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
  65 #define MSR_K7_HWCR_CPB_DIS     (1ULL << 25)
  66
  67 struct acpi_cpufreq_data {
  68         unsigned int resume;
  69         unsigned int cpu_feature;
  70         unsigned int acpi_perf_cpu;
  71         cpumask_var_t freqdomain_cpus;
  72         void (*cpu_freq_write)(struct acpi_pct_register *reg, u32 val);
  73         u32 (*cpu_freq_read)(struct acpi_pct_register *reg);
  74 };
  75
  76 /* acpi_perf_data is a pointer to percpu data. */
  77 static struct acpi_processor_performance __percpu *acpi_perf_data;
  78
  79 static inline struct acpi_processor_performance *to_perf_data(struct acpi_cpufreq_data *data)
  80 {
  81         return per_cpu_ptr(acpi_perf_data, data->acpi_perf_cpu);
  82 }
  83
  84 static struct cpufreq_driver acpi_cpufreq_driver;
  85
  86 static unsigned int acpi_pstate_strict;
  87 static struct msr __percpu *msrs;
  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_AMD:
 100                 rdmsr_on_cpu(cpu, MSR_K7_HWCR, &lo, &hi);
 101                 msr = lo | ((u64)hi << 32);
 102                 return !(msr & MSR_K7_HWCR_CPB_DIS);
 103         }
 104         return false;
 105 }
 106
 107 static void boost_set_msrs(bool enable, const struct cpumask *cpumask)
 108 {
 109         u32 cpu;
 110         u32 msr_addr;
 111         u64 msr_mask;
 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_AMD:
 119                 msr_addr = MSR_K7_HWCR;
 120                 msr_mask = MSR_K7_HWCR_CPB_DIS;
 121                 break;
 122         default:
 123                 return;
 124         }
 125
 126         rdmsr_on_cpus(cpumask, msr_addr, msrs);
 127
 128         for_each_cpu(cpu, cpumask) {
 129                 struct msr *reg = per_cpu_ptr(msrs, cpu);
 130                 if (enable)
 131                         reg->q &= ~msr_mask;
 132                 else
 133                         reg->q |= msr_mask;
 134         }
 135
 136         wrmsr_on_cpus(cpumask, msr_addr, msrs);
 137 }
 138
 139 static int set_boost(int val)
 140 {
 141         get_online_cpus();
 142         boost_set_msrs(val, cpu_online_mask);
 143         put_online_cpus();
 144         pr_debug("Core Boosting %sabled.\n", val ? "en" : "dis");
 145
 146         return 0;
 147 }
 148
 149 static ssize_t show_freqdomain_cpus(struct cpufreq_policy *policy, char *buf)
 150 {
 151         struct acpi_cpufreq_data *data = policy->driver_data;
 152
 153         if (unlikely(!data))
 154                 return -ENODEV;
 155
 156         return cpufreq_show_cpus(data->freqdomain_cpus, buf);
 157 }
 158
 159 cpufreq_freq_attr_ro(freqdomain_cpus);
 160
 161 #ifdef CONFIG_X86_ACPI_CPUFREQ_CPB
 162 static ssize_t store_cpb(struct cpufreq_policy *policy, const char *buf,
 163                          size_t count)
 164 {
 165         int ret;
 166         unsigned int val = 0;
 167
 168         if (!acpi_cpufreq_driver.set_boost)
 169                 return -EINVAL;
 170
 171         ret = kstrtouint(buf, 10, &val);
 172         if (ret || val > 1)
 173                 return -EINVAL;
 174
 175         set_boost(val);
 176
 177         return count;
 178 }
 179
 180 static ssize_t show_cpb(struct cpufreq_policy *policy, char *buf)
 181 {
 182         return sprintf(buf, "%u\n", acpi_cpufreq_driver.boost_enabled);
 183 }
 184
 185 cpufreq_freq_attr_rw(cpb);
 186 #endif
 187
 188 static int check_est_cpu(unsigned int cpuid)
 189 {
 190         struct cpuinfo_x86 *cpu = &cpu_data(cpuid);
 191
 192         return cpu_has(cpu, X86_FEATURE_EST);
 193 }
 194
 195 static int check_amd_hwpstate_cpu(unsigned int cpuid)
 196 {
 197         struct cpuinfo_x86 *cpu = &cpu_data(cpuid);
 198
 199         return cpu_has(cpu, X86_FEATURE_HW_PSTATE);
 200 }
 201
 202 static unsigned extract_io(struct cpufreq_policy *policy, u32 value)
 203 {
 204         struct acpi_cpufreq_data *data = policy->driver_data;
 205         struct acpi_processor_performance *perf;
 206         int i;
 207
 208         perf = to_perf_data(data);
 209
 210         for (i = 0; i < perf->state_count; i++) {
 211                 if (value == perf->states[i].status)
 212                         return policy->freq_table[i].frequency;
 213         }
 214         return 0;
 215 }
 216
 217 static unsigned extract_msr(struct cpufreq_policy *policy, u32 msr)
 218 {
 219         struct acpi_cpufreq_data *data = policy->driver_data;
