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 acpi_processor_performance *acpi_data;
  69         struct cpufreq_frequency_table *freq_table;
  70         unsigned int resume;
  71         unsigned int cpu_feature;
  72         cpumask_var_t freqdomain_cpus;
  73 };
  74
  75 static DEFINE_PER_CPU(struct acpi_cpufreq_data *, acfreq_data);
  76
  77 /* acpi_perf_data is a pointer to percpu data. */
  78 static struct acpi_processor_performance __percpu *acpi_perf_data;
  79
  80 static struct cpufreq_driver acpi_cpufreq_driver;
  81
  82 static unsigned int acpi_pstate_strict;
  83 static struct msr __percpu *msrs;
  84
  85 static bool boost_state(unsigned int cpu)
  86 {
  87         u32 lo, hi;
  88         u64 msr;
  89
  90         switch (boot_cpu_data.x86_vendor) {
  91         case X86_VENDOR_INTEL:
  92                 rdmsr_on_cpu(cpu, MSR_IA32_MISC_ENABLE, &lo, &hi);
  93                 msr = lo | ((u64)hi << 32);
  94                 return !(msr & MSR_IA32_MISC_ENABLE_TURBO_DISABLE);
  95         case X86_VENDOR_AMD:
  96                 rdmsr_on_cpu(cpu, MSR_K7_HWCR, &lo, &hi);
  97                 msr = lo | ((u64)hi << 32);
  98                 return !(msr & MSR_K7_HWCR_CPB_DIS);
  99         }
 100         return false;
 101 }
 102
 103 static void boost_set_msrs(bool enable, const struct cpumask *cpumask)
 104 {
 105         u32 cpu;
 106         u32 msr_addr;
 107         u64 msr_mask;
 108
 109         switch (boot_cpu_data.x86_vendor) {
 110         case X86_VENDOR_INTEL:
 111                 msr_addr = MSR_IA32_MISC_ENABLE;
 112                 msr_mask = MSR_IA32_MISC_ENABLE_TURBO_DISABLE;
 113                 break;
 114         case X86_VENDOR_AMD:
 115                 msr_addr = MSR_K7_HWCR;
 116                 msr_mask = MSR_K7_HWCR_CPB_DIS;
 117                 break;
 118         default:
 119                 return;
 120         }
 121
 122         rdmsr_on_cpus(cpumask, msr_addr, msrs);
 123
 124         for_each_cpu(cpu, cpumask) {
 125                 struct msr *reg = per_cpu_ptr(msrs, cpu);
 126                 if (enable)
 127                         reg->q &= ~msr_mask;
 128                 else
 129                         reg->q |= msr_mask;
 130         }
 131
 132         wrmsr_on_cpus(cpumask, msr_addr, msrs);
 133 }
 134
 135 static int _store_boost(int val)
 136 {
 137         get_online_cpus();
 138         boost_set_msrs(val, cpu_online_mask);
 139         put_online_cpus();
 140         pr_debug("Core Boosting %sabled.\n", val ? "en" : "dis");
 141
 142         return 0;
 143 }
 144
 145 static ssize_t show_freqdomain_cpus(struct cpufreq_policy *policy, char *buf)
 146 {
 147         struct acpi_cpufreq_data *data = per_cpu(acfreq_data, policy->cpu);
 148
 149         return cpufreq_show_cpus(data->freqdomain_cpus, buf);
 150 }
 151
 152 cpufreq_freq_attr_ro(freqdomain_cpus);
 153
 154 #ifdef CONFIG_X86_ACPI_CPUFREQ_CPB
 155 static ssize_t store_boost(const char *buf, size_t count)
 156 {
 157         int ret;
 158         unsigned long val = 0;
 159
 160         if (!acpi_cpufreq_driver.boost_supported)
 161                 return -EINVAL;
 162
 163         ret = kstrtoul(buf, 10, &val);
 164         if (ret || (val > 1))
 165                 return -EINVAL;
 166
 167         _store_boost((int) val);
 168
 169         return count;
 170 }
 171
 172 static ssize_t store_cpb(struct cpufreq_policy *policy, const char *buf,
 173                          size_t count)
 174 {
 175         return store_boost(buf, count);
 176 }
 177
 178 static ssize_t show_cpb(struct cpufreq_policy *policy, char *buf)
 179 {
 180         return sprintf(buf, "%u\n", acpi_cpufreq_driver.boost_enabled);
 181 }
 182
 183 cpufreq_freq_attr_rw(cpb);
 184 #endif
 185
 186 static int check_est_cpu(unsigned int cpuid)
 187 {
 188         struct cpuinfo_x86 *cpu = &cpu_data(cpuid);
 189
 190         return cpu_has(cpu, X86_FEATURE_EST);
 191 }
 192
 193 static int check_amd_hwpstate_cpu(unsigned int cpuid)
 194 {
 195         struct cpuinfo_x86 *cpu = &cpu_data(cpuid);
 196
 197         return cpu_has(cpu, X86_FEATURE_HW_PSTATE);
 198 }
 199
 200 static unsigned extract_io(u32 value, struct acpi_cpufreq_data *data)
 201 {
 202         struct acpi_processor_performance *perf;
 203         int i;
 204
