drivers/cpufreq/acpi-cpufreq.c

   1 /*
   2  * acpi-cpufreq.c - ACPI Processor P-States Driver
   3  *
   4  *  Copyright (C) 2001, 2002 Andy Grover <andrew.grover@intel.com>
   5  *  Copyright (C) 2001, 2002 Paul Diefenbaugh <paul.s.diefenbaugh@intel.com>
   6  *  Copyright (C) 2002 - 2004 Dominik Brodowski <linux@brodo.de>
   7  *  Copyright (C) 2006       Denis Sadykov <denis.m.sadykov@intel.com>
   8  *
   9  * ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
  10  *
  11  *  This program is free software; you can redistribute it and/or modify
  12  *  it under the terms of the GNU General Public License as published by
  13  *  the Free Software Foundation; either version 2 of the License, or (at
  14  *  your option) any later version.
  15  *
  16  *  This program is distributed in the hope that it will be useful, but
  17  *  WITHOUT ANY WARRANTY; without even the implied warranty of
  18  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  19  *  General Public License for more details.
  20  *
  21  *  You should have received a copy of the GNU General Public License along
  22  *  with this program; if not, write to the Free Software Foundation, Inc.,
  23  *  59 Temple Place, Suite 330, Boston, MA 02111-1307 USA.
  24  *
  25  * ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
  26  */
  27
  28 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  29
  30 #include <linux/kernel.h>
  31 #include <linux/module.h>
  32 #include <linux/init.h>
  33 #include <linux/smp.h>
  34 #include <linux/sched.h>
  35 #include <linux/cpufreq.h>
  36 #include <linux/compiler.h>
  37 #include <linux/dmi.h>
  38 #include <linux/slab.h>
  39
  40 #include <linux/acpi.h>
  41 #include <linux/io.h>
  42 #include <linux/delay.h>
  43 #include <linux/uaccess.h>
  44
  45 #include <acpi/processor.h>
  46
  47 #include <asm/msr.h>
  48 #include <asm/processor.h>
  49 #include <asm/cpufeature.h>
  50
  51 MODULE_AUTHOR("Paul Diefenbaugh, Dominik Brodowski");
  52 MODULE_DESCRIPTION("ACPI Processor P-States Driver");
  53 MODULE_LICENSE("GPL");
  54
  55 enum {
  56         UNDEFINED_CAPABLE = 0,
  57         SYSTEM_INTEL_MSR_CAPABLE,
  58         SYSTEM_AMD_MSR_CAPABLE,
  59         SYSTEM_IO_CAPABLE,
  60 };
  61
  62 #define INTEL_MSR_RANGE         (0xffff)
  63 #define AMD_MSR_RANGE           (0x7)
  64 #define HYGON_MSR_RANGE         (0x7)
  65
  66 #define MSR_K7_HWCR_CPB_DIS     (1ULL << 25)
  67
  68 struct acpi_cpufreq_data {
  69         unsigned int resume;
  70         unsigned int cpu_feature;
  71         unsigned int acpi_perf_cpu;
  72         cpumask_var_t freqdomain_cpus;
  73         void (*cpu_freq_write)(struct acpi_pct_register *reg, u32 val);
  74         u32 (*cpu_freq_read)(struct acpi_pct_register *reg);
  75 };
  76
  77 /* acpi_perf_data is a pointer to percpu data. */
  78 static struct acpi_processor_performance __percpu *acpi_perf_data;
  79
  80 static inline struct acpi_processor_performance *to_perf_data(struct acpi_cpufreq_data *data)
  81 {
  82         return per_cpu_ptr(acpi_perf_data, data->acpi_perf_cpu);
  83 }
  84
  85 static struct cpufreq_driver acpi_cpufreq_driver;
  86
  87 static unsigned int acpi_pstate_strict;
  88
  89 static bool boost_state(unsigned int cpu)
  90 {
  91         u32 lo, hi;
  92         u64 msr;
  93
  94         switch (boot_cpu_data.x86_vendor) {
  95         case X86_VENDOR_INTEL:
  96                 rdmsr_on_cpu(cpu, MSR_IA32_MISC_ENABLE, &lo, &hi);
  97                 msr = lo | ((u64)hi << 32);
  98                 return !(msr & MSR_IA32_MISC_ENABLE_TURBO_DISABLE);
  99         case X86_VENDOR_HYGON:
 100         case X86_VENDOR_AMD:
 101                 rdmsr_on_cpu(cpu, MSR_K7_HWCR, &lo, &hi);
 102                 msr = lo | ((u64)hi << 32);
 103                 return !(msr & MSR_K7_HWCR_CPB_DIS);
 104         }
 105         return false;
 106 }
 107
 108 static int boost_set_msr(bool enable)
 109 {
 110         u32 msr_addr;
 111         u64 msr_mask, val;
 112
 113         switch (boot_cpu_data.x86_vendor) {
 114         case X86_VENDOR_INTEL:
 115                 msr_addr = MSR_IA32_MISC_ENABLE;
 116                 msr_mask = MSR_IA32_MISC_ENABLE_TURBO_DISABLE;
 117                 break;
 118         case X86_VENDOR_HYGON:
