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 #include <linux/kernel.h>
  29 #include <linux/module.h>
  30 #include <linux/init.h>
  31 #include <linux/smp.h>
  32 #include <linux/sched.h>
  33 #include <linux/cpufreq.h>
  34 #include <linux/compiler.h>
  35 #include <linux/dmi.h>
  36 #include <linux/slab.h>
  37
  38 #include <linux/acpi.h>
  39 #include <linux/io.h>
  40 #include <linux/delay.h>
  41 #include <linux/uaccess.h>
  42
  43 #include <acpi/processor.h>
  44
  45 #include <asm/msr.h>
  46 #include <asm/processor.h>
  47 #include <asm/cpufeature.h>
  48
  49 MODULE_AUTHOR("Paul Diefenbaugh, Dominik Brodowski");
  50 MODULE_DESCRIPTION("ACPI Processor P-States Driver");
  51 MODULE_LICENSE("GPL");
  52
  53 #define PFX "acpi-cpufreq: "
  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         struct cpufreq_frequency_table *freq_table;
  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 static struct msr __percpu *msrs;
  89
  90 static bool boost_state(unsigned int cpu)
  91 {
  92         u32 lo, hi;
  93         u64 msr;
  94
  95         switch (boot_cpu_data.x86_vendor) {
  96         case X86_VENDOR_INTEL:
  97                 rdmsr_on_cpu(cpu, MSR_IA32_MISC_ENABLE, &lo, &hi);
  98                 msr = lo | ((u64)hi << 32);
  99                 return !(msr & MSR_IA32_MISC_ENABLE_TURBO_DISABLE);
 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 void boost_set_msrs(bool enable, const struct cpumask *cpumask)
 109 {
 110         u32 cpu;
 111         u32 msr_addr;
 112         u64 msr_mask;
 113
 114         switch (boot_cpu_data.x86_vendor) {
 115         case X86_VENDOR_INTEL:
 116                 msr_addr = MSR_IA32_MISC_ENABLE;
 117                 msr_mask = MSR_IA32_MISC_ENABLE_TURBO_DISABLE;
 118                 break;
 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;
 125         }
 126
 127         rdmsr_on_cpus(cpumask, msr_addr, msrs);
 128
 129         for_each_cpu(cpu, cpumask) {
 130                 struct msr *reg = per_cpu_ptr(msrs, cpu);
 131                 if (enable)
 132                         reg->q &= ~msr_mask;
 133                 else
 134                         reg->q |= msr_mask;
 135         }
 136
 137         wrmsr_on_cpus(cpumask, msr_addr, msrs);
 138 }
 139
 140 static int set_boost(int val)
 141 {
 142         get_online_cpus();
 143         boost_set_msrs(val, cpu_online_mask);
 144         put_online_cpus();
 145         pr_debug("Core Boosting %sabled.\n", val ? "en" : "dis");
 146
 147         return 0;
 148 }
 149
 150 static ssize_t show_freqdomain_cpus(struct cpufreq_policy *policy, char *buf)
 151 {
 152         struct acpi_cpufreq_data *data = policy->driver_data;
 153
 154         if (unlikely(!data))
 155                 return -ENODEV;
 156
 157         return cpufreq_show_cpus(data->freqdomain_cpus, buf);
 158 }
 159
 160 cpufreq_freq_attr_ro(freqdomain_cpus);
 161
 162 #ifdef CONFIG_X86_ACPI_CPUFREQ_CPB
 163 static ssize_t store_cpb(struct cpufreq_policy *policy, const char *buf,
 164                          size_t count)
 165 {
 166         int ret;
 167         unsigned int val = 0;
 168
 169         if (!acpi_cpufreq_driver.set_boost)
 170                 return -EINVAL;
 171
