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 };
  74
  75 /* acpi_perf_data is a pointer to percpu data. */
  76 static struct acpi_processor_performance __percpu *acpi_perf_data;
  77
  78 static inline struct acpi_processor_performance *to_perf_data(struct acpi_cpufreq_data *data)
  79 {
  80         return per_cpu_ptr(acpi_perf_data, data->acpi_perf_cpu);
  81 }
  82
  83 static struct cpufreq_driver acpi_cpufreq_driver;
  84
  85 static unsigned int acpi_pstate_strict;
  86 static struct msr __percpu *msrs;
  87
  88 static bool boost_state(unsigned int cpu)
  89 {
  90         u32 lo, hi;
  91         u64 msr;
  92
  93         switch (boot_cpu_data.x86_vendor) {
  94         case X86_VENDOR_INTEL:
  95                 rdmsr_on_cpu(cpu, MSR_IA32_MISC_ENABLE, &lo, &hi);
  96                 msr = lo | ((u64)hi << 32);
  97                 return !(msr & MSR_IA32_MISC_ENABLE_TURBO_DISABLE);
  98         case X86_VENDOR_AMD:
  99                 rdmsr_on_cpu(cpu, MSR_K7_HWCR, &lo, &hi);
 100                 msr = lo | ((u64)hi << 32);
 101                 return !(msr & MSR_K7_HWCR_CPB_DIS);
 102         }
 103         return false;
 104 }
 105
 106 static void boost_set_msrs(bool enable, const struct cpumask *cpumask)
 107 {
 108         u32 cpu;
 109         u32 msr_addr;
 110         u64 msr_mask;
 111
 112         switch (boot_cpu_data.x86_vendor) {
 113         case X86_VENDOR_INTEL:
 114                 msr_addr = MSR_IA32_MISC_ENABLE;
 115                 msr_mask = MSR_IA32_MISC_ENABLE_TURBO_DISABLE;
 116                 break;
 117         case X86_VENDOR_AMD:
 118                 msr_addr = MSR_K7_HWCR;
 119                 msr_mask = MSR_K7_HWCR_CPB_DIS;
 120                 break;
 121         default:
 122                 return;
 123         }
 124
 125         rdmsr_on_cpus(cpumask, msr_addr, msrs);
 126
 127         for_each_cpu(cpu, cpumask) {
 128                 struct msr *reg = per_cpu_ptr(msrs, cpu);
 129                 if (enable)
 130                         reg->q &= ~msr_mask;
 131                 else
 132                         reg->q |= msr_mask;
 133         }
 134
 135         wrmsr_on_cpus(cpumask, msr_addr, msrs);
 136 }
 137
 138 static int _store_boost(int val)
 139 {
 140         get_online_cpus();
 141         boost_set_msrs(val, cpu_online_mask);
 142         put_online_cpus();
 143         pr_debug("Core Boosting %sabled.\n", val ? "en" : "dis");
 144
 145         return 0;
 146 }
 147
 148 static ssize_t show_freqdomain_cpus(struct cpufreq_policy *policy, char *buf)
 149 {
 150         struct acpi_cpufreq_data *data = policy->driver_data;
 151
 152         if (unlikely(!data))
 153                 return -ENODEV;
 154
 155         return cpufreq_show_cpus(data->freqdomain_cpus, buf);
 156 }
 157
 158 cpufreq_freq_attr_ro(freqdomain_cpus);
 159
 160 #ifdef CONFIG_X86_ACPI_CPUFREQ_CPB
 161 static ssize_t store_boost(const char *buf, size_t count)
 162 {
 163         int ret;
 164         unsigned long val = 0;
 165
 166         if (!acpi_cpufreq_driver.boost_supported)
 167                 return -EINVAL;
 168
 169         ret = kstrtoul(buf, 10, &val);
 170         if (ret || (val > 1))
 171                 return -EINVAL;
 172
 173         _store_boost((int) val);
 174
 175         return count;
 176 }
 177
 178 static ssize_t store_cpb(struct cpufreq_policy *policy, const char *buf,
 179                          size_t count)
 180 {
 181         return store_boost(buf, count);
 182 }
 183
 184 static ssize_t show_cpb(struct cpufreq_policy *policy, char *buf)
 185 {
 186         return sprintf(buf, "%u\n", acpi_cpufreq_driver.boost_enabled);
 187 }
 188
 189 cpufreq_freq_attr_rw(cpb);
 190 #endif
 191
 192 static int check_est_cpu(unsigned int cpuid)
 193 {
 194         struct cpuinfo_x86 *cpu = &cpu_data(cpuid);
 195
 196         return cpu_has(cpu, X86_FEATURE_EST);
 197 }
 198
 199 static int check_amd_hwpstate_cpu(unsigned int cpuid)
 200 {
 201         struct cpuinfo_x86 *cpu = &cpu_data(cpuid);
 202
 203         return cpu_has(cpu, X86_FEATURE_HW_PSTATE);
 204 }
 205
 206 static unsigned extract_io(u32 value, struct acpi_cpufreq_data *data)
 207 {
 208         struct acpi_processor_performance *perf;
 209         int i;
 210
 211         perf = to_perf_data(data);
 212
