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         case X86_VENDOR_CENTAUR:
  82         case X86_VENDOR_ZHAOXIN:
  83                 rdmsr_on_cpu(cpu, MSR_IA32_MISC_ENABLE, &lo, &hi);
  84                 msr = lo | ((u64)hi << 32);
  85                 return !(msr & MSR_IA32_MISC_ENABLE_TURBO_DISABLE);
  86         case X86_VENDOR_HYGON:
  87         case X86_VENDOR_AMD:
  88                 rdmsr_on_cpu(cpu, MSR_K7_HWCR, &lo, &hi);
  89                 msr = lo | ((u64)hi << 32);
  90                 return !(msr & MSR_K7_HWCR_CPB_DIS);
  91         }
  92         return false;
  93 }
  94
  95 static int boost_set_msr(bool enable)
  96 {
  97         u32 msr_addr;
  98         u64 msr_mask, val;
  99
 100         switch (boot_cpu_data.x86_vendor) {
 101         case X86_VENDOR_INTEL:
 102         case X86_VENDOR_CENTAUR:
 103         case X86_VENDOR_ZHAOXIN:
 104                 msr_addr = MSR_IA32_MISC_ENABLE;
 105                 msr_mask = MSR_IA32_MISC_ENABLE_TURBO_DISABLE;
 106                 break;
 107         case X86_VENDOR_HYGON:
 108         case X86_VENDOR_AMD:
 109                 msr_addr = MSR_K7_HWCR;
 110                 msr_mask = MSR_K7_HWCR_CPB_DIS;
 111                 break;
 112         default:
 113                 return -EINVAL;
 114         }
 115
 116         rdmsrl(msr_addr, val);
 117
 118         if (enable)
 119                 val &= ~msr_mask;
 120         else
 121                 val |= msr_mask;
 122
 123         wrmsrl(msr_addr, val);
 124         return 0;
 125 }
 126
 127 static void boost_set_msr_each(void *p_en)
 128 {
 129         bool enable = (bool) p_en;
 130
 131         boost_set_msr(enable);
 132 }
 133
 134 static int set_boost(struct cpufreq_policy *policy, int val)
 135 {
 136         on_each_cpu_mask(policy->cpus, boost_set_msr_each,
 137                          (void *)(long)val, 1);
 138         pr_debug("CPU %*pbl: Core Boosting %sabled.\n",
 139                  cpumask_pr_args(policy->cpus), val ? "en" : "dis");
 140
 141         return 0;
 142 }
 143
 144 static ssize_t show_freqdomain_cpus(struct cpufreq_policy *policy, char *buf)
 145 {
 146         struct acpi_cpufreq_data *data = policy->driver_data;
 147
 148         if (unlikely(!data))
 149                 return -ENODEV;
 150
 151         return cpufreq_show_cpus(data->freqdomain_cpus, buf);
 152 }
 153
 154 cpufreq_freq_attr_ro(freqdomain_cpus);
 155
 156 #ifdef CONFIG_X86_ACPI_CPUFREQ_CPB
 157 static ssize_t store_cpb(struct cpufreq_policy *policy, const char *buf,
 158                          size_t count)
 159 {
 160         int ret;
 161         unsigned int val = 0;
 162
 163         if (!acpi_cpufreq_driver.set_boost)
 164                 return -EINVAL;
 165
 166         ret = kstrtouint(buf, 10, &val);
 167         if (ret || val > 1)
 168                 return -EINVAL;
 169
 170         cpus_read_lock();
 171         set_boost(policy, val);
 172         cpus_read_unlock();
 173
 174         return count;
 175 }
 176
 177 static ssize_t show_cpb(struct cpufreq_policy *policy, char *buf)
 178 {
 179         return sprintf(buf, "%u\n", acpi_cpufreq_driver.boost_enabled);
 180 }
 181
 182 cpufreq_freq_attr_rw(cpb);
 183 #endif
 184
 185 static int check_est_cpu(unsigned int cpuid)
 186 {
 187         struct cpuinfo_x86 *cpu = &cpu_data(cpuid);
 188
 189         return cpu_has(cpu, X86_FEATURE_EST);
 190 }
 191
 192 static int check_amd_hwpstate_cpu(unsigned int cpuid)
 193 {
 194         struct cpuinfo_x86 *cpu = &cpu_data(cpuid);
 195
 196         return cpu_has(cpu, X86_FEATURE_HW_PSTATE);
 197 }
 198
 199 static unsigned extract_io(struct cpufreq_policy *policy, u32 value)
 200 {
 201         struct acpi_cpufreq_data *data = policy->driver_data;
 202         struct acpi_processor_performance *perf;
 203         int i;
 204
 205         perf = to_perf_data(data);
 206
 207         for (i = 0; i < perf->state_count; i++) {
 208                 if (value == perf->states[i].status)
 209                         return policy->freq_table[i].frequency;
 210         }
 211         return 0;
 212 }
 213
 214 static unsigned extract_msr(struct cpufreq_policy *policy, u32 msr)
 215 {
 216         struct acpi_cpufreq_data *data = policy->driver_data;
 217         struct cpufreq_frequency_table *pos;
 218         struct acpi_processor_performance *perf;
 219