 172         ret = kstrtouint(buf, 10, &val);
 173         if (ret || val > 1)
 174                 return -EINVAL;
 175
 176         set_boost(val);
 177
 178         return count;
 179 }
 180
 181 static ssize_t show_cpb(struct cpufreq_policy *policy, char *buf)
 182 {
 183         return sprintf(buf, "%u\n", acpi_cpufreq_driver.boost_enabled);
 184 }
 185
 186 cpufreq_freq_attr_rw(cpb);
 187 #endif
 188
 189 static int check_est_cpu(unsigned int cpuid)
 190 {
 191         struct cpuinfo_x86 *cpu = &cpu_data(cpuid);
 192
 193         return cpu_has(cpu, X86_FEATURE_EST);
 194 }
 195
 196 static int check_amd_hwpstate_cpu(unsigned int cpuid)
 197 {
 198         struct cpuinfo_x86 *cpu = &cpu_data(cpuid);
 199
 200         return cpu_has(cpu, X86_FEATURE_HW_PSTATE);
 201 }
 202
 203 static unsigned extract_io(u32 value, struct acpi_cpufreq_data *data)
 204 {
 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 data->freq_table[i].frequency;
 213         }
 214         return 0;
 215 }
 216
 217 static unsigned extract_msr(u32 msr, struct acpi_cpufreq_data *data)
 218 {
 219         struct cpufreq_frequency_table *pos;
 220         struct acpi_processor_performance *perf;
 221
 222         if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD)
 223                 msr &= AMD_MSR_RANGE;
 224         else
 225                 msr &= INTEL_MSR_RANGE;
 226
 227         perf = to_perf_data(data);
 228
 229         cpufreq_for_each_entry(pos, data->freq_table)
 230                 if (msr == perf->states[pos->driver_data].status)
 231                         return pos->frequency;
 232         return data->freq_table[0].frequency;
 233 }
 234
 235 static unsigned extract_freq(u32 val, struct acpi_cpufreq_data *data)
 236 {
 237         switch (data->cpu_feature) {
 238         case SYSTEM_INTEL_MSR_CAPABLE:
 239         case SYSTEM_AMD_MSR_CAPABLE:
 240                 return extract_msr(val, data);
 241         case SYSTEM_IO_CAPABLE:
 242                 return extract_io(val, data);
 243         default:
 244                 return 0;
 245         }
 246 }
 247
 248 static u32 cpu_freq_read_intel(struct acpi_pct_register *not_used)
 249 {
 250         u32 val, dummy;
 251
 252         rdmsr(MSR_IA32_PERF_CTL, val, dummy);
 253         return val;
 254 }
 255
 256 static void cpu_freq_write_intel(struct acpi_pct_register *not_used, u32 val)
 257 {
 258         u32 lo, hi;
 259
 260         rdmsr(MSR_IA32_PERF_CTL, lo, hi);
 261         lo = (lo & ~INTEL_MSR_RANGE) | (val & INTEL_MSR_RANGE);
 262         wrmsr(MSR_IA32_PERF_CTL, lo, hi);
 263 }
 264
 265 static u32 cpu_freq_read_amd(struct acpi_pct_register *not_used)
 266 {
 267         u32 val, dummy;
 268
 269         rdmsr(MSR_AMD_PERF_CTL, val, dummy);
 270         return val;
 271 }
 272
 273 static void cpu_freq_write_amd(struct acpi_pct_register *not_used, u32 val)
 274 {
 275         wrmsr(MSR_AMD_PERF_CTL, val, 0);
 276 }
 277
 278 static u32 cpu_freq_read_io(struct acpi_pct_register *reg)
 279 {
 280         u32 val;
 281
 282         acpi_os_read_port(reg->address, &val, reg->bit_width);
 283         return val;
 284 }
 285
 286 static void cpu_freq_write_io(struct acpi_pct_register *reg, u32 val)
 287 {
 288         acpi_os_write_port(reg->address, val, reg->bit_width);
