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 #include <linux/string_helpers.h>
  23 #include <linux/platform_device.h>
  24
  25 #include <linux/acpi.h>
  26 #include <linux/io.h>
  27 #include <linux/delay.h>
  28 #include <linux/uaccess.h>
  29
  30 #include <acpi/processor.h>
  31 #include <acpi/cppc_acpi.h>
  32
  33 #include <asm/msr.h>
  34 #include <asm/processor.h>
  35 #include <asm/cpufeature.h>
  36 #include <asm/cpu_device_id.h>
  37
  38 MODULE_AUTHOR("Paul Diefenbaugh, Dominik Brodowski");
  39 MODULE_DESCRIPTION("ACPI Processor P-States Driver");
  40 MODULE_LICENSE("GPL");
  41
  42 enum {
  43         UNDEFINED_CAPABLE = 0,
  44         SYSTEM_INTEL_MSR_CAPABLE,
  45         SYSTEM_AMD_MSR_CAPABLE,
  46         SYSTEM_IO_CAPABLE,
  47 };
  48
  49 #define INTEL_MSR_RANGE         (0xffff)
  50 #define AMD_MSR_RANGE           (0x7)
  51 #define HYGON_MSR_RANGE         (0x7)
  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         u64 msr;
  77
  78         switch (boot_cpu_data.x86_vendor) {
  79         case X86_VENDOR_INTEL:
  80         case X86_VENDOR_CENTAUR:
  81         case X86_VENDOR_ZHAOXIN:
  82                 rdmsrl_on_cpu(cpu, MSR_IA32_MISC_ENABLE, &msr);
  83                 return !(msr & MSR_IA32_MISC_ENABLE_TURBO_DISABLE);
  84         case X86_VENDOR_HYGON:
  85         case X86_VENDOR_AMD:
  86                 rdmsrl_on_cpu(cpu, MSR_K7_HWCR, &msr);
  87                 return !(msr & MSR_K7_HWCR_CPB_DIS);
  88         }
  89         return false;
  90 }
  91
  92 static int boost_set_msr(bool enable)
  93 {
  94         u32 msr_addr;
  95         u64 msr_mask, val;
  96
  97         switch (boot_cpu_data.x86_vendor) {
  98         case X86_VENDOR_INTEL:
  99         case X86_VENDOR_CENTAUR:
 100         case X86_VENDOR_ZHAOXIN:
 101                 msr_addr = MSR_IA32_MISC_ENABLE;
 102                 msr_mask = MSR_IA32_MISC_ENABLE_TURBO_DISABLE;
 103                 break;
 104         case X86_VENDOR_HYGON:
 105         case X86_VENDOR_AMD:
 106                 msr_addr = MSR_K7_HWCR;
 107                 msr_mask = MSR_K7_HWCR_CPB_DIS;
 108                 break;
 109         default:
 110                 return -EINVAL;
 111         }
 112
 113         rdmsrl(msr_addr, val);
 114
 115         if (enable)
 116                 val &= ~msr_mask;
 117         else
 118                 val |= msr_mask;
 119
 120         wrmsrl(msr_addr, val);
 121         return 0;
 122 }
 123
 124 static void boost_set_msr_each(void *p_en)
 125 {
 126         bool enable = (bool) p_en;
 127
 128         boost_set_msr(enable);
 129 }
 130
 131 static int set_boost(struct cpufreq_policy *policy, int val)
 132 {
 133         on_each_cpu_mask(policy->cpus, boost_set_msr_each,
 134                          (void *)(long)val, 1);
 135         pr_debug("CPU %*pbl: Core Boosting %s.\n",
 136                  cpumask_pr_args(policy->cpus), str_enabled_disabled(val));
 137
 138         return 0;
 139 }
 140
 141 static ssize_t show_freqdomain_cpus(struct cpufreq_policy *policy, char *buf)
 142 {
 143         struct acpi_cpufreq_data *data = policy->driver_data;
 144
 145         if (unlikely(!data))
 146                 return -ENODEV;
 147
 148         return cpufreq_show_cpus(data->freqdomain_cpus, buf);
 149 }
 150
 151 cpufreq_freq_attr_ro(freqdomain_cpus);
 152
 153 #ifdef CONFIG_X86_ACPI_CPUFREQ_CPB
 154 static ssize_t store_cpb(struct cpufreq_policy *policy, const char *buf,
 155                          size_t count)
 156 {
 157         int ret;
 158         unsigned int val = 0;
 159
 160         if (!acpi_cpufreq_driver.set_boost)
 161                 return -EINVAL;
 162
 163         ret = kstrtouint(buf, 10, &val);
 164         if (ret || val > 1)
 165                 return -EINVAL;
 166
 167         cpus_read_lock();
 168         set_boost(policy, val);
 169         cpus_read_unlock();
 170
 171         return count;
 172 }
 173
 174 static ssize_t show_cpb(struct cpufreq_policy *policy, char *buf)
 175 {
 176         return sprintf(buf, "%u\n", acpi_cpufreq_driver.boost_enabled);
 177 }
 178
 179 cpufreq_freq_attr_rw(cpb);
 180 #endif
 181
 182 static int check_est_cpu(unsigned int cpuid)
 183 {
 184         struct cpuinfo_x86 *cpu = &cpu_data(cpuid);
 185