 220         if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD)
 221                 msr &= AMD_MSR_RANGE;
 222         else if (boot_cpu_data.x86_vendor == X86_VENDOR_HYGON)
 223                 msr &= HYGON_MSR_RANGE;
 224         else
 225                 msr &= INTEL_MSR_RANGE;
 226
 227         perf = to_perf_data(data);
 228
 229         cpufreq_for_each_entry(pos, policy->freq_table)
 230                 if (msr == perf->states[pos->driver_data].status)
 231                         return pos->frequency;
 232         return policy->freq_table[0].frequency;
 233 }
 234
 235 static unsigned extract_freq(struct cpufreq_policy *policy, u32 val)
 236 {
 237         struct acpi_cpufreq_data *data = policy->driver_data;
 238
 239         switch (data->cpu_feature) {
 240         case SYSTEM_INTEL_MSR_CAPABLE:
 241         case SYSTEM_AMD_MSR_CAPABLE:
 242                 return extract_msr(policy, val);
 243         case SYSTEM_IO_CAPABLE:
 244                 return extract_io(policy, val);
 245         default:
 246                 return 0;
 247         }
 248 }
 249
 250 static u32 cpu_freq_read_intel(struct acpi_pct_register *not_used)
 251 {
 252         u32 val, dummy __always_unused;
 253
 254         rdmsr(MSR_IA32_PERF_CTL, val, dummy);
 255         return val;
 256 }
 257
 258 static void cpu_freq_write_intel(struct acpi_pct_register *not_used, u32 val)
 259 {
 260         u32 lo, hi;
 261
 262         rdmsr(MSR_IA32_PERF_CTL, lo, hi);
 263         lo = (lo & ~INTEL_MSR_RANGE) | (val & INTEL_MSR_RANGE);
 264         wrmsr(MSR_IA32_PERF_CTL, lo, hi);
 265 }
 266
 267 static u32 cpu_freq_read_amd(struct acpi_pct_register *not_used)
 268 {
 269         u32 val, dummy __always_unused;
 270
 271         rdmsr(MSR_AMD_PERF_CTL, val, dummy);
 272         return val;
 273 }
 274
 275 static void cpu_freq_write_amd(struct acpi_pct_register *not_used, u32 val)
 276 {
 277         wrmsr(MSR_AMD_PERF_CTL, val, 0);
 278 }
 279
 280 static u32 cpu_freq_read_io(struct acpi_pct_register *reg)
 281 {
 282         u32 val;
 283
 284         acpi_os_read_port(reg->address, &val, reg->bit_width);
 285         return val;
 286 }
 287
 288 static void cpu_freq_write_io(struct acpi_pct_register *reg, u32 val)
 289 {
 290         acpi_os_write_port(reg->address, val, reg->bit_width);
 291 }
 292
 293 struct drv_cmd {
 294         struct acpi_pct_register *reg;
 295         u32 val;
 296         union {
 297                 void (*write)(struct acpi_pct_register *reg, u32 val);
 298                 u32 (*read)(struct acpi_pct_register *reg);
 299         } func;
 300 };
 301
 302 /* Called via smp_call_function_single(), on the target CPU */
 303 static void do_drv_read(void *_cmd)
 304 {
 305         struct drv_cmd *cmd = _cmd;
 306
 307         cmd->val = cmd->func.read(cmd->reg);
 308 }
 309
 310 static u32 drv_read(struct acpi_cpufreq_data *data, const struct cpumask *mask)
 311 {
 312         struct acpi_processor_performance *perf = to_perf_data(data);
 313         struct drv_cmd cmd = {
 314                 .reg = &perf->control_register,
 315                 .func.read = data->cpu_freq_read,
 316         };
 317         int err;
 318
 319         err = smp_call_function_any(mask, do_drv_read, &cmd, 1);
 320         WARN_ON_ONCE(err);      /* smp_call_function_any() was buggy? */
 321         return cmd.val;
 322 }
 323
 324 /* Called via smp_call_function_many(), on the target CPUs */
 325 static void do_drv_write(void *_cmd)
 326 {
 327         struct drv_cmd *cmd = _cmd;
 328
 329         cmd->func.write(cmd->reg, cmd->val);
 330 }
 331
 332 static void drv_write(struct acpi_cpufreq_data *data,
 333                       const struct cpumask *mask, u32 val)
 334 {
 335         struct acpi_processor_performance *perf = to_perf_data(data);
 336         struct drv_cmd cmd = {
 337                 .reg = &perf->control_register,
 338                 .val = val,
 339                 .func.write = data->cpu_freq_write,
 340         };
 341         int this_cpu;
 342
 343         this_cpu = get_cpu();
 344         if (cpumask_test_cpu(this_cpu, mask))
 345                 do_drv_write(&cmd);
 346
 347         smp_call_function_many(mask, do_drv_write, &cmd, 1);
 348         put_cpu();
 349 }
 350
 351 static u32 get_cur_val(const struct cpumask *mask, struct acpi_cpufreq_data *data)
 352 {
 353         u32 val;
