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
  30 #include <asm/msr.h>
  31 #include <asm/processor.h>
  32 #include <asm/cpufeature.h>
  33 #include <asm/cpu_device_id.h>
  34
  35 MODULE_AUTHOR("Paul Diefenbaugh, Dominik Brodowski");
  36 MODULE_DESCRIPTION("ACPI Processor P-States Driver");
  37 MODULE_LICENSE("GPL");
  38
  39 enum {
  40         UNDEFINED_CAPABLE = 0,
  41         SYSTEM_INTEL_MSR_CAPABLE,
  42         SYSTEM_AMD_MSR_CAPABLE,
  43         SYSTEM_IO_CAPABLE,
  44 };
  45
  46 #define INTEL_MSR_RANGE         (0xffff)
  47 #define AMD_MSR_RANGE           (0x7)
  48 #define HYGON_MSR_RANGE         (0x7)
  49
  50 #define MSR_K7_HWCR_CPB_DIS     (1ULL << 25)
  51
  52 struct acpi_cpufreq_data {
  53         unsigned int resume;
  54         unsigned int cpu_feature;
  55         unsigned int acpi_perf_cpu;
  56         cpumask_var_t freqdomain_cpus;
  57         void (*cpu_freq_write)(struct acpi_pct_register *reg, u32 val);
  58         u32 (*cpu_freq_read)(struct acpi_pct_register *reg);
  59 };
  60
  61 /* acpi_perf_data is a pointer to percpu data. */
  62 static struct acpi_processor_performance __percpu *acpi_perf_data;
  63
  64 static inline struct acpi_processor_performance *to_perf_data(struct acpi_cpufreq_data *data)
  65 {
  66         return per_cpu_ptr(acpi_perf_data, data->acpi_perf_cpu);
  67 }
  68
  69 static struct cpufreq_driver acpi_cpufreq_driver;
  70
  71 static unsigned int acpi_pstate_strict;
  72
  73 static bool boost_state(unsigned int cpu)
  74 {
  75         u32 lo, hi;
  76         u64 msr;
  77
  78         switch (boot_cpu_data.x86_vendor) {
  79         case X86_VENDOR_INTEL:
  80                 rdmsr_on_cpu(cpu, MSR_IA32_MISC_ENABLE, &lo, &hi);
  81                 msr = lo | ((u64)hi << 32);
  82                 return !(msr & MSR_IA32_MISC_ENABLE_TURBO_DISABLE);
  83         case X86_VENDOR_HYGON:
  84         case X86_VENDOR_AMD:
  85                 rdmsr_on_cpu(cpu, MSR_K7_HWCR, &lo, &hi);
  86                 msr = lo | ((u64)hi << 32);
  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                 msr_addr = MSR_IA32_MISC_ENABLE;
 100                 msr_mask = MSR_IA32_MISC_ENABLE_TURBO_DISABLE;
 101                 break;
 102         case X86_VENDOR_HYGON:
 103         case X86_VENDOR_AMD:
 104                 msr_addr = MSR_K7_HWCR;
 105                 msr_mask = MSR_K7_HWCR_CPB_DIS;
 106                 break;
 107         default:
 108                 return -EINVAL;
 109         }
 110
 111         rdmsrl(msr_addr, val);
 112
 113         if (enable)
 114                 val &= ~msr_mask;
 115         else
 116                 val |= msr_mask;
 117
 118         wrmsrl(msr_addr, val);
 119         return 0;
 120 }
 121
 122 static void boost_set_msr_each(void *p_en)
 123 {
 124         bool enable = (bool) p_en;
 125
 126         boost_set_msr(enable);
 127 }
 128
 129 static int set_boost(int val)
 130 {
 131         get_online_cpus();
 132         on_each_cpu(boost_set_msr_each, (void *)(long)val, 1);
 133         put_online_cpus();
 134         pr_debug("Core Boosting %sabled.\n", val ? "en" : "dis");
 135
 136         return 0;
 137 }
 138
 139 static ssize_t show_freqdomain_cpus(struct cpufreq_policy *policy, char *buf)
 140 {
 141         struct acpi_cpufreq_data *data = policy->driver_data;
 142
 143         if (unlikely(!data))
 144                 return -ENODEV;
 145
 146         return cpufreq_show_cpus(data->freqdomain_cpus, buf);
 147 }
 148
 149 cpufreq_freq_attr_ro(freqdomain_cpus);
 150
 151 #ifdef CONFIG_X86_ACPI_CPUFREQ_CPB
