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 <[email protected]>
   6  *  Copyright (C) 2001, 2002 Paul Diefenbaugh <[email protected]>
   7  *  Copyright (C) 2002 - 2004 Dominik Brodowski <[email protected]>
   8  *  Copyright (C) 2006       Denis Sadykov <[email protected]>
   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 #include <linux/string_helpers.h>
  23
  24 #include <linux/acpi.h>
  25 #include <linux/io.h>
  26 #include <linux/delay.h>
  27 #include <linux/uaccess.h>
  28
  29 #include <acpi/processor.h>
  30 #include <acpi/cppc_acpi.h>
  31
  32 #include <asm/msr.h>
  33 #include <asm/processor.h>
  34 #include <asm/cpufeature.h>
  35 #include <asm/cpu_device_id.h>
  36
  37 MODULE_AUTHOR("Paul Diefenbaugh, Dominik Brodowski");
  38 MODULE_DESCRIPTION("ACPI Processor P-States Driver");
  39 MODULE_LICENSE("GPL");
  40
  41 enum {
  42         UNDEFINED_CAPABLE = 0,
  43         SYSTEM_INTEL_MSR_CAPABLE,
  44         SYSTEM_AMD_MSR_CAPABLE,
  45         SYSTEM_IO_CAPABLE,
  46 };
  47
  48 #define INTEL_MSR_RANGE         (0xffff)
  49 #define AMD_MSR_RANGE           (0x7)
  50 #define HYGON_MSR_RANGE         (0x7)
  51
  52 #define MSR_K7_HWCR_CPB_DIS     (1ULL << 25)
  53
  54 struct acpi_cpufreq_data {
  55         unsigned int resume;
  56         unsigned int cpu_feature;
  57         unsigned int acpi_perf_cpu;
  58         cpumask_var_t freqdomain_cpus;
  59         void (*cpu_freq_write)(struct acpi_pct_register *reg, u32 val);
  60         u32 (*cpu_freq_read)(struct acpi_pct_register *reg);
  61 };
  62
  63 /* acpi_perf_data is a pointer to percpu data. */
  64 static struct acpi_processor_performance __percpu *acpi_perf_data;
  65
  66 static inline struct acpi_processor_performance *to_perf_data(struct acpi_cpufreq_data *data)
  67 {
  68         return per_cpu_ptr(acpi_perf_data, data->acpi_perf_cpu);
  69 }
  70
  71 static struct cpufreq_driver acpi_cpufreq_driver;
  72
  73 static unsigned int acpi_pstate_strict;
  74
  75 static bool boost_state(unsigned int cpu)
  76 {
  77         u32 lo, hi;
  78         u64 msr;
  79
  80         switch (boot_cpu_data.x86_vendor) {
  81         case X86_VENDOR_INTEL:
  82         case X86_VENDOR_CENTAUR:
  83         case X86_VENDOR_ZHAOXIN:
  84                 rdmsr_on_cpu(cpu, MSR_IA32_MISC_ENABLE, &lo, &hi);
  85                 msr = lo | ((u64)hi << 32);
  86                 return !(msr & MSR_IA32_MISC_ENABLE_TURBO_DISABLE);
  87         case X86_VENDOR_HYGON:
  88         case X86_VENDOR_AMD:
  89                 rdmsr_on_cpu(cpu, MSR_K7_HWCR, &lo, &hi);
  90                 msr = lo | ((u64)hi << 32);
  91                 return !(msr & MSR_K7_HWCR_CPB_DIS);
  92         }
  93         return false;
  94 }
  95
  96 static int boost_set_msr(bool enable)
  97 {
  98         u32 msr_addr;
  99         u64 msr_mask, val;
 100
 101         switch (boot_cpu_data.x86_vendor) {
 102         case X86_VENDOR_INTEL:
 103         case X86_VENDOR_CENTAUR:
 104         case X86_VENDOR_ZHAOXIN:
 105                 msr_addr = MSR_IA32_MISC_ENABLE;
 106                 msr_mask = MSR_IA32_MISC_ENABLE_TURBO_DISABLE;
 107                 break;
 108         case X86_VENDOR_HYGON:
 109         case X86_VENDOR_AMD:
 110                 msr_addr = MSR_K7_HWCR;
 111                 msr_mask = MSR_K7_HWCR_CPB_DIS;
 112                 break;
 113         default:
 114                 return -EINVAL;
 115         }
 116
 117         rdmsrl(msr_addr, val);
 118
 119         if (enable)
 120                 val &= ~msr_mask;
 121         else
 122                 val |= msr_mask;
 123
 124         wrmsrl(msr_addr, val);
 125         return 0;
 126 }
 127
 128 static void boost_set_msr_each(void *p_en)
 129 {
 130         bool enable = (bool) p_en;
 131
 132         boost_set_msr(enable);
 133 }
 134
 135 static int set_boost(struct cpufreq_policy *policy, int val)
 136 {
 137         on_each_cpu_mask(policy->cpus, boost_set_msr_each,
 138                          (void *)(long)val, 1);