 354
 355         if (unlikely(cpumask_empty(mask)))
 356                 return 0;
 357
 358         val = drv_read(data, mask);
 359
 360         pr_debug("%s = %u\n", __func__, val);
 361
 362         return val;
 363 }
 364
 365 static unsigned int get_cur_freq_on_cpu(unsigned int cpu)
 366 {
 367         struct acpi_cpufreq_data *data;
 368         struct cpufreq_policy *policy;
 369         unsigned int freq;
 370         unsigned int cached_freq;
 371
 372         pr_debug("%s (%d)\n", __func__, cpu);
 373
 374         policy = cpufreq_cpu_get_raw(cpu);
 375         if (unlikely(!policy))
 376                 return 0;
 377
 378         data = policy->driver_data;
 379         if (unlikely(!data || !policy->freq_table))
 380                 return 0;
 381
 382         cached_freq = policy->freq_table[to_perf_data(data)->state].frequency;
 383         freq = extract_freq(policy, get_cur_val(cpumask_of(cpu), data));
 384         if (freq != cached_freq) {
 385                 /*
 386                  * The dreaded BIOS frequency change behind our back.
 387                  * Force set the frequency on next target call.
 388                  */
 389                 data->resume = 1;
 390         }
 391
 392         pr_debug("cur freq = %u\n", freq);
 393
 394         return freq;
 395 }
 396
 397 static unsigned int check_freqs(struct cpufreq_policy *policy,
 398                                 const struct cpumask *mask, unsigned int freq)
 399 {
 400         struct acpi_cpufreq_data *data = policy->driver_data;
 401         unsigned int cur_freq;
 402         unsigned int i;
 403
 404         for (i = 0; i < 100; i++) {
 405                 cur_freq = extract_freq(policy, get_cur_val(mask, data));
 406                 if (cur_freq == freq)
 407                         return 1;
 408                 udelay(10);
 409         }
 410         return 0;
 411 }
 412
 413 static int acpi_cpufreq_target(struct cpufreq_policy *policy,
 414                                unsigned int index)
 415 {
 416         struct acpi_cpufreq_data *data = policy->driver_data;
 417         struct acpi_processor_performance *perf;
 418         const struct cpumask *mask;
 419         unsigned int next_perf_state = 0; /* Index into perf table */
 420         int result = 0;
 421
 422         if (unlikely(!data)) {
 423                 return -ENODEV;
 424         }
 425
 426         perf = to_perf_data(data);
 427         next_perf_state = policy->freq_table[index].driver_data;
 428         if (perf->state == next_perf_state) {
 429                 if (unlikely(data->resume)) {
 430                         pr_debug("Called after resume, resetting to P%d\n",
 431                                 next_perf_state);
 432                         data->resume = 0;
 433                 } else {
 434                         pr_debug("Already at target state (P%d)\n",
 435                                 next_perf_state);
 436                         return 0;
 437                 }
 438         }
 439
 440         /*
 441          * The core won't allow CPUs to go away until the governor has been
 442          * stopped, so we can rely on the stability of policy->cpus.
 443          */
 444         mask = policy->shared_type == CPUFREQ_SHARED_TYPE_ANY ?
 445                 cpumask_of(policy->cpu) : policy->cpus;
 446
 447         drv_write(data, mask, perf->states[next_perf_state].control);
 448
 449         if (acpi_pstate_strict) {
 450                 if (!check_freqs(policy, mask,
 451                                  policy->freq_table[index].frequency)) {
 452                         pr_debug("%s (%d)\n", __func__, policy->cpu);
 453                         result = -EAGAIN;
 454                 }
 455         }
 456
 457         if (!result)
 458                 perf->state = next_perf_state;
 459
 460         return result;
 461 }
 462
 463 static unsigned int acpi_cpufreq_fast_switch(struct cpufreq_policy *policy,
 464                                              unsigned int target_freq)
 465 {
 466         struct acpi_cpufreq_data *data = policy->driver_data;
 467         struct acpi_processor_performance *perf;
 468         struct cpufreq_frequency_table *entry;
 469         unsigned int next_perf_state, next_freq, index;
 470
 471         /*
 472          * Find the closest frequency above target_freq.
