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

   1 /*
   2  * acpi-cpufreq.c - ACPI Processor P-States Driver ($Revision: 1.4 $)
   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 <acpi/processor.h>
  40
  41 #include <asm/io.h>
  42 #include <asm/msr.h>
  43 #include <asm/processor.h>
  44 #include <asm/cpufeature.h>
  45 #include <asm/delay.h>
  46 #include <asm/uaccess.h>
  47
  48 #define dprintk(msg...) cpufreq_debug_printk(CPUFREQ_DEBUG_DRIVER, "acpi-cpufreq", msg)
  49
  50 MODULE_AUTHOR("Paul Diefenbaugh, Dominik Brodowski");
  51 MODULE_DESCRIPTION("ACPI Processor P-States Driver");
  52 MODULE_LICENSE("GPL");
  53
  54 enum {
  55         UNDEFINED_CAPABLE = 0,
  56         SYSTEM_INTEL_MSR_CAPABLE,
  57         SYSTEM_IO_CAPABLE,
  58 };
  59
  60 #define INTEL_MSR_RANGE         (0xffff)
  61 #define CPUID_6_ECX_APERFMPERF_CAPABILITY       (0x1)
  62
  63 struct acpi_cpufreq_data {
  64         struct acpi_processor_performance *acpi_data;
  65         struct cpufreq_frequency_table *freq_table;
  66         unsigned int max_freq;
  67         unsigned int resume;
  68         unsigned int cpu_feature;
  69 };
  70
  71 static DEFINE_PER_CPU(struct acpi_cpufreq_data *, drv_data);
  72
  73 DEFINE_TRACE(power_mark);
  74
  75 /* acpi_perf_data is a pointer to percpu data. */
  76 static struct acpi_processor_performance *acpi_perf_data;
  77
  78 static struct cpufreq_driver acpi_cpufreq_driver;
  79
  80 static unsigned int acpi_pstate_strict;
  81
  82 static int check_est_cpu(unsigned int cpuid)
  83 {
  84         struct cpuinfo_x86 *cpu = &cpu_data(cpuid);
  85
  86         if (cpu->x86_vendor != X86_VENDOR_INTEL ||
  87             !cpu_has(cpu, X86_FEATURE_EST))
  88                 return 0;
  89
  90         return 1;
  91 }
  92
  93 static unsigned extract_io(u32 value, struct acpi_cpufreq_data *data)
  94 {
  95         struct acpi_processor_performance *perf;
  96         int i;
  97
  98         perf = data->acpi_data;
  99
 100         for (i=0; i<perf->state_count; i++) {
 101                 if (value == perf->states[i].status)
 102                         return data->freq_table[i].frequency;
 103         }
 104         return 0;
 105 }
 106
 107 static unsigned extract_msr(u32 msr, struct acpi_cpufreq_data *data)
 108 {
 109         int i;
 110         struct acpi_processor_performance *perf;
 111
 112         msr &= INTEL_MSR_RANGE;
 113         perf = data->acpi_data;
 114
 115         for (i=0; data->freq_table[i].frequency != CPUFREQ_TABLE_END; i++) {
 116                 if (msr == perf->states[data->freq_table[i].index].status)
 117                         return data->freq_table[i].frequency;
 118         }
 119         return data->freq_table[0].frequency;
 120 }
 121
 122 static unsigned extract_freq(u32 val, struct acpi_cpufreq_data *data)
 123 {
 124         switch (data->cpu_feature) {
 125         case SYSTEM_INTEL_MSR_CAPABLE:
 126                 return extract_msr(val, data);
 127         case SYSTEM_IO_CAPABLE:
 128                 return extract_io(val, data);
 129         default:
 130                 return 0;
 131         }
 132 }
 133
 134 struct msr_addr {
 135         u32 reg;
 136 };
 137
 138 struct io_addr {
 139         u16 port;
 140         u8 bit_width;
 141 };
 142
 143 typedef union {
 144         struct msr_addr msr;
 145         struct io_addr io;
 146 } drv_addr_union;
 147
 148 struct drv_cmd {
 149         unsigned int type;
 150         const struct cpumask *mask;
 151         drv_addr_union addr;
 152         u32 val;
 153 };
 154
 155 static long do_drv_read(void *_cmd)