 220         struct cpufreq_frequency_table *pos;
 221         struct acpi_processor_performance *perf;
 222
 223         if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD)
 224                 msr &= AMD_MSR_RANGE;
 225         else
 226                 msr &= INTEL_MSR_RANGE;
 227
 228         perf = to_perf_data(data);
 229
 230         cpufreq_for_each_entry(pos, policy->freq_table)
 231                 if (msr == perf->states[pos->driver_data].status)
 232                         return pos->frequency;
 233         return policy->freq_table[0].frequency;
 234 }
 235
 236 static unsigned extract_freq(struct cpufreq_policy *policy, u32 val)
 237 {
 238         struct acpi_cpufreq_data *data = policy->driver_data;
 239
 240         switch (data->cpu_feature) {
 241         case SYSTEM_INTEL_MSR_CAPABLE:
 242         case SYSTEM_AMD_MSR_CAPABLE:
 243                 return extract_msr(policy, val);
 244         case SYSTEM_IO_CAPABLE:
 245                 return extract_io(policy, val);
 246         default:
 247                 return 0;
 248         }
 249 }
 250
 251 static u32 cpu_freq_read_intel(struct acpi_pct_register *not_used)
 252 {
 253         u32 val, dummy;
 254
 255         rdmsr(MSR_IA32_PERF_CTL, val, dummy);
 256         return val;
 257 }
 258
 259 static void cpu_freq_write_intel(struct acpi_pct_register *not_used, u32 val)
 260 {
 261         u32 lo, hi;
 262
 263         rdmsr(MSR_IA32_PERF_CTL, lo, hi);
 264         lo = (lo & ~INTEL_MSR_RANGE) | (val & INTEL_MSR_RANGE);
 265         wrmsr(MSR_IA32_PERF_CTL, lo, hi);
 266 }
 267
 268 static u32 cpu_freq_read_amd(struct acpi_pct_register *not_used)
 269 {
 270         u32 val, dummy;
 271
 272         rdmsr(MSR_AMD_PERF_CTL, val, dummy);
 273         return val;
 274 }
 275
 276 static void cpu_freq_write_amd(struct acpi_pct_register *not_used, u32 val)
 277 {
 278         wrmsr(MSR_AMD_PERF_CTL, val, 0);
 279 }
 280
 281 static u32 cpu_freq_read_io(struct acpi_pct_register *reg)
 282 {
 283         u32 val;
 284
 285         acpi_os_read_port(reg->address, &val, reg->bit_width);
 286         return val;
 287 }
 288
 289 static void cpu_freq_write_io(struct acpi_pct_register *reg, u32 val)
 290 {
 291         acpi_os_write_port(reg->address, val, reg->bit_width);
 292 }
 293
 294 struct drv_cmd {
 295         struct acpi_pct_register *reg;
 296         u32 val;
 297         union {
 298                 void (*write)(struct acpi_pct_register *reg, u32 val);
 299                 u32 (*read)(struct acpi_pct_register *reg);
 300         } func;
 301 };
 302
 303 /* Called via smp_call_function_single(), on the target CPU */
 304 static void do_drv_read(void *_cmd)
 305 {
 306         struct drv_cmd *cmd = _cmd;
 307
 308         cmd->val = cmd->func.read(cmd->reg);
 309 }
 310
 311 static u32 drv_read(struct acpi_cpufreq_data *data, const struct cpumask *mask)
 312 {
 313         struct acpi_processor_performance *perf = to_perf_data(data);
 314         struct drv_cmd cmd = {
 315                 .reg = &perf->control_register,
 316                 .func.read = data->cpu_freq_read,
 317         };
 318         int err;
 319
 320         err = smp_call_function_any(mask, do_drv_read, &cmd, 1);
 321         WARN_ON_ONCE(err);      /* smp_call_function_any() was buggy? */
 322         return cmd.val;
 323 }
 324
 325 /* Called via smp_call_function_many(), on the target CPUs */
 326 static void do_drv_write(void *_cmd)
 327 {
 328         struct drv_cmd *cmd = _cmd;
 329
 330         cmd->func.write(cmd->reg, cmd->val);
 331 }
 332
 333 static void drv_write(struct acpi_cpufreq_data *data,
 334                       const struct cpumask *mask, u32 val)
 335 {
 336         struct acpi_processor_performance *perf = to_perf_data(data);
 337         struct drv_cmd cmd = {
 338                 .reg = &perf->control_register,
 339                 .val = val,
 340                 .func.write = data->cpu_freq_write,
 341         };
 342         int this_cpu;
 343
 344         this_cpu = get_cpu();
 345         if (cpumask_test_cpu(this_cpu, mask))
 346                 do_drv_write(&cmd);
 347
 348         smp_call_function_many(mask, do_drv_write, &cmd, 1);
 349         put_cpu();
 350 }
 351
 352 static u32 get_cur_val(const struct cpumask *mask, struct acpi_cpufreq_data *data)
 353 {
 354         u32 val;
 355