 289 }
 290
 291 struct drv_cmd {
 292         struct acpi_pct_register *reg;
 293         u32 val;
 294         union {
 295                 void (*write)(struct acpi_pct_register *reg, u32 val);
 296                 u32 (*read)(struct acpi_pct_register *reg);
 297         } func;
 298 };
 299
 300 /* Called via smp_call_function_single(), on the target CPU */
 301 static void do_drv_read(void *_cmd)
 302 {
 303         struct drv_cmd *cmd = _cmd;
 304
 305         cmd->val = cmd->func.read(cmd->reg);
 306 }
 307
 308 static u32 drv_read(struct acpi_cpufreq_data *data, const struct cpumask *mask)
 309 {
 310         struct acpi_processor_performance *perf = to_perf_data(data);
 311         struct drv_cmd cmd = {
 312                 .reg = &perf->control_register,
 313                 .func.read = data->cpu_freq_read,
 314         };
 315         int err;
 316
 317         err = smp_call_function_any(mask, do_drv_read, &cmd, 1);
 318         WARN_ON_ONCE(err);      /* smp_call_function_any() was buggy? */
 319         return cmd.val;
 320 }
 321
 322 /* Called via smp_call_function_many(), on the target CPUs */
 323 static void do_drv_write(void *_cmd)
 324 {
 325         struct drv_cmd *cmd = _cmd;
 326
 327         cmd->func.write(cmd->reg, cmd->val);
 328 }
 329
 330 static void drv_write(struct acpi_cpufreq_data *data,
 331                       const struct cpumask *mask, u32 val)
 332 {
 333         struct acpi_processor_performance *perf = to_perf_data(data);
 334         struct drv_cmd cmd = {
 335                 .reg = &perf->control_register,
 336                 .val = val,
 337                 .func.write = data->cpu_freq_write,
 338         };
 339         int this_cpu;
 340
 341         this_cpu = get_cpu();
 342         if (cpumask_test_cpu(this_cpu, mask))
 343                 do_drv_write(&cmd);
 344
 345         smp_call_function_many(mask, do_drv_write, &cmd, 1);
 346         put_cpu();
 347 }
 348
 349 static u32 get_cur_val(const struct cpumask *mask, struct acpi_cpufreq_data *data)
 350 {
 351         u32 val;
 352
 353         if (unlikely(cpumask_empty(mask)))
 354                 return 0;
 355
 356         val = drv_read(data, mask);
 357
 358         pr_debug("get_cur_val = %u\n", val);
 359
 360         return val;
 361 }
 362
 363 static unsigned int get_cur_freq_on_cpu(unsigned int cpu)
 364 {
 365         struct acpi_cpufreq_data *data;
 366         struct cpufreq_policy *policy;
 367         unsigned int freq;
 368         unsigned int cached_freq;
 369
 370         pr_debug("get_cur_freq_on_cpu (%d)\n", cpu);
 371
 372         policy = cpufreq_cpu_get_raw(cpu);
 373         if (unlikely(!policy))
 374                 return 0;
 375
 376         data = policy->driver_data;
 377         if (unlikely(!data || !data->freq_table))
 378                 return 0;
 379
 380         cached_freq = data->freq_table[to_perf_data(data)->state].frequency;
 381         freq = extract_freq(get_cur_val(cpumask_of(cpu), data), data);
 382         if (freq != cached_freq) {
 383                 /*
 384                  * The dreaded BIOS frequency change behind our back.
 385                  * Force set the frequency on next target call.