 205         perf = data->acpi_data;
 206
 207         for (i = 0; i < perf->state_count; i++) {
 208                 if (value == perf->states[i].status)
 209                         return data->freq_table[i].frequency;
 210         }
 211         return 0;
 212 }
 213
 214 static unsigned extract_msr(u32 msr, struct acpi_cpufreq_data *data)
 215 {
 216         struct cpufreq_frequency_table *pos;
 217         struct acpi_processor_performance *perf;
 218
 219         if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD)
 220                 msr &= AMD_MSR_RANGE;
 221         else
 222                 msr &= INTEL_MSR_RANGE;
 223
 224         perf = data->acpi_data;
 225
 226         cpufreq_for_each_entry(pos, data->freq_table)
 227                 if (msr == perf->states[pos->driver_data].status)
 228                         return pos->frequency;
 229         return data->freq_table[0].frequency;
 230 }
 231
 232 static unsigned extract_freq(u32 val, struct acpi_cpufreq_data *data)
 233 {
 234         switch (data->cpu_feature) {
 235         case SYSTEM_INTEL_MSR_CAPABLE:
 236         case SYSTEM_AMD_MSR_CAPABLE:
 237                 return extract_msr(val, data);
 238         case SYSTEM_IO_CAPABLE:
 239                 return extract_io(val, data);
 240         default:
 241                 return 0;
 242         }
 243 }
 244
 245 struct msr_addr {
 246         u32 reg;
 247 };
 248
 249 struct io_addr {
 250         u16 port;
 251         u8 bit_width;
 252 };
 253
 254 struct drv_cmd {
 255         unsigned int type;
 256         const struct cpumask *mask;
 257         union {
 258                 struct msr_addr msr;
 259                 struct io_addr io;
 260         } addr;
 261         u32 val;
 262 };
 263
 264 /* Called via smp_call_function_single(), on the target CPU */
 265 static void do_drv_read(void *_cmd)
 266 {
 267         struct drv_cmd *cmd = _cmd;
 268         u32 h;
 269
 270         switch (cmd->type) {
 271         case SYSTEM_INTEL_MSR_CAPABLE:
 272         case SYSTEM_AMD_MSR_CAPABLE:
 273                 rdmsr(cmd->addr.msr.reg, cmd->val, h);
 274                 break;
 275         case SYSTEM_IO_CAPABLE:
 276                 acpi_os_read_port((acpi_io_address)cmd->addr.io.port,
 277                                 &cmd->val,
 278                                 (u32)cmd->addr.io.bit_width);
 279                 break;
 280         default:
 281                 break;
 282         }
 283 }
 284
 285 /* Called via smp_call_function_many(), on the target CPUs */
 286 static void do_drv_write(void *_cmd)
 287 {
 288         struct drv_cmd *cmd = _cmd;
 289         u32 lo, hi;
 290
 291         switch (cmd->type) {
 292         case SYSTEM_INTEL_MSR_CAPABLE:
 293                 rdmsr(cmd->addr.msr.reg, lo, hi);
 294                 lo = (lo & ~INTEL_MSR_RANGE) | (cmd->val & INTEL_MSR_RANGE);
 295                 wrmsr(cmd->addr.msr.reg, lo, hi);
 296                 break;
 297         case SYSTEM_AMD_MSR_CAPABLE:
 298                 wrmsr(cmd->addr.msr.reg, cmd->val, 0);
 299                 break;
 300         case SYSTEM_IO_CAPABLE:
 301                 acpi_os_write_port((acpi_io_address)cmd->addr.io.port,
 302                                 cmd->val,
 303                                 (u32)cmd->addr.io.bit_width);
 304                 break;
 305         default:
 306                 break;
 307         }
 308 }
 309
 310 static void drv_read(struct drv_cmd *cmd)
 311 {
 312         int err;
 313         cmd->val = 0;
 314
 315         err = smp_call_function_any(cmd->mask, do_drv_read, cmd, 1);
 316         WARN_ON_ONCE(err);      /* smp_call_function_any() was buggy? */
 317 }
 318
 319 static void drv_write(struct drv_cmd *cmd)
 320 {
 321         int this_cpu;
 322
 323         this_cpu = get_cpu();
 324         if (cpumask_test_cpu(this_cpu, cmd->mask))
 325                 do_drv_write(cmd);
 326         smp_call_function_many(cmd->mask, do_drv_write, cmd, 1);
 327         put_cpu();
 328 }
 329
 330 static u32 get_cur_val(const struct cpumask *mask)
 331 {
 332         struct acpi_processor_performance *perf;
 333         struct drv_cmd cmd;
 334
 335         if (unlikely(cpumask_empty(mask)))
 336                 return 0;
 337
 338         switch (per_cpu(acfreq_data, cpumask_first(mask))->cpu_feature) {
