drivers/cpufreq/acpi-cpufreq.c

   1 /*
   2  * acpi-cpufreq.c - ACPI Processor P-States Driver
   3  *
   4  *  Copyright (C) 2001, 2002 Andy Grover <andrew.grover@intel.com>
   5  *  Copyright (C) 2001, 2002 Paul Diefenbaugh <paul.s.diefenbaugh@intel.com>
   6  *  Copyright (C) 2002 - 2004 Dominik Brodowski <linux@brodo.de>
   7  *  Copyright (C) 2006       Denis Sadykov <denis.m.sadykov@intel.com>
   8  *
   9  * ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
  10  *
  11  *  This program is free software; you can redistribute it and/or modify
  12  *  it under the terms of the GNU General Public License as published by
  13  *  the Free Software Foundation; either version 2 of the License, or (at
  14  *  your option) any later version.
  15  *
  16  *  This program is distributed in the hope that it will be useful, but
  17  *  WITHOUT ANY WARRANTY; without even the implied warranty of
  18  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  19  *  General Public License for more details.
  20  *
  21  *  You should have received a copy of the GNU General Public License along
  22  *  with this program; if not, write to the Free Software Foundation, Inc.,
  23  *  59 Temple Place, Suite 330, Boston, MA 02111-1307 USA.
  24  *
  25  * ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
  26  */
  27
  28 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
  29
  30 #include <linux/kernel.h>
  31 #include <linux/module.h>
  32 #include <linux/init.h>
  33 #include <linux/smp.h>
  34 #include <linux/sched.h>
  35 #include <linux/cpufreq.h>
  36 #include <linux/compiler.h>
  37 #include <linux/dmi.h>
  38 #include <linux/slab.h>
  39
  40 #include <linux/acpi.h>
  41 #include <linux/io.h>
  42 #include <linux/delay.h>
  43 #include <linux/uaccess.h>
  44
  45 #include <acpi/processor.h>
  46
  47 #include <asm/msr.h>
  48 #include <asm/processor.h>
  49 #include <asm/cpufeature.h>
  50
  51 MODULE_AUTHOR("Paul Diefenbaugh, Dominik Brodowski");
  52 MODULE_DESCRIPTION("ACPI Processor P-States Driver");
  53 MODULE_LICENSE("GPL");
  54
  55 enum {
  56         UNDEFINED_CAPABLE = 0,
  57         SYSTEM_INTEL_MSR_CAPABLE,
  58         SYSTEM_AMD_MSR_CAPABLE,
  59         SYSTEM_IO_CAPABLE,
  60 };
  61
  62 #define INTEL_MSR_RANGE         (0xffff)
  63 #define AMD_MSR_RANGE           (0x7)
  64
  65 #define MSR_K7_HWCR_CPB_DIS     (1ULL << 25)
  66
  67 struct acpi_cpufreq_data {
  68         unsigned int resume;
  69         unsigned int cpu_feature;
  70         unsigned int acpi_perf_cpu;
  71         cpumask_var_t freqdomain_cpus;
  72         void (*cpu_freq_write)(struct acpi_pct_register *reg, u32 val);
  73         u32 (*cpu_freq_read)(struct acpi_pct_register *reg);
  74 };
  75
  76 /* acpi_perf_data is a pointer to percpu data. */
  77 static struct acpi_processor_performance __percpu *acpi_perf_data;
  78
  79 static inline struct acpi_processor_performance *to_perf_data(struct acpi_cpufreq_data *data)
  80 {
  81         return per_cpu_ptr(acpi_perf_data, data->acpi_perf_cpu);
  82 }
  83
  84 static struct cpufreq_driver acpi_cpufreq_driver;
  85
  86 static unsigned int acpi_pstate_strict;
  87 static struct msr __percpu *msrs;
  88
  89 static bool boost_state(unsigned int cpu)
  90 {
  91         u32 lo, hi;
  92         u64 msr;
  93
  94         switch (boot_cpu_data.x86_vendor) {
  95         case X86_VENDOR_INTEL:
  96                 rdmsr_on_cpu(cpu, MSR_IA32_MISC_ENABLE, &lo, &hi);
  97                 msr = lo | ((u64)hi << 32);
  98                 return !(msr & MSR_IA32_MISC_ENABLE_TURBO_DISABLE);
  99         case X86_VENDOR_AMD:
 100                 rdmsr_on_cpu(cpu, MSR_K7_HWCR, &lo, &hi);
 101                 msr = lo | ((u64)hi << 32);
 102                 return !(msr & MSR_K7_HWCR_CPB_DIS);
 103         }
 104         return false;
 105 }
 106
 107 static void boost_set_msrs(bool enable, const struct cpumask *cpumask)
 108 {
 109         u32 cpu;
 110         u32 msr_addr;
 111         u64 msr_mask;
 112
 113         switch (boot_cpu_data.x86_vendor) {
 114         case X86_VENDOR_INTEL:
 115                 msr_addr = MSR_IA32_MISC_ENABLE;
 116                 msr_mask = MSR_IA32_MISC_ENABLE_TURBO_DISABLE;