 152 static ssize_t store_cpb(struct cpufreq_policy *policy, const char *buf,
 153                          size_t count)
 154 {
 155         int ret;
 156         unsigned int val = 0;
 157
 158         if (!acpi_cpufreq_driver.set_boost)
 159                 return -EINVAL;
 160
 161         ret = kstrtouint(buf, 10, &val);
 162         if (ret || val > 1)
 163                 return -EINVAL;
 164
 165         set_boost(val);
 166
 167         return count;
 168 }
 169
 170 static ssize_t show_cpb(struct cpufreq_policy *policy, char *buf)
 171 {
 172         return sprintf(buf, "%u\n", acpi_cpufreq_driver.boost_enabled);
 173 }
 174
 175 cpufreq_freq_attr_rw(cpb);
 176 #endif
 177
 178 static int check_est_cpu(unsigned int cpuid)
 179 {
 180         struct cpuinfo_x86 *cpu = &cpu_data(cpuid);
 181
 182         return cpu_has(cpu, X86_FEATURE_EST);
 183 }
 184
 185 static int check_amd_hwpstate_cpu(unsigned int cpuid)
 186 {
 187         struct cpuinfo_x86 *cpu = &cpu_data(cpuid);
 188
 189         return cpu_has(cpu, X86_FEATURE_HW_PSTATE);
 190 }
 191
 192 static unsigned extract_io(struct cpufreq_policy *policy, u32 value)
 193 {
 194         struct acpi_cpufreq_data *data = policy->driver_data;
 195         struct acpi_processor_performance *perf;
 196         int i;
 197
 198         perf = to_perf_data(data);
 199
 200         for (i = 0; i < perf->state_count; i++) {
 201                 if (value == perf->states[i].status)
 202                         return policy->freq_table[i].frequency;
 203         }
 204         return 0;
 205 }
 206
 207 static unsigned extract_msr(struct cpufreq_policy *policy, u32 msr)
 208 {
 209         struct acpi_cpufreq_data *data = policy->driver_data;
 210         struct cpufreq_frequency_table *pos;
 211         struct acpi_processor_performance *perf;
 212
 213         if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD)
 214                 msr &= AMD_MSR_RANGE;
 215         else if (boot_cpu_data.x86_vendor == X86_VENDOR_HYGON)
 216                 msr &= HYGON_MSR_RANGE;
 217         else
 218                 msr &= INTEL_MSR_RANGE;
 219
 220         perf = to_perf_data(data);
 221
 222         cpufreq_for_each_entry(pos, policy->freq_table)
 223                 if (msr == perf->states[pos->driver_data].status)
 224                         return pos->frequency;
 225         return policy->freq_table[0].frequency;
 226 }
 227
 228 static unsigned extract_freq(struct cpufreq_policy *policy, u32 val)
 229 {
 230         struct acpi_cpufreq_data *data = policy->driver_data;
 231
 232         switch (data->cpu_feature) {
 233         case SYSTEM_INTEL_MSR_CAPABLE:
 234         case SYSTEM_AMD_MSR_CAPABLE:
 235                 return extract_msr(policy, val);
 236         case SYSTEM_IO_CAPABLE:
 237                 return extract_io(policy, val);
 238         default:
 239                 return 0;
 240         }
 241 }
 242
 243 static u32 cpu_freq_read_intel(struct acpi_pct_register *not_used)
 244 {
 245         u32 val, dummy;
 246
 247         rdmsr(MSR_IA32_PERF_CTL, val, dummy);
 248         return val;
 249 }
 250
 251 static void cpu_freq_write_intel(struct acpi_pct_register *not_used, u32 val)
 252 {
 253         u32 lo, hi;
 254
 255         rdmsr(MSR_IA32_PERF_CTL, lo, hi);
 256         lo = (lo & ~INTEL_MSR_RANGE) | (val & INTEL_MSR_RANGE);
 257         wrmsr(MSR_IA32_PERF_CTL, lo, hi);
 258 }
 259
 260 static u32 cpu_freq_read_amd(struct acpi_pct_register *not_used)
 261 {
 262         u32 val, dummy;
 263
 264         rdmsr(MSR_AMD_PERF_CTL, val, dummy);
 265         return val;
 266 }
 267