 156 {
 157         struct drv_cmd *cmd = _cmd;
 158         u32 h;
 159
 160         switch (cmd->type) {
 161         case SYSTEM_INTEL_MSR_CAPABLE:
 162                 rdmsr(cmd->addr.msr.reg, cmd->val, h);
 163                 break;
 164         case SYSTEM_IO_CAPABLE:
 165                 acpi_os_read_port((acpi_io_address)cmd->addr.io.port,
 166                                 &cmd->val,
 167                                 (u32)cmd->addr.io.bit_width);
 168                 break;
 169         default:
 170                 break;
 171         }
 172         return 0;
 173 }
 174
 175 static long do_drv_write(void *_cmd)
 176 {
 177         struct drv_cmd *cmd = _cmd;
 178         u32 lo, hi;
 179
 180         switch (cmd->type) {
 181         case SYSTEM_INTEL_MSR_CAPABLE:
 182                 rdmsr(cmd->addr.msr.reg, lo, hi);
 183                 lo = (lo & ~INTEL_MSR_RANGE) | (cmd->val & INTEL_MSR_RANGE);
 184                 wrmsr(cmd->addr.msr.reg, lo, hi);
 185                 break;
 186         case SYSTEM_IO_CAPABLE:
 187                 acpi_os_write_port((acpi_io_address)cmd->addr.io.port,
 188                                 cmd->val,
 189                                 (u32)cmd->addr.io.bit_width);
 190                 break;
 191         default:
 192                 break;
 193         }
 194         return 0;
 195 }
 196
 197 static void drv_read(struct drv_cmd *cmd)
 198 {
 199         cmd->val = 0;
 200
 201         work_on_cpu(cpumask_any(cmd->mask), do_drv_read, cmd);
 202 }
 203
 204 static void drv_write(struct drv_cmd *cmd)
 205 {
 206         unsigned int i;
 207
 208         for_each_cpu(i, cmd->mask) {
 209                 work_on_cpu(i, do_drv_write, cmd);
 210         }
 211 }
 212
 213 static u32 get_cur_val(const struct cpumask *mask)
 214 {
 215         struct acpi_processor_performance *perf;
 216         struct drv_cmd cmd;
 217
 218         if (unlikely(cpumask_empty(mask)))
 219                 return 0;
 220
 221         switch (per_cpu(drv_data, cpumask_first(mask))->cpu_feature) {
 222         case SYSTEM_INTEL_MSR_CAPABLE:
 223                 cmd.type = SYSTEM_INTEL_MSR_CAPABLE;
 224                 cmd.addr.msr.reg = MSR_IA32_PERF_STATUS;
 225                 break;
 226         case SYSTEM_IO_CAPABLE:
 227                 cmd.type = SYSTEM_IO_CAPABLE;
 228                 perf = per_cpu(drv_data, cpumask_first(mask))->acpi_data;
 229                 cmd.addr.io.port = perf->control_register.address;
 230                 cmd.addr.io.bit_width = perf->control_register.bit_width;
 231                 break;
 232         default:
 233                 return 0;
 234         }
 235
 236         cmd.mask = mask;
 237         drv_read(&cmd);
 238
 239         dprintk("get_cur_val = %u\n", cmd.val);
 240
 241         return cmd.val;
 242 }
 243
 244 struct perf_cur {
 245         union {
 246                 struct {
 247                         u32 lo;
 248                         u32 hi;
 249                 } split;
 250                 u64 whole;
 251         } aperf_cur, mperf_cur;
 252 };
 253
 254
 255 static long read_measured_perf_ctrs(void *_cur)
 256 {
 257         struct perf_cur *cur = _cur;
 258
 259         rdmsr(MSR_IA32_APERF, cur->aperf_cur.split.lo, cur->aperf_cur.split.hi);
 260         rdmsr(MSR_IA32_MPERF, cur->mperf_cur.split.lo, cur->mperf_cur.split.hi);
 261
 262         wrmsr(MSR_IA32_APERF, 0, 0);
 263         wrmsr(MSR_IA32_MPERF, 0, 0);
 264
 265         return 0;
 266 }
 267
 268 /*
 269  * Return the measured active (C0) frequency on this CPU since last call
 270  * to this function.
