git.oblomov.eu Git - linux-2.6/blob - 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 max_freq;
  69         unsigned int resume;
  70         unsigned int cpu_feature;
  71         u64 saved_aperf, saved_mperf;
  72 };
  73
  74 static DEFINE_PER_CPU(struct acpi_cpufreq_data *, drv_data);
  75
  76 DEFINE_TRACE(power_mark);
  77
  78 /* acpi_perf_data is a pointer to percpu data. */
  79 static struct acpi_processor_performance *acpi_perf_data;
  80
  81 static struct cpufreq_driver acpi_cpufreq_driver;
  82
  83 static unsigned int acpi_pstate_strict;
  84
  85 static int check_est_cpu(unsigned int cpuid)
  86 {
  87         struct cpuinfo_x86 *cpu = &cpu_data(cpuid);
  88
  89         if (cpu->x86_vendor != X86_VENDOR_INTEL ||
  90             !cpu_has(cpu, X86_FEATURE_EST))
  91                 return 0;
  92
  93         return 1;
  94 }
  95
  96 static unsigned extract_io(u32 value, struct acpi_cpufreq_data *data)
  97 {
  98         struct acpi_processor_performance *perf;
  99         int i;
 100
 101         perf = data->acpi_data;
 102
 103         for (i = 0; i < perf->state_count; i++) {
 104                 if (value == perf->states[i].status)
 105                         return data->freq_table[i].frequency;
 106         }
 107         return 0;
 108 }
 109
 110 static unsigned extract_msr(u32 msr, struct acpi_cpufreq_data *data)
 111 {
 112         int i;
 113         struct acpi_processor_performance *perf;
 114
 115         msr &= INTEL_MSR_RANGE;
 116         perf = data->acpi_data;
 117
 118         for (i = 0; data->freq_table[i].frequency != CPUFREQ_TABLE_END; i++) {
 119                 if (msr == perf->states[data->freq_table[i].index].status)
 120                         return data->freq_table[i].frequency;
 121         }
 122         return data->freq_table[0].frequency;
 123 }
 124
 125 static unsigned extract_freq(u32 val, struct acpi_cpufreq_data *data)
 126 {
 127         switch (data->cpu_feature) {
 128         case SYSTEM_INTEL_MSR_CAPABLE:
 129                 return extract_msr(val, data);
 130         case SYSTEM_IO_CAPABLE:
 131                 return extract_io(val, data);
 132         default:
 133                 return 0;
 134         }
 135 }
 136
 137 struct msr_addr {
 138         u32 reg;
 139 };
 140
 141 struct io_addr {
 142         u16 port;
 143         u8 bit_width;
 144 };
 145
 146 struct drv_cmd {
 147         unsigned int type;
 148         const struct cpumask *mask;
 149         union {
 150                 struct msr_addr msr;
 151                 struct io_addr io;
 152         } addr;
 153         u32 val;
 154 };
 155
 156 /* Called via smp_call_function_single(), on the target CPU */
 157 static void do_drv_read(void *_cmd)
 158 {
 159         struct drv_cmd *cmd = _cmd;
 160         u32 h;
 161
 162         switch (cmd->type) {
 163         case SYSTEM_INTEL_MSR_CAPABLE:
 164                 rdmsr(cmd->addr.msr.reg, cmd->val, h);
 165                 break;
 166         case SYSTEM_IO_CAPABLE:
 167                 acpi_os_read_port((acpi_io_address)cmd->addr.io.port,
 168                                 &cmd->val,
 169                                 (u32)cmd->addr.io.bit_width);
 170                 break;
 171         default:
 172                 break;
 173         }
 174 }
 175
 176 /* Called via smp_call_function_many(), on the target CPUs */
 177 static void do_drv_write(void *_cmd)
 178 {
 179         struct drv_cmd *cmd = _cmd;
 180         u32 lo, hi;
 181
 182         switch (cmd->type) {
 183         case SYSTEM_INTEL_MSR_CAPABLE:
 184                 rdmsr(cmd->addr.msr.reg, lo, hi);
 185                 lo = (lo & ~INTEL_MSR_RANGE) | (cmd->val & INTEL_MSR_RANGE);
 186                 wrmsr(cmd->addr.msr.reg, lo, hi);
 187                 break;
 188         case SYSTEM_IO_CAPABLE:
 189                 acpi_os_write_port((acpi_io_address)cmd->addr.io.port,
