drivers/cpufreq/acpi-cpufreq.c

   1 // SPDX-License-Identifier: GPL-2.0-or-later
   2 /*
   3  * acpi-cpufreq.c - ACPI Processor P-States Driver
   4  *
   5  *  Copyright (C) 2001, 2002 Andy Grover <andrew.grover@intel.com>
   6  *  Copyright (C) 2001, 2002 Paul Diefenbaugh <paul.s.diefenbaugh@intel.com>
   7  *  Copyright (C) 2002 - 2004 Dominik Brodowski <linux@brodo.de>
   8  *  Copyright (C) 2006       Denis Sadykov <denis.m.sadykov@intel.com>
   9  */
  10
  11 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  12
  13 #include <linux/kernel.h>
  14 #include <linux/module.h>
  15 #include <linux/init.h>
  16 #include <linux/smp.h>
  17 #include <linux/sched.h>
  18 #include <linux/cpufreq.h>
  19 #include <linux/compiler.h>
  20 #include <linux/dmi.h>
  21 #include <linux/slab.h>
  22
  23 #include <linux/acpi.h>
  24 #include <linux/io.h>
  25 #include <linux/delay.h>
  26 #include <linux/uaccess.h>
  27
  28 #include <acpi/processor.h>
  29 #include <acpi/cppc_acpi.h>
  30
  31 #include <asm/msr.h>
  32 #include <asm/processor.h>
  33 #include <asm/cpufeature.h>
  34 #include <asm/cpu_device_id.h>
  35
  36 MODULE_AUTHOR("Paul Diefenbaugh, Dominik Brodowski");
  37 MODULE_DESCRIPTION("ACPI Processor P-States Driver");
  38 MODULE_LICENSE("GPL");
  39
  40 enum {
  41         UNDEFINED_CAPABLE = 0,
  42         SYSTEM_INTEL_MSR_CAPABLE,
  43         SYSTEM_AMD_MSR_CAPABLE,
  44         SYSTEM_IO_CAPABLE,
  45 };
  46
  47 #define INTEL_MSR_RANGE         (0xffff)
  48 #define AMD_MSR_RANGE           (0x7)
  49 #define HYGON_MSR_RANGE         (0x7)
  50
  51 #define MSR_K7_HWCR_CPB_DIS     (1ULL << 25)
  52
  53 struct acpi_cpufreq_data {
  54         unsigned int resume;
  55         unsigned int cpu_feature;
  56         unsigned int acpi_perf_cpu;
  57         cpumask_var_t freqdomain_cpus;
  58         void (*cpu_freq_write)(struct acpi_pct_register *reg, u32 val);
  59         u32 (*cpu_freq_read)(struct acpi_pct_register *reg);
  60 };
  61
  62 /* acpi_perf_data is a pointer to percpu data. */
  63 static struct acpi_processor_performance __percpu *acpi_perf_data;
  64
  65 static inline struct acpi_processor_performance *to_perf_data(struct acpi_cpufreq_data *data)
  66 {
  67         return per_cpu_ptr(acpi_perf_data, data->acpi_perf_cpu);
  68 }
  69
  70 static struct cpufreq_driver acpi_cpufreq_driver;
  71
  72 static unsigned int acpi_pstate_strict;
  73
  74 static bool boost_state(unsigned int cpu)
  75 {
  76         u32 lo, hi;
  77         u64 msr;
  78
  79         switch (boot_cpu_data.x86_vendor) {
  80         case X86_VENDOR_INTEL:
  81                 rdmsr_on_cpu(cpu, MSR_IA32_MISC_ENABLE, &lo, &hi);
  82                 msr = lo | ((u64)hi << 32);
  83                 return !(msr & MSR_IA32_MISC_ENABLE_TURBO_DISABLE);
  84         case X86_VENDOR_HYGON:
  85         case X86_VENDOR_AMD:
  86                 rdmsr_on_cpu(cpu, MSR_K7_HWCR, &lo, &hi);
  87                 msr = lo | ((u64)hi << 32);
  88                 return !(msr & MSR_K7_HWCR_CPB_DIS);
  89         }
  90         return false;
  91 }
  92
  93 static int boost_set_msr(bool enable)
  94 {
  95         u32 msr_addr;
  96         u64 msr_mask, val;
  97
  98         switch (boot_cpu_data.x86_vendor) {
  99         case X86_VENDOR_INTEL:
 100                 msr_addr = MSR_IA32_MISC_ENABLE;
 101                 msr_mask = MSR_IA32_MISC_ENABLE_TURBO_DISABLE;
 102                 break;
 103         case X86_VENDOR_HYGON:
 104         case X86_VENDOR_AMD:
 105                 msr_addr = MSR_K7_HWCR;
 106                 msr_mask = MSR_K7_HWCR_CPB_DIS;
 107                 break;
 108         default:
 109                 return -EINVAL;
 110         }
 111
 112         rdmsrl(msr_addr, val);
 113
 114         if (enable)
 115                 val &= ~msr_mask;
 116         else
 117                 val |= msr_mask;
 118
 119         wrmsrl(msr_addr, val);
 120         return 0;
 121 }
 122
 123 static void boost_set_msr_each(void *p_en)
 124 {
 125         bool enable = (bool) p_en;
 126
 127         boost_set_msr(enable);
 128 }
 129
 130 static int set_boost(struct cpufreq_policy *policy, int val)
 131 {
 132         on_each_cpu_mask(policy->cpus, boost_set_msr_each,
 133                          (void *)(long)val, 1);
 134         pr_debug("CPU %*pbl: Core Boosting %sabled.\n",
 135                  cpumask_pr_args(policy->cpus), val ? "en" : "dis");
 136
 137         return 0;
 138 }
 139
 140 static ssize_t show_freqdomain_cpus(struct cpufreq_policy *policy, char *buf)
 141 {
 142         struct acpi_cpufreq_data *data = policy->driver_data;
 143
 144         if (unlikely(!data))
 145                 return -ENODEV;
 146
 147         return cpufreq_show_cpus(data->freqdomain_cpus, buf);
 148 }
 149