 186         return cpu_has(cpu, X86_FEATURE_EST);
 187 }
 188
 189 static int check_amd_hwpstate_cpu(unsigned int cpuid)
 190 {
 191         struct cpuinfo_x86 *cpu = &cpu_data(cpuid);
 192
 193         return cpu_has(cpu, X86_FEATURE_HW_PSTATE);
 194 }
 195
 196 static unsigned extract_io(struct cpufreq_policy *policy, u32 value)
 197 {
 198         struct acpi_cpufreq_data *data = policy->driver_data;
 199         struct acpi_processor_performance *perf;
 200         int i;
 201
 202         perf = to_perf_data(data);
 203
 204         for (i = 0; i < perf->state_count; i++) {
 205                 if (value == perf->states[i].status)
 206                         return policy->freq_table[i].frequency;
 207         }
 208         return 0;
 209 }
 210
 211 static unsigned extract_msr(struct cpufreq_policy *policy, u32 msr)
 212 {
 213         struct acpi_cpufreq_data *data = policy->driver_data;
 214         struct cpufreq_frequency_table *pos;
 215         struct acpi_processor_performance *perf;
 216
 217         if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD)
 218                 msr &= AMD_MSR_RANGE;
 219         else if (boot_cpu_data.x86_vendor == X86_VENDOR_HYGON)
 220                 msr &= HYGON_MSR_RANGE;
 221         else
 222                 msr &= INTEL_MSR_RANGE;
 223
 224         perf = to_perf_data(data);
 225
 226         cpufreq_for_each_entry(pos, policy->freq_table)
 227                 if (msr == perf->states[pos->driver_data].status)
 228                         return pos->frequency;
 229         return policy->freq_table[0].frequency;
 230 }
 231
 232 static unsigned extract_freq(struct cpufreq_policy *policy, u32 val)
 233 {
 234         struct acpi_cpufreq_data *data = policy->driver_data;
 235
 236         switch (data->cpu_feature) {
 237         case SYSTEM_INTEL_MSR_CAPABLE:
 238         case SYSTEM_AMD_MSR_CAPABLE:
 239                 return extract_msr(policy, val);
 240         case SYSTEM_IO_CAPABLE:
 241                 return extract_io(policy, val);
 242         default:
 243                 return 0;
 244         }
 245 }
 246
 247 static u32 cpu_freq_read_intel(struct acpi_pct_register *not_used)
 248 {
 249         u32 val, dummy __always_unused;
 250
 251         rdmsr(MSR_IA32_PERF_CTL, val, dummy);
 252         return val;
 253 }
 254
 255 static void cpu_freq_write_intel(struct acpi_pct_register *not_used, u32 val)
 256 {
 257         u32 lo, hi;
 258
 259         rdmsr(MSR_IA32_PERF_CTL, lo, hi);
 260         lo = (lo & ~INTEL_MSR_RANGE) | (val & INTEL_MSR_RANGE);
 261         wrmsr(MSR_IA32_PERF_CTL, lo, hi);
 262 }
 263
 264 static u32 cpu_freq_read_amd(struct acpi_pct_register *not_used)
 265 {
 266         u32 val, dummy __always_unused;
 267
 268         rdmsr(MSR_AMD_PERF_CTL, val, dummy);
 269         return val;
 270 }
 271
 272 static void cpu_freq_write_amd(struct acpi_pct_register *not_used, u32 val)
 273 {
 274         wrmsr(MSR_AMD_PERF_CTL, val, 0);
 275 }
 276
 277 static u32 cpu_freq_read_io(struct acpi_pct_register *reg)
 278 {
 279         u32 val;
 280
 281         acpi_os_read_port(reg->address, &val, reg->bit_width);
 282         return val;
 283 }
 284
 285 static void cpu_freq_write_io(struct acpi_pct_register *reg, u32 val)
 286 {
 287         acpi_os_write_port(reg->address, val, reg->bit_width);
 288 }
 289
 290 struct drv_cmd {
 291         struct acpi_pct_register *reg;
 292         u32 val;
 293         union {
 294                 void (*write)(struct acpi_pct_register *reg, u32 val);
 295                 u32 (*read)(struct acpi_pct_register *reg);
 296         } func;
 297 };
 298
 299 /* Called via smp_call_function_single(), on the target CPU */
 300 static void do_drv_read(void *_cmd)
 301 {
 302         struct drv_cmd *cmd = _cmd;
 303
 304         cmd->val = cmd->func.read(cmd->reg);
 305 }
 306
 307 static u32 drv_read(struct acpi_cpufreq_data *data, const struct cpumask *mask)
 308 {
 309         struct acpi_processor_performance *perf = to_perf_data(data);
 310         struct drv_cmd cmd = {
 311                 .reg = &perf->control_register,
 312                 .func.read = data->cpu_freq_read,
 313         };
 314         int err;
 315
 316         err = smp_call_function_any(mask, do_drv_read, &cmd, 1);