 190                                 cmd->val,
 191                                 (u32)cmd->addr.io.bit_width);
 192                 break;
 193         default:
 194                 break;
 195         }
 196 }
 197
 198 static void drv_read(struct drv_cmd *cmd)
 199 {
 200         cmd->val = 0;
 201
 202         smp_call_function_single(cpumask_any(cmd->mask), do_drv_read, cmd, 1);
 203 }
 204
 205 static void drv_write(struct drv_cmd *cmd)
 206 {
 207         smp_call_function_many(cmd->mask, do_drv_write, cmd, 1);
 208 }
 209
 210 static u32 get_cur_val(const struct cpumask *mask)
 211 {
 212         struct acpi_processor_performance *perf;
 213         struct drv_cmd cmd;
 214
 215         if (unlikely(cpumask_empty(mask)))
 216                 return 0;
 217
 218         switch (per_cpu(drv_data, cpumask_first(mask))->cpu_feature) {
 219         case SYSTEM_INTEL_MSR_CAPABLE:
 220                 cmd.type = SYSTEM_INTEL_MSR_CAPABLE;
 221                 cmd.addr.msr.reg = MSR_IA32_PERF_STATUS;
 222                 break;
 223         case SYSTEM_IO_CAPABLE:
 224                 cmd.type = SYSTEM_IO_CAPABLE;
 225                 perf = per_cpu(drv_data, cpumask_first(mask))->acpi_data;
 226                 cmd.addr.io.port = perf->control_register.address;
 227                 cmd.addr.io.bit_width = perf->control_register.bit_width;
 228                 break;
 229         default:
 230                 return 0;
 231         }
 232
 233         cmd.mask = mask;
 234         drv_read(&cmd);
 235
 236         dprintk("get_cur_val = %u\n", cmd.val);
 237
 238         return cmd.val;
 239 }
 240
 241 struct perf_pair {
 242         union {
 243                 struct {
 244                         u32 lo;
 245                         u32 hi;
 246                 } split;
 247                 u64 whole;
 248         } aperf, mperf;
 249 };
 250
 251 /* Called via smp_call_function_single(), on the target CPU */
 252 static void read_measured_perf_ctrs(void *_cur)
 253 {
 254         struct perf_pair *cur = _cur;
 255
 256         rdmsr(MSR_IA32_APERF, cur->aperf.split.lo, cur->aperf.split.hi);
 257         rdmsr(MSR_IA32_MPERF, cur->mperf.split.lo, cur->mperf.split.hi);
 258 }
 259
 260 /*
 261  * Return the measured active (C0) frequency on this CPU since last call
 262  * to this function.
 263  * Input: cpu number
 264  * Return: Average CPU frequency in terms of max frequency (zero on error)
 265  *
 266  * We use IA32_MPERF and IA32_APERF MSRs to get the measured performance
 267  * over a period of time, while CPU is in C0 state.
 268  * IA32_MPERF counts at the rate of max advertised frequency
 269  * IA32_APERF counts at the rate of actual CPU frequency
 270  * Only IA32_APERF/IA32_MPERF ratio is architecturally defined and
 271  * no meaning should be associated with absolute values of these MSRs.
 272  */
 273 static unsigned int get_measured_perf(struct cpufreq_policy *policy,
 274                                       unsigned int cpu)
 275 {
 276         struct perf_pair readin, cur;
 277         unsigned int perf_percent;
 278         unsigned int retval;
 279
 280         if (smp_call_function_single(cpu, read_measured_perf_ctrs, &readin, 1))
 281                 return 0;
 282
 283         cur.aperf.whole = readin.aperf.whole -
 284                                 per_cpu(drv_data, cpu)->saved_aperf;
 285         cur.mperf.whole = readin.mperf.whole -
 286                                 per_cpu(drv_data, cpu)->saved_mperf;
 287         per_cpu(drv_data, cpu)->saved_aperf = readin.aperf.whole;
 288         per_cpu(drv_data, cpu)->saved_mperf = readin.mperf.whole;
 289
 290 #ifdef __i386__
 291         /*
 292          * We dont want to do 64 bit divide with 32 bit kernel
 293          * Get an approximate value. Return failure in case we cannot get
 294          * an approximate value.