 268 static void cpu_freq_write_amd(struct acpi_pct_register *not_used, u32 val)
 269 {
 270         wrmsr(MSR_AMD_PERF_CTL, val, 0);
 271 }
 272
 273 static u32 cpu_freq_read_io(struct acpi_pct_register *reg)
 274 {
 275         u32 val;
 276
 277         acpi_os_read_port(reg->address, &val, reg->bit_width);
 278         return val;
 279 }
 280
 281 static void cpu_freq_write_io(struct acpi_pct_register *reg, u32 val)
 282 {
 283         acpi_os_write_port(reg->address, val, reg->bit_width);
 284 }
 285
 286 struct drv_cmd {
 287         struct acpi_pct_register *reg;
 288         u32 val;
 289         union {
 290                 void (*write)(struct acpi_pct_register *reg, u32 val);
 291                 u32 (*read)(struct acpi_pct_register *reg);
 292         } func;
 293 };
 294
 295 /* Called via smp_call_function_single(), on the target CPU */
 296 static void do_drv_read(void *_cmd)
 297 {
 298         struct drv_cmd *cmd = _cmd;
 299
 300         cmd->val = cmd->func.read(cmd->reg);
 301 }
 302
 303 static u32 drv_read(struct acpi_cpufreq_data *data, const struct cpumask *mask)
 304 {
 305         struct acpi_processor_performance *perf = to_perf_data(data);
 306         struct drv_cmd cmd = {
 307                 .reg = &perf->control_register,
 308                 .func.read = data->cpu_freq_read,
 309         };
 310         int err;
 311
 312         err = smp_call_function_any(mask, do_drv_read, &cmd, 1);
 313         WARN_ON_ONCE(err);      /* smp_call_function_any() was buggy? */
 314         return cmd.val;
 315 }
 316
 317 /* Called via smp_call_function_many(), on the target CPUs */
 318 static void do_drv_write(void *_cmd)
 319 {
 320         struct drv_cmd *cmd = _cmd;
 321
 322         cmd->func.write(cmd->reg, cmd->val);
 323 }
 324
 325 static void drv_write(struct acpi_cpufreq_data *data,
 326                       const struct cpumask *mask, u32 val)
 327 {
 328         struct acpi_processor_performance *perf = to_perf_data(data);
 329         struct drv_cmd cmd = {
 330                 .reg = &perf->control_register,
 331                 .val = val,
 332                 .func.write = data->cpu_freq_write,
 333         };
 334         int this_cpu;
 335
 336         this_cpu = get_cpu();
 337         if (cpumask_test_cpu(this_cpu, mask))
 338                 do_drv_write(&cmd);
 339
 340         smp_call_function_many(mask, do_drv_write, &cmd, 1);
 341         put_cpu();
 342 }
 343
 344 static u32 get_cur_val(const struct cpumask *mask, struct acpi_cpufreq_data *data)
 345 {
 346         u32 val;
 347
 348         if (unlikely(cpumask_empty(mask)))
 349                 return 0;
 350
 351         val = drv_read(data, mask);
 352
 353         pr_debug("%s = %u\n", __func__, val);
 354
 355         return val;
 356 }
 357
 358 static unsigned int get_cur_freq_on_cpu(unsigned int cpu)
 359 {
 360         struct acpi_cpufreq_data *data;
 361         struct cpufreq_policy *policy;
 362         unsigned int freq;
 363         unsigned int cached_freq;
 364
 365         pr_debug("%s (%d)\n", __func__, cpu);
 366
 367         policy = cpufreq_cpu_get_raw(cpu);
 368         if (unlikely(!policy))
 369                 return 0;
 370
 371         data = policy->driver_data;
 372         if (unlikely(!data || !policy->freq_table))
 373                 return 0;
 374
 375         cached_freq = policy->freq_table[to_perf_data(data)->state].frequency;
 376         freq = extract_freq(policy, get_cur_val(cpumask_of(cpu), data));
 377         if (freq != cached_freq) {
 378                 /*
 379                  * The dreaded BIOS frequency change behind our back.