 356         if (unlikely(cpumask_empty(mask)))
 357                 return 0;
 358
 359         val = drv_read(data, mask);
 360
 361         pr_debug("get_cur_val = %u\n", val);
 362
 363         return val;
 364 }
 365
 366 static unsigned int get_cur_freq_on_cpu(unsigned int cpu)
 367 {
 368         struct acpi_cpufreq_data *data;
 369         struct cpufreq_policy *policy;
 370         unsigned int freq;
 371         unsigned int cached_freq;
 372
 373         pr_debug("get_cur_freq_on_cpu (%d)\n", cpu);
 374
 375         policy = cpufreq_cpu_get_raw(cpu);
 376         if (unlikely(!policy))
 377                 return 0;
 378
 379         data = policy->driver_data;
 380         if (unlikely(!data || !policy->freq_table))
 381                 return 0;
 382
 383         cached_freq = policy->freq_table[to_perf_data(data)->state].frequency;
 384         freq = extract_freq(policy, get_cur_val(cpumask_of(cpu), data));
 385         if (freq != cached_freq) {
 386                 /*
 387                  * The dreaded BIOS frequency change behind our back.
 388                  * Force set the frequency on next target call.
 389                  */
 390                 data->resume = 1;
 391         }
 392
 393         pr_debug("cur freq = %u\n", freq);
 394
 395         return freq;
 396 }
 397
 398 static unsigned int check_freqs(struct cpufreq_policy *policy,
 399                                 const struct cpumask *mask, unsigned int freq)
 400 {
 401         struct acpi_cpufreq_data *data = policy->driver_data;
 402         unsigned int cur_freq;
 403         unsigned int i;
 404
 405         for (i = 0; i < 100; i++) {
 406                 cur_freq = extract_freq(policy, get_cur_val(mask, data));
 407                 if (cur_freq == freq)
 408                         return 1;
 409                 udelay(10);
 410         }
 411         return 0;
 412 }
 413
 414 static int acpi_cpufreq_target(struct cpufreq_policy *policy,
 415                                unsigned int index)
 416 {
 417         struct acpi_cpufreq_data *data = policy->driver_data;
 418         struct acpi_processor_performance *perf;
 419         const struct cpumask *mask;
 420         unsigned int next_perf_state = 0; /* Index into perf table */
 421         int result = 0;
 422
 423         if (unlikely(!data)) {
 424                 return -ENODEV;
 425         }
 426
 427         perf = to_perf_data(data);
 428         next_perf_state = policy->freq_table[index].driver_data;
 429         if (perf->state == next_perf_state) {
 430                 if (unlikely(data->resume)) {
 431                         pr_debug("Called after resume, resetting to P%d\n",
 432                                 next_perf_state);
 433                         data->resume = 0;
 434                 } else {
 435                         pr_debug("Already at target state (P%d)\n",
 436                                 next_perf_state);
 437                         return 0;
 438                 }
 439         }
 440
 441         /*
 442          * The core won't allow CPUs to go away until the governor has been
 443          * stopped, so we can rely on the stability of policy->cpus.
 444          */
 445         mask = policy->shared_type == CPUFREQ_SHARED_TYPE_ANY ?
 446                 cpumask_of(policy->cpu) : policy->cpus;
 447
 448         drv_write(data, mask, perf->states[next_perf_state].control);
 449
 450         if (acpi_pstate_strict) {
 451                 if (!check_freqs(policy, mask,
 452                                  policy->freq_table[index].frequency)) {
 453                         pr_debug("acpi_cpufreq_target failed (%d)\n",
 454                                 policy->cpu);
 455                         result = -EAGAIN;
 456                 }
 457         }
 458
 459         if (!result)
 460                 perf->state = next_perf_state;
 461
 462         return result;
 463 }
 464
 465 unsigned int acpi_cpufreq_fast_switch(struct cpufreq_policy *policy,
 466                                       unsigned int target_freq)
 467 {
 468         struct acpi_cpufreq_data *data = policy->driver_data;
 469         struct acpi_processor_performance *perf;
 470         struct cpufreq_frequency_table *entry;
 471         unsigned int next_perf_state, next_freq, index;
 472
 473         /*
 474          * Find the closest frequency above target_freq.
