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 #include "mperf.h"
  49
  50 MODULE_AUTHOR("Paul Diefenbaugh, Dominik Brodowski");
  51 MODULE_DESCRIPTION("ACPI Processor P-States Driver");
  52 MODULE_LICENSE("GPL");
  53
  54 #define PFX "acpi-cpufreq: "
  55
  56 enum {
  57         UNDEFINED_CAPABLE = 0,
  58         SYSTEM_INTEL_MSR_CAPABLE,
  59         SYSTEM_AMD_MSR_CAPABLE,
  60         SYSTEM_IO_CAPABLE,
  61 };
  62
  63 #define INTEL_MSR_RANGE         (0xffff)
  64 #define AMD_MSR_RANGE           (0x7)
  65
  66 #define MSR_K7_HWCR_CPB_DIS     (1ULL << 25)
  67
  68 struct acpi_cpufreq_data {
  69         struct acpi_processor_performance *acpi_data;
  70         struct cpufreq_frequency_table *freq_table;
  71         unsigned int resume;
  72         unsigned int cpu_feature;
  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 bool boost_enabled, boost_supported;
  84 static struct msr __percpu *msrs;
  85
  86 static bool boost_state(unsigned int cpu)
  87 {
  88         u32 lo, hi;
  89         u64 msr;
  90
  91         switch (boot_cpu_data.x86_vendor) {
  92         case X86_VENDOR_INTEL:
  93                 rdmsr_on_cpu(cpu, MSR_IA32_MISC_ENABLE, &lo, &hi);
  94                 msr = lo | ((u64)hi << 32);
  95                 return !(msr & MSR_IA32_MISC_ENABLE_TURBO_DISABLE);
  96         case X86_VENDOR_AMD:
  97                 rdmsr_on_cpu(cpu, MSR_K7_HWCR, &lo, &hi);
  98                 msr = lo | ((u64)hi << 32);
  99                 return !(msr & MSR_K7_HWCR_CPB_DIS);
 100         }
 101         return false;
 102 }
 103
 104 static void boost_set_msrs(bool enable, const struct cpumask *cpumask)
 105 {
 106         u32 cpu;
 107         u32 msr_addr;
 108         u64 msr_mask;
 109
 110         switch (boot_cpu_data.x86_vendor) {
 111         case X86_VENDOR_INTEL:
 112                 msr_addr = MSR_IA32_MISC_ENABLE;
 113                 msr_mask = MSR_IA32_MISC_ENABLE_TURBO_DISABLE;
 114                 break;
 115         case X86_VENDOR_AMD:
 116                 msr_addr = MSR_K7_HWCR;
 117                 msr_mask = MSR_K7_HWCR_CPB_DIS;
 118                 break;
 119         default:
 120                 return;
 121         }
 122
 123         rdmsr_on_cpus(cpumask, msr_addr, msrs);
 124
 125         for_each_cpu(cpu, cpumask) {
 126                 struct msr *reg = per_cpu_ptr(msrs, cpu);
 127                 if (enable)
 128                         reg->q &= ~msr_mask;
 129                 else
 130                         reg->q |= msr_mask;
 131         }
 132
 133         wrmsr_on_cpus(cpumask, msr_addr, msrs);
 134 }
 135
 136 static ssize_t _store_boost(const char *buf, size_t count)
 137 {
 138         int ret;
 139         unsigned long val = 0;
 140
 141         if (!boost_supported)
 142                 return -EINVAL;
 143
 144         ret = kstrtoul(buf, 10, &val);
 145         if (ret || (val > 1))
 146                 return -EINVAL;
 147
 148         if ((val && boost_enabled) || (!val && !boost_enabled))
 149                 return count;
 150
 151         get_online_cpus();
 152
 153         boost_set_msrs(val, cpu_online_mask);
 154
 155         put_online_cpus();
 156
 157         boost_enabled = val;
 158         pr_debug("Core Boosting %sabled.\n", val ? "en" : "dis");
 159
 160         return count;
 161 }
 162
 163 static ssize_t store_global_boost(struct kobject *kobj, struct attribute *attr,
 164                                   const char *buf, size_t count)
 165 {
 166         return _store_boost(buf, count);
 167 }
 168
 169 static ssize_t show_global_boost(struct kobject *kobj,
 170                                  struct attribute *attr, char *buf)
 171 {
 172         return sprintf(buf, "%u\n", boost_enabled);
 173 }
 174
 175 static struct global_attr global_boost = __ATTR(boost, 0644,
 176                                                 show_global_boost,
 177                                                 store_global_boost);
 178
 179 #ifdef CONFIG_X86_ACPI_CPUFREQ_CPB
 180 static ssize_t store_cpb(struct cpufreq_policy *policy, const char *buf,
 181                          size_t count)
 182 {
 183         return _store_boost(buf, count);
 184 }
 185
 186 static ssize_t show_cpb(struct cpufreq_policy *policy, char *buf)
 187 {
 188         return sprintf(buf, "%u\n", boost_enabled);
 189 }
 190
 191 static struct freq_attr cpb = __ATTR(cpb, 0644, show_cpb, store_cpb);
 192 #endif
 193
 194 static int check_est_cpu(unsigned int cpuid)
 195 {