 213         for (i = 0; i < perf->state_count; i++) {
 214                 if (value == perf->states[i].status)
 215                         return data->freq_table[i].frequency;
 216         }
 217         return 0;
 218 }
 219
 220 static unsigned extract_msr(u32 msr, struct acpi_cpufreq_data *data)
 221 {
 222         struct cpufreq_frequency_table *pos;
 223         struct acpi_processor_performance *perf;
 224
 225         if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD)
 226                 msr &= AMD_MSR_RANGE;
 227         else
 228                 msr &= INTEL_MSR_RANGE;
 229
 230         perf = to_perf_data(data);
 231
 232         cpufreq_for_each_entry(pos, data->freq_table)
 233                 if (msr == perf->states[pos->driver_data].status)
 234                         return pos->frequency;
 235         return data->freq_table[0].frequency;
 236 }
 237
 238 static unsigned extract_freq(u32 val, struct acpi_cpufreq_data *data)
 239 {
 240         switch (data->cpu_feature) {
 241         case SYSTEM_INTEL_MSR_CAPABLE:
 242         case SYSTEM_AMD_MSR_CAPABLE:
 243                 return extract_msr(val, data);
 244         case SYSTEM_IO_CAPABLE:
 245                 return extract_io(val, data);
 246         default:
 247                 return 0;
 248         }
 249 }
 250
 251 struct msr_addr {
 252         u32 reg;
 253 };
 254
 255 struct io_addr {
 256         u16 port;
 257         u8 bit_width;
 258 };
 259
 260 struct drv_cmd {
 261         unsigned int type;
 262         const struct cpumask *mask;
 263         union {
 264                 struct msr_addr msr;
 265                 struct io_addr io;
 266         } addr;
 267         u32 val;
 268 };
 269
 270 /* Called via smp_call_function_single(), on the target CPU */
 271 static void do_drv_read(void *_cmd)
 272 {
 273         struct drv_cmd *cmd = _cmd;
 274         u32 h;
 275
 276         switch (cmd->type) {
 277         case SYSTEM_INTEL_MSR_CAPABLE:
 278         case SYSTEM_AMD_MSR_CAPABLE:
 279                 rdmsr(cmd->addr.msr.reg, cmd->val, h);
 280                 break;
 281         case SYSTEM_IO_CAPABLE:
 282                 acpi_os_read_port((acpi_io_address)cmd->addr.io.port,
 283                                 &cmd->val,
 284                                 (u32)cmd->addr.io.bit_width);
 285                 break;
 286         default:
 287                 break;
 288         }
 289 }
 290
 291 /* Called via smp_call_function_many(), on the target CPUs */
 292 static void do_drv_write(void *_cmd)
 293 {
 294         struct drv_cmd *cmd = _cmd;
 295         u32 lo, hi;
 296
 297         switch (cmd->type) {
 298         case SYSTEM_INTEL_MSR_CAPABLE:
 299                 rdmsr(cmd->addr.msr.reg, lo, hi);
 300                 lo = (lo & ~INTEL_MSR_RANGE) | (cmd->val & INTEL_MSR_RANGE);
 301                 wrmsr(cmd->addr.msr.reg, lo, hi);
 302                 break;
 303         case SYSTEM_AMD_MSR_CAPABLE:
 304                 wrmsr(cmd->addr.msr.reg, cmd->val, 0);
 305                 break;
 306         case SYSTEM_IO_CAPABLE:
 307                 acpi_os_write_port((acpi_io_address)cmd->addr.io.port,
 308                                 cmd->val,
 309                                 (u32)cmd->addr.io.bit_width);
 310                 break;
 311         default:
 312                 break;
 313         }
 314 }
 315
 316 static void drv_read(struct drv_cmd *cmd)
 317 {
 318         int err;
 319         cmd->val = 0;
 320
 321         err = smp_call_function_any(cmd->mask, do_drv_read, cmd, 1);
 322         WARN_ON_ONCE(err);      /* smp_call_function_any() was buggy? */
 323 }
 324
 325 static void drv_write(struct drv_cmd *cmd)
 326 {
 327         int this_cpu;
 328
 329         this_cpu = get_cpu();
 330         if (cpumask_test_cpu(this_cpu, cmd->mask))
 331                 do_drv_write(cmd);
 332         smp_call_function_many(cmd->mask, do_drv_write, cmd, 1);
 333         put_cpu();
 334 }
 335
 336 static u32
 337 get_cur_val(const struct cpumask *mask, struct acpi_cpufreq_data *data)
 338 {
 339         struct acpi_processor_performance *perf;
 340         struct drv_cmd cmd;
 341
 342         if (unlikely(cpumask_empty(mask)))
 343                 return 0;
 344
 345         switch (data->cpu_feature) {