 473          */
 474         if (policy->cached_target_freq == target_freq)
 475                 index = policy->cached_resolved_idx;
 476         else
 477                 index = cpufreq_table_find_index_dl(policy, target_freq,
 478                                                     false);
 479
 480         entry = &policy->freq_table[index];
 481         next_freq = entry->frequency;
 482         next_perf_state = entry->driver_data;
 483
 484         perf = to_perf_data(data);
 485         if (perf->state == next_perf_state) {
 486                 if (unlikely(data->resume))
 487                         data->resume = 0;
 488                 else
 489                         return next_freq;
 490         }
 491
 492         data->cpu_freq_write(&perf->control_register,
 493                              perf->states[next_perf_state].control);
 494         perf->state = next_perf_state;
 495         return next_freq;
 496 }
 497
 498 static unsigned long
 499 acpi_cpufreq_guess_freq(struct acpi_cpufreq_data *data, unsigned int cpu)
 500 {
 501         struct acpi_processor_performance *perf;
 502
 503         perf = to_perf_data(data);
 504         if (cpu_khz) {
 505                 /* search the closest match to cpu_khz */
 506                 unsigned int i;
 507                 unsigned long freq;
 508                 unsigned long freqn = perf->states[0].core_frequency * 1000;
 509
 510                 for (i = 0; i < (perf->state_count-1); i++) {
 511                         freq = freqn;
 512                         freqn = perf->states[i+1].core_frequency * 1000;
 513                         if ((2 * cpu_khz) > (freqn + freq)) {
 514                                 perf->state = i;
 515                                 return freq;
 516                         }
 517                 }
 518                 perf->state = perf->state_count-1;
 519                 return freqn;
 520         } else {
 521                 /* assume CPU is at P0... */
 522                 perf->state = 0;
 523                 return perf->states[0].core_frequency * 1000;
 524         }
 525 }
 526
 527 static void free_acpi_perf_data(void)
 528 {
 529         unsigned int i;
 530
 531         /* Freeing a NULL pointer is OK, and alloc_percpu zeroes. */
 532         for_each_possible_cpu(i)
 533                 free_cpumask_var(per_cpu_ptr(acpi_perf_data, i)
 534                                  ->shared_cpu_map);
 535         free_percpu(acpi_perf_data);
 536 }
 537
 538 static int cpufreq_boost_online(unsigned int cpu)
 539 {
 540         /*
 541          * On the CPU_UP path we simply keep the boost-disable flag
 542          * in sync with the current global state.
 543          */
 544         return boost_set_msr(acpi_cpufreq_driver.boost_enabled);
 545 }
 546
 547 static int cpufreq_boost_down_prep(unsigned int cpu)
 548 {
 549         /*
 550          * Clear the boost-disable bit on the CPU_DOWN path so that
 551          * this cpu cannot block the remaining ones from boosting.
 552          */
 553         return boost_set_msr(1);
 554 }
 555
 556 /*
 557  * acpi_cpufreq_early_init - initialize ACPI P-States library
 558  *
 559  * Initialize the ACPI P-States library (drivers/acpi/processor_perflib.c)
 560  * in order to determine correct frequency and voltage pairings. We can
 561  * do _PDC and _PSD and find out the processor dependency for the
 562  * actual init that will happen later...
 563  */
 564 static int __init acpi_cpufreq_early_init(void)
 565 {
 566         unsigned int i;
 567         pr_debug("%s\n", __func__);
 568
 569         acpi_perf_data = alloc_percpu(struct acpi_processor_performance);
 570         if (!acpi_perf_data) {
 571                 pr_debug("Memory allocation error for acpi_perf_data.\n");
 572                 return -ENOMEM;
 573         }
 574         for_each_possible_cpu(i) {
 575                 if (!zalloc_cpumask_var_node(
 576                         &per_cpu_ptr(acpi_perf_data, i)->shared_cpu_map,
 577                         GFP_KERNEL, cpu_to_node(i))) {
 578
 579                         /* Freeing a NULL pointer is OK: alloc_percpu zeroes. */
 580                         free_acpi_perf_data();
 581                         return -ENOMEM;
 582                 }
 583         }
 584
 585         /* Do initialization in ACPI core */
 586         acpi_processor_preregister_performance(acpi_perf_data);
 587         return 0;
 588 }
 589
 590 #ifdef CONFIG_SMP
 591 /*
 592  * Some BIOSes do SW_ANY coordination internally, either set it up in hw
 593  * or do it in BIOS firmware and won't inform about it to OS. If not
 594  * detected, this has a side effect of making CPU run at a different speed
 595  * than OS intended it to run at. Detect it and handle it cleanly.