 139         pr_debug("CPU %*pbl: Core Boosting %s.\n",
 140                  cpumask_pr_args(policy->cpus), str_enabled_disabled(val));
 141
 142         return 0;
 143 }
 144
 145 static ssize_t show_freqdomain_cpus(struct cpufreq_policy *policy, char *buf)
 146 {
 147         struct acpi_cpufreq_data *data = policy->driver_data;
 148
 149         if (unlikely(!data))
 150                 return -ENODEV;
 151
 152         return cpufreq_show_cpus(data->freqdomain_cpus, buf);
 153 }
 154
 155 cpufreq_freq_attr_ro(freqdomain_cpus);
 156
 157 #ifdef CONFIG_X86_ACPI_CPUFREQ_CPB
 158 static ssize_t store_cpb(struct cpufreq_policy *policy, const char *buf,
 159                          size_t count)
 160 {
 161         int ret;
 162         unsigned int val = 0;
 163
 164         if (!acpi_cpufreq_driver.set_boost)
 165                 return -EINVAL;
 166
 167         ret = kstrtouint(buf, 10, &val);
 168         if (ret || val > 1)
 169                 return -EINVAL;
 170
 171         cpus_read_lock();
 172         set_boost(policy, val);
 173         cpus_read_unlock();
 174
 175         return count;
 176 }
 177
 178 static ssize_t show_cpb(struct cpufreq_policy *policy, char *buf)
 179 {
 180         return sprintf(buf, "%u\n", acpi_cpufreq_driver.boost_enabled);
 181 }
 182
 183 cpufreq_freq_attr_rw(cpb);
 184 #endif
 185
 186 static int check_est_cpu(unsigned int cpuid)
 187 {
 188         struct cpuinfo_x86 *cpu = &cpu_data(cpuid);
 189
 190         return cpu_has(cpu, X86_FEATURE_EST);
 191 }
 192
 193 static int check_amd_hwpstate_cpu(unsigned int cpuid)
 194 {
 195         struct cpuinfo_x86 *cpu = &cpu_data(cpuid);
 196
 197         return cpu_has(cpu, X86_FEATURE_HW_PSTATE);
 198 }
 199
 200 static unsigned extract_io(struct cpufreq_policy *policy, u32 value)
 201 {
 202         struct acpi_cpufreq_data *data = policy->driver_data;
 203         struct acpi_processor_performance *perf;
 204         int i;
 205
 206         perf = to_perf_data(data);
 207
 208         for (i = 0; i < perf->state_count; i++) {
 209                 if (value == perf->states[i].status)
 210                         return policy->freq_table[i].frequency;
 211         }
 212         return 0;
 213 }
 214
 215 static unsigned extract_msr(struct cpufreq_policy *policy, u32 msr)
 216 {
 217         struct acpi_cpufreq_data *data = policy->driver_data;
 218         struct cpufreq_frequency_table *pos;
 219         struct acpi_processor_performance *perf;
 220
 221         if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD)
 222                 msr &= AMD_MSR_RANGE;
 223         else if (boot_cpu_data.x86_vendor == X86_VENDOR_HYGON)
 224                 msr &= HYGON_MSR_RANGE;
 225         else
 226                 msr &= INTEL_MSR_RANGE;
 227
 228         perf = to_perf_data(data);
 229
 230         cpufreq_for_each_entry(pos, policy->freq_table)
 231                 if (msr == perf->states[pos->driver_data].status)
 232                         return pos->frequency;
 233         return policy->freq_table[0].frequency;
 234 }
 235
 236 static unsigned extract_freq(struct cpufreq_policy *policy, u32 val)
 237 {
 238         struct acpi_cpufreq_data *data = policy->driver_data;
 239
 240         switch (data->cpu_feature) {
 241         case SYSTEM_INTEL_MSR_CAPABLE:
 242         case SYSTEM_AMD_MSR_CAPABLE:
 243                 return extract_msr(policy, val);
 244         case SYSTEM_IO_CAPABLE:
 245                 return extract_io(policy, val);
 246         default:
 247                 return 0;
 248         }
 249 }
 250
 251 static u32 cpu_freq_read_intel(struct acpi_pct_register *not_used)
 252 {
 253         u32 val, dummy __always_unused;
 254
 255         rdmsr(MSR_IA32_PERF_CTL, val, dummy);
 256         return val;
 257 }
 258
 259 static void cpu_freq_write_intel(struct acpi_pct_register *not_used, u32 val)
 260 {
 261         u32 lo, hi;
 262
 263         rdmsr(MSR_IA32_PERF_CTL, lo, hi);
 264         lo = (lo & ~INTEL_MSR_RANGE) | (val & INTEL_MSR_RANGE);
