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 #include <asm/cpu_device_id.h>
  51
  52 MODULE_AUTHOR("Paul Diefenbaugh, Dominik Brodowski");
  53 MODULE_DESCRIPTION("ACPI Processor P-States Driver");
  54 MODULE_LICENSE("GPL");
  55
  56 enum {
  57         UNDEFINED_CAPABLE = 0,
  58         SYSTEM_INTEL_MSR_CAPABLE,
  59         SYSTEM_AMD_MSR_CAPABLE,
  60         SYSTEM_IO_CAPABLE,
  61 };
  62
  63 #define INTEL_MSR_RANGE         (0xffff)
  64 #define AMD_MSR_RANGE           (0x7)
  65
  66 #define MSR_K7_HWCR_CPB_DIS     (1ULL << 25)
  67
  68 struct acpi_cpufreq_data {
  69         unsigned int resume;
  70         unsigned int cpu_feature;
  71         unsigned int acpi_perf_cpu;
  72         cpumask_var_t freqdomain_cpus;
  73         void (*cpu_freq_write)(struct acpi_pct_register *reg, u32 val);
  74         u32 (*cpu_freq_read)(struct acpi_pct_register *reg);
  75 };
  76
  77 /* acpi_perf_data is a pointer to percpu data. */
  78 static struct acpi_processor_performance __percpu *acpi_perf_data;
  79
  80 static inline struct acpi_processor_performance *to_perf_data(struct acpi_cpufreq_data *data)
  81 {
  82         return per_cpu_ptr(acpi_perf_data, data->acpi_perf_cpu);
  83 }
  84
  85 static struct cpufreq_driver acpi_cpufreq_driver;
  86
  87 static unsigned int acpi_pstate_strict;
  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 int boost_set_msr(bool enable)
 108 {
 109         u32 msr_addr;
 110         u64 msr_mask, val;
 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 -EINVAL;
 123         }
 124
 125         rdmsrl(msr_addr, val);
 126
 127         if (enable)
 128                 val &= ~msr_mask;
 129         else
 130                 val |= msr_mask;
 131
 132         wrmsrl(msr_addr, val);
 133         return 0;
 134 }
 135
 136 static void boost_set_msr_each(void *p_en)
 137 {
 138         bool enable = (bool) p_en;
 139
 140         boost_set_msr(enable);
 141 }
 142
 143 static int set_boost(int val)
 144 {
 145         get_online_cpus();
 146         on_each_cpu(boost_set_msr_each, (void *)(long)val, 1);
 147         put_online_cpus();
 148         pr_debug("Core Boosting %sabled.\n", val ? "en" : "dis");
 149
 150         return 0;
 151 }
 152
 153 static ssize_t show_freqdomain_cpus(struct cpufreq_policy *policy, char *buf)
 154 {
 155         struct acpi_cpufreq_data *data = policy->driver_data;
 156
 157         if (unlikely(!data))
 158                 return -ENODEV;
 159
 160         return cpufreq_show_cpus(data->freqdomain_cpus, buf);
 161 }
 162
 163 cpufreq_freq_attr_ro(freqdomain_cpus);
 164
 165 #ifdef CONFIG_X86_ACPI_CPUFREQ_CPB
 166 static ssize_t store_cpb(struct cpufreq_policy *policy, const char *buf,
 167                          size_t count)
 168 {
 169         int ret;
 170         unsigned int val = 0;
 171
 172         if (!acpi_cpufreq_driver.set_boost)
 173                 return -EINVAL;
 174
 175         ret = kstrtouint(buf, 10, &val);
 176         if (ret || val > 1)
 177                 return -EINVAL;
 178
 179         set_boost(val);
 180
 181         return 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(struct cpufreq_policy *policy, u32 value)
 207 {
 208         struct acpi_cpufreq_data *data = policy->driver_data;
 209         struct acpi_processor_performance *perf;
 210         int i;
 211
 212         perf = to_perf_data(data);
 213
 214         for (i = 0; i < perf->state_count; i++) {
 215                 if (value == perf->states[i].status)
 216                         return policy->freq_table[i].frequency;
 217         }
 218         return 0;
 219 }
 220
 221 static unsigned extract_msr(struct cpufreq_policy *policy, u32 msr)
 222 {
 223         struct acpi_cpufreq_data *data = policy->driver_data;
 224         struct cpufreq_frequency_table *pos;
 225         struct acpi_processor_performance *perf;
 226
 227         if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD)
 228                 msr &= AMD_MSR_RANGE;
 229         else
 230                 msr &= INTEL_MSR_RANGE;
 231
 232         perf = to_perf_data(data);
 233
 234         cpufreq_for_each_entry(pos, policy->freq_table)