 596  */
 597 static int bios_with_sw_any_bug;
 598
 599 static int sw_any_bug_found(const struct dmi_system_id *d)
 600 {
 601         bios_with_sw_any_bug = 1;
 602         return 0;
 603 }
 604
 605 static const struct dmi_system_id sw_any_bug_dmi_table[] = {
 606         {
 607                 .callback = sw_any_bug_found,
 608                 .ident = "Supermicro Server X6DLP",
 609                 .matches = {
 610                         DMI_MATCH(DMI_SYS_VENDOR, "Supermicro"),
 611                         DMI_MATCH(DMI_BIOS_VERSION, "080010"),
 612                         DMI_MATCH(DMI_PRODUCT_NAME, "X6DLP"),
 613                 },
 614         },
 615         { }
 616 };
 617
 618 static int acpi_cpufreq_blacklist(struct cpuinfo_x86 *c)
 619 {
 620         /* Intel Xeon Processor 7100 Series Specification Update
 621          * https://www.intel.com/Assets/PDF/specupdate/314554.pdf
 622          * AL30: A Machine Check Exception (MCE) Occurring during an
 623          * Enhanced Intel SpeedStep Technology Ratio Change May Cause
 624          * Both Processor Cores to Lock Up. */
 625         if (c->x86_vendor == X86_VENDOR_INTEL) {
 626                 if ((c->x86 == 15) &&
 627                     (c->x86_model == 6) &&
 628                     (c->x86_stepping == 8)) {
 629                         pr_info("Intel(R) Xeon(R) 7100 Errata AL30, processors may lock up on frequency changes: disabling acpi-cpufreq\n");
 630                         return -ENODEV;
 631                     }
 632                 }
 633         return 0;
 634 }
 635 #endif
 636
 637 #ifdef CONFIG_ACPI_CPPC_LIB
 638 static u64 get_max_boost_ratio(unsigned int cpu)
 639 {
 640         struct cppc_perf_caps perf_caps;
 641         u64 highest_perf, nominal_perf;
 642         int ret;
 643
 644         if (acpi_pstate_strict)
 645                 return 0;
 646
 647         ret = cppc_get_perf_caps(cpu, &perf_caps);
 648         if (ret) {
 649                 pr_debug("CPU%d: Unable to get performance capabilities (%d)\n",
 650                          cpu, ret);
 651                 return 0;
 652         }
 653
 654         if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD)
 655                 highest_perf = amd_get_highest_perf();
 656         else
 657                 highest_perf = perf_caps.highest_perf;
 658
 659         nominal_perf = perf_caps.nominal_perf;
 660
 661         if (!highest_perf || !nominal_perf) {
 662                 pr_debug("CPU%d: highest or nominal performance missing\n", cpu);
 663                 return 0;
 664         }
 665
 666         if (highest_perf < nominal_perf) {
 667                 pr_debug("CPU%d: nominal performance above highest\n", cpu);
 668                 return 0;
 669         }
 670
 671         return div_u64(highest_perf << SCHED_CAPACITY_SHIFT, nominal_perf);
 672 }
 673 #else
 674 static inline u64 get_max_boost_ratio(unsigned int cpu) { return 0; }
 675 #endif
 676
 677 static int acpi_cpufreq_cpu_init(struct cpufreq_policy *policy)
 678 {
 679         struct cpufreq_frequency_table *freq_table;
 680         struct acpi_processor_performance *perf;
 681         struct acpi_cpufreq_data *data;
 682         unsigned int cpu = policy->cpu;
 683         struct cpuinfo_x86 *c = &cpu_data(cpu);
 684         unsigned int valid_states = 0;
 685         unsigned int result = 0;
 686         u64 max_boost_ratio;
 687         unsigned int i;
 688 #ifdef CONFIG_SMP
 689         static int blacklisted;
 690 #endif
 691
 692         pr_debug("%s\n", __func__);
 693
 694 #ifdef CONFIG_SMP
 695         if (blacklisted)
 696                 return blacklisted;
 697         blacklisted = acpi_cpufreq_blacklist(c);
 698         if (blacklisted)
 699                 return blacklisted;
 700 #endif
 701
 702         data = kzalloc(sizeof(*data), GFP_KERNEL);
 703         if (!data)
 704                 return -ENOMEM;
 705
 706         if (!zalloc_cpumask_var(&data->freqdomain_cpus, GFP_KERNEL)) {
 707                 result = -ENOMEM;
 708                 goto err_free;
 709         }
 710
 711         perf = per_cpu_ptr(acpi_perf_data, cpu);
 712         data->acpi_perf_cpu = cpu;
 713         policy->driver_data = data;
 714
 715         if (cpu_has(c, X86_FEATURE_CONSTANT_TSC))
 716                 acpi_cpufreq_driver.flags |= CPUFREQ_CONST_LOOPS;
 717
 718         result = acpi_processor_register_performance(perf, cpu);
 719         if (result)
 720                 goto err_free_mask;
 721
 722         policy->shared_type = perf->shared_type;
 723
 724         /*
 725          * Will let policy->cpus know about dependency only when software
 726          * coordination is required.