 119         case X86_VENDOR_AMD:
 120                 msr_addr = MSR_K7_HWCR;
 121                 msr_mask = MSR_K7_HWCR_CPB_DIS;
 122                 break;
 123         default:
 124                 return -EINVAL;
 125         }
 126
 127         rdmsrl(msr_addr, val);
 128
 129         if (enable)
 130                 val &= ~msr_mask;
 131         else
 132                 val |= msr_mask;
 133
 134         wrmsrl(msr_addr, val);
 135         return 0;
 136 }
 137
 138 static void boost_set_msr_each(void *p_en)
 139 {
 140         bool enable = (bool) p_en;
 141
 142         boost_set_msr(enable);
 143 }
 144
 145 static int set_boost(int val)
 146 {
 147         get_online_cpus();
 148         on_each_cpu(boost_set_msr_each, (void *)(long)val, 1);
 149         put_online_cpus();
 150         pr_debug("Core Boosting %sabled.\n", val ? "en" : "dis");
 151
 152         return 0;
 153 }
 154
 155 static ssize_t show_freqdomain_cpus(struct cpufreq_policy *policy, char *buf)
 156 {
 157         struct acpi_cpufreq_data *data = policy->driver_data;
 158
 159         if (unlikely(!data))
 160                 return -ENODEV;
 161
 162         return cpufreq_show_cpus(data->freqdomain_cpus, buf);
 163 }
 164
 165 cpufreq_freq_attr_ro(freqdomain_cpus);
 166
 167 #ifdef CONFIG_X86_ACPI_CPUFREQ_CPB
 168 static ssize_t store_cpb(struct cpufreq_policy *policy, const char *buf,
 169                          size_t count)
 170 {
 171         int ret;
 172         unsigned int val = 0;
 173
 174         if (!acpi_cpufreq_driver.set_boost)
 175                 return -EINVAL;
 176
 177         ret = kstrtouint(buf, 10, &val);
 178         if (ret || val > 1)
 179                 return -EINVAL;
 180
 181         set_boost(val);
 182
 183         return count;
 184 }
 185
 186 static ssize_t show_cpb(struct cpufreq_policy *policy, char *buf)
 187 {
 188         return sprintf(buf, "%u\n", acpi_cpufreq_driver.boost_enabled);
 189 }
 190
 191 cpufreq_freq_attr_rw(cpb);
 192 #endif
 193
 194 static int check_est_cpu(unsigned int cpuid)
 195 {
 196         struct cpuinfo_x86 *cpu = &cpu_data(cpuid);
 197
 198         return cpu_has(cpu, X86_FEATURE_EST);
 199 }
 200
 201 static int check_amd_hwpstate_cpu(unsigned int cpuid)
 202 {
 203         struct cpuinfo_x86 *cpu = &cpu_data(cpuid);
 204
 205         return cpu_has(cpu, X86_FEATURE_HW_PSTATE);
 206 }
 207
 208 static unsigned extract_io(struct cpufreq_policy *policy, u32 value)
 209 {
 210         struct acpi_cpufreq_data *data = policy->driver_data;
 211         struct acpi_processor_performance *perf;
 212         int i;
 213
 214         perf = to_perf_data(data);
 215
 216         for (i = 0; i < perf->state_count; i++) {
 217                 if (value == perf->states[i].status)
 218                         return policy->freq_table[i].frequency;
 219         }
 220         return 0;
 221 }
 222
 223 static unsigned extract_msr(struct cpufreq_policy *policy, u32 msr)
 224 {
 225         struct acpi_cpufreq_data *data = policy->driver_data;
 226         struct cpufreq_frequency_table *pos;
 227         struct acpi_processor_performance *perf;
 228
 229         if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD)
 230                 msr &= AMD_MSR_RANGE;
 231         else if (boot_cpu_data.x86_vendor == X86_VENDOR_HYGON)
 232                 msr &= HYGON_MSR_RANGE;
 233         else
 234                 msr &= INTEL_MSR_RANGE;
 235
 236         perf = to_perf_data(data);
 237
 238         cpufreq_for_each_entry(pos, policy->freq_table)
 239                 if (msr == perf->states[pos->driver_data].status)
 240                         return pos->frequency;
 241         return policy->freq_table[0].frequency;
 242 }
 243
 244 static unsigned extract_freq(struct cpufreq_policy *policy, u32 val)
 245 {
 246         struct acpi_cpufreq_data *data = policy->driver_data;
 247
 248         switch (data->cpu_feature) {
 249         case SYSTEM_INTEL_MSR_CAPABLE:
 250         case SYSTEM_AMD_MSR_CAPABLE:
 251                 return extract_msr(policy, val);
 252         case SYSTEM_IO_CAPABLE:
 253                 return extract_io(policy, val);
 254         default:
 255                 return 0;
 256         }
 257 }
 258
 259 static u32 cpu_freq_read_intel(struct acpi_pct_register *not_used)