 265         wrmsr(MSR_IA32_PERF_CTL, lo, hi);
 266 }
 267
 268 static u32 cpu_freq_read_amd(struct acpi_pct_register *not_used)
 269 {
 270         u32 val, dummy __always_unused;
 271
 272         rdmsr(MSR_AMD_PERF_CTL, val, dummy);
 273         return val;
 274 }
 275
 276 static void cpu_freq_write_amd(struct acpi_pct_register *not_used, u32 val)
 277 {
 278         wrmsr(MSR_AMD_PERF_CTL, val, 0);
 279 }
 280
 281 static u32 cpu_freq_read_io(struct acpi_pct_register *reg)
 282 {
 283         u32 val;
 284
 285         acpi_os_read_port(reg->address, &val, reg->bit_width);
 286         return val;
 287 }
 288
 289 static void cpu_freq_write_io(struct acpi_pct_register *reg, u32 val)
 290 {
 291         acpi_os_write_port(reg->address, val, reg->bit_width);
 292 }
 293
 294 struct drv_cmd {
 295         struct acpi_pct_register *reg;
 296         u32 val;
 297         union {
 298                 void (*write)(struct acpi_pct_register *reg, u32 val);
 299                 u32 (*read)(struct acpi_pct_register *reg);
 300         } func;
 301 };
 302
 303 /* Called via smp_call_function_single(), on the target CPU */
 304 static void do_drv_read(void *_cmd)
 305 {
 306         struct drv_cmd *cmd = _cmd;
 307
 308         cmd->val = cmd->func.read(cmd->reg);
 309 }
 310
 311 static u32 drv_read(struct acpi_cpufreq_data *data, const struct cpumask *mask)
 312 {
 313         struct acpi_processor_performance *perf = to_perf_data(data);
 314         struct drv_cmd cmd = {
 315                 .reg = &perf->control_register,
 316                 .func.read = data->cpu_freq_read,
 317         };
 318         int err;
 319
 320         err = smp_call_function_any(mask, do_drv_read, &cmd, 1);
 321         WARN_ON_ONCE(err);      /* smp_call_function_any() was buggy? */
 322         return cmd.val;
 323 }
 324
 325 /* Called via smp_call_function_many(), on the target CPUs */
 326 static void do_drv_write(void *_cmd)
 327 {
 328         struct drv_cmd *cmd = _cmd;
 329
 330         cmd->func.write(cmd->reg, cmd->val);
 331 }
 332
 333 static void drv_write(struct acpi_cpufreq_data *data,
 334                       const struct cpumask *mask, u32 val)
 335 {
 336         struct acpi_processor_performance *perf = to_perf_data(data);
 337         struct drv_cmd cmd = {
 338                 .reg = &perf->control_register,
 339                 .val = val,
 340                 .func.write = data->cpu_freq_write,
 341         };
 342         int this_cpu;
 343
 344         this_cpu = get_cpu();
 345         if (cpumask_test_cpu(this_cpu, mask))
 346                 do_drv_write(&cmd);
 347
 348         smp_call_function_many(mask, do_drv_write, &cmd, 1);
 349         put_cpu();
 350 }
 351
 352 static u32 get_cur_val(const struct cpumask *mask, struct acpi_cpufreq_data *data)
 353 {
 354         u32 val;
 355
 356         if (unlikely(cpumask_empty(mask)))
 357                 return 0;
 358
 359         val = drv_read(data, mask);
 360
 361         pr_debug("%s = %u\n", __func__, val);
 362
 363         return val;
 364 }
 365
 366 static unsigned int get_cur_freq_on_cpu(unsigned int cpu)
 367 {
 368         struct acpi_cpufreq_data *data;
 369         struct cpufreq_policy *policy;
 370         unsigned int freq;
 371         unsigned int cached_freq;
 372
 373         pr_debug("%s (%d)\n", __func__, cpu);
 374
 375         policy = cpufreq_cpu_get_raw(cpu);
 376         if (unlikely(!policy))
 377                 return 0;
 378
 379         data = policy->driver_data;
 380         if (unlikely(!data || !policy->freq_table))
 381                 return 0;
 382
 383         cached_freq = policy->freq_table[to_perf_data(data)->state].frequency;
 384         freq = extract_freq(policy, get_cur_val(cpumask_of(cpu), data));
 385         if (freq != cached_freq) {
 386                 /*
 387                  * The dreaded BIOS frequency change behind our back.
 388                  * Force set the frequency on next target call.