 727          */
 728         if (policy->shared_type == CPUFREQ_SHARED_TYPE_ALL ||
 729             policy->shared_type == CPUFREQ_SHARED_TYPE_ANY) {
 730                 cpumask_copy(policy->cpus, perf->shared_cpu_map);
 731         }
 732         cpumask_copy(data->freqdomain_cpus, perf->shared_cpu_map);
 733
 734 #ifdef CONFIG_SMP
 735         dmi_check_system(sw_any_bug_dmi_table);
 736         if (bios_with_sw_any_bug && !policy_is_shared(policy)) {
 737                 policy->shared_type = CPUFREQ_SHARED_TYPE_ALL;
 738                 cpumask_copy(policy->cpus, topology_core_cpumask(cpu));
 739         }
 740
 741         if (check_amd_hwpstate_cpu(cpu) && boot_cpu_data.x86 < 0x19 &&
 742             !acpi_pstate_strict) {
 743                 cpumask_clear(policy->cpus);
 744                 cpumask_set_cpu(cpu, policy->cpus);
 745                 cpumask_copy(data->freqdomain_cpus,
 746                              topology_sibling_cpumask(cpu));
 747                 policy->shared_type = CPUFREQ_SHARED_TYPE_HW;
 748                 pr_info_once("overriding BIOS provided _PSD data\n");
 749         }
 750 #endif
 751
 752         /* capability check */
 753         if (perf->state_count <= 1) {
 754                 pr_debug("No P-States\n");
 755                 result = -ENODEV;
 756                 goto err_unreg;
 757         }
 758
 759         if (perf->control_register.space_id != perf->status_register.space_id) {
 760                 result = -ENODEV;
 761                 goto err_unreg;
 762         }
 763
 764         switch (perf->control_register.space_id) {
 765         case ACPI_ADR_SPACE_SYSTEM_IO:
 766                 if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD &&
 767                     boot_cpu_data.x86 == 0xf) {
 768                         pr_debug("AMD K8 systems must use native drivers.\n");
 769                         result = -ENODEV;
 770                         goto err_unreg;
 771                 }
 772                 pr_debug("SYSTEM IO addr space\n");
 773                 data->cpu_feature = SYSTEM_IO_CAPABLE;
 774                 data->cpu_freq_read = cpu_freq_read_io;
 775                 data->cpu_freq_write = cpu_freq_write_io;
 776                 break;
 777         case ACPI_ADR_SPACE_FIXED_HARDWARE:
 778                 pr_debug("HARDWARE addr space\n");
 779                 if (check_est_cpu(cpu)) {
 780                         data->cpu_feature = SYSTEM_INTEL_MSR_CAPABLE;
 781                         data->cpu_freq_read = cpu_freq_read_intel;
 782                         data->cpu_freq_write = cpu_freq_write_intel;
 783                         break;
 784                 }
 785                 if (check_amd_hwpstate_cpu(cpu)) {
 786                         data->cpu_feature = SYSTEM_AMD_MSR_CAPABLE;
 787                         data->cpu_freq_read = cpu_freq_read_amd;
 788                         data->cpu_freq_write = cpu_freq_write_amd;
 789                         break;
 790                 }
 791                 result = -ENODEV;
 792                 goto err_unreg;
 793         default:
 794                 pr_debug("Unknown addr space %d\n",
 795                         (u32) (perf->control_register.space_id));
 796                 result = -ENODEV;
 797                 goto err_unreg;
 798         }
 799
 800         freq_table = kcalloc(perf->state_count + 1, sizeof(*freq_table),
 801                              GFP_KERNEL);
 802         if (!freq_table) {
 803                 result = -ENOMEM;
 804                 goto err_unreg;
 805         }
 806
 807         /* detect transition latency */
 808         policy->cpuinfo.transition_latency = 0;
 809         for (i = 0; i < perf->state_count; i++) {
 810                 if ((perf->states[i].transition_latency * 1000) >
 811                     policy->cpuinfo.transition_latency)
 812                         policy->cpuinfo.transition_latency =
 813                             perf->states[i].transition_latency * 1000;
 814         }
 815
 816         /* Check for high latency (>20uS) from buggy BIOSes, like on T42 */
 817         if (perf->control_register.space_id == ACPI_ADR_SPACE_FIXED_HARDWARE &&
 818             policy->cpuinfo.transition_latency > 20 * 1000) {
 819                 policy->cpuinfo.transition_latency = 20 * 1000;
 820                 pr_info_once("P-state transition latency capped at 20 uS\n");
 821         }
 822
 823         /* table init */
 824         for (i = 0; i < perf->state_count; i++) {
 825                 if (i > 0 && perf->states[i].core_frequency >=
 826                     freq_table[valid_states-1].frequency / 1000)
 827                         continue;
 828
 829                 freq_table[valid_states].driver_data = i;
 830                 freq_table[valid_states].frequency =
 831                     perf->states[i].core_frequency * 1000;
 832                 valid_states++;
 833         }
 834         freq_table[valid_states].frequency = CPUFREQ_TABLE_END;
 835
 836         max_boost_ratio = get_max_boost_ratio(cpu);
 837         if (max_boost_ratio) {
 838                 unsigned int freq = freq_table[0].frequency;
 839
 840                 /*
 841                  * Because the loop above sorts the freq_table entries in the
 842                  * descending order, freq is the maximum frequency in the table.