 295          */
 296         if (unlikely(cur.aperf.split.hi || cur.mperf.split.hi)) {
 297                 int shift_count;
 298                 u32 h;
 299
 300                 h = max_t(u32, cur.aperf.split.hi, cur.mperf.split.hi);
 301                 shift_count = fls(h);
 302
 303                 cur.aperf.whole >>= shift_count;
 304                 cur.mperf.whole >>= shift_count;
 305         }
 306
 307         if (((unsigned long)(-1) / 100) < cur.aperf.split.lo) {
 308                 int shift_count = 7;
 309                 cur.aperf.split.lo >>= shift_count;
 310                 cur.mperf.split.lo >>= shift_count;
 311         }
 312
 313         if (cur.aperf.split.lo && cur.mperf.split.lo)
 314                 perf_percent = (cur.aperf.split.lo * 100) / cur.mperf.split.lo;
 315         else
 316                 perf_percent = 0;
 317
 318 #else
 319         if (unlikely(((unsigned long)(-1) / 100) < cur.aperf.whole)) {
 320                 int shift_count = 7;
 321                 cur.aperf.whole >>= shift_count;
 322                 cur.mperf.whole >>= shift_count;
 323         }
 324
 325         if (cur.aperf.whole && cur.mperf.whole)
 326                 perf_percent = (cur.aperf.whole * 100) / cur.mperf.whole;
 327         else
 328                 perf_percent = 0;
 329
 330 #endif
 331
 332         retval = per_cpu(drv_data, policy->cpu)->max_freq * perf_percent / 100;
 333
 334         return retval;
 335 }
 336
 337 static unsigned int get_cur_freq_on_cpu(unsigned int cpu)
 338 {
 339         struct acpi_cpufreq_data *data = per_cpu(drv_data, cpu);
 340         unsigned int freq;
 341         unsigned int cached_freq;
 342
 343         dprintk("get_cur_freq_on_cpu (%d)\n", cpu);
 344
 345         if (unlikely(data == NULL ||
 346                      data->acpi_data == NULL || data->freq_table == NULL)) {
 347                 return 0;
 348         }
 349
 350         cached_freq = data->freq_table[data->acpi_data->state].frequency;
 351         freq = extract_freq(get_cur_val(cpumask_of(cpu)), data);
 352         if (freq != cached_freq) {
 353                 /*
 354                  * The dreaded BIOS frequency change behind our back.
 355                  * Force set the frequency on next target call.
 356                  */
 357                 data->resume = 1;
 358         }
 359
 360         dprintk("cur freq = %u\n", freq);
 361
 362         return freq;
 363 }
 364
 365 static unsigned int check_freqs(const struct cpumask *mask, unsigned int freq,
 366                                 struct acpi_cpufreq_data *data)
 367 {
 368         unsigned int cur_freq;
 369         unsigned int i;
 370
 371         for (i = 0; i < 100; i++) {
 372                 cur_freq = extract_freq(get_cur_val(mask), data);
 373                 if (cur_freq == freq)
 374                         return 1;
 375                 udelay(10);
 376         }
 377         return 0;
 378 }
 379
 380 static int acpi_cpufreq_target(struct cpufreq_policy *policy,
 381                                unsigned int target_freq, unsigned int relation)
 382 {
 383         struct acpi_cpufreq_data *data = per_cpu(drv_data, policy->cpu);
 384         struct acpi_processor_performance *perf;
 385         struct cpufreq_freqs freqs;
 386         struct drv_cmd cmd;
 387         unsigned int next_state = 0; /* Index into freq_table */
 388         unsigned int next_perf_state = 0; /* Index into perf table */
 389         unsigned int i;
 390         int result = 0;
 391         struct power_trace it;
 392
 393         dprintk("acpi_cpufreq_target %d (%d)\n", target_freq, policy->cpu);
 394
 395         if (unlikely(data == NULL ||
 396              data->acpi_data == NULL || data->freq_table == NULL)) {
 397                 return -ENODEV;
 398         }
 399
 400         perf = data->acpi_data;
 401         result = cpufreq_frequency_table_target(policy,
 402                                                 data->freq_table,
 403                                                 target_freq,
 404                                                 relation, &next_state);
 405         if (unlikely(result)) {
 406                 result = -ENODEV;
 407                 goto out;
 408         }
 409
 410         next_perf_state = data->freq_table[next_state].index;
 411         if (perf->state == next_perf_state) {
 412                 if (unlikely(data->resume)) {
 413                         dprintk("Called after resume, resetting to P%d\n",
 414                                 next_perf_state);