 386                  */
 387                 data->resume = 1;
 388         }
 389
 390         pr_debug("cur freq = %u\n", freq);
 391
 392         return freq;
 393 }
 394
 395 static unsigned int check_freqs(const struct cpumask *mask, unsigned int freq,
 396                                 struct acpi_cpufreq_data *data)
 397 {
 398         unsigned int cur_freq;
 399         unsigned int i;
 400
 401         for (i = 0; i < 100; i++) {
 402                 cur_freq = extract_freq(get_cur_val(mask, data), data);
 403                 if (cur_freq == freq)
 404                         return 1;
 405                 udelay(10);
 406         }
 407         return 0;
 408 }
 409
 410 static int acpi_cpufreq_target(struct cpufreq_policy *policy,
 411                                unsigned int index)
 412 {
 413         struct acpi_cpufreq_data *data = policy->driver_data;
 414         struct acpi_processor_performance *perf;
 415         const struct cpumask *mask;
 416         unsigned int next_perf_state = 0; /* Index into perf table */
 417         int result = 0;
 418
 419         if (unlikely(data == NULL || data->freq_table == NULL)) {
 420                 return -ENODEV;
 421         }
 422
 423         perf = to_perf_data(data);
 424         next_perf_state = data->freq_table[index].driver_data;
 425         if (perf->state == next_perf_state) {
 426                 if (unlikely(data->resume)) {
 427                         pr_debug("Called after resume, resetting to P%d\n",
 428                                 next_perf_state);
 429                         data->resume = 0;
 430                 } else {
 431                         pr_debug("Already at target state (P%d)\n",
 432                                 next_perf_state);
 433                         return 0;
 434                 }
 435         }
 436
 437         /*
 438          * The core won't allow CPUs to go away until the governor has been
 439          * stopped, so we can rely on the stability of policy->cpus.
 440          */
 441         mask = policy->shared_type == CPUFREQ_SHARED_TYPE_ANY ?
 442                 cpumask_of(policy->cpu) : policy->cpus;
 443
 444         drv_write(data, mask, perf->states[next_perf_state].control);
 445
 446         if (acpi_pstate_strict) {
 447                 if (!check_freqs(mask, data->freq_table[index].frequency,
 448                                         data)) {
 449                         pr_debug("acpi_cpufreq_target failed (%d)\n",
 450                                 policy->cpu);
 451                         result = -EAGAIN;
 452                 }
 453         }
 454
 455         if (!result)
 456                 perf->state = next_perf_state;
 457
 458         return result;
 459 }
 460
 461 unsigned int acpi_cpufreq_fast_switch(struct cpufreq_policy *policy,
 462                                       unsigned int target_freq)
 463 {
 464         struct acpi_cpufreq_data *data = policy->driver_data;
 465         struct acpi_processor_performance *perf;
 466         struct cpufreq_frequency_table *entry;
 467         unsigned int next_perf_state, next_freq, freq;
 468
 469         /*
 470          * Find the closest frequency above target_freq.
 471          *
 472          * The table is sorted in the reverse order with respect to the
 473          * frequency and all of the entries are valid (see the initialization).
 474          */
 475         entry = data->freq_table;
 476         do {
 477                 entry++;
 478                 freq = entry->frequency;
 479         } while (freq >= target_freq && freq != CPUFREQ_TABLE_END);
 480         entry--;
 481         next_freq = entry->frequency;
 482         next_perf_state = entry->driver_data;
 483
 484         perf = to_perf_data(data);
 485         if (perf->state == next_perf_state) {
 486                 if (unlikely(data->resume))
 487                         data->resume = 0;
 488                 else
 489                         return next_freq;
 490         }
 491
 492         data->cpu_freq_write(&perf->control_register,
 493                              perf->states[next_perf_state].control);
 494         perf->state = next_perf_state;
 495         return next_freq;
 496 }
 497
 498 static unsigned long
 499 acpi_cpufreq_guess_freq(struct acpi_cpufreq_data *data, unsigned int cpu)
 500 {