 380                  * Force set the frequency on next target call.
 381                  */
 382                 data->resume = 1;
 383         }
 384
 385         pr_debug("cur freq = %u\n", freq);
 386
 387         return freq;
 388 }
 389
 390 static unsigned int check_freqs(struct cpufreq_policy *policy,
 391                                 const struct cpumask *mask, unsigned int freq)
 392 {
 393         struct acpi_cpufreq_data *data = policy->driver_data;
 394         unsigned int cur_freq;
 395         unsigned int i;
 396
 397         for (i = 0; i < 100; i++) {
 398                 cur_freq = extract_freq(policy, get_cur_val(mask, data));
 399                 if (cur_freq == freq)
 400                         return 1;
 401                 udelay(10);
 402         }
 403         return 0;
 404 }
 405
 406 static int acpi_cpufreq_target(struct cpufreq_policy *policy,
 407                                unsigned int index)
 408 {
 409         struct acpi_cpufreq_data *data = policy->driver_data;
 410         struct acpi_processor_performance *perf;
 411         const struct cpumask *mask;
 412         unsigned int next_perf_state = 0; /* Index into perf table */
 413         int result = 0;
 414
 415         if (unlikely(!data)) {
 416                 return -ENODEV;
 417         }
 418
 419         perf = to_perf_data(data);
 420         next_perf_state = policy->freq_table[index].driver_data;
 421         if (perf->state == next_perf_state) {
 422                 if (unlikely(data->resume)) {
 423                         pr_debug("Called after resume, resetting to P%d\n",
 424                                 next_perf_state);
 425                         data->resume = 0;
 426                 } else {
 427                         pr_debug("Already at target state (P%d)\n",
 428                                 next_perf_state);
 429                         return 0;
 430                 }
 431         }
 432
 433         /*
 434          * The core won't allow CPUs to go away until the governor has been
 435          * stopped, so we can rely on the stability of policy->cpus.
 436          */
 437         mask = policy->shared_type == CPUFREQ_SHARED_TYPE_ANY ?
 438                 cpumask_of(policy->cpu) : policy->cpus;
 439
 440         drv_write(data, mask, perf->states[next_perf_state].control);
 441
 442         if (acpi_pstate_strict) {
 443                 if (!check_freqs(policy, mask,
 444                                  policy->freq_table[index].frequency)) {
 445                         pr_debug("%s (%d)\n", __func__, policy->cpu);
 446                         result = -EAGAIN;
 447                 }
 448         }
 449
 450         if (!result)
 451                 perf->state = next_perf_state;
 452
 453         return result;
 454 }
 455
 456 static unsigned int acpi_cpufreq_fast_switch(struct cpufreq_policy *policy,
 457                                              unsigned int target_freq)
 458 {
 459         struct acpi_cpufreq_data *data = policy->driver_data;
 460         struct acpi_processor_performance *perf;
 461         struct cpufreq_frequency_table *entry;
 462         unsigned int next_perf_state, next_freq, index;
 463
 464         /*
 465          * Find the closest frequency above target_freq.
 466          */
 467         if (policy->cached_target_freq == target_freq)
 468                 index = policy->cached_resolved_idx;
 469         else
 470                 index = cpufreq_table_find_index_dl(policy, target_freq);
 471
 472         entry = &policy->freq_table[index];
 473         next_freq = entry->frequency;
 474         next_perf_state = entry->driver_data;
 475
 476         perf = to_perf_data(data);
 477         if (perf->state == next_perf_state) {
 478                 if (unlikely(data->resume))
 479                         data->resume = 0;
 480                 else
 481                         return next_freq;
 482         }
 483
 484         data->cpu_freq_write(&perf->control_register,
 485                              perf->states[next_perf_state].control);
 486         perf->state = next_perf_state;
 487         return next_freq;
 488 }
 489
 490 static unsigned long
 491 acpi_cpufreq_guess_freq(struct acpi_cpufreq_data *data, unsigned int cpu)
 492 {
 493         struct acpi_processor_performance *perf;
 494
 495         perf = to_perf_data(data);
 496         if (cpu_khz) {
 497                 /* search the closest match to cpu_khz */
 498                 unsigned int i;
 499                 unsigned long freq;
 500                 unsigned long freqn = perf->states[0].core_frequency * 1000;
 501
 502                 for (i = 0; i < (perf->state_count-1); i++) {
 503                         freq = freqn;
 504                         freqn = perf->states[i+1].core_frequency * 1000;
 505                         if ((2 * cpu_khz) > (freqn + freq)) {
 506                                 perf->state = i;
 507                                 return freq;
 508                         }
 509                 }
 510                 perf->state = perf->state_count-1;
 511                 return freqn;
 512         } else {
 513                 /* assume CPU is at P0... */
 514                 perf->state = 0;
 515                 return perf->states[0].core_frequency * 1000;
 516         }
 517 }
 518
 519 static void free_acpi_perf_data(void)
 520 {
 521         unsigned int i;
 522
 523         /* Freeing a NULL pointer is OK, and alloc_percpu zeroes. */
 524         for_each_possible_cpu(i)
 525                 free_cpumask_var(per_cpu_ptr(acpi_perf_data, i)
 526                                  ->shared_cpu_map);
 527         free_percpu(acpi_perf_data);
 528 }
 529
 530 static int cpufreq_boost_online(unsigned int cpu)
 531 {
 532         /*
 533          * On the CPU_UP path we simply keep the boost-disable flag
 534          * in sync with the current global state.
