_limit_perf); 462 max_perf = clamp_t(unsigned long, max_perf, cpudata->min_limit_perf, 463 cpudata->max_limit_perf); 464 des_perf = clamp_t(unsigned long, des_perf, min_perf, max_perf); 465 466 if ((cppc_state == AMD_PSTATE_GUIDED) && (gov_flags & CPUFREQ_GOV_DYNAMIC_SWITCHING)) { 467 min_perf = des_perf; 468 des_perf = 0; 469 } 470 471 value &= ~AMD_CPPC_MIN_PERF(~0L); 472 value |= AMD_CPPC_MIN_PERF(min_perf); 473 474 value &= ~AMD_CPPC_DES_PERF(~0L); 475 value |= AMD_CPPC_DES_PERF(des_perf); 476 477 value &= ~AMD_CPPC_MAX_PERF(~0L); 478 value |= AMD_CPPC_MAX_PERF(max_perf); 479 480 if (trace_amd_pstate_perf_enabled() && amd_pstate_sample(cpudata)) { 481 trace_amd_pstate_perf(min_perf, des_perf, max_perf, cpudata->freq, 482 cpudata->cur.mperf, cpudata->cur.aperf, cpudata->cur.tsc, 483 cpudata->cpu, (value != prev), fast_switch); 484 } 485 486 if (value == prev) 487 return; 488 489 WRITE_ONCE(cpudata->cppc_req_cached, value); 490 491 amd_pstate_update_perf(cpudata, min_perf, des_perf, 492 max_perf, fast_switch); 493 } 494 495 static int amd_pstate_verify(struct cpufreq_policy_data *policy) 496 { 497 cpufreq_verify_within_cpu_limits(policy); 498 499 return 0; 500 } 501 502 static int amd_pstate_update_min_max_limit(struct cpufreq_policy *policy) 503 { 504 u32 max_limit_perf, min_limit_perf; 505 struct amd_cpudata *cpudata = policy->driver_data; 506 507 max_limit_perf = div_u64(policy->max * cpudata->highest_perf, cpudata->max_freq); 508 min_limit_perf = div_u64(policy->min * cpudata->highest_perf, cpudata->max_freq); 509 510 WRITE_ONCE(cpudata->max_limit_perf, max_limit_perf); 511 WRITE_ONCE(cpudata->min_limit_perf, min_limit_perf); 512 WRITE_ONCE(cpudata->max_limit_freq, policy->max); 513 WRITE_ONCE(cpudata->min_limit_freq, policy->min); 514 515 return 0; 516 } 517 518 static int amd_pstate_update_freq(struct cpufreq_policy *policy, 519 unsigned int target_freq, bool fast_switch) 520 { 521 struct cpufreq_freqs freqs; 522 struct amd_cpudata *cpudata = policy->driver_data; 523 unsigned long max_perf, min_perf, des_perf, cap_perf; 524 525 if (!cpudata->max_freq) 526 return -ENODEV; 527 528 if (policy->min != cpudata->min_limit_freq || policy->max != cpudata->max_limit_freq) 529 amd_pstate_update_min_max_limit(policy); 530 531 cap_perf = READ_ONCE(cpudata->highest_perf); 532 min_perf = READ_ONCE(cpudata->lowest_perf); 533 max_perf = cap_perf; 534 535 freqs.old = policy->cur; 536 freqs.new = target_freq; 537 538 des_perf = DIV_ROUND_CLOSEST(target_freq * cap_perf, 539 cpudata->max_freq); 540 541 WARN_ON(fast_switch && !policy->fast_switch_enabled); 542 /* 543 * If fast_switch is desired, then there aren't any registered 544 * transition notifiers. See comment for 545 * cpufreq_enable_fast_switch(). 546 */ 547 if (!fast_switch) 548 cpufreq_freq_transition_begin(policy, &freqs); 549 550 amd_pstate_update(cpudata, min_perf, des_perf, 551 max_perf, fast_switch, policy->governor->flags); 552 553 if (!fast_switch) 554 cpufreq_freq_transition_end(policy, &freqs, false); 555 556 return 0; 557 } 558 559 static int amd_pstate_target(struct cpufreq_policy *policy, 560 unsigned int target_freq, 561 unsigned int relation) 562 { 563 return amd_pstate_update_freq(policy, target_freq, false); 564 } 565 566 static unsigned int amd_pstate_fast_switch(struct cpufreq_policy *policy, 567 unsigned int target_freq) 568 { 569 if (!amd_pstate_update_freq(policy, target_freq, true)) 570 return target_freq; 571 return policy->cur; 572 } 573 574 static void amd_pstate_adjust_perf(unsigned int cpu, 575 unsigned long _min_perf, 576 unsigned long target_perf, 577 unsigned long capacity) 578 { 579 unsigned long max_perf, min_perf, des_perf, 580 cap_perf, lowest_nonlinear_perf, max_freq; 581 struct cpufreq_policy *policy = cpufreq_cpu_get(cpu); 582 struct amd_cpudata *cpudata = policy->driver_data; 583 unsigned int target_freq; 584 585 if (policy->min != cpudata->min_limit_freq || policy->max != cpudata->max_limit_freq) 586 amd_pstate_update_min_max_limit(policy); 587 588 589 cap_perf = READ_ONCE(cpudata->highest_perf); 590 lowest_nonlinear_perf = READ_ONCE(cpudata->lowest_nonlinear_perf); 591 max_freq = READ_ONCE(cpudata->max_freq); 592 593 des_perf = cap_perf; 594 if (target_perf < capacity) 595 des_perf = DIV_ROUND_UP(cap_perf * target_perf, capacity); 596 597 min_perf = READ_ONCE(cpudata->lowest_perf); 598 if (_min_perf < capacity) 599 min_perf = DIV_ROUND_UP(cap_perf * _min_perf, capacity); 600 601 if (min_perf < lowest_nonlinear_perf) 602 min_perf = lowest_nonlinear_perf; 603 604 max_perf = cap_perf; 605 if (max_perf < min_perf) 606 max_perf = min_perf; 607 608 des_perf = clamp_t(unsigned long, des_perf, min_perf, max_perf); 609 target_freq = div_u64(des_perf * max_freq, max_perf); 610 policy->cur = target_freq; 611 612 amd_pstate_update(cpudata, min_perf, des_perf, max_perf, true, 613 policy->governor->flags); 614 cpufreq_cpu_put(policy); 615 } 616 617 static int amd_get_min_freq(struct amd_cpudata *cpudata) 618 { 619 struct cppc_perf_caps cppc_perf; 620 621 int ret = cppc_get_perf_caps(cpudata->cpu, &cppc_perf); 622 if (ret) 623 return ret; 624 625 /* Switch to khz */ 626 return cppc_perf.lowest_freq * 1000; 627 } 628 629 static int amd_get_max_freq(struct amd_cpudata *cpudata) 630 { 631 struct cppc_perf_caps cppc_perf; 632 u32 max_perf, max_freq, nominal_freq, nominal_perf; 633 u64 boost_ratio; 634 635 int ret = cppc_get_perf_caps(cpudata->cpu, &cppc_perf); 636 if (ret) 637 return ret; 638 639 nominal_freq = cppc_perf.nominal_freq; 640 nominal_perf = READ_ONCE(cpudata->nominal_perf); 641 max_perf = READ_ONCE(cpudata->highest_perf); 642 643 boost_ratio = div_u64(max_perf << SCHED_CAPACITY_SHIFT, 644 nominal_perf); 645 646 max_freq = nominal_freq * boost_ratio >> SCHED_CAPACITY_SHIFT; 647 648 /* Switch to khz */ 649 return max_freq * 1000; 650 } 651 652 static int amd_get_nominal_freq(struct amd_cpudata *cpudata) 653 { 654 struct cppc_perf_caps cppc_perf; 655 656 int ret = cppc_get_perf_caps(cpudata->cpu, &cppc_perf); 657 if (ret) 658 return ret; 659 660 /* Switch to khz */ 661 return cppc_perf.nominal_freq * 1000; 662 } 663 664 static int amd_get_lowest_nonlinear_freq(struct amd_cpudata *cpudata) 665 { 666 struct cppc_perf_caps cppc_perf; 667 u32 lowest_nonlinear_freq, lowest_nonlinear_perf, 668 nominal_freq, nominal_perf; 669 u64 lowest_nonlinear_ratio; 670 671 int ret = cppc_get_perf_caps(cpudata->cpu, &cppc_perf); 672 if (ret) 673 return ret; 674 675 nominal_freq = cppc_perf.nominal_freq; 676 nominal_perf = READ_ONCE(cpudata->nominal_perf); 677 678 lowest_nonlinear_perf = cppc_perf.lowest_nonlinear_perf; 679 680 lowest_nonlinear_ratio = div_u64(lowest_nonlinear_perf << SCHED_CAPACITY_SHIFT, 681 nominal_perf); 682 683 lowest_nonlinear_freq = nominal_freq * lowest_nonlinear_ratio >> SCHED_CAPACITY_SHIFT; 684 685 /* Switch to khz */ 686 return lowest_nonlinear_freq * 1000; 687 } 688 689 static int amd_pstate_set_boost(struct cpufreq_policy *policy, int state) 690 { 691 struct amd_cpudata *cpudata = policy->driver_data; 692 int ret; 693 694 if (!cpudata->boost_supported) { 695 pr_err("Boost mode is not supported by this processor or SBIOS\n"); 696 return -EINVAL; 697 } 698 699 if (state) 700 policy->cpuinfo.max_freq = cpudata->max_freq; 701 else 702 policy->cpuinfo.max_freq = cpudata->nominal_freq; 703 704 policy->max = policy->cpuinfo.max_freq; 705 706 ret = freq_qos_update_request(&cpudata->req[1], 707 policy->cpuinfo.max_freq); 708 if (ret < 0) 709 return ret; 710 711 return 0; 712 } 713 714 static void amd_pstate_boost_init(struct amd_cpudata *cpudata) 715 { 716 u32 highest_perf, nominal_perf; 717 718 highest_perf = READ_ONCE(cpudata->highest_perf); 719 nominal_perf = READ_ONCE(cpudata->nominal_perf); 720 721 if (highest_perf <= nominal_perf) 722 return; 723 724 cpudata->boost_supported = true; 725 current_pstate_driver->boost_enabled = true; 726 } 727 728 static void amd_perf_ctl_reset(unsigned int cpu) 729 { 730 wrmsrl_on_cpu(cpu, MSR_AMD_PERF_CTL, 0); 731 } 732 733 /* 734 * Set amd-pstate preferred core enable can't be done directly from cpufreq callbacks 735 * due to locking, so queue the work for later. 736 */ 737 static void amd_pstste_sched_prefcore_workfn(struct work_struct *work) 738 { 739 sched_set_itmt_support(); 740 } 741 static DECLARE_WORK(sched_prefcore_work, amd_pstste_sched_prefcore_workfn); 742 743 /* 744 * Get the highest performance register value. 745 * @cpu: CPU from which to get highest performance. 746 * @highest_perf: Return address. 747 * 748 * Return: 0 for success, -EIO otherwise. 749 */ 750 static int amd_pstate_get_highest_perf(int cpu, u32 *highest_perf) 751 { 752 int ret; 753 754 if (boot_cpu_has(X86_FEATURE_CPPC)) { 755 u64 cap1; 756 757 ret = rdmsrl_safe_on_cpu(cpu, MSR_AMD_CPPC_CAP1, &cap1); 758 if (ret) 759 return ret; 760 WRITE_ONCE(*highest_perf, AMD_CPPC_HIGHEST_PERF(cap1)); 761 } else { 762 u64 cppc_highest_perf; 763 764 ret = cppc_get_highest_perf(cpu, &cppc_highest_perf); 765 if (ret) 766 return ret; 767 WRITE_ONCE(*highest_perf, cppc_highest_perf); 768 } 769 770 return (ret); 771 } 772 773 #define CPPC_MAX_PERF U8_MAX 774 775 static void amd_pstate_init_prefcore(struct amd_cpudata *cpudata) 776 { 777 int ret, prio; 778 u32 highest_perf; 779 780 ret = amd_pstate_get_highest_perf(cpudata->cpu, &highest_perf); 781 if (ret) 782 return; 783 784 cpudata->hw_prefcore = true; 785 /* check if CPPC preferred core feature is enabled*/ 786 if (highest_perf < CPPC_MAX_PERF) 787 prio = (int)highest_perf; 788 else { 789 pr_debug("AMD CPPC preferred core is unsupported!