 150 cpufreq_freq_attr_ro(freqdomain_cpus);
 151
 152 #ifdef CONFIG_X86_ACPI_CPUFREQ_CPB
 153 static ssize_t store_cpb(struct cpufreq_policy *policy, const char *buf,
 154                          size_t count)
 155 {
 156         int ret;
 157         unsigned int val = 0;
 158
 159         if (!acpi_cpufreq_driver.set_boost)
 160                 return -EINVAL;
 161
 162         ret = kstrtouint(buf, 10, &val);
 163         if (ret || val > 1)
 164                 return -EINVAL;
 165
 166         cpus_read_lock();
 167         set_boost(policy, val);
 168         cpus_read_unlock();
 169
 170         return count;
 171 }
 172
 173 static ssize_t show_cpb(struct cpufreq_policy *policy, char *buf)
 174 {
 175         return sprintf(buf, "%u\n", acpi_cpufreq_driver.boost_enabled);
 176 }
 177
 178 cpufreq_freq_attr_rw(cpb);
 179 #endif
 180
 181 static int check_est_cpu(unsigned int cpuid)
 182 {
 183         struct cpuinfo_x86 *cpu = &cpu_data(cpuid);
 184
 185         return cpu_has(cpu, X86_FEATURE_EST);
 186 }
 187
 188 static int check_amd_hwpstate_cpu(unsigned int cpuid)
 189 {
 190         struct cpuinfo_x86 *cpu = &cpu_data(cpuid);
 191
 192         return cpu_has(cpu, X86_FEATURE_HW_PSTATE);
 193 }
 194
 195 static unsigned extract_io(struct cpufreq_policy *policy, u32 value)
 196 {
 197         struct acpi_cpufreq_data *data = policy->driver_data;
 198         struct acpi_processor_performance *perf;
 199         int i;
 200
 201         perf = to_perf_data(data);
 202
 203         for (i = 0; i < perf->state_count; i++) {
 204                 if (value == perf->states[i].status)
 205                         return policy->freq_table[i].frequency;
 206         }
 207         return 0;
 208 }
 209
 210 static unsigned extract_msr(struct cpufreq_policy *policy, u32 msr)
 211 {
 212         struct acpi_cpufreq_data *data = policy->driver_data;
 213         struct cpufreq_frequency_table *pos;
 214         struct acpi_processor_performance *perf;
 215
 216         if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD)
 217                 msr &= AMD_MSR_RANGE;
 218         else if (boot_cpu_data.x86_vendor == X86_VENDOR_HYGON)
 219                 msr &= HYGON_MSR_RANGE;
 220         else
 221                 msr &= INTEL_MSR_RANGE;
 222
 223         perf = to_perf_data(data);
 224
 225         cpufreq_for_each_entry(pos, policy->freq_table)
 226                 if (msr == perf->states[pos->driver_data].status)
 227                         return pos->frequency;
 228         return policy->freq_table[0].frequency;
 229 }
 230
 231 static unsigned extract_freq(struct cpufreq_policy *policy, u32 val)
 232 {
 233         struct acpi_cpufreq_data *data = policy->driver_data;
 234
 235         switch (data->cpu_feature) {
 236         case SYSTEM_INTEL_MSR_CAPABLE:
 237         case SYSTEM_AMD_MSR_CAPABLE:
 238                 return extract_msr(policy, val);
 239         case SYSTEM_IO_CAPABLE:
 240                 return extract_io(policy, val);
 241         default:
 242                 return 0;
 243         }
 244 }
 245
 246 static u32 cpu_freq_read_intel(struct acpi_pct_register *not_used)
 247 {
 248         u32 val, dummy __always_unused;
 249
 250         rdmsr(MSR_IA32_PERF_CTL, val, dummy);
 251         return val;
 252 }
 253
 254 static void cpu_freq_write_intel(struct acpi_pct_register *not_used, u32 val)
 255 {
 256         u32 lo, hi;
 257
 258         rdmsr(MSR_IA32_PERF_CTL, lo, hi);
 259         lo = (lo & ~INTEL_MSR_RANGE) | (val & INTEL_MSR_RANGE);
 260         wrmsr(MSR_IA32_PERF_CTL, lo, hi);
 261 }
 262
 263 static u32 cpu_freq_read_amd(struct acpi_pct_register *not_used)
 264 {
 265         u32 val, dummy __always_unused;
 266
 267         rdmsr(MSR_AMD_PERF_CTL, val, dummy);
 268         return val;
 269 }
 270
 271 static void cpu_freq_write_amd(struct acpi_pct_register *not_used, u32 val)
 272 {
 273         wrmsr(MSR_AMD_PERF_CTL, val, 0);
 274 }
 275
 276 static u32 cpu_freq_read_io(struct acpi_pct_register *reg)
 277 {
 278         u32 val;
 279
 280         acpi_os_read_port(reg->address, &val, reg->bit_width);
 281         return val;
 282 }
 283
 284 static void cpu_freq_write_io(struct acpi_pct_register *reg, u32 val)
 285 {
 286         acpi_os_write_port(reg->address, val, reg->bit_width);
 287 }
 288
 289 struct drv_cmd {
 290         struct acpi_pct_register *reg;
 291         u32 val;
 292         union {
 293                 void (*write)(struct acpi_pct_register *reg, u32 val);
 294                 u32 (*read)(struct acpi_pct_register *reg);
 295         } func;
 296 };
 297
 298 /* Called via smp_call_function_single(), on the target CPU */
 299 static void do_drv_read(void *_cmd)
 300 {
 301         struct drv_cmd *cmd = _cmd;
 302