 317         WARN_ON_ONCE(err);      /* smp_call_function_any() was buggy? */
 318         return cmd.val;
 319 }
 320
 321 /* Called via smp_call_function_many(), on the target CPUs */
 322 static void do_drv_write(void *_cmd)
 323 {
 324         struct drv_cmd *cmd = _cmd;
 325
 326         cmd->func.write(cmd->reg, cmd->val);
 327 }
 328
 329 static void drv_write(struct acpi_cpufreq_data *data,
 330                       const struct cpumask *mask, u32 val)
 331 {
 332         struct acpi_processor_performance *perf = to_perf_data(data);
 333         struct drv_cmd cmd = {
 334                 .reg = &perf->control_register,
 335                 .val = val,
 336                 .func.write = data->cpu_freq_write,
 337         };
 338         int this_cpu;
 339
 340         this_cpu = get_cpu();
 341         if (cpumask_test_cpu(this_cpu, mask))
 342                 do_drv_write(&cmd);
 343
 344         smp_call_function_many(mask, do_drv_write, &cmd, 1);
 345         put_cpu();
 346 }
 347
 348 static u32 get_cur_val(const struct cpumask *mask, struct acpi_cpufreq_data *data)
 349 {
 350         u32 val;
 351
 352         if (unlikely(cpumask_empty(mask)))
 353                 return 0;
 354
 355         val = drv_read(data, mask);
 356
 357         pr_debug("%s = %u\n", __func__, val);
 358
 359         return val;
 360 }
 361
 362 static unsigned int get_cur_freq_on_cpu(unsigned int cpu)
 363 {
 364         struct acpi_cpufreq_data *data;
 365         struct cpufreq_policy *policy;
 366         unsigned int freq;
 367         unsigned int cached_freq;
 368
 369         pr_debug("%s (%d)\n", __func__, cpu);
 370
 371         policy = cpufreq_cpu_get_raw(cpu);
 372         if (unlikely(!policy))
 373                 return 0;
 374
 375         data = policy->driver_data;
 376         if (unlikely(!data || !policy->freq_table))
 377                 return 0;
 378
 379         cached_freq = policy->freq_table[to_perf_data(data)->state].frequency;
 380         freq = extract_freq(policy, get_cur_val(cpumask_of(cpu), data));
 381         if (freq != cached_freq) {
 382                 /*
 383                  * The dreaded BIOS frequency change behind our back.
 384                  * Force set the frequency on next target call.
 385                  */
 386                 data->resume = 1;
 387         }
 388
 389         pr_debug("cur freq = %u\n", freq);
 390
 391         return freq;
 392 }
 393
 394 static unsigned int check_freqs(struct cpufreq_policy *policy,
 395                                 const struct cpumask *mask, unsigned int freq)
 396 {
 397         struct acpi_cpufreq_data *data = policy->driver_data;
 398         unsigned int cur_freq;
 399         unsigned int i;
 400
 401         for (i = 0; i < 100; i++) {
 402                 cur_freq = extract_freq(policy, get_cur_val(mask, data));
 403                 if (cur_freq == freq)
 404                         return 1;
 405                 udelay(10);
 406         }
 407         return 0;
 408 }
 409
 410 static int acpi_cpufreq_target(struct cpufreq_policy *policy,
 411                                unsigned int index)
 412 {
 413         struct acpi_cpufreq_data *data = policy->driver_data;
 414         struct acpi_processor_performance *perf;
 415         const struct cpumask *mask;
 416         unsigned int next_perf_state = 0; /* Index into perf table */
 417         int result = 0;
 418
 419         if (unlikely(!data)) {
 420                 return -ENODEV;
 421         }
 422
 423         perf = to_perf_data(data);
 424         next_perf_state = policy->freq_table[index].driver_data;
 425         if (perf->state == next_perf_state) {
 426                 if (unlikely(data->resume)) {
 427                         pr_debug("Called after resume, resetting to P%d\n",
 428                                 next_perf_state);
 429                         data->resume = 0;
 430                 } else {
 431                         pr_debug("Already at target state (P%d)\n",
 432                                 next_perf_state);
 433                         return 0;
 434                 }
 435         }
 436
 437         /*
 438          * The core won't allow CPUs to go away until the governor has been
 439          * stopped, so we can rely on the stability of policy->cpus.