\n"); 790 cpudata->hw_prefcore = false; 791 return; 792 } 793 794 if (!amd_pstate_prefcore) 795 return; 796 797 /* 798 * The priorities can be set regardless of whether or not 799 * sched_set_itmt_support(true) has been called and it is valid to 800 * update them at any time after it has been called. 801 */ 802 sched_set_itmt_core_prio(prio, cpudata->cpu); 803 804 schedule_work(&sched_prefcore_work); 805 } 806 807 static int amd_pstate_cpu_init(struct cpufreq_policy *policy) 808 { 809 int min_freq, max_freq, nominal_freq, lowest_nonlinear_freq, ret; 810 struct device *dev; 811 struct amd_cpudata *cpudata; 812 813 /* 814 * Resetting PERF_CTL_MSR will put the CPU in P0 frequency, 815 * which is ideal for initialization process. 816 */ 817 amd_perf_ctl_reset(policy->cpu); 818 dev = get_cpu_device(policy->cpu); 819 if (!dev) 820 return -ENODEV; 821 822 cpudata = kzalloc(sizeof(*cpudata), GFP_KERNEL); 823 if (!cpudata) 824 return -ENOMEM; 825 826 cpudata->cpu = policy->cpu; 827 828 amd_pstate_init_prefcore(cpudata); 829 830 ret = amd_pstate_init_perf(cpudata); 831 if (ret) 832 goto free_cpudata1; 833 834 min_freq = amd_get_min_freq(cpudata); 835 max_freq = amd_get_max_freq(cpudata); 836 nominal_freq = amd_get_nominal_freq(cpudata); 837 lowest_nonlinear_freq = amd_get_lowest_nonlinear_freq(cpudata); 838 839 if (min_freq < 0 || max_freq < 0 || min_freq > max_freq) { 840 dev_err(dev, "min_freq(%d) or max_freq(%d) value is incorrect\n", 841 min_freq, max_freq); 842 ret = -EINVAL; 843 goto free_cpudata1; 844 } 845 846 policy->cpuinfo.transition_latency = AMD_PSTATE_TRANSITION_LATENCY; 847 policy->transition_delay_us = AMD_PSTATE_TRANSITION_DELAY; 848 849 policy->min = min_freq; 850 policy->max = max_freq; 851 852 policy->cpuinfo.min_freq = min_freq; 853 policy->cpuinfo.max_freq = max_freq; 854 855 /* It will be updated by governor */ 856 policy->cur = policy->cpuinfo.min_freq; 857 858 if (boot_cpu_has(X86_FEATURE_CPPC)) 859 policy->fast_switch_possible = true; 860 861 ret = freq_qos_add_request(&policy->constraints, &cpudata->req[0], 862 FREQ_QOS_MIN, policy->cpuinfo.min_freq); 863 if (ret < 0) { 864 dev_err(dev, "Failed to add min-freq constraint (%d)\n", ret); 865 goto free_cpudata1; 866 } 867 868 ret = freq_qos_add_request(&policy->constraints, &cpudata->req[1], 869 FREQ_QOS_MAX, policy->cpuinfo.max_freq); 870 if (ret < 0) { 871 dev_err(dev, "Failed to add max-freq constraint (%d)\n", ret); 872 goto free_cpudata2; 873 } 874 875 /* Initial processor data capability frequencies */ 876 cpudata->max_freq = max_freq; 877 cpudata->min_freq = min_freq; 878 cpudata->max_limit_freq = max_freq; 879 cpudata->min_limit_freq = min_freq; 880 cpudata->nominal_freq = nominal_freq; 881 cpudata->lowest_nonlinear_freq = lowest_nonlinear_freq; 882 883 policy->driver_data = cpudata; 884 885 amd_pstate_boost_init(cpudata); 886 if (!current_pstate_driver->adjust_perf) 887 current_pstate_driver->adjust_perf = amd_pstate_adjust_perf; 888 889 return 0; 890 891 free_cpudata2: 892 freq_qos_remove_request(&cpudata->req[0]); 893 free_cpudata1: 894 kfree(cpudata); 895 return ret; 896 } 897 898 static int amd_pstate_cpu_exit(struct cpufreq_policy *policy) 899 { 900 struct amd_cpudata *cpudata = policy->driver_data; 901 902 freq_qos_remove_request(&cpudata->req[1]); 903 freq_qos_remove_request(&cpudata->req[0]); 904 policy->fast_switch_possible = false; 905 kfree(cpudata); 906 907 return 0; 908 } 909 910 static int amd_pstate_cpu_resume(struct cpufreq_policy *policy) 911 { 912 int ret; 913 914 ret = amd_pstate_enable(true); 915 if (ret) 916 pr_err("failed to enable amd-pstate during resume, return %d\n", ret); 917 918 return ret; 919 } 920 921 static int amd_pstate_cpu_suspend(struct cpufreq_policy *policy) 922 { 923 int ret; 924 925 ret = amd_pstate_enable(false); 926 if (ret) 927 pr_err("failed to disable amd-pstate during suspend, return %d\n", ret); 928 929 return ret; 930 } 931 932 /* Sysfs attributes */ 933 934 /* 935 * This frequency is to indicate the maximum hardware frequency. 936 * If boost is not active but supported, the frequency will be larger than the 937 * one in cpuinfo. 