 475          */
 476         if (policy->cached_target_freq == target_freq)
 477                 index = policy->cached_resolved_idx;
 478         else
 479                 index = cpufreq_table_find_index_dl(policy, target_freq);
 480
 481         entry = &policy->freq_table[index];
 482         next_freq = entry->frequency;
 483         next_perf_state = entry->driver_data;
 484
 485         perf = to_perf_data(data);
 486         if (perf->state == next_perf_state) {
 487                 if (unlikely(data->resume))
 488                         data->resume = 0;
 489                 else
 490                         return next_freq;
 491         }
 492
 493         data->cpu_freq_write(&perf->control_register,
 494                              perf->states[next_perf_state].control);
 495         perf->state = next_perf_state;
 496         return next_freq;
 497 }
 498
 499 static unsigned long
 500 acpi_cpufreq_guess_freq(struct acpi_cpufreq_data *data, unsigned int cpu)
 501 {
 502         struct acpi_processor_performance *perf;
 503
 504         perf = to_perf_data(data);
 505         if (cpu_khz) {
 506                 /* search the closest match to cpu_khz */
 507                 unsigned int i;
 508                 unsigned long freq;
 509                 unsigned long freqn = perf->states[0].core_frequency * 1000;
 510
 511                 for (i = 0; i < (perf->state_count-1); i++) {
 512                         freq = freqn;
 513                         freqn = perf->states[i+1].core_frequency * 1000;
 514                         if ((2 * cpu_khz) > (freqn + freq)) {
 515                                 perf->state = i;
 516                                 return freq;
 517                         }
 518                 }
 519                 perf->state = perf->state_count-1;
 520                 return freqn;
 521         } else {
 522                 /* assume CPU is at P0... */
 523                 perf->state = 0;
 524                 return perf->states[0].core_frequency * 1000;
 525         }
 526 }
 527
 528 static void free_acpi_perf_data(void)
 529 {
 530         unsigned int i;
 531
 532         /* Freeing a NULL pointer is OK, and alloc_percpu zeroes. */
 533         for_each_possible_cpu(i)
 534                 free_cpumask_var(per_cpu_ptr(acpi_perf_data, i)
 535                                  ->shared_cpu_map);
 536         free_percpu(acpi_perf_data);
 537 }
 538
 539 static int boost_notify(struct notifier_block *nb, unsigned long action,
 540                       void *hcpu)
 541 {
 542         unsigned cpu = (long)hcpu;
 543         const struct cpumask *cpumask;
 544
 545         cpumask = get_cpu_mask(cpu);
 546
 547         /*
 548          * Clear the boost-disable bit on the CPU_DOWN path so that
 549          * this cpu cannot block the remaining ones from boosting. On
 550          * the CPU_UP path we simply keep the boost-disable flag in
 551          * sync with the current global state.
 552          */
 553
 554         switch (action) {
 555         case CPU_DOWN_FAILED:
 556         case CPU_DOWN_FAILED_FROZEN:
 557         case CPU_ONLINE:
 558         case CPU_ONLINE_FROZEN:
 559                 boost_set_msrs(acpi_cpufreq_driver.boost_enabled, cpumask);
 560                 break;
 561
 562         case CPU_DOWN_PREPARE:
 563         case CPU_DOWN_PREPARE_FROZEN:
 564                 boost_set_msrs(1, cpumask);
 565                 break;
 566
 567         default:
 568                 break;
 569         }
 570
 571         return NOTIFY_OK;
 572 }
 573
 574
 575 static struct notifier_block boost_nb = {
 576         .notifier_call          = boost_notify,
 577 };
 578
 579 /*
 580  * acpi_cpufreq_early_init - initialize ACPI P-States library
 581  *
 582  * Initialize the ACPI P-States library (drivers/acpi/processor_perflib.c)
 583  * in order to determine correct frequency and voltage pairings. We can
 584  * do _PDC and _PSD and find out the processor dependency for the
 585  * actual init that will happen later...