 196         struct cpuinfo_x86 *cpu = &cpu_data(cpuid);
 197
 198         return cpu_has(cpu, X86_FEATURE_EST);
 199 }
 200
 201 static int check_amd_hwpstate_cpu(unsigned int cpuid)
 202 {
 203         struct cpuinfo_x86 *cpu = &cpu_data(cpuid);
 204
 205         return cpu_has(cpu, X86_FEATURE_HW_PSTATE);
 206 }
 207
 208 static unsigned extract_io(u32 value, struct acpi_cpufreq_data *data)
 209 {
 210         struct acpi_processor_performance *perf;
 211         int i;
 212
 213         perf = data->acpi_data;
 214
 215         for (i = 0; i < perf->state_count; i++) {
 216                 if (value == perf->states[i].status)
 217                         return data->freq_table[i].frequency;
 218         }
 219         return 0;
 220 }
 221
 222 static unsigned extract_msr(u32 msr, struct acpi_cpufreq_data *data)
 223 {
 224         int i;
 225         struct acpi_processor_performance *perf;
 226
 227         if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD)
 228                 msr &= AMD_MSR_RANGE;
 229         else
 230                 msr &= INTEL_MSR_RANGE;
 231
 232         perf = data->acpi_data;
 233
 234         for (i = 0; data->freq_table[i].frequency != CPUFREQ_TABLE_END; i++) {
 235                 if (msr == perf->states[data->freq_table[i].index].status)
 236                         return data->freq_table[i].frequency;
 237         }
 238         return data->freq_table[0].frequency;
 239 }
 240
 241 static unsigned extract_freq(u32 val, struct acpi_cpufreq_data *data)
 242 {
 243         switch (data->cpu_feature) {
 244         case SYSTEM_INTEL_MSR_CAPABLE:
 245         case SYSTEM_AMD_MSR_CAPABLE:
 246                 return extract_msr(val, data);
 247         case SYSTEM_IO_CAPABLE:
 248                 return extract_io(val, data);
 249         default:
 250                 return 0;
 251         }
 252 }
 253
 254 struct msr_addr {
 255         u32 reg;
 256 };
 257
 258 struct io_addr {
 259         u16 port;
 260         u8 bit_width;
 261 };
 262
 263 struct drv_cmd {
 264         unsigned int type;
 265         const struct cpumask *mask;
 266         union {
 267                 struct msr_addr msr;
 268                 struct io_addr io;
 269         } addr;
 270         u32 val;
 271 };
 272
 273 /* Called via smp_call_function_single(), on the target CPU */
 274 static void do_drv_read(void *_cmd)
 275 {
 276         struct drv_cmd *cmd = _cmd;
 277         u32 h;
 278
 279         switch (cmd->type) {
 280         case SYSTEM_INTEL_MSR_CAPABLE:
 281         case SYSTEM_AMD_MSR_CAPABLE:
 282                 rdmsr(cmd->addr.msr.reg, cmd->val, h);
 283                 break;
 284         case SYSTEM_IO_CAPABLE:
 285                 acpi_os_read_port((acpi_io_address)cmd->addr.io.port,
 286                                 &cmd->val,
 287                                 (u32)cmd->addr.io.bit_width);
 288                 break;
 289         default:
 290                 break;
 291         }
 292 }
 293
 294 /* Called via smp_call_function_many(), on the target CPUs */
 295 static void do_drv_write(void *_cmd)
 296 {
 297         struct drv_cmd *cmd = _cmd;
 298         u32 lo, hi;
 299
 300         switch (cmd->type) {
 301         case SYSTEM_INTEL_MSR_CAPABLE:
 302                 rdmsr(cmd->addr.msr.reg, lo, hi);
 303                 lo = (lo & ~INTEL_MSR_RANGE) | (cmd->val & INTEL_MSR_RANGE);
 304                 wrmsr(cmd->addr.msr.reg, lo, hi);
 305                 break;
 306         case SYSTEM_AMD_MSR_CAPABLE:
 307                 wrmsr(cmd->addr.msr.reg, cmd->val, 0);
 308                 break;
 309         case SYSTEM_IO_CAPABLE:
 310                 acpi_os_write_port((acpi_io_address)cmd->addr.io.port,
 311                                 cmd->val,
 312                                 (u32)cmd->addr.io.bit_width);
 313                 break;
 314         default:
 315                 break;
 316         }
 317 }
 318
 319 static void drv_read(struct drv_cmd *cmd)
 320 {
 321         int err;
 322         cmd->val = 0;
 323
 324         err = smp_call_function_any(cmd->mask, do_drv_read, cmd, 1);
 325         WARN_ON_ONCE(err);      /* smp_call_function_any() was buggy? */
 326 }
 327
 328 static void drv_write(struct drv_cmd *cmd)
 329 {
 330         int this_cpu;
 331
 332         this_cpu = get_cpu();
 333         if (cpumask_test_cpu(this_cpu, cmd->mask))
 334                 do_drv_write(cmd);
 335         smp_call_function_many(cmd->mask, do_drv_write, cmd, 1);
 336         put_cpu();
 337 }
 338