 117                 break;
 118         case X86_VENDOR_AMD:
 119                 msr_addr = MSR_K7_HWCR;
 120                 msr_mask = MSR_K7_HWCR_CPB_DIS;
 121                 break;
 122         default:
 123                 return;
 124         }
 125
 126         rdmsr_on_cpus(cpumask, msr_addr, msrs);
 127
 128         for_each_cpu(cpu, cpumask) {
 129                 struct msr *reg = per_cpu_ptr(msrs, cpu);
 130                 if (enable)
 131                         reg->q &= ~msr_mask;
 132                 else
 133                         reg->q |= msr_mask;
 134         }
 135
 136         wrmsr_on_cpus(cpumask, msr_addr, msrs);
 137 }
 138
 139 static int set_boost(int val)
 140 {
 141         get_online_cpus();
 142         boost_set_msrs(val, cpu_online_mask);
 143         put_online_cpus();
 144         pr_debug("Core Boosting %sabled.\n", val ? "en" : "dis");
 145
 146         return 0;
 147 }
 148
 149 static ssize_t show_freqdomain_cpus(struct cpufreq_policy *policy, char *buf)
 150 {
 151         struct acpi_cpufreq_data *data = policy->driver_data;
 152
 153         if (unlikely(!data))
 154                 return -ENODEV;
 155
 156         return cpufreq_show_cpus(data->freqdomain_cpus, buf);
 157 }
 158
 159 cpufreq_freq_attr_ro(freqdomain_cpus);
 160
 161 #ifdef CONFIG_X86_ACPI_CPUFREQ_CPB
 162 static ssize_t store_cpb(struct cpufreq_policy *policy, const char *buf,
 163                          size_t count)
 164 {
 165         int ret;
 166         unsigned int val = 0;
 167
 168         if (!acpi_cpufreq_driver.set_boost)
 169                 return -EINVAL;
 170
 171         ret = kstrtouint(buf, 10, &val);
 172         if (ret || val > 1)
 173                 return -EINVAL;
 174
 175         set_boost(val);
 176
 177         return count;
 178 }
 179
 180 static ssize_t show_cpb(struct cpufreq_policy *policy, char *buf)
 181 {
 182         return sprintf(buf, "%u\n", acpi_cpufreq_driver.boost_enabled);
 183 }
 184
 185 cpufreq_freq_attr_rw(cpb);
 186 #endif
 187
 188 static int check_est_cpu(unsigned int cpuid)
 189 {
 190         struct cpuinfo_x86 *cpu = &cpu_data(cpuid);
 191
 192         return cpu_has(cpu, X86_FEATURE_EST);
 193 }
 194
 195 static int check_amd_hwpstate_cpu(unsigned int cpuid)
 196 {
 197         struct cpuinfo_x86 *cpu = &cpu_data(cpuid);
 198
 199         return cpu_has(cpu, X86_FEATURE_HW_PSTATE);
 200 }
 201
 202 static unsigned extract_io(struct cpufreq_policy *policy, u32 value)
 203 {
 204         struct acpi_cpufreq_data *data = policy->driver_data;
 205         struct acpi_processor_performance *perf;
 206         int i;
 207
 208         perf = to_perf_data(data);
 209
 210         for (i = 0; i < perf->state_count; i++) {
 211                 if (value == perf->states[i].status)
 212                         return policy->freq_table[i].frequency;
 213         }
 214         return 0;
 215 }
 216
 217 static unsigned extract_msr(struct cpufreq_policy *policy, u32 msr)
 218 {
 219         struct acpi_cpufreq_data *data = policy->driver_data;
 220         struct cpufreq_frequency_table *pos;
 221         struct acpi_processor_performance *perf;
 222
 223         if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD)
 224                 msr &= AMD_MSR_RANGE;
 225         else
 226                 msr &= INTEL_MSR_RANGE;
 227
 228         perf = to_perf_data(data);
 229
 230         cpufreq_for_each_entry(pos, policy->freq_table)
 231                 if (msr == perf->states[pos->driver_data].status)
 232                         return pos->frequency;
 233         return policy->freq_table[0].frequency;
 234 }
 235
 236 static unsigned extract_freq(struct cpufreq_policy *policy, u32 val)
 237 {
 238         struct acpi_cpufreq_data *data = policy->driver_data;
 239
 240         switch (data->cpu_feature) {
 241         case SYSTEM_INTEL_MSR_CAPABLE:
 242         case SYSTEM_AMD_MSR_CAPABLE:
 243                 return extract_msr(policy, val);
 244         case SYSTEM_IO_CAPABLE:
 245                 return extract_io(policy, val);
 246         default:
 247                 return 0;
 248         }
 249 }
 250
 251 static u32 cpu_freq_read_intel(struct acpi_pct_register *not_used)
 252 {
 253         u32 val, dummy;