 339         case SYSTEM_INTEL_MSR_CAPABLE:
 340                 cmd.type = SYSTEM_INTEL_MSR_CAPABLE;
 341                 cmd.addr.msr.reg = MSR_IA32_PERF_CTL;
 342                 break;
 343         case SYSTEM_AMD_MSR_CAPABLE:
 344                 cmd.type = SYSTEM_AMD_MSR_CAPABLE;
 345                 cmd.addr.msr.reg = MSR_AMD_PERF_CTL;
 346                 break;
 347         case SYSTEM_IO_CAPABLE:
 348                 cmd.type = SYSTEM_IO_CAPABLE;
 349                 perf = per_cpu(acfreq_data, cpumask_first(mask))->acpi_data;
 350                 cmd.addr.io.port = perf->control_register.address;
 351                 cmd.addr.io.bit_width = perf->control_register.bit_width;
 352                 break;
 353         default:
 354                 return 0;
 355         }
 356
 357         cmd.mask = mask;
 358         drv_read(&cmd);
 359
 360         pr_debug("get_cur_val = %u\n", cmd.val);
 361
 362         return cmd.val;
 363 }
 364
 365 static unsigned int get_cur_freq_on_cpu(unsigned int cpu)
 366 {
 367         struct acpi_cpufreq_data *data = per_cpu(acfreq_data, cpu);
 368         unsigned int freq;
 369         unsigned int cached_freq;
 370
 371         pr_debug("get_cur_freq_on_cpu (%d)\n", cpu);
 372
 373         if (unlikely(data == NULL ||
 374                      data->acpi_data == NULL || data->freq_table == NULL)) {
 375                 return 0;
 376         }
 377
 378         cached_freq = data->freq_table[data->acpi_data->state].frequency;
 379         freq = extract_freq(get_cur_val(cpumask_of(cpu)), data);
 380         if (freq != cached_freq) {
 381                 /*
 382                  * The dreaded BIOS frequency change behind our back.
 383                  * Force set the frequency on next target call.
 384                  */
 385                 data->resume = 1;
 386         }
 387
 388         pr_debug("cur freq = %u\n", freq);
 389
 390         return freq;
 391 }
 392
 393 static unsigned int check_freqs(const struct cpumask *mask, unsigned int freq,
 394                                 struct acpi_cpufreq_data *data)
 395 {
 396         unsigned int cur_freq;
 397         unsigned int i;
 398
 399         for (i = 0; i < 100; i++) {
 400                 cur_freq = extract_freq(get_cur_val(mask), data);
 401                 if (cur_freq == freq)
 402                         return 1;
 403                 udelay(10);
 404         }
 405         return 0;
 406 }
 407
 408 static int acpi_cpufreq_target(struct cpufreq_policy *policy,
 409                                unsigned int index)
 410 {
 411         struct acpi_cpufreq_data *data = per_cpu(acfreq_data, policy->cpu);
 412         struct acpi_processor_performance *perf;
 413         struct drv_cmd cmd;
 414         unsigned int next_perf_state = 0; /* Index into perf table */
 415         int result = 0;
 416
 417         if (unlikely(data == NULL ||
 418              data->acpi_data == NULL || data->freq_table == NULL)) {
 419                 return -ENODEV;
 420         }
 421
 422         perf = data->acpi_data;
 423         next_perf_state = data->freq_table[index].driver_data;
 424         if (perf->state == next_perf_state) {
 425                 if (unlikely(data->resume)) {
 426                         pr_debug("Called after resume, resetting to P%d\n",
 427                                 next_perf_state);
 428                         data->resume = 0;
 429                 } else {
 430                         pr_debug("Already at target state (P%d)\n",
 431                                 next_perf_state);
 432                         goto out;
 433                 }
 434         }
 435
 436         switch (data->cpu_feature) {
 437         case SYSTEM_INTEL_MSR_CAPABLE:
 438                 cmd.type = SYSTEM_INTEL_MSR_CAPABLE;
 439                 cmd.addr.msr.reg = MSR_IA32_PERF_CTL;
 440                 cmd.val = (u32) perf->states[next_perf_state].control;
 441                 break;
 442         case SYSTEM_AMD_MSR_CAPABLE:
 443                 cmd.type = SYSTEM_AMD_MSR_CAPABLE;
 444                 cmd.addr.msr.reg = MSR_AMD_PERF_CTL;
 445                 cmd.val = (u32) perf->states[next_perf_state].control;
 446                 break;
 447         case SYSTEM_IO_CAPABLE:
 448                 cmd.type = SYSTEM_IO_CAPABLE;
 449                 cmd.addr.io.port = perf->control_register.address;
 450                 cmd.addr.io.bit_width = perf->control_register.bit_width;
 451                 cmd.val = (u32) perf->states[next_perf_state].control;
 452                 break;
 453         default:
 454                 result = -ENODEV;
 455                 goto out;
 456         }
 457
 458         /* cpufreq holds the hotplug lock, so we are safe from here on */
 459         if (policy->shared_type != CPUFREQ_SHARED_TYPE_ANY)
 460                 cmd.mask = policy->cpus;
 461         else
 462                 cmd.mask = cpumask_of(policy->cpu);
 463
 464         drv_write(&cmd);
 465
 466         if (acpi_pstate_strict) {
 467                 if (!check_freqs(cmd.mask, data->freq_table[index].frequency,
 468                                         data)) {
 469                         pr_debug("acpi_cpufreq_target failed (%d)\n",
 470                                 policy->cpu);
 471                         result = -EAGAIN;
 472                 }
 473         }
 474
 475         if (!result)
 476                 perf->state = next_perf_state;
 477
 478 out:
 479         return result;
 480 }
 481
 482 static unsigned long
 483 acpi_cpufreq_guess_freq(struct acpi_cpufreq_data *data, unsigned int cpu)
 484 {
 485         struct acpi_processor_performance *perf = data->acpi_data;
 486
 487         if (cpu_khz) {
 488                 /* search the closest match to cpu_khz */
 489                 unsigned int i;
 490                 unsigned long freq;
 491                 unsigned long freqn = perf->states[0].core_frequency * 1000;
 492
 493                 for (i = 0; i < (perf->state_count-1); i++) {
 494                         freq = freqn;
 495                         freqn = perf->states[i+1].core_frequency * 1000;
 496                         if ((2 * cpu_khz) > (freqn + freq)) {
 497                                 perf->state = i;
 498                                 return freq;
 499                         }
 500                 }
 501                 perf->state = perf->state_count-1;
 502                 return freqn;
 503         } else {
 504                 /* assume CPU is at P0... */
 505                 perf->state = 0;
 506                 return perf->states[0].core_frequency * 1000;
 507         }
 508 }
 509
 510 static void free_acpi_perf_data(void)
 511 {
 512         unsigned int i;
 513
 514         /* Freeing a NULL pointer is OK, and alloc_percpu zeroes. */
 515         for_each_possible_cpu(i)
 516                 free_cpumask_var(per_cpu_ptr(acpi_perf_data, i)
 517                                  ->shared_cpu_map);
 518         free_percpu(acpi_perf_data);
 519 }
 520
 521 static int boost_notify(struct notifier_block *nb, unsigned long action,
 522                       void *hcpu)
 523 {
 524         unsigned cpu = (long)hcpu;
 525         const struct cpumask *cpumask;
 526
 527         cpumask = get_cpu_mask(cpu);
 528
 529         /*
 530          * Clear the boost-disable bit on the CPU_DOWN path so that
 531          * this cpu cannot block the remaining ones from boosting. On
 532          * the CPU_UP path we simply keep the boost-disable flag in
 533          * sync with the current global state.
 534          */
 535
 536         switch (action) {
 537         case CPU_UP_PREPARE:
 538         case CPU_UP_PREPARE_FROZEN:
 539                 boost_set_msrs(acpi_cpufreq_driver.boost_enabled, cpumask);
 540                 break;
 541
 542         case CPU_DOWN_PREPARE:
 543         case CPU_DOWN_PREPARE_FROZEN:
 544                 boost_set_msrs(1, cpumask);
 545                 break;
 546
 547         default:
 548                 break;
 549         }
 550
 551         return NOTIFY_OK;
 552 }
 553
 554
 555 static struct notifier_block boost_nb = {
 556         .notifier_call          = boost_notify,
 557 };
 558
 559 /*
 560  * acpi_cpufreq_early_init - initialize ACPI P-States library
 561  *
 562  * Initialize the ACPI P-States library (drivers/acpi/processor_perflib.c)
 563  * in order to determine correct frequency and voltage pairings. We can
 564  * do _PDC and _PSD and find out the processor dependency for the
 565  * actual init that will happen later...
