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, index;
 472
 473         /*
 474          * Find the closest frequency above target_freq.
 475          */
 476         index = cpufreq_table_find_index_dl(policy, target_freq);
 477
 478         entry = &policy->freq_table[index];
 479         next_freq = entry->frequency;
 480         next_perf_state = entry->driver_data;
 481
 482         perf = to_perf_data(data);
 483         if (perf->state == next_perf_state) {
 484                 if (unlikely(data->resume))
 485                         data->resume = 0;
 486                 else
 487                         return next_freq;
 488         }
 489
 490         data->cpu_freq_write(&perf->control_register,
 491                              perf->states[next_perf_state].control);
 492         perf->state = next_perf_state;
 493         return next_freq;
 494 }
 495
 496 static unsigned long
 497 acpi_cpufreq_guess_freq(struct acpi_cpufreq_data *data, unsigned int cpu)
 498 {
 499         struct acpi_processor_performance *perf;
 500
 501         perf = to_perf_data(data);
 502         if (cpu_khz) {
 503                 /* search the closest match to cpu_khz */
 504                 unsigned int i;
 505                 unsigned long freq;
 506                 unsigned long freqn = perf->states[0].core_frequency * 1000;
 507
 508                 for (i = 0; i < (perf->state_count-1); i++) {
 509                         freq = freqn;
 510                         freqn = perf->states[i+1].core_frequency * 1000;
 511                         if ((2 * cpu_khz) > (freqn + freq)) {
 512                                 perf->state = i;
 513                                 return freq;
 514                         }
 515                 }
 516                 perf->state = perf->state_count-1;
 517                 return freqn;
 518         } else {
 519                 /* assume CPU is at P0... */
 520                 perf->state = 0;
 521                 return perf->states[0].core_frequency * 1000;
 522         }
 523 }
 524
 525 static void free_acpi_perf_data(void)
 526 {
 527         unsigned int i;
 528
 529         /* Freeing a NULL pointer is OK, and alloc_percpu zeroes. */
 530         for_each_possible_cpu(i)
 531                 free_cpumask_var(per_cpu_ptr(acpi_perf_data, i)
 532                                  ->shared_cpu_map);
 533         free_percpu(acpi_perf_data);
 534 }
 535
 536 static int boost_notify(struct notifier_block *nb, unsigned long action,
 537                       void *hcpu)
 538 {
 539         unsigned cpu = (long)hcpu;
 540         const struct cpumask *cpumask;
 541
 542         cpumask = get_cpu_mask(cpu);
 543
 544         /*
 545          * Clear the boost-disable bit on the CPU_DOWN path so that
 546          * this cpu cannot block the remaining ones from boosting. On
 547          * the CPU_UP path we simply keep the boost-disable flag in
 548          * sync with the current global state.
 549          */
 550
 551         switch (action) {
 552         case CPU_DOWN_FAILED:
 553         case CPU_DOWN_FAILED_FROZEN:
 554         case CPU_ONLINE:
 555         case CPU_ONLINE_FROZEN:
 556                 boost_set_msrs(acpi_cpufreq_driver.boost_enabled, cpumask);
 557                 break;
 558
 559         case CPU_DOWN_PREPARE:
 560         case CPU_DOWN_PREPARE_FROZEN:
 561                 boost_set_msrs(1, cpumask);
 562                 break;
 563
 564         default:
 565                 break;
 566         }
 567
 568         return NOTIFY_OK;
 569 }
 570
 571
 572 static struct notifier_block boost_nb = {
 573         .notifier_call          = boost_notify,
 574 };
 575
 576 /*
 577  * acpi_cpufreq_early_init - initialize ACPI P-States library
 578  *
 579  * Initialize the ACPI P-States library (drivers/acpi/processor_perflib.c)
 580  * in order to determine correct frequency and voltage pairings. We can
 581  * do _PDC and _PSD and find out the processor dependency for the
 582  * actual init that will happen later...