938 */ 939 static ssize_t show_amd_pstate_max_freq(struct cpufreq_policy *policy, 940 char *buf) 941 { 942 int max_freq; 943 struct amd_cpudata *cpudata = policy->driver_data; 944 945 max_freq = amd_get_max_freq(cpudata); 946 if (max_freq < 0) 947 return max_freq; 948 949 return sysfs_emit(buf, "%u\n", max_freq); 950 } 951 952 static ssize_t show_amd_pstate_lowest_nonlinear_freq(struct cpufreq_policy *policy, 953 char *buf) 954 { 955 int freq; 956 struct amd_cpudata *cpudata = policy->driver_data; 957 958 freq = amd_get_lowest_nonlinear_freq(cpudata); 959 if (freq < 0) 960 return freq; 961 962 return sysfs_emit(buf, "%u\n", freq); 963 } 964 965 /* 966 * In some of ASICs, the highest_perf is not the one in the _CPC table, so we 967 * need to expose it to sysfs. 968 */ 969 static ssize_t show_amd_pstate_highest_perf(struct cpufreq_policy *policy, 970 char *buf) 971 { 972 u32 perf; 973 struct amd_cpudata *cpudata = policy->driver_data; 974 975 perf = READ_ONCE(cpudata->highest_perf); 976 977 return sysfs_emit(buf, "%u\n", perf); 978 } 979 980 static ssize_t show_amd_pstate_hw_prefcore(struct cpufreq_policy *policy, 981 char *buf) 982 { 983 bool hw_prefcore; 984 struct amd_cpudata *cpudata = policy->driver_data; 985 986 hw_prefcore = READ_ONCE(cpudata->hw_prefcore); 987 988 return sysfs_emit(buf, "%s\n", str_enabled_disabled(hw_prefcore)); 989 } 990 991 static ssize_t show_energy_performance_available_preferences( 992 struct cpufreq_policy *policy, char *buf) 993 { 994 int i = 0; 995 int offset = 0; 996 struct amd_cpudata *cpudata = policy->driver_data; 997 998 if (cpudata->policy == CPUFREQ_POLICY_PERFORMANCE) 999 return sysfs_emit_at(buf, offset, "%s\n", 1000 energy_perf_strings[EPP_INDEX_PERFORMANCE]); 1001 1002 while (energy_perf_strings[i] != NULL) 1003 offset += sysfs_emit_at(buf, offset, "%s ", energy_perf_strings[i++]); 1004 1005 offset += sysfs_emit_at(buf, offset, "\n"); 1006 1007 return offset; 1008 } 1009 1010 static ssize_t store_energy_performance_preference( 1011 struct cpufreq_policy *policy, const char *buf, size_t count) 1012 { 1013 struct amd_cpudata *cpudata = policy->driver_data; 1014 char str_preference[21]; 1015 ssize_t ret; 1016 1017 ret = sscanf(buf, "%20s", str_preference); 1018 if (ret != 1) 1019 return -EINVAL; 1020 1021 ret = match_string(energy_perf_strings, -1, str_preference); 1022 if (ret < 0) 1023 return -EINVAL; 1024 1025 mutex_lock(&amd_pstate_limits_lock); 1026 ret = amd_pstate_set_energy_pref_index(cpudata, ret); 1027 mutex_unlock(&amd_pstate_limits_lock); 1028 1029 return ret ?: count; 1030 } 1031 1032 static ssize_t show_energy_performance_preference( 1033 struct cpufreq_policy *policy, char *buf) 1034 { 1035 struct amd_cpudata *cpudata = policy->driver_data; 1036 int preference; 1037 1038 preference = amd_pstate_get_energy_pref_index(cpudata); 1039 if (preference < 0) 1040 return preference; 1041 1042 return sysfs_emit(buf, "%s\n", energy_perf_strings[preference]); 1043 } 1044 1045 static void amd_pstate_driver_cleanup(void) 1046 { 1047 amd_pstate_enable(false); 1048 cppc_state = AMD_PSTATE_DISABLE; 1049 current_pstate_driver = NULL; 1050 } 1051 1052 static int amd_pstate_register_driver(int mode) 1053 { 1054 int ret; 1055 1056 if (mode == AMD_PSTATE_PASSIVE || mode == AMD_PSTATE_GUIDED) 1057 current_pstate_driver = &amd_pstate_driver; 1058 else if (mode == AMD_PSTATE_ACTIVE) 1059 current_pstate_driver = &amd_pstate_epp_driver; 1060 else 1061 return -EINVAL; 1062 1063 cppc_state = mode; 1064 1065 ret = amd_pstate_enable(true); 1066 if (ret) { 1067 pr_err("failed to enable cppc during amd-pstate driver registration, return %d\n", 1068 ret); 1069 amd_pstate_driver_cleanup(); 1070 return ret; 1071 } 1072 1073 ret = cpufreq_register_driver(current_pstate_driver); 1074 if (ret) { 1075 amd_pstate_driver_cleanup(); 1076 return ret; 1077 } 1078 1079 return 0; 1080 } 1081 1082 static int amd_pstate_unregister_driver(int dummy) 1083 { 1084 cpufreq_unregister_driver(current_pstate_driver); 1085 amd_pstate_driver_cleanup(); 1086 return 0; 1087 } 1088 1089 static int amd_pstate_change_mode_without_dvr_change(int mode) 1090 { 1091 int cpu = 0; 1092 1093 cppc_state = mode; 1094 1095 if (boot_cpu_has(X86_FEATURE_CPPC) || cppc_state == AMD_PSTATE_ACTIVE) 1096 return 0; 1097 1098 for_each_present_cpu(cpu) { 1099 cppc_set_auto_sel(cpu, (cppc_state == AMD_PSTATE_PASSIVE) ? 