 566  */
 567 static int __init acpi_cpufreq_early_init(void)
 568 {
 569         unsigned int i;
 570         pr_debug("acpi_cpufreq_early_init\n");
 571
 572         acpi_perf_data = alloc_percpu(struct acpi_processor_performance);
 573         if (!acpi_perf_data) {
 574                 pr_debug("Memory allocation error for acpi_perf_data.\n");
 575                 return -ENOMEM;
 576         }
 577         for_each_possible_cpu(i) {
 578                 if (!zalloc_cpumask_var_node(
 579                         &per_cpu_ptr(acpi_perf_data, i)->shared_cpu_map,
 580                         GFP_KERNEL, cpu_to_node(i))) {
 581
 582                         /* Freeing a NULL pointer is OK: alloc_percpu zeroes. */
 583                         free_acpi_perf_data();
 584                         return -ENOMEM;
 585                 }
 586         }
 587
 588         /* Do initialization in ACPI core */
 589         acpi_processor_preregister_performance(acpi_perf_data);
 590         return 0;
 591 }
 592
 593 #ifdef CONFIG_SMP
 594 /*
 595  * Some BIOSes do SW_ANY coordination internally, either set it up in hw
 596  * or do it in BIOS firmware and won't inform about it to OS. If not
 597  * detected, this has a side effect of making CPU run at a different speed
 598  * than OS intended it to run at. Detect it and handle it cleanly.
 599  */
 600 static int bios_with_sw_any_bug;
 601
 602 static int sw_any_bug_found(const struct dmi_system_id *d)
 603 {
 604         bios_with_sw_any_bug = 1;
 605         return 0;
 606 }
 607
 608 static const struct dmi_system_id sw_any_bug_dmi_table[] = {
 609         {
 610                 .callback = sw_any_bug_found,
 611                 .ident = "Supermicro Server X6DLP",
 612                 .matches = {
 613                         DMI_MATCH(DMI_SYS_VENDOR, "Supermicro"),
 614                         DMI_MATCH(DMI_BIOS_VERSION, "080010"),
 615                         DMI_MATCH(DMI_PRODUCT_NAME, "X6DLP"),
 616                 },
 617         },
 618         { }
 619 };
 620
 621 static int acpi_cpufreq_blacklist(struct cpuinfo_x86 *c)
 622 {
 623         /* Intel Xeon Processor 7100 Series Specification Update
 624          * http://www.intel.com/Assets/PDF/specupdate/314554.pdf
 625          * AL30: A Machine Check Exception (MCE) Occurring during an
 626          * Enhanced Intel SpeedStep Technology Ratio Change May Cause
 627          * Both Processor Cores to Lock Up. */
 628         if (c->x86_vendor == X86_VENDOR_INTEL) {
 629                 if ((c->x86 == 15) &&
 630                     (c->x86_model == 6) &&
 631                     (c->x86_mask == 8)) {
 632                         printk(KERN_INFO "acpi-cpufreq: Intel(R) "
 633                             "Xeon(R) 7100 Errata AL30, processors may "
 634                             "lock up on frequency changes: disabling "
 635                             "acpi-cpufreq.\n");
 636                         return -ENODEV;
 637                     }
 638                 }
 639         return 0;
 640 }
 641 #endif
 642
 643 static int acpi_cpufreq_cpu_init(struct cpufreq_policy *policy)
 644 {
 645         unsigned int i;
 646         unsigned int valid_states = 0;
 647         unsigned int cpu = policy->cpu;
 648         struct acpi_cpufreq_data *data;
 649         unsigned int result = 0;
 650         struct cpuinfo_x86 *c = &cpu_data(policy->cpu);
 651         struct acpi_processor_performance *perf;
 652 #ifdef CONFIG_SMP
 653         static int blacklisted;
 654 #endif
 655
 656         pr_debug("acpi_cpufreq_cpu_init\n");
 657
 658 #ifdef CONFIG_SMP
 659         if (blacklisted)
 660                 return blacklisted;
 661         blacklisted = acpi_cpufreq_blacklist(c);
 662         if (blacklisted)
 663                 return blacklisted;
 664 #endif
 665
 666         data = kzalloc(sizeof(*data), GFP_KERNEL);
 667         if (!data)
 668                 return -ENOMEM;
 669
 670         if (!zalloc_cpumask_var(&data->freqdomain_cpus, GFP_KERNEL)) {
 671                 result = -ENOMEM;
 672                 goto err_free;
 673         }
 674
 675         data->acpi_data = per_cpu_ptr(acpi_perf_data, cpu);
 676         per_cpu(acfreq_data, cpu) = data;
 677
 678         if (cpu_has(c, X86_FEATURE_CONSTANT_TSC))
 679                 acpi_cpufreq_driver.flags |= CPUFREQ_CONST_LOOPS;
 680
 681         result = acpi_processor_register_performance(data->acpi_data, cpu);
 682         if (result)
 683                 goto err_free_mask;
 684
 685         perf = data->acpi_data;
 686         policy->shared_type = perf->shared_type;
 687
 688         /*
 689          * Will let policy->cpus know about dependency only when software
 690          * coordination is required.