 271  * Input: cpu number
 272  * Return: Average CPU frequency in terms of max frequency (zero on error)
 273  *
 274  * We use IA32_MPERF and IA32_APERF MSRs to get the measured performance
 275  * over a period of time, while CPU is in C0 state.
 276  * IA32_MPERF counts at the rate of max advertised frequency
 277  * IA32_APERF counts at the rate of actual CPU frequency
 278  * Only IA32_APERF/IA32_MPERF ratio is architecturally defined and
 279  * no meaning should be associated with absolute values of these MSRs.
 280  */
 281 static unsigned int get_measured_perf(struct cpufreq_policy *policy,
 282                                       unsigned int cpu)
 283 {
 284         struct perf_cur cur;
 285         unsigned int perf_percent;
 286         unsigned int retval;
 287
 288         if (!work_on_cpu(cpu, read_measured_perf_ctrs, &cur))
 289                 return 0;
 290
 291 #ifdef __i386__
 292         /*
 293          * We dont want to do 64 bit divide with 32 bit kernel
 294          * Get an approximate value. Return failure in case we cannot get
 295          * an approximate value.
 296          */
 297         if (unlikely(cur.aperf_cur.split.hi || cur.mperf_cur.split.hi)) {
 298                 int shift_count;
 299                 u32 h;
 300
 301                 h = max_t(u32, cur.aperf_cur.split.hi, cur.mperf_cur.split.hi);
 302                 shift_count = fls(h);
 303
 304                 cur.aperf_cur.whole >>= shift_count;
 305                 cur.mperf_cur.whole >>= shift_count;
 306         }
 307
 308         if (((unsigned long)(-1) / 100) < cur.aperf_cur.split.lo) {
 309                 int shift_count = 7;
 310                 cur.aperf_cur.split.lo >>= shift_count;
 311                 cur.mperf_cur.split.lo >>= shift_count;
 312         }
 313
 314         if (cur.aperf_cur.split.lo && cur.mperf_cur.split.lo)
 315                 perf_percent = (cur.aperf_cur.split.lo * 100) /
 316                                 cur.mperf_cur.split.lo;
 317         else
 318                 perf_percent = 0;
 319
 320 #else
 321         if (unlikely(((unsigned long)(-1) / 100) < cur.aperf_cur.whole)) {
 322                 int shift_count = 7;
 323                 cur.aperf_cur.whole >>= shift_count;
 324                 cur.mperf_cur.whole >>= shift_count;
 325         }
 326
 327         if (cur.aperf_cur.whole && cur.mperf_cur.whole)
 328                 perf_percent = (cur.aperf_cur.whole * 100) /
 329                                 cur.mperf_cur.whole;
 330         else
 331                 perf_percent = 0;
 332
 333 #endif
 334
 335         retval = per_cpu(drv_data, policy->cpu)->max_freq * perf_percent / 100;
 336
 337         return retval;
 338 }
 339
 340 static unsigned int get_cur_freq_on_cpu(unsigned int cpu)
 341 {
 342         struct acpi_cpufreq_data *data = per_cpu(drv_data, cpu);
 343         unsigned int freq;
 344         unsigned int cached_freq;
 345
 346         dprintk("get_cur_freq_on_cpu (%d)\n", cpu);
 347
 348         if (unlikely(data == NULL ||
 349                      data->acpi_data == NULL || data->freq_table == NULL)) {
 350                 return 0;
 351         }
 352
 353         cached_freq = data->freq_table[data->acpi_data->state].frequency;
 354         freq = extract_freq(get_cur_val(cpumask_of(cpu)), data);
 355         if (freq != cached_freq) {
 356                 /*
 357                  * The dreaded BIOS frequency change behind our back.