0 : 1); 1100 } 1101 1102 return 0; 1103 } 1104 1105 static int amd_pstate_change_driver_mode(int mode) 1106 { 1107 int ret; 1108 1109 ret = amd_pstate_unregister_driver(0); 1110 if (ret) 1111 return ret; 1112 1113 ret = amd_pstate_register_driver(mode); 1114 if (ret) 1115 return ret; 1116 1117 return 0; 1118 } 1119 1120 static cppc_mode_transition_fn mode_state_machine[AMD_PSTATE_MAX][AMD_PSTATE_MAX] = { 1121 [AMD_PSTATE_DISABLE] = { 1122 [AMD_PSTATE_DISABLE] = NULL, 1123 [AMD_PSTATE_PASSIVE] = amd_pstate_register_driver, 1124 [AMD_PSTATE_ACTIVE] = amd_pstate_register_driver, 1125 [AMD_PSTATE_GUIDED] = amd_pstate_register_driver, 1126 }, 1127 [AMD_PSTATE_PASSIVE] = { 1128 [AMD_PSTATE_DISABLE] = amd_pstate_unregister_driver, 1129 [AMD_PSTATE_PASSIVE] = NULL, 1130 [AMD_PSTATE_ACTIVE] = amd_pstate_change_driver_mode, 1131 [AMD_PSTATE_GUIDED] = amd_pstate_change_mode_without_dvr_change, 1132 }, 1133 [AMD_PSTATE_ACTIVE] = { 1134 [AMD_PSTATE_DISABLE] = amd_pstate_unregister_driver, 1135 [AMD_PSTATE_PASSIVE] = amd_pstate_change_driver_mode, 1136 [AMD_PSTATE_ACTIVE] = NULL, 1137 [AMD_PSTATE_GUIDED] = amd_pstate_change_driver_mode, 1138 }, 1139 [AMD_PSTATE_GUIDED] = { 1140 [AMD_PSTATE_DISABLE] = amd_pstate_unregister_driver, 1141 [AMD_PSTATE_PASSIVE] = amd_pstate_change_mode_without_dvr_change, 1142 [AMD_PSTATE_ACTIVE] = amd_pstate_change_driver_mode, 1143 [AMD_PSTATE_GUIDED] = NULL, 1144 }, 1145 }; 1146 1147 static ssize_t amd_pstate_show_status(char *buf) 1148 { 1149 if (!current_pstate_driver) 1150 return sysfs_emit(buf, "disable\n"); 1151 1152 return sysfs_emit(buf, "%s\n", amd_pstate_mode_string[cppc_state]); 1153 } 1154 1155 static int amd_pstate_update_status(const char *buf, size_t size) 1156 { 1157 int mode_idx; 1158 1159 if (size > strlen("passive") || size < strlen("active")) 1160 return -EINVAL; 1161 1162 mode_idx = get_mode_idx_from_str(buf, size); 1163 1164 if (mode_idx < 0 || mode_idx >= AMD_PSTATE_MAX) 1165 return -EINVAL; 1166 1167 if (mode_state_machine[cppc_state][mode_idx]) 1168 return mode_state_machine[cppc_state][mode_idx](mode_idx); 1169 1170 return 0; 1171 } 1172 1173 static ssize_t status_show(struct device *dev, 1174 struct device_attribute *attr, char *buf) 1175 { 1176 ssize_t ret; 1177 1178 mutex_lock(&amd_pstate_driver_lock); 1179 ret = amd_pstate_show_status(buf); 1180 mutex_unlock(&amd_pstate_driver_lock); 1181 1182 return ret; 1183 } 1184 1185 static ssize_t status_store(struct device *a, struct device_attribute *b, 1186 const char *buf, size_t count) 1187 { 1188 char *p = memchr(buf, '\n', count); 1189 int ret; 1190 1191 mutex_lock(&amd_pstate_driver_lock); 1192 ret = amd_pstate_update_status(buf, p ? p - buf : count); 1193 mutex_unlock(&amd_pstate_driver_lock); 1194 1195 return ret < 0 ? ret : count; 1196 } 1197 1198 static ssize_t prefcore_show(struct device *dev, 1199 struct device_attribute *attr, char *buf) 1200 { 1201 return sysfs_emit(buf, "%s\n", str_enabled_disabled(amd_pstate_prefcore)); 1202 } 1203 1204 cpufreq_freq_attr_ro(amd_pstate_max_freq); 1205 cpufreq_freq_attr_ro(amd_pstate_lowest_nonlinear_freq); 1206 1207 cpufreq_freq_attr_ro(amd_pstate_highest_perf); 1208 cpufreq_freq_attr_ro(amd_pstate_hw_prefcore); 1209 cpufreq_freq_attr_rw(energy_performance_preference); 1210 cpufreq_freq_attr_ro(energy_performance_available_preferences); 1211 static DEVICE_ATTR_RW(status); 1212 static DEVICE_ATTR_RO(prefcore); 1213 1214 static struct freq_attr *amd_pstate_attr[] = { 1215 &amd_pstate_max_freq, 1216 &amd_pstate_lowest_nonlinear_freq, 1217 &amd_pstate_highest_perf, 1218 &amd_pstate_hw_prefcore, 1219 NULL, 1220 }; 1221 1222 static struct freq_attr *amd_pstate_epp_attr[] = { 1223 &amd_pstate_max_freq, 1224 &amd_pstate_lowest_nonlinear_freq, 1225 &amd_pstate_highest_perf, 1226 &amd_pstate_hw_prefcore, 1227 &energy_performance_preference, 1228 &energy_performance_available_preferences, 1229 NULL, 1230 }; 1231 1232 static struct attribute *pstate_global_attributes[] = { 1233 &dev_attr_status.attr, 1234 &dev_attr_prefcore.attr, 1235 NULL 1236 }; 1237 1238 static const struct attribute_group amd_pstate_global_attr_group = { 1239 .name = "amd_pstate", 1240 .attrs = pstate_global_attributes, 1241 }; 1242 1243 static bool amd_pstate_acpi_pm_profile_server(void) 1244 { 1245 switch (acpi_gbl_FADT.preferred_profile) { 1246 case PM_ENTERPRISE_SERVER: 1247 case PM_SOHO_SERVER: 1248 case PM_PERFORMANCE_SERVER: 1249 return true; 1250 } 1251 return false; 1252 } 1253 1254 static bool amd_pstate_acpi_pm_profile_undefined(void) 1255 { 1256 if (acpi_gbl_FADT.preferred_profile == PM_UNSPECIFIED) 1257 return true; 1258 if (acpi_gbl_FADT.preferred_profile >= NR_PM_PROFILES) 1259 return true; 1260 return false; 1261 } 1262 1263 static int amd_pstate_epp_cpu_init(struct cpufreq_policy *policy) 1264 { 1265 int min_freq, max_freq, nominal_freq, lowest_nonlinear_freq, ret; 1266 struct amd_cpudata *cpudata; 1267 struct device *dev; 1268 u64 value; 1269 1270 /* 1271 * Resetting PERF_CTL_MSR will put the CPU in P0 frequency, 1272 * which is ideal for initialization process. 