 415                         data->resume = 0;
 416                 } else {
 417                         dprintk("Already at target state (P%d)\n",
 418                                 next_perf_state);
 419                         goto out;
 420                 }
 421         }
 422
 423         trace_power_mark(&it, POWER_PSTATE, next_perf_state);
 424
 425         switch (data->cpu_feature) {
 426         case SYSTEM_INTEL_MSR_CAPABLE:
 427                 cmd.type = SYSTEM_INTEL_MSR_CAPABLE;
 428                 cmd.addr.msr.reg = MSR_IA32_PERF_CTL;
 429                 cmd.val = (u32) perf->states[next_perf_state].control;
 430                 break;
 431         case SYSTEM_IO_CAPABLE:
 432                 cmd.type = SYSTEM_IO_CAPABLE;
 433                 cmd.addr.io.port = perf->control_register.address;
 434                 cmd.addr.io.bit_width = perf->control_register.bit_width;
 435                 cmd.val = (u32) perf->states[next_perf_state].control;
 436                 break;
 437         default:
 438                 result = -ENODEV;
 439                 goto out;
 440         }
 441
 442         /* cpufreq holds the hotplug lock, so we are safe from here on */
 443         if (policy->shared_type != CPUFREQ_SHARED_TYPE_ANY)
 444                 cmd.mask = policy->cpus;
 445         else
 446                 cmd.mask = cpumask_of(policy->cpu);
 447
 448         freqs.old = perf->states[perf->state].core_frequency * 1000;
 449         freqs.new = data->freq_table[next_state].frequency;
 450         for_each_cpu(i, cmd.mask) {
 451                 freqs.cpu = i;
 452                 cpufreq_notify_transition(&freqs, CPUFREQ_PRECHANGE);
 453         }
 454
 455         drv_write(&cmd);
 456
 457         if (acpi_pstate_strict) {
 458                 if (!check_freqs(cmd.mask, freqs.new, data)) {
 459                         dprintk("acpi_cpufreq_target failed (%d)\n",
 460                                 policy->cpu);
 461                         result = -EAGAIN;
 462                         goto out;
 463                 }
 464         }
 465
 466         for_each_cpu(i, cmd.mask) {
 467                 freqs.cpu = i;
 468                 cpufreq_notify_transition(&freqs, CPUFREQ_POSTCHANGE);
 469         }
 470         perf->state = next_perf_state;
 471
 472 out:
 473         return result;
 474 }
 475
 476 static int acpi_cpufreq_verify(struct cpufreq_policy *policy)
 477 {
 478         struct acpi_cpufreq_data *data = per_cpu(drv_data, policy->cpu);
 479
 480         dprintk("acpi_cpufreq_verify\n");
 481
 482         return cpufreq_frequency_table_verify(policy, data->freq_table);
 483 }
 484
 485 static unsigned long
 486 acpi_cpufreq_guess_freq(struct acpi_cpufreq_data *data, unsigned int cpu)
 487 {
 488         struct acpi_processor_performance *perf = data->acpi_data;
 489
 490         if (cpu_khz) {
 491                 /* search the closest match to cpu_khz */
 492                 unsigned int i;
 493                 unsigned long freq;
 494                 unsigned long freqn = perf->states[0].core_frequency * 1000;
 495
 496                 for (i = 0; i < (perf->state_count-1); i++) {
 497                         freq = freqn;
 498                         freqn = perf->states[i+1].core_frequency * 1000;
 499                         if ((2 * cpu_khz) > (freqn + freq)) {
 500                                 perf->state = i;
 501                                 return freq;
 502                         }
 503                 }
 504                 perf->state = perf->state_count-1;
 505                 return freqn;
 506         } else {
 507                 /* assume CPU is at P0... */
 508                 perf->state = 0;
 509                 return perf->states[0].core_frequency * 1000;
 510         }
 511 }
 512
 513 static void free_acpi_perf_data(void)
 514 {
 515         unsigned int i;
 516
 517         /* Freeing a NULL pointer is OK, and alloc_percpu zeroes. */
 518         for_each_possible_cpu(i)
 519                 free_cpumask_var(per_cpu_ptr(acpi_perf_data, i)
 520                                  ->shared_cpu_map);
 521         free_percpu(acpi_perf_data);
 522 }
 523
 524 /*
 525  * acpi_cpufreq_early_init - initialize ACPI P-States library
 526  *
 527  * Initialize the ACPI P-States library (drivers/acpi/processor_perflib.c)
 528  * in order to determine correct frequency and voltage pairings. We can
 529  * do _PDC and _PSD and find out the processor dependency for the
 530  * actual init that will happen later...