 303         cmd->val = cmd->func.read(cmd->reg);
 304 }
 305
 306 static u32 drv_read(struct acpi_cpufreq_data *data, const struct cpumask *mask)
 307 {
 308         struct acpi_processor_performance *perf = to_perf_data(data);
 309         struct drv_cmd cmd = {
 310                 .reg = &perf->control_register,
 311                 .func.read = data->cpu_freq_read,
 312         };
 313         int err;
 314
 315         err = smp_call_function_any(mask, do_drv_read, &cmd, 1);
 316         WARN_ON_ONCE(err);      /* smp_call_function_any() was buggy? */
 317         return cmd.val;
 318 }
 319
 320 /* Called via smp_call_function_many(), on the target CPUs */
 321 static void do_drv_write(void *_cmd)
 322 {
 323         struct drv_cmd *cmd = _cmd;
 324
 325         cmd->func.write(cmd->reg, cmd->val);
 326 }
 327
 328 static void drv_write(struct acpi_cpufreq_data *data,
 329                       const struct cpumask *mask, u32 val)
 330 {
 331         struct acpi_processor_performance *perf = to_perf_data(data);
 332         struct drv_cmd cmd = {
 333                 .reg = &perf->control_register,
 334                 .val = val,
 335                 .func.write = data->cpu_freq_write,
 336         };
 337         int this_cpu;
 338
 339         this_cpu = get_cpu();
 340         if (cpumask_test_cpu(this_cpu, mask))
 341                 do_drv_write(&cmd);
 342
 343         smp_call_function_many(mask, do_drv_write, &cmd, 1);
 344         put_cpu();
 345 }
 346
 347 static u32 get_cur_val(const struct cpumask *mask, struct acpi_cpufreq_data *data)
 348 {
 349         u32 val;
 350
 351         if (unlikely(cpumask_empty(mask)))
 352                 return 0;
 353
 354         val = drv_read(data, mask);
 355
 356         pr_debug("%s = %u\n", __func__, val);
 357
 358         return val;
 359 }
 360
 361 static unsigned int get_cur_freq_on_cpu(unsigned int cpu)
 362 {
 363         struct acpi_cpufreq_data *data;
 364         struct cpufreq_policy *policy;
 365         unsigned int freq;
 366         unsigned int cached_freq;
 367
 368         pr_debug("%s (%d)\n", __func__, cpu);
 369
 370         policy = cpufreq_cpu_get_raw(cpu);
 371         if (unlikely(!policy))
 372                 return 0;
 373
 374         data = policy->driver_data;
 375         if (unlikely(!data || !policy->freq_table))
 376                 return 0;
 377
 378         cached_freq = policy->freq_table[to_perf_data(data)->state].frequency;
 379         freq = extract_freq(policy, get_cur_val(cpumask_of(cpu), data));
 380         if (freq != cached_freq) {
 381                 /*
 382                  * The dreaded BIOS frequency change behind our back.
 383                  * Force set the frequency on next target call.
 384                  */
 385                 data->resume = 1;
 386         }
 387
 388         pr_debug("cur freq = %u\n", freq);
 389
 390         return freq;
 391 }
 392
 393 static unsigned int check_freqs(struct cpufreq_policy *policy,
 394                                 const struct cpumask *mask, unsigned int freq)
 395 {
 396         struct acpi_cpufreq_data *data = policy->driver_data;
 397         unsigned int cur_freq;
 398         unsigned int i;
 399
 400         for (i = 0; i < 100; i++) {
 401                 cur_freq = extract_freq(policy, get_cur_val(mask, data));
 402                 if (cur_freq == freq)
 403                         return 1;
 404                 udelay(10);
 405         }
 406         return 0;
 407 }
 408
 409 static int acpi_cpufreq_target(struct cpufreq_policy *policy,
 410                                unsigned int index)
 411 {
 412         struct acpi_cpufreq_data *data = policy->driver_data;
 413         struct acpi_processor_performance *perf;
 414         const struct cpumask *mask;
 415         unsigned int next_perf_state = 0; /* Index into perf table */
 416         int result = 0;
 417
 418         if (unlikely(!data)) {
 419                 return -ENODEV;
 420         }
 421
 422         perf = to_perf_data(data);
 423         next_perf_state = policy->freq_table[index].driver_data;
 424         if (perf->state == next_perf_state) {
 425                 if (unlikely(data->resume)) {
 426                         pr_debug("Called after resume, resetting to P%d\n",
 427                                 next_perf_state);
 428                         data->resume = 0;
 429                 } else {
 430                         pr_debug("Already at target state (P%d)\n",
 431                                 next_perf_state);
 432                         return 0;
 433                 }
 434         }
 435
 436         /*
 437          * The core won't allow CPUs to go away until the governor has been
 438          * stopped, so we can rely on the stability of policy->cpus.
