drivers/cpufreq/amd-pstate.c

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