 586  */
 587 static int __init acpi_cpufreq_early_init(void)
 588 {
 589         unsigned int i;
 590         pr_debug("acpi_cpufreq_early_init\n");
 591
 592         acpi_perf_data = alloc_percpu(struct acpi_processor_performance);
 593         if (!acpi_perf_data) {
 594                 pr_debug("Memory allocation error for acpi_perf_data.\n");
 595                 return -ENOMEM;
 596         }
 597         for_each_possible_cpu(i) {
 598                 if (!zalloc_cpumask_var_node(
 599                         &per_cpu_ptr(acpi_perf_data, i)->shared_cpu_map,
 600                         GFP_KERNEL, cpu_to_node(i))) {
 601
 602                         /* Freeing a NULL pointer is OK: alloc_percpu zeroes. */
 603                         free_acpi_perf_data();
 604                         return -ENOMEM;
 605                 }
 606         }
 607
 608         /* Do initialization in ACPI core */
 609         acpi_processor_preregister_performance(acpi_perf_data);
 610         return 0;
 611 }
 612
 613 #ifdef CONFIG_SMP
 614 /*
 615  * Some BIOSes do SW_ANY coordination internally, either set it up in hw
 616  * or do it in BIOS firmware and won't inform about it to OS. If not
 617  * detected, this has a side effect of making CPU run at a different speed
 618  * than OS intended it to run at. Detect it and handle it cleanly.
 619  */
 620 static int bios_with_sw_any_bug;
 621
 622 static int sw_any_bug_found(const struct dmi_system_id *d)
 623 {
 624         bios_with_sw_any_bug = 1;
 625         return 0;
 626 }
 627
 628 static const struct dmi_system_id sw_any_bug_dmi_table[] = {
 629         {
 630                 .callback = sw_any_bug_found,
 631                 .ident = "Supermicro Server X6DLP",
 632                 .matches = {
 633                         DMI_MATCH(DMI_SYS_VENDOR, "Supermicro"),
 634                         DMI_MATCH(DMI_BIOS_VERSION, "080010"),
 635                         DMI_MATCH(DMI_PRODUCT_NAME, "X6DLP"),
 636                 },
 637         },
 638         { }
 639 };
 640
 641 static int acpi_cpufreq_blacklist(struct cpuinfo_x86 *c)
 642 {
 643         /* Intel Xeon Processor 7100 Series Specification Update
 644          * http://www.intel.com/Assets/PDF/specupdate/314554.pdf
 645          * AL30: A Machine Check Exception (MCE) Occurring during an
 646          * Enhanced Intel SpeedStep Technology Ratio Change May Cause
 647          * Both Processor Cores to Lock Up. */
 648         if (c->x86_vendor == X86_VENDOR_INTEL) {
 649                 if ((c->x86 == 15) &&
 650                     (c->x86_model == 6) &&
 651                     (c->x86_mask == 8)) {
 652                         pr_info("Intel(R) Xeon(R) 7100 Errata AL30, processors may lock up on frequency changes: disabling acpi-cpufreq\n");
 653                         return -ENODEV;
 654                     }
 655                 }
 656         return 0;
 657 }
 658 #endif
 659
 660 static int acpi_cpufreq_cpu_init(struct cpufreq_policy *policy)
 661 {
 662         unsigned int i;
 663         unsigned int valid_states = 0;
 664         unsigned int cpu = policy->cpu;
 665         struct acpi_cpufreq_data *data;
 666         unsigned int result = 0;
 667         struct cpuinfo_x86 *c = &cpu_data(policy->cpu);
 668         struct acpi_processor_performance *perf;
 669         struct cpufreq_frequency_table *freq_table;
 670 #ifdef CONFIG_SMP
 671         static int blacklisted;
 672 #endif
 673
 674         pr_debug("acpi_cpufreq_cpu_init\n");
 675
 676 #ifdef CONFIG_SMP
 677         if (blacklisted)
 678                 return blacklisted;
 679         blacklisted = acpi_cpufreq_blacklist(c);
 680         if (blacklisted)
 681                 return blacklisted;
 682 #endif
 683
 684         data = kzalloc(sizeof(*data), GFP_KERNEL);
 685         if (!data)
 686                 return -ENOMEM;
 687
 688         if (!zalloc_cpumask_var(&data->freqdomain_cpus, GFP_KERNEL)) {
 689                 result = -ENOMEM;
 690                 goto err_free;
 691         }
 692
 693         perf = per_cpu_ptr(acpi_perf_data, cpu);
 694         data->acpi_perf_cpu = cpu;
 695         policy->driver_data = data;
 696
 697         if (cpu_has(c, X86_FEATURE_CONSTANT_TSC))
 698                 acpi_cpufreq_driver.flags |= CPUFREQ_CONST_LOOPS;
 699
 700         result = acpi_processor_register_performance(perf, cpu);
 701         if (result)
 702                 goto err_free_mask;
 703
 704         policy->shared_type = perf->shared_type;
 705
 706         /*
 707          * Will let policy->cpus know about dependency only when software
 708          * coordination is required.