 531  */
 532 static int __init acpi_cpufreq_early_init(void)
 533 {
 534         unsigned int i;
 535         dprintk("acpi_cpufreq_early_init\n");
 536
 537         acpi_perf_data = alloc_percpu(struct acpi_processor_performance);
 538         if (!acpi_perf_data) {
 539                 dprintk("Memory allocation error for acpi_perf_data.\n");
 540                 return -ENOMEM;
 541         }
 542         for_each_possible_cpu(i) {
 543                 if (!alloc_cpumask_var_node(
 544                         &per_cpu_ptr(acpi_perf_data, i)->shared_cpu_map,
 545                         GFP_KERNEL, cpu_to_node(i))) {
 546
 547                         /* Freeing a NULL pointer is OK: alloc_percpu zeroes. */
 548                         free_acpi_perf_data();
 549                         return -ENOMEM;
 550                 }
 551         }
 552
 553         /* Do initialization in ACPI core */
 554         acpi_processor_preregister_performance(acpi_perf_data);
 555         return 0;
 556 }
 557
 558 #ifdef CONFIG_SMP
 559 /*
 560  * Some BIOSes do SW_ANY coordination internally, either set it up in hw
 561  * or do it in BIOS firmware and won't inform about it to OS. If not
 562  * detected, this has a side effect of making CPU run at a different speed
 563  * than OS intended it to run at. Detect it and handle it cleanly.
 564  */
 565 static int bios_with_sw_any_bug;
 566
 567 static int sw_any_bug_found(const struct dmi_system_id *d)
 568 {
 569         bios_with_sw_any_bug = 1;
 570         return 0;
 571 }
 572
 573 static const struct dmi_system_id sw_any_bug_dmi_table[] = {
 574         {
 575                 .callback = sw_any_bug_found,
 576                 .ident = "Supermicro Server X6DLP",
 577                 .matches = {
 578                         DMI_MATCH(DMI_SYS_VENDOR, "Supermicro"),
 579                         DMI_MATCH(DMI_BIOS_VERSION, "080010"),
 580                         DMI_MATCH(DMI_PRODUCT_NAME, "X6DLP"),
 581                 },
 582         },
 583         { }
 584 };
 585 #endif
 586
 587 static int acpi_cpufreq_cpu_init(struct cpufreq_policy *policy)
 588 {
 589         unsigned int i;
 590         unsigned int valid_states = 0;
 591         unsigned int cpu = policy->cpu;
 592         struct acpi_cpufreq_data *data;
 593         unsigned int result = 0;
 594         struct cpuinfo_x86 *c = &cpu_data(policy->cpu);
 595         struct acpi_processor_performance *perf;
 596
 597         dprintk("acpi_cpufreq_cpu_init\n");
 598
 599         data = kzalloc(sizeof(struct acpi_cpufreq_data), GFP_KERNEL);
 600         if (!data)
 601                 return -ENOMEM;
 602
 603         data->acpi_data = per_cpu_ptr(acpi_perf_data, cpu);
 604         per_cpu(drv_data, cpu) = data;
 605
 606         if (cpu_has(c, X86_FEATURE_CONSTANT_TSC))
 607                 acpi_cpufreq_driver.flags |= CPUFREQ_CONST_LOOPS;
 608
 609         result = acpi_processor_register_performance(data->acpi_data, cpu);
 610         if (result)
 611                 goto err_free;
 612
 613         perf = data->acpi_data;
 614         policy->shared_type = perf->shared_type;
 615
 616         /*
 617          * Will let policy->cpus know about dependency only when software
 618          * coordination is required.