 339 static u32 get_cur_val(const struct cpumask *mask)
 340 {
 341         struct acpi_processor_performance *perf;
 342         struct drv_cmd cmd;
 343
 344         if (unlikely(cpumask_empty(mask)))
 345                 return 0;
 346
 347         switch (per_cpu(acfreq_data, cpumask_first(mask))->cpu_feature) {
 348         case SYSTEM_INTEL_MSR_CAPABLE:
 349                 cmd.type = SYSTEM_INTEL_MSR_CAPABLE;
 350                 cmd.addr.msr.reg = MSR_IA32_PERF_CTL;
 351                 break;
 352         case SYSTEM_AMD_MSR_CAPABLE:
 353                 cmd.type = SYSTEM_AMD_MSR_CAPABLE;
 354                 cmd.addr.msr.reg = MSR_AMD_PERF_CTL;
 355                 break;
 356         case SYSTEM_IO_CAPABLE:
 357                 cmd.type = SYSTEM_IO_CAPABLE;
 358                 perf = per_cpu(acfreq_data, cpumask_first(mask))->acpi_data;
 359                 cmd.addr.io.port = perf->control_register.address;
 360                 cmd.addr.io.bit_width = perf->control_register.bit_width;
 361                 break;
 362         default:
 363                 return 0;
 364         }
 365
 366         cmd.mask = mask;
 367         drv_read(&cmd);
 368
 369         pr_debug("get_cur_val = %u\n", cmd.val);
 370
 371         return cmd.val;
 372 }
 373
 374 static unsigned int get_cur_freq_on_cpu(unsigned int cpu)
 375 {
 376         struct acpi_cpufreq_data *data = per_cpu(acfreq_data, cpu);
 377         unsigned int freq;
 378         unsigned int cached_freq;
 379
 380         pr_debug("get_cur_freq_on_cpu (%d)\n", cpu);
 381
 382         if (unlikely(data == NULL ||
 383                      data->acpi_data == NULL || data->freq_table == NULL)) {
 384                 return 0;
 385         }
 386
 387         cached_freq = data->freq_table[data->acpi_data->state].frequency;
 388         freq = extract_freq(get_cur_val(cpumask_of(cpu)), data);
 389         if (freq != cached_freq) {
 390                 /*
 391                  * The dreaded BIOS frequency change behind our back.
 392                  * Force set the frequency on next target call.
 393                  */
 394                 data->resume = 1;
 395         }
 396
 397         pr_debug("cur freq = %u\n", freq);
 398
 399         return freq;
 400 }
 401
 402 static unsigned int check_freqs(const struct cpumask *mask, unsigned int freq,
 403                                 struct acpi_cpufreq_data *data)
 404 {
 405         unsigned int cur_freq;
 406         unsigned int i;
 407
 408         for (i = 0; i < 100; i++) {
 409                 cur_freq = extract_freq(get_cur_val(mask), data);
 410                 if (cur_freq == freq)
 411                         return 1;
 412                 udelay(10);
 413         }
 414         return 0;
 415 }
 416
 417 static int acpi_cpufreq_target(struct cpufreq_policy *policy,
 418                                unsigned int target_freq, unsigned int relation)
 419 {
 420         struct acpi_cpufreq_data *data = per_cpu(acfreq_data, policy->cpu);
 421         struct acpi_processor_performance *perf;
 422         struct cpufreq_freqs freqs;
 423         struct drv_cmd cmd;
 424         unsigned int next_state = 0; /* Index into freq_table */
 425         unsigned int next_perf_state = 0; /* Index into perf table */
 426         int result = 0;
 427
 428         pr_debug("acpi_cpufreq_target %d (%d)\n", target_freq, policy->cpu);
 429
 430         if (unlikely(data == NULL ||
 431              data->acpi_data == NULL || data->freq_table == NULL)) {
 432                 return -ENODEV;
 433         }
 434
 435         perf = data->acpi_data;
 436         result = cpufreq_frequency_table_target(policy,
 437                                                 data->freq_table,
 438                                                 target_freq,
 439                                                 relation, &next_state);
 440         if (unlikely(result)) {
 441                 result = -ENODEV;
 442                 goto out;
 443         }
 444
 445         next_perf_state = data->freq_table[next_state].index;
 446         if (perf->state == next_perf_state) {
 447                 if (unlikely(data->resume)) {
 448                         pr_debug("Called after resume, resetting to P%d\n",
 449                                 next_perf_state);
 450                         data->resume = 0;
 451                 } else {
 452                         pr_debug("Already at target state (P%d)\n",
 453                                 next_perf_state);