 346         case SYSTEM_INTEL_MSR_CAPABLE:
 347                 cmd.type = SYSTEM_INTEL_MSR_CAPABLE;
 348                 cmd.addr.msr.reg = MSR_IA32_PERF_CTL;
 349                 break;
 350         case SYSTEM_AMD_MSR_CAPABLE:
 351                 cmd.type = SYSTEM_AMD_MSR_CAPABLE;
 352                 cmd.addr.msr.reg = MSR_AMD_PERF_CTL;
 353                 break;
 354         case SYSTEM_IO_CAPABLE:
 355                 cmd.type = SYSTEM_IO_CAPABLE;
 356                 perf = to_perf_data(data);
 357                 cmd.addr.io.port = perf->control_register.address;
 358                 cmd.addr.io.bit_width = perf->control_register.bit_width;
 359                 break;
 360         default:
 361                 return 0;
 362         }
 363
 364         cmd.mask = mask;
 365         drv_read(&cmd);
 366
 367         pr_debug("get_cur_val = %u\n", cmd.val);
 368
 369         return cmd.val;
 370 }
 371
 372 static unsigned int get_cur_freq_on_cpu(unsigned int cpu)
 373 {
 374         struct acpi_cpufreq_data *data;
 375         struct cpufreq_policy *policy;
 376         unsigned int freq;
 377         unsigned int cached_freq;
 378
 379         pr_debug("get_cur_freq_on_cpu (%d)\n", cpu);
 380
 381         policy = cpufreq_cpu_get_raw(cpu);
 382         if (unlikely(!policy))
 383                 return 0;
 384
 385         data = policy->driver_data;
 386         if (unlikely(!data || !data->freq_table))
 387                 return 0;
 388
 389         cached_freq = data->freq_table[to_perf_data(data)->state].frequency;
 390         freq = extract_freq(get_cur_val(cpumask_of(cpu), data), data);
 391         if (freq != cached_freq) {
 392                 /*
 393                  * The dreaded BIOS frequency change behind our back.
 394                  * Force set the frequency on next target call.
 395                  */
 396                 data->resume = 1;
 397         }
 398
 399         pr_debug("cur freq = %u\n", freq);
 400
 401         return freq;
 402 }
 403
 404 static unsigned int check_freqs(const struct cpumask *mask, unsigned int freq,
 405                                 struct acpi_cpufreq_data *data)
 406 {
 407         unsigned int cur_freq;
 408         unsigned int i;
 409
 410         for (i = 0; i < 100; i++) {
 411                 cur_freq = extract_freq(get_cur_val(mask, data), data);
 412                 if (cur_freq == freq)
 413                         return 1;
 414                 udelay(10);
 415         }
 416         return 0;
 417 }
 418
 419 static int acpi_cpufreq_target(struct cpufreq_policy *policy,
 420                                unsigned int index)
 421 {
 422         struct acpi_cpufreq_data *data = policy->driver_data;
 423         struct acpi_processor_performance *perf;
 424         struct drv_cmd cmd;
 425         unsigned int next_perf_state = 0; /* Index into perf table */
 426         int result = 0;
 427
 428         if (unlikely(data == NULL || data->freq_table == NULL)) {
 429                 return -ENODEV;
 430         }
 431
 432         perf = to_perf_data(data);
 433         next_perf_state = data->freq_table[index].driver_data;
 434         if (perf->state == next_perf_state) {
 435                 if (unlikely(data->resume)) {
 436                         pr_debug("Called after resume, resetting to P%d\n",
 437                                 next_perf_state);
 438                         data->resume = 0;
 439                 } else {
 440                         pr_debug("Already at target state (P%d)\n",
 441                                 next_perf_state);
 442                         goto out;
 443                 }
 444         }
 445
 446         switch (data->cpu_feature) {
 447         case SYSTEM_INTEL_MSR_CAPABLE:
 448                 cmd.type = SYSTEM_INTEL_MSR_CAPABLE;
 449                 cmd.addr.msr.reg = MSR_IA32_PERF_CTL;
 450                 cmd.val = (u32) perf->states[next_perf_state].control;
 451                 break;
 452         case SYSTEM_AMD_MSR_CAPABLE:
 453                 cmd.type = SYSTEM_AMD_MSR_CAPABLE;
 454                 cmd.addr.msr.reg = MSR_AMD_PERF_CTL;
 455                 cmd.val = (u32) perf->states[next_perf_state].control;
 456                 break;
 457         case SYSTEM_IO_CAPABLE:
 458                 cmd.type = SYSTEM_IO_CAPABLE;