 439          */
 440         mask = policy->shared_type == CPUFREQ_SHARED_TYPE_ANY ?
 441                 cpumask_of(policy->cpu) : policy->cpus;
 442
 443         drv_write(data, mask, perf->states[next_perf_state].control);
 444
 445         if (acpi_pstate_strict) {
 446                 if (!check_freqs(policy, mask,
 447                                  policy->freq_table[index].frequency)) {
 448                         pr_debug("%s (%d)\n", __func__, policy->cpu);
 449                         result = -EAGAIN;
 450                 }
 451         }
 452
 453         if (!result)
 454                 perf->state = next_perf_state;
 455
 456         return result;
 457 }
 458
 459 static unsigned int acpi_cpufreq_fast_switch(struct cpufreq_policy *policy,
 460                                              unsigned int target_freq)
 461 {
 462         struct acpi_cpufreq_data *data = policy->driver_data;
 463         struct acpi_processor_performance *perf;
 464         struct cpufreq_frequency_table *entry;
 465         unsigned int next_perf_state, next_freq, index;
 466
 467         /*
 468          * Find the closest frequency above target_freq.
 469          */
 470         if (policy->cached_target_freq == target_freq)
 471                 index = policy->cached_resolved_idx;
 472         else
 473                 index = cpufreq_table_find_index_dl(policy, target_freq,
 474                                                     false);
 475
 476         entry = &policy->freq_table[index];
 477         next_freq = entry->frequency;
 478         next_perf_state = entry->driver_data;
 479
 480         perf = to_perf_data(data);
 481         if (perf->state == next_perf_state) {
 482                 if (unlikely(data->resume))
 483                         data->resume = 0;
 484                 else
 485                         return next_freq;
 486         }
 487
 488         data->cpu_freq_write(&perf->control_register,
 489                              perf->states[next_perf_state].control);
 490         perf->state = next_perf_state;
 491         return next_freq;
 492 }
 493
 494 static unsigned long
 495 acpi_cpufreq_guess_freq(struct acpi_cpufreq_data *data, unsigned int cpu)
 496 {
 497         struct acpi_processor_performance *perf;
 498
 499         perf = to_perf_data(data);
 500         if (cpu_khz) {
 501                 /* search the closest match to cpu_khz */
 502                 unsigned int i;
 503                 unsigned long freq;
 504                 unsigned long freqn = perf->states[0].core_frequency * 1000;
 505
 506                 for (i = 0; i < (perf->state_count-1); i++) {
 507                         freq = freqn;
 508                         freqn = perf->states[i+1].core_frequency * 1000;
 509                         if ((2 * cpu_khz) > (freqn + freq)) {
 510                                 perf->state = i;
 511                                 return freq;
 512                         }
 513                 }
 514                 perf->state = perf->state_count-1;
 515                 return freqn;
 516         } else {
 517                 /* assume CPU is at P0... */
 518                 perf->state = 0;
 519                 return perf->states[0].core_frequency * 1000;
 520         }
 521 }
 522
 523 static void free_acpi_perf_data(void)
 524 {
 525         unsigned int i;
 526
 527         /* Freeing a NULL pointer is OK, and alloc_percpu zeroes. */
 528         for_each_possible_cpu(i)
 529                 free_cpumask_var(per_cpu_ptr(acpi_perf_data, i)
 530                                  ->shared_cpu_map);
 531         free_percpu(acpi_perf_data);
 532 }
 533
 534 static int cpufreq_boost_online(unsigned int cpu)
 535 {
 536         /*
 537          * On the CPU_UP path we simply keep the boost-disable flag
 538          * in sync with the current global state.
 539          */
 540         return boost_set_msr(acpi_cpufreq_driver.boost_enabled);
 541 }
 542
 543 static int cpufreq_boost_down_prep(unsigned int cpu)
 544 {
 545         /*
 546          * Clear the boost-disable bit on the CPU_DOWN path so that
 547          * this cpu cannot block the remaining ones from boosting.
 548          */
 549         return boost_set_msr(1);
 550 }
 551
 552 /*
 553  * acpi_cpufreq_early_init - initialize ACPI P-States library
 554  *
 555  * Initialize the ACPI P-States library (drivers/acpi/processor_perflib.c)
 556  * in order to determine correct frequency and voltage pairings. We can
 557  * do _PDC and _PSD and find out the processor dependency for the
 558  * actual init that will happen later...