 235                 if (msr == perf->states[pos->driver_data].status)
 236                         return pos->frequency;
 237         return policy->freq_table[0].frequency;
 238 }
 239
 240 static unsigned extract_freq(struct cpufreq_policy *policy, u32 val)
 241 {
 242         struct acpi_cpufreq_data *data = policy->driver_data;
 243
 244         switch (data->cpu_feature) {
 245         case SYSTEM_INTEL_MSR_CAPABLE:
 246         case SYSTEM_AMD_MSR_CAPABLE:
 247                 return extract_msr(policy, val);
 248         case SYSTEM_IO_CAPABLE:
 249                 return extract_io(policy, val);
 250         default:
 251                 return 0;
 252         }
 253 }
 254
 255 static u32 cpu_freq_read_intel(struct acpi_pct_register *not_used)
 256 {
 257         u32 val, dummy;
 258
 259         rdmsr(MSR_IA32_PERF_CTL, val, dummy);
 260         return val;
 261 }
 262
 263 static void cpu_freq_write_intel(struct acpi_pct_register *not_used, u32 val)
 264 {
 265         u32 lo, hi;
 266
 267         rdmsr(MSR_IA32_PERF_CTL, lo, hi);
 268         lo = (lo & ~INTEL_MSR_RANGE) | (val & INTEL_MSR_RANGE);
 269         wrmsr(MSR_IA32_PERF_CTL, lo, hi);
 270 }
 271
 272 static u32 cpu_freq_read_amd(struct acpi_pct_register *not_used)
 273 {
 274         u32 val, dummy;
 275
 276         rdmsr(MSR_AMD_PERF_CTL, val, dummy);
 277         return val;
 278 }
 279
 280 static void cpu_freq_write_amd(struct acpi_pct_register *not_used, u32 val)
 281 {
 282         wrmsr(MSR_AMD_PERF_CTL, val, 0);
 283 }
 284
 285 static u32 cpu_freq_read_io(struct acpi_pct_register *reg)
 286 {
 287         u32 val;
 288
 289         acpi_os_read_port(reg->address, &val, reg->bit_width);
 290         return val;
 291 }
 292
 293 static void cpu_freq_write_io(struct acpi_pct_register *reg, u32 val)
 294 {
 295         acpi_os_write_port(reg->address, val, reg->bit_width);
 296 }
 297
 298 struct drv_cmd {
 299         struct acpi_pct_register *reg;
 300         u32 val;
 301         union {
 302                 void (*write)(struct acpi_pct_register *reg, u32 val);
 303                 u32 (*read)(struct acpi_pct_register *reg);
 304         } func;
 305 };
 306
 307 /* Called via smp_call_function_single(), on the target CPU */
 308 static void do_drv_read(void *_cmd)
 309 {
 310         struct drv_cmd *cmd = _cmd;
 311
 312         cmd->val = cmd->func.read(cmd->reg);
 313 }
 314
 315 static u32 drv_read(struct acpi_cpufreq_data *data, const struct cpumask *mask)
 316 {
 317         struct acpi_processor_performance *perf = to_perf_data(data);
 318         struct drv_cmd cmd = {
 319                 .reg = &perf->control_register,
 320                 .func.read = data->cpu_freq_read,
 321         };
 322         int err;
 323
 324         err = smp_call_function_any(mask, do_drv_read, &cmd, 1);
 325         WARN_ON_ONCE(err);      /* smp_call_function_any() was buggy? */
 326         return cmd.val;
 327 }
 328
 329 /* Called via smp_call_function_many(), on the target CPUs */
 330 static void do_drv_write(void *_cmd)
 331 {
 332         struct drv_cmd *cmd = _cmd;
 333
 334         cmd->func.write(cmd->reg, cmd->val);
 335 }
 336
 337 static void drv_write(struct acpi_cpufreq_data *data,
 338                       const struct cpumask *mask, u32 val)
 339 {
 340         struct acpi_processor_performance *perf = to_perf_data(data);
 341         struct drv_cmd cmd = {
 342                 .reg = &perf->control_register,
 343                 .val = val,
 344                 .func.write = data->cpu_freq_write,
 345         };
 346         int this_cpu;
 347
 348         this_cpu = get_cpu();
 349         if (cpumask_test_cpu(this_cpu, mask))
 350                 do_drv_write(&cmd);
 351
 352         smp_call_function_many(mask, do_drv_write, &cmd, 1);
 353         put_cpu();
 354 }
 355
 356 static u32 get_cur_val(const struct cpumask *mask, struct acpi_cpufreq_data *data)
 357 {
 358         u32 val;
 359
 360         if (unlikely(cpumask_empty(mask)))
 361                 return 0;
 362
 363         val = drv_read(data, mask);
 364
 365         pr_debug("get_cur_val = %u\n", val);
 366
 367         return val;
 368 }
 369