 254
 255         rdmsr(MSR_IA32_PERF_CTL, val, dummy);
 256         return val;
 257 }
 258
 259 static void cpu_freq_write_intel(struct acpi_pct_register *not_used, u32 val)
 260 {
 261         u32 lo, hi;
 262
 263         rdmsr(MSR_IA32_PERF_CTL, lo, hi);
 264         lo = (lo & ~INTEL_MSR_RANGE) | (val & INTEL_MSR_RANGE);
 265         wrmsr(MSR_IA32_PERF_CTL, lo, hi);
 266 }
 267
 268 static u32 cpu_freq_read_amd(struct acpi_pct_register *not_used)
 269 {
 270         u32 val, dummy;
 271
 272         rdmsr(MSR_AMD_PERF_CTL, val, dummy);
 273         return val;
 274 }
 275
 276 static void cpu_freq_write_amd(struct acpi_pct_register *not_used, u32 val)
 277 {
 278         wrmsr(MSR_AMD_PERF_CTL, val, 0);
 279 }
 280
 281 static u32 cpu_freq_read_io(struct acpi_pct_register *reg)
 282 {
 283         u32 val;
 284
 285         acpi_os_read_port(reg->address, &val, reg->bit_width);
 286         return val;
 287 }
 288
 289 static void cpu_freq_write_io(struct acpi_pct_register *reg, u32 val)
 290 {
 291         acpi_os_write_port(reg->address, val, reg->bit_width);
 292 }
 293
 294 struct drv_cmd {
 295         struct acpi_pct_register *reg;
 296         u32 val;
 297         union {
 298                 void (*write)(struct acpi_pct_register *reg, u32 val);
 299                 u32 (*read)(struct acpi_pct_register *reg);
 300         } func;
 301 };
 302
 303 /* Called via smp_call_function_single(), on the target CPU */
 304 static void do_drv_read(void *_cmd)
 305 {
 306         struct drv_cmd *cmd = _cmd;
 307
 308         cmd->val = cmd->func.read(cmd->reg);
 309 }
 310
 311 static u32 drv_read(struct acpi_cpufreq_data *data, const struct cpumask *mask)
 312 {
 313         struct acpi_processor_performance *perf = to_perf_data(data);
 314         struct drv_cmd cmd = {
 315                 .reg = &perf->control_register,
 316                 .func.read = data->cpu_freq_read,
 317         };
 318         int err;
 319
 320         err = smp_call_function_any(mask, do_drv_read, &cmd, 1);
 321         WARN_ON_ONCE(err);      /* smp_call_function_any() was buggy? */
 322         return cmd.val;
 323 }
 324
 325 /* Called via smp_call_function_many(), on the target CPUs */
 326 static void do_drv_write(void *_cmd)
 327 {
 328         struct drv_cmd *cmd = _cmd;
 329
 330         cmd->func.write(cmd->reg, cmd->val);
 331 }
 332
 333 static void drv_write(struct acpi_cpufreq_data *data,
 334                       const struct cpumask *mask, u32 val)
 335 {
 336         struct acpi_processor_performance *perf = to_perf_data(data);
 337         struct drv_cmd cmd = {
 338                 .reg = &perf->control_register,
 339                 .val = val,
 340                 .func.write = data->cpu_freq_write,
 341         };
 342         int this_cpu;
 343
 344         this_cpu = get_cpu();
 345         if (cpumask_test_cpu(this_cpu, mask))
 346                 do_drv_write(&cmd);
 347
 348         smp_call_function_many(mask, do_drv_write, &cmd, 1);
 349         put_cpu();
 350 }
 351
 352 static u32 get_cur_val(const struct cpumask *mask, struct acpi_cpufreq_data *data)
 353 {
 354         u32 val;
 355
 356         if (unlikely(cpumask_empty(mask)))
 357                 return 0;
 358
 359         val = drv_read(data, mask);
 360
 361         pr_debug("get_cur_val = %u\n", val);
 362
 363         return val;
 364 }
 365
 366 static unsigned int get_cur_freq_on_cpu(unsigned int cpu)
 367 {
 368         struct acpi_cpufreq_data *data;
 369         struct cpufreq_policy *policy;
 370         unsigned int freq;
 371         unsigned int cached_freq;
 372
 373         pr_debug("get_cur_freq_on_cpu (%d)\n", cpu);
 374
 375         policy = cpufreq_cpu_get_raw(cpu);
 376         if (unlikely(!policy))
 377                 return 0;
 378
 379         data = policy->driver_data;
 380         if (unlikely(!data || !policy->freq_table))
 381                 return 0;
 382
 383         cached_freq = policy->freq_table[to_perf_data(data)->state].frequency;
 384         freq = extract_freq(policy, get_cur_val(cpumask_of(cpu), data));
 385         if (freq != cached_freq) {
 386                 /*
 387                  * The dreaded BIOS frequency change behind our back.