 691          */
 692         if (policy->shared_type == CPUFREQ_SHARED_TYPE_ALL ||
 693             policy->shared_type == CPUFREQ_SHARED_TYPE_ANY) {
 694                 cpumask_copy(policy->cpus, perf->shared_cpu_map);
 695         }
 696         cpumask_copy(data->freqdomain_cpus, perf->shared_cpu_map);
 697
 698 #ifdef CONFIG_SMP
 699         dmi_check_system(sw_any_bug_dmi_table);
 700         if (bios_with_sw_any_bug && !policy_is_shared(policy)) {
 701                 policy->shared_type = CPUFREQ_SHARED_TYPE_ALL;
 702                 cpumask_copy(policy->cpus, topology_core_cpumask(cpu));
 703         }
 704
 705         if (check_amd_hwpstate_cpu(cpu) && !acpi_pstate_strict) {
 706                 cpumask_clear(policy->cpus);
 707                 cpumask_set_cpu(cpu, policy->cpus);
 708                 cpumask_copy(data->freqdomain_cpus,
 709                              topology_sibling_cpumask(cpu));
 710                 policy->shared_type = CPUFREQ_SHARED_TYPE_HW;
 711                 pr_info_once(PFX "overriding BIOS provided _PSD data\n");
 712         }
 713 #endif
 714
 715         /* capability check */
 716         if (perf->state_count <= 1) {
 717                 pr_debug("No P-States\n");
 718                 result = -ENODEV;
 719                 goto err_unreg;
 720         }
 721
 722         if (perf->control_register.space_id != perf->status_register.space_id) {
 723                 result = -ENODEV;
 724                 goto err_unreg;
 725         }
 726
 727         switch (perf->control_register.space_id) {
 728         case ACPI_ADR_SPACE_SYSTEM_IO:
 729                 if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD &&
 730                     boot_cpu_data.x86 == 0xf) {
 731                         pr_debug("AMD K8 systems must use native drivers.\n");
 732                         result = -ENODEV;
 733                         goto err_unreg;
 734                 }
 735                 pr_debug("SYSTEM IO addr space\n");
 736                 data->cpu_feature = SYSTEM_IO_CAPABLE;
 737                 break;
 738         case ACPI_ADR_SPACE_FIXED_HARDWARE:
 739                 pr_debug("HARDWARE addr space\n");
 740                 if (check_est_cpu(cpu)) {
 741                         data->cpu_feature = SYSTEM_INTEL_MSR_CAPABLE;
 742                         break;
 743                 }
 744                 if (check_amd_hwpstate_cpu(cpu)) {
 745                         data->cpu_feature = SYSTEM_AMD_MSR_CAPABLE;
 746                         break;
 747                 }
 748                 result = -ENODEV;
 749                 goto err_unreg;
 750         default:
 751                 pr_debug("Unknown addr space %d\n",
 752                         (u32) (perf->control_register.space_id));
 753                 result = -ENODEV;
 754                 goto err_unreg;
 755         }
 756
 757         data->freq_table = kzalloc(sizeof(*data->freq_table) *
 758                     (perf->state_count+1), GFP_KERNEL);
 759         if (!data->freq_table) {
 760                 result = -ENOMEM;
 761                 goto err_unreg;
 762         }
 763
 764         /* detect transition latency */
 765         policy->cpuinfo.transition_latency = 0;
 766         for (i = 0; i < perf->state_count; i++) {
 767                 if ((perf->states[i].transition_latency * 1000) >
 768                     policy->cpuinfo.transition_latency)
 769                         policy->cpuinfo.transition_latency =
 770                             perf->states[i].transition_latency * 1000;
 771         }
 772
 773         /* Check for high latency (>20uS) from buggy BIOSes, like on T42 */
 774         if (perf->control_register.space_id == ACPI_ADR_SPACE_FIXED_HARDWARE &&
 775             policy->cpuinfo.transition_latency > 20 * 1000) {
 776                 policy->cpuinfo.transition_latency = 20 * 1000;
 777                 printk_once(KERN_INFO
 778                             "P-state transition latency capped at 20 uS\n");
 779         }
 780
 781         /* table init */
 782         for (i = 0; i < perf->state_count; i++) {