 559  */
 560 static int __init acpi_cpufreq_early_init(void)
 561 {
 562         unsigned int i;
 563         pr_debug("%s\n", __func__);
 564
 565         acpi_perf_data = alloc_percpu(struct acpi_processor_performance);
 566         if (!acpi_perf_data) {
 567                 pr_debug("Memory allocation error for acpi_perf_data.\n");
 568                 return -ENOMEM;
 569         }
 570         for_each_possible_cpu(i) {
 571                 if (!zalloc_cpumask_var_node(
 572                         &per_cpu_ptr(acpi_perf_data, i)->shared_cpu_map,
 573                         GFP_KERNEL, cpu_to_node(i))) {
 574
 575                         /* Freeing a NULL pointer is OK: alloc_percpu zeroes. */
 576                         free_acpi_perf_data();
 577                         return -ENOMEM;
 578                 }
 579         }
 580
 581         /* Do initialization in ACPI core */
 582         acpi_processor_preregister_performance(acpi_perf_data);
 583         return 0;
 584 }
 585
 586 #ifdef CONFIG_SMP
 587 /*
 588  * Some BIOSes do SW_ANY coordination internally, either set it up in hw
 589  * or do it in BIOS firmware and won't inform about it to OS. If not
 590  * detected, this has a side effect of making CPU run at a different speed
 591  * than OS intended it to run at. Detect it and handle it cleanly.
 592  */
 593 static int bios_with_sw_any_bug;
 594
 595 static int sw_any_bug_found(const struct dmi_system_id *d)
 596 {
 597         bios_with_sw_any_bug = 1;
 598         return 0;
 599 }
 600
 601 static const struct dmi_system_id sw_any_bug_dmi_table[] = {
 602         {
 603                 .callback = sw_any_bug_found,
 604                 .ident = "Supermicro Server X6DLP",
 605                 .matches = {
 606                         DMI_MATCH(DMI_SYS_VENDOR, "Supermicro"),
 607                         DMI_MATCH(DMI_BIOS_VERSION, "080010"),
 608                         DMI_MATCH(DMI_PRODUCT_NAME, "X6DLP"),
 609                 },
 610         },
 611         { }
 612 };
 613
 614 static int acpi_cpufreq_blacklist(struct cpuinfo_x86 *c)
 615 {
 616         /* Intel Xeon Processor 7100 Series Specification Update
 617          * https://www.intel.com/Assets/PDF/specupdate/314554.pdf
 618          * AL30: A Machine Check Exception (MCE) Occurring during an
 619          * Enhanced Intel SpeedStep Technology Ratio Change May Cause
 620          * Both Processor Cores to Lock Up. */
 621         if (c->x86_vendor == X86_VENDOR_INTEL) {
 622                 if ((c->x86 == 15) &&
 623                     (c->x86_model == 6) &&
 624                     (c->x86_stepping == 8)) {
 625                         pr_info("Intel(R) Xeon(R) 7100 Errata AL30, processors may lock up on frequency changes: disabling acpi-cpufreq\n");
 626                         return -ENODEV;
 627                     }
 628                 }
 629         return 0;
 630 }
 631 #endif
 632
 633 #ifdef CONFIG_ACPI_CPPC_LIB
 634 static u64 get_max_boost_ratio(unsigned int cpu)
 635 {
 636         struct cppc_perf_caps perf_caps;
 637         u64 highest_perf, nominal_perf;
 638         int ret;
 639
 640         if (acpi_pstate_strict)
 641                 return 0;
 642
 643         ret = cppc_get_perf_caps(cpu, &perf_caps);
 644         if (ret) {
 645                 pr_debug("CPU%d: Unable to get performance capabilities (%d)\n",
 646                          cpu, ret);
 647                 return 0;
 648         }
 649
 650         if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD)
 651                 highest_perf = amd_get_highest_perf();
 652         else
 653                 highest_perf = perf_caps.highest_perf;
 654
 655         nominal_perf = perf_caps.nominal_perf;
 656
 657         if (!highest_perf || !nominal_perf) {
 658                 pr_debug("CPU%d: highest or nominal performance missing\n", cpu);
 659                 return 0;
 660         }
 661
 662         if (highest_perf < nominal_perf) {
 663                 pr_debug("CPU%d: nominal performance above highest\n", cpu);
 664                 return 0;
 665         }
 666
 667         return div_u64(highest_perf << SCHED_CAPACITY_SHIFT, nominal_perf);
 668 }
 669 #else
 670 static inline u64 get_max_boost_ratio(unsigned int cpu) { return 0; }
 671 #endif
 672
 673 static int acpi_cpufreq_cpu_init(struct cpufreq_policy *policy)
 674 {
 675         struct cpufreq_frequency_table *freq_table;
 676         struct acpi_processor_performance *perf;
 677         struct acpi_cpufreq_data *data;
 678         unsigned int cpu = policy->cpu;
 679         struct cpuinfo_x86 *c = &cpu_data(cpu);
 680         unsigned int valid_states = 0;
 681         unsigned int result = 0;
 682         u64 max_boost_ratio;
 683         unsigned int i;
 684 #ifdef CONFIG_SMP
 685         static int blacklisted;
 686 #endif
 687
 688         pr_debug("%s\n", __func__);
 689
 690 #ifdef CONFIG_SMP
 691         if (blacklisted)
 692                 return blacklisted;
 693         blacklisted = acpi_cpufreq_blacklist(c);
 694         if (blacklisted)
 695                 return blacklisted;
 696 #endif
 697
 698         data = kzalloc(sizeof(*data), GFP_KERNEL);
 699         if (!data)
 700                 return -ENOMEM;
 701
 702         if (!zalloc_cpumask_var(&data->freqdomain_cpus, GFP_KERNEL)) {
 703                 result = -ENOMEM;
 704                 goto err_free;
 705         }
 706
 707         perf = per_cpu_ptr(acpi_perf_data, cpu);
 708         data->acpi_perf_cpu = cpu;
 709         policy->driver_data = data;
 710
 711         if (cpu_has(c, X86_FEATURE_CONSTANT_TSC))
 712                 acpi_cpufreq_driver.flags |= CPUFREQ_CONST_LOOPS;
 713
 714         result = acpi_processor_register_performance(perf, cpu);
 715         if (result)
 716                 goto err_free_mask;
 717
 718         policy->shared_type = perf->shared_type;
 719
 720         /*
 721          * Will let policy->cpus know about dependency only when software
 722          * coordination is required.