 370 static unsigned int get_cur_freq_on_cpu(unsigned int cpu)
 371 {
 372         struct acpi_cpufreq_data *data;
 373         struct cpufreq_policy *policy;
 374         unsigned int freq;
 375         unsigned int cached_freq;
 376
 377         pr_debug("get_cur_freq_on_cpu (%d)\n", cpu);
 378
 379         policy = cpufreq_cpu_get_raw(cpu);
 380         if (unlikely(!policy))
 381                 return 0;
 382
 383         data = policy->driver_data;
 384         if (unlikely(!data || !policy->freq_table))
 385                 return 0;
 386
 387         cached_freq = policy->freq_table[to_perf_data(data)->state].frequency;
 388         freq = extract_freq(policy, get_cur_val(cpumask_of(cpu), data));
 389         if (freq != cached_freq) {
 390                 /*
 391                  * The dreaded BIOS frequency change behind our back.
 392                  * Force set the frequency on next target call.
 393                  */
 394                 data->resume = 1;
 395         }
 396
 397         pr_debug("cur freq = %u\n", freq);
 398
 399         return freq;
 400 }
 401
 402 static unsigned int check_freqs(struct cpufreq_policy *policy,
 403                                 const struct cpumask *mask, unsigned int freq)
 404 {
 405         struct acpi_cpufreq_data *data = policy->driver_data;
 406         unsigned int cur_freq;
 407         unsigned int i;
 408
 409         for (i = 0; i < 100; i++) {
 410                 cur_freq = extract_freq(policy, get_cur_val(mask, data));
 411                 if (cur_freq == freq)
 412                         return 1;
 413                 udelay(10);
 414         }
 415         return 0;
 416 }
 417
 418 static int acpi_cpufreq_target(struct cpufreq_policy *policy,
 419                                unsigned int index)
 420 {
 421         struct acpi_cpufreq_data *data = policy->driver_data;
 422         struct acpi_processor_performance *perf;
 423         const struct cpumask *mask;
 424         unsigned int next_perf_state = 0; /* Index into perf table */
 425         int result = 0;
 426
 427         if (unlikely(!data)) {
 428                 return -ENODEV;
 429         }
 430
 431         perf = to_perf_data(data);
 432         next_perf_state = policy->freq_table[index].driver_data;
 433         if (perf->state == next_perf_state) {
 434                 if (unlikely(data->resume)) {
 435                         pr_debug("Called after resume, resetting to P%d\n",
 436                                 next_perf_state);
 437                         data->resume = 0;
 438                 } else {
 439                         pr_debug("Already at target state (P%d)\n",
 440                                 next_perf_state);
 441                         return 0;
 442                 }
 443         }
 444
 445         /*
 446          * The core won't allow CPUs to go away until the governor has been
 447          * stopped, so we can rely on the stability of policy->cpus.
 448          */
 449         mask = policy->shared_type == CPUFREQ_SHARED_TYPE_ANY ?
 450                 cpumask_of(policy->cpu) : policy->cpus;
 451
 452         drv_write(data, mask, perf->states[next_perf_state].control);
 453
 454         if (acpi_pstate_strict) {
 455                 if (!check_freqs(policy, mask,
 456                                  policy->freq_table[index].frequency)) {
 457                         pr_debug("acpi_cpufreq_target failed (%d)\n",
 458                                 policy->cpu);
 459                         result = -EAGAIN;
 460                 }
 461         }
 462
 463         if (!result)
 464                 perf->state = next_perf_state;
 465
 466         return result;
 467 }
 468
 469 static unsigned int acpi_cpufreq_fast_switch(struct cpufreq_policy *policy,
 470                                              unsigned int target_freq)
 471 {
 472         struct acpi_cpufreq_data *data = policy->driver_data;
 473         struct acpi_processor_performance *perf;
 474         struct cpufreq_frequency_table *entry;
 475         unsigned int next_perf_state, next_freq, index;
 476
 477         /*
 478          * Find the closest frequency above target_freq.