 260 {
 261         u32 val, dummy;
 262
 263         rdmsr(MSR_IA32_PERF_CTL, val, dummy);
 264         return val;
 265 }
 266
 267 static void cpu_freq_write_intel(struct acpi_pct_register *not_used, u32 val)
 268 {
 269         u32 lo, hi;
 270
 271         rdmsr(MSR_IA32_PERF_CTL, lo, hi);
 272         lo = (lo & ~INTEL_MSR_RANGE) | (val & INTEL_MSR_RANGE);
 273         wrmsr(MSR_IA32_PERF_CTL, lo, hi);
 274 }
 275
 276 static u32 cpu_freq_read_amd(struct acpi_pct_register *not_used)
 277 {
 278         u32 val, dummy;
 279
 280         rdmsr(MSR_AMD_PERF_CTL, val, dummy);
 281         return val;
 282 }
 283
 284 static void cpu_freq_write_amd(struct acpi_pct_register *not_used, u32 val)
 285 {
 286         wrmsr(MSR_AMD_PERF_CTL, val, 0);
 287 }
 288
 289 static u32 cpu_freq_read_io(struct acpi_pct_register *reg)
 290 {
 291         u32 val;
 292
 293         acpi_os_read_port(reg->address, &val, reg->bit_width);
 294         return val;
 295 }
 296
 297 static void cpu_freq_write_io(struct acpi_pct_register *reg, u32 val)
 298 {
 299         acpi_os_write_port(reg->address, val, reg->bit_width);
 300 }
 301
 302 struct drv_cmd {
 303         struct acpi_pct_register *reg;
 304         u32 val;
 305         union {
 306                 void (*write)(struct acpi_pct_register *reg, u32 val);
 307                 u32 (*read)(struct acpi_pct_register *reg);
 308         } func;
 309 };
 310
 311 /* Called via smp_call_function_single(), on the target CPU */
 312 static void do_drv_read(void *_cmd)
 313 {
 314         struct drv_cmd *cmd = _cmd;
 315
 316         cmd->val = cmd->func.read(cmd->reg);
 317 }
 318
 319 static u32 drv_read(struct acpi_cpufreq_data *data, const struct cpumask *mask)
 320 {
 321         struct acpi_processor_performance *perf = to_perf_data(data);
 322         struct drv_cmd cmd = {
 323                 .reg = &perf->control_register,
 324                 .func.read = data->cpu_freq_read,
 325         };
 326         int err;
 327
 328         err = smp_call_function_any(mask, do_drv_read, &cmd, 1);
 329         WARN_ON_ONCE(err);      /* smp_call_function_any() was buggy? */
 330         return cmd.val;
 331 }
 332
 333 /* Called via smp_call_function_many(), on the target CPUs */
 334 static void do_drv_write(void *_cmd)
 335 {
 336         struct drv_cmd *cmd = _cmd;
 337
 338         cmd->func.write(cmd->reg, cmd->val);
 339 }
 340
 341 static void drv_write(struct acpi_cpufreq_data *data,
 342                       const struct cpumask *mask, u32 val)
 343 {
 344         struct acpi_processor_performance *perf = to_perf_data(data);
 345         struct drv_cmd cmd = {
 346                 .reg = &perf->control_register,
 347                 .val = val,
 348                 .func.write = data->cpu_freq_write,
 349         };
 350         int this_cpu;
 351
 352         this_cpu = get_cpu();
 353         if (cpumask_test_cpu(this_cpu, mask))
 354                 do_drv_write(&cmd);
 355
 356         smp_call_function_many(mask, do_drv_write, &cmd, 1);
 357         put_cpu();
 358 }
 359
 360 static u32 get_cur_val(const struct cpumask *mask, struct acpi_cpufreq_data *data)
 361 {
 362         u32 val;
 363
 364         if (unlikely(cpumask_empty(mask)))
 365                 return 0;
 366
 367         val = drv_read(data, mask);
 368
 369         pr_debug("%s = %u\n", __func__, val);
 370
 371         return val;
 372 }
 373
 374 static unsigned int get_cur_freq_on_cpu(unsigned int cpu)
 375 {
 376         struct acpi_cpufreq_data *data;
 377         struct cpufreq_policy *policy;
 378         unsigned int freq;
 379         unsigned int cached_freq;
 380
 381         pr_debug("%s (%d)\n", __func__, cpu);
 382
 383         policy = cpufreq_cpu_get_raw(cpu);
 384         if (unlikely(!policy))
 385                 return 0;
 386
 387         data = policy->driver_data;
 388         if (unlikely(!data || !policy->freq_table))
 389                 return 0;
 390
 391         cached_freq = policy->freq_table[to_perf_data(data)->state].frequency;
 392         freq = extract_freq(policy, get_cur_val(cpumask_of(cpu), data));
 393         if (freq != cached_freq) {
 394                 /*
 395                  * The dreaded BIOS frequency change behind our back.