 459                 cmd.addr.io.port = perf->control_register.address;
 460                 cmd.addr.io.bit_width = perf->control_register.bit_width;
 461                 cmd.val = (u32) perf->states[next_perf_state].control;
 462                 break;
 463         default:
 464                 result = -ENODEV;
 465                 goto out;
 466         }
 467
 468         /* cpufreq holds the hotplug lock, so we are safe from here on */
 469         if (policy->shared_type != CPUFREQ_SHARED_TYPE_ANY)
 470                 cmd.mask = policy->cpus;
 471         else
 472                 cmd.mask = cpumask_of(policy->cpu);
 473
 474         drv_write(&cmd);
 475
 476         if (acpi_pstate_strict) {
 477                 if (!check_freqs(cmd.mask, data->freq_table[index].frequency,
 478                                         data)) {
 479                         pr_debug("acpi_cpufreq_target failed (%d)\n",
 480                                 policy->cpu);
 481                         result = -EAGAIN;
 482                 }
 483         }
 484
 485         if (!result)
 486                 perf->state = next_perf_state;
 487
 488 out:
 489         return result;
 490 }
 491
 492 static unsigned long
 493 acpi_cpufreq_guess_freq(struct acpi_cpufreq_data *data, unsigned int cpu)
 494 {
 495         struct acpi_processor_performance *perf;
 496
 497         perf = to_perf_data(data);
 498         if (cpu_khz) {
 499                 /* search the closest match to cpu_khz */
 500                 unsigned int i;
 501                 unsigned long freq;
 502                 unsigned long freqn = perf->states[0].core_frequency * 1000;
 503
 504                 for (i = 0; i < (perf->state_count-1); i++) {
 505                         freq = freqn;
 506                         freqn = perf->states[i+1].core_frequency * 1000;
 507                         if ((2 * cpu_khz) > (freqn + freq)) {
 508                                 perf->state = i;
 509                                 return freq;
 510                         }
 511                 }
 512                 perf->state = perf->state_count-1;
 513                 return freqn;
 514         } else {
 515                 /* assume CPU is at P0... */
 516                 perf->state = 0;
 517                 return perf->states[0].core_frequency * 1000;
 518         }
 519 }
 520
 521 static void free_acpi_perf_data(void)
 522 {
 523         unsigned int i;
 524
 525         /* Freeing a NULL pointer is OK, and alloc_percpu zeroes. */
 526         for_each_possible_cpu(i)
 527                 free_cpumask_var(per_cpu_ptr(acpi_perf_data, i)
 528                                  ->shared_cpu_map);
 529         free_percpu(acpi_perf_data);
 530 }
 531
 532 static int boost_notify(struct notifier_block *nb, unsigned long action,
 533                       void *hcpu)
 534 {
 535         unsigned cpu = (long)hcpu;
 536         const struct cpumask *cpumask;
 537
 538         cpumask = get_cpu_mask(cpu);
 539
 540         /*
 541          * Clear the boost-disable bit on the CPU_DOWN path so that
 542          * this cpu cannot block the remaining ones from boosting. On
 543          * the CPU_UP path we simply keep the boost-disable flag in
 544          * sync with the current global state.
 545          */
 546
 547         switch (action) {
 548         case CPU_UP_PREPARE:
 549         case CPU_UP_PREPARE_FROZEN:
 550                 boost_set_msrs(acpi_cpufreq_driver.boost_enabled, cpumask);
 551                 break;
 552
 553         case CPU_DOWN_PREPARE:
 554         case CPU_DOWN_PREPARE_FROZEN:
 555                 boost_set_msrs(1, cpumask);
 556                 break;
 557
 558         default:
 559                 break;
 560         }
 561
 562         return NOTIFY_OK;
 563 }
 564
 565
 566 static struct notifier_block boost_nb = {
 567         .notifier_call          = boost_notify,
 568 };
 569
 570 /*
 571  * acpi_cpufreq_early_init - initialize ACPI P-States library
 572  *
 573  * Initialize the ACPI P-States library (drivers/acpi/processor_perflib.c)
 574  * in order to determine correct frequency and voltage pairings. We can
 575  * do _PDC and _PSD and find out the processor dependency for the
 576  * actual init that will happen later...