 583  */
 584 static int __init acpi_cpufreq_early_init(void)
 585 {
 586         unsigned int i;
 587         pr_debug("acpi_cpufreq_early_init\n");
 588
 589         acpi_perf_data = alloc_percpu(struct acpi_processor_performance);
 590         if (!acpi_perf_data) {
 591                 pr_debug("Memory allocation error for acpi_perf_data.\n");
 592                 return -ENOMEM;
 593         }
 594         for_each_possible_cpu(i) {
 595                 if (!zalloc_cpumask_var_node(
 596                         &per_cpu_ptr(acpi_perf_data, i)->shared_cpu_map,
 597                         GFP_KERNEL, cpu_to_node(i))) {
 598
 599                         /* Freeing a NULL pointer is OK: alloc_percpu zeroes. */
 600                         free_acpi_perf_data();
 601                         return -ENOMEM;
 602                 }
 603         }
 604
 605         /* Do initialization in ACPI core */
 606         acpi_processor_preregister_performance(acpi_perf_data);
 607         return 0;
 608 }
 609
 610 #ifdef CONFIG_SMP
 611 /*
 612  * Some BIOSes do SW_ANY coordination internally, either set it up in hw
 613  * or do it in BIOS firmware and won't inform about it to OS. If not
 614  * detected, this has a side effect of making CPU run at a different speed
 615  * than OS intended it to run at. Detect it and handle it cleanly.
 616  */
 617 static int bios_with_sw_any_bug;
 618
 619 static int sw_any_bug_found(const struct dmi_system_id *d)
 620 {
 621         bios_with_sw_any_bug = 1;
 622         return 0;
 623 }
 624
 625 static const struct dmi_system_id sw_any_bug_dmi_table[] = {
 626         {
 627                 .callback = sw_any_bug_found,
 628                 .ident = "Supermicro Server X6DLP",
 629                 .matches = {
 630                         DMI_MATCH(DMI_SYS_VENDOR, "Supermicro"),
 631                         DMI_MATCH(DMI_BIOS_VERSION, "080010"),
 632                         DMI_MATCH(DMI_PRODUCT_NAME, "X6DLP"),
 633                 },
 634         },
 635         { }
 636 };
 637
 638 static int acpi_cpufreq_blacklist(struct cpuinfo_x86 *c)
 639 {
 640         /* Intel Xeon Processor 7100 Series Specification Update
 641          * http://www.intel.com/Assets/PDF/specupdate/314554.pdf
 642          * AL30: A Machine Check Exception (MCE) Occurring during an
 643          * Enhanced Intel SpeedStep Technology Ratio Change May Cause
 644          * Both Processor Cores to Lock Up. */
 645         if (c->x86_vendor == X86_VENDOR_INTEL) {
 646                 if ((c->x86 == 15) &&
 647                     (c->x86_model == 6) &&
 648                     (c->x86_mask == 8)) {
 649                         pr_info("Intel(R) Xeon(R) 7100 Errata AL30, processors may lock up on frequency changes: disabling acpi-cpufreq\n");
 650                         return -ENODEV;
 651                     }
 652                 }
 653         return 0;
 654 }
 655 #endif
 656
 657 static int acpi_cpufreq_cpu_init(struct cpufreq_policy *policy)
 658 {
 659         unsigned int i;
 660         unsigned int valid_states = 0;
 661         unsigned int cpu = policy->cpu;
 662         struct acpi_cpufreq_data *data;
 663         unsigned int result = 0;
 664         struct cpuinfo_x86 *c = &cpu_data(policy->cpu);
 665         struct acpi_processor_performance *perf;
 666         struct cpufreq_frequency_table *freq_table;
 667 #ifdef CONFIG_SMP
 668         static int blacklisted;
 669 #endif
 670
 671         pr_debug("acpi_cpufreq_cpu_init\n");
 672
 673 #ifdef CONFIG_SMP
 674         if (blacklisted)
 675                 return blacklisted;
 676         blacklisted = acpi_cpufreq_blacklist(c);
 677         if (blacklisted)
 678                 return blacklisted;
 679 #endif
 680
 681         data = kzalloc(sizeof(*data), GFP_KERNEL);
 682         if (!data)
 683                 return -ENOMEM;
 684
 685         if (!zalloc_cpumask_var(&data->freqdomain_cpus, GFP_KERNEL)) {
 686                 result = -ENOMEM;
 687                 goto err_free;
 688         }
 689
 690         perf = per_cpu_ptr(acpi_perf_data, cpu);
 691         data->acpi_perf_cpu = cpu;
 692         policy->driver_data = data;
 693
 694         if (cpu_has(c, X86_FEATURE_CONSTANT_TSC))
 695                 acpi_cpufreq_driver.flags |= CPUFREQ_CONST_LOOPS;
 696
 697         result = acpi_processor_register_performance(perf, cpu);
 698         if (result)
 699                 goto err_free_mask;
 700
 701         policy->shared_type = perf->shared_type;
 702
 703         /*
 704          * Will let policy->cpus know about dependency only when software
 705          * coordination is required.