 535          */
 536         return boost_set_msr(acpi_cpufreq_driver.boost_enabled);
 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          * http://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 static int acpi_cpufreq_cpu_init(struct cpufreq_policy *policy)
 630 {
 631         unsigned int i;
 632         unsigned int valid_states = 0;
 633         unsigned int cpu = policy->cpu;
 634         struct acpi_cpufreq_data *data;
 635         unsigned int result = 0;
 636         struct cpuinfo_x86 *c = &cpu_data(policy->cpu);
 637         struct acpi_processor_performance *perf;
 638         struct cpufreq_frequency_table *freq_table;
 639 #ifdef CONFIG_SMP
 640         static int blacklisted;
 641 #endif
 642
 643         pr_debug("%s\n", __func__);
 644
 645 #ifdef CONFIG_SMP
 646         if (blacklisted)
 647                 return blacklisted;
 648         blacklisted = acpi_cpufreq_blacklist(c);
 649         if (blacklisted)
 650                 return blacklisted;
 651 #endif
 652
 653         data = kzalloc(sizeof(*data), GFP_KERNEL);
 654         if (!data)
 655                 return -ENOMEM;
 656
 657         if (!zalloc_cpumask_var(&data->freqdomain_cpus, GFP_KERNEL)) {
 658                 result = -ENOMEM;
 659                 goto err_free;
 660         }
 661
 662         perf = per_cpu_ptr(acpi_perf_data, cpu);
 663         data->acpi_perf_cpu = cpu;
 664         policy->driver_data = data;
 665
 666         if (cpu_has(c, X86_FEATURE_CONSTANT_TSC))
 667                 acpi_cpufreq_driver.flags |= CPUFREQ_CONST_LOOPS;
 668
 669         result = acpi_processor_register_performance(perf, cpu);
 670         if (result)
 671                 goto err_free_mask;
 672
 673         policy->shared_type = perf->shared_type;
 674
 675         /*
 676          * Will let policy->cpus know about dependency only when software
 677          * coordination is required.