 358                  * Force set the frequency on next target call.
 359                  */
 360                 data->resume = 1;
 361         }
 362
 363         dprintk("cur freq = %u\n", freq);
 364
 365         return freq;
 366 }
 367
 368 static unsigned int check_freqs(const struct cpumask *mask, unsigned int freq,
 369                                 struct acpi_cpufreq_data *data)
 370 {
 371         unsigned int cur_freq;
 372         unsigned int i;
 373
 374         for (i=0; i<100; i++) {
 375                 cur_freq = extract_freq(get_cur_val(mask), data);
 376                 if (cur_freq == freq)
 377                         return 1;
 378                 udelay(10);
 379         }
 380         return 0;
 381 }
 382
 383 static int acpi_cpufreq_target(struct cpufreq_policy *policy,
 384                                unsigned int target_freq, unsigned int relation)
 385 {
 386         struct acpi_cpufreq_data *data = per_cpu(drv_data, policy->cpu);
 387         struct acpi_processor_performance *perf;
 388         struct cpufreq_freqs freqs;
 389         struct drv_cmd cmd;
 390         unsigned int next_state = 0; /* Index into freq_table */
 391         unsigned int next_perf_state = 0; /* Index into perf table */
 392         unsigned int i;
 393         int result = 0;
 394         struct power_trace it;
 395
 396         dprintk("acpi_cpufreq_target %d (%d)\n", target_freq, policy->cpu);
 397
 398         if (unlikely(data == NULL ||
 399              data->acpi_data == NULL || data->freq_table == NULL)) {
 400                 return -ENODEV;
 401         }
 402
 403         perf = data->acpi_data;
 404         result = cpufreq_frequency_table_target(policy,
 405                                                 data->freq_table,
 406                                                 target_freq,
 407                                                 relation, &next_state);
 408         if (unlikely(result)) {
 409                 result = -ENODEV;
 410                 goto out;
 411         }
 412
 413         next_perf_state = data->freq_table[next_state].index;
 414         if (perf->state == next_perf_state) {
 415                 if (unlikely(data->resume)) {
 416                         dprintk("Called after resume, resetting to P%d\n",
 417                                 next_perf_state);
 418                         data->resume = 0;
 419                 } else {
 420                         dprintk("Already at target state (P%d)\n",
 421                                 next_perf_state);
 422                         goto out;
 423                 }
 424         }
 425
 426         trace_power_mark(&it, POWER_PSTATE, next_perf_state);
 427
 428         switch (data->cpu_feature) {
 429         case SYSTEM_INTEL_MSR_CAPABLE:
 430                 cmd.type = SYSTEM_INTEL_MSR_CAPABLE;
 431                 cmd.addr.msr.reg = MSR_IA32_PERF_CTL;
 432                 cmd.val = (u32) perf->states[next_perf_state].control;
 433                 break;
 434         case SYSTEM_IO_CAPABLE:
 435                 cmd.type = SYSTEM_IO_CAPABLE;
 436                 cmd.addr.io.port = perf->control_register.address;
 437                 cmd.addr.io.bit_width = perf->control_register.bit_width;
 438                 cmd.val = (u32) perf->states[next_perf_state].control;
 439                 break;
 440         default:
 441                 result = -ENODEV;
 442                 goto out;
