arch/x86/kernel/cpu/cpufreq/acpi-cpufreq.c

   1 /*
   2  * acpi-cpufreq.c - ACPI Processor P-States Driver
   3  *
   4  *  Copyright (C) 2001, 2002 Andy Grover <andrew.grover@intel.com>
   5  *  Copyright (C) 2001, 2002 Paul Diefenbaugh <paul.s.diefenbaugh@intel.com>
   6  *  Copyright (C) 2002 - 2004 Dominik Brodowski <linux@brodo.de>
   7  *  Copyright (C) 2006       Denis Sadykov <denis.m.sadykov@intel.com>
   8  *
   9  * ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
  10  *
  11  *  This program is free software; you can redistribute it and/or modify
  12  *  it under the terms of the GNU General Public License as published by
  13  *  the Free Software Foundation; either version 2 of the License, or (at
  14  *  your option) any later version.
  15  *
  16  *  This program is distributed in the hope that it will be useful, but
  17  *  WITHOUT ANY WARRANTY; without even the implied warranty of
  18  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  19  *  General Public License for more details.
  20  *
  21  *  You should have received a copy of the GNU General Public License along
  22  *  with this program; if not, write to the Free Software Foundation, Inc.,
  23  *  59 Temple Place, Suite 330, Boston, MA 02111-1307 USA.
  24  *
  25  * ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
  26  */
  27
  28 #include <linux/kernel.h>
  29 #include <linux/module.h>
  30 #include <linux/init.h>
  31 #include <linux/smp.h>
  32 #include <linux/sched.h>
  33 #include <linux/cpufreq.h>
  34 #include <linux/compiler.h>
  35 #include <linux/dmi.h>
  36 #include <trace/power.h>
  37
  38 #include <linux/acpi.h>
  39 #include <linux/io.h>
  40 #include <linux/delay.h>
  41 #include <linux/uaccess.h>
  42
  43 #include <acpi/processor.h>
  44
  45 #include <asm/msr.h>
  46 #include <asm/processor.h>
  47 #include <asm/cpufeature.h>
  48
  49 #define dprintk(msg...) cpufreq_debug_printk(CPUFREQ_DEBUG_DRIVER, \
  50                 "acpi-cpufreq", msg)
  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_IO_CAPABLE,
  60 };
  61
  62 #define INTEL_MSR_RANGE         (0xffff)
  63 #define CPUID_6_ECX_APERFMPERF_CAPABILITY       (0x1)
  64
  65 struct acpi_cpufreq_data {
  66         struct acpi_processor_performance *acpi_data;
  67         struct cpufreq_frequency_table *freq_table;
  68         unsigned int resume;
  69         unsigned int cpu_feature;
  70 };
  71
  72 static DEFINE_PER_CPU(struct acpi_cpufreq_data *, drv_data);
  73
  74 struct acpi_msr_data {
  75         u64 saved_aperf, saved_mperf;
  76 };
  77
  78 static DEFINE_PER_CPU(struct acpi_msr_data, msr_data);
  79
  80 DEFINE_TRACE(power_mark);
  81
  82 /* acpi_perf_data is a pointer to percpu data. */
  83 static struct acpi_processor_performance *acpi_perf_data;
  84
  85 static struct cpufreq_driver acpi_cpufreq_driver;
  86
  87 static unsigned int acpi_pstate_strict;
  88
  89 static int check_est_cpu(unsigned int cpuid)
  90 {
  91         struct cpuinfo_x86 *cpu = &cpu_data(cpuid);
  92
  93         if (cpu->x86_vendor != X86_VENDOR_INTEL ||
  94             !cpu_has(cpu, X86_FEATURE_EST))
  95                 return 0;
  96
  97         return 1;
  98 }
  99
 100 static unsigned extract_io(u32 value, struct acpi_cpufreq_data *data)
 101 {
 102         struct acpi_processor_performance *perf;
 103         int i;
 104
 105         perf = data->acpi_data;
 106