 709          */
 710         if (policy->shared_type == CPUFREQ_SHARED_TYPE_ALL ||
 711             policy->shared_type == CPUFREQ_SHARED_TYPE_ANY) {
 712                 cpumask_copy(policy->cpus, perf->shared_cpu_map);
 713         }
 714         cpumask_copy(data->freqdomain_cpus, perf->shared_cpu_map);
 715
 716 #ifdef CONFIG_SMP
 717         dmi_check_system(sw_any_bug_dmi_table);
 718         if (bios_with_sw_any_bug && !policy_is_shared(policy)) {
 719                 policy->shared_type = CPUFREQ_SHARED_TYPE_ALL;
 720                 cpumask_copy(policy->cpus, topology_core_cpumask(cpu));
 721         }
 722
 723         if (check_amd_hwpstate_cpu(cpu) && !acpi_pstate_strict) {
 724                 cpumask_clear(policy->cpus);
 725                 cpumask_set_cpu(cpu, policy->cpus);
 726                 cpumask_copy(data->freqdomain_cpus,
 727                              topology_sibling_cpumask(cpu));
 728                 policy->shared_type = CPUFREQ_SHARED_TYPE_HW;
 729                 pr_info_once("overriding BIOS provided _PSD data\n");
 730         }
 731 #endif
 732
 733         /* capability check */
 734         if (perf->state_count <= 1) {
 735                 pr_debug("No P-States\n");
 736                 result = -ENODEV;
 737                 goto err_unreg;
 738         }
 739
 740         if (perf->control_register.space_id != perf->status_register.space_id) {
 741                 result = -ENODEV;
 742                 goto err_unreg;
 743         }
 744
 745         switch (perf->control_register.space_id) {
 746         case ACPI_ADR_SPACE_SYSTEM_IO:
 747                 if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD &&
 748                     boot_cpu_data.x86 == 0xf) {
 749                         pr_debug("AMD K8 systems must use native drivers.\n");
 750                         result = -ENODEV;
 751                         goto err_unreg;
 752                 }
 753                 pr_debug("SYSTEM IO addr space\n");
 754                 data->cpu_feature = SYSTEM_IO_CAPABLE;
 755                 data->cpu_freq_read = cpu_freq_read_io;
 756                 data->cpu_freq_write = cpu_freq_write_io;
 757                 break;
 758         case ACPI_ADR_SPACE_FIXED_HARDWARE:
 759                 pr_debug("HARDWARE addr space\n");
 760                 if (check_est_cpu(cpu)) {
 761                         data->cpu_feature = SYSTEM_INTEL_MSR_CAPABLE;
 762                         data->cpu_freq_read = cpu_freq_read_intel;
 763                         data->cpu_freq_write = cpu_freq_write_intel;
 764                         break;
 765                 }
 766                 if (check_amd_hwpstate_cpu(cpu)) {
 767                         data->cpu_feature = SYSTEM_AMD_MSR_CAPABLE;
 768                         data->cpu_freq_read = cpu_freq_read_amd;
 769                         data->cpu_freq_write = cpu_freq_write_amd;
 770                         break;
 771                 }
 772                 result = -ENODEV;
 773                 goto err_unreg;
 774         default:
 775                 pr_debug("Unknown addr space %d\n",
 776                         (u32) (perf->control_register.space_id));
 777                 result = -ENODEV;
 778                 goto err_unreg;
 779         }
 780
 781         freq_table = kzalloc(sizeof(*freq_table) *
 782                     (perf->state_count+1), GFP_KERNEL);
 783         if (!freq_table) {
 784                 result = -ENOMEM;
 785                 goto err_unreg;
 786         }
 787
 788         /* detect transition latency */
 789         policy->cpuinfo.transition_latency = 0;
 790         for (i = 0; i < perf->state_count; i++) {
 791                 if ((perf->states[i].transition_latency * 1000) >
 792                     policy->cpuinfo.transition_latency)
 793                         policy->cpuinfo.transition_latency =
 794                             perf->states[i].transition_latency * 1000;
 795         }
 796
 797         /* Check for high latency (>20uS) from buggy BIOSes, like on T42 */
 798         if (perf->control_register.space_id == ACPI_ADR_SPACE_FIXED_HARDWARE &&
 799             policy->cpuinfo.transition_latency > 20 * 1000) {
 800                 policy->cpuinfo.transition_latency = 20 * 1000;
 801                 pr_info_once("P-state transition latency capped at 20 uS\n");
 802         }
 803
 804         /* table init */
 805         for (i = 0; i < perf->state_count; i++) {
 806                 if (i > 0 && perf->states[i].core_frequency >=
 807                     freq_table[valid_states-1].frequency / 1000)
 808                         continue;
 809
 810                 freq_table[valid_states].driver_data = i;
 811                 freq_table[valid_states].frequency =
 812                     perf->states[i].core_frequency * 1000;
 813                 valid_states++;
 814         }
 815         freq_table[valid_states].frequency = CPUFREQ_TABLE_END;
 816         perf->state = 0;
 817
 818         result = cpufreq_table_validate_and_show(policy, freq_table);
 819         if (result)
 820                 goto err_freqfree;
 821
 822         if (perf->states[0].core_frequency * 1000 != policy->cpuinfo.max_freq)
 823                 pr_warn(FW_WARN "P-state 0 is not max freq\n");
 824
 825         switch (perf->control_register.space_id) {
 826         case ACPI_ADR_SPACE_SYSTEM_IO:
 827                 /*
 828                  * The core will not set policy->cur, because
 829                  * cpufreq_driver->get is NULL, so we need to set it here.