 479          */
 480         if (policy->cached_target_freq == target_freq)
 481                 index = policy->cached_resolved_idx;
 482         else
 483                 index = cpufreq_table_find_index_dl(policy, target_freq);
 484
 485         entry = &policy->freq_table[index];
 486         next_freq = entry->frequency;
 487         next_perf_state = entry->driver_data;
 488
 489         perf = to_perf_data(data);
 490         if (perf->state == next_perf_state) {
 491                 if (unlikely(data->resume))
 492                         data->resume = 0;
 493                 else
 494                         return next_freq;
 495         }
 496
 497         data->cpu_freq_write(&perf->control_register,
 498                              perf->states[next_perf_state].control);
 499         perf->state = next_perf_state;
 500         return next_freq;
 501 }
 502
 503 static unsigned long
 504 acpi_cpufreq_guess_freq(struct acpi_cpufreq_data *data, unsigned int cpu)
 505 {
 506         struct acpi_processor_performance *perf;
 507
 508         perf = to_perf_data(data);
 509         if (cpu_khz) {
 510                 /* search the closest match to cpu_khz */
 511                 unsigned int i;
 512                 unsigned long freq;
 513                 unsigned long freqn = perf->states[0].core_frequency * 1000;
 514
 515                 for (i = 0; i < (perf->state_count-1); i++) {
 516                         freq = freqn;
 517                         freqn = perf->states[i+1].core_frequency * 1000;
 518                         if ((2 * cpu_khz) > (freqn + freq)) {
 519                                 perf->state = i;
 520                                 return freq;
 521                         }
 522                 }
 523                 perf->state = perf->state_count-1;
 524                 return freqn;
 525         } else {
 526                 /* assume CPU is at P0... */
 527                 perf->state = 0;
 528                 return perf->states[0].core_frequency * 1000;
 529         }
 530 }
 531
 532 static void free_acpi_perf_data(void)
 533 {
 534         unsigned int i;
 535
 536         /* Freeing a NULL pointer is OK, and alloc_percpu zeroes. */
 537         for_each_possible_cpu(i)
 538                 free_cpumask_var(per_cpu_ptr(acpi_perf_data, i)
 539                                  ->shared_cpu_map);
 540         free_percpu(acpi_perf_data);
 541 }
 542
 543 static int cpufreq_boost_online(unsigned int cpu)
 544 {
 545         /*
 546          * On the CPU_UP path we simply keep the boost-disable flag
 547          * in sync with the current global state.
 548          */
 549         return boost_set_msr(acpi_cpufreq_driver.boost_enabled);
 550 }
 551
 552 static int cpufreq_boost_down_prep(unsigned int cpu)
 553 {
 554         /*
 555          * Clear the boost-disable bit on the CPU_DOWN path so that
 556          * this cpu cannot block the remaining ones from boosting.
 557          */
 558         return boost_set_msr(1);
 559 }
 560
 561 /*
 562  * acpi_cpufreq_early_init - initialize ACPI P-States library
 563  *
 564  * Initialize the ACPI P-States library (drivers/acpi/processor_perflib.c)
 565  * in order to determine correct frequency and voltage pairings. We can
 566  * do _PDC and _PSD and find out the processor dependency for the
 567  * actual init that will happen later...
 568  */
 569 static int __init acpi_cpufreq_early_init(void)
 570 {
 571         unsigned int i;
 572         pr_debug("acpi_cpufreq_early_init\n");
 573
 574         acpi_perf_data = alloc_percpu(struct acpi_processor_performance);
 575         if (!acpi_perf_data) {
 576                 pr_debug("Memory allocation error for acpi_perf_data.\n");
 577                 return -ENOMEM;
 578         }
 579         for_each_possible_cpu(i) {
 580                 if (!zalloc_cpumask_var_node(
 581                         &per_cpu_ptr(acpi_perf_data, i)->shared_cpu_map,
 582                         GFP_KERNEL, cpu_to_node(i))) {
 583
 584                         /* Freeing a NULL pointer is OK: alloc_percpu zeroes. */
 585                         free_acpi_perf_data();
 586                         return -ENOMEM;
 587                 }
 588         }
 589
 590         /* Do initialization in ACPI core */
 591         acpi_processor_preregister_performance(acpi_perf_data);
 592         return 0;
 593 }
 594
 595 #ifdef CONFIG_SMP
 596 /*
 597  * Some BIOSes do SW_ANY coordination internally, either set it up in hw
 598  * or do it in BIOS firmware and won't inform about it to OS. If not
 599  * detected, this has a side effect of making CPU run at a different speed
 600  * than OS intended it to run at. Detect it and handle it cleanly.