 443         }
 444
 445         /* cpufreq holds the hotplug lock, so we are safe from here on */
 446         if (policy->shared_type != CPUFREQ_SHARED_TYPE_ANY)
 447                 cmd.mask = policy->cpus;
 448         else
 449                 cmd.mask = cpumask_of(policy->cpu);
 450
 451         freqs.old = perf->states[perf->state].core_frequency * 1000;
 452         freqs.new = data->freq_table[next_state].frequency;
 453         for_each_cpu(i, cmd.mask) {
 454                 freqs.cpu = i;
 455                 cpufreq_notify_transition(&freqs, CPUFREQ_PRECHANGE);
 456         }
 457
 458         drv_write(&cmd);
 459
 460         if (acpi_pstate_strict) {
 461                 if (!check_freqs(cmd.mask, freqs.new, data)) {
 462                         dprintk("acpi_cpufreq_target failed (%d)\n",
 463                                 policy->cpu);
 464                         result = -EAGAIN;
 465                         goto out;
 466                 }
 467         }
 468
 469         for_each_cpu(i, cmd.mask) {
 470                 freqs.cpu = i;
 471                 cpufreq_notify_transition(&freqs, CPUFREQ_POSTCHANGE);
 472         }
 473         perf->state = next_perf_state;
 474
 475 out:
 476         return result;
 477 }
 478
 479 static int acpi_cpufreq_verify(struct cpufreq_policy *policy)
 480 {
 481         struct acpi_cpufreq_data *data = per_cpu(drv_data, policy->cpu);
 482
 483         dprintk("acpi_cpufreq_verify\n");
 484
 485         return cpufreq_frequency_table_verify(policy, data->freq_table);
 486 }
 487
 488 static unsigned long
 489 acpi_cpufreq_guess_freq(struct acpi_cpufreq_data *data, unsigned int cpu)
 490 {
 491         struct acpi_processor_performance *perf = data->acpi_data;
 492
 493         if (cpu_khz) {
 494                 /* search the closest match to cpu_khz */
 495                 unsigned int i;
 496                 unsigned long freq;
 497                 unsigned long freqn = perf->states[0].core_frequency * 1000;
 498
 499                 for (i=0; i<(perf->state_count-1); i++) {
 500                         freq = freqn;
 501                         freqn = perf->states[i+1].core_frequency * 1000;
 502                         if ((2 * cpu_khz) > (freqn + freq)) {
 503                                 perf->state = i;
 504                                 return freq;
 505                         }
 506                 }
 507                 perf->state = perf->state_count-1;
 508                 return freqn;
 509         } else {
 510                 /* assume CPU is at P0... */
 511                 perf->state = 0;
 512                 return perf->states[0].core_frequency * 1000;
 513         }
 514 }
 515
 516 static void free_acpi_perf_data(void)
 517 {
 518         unsigned int i;
 519
 520         /* Freeing a NULL pointer is OK, and alloc_percpu zeroes. */
 521         for_each_possible_cpu(i)
 522                 free_cpumask_var(per_cpu_ptr(acpi_perf_data, i)
 523                                  ->shared_cpu_map);
 524         free_percpu(acpi_perf_data);
 525 }
 526
 527 /*
 528  * acpi_cpufreq_early_init - initialize ACPI P-States library
 529  *
 530  * Initialize the ACPI P-States library (drivers/acpi/processor_perflib.c)
 531  * in order to determine correct frequency and voltage pairings. We can
 532  * do _PDC and _PSD and find out the processor dependency for the
 533  * actual init that will happen later...