 830                  * However, we have to guess it, because the current speed is
 831                  * unknown and not detectable via IO ports.
 832                  */
 833                 policy->cur = acpi_cpufreq_guess_freq(data, policy->cpu);
 834                 break;
 835         case ACPI_ADR_SPACE_FIXED_HARDWARE:
 836                 acpi_cpufreq_driver.get = get_cur_freq_on_cpu;
 837                 break;
 838         default:
 839                 break;
 840         }
 841
 842         /* notify BIOS that we exist */
 843         acpi_processor_notify_smm(THIS_MODULE);
 844
 845         pr_debug("CPU%u - ACPI performance management activated.\n", cpu);
 846         for (i = 0; i < perf->state_count; i++)
 847                 pr_debug("     %cP%d: %d MHz, %d mW, %d uS\n",
 848                         (i == perf->state ? '*' : ' '), i,
 849                         (u32) perf->states[i].core_frequency,
 850                         (u32) perf->states[i].power,
 851                         (u32) perf->states[i].transition_latency);
 852
 853         /*
 854          * the first call to ->target() should result in us actually
 855          * writing something to the appropriate registers.
 856          */
 857         data->resume = 1;
 858
 859         policy->fast_switch_possible = !acpi_pstate_strict &&
 860                 !(policy_is_shared(policy) && policy->shared_type != CPUFREQ_SHARED_TYPE_ANY);
 861
 862         return result;
 863
 864 err_freqfree:
 865         kfree(freq_table);
 866 err_unreg:
 867         acpi_processor_unregister_performance(cpu);
 868 err_free_mask:
 869         free_cpumask_var(data->freqdomain_cpus);
 870 err_free:
 871         kfree(data);
 872         policy->driver_data = NULL;
 873
 874         return result;
 875 }
 876
 877 static int acpi_cpufreq_cpu_exit(struct cpufreq_policy *policy)
 878 {
 879         struct acpi_cpufreq_data *data = policy->driver_data;
 880
 881         pr_debug("acpi_cpufreq_cpu_exit\n");
 882
 883         policy->fast_switch_possible = false;
 884         policy->driver_data = NULL;
 885         acpi_processor_unregister_performance(data->acpi_perf_cpu);
 886         free_cpumask_var(data->freqdomain_cpus);
 887         kfree(policy->freq_table);
 888         kfree(data);
 889
 890         return 0;
 891 }
 892
 893 static int acpi_cpufreq_resume(struct cpufreq_policy *policy)
 894 {
 895         struct acpi_cpufreq_data *data = policy->driver_data;
 896
 897         pr_debug("acpi_cpufreq_resume\n");
 898
 899         data->resume = 1;
 900
 901         return 0;
 902 }
 903
 904 static struct freq_attr *acpi_cpufreq_attr[] = {
 905         &cpufreq_freq_attr_scaling_available_freqs,
 906         &freqdomain_cpus,
 907 #ifdef CONFIG_X86_ACPI_CPUFREQ_CPB
 908         &cpb,
 909 #endif
 910         NULL,
 911 };
 912
 913 static struct cpufreq_driver acpi_cpufreq_driver = {
 914         .verify         = cpufreq_generic_frequency_table_verify,
 915         .target_index   = acpi_cpufreq_target,
 916         .fast_switch    = acpi_cpufreq_fast_switch,
 917         .bios_limit     = acpi_processor_get_bios_limit,
 918         .init           = acpi_cpufreq_cpu_init,
 919         .exit           = acpi_cpufreq_cpu_exit,
 920         .resume         = acpi_cpufreq_resume,
 921         .name           = "acpi-cpufreq",
 922         .attr           = acpi_cpufreq_attr,
 923 };
 924
 925 static void __init acpi_cpufreq_boost_init(void)
 926 {
 927         if (boot_cpu_has(X86_FEATURE_CPB) || boot_cpu_has(X86_FEATURE_IDA)) {