 601  */
 602 static int bios_with_sw_any_bug;
 603
 604 static int sw_any_bug_found(const struct dmi_system_id *d)
 605 {
 606         bios_with_sw_any_bug = 1;
 607         return 0;
 608 }
 609
 610 static const struct dmi_system_id sw_any_bug_dmi_table[] = {
 611         {
 612                 .callback = sw_any_bug_found,
 613                 .ident = "Supermicro Server X6DLP",
 614                 .matches = {
 615                         DMI_MATCH(DMI_SYS_VENDOR, "Supermicro"),
 616                         DMI_MATCH(DMI_BIOS_VERSION, "080010"),
 617                         DMI_MATCH(DMI_PRODUCT_NAME, "X6DLP"),
 618                 },
 619         },
 620         { }
 621 };
 622
 623 static int acpi_cpufreq_blacklist(struct cpuinfo_x86 *c)
 624 {
 625         /* Intel Xeon Processor 7100 Series Specification Update
 626          * http://www.intel.com/Assets/PDF/specupdate/314554.pdf
 627          * AL30: A Machine Check Exception (MCE) Occurring during an
 628          * Enhanced Intel SpeedStep Technology Ratio Change May Cause
 629          * Both Processor Cores to Lock Up. */
 630         if (c->x86_vendor == X86_VENDOR_INTEL) {
 631                 if ((c->x86 == 15) &&
 632                     (c->x86_model == 6) &&
 633                     (c->x86_stepping == 8)) {
 634                         pr_info("Intel(R) Xeon(R) 7100 Errata AL30, processors may lock up on frequency changes: disabling acpi-cpufreq\n");
 635                         return -ENODEV;
 636                     }
 637                 }
 638         return 0;
 639 }
 640 #endif
 641
 642 static int acpi_cpufreq_cpu_init(struct cpufreq_policy *policy)
 643 {
 644         unsigned int i;
 645         unsigned int valid_states = 0;
 646         unsigned int cpu = policy->cpu;
 647         struct acpi_cpufreq_data *data;
 648         unsigned int result = 0;
 649         struct cpuinfo_x86 *c = &cpu_data(policy->cpu);
 650         struct acpi_processor_performance *perf;
 651         struct cpufreq_frequency_table *freq_table;
 652 #ifdef CONFIG_SMP
 653         static int blacklisted;
 654 #endif
 655
 656         pr_debug("acpi_cpufreq_cpu_init\n");
 657
 658 #ifdef CONFIG_SMP
 659         if (blacklisted)
 660                 return blacklisted;
 661         blacklisted = acpi_cpufreq_blacklist(c);
 662         if (blacklisted)
 663                 return blacklisted;
 664 #endif
 665
 666         data = kzalloc(sizeof(*data), GFP_KERNEL);
 667         if (!data)
 668                 return -ENOMEM;
 669
 670         if (!zalloc_cpumask_var(&data->freqdomain_cpus, GFP_KERNEL)) {
 671                 result = -ENOMEM;
 672                 goto err_free;
 673         }
 674
 675         perf = per_cpu_ptr(acpi_perf_data, cpu);
 676         data->acpi_perf_cpu = cpu;
 677         policy->driver_data = data;
 678
 679         if (cpu_has(c, X86_FEATURE_CONSTANT_TSC))
 680                 acpi_cpufreq_driver.flags |= CPUFREQ_CONST_LOOPS;
 681
 682         result = acpi_processor_register_performance(perf, cpu);
 683         if (result)
 684                 goto err_free_mask;
 685
 686         policy->shared_type = perf->shared_type;
 687
 688         /*
 689          * Will let policy->cpus know about dependency only when software
 690          * coordination is required.