 396                  * Force set the frequency on next target call.
 397                  */
 398                 data->resume = 1;
 399         }
 400
 401         pr_debug("cur freq = %u\n", freq);
 402
 403         return freq;
 404 }
 405
 406 static unsigned int check_freqs(struct cpufreq_policy *policy,
 407                                 const struct cpumask *mask, unsigned int freq)
 408 {
 409         struct acpi_cpufreq_data *data = policy->driver_data;
 410         unsigned int cur_freq;
 411         unsigned int i;
 412
 413         for (i = 0; i < 100; i++) {
 414                 cur_freq = extract_freq(policy, get_cur_val(mask, data));
 415                 if (cur_freq == freq)
 416                         return 1;
 417                 udelay(10);
 418         }
 419         return 0;
 420 }
 421
 422 static int acpi_cpufreq_target(struct cpufreq_policy *policy,
 423                                unsigned int index)
 424 {
 425         struct acpi_cpufreq_data *data = policy->driver_data;
 426         struct acpi_processor_performance *perf;
 427         const struct cpumask *mask;
 428         unsigned int next_perf_state = 0; /* Index into perf table */
 429         int result = 0;
 430
 431         if (unlikely(!data)) {
 432                 return -ENODEV;
 433         }
 434
 435         perf = to_perf_data(data);
 436         next_perf_state = policy->freq_table[index].driver_data;
 437         if (perf->state == next_perf_state) {
 438                 if (unlikely(data->resume)) {
 439                         pr_debug("Called after resume, resetting to P%d\n",
 440                                 next_perf_state);
 441                         data->resume = 0;
 442                 } else {
 443                         pr_debug("Already at target state (P%d)\n",
 444                                 next_perf_state);
 445                         return 0;
 446                 }
 447         }
 448
 449         /*
 450          * The core won't allow CPUs to go away until the governor has been
 451          * stopped, so we can rely on the stability of policy->cpus.
 452          */
 453         mask = policy->shared_type == CPUFREQ_SHARED_TYPE_ANY ?
 454                 cpumask_of(policy->cpu) : policy->cpus;
 455
 456         drv_write(data, mask, perf->states[next_perf_state].control);
 457
 458         if (acpi_pstate_strict) {
 459                 if (!check_freqs(policy, mask,
 460                                  policy->freq_table[index].frequency)) {
 461                         pr_debug("%s (%d)\n", __func__, policy->cpu);
 462                         result = -EAGAIN;
 463                 }
 464         }
 465
 466         if (!result)
 467                 perf->state = next_perf_state;
 468
 469         return result;
 470 }
 471
 472 static unsigned int acpi_cpufreq_fast_switch(struct cpufreq_policy *policy,
 473                                              unsigned int target_freq)
 474 {
 475         struct acpi_cpufreq_data *data = policy->driver_data;
 476         struct acpi_processor_performance *perf;
 477         struct cpufreq_frequency_table *entry;
 478         unsigned int next_perf_state, next_freq, index;
 479
 480         /*
 481          * Find the closest frequency above target_freq.
 482          */
 483         if (policy->cached_target_freq == target_freq)
 484                 index = policy->cached_resolved_idx;
 485         else
 486                 index = cpufreq_table_find_index_dl(policy, target_freq);
 487
 488         entry = &policy->freq_table[index];
 489         next_freq = entry->frequency;
 490         next_perf_state = entry->driver_data;
 491
 492         perf = to_perf_data(data);
 493         if (perf->state == next_perf_state) {
 494                 if (unlikely(data->resume))
 495                         data->resume = 0;
 496                 else
 497                         return next_freq;
 498         }
 499
 500         data->cpu_freq_write(&perf->control_register,
 501                              perf->states[next_perf_state].control);
 502         perf->state = next_perf_state;
 503         return next_freq;
 504 }
 505
 506 static unsigned long
 507 acpi_cpufreq_guess_freq(struct acpi_cpufreq_data *data, unsigned int cpu)
 508 {
 509         struct acpi_processor_performance *perf;
 510
 511         perf = to_perf_data(data);
 512         if (cpu_khz) {
 513                 /* search the closest match to cpu_khz */
 514                 unsigned int i;
 515                 unsigned long freq;
 516                 unsigned long freqn = perf->states[0].core_frequency * 1000;
 517
 518                 for (i = 0; i < (perf->state_count-1); i++) {
 519                         freq = freqn;
 520                         freqn = perf->states[i+1].core_frequency * 1000;
 521                         if ((2 * cpu_khz) > (freqn + freq)) {
 522                                 perf->state = i;
 523                                 return freq;
 524                         }
 525                 }
 526                 perf->state = perf->state_count-1;
 527                 return freqn;
 528         } else {
 529                 /* assume CPU is at P0... */
 530                 perf->state = 0;
 531                 return perf->states[0].core_frequency * 1000;
 532         }
 533 }
 534
 535 static void free_acpi_perf_data(void)
 536 {
 537         unsigned int i;
 538
 539         /* Freeing a NULL pointer is OK, and alloc_percpu zeroes. */
 540         for_each_possible_cpu(i)
 541                 free_cpumask_var(per_cpu_ptr(acpi_perf_data, i)
 542                                  ->shared_cpu_map);
 543         free_percpu(acpi_perf_data);
 544 }
 545
 546 static int cpufreq_boost_online(unsigned int cpu)
 547 {
 548         /*
 549          * On the CPU_UP path we simply keep the boost-disable flag
 550          * in sync with the current global state.