 107         for (i = 0; i < perf->state_count; i++) {
 108                 if (value == perf->states[i].status)
 109                         return data->freq_table[i].frequency;
 110         }
 111         return 0;
 112 }
 113
 114 static unsigned extract_msr(u32 msr, struct acpi_cpufreq_data *data)
 115 {
 116         int i;
 117         struct acpi_processor_performance *perf;
 118
 119         msr &= INTEL_MSR_RANGE;
 120         perf = data->acpi_data;
 121
 122         for (i = 0; data->freq_table[i].frequency != CPUFREQ_TABLE_END; i++) {
 123                 if (msr == perf->states[data->freq_table[i].index].status)
 124                         return data->freq_table[i].frequency;
 125         }
 126         return data->freq_table[0].frequency;
 127 }
 128
 129 static unsigned extract_freq(u32 val, struct acpi_cpufreq_data *data)
 130 {
 131         switch (data->cpu_feature) {
 132         case SYSTEM_INTEL_MSR_CAPABLE:
 133                 return extract_msr(val, data);
 134         case SYSTEM_IO_CAPABLE:
 135                 return extract_io(val, data);
 136         default:
 137                 return 0;
 138         }
 139 }
 140
 141 struct msr_addr {
 142         u32 reg;
 143 };
 144
 145 struct io_addr {
 146         u16 port;
 147         u8 bit_width;
 148 };
 149
 150 struct drv_cmd {
 151         unsigned int type;
 152         const struct cpumask *mask;
 153         union {
 154                 struct msr_addr msr;
 155                 struct io_addr io;
 156         } addr;
 157         u32 val;
 158 };
 159
 160 /* Called via smp_call_function_single(), on the target CPU */
 161 static void do_drv_read(void *_cmd)
 162 {
 163         struct drv_cmd *cmd = _cmd;
 164         u32 h;
 165
 166         switch (cmd->type) {
 167         case SYSTEM_INTEL_MSR_CAPABLE:
 168                 rdmsr(cmd->addr.msr.reg, cmd->val, h);
 169                 break;
 170         case SYSTEM_IO_CAPABLE:
 171                 acpi_os_read_port((acpi_io_address)cmd->addr.io.port,
 172                                 &cmd->val,
 173                                 (u32)cmd->addr.io.bit_width);
 174                 break;
 175         default:
 176                 break;
 177         }
 178 }
 179
 180 /* Called via smp_call_function_many(), on the target CPUs */
 181 static void do_drv_write(void *_cmd)
 182 {
 183         struct drv_cmd *cmd = _cmd;
 184         u32 lo, hi;
 185
 186         switch (cmd->type) {
 187         case SYSTEM_INTEL_MSR_CAPABLE:
 188                 rdmsr(cmd->addr.msr.reg, lo, hi);
 189                 lo = (lo & ~INTEL_MSR_RANGE) | (cmd->val & INTEL_MSR_RANGE);
 190                 wrmsr(cmd->addr.msr.reg, lo, hi);
 191                 break;
 192         case SYSTEM_IO_CAPABLE:
 193                 acpi_os_write_port((acpi_io_address)cmd->addr.io.port,
 194                                 cmd->val,
 195                                 (u32)cmd->addr.io.bit_width);
 196                 break;
 197         default:
 198                 break;
 199         }
 200 }
 201
 202 static void drv_read(struct drv_cmd *cmd)
 203 {
 204         cmd->val = 0;
 205
 206         smp_call_function_single(cpumask_any(cmd->mask), do_drv_read, cmd, 1);
 207 }
 208
 209 static void drv_write(struct drv_cmd *cmd)
 210 {
 211         int this_cpu;
 212
 213         this_cpu = get_cpu();
 214         if (cpumask_test_cpu(this_cpu, cmd->mask))
 215                 do_drv_write(cmd);