 454                         goto out;
 455                 }
 456         }
 457
 458         switch (data->cpu_feature) {
 459         case SYSTEM_INTEL_MSR_CAPABLE:
 460                 cmd.type = SYSTEM_INTEL_MSR_CAPABLE;
 461                 cmd.addr.msr.reg = MSR_IA32_PERF_CTL;
 462                 cmd.val = (u32) perf->states[next_perf_state].control;
 463                 break;
 464         case SYSTEM_AMD_MSR_CAPABLE:
 465                 cmd.type = SYSTEM_AMD_MSR_CAPABLE;
 466                 cmd.addr.msr.reg = MSR_AMD_PERF_CTL;
 467                 cmd.val = (u32) perf->states[next_perf_state].control;
 468                 break;
 469         case SYSTEM_IO_CAPABLE:
 470                 cmd.type = SYSTEM_IO_CAPABLE;
 471                 cmd.addr.io.port = perf->control_register.address;
 472                 cmd.addr.io.bit_width = perf->control_register.bit_width;
 473                 cmd.val = (u32) perf->states[next_perf_state].control;
 474                 break;
 475         default:
 476                 result = -ENODEV;
 477                 goto out;
 478         }
 479
 480         /* cpufreq holds the hotplug lock, so we are safe from here on */
 481         if (policy->shared_type != CPUFREQ_SHARED_TYPE_ANY)
 482                 cmd.mask = policy->cpus;
 483         else
 484                 cmd.mask = cpumask_of(policy->cpu);
 485
 486         freqs.old = perf->states[perf->state].core_frequency * 1000;
 487         freqs.new = data->freq_table[next_state].frequency;
 488         cpufreq_notify_transition(policy, &freqs, CPUFREQ_PRECHANGE);
 489
 490         drv_write(&cmd);
 491
 492         if (acpi_pstate_strict) {
 493                 if (!check_freqs(cmd.mask, freqs.new, data)) {
 494                         pr_debug("acpi_cpufreq_target failed (%d)\n",
 495                                 policy->cpu);
 496                         result = -EAGAIN;
 497                         goto out;
 498                 }
 499         }
 500
 501         cpufreq_notify_transition(policy, &freqs, CPUFREQ_POSTCHANGE);
 502         perf->state = next_perf_state;
 503
 504 out:
 505         return result;
 506 }
 507
 508 static int acpi_cpufreq_verify(struct cpufreq_policy *policy)
 509 {
 510         struct acpi_cpufreq_data *data = per_cpu(acfreq_data, policy->cpu);
 511
 512         pr_debug("acpi_cpufreq_verify\n");
 513
 514         return cpufreq_frequency_table_verify(policy, data->freq_table);
 515 }
 516
 517 static unsigned long
 518 acpi_cpufreq_guess_freq(struct acpi_cpufreq_data *data, unsigned int cpu)
 519 {
 520         struct acpi_processor_performance *perf = data->acpi_data;
 521
 522         if (cpu_khz) {
 523                 /* search the closest match to cpu_khz */
 524                 unsigned int i;
 525                 unsigned long freq;
 526                 unsigned long freqn = perf->states[0].core_frequency * 1000;
 527
 528                 for (i = 0; i < (perf->state_count-1); i++) {
 529                         freq = freqn;
 530                         freqn = perf->states[i+1].core_frequency * 1000;
 531                         if ((2 * cpu_khz) > (freqn + freq)) {
 532                                 perf->state = i;
 533                                 return freq;
 534                         }
 535                 }
 536                 perf->state = perf->state_count-1;
 537                 return freqn;
 538         } else {
 539                 /* assume CPU is at P0... */
 540                 perf->state = 0;
 541                 return perf->states[0].core_frequency * 1000;
 542         }
 543 }
 544
 545 static void free_acpi_perf_data(void)
 546 {
 547         unsigned int i;
 548
 549         /* Freeing a NULL pointer is OK, and alloc_percpu zeroes. */
 550         for_each_possible_cpu(i)
 551                 free_cpumask_var(per_cpu_ptr(acpi_perf_data, i)
 552                                  ->shared_cpu_map);
 553         free_percpu(acpi_perf_data);
 554 }
 555
 556 static int boost_notify(struct notifier_block *nb, unsigned long action,
 557                       void *hcpu)
 558 {
 559         unsigned cpu = (long)hcpu;
 560         const struct cpumask *cpumask;
 561
 562         cpumask = get_cpu_mask(cpu);
 563
 564         /*
 565          * Clear the boost-disable bit on the CPU_DOWN path so that
 566          * this cpu cannot block the remaining ones from boosting. On
 567          * the CPU_UP path we simply keep the boost-disable flag in
 568          * sync with the current global state.