 551          */
 552         return boost_set_msr(acpi_cpufreq_driver.boost_enabled);
 553 }
 554
 555 static int cpufreq_boost_down_prep(unsigned int cpu)
 556 {
 557         /*
 558          * Clear the boost-disable bit on the CPU_DOWN path so that
 559          * this cpu cannot block the remaining ones from boosting.
 560          */
 561         return boost_set_msr(1);
 562 }
 563
 564 /*
 565  * acpi_cpufreq_early_init - initialize ACPI P-States library
 566  *
 567  * Initialize the ACPI P-States library (drivers/acpi/processor_perflib.c)
 568  * in order to determine correct frequency and voltage pairings. We can
 569  * do _PDC and _PSD and find out the processor dependency for the
 570  * actual init that will happen later...
 571  */
 572 static int __init acpi_cpufreq_early_init(void)
 573 {
 574         unsigned int i;
 575         pr_debug("%s\n", __func__);
 576
 577         acpi_perf_data = alloc_percpu(struct acpi_processor_performance);
 578         if (!acpi_perf_data) {
 579                 pr_debug("Memory allocation error for acpi_perf_data.\n");
 580                 return -ENOMEM;
 581         }
 582         for_each_possible_cpu(i) {
 583                 if (!zalloc_cpumask_var_node(
 584                         &per_cpu_ptr(acpi_perf_data, i)->shared_cpu_map,
 585                         GFP_KERNEL, cpu_to_node(i))) {
 586
 587                         /* Freeing a NULL pointer is OK: alloc_percpu zeroes. */
 588                         free_acpi_perf_data();
 589                         return -ENOMEM;
 590                 }
 591         }
 592
 593         /* Do initialization in ACPI core */
 594         acpi_processor_preregister_performance(acpi_perf_data);
 595         return 0;
 596 }
 597
 598 #ifdef CONFIG_SMP
 599 /*
 600  * Some BIOSes do SW_ANY coordination internally, either set it up in hw
 601  * or do it in BIOS firmware and won't inform about it to OS. If not
 602  * detected, this has a side effect of making CPU run at a different speed
 603  * than OS intended it to run at. Detect it and handle it cleanly.
 604  */
 605 static int bios_with_sw_any_bug;
 606
 607 static int sw_any_bug_found(const struct dmi_system_id *d)
 608 {
 609         bios_with_sw_any_bug = 1;
 610         return 0;
 611 }
 612
 613 static const struct dmi_system_id sw_any_bug_dmi_table[] = {
 614         {
 615                 .callback = sw_any_bug_found,
 616                 .ident = "Supermicro Server X6DLP",
 617                 .matches = {
 618                         DMI_MATCH(DMI_SYS_VENDOR, "Supermicro"),
 619                         DMI_MATCH(DMI_BIOS_VERSION, "080010"),
 620                         DMI_MATCH(DMI_PRODUCT_NAME, "X6DLP"),
 621                 },
 622         },
 623         { }
 624 };
 625
 626 static int acpi_cpufreq_blacklist(struct cpuinfo_x86 *c)
 627 {
 628         /* Intel Xeon Processor 7100 Series Specification Update
 629          * http://www.intel.com/Assets/PDF/specupdate/314554.pdf
 630          * AL30: A Machine Check Exception (MCE) Occurring during an
 631          * Enhanced Intel SpeedStep Technology Ratio Change May Cause
 632          * Both Processor Cores to Lock Up. */
 633         if (c->x86_vendor == X86_VENDOR_INTEL) {
 634                 if ((c->x86 == 15) &&
 635                     (c->x86_model == 6) &&
 636                     (c->x86_stepping == 8)) {
 637                         pr_info("Intel(R) Xeon(R) 7100 Errata AL30, processors may lock up on frequency changes: disabling acpi-cpufreq\n");
 638                         return -ENODEV;
 639                     }
 640                 }
 641         return 0;
 642 }
 643 #endif
 644
 645 static int acpi_cpufreq_cpu_init(struct cpufreq_policy *policy)
 646 {
 647         unsigned int i;
 648         unsigned int valid_states = 0;
 649         unsigned int cpu = policy->cpu;
 650         struct acpi_cpufreq_data *data;
 651         unsigned int result = 0;
 652         struct cpuinfo_x86 *c = &cpu_data(policy->cpu);
 653         struct acpi_processor_performance *perf;
 654         struct cpufreq_frequency_table *freq_table;
 655 #ifdef CONFIG_SMP
 656         static int blacklisted;
 657 #endif
 658
 659         pr_debug("%s\n", __func__);
 660
 661 #ifdef CONFIG_SMP
 662         if (blacklisted)
 663                 return blacklisted;
 664         blacklisted = acpi_cpufreq_blacklist(c);
 665         if (blacklisted)
 666                 return blacklisted;
 667 #endif
 668
 669         data = kzalloc(sizeof(*data), GFP_KERNEL);
 670         if (!data)
 671                 return -ENOMEM;
 672
 673         if (!zalloc_cpumask_var(&data->freqdomain_cpus, GFP_KERNEL)) {
 674                 result = -ENOMEM;
 675                 goto err_free;
 676         }
 677
 678         perf = per_cpu_ptr(acpi_perf_data, cpu);
 679         data->acpi_perf_cpu = cpu;
 680         policy->driver_data = data;
 681
 682         if (cpu_has(c, X86_FEATURE_CONSTANT_TSC))
 683                 acpi_cpufreq_driver.flags |= CPUFREQ_CONST_LOOPS;
 684
 685         result = acpi_processor_register_performance(perf, cpu);
 686         if (result)
 687                 goto err_free_mask;
 688
 689         policy->shared_type = perf->shared_type;
 690
 691         /*
 692          * Will let policy->cpus know about dependency only when software
 693          * coordination is required.