 619          */
 620         if (policy->shared_type == CPUFREQ_SHARED_TYPE_ALL ||
 621             policy->shared_type == CPUFREQ_SHARED_TYPE_ANY) {
 622                 cpumask_copy(policy->cpus, perf->shared_cpu_map);
 623         }
 624         cpumask_copy(policy->related_cpus, perf->shared_cpu_map);
 625
 626 #ifdef CONFIG_SMP
 627         dmi_check_system(sw_any_bug_dmi_table);
 628         if (bios_with_sw_any_bug && cpumask_weight(policy->cpus) == 1) {
 629                 policy->shared_type = CPUFREQ_SHARED_TYPE_ALL;
 630                 cpumask_copy(policy->cpus, cpu_core_mask(cpu));
 631         }
 632 #endif
 633
 634         /* capability check */
 635         if (perf->state_count <= 1) {
 636                 dprintk("No P-States\n");
 637                 result = -ENODEV;
 638                 goto err_unreg;
 639         }
 640
 641         if (perf->control_register.space_id != perf->status_register.space_id) {
 642                 result = -ENODEV;
 643                 goto err_unreg;
 644         }
 645
 646         switch (perf->control_register.space_id) {
 647         case ACPI_ADR_SPACE_SYSTEM_IO:
 648                 dprintk("SYSTEM IO addr space\n");
 649                 data->cpu_feature = SYSTEM_IO_CAPABLE;
 650                 break;
 651         case ACPI_ADR_SPACE_FIXED_HARDWARE:
 652                 dprintk("HARDWARE addr space\n");
 653                 if (!check_est_cpu(cpu)) {
 654                         result = -ENODEV;
 655                         goto err_unreg;
 656                 }
 657                 data->cpu_feature = SYSTEM_INTEL_MSR_CAPABLE;
 658                 break;
 659         default:
 660                 dprintk("Unknown addr space %d\n",
 661                         (u32) (perf->control_register.space_id));
 662                 result = -ENODEV;
 663                 goto err_unreg;
 664         }
 665
 666         data->freq_table = kmalloc(sizeof(struct cpufreq_frequency_table) *
 667                     (perf->state_count+1), GFP_KERNEL);
 668         if (!data->freq_table) {
 669                 result = -ENOMEM;
 670                 goto err_unreg;
 671         }
 672
 673         /* detect transition latency */
 674         policy->cpuinfo.transition_latency = 0;
 675         for (i = 0; i < perf->state_count; i++) {
 676                 if ((perf->states[i].transition_latency * 1000) >
 677                     policy->cpuinfo.transition_latency)
 678                         policy->cpuinfo.transition_latency =
 679                             perf->states[i].transition_latency * 1000;
 680         }
 681
 682         /* Check for high latency (>20uS) from buggy BIOSes, like on T42 */
 683         if (perf->control_register.space_id == ACPI_ADR_SPACE_FIXED_HARDWARE &&
 684             policy->cpuinfo.transition_latency > 20 * 1000) {
 685                 static int print_once;
 686                 policy->cpuinfo.transition_latency = 20 * 1000;
 687                 if (!print_once) {
 688                         print_once = 1;
 689                         printk(KERN_INFO "Capping off P-state tranision latency"
 690                                 " at 20 uS\n");
 691                 }
 692         }
 693
 694         data->max_freq = perf->states[0].core_frequency * 1000;
 695         /* table init */
 696         for (i = 0; i < perf->state_count; i++) {
 697                 if (i > 0 && perf->states[i].core_frequency >=
 698                     data->freq_table[valid_states-1].frequency / 1000)
 699                         continue;
 700
 701                 data->freq_table[valid_states].index = i;
 702                 data->freq_table[valid_states].frequency =
 703                     perf->states[i].core_frequency * 1000;
 704                 valid_states++;
 705         }
 706         data->freq_table[valid_states].frequency = CPUFREQ_TABLE_END;
 707         perf->state = 0;
 708
 709         result = cpufreq_frequency_table_cpuinfo(policy, data->freq_table);
 710         if (result)
 711                 goto err_freqfree;
 712
 713         switch (perf->control_register.space_id) {
 714         case ACPI_ADR_SPACE_SYSTEM_IO:
 715                 /* Current speed is unknown and not detectable by IO port */
 716                 policy->cur = acpi_cpufreq_guess_freq(data, policy->cpu);
 717                 break;
 718         case ACPI_ADR_SPACE_FIXED_HARDWARE:
 719                 acpi_cpufreq_driver.get = get_cur_freq_on_cpu;
 720                 policy->cur = get_cur_freq_on_cpu(cpu);
 721                 break;
 722         default:
 723                 break;
 724         }
 725
 726         /* notify BIOS that we exist */
 727         acpi_processor_notify_smm(THIS_MODULE);
 728
 729         /* Check for APERF/MPERF support in hardware */
 730         if (c->x86_vendor == X86_VENDOR_INTEL && c->cpuid_level >= 6) {
 731                 unsigned int ecx;
 732                 ecx = cpuid_ecx(6);
 733                 if (ecx & CPUID_6_ECX_APERFMPERF_CAPABILITY)
 734                         acpi_cpufreq_driver.getavg = get_measured_perf;
 735         }
 736
 737         dprintk("CPU%u - ACPI performance management activated.\n", cpu);
 738         for (i = 0; i < perf->state_count; i++)
 739                 dprintk("     %cP%d: %d MHz, %d mW, %d uS\n",
 740                         (i == perf->state ? '*' : ' '), i,
 741                         (u32) perf->states[i].core_frequency,
 742                         (u32) perf->states[i].power,
 743                         (u32) perf->states[i].transition_latency);
 744
 745         cpufreq_frequency_table_get_attr(data->freq_table, policy->cpu);
 746
 747         /*
 748          * the first call to ->target() should result in us actually
 749          * writing something to the appropriate registers.