 691          */
 692         if (policy->shared_type == CPUFREQ_SHARED_TYPE_ALL ||
 693             policy->shared_type == CPUFREQ_SHARED_TYPE_ANY) {
 694                 cpumask_copy(policy->cpus, perf->shared_cpu_map);
 695         }
 696         cpumask_copy(data->freqdomain_cpus, perf->shared_cpu_map);
 697
 698 #ifdef CONFIG_SMP
 699         dmi_check_system(sw_any_bug_dmi_table);
 700         if (bios_with_sw_any_bug && !policy_is_shared(policy)) {
 701                 policy->shared_type = CPUFREQ_SHARED_TYPE_ALL;
 702                 cpumask_copy(policy->cpus, topology_core_cpumask(cpu));
 703         }
 704
 705         if (check_amd_hwpstate_cpu(cpu) && boot_cpu_data.x86 < 0x19 &&
 706             !acpi_pstate_strict) {
 707                 cpumask_clear(policy->cpus);
 708                 cpumask_set_cpu(cpu, policy->cpus);
 709                 cpumask_copy(data->freqdomain_cpus,
 710                              topology_sibling_cpumask(cpu));
 711                 policy->shared_type = CPUFREQ_SHARED_TYPE_HW;
 712                 pr_info_once("overriding BIOS provided _PSD data\n");
 713         }
 714 #endif
 715
 716         /* capability check */
 717         if (perf->state_count <= 1) {
 718                 pr_debug("No P-States\n");
 719                 result = -ENODEV;
 720                 goto err_unreg;
 721         }
 722
 723         if (perf->control_register.space_id != perf->status_register.space_id) {
 724                 result = -ENODEV;
 725                 goto err_unreg;
 726         }
 727
 728         switch (perf->control_register.space_id) {
 729         case ACPI_ADR_SPACE_SYSTEM_IO:
 730                 if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD &&
 731                     boot_cpu_data.x86 == 0xf) {
 732                         pr_debug("AMD K8 systems must use native drivers.\n");
 733                         result = -ENODEV;
 734                         goto err_unreg;
 735                 }
 736                 pr_debug("SYSTEM IO addr space\n");
 737                 data->cpu_feature = SYSTEM_IO_CAPABLE;
 738                 data->cpu_freq_read = cpu_freq_read_io;
 739                 data->cpu_freq_write = cpu_freq_write_io;
 740                 break;
 741         case ACPI_ADR_SPACE_FIXED_HARDWARE:
 742                 pr_debug("HARDWARE addr space\n");
 743                 if (check_est_cpu(cpu)) {
 744                         data->cpu_feature = SYSTEM_INTEL_MSR_CAPABLE;
 745                         data->cpu_freq_read = cpu_freq_read_intel;
 746                         data->cpu_freq_write = cpu_freq_write_intel;
 747                         break;
 748                 }
 749                 if (check_amd_hwpstate_cpu(cpu)) {
 750                         data->cpu_feature = SYSTEM_AMD_MSR_CAPABLE;
 751                         data->cpu_freq_read = cpu_freq_read_amd;
 752                         data->cpu_freq_write = cpu_freq_write_amd;
 753                         break;
 754                 }
 755                 result = -ENODEV;
 756                 goto err_unreg;
 757         default:
 758                 pr_debug("Unknown addr space %d\n",
 759                         (u32) (perf->control_register.space_id));
 760                 result = -ENODEV;
 761                 goto err_unreg;
 762         }
 763
 764         freq_table = kcalloc(perf->state_count + 1, sizeof(*freq_table),
 765                              GFP_KERNEL);
 766         if (!freq_table) {
 767                 result = -ENOMEM;
 768                 goto err_unreg;
 769         }
 770
 771         /* detect transition latency */
 772         policy->cpuinfo.transition_latency = 0;
 773         for (i = 0; i < perf->state_count; i++) {
 774                 if ((perf->states[i].transition_latency * 1000) >
 775                     policy->cpuinfo.transition_latency)
 776                         policy->cpuinfo.transition_latency =
 777                             perf->states[i].transition_latency * 1000;
 778         }
 779
 780         /* Check for high latency (>20uS) from buggy BIOSes, like on T42 */
 781         if (perf->control_register.space_id == ACPI_ADR_SPACE_FIXED_HARDWARE &&
 782             policy->cpuinfo.transition_latency > 20 * 1000) {
 783                 policy->cpuinfo.transition_latency = 20 * 1000;
 784                 pr_info_once("P-state transition latency capped at 20 uS\n");
 785         }
 786
 787         /* table init */
 788         for (i = 0; i < perf->state_count; i++) {
 789                 if (i > 0 && perf->states[i].core_frequency >=
 790                     freq_table[valid_states-1].frequency / 1000)
 791                         continue;
 792
 793                 freq_table[valid_states].driver_data = i;
 794                 freq_table[valid_states].frequency =
 795                     perf->states[i].core_frequency * 1000;
 796                 valid_states++;
 797         }
 798         freq_table[valid_states].frequency = CPUFREQ_TABLE_END;
 799         policy->freq_table = freq_table;
 800         perf->state = 0;
 801
 802         switch (perf->control_register.space_id) {
 803         case ACPI_ADR_SPACE_SYSTEM_IO:
 804                 /*
 805                  * The core will not set policy->cur, because
 806                  * cpufreq_driver->get is NULL, so we need to set it here.