 706          */
 707         if (policy->shared_type == CPUFREQ_SHARED_TYPE_ALL ||
 708             policy->shared_type == CPUFREQ_SHARED_TYPE_ANY) {
 709                 cpumask_copy(policy->cpus, perf->shared_cpu_map);
 710         }
 711         cpumask_copy(data->freqdomain_cpus, perf->shared_cpu_map);
 712
 713 #ifdef CONFIG_SMP
 714         dmi_check_system(sw_any_bug_dmi_table);
 715         if (bios_with_sw_any_bug && !policy_is_shared(policy)) {
 716                 policy->shared_type = CPUFREQ_SHARED_TYPE_ALL;
 717                 cpumask_copy(policy->cpus, topology_core_cpumask(cpu));
 718         }
 719
 720         if (check_amd_hwpstate_cpu(cpu) && !acpi_pstate_strict) {
 721                 cpumask_clear(policy->cpus);
 722                 cpumask_set_cpu(cpu, policy->cpus);
 723                 cpumask_copy(data->freqdomain_cpus,
 724                              topology_sibling_cpumask(cpu));
 725                 policy->shared_type = CPUFREQ_SHARED_TYPE_HW;
 726                 pr_info_once("overriding BIOS provided _PSD data\n");
 727         }
 728 #endif
 729
 730         /* capability check */
 731         if (perf->state_count <= 1) {
 732                 pr_debug("No P-States\n");
 733                 result = -ENODEV;
 734                 goto err_unreg;
 735         }
 736
 737         if (perf->control_register.space_id != perf->status_register.space_id) {
 738                 result = -ENODEV;
 739                 goto err_unreg;
 740         }
 741
 742         switch (perf->control_register.space_id) {
 743         case ACPI_ADR_SPACE_SYSTEM_IO:
 744                 if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD &&
 745                     boot_cpu_data.x86 == 0xf) {
 746                         pr_debug("AMD K8 systems must use native drivers.\n");
 747                         result = -ENODEV;
 748                         goto err_unreg;
 749                 }
 750                 pr_debug("SYSTEM IO addr space\n");
 751                 data->cpu_feature = SYSTEM_IO_CAPABLE;
 752                 data->cpu_freq_read = cpu_freq_read_io;
 753                 data->cpu_freq_write = cpu_freq_write_io;
 754                 break;
 755         case ACPI_ADR_SPACE_FIXED_HARDWARE:
 756                 pr_debug("HARDWARE addr space\n");
 757                 if (check_est_cpu(cpu)) {
 758                         data->cpu_feature = SYSTEM_INTEL_MSR_CAPABLE;
 759                         data->cpu_freq_read = cpu_freq_read_intel;
 760                         data->cpu_freq_write = cpu_freq_write_intel;
 761                         break;
 762                 }
 763                 if (check_amd_hwpstate_cpu(cpu)) {
 764                         data->cpu_feature = SYSTEM_AMD_MSR_CAPABLE;
 765                         data->cpu_freq_read = cpu_freq_read_amd;
 766                         data->cpu_freq_write = cpu_freq_write_amd;
 767                         break;
 768                 }
 769                 result = -ENODEV;
 770                 goto err_unreg;
 771         default:
 772                 pr_debug("Unknown addr space %d\n",
 773                         (u32) (perf->control_register.space_id));
 774                 result = -ENODEV;
 775                 goto err_unreg;
 776         }
 777
 778         freq_table = kzalloc(sizeof(*freq_table) *
 779                     (perf->state_count+1), GFP_KERNEL);
 780         if (!freq_table) {
 781                 result = -ENOMEM;
 782                 goto err_unreg;
 783         }
 784
 785         /* detect transition latency */
 786         policy->cpuinfo.transition_latency = 0;
 787         for (i = 0; i < perf->state_count; i++) {
 788                 if ((perf->states[i].transition_latency * 1000) >
 789                     policy->cpuinfo.transition_latency)
 790                         policy->cpuinfo.transition_latency =
 791                             perf->states[i].transition_latency * 1000;
 792         }
 793
 794         /* Check for high latency (>20uS) from buggy BIOSes, like on T42 */
 795         if (perf->control_register.space_id == ACPI_ADR_SPACE_FIXED_HARDWARE &&
 796             policy->cpuinfo.transition_latency > 20 * 1000) {
 797                 policy->cpuinfo.transition_latency = 20 * 1000;
 798                 pr_info_once("P-state transition latency capped at 20 uS\n");
 799         }
 800
 801         /* table init */
 802         for (i = 0; i < perf->state_count; i++) {
 803                 if (i > 0 && perf->states[i].core_frequency >=
 804                     freq_table[valid_states-1].frequency / 1000)
 805                         continue;
 806
 807                 freq_table[valid_states].driver_data = i;
 808                 freq_table[valid_states].frequency =
 809                     perf->states[i].core_frequency * 1000;
 810                 valid_states++;
 811         }
 812         freq_table[valid_states].frequency = CPUFREQ_TABLE_END;
 813         perf->state = 0;
 814
 815         result = cpufreq_table_validate_and_show(policy, freq_table);
 816         if (result)
 817                 goto err_freqfree;
 818
 819         if (perf->states[0].core_frequency * 1000 != policy->cpuinfo.max_freq)
 820                 pr_warn(FW_WARN "P-state 0 is not max freq\n");
 821
 822         switch (perf->control_register.space_id) {
 823         case ACPI_ADR_SPACE_SYSTEM_IO:
 824                 /*
 825                  * The core will not set policy->cur, because
 826                  * cpufreq_driver->get is NULL, so we need to set it here.