 843                  * Assume that it corresponds to the CPPC nominal frequency and
 844                  * use it to set cpuinfo.max_freq.
 845                  */
 846                 policy->cpuinfo.max_freq = freq * max_boost_ratio >> SCHED_CAPACITY_SHIFT;
 847         } else {
 848                 /*
 849                  * If the maximum "boost" frequency is unknown, ask the arch
 850                  * scale-invariance code to use the "nominal" performance for
 851                  * CPU utilization scaling so as to prevent the schedutil
 852                  * governor from selecting inadequate CPU frequencies.
 853                  */
 854                 arch_set_max_freq_ratio(true);
 855         }
 856
 857         policy->freq_table = freq_table;
 858         perf->state = 0;
 859
 860         switch (perf->control_register.space_id) {
 861         case ACPI_ADR_SPACE_SYSTEM_IO:
 862                 /*
 863                  * The core will not set policy->cur, because
 864                  * cpufreq_driver->get is NULL, so we need to set it here.
 865                  * However, we have to guess it, because the current speed is
 866                  * unknown and not detectable via IO ports.
 867                  */
 868                 policy->cur = acpi_cpufreq_guess_freq(data, policy->cpu);
 869                 break;
 870         case ACPI_ADR_SPACE_FIXED_HARDWARE:
 871                 acpi_cpufreq_driver.get = get_cur_freq_on_cpu;
 872                 break;
 873         default:
 874                 break;
 875         }
 876
 877         /* notify BIOS that we exist */
 878         acpi_processor_notify_smm(THIS_MODULE);
 879
 880         pr_debug("CPU%u - ACPI performance management activated.\n", cpu);
 881         for (i = 0; i < perf->state_count; i++)
 882                 pr_debug("     %cP%d: %d MHz, %d mW, %d uS\n",
 883                         (i == perf->state ? '*' : ' '), i,
 884                         (u32) perf->states[i].core_frequency,
 885                         (u32) perf->states[i].power,
 886                         (u32) perf->states[i].transition_latency);
 887
 888         /*
 889          * the first call to ->target() should result in us actually
 890          * writing something to the appropriate registers.
 891          */
 892         data->resume = 1;
 893
 894         policy->fast_switch_possible = !acpi_pstate_strict &&
 895                 !(policy_is_shared(policy) && policy->shared_type != CPUFREQ_SHARED_TYPE_ANY);
 896
 897         if (perf->states[0].core_frequency * 1000 != freq_table[0].frequency)
 898                 pr_warn(FW_WARN "P-state 0 is not max freq\n");
 899
 900         return result;
 901
 902 err_unreg:
 903         acpi_processor_unregister_performance(cpu);
 904 err_free_mask:
 905         free_cpumask_var(data->freqdomain_cpus);
 906 err_free:
 907         kfree(data);
 908         policy->driver_data = NULL;
 909
 910         return result;
 911 }
 912
 913 static int acpi_cpufreq_cpu_exit(struct cpufreq_policy *policy)
 914 {
 915         struct acpi_cpufreq_data *data = policy->driver_data;
 916
 917         pr_debug("%s\n", __func__);
 918
 919         policy->fast_switch_possible = false;
 920         policy->driver_data = NULL;
 921         acpi_processor_unregister_performance(data->acpi_perf_cpu);
 922         free_cpumask_var(data->freqdomain_cpus);
 923         kfree(policy->freq_table);
 924         kfree(data);
 925
 926         return 0;
 927 }
 928
 929 static int acpi_cpufreq_resume(struct cpufreq_policy *policy)
 930 {
 931         struct acpi_cpufreq_data *data = policy->driver_data;
 932
 933         pr_debug("%s\n", __func__);
 934
 935         data->resume = 1;
 936
 937         return 0;
 938 }
 939
 940 static struct freq_attr *acpi_cpufreq_attr[] = {
 941         &cpufreq_freq_attr_scaling_available_freqs,
 942         &freqdomain_cpus,
 943 #ifdef CONFIG_X86_ACPI_CPUFREQ_CPB
 944         &cpb,
 945 #endif
 946         NULL,
 947 };
 948
 949 static struct cpufreq_driver acpi_cpufreq_driver = {
 950         .verify         = cpufreq_generic_frequency_table_verify,
 951         .target_index   = acpi_cpufreq_target,
 952         .fast_switch    = acpi_cpufreq_fast_switch,
 953         .bios_limit     = acpi_processor_get_bios_limit,
 954         .init           = acpi_cpufreq_cpu_init,
 955         .exit           = acpi_cpufreq_cpu_exit,
 956         .resume         = acpi_cpufreq_resume,
 957         .name           = "acpi-cpufreq",
 958         .attr           = acpi_cpufreq_attr,
 959 };
 960
 961 static enum cpuhp_state acpi_cpufreq_online;
 962
 963 static void __init acpi_cpufreq_boost_init(void)
 964 {
 965         int ret;
 966
 967         if (!(boot_cpu_has(X86_FEATURE_CPB) || boot_cpu_has(X86_FEATURE_IDA))) {
 968                 pr_debug("Boost capabilities not present in the processor\n");
 969                 return;
 970         }
 971
 972         acpi_cpufreq_driver.set_boost = set_boost;
 973         acpi_cpufreq_driver.boost_enabled = boost_state(0);
 974
 975         /*
 976          * This calls the online callback on all online cpu and forces all
 977          * MSRs to the same value.