 534  */
 535 static int __init acpi_cpufreq_early_init(void)
 536 {
 537         unsigned int i;
 538         dprintk("acpi_cpufreq_early_init\n");
 539
 540         acpi_perf_data = alloc_percpu(struct acpi_processor_performance);
 541         if (!acpi_perf_data) {
 542                 dprintk("Memory allocation error for acpi_perf_data.\n");
 543                 return -ENOMEM;
 544         }
 545         for_each_possible_cpu(i) {
 546                 if (!alloc_cpumask_var_node(
 547                         &per_cpu_ptr(acpi_perf_data, i)->shared_cpu_map,
 548                         GFP_KERNEL, cpu_to_node(i))) {
 549
 550                         /* Freeing a NULL pointer is OK: alloc_percpu zeroes. */
 551                         free_acpi_perf_data();
 552                         return -ENOMEM;
 553                 }
 554         }
 555
 556         /* Do initialization in ACPI core */
 557         acpi_processor_preregister_performance(acpi_perf_data);
 558         return 0;
 559 }
 560
 561 #ifdef CONFIG_SMP
 562 /*
 563  * Some BIOSes do SW_ANY coordination internally, either set it up in hw
 564  * or do it in BIOS firmware and won't inform about it to OS. If not
 565  * detected, this has a side effect of making CPU run at a different speed
 566  * than OS intended it to run at. Detect it and handle it cleanly.
 567  */
 568 static int bios_with_sw_any_bug;
 569
 570 static int sw_any_bug_found(const struct dmi_system_id *d)
 571 {
 572         bios_with_sw_any_bug = 1;
 573         return 0;
 574 }
 575
 576 static const struct dmi_system_id sw_any_bug_dmi_table[] = {
 577         {
 578                 .callback = sw_any_bug_found,
 579                 .ident = "Supermicro Server X6DLP",
 580                 .matches = {
 581                         DMI_MATCH(DMI_SYS_VENDOR, "Supermicro"),
 582                         DMI_MATCH(DMI_BIOS_VERSION, "080010"),
 583                         DMI_MATCH(DMI_PRODUCT_NAME, "X6DLP"),
 584                 },
 585         },
 586         { }
 587 };
 588 #endif
 589
 590 static int acpi_cpufreq_cpu_init(struct cpufreq_policy *policy)
 591 {
 592         unsigned int i;
 593         unsigned int valid_states = 0;
 594         unsigned int cpu = policy->cpu;
 595         struct acpi_cpufreq_data *data;
 596         unsigned int result = 0;
 597         struct cpuinfo_x86 *c = &cpu_data(policy->cpu);
 598         struct acpi_processor_performance *perf;
 599
 600         dprintk("acpi_cpufreq_cpu_init\n");
 601
 602         data = kzalloc(sizeof(struct acpi_cpufreq_data), GFP_KERNEL);
 603         if (!data)
 604                 return -ENOMEM;
 605
 606         data->acpi_data = per_cpu_ptr(acpi_perf_data, cpu);
 607         per_cpu(drv_data, cpu) = data;
 608
 609         if (cpu_has(c, X86_FEATURE_CONSTANT_TSC))
 610                 acpi_cpufreq_driver.flags |= CPUFREQ_CONST_LOOPS;
 611
 612         result = acpi_processor_register_performance(data->acpi_data, cpu);
 613         if (result)
 614                 goto err_free;
 615
 616         perf = data->acpi_data;
 617         policy->shared_type = perf->shared_type;
 618
 619         /*
 620          * Will let policy->cpus know about dependency only when software
 621          * coordination is required.