 440          */
 441         mask = policy->shared_type == CPUFREQ_SHARED_TYPE_ANY ?
 442                 cpumask_of(policy->cpu) : policy->cpus;
 443
 444         drv_write(data, mask, perf->states[next_perf_state].control);
 445
 446         if (acpi_pstate_strict) {
 447                 if (!check_freqs(policy, mask,
 448                                  policy->freq_table[index].frequency)) {
 449                         pr_debug("%s (%d)\n", __func__, policy->cpu);
 450                         result = -EAGAIN;
 451                 }
 452         }
 453
 454         if (!result)
 455                 perf->state = next_perf_state;
 456
 457         return result;
 458 }
 459
 460 static unsigned int acpi_cpufreq_fast_switch(struct cpufreq_policy *policy,
 461                                              unsigned int target_freq)
 462 {
 463         struct acpi_cpufreq_data *data = policy->driver_data;
 464         struct acpi_processor_performance *perf;
 465         struct cpufreq_frequency_table *entry;
 466         unsigned int next_perf_state, next_freq, index;
 467
 468         /*
 469          * Find the closest frequency above target_freq.
 470          */
 471         if (policy->cached_target_freq == target_freq)
 472                 index = policy->cached_resolved_idx;
 473         else
 474                 index = cpufreq_table_find_index_dl(policy, target_freq,
 475                                                     false);
 476
 477         entry = &policy->freq_table[index];
 478         next_freq = entry->frequency;
 479         next_perf_state = entry->driver_data;
 480
 481         perf = to_perf_data(data);
 482         if (perf->state == next_perf_state) {
 483                 if (unlikely(data->resume))
 484                         data->resume = 0;
 485                 else
 486                         return next_freq;
 487         }
 488
 489         data->cpu_freq_write(&perf->control_register,
 490                              perf->states[next_perf_state].control);
 491         perf->state = next_perf_state;
 492         return next_freq;
 493 }
 494
 495 static unsigned long
 496 acpi_cpufreq_guess_freq(struct acpi_cpufreq_data *data, unsigned int cpu)
 497 {
 498         struct acpi_processor_performance *perf;
 499
 500         perf = to_perf_data(data);
 501         if (cpu_khz) {
 502                 /* search the closest match to cpu_khz */
 503                 unsigned int i;
 504                 unsigned long freq;
 505                 unsigned long freqn = perf->states[0].core_frequency * 1000;
 506
 507                 for (i = 0; i < (perf->state_count-1); i++) {
 508                         freq = freqn;
 509                         freqn = perf->states[i+1].core_frequency * 1000;
 510                         if ((2 * cpu_khz) > (freqn + freq)) {
 511                                 perf->state = i;
 512                                 return freq;
 513                         }
 514                 }
 515                 perf->state = perf->state_count-1;
 516                 return freqn;
 517         } else {
 518                 /* assume CPU is at P0... */
 519                 perf->state = 0;
 520                 return perf->states[0].core_frequency * 1000;
 521         }
 522 }
 523
 524 static void free_acpi_perf_data(void)
 525 {
 526         unsigned int i;
 527
 528         /* Freeing a NULL pointer is OK, and alloc_percpu zeroes. */
 529         for_each_possible_cpu(i)
 530                 free_cpumask_var(per_cpu_ptr(acpi_perf_data, i)
 531                                  ->shared_cpu_map);
 532         free_percpu(acpi_perf_data);
 533 }
 534
 535 static int cpufreq_boost_down_prep(unsigned int cpu)
 536 {
 537         /*
 538          * Clear the boost-disable bit on the CPU_DOWN path so that
 539          * this cpu cannot block the remaining ones from boosting.
 540          */
 541         return boost_set_msr(1);
 542 }
 543
 544 /*
 545  * acpi_cpufreq_early_init - initialize ACPI P-States library
 546  *
 547  * Initialize the ACPI P-States library (drivers/acpi/processor_perflib.c)
 548  * in order to determine correct frequency and voltage pairings. We can
 549  * do _PDC and _PSD and find out the processor dependency for the
 550  * actual init that will happen later...
 551  */
 552 static int __init acpi_cpufreq_early_init(void)
 553 {
 554         unsigned int i;
 555         pr_debug("%s\n", __func__);
 556
 557         acpi_perf_data = alloc_percpu(struct acpi_processor_performance);
 558         if (!acpi_perf_data) {
 559                 pr_debug("Memory allocation error for acpi_perf_data.\n");
 560                 return -ENOMEM;
 561         }
 562         for_each_possible_cpu(i) {
 563                 if (!zalloc_cpumask_var_node(
 564                         &per_cpu_ptr(acpi_perf_data, i)->shared_cpu_map,
 565                         GFP_KERNEL, cpu_to_node(i))) {
 566
 567                         /* Freeing a NULL pointer is OK: alloc_percpu zeroes. */
 568                         free_acpi_perf_data();
 569                         return -ENOMEM;
 570                 }
 571         }
 572
 573         /* Do initialization in ACPI core */
 574         acpi_processor_preregister_performance(acpi_perf_data);
 575         return 0;
 576 }
 577
 578 #ifdef CONFIG_SMP
 579 /*
 580  * Some BIOSes do SW_ANY coordination internally, either set it up in hw
 581  * or do it in BIOS firmware and won't inform about it to OS. If not
 582  * detected, this has a side effect of making CPU run at a different speed
 583  * than OS intended it to run at. Detect it and handle it cleanly.