 501         struct acpi_processor_performance *perf;
 502
 503         perf = to_perf_data(data);
 504         if (cpu_khz) {
 505                 /* search the closest match to cpu_khz */
 506                 unsigned int i;
 507                 unsigned long freq;
 508                 unsigned long freqn = perf->states[0].core_frequency * 1000;
 509
 510                 for (i = 0; i < (perf->state_count-1); i++) {
 511                         freq = freqn;
 512                         freqn = perf->states[i+1].core_frequency * 1000;
 513                         if ((2 * cpu_khz) > (freqn + freq)) {
 514                                 perf->state = i;
 515                                 return freq;
 516                         }
 517                 }
 518                 perf->state = perf->state_count-1;
 519                 return freqn;
 520         } else {
 521                 /* assume CPU is at P0... */
 522                 perf->state = 0;
 523                 return perf->states[0].core_frequency * 1000;
 524         }
 525 }
 526
 527 static void free_acpi_perf_data(void)
 528 {
 529         unsigned int i;
 530
 531         /* Freeing a NULL pointer is OK, and alloc_percpu zeroes. */
 532         for_each_possible_cpu(i)
 533                 free_cpumask_var(per_cpu_ptr(acpi_perf_data, i)
 534                                  ->shared_cpu_map);
 535         free_percpu(acpi_perf_data);
 536 }
 537
 538 static int boost_notify(struct notifier_block *nb, unsigned long action,
 539                       void *hcpu)
 540 {
 541         unsigned cpu = (long)hcpu;
 542         const struct cpumask *cpumask;
 543
 544         cpumask = get_cpu_mask(cpu);
 545
 546         /*
 547          * Clear the boost-disable bit on the CPU_DOWN path so that
 548          * this cpu cannot block the remaining ones from boosting. On
 549          * the CPU_UP path we simply keep the boost-disable flag in
 550          * sync with the current global state.
 551          */
 552
 553         switch (action) {
 554         case CPU_DOWN_FAILED:
 555         case CPU_DOWN_FAILED_FROZEN:
 556         case CPU_ONLINE:
 557         case CPU_ONLINE_FROZEN:
 558                 boost_set_msrs(acpi_cpufreq_driver.boost_enabled, cpumask);
 559                 break;
 560
 561         case CPU_DOWN_PREPARE:
 562         case CPU_DOWN_PREPARE_FROZEN:
 563                 boost_set_msrs(1, cpumask);
 564                 break;
 565
 566         default:
 567                 break;
 568         }
 569
 570         return NOTIFY_OK;
 571 }
 572
 573
 574 static struct notifier_block boost_nb = {
 575         .notifier_call          = boost_notify,
 576 };
 577
 578 /*
 579  * acpi_cpufreq_early_init - initialize ACPI P-States library
 580  *
 581  * Initialize the ACPI P-States library (drivers/acpi/processor_perflib.c)
 582  * in order to determine correct frequency and voltage pairings. We can
 583  * do _PDC and _PSD and find out the processor dependency for the
 584  * actual init that will happen later...
 585  */
 586 static int __init acpi_cpufreq_early_init(void)
 587 {
 588         unsigned int i;
 589         pr_debug("acpi_cpufreq_early_init\n");
 590
 591         acpi_perf_data = alloc_percpu(struct acpi_processor_performance);
 592         if (!acpi_perf_data) {
 593                 pr_debug("Memory allocation error for acpi_perf_data.\n");
 594                 return -ENOMEM;
 595         }
 596         for_each_possible_cpu(i) {
 597                 if (!zalloc_cpumask_var_node(
 598                         &per_cpu_ptr(acpi_perf_data, i)->shared_cpu_map,
 599                         GFP_KERNEL, cpu_to_node(i))) {
 600
 601                         /* Freeing a NULL pointer is OK: alloc_percpu zeroes. */
 602                         free_acpi_perf_data();
 603                         return -ENOMEM;
 604                 }
 605         }
 606
 607         /* Do initialization in ACPI core */
 608         acpi_processor_preregister_performance(acpi_perf_data);
 609         return 0;
 610 }
 611
 612 #ifdef CONFIG_SMP
 613 /*
 614  * Some BIOSes do SW_ANY coordination internally, either set it up in hw
 615  * or do it in BIOS firmware and won't inform about it to OS. If not
 616  * detected, this has a side effect of making CPU run at a different speed
 617  * than OS intended it to run at. Detect it and handle it cleanly.