 389                  */
 390                 data->resume = 1;
 391         }
 392
 393         pr_debug("cur freq = %u\n", freq);
 394
 395         return freq;
 396 }
 397
 398 static unsigned int check_freqs(struct cpufreq_policy *policy,
 399                                 const struct cpumask *mask, unsigned int freq)
 400 {
 401         struct acpi_cpufreq_data *data = policy->driver_data;
 402         unsigned int cur_freq;
 403         unsigned int i;
 404
 405         for (i = 0; i < 100; i++) {
 406                 cur_freq = extract_freq(policy, get_cur_val(mask, data));
 407                 if (cur_freq == freq)
 408                         return 1;
 409                 udelay(10);
 410         }
 411         return 0;
 412 }
 413
 414 static int acpi_cpufreq_target(struct cpufreq_policy *policy,
 415                                unsigned int index)
 416 {
 417         struct acpi_cpufreq_data *data = policy->driver_data;
 418         struct acpi_processor_performance *perf;
 419         const struct cpumask *mask;
 420         unsigned int next_perf_state = 0; /* Index into perf table */
 421         int result = 0;
 422
 423         if (unlikely(!data)) {
 424                 return -ENODEV;
 425         }
 426
 427         perf = to_perf_data(data);
 428         next_perf_state = policy->freq_table[index].driver_data;
 429         if (perf->state == next_perf_state) {
 430                 if (unlikely(data->resume)) {
 431                         pr_debug("Called after resume, resetting to P%d\n",
 432                                 next_perf_state);
 433                         data->resume = 0;
 434                 } else {
 435                         pr_debug("Already at target state (P%d)\n",
 436                                 next_perf_state);
 437                         return 0;
 438                 }
 439         }
 440
 441         /*
 442          * The core won't allow CPUs to go away until the governor has been
 443          * stopped, so we can rely on the stability of policy->cpus.
 444          */
 445         mask = policy->shared_type == CPUFREQ_SHARED_TYPE_ANY ?
 446                 cpumask_of(policy->cpu) : policy->cpus;
 447
 448         drv_write(data, mask, perf->states[next_perf_state].control);
 449
 450         if (acpi_pstate_strict) {
 451                 if (!check_freqs(policy, mask,
 452                                  policy->freq_table[index].frequency)) {
 453                         pr_debug("%s (%d)\n", __func__, policy->cpu);
 454                         result = -EAGAIN;
 455                 }
 456         }
 457
 458         if (!result)
 459                 perf->state = next_perf_state;
 460
 461         return result;
 462 }
 463
 464 static unsigned int acpi_cpufreq_fast_switch(struct cpufreq_policy *policy,
 465                                              unsigned int target_freq)
 466 {
 467         struct acpi_cpufreq_data *data = policy->driver_data;
 468         struct acpi_processor_performance *perf;
 469         struct cpufreq_frequency_table *entry;
 470         unsigned int next_perf_state, next_freq, index;
 471
 472         /*
 473          * Find the closest frequency above target_freq.
 474          */
 475         if (policy->cached_target_freq == target_freq)
 476                 index = policy->cached_resolved_idx;
 477         else
 478                 index = cpufreq_table_find_index_dl(policy, target_freq,
 479                                                     false);
 480
 481         entry = &policy->freq_table[index];
 482         next_freq = entry->frequency;
 483         next_perf_state = entry->driver_data;
 484
 485         perf = to_perf_data(data);
 486         if (perf->state == next_perf_state) {
 487                 if (unlikely(data->resume))
 488                         data->resume = 0;
 489                 else
 490                         return next_freq;
 491         }
 492
 493         data->cpu_freq_write(&perf->control_register,
 494                              perf->states[next_perf_state].control);
 495         perf->state = next_perf_state;
 496         return next_freq;
 497 }
 498
 499 static unsigned long
 500 acpi_cpufreq_guess_freq(struct acpi_cpufreq_data *data, unsigned int cpu)
 501 {
 502         struct acpi_processor_performance *perf;
 503
 504         perf = to_perf_data(data);
 505         if (cpu_khz) {
 506                 /* search the closest match to cpu_khz */
 507                 unsigned int i;
 508                 unsigned long freq;
 509                 unsigned long freqn = perf->states[0].core_frequency * 1000;
 510
 511                 for (i = 0; i < (perf->state_count-1); i++) {
 512                         freq = freqn;
 513                         freqn = perf->states[i+1].core_frequency * 1000;
 514                         if ((2 * cpu_khz) > (freqn + freq)) {
 515                                 perf->state = i;
 516                                 return freq;
 517                         }
 518                 }
 519                 perf->state = perf->state_count-1;
 520                 return freqn;
 521         } else {
 522                 /* assume CPU is at P0... */
 523                 perf->state = 0;
 524                 return perf->states[0].core_frequency * 1000;
 525         }
 526 }
 527
 528 static void free_acpi_perf_data(void)
 529 {
 530         unsigned int i;
 531
 532         /* Freeing a NULL pointer is OK, and alloc_percpu zeroes. */
 533         for_each_possible_cpu(i)
 534                 free_cpumask_var(per_cpu_ptr(acpi_perf_data, i)
 535                                  ->shared_cpu_map);
 536         free_percpu(acpi_perf_data);
 537 }
 538
 539 static int cpufreq_boost_down_prep(unsigned int cpu)
 540 {
 541         /*
 542          * Clear the boost-disable bit on the CPU_DOWN path so that
 543          * this cpu cannot block the remaining ones from boosting.
