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/events/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
  64 struct acpi_cpufreq_data {
  65         struct acpi_processor_performance *acpi_data;
  66         struct cpufreq_frequency_table *freq_table;
  67         unsigned int resume;
  68         unsigned int cpu_feature;
  69 };
  70
  71 static DEFINE_PER_CPU(struct acpi_cpufreq_data *, acfreq_data);
  72
  73 static DEFINE_PER_CPU(struct aperfmperf, acfreq_old_perf);
  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         return cpu_has(cpu, X86_FEATURE_EST);
  87 }
  88
  89 static unsigned extract_io(u32 value, struct acpi_cpufreq_data *data)
  90 {
  91         struct acpi_processor_performance *perf;
  92         int i;
  93
  94         perf = data->acpi_data;
  95
  96         for (i = 0; i < perf->state_count; i++) {
  97                 if (value == perf->states[i].status)
  98                         return data->freq_table[i].frequency;
  99         }
 100         return 0;
 101 }
 102
 103 static unsigned extract_msr(u32 msr, struct acpi_cpufreq_data *data)
 104 {
 105         int i;
 106         struct acpi_processor_performance *perf;
 107
 108         msr &= INTEL_MSR_RANGE;
 109         perf = data->acpi_data;
 110
 111         for (i = 0; data->freq_table[i].frequency != CPUFREQ_TABLE_END; i++) {
 112                 if (msr == perf->states[data->freq_table[i].index].status)
 113                         return data->freq_table[i].frequency;
 114         }
 115         return data->freq_table[0].frequency;
 116 }
 117
 118 static unsigned extract_freq(u32 val, struct acpi_cpufreq_data *data)
 119 {
 120         switch (data->cpu_feature) {
 121         case SYSTEM_INTEL_MSR_CAPABLE:
 122                 return extract_msr(val, data);
 123         case SYSTEM_IO_CAPABLE:
 124                 return extract_io(val, data);
 125         default:
 126                 return 0;
 127         }
 128 }
 129
 130 struct msr_addr {
 131         u32 reg;
 132 };
 133
 134 struct io_addr {
 135         u16 port;
 136         u8 bit_width;
 137 };
 138
 139 struct drv_cmd {
 140         unsigned int type;
 141         const struct cpumask *mask;
 142         union {
 143                 struct msr_addr msr;
 144                 struct io_addr io;
 145         } addr;
 146         u32 val;
 147 };
 148
 149 /* Called via smp_call_function_single(), on the target CPU */
 150 static void do_drv_read(void *_cmd)
 151 {
 152         struct drv_cmd *cmd = _cmd;
 153         u32 h;
 154
 155         switch (cmd->type) {
 156         case SYSTEM_INTEL_MSR_CAPABLE:
 157                 rdmsr(cmd->addr.msr.reg, cmd->val, h);
 158                 break;
 159         case SYSTEM_IO_CAPABLE:
 160                 acpi_os_read_port((acpi_io_address)cmd->addr.io.port,
 161                                 &cmd->val,
 162                                 (u32)cmd->addr.io.bit_width);
 163                 break;
 164         default:
 165                 break;
 166         }
 167 }
 168
 169 /* Called via smp_call_function_many(), on the target CPUs */
 170 static void do_drv_write(void *_cmd)
 171 {
 172         struct drv_cmd *cmd = _cmd;
 173         u32 lo, hi;
 174
 175         switch (cmd->type) {
 176         case SYSTEM_INTEL_MSR_CAPABLE:
 177                 rdmsr(cmd->addr.msr.reg, lo, hi);
 178                 lo = (lo & ~INTEL_MSR_RANGE) | (cmd->val & INTEL_MSR_RANGE);
 179                 wrmsr(cmd->addr.msr.reg, lo, hi);
 180                 break;
 181         case SYSTEM_IO_CAPABLE:
 182                 acpi_os_write_port((acpi_io_address)cmd->addr.io.port,
 183                                 cmd->val,
 184                                 (u32)cmd->addr.io.bit_width);
 185                 break;
 186         default:
 187                 break;
 188         }
 189 }
 190
 191 static void drv_read(struct drv_cmd *cmd)
 192 {
 193         cmd->val = 0;
 194
 195         smp_call_function_single(cpumask_any(cmd->mask), do_drv_read, cmd, 1);
 196 }
 197
 198 static void drv_write(struct