 584  */
 585 static int bios_with_sw_any_bug;
 586
 587 static int sw_any_bug_found(const struct dmi_system_id *d)
 588 {
 589         bios_with_sw_any_bug = 1;
 590         return 0;
 591 }
 592
 593 static const struct dmi_system_id sw_any_bug_dmi_table[] = {
 594         {
 595                 .callback = sw_any_bug_found,
 596                 .ident = "Supermicro Server X6DLP",
 597                 .matches = {
 598                         DMI_MATCH(DMI_SYS_VENDOR, "Supermicro"),
 599                         DMI_MATCH(DMI_BIOS_VERSION, "080010"),
 600                         DMI_MATCH(DMI_PRODUCT_NAME, "X6DLP"),
 601                 },
 602         },
 603         { }
 604 };
 605
 606 static int acpi_cpufreq_blacklist(struct cpuinfo_x86 *c)
 607 {
 608         /* Intel Xeon Processor 7100 Series Specification Update
 609          * https://www.intel.com/Assets/PDF/specupdate/314554.pdf
 610          * AL30: A Machine Check Exception (MCE) Occurring during an
 611          * Enhanced Intel SpeedStep Technology Ratio Change May Cause
 612          * Both Processor Cores to Lock Up. */
 613         if (c->x86_vendor == X86_VENDOR_INTEL) {
 614                 if ((c->x86 == 15) &&
 615                     (c->x86_model == 6) &&
 616                     (c->x86_stepping == 8)) {
 617                         pr_info("Intel(R) Xeon(R) 7100 Errata AL30, processors may lock up on frequency changes: disabling acpi-cpufreq\n");
 618                         return -ENODEV;
 619                     }
 620                 }
 621         return 0;
 622 }
 623 #endif
 624
 625 #ifdef CONFIG_ACPI_CPPC_LIB
 626 static u64 get_max_boost_ratio(unsigned int cpu)
 627 {
 628         struct cppc_perf_caps perf_caps;
 629         u64 highest_perf, nominal_perf;
 630         int ret;
 631
 632         if (acpi_pstate_strict)
 633                 return 0;
 634
 635         ret = cppc_get_perf_caps(cpu, &perf_caps);
 636         if (ret) {
 637                 pr_debug("CPU%d: Unable to get performance capabilities (%d)\n",
 638                          cpu, ret);
 639                 return 0;
 640         }
 641
 642         if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD) {
 643                 ret = amd_get_boost_ratio_numerator(cpu, &highest_perf);
 644                 if (ret) {
 645                         pr_debug("CPU%d: Unable to get boost ratio numerator (%d)\n",
 646                                  cpu, ret);
 647                         return 0;
 648                 }
 649         } else {
 650                 highest_perf = perf_caps.highest_perf;
 651         }
 652
 653         nominal_perf = perf_caps.nominal_perf;
 654
 655         if (!highest_perf || !nominal_perf) {
 656                 pr_debug("CPU%d: highest or nominal performance missing\n", cpu);
 657                 return 0;
 658         }
 659
 660         if (highest_perf < nominal_perf) {
 661                 pr_debug("CPU%d: nominal performance above highest\n", cpu);
 662                 return 0;
 663         }
 664
 665         return div_u64(highest_perf << SCHED_CAPACITY_SHIFT, nominal_perf);
 666 }
 667 #else
 668 static inline u64 get_max_boost_ratio(unsigned int cpu) { return 0; }
 669 #endif
 670
 671 static int acpi_cpufreq_cpu_init(struct cpufreq_policy *policy)
 672 {
 673         struct cpufreq_frequency_table *freq_table;
 674         struct acpi_processor_performance *perf;
 675         struct acpi_cpufreq_data *data;
 676         unsigned int cpu = policy->cpu;
 677         struct cpuinfo_x86 *c = &cpu_data(cpu);
 678         unsigned int valid_states = 0;
 679         unsigned int result = 0;
 680         u64 max_boost_ratio;
 681         unsigned int i;
 682 #ifdef CONFIG_SMP
 683         static int blacklisted;
 684 #endif
 685
 686         pr_debug("%s\n", __func__);
 687
 688 #ifdef CONFIG_SMP
 689         if (blacklisted)
 690                 return blacklisted;
 691         blacklisted = acpi_cpufreq_blacklist(c);
 692         if (blacklisted)
 693                 return blacklisted;
 694 #endif
 695
 696         data = kzalloc(sizeof(*data), GFP_KERNEL);
 697         if (!data)
 698                 return -ENOMEM;
 699
 700         if (!zalloc_cpumask_var(&data->freqdomain_cpus, GFP_KERNEL)) {
 701                 result = -ENOMEM;
 702                 goto err_free;
 703         }
 704
 705         perf = per_cpu_ptr(acpi_perf_data, cpu);
 706         data->acpi_perf_cpu = cpu;
 707         policy->driver_data = data;
 708
 709         if (cpu_has(c, X86_FEATURE_CONSTANT_TSC))
 710                 acpi_cpufreq_driver.flags |= CPUFREQ_CONST_LOOPS;
 711
 712         result = acpi_processor_register_performance(perf, cpu);
 713         if (result)
 714                 goto err_free_mask;
 715
 716         policy->shared_type = perf->shared_type;
 717
 718         /*
 719          * Will let policy->cpus know about dependency only when software
 720          * coordination is required.