 723          */
 724         if (policy->shared_type == CPUFREQ_SHARED_TYPE_ALL ||
 725             policy->shared_type == CPUFREQ_SHARED_TYPE_ANY) {
 726                 cpumask_copy(policy->cpus, perf->shared_cpu_map);
 727         }
 728         cpumask_copy(data->freqdomain_cpus, perf->shared_cpu_map);
 729
 730 #ifdef CONFIG_SMP
 731         dmi_check_system(sw_any_bug_dmi_table);
 732         if (bios_with_sw_any_bug && !policy_is_shared(policy)) {
 733                 policy->shared_type = CPUFREQ_SHARED_TYPE_ALL;
 734                 cpumask_copy(policy->cpus, topology_core_cpumask(cpu));
 735         }
 736
 737         if (check_amd_hwpstate_cpu(cpu) && boot_cpu_data.x86 < 0x19 &&
 738             !acpi_pstate_strict) {
 739                 cpumask_clear(policy->cpus);
 740                 cpumask_set_cpu(cpu, policy->cpus);
 741                 cpumask_copy(data->freqdomain_cpus,
 742                              topology_sibling_cpumask(cpu));
 743                 policy->shared_type = CPUFREQ_SHARED_TYPE_HW;
 744                 pr_info_once("overriding BIOS provided _PSD data\n");
 745         }
 746 #endif
 747
 748         /* capability check */
 749         if (perf->state_count <= 1) {
 750                 pr_debug("No P-States\n");
 751                 result = -ENODEV;
 752                 goto err_unreg;
 753         }
 754
 755         if (perf->control_register.space_id != perf->status_register.space_id) {
 756                 result = -ENODEV;
 757                 goto err_unreg;
 758         }
 759
 760         switch (perf->control_register.space_id) {
 761         case ACPI_ADR_SPACE_SYSTEM_IO:
 762                 if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD &&
 763                     boot_cpu_data.x86 == 0xf) {
 764                         pr_debug("AMD K8 systems must use native drivers.\n");
 765                         result = -ENODEV;
 766                         goto err_unreg;
 767                 }
 768                 pr_debug("SYSTEM IO addr space\n");
 769                 data->cpu_feature = SYSTEM_IO_CAPABLE;
 770                 data->cpu_freq_read = cpu_freq_read_io;
 771                 data->cpu_freq_write = cpu_freq_write_io;
 772                 break;
 773         case ACPI_ADR_SPACE_FIXED_HARDWARE:
 774                 pr_debug("HARDWARE addr space\n");
 775                 if (check_est_cpu(cpu)) {
 776                         data->cpu_feature = SYSTEM_INTEL_MSR_CAPABLE;
 777                         data->cpu_freq_read = cpu_freq_read_intel;
 778                         data->cpu_freq_write = cpu_freq_write_intel;
 779                         break;
 780                 }
 781                 if (check_amd_hwpstate_cpu(cpu)) {
 782                         data->cpu_feature = SYSTEM_AMD_MSR_CAPABLE;
 783                         data->cpu_freq_read = cpu_freq_read_amd;
 784                         data->cpu_freq_write = cpu_freq_write_amd;
 785                         break;
 786                 }
 787                 result = -ENODEV;
 788                 goto err_unreg;
 789         default:
 790                 pr_debug("Unknown addr space %d\n",
 791                         (u32) (perf->control_register.space_id));
 792                 result = -ENODEV;
 793                 goto err_unreg;
 794         }
 795
 796         freq_table = kcalloc(perf->state_count + 1, sizeof(*freq_table),
 797                              GFP_KERNEL);
 798         if (!freq_table) {
 799                 result = -ENOMEM;
 800                 goto err_unreg;
 801         }
 802
 803         /* detect transition latency */
 804         policy->cpuinfo.transition_latency = 0;
 805         for (i = 0; i < perf->state_count; i++) {
 806                 if ((perf->states[i].transition_latency * 1000) >
 807                     policy->cpuinfo.transition_latency)
 808                         policy->cpuinfo.transition_latency =
 809                             perf->states[i].transition_latency * 1000;
 810         }
 811
 812         /* Check for high latency (>20uS) from buggy BIOSes, like on T42 */
 813         if (perf->control_register.space_id == ACPI_ADR_SPACE_FIXED_HARDWARE &&
 814             policy->cpuinfo.transition_latency > 20 * 1000) {
 815                 policy->cpuinfo.transition_latency = 20 * 1000;
 816                 pr_info_once("P-state transition latency capped at 20 uS\n");
 817         }
 818
 819         /* table init */
 820         for (i = 0; i < perf->state_count; i++) {
 821                 if (i > 0 && perf->states[i].core_frequency >=
 822                     freq_table[valid_states-1].frequency / 1000)
 823                         continue;
 824
 825                 freq_table[valid_states].driver_data = i;
 826                 freq_table[valid_states].frequency =
 827                     perf->states[i].core_frequency * 1000;
 828                 valid_states++;
 829         }
 830         freq_table[valid_states].frequency = CPUFREQ_TABLE_END;
 831
 832         max_boost_ratio = get_max_boost_ratio(cpu);
 833         if (max_boost_ratio) {
 834                 unsigned int freq = freq_table[0].frequency;
 835
 836                 /*
 837                  * Because the loop above sorts the freq_table entries in the
 838                  * descending order, freq is the maximum frequency in the table.