 569          */
 570
 571         switch (action) {
 572         case CPU_UP_PREPARE:
 573         case CPU_UP_PREPARE_FROZEN:
 574                 boost_set_msrs(boost_enabled, cpumask);
 575                 break;
 576
 577         case CPU_DOWN_PREPARE:
 578         case CPU_DOWN_PREPARE_FROZEN:
 579                 boost_set_msrs(1, cpumask);
 580                 break;
 581
 582         default:
 583                 break;
 584         }
 585
 586         return NOTIFY_OK;
 587 }
 588
 589
 590 static struct notifier_block boost_nb = {
 591         .notifier_call          = boost_notify,
 592 };
 593
 594 /*
 595  * acpi_cpufreq_early_init - initialize ACPI P-States library
 596  *
 597  * Initialize the ACPI P-States library (drivers/acpi/processor_perflib.c)
 598  * in order to determine correct frequency and voltage pairings. We can
 599  * do _PDC and _PSD and find out the processor dependency for the
 600  * actual init that will happen later...
 601  */
 602 static int __init acpi_cpufreq_early_init(void)
 603 {
 604         unsigned int i;
 605         pr_debug("acpi_cpufreq_early_init\n");
 606
 607         acpi_perf_data = alloc_percpu(struct acpi_processor_performance);
 608         if (!acpi_perf_data) {
 609                 pr_debug("Memory allocation error for acpi_perf_data.\n");
 610                 return -ENOMEM;
 611         }
 612         for_each_possible_cpu(i) {
 613                 if (!zalloc_cpumask_var_node(
 614                         &per_cpu_ptr(acpi_perf_data, i)->shared_cpu_map,
 615                         GFP_KERNEL, cpu_to_node(i))) {
 616
 617                         /* Freeing a NULL pointer is OK: alloc_percpu zeroes. */
 618                         free_acpi_perf_data();
 619                         return -ENOMEM;
 620                 }
 621         }
 622
 623         /* Do initialization in ACPI core */
 624         acpi_processor_preregister_performance(acpi_perf_data);
 625         return 0;
 626 }
 627
 628 #ifdef CONFIG_SMP
 629 /*
 630  * Some BIOSes do SW_ANY coordination internally, either set it up in hw
 631  * or do it in BIOS firmware and won't inform about it to OS. If not
 632  * detected, this has a side effect of making CPU run at a different speed
 633  * than OS intended it to run at. Detect it and handle it cleanly.
 634  */
 635 static int bios_with_sw_any_bug;
 636
 637 static int sw_any_bug_found(const struct dmi_system_id *d)
 638 {
 639         bios_with_sw_any_bug = 1;
 640         return 0;
 641 }
 642
 643 static const struct dmi_system_id sw_any_bug_dmi_table[] = {
 644         {
 645                 .callback = sw_any_bug_found,
 646                 .ident = "Supermicro Server X6DLP",
 647                 .matches = {
 648                         DMI_MATCH(DMI_SYS_VENDOR, "Supermicro"),
 649                         DMI_MATCH(DMI_BIOS_VERSION, "080010"),
 650                         DMI_MATCH(DMI_PRODUCT_NAME, "X6DLP"),
 651                 },
 652         },
 653         { }
 654 };
 655
 656 static int acpi_cpufreq_blacklist(struct cpuinfo_x86 *c)
 657 {
 658         /* Intel Xeon Processor 7100 Series Specification Update
 659          * http://www.intel.com/Assets/PDF/specupdate/314554.pdf
 660          * AL30: A Machine Check Exception (MCE) Occurring during an
 661          * Enhanced Intel SpeedStep Technology Ratio Change May Cause
 662          * Both Processor Cores to Lock Up. */
 663         if (c->x86_vendor == X86_VENDOR_INTEL) {
 664                 if ((c->x86 == 15) &&
 665                     (c->x86_model == 6) &&
 666                     (c->x86_mask == 8)) {
 667                         printk(KERN_INFO "acpi-cpufreq: Intel(R) "
 668                             "Xeon(R) 7100 Errata AL30, processors may "
 669                             "lock up on frequency changes: disabling "
 670                             "acpi-cpufreq.\n");
 671                         return -ENODEV;
 672                     }