 618  */
 619 static int bios_with_sw_any_bug;
 620
 621 static int sw_any_bug_found(const struct dmi_system_id *d)
 622 {
 623         bios_with_sw_any_bug = 1;
 624         return 0;
 625 }
 626
 627 static const struct dmi_system_id sw_any_bug_dmi_table[] = {
 628         {
 629                 .callback = sw_any_bug_found,
 630                 .ident = "Supermicro Server X6DLP",
 631                 .matches = {
 632                         DMI_MATCH(DMI_SYS_VENDOR, "Supermicro"),
 633                         DMI_MATCH(DMI_BIOS_VERSION, "080010"),
 634                         DMI_MATCH(DMI_PRODUCT_NAME, "X6DLP"),
 635                 },
 636         },
 637         { }
 638 };
 639
 640 static int acpi_cpufreq_blacklist(struct cpuinfo_x86 *c)
 641 {
 642         /* Intel Xeon Processor 7100 Series Specification Update
 643          * http://www.intel.com/Assets/PDF/specupdate/314554.pdf
 644          * AL30: A Machine Check Exception (MCE) Occurring during an
 645          * Enhanced Intel SpeedStep Technology Ratio Change May Cause
 646          * Both Processor Cores to Lock Up. */
 647         if (c->x86_vendor == X86_VENDOR_INTEL) {
 648                 if ((c->x86 == 15) &&
 649                     (c->x86_model == 6) &&
 650                     (c->x86_mask == 8)) {
 651                         pr_info("acpi-cpufreq: Intel(R) Xeon(R) 7100 Errata AL30, processors may lock up on frequency changes: disabling acpi-cpufreq\n");
 652                         return -ENODEV;
 653                     }
 654                 }
 655         return 0;
 656 }
 657 #endif
 658
 659 static int acpi_cpufreq_cpu_init(struct cpufreq_policy *policy)
 660 {
 661         unsigned int i;
 662         unsigned int valid_states = 0;
 663         unsigned int cpu = policy->cpu;
 664         struct acpi_cpufreq_data *data;
 665         unsigned int result = 0;
 666         struct cpuinfo_x86 *c = &cpu_data(policy->cpu);
 667         struct acpi_processor_performance *perf;
 668 #ifdef CONFIG_SMP
 669         static int blacklisted;
 670 #endif
 671
 672         pr_debug("acpi_cpufreq_cpu_init\n");
 673
 674 #ifdef CONFIG_SMP
 675         if (blacklisted)
 676                 return blacklisted;
 677         blacklisted = acpi_cpufreq_blacklist(c);
 678         if (blacklisted)
 679                 return blacklisted;
 680 #endif
 681
 682         data = kzalloc(sizeof(*data), GFP_KERNEL);
 683         if (!data)
 684                 return -ENOMEM;
 685
 686         if (!zalloc_cpumask_var(&data->freqdomain_cpus, GFP_KERNEL)) {
 687                 result = -ENOMEM;
 688                 goto err_free;
 689         }
 690
 691         perf = per_cpu_ptr(acpi_perf_data, cpu);
 692         data->acpi_perf_cpu = cpu;
 693         policy->driver_data = data;
 694
 695         if (cpu_has(c, X86_FEATURE_CONSTANT_TSC))
 696                 acpi_cpufreq_driver.flags |= CPUFREQ_CONST_LOOPS;
 697
 698         result = acpi_processor_register_performance(perf, cpu);
 699         if (result)
 700                 goto err_free_mask;
 701
 702         policy->shared_type = perf->shared_type;
 703
 704         /*
 705          * Will let policy->cpus know about dependency only when software
 706          * coordination is required.