 216         smp_call_function_many(cmd->mask, do_drv_write, cmd, 1);
 217         put_cpu();
 218 }
 219
 220 static u32 get_cur_val(const struct cpumask *mask)
 221 {
 222         struct acpi_processor_performance *perf;
 223         struct drv_cmd cmd;
 224
 225         if (unlikely(cpumask_empty(mask)))
 226                 return 0;
 227
 228         switch (per_cpu(drv_data, cpumask_first(mask))->cpu_feature) {
 229         case SYSTEM_INTEL_MSR_CAPABLE:
 230                 cmd.type = SYSTEM_INTEL_MSR_CAPABLE;
 231                 cmd.addr.msr.reg = MSR_IA32_PERF_STATUS;
 232                 break;
 233         case SYSTEM_IO_CAPABLE:
 234                 cmd.type = SYSTEM_IO_CAPABLE;
 235                 perf = per_cpu(drv_data, cpumask_first(mask))->acpi_data;
 236                 cmd.addr.io.port = perf->control_register.address;
 237                 cmd.addr.io.bit_width = perf->control_register.bit_width;
 238                 break;
 239         default:
 240                 return 0;
 241         }
 242
 243         cmd.mask = mask;
 244         drv_read(&cmd);
 245
 246         dprintk("get_cur_val = %u\n", cmd.val);
 247
 248         return cmd.val;
 249 }
 250
 251 struct perf_pair {
 252         union {
 253                 struct {
 254                         u32 lo;
 255                         u32 hi;
 256                 } split;
 257                 u64 whole;
 258         } aperf, mperf;
 259 };
 260
 261 /* Called via smp_call_function_single(), on the target CPU */
 262 static void read_measured_perf_ctrs(void *_cur)
 263 {
 264         struct perf_pair *cur = _cur;
 265
 266         rdmsr(MSR_IA32_APERF, cur->aperf.split.lo, cur->aperf.split.hi);
 267         rdmsr(MSR_IA32_MPERF, cur->mperf.split.lo, cur->mperf.split.hi);
 268 }
 269
 270 /*
 271  * Return the measured active (C0) frequency on this CPU since last call
 272  * to this function.
 273  * Input: cpu number
 274  * Return: Average CPU frequency in terms of max frequency (zero on error)
 275  *
 276  * We use IA32_MPERF and IA32_APERF MSRs to get the measured performance
 277  * over a period of time, while CPU is in C0 state.
 278  * IA32_MPERF counts at the rate of max advertised frequency
 279  * IA32_APERF counts at the rate of actual CPU frequency
 280  * Only IA32_APERF/IA32_MPERF ratio is architecturally defined and
 281  * no meaning should be associated with absolute values of these MSRs.
 282  */
 283 static unsigned int get_measured_perf(struct cpufreq_policy *policy,
 284                                       unsigned int cpu)
 285 {
 286         struct perf_pair readin, cur;
 287         unsigned int perf_percent;
 288         unsigned int retval;
 289
 290         if (smp_call_function_single(cpu, read_measured_perf_ctrs, &readin, 1))
 291                 return 0;
 292
 293         cur.aperf.whole = readin.aperf.whole -
 294                                 per_cpu(msr_data, cpu).saved_aperf;
 295         cur.mperf.whole = readin.mperf.whole -
 296                                 per_cpu(msr_data, cpu).saved_mperf;
 297         per_cpu(msr_data, cpu).saved_aperf = readin.aperf.whole;
 298         per_cpu(msr_data, cpu).saved_mperf = readin.mperf.whole;
 299
 300 #ifdef __i386__
 301         /*
 302          * We dont want to do 64 bit divide with 32 bit kernel
 303          * Get an approximate value. Return failure in case we cannot get
 304          * an approximate value.