 544          */
 545         return boost_set_msr(1);
 546 }
 547
 548 /*
 549  * acpi_cpufreq_early_init - initialize ACPI P-States library
 550  *
 551  * Initialize the ACPI P-States library (drivers/acpi/processor_perflib.c)
 552  * in order to determine correct frequency and voltage pairings. We can
 553  * do _PDC and _PSD and find out the processor dependency for the
 554  * actual init that will happen later...
 555  */
 556 static int __init acpi_cpufreq_early_init(void)
 557 {
 558         unsigned int i;
 559         pr_debug("%s\n", __func__);
 560
 561         acpi_perf_data = alloc_percpu(struct acpi_processor_performance);
 562         if (!acpi_perf_data) {
 563                 pr_debug("Memory allocation error for acpi_perf_data.\n");
 564                 return -ENOMEM;
 565         }
 566         for_each_possible_cpu(i) {
 567                 if (!zalloc_cpumask_var_node(
 568                         &per_cpu_ptr(acpi_perf_data, i)->shared_cpu_map,
 569                         GFP_KERNEL, cpu_to_node(i))) {
 570
 571                         /* Freeing a NULL pointer is OK: alloc_percpu zeroes. */
 572                         free_acpi_perf_data();
 573                         return -ENOMEM;
 574                 }
 575         }
 576
 577         /* Do initialization in ACPI core */
 578         acpi_processor_preregister_performance(acpi_perf_data);
 579         return 0;
 580 }
 581
 582 #ifdef CONFIG_SMP
 583 /*
 584  * Some BIOSes do SW_ANY coordination internally, either set it up in hw
 585  * or do it in BIOS firmware and won't inform about it to OS. If not
 586  * detected, this has a side effect of making CPU run at a different speed
 587  * than OS intended it to run at. Detect it and handle it cleanly.
 588  */
 589 static int bios_with_sw_any_bug;
 590
 591 static int sw_any_bug_found(const struct dmi_system_id *d)
 592 {
 593         bios_with_sw_any_bug = 1;
 594         return 0;
 595 }
 596
 597 static const struct dmi_system_id sw_any_bug_dmi_table[] = {
 598         {
 599                 .callback = sw_any_bug_found,
 600                 .ident = "Supermicro Server X6DLP",
 601                 .matches = {
 602                         DMI_MATCH(DMI_SYS_VENDOR, "Supermicro"),
 603                         DMI_MATCH(DMI_BIOS_VERSION, "080010"),
 604                         DMI_MATCH(DMI_PRODUCT_NAME, "X6DLP"),
 605                 },
 606         },
 607         { }
 608 };
 609
 610 static int acpi_cpufreq_blacklist(struct cpuinfo_x86 *c)
 611 {
 612         /* Intel Xeon Processor 7100 Series Specification Update
 613          * https://www.intel.com/Assets/PDF/specupdate/314554.pdf
 614          * AL30: A Machine Check Exception (MCE) Occurring during an
 615          * Enhanced Intel SpeedStep Technology Ratio Change May Cause
 616          * Both Processor Cores to Lock Up. */
 617         if (c->x86_vendor == X86_VENDOR_INTEL) {
 618                 if ((c->x86 == 15) &&
 619                     (c->x86_model == 6) &&
 620                     (c->x86_stepping == 8)) {
 621                         pr_info("Intel(R) Xeon(R) 7100 Errata AL30, processors may lock up on frequency changes: disabling acpi-cpufreq\n");
 622                         return -ENODEV;
 623                     }
 624                 }
 625         return 0;
 626 }
 627 #endif
 628
 629 #ifdef CONFIG_ACPI_CPPC_LIB
 630 static u64 get_max_boost_ratio(unsigned int cpu)
 631 {
 632         struct cppc_perf_caps perf_caps;
 633         u64 highest_perf, nominal_perf;
 634         int ret;
 635
 636         if (acpi_pstate_strict)
 637                 return 0;
 638
 639         ret = cppc_get_perf_caps(cpu, &perf_caps);
 640         if (ret) {
 641                 pr_debug("CPU%d: Unable to get performance capabilities (%d)\n",
 642                          cpu, ret);
 643                 return 0;
 644         }
 645
 646         if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD)
 647                 highest_perf = amd_get_highest_perf();
 648         else
 649                 highest_perf = perf_caps.highest_perf;
 650
 651         nominal_perf = perf_caps.nominal_perf;
 652
 653         if (!highest_perf || !nominal_perf) {
 654                 pr_debug("CPU%d: highest or nominal performance missing\n", cpu);
 655                 return 0;
 656         }
 657
 658         if (highest_perf < nominal_perf) {
 659                 pr_debug("CPU%d: nominal performance above highest\n", cpu);
 660                 return 0;
 661         }
 662
 663         return div_u64(highest_perf << SCHED_CAPACITY_SHIFT, nominal_perf);
 664 }
 665 #else
 666 static inline u64 get_max_boost_ratio(unsigned int cpu) { return 0; }