1273 */ 1274 amd_perf_ctl_reset(policy->cpu); 1275 dev = get_cpu_device(policy->cpu); 1276 if (!dev) 1277 return -ENODEV; 1278 1279 cpudata = kzalloc(sizeof(*cpudata), GFP_KERNEL); 1280 if (!cpudata) 1281 return -ENOMEM; 1282 1283 cpudata->cpu = policy->cpu; 1284 cpudata->epp_policy = 0; 1285 1286 amd_pstate_init_prefcore(cpudata); 1287 1288 ret = amd_pstate_init_perf(cpudata); 1289 if (ret) 1290 goto free_cpudata1; 1291 1292 min_freq = amd_get_min_freq(cpudata); 1293 max_freq = amd_get_max_freq(cpudata); 1294 nominal_freq = amd_get_nominal_freq(cpudata); 1295 lowest_nonlinear_freq = amd_get_lowest_nonlinear_freq(cpudata); 1296 if (min_freq < 0 || max_freq < 0 || min_freq > max_freq) { 1297 dev_err(dev, "min_freq(%d) or max_freq(%d) value is incorrect\n", 1298 min_freq, max_freq); 1299 ret = -EINVAL; 1300 goto free_cpudata1; 1301 } 1302 1303 policy->cpuinfo.min_freq = min_freq; 1304 policy->cpuinfo.max_freq = max_freq; 1305 /* It will be updated by governor */ 1306 policy->cur = policy->cpuinfo.min_freq; 1307 1308 /* Initial processor data capability frequencies */ 1309 cpudata->max_freq = max_freq; 1310 cpudata->min_freq = min_freq; 1311 cpudata->nominal_freq = nominal_freq; 1312 cpudata->lowest_nonlinear_freq = lowest_nonlinear_freq; 1313 1314 policy->driver_data = cpudata; 1315 1316 cpudata->epp_cached = amd_pstate_get_epp(cpudata, 0); 1317 1318 policy->min = policy->cpuinfo.min_freq; 1319 policy->max = policy->cpuinfo.max_freq; 1320 1321 /* 1322 * Set the policy to provide a valid fallback value in case 1323 * the default cpufreq governor is neither powersave nor performance. 1324 */ 1325 if (amd_pstate_acpi_pm_profile_server() || 1326 amd_pstate_acpi_pm_profile_undefined()) 1327 policy->policy = CPUFREQ_POLICY_PERFORMANCE; 1328 else 1329 policy->policy = CPUFREQ_POLICY_POWERSAVE; 1330 1331 if (boot_cpu_has(X86_FEATURE_CPPC)) { 1332 ret = rdmsrl_on_cpu(cpudata->cpu, MSR_AMD_CPPC_REQ, &value); 1333 if (ret) 1334 return ret; 1335 WRITE_ONCE(cpudata->cppc_req_cached, value); 1336 1337 ret = rdmsrl_on_cpu(cpudata->cpu, MSR_AMD_CPPC_CAP1, &value); 1338 if (ret) 1339 return ret; 1340 WRITE_ONCE(cpudata->cppc_cap1_cached, value); 1341 } 1342 amd_pstate_boost_init(cpudata); 1343 1344 return 0; 1345 1346 free_cpudata1: 1347 kfree(cpudata); 1348 return ret; 1349 } 1350 1351 static int amd_pstate_epp_cpu_exit(struct cpufreq_policy *policy) 1352 { 1353 struct amd_cpudata *cpudata = policy->driver_data; 1354 1355 if (cpudata) { 1356 kfree(cpudata); 1357 policy->driver_data = NULL; 1358 } 1359 1360 pr_debug("CPU %d exiting\n", policy->cpu); 1361 return 0; 1362 } 1363 1364 static void amd_pstate_epp_update_limit(struct cpufreq_policy *policy) 1365 { 1366 struct amd_cpudata *cpudata = policy->driver_data; 1367 u32 max_perf, min_perf, min_limit_perf, max_limit_perf; 1368 u64 value; 1369 s16 epp; 1370 1371 max_perf = READ_ONCE(cpudata->highest_perf); 1372 min_perf = READ_ONCE(cpudata->lowest_perf); 1373 max_limit_perf = div_u64(policy->max * cpudata->highest_perf, cpudata->max_freq); 1374 min_limit_perf = div_u64(policy->min * cpudata->highest_perf, cpudata->max_freq); 1375 1376 WRITE_ONCE(cpudata->max_limit_perf, max_limit_perf); 1377 WRITE_ONCE(cpudata->min_limit_perf, min_limit_perf); 1378 1379 max_perf = clamp_t(unsigned long, max_perf, cpudata->min_limit_perf, 1380 cpudata->max_limit_perf); 1381 min_perf = clamp_t(unsigned long, min_perf, cpudata->min_limit_perf, 1382 cpudata->max_limit_perf); 1383 value = READ_ONCE(cpudata->cppc_req_cached); 1384 1385 if (cpudata->policy == CPUFREQ_POLICY_PERFORMANCE) 1386 min_perf = min(cpudata->nominal_perf, max_perf); 1387 1388 /* Initial min/max values for CPPC Performance Controls Register */ 1389 value &= ~AMD_CPPC_MIN_PERF(~0L); 1390 value |= AMD_CPPC_MIN_PERF(min_perf); 1391 1392 value &= ~AMD_CPPC_MAX_PERF(~0L); 1393 value |= AMD_CPPC_MAX_PERF(max_perf); 1394 1395 /* CPPC EPP feature require to set zero to the desire perf bit */ 1396 value &= ~AMD_CPPC_DES_PERF(~0L); 1397 value |= AMD_CPPC_DES_PERF(0); 1398 1399 cpudata->epp_policy = cpudata->policy; 1400 1401 /* Get BIOS pre-defined epp value */ 1402 epp = amd_pstate_get_epp(cpudata, value); 1403 if (epp < 0) { 1404 /** 1405 * This return value can only be negative for shared_memory 1406 * systems where EPP register read/write not supported. 