 721          */
 722         if (policy->shared_type == CPUFREQ_SHARED_TYPE_ALL ||
 723             policy->shared_type == CPUFREQ_SHARED_TYPE_ANY) {
 724                 cpumask_copy(policy->cpus, perf->shared_cpu_map);
 725         }
 726         cpumask_copy(data->freqdomain_cpus, perf->shared_cpu_map);
 727
 728 #ifdef CONFIG_SMP
 729         dmi_check_system(sw_any_bug_dmi_table);
 730         if (bios_with_sw_any_bug && !policy_is_shared(policy)) {
 731                 policy->shared_type = CPUFREQ_SHARED_TYPE_ALL;
 732                 cpumask_copy(policy->cpus, topology_core_cpumask(cpu));
 733         }
 734
 735         if (check_amd_hwpstate_cpu(cpu) && boot_cpu_data.x86 < 0x19 &&
 736             !acpi_pstate_strict) {
 737                 cpumask_clear(policy->cpus);
 738                 cpumask_set_cpu(cpu, policy->cpus);
 739                 cpumask_copy(data->freqdomain_cpus,
 740                              topology_sibling_cpumask(cpu));
 741                 policy->shared_type = CPUFREQ_SHARED_TYPE_HW;
 742                 pr_info_once("overriding BIOS provided _PSD data\n");
 743         }
 744 #endif
 745
 746         /* capability check */
 747         if (perf->state_count <= 1) {
 748                 pr_debug("No P-States\n");
 749                 result = -ENODEV;
 750                 goto err_unreg;
 751         }
 752
 753         if (perf->control_register.space_id != perf->status_register.space_id) {
 754                 result = -ENODEV;
 755                 goto err_unreg;
 756         }
 757
 758         switch (perf->control_register.space_id) {
 759         case ACPI_ADR_SPACE_SYSTEM_IO:
 760                 if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD &&
 761                     boot_cpu_data.x86 == 0xf) {
 762                         pr_debug("AMD K8 systems must use native drivers.\n");
 763                         result = -ENODEV;
 764                         goto err_unreg;
 765                 }
 766                 pr_debug("SYSTEM IO addr space\n");
 767                 data->cpu_feature = SYSTEM_IO_CAPABLE;
 768                 data->cpu_freq_read = cpu_freq_read_io;
 769                 data->cpu_freq_write = cpu_freq_write_io;
 770                 break;
 771         case ACPI_ADR_SPACE_FIXED_HARDWARE:
 772                 pr_debug("HARDWARE addr space\n");
 773                 if (check_est_cpu(cpu)) {
 774                         data->cpu_feature = SYSTEM_INTEL_MSR_CAPABLE;
 775                         data->cpu_freq_read = cpu_freq_read_intel;
 776                         data->cpu_freq_write = cpu_freq_write_intel;
 777                         break;
 778                 }
 779                 if (check_amd_hwpstate_cpu(cpu)) {
 780                         data->cpu_feature = SYSTEM_AMD_MSR_CAPABLE;
 781                         data->cpu_freq_read = cpu_freq_read_amd;
 782                         data->cpu_freq_write = cpu_freq_write_amd;
 783                         break;
 784                 }
 785                 result = -ENODEV;
 786                 goto err_unreg;
 787         default:
 788                 pr_debug("Unknown addr space %d\n",
 789                         (u32) (perf->control_register.space_id));
 790                 result = -ENODEV;
 791                 goto err_unreg;
 792         }
 793
 794         freq_table = kcalloc(perf->state_count + 1, sizeof(*freq_table),
 795                              GFP_KERNEL);
 796         if (!freq_table) {
 797                 result = -ENOMEM;
 798                 goto err_unreg;
 799         }
 800
 801         /* detect transition latency */
 802         policy->cpuinfo.transition_latency = 0;
 803         for (i = 0; i < perf->state_count; i++) {
 804                 if ((perf->states[i].transition_latency * 1000) >