 667 #endif
 668
 669 static int acpi_cpufreq_cpu_init(struct cpufreq_policy *policy)
 670 {
 671         struct cpufreq_frequency_table *freq_table;
 672         struct acpi_processor_performance *perf;
 673         struct acpi_cpufreq_data *data;
 674         unsigned int cpu = policy->cpu;
 675         struct cpuinfo_x86 *c = &cpu_data(cpu);
 676         unsigned int valid_states = 0;
 677         unsigned int result = 0;
 678         u64 max_boost_ratio;
 679         unsigned int i;
 680 #ifdef CONFIG_SMP
 681         static int blacklisted;
 682 #endif
 683
 684         pr_debug("%s\n", __func__);
 685
 686 #ifdef CONFIG_SMP
 687         if (blacklisted)
 688                 return blacklisted;
 689         blacklisted = acpi_cpufreq_blacklist(c);
 690         if (blacklisted)
 691                 return blacklisted;
 692 #endif
 693
 694         data = kzalloc(sizeof(*data), GFP_KERNEL);
 695         if (!data)
 696                 return -ENOMEM;
 697
 698         if (!zalloc_cpumask_var(&data->freqdomain_cpus, GFP_KERNEL)) {
 699                 result = -ENOMEM;
 700                 goto err_free;
 701         }
 702
 703         perf = per_cpu_ptr(acpi_perf_data, cpu);
 704         data->acpi_perf_cpu = cpu;
 705         policy->driver_data = data;
 706
 707         if (cpu_has(c, X86_FEATURE_CONSTANT_TSC))
 708                 acpi_cpufreq_driver.flags |= CPUFREQ_CONST_LOOPS;
 709
 710         result = acpi_processor_register_performance(perf, cpu);
 711         if (result)
 712                 goto err_free_mask;
 713
 714         policy->shared_type = perf->shared_type;
 715
 716         /*
 717          * Will let policy->cpus know about dependency only when software
 718          * coordination is required.
 719          */
 720         if (policy->shared_type == CPUFREQ_SHARED_TYPE_ALL ||
 721             policy->shared_type == CPUFREQ_SHARED_TYPE_ANY) {
 722                 cpumask_copy(policy->cpus, perf->shared_cpu_map);
 723         }
 724         cpumask_copy(data->freqdomain_cpus, perf->shared_cpu_map);
 725
 726 #ifdef CONFIG_SMP
 727         dmi_check_system(sw_any_bug_dmi_table);
 728         if (bios_with_sw_any_bug && !policy_is_shared(policy)) {
 729                 policy->shared_type = CPUFREQ_SHARED_TYPE_ALL;
 730                 cpumask_copy(policy->cpus, topology_core_cpumask(cpu));
 731         }
 732
 733         if (check_amd_hwpstate_cpu(cpu) && boot_cpu_data.x86 < 0x19 &&
 734             !acpi_pstate_strict) {
 735                 cpumask_clear(policy->cpus);
 736                 cpumask_set_cpu(cpu, policy->cpus);
 737                 cpumask_copy(data->freqdomain_cpus,
 738                              topology_sibling_cpumask(cpu));
 739                 policy->shared_type = CPUFREQ_SHARED_TYPE_HW;
 740                 pr_info_once("overriding BIOS provided _PSD data\n");
 741         }
 742 #endif
 743
 744         /* capability check */
 745         if (perf->state_count <= 1) {
 746                 pr_debug("No P-States\n");
 747                 result = -ENODEV;
 748                 goto err_unreg;
 749         }
 750
 751         if (perf->control_register.space_id != perf->status_register.space_id) {
 752                 result = -ENODEV;
 753                 goto err_unreg;
 754         }
 755
 756         switch (perf->control_register.space_id) {
 757         case ACPI_ADR_SPACE_SYSTEM_IO:
 758                 if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD &&
 759                     boot_cpu_data.x86 == 0xf) {
 760                         pr_debug("AMD K8 systems must use native drivers.\n");
 761                         result = -ENODEV;
 762                         goto err_unreg;
 763                 }
 764                 pr_debug("SYSTEM IO addr space\n");
 765                 data->cpu_feature = SYSTEM_IO_CAPABLE;
 766                 data->cpu_freq_read = cpu_freq_read_io;
 767                 data->cpu_freq_write = cpu_freq_write_io;
 768                 break;
 769         case ACPI_ADR_SPACE_FIXED_HARDWARE:
 770                 pr_debug("HARDWARE addr space\n");
 771                 if (check_est_cpu(cpu)) {
 772                         data->cpu_feature = SYSTEM_INTEL_MSR_CAPABLE;
 773                         data->cpu_freq_read = cpu_freq_read_intel;
 774                         data->cpu_freq_write = cpu_freq_write_intel;
 775                         break;
 776                 }
 777                 if (check_amd_hwpstate_cpu(cpu)) {
 778                         data->cpu_feature = SYSTEM_AMD_MSR_CAPABLE;
 779                         data->cpu_freq_read = cpu_freq_read_amd;
 780                         data->cpu_freq_write = cpu_freq_write_amd;