 577  */
 578 static int __init acpi_cpufreq_early_init(void)
 579 {
 580         unsigned int i;
 581         pr_debug("acpi_cpufreq_early_init\n");
 582
 583         acpi_perf_data = alloc_percpu(struct acpi_processor_performance);
 584         if (!acpi_perf_data) {
 585                 pr_debug("Memory allocation error for acpi_perf_data.\n");
 586                 return -ENOMEM;
 587         }
 588         for_each_possible_cpu(i) {
 589                 if (!zalloc_cpumask_var_node(
 590                         &per_cpu_ptr(acpi_perf_data, i)->shared_cpu_map,
 591                         GFP_KERNEL, cpu_to_node(i))) {
 592
 593                         /* Freeing a NULL pointer is OK: alloc_percpu zeroes. */
 594                         free_acpi_perf_data();
 595                         return -ENOMEM;
 596                 }
 597         }
 598
 599         /* Do initialization in ACPI core */
 600         acpi_processor_preregister_performance(acpi_perf_data);
 601         return 0;
 602 }
 603
 604 #ifdef CONFIG_SMP
 605 /*
 606  * Some BIOSes do SW_ANY coordination internally, either set it up in hw
 607  * or do it in BIOS firmware and won't inform about it to OS. If not
 608  * detected, this has a side effect of making CPU run at a different speed
 609  * than OS intended it to run at. Detect it and handle it cleanly.
 610  */
 611 static int bios_with_sw_any_bug;
 612
 613 static int sw_any_bug_found(const struct dmi_system_id *d)
 614 {
 615         bios_with_sw_any_bug = 1;
 616         return 0;
 617 }
 618
 619 static const struct dmi_system_id sw_any_bug_dmi_table[] = {
 620         {
 621                 .callback = sw_any_bug_found,
 622                 .ident = "Supermicro Server X6DLP",
 623                 .matches = {
 624                         DMI_MATCH(DMI_SYS_VENDOR, "Supermicro"),
 625                         DMI_MATCH(DMI_BIOS_VERSION, "080010"),
 626                         DMI_MATCH(DMI_PRODUCT_NAME, "X6DLP"),
 627                 },
 628         },
 629         { }
 630 };
 631
 632 static int acpi_cpufreq_blacklist(struct cpuinfo_x86 *c)
 633 {
 634         /* Intel Xeon Processor 7100 Series Specification Update
 635          * http://www.intel.com/Assets/PDF/specupdate/314554.pdf
 636          * AL30: A Machine Check Exception (MCE) Occurring during an
 637          * Enhanced Intel SpeedStep Technology Ratio Change May Cause
 638          * Both Processor Cores to Lock Up. */
 639         if (c->x86_vendor == X86_VENDOR_INTEL) {
 640                 if ((c->x86 == 15) &&
 641                     (c->x86_model == 6) &&
 642                     (c->x86_mask == 8)) {
 643                         printk(KERN_INFO "acpi-cpufreq: Intel(R) "
 644                             "Xeon(R) 7100 Errata AL30, processors may "
 645                             "lock up on frequency changes: disabling "
 646                             "acpi-cpufreq.\n");
 647                         return -ENODEV;
 648                     }
 649                 }
 650         return 0;
 651 }
 652 #endif
 653
 654 static int acpi_cpufreq_cpu_init(struct cpufreq_policy *policy)
 655 {
 656         unsigned int i;
 657         unsigned int valid_states = 0;
 658         unsigned int cpu = policy->cpu;
 659         struct acpi_cpufreq_data *data;
 660         unsigned int result = 0;
 661         struct cpuinfo_x86 *c = &cpu_data(policy->cpu);
 662         struct acpi_processor_performance *perf;
 663 #ifdef CONFIG_SMP
 664         static int blacklisted;
 665 #endif
 666
 667         pr_debug("acpi_cpufreq_cpu_init\n");
 668
 669 #ifdef CONFIG_SMP
 670         if (blacklisted)
 671                 return blacklisted;
 672         blacklisted = acpi_cpufreq_blacklist(c);
 673         if (blacklisted)
 674                 return blacklisted;
 675 #endif
 676
 677         data = kzalloc(sizeof(*data), GFP_KERNEL);
 678         if (!data)
 679                 return -ENOMEM;
 680
 681         if (!zalloc_cpumask_var(&data->freqdomain_cpus, GFP_KERNEL)) {
 682                 result = -ENOMEM;
 683                 goto err_free;
 684         }
 685
 686         perf = per_cpu_ptr(acpi_perf_data, cpu);
 687         data->acpi_perf_cpu = cpu;
 688         policy->driver_data = data;
 689
 690         if (cpu_has(c, X86_FEATURE_CONSTANT_TSC))
 691                 acpi_cpufreq_driver.flags |= CPUFREQ_CONST_LOOPS;
 692
 693         result = acpi_processor_register_performance(perf, cpu);
 694         if (result)
 695                 goto err_free_mask;
 696
 697         policy->shared_type = perf->shared_type;
 698
 699         /*
 700          * Will let policy->cpus know about dependency only when software
 701          * coordination is required.