 388                  * Force set the frequency on next target call.
 389                  */
 390                 data->resume = 1;
 391         }
 392
 393         pr_debug("cur freq = %u\n", freq);
 394
 395         return freq;
 396 }
 397
 398 static unsigned int check_freqs(struct cpufreq_policy *policy,
 399                                 const struct cpumask *mask, unsigned int freq)
 400 {
 401         struct acpi_cpufreq_data *data = policy->driver_data;
 402         unsigned int cur_freq;
 403         unsigned int i;
 404
 405         for (i = 0; i < 100; i++) {
 406                 cur_freq = extract_freq(policy, get_cur_val(mask, data));
 407                 if (cur_freq == freq)
 408                         return 1;
 409                 udelay(10);
 410         }
 411         return 0;
 412 }
 413
 414 static int acpi_cpufreq_target(struct cpufreq_policy *policy,
 415                                unsigned int index)
 416 {
 417         struct acpi_cpufreq_data *data = policy->driver_data;
 418         struct acpi_processor_performance *perf;
 419         const struct cpumask *mask;
 420         unsigned int next_perf_state = 0; /* Index into perf table */
 421         int result = 0;
 422
 423         if (unlikely(!data)) {
 424                 return -ENODEV;
 425         }
 426
 427         perf = to_perf_data(data);
 428         next_perf_state = policy->freq_table[index].driver_data;
 429         if (perf->state == next_perf_state) {
 430                 if (unlikely(data->resume)) {
 431                         pr_debug("Called after resume, resetting to P%d\n",
 432                                 next_perf_state);
 433                         data->resume = 0;
 434                 } else {
 435                         pr_debug("Already at target state (P%d)\n",
 436                                 next_perf_state);
 437                         return 0;
 438                 }
 439         }
 440
 441         /*
 442          * The core won't allow CPUs to go away until the governor has been
 443          * stopped, so we can rely on the stability of policy->cpus.
 444          */
 445         mask = policy->shared_type == CPUFREQ_SHARED_TYPE_ANY ?
 446                 cpumask_of(policy->cpu) : policy->cpus;
 447
 448         drv_write(data, mask, perf->states[next_perf_state].control);
 449
 450         if (acpi_pstate_strict) {
 451                 if (!check_freqs(policy, mask,
 452                                  policy->freq_table[index].frequency)) {
 453                         pr_debug("acpi_cpufreq_target failed (%d)\n",
 454                                 policy->cpu);
 455                         result = -EAGAIN;
 456                 }
 457         }
 458
 459         if (!result)
 460                 perf->state = next_perf_state;
 461
 462         return result;
 463 }
 464
 465 unsigned int acpi_cpufreq_fast_switch(struct cpufreq_policy *policy,
 466                                       unsigned int target_freq)
 467 {
 468         struct acpi_cpufreq_data *data = policy->driver_data;
 469         struct acpi_processor_performance *perf;
 470         struct cpufreq_frequency_table *entry;
 471         unsigned int next_perf_state, next_freq, freq;
 472
 473         /*
 474          * Find the closest frequency above target_freq.
 475          *
 476          * The table is sorted in the reverse order with respect to the
 477          * frequency and all of the entries are valid (see the initialization).
 478          */
 479         entry = policy->freq_table;
 480         do {
 481                 entry++;
 482                 freq = entry->frequency;
 483         } while (freq >= target_freq && freq != CPUFREQ_TABLE_END);
 484         entry--;
 485         next_freq = entry->frequency;
 486         next_perf_state = entry->driver_data;
 487
 488         perf = to_perf_data(data);
 489         if (perf->state == next_perf_state) {
 490                 if (unlikely(data->resume))
 491                         data->resume = 0;
 492                 else
 493                         return next_freq;
 494         }
 495
 496         data->cpu_freq_write(&perf->control_register,
 497                              perf->states[next_perf_state].control);
 498         perf->state = next_perf_state;
 499         return next_freq;
 500 }
 501
 502 static unsigned long
 503 acpi_cpufreq_guess_freq(struct acpi_cpufreq_data *data, unsigned int cpu)
 504 {
 505         struct acpi_processor_performance *perf;
 506
 507         perf = to_perf_data(data);
 508         if (cpu_khz) {
 509                 /* search the closest match to cpu_khz */
 510                 unsigned int i;
 511                 unsigned long freq;
 512                 unsigned long freqn = perf->states[0].core_frequency * 1000;
 513
 514                 for (i = 0; i < (perf->state_count-1); i++) {
 515                         freq = freqn;
 516                         freqn = perf->states[i+1].core_frequency * 1000;
 517                         if ((2 * cpu_khz) > (freqn + freq)) {
 518                                 perf->state = i;
 519                                 return freq;
 520                         }
 521                 }
 522                 perf->state = perf->state_count-1;
 523                 return freqn;
 524         } else {
 525                 /* assume CPU is at P0... */
 526                 perf->state = 0;
 527                 return perf->states[0].core_frequency * 1000;
 528         }
 529 }
 530
 531 static void free_acpi_perf_data(void)
 532 {
 533         unsigned int i;
 534
 535         /* Freeing a NULL pointer is OK, and alloc_percpu zeroes. */
 536         for_each_possible_cpu(i)
 537                 free_cpumask_var(per_cpu_ptr(acpi_perf_data, i)
 538                                  ->shared_cpu_map);
 539         free_percpu(acpi_perf_data);
 540 }
 541
 542 static int boost_notify(struct notifier_block *nb, unsigned long action,
 543                       void *hcpu)
 544 {
 545         unsigned cpu = (long)hcpu;
 546         const struct cpumask *cpumask;
 547
 548         cpumask = get_cpu_mask(cpu);
 549
 550         /*
 551          * Clear the boost-disable bit on the CPU_DOWN path so that
 552          * this cpu cannot block the remaining ones from boosting. On
 553          * the CPU_UP path we simply keep the boost-disable flag in
 554          * sync with the current global state.