 781                         break;
 782                 }
 783                 result = -ENODEV;
 784                 goto err_unreg;
 785         default:
 786                 pr_debug("Unknown addr space %d\n",
 787                         (u32) (perf->control_register.space_id));
 788                 result = -ENODEV;
 789                 goto err_unreg;
 790         }
 791
 792         freq_table = kcalloc(perf->state_count + 1, sizeof(*freq_table),
 793                              GFP_KERNEL);
 794         if (!freq_table) {
 795                 result = -ENOMEM;
 796                 goto err_unreg;
 797         }
 798
 799         /* detect transition latency */
 800         policy->cpuinfo.transition_latency = 0;
 801         for (i = 0; i < perf->state_count; i++) {
 802                 if ((perf->states[i].transition_latency * 1000) >
 803                     policy->cpuinfo.transition_latency)
 804                         policy->cpuinfo.transition_latency =
 805                             perf->states[i].transition_latency * 1000;
 806         }
 807
 808         /* Check for high latency (>20uS) from buggy BIOSes, like on T42 */
 809         if (perf->control_register.space_id == ACPI_ADR_SPACE_FIXED_HARDWARE &&
 810             policy->cpuinfo.transition_latency > 20 * 1000) {
 811                 policy->cpuinfo.transition_latency = 20 * 1000;
 812                 pr_info_once("P-state transition latency capped at 20 uS\n");
 813         }
 814
 815         /* table init */
 816         for (i = 0; i < perf->state_count; i++) {
 817                 if (i > 0 && perf->states[i].core_frequency >=
 818                     freq_table[valid_states-1].frequency / 1000)
 819                         continue;
 820
 821                 freq_table[valid_states].driver_data = i;
 822                 freq_table[valid_states].frequency =
 823                     perf->states[i].core_frequency * 1000;
 824                 valid_states++;
 825         }
 826         freq_table[valid_states].frequency = CPUFREQ_TABLE_END;
 827
 828         max_boost_ratio = get_max_boost_ratio(cpu);
 829         if (max_boost_ratio) {
 830                 unsigned int freq = freq_table[0].frequency;
 831
 832                 /*
 833                  * Because the loop above sorts the freq_table entries in the
 834                  * descending order, freq is the maximum frequency in the table.
 835                  * Assume that it corresponds to the CPPC nominal frequency and
 836                  * use it to set cpuinfo.max_freq.
 837                  */
 838                 policy->cpuinfo.max_freq = freq * max_boost_ratio >> SCHED_CAPACITY_SHIFT;
 839         } else {
 840                 /*
 841                  * If the maximum "boost" frequency is unknown, ask the arch
 842                  * scale-invariance code to use the "nominal" performance for
 843                  * CPU utilization scaling so as to prevent the schedutil
 844                  * governor from selecting inadequate CPU frequencies.
 845                  */
 846                 arch_set_max_freq_ratio(true);
 847         }
 848
 849         policy->freq_table = freq_table;
 850         perf->state = 0;
 851
 852         switch (perf->control_register.space_id) {
 853         case ACPI_ADR_SPACE_SYSTEM_IO:
 854                 /*
 855                  * The core will not set policy->cur, because
 856                  * cpufreq_driver->get is NULL, so we need to set it here.
 857                  * However, we have to guess it, because the current speed is
 858                  * unknown and not detectable via IO ports.
 859                  */
 860                 policy->cur = acpi_cpufreq_guess_freq(data, policy->cpu);
 861                 break;
 862         case ACPI_ADR_SPACE_FIXED_HARDWARE:
 863                 acpi_cpufreq_driver.get = get_cur_freq_on_cpu;
 864                 break;
 865         default:
 866                 break;
 867         }
 868
 869         /* notify BIOS that we exist */
 870         acpi_processor_notify_smm(THIS_MODULE);
 871
 872         pr_debug("CPU%u - ACPI performance management activated.\n", cpu);
 873         for (i = 0; i < perf->state_count; i++)
 874                 pr_debug("     %cP%d: %d MHz, %d mW, %d uS\n",
 875                         (i == perf->state ? '*' : ' '), i,
 876                         (u32) perf->states[i].core_frequency,
 877                         (u32) perf->states[i].power,
 878                         (u32) perf->states[i].transition_latency);
 879
 880         /*
 881          * the first call to ->target() should result in us actually
 882          * writing something to the appropriate registers.