 702          */
 703         if (policy->shared_type == CPUFREQ_SHARED_TYPE_ALL ||
 704             policy->shared_type == CPUFREQ_SHARED_TYPE_ANY) {
 705                 cpumask_copy(policy->cpus, perf->shared_cpu_map);
 706         }
 707         cpumask_copy(data->freqdomain_cpus, perf->shared_cpu_map);
 708
 709 #ifdef CONFIG_SMP
 710         dmi_check_system(sw_any_bug_dmi_table);
 711         if (bios_with_sw_any_bug && !policy_is_shared(policy)) {
 712                 policy->shared_type = CPUFREQ_SHARED_TYPE_ALL;
 713                 cpumask_copy(policy->cpus, topology_core_cpumask(cpu));
 714         }
 715
 716         if (check_amd_hwpstate_cpu(cpu) && !acpi_pstate_strict) {
 717                 cpumask_clear(policy->cpus);
 718                 cpumask_set_cpu(cpu, policy->cpus);
 719                 cpumask_copy(data->freqdomain_cpus,
 720                              topology_sibling_cpumask(cpu));
 721                 policy->shared_type = CPUFREQ_SHARED_TYPE_HW;
 722                 pr_info_once(PFX "overriding BIOS provided _PSD data\n");
 723         }
 724 #endif
 725
 726         /* capability check */
 727         if (perf->state_count <= 1) {
 728                 pr_debug("No P-States\n");
 729                 result = -ENODEV;
 730                 goto err_unreg;
 731         }
 732
 733         if (perf->control_register.space_id != perf->status_register.space_id) {
 734                 result = -ENODEV;
 735                 goto err_unreg;
 736         }
 737
 738         switch (perf->control_register.space_id) {
 739         case ACPI_ADR_SPACE_SYSTEM_IO:
 740                 if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD &&
 741                     boot_cpu_data.x86 == 0xf) {
 742                         pr_debug("AMD K8 systems must use native drivers.\n");
 743                         result = -ENODEV;
 744                         goto err_unreg;
 745                 }
 746                 pr_debug("SYSTEM IO addr space\n");
 747                 data->cpu_feature = SYSTEM_IO_CAPABLE;
 748                 break;
 749         case ACPI_ADR_SPACE_FIXED_HARDWARE:
 750                 pr_debug("HARDWARE addr space\n");
 751                 if (check_est_cpu(cpu)) {
 752                         data->cpu_feature = SYSTEM_INTEL_MSR_CAPABLE;
 753                         break;
 754                 }
 755                 if (check_amd_hwpstate_cpu(cpu)) {
 756                         data->cpu_feature = SYSTEM_AMD_MSR_CAPABLE;
 757                         break;
 758                 }
 759                 result = -ENODEV;
 760                 goto err_unreg;
 761         default:
 762                 pr_debug("Unknown addr space %d\n",
 763                         (u32) (perf->control_register.space_id));
 764                 result = -ENODEV;
 765                 goto err_unreg;
 766         }
 767
 768         data->freq_table = kzalloc(sizeof(*data->freq_table) *
 769                     (perf->state_count+1), GFP_KERNEL);
 770         if (!data->freq_table) {
 771                 result = -ENOMEM;
 772                 goto err_unreg;
 773         }
 774
 775         /* detect transition latency */
 776         policy->cpuinfo.transition_latency = 0;
 777         for (i = 0; i < perf->state_count; i++) {
 778                 if ((perf->states[i].transition_latency * 1000) >
 779                     policy->cpuinfo.transition_latency)
 780                         policy->cpuinfo.transition_latency =
 781                             perf->states[i].transition_latency * 1000;
 782         }
 783
 784         /* Check for high latency (>20uS) from buggy BIOSes, like on T42 */
 785         if (perf->control_register.space_id == ACPI_ADR_SPACE_FIXED_HARDWARE &&
 786             policy->cpuinfo.transition_latency > 20 * 1000) {
 787                 policy->cpuinfo.transition_latency = 20 * 1000;
 788                 printk_once(KERN_INFO
 789                             "P-state transition latency capped at 20 uS\n");
 790         }
 791
 792         /* table init */