 807                  * However, we have to guess it, because the current speed is
 808                  * unknown and not detectable via IO ports.
 809                  */
 810                 policy->cur = acpi_cpufreq_guess_freq(data, policy->cpu);
 811                 break;
 812         case ACPI_ADR_SPACE_FIXED_HARDWARE:
 813                 acpi_cpufreq_driver.get = get_cur_freq_on_cpu;
 814                 break;
 815         default:
 816                 break;
 817         }
 818
 819         /* notify BIOS that we exist */
 820         acpi_processor_notify_smm(THIS_MODULE);
 821
 822         pr_debug("CPU%u - ACPI performance management activated.\n", cpu);
 823         for (i = 0; i < perf->state_count; i++)
 824                 pr_debug("     %cP%d: %d MHz, %d mW, %d uS\n",
 825                         (i == perf->state ? '*' : ' '), i,
 826                         (u32) perf->states[i].core_frequency,
 827                         (u32) perf->states[i].power,
 828                         (u32) perf->states[i].transition_latency);
 829
 830         /*
 831          * the first call to ->target() should result in us actually
 832          * writing something to the appropriate registers.
 833          */
 834         data->resume = 1;
 835
 836         policy->fast_switch_possible = !acpi_pstate_strict &&
 837                 !(policy_is_shared(policy) && policy->shared_type != CPUFREQ_SHARED_TYPE_ANY);
 838
 839         return result;
 840
 841 err_unreg:
 842         acpi_processor_unregister_performance(cpu);
 843 err_free_mask:
 844         free_cpumask_var(data->freqdomain_cpus);
 845 err_free:
 846         kfree(data);
 847         policy->driver_data = NULL;
 848
 849         return result;
 850 }
 851
 852 static int acpi_cpufreq_cpu_exit(struct cpufreq_policy *policy)
 853 {
 854         struct acpi_cpufreq_data *data = policy->driver_data;
 855
 856         pr_debug("acpi_cpufreq_cpu_exit\n");
 857
 858         policy->fast_switch_possible = false;
 859         policy->driver_data = NULL;
 860         acpi_processor_unregister_performance(data->acpi_perf_cpu);
 861         free_cpumask_var(data->freqdomain_cpus);
 862         kfree(policy->freq_table);
 863         kfree(data);
 864
 865         return 0;
 866 }
 867
 868 static void acpi_cpufreq_cpu_ready(struct cpufreq_policy *policy)
 869 {
 870         struct acpi_processor_performance *perf = per_cpu_ptr(acpi_perf_data,
 871                                                               policy->cpu);
 872
 873         if (perf->states[0].core_frequency * 1000 != policy->cpuinfo.max_freq)
 874                 pr_warn(FW_WARN "P-state 0 is not max freq\n");
 875 }
 876
 877 static int acpi_cpufreq_resume(struct cpufreq_policy *policy)
 878 {
 879         struct acpi_cpufreq_data *data = policy->driver_data;
 880
 881         pr_debug("acpi_cpufreq_resume\n");
 882
 883         data->resume = 1;
 884
 885         return 0;
 886 }
 887
 888 static struct freq_attr *acpi_cpufreq_attr[] = {
 889         &cpufreq_freq_attr_scaling_available_freqs,
 890         &freqdomain_cpus,
 891 #ifdef CONFIG_X86_ACPI_CPUFREQ_CPB
 892         &cpb,
 893 #endif
 894         NULL,
 895 };
 896
 897 static struct cpufreq_driver acpi_cpufreq_driver = {
 898         .verify         = cpufreq_generic_frequency_table_verify,
 899         .target_index   = acpi_cpufreq_target,
 900         .fast_switch    = acpi_cpufreq_fast_switch,
 901         .bios_limit     = acpi_processor_get_bios_limit,
 902         .init           = acpi_cpufreq_cpu_init,
 903         .exit           = acpi_cpufreq_cpu_exit,
 904         .ready          = acpi_cpufreq_cpu_ready,
 905         .resume         = acpi_cpufreq_resume,
 906         .name           = "acpi-cpufreq",
 907         .attr           = acpi_cpufreq_attr,
 908 };
 909
 910 static enum cpuhp_state acpi_cpufreq_online;
 911
 912 static void __init acpi_cpufreq_boost_init(void)
 913 {
 914         int ret;
 915
 916         if (!(boot_cpu_has(X86_FEATURE_CPB) || boot_cpu_has(X86_FEATURE_IDA))) {
 917                 pr_debug("Boost capabilities not present in the processor\n");
 918                 return;
 919         }
 920
 921         acpi_cpufreq_driver.set_boost = set_boost;
 922         acpi_cpufreq_driver.boost_enabled = boost_state(0);
 923
 924         /*
 925          * This calls the online callback on all online cpu and forces all
 926          * MSRs to the same value.