 707          */
 708         if (policy->shared_type == CPUFREQ_SHARED_TYPE_ALL ||
 709             policy->shared_type == CPUFREQ_SHARED_TYPE_ANY) {
 710                 cpumask_copy(policy->cpus, perf->shared_cpu_map);
 711         }
 712         cpumask_copy(data->freqdomain_cpus, perf->shared_cpu_map);
 713
 714 #ifdef CONFIG_SMP
 715         dmi_check_system(sw_any_bug_dmi_table);
 716         if (bios_with_sw_any_bug && !policy_is_shared(policy)) {
 717                 policy->shared_type = CPUFREQ_SHARED_TYPE_ALL;
 718                 cpumask_copy(policy->cpus, topology_core_cpumask(cpu));
 719         }
 720
 721         if (check_amd_hwpstate_cpu(cpu) && !acpi_pstate_strict) {
 722                 cpumask_clear(policy->cpus);
 723                 cpumask_set_cpu(cpu, policy->cpus);
 724                 cpumask_copy(data->freqdomain_cpus,
 725                              topology_sibling_cpumask(cpu));
 726                 policy->shared_type = CPUFREQ_SHARED_TYPE_HW;
 727                 pr_info_once(PFX "overriding BIOS provided _PSD data\n");
 728         }
 729 #endif
 730
 731         /* capability check */
 732         if (perf->state_count <= 1) {
 733                 pr_debug("No P-States\n");
 734                 result = -ENODEV;
 735                 goto err_unreg;
 736         }
 737
 738         if (perf->control_register.space_id != perf->status_register.space_id) {
 739                 result = -ENODEV;
 740                 goto err_unreg;
 741         }
 742
 743         switch (perf->control_register.space_id) {
 744         case ACPI_ADR_SPACE_SYSTEM_IO:
 745                 if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD &&
 746                     boot_cpu_data.x86 == 0xf) {
 747                         pr_debug("AMD K8 systems must use native drivers.\n");
 748                         result = -ENODEV;
 749                         goto err_unreg;
 750                 }
 751                 pr_debug("SYSTEM IO addr space\n");
 752                 data->cpu_feature = SYSTEM_IO_CAPABLE;
 753                 data->cpu_freq_read = cpu_freq_read_io;
 754                 data->cpu_freq_write = cpu_freq_write_io;
 755                 break;
 756         case ACPI_ADR_SPACE_FIXED_HARDWARE:
 757                 pr_debug("HARDWARE addr space\n");
 758                 if (check_est_cpu(cpu)) {
 759                         data->cpu_feature = SYSTEM_INTEL_MSR_CAPABLE;
 760                         data->cpu_freq_read = cpu_freq_read_intel;
 761                         data->cpu_freq_write = cpu_freq_write_intel;
 762                         break;
 763                 }
 764                 if (check_amd_hwpstate_cpu(cpu)) {
 765                         data->cpu_feature = SYSTEM_AMD_MSR_CAPABLE;
 766                         data->cpu_freq_read = cpu_freq_read_amd;
 767                         data->cpu_freq_write = cpu_freq_write_amd;
 768                         break;
 769                 }
 770                 result = -ENODEV;
 771                 goto err_unreg;
 772         default:
 773                 pr_debug("Unknown addr space %d\n",
 774                         (u32) (perf->control_register.space_id));
 775                 result = -ENODEV;
 776                 goto err_unreg;
 777         }
 778
 779         data->freq_table = kzalloc(sizeof(*data->freq_table) *
 780                     (perf->state_count+1), GFP_KERNEL);
 781         if (!data->freq_table) {
 782                 result = -ENOMEM;
 783                 goto err_unreg;
 784         }
 785
 786         /* detect transition latency */
 787         policy->cpuinfo.transition_latency = 0;
 788         for (i = 0; i < perf->state_count; i++) {
 789                 if ((perf->states[i].transition_latency * 1000) >
 790                     policy->cpuinfo.transition_latency)
 791                         policy->cpuinfo.transition_latency =
 792                             perf->states[i].transition_latency * 1000;
 793         }
 794