 694          */
 695         if (policy->shared_type == CPUFREQ_SHARED_TYPE_ALL ||
 696             policy->shared_type == CPUFREQ_SHARED_TYPE_ANY) {
 697                 cpumask_copy(policy->cpus, perf->shared_cpu_map);
 698         }
 699         cpumask_copy(data->freqdomain_cpus, perf->shared_cpu_map);
 700
 701 #ifdef CONFIG_SMP
 702         dmi_check_system(sw_any_bug_dmi_table);
 703         if (bios_with_sw_any_bug && !policy_is_shared(policy)) {
 704                 policy->shared_type = CPUFREQ_SHARED_TYPE_ALL;
 705                 cpumask_copy(policy->cpus, topology_core_cpumask(cpu));
 706         }
 707
 708         if (check_amd_hwpstate_cpu(cpu) && !acpi_pstate_strict) {
 709                 cpumask_clear(policy->cpus);
 710                 cpumask_set_cpu(cpu, policy->cpus);
 711                 cpumask_copy(data->freqdomain_cpus,
 712                              topology_sibling_cpumask(cpu));
 713                 policy->shared_type = CPUFREQ_SHARED_TYPE_HW;
 714                 pr_info_once("overriding BIOS provided _PSD data\n");
 715         }
 716 #endif
 717
 718         /* capability check */
 719         if (perf->state_count <= 1) {
 720                 pr_debug("No P-States\n");
 721                 result = -ENODEV;
 722                 goto err_unreg;
 723         }
 724
 725         if (perf->control_register.space_id != perf->status_register.space_id) {
 726                 result = -ENODEV;
 727                 goto err_unreg;
 728         }
 729
 730         switch (perf->control_register.space_id) {
 731         case ACPI_ADR_SPACE_SYSTEM_IO:
 732                 if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD &&
 733                     boot_cpu_data.x86 == 0xf) {
 734                         pr_debug("AMD K8 systems must use native drivers.\n");
 735                         result = -ENODEV;
 736                         goto err_unreg;
 737                 }
 738                 pr_debug("SYSTEM IO addr space\n");
 739                 data->cpu_feature = SYSTEM_IO_CAPABLE;
 740                 data->cpu_freq_read = cpu_freq_read_io;
 741                 data->cpu_freq_write = cpu_freq_write_io;
 742                 break;
 743         case ACPI_ADR_SPACE_FIXED_HARDWARE:
 744                 pr_debug("HARDWARE addr space\n");
 745                 if (check_est_cpu(cpu)) {
 746                         data->cpu_feature = SYSTEM_INTEL_MSR_CAPABLE;
 747                         data->cpu_freq_read = cpu_freq_read_intel;
 748                         data->cpu_freq_write = cpu_freq_write_intel;
 749                         break;
 750                 }
 751                 if (check_amd_hwpstate_cpu(cpu)) {
 752                         data->cpu_feature = SYSTEM_AMD_MSR_CAPABLE;
 753                         data->cpu_freq_read = cpu_freq_read_amd;
 754                         data->cpu_freq_write = cpu_freq_write_amd;
 755                         break;
 756                 }
 757                 result = -ENODEV;
 758                 goto err_unreg;
 759         default:
 760                 pr_debug("Unknown addr space %d\n",
 761                         (u32) (perf->control_register.space_id));
 762                 result = -ENODEV;
 763                 goto err_unreg;
 764         }
 765
 766         freq_table = kcalloc(perf->state_count + 1, sizeof(*freq_table),
 767                              GFP_KERNEL);
 768         if (!freq_table) {
 769                 result = -ENOMEM;
 770                 goto err_unreg;
 771         }
 772
 773         /* detect transition latency */
 774         policy->cpuinfo.transition_latency = 0;
 775         for (i = 0; i < perf->state_count; i++) {
 776                 if ((perf->states[i].transition_latency * 1000) >
 777                     policy->cpuinfo.transition_latency)
 778                         policy->cpuinfo.transition_latency =
 779                             perf->states[i].transition_latency * 1000;
 780         }
 781
 782         /* Check for high latency (>20uS) from buggy BIOSes, like on T42 */
 783         if (perf->control_register.space_id == ACPI_ADR_SPACE_FIXED_HARDWARE &&
 784             policy->cpuinfo.transition_latency > 20 * 1000) {
 785                 policy->cpuinfo.transition_latency = 20 * 1000;
 786                 pr_info_once("P-state transition latency capped at 20 uS\n");
 787         }
 788
 789         /* table init */
 790         for (i = 0; i < perf->state_count; i++) {
 791                 if (i > 0 && perf->states[i].core_frequency >=
 792                     freq_table[valid_states-1].frequency / 1000)
 793                         continue;
 794
 795                 freq_table[valid_states].driver_data = i;
 796                 freq_table[valid_states].frequency =
 797                     perf->states[i].core_frequency * 1000;
 798                 valid_states++;