 305          */
 306         if (unlikely(cur.aperf.split.hi || cur.mperf.split.hi)) {
 307                 int shift_count;
 308                 u32 h;
 309
 310                 h = max_t(u32, cur.aperf.split.hi, cur.mperf.split.hi);
 311                 shift_count = fls(h);
 312
 313                 cur.aperf.whole >>= shift_count;
 314                 cur.mperf.whole >>= shift_count;
 315         }
 316
 317         if (((unsigned long)(-1) / 100) < cur.aperf.split.lo) {
 318                 int shift_count = 7;
 319                 cur.aperf.split.lo >>= shift_count;
 320                 cur.mperf.split.lo >>= shift_count;
 321         }
 322
 323         if (cur.aperf.split.lo && cur.mperf.split.lo)
 324                 perf_percent = (cur.aperf.split.lo * 100) / cur.mperf.split.lo;
 325         else
 326                 perf_percent = 0;
 327
 328 #else
 329         if (unlikely(((unsigned long)(-1) / 100) < cur.aperf.whole)) {
 330                 int shift_count = 7;
 331                 cur.aperf.whole >>= shift_count;
 332                 cur.mperf.whole >>= shift_count;
 333         }
 334
 335         if (cur.aperf.whole && cur.mperf.whole)
 336                 perf_percent = (cur.aperf.whole * 100) / cur.mperf.whole;
 337         else
 338                 perf_percent = 0;
 339
 340 #endif
 341
 342         retval = (policy->cpuinfo.max_freq * perf_percent) / 100;
 343
 344         return retval;
 345 }
 346
 347 static unsigned int get_cur_freq_on_cpu(unsigned int cpu)
 348 {
 349         struct acpi_cpufreq_data *data = per_cpu(drv_data, cpu);
 350         unsigned int freq;
 351         unsigned int cached_freq;
 352
 353         dprintk("get_cur_freq_on_cpu (%d)\n", cpu);
 354
 355         if (unlikely(data == NULL ||
 356                      data->acpi_data == NULL || data->freq_table == NULL)) {
 357                 return 0;
 358         }
 359
 360         cached_freq = data->freq_table[data->acpi_data->state].frequency;
 361         freq = extract_freq(get_cur_val(cpumask_of(cpu)), data);
 362         if (freq != cached_freq) {
 363                 /*
 364                  * The dreaded BIOS frequency change behind our back.
 365                  * Force set the frequency on next target call.
 366                  */
 367                 data->resume = 1;
 368         }
 369
 370         dprintk("cur freq = %u\n", freq);
 371
 372         return freq;
 373 }
 374
 375 static unsigned int check_freqs(const struct cpumask *mask, unsigned int freq,
 376                                 struct acpi_cpufreq_data *data)
 377 {
 378         unsigned int cur_freq;
 379         unsigned int i;
 380
 381         for (i = 0; i < 100; i++) {
 382                 cur_freq = extract_freq(get_cur_val(mask), data);
 383                 if (cur_freq == freq)
 384                         return 1;
 385                 udelay(10);
 386         }
 387         return 0;
 388 }
 389
 390 static int acpi_cpufreq_target(struct cpufreq_policy *policy,
 391                                unsigned int target_freq, unsigned int relation)
 392 {
 393         struct acpi_cpufreq_data *data = per_cpu(drv_data, policy->cpu);
 394         struct acpi_processor_performance *perf;
 395         struct cpufreq_freqs freqs;
 396         struct drv_cmd cmd;
 397         unsigned int next_state = 0; /* Index into freq_table */
 398         unsigned int next_perf_state = 0; /* Index into perf table */
 399         unsigned int i;
 400         int result = 0;
 401         struct power_trace it;
 402
 403         dprintk("acpi_cpufreq_target %d (%d)\n", target_freq, policy->cpu);
 404
 405         if (unlikely(data == NULL ||
 406              data->acpi_data == NULL || data->freq_table == NULL)) {
 407                 return -ENODEV;
 408         }
 409
 410         perf = data->acpi_data;
 411         result = cpufreq_frequency_table_target(policy,
 412                                                 data->freq_table,
 413                                                 target_freq,
 414                                                 relation, &next_state);
 415         if (unlikely(result)) {
 416                 result = -ENODEV;
 417                 goto out;
 418         }
 419
 420         next_perf_state = data->freq_table[next_state].index;
 421         if (perf->state == next_perf_state) {
 422                 if (unlikely(data->resume)) {
 423                         dprintk("Called after resume, resetting to P%d\n",
 424                                 next_perf_state);