 978          */
 979         ret = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "cpufreq/acpi:online",
 980                                 cpufreq_boost_online, cpufreq_boost_down_prep);
 981         if (ret < 0) {
 982                 pr_err("acpi_cpufreq: failed to register hotplug callbacks\n");
 983                 return;
 984         }
 985         acpi_cpufreq_online = ret;
 986 }
 987
 988 static void acpi_cpufreq_boost_exit(void)
 989 {
 990         if (acpi_cpufreq_online > 0)
 991                 cpuhp_remove_state_nocalls(acpi_cpufreq_online);
 992 }
 993
 994 static int __init acpi_cpufreq_init(void)
 995 {
 996         int ret;
 997
 998         if (acpi_disabled)
 999                 return -ENODEV;
1000
1001         /* don't keep reloading if cpufreq_driver exists */
1002         if (cpufreq_get_current_driver())
1003                 return -EEXIST;
1004
1005         pr_debug("%s\n", __func__);
1006
1007         ret = acpi_cpufreq_early_init();
1008         if (ret)
1009                 return ret;
1010
1011 #ifdef CONFIG_X86_ACPI_CPUFREQ_CPB
1012         /* this is a sysfs file with a strange name and an even stranger
1013          * semantic - per CPU instantiation, but system global effect.
1014          * Lets enable it only on AMD CPUs for compatibility reasons and
1015          * only if configured. This is considered legacy code, which
1016          * will probably be removed at some point in the future.
1017          */
1018         if (!check_amd_hwpstate_cpu(0)) {
1019                 struct freq_attr **attr;
1020
1021                 pr_debug("CPB unsupported, do not expose it\n");
1022
1023                 for (attr = acpi_cpufreq_attr; *attr; attr++)
1024                         if (*attr == &cpb) {
1025                                 *attr = NULL;
1026                                 break;
1027                         }
1028         }
1029 #endif
1030         acpi_cpufreq_boost_init();
1031
1032         ret = cpufreq_register_driver(&acpi_cpufreq_driver);
1033         if (ret) {
1034                 free_acpi_perf_data();
1035                 acpi_cpufreq_boost_exit();
1036         }
1037         return ret;
1038 }
1039
1040 static void __exit acpi_cpufreq_exit(void)
1041 {
1042         pr_debug("%s\n", __func__);
1043
1044         acpi_cpufreq_boost_exit();
1045
1046         cpufreq_unregister_driver(&acpi_cpufreq_driver);
1047
1048         free_acpi_perf_data();
1049 }
1050
1051 module_param(acpi_pstate_strict, uint, 0644);
1052 MODULE_PARM_DESC(acpi_pstate_strict,
1053         "value 0 or non-zero. non-zero -> strict ACPI checks are "
1054         "performed during frequency changes.");
1055
1056 late_initcall(acpi_cpufreq_init);
1057 module_exit(acpi_cpufreq_exit);
1058
1059 static const struct x86_cpu_id __maybe_unused acpi_cpufreq_ids[] = {
1060         X86_MATCH_FEATURE(X86_FEATURE_ACPI, NULL),
1061         X86_MATCH_FEATURE(X86_FEATURE_HW_PSTATE, NULL),
1062         {}
1063 };
1064 MODULE_DEVICE_TABLE(x86cpu, acpi_cpufreq_ids);
1065
1066 static const struct acpi_device_id __maybe_unused processor_device_ids[] = {
1067         {ACPI_PROCESSOR_OBJECT_HID, },
1068         {ACPI_PROCESSOR_DEVICE_HID, },
1069         {},
1070 };
1071 MODULE_DEVICE_TABLE(acpi, processor_device_ids);
1072
1073 MODULE_ALIAS("acpi");