 555          */
 556
 557         switch (action) {
 558         case CPU_DOWN_FAILED:
 559         case CPU_DOWN_FAILED_FROZEN:
 560         case CPU_ONLINE:
 561         case CPU_ONLINE_FROZEN:
 562                 boost_set_msrs(acpi_cpufreq_driver.boost_enabled, cpumask);
 563                 break;
 564
 565         case CPU_DOWN_PREPARE:
 566         case CPU_DOWN_PREPARE_FROZEN:
 567                 boost_set_msrs(1, cpumask);
 568                 break;
 569
 570         default:
 571                 break;
 572         }
 573
 574         return NOTIFY_OK;
 575 }
 576
 577
 578 static struct notifier_block boost_nb = {
 579         .notifier_call          = boost_notify,
 580 };
 581
 582 /*
 583  * acpi_cpufreq_early_init - initialize ACPI P-States library
 584  *
 585  * Initialize the ACPI P-States library (drivers/acpi/processor_perflib.c)
 586  * in order to determine correct frequency and voltage pairings. We can
 587  * do _PDC and _PSD and find out the processor dependency for the
 588  * actual init that will happen later...
 589  */
 590 static int __init acpi_cpufreq_early_init(void)
 591 {
 592         unsigned int i;
 593         pr_debug("acpi_cpufreq_early_init\n");
 594
 595         acpi_perf_data = alloc_percpu(struct acpi_processor_performance);
 596         if (!acpi_perf_data) {
 597                 pr_debug("Memory allocation error for acpi_perf_data.\n");
 598                 return -ENOMEM;
 599         }
 600         for_each_possible_cpu(i) {
 601                 if (!zalloc_cpumask_var_node(
 602                         &per_cpu_ptr(acpi_perf_data, i)->shared_cpu_map,
 603                         GFP_KERNEL, cpu_to_node(i))) {
 604
 605                         /* Freeing a NULL pointer is OK: alloc_percpu zeroes. */
 606                         free_acpi_perf_data();
 607                         return -ENOMEM;
 608                 }
 609         }
 610
 611         /* Do initialization in ACPI core */
 612         acpi_processor_preregister_performance(acpi_perf_data);
 613         return 0;
 614 }
 615
 616 #ifdef CONFIG_SMP
 617 /*
 618  * Some BIOSes do SW_ANY coordination internally, either set it up in hw
 619  * or do it in BIOS firmware and won't inform about it to OS. If not
 620  * detected, this has a side effect of making CPU run at a different speed
 621  * than OS intended it to run at. Detect it and handle it cleanly.