 425                         data->resume = 0;
 426                 } else {
 427                         dprintk("Already at target state (P%d)\n",
 428                                 next_perf_state);
 429                         goto out;
 430                 }
 431         }
 432
 433         trace_power_mark(&it, POWER_PSTATE, next_perf_state);
 434
 435         switch (data->cpu_feature) {
 436         case SYSTEM_INTEL_MSR_CAPABLE:
 437                 cmd.type = SYSTEM_INTEL_MSR_CAPABLE;
 438                 cmd.addr.msr.reg = MSR_IA32_PERF_CTL;
 439                 cmd.val = (u32) perf->states[next_perf_state].control;
 440                 break;
 441         case SYSTEM_IO_CAPABLE:
 442                 cmd.type = SYSTEM_IO_CAPABLE;
 443                 cmd.addr.io.port = perf->control_register.address;
 444                 cmd.addr.io.bit_width = perf->control_register.bit_width;
 445                 cmd.val = (u32) perf->states[next_perf_state].control;
 446                 break;
 447         default:
 448                 result = -ENODEV;
 449                 goto out;
 450         }
 451
 452         /* cpufreq holds the hotplug lock, so we are safe from here on */
 453         if (policy->shared_type != CPUFREQ_SHARED_TYPE_ANY)
 454                 cmd.mask = policy->cpus;
 455         else
 456                 cmd.mask = cpumask_of(policy->cpu);
 457
 458         freqs.old = perf->states[perf->state].core_frequency * 1000;
 459         freqs.new = data->freq_table[next_state].frequency;
 460         for_each_cpu(i, cmd.mask) {
 461                 freqs.cpu = i;
 462                 cpufreq_notify_transition(&freqs, CPUFREQ_PRECHANGE);
 463         }
 464
 465         drv_write(&cmd);
 466
 467         if (acpi_pstate_strict) {
 468                 if (!check_freqs(cmd.mask, freqs.new, data)) {
 469                         dprintk("acpi_cpufreq_target failed (%d)\n",
 470                                 policy->cpu);
 471                         result = -EAGAIN;
 472                         goto out;
 473                 }
 474         }
 475
 476         for_each_cpu(i, cmd.mask) {
 477                 freqs.cpu = i;
 478                 cpufreq_notify_transition(&freqs, CPUFREQ_POSTCHANGE);
 479         }
 480         perf->state = next_perf_state;
 481
 482 out:
 483         return result;
 484 }
 485
 486 static int acpi_cpufreq_verify(struct cpufreq_policy *policy)
 487 {
 488         struct acpi_cpufreq_data *data = per_cpu(drv_data, policy->cpu);
 489
 490         dprintk("acpi_cpufreq_verify\n");
 491
 492         return cpufreq_frequency_table_verify(policy, data->freq_table);
 493 }
 494
 495 static unsigned long
 496 acpi_cpufreq_guess_freq(struct acpi_cpufreq_data *data, unsigned int cpu)
 497 {
 498         struct acpi_processor_performance *perf = data->acpi_data;
 499
 500         if (cpu_khz) {
 501                 /* search the closest match to cpu_khz */
 502                 unsigned int i;
 503                 unsigned long freq;
 504                 unsigned long freqn = perf->states[0].core_frequency * 1000;
 505
 506                 for (i = 0; i < (perf->state_count-1); i++) {
 507                         freq = freqn;
 508                         freqn = perf->states[i+1].core_frequency * 1000;
 509                         if ((2 * cpu_khz) > (freqn + freq)) {
 510                                 perf->state = i;
 511                                 return freq;
 512                         }
 513                 }
 514                 perf->state = perf->state_count-1;
 515                 return freqn;
 516         } else {
 517                 /* assume CPU is at P0... */
 518                 perf->state = 0;
 519                 return perf->states[0].core_frequency * 1000;
 520         }
 521 }
 522
 523 static void free_acpi_perf_data(void)
 524 {
 525         unsigned int i;
 526
 527         /* Freeing a NULL pointer is OK, and alloc_percpu zeroes. */
 528         for_each_possible_cpu(i)
 529                 free_cpumask_var(per_cpu_ptr(acpi_perf_data, i)
 530                                  ->shared_cpu_map);
 531         free_percpu(acpi_perf_data);
 532 }
 533
 534 /*
 535  * acpi_cpufreq_early_init - initialize ACPI P-States library
 536  *
 537  * Initialize the ACPI P-States library (drivers/acpi/processor_perflib.c)
 538  * in order to determine correct frequency and voltage pairings. We can
 539  * do _PDC and _PSD and find out the processor dependency for the
 540  * actual init that will happen later...