 793         for (i = 0; i < perf->state_count; i++) {
 794                 if (i > 0 && perf->states[i].core_frequency >=
 795                     data->freq_table[valid_states-1].frequency / 1000)
 796                         continue;
 797
 798                 data->freq_table[valid_states].driver_data = i;
 799                 data->freq_table[valid_states].frequency =
 800                     perf->states[i].core_frequency * 1000;
 801                 valid_states++;
 802         }
 803         data->freq_table[valid_states].frequency = CPUFREQ_TABLE_END;
 804         perf->state = 0;
 805
 806         result = cpufreq_table_validate_and_show(policy, data->freq_table);
 807         if (result)
 808                 goto err_freqfree;
 809
 810         if (perf->states[0].core_frequency * 1000 != policy->cpuinfo.max_freq)
 811                 printk(KERN_WARNING FW_WARN "P-state 0 is not max freq\n");
 812
 813         switch (perf->control_register.space_id) {
 814         case ACPI_ADR_SPACE_SYSTEM_IO:
 815                 /*
 816                  * The core will not set policy->cur, because
 817                  * cpufreq_driver->get is NULL, so we need to set it here.
 818                  * However, we have to guess it, because the current speed is
 819                  * unknown and not detectable via IO ports.
 820                  */
 821                 policy->cur = acpi_cpufreq_guess_freq(data, policy->cpu);
 822                 break;
 823         case ACPI_ADR_SPACE_FIXED_HARDWARE:
 824                 acpi_cpufreq_driver.get = get_cur_freq_on_cpu;
 825                 break;
 826         default:
 827                 break;
 828         }
 829
 830         /* notify BIOS that we exist */
 831         acpi_processor_notify_smm(THIS_MODULE);
 832
 833         pr_debug("CPU%u - ACPI performance management activated.\n", cpu);
 834         for (i = 0; i < perf->state_count; i++)
 835                 pr_debug("     %cP%d: %d MHz, %d mW, %d uS\n",
 836                         (i == perf->state ? '*' : ' '), i,
 837                         (u32) perf->states[i].core_frequency,
 838                         (u32) perf->states[i].power,
 839                         (u32) perf->states[i].transition_latency);
 840
 841         /*
 842          * the first call to ->target() should result in us actually
 843          * writing something to the appropriate registers.
 844          */
 845         data->resume = 1;
 846
 847         return result;
 848
 849 err_freqfree:
 850         kfree(data->freq_table);
 851 err_unreg:
 852         acpi_processor_unregister_performance(cpu);
 853 err_free_mask:
 854         free_cpumask_var(data->freqdomain_cpus);
 855 err_free:
 856         kfree(data);
 857         policy->driver_data = NULL;
 858
 859         return result;
 860 }
 861
 862 static int acpi_cpufreq_cpu_exit(struct cpufreq_policy *policy)
 863 {
 864         struct acpi_cpufreq_data *data = policy->driver_data;
 865
 866         pr_debug("acpi_cpufreq_cpu_exit\n");
 867
 868         if (data) {
 869                 policy->driver_data = NULL;
 870                 acpi_processor_unregister_performance(data->acpi_perf_cpu);
 871                 free_cpumask_var(data->freqdomain_cpus);
 872                 kfree(data->freq_table);
 873                 kfree(data);
 874         }
 875
 876         return 0;
 877 }
 878
 879 static int acpi_cpufreq_resume(struct cpufreq_policy *policy)
 880 {
 881         struct acpi_cpufreq_data *data = policy->driver_data;
 882
 883         pr_debug("acpi_cpufreq_resume\n");
 884
 885         data->resume = 1;
 886
 887         return 0;
 888 }
 889
 890 static struct freq_attr *acpi_cpufreq_attr[] = {
 891         &cpufreq_freq_attr_scaling_available_freqs,
 892         &freqdomain_cpus,
 893 #ifdef CONFIG_X86_ACPI_CPUFREQ_CPB
 894         &cpb,
 895 #endif
 896         NULL,
 897 };
 898
 899 static struct cpufreq_driver acpi_cpufreq_driver = {
 900         .verify         = cpufreq_generic_frequency_table_verify,
 901         .target_index   = acpi_cpufreq_target,