1407 */ 1408 return; 1409 } 1410 1411 if (cpudata->policy == CPUFREQ_POLICY_PERFORMANCE) 1412 epp = 0; 1413 1414 /* Set initial EPP value */ 1415 if (boot_cpu_has(X86_FEATURE_CPPC)) { 1416 value &= ~GENMASK_ULL(31, 24); 1417 value |= (u64)epp << 24; 1418 } 1419 1420 WRITE_ONCE(cpudata->cppc_req_cached, value); 1421 amd_pstate_set_epp(cpudata, epp); 1422 } 1423 1424 static int amd_pstate_epp_set_policy(struct cpufreq_policy *policy) 1425 { 1426 struct amd_cpudata *cpudata = policy->driver_data; 1427 1428 if (!policy->cpuinfo.max_freq) 1429 return -ENODEV; 1430 1431 pr_debug("set_policy: cpuinfo.max %u policy->max %u\n", 1432 policy->cpuinfo.max_freq, policy->max); 1433 1434 cpudata->policy = policy->policy; 1435 1436 amd_pstate_epp_update_limit(policy); 1437 1438 return 0; 1439 } 1440 1441 static void amd_pstate_epp_reenable(struct amd_cpudata *cpudata) 1442 { 1443 struct cppc_perf_ctrls perf_ctrls; 1444 u64 value, max_perf; 1445 int ret; 1446 1447 ret = amd_pstate_enable(true); 1448 if (ret) 1449 pr_err("failed to enable amd pstate during resume, return %d\n", ret); 1450 1451 value = READ_ONCE(cpudata->cppc_req_cached); 1452 max_perf = READ_ONCE(cpudata->highest_perf); 1453 1454 if (boot_cpu_has(X86_FEATURE_CPPC)) { 1455 wrmsrl_on_cpu(cpudata->cpu, MSR_AMD_CPPC_REQ, value); 1456 } else { 1457 perf_ctrls.max_perf = max_perf; 1458 perf_ctrls.energy_perf = AMD_CPPC_ENERGY_PERF_PREF(cpudata->epp_cached); 1459 cppc_set_perf(cpudata->cpu, &perf_ctrls); 1460 } 1461 } 1462 1463 static int amd_pstate_epp_cpu_online(struct cpufreq_policy *policy) 1464 { 1465 struct amd_cpudata *cpudata = policy->driver_data; 1466 1467 pr_debug("AMD CPU Core %d going online\n", cpudata->cpu); 1468 1469 if (cppc_state == AMD_PSTATE_ACTIVE) { 1470 amd_pstate_epp_reenable(cpudata); 1471 cpudata->suspended = false; 1472 } 1473 1474 return 0; 1475 } 1476 1477 static void amd_pstate_epp_offline(struct cpufreq_policy *policy) 1478 { 1479 struct amd_cpudata *cpudata = policy->driver_data; 1480 struct cppc_perf_ctrls perf_ctrls; 1481 int min_perf; 1482 u64 value; 1483 1484 min_perf = READ_ONCE(cpudata->lowest_perf); 1485 value = READ_ONCE(cpudata->cppc_req_cached); 1486 1487 mutex_lock(&amd_pstate_limits_lock); 1488 if (boot_cpu_has(X86_FEATURE_CPPC)) { 1489 cpudata->epp_policy = CPUFREQ_POLICY_UNKNOWN; 1490 1491 /* Set max perf same as min perf */ 1492 value &= ~AMD_CPPC_MAX_PERF(~0L); 1493 value |= AMD_CPPC_MAX_PERF(min_perf); 1494 value &= ~AMD_CPPC_MIN_PERF(~0L); 1495 value |= AMD_CPPC_MIN_PERF(min_perf); 1496 wrmsrl_on_cpu(cpudata->cpu, MSR_AMD_CPPC_REQ, value); 1497 } else { 1498 perf_ctrls.desired_perf = 0; 1499 perf_ctrls.max_perf = min_perf; 1500 perf_ctrls.energy_perf = AMD_CPPC_ENERGY_PERF_PREF(HWP_EPP_BALANCE_POWERSAVE); 1501 cppc_set_perf(cpudata->cpu, &perf_ctrls); 1502 } 1503 mutex_unlock(&amd_pstate_limits_lock); 1504 } 1505 1506 static int amd_pstate_epp_cpu_offline(struct cpufreq_policy *policy) 1507 { 1508 struct amd_cpudata *cpudata = policy->driver_data; 1509 1510 pr_debug("AMD CPU Core %d going offline\n", cpudata->cpu); 1511 1512 if (cpudata->suspended) 1513 return 0; 1514 1515 if (cppc_state == AMD_PSTATE_ACTIVE) 1516 amd_pstate_epp_offline(policy); 1517 1518 return 0; 1519 } 1520 1521 static int amd_pstate_epp_verify_policy(struct cpufreq_policy_data *policy) 1522 { 1523 cpufreq_verify_within_cpu_limits(policy); 1524 pr_debug("policy_max =%d, policy_min=%d\n", policy->max, policy->min); 1525 return 0; 1526 } 1527 1528 static int amd_pstate_epp_suspend(struct cpufreq_policy *policy) 1529 { 1530 struct amd_cpudata *cpudata = policy->driver_data; 1531 int ret; 1532 1533 /* avoid suspending when EPP is not enabled */ 1534 if (cppc_state != AMD_PSTATE_ACTIVE) 1535 return 0; 1536 1537 /* set this flag to avoid setting core offline*/ 1538 cpudata->suspended = true; 1539 1540 /* disable CPPC in lowlevel firmware */ 1541 ret = amd_pstate_enable(false); 1542 if (ret) 1543 pr_err("failed to suspend, return %d\n", ret); 1544 1545 return 0; 1546 } 1547 1548 static int amd_pstate_epp_resume(struct cpufreq_policy *policy) 1549 { 1550 struct amd_cpudata *cpudata = policy->driver_data; 1551 1552 if (cpudata->suspended) { 1553 mutex_lock(&amd_pstate_limits_lock); 1554 1555 /* enable amd pstate from suspend state*/ 1556 amd_pstate_epp_reenable(cpudata); 1557 1558 mutex_unlock(&amd_pstate_limits_lock); 1559 1560 cpudata->suspended = false; 1561 } 1562 1563 return 0; 1564 } 1565 1566 static struct cpufreq_driver amd_pstate_driver = { 1567 .flags = CPUFREQ_CONST_LOOPS | CPUFREQ_NEED_UPDATE_LIMITS, 1568 .verify = amd_pstate_verify, 1569 .target = amd_pstate_target, 1570 .fast_switch = amd_pstate_fast_switch, 1571 .init = amd_pstate_cpu_init, 1572 .exit = amd_pstate_cpu_exit, 1573 .suspend = amd_pstate_cpu_suspend, 1574 .resume = amd_pstate_cpu_resume, 1575 .set_boost = amd_pstate_set_boost, 1576 .name = "amd-pstate", 1577 .attr = amd_pstate_attr, 1578 }; 1579 1580 static struct cpufreq_driver amd_pstate_epp_driver = { 1581 .flags = CPUFREQ_CONST_LOOPS, 1582 .verify = amd_pstate_epp_verify_policy, 1583 .setpolicy = amd_pstate_epp_set_policy, 1584 .init = amd_pstate_epp_cpu_init, 1585 .exit = amd_pstate_epp_cpu_exit, 1586 .offline = amd_pstate_epp_cpu_offline, 1587 .online = amd_pstate_epp_cpu_online, 1588 .suspend = amd_pstate_epp_suspend, 1589 .resume = amd_pstate_epp_resume, 1590 .name = "amd-pstate-epp", 1591 .attr = amd_pstate_epp_attr, 1592 }; 1593 1594 static int __init amd_pstate_set_driver(int mode_idx) 1595 { 1596 if (mode_idx >= AMD_PSTATE_DISABLE && mode_idx < AMD_PSTATE_MAX) { 1597 cppc_state = mode_idx; 1598 if (cppc_state == AMD_PSTATE_DISABLE) 1599 pr_info("driver is explicitly disabled\n"); 1600 1601 if (cppc_state == AMD_PSTATE_ACTIVE) 1602 current_pstate_driver = &amd_pstate_epp_driver; 1603 1604 if (cppc_state == AMD_PSTATE_PASSIVE || cppc_state == AMD_PSTATE_GUIDED) 1605 current_pstate_driver = &amd_pstate_driver; 1606 1607 return 0; 1608 } 1609 1610 return -EINVAL; 1611 } 1612 1613 static int __init amd_pstate_init(void) 1614 { 1615 struct device *dev_root; 1616 int ret; 1617 1618 if (boot_cpu_data.x86_vendor != X86_VENDOR_AMD) 1619 return -ENODEV; 1620 1621 if (!acpi_cpc_valid()) { 1622 pr_warn_once("the _CPC object is not present in SBIOS or ACPI disabled\n"); 1623 return -ENODEV; 1624 } 1625 1626 /* don't keep reloading if cpufreq_driver exists */ 1627 if (cpufreq_get_current_driver()) 1628 return -EEXIST; 1629 1630 switch (cppc_state) { 1631 case AMD_PSTATE_UNDEFINED: 1632 /* Disable on the following configs by default: 1633 * 1. Undefined platforms 1634 * 2. Server platforms 1635 * 3. Shared memory designs 1636 */ 1637 if (amd_pstate_acpi_pm_profile_undefined() || 1638 amd_pstate_acpi_pm_profile_server() || 1639 !boot_cpu_has(X86_FEATURE_CPPC)) { 1640 pr_info("driver load is disabled, boot with specific mode to enable this\n"); 1641 return -ENODEV; 1642 } 1643 ret = amd_pstate_set_driver(CONFIG_X86_AMD_PSTATE_DEFAULT_MODE); 1644 if (ret) 1645 return ret; 1646 break; 1647 case AMD_PSTATE_DISABLE: 1648 return -ENODEV; 1649 case AMD_PSTATE_PASSIVE: 1650 case AMD_PSTATE_ACTIVE: 1651 case AMD_PSTATE_GUIDED: 1652 break; 1653 default: 1654 return -EINVAL; 1655 } 1656 1657 /* capability check */ 1658 if (boot_cpu_has(X86_FEATURE_CPPC)) { 1659 pr_debug("AMD CPPC MSR based functionality is supported\n"); 1660 if (cppc_state != AMD_PSTATE_ACTIVE) 1661 current_pstate_driver->adjust_perf = amd_pstate_adjust_perf; 1662 } else { 1663 pr_debug("AMD CPPC shared memory based functionality is supported\n"); 1664 static_call_update(amd_pstate_enable, cppc_enable); 1665 static_call_update(amd_pstate_init_perf, cppc_init_perf); 1666 static_call_update(amd_pstate_update_perf, cppc_update_perf); 1667 } 1668 1669 /* enable amd pstate feature */ 1670 ret = amd_pstate_enable(true); 1671 if (ret) { 1672 pr_err("failed to enable with return %d\n", ret); 1673 return ret; 1674 } 1675 1676 ret = cpufreq_register_driver(current_pstate_driver); 1677 if (ret) 1678 pr_err("failed to register with return %d\n", ret); 1679 1680 dev_root = bus_get_dev_root(&cpu_subsys); 1681 if (dev_root) { 1682 ret = sysfs_create_group(&dev_root->kobj, &amd_pstate_global_attr_group); 1683 put_device(dev_root); 1684 if (ret) { 1685 pr_err("sysfs attribute export failed with error %d.\n", ret); 1686 goto global_attr_free; 1687 } 1688 } 1689 1690 return ret; 1691 1692 global_attr_free: 1693 cpufreq_unregister_driver(current_pstate_driver); 1694 return ret; 1695 } 1696 device_initcall(amd_pstate_init); 1697 1698 static int __init amd_pstate_param(char *str) 1699 { 1700 size_t size; 1701 int mode_idx; 1702 1703 if (!str) 1704 return -EINVAL; 1705 1706 size = strlen(str); 1707 mode_idx = get_mode_idx_from_str(str, size); 1708 1709 return amd_pstate_set_driver(mode_idx); 1710 } 1711 1712 static int __init amd_prefcore_param(char *str) 1713 { 1714 if (!strcmp(str, "disable")) 1715 amd_pstate_prefcore = false; 1716 1717 return 0; 1718 } 1719 1720 early_param("amd_pstate", amd_pstate_param); 1721 early_param("amd_prefcore", amd_prefcore_param); 1722 1723 MODULE_AUTHOR("Huang Rui <ray.huang@amd.com>"); 1724 MODULE_DESCRIPTION("AMD Processor P-state Frequency Driver"); 1725