 673                 }
 674         return 0;
 675 }
 676 #endif
 677
 678 static int acpi_cpufreq_cpu_init(struct cpufreq_policy *policy)
 679 {
 680         unsigned int i;
 681         unsigned int valid_states = 0;
 682         unsigned int cpu = policy->cpu;
 683         struct acpi_cpufreq_data *data;
 684         unsigned int result = 0;
 685         struct cpuinfo_x86 *c = &cpu_data(policy->cpu);
 686         struct acpi_processor_performance *perf;
 687 #ifdef CONFIG_SMP
 688         static int blacklisted;
 689 #endif
 690
 691         pr_debug("acpi_cpufreq_cpu_init\n");
 692
 693 #ifdef CONFIG_SMP
 694         if (blacklisted)
 695                 return blacklisted;
 696         blacklisted = acpi_cpufreq_blacklist(c);
 697         if (blacklisted)
 698                 return blacklisted;
 699 #endif
 700
 701         data = kzalloc(sizeof(struct acpi_cpufreq_data), GFP_KERNEL);
 702         if (!data)
 703                 return -ENOMEM;
 704
 705         data->acpi_data = per_cpu_ptr(acpi_perf_data, cpu);
 706         per_cpu(acfreq_data, cpu) = data;
 707
 708         if (cpu_has(c, X86_FEATURE_CONSTANT_TSC))
 709                 acpi_cpufreq_driver.flags |= CPUFREQ_CONST_LOOPS;
 710
 711         result = acpi_processor_register_performance(data->acpi_data, cpu);
 712         if (result)
 713                 goto err_free;
 714
 715         perf = data->acpi_data;
 716         policy->shared_type = perf->shared_type;
 717
 718         /*
 719          * Will let policy->cpus know about dependency only when software
 720          * coordination is required.
 721          */
 722         if (policy->shared_type == CPUFREQ_SHARED_TYPE_ALL ||
 723             policy->shared_type == CPUFREQ_SHARED_TYPE_ANY) {
 724                 cpumask_copy(policy->cpus, perf->shared_cpu_map);
 725         }
 726
 727 #ifdef CONFIG_SMP
 728         dmi_check_system(sw_any_bug_dmi_table);
 729         if (bios_with_sw_any_bug && !policy_is_shared(policy)) {
 730                 policy->shared_type = CPUFREQ_SHARED_TYPE_ALL;
 731                 cpumask_copy(policy->cpus, cpu_core_mask(cpu));
 732         }
 733
 734         if (check_amd_hwpstate_cpu(cpu) && !acpi_pstate_strict) {
 735                 cpumask_clear(policy->cpus);
 736                 cpumask_set_cpu(cpu, policy->cpus);
 737                 policy->shared_type = CPUFREQ_SHARED_TYPE_HW;
 738                 pr_info_once(PFX "overriding BIOS provided _PSD data\n");
 739         }
 740 #endif
 741
 742         /* capability check */
 743         if (perf->state_count <= 1) {
 744                 pr_debug("No P-States\n");
 745                 result = -ENODEV;
 746                 goto err_unreg;
 747         }
 748
 749         if (perf->control_register.space_id != perf->status_register.space_id) {
 750                 result = -ENODEV;
 751                 goto err_unreg;
 752         }
 753
 754         switch (perf->control_register.space_id) {
 755         case ACPI_ADR_SPACE_SYSTEM_IO:
 756                 if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD &&
 757                     boot_cpu_data.x86 == 0xf) {
 758                         pr_debug("AMD K8 systems must use native drivers.\n");
 759                         result = -ENODEV;
 760                         goto err_unreg;
 761                 }
 762                 pr_debug("SYSTEM IO addr space\n");
 763                 data->cpu_feature = SYSTEM_IO_CAPABLE;
 764                 break;
 765         case ACPI_ADR_SPACE_FIXED_HARDWARE:
 766                 pr_debug("HARDWARE addr space\n");
 767                 if (check_est_cpu(cpu)) {
 768                         data->cpu_feature = SYSTEM_INTEL_MSR_CAPABLE;
 769                         break;
 770                 }
 771                 if (check_amd_hwpstate_cpu(cpu)) {
 772                         data->cpu_feature = SYSTEM_AMD_MSR_CAPABLE;
 773                         break;
 774                 }
 775                 result = -ENODEV;