 541  */
 542 static int __init acpi_cpufreq_early_init(void)
 543 {
 544         unsigned int i;
 545         dprintk("acpi_cpufreq_early_init\n");
 546
 547         acpi_perf_data = alloc_percpu(struct acpi_processor_performance);
 548         if (!acpi_perf_data) {
 549                 dprintk("Memory allocation error for acpi_perf_data.\n");
 550                 return -ENOMEM;
 551         }
 552         for_each_possible_cpu(i) {
 553                 if (!alloc_cpumask_var_node(
 554                         &per_cpu_ptr(acpi_perf_data, i)->shared_cpu_map,
 555                         GFP_KERNEL, cpu_to_node(i))) {
 556
 557                         /* Freeing a NULL pointer is OK: alloc_percpu zeroes. */
 558                         free_acpi_perf_data();
 559                         return -ENOMEM;
 560                 }
 561         }
 562
 563         /* Do initialization in ACPI core */
 564         acpi_processor_preregister_performance(acpi_perf_data);
 565         return 0;
 566 }
 567
 568 #ifdef CONFIG_SMP
 569 /*
 570  * Some BIOSes do SW_ANY coordination internally, either set it up in hw
 571  * or do it in BIOS firmware and won't inform about it to OS. If not
 572  * detected, this has a side effect of making CPU run at a different speed
 573  * than OS intended it to run at. Detect it and handle it cleanly.
 574  */
 575 static int bios_with_sw_any_bug;
 576
 577 static int sw_any_bug_found(const struct dmi_system_id *d)
 578 {
 579         bios_with_sw_any_bug = 1;
 580         return 0;
 581 }
 582
 583 static const struct dmi_system_id sw_any_bug_dmi_table[] = {
 584         {
 585                 .callback = sw_any_bug_found,
 586                 .ident = "Supermicro Server X6DLP",
 587                 .matches = {
 588                         DMI_MATCH(DMI_SYS_VENDOR, "Supermicro"),
 589                         DMI_MATCH(DMI_BIOS_VERSION, "080010"),
 590                         DMI_MATCH(DMI_PRODUCT_NAME, "X6DLP"),
 591                 },
 592         },
 593         { }
 594 };
 595 #endif
 596
 597 static int acpi_cpufreq_cpu_init(struct cpufreq_policy *policy)
 598 {
 599         unsigned int i;
 600         unsigned int valid_states = 0;
 601         unsigned int cpu = policy->cpu;
 602         struct acpi_cpufreq_data *data;
 603         unsigned int result = 0;
 604         struct cpuinfo_x86 *c = &cpu_data(policy->cpu);
 605         struct acpi_processor_performance *perf;
 606
 607         dprintk("acpi_cpufreq_cpu_init\n");
 608
 609         data = kzalloc(sizeof(struct acpi_cpufreq_data), GFP_KERNEL);
 610         if (!data)
 611                 return -ENOMEM;
 612
 613         data->acpi_data = per_cpu_ptr(acpi_perf_data, cpu);
 614         per_cpu(drv_data, cpu) = data;
 615
 616         if (cpu_has(c, X86_FEATURE_CONSTANT_TSC))
 617                 acpi_cpufreq_driver.flags |= CPUFREQ_CONST_LOOPS;
 618
 619         result = acpi_processor_register_performance(data->acpi_data, cpu);
 620         if (result)
 621                 goto err_free;
 622
 623         perf = data->acpi_data;
 624         policy->shared_type = perf->shared_type;
 625
 626         /*
 627          * Will let policy->cpus know about dependency only when software
 628          * coordination is required.