 783                 if (i > 0 && perf->states[i].core_frequency >=
 784                     data->freq_table[valid_states-1].frequency / 1000)
 785                         continue;
 786
 787                 data->freq_table[valid_states].driver_data = i;
 788                 data->freq_table[valid_states].frequency =
 789                     perf->states[i].core_frequency * 1000;
 790                 valid_states++;
 791         }
 792         data->freq_table[valid_states].frequency = CPUFREQ_TABLE_END;
 793         perf->state = 0;
 794
 795         result = cpufreq_table_validate_and_show(policy, data->freq_table);
 796         if (result)
 797                 goto err_freqfree;
 798
 799         if (perf->states[0].core_frequency * 1000 != policy->cpuinfo.max_freq)
 800                 printk(KERN_WARNING FW_WARN "P-state 0 is not max freq\n");
 801
 802         switch (perf->control_register.space_id) {
 803         case ACPI_ADR_SPACE_SYSTEM_IO:
 804                 /*
 805                  * The core will not set policy->cur, because
 806                  * cpufreq_driver->get is NULL, so we need to set it here.
 807                  * However, we have to guess it, because the current speed is
 808                  * unknown and not detectable via IO ports.
 809                  */
 810                 policy->cur = acpi_cpufreq_guess_freq(data, policy->cpu);
 811                 break;
 812         case ACPI_ADR_SPACE_FIXED_HARDWARE:
 813                 acpi_cpufreq_driver.get = get_cur_freq_on_cpu;
 814                 break;
 815         default:
 816                 break;
 817         }
 818
 819         /* notify BIOS that we exist */
 820         acpi_processor_notify_smm(THIS_MODULE);
 821
 822         pr_debug("CPU%u - ACPI performance management activated.\n", cpu);
 823         for (i = 0; i < perf->state_count; i++)
 824                 pr_debug("     %cP%d: %d MHz, %d mW, %d uS\n",
 825                         (i == perf->state ? '*' : ' '), i,
 826                         (u32) perf->states[i].core_frequency,
 827                         (u32) perf->states[i].power,
 828                         (u32) perf->states[i].transition_latency);
 829
 830         /*
 831          * the first call to ->target() should result in us actually
 832          * writing something to the appropriate registers.
 833          */
 834         data->resume = 1;
 835
 836         return result;
 837
 838 err_freqfree:
 839         kfree(data->freq_table);
 840 err_unreg:
 841         acpi_processor_unregister_performance(perf, cpu);
 842 err_free_mask:
 843         free_cpumask_var(data->freqdomain_cpus);
 844 err_free:
 845         kfree(data);
 846         per_cpu(acfreq_data, cpu) = NULL;
 847
 848         return result;
 849 }
 850
 851 static int acpi_cpufreq_cpu_exit(struct cpufreq_policy *policy)
 852 {
 853         struct acpi_cpufreq_data *data = per_cpu(acfreq_data, policy->cpu);
 854
 855         pr_debug("acpi_cpufreq_cpu_exit\n");
 856
 857         if (data) {
 858                 per_cpu(acfreq_data, policy->cpu) = NULL;
 859                 acpi_processor_unregister_performance(data->acpi_data,
 860                                                       policy->cpu);
 861                 free_cpumask_var(data->freqdomain_cpus);
 862                 kfree(data->freq_table);
 863                 kfree(data);
 864         }
 865
 866         return 0;
 867 }
 868
 869 static int acpi_cpufreq_resume(struct cpufreq_policy *policy)
 870 {
 871         struct acpi_cpufreq_data *data = per_cpu(acfreq_data, policy->cpu);
 872
 873         pr_debug("acpi_cpufreq_resume\n");
 874
 875         data->resume = 1;
 876
 877         return 0;
 878 }
 879
 880 static struct freq_attr *acpi_cpufreq_attr[] = {
 881         &cpufreq_freq_attr_scaling_available_freqs,
 882         &freqdomain_cpus,
 883         NULL,   /* this is a placeholder for cpb, do not remove */
 884         NULL,
 885 };
 886
 887 static struct cpufreq_driver acpi_cpufreq_driver = {
 888         .verify         = cpufreq_generic_frequency_table_verify,
 889         .target_index   = acpi_cpufreq_target,
