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