 776                 goto err_unreg;
 777         default:
 778                 pr_debug("Unknown addr space %d\n",
 779                         (u32) (perf->control_register.space_id));
 780                 result = -ENODEV;
 781                 goto err_unreg;
 782         }
 783
 784         data->freq_table = kmalloc(sizeof(struct cpufreq_frequency_table) *
 785                     (perf->state_count+1), GFP_KERNEL);
 786         if (!data->freq_table) {
 787                 result = -ENOMEM;
 788                 goto err_unreg;
 789         }
 790
 791         /* detect transition latency */
 792         policy->cpuinfo.transition_latency = 0;
 793         for (i = 0; i < perf->state_count; i++) {
 794                 if ((perf->states[i].transition_latency * 1000) >
 795                     policy->cpuinfo.transition_latency)
 796                         policy->cpuinfo.transition_latency =
 797                             perf->states[i].transition_latency * 1000;
 798         }
 799
 800         /* Check for high latency (>20uS) from buggy BIOSes, like on T42 */
 801         if (perf->control_register.space_id == ACPI_ADR_SPACE_FIXED_HARDWARE &&
 802             policy->cpuinfo.transition_latency > 20 * 1000) {
 803                 policy->cpuinfo.transition_latency = 20 * 1000;
 804                 printk_once(KERN_INFO
 805                             "P-state transition latency capped at 20 uS\n");
 806         }
 807
 808         /* table init */
 809         for (i = 0; i < perf->state_count; i++) {
 810                 if (i > 0 && perf->states[i].core_frequency >=
 811                     data->freq_table[valid_states-1].frequency / 1000)
 812                         continue;
 813
 814                 data->freq_table[valid_states].index = i;
 815                 data->freq_table[valid_states].frequency =
 816                     perf->states[i].core_frequency * 1000;
 817                 valid_states++;
 818         }
 819         data->freq_table[valid_states].frequency = CPUFREQ_TABLE_END;
 820         perf->state = 0;
 821
 822         result = cpufreq_frequency_table_cpuinfo(policy, data->freq_table);
 823         if (result)
 824                 goto err_freqfree;
 825
 826         if (perf->states[0].core_frequency * 1000 != policy->cpuinfo.max_freq)
 827                 printk(KERN_WARNING FW_WARN "P-state 0 is not max freq\n");
 828
 829         switch (perf->control_register.space_id) {
 830         case ACPI_ADR_SPACE_SYSTEM_IO:
 831                 /* Current speed is unknown and not detectable by IO port */
 832                 policy->cur = acpi_cpufreq_guess_freq(data, policy->cpu);
 833                 break;
 834         case ACPI_ADR_SPACE_FIXED_HARDWARE:
 835                 acpi_cpufreq_driver.get = get_cur_freq_on_cpu;
 836                 policy->cur = get_cur_freq_on_cpu(cpu);
 837                 break;
 838         default:
 839                 break;
 840         }
 841
 842         /* notify BIOS that we exist */
 843         acpi_processor_notify_smm(THIS_MODULE);
 844
 845         /* Check for APERF/MPERF support in hardware */
 846         if (boot_cpu_has(X86_FEATURE_APERFMPERF))
 847                 acpi_cpufreq_driver.getavg = cpufreq_get_measured_perf;
 848
 849         pr_debug("CPU%u - ACPI performance management activated.\n", cpu);
 850         for (i = 0; i < perf->state_count; i++)
 851                 pr_debug("     %cP%d: %d MHz, %d mW, %d uS\n",
 852                         (i == perf->state ? '*' : ' '), i,
 853                         (u32) perf->states[i].core_frequency,
 854                         (u32) perf->states[i].power,
 855                         (u32) perf->states[i].transition_latency);
 856
 857         cpufreq_frequency_table_get_attr(data->freq_table, policy->cpu);
 858
 859         /*
 860          * the first call to ->target() should result in us actually
 861          * writing something to the appropriate registers.