 678          */
 679         if (policy->shared_type == CPUFREQ_SHARED_TYPE_ALL ||
 680             policy->shared_type == CPUFREQ_SHARED_TYPE_ANY) {
 681                 cpumask_copy(policy->cpus, perf->shared_cpu_map);
 682         }
 683         cpumask_copy(data->freqdomain_cpus, perf->shared_cpu_map);
 684
 685 #ifdef CONFIG_SMP
 686         dmi_check_system(sw_any_bug_dmi_table);
 687         if (bios_with_sw_any_bug && !policy_is_shared(policy)) {
 688                 policy->shared_type = CPUFREQ_SHARED_TYPE_ALL;
 689                 cpumask_copy(policy->cpus, topology_core_cpumask(cpu));
 690         }
 691
 692         if (check_amd_hwpstate_cpu(cpu) && boot_cpu_data.x86 < 0x19 &&
 693             !acpi_pstate_strict) {
 694                 cpumask_clear(policy->cpus);
 695                 cpumask_set_cpu(cpu, policy->cpus);
 696                 cpumask_copy(data->freqdomain_cpus,
 697                              topology_sibling_cpumask(cpu));
 698                 policy->shared_type = CPUFREQ_SHARED_TYPE_HW;
 699                 pr_info_once("overriding BIOS provided _PSD data\n");
 700         }
 701 #endif
 702
 703         /* capability check */
 704         if (perf->state_count <= 1) {
 705                 pr_debug("No P-States\n");
 706                 result = -ENODEV;
 707                 goto err_unreg;
 708         }
 709
 710         if (perf->control_register.space_id != perf->status_register.space_id) {
 711                 result = -ENODEV;
 712                 goto err_unreg;
 713         }
 714
 715         switch (perf->control_register.space_id) {
 716         case ACPI_ADR_SPACE_SYSTEM_IO:
 717                 if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD &&
 718                     boot_cpu_data.x86 == 0xf) {
 719                         pr_debug("AMD K8 systems must use native drivers.\n");
 720                         result = -ENODEV;
 721                         goto err_unreg;
 722                 }
 723                 pr_debug("SYSTEM IO addr space\n");
 724                 data->cpu_feature = SYSTEM_IO_CAPABLE;
 725                 data->cpu_freq_read = cpu_freq_read_io;
 726                 data->cpu_freq_write = cpu_freq_write_io;
 727                 break;
 728         case ACPI_ADR_SPACE_FIXED_HARDWARE:
 729                 pr_debug("HARDWARE addr space\n");
 730                 if (check_est_cpu(cpu)) {
 731                         data->cpu_feature = SYSTEM_INTEL_MSR_CAPABLE;
 732                         data->cpu_freq_read = cpu_freq_read_intel;
 733                         data->cpu_freq_write = cpu_freq_write_intel;
 734                         break;
 735                 }
 736                 if (check_amd_hwpstate_cpu(cpu)) {
 737                         data->cpu_feature = SYSTEM_AMD_MSR_CAPABLE;
 738                         data->cpu_freq_read = cpu_freq_read_amd;
 739                         data->cpu_freq_write = cpu_freq_write_amd;
 740                         break;
 741                 }
 742                 result = -ENODEV;
 743                 goto err_unreg;
 744         default:
 745                 pr_debug("Unknown addr space %d\n",
 746                         (u32) (perf->control_register.space_id));
 747                 result = -ENODEV;
 748                 goto err_unreg;
 749         }
 750
 751         freq_table = kcalloc(perf->state_count + 1, sizeof(*freq_table),
 752                              GFP_KERNEL);
 753         if (!freq_table) {
 754                 result = -ENOMEM;
 755                 goto err_unreg;
 756         }
 757
 758         /* detect transition latency */
 759         policy->cpuinfo.transition_latency = 0;
 760         for (i = 0; i < perf->state_count; i++) {
 761                 if ((perf->states[i].transition_latency * 1000) >
 762                     policy->cpuinfo.transition_latency)
 763                         policy->cpuinfo.transition_latency =
 764                             perf->states[i].transition_latency * 1000;
 765         }
 766
 767         /* Check for high latency (>20uS) from buggy BIOSes, like on T42 */
 768         if (perf->control_register.space_id == ACPI_ADR_SPACE_FIXED_HARDWARE &&
 769             policy->cpuinfo.transition_latency > 20 * 1000) {
 770                 policy->cpuinfo.transition_latency = 20 * 1000;
 771                 pr_info_once("P-state transition latency capped at 20 uS\n");
 772         }
 773
 774         /* table init */
 775         for (i = 0; i < perf->state_count; i++) {
 776                 if (i > 0 && perf->states[i].core_frequency >=
 777                     freq_table[valid_states-1].frequency / 1000)
 778                         continue;
 779
 780                 freq_table[valid_states].driver_data = i;
 781                 freq_table[valid_states].frequency =
 782                     perf->states[i].core_frequency * 1000;
 783                 valid_states++;
 784         }
 785         freq_table[valid_states].frequency = CPUFREQ_TABLE_END;
 786         policy->freq_table = freq_table;
 787         perf->state = 0;
 788
 789         switch (perf->control_register.space_id) {
 790         case ACPI_ADR_SPACE_SYSTEM_IO:
 791                 /*
 792                  * The core will not set policy->cur, because
 793                  * cpufreq_driver->get is NULL, so we need to set it here.