 799         }
 800         freq_table[valid_states].frequency = CPUFREQ_TABLE_END;
 801         policy->freq_table = freq_table;
 802         perf->state = 0;
 803
 804         switch (perf->control_register.space_id) {
 805         case ACPI_ADR_SPACE_SYSTEM_IO:
 806                 /*
 807                  * The core will not set policy->cur, because
 808                  * cpufreq_driver->get is NULL, so we need to set it here.
 809                  * However, we have to guess it, because the current speed is
 810                  * unknown and not detectable via IO ports.
 811                  */
 812                 policy->cur = acpi_cpufreq_guess_freq(data, policy->cpu);
 813                 break;
 814         case ACPI_ADR_SPACE_FIXED_HARDWARE:
 815                 acpi_cpufreq_driver.get = get_cur_freq_on_cpu;
 816                 break;
 817         default:
 818                 break;
 819         }
 820
 821         /* notify BIOS that we exist */
 822         acpi_processor_notify_smm(THIS_MODULE);
 823
 824         pr_debug("CPU%u - ACPI performance management activated.\n", cpu);
 825         for (i = 0; i < perf->state_count; i++)
 826                 pr_debug("     %cP%d: %d MHz, %d mW, %d uS\n",
 827                         (i == perf->state ? '*' : ' '), i,
 828                         (u32) perf->states[i].core_frequency,
 829                         (u32) perf->states[i].power,
 830                         (u32) perf->states[i].transition_latency);
 831
 832         /*
 833          * the first call to ->target() should result in us actually
 834          * writing something to the appropriate registers.
 835          */
 836         data->resume = 1;
 837
 838         policy->fast_switch_possible = !acpi_pstate_strict &&
 839                 !(policy_is_shared(policy) && policy->shared_type != CPUFREQ_SHARED_TYPE_ANY);
 840
 841         return result;
 842
 843 err_unreg:
 844         acpi_processor_unregister_performance(cpu);
 845 err_free_mask:
 846         free_cpumask_var(data->freqdomain_cpus);
 847 err_free:
 848         kfree(data);
 849         policy->driver_data = NULL;
 850
 851         return result;
 852 }
 853
 854 static int acpi_cpufreq_cpu_exit(struct cpufreq_policy *policy)
 855 {
 856         struct acpi_cpufreq_data *data = policy->driver_data;
 857
 858         pr_debug("%s\n", __func__);
 859
 860         policy->fast_switch_possible = false;
 861         policy->driver_data = NULL;
 862         acpi_processor_unregister_performance(data->acpi_perf_cpu);
 863         free_cpumask_var(data->freqdomain_cpus);
 864         kfree(policy->freq_table);
 865         kfree(data);
 866
 867         return 0;
 868 }
 869
 870 static void acpi_cpufreq_cpu_ready(struct cpufreq_policy *policy)
 871 {
 872         struct acpi_processor_performance *perf = per_cpu_ptr(acpi_perf_data,
 873                                                               policy->cpu);
 874
 875         if (perf->states[0].core_frequency * 1000 != policy->cpuinfo.max_freq)
 876                 pr_warn(FW_WARN "P-state 0 is not max freq\n");
 877 }
 878
 879 static int acpi_cpufreq_resume(struct cpufreq_policy *policy)
 880 {
 881         struct acpi_cpufreq_data *data = policy->driver_data;
 882
 883         pr_debug("%s\n", __func__);
 884
 885         data->resume = 1;
 886
 887         return 0;
 888 }
 889
 890 static struct freq_attr *acpi_cpufreq_attr[] = {
 891         &cpufreq_freq_attr_scaling_available_freqs,
 892         &freqdomain_cpus,
 893 #ifdef CONFIG_X86_ACPI_CPUFREQ_CPB
 894         &cpb,
 895 #endif
 896         NULL,
 897 };
 898
 899 static struct cpufreq_driver acpi_cpufreq_driver = {
 900         .verify         = cpufreq_generic_frequency_table_verify,
 901         .target_index   = acpi_cpufreq_target,
 902         .fast_switch    = acpi_cpufreq_fast_switch,
 903         .bios_limit     = acpi_processor_get_bios_limit,
 904         .init           = acpi_cpufreq_cpu_init,
 905         .exit           = acpi_cpufreq_cpu_exit,
 906         .ready          = acpi_cpufreq_cpu_ready,
 907         .resume         = acpi_cpufreq_resume,
 908         .name           = "acpi-cpufreq",
 909         .attr           = acpi_cpufreq_attr,
 910 };
 911
 912 static enum cpuhp_state acpi_cpufreq_online;
 913
 914 static void __init acpi_cpufreq_boost_init(void)
 915 {
 916         int ret;
 917
 918         if (!(boot_cpu_has(X86_FEATURE_CPB) || boot_cpu_has(X86_FEATURE_IDA))) {
 919                 pr_debug("Boost capabilities not present in the processor\n");
 920                 return;
 921         }
 922
 923         acpi_cpufreq_driver.set_boost = set_boost;
 924         acpi_cpufreq_driver.boost_enabled = boost_state(0);
 925
 926         /*
 927          * This calls the online callback on all online cpu and forces all
 928          * MSRs to the same value.