 622  */
 623 static int bios_with_sw_any_bug;
 624
 625 static int sw_any_bug_found(const struct dmi_system_id *d)
 626 {
 627         bios_with_sw_any_bug = 1;
 628         return 0;
 629 }
 630
 631 static const struct dmi_system_id sw_any_bug_dmi_table[] = {
 632         {
 633                 .callback = sw_any_bug_found,
 634                 .ident = "Supermicro Server X6DLP",
 635                 .matches = {
 636                         DMI_MATCH(DMI_SYS_VENDOR, "Supermicro"),
 637                         DMI_MATCH(DMI_BIOS_VERSION, "080010"),
 638                         DMI_MATCH(DMI_PRODUCT_NAME, "X6DLP"),
 639                 },
 640         },
 641         { }
 642 };
 643
 644 static int acpi_cpufreq_blacklist(struct cpuinfo_x86 *c)
 645 {
 646         /* Intel Xeon Processor 7100 Series Specification Update
 647          * http://www.intel.com/Assets/PDF/specupdate/314554.pdf
 648          * AL30: A Machine Check Exception (MCE) Occurring during an
 649          * Enhanced Intel SpeedStep Technology Ratio Change May Cause
 650          * Both Processor Cores to Lock Up. */
 651         if (c->x86_vendor == X86_VENDOR_INTEL) {
 652                 if ((c->x86 == 15) &&
 653                     (c->x86_model == 6) &&
 654                     (c->x86_mask == 8)) {
 655                         pr_info("Intel(R) Xeon(R) 7100 Errata AL30, processors may lock up on frequency changes: disabling acpi-cpufreq\n");
 656                         return -ENODEV;
 657                     }
 658                 }
 659         return 0;
 660 }
 661 #endif
 662
 663 static int acpi_cpufreq_cpu_init(struct cpufreq_policy *policy)
 664 {
 665         unsigned int i;
 666         unsigned int valid_states = 0;
 667         unsigned int cpu = policy->cpu;
 668         struct acpi_cpufreq_data *data;
 669         unsigned int result = 0;
 670         struct cpuinfo_x86 *c = &cpu_data(policy->cpu);
 671         struct acpi_processor_performance *perf;
 672         struct cpufreq_frequency_table *freq_table;
 673 #ifdef CONFIG_SMP
 674         static int blacklisted;
 675 #endif
 676
 677         pr_debug("acpi_cpufreq_cpu_init\n");
 678
 679 #ifdef CONFIG_SMP
 680         if (blacklisted)
 681                 return blacklisted;
 682         blacklisted = acpi_cpufreq_blacklist(c);
 683         if (blacklisted)
 684                 return blacklisted;
 685 #endif
 686
 687         data = kzalloc(sizeof(*data), GFP_KERNEL);
 688         if (!data)
 689                 return -ENOMEM;
 690
 691         if (!zalloc_cpumask_var(&data->freqdomain_cpus, GFP_KERNEL)) {
 692                 result = -ENOMEM;
 693                 goto err_free;
 694         }
 695
 696         perf = per_cpu_ptr(acpi_perf_data, cpu);
 697         data->acpi_perf_cpu = cpu;
 698         policy->driver_data = data;
 699
 700         if (cpu_has(c, X86_FEATURE_CONSTANT_TSC))
 701                 acpi_cpufreq_driver.flags |= CPUFREQ_CONST_LOOPS;
 702
 703         result = acpi_processor_register_performance(perf, cpu);
 704         if (result)
 705                 goto err_free_mask;
 706
 707         policy->shared_type = perf->shared_type;
 708
 709         /*
 710          * Will let policy->cpus know about dependency only when software
 711          * coordination is required.
 712          */
 713         if (policy->shared_type == CPUFREQ_SHARED_TYPE_ALL ||
 714             policy->shared_type == CPUFREQ_SHARED_TYPE_ANY) {
 715                 cpumask_copy(policy->cpus, perf->shared_cpu_map);
 716         }
 717         cpumask_copy(data->freqdomain_cpus, perf->shared_cpu_map);
 718
 719 #ifdef CONFIG_SMP
 720         dmi_check_system(sw_any_bug_dmi_table);
 721         if (bios_with_sw_any_bug && !policy_is_shared(policy)) {
 722                 policy->shared_type = CPUFREQ_SHARED_TYPE_ALL;
 723                 cpumask_copy(policy->cpus, topology_core_cpumask(cpu));
 724         }
 725
 726         if (check_amd_hwpstate_cpu(cpu) && !acpi_pstate_strict) {
 727                 cpumask_clear(policy->cpus);
 728                 cpumask_set_cpu(cpu, policy->cpus);
 729                 cpumask_copy(data->freqdomain_cpus,
 730                              topology_sibling_cpumask(cpu));
 731                 policy->shared_type = CPUFREQ_SHARED_TYPE_HW;
 732                 pr_info_once("overriding BIOS provided _PSD data\n");
 733         }
 734 #endif
 735
 736         /* capability check */
 737         if (perf->state_count <= 1) {
 738                 pr_debug("No P-States\n");
 739                 result = -ENODEV;
 740                 goto err_unreg;
 741         }
 742
 743         if (perf->control_register.space_id != perf->status_register.space_id) {
 744                 result = -ENODEV;
 745                 goto err_unreg;
 746         }
 747
 748         switch (perf->control_register.space_id) {
 749         case ACPI_ADR_SPACE_SYSTEM_IO:
 750                 if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD &&
 751                     boot_cpu_data.x86 == 0xf) {
 752                         pr_debug("AMD K8 systems must use native drivers.\n");
 753                         result = -ENODEV;
 754                         goto err_unreg;
 755                 }
 756                 pr_debug("SYSTEM IO addr space\n");
 757                 data->cpu_feature = SYSTEM_IO_CAPABLE;
 758                 data->cpu_freq_read = cpu_freq_read_io;
 759                 data->cpu_freq_write = cpu_freq_write_io;
 760                 break;
 761         case ACPI_ADR_SPACE_FIXED_HARDWARE:
 762                 pr_debug("HARDWARE addr space\n");
 763                 if (check_est_cpu(cpu)) {
 764                         data->cpu_feature = SYSTEM_INTEL_MSR_CAPABLE;
 765                         data->cpu_freq_read = cpu_freq_read_intel;
 766                         data->cpu_freq_write = cpu_freq_write_intel;
 767                         break;
 768                 }
 769                 if (check_amd_hwpstate_cpu(cpu)) {
 770                         data->cpu_feature = SYSTEM_AMD_MSR_CAPABLE;
 771                         data->cpu_freq_read = cpu_freq_read_amd;
 772                         data->cpu_freq_write = cpu_freq_write_amd;
 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         freq_table = kzalloc(sizeof(*freq_table) *
 785                     (perf->state_count+1), GFP_KERNEL);
 786         if (!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                 pr_info_once("P-state transition latency capped at 20 uS\n");
 805         }
 806
 807         /* table init */
 808         for (i = 0; i < perf->state_count; i++) {
 809                 if (i > 0 && perf->states[i].core_frequency >=
 810                     freq_table[valid_states-1].frequency / 1000)
 811                         continue;
 812
 813                 freq_table[valid_states].driver_data = i;
 814                 freq_table[valid_states].frequency =
 815                     perf->states[i].core_frequency * 1000;
 816                 valid_states++;
 817         }
 818         freq_table[valid_states].frequency = CPUFREQ_TABLE_END;
 819         perf->state = 0;
 820
 821         result = cpufreq_table_validate_and_show(policy, freq_table);
 822         if (result)
 823                 goto err_freqfree;
 824
 825         if (perf->states[0].core_frequency * 1000 != policy->cpuinfo.max_freq)
 826                 pr_warn(FW_WARN "P-state 0 is not max freq\n");
 827
 828         switch (perf->control_register.space_id) {
 829         case ACPI_ADR_SPACE_SYSTEM_IO:
 830                 /*
 831                  * The core will not set policy->cur, because
 832                  * cpufreq_driver->get is NULL, so we need to set it here.
 833                  * However, we have to guess it, because the current speed is
 834                  * unknown and not detectable via IO ports.
 835                  */
 836                 policy->cur = acpi_cpufreq_guess_freq(data, policy->cpu);
 837                 break;
 838         case ACPI_ADR_SPACE_FIXED_HARDWARE:
 839                 acpi_cpufreq_driver.get = get_cur_freq_on_cpu;
 840                 break;
 841         default:
 842                 break;
 843         }
 844
 845         /* notify BIOS that we exist */
 846         acpi_processor_notify_smm(THIS_MODULE);
 847
 848         pr_debug("CPU%u - ACPI performance management activated.\n", cpu);
 849         for (i = 0; i < perf->state_count; i++)
 850                 pr_debug("     %cP%d: %d MHz, %d mW, %d uS\n",
 851                         (i == perf->state ? '*' : ' '), i,
 852                         (u32) perf->states[i].core_frequency,
 853                         (u32) perf->states[i].power,
 854                         (u32) perf->states[i].transition_latency);
 855
 856         /*
 857          * the first call to ->target() should result in us actually
 858          * writing something to the appropriate registers.
 859          */
 860         data->resume = 1;
 861
 862         policy->fast_switch_possible = !acpi_pstate_strict &&
 863                 !(policy_is_shared(policy) && policy->shared_type != CPUFREQ_SHARED_TYPE_ANY);
 864
 865         return result;
 866
 867 err_freqfree:
 868         kfree(freq_table);
 869 err_unreg:
 870         acpi_processor_unregister_performance(cpu);
 871 err_free_mask:
 872         free_cpumask_var(data->freqdomain_cpus);
 873 err_free:
 874         kfree(data);
 875         policy->driver_data = NULL;
 876
 877         return result;
 878 }
 879
 880 static int acpi_cpufreq_cpu_exit(struct cpufreq_policy *policy)
 881 {
 882         struct acpi_cpufreq_data *data = policy->driver_data;
 883
 884         pr_debug("acpi_cpufreq_cpu_exit\n");
 885
 886         policy->fast_switch_possible = false;