 622          */
 623         if (policy->shared_type == CPUFREQ_SHARED_TYPE_ALL ||
 624             policy->shared_type == CPUFREQ_SHARED_TYPE_ANY) {
 625                 cpumask_copy(policy->cpus, perf->shared_cpu_map);
 626         }
 627         cpumask_copy(policy->related_cpus, perf->shared_cpu_map);
 628
 629 #ifdef CONFIG_SMP
 630         dmi_check_system(sw_any_bug_dmi_table);
 631         if (bios_with_sw_any_bug && cpumask_weight(policy->cpus) == 1) {
 632                 policy->shared_type = CPUFREQ_SHARED_TYPE_ALL;
 633                 cpumask_copy(policy->cpus, cpu_core_mask(cpu));
 634         }
 635 #endif
 636
 637         /* capability check */
 638         if (perf->state_count <= 1) {
 639                 dprintk("No P-States\n");
 640                 result = -ENODEV;
 641                 goto err_unreg;
 642         }
 643
 644         if (perf->control_register.space_id != perf->status_register.space_id) {
 645                 result = -ENODEV;
 646                 goto err_unreg;
 647         }
 648
 649         switch (perf->control_register.space_id) {
 650         case ACPI_ADR_SPACE_SYSTEM_IO:
 651                 dprintk("SYSTEM IO addr space\n");
 652                 data->cpu_feature = SYSTEM_IO_CAPABLE;
 653                 break;
 654         case ACPI_ADR_SPACE_FIXED_HARDWARE:
 655                 dprintk("HARDWARE addr space\n");
 656                 if (!check_est_cpu(cpu)) {
 657                         result = -ENODEV;
 658                         goto err_unreg;
 659                 }
 660                 data->cpu_feature = SYSTEM_INTEL_MSR_CAPABLE;
 661                 break;
 662         default:
 663                 dprintk("Unknown addr space %d\n",
 664                         (u32) (perf->control_register.space_id));
 665                 result = -ENODEV;
 666                 goto err_unreg;
 667         }
 668
 669         data->freq_table = kmalloc(sizeof(struct cpufreq_frequency_table) *
 670                     (perf->state_count+1), GFP_KERNEL);
 671         if (!data->freq_table) {
 672                 result = -ENOMEM;
 673                 goto err_unreg;
 674         }
 675
 676         /* detect transition latency */
 677         policy->cpuinfo.transition_latency = 0;
 678         for (i=0; i<perf->state_count; i++) {
 679                 if ((perf->states[i].transition_latency * 1000) >
 680                     policy->cpuinfo.transition_latency)
 681                         policy->cpuinfo.transition_latency =
 682                             perf->states[i].transition_latency * 1000;
 683         }
 684
 685         data->max_freq = perf->states[0].core_frequency * 1000;
 686         /* table init */
 687         for (i=0; i<perf->state_count; i++) {
 688                 if (i>0 && perf->states[i].core_frequency >=
 689                     data->freq_table[valid_states-1].frequency / 1000)
 690                         continue;
 691
 692                 data->freq_table[valid_states].index = i;
 693                 data->freq_table[valid_states].frequency =
 694                     perf->states[i].core_frequency * 1000;
 695                 valid_states++;
 696         }
 697         data->freq_table[valid_states].frequency = CPUFREQ_TABLE_END;
 698         perf->state = 0;
 699
 700         result = cpufreq_frequency_table_cpuinfo(policy, data->freq_table);
 701         if (result)
 702                 goto err_freqfree;
 703
 704         switch (perf->control_register.space_id) {
 705         case ACPI_ADR_SPACE_SYSTEM_IO:
 706                 /* Current speed is unknown and not detectable by IO port */
 707                 policy->cur = acpi_cpufreq_guess_freq(data, policy->cpu);
 708                 break;
 709         case ACPI_ADR_SPACE_FIXED_HARDWARE:
 710                 acpi_cpufreq_driver.get = get_cur_freq_on_cpu;
 711                 policy->cur = get_cur_freq_on_cpu(cpu);
 712                 break;
 713         default:
 714                 break;
 715         }
 716
 717         /* notify BIOS that we exist */
 718         acpi_processor_notify_smm(THIS_MODULE);
 719
 720         /* Check for APERF/MPERF support in hardware */
 721         if (c->x86_vendor == X86_VENDOR_INTEL && c->cpuid_level >= 6) {
 722                 unsigned int ecx;
 723                 ecx = cpuid_ecx(6);
 724                 if (ecx & CPUID_6_ECX_APERFMPERF_CAPABILITY)
 725                         acpi_cpufreq_driver.getavg = get_measured_perf;
 726         }
 727
 728         dprintk("CPU%u - ACPI performance management activated.\n", cpu);
 729         for (i = 0; i < perf->state_count; i++)
 730                 dprintk("     %cP%d: %d MHz, %d mW, %d uS\n",
 731                         (i == perf->state ? '*' : ' '), i,
 732                         (u32) perf->states[i].core_frequency,
 733                         (u32) perf->states[i].power,
 734                         (u32) perf->states[i].transition_latency);
 735
 736         cpufreq_frequency_table_get_attr(data->freq_table, policy->cpu);
 737
 738         /*
 739          * the first call to ->target() should result in us actually
 740          * writing something to the appropriate registers.