 927          */
 928         ret = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "cpufreq/acpi:online",
 929                                 cpufreq_boost_online, cpufreq_boost_down_prep);
 930         if (ret < 0) {
 931                 pr_err("acpi_cpufreq: failed to register hotplug callbacks\n");
 932                 return;
 933         }
 934         acpi_cpufreq_online = ret;
 935 }
 936
 937 static void acpi_cpufreq_boost_exit(void)
 938 {
 939         if (acpi_cpufreq_online > 0)
 940                 cpuhp_remove_state_nocalls(acpi_cpufreq_online);
 941 }
 942
 943 static int __init acpi_cpufreq_init(void)
 944 {
 945         int ret;
 946
 947         if (acpi_disabled)
 948                 return -ENODEV;
 949
 950         /* don't keep reloading if cpufreq_driver exists */
 951         if (cpufreq_get_current_driver())
 952                 return -EEXIST;
 953
 954         pr_debug("acpi_cpufreq_init\n");
 955
 956         ret = acpi_cpufreq_early_init();
 957         if (ret)
 958                 return ret;
 959
 960 #ifdef CONFIG_X86_ACPI_CPUFREQ_CPB
 961         /* this is a sysfs file with a strange name and an even stranger
 962          * semantic - per CPU instantiation, but system global effect.
 963          * Lets enable it only on AMD CPUs for compatibility reasons and
 964          * only if configured. This is considered legacy code, which
 965          * will probably be removed at some point in the future.
 966          */
 967         if (!check_amd_hwpstate_cpu(0)) {
 968                 struct freq_attr **attr;
 969
 970                 pr_debug("CPB unsupported, do not expose it\n");
 971
 972                 for (attr = acpi_cpufreq_attr; *attr; attr++)
 973                         if (*attr == &cpb) {
 974                                 *attr = NULL;
 975                                 break;
 976                         }
 977         }
 978 #endif
 979         acpi_cpufreq_boost_init();
 980
 981         ret = cpufreq_register_driver(&acpi_cpufreq_driver);
 982         if (ret) {
 983                 free_acpi_perf_data();
 984                 acpi_cpufreq_boost_exit();
 985         }
 986         return ret;
 987 }
 988
 989 static void __exit acpi_cpufreq_exit(void)
 990 {
 991         pr_debug("acpi_cpufreq_exit\n");
 992
 993         acpi_cpufreq_boost_exit();
 994
 995         cpufreq_unregister_driver(&acpi_cpufreq_driver);
 996
 997         free_acpi_perf_data();
 998 }
 999
1000 module_param(acpi_pstate_strict, uint, 0644);
1001 MODULE_PARM_DESC(acpi_pstate_strict,
1002         "value 0 or non-zero. non-zero -> strict ACPI checks are "
1003         "performed during frequency changes.");
1004
1005 late_initcall(acpi_cpufreq_init);
1006 module_exit(acpi_cpufreq_exit);
1007
1008 static const struct x86_cpu_id acpi_cpufreq_ids[] = {
1009         X86_FEATURE_MATCH(X86_FEATURE_ACPI),
1010         X86_FEATURE_MATCH(X86_FEATURE_HW_PSTATE),
1011         {}
1012 };
1013 MODULE_DEVICE_TABLE(x86cpu, acpi_cpufreq_ids);
1014
1015 static const struct acpi_device_id processor_device_ids[] = {
1016         {ACPI_PROCESSOR_OBJECT_HID, },
1017         {ACPI_PROCESSOR_DEVICE_HID, },
1018         {},
1019 };
1020 MODULE_DEVICE_TABLE(acpi, processor_device_ids);
1021
1022 MODULE_ALIAS("acpi");