 887         policy->driver_data = NULL;
 888         acpi_processor_unregister_performance(data->acpi_perf_cpu);
 889         free_cpumask_var(data->freqdomain_cpus);
 890         kfree(policy->freq_table);
 891         kfree(data);
 892
 893         return 0;
 894 }
 895
 896 static int acpi_cpufreq_resume(struct cpufreq_policy *policy)
 897 {
 898         struct acpi_cpufreq_data *data = policy->driver_data;
 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         &freqdomain_cpus,
 910 #ifdef CONFIG_X86_ACPI_CPUFREQ_CPB
 911         &cpb,
 912 #endif
 913         NULL,
 914 };
 915
 916 static struct cpufreq_driver acpi_cpufreq_driver = {
 917         .verify         = cpufreq_generic_frequency_table_verify,
 918         .target_index   = acpi_cpufreq_target,
 919         .fast_switch    = acpi_cpufreq_fast_switch,
 920         .bios_limit     = acpi_processor_get_bios_limit,
 921         .init           = acpi_cpufreq_cpu_init,
 922         .exit           = acpi_cpufreq_cpu_exit,
 923         .resume         = acpi_cpufreq_resume,
 924         .name           = "acpi-cpufreq",
 925         .attr           = acpi_cpufreq_attr,
 926 };
 927
 928 static void __init acpi_cpufreq_boost_init(void)
 929 {
 930         if (boot_cpu_has(X86_FEATURE_CPB) || boot_cpu_has(X86_FEATURE_IDA)) {
 931                 msrs = msrs_alloc();
 932
 933                 if (!msrs)
 934                         return;
 935
 936                 acpi_cpufreq_driver.set_boost = set_boost;
 937                 acpi_cpufreq_driver.boost_enabled = boost_state(0);
 938
 939                 cpu_notifier_register_begin();
 940
 941                 /* Force all MSRs to the same value */
 942                 boost_set_msrs(acpi_cpufreq_driver.boost_enabled,
 943                                cpu_online_mask);
 944
 945                 __register_cpu_notifier(&boost_nb);
 946
 947                 cpu_notifier_register_done();
 948         }
 949 }
 950
 951 static void acpi_cpufreq_boost_exit(void)
 952 {
 953         if (msrs) {
 954                 unregister_cpu_notifier(&boost_nb);
 955
 956                 msrs_free(msrs);
 957                 msrs = NULL;
 958         }
 959 }
 960
 961 static int __init acpi_cpufreq_init(void)
 962 {
 963         int ret;
 964
 965         if (acpi_disabled)
 966                 return -ENODEV;
 967
 968         /* don't keep reloading if cpufreq_driver exists */
 969         if (cpufreq_get_current_driver())
 970                 return -EEXIST;
 971
 972         pr_debug("acpi_cpufreq_init\n");
 973
 974         ret = acpi_cpufreq_early_init();
 975         if (ret)
 976                 return ret;
 977
 978 #ifdef CONFIG_X86_ACPI_CPUFREQ_CPB
 979         /* this is a sysfs file with a strange name and an even stranger
 980          * semantic - per CPU instantiation, but system global effect.
 981          * Lets enable it only on AMD CPUs for compatibility reasons and
 982          * only if configured. This is considered legacy code, which
 983          * will probably be removed at some point in the future.
 984          */
 985         if (!check_amd_hwpstate_cpu(0)) {
 986                 struct freq_attr **attr;
 987
 988                 pr_debug("CPB unsupported, do not expose it\n");
 989
 990                 for (attr = acpi_cpufreq_attr; *attr; attr++)
 991                         if (*attr == &cpb) {
 992                                 *attr = NULL;
 993                                 break;
 994                         }
 995         }
 996 #endif
 997         acpi_cpufreq_boost_init();
 998
 999         ret = cpufreq_register_driver(&acpi_cpufreq_driver);
1000         if (ret) {
1001                 free_acpi_perf_data();
1002                 acpi_cpufreq_boost_exit();
1003         }
1004         return ret;
1005 }
1006
1007 static void __exit acpi_cpufreq_exit(void)
1008 {
1009         pr_debug("acpi_cpufreq_exit\n");
1010
1011         acpi_cpufreq_boost_exit();
1012
1013         cpufreq_unregister_driver(&acpi_cpufreq_driver);
1014
1015         free_acpi_perf_data();
1016 }
1017
1018 module_param(acpi_pstate_strict, uint, 0644);
1019 MODULE_PARM_DESC(acpi_pstate_strict,
1020         "value 0 or non-zero. non-zero -> strict ACPI checks are "
1021         "performed during frequency changes.");
1022
1023 late_initcall(acpi_cpufreq_init);
1024 module_exit(acpi_cpufreq_exit);
1025
1026 static const struct x86_cpu_id acpi_cpufreq_ids[] = {
1027         X86_FEATURE_MATCH(X86_FEATURE_ACPI),
1028         X86_FEATURE_MATCH(X86_FEATURE_HW_PSTATE),
1029         {}
1030 };
1031 MODULE_DEVICE_TABLE(x86cpu, acpi_cpufreq_ids);
1032
1033 static const struct acpi_device_id processor_device_ids[] = {
1034         {ACPI_PROCESSOR_OBJECT_HID, },
1035         {ACPI_PROCESSOR_DEVICE_HID, },
1036         {},
1037 };
1038 MODULE_DEVICE_TABLE(acpi, processor_device_ids);
1039
1040 MODULE_ALIAS("acpi");