 839                  * Assume that it corresponds to the CPPC nominal frequency and
 840                  * use it to set cpuinfo.max_freq.
 841                  */
 842                 policy->cpuinfo.max_freq = freq * max_boost_ratio >> SCHED_CAPACITY_SHIFT;
 843         } else {
 844                 /*
 845                  * If the maximum "boost" frequency is unknown, ask the arch
 846                  * scale-invariance code to use the "nominal" performance for
 847                  * CPU utilization scaling so as to prevent the schedutil
 848                  * governor from selecting inadequate CPU frequencies.
 849                  */
 850                 arch_set_max_freq_ratio(true);
 851         }
 852
 853         policy->freq_table = freq_table;
 854         perf->state = 0;
 855
 856         switch (perf->control_register.space_id) {
 857         case ACPI_ADR_SPACE_SYSTEM_IO:
 858                 /*
 859                  * The core will not set policy->cur, because
 860                  * cpufreq_driver->get is NULL, so we need to set it here.
 861                  * However, we have to guess it, because the current speed is
 862                  * unknown and not detectable via IO ports.
 863                  */
 864                 policy->cur = acpi_cpufreq_guess_freq(data, policy->cpu);
 865                 break;
 866         case ACPI_ADR_SPACE_FIXED_HARDWARE:
 867                 acpi_cpufreq_driver.get = get_cur_freq_on_cpu;
 868                 break;
 869         default:
 870                 break;
 871         }
 872
 873         /* notify BIOS that we exist */
 874         acpi_processor_notify_smm(THIS_MODULE);
 875
 876         pr_debug("CPU%u - ACPI performance management activated.\n", cpu);
 877         for (i = 0; i < perf->state_count; i++)
 878                 pr_debug("     %cP%d: %d MHz, %d mW, %d uS\n",
 879                         (i == perf->state ? '*' : ' '), i,
 880                         (u32) perf->states[i].core_frequency,
 881                         (u32) perf->states[i].power,
 882                         (u32) perf->states[i].transition_latency);
 883
 884         /*
 885          * the first call to ->target() should result in us actually
 886          * writing something to the appropriate registers.
 887          */
 888         data->resume = 1;
 889
 890         policy->fast_switch_possible = !acpi_pstate_strict &&
 891                 !(policy_is_shared(policy) && policy->shared_type != CPUFREQ_SHARED_TYPE_ANY);
 892
 893         if (perf->states[0].core_frequency * 1000 != freq_table[0].frequency)
 894                 pr_warn(FW_WARN "P-state 0 is not max freq\n");
 895
 896         return result;
 897
 898 err_unreg:
 899         acpi_processor_unregister_performance(cpu);
 900 err_free_mask:
 901         free_cpumask_var(data->freqdomain_cpus);
 902 err_free:
 903         kfree(data);
 904         policy->driver_data = NULL;
 905
 906         return result;
 907 }
 908
 909 static int acpi_cpufreq_cpu_exit(struct cpufreq_policy *policy)
 910 {
 911         struct acpi_cpufreq_data *data = policy->driver_data;
 912
 913         pr_debug("%s\n", __func__);
 914
 915         policy->fast_switch_possible = false;
 916         policy->driver_data = NULL;
 917         acpi_processor_unregister_performance(data->acpi_perf_cpu);
 918         free_cpumask_var(data->freqdomain_cpus);
 919         kfree(policy->freq_table);
 920         kfree(data);
 921
 922         return 0;
 923 }
 924
 925 static int acpi_cpufreq_resume(struct cpufreq_policy *policy)
 926 {
 927         struct acpi_cpufreq_data *data = policy->driver_data;
 928
 929         pr_debug("%s\n", __func__);
 930
 931         data->resume = 1;
 932
 933         return 0;
 934 }
 935
 936 static struct freq_attr *acpi_cpufreq_attr[] = {
 937         &cpufreq_freq_attr_scaling_available_freqs,
 938         &freqdomain_cpus,
 939 #ifdef CONFIG_X86_ACPI_CPUFREQ_CPB
 940         &cpb,
 941 #endif
 942         NULL,
 943 };
 944
 945 static struct cpufreq_driver acpi_cpufreq_driver = {
 946         .verify         = cpufreq_generic_frequency_table_verify,
 947         .target_index   = acpi_cpufreq_target,
 948         .fast_switch    = acpi_cpufreq_fast_switch,
 949         .bios_limit     = acpi_processor_get_bios_limit,
 950         .init           = acpi_cpufreq_cpu_init,
 951         .exit           = acpi_cpufreq_cpu_exit,
 952         .resume         = acpi_cpufreq_resume,
 953         .name           = "acpi-cpufreq",
 954         .attr           = acpi_cpufreq_attr,
 955 };
 956
 957 static enum cpuhp_state acpi_cpufreq_online;
 958
 959 static void __init acpi_cpufreq_boost_init(void)
 960 {
 961         int ret;
 962
 963         if (!(boot_cpu_has(X86_FEATURE_CPB) || boot_cpu_has(X86_FEATURE_IDA))) {
 964                 pr_debug("Boost capabilities not present in the processor\n");
 965                 return;
 966         }
 967
 968         acpi_cpufreq_driver.set_boost = set_boost;
 969         acpi_cpufreq_driver.boost_enabled = boost_state(0);
 970
 971         /*
 972          * This calls the online callback on all online cpu and forces all
 973          * MSRs to the same value.