 629          */
 630         if (policy->shared_type == CPUFREQ_SHARED_TYPE_ALL ||
 631             policy->shared_type == CPUFREQ_SHARED_TYPE_ANY) {
 632                 cpumask_copy(policy->cpus, perf->shared_cpu_map);
 633         }
 634         cpumask_copy(policy->related_cpus, perf->shared_cpu_map);
 635
 636 #ifdef CONFIG_SMP
 637         dmi_check_system(sw_any_bug_dmi_table);
 638         if (bios_with_sw_any_bug && cpumask_weight(policy->cpus) == 1) {
 639                 policy->shared_type = CPUFREQ_SHARED_TYPE_ALL;
 640                 cpumask_copy(policy->cpus, cpu_core_mask(cpu));
 641         }
 642 #endif
 643
 644         /* capability check */
 645         if (perf->state_count <= 1) {
 646                 dprintk("No P-States\n");
 647                 result = -ENODEV;
 648                 goto err_unreg;
 649         }
 650
 651         if (perf->control_register.space_id != perf->status_register.space_id) {
 652                 result = -ENODEV;
 653                 goto err_unreg;
 654         }
 655
 656         switch (perf->control_register.space_id) {
 657         case ACPI_ADR_SPACE_SYSTEM_IO:
 658                 dprintk("SYSTEM IO addr space\n");
 659                 data->cpu_feature = SYSTEM_IO_CAPABLE;
 660                 break;
 661         case ACPI_ADR_SPACE_FIXED_HARDWARE:
 662                 dprintk("HARDWARE addr space\n");
 663                 if (!check_est_cpu(cpu)) {
 664                         result = -ENODEV;
 665                         goto err_unreg;
 666                 }
 667                 data->cpu_feature = SYSTEM_INTEL_MSR_CAPABLE;
 668                 break;
 669         default:
 670                 dprintk("Unknown addr space %d\n",
 671                         (u32) (perf->control_register.space_id));
 672                 result = -ENODEV;
 673                 goto err_unreg;
 674         }
 675
 676         data->freq_table = kmalloc(sizeof(struct cpufreq_frequency_table) *
 677                     (perf->state_count+1), GFP_KERNEL);
 678         if (!data->freq_table) {
 679                 result = -ENOMEM;
 680                 goto err_unreg;
 681         }
 682
 683         /* detect transition latency */
 684         policy->cpuinfo.transition_latency = 0;
 685         for (i = 0; i < perf->state_count; i++) {
 686                 if ((perf->states[i].transition_latency * 1000) >
 687                     policy->cpuinfo.transition_latency)
 688                         policy->cpuinfo.transition_latency =
 689                             perf->states[i].transition_latency * 1000;
 690         }
 691
 692         /* Check for high latency (>20uS) from buggy BIOSes, like on T42 */
 693         if (perf->control_register.space_id == ACPI_ADR_SPACE_FIXED_HARDWARE &&
 694             policy->cpuinfo.transition_latency > 20 * 1000) {
 695                 policy->cpuinfo.transition_latency = 20 * 1000;
 696                         printk_once(KERN_INFO "Capping off P-state tranision"
 697                                     " latency at 20 uS\n");
 698         }
 699
 700         /* table init */
 701         for (i = 0; i < perf->state_count; i++) {
 702                 if (i > 0 && perf->states[i].core_frequency >=
 703                     data->freq_table[valid_states-1].frequency / 1000)
 704                         continue;
 705
 706                 data->freq_table[valid_states].index = i;
 707                 data->freq_table[valid_states].frequency =
 708                     perf->states[i].core_frequency * 1000;
 709                 valid_states++;
 710         }
 711         data->freq_table[valid_states].frequency = CPUFREQ_TABLE_END;
 712         perf->state = 0;
 713
 714         result = cpufreq_frequency_table_cpuinfo(policy, data->freq_table);
 715         if (result)
 716                 goto err_freqfree;
 717
 718         if (perf->states[0].core_frequency * 1000 != policy->cpuinfo.max_freq)
 719                 printk(KERN_WARNING FW_WARN "P-state 0 is not max freq\n");
 720
 721         switch (perf->control_register.space_id) {
 722         case ACPI_ADR_SPACE_SYSTEM_IO:
 723                 /* Current speed is unknown and not detectable by IO port */
 724                 policy->cur = acpi_cpufreq_guess_freq(data, policy->cpu);
 725                 break;
 726         case ACPI_ADR_SPACE_FIXED_HARDWARE:
 727                 acpi_cpufreq_driver.get = get_cur_freq_on_cpu;
 728                 policy->cur = get_cur_freq_on_cpu(cpu);
 729                 break;
 730         default:
 731                 break;
 732         }
 733
 734         /* notify BIOS that we exist */
 735         acpi_processor_notify_smm(THIS_MODULE);
 736
 737         /* Check for APERF/MPERF support in hardware */
 738         if (c->x86_vendor == X86_VENDOR_INTEL && c->cpuid_level >= 6) {
 739                 unsigned int ecx;
 740                 ecx = cpuid_ecx(6);
 741                 if (ecx & CPUID_6_ECX_APERFMPERF_CAPABILITY)
 742                         acpi_cpufreq_driver.getavg = get_measured_perf;
 743         }
 744
 745         dprintk("CPU%u - ACPI performance management activated.\n", cpu);
 746         for (i = 0; i < perf->state_count; i++)
 747                 dprintk("     %cP%d: %d MHz, %d mW, %d uS\n",
 748                         (i == perf->state ? '*' : ' '), i,
 749                         (u32) perf->states[i].core_frequency,
 750                         (u32) perf->states[i].power,
 751                         (u32) perf->states[i].transition_latency);
 752
 753         cpufreq_frequency_table_get_attr(data->freq_table, policy->cpu);
 754
 755         /*
 756          * the first call to ->target() should result in us actually
 757          * writing something to the appropriate registers.