 883          */
 884         data->resume = 1;
 885
 886         policy->fast_switch_possible = !acpi_pstate_strict &&
 887                 !(policy_is_shared(policy) && policy->shared_type != CPUFREQ_SHARED_TYPE_ANY);
 888
 889         if (perf->states[0].core_frequency * 1000 != freq_table[0].frequency)
 890                 pr_warn(FW_WARN "P-state 0 is not max freq\n");
 891
 892         if (acpi_cpufreq_driver.set_boost)
 893                 set_boost(policy, acpi_cpufreq_driver.boost_enabled);
 894
 895         return result;
 896
 897 err_unreg:
 898         acpi_processor_unregister_performance(cpu);
 899 err_free_mask:
 900         free_cpumask_var(data->freqdomain_cpus);
 901 err_free:
 902         kfree(data);
 903         policy->driver_data = NULL;
 904
 905         return result;
 906 }
 907
 908 static int acpi_cpufreq_cpu_exit(struct cpufreq_policy *policy)
 909 {
 910         struct acpi_cpufreq_data *data = policy->driver_data;
 911
 912         pr_debug("%s\n", __func__);
 913
 914         cpufreq_boost_down_prep(policy->cpu);
 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 void __init acpi_cpufreq_boost_init(void)
 958 {
 959         if (!(boot_cpu_has(X86_FEATURE_CPB) || boot_cpu_has(X86_FEATURE_IDA))) {
 960                 pr_debug("Boost capabilities not present in the processor\n");
 961                 return;
 962         }
 963
 964         acpi_cpufreq_driver.set_boost = set_boost;
 965         acpi_cpufreq_driver.boost_enabled = boost_state(0);
 966 }
 967
 968 static int __init acpi_cpufreq_init(void)
 969 {
 970         int ret;
 971
 972         if (acpi_disabled)
 973                 return -ENODEV;
 974
 975         /* don't keep reloading if cpufreq_driver exists */
 976         if (cpufreq_get_current_driver())
 977                 return -EEXIST;
 978
 979         pr_debug("%s\n", __func__);
 980
 981         ret = acpi_cpufreq_early_init();
 982         if (ret)
 983                 return ret;
 984
 985 #ifdef CONFIG_X86_ACPI_CPUFREQ_CPB
 986         /* this is a sysfs file with a strange name and an even stranger
 987          * semantic - per CPU instantiation, but system global effect.
 988          * Lets enable it only on AMD CPUs for compatibility reasons and
 989          * only if configured. This is considered legacy code, which
 990          * will probably be removed at some point in the future.
 991          */
 992         if (!check_amd_hwpstate_cpu(0)) {
 993                 struct freq_attr **attr;
 994
 995                 pr_debug("CPB unsupported, do not expose it\n");
 996
 997                 for (attr = acpi_cpufreq_attr; *attr; attr++)
 998                         if (*attr == &cpb) {
 999                                 *attr = NULL;
1000                                 break;
1001                         }
1002         }
1003 #endif
1004         acpi_cpufreq_boost_init();
1005
1006         ret = cpufreq_register_driver(&acpi_cpufreq_driver);
1007         if (ret) {
1008                 free_acpi_perf_data();
1009         }
1010         return ret;
1011 }
1012
1013 static void __exit acpi_cpufreq_exit(void)
1014 {
1015         pr_debug("%s\n", __func__);
1016
1017         cpufreq_unregister_driver(&acpi_cpufreq_driver);
1018
1019         free_acpi_perf_data();
1020 }
1021
1022 module_param(acpi_pstate_strict, uint, 0644);
1023 MODULE_PARM_DESC(acpi_pstate_strict,
1024         "value 0 or non-zero. non-zero -> strict ACPI checks are "
1025         "performed during frequency changes.");
1026
1027 late_initcall(acpi_cpufreq_init);
1028 module_exit(acpi_cpufreq_exit);
1029
1030 static const struct x86_cpu_id __maybe_unused acpi_cpufreq_ids[] = {
1031         X86_MATCH_FEATURE(X86_FEATURE_ACPI, NULL),
1032         X86_MATCH_FEATURE(X86_FEATURE_HW_PSTATE, NULL),
1033         {}
1034 };
1035 MODULE_DEVICE_TABLE(x86cpu, acpi_cpufreq_ids);
1036
1037 static const struct acpi_device_id __maybe_unused processor_device_ids[] = {
1038         {ACPI_PROCESSOR_OBJECT_HID, },
1039         {ACPI_PROCESSOR_DEVICE_HID, },
1040         {},
1041 };
1042 MODULE_DEVICE_TABLE(acpi, processor_device_ids);
1043
1044 MODULE_ALIAS("acpi");