 827                  * However, we have to guess it, because the current speed is
 828                  * unknown and not detectable via IO ports.
 829                  */
 830                 policy->cur = acpi_cpufreq_guess_freq(data, policy->cpu);
 831                 break;
 832         case ACPI_ADR_SPACE_FIXED_HARDWARE:
 833                 acpi_cpufreq_driver.get = get_cur_freq_on_cpu;
 834                 break;
 835         default:
 836                 break;
 837         }
 838
 839         /* notify BIOS that we exist */
 840         acpi_processor_notify_smm(THIS_MODULE);
 841
 842         pr_debug("CPU%u - ACPI performance management activated.\n", cpu);
 843         for (i = 0; i < perf->state_count; i++)
 844                 pr_debug("     %cP%d: %d MHz, %d mW, %d uS\n",
 845                         (i == perf->state ? '*' : ' '), i,
 846                         (u32) perf->states[i].core_frequency,
 847                         (u32) perf->states[i].power,
 848                         (u32) perf->states[i].transition_latency);
 849
 850         /*
 851          * the first call to ->target() should result in us actually
 852          * writing something to the appropriate registers.
 853          */
 854         data->resume = 1;
 855
 856         policy->fast_switch_possible = !acpi_pstate_strict &&
 857                 !(policy_is_shared(policy) && policy->shared_type != CPUFREQ_SHARED_TYPE_ANY);
 858
 859         return result;
 860
 861 err_freqfree:
 862         kfree(freq_table);
 863 err_unreg:
 864         acpi_processor_unregister_performance(cpu);
 865 err_free_mask:
 866         free_cpumask_var(data->freqdomain_cpus);
 867 err_free:
 868         kfree(data);
 869         policy->driver_data = NULL;
 870
 871         return result;
 872 }
 873
 874 static int acpi_cpufreq_cpu_exit(struct cpufreq_policy *policy)
 875 {
 876         struct acpi_cpufreq_data *data = policy->driver_data;
 877
 878         pr_debug("acpi_cpufreq_cpu_exit\n");
 879
 880         policy->fast_switch_possible = false;
 881         policy->driver_data = NULL;
 882         acpi_processor_unregister_performance(data->acpi_perf_cpu);
 883         free_cpumask_var(data->freqdomain_cpus);
 884         kfree(policy->freq_table);
 885         kfree(data);
 886
 887         return 0;
 888 }
 889
 890 static int acpi_cpufreq_resume(struct cpufreq_policy *policy)
 891 {
 892         struct acpi_cpufreq_data *data = policy->driver_data;
 893
 894         pr_debug("acpi_cpufreq_resume\n");
 895
 896         data->resume = 1;
 897
 898         return 0;
 899 }
 900
 901 static struct freq_attr *acpi_cpufreq_attr[] = {
 902         &cpufreq_freq_attr_scaling_available_freqs,
 903         &freqdomain_cpus,
 904 #ifdef CONFIG_X86_ACPI_CPUFREQ_CPB
 905         &cpb,
 906 #endif
 907         NULL,
 908 };
 909
 910 static struct cpufreq_driver acpi_cpufreq_driver = {
 911         .verify         = cpufreq_generic_frequency_table_verify,
 912         .target_index   = acpi_cpufreq_target,
 913         .fast_switch    = acpi_cpufreq_fast_switch,
 914         .bios_limit     = acpi_processor_get_bios_limit,
 915         .init           = acpi_cpufreq_cpu_init,
 916         .exit           = acpi_cpufreq_cpu_exit,
 917         .resume         = acpi_cpufreq_resume,
 918         .name           = "acpi-cpufreq",
 919         .attr           = acpi_cpufreq_attr,
 920 };
 921
 922 static void __init acpi_cpufreq_boost_init(void)
 923 {
 924         if (boot_cpu_has(X86_FEATURE_CPB) || boot_cpu_has(X86_FEATURE_IDA)) {