 795         /* Check for high latency (>20uS) from buggy BIOSes, like on T42 */
 796         if (perf->control_register.space_id == ACPI_ADR_SPACE_FIXED_HARDWARE &&
 797             policy->cpuinfo.transition_latency > 20 * 1000) {
 798                 policy->cpuinfo.transition_latency = 20 * 1000;
 799                 pr_info_once("P-state transition latency capped at 20 uS\n");
 800         }
 801
 802         /* table init */
 803         for (i = 0; i < perf->state_count; i++) {
 804                 if (i > 0 && perf->states[i].core_frequency >=
 805                     data->freq_table[valid_states-1].frequency / 1000)
 806                         continue;
 807
 808                 data->freq_table[valid_states].driver_data = i;
 809                 data->freq_table[valid_states].frequency =
 810                     perf->states[i].core_frequency * 1000;
 811                 valid_states++;
 812         }
 813         data->freq_table[valid_states].frequency = CPUFREQ_TABLE_END;
 814         perf->state = 0;
 815
 816         result = cpufreq_table_validate_and_show(policy, data->freq_table);
 817         if (result)
 818                 goto err_freqfree;
 819
 820         if (perf->states[0].core_frequency * 1000 != policy->cpuinfo.max_freq)
 821                 pr_warn(FW_WARN "P-state 0 is not max freq\n");
 822
 823         switch (perf->control_register.space_id) {
 824         case ACPI_ADR_SPACE_SYSTEM_IO:
 825                 /*
 826                  * The core will not set policy->cur, because
 827                  * cpufreq_driver->get is NULL, so we need to set it here.
 828                  * However, we have to guess it, because the current speed is
 829                  * unknown and not detectable via IO ports.
 830                  */
 831                 policy->cur = acpi_cpufreq_guess_freq(data, policy->cpu);
 832                 break;
 833         case ACPI_ADR_SPACE_FIXED_HARDWARE:
 834                 acpi_cpufreq_driver.get = get_cur_freq_on_cpu;
 835                 break;
 836         default:
 837                 break;
 838         }
 839
 840         /* notify BIOS that we exist */
 841         acpi_processor_notify_smm(THIS_MODULE);
 842
 843         pr_debug("CPU%u - ACPI performance management activated.\n", cpu);
 844         for (i = 0; i < perf->state_count; i++)
 845                 pr_debug("     %cP%d: %d MHz, %d mW, %d uS\n",
 846                         (i == perf->state ? '*' : ' '), i,
 847                         (u32) perf->states[i].core_frequency,
 848                         (u32) perf->states[i].power,
 849                         (u32) perf->states[i].transition_latency);
 850
 851         /*
 852          * the first call to ->target() should result in us actually
 853          * writing something to the appropriate registers.
 854          */
 855         data->resume = 1;
 856
 857         policy->fast_switch_possible = !acpi_pstate_strict &&
 858                 !(policy_is_shared(policy) && policy->shared_type != CPUFREQ_SHARED_TYPE_ANY);
 859
 860         return result;
 861
 862 err_freqfree:
 863         kfree(data->freq_table);
 864 err_unreg:
 865         acpi_processor_unregister_performance(cpu);
 866 err_free_mask:
 867         free_cpumask_var(data->freqdomain_cpus);
 868 err_free:
 869         kfree(data);
 870         policy->driver_data = NULL;
 871
 872         return result;
 873 }
 874
 875 static int acpi_cpufreq_cpu_exit(struct cpufreq_policy *policy)
 876 {
 877         struct acpi_cpufreq_data *data = policy->driver_data;
 878
 879         pr_debug("acpi_cpufreq_cpu_exit\n");
 880
 881         if (data) {
 882                 policy->fast_switch_possible = false;
 883                 policy->driver_data = NULL;
 884                 acpi_processor_unregister_performance(data->acpi_perf_cpu);
 885                 free_cpumask_var(data->freqdomain_cpus);
 886                 kfree(data->freq_table);
 887                 kfree(data);
 888         }
 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");