 794                  * However, we have to guess it, because the current speed is
 795                  * unknown and not detectable via IO ports.
 796                  */
 797                 policy->cur = acpi_cpufreq_guess_freq(data, policy->cpu);
 798                 break;
 799         case ACPI_ADR_SPACE_FIXED_HARDWARE:
 800                 acpi_cpufreq_driver.get = get_cur_freq_on_cpu;
 801                 break;
 802         default:
 803                 break;
 804         }
 805
 806         /* notify BIOS that we exist */
 807         acpi_processor_notify_smm(THIS_MODULE);
 808
 809         pr_debug("CPU%u - ACPI performance management activated.\n", cpu);
 810         for (i = 0; i < perf->state_count; i++)
 811                 pr_debug("     %cP%d: %d MHz, %d mW, %d uS\n",
 812                         (i == perf->state ? '*' : ' '), i,
 813                         (u32) perf->states[i].core_frequency,
 814                         (u32) perf->states[i].power,
 815                         (u32) perf->states[i].transition_latency);
 816
 817         /*
 818          * the first call to ->target() should result in us actually
 819          * writing something to the appropriate registers.
 820          */
 821         data->resume = 1;
 822
 823         policy->fast_switch_possible = !acpi_pstate_strict &&
 824                 !(policy_is_shared(policy) && policy->shared_type != CPUFREQ_SHARED_TYPE_ANY);
 825
 826         return result;
 827
 828 err_unreg:
 829         acpi_processor_unregister_performance(cpu);
 830 err_free_mask:
 831         free_cpumask_var(data->freqdomain_cpus);
 832 err_free:
 833         kfree(data);
 834         policy->driver_data = NULL;
 835
 836         return result;
 837 }
 838
 839 static int acpi_cpufreq_cpu_exit(struct cpufreq_policy *policy)
 840 {
 841         struct acpi_cpufreq_data *data = policy->driver_data;
 842
 843         pr_debug("%s\n", __func__);
 844
 845         policy->fast_switch_possible = false;
 846         policy->driver_data = NULL;
 847         acpi_processor_unregister_performance(data->acpi_perf_cpu);
 848         free_cpumask_var(data->freqdomain_cpus);
 849         kfree(policy->freq_table);
 850         kfree(data);
 851
 852         return 0;
 853 }
 854
 855 static void acpi_cpufreq_cpu_ready(struct cpufreq_policy *policy)
 856 {
 857         struct acpi_processor_performance *perf = per_cpu_ptr(acpi_perf_data,
 858                                                               policy->cpu);
 859
 860         if (perf->states[0].core_frequency * 1000 != policy->cpuinfo.max_freq)
 861                 pr_warn(FW_WARN "P-state 0 is not max freq\n");
 862 }
 863
 864 static int acpi_cpufreq_resume(struct cpufreq_policy *policy)
 865 {
 866         struct acpi_cpufreq_data *data = policy->driver_data;
 867
 868         pr_debug("%s\n", __func__);
 869
 870         data->resume = 1;
 871
 872         return 0;
 873 }
 874
 875 static struct freq_attr *acpi_cpufreq_attr[] = {
 876         &cpufreq_freq_attr_scaling_available_freqs,
 877         &freqdomain_cpus,
 878 #ifdef CONFIG_X86_ACPI_CPUFREQ_CPB
 879         &cpb,
 880 #endif
 881         NULL,
 882 };
 883
 884 static struct cpufreq_driver acpi_cpufreq_driver = {
 885         .verify         = cpufreq_generic_frequency_table_verify,
 886         .target_index   = acpi_cpufreq_target,
 887         .fast_switch    = acpi_cpufreq_fast_switch,
 888         .bios_limit     = acpi_processor_get_bios_limit,
 889         .init           = acpi_cpufreq_cpu_init,
 890         .exit           = acpi_cpufreq_cpu_exit,
 891         .ready          = acpi_cpufreq_cpu_ready,
 892         .resume         = acpi_cpufreq_resume,
 893         .name           = "acpi-cpufreq",
 894         .attr           = acpi_cpufreq_attr,
 895 };
 896
 897 static enum cpuhp_state acpi_cpufreq_online;
 898
 899 static void __init acpi_cpufreq_boost_init(void)
 900 {
 901         int ret;
 902
 903         if (!(boot_cpu_has(X86_FEATURE_CPB) || boot_cpu_has(X86_FEATURE_IDA))) {
 904                 pr_debug("Boost capabilities not present in the processor\n");
 905                 return;
 906         }
 907
 908         acpi_cpufreq_driver.set_boost = set_boost;
 909         acpi_cpufreq_driver.boost_enabled = boost_state(0);
 910
 911         /*
 912          * This calls the online callback on all online cpu and forces all
 913          * MSRs to the same value.