 805                     policy->cpuinfo.transition_latency)
 806                         policy->cpuinfo.transition_latency =
 807                             perf->states[i].transition_latency * 1000;
 808         }
 809
 810         /* Check for high latency (>20uS) from buggy BIOSes, like on T42 */
 811         if (perf->control_register.space_id == ACPI_ADR_SPACE_FIXED_HARDWARE &&
 812             policy->cpuinfo.transition_latency > 20 * 1000) {
 813                 policy->cpuinfo.transition_latency = 20 * 1000;
 814                 pr_info_once("P-state transition latency capped at 20 uS\n");
 815         }
 816
 817         /* table init */
 818         for (i = 0; i < perf->state_count; i++) {
 819                 if (i > 0 && perf->states[i].core_frequency >=
 820                     freq_table[valid_states-1].frequency / 1000)
 821                         continue;
 822
 823                 freq_table[valid_states].driver_data = i;
 824                 freq_table[valid_states].frequency =
 825                     perf->states[i].core_frequency * 1000;
 826                 valid_states++;
 827         }
 828         freq_table[valid_states].frequency = CPUFREQ_TABLE_END;
 829
 830         max_boost_ratio = get_max_boost_ratio(cpu);
 831         if (max_boost_ratio) {
 832                 unsigned int freq = freq_table[0].frequency;
 833
 834                 /*
 835                  * Because the loop above sorts the freq_table entries in the
 836                  * descending order, freq is the maximum frequency in the table.
 837                  * Assume that it corresponds to the CPPC nominal frequency and
 838                  * use it to set cpuinfo.max_freq.
 839                  */
 840                 policy->cpuinfo.max_freq = freq * max_boost_ratio >> SCHED_CAPACITY_SHIFT;
 841         } else {
 842                 /*
 843                  * If the maximum "boost" frequency is unknown, ask the arch
 844                  * scale-invariance code to use the "nominal" performance for
 845                  * CPU utilization scaling so as to prevent the schedutil
 846                  * governor from selecting inadequate CPU frequencies.
 847                  */
 848                 arch_set_max_freq_ratio(true);
 849         }
 850
 851         policy->freq_table = freq_table;
 852         perf->state = 0;
 853
 854         switch (perf->control_register.space_id) {
 855         case ACPI_ADR_SPACE_SYSTEM_IO:
 856                 /*
 857                  * The core will not set policy->cur, because
 858                  * cpufreq_driver->get is NULL, so we need to set it here.
 859                  * However, we have to guess it, because the current speed is
 860                  * unknown and not detectable via IO ports.
 861                  */
 862                 policy->cur = acpi_cpufreq_guess_freq(data, policy->cpu);
 863                 break;
 864         case ACPI_ADR_SPACE_FIXED_HARDWARE:
 865                 acpi_cpufreq_driver.get = get_cur_freq_on_cpu;
 866                 break;
 867         default:
 868                 break;
 869         }
 870
 871         /* notify BIOS that we exist */
 872         acpi_processor_notify_smm(THIS_MODULE);
 873
 874         pr_debug("CPU%u - ACPI performance management activated.\n", cpu);
 875         for (i = 0; i < perf->state_count; i++)
 876                 pr_debug("     %cP%d: %d MHz, %d mW, %d uS\n",
 877                         (i == perf->state ? '*' : ' '), i,
 878                         (u32) perf->states[i].core_frequency,
 879                         (u32) perf->states[i].power,
 880                         (u32) perf->states[i].transition_latency);
 881
 882         /*
 883          * the first call to ->target() should result in us actually
 884          * writing something to the appropriate registers.