 928                 msrs = msrs_alloc();
 929
 930                 if (!msrs)
 931                         return;
 932
 933                 acpi_cpufreq_driver.set_boost = set_boost;
 934                 acpi_cpufreq_driver.boost_enabled = boost_state(0);
 935
 936                 cpu_notifier_register_begin();
 937
 938                 /* Force all MSRs to the same value */
 939                 boost_set_msrs(acpi_cpufreq_driver.boost_enabled,
 940                                cpu_online_mask);
 941
 942                 __register_cpu_notifier(&boost_nb);
 943
 944                 cpu_notifier_register_done();
 945         }
 946 }
 947
 948 static void acpi_cpufreq_boost_exit(void)
 949 {
 950         if (msrs) {
 951                 unregister_cpu_notifier(&boost_nb);
 952
 953                 msrs_free(msrs);
 954                 msrs = NULL;
 955         }
 956 }
 957
 958 static int __init acpi_cpufreq_init(void)
 959 {
 960         int ret;
 961
 962         if (acpi_disabled)
 963                 return -ENODEV;
 964
 965         /* don't keep reloading if cpufreq_driver exists */
 966         if (cpufreq_get_current_driver())
 967                 return -EEXIST;
 968
 969         pr_debug("acpi_cpufreq_init\n");
 970
 971         ret = acpi_cpufreq_early_init();
 972         if (ret)
 973                 return ret;
 974
 975 #ifdef CONFIG_X86_ACPI_CPUFREQ_CPB
 976         /* this is a sysfs file with a strange name and an even stranger
 977          * semantic - per CPU instantiation, but system global effect.
 978          * Lets enable it only on AMD CPUs for compatibility reasons and
 979          * only if configured. This is considered legacy code, which
 980          * will probably be removed at some point in the future.
 981          */
 982         if (!check_amd_hwpstate_cpu(0)) {
 983                 struct freq_attr **attr;
 984
 985                 pr_debug("CPB unsupported, do not expose it\n");
 986
 987                 for (attr = acpi_cpufreq_attr; *attr; attr++)
 988                         if (*attr == &cpb) {
 989                                 *attr = NULL;
 990                                 break;
 991                         }
 992         }
 993 #endif
 994         acpi_cpufreq_boost_init();
 995
 996         ret = cpufreq_register_driver(&acpi_cpufreq_driver);
 997         if (ret) {
 998                 free_acpi_perf_data();
 999                 acpi_cpufreq_boost_exit();
1000         }
1001         return ret;
1002 }
1003
1004 static void __exit acpi_cpufreq_exit(void)
1005 {
1006         pr_debug("acpi_cpufreq_exit\n");
1007
1008         acpi_cpufreq_boost_exit();
1009
1010         cpufreq_unregister_driver(&acpi_cpufreq_driver);
1011
1012         free_acpi_perf_data();
1013 }
1014
1015 module_param(acpi_pstate_strict, uint, 0644);
1016 MODULE_PARM_DESC(acpi_pstate_strict,
1017         "value 0 or non-zero. non-zero -> strict ACPI checks are "
1018         "performed during frequency changes.");
1019
1020 late_initcall(acpi_cpufreq_init);
1021 module_exit(acpi_cpufreq_exit);
1022
1023 static const struct x86_cpu_id acpi_cpufreq_ids[] = {
1024         X86_FEATURE_MATCH(X86_FEATURE_ACPI),
1025         X86_FEATURE_MATCH(X86_FEATURE_HW_PSTATE),
1026         {}
1027 };
1028 MODULE_DEVICE_TABLE(x86cpu, acpi_cpufreq_ids);
1029
1030 static const struct acpi_device_id processor_device_ids[] = {
1031         {ACPI_PROCESSOR_OBJECT_HID, },
1032         {ACPI_PROCESSOR_DEVICE_HID, },
1033         {},
1034 };
1035 MODULE_DEVICE_TABLE(acpi, processor_device_ids);
1036
1037 MODULE_ALIAS("acpi");