 974          */
 975         ret = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "cpufreq/acpi:online",
 976                                 cpufreq_boost_online, cpufreq_boost_down_prep);
 977         if (ret < 0) {
 978                 pr_err("acpi_cpufreq: failed to register hotplug callbacks\n");
 979                 return;
 980         }
 981         acpi_cpufreq_online = ret;
 982 }
 983
 984 static void acpi_cpufreq_boost_exit(void)
 985 {
 986         if (acpi_cpufreq_online > 0)
 987                 cpuhp_remove_state_nocalls(acpi_cpufreq_online);
 988 }
 989
 990 static int __init acpi_cpufreq_init(void)
 991 {
 992         int ret;
 993
 994         if (acpi_disabled)
 995                 return -ENODEV;
 996
 997         /* don't keep reloading if cpufreq_driver exists */
 998         if (cpufreq_get_current_driver())
 999                 return -EEXIST;
1000
1001         pr_debug("%s\n", __func__);
1002
1003         ret = acpi_cpufreq_early_init();
1004         if (ret)
1005                 return ret;
1006
1007 #ifdef CONFIG_X86_ACPI_CPUFREQ_CPB
1008         /* this is a sysfs file with a strange name and an even stranger
1009          * semantic - per CPU instantiation, but system global effect.
1010          * Lets enable it only on AMD CPUs for compatibility reasons and
1011          * only if configured. This is considered legacy code, which
1012          * will probably be removed at some point in the future.
1013          */
1014         if (!check_amd_hwpstate_cpu(0)) {
1015                 struct freq_attr **attr;
1016
1017                 pr_debug("CPB unsupported, do not expose it\n");
1018
1019                 for (attr = acpi_cpufreq_attr; *attr; attr++)
1020                         if (*attr == &cpb) {
1021                                 *attr = NULL;
1022                                 break;
1023                         }
1024         }
1025 #endif
1026         acpi_cpufreq_boost_init();
1027
1028         ret = cpufreq_register_driver(&acpi_cpufreq_driver);
1029         if (ret) {
1030                 free_acpi_perf_data();
1031                 acpi_cpufreq_boost_exit();
1032         }
1033         return ret;
1034 }
1035
1036 static void __exit acpi_cpufreq_exit(void)
1037 {
1038         pr_debug("%s\n", __func__);
1039
1040         acpi_cpufreq_boost_exit();
1041
1042         cpufreq_unregister_driver(&acpi_cpufreq_driver);
1043
1044         free_acpi_perf_data();
1045 }
1046
1047 module_param(acpi_pstate_strict, uint, 0644);
1048 MODULE_PARM_DESC(acpi_pstate_strict,
1049         "value 0 or non-zero. non-zero -> strict ACPI checks are "
1050         "performed during frequency changes.");
1051
1052 late_initcall(acpi_cpufreq_init);
1053 module_exit(acpi_cpufreq_exit);
1054
1055 static const struct x86_cpu_id __maybe_unused acpi_cpufreq_ids[] = {
1056         X86_MATCH_FEATURE(X86_FEATURE_ACPI, NULL),
1057         X86_MATCH_FEATURE(X86_FEATURE_HW_PSTATE, NULL),
1058         {}
1059 };
1060 MODULE_DEVICE_TABLE(x86cpu, acpi_cpufreq_ids);
1061
1062 static const struct acpi_device_id __maybe_unused processor_device_ids[] = {
1063         {ACPI_PROCESSOR_OBJECT_HID, },
1064         {ACPI_PROCESSOR_DEVICE_HID, },
1065         {},
1066 };
1067 MODULE_DEVICE_TABLE(acpi, processor_device_ids);
1068
1069 MODULE_ALIAS("acpi");