 925                 msrs = msrs_alloc();
 926
 927                 if (!msrs)
 928                         return;
 929
 930                 acpi_cpufreq_driver.set_boost = set_boost;
 931                 acpi_cpufreq_driver.boost_enabled = boost_state(0);
 932
 933                 cpu_notifier_register_begin();
 934
 935                 /* Force all MSRs to the same value */
 936                 boost_set_msrs(acpi_cpufreq_driver.boost_enabled,
 937                                cpu_online_mask);
 938
 939                 __register_cpu_notifier(&boost_nb);
 940
 941                 cpu_notifier_register_done();
 942         }
 943 }
 944
 945 static void acpi_cpufreq_boost_exit(void)
 946 {
 947         if (msrs) {
 948                 unregister_cpu_notifier(&boost_nb);
 949
 950                 msrs_free(msrs);
 951                 msrs = NULL;
 952         }
 953 }
 954
 955 static int __init acpi_cpufreq_init(void)
 956 {
 957         int ret;
 958
 959         if (acpi_disabled)
 960                 return -ENODEV;
 961
 962         /* don't keep reloading if cpufreq_driver exists */
 963         if (cpufreq_get_current_driver())
 964                 return -EEXIST;
 965
 966         pr_debug("acpi_cpufreq_init\n");
 967
 968         ret = acpi_cpufreq_early_init();
 969         if (ret)
 970                 return ret;
 971
 972 #ifdef CONFIG_X86_ACPI_CPUFREQ_CPB
 973         /* this is a sysfs file with a strange name and an even stranger
 974          * semantic - per CPU instantiation, but system global effect.
 975          * Lets enable it only on AMD CPUs for compatibility reasons and
 976          * only if configured. This is considered legacy code, which
 977          * will probably be removed at some point in the future.
 978          */
 979         if (!check_amd_hwpstate_cpu(0)) {
 980                 struct freq_attr **attr;
 981
 982                 pr_debug("CPB unsupported, do not expose it\n");
 983
 984                 for (attr = acpi_cpufreq_attr; *attr; attr++)
 985                         if (*attr == &cpb) {
 986                                 *attr = NULL;
 987                                 break;
 988                         }
 989         }
 990 #endif
 991         acpi_cpufreq_boost_init();
 992
 993         ret = cpufreq_register_driver(&acpi_cpufreq_driver);
 994         if (ret) {
 995                 free_acpi_perf_data();
 996                 acpi_cpufreq_boost_exit();
 997         }
 998         return ret;
 999 }
1000
1001 static void __exit acpi_cpufreq_exit(void)
1002 {
1003         pr_debug("acpi_cpufreq_exit\n");
1004
1005         acpi_cpufreq_boost_exit();
1006
1007         cpufreq_unregister_driver(&acpi_cpufreq_driver);
1008
1009         free_acpi_perf_data();
1010 }
1011
1012 module_param(acpi_pstate_strict, uint, 0644);
1013 MODULE_PARM_DESC(acpi_pstate_strict,
1014         "value 0 or non-zero. non-zero -> strict ACPI checks are "
1015         "performed during frequency changes.");
1016
1017 late_initcall(acpi_cpufreq_init);
1018 module_exit(acpi_cpufreq_exit);
1019
1020 static const struct x86_cpu_id acpi_cpufreq_ids[] = {
1021         X86_FEATURE_MATCH(X86_FEATURE_ACPI),
1022         X86_FEATURE_MATCH(X86_FEATURE_HW_PSTATE),
1023         {}
1024 };
1025 MODULE_DEVICE_TABLE(x86cpu, acpi_cpufreq_ids);
1026
1027 static const struct acpi_device_id processor_device_ids[] = {
1028         {ACPI_PROCESSOR_OBJECT_HID, },
1029         {ACPI_PROCESSOR_DEVICE_HID, },
1030         {},
1031 };
1032 MODULE_DEVICE_TABLE(acpi, processor_device_ids);
1033
1034 MODULE_ALIAS("acpi");