 862          */
 863         data->resume = 1;
 864
 865         return result;
 866
 867 err_freqfree:
 868         kfree(data->freq_table);
 869 err_unreg:
 870         acpi_processor_unregister_performance(perf, cpu);
 871 err_free:
 872         kfree(data);
 873         per_cpu(acfreq_data, cpu) = NULL;
 874
 875         return result;
 876 }
 877
 878 static int acpi_cpufreq_cpu_exit(struct cpufreq_policy *policy)
 879 {
 880         struct acpi_cpufreq_data *data = per_cpu(acfreq_data, policy->cpu);
 881
 882         pr_debug("acpi_cpufreq_cpu_exit\n");
 883
 884         if (data) {
 885                 cpufreq_frequency_table_put_attr(policy->cpu);
 886                 per_cpu(acfreq_data, policy->cpu) = NULL;
 887                 acpi_processor_unregister_performance(data->acpi_data,
 888                                                       policy->cpu);
 889                 kfree(data->freq_table);
 890                 kfree(data);
 891         }
 892
 893         return 0;
 894 }
 895
 896 static int acpi_cpufreq_resume(struct cpufreq_policy *policy)
 897 {
 898         struct acpi_cpufreq_data *data = per_cpu(acfreq_data, policy->cpu);
 899
 900         pr_debug("acpi_cpufreq_resume\n");
 901
 902         data->resume = 1;
 903
 904         return 0;
 905 }
 906
 907 static struct freq_attr *acpi_cpufreq_attr[] = {
 908         &cpufreq_freq_attr_scaling_available_freqs,
 909         NULL,   /* this is a placeholder for cpb, do not remove */
 910         NULL,
 911 };
 912
 913 static struct cpufreq_driver acpi_cpufreq_driver = {
 914         .verify         = acpi_cpufreq_verify,
 915         .target         = acpi_cpufreq_target,
 916         .bios_limit     = acpi_processor_get_bios_limit,
 917         .init           = acpi_cpufreq_cpu_init,
 918         .exit           = acpi_cpufreq_cpu_exit,
 919         .resume         = acpi_cpufreq_resume,
 920         .name           = "acpi-cpufreq",
 921         .owner          = THIS_MODULE,
 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                 boost_supported = true;
 934                 boost_enabled = boost_state(0);
 935
 936                 get_online_cpus();
 937
 938                 /* Force all MSRs to the same value */
 939                 boost_set_msrs(boost_enabled, cpu_online_mask);
 940
 941                 register_cpu_notifier(&boost_nb);
 942
 943                 put_online_cpus();
 944         } else
 945                 global_boost.attr.mode = 0444;
 946
 947         /* We create the boost file in any case, though for systems without
 948          * hardware support it will be read-only and hardwired to return 0.
 949          */
 950         if (sysfs_create_file(cpufreq_global_kobject, &(global_boost.attr)))
 951                 pr_warn(PFX "could not register global boost sysfs file\n");
 952         else
 953                 pr_debug("registered global boost sysfs file\n");
 954 }
 955
 956 static void __exit acpi_cpufreq_boost_exit(void)
 957 {
 958         sysfs_remove_file(cpufreq_global_kobject, &(global_boost.attr));
 959
 960         if (msrs) {
 961                 unregister_cpu_notifier(&boost_nb);
 962
 963                 msrs_free(msrs);
 964                 msrs = NULL;
 965         }
 966 }
 967
 968 static int __init acpi_cpufreq_init(void)
 969 {
 970         int ret;
 971
 972         if (acpi_disabled)
 973                 return 0;
 974
 975         pr_debug("acpi_cpufreq_init\n");
 976
 977         ret = acpi_cpufreq_early_init();
 978         if (ret)
 979                 return ret;
 980
 981 #ifdef CONFIG_X86_ACPI_CPUFREQ_CPB
 982         /* this is a sysfs file with a strange name and an even stranger
 983          * semantic - per CPU instantiation, but system global effect.
 984          * Lets enable it only on AMD CPUs for compatibility reasons and
 985          * only if configured. This is considered legacy code, which
 986          * will probably be removed at some point in the future.
 987          */
 988         if (check_amd_hwpstate_cpu(0)) {
 989                 struct freq_attr **iter;
 990
 991                 pr_debug("adding sysfs entry for cpb\n");
 992
 993                 for (iter = acpi_cpufreq_attr; *iter != NULL; iter++)
 994                         ;
 995
 996                 /* make sure there is a terminator behind it */
 997                 if (iter[1] == NULL)
 998                         *iter = &cpb;
 999         }
1000 #endif
1001
1002         ret = cpufreq_register_driver(&acpi_cpufreq_driver);
1003         if (ret)
1004                 free_acpi_perf_data();
1005         else
1006                 acpi_cpufreq_boost_init();
1007
1008         return ret;
1009 }
1010
1011 static void __exit acpi_cpufreq_exit(void)
1012 {
1013         pr_debug("acpi_cpufreq_exit\n");
1014
1015         acpi_cpufreq_boost_exit();
1016
1017         cpufreq_unregister_driver(&acpi_cpufreq_driver);
1018
1019         free_acpi_perf_data();
1020 }
1021
1022 module_param(acpi_pstate_strict, uint, 0644);
1023 MODULE_PARM_DESC(acpi_pstate_strict,
1024         "value 0 or non-zero. non-zero -> strict ACPI checks are "
1025         "performed during frequency changes.");
1026
1027 late_initcall(acpi_cpufreq_init);
1028 module_exit(acpi_cpufreq_exit);
1029
1030 static const struct x86_cpu_id acpi_cpufreq_ids[] = {
1031         X86_FEATURE_MATCH(X86_FEATURE_ACPI),
1032         X86_FEATURE_MATCH(X86_FEATURE_HW_PSTATE),
1033         {}
1034 };
1035 MODULE_DEVICE_TABLE(x86cpu, acpi_cpufreq_ids);
1036
1037 MODULE_ALIAS("acpi");