 890         .bios_limit     = acpi_processor_get_bios_limit,
 891         .init           = acpi_cpufreq_cpu_init,
 892         .exit           = acpi_cpufreq_cpu_exit,
 893         .resume         = acpi_cpufreq_resume,
 894         .name           = "acpi-cpufreq",
 895         .attr           = acpi_cpufreq_attr,
 896         .set_boost      = _store_boost,
 897 };
 898
 899 static void __init acpi_cpufreq_boost_init(void)
 900 {
 901         if (boot_cpu_has(X86_FEATURE_CPB) || boot_cpu_has(X86_FEATURE_IDA)) {
 902                 msrs = msrs_alloc();
 903
 904                 if (!msrs)
 905                         return;
 906
 907                 acpi_cpufreq_driver.boost_supported = true;
 908                 acpi_cpufreq_driver.boost_enabled = boost_state(0);
 909
 910                 cpu_notifier_register_begin();
 911
 912                 /* Force all MSRs to the same value */
 913                 boost_set_msrs(acpi_cpufreq_driver.boost_enabled,
 914                                cpu_online_mask);
 915
 916                 __register_cpu_notifier(&boost_nb);
 917
 918                 cpu_notifier_register_done();
 919         }
 920 }
 921
 922 static void acpi_cpufreq_boost_exit(void)
 923 {
 924         if (msrs) {
 925                 unregister_cpu_notifier(&boost_nb);
 926
 927                 msrs_free(msrs);
 928                 msrs = NULL;
 929         }
 930 }
 931
 932 static int __init acpi_cpufreq_init(void)
 933 {
 934         int ret;
 935
 936         if (acpi_disabled)
 937                 return -ENODEV;
 938
 939         /* don't keep reloading if cpufreq_driver exists */
 940         if (cpufreq_get_current_driver())
 941                 return -EEXIST;
 942
 943         pr_debug("acpi_cpufreq_init\n");
 944
 945         ret = acpi_cpufreq_early_init();
 946         if (ret)
 947                 return ret;
 948
 949 #ifdef CONFIG_X86_ACPI_CPUFREQ_CPB
 950         /* this is a sysfs file with a strange name and an even stranger
 951          * semantic - per CPU instantiation, but system global effect.
 952          * Lets enable it only on AMD CPUs for compatibility reasons and
 953          * only if configured. This is considered legacy code, which
 954          * will probably be removed at some point in the future.
 955          */
 956         if (check_amd_hwpstate_cpu(0)) {
 957                 struct freq_attr **iter;
 958
 959                 pr_debug("adding sysfs entry for cpb\n");
 960
 961                 for (iter = acpi_cpufreq_attr; *iter != NULL; iter++)
 962                         ;
 963
 964                 /* make sure there is a terminator behind it */
 965                 if (iter[1] == NULL)
 966                         *iter = &cpb;
 967         }
 968 #endif
 969         acpi_cpufreq_boost_init();
 970
 971         ret = cpufreq_register_driver(&acpi_cpufreq_driver);
 972         if (ret) {
 973                 free_acpi_perf_data();
 974                 acpi_cpufreq_boost_exit();
 975         }
 976         return ret;
 977 }
 978
 979 static void __exit acpi_cpufreq_exit(void)
 980 {
 981         pr_debug("acpi_cpufreq_exit\n");
 982
 983         acpi_cpufreq_boost_exit();
 984
 985         cpufreq_unregister_driver(&acpi_cpufreq_driver);
 986
 987         free_acpi_perf_data();
 988 }
 989
 990 module_param(acpi_pstate_strict, uint, 0644);
 991 MODULE_PARM_DESC(acpi_pstate_strict,
 992         "value 0 or non-zero. non-zero -> strict ACPI checks are "
 993         "performed during frequency changes.");
 994
 995 late_initcall(acpi_cpufreq_init);
 996 module_exit(acpi_cpufreq_exit);
 997
 998 static const struct x86_cpu_id acpi_cpufreq_ids[] = {
 999         X86_FEATURE_MATCH(X86_FEATURE_ACPI),
1000         X86_FEATURE_MATCH(X86_FEATURE_HW_PSTATE),
1001         {}
1002 };
1003 MODULE_DEVICE_TABLE(x86cpu, acpi_cpufreq_ids);
1004
1005 static const struct acpi_device_id processor_device_ids[] = {
1006         {ACPI_PROCESSOR_OBJECT_HID, },
1007         {ACPI_PROCESSOR_DEVICE_HID, },
1008         {},
1009 };
1010 MODULE_DEVICE_TABLE(acpi, processor_device_ids);
1011
1012 MODULE_ALIAS("acpi");