 885          */
 886         data->resume = 1;
 887
 888         policy->fast_switch_possible = !acpi_pstate_strict &&
 889                 !(policy_is_shared(policy) && policy->shared_type != CPUFREQ_SHARED_TYPE_ANY);
 890
 891         if (perf->states[0].core_frequency * 1000 != freq_table[0].frequency)
 892                 pr_warn(FW_WARN "P-state 0 is not max freq\n");
 893
 894         if (acpi_cpufreq_driver.set_boost) {
 895                 set_boost(policy, acpi_cpufreq_driver.boost_enabled);
 896                 policy->boost_enabled = acpi_cpufreq_driver.boost_enabled;
 897         }
 898
 899         return result;
 900
 901 err_unreg:
 902         acpi_processor_unregister_performance(cpu);
 903 err_free_mask:
 904         free_cpumask_var(data->freqdomain_cpus);
 905 err_free:
 906         kfree(data);
 907         policy->driver_data = NULL;
 908
 909         return result;
 910 }
 911
 912 static void acpi_cpufreq_cpu_exit(struct cpufreq_policy *policy)
 913 {
 914         struct acpi_cpufreq_data *data = policy->driver_data;
 915
 916         pr_debug("%s\n", __func__);
 917
 918         cpufreq_boost_down_prep(policy->cpu);
 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
 927 static int acpi_cpufreq_resume(struct cpufreq_policy *policy)
 928 {
 929         struct acpi_cpufreq_data *data = policy->driver_data;
 930
 931         pr_debug("%s\n", __func__);
 932
 933         data->resume = 1;
 934
 935         return 0;
 936 }
 937
 938 static struct freq_attr *acpi_cpufreq_attr[] = {
 939         &cpufreq_freq_attr_scaling_available_freqs,
 940         &freqdomain_cpus,
 941 #ifdef CONFIG_X86_ACPI_CPUFREQ_CPB
 942         &cpb,
 943 #endif
 944         NULL,
 945 };
 946
 947 static struct cpufreq_driver acpi_cpufreq_driver = {
 948         .verify         = cpufreq_generic_frequency_table_verify,
 949         .target_index   = acpi_cpufreq_target,
 950         .fast_switch    = acpi_cpufreq_fast_switch,
 951         .bios_limit     = acpi_processor_get_bios_limit,
 952         .init           = acpi_cpufreq_cpu_init,
 953         .exit           = acpi_cpufreq_cpu_exit,
 954         .resume         = acpi_cpufreq_resume,
 955         .name           = "acpi-cpufreq",
 956         .attr           = acpi_cpufreq_attr,
 957 };
 958
 959 static void __init acpi_cpufreq_boost_init(void)
 960 {
 961         if (!(boot_cpu_has(X86_FEATURE_CPB) || boot_cpu_has(X86_FEATURE_IDA))) {
 962                 pr_debug("Boost capabilities not present in the processor\n");
 963                 return;
 964         }
 965
 966         acpi_cpufreq_driver.set_boost = set_boost;
 967         acpi_cpufreq_driver.boost_enabled = boost_state(0);
 968 }
 969
 970 static int __init acpi_cpufreq_probe(struct platform_device *pdev)
 971 {
 972         int ret;
 973
 974         if (acpi_disabled)
 975                 return -ENODEV;
 976
 977         /* don't keep reloading if cpufreq_driver exists */
 978         if (cpufreq_get_current_driver())
 979                 return -ENODEV;
 980
 981         pr_debug("%s\n", __func__);
 982
 983         ret = acpi_cpufreq_early_init();
 984         if (ret)
 985                 return ret;
 986
 987 #ifdef CONFIG_X86_ACPI_CPUFREQ_CPB
 988         /* this is a sysfs file with a strange name and an even stranger
 989          * semantic - per CPU instantiation, but system global effect.
 990          * Lets enable it only on AMD CPUs for compatibility reasons and
 991          * only if configured. This is considered legacy code, which
 992          * will probably be removed at some point in the future.
 993          */
 994         if (!check_amd_hwpstate_cpu(0)) {
 995                 struct freq_attr **attr;
 996
 997                 pr_debug("CPB unsupported, do not expose it\n");
 998
 999                 for (attr = acpi_cpufreq_attr; *attr; attr++)
1000                         if (*attr == &cpb) {
1001                                 *attr = NULL;
1002                                 break;
1003                         }
1004         }
1005 #endif
1006         acpi_cpufreq_boost_init();
1007
1008         ret = cpufreq_register_driver(&acpi_cpufreq_driver);
1009         if (ret) {
1010                 free_acpi_perf_data();
1011         }
1012         return ret;
1013 }
1014
1015 static void acpi_cpufreq_remove(struct platform_device *pdev)
1016 {
1017         pr_debug("%s\n", __func__);
1018
1019         cpufreq_unregister_driver(&acpi_cpufreq_driver);
1020
1021         free_acpi_perf_data();
1022 }
1023
1024 static struct platform_driver acpi_cpufreq_platdrv = {
1025         .driver = {
1026                 .name   = "acpi-cpufreq",
1027         },
1028         .remove = acpi_cpufreq_remove,
1029 };
1030
1031 static int __init acpi_cpufreq_init(void)
1032 {
1033         return platform_driver_probe(&acpi_cpufreq_platdrv, acpi_cpufreq_probe);
1034 }
1035
1036 static void __exit acpi_cpufreq_exit(void)
1037 {
1038         platform_driver_unregister(&acpi_cpufreq_platdrv);
1039 }
1040
1041 module_param(acpi_pstate_strict, uint, 0644);
1042 MODULE_PARM_DESC(acpi_pstate_strict,
1043         "value 0 or non-zero. non-zero -> strict ACPI checks are "
1044         "performed during frequency changes.");
1045
1046 late_initcall(acpi_cpufreq_init);
1047 module_exit(acpi_cpufreq_exit);
1048
1049 MODULE_ALIAS("platform:acpi-cpufreq");