 914          */
 915         ret = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "cpufreq/acpi:online",
 916                                 cpufreq_boost_online, cpufreq_boost_down_prep);
 917         if (ret < 0) {
 918                 pr_err("acpi_cpufreq: failed to register hotplug callbacks\n");
 919                 return;
 920         }
 921         acpi_cpufreq_online = ret;
 922 }
 923
 924 static void acpi_cpufreq_boost_exit(void)
 925 {
 926         if (acpi_cpufreq_online > 0)
 927                 cpuhp_remove_state_nocalls(acpi_cpufreq_online);
 928 }
 929
 930 static int __init acpi_cpufreq_init(void)
 931 {
 932         int ret;
 933
 934         if (acpi_disabled)
 935                 return -ENODEV;
 936
 937         /* don't keep reloading if cpufreq_driver exists */
 938         if (cpufreq_get_current_driver())
 939                 return -EEXIST;
 940
 941         pr_debug("%s\n", __func__);
 942
 943         ret = acpi_cpufreq_early_init();
 944         if (ret)
 945                 return ret;
 946
 947 #ifdef CONFIG_X86_ACPI_CPUFREQ_CPB
 948         /* this is a sysfs file with a strange name and an even stranger
 949          * semantic - per CPU instantiation, but system global effect.
 950          * Lets enable it only on AMD CPUs for compatibility reasons and
 951          * only if configured. This is considered legacy code, which
 952          * will probably be removed at some point in the future.
 953          */
 954         if (!check_amd_hwpstate_cpu(0)) {
 955                 struct freq_attr **attr;
 956
 957                 pr_debug("CPB unsupported, do not expose it\n");
 958
 959                 for (attr = acpi_cpufreq_attr; *attr; attr++)
 960                         if (*attr == &cpb) {
 961                                 *attr = NULL;
 962                                 break;
 963                         }
 964         }
 965 #endif
 966         acpi_cpufreq_boost_init();
 967
 968         ret = cpufreq_register_driver(&acpi_cpufreq_driver);
 969         if (ret) {
 970                 free_acpi_perf_data();
 971                 acpi_cpufreq_boost_exit();
 972         }
 973         return ret;
 974 }
 975
 976 static void __exit acpi_cpufreq_exit(void)
 977 {
 978         pr_debug("%s\n", __func__);
 979
 980         acpi_cpufreq_boost_exit();
 981
 982         cpufreq_unregister_driver(&acpi_cpufreq_driver);
 983
 984         free_acpi_perf_data();
 985 }
 986
 987 module_param(acpi_pstate_strict, uint, 0644);
 988 MODULE_PARM_DESC(acpi_pstate_strict,
 989         "value 0 or non-zero. non-zero -> strict ACPI checks are "
 990         "performed during frequency changes.");
 991
 992 late_initcall(acpi_cpufreq_init);
 993 module_exit(acpi_cpufreq_exit);
 994
 995 static const struct x86_cpu_id acpi_cpufreq_ids[] = {
 996         X86_FEATURE_MATCH(X86_FEATURE_ACPI),
 997         X86_FEATURE_MATCH(X86_FEATURE_HW_PSTATE),
 998         {}
 999 };
1000 MODULE_DEVICE_TABLE(x86cpu, acpi_cpufreq_ids);
1001
1002 static const struct acpi_device_id processor_device_ids[] = {
1003         {ACPI_PROCESSOR_OBJECT_HID, },
1004         {ACPI_PROCESSOR_DEVICE_HID, },
1005         {},
1006 };
1007 MODULE_DEVICE_TABLE(acpi, processor_device_ids);
1008
1009 MODULE_ALIAS("acpi");