 929          */
 930         ret = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "cpufreq/acpi:online",
 931                                 cpufreq_boost_online, cpufreq_boost_down_prep);
 932         if (ret < 0) {
 933                 pr_err("acpi_cpufreq: failed to register hotplug callbacks\n");
 934                 return;
 935         }
 936         acpi_cpufreq_online = ret;
 937 }
 938
 939 static void acpi_cpufreq_boost_exit(void)
 940 {
 941         if (acpi_cpufreq_online > 0)
 942                 cpuhp_remove_state_nocalls(acpi_cpufreq_online);
 943 }
 944
 945 static int __init acpi_cpufreq_init(void)
 946 {
 947         int ret;
 948
 949         if (acpi_disabled)
 950                 return -ENODEV;
 951
 952         /* don't keep reloading if cpufreq_driver exists */
 953         if (cpufreq_get_current_driver())
 954                 return -EEXIST;
 955
 956         pr_debug("%s\n", __func__);
 957
 958         ret = acpi_cpufreq_early_init();
 959         if (ret)
 960                 return ret;
 961
 962 #ifdef CONFIG_X86_ACPI_CPUFREQ_CPB
 963         /* this is a sysfs file with a strange name and an even stranger
 964          * semantic - per CPU instantiation, but system global effect.
 965          * Lets enable it only on AMD CPUs for compatibility reasons and
 966          * only if configured. This is considered legacy code, which
 967          * will probably be removed at some point in the future.
 968          */
 969         if (!check_amd_hwpstate_cpu(0)) {
 970                 struct freq_attr **attr;
 971
 972                 pr_debug("CPB unsupported, do not expose it\n");
 973
 974                 for (attr = acpi_cpufreq_attr; *attr; attr++)
 975                         if (*attr == &cpb) {
 976                                 *attr = NULL;
 977                                 break;
 978                         }
 979         }
 980 #endif
 981         acpi_cpufreq_boost_init();
 982
 983         ret = cpufreq_register_driver(&acpi_cpufreq_driver);
 984         if (ret) {
 985                 free_acpi_perf_data();
 986                 acpi_cpufreq_boost_exit();
 987         }
 988         return ret;
 989 }
 990
 991 static void __exit acpi_cpufreq_exit(void)
 992 {
 993         pr_debug("%s\n", __func__);
 994
 995         acpi_cpufreq_boost_exit();
 996
 997         cpufreq_unregister_driver(&acpi_cpufreq_driver);
 998
 999         free_acpi_perf_data();
1000 }
1001
1002 module_param(acpi_pstate_strict, uint, 0644);
1003 MODULE_PARM_DESC(acpi_pstate_strict,
1004         "value 0 or non-zero. non-zero -> strict ACPI checks are "
1005         "performed during frequency changes.");
1006
1007 late_initcall(acpi_cpufreq_init);
1008 module_exit(acpi_cpufreq_exit);
1009
1010 static const struct x86_cpu_id acpi_cpufreq_ids[] = {
1011         X86_FEATURE_MATCH(X86_FEATURE_ACPI),
1012         X86_FEATURE_MATCH(X86_FEATURE_HW_PSTATE),
1013         {}
1014 };
1015 MODULE_DEVICE_TABLE(x86cpu, acpi_cpufreq_ids);
1016
1017 static const struct acpi_device_id processor_device_ids[] = {
1018         {ACPI_PROCESSOR_OBJECT_HID, },
1019         {ACPI_PROCESSOR_DEVICE_HID, },
1020         {},
1021 };
1022 MODULE_DEVICE_TABLE(acpi, processor_device_ids);
1023
1024 MODULE_ALIAS("acpi");