 902         .bios_limit     = acpi_processor_get_bios_limit,
 903         .init           = acpi_cpufreq_cpu_init,
 904         .exit           = acpi_cpufreq_cpu_exit,
 905         .resume         = acpi_cpufreq_resume,
 906         .name           = "acpi-cpufreq",
 907         .attr           = acpi_cpufreq_attr,
 908         .set_boost      = _store_boost,
 909 };
 910
 911 static void __init acpi_cpufreq_boost_init(void)
 912 {
 913         if (boot_cpu_has(X86_FEATURE_CPB) || boot_cpu_has(X86_FEATURE_IDA)) {
 914                 msrs = msrs_alloc();
 915
 916                 if (!msrs)
 917                         return;
 918
 919                 acpi_cpufreq_driver.boost_supported = true;
 920                 acpi_cpufreq_driver.boost_enabled = boost_state(0);
 921
 922                 cpu_notifier_register_begin();
 923
 924                 /* Force all MSRs to the same value */
 925                 boost_set_msrs(acpi_cpufreq_driver.boost_enabled,
 926                                cpu_online_mask);
 927
 928                 __register_cpu_notifier(&boost_nb);
 929
 930                 cpu_notifier_register_done();
 931         }
 932 }
 933
 934 static void acpi_cpufreq_boost_exit(void)
 935 {
 936         if (msrs) {
 937                 unregister_cpu_notifier(&boost_nb);
 938
 939                 msrs_free(msrs);
 940                 msrs = NULL;
 941         }
 942 }
 943
 944 static int __init acpi_cpufreq_init(void)
 945 {
 946         int ret;
 947
 948         if (acpi_disabled)
 949                 return -ENODEV;
 950
 951         /* don't keep reloading if cpufreq_driver exists */
 952         if (cpufreq_get_current_driver())
 953                 return -EEXIST;
 954
 955         pr_debug("acpi_cpufreq_init\n");
 956
 957         ret = acpi_cpufreq_early_init();
 958         if (ret)
 959                 return ret;
 960
 961 #ifdef CONFIG_X86_ACPI_CPUFREQ_CPB
 962         /* this is a sysfs file with a strange name and an even stranger
 963          * semantic - per CPU instantiation, but system global effect.
 964          * Lets enable it only on AMD CPUs for compatibility reasons and
 965          * only if configured. This is considered legacy code, which
 966          * will probably be removed at some point in the future.
 967          */
 968         if (!check_amd_hwpstate_cpu(0)) {
 969                 struct freq_attr **attr;
 970
 971                 pr_debug("CPB unsupported, do not expose it\n");
 972
 973                 for (attr = acpi_cpufreq_attr; *attr; attr++)
 974                         if (*attr == &cpb) {
 975                                 *attr = NULL;
 976                                 break;
 977                         }
 978         }
 979 #endif
 980         acpi_cpufreq_boost_init();
 981
 982         ret = cpufreq_register_driver(&acpi_cpufreq_driver);
 983         if (ret) {
 984                 free_acpi_perf_data();
 985                 acpi_cpufreq_boost_exit();
 986         }
 987         return ret;
 988 }
 989
 990 static void __exit acpi_cpufreq_exit(void)
 991 {
 992         pr_debug("acpi_cpufreq_exit\n");
 993
 994         acpi_cpufreq_boost_exit();
 995
 996         cpufreq_unregister_driver(&acpi_cpufreq_driver);
 997
 998         free_acpi_perf_data();
 999 }
1000
1001 module_param(acpi_pstate_strict, uint, 0644);
1002 MODULE_PARM_DESC(acpi_pstate_strict,
1003         "value 0 or non-zero. non-zero -> strict ACPI checks are "
1004         "performed during frequency changes.");
1005
1006 late_initcall(acpi_cpufreq_init);
1007 module_exit(acpi_cpufreq_exit);
1008
1009 static const struct x86_cpu_id acpi_cpufreq_ids[] = {
1010         X86_FEATURE_MATCH(X86_FEATURE_ACPI),
1011         X86_FEATURE_MATCH(X86_FEATURE_HW_PSTATE),
1012         {}
1013 };
1014 MODULE_DEVICE_TABLE(x86cpu, acpi_cpufreq_ids);
1015
1016 static const struct acpi_device_id processor_device_ids[] = {
1017         {ACPI_PROCESSOR_OBJECT_HID, },
1018         {ACPI_PROCESSOR_DEVICE_HID, },
1019         {},
1020 };
1021 MODULE_DEVICE_TABLE(acpi, processor_device_ids);
1022
1023 MODULE_ALIAS("acpi");