 741          */
 742         data->resume = 1;
 743
 744         return result;
 745
 746 err_freqfree:
 747         kfree(data->freq_table);
 748 err_unreg:
 749         acpi_processor_unregister_performance(perf, cpu);
 750 err_free:
 751         kfree(data);
 752         per_cpu(drv_data, cpu) = NULL;
 753
 754         return result;
 755 }
 756
 757 static int acpi_cpufreq_cpu_exit(struct cpufreq_policy *policy)
 758 {
 759         struct acpi_cpufreq_data *data = per_cpu(drv_data, policy->cpu);
 760
 761         dprintk("acpi_cpufreq_cpu_exit\n");
 762
 763         if (data) {
 764                 cpufreq_frequency_table_put_attr(policy->cpu);
 765                 per_cpu(drv_data, policy->cpu) = NULL;
 766                 acpi_processor_unregister_performance(data->acpi_data,
 767                                                       policy->cpu);
 768                 kfree(data);
 769         }
 770
 771         return 0;
 772 }
 773
 774 static int acpi_cpufreq_resume(struct cpufreq_policy *policy)
 775 {
 776         struct acpi_cpufreq_data *data = per_cpu(drv_data, policy->cpu);
 777
 778         dprintk("acpi_cpufreq_resume\n");
 779
 780         data->resume = 1;
 781
 782         return 0;
 783 }
 784
 785 static struct freq_attr *acpi_cpufreq_attr[] = {
 786         &cpufreq_freq_attr_scaling_available_freqs,
 787         NULL,
 788 };
 789
 790 static struct cpufreq_driver acpi_cpufreq_driver = {
 791         .verify = acpi_cpufreq_verify,
 792         .target = acpi_cpufreq_target,
 793         .init = acpi_cpufreq_cpu_init,
 794         .exit = acpi_cpufreq_cpu_exit,
 795         .resume = acpi_cpufreq_resume,
 796         .name = "acpi-cpufreq",
 797         .owner = THIS_MODULE,
 798         .attr = acpi_cpufreq_attr,
 799 };
 800
 801 static int __init acpi_cpufreq_init(void)
 802 {
 803         int ret;
 804
 805         if (acpi_disabled)
 806                 return 0;
 807
 808         dprintk("acpi_cpufreq_init\n");
 809
 810         ret = acpi_cpufreq_early_init();
 811         if (ret)
 812                 return ret;
 813
 814         ret = cpufreq_register_driver(&acpi_cpufreq_driver);
 815         if (ret)
 816                 free_acpi_perf_data();
 817
 818         return ret;
 819 }
 820
 821 static void __exit acpi_cpufreq_exit(void)
 822 {
 823         dprintk("acpi_cpufreq_exit\n");
 824
 825         cpufreq_unregister_driver(&acpi_cpufreq_driver);
 826
 827         free_percpu(acpi_perf_data);
 828 }
 829
 830 module_param(acpi_pstate_strict, uint, 0644);
 831 MODULE_PARM_DESC(acpi_pstate_strict,
 832         "value 0 or non-zero. non-zero -> strict ACPI checks are "
 833         "performed during frequency changes.");
 834
 835 late_initcall(acpi_cpufreq_init);
 836 module_exit(acpi_cpufreq_exit);
 837
 838 MODULE_ALIAS("acpi");