 758          */
 759         data->resume = 1;
 760
 761         return result;
 762
 763 err_freqfree:
 764         kfree(data->freq_table);
 765 err_unreg:
 766         acpi_processor_unregister_performance(perf, cpu);
 767 err_free:
 768         kfree(data);
 769         per_cpu(drv_data, cpu) = NULL;
 770
 771         return result;
 772 }
 773
 774 static int acpi_cpufreq_cpu_exit(struct cpufreq_policy *policy)
 775 {
 776         struct acpi_cpufreq_data *data = per_cpu(drv_data, policy->cpu);
 777
 778         dprintk("acpi_cpufreq_cpu_exit\n");
 779
 780         if (data) {
 781                 cpufreq_frequency_table_put_attr(policy->cpu);
 782                 per_cpu(drv_data, policy->cpu) = NULL;
 783                 acpi_processor_unregister_performance(data->acpi_data,
 784                                                       policy->cpu);
 785                 kfree(data);
 786         }
 787
 788         return 0;
 789 }
 790
 791 static int acpi_cpufreq_resume(struct cpufreq_policy *policy)
 792 {
 793         struct acpi_cpufreq_data *data = per_cpu(drv_data, policy->cpu);
 794
 795         dprintk("acpi_cpufreq_resume\n");
 796
 797         data->resume = 1;
 798
 799         return 0;
 800 }
 801
 802 static struct freq_attr *acpi_cpufreq_attr[] = {
 803         &cpufreq_freq_attr_scaling_available_freqs,
 804         NULL,
 805 };
 806
 807 static struct cpufreq_driver acpi_cpufreq_driver = {
 808         .verify = acpi_cpufreq_verify,
 809         .target = acpi_cpufreq_target,
 810         .init = acpi_cpufreq_cpu_init,
 811         .exit = acpi_cpufreq_cpu_exit,
 812         .resume = acpi_cpufreq_resume,
 813         .name = "acpi-cpufreq",
 814         .owner = THIS_MODULE,
 815         .attr = acpi_cpufreq_attr,
 816 };
 817
 818 static int __init acpi_cpufreq_init(void)
 819 {
 820         int ret;
 821
 822         if (acpi_disabled)
 823                 return 0;
 824
 825         dprintk("acpi_cpufreq_init\n");
 826
 827         ret = acpi_cpufreq_early_init();
 828         if (ret)
 829                 return ret;
 830
 831         ret = cpufreq_register_driver(&acpi_cpufreq_driver);
 832         if (ret)
 833                 free_acpi_perf_data();
 834
 835         return ret;
 836 }
 837
 838 static void __exit acpi_cpufreq_exit(void)
 839 {
 840         dprintk("acpi_cpufreq_exit\n");
 841
 842         cpufreq_unregister_driver(&acpi_cpufreq_driver);
 843
 844         free_percpu(acpi_perf_data);
 845 }
 846
 847 module_param(acpi_pstate_strict, uint, 0644);
 848 MODULE_PARM_DESC(acpi_pstate_strict,
 849         "value 0 or non-zero. non-zero -> strict ACPI checks are "
 850         "performed during frequency changes.");
 851
 852 late_initcall(acpi_cpufreq_init);
 853 module_exit(acpi_cpufreq_exit);
 854
 855 MODULE_ALIAS("acpi");