 750          */
 751         data->resume = 1;
 752
 753         return result;
 754
 755 err_freqfree:
 756         kfree(data->freq_table);
 757 err_unreg:
 758         acpi_processor_unregister_performance(perf, cpu);
 759 err_free:
 760         kfree(data);
 761         per_cpu(drv_data, cpu) = NULL;
 762
 763         return result;
 764 }
 765
 766 static int acpi_cpufreq_cpu_exit(struct cpufreq_policy *policy)
 767 {
 768         struct acpi_cpufreq_data *data = per_cpu(drv_data, policy->cpu);
 769
 770         dprintk("acpi_cpufreq_cpu_exit\n");
 771
 772         if (data) {
 773                 cpufreq_frequency_table_put_attr(policy->cpu);
 774                 per_cpu(drv_data, policy->cpu) = NULL;
 775                 acpi_processor_unregister_performance(data->acpi_data,
 776                                                       policy->cpu);
 777                 kfree(data);
 778         }
 779
 780         return 0;
 781 }
 782
 783 static int acpi_cpufreq_resume(struct cpufreq_policy *policy)
 784 {
 785         struct acpi_cpufreq_data *data = per_cpu(drv_data, policy->cpu);
 786
 787         dprintk("acpi_cpufreq_resume\n");
 788
 789         data->resume = 1;
 790
 791         return 0;
 792 }
 793
 794 static struct freq_attr *acpi_cpufreq_attr[] = {
 795         &cpufreq_freq_attr_scaling_available_freqs,
 796         NULL,
 797 };
 798
 799 static struct cpufreq_driver acpi_cpufreq_driver = {
 800         .verify = acpi_cpufreq_verify,
 801         .target = acpi_cpufreq_target,
 802         .init = acpi_cpufreq_cpu_init,
 803         .exit = acpi_cpufreq_cpu_exit,
 804         .resume = acpi_cpufreq_resume,
 805         .name = "acpi-cpufreq",
 806         .owner = THIS_MODULE,
 807         .attr = acpi_cpufreq_attr,
 808 };
 809
 810 static int __init acpi_cpufreq_init(void)
 811 {
 812         int ret;
 813
 814         if (acpi_disabled)
 815                 return 0;
 816
 817         dprintk("acpi_cpufreq_init\n");
 818
 819         ret = acpi_cpufreq_early_init();
 820         if (ret)
 821                 return ret;
 822
 823         ret = cpufreq_register_driver(&acpi_cpufreq_driver);
 824         if (ret)
 825                 free_acpi_perf_data();
 826
 827         return ret;
 828 }
 829
 830 static void __exit acpi_cpufreq_exit(void)
 831 {
 832         dprintk("acpi_cpufreq_exit\n");
 833
 834         cpufreq_unregister_driver(&acpi_cpufreq_driver);
 835
 836         free_percpu(acpi_perf_data);
 837 }
 838
 839 module_param(acpi_pstate_strict, uint, 0644);
 840 MODULE_PARM_DESC(acpi_pstate_strict,
 841         "value 0 or non-zero. non-zero -> strict ACPI checks are "
 842         "performed during frequency changes.");
 843
 844 late_initcall(acpi_cpufreq_init);
 845 module_exit(acpi_cpufreq_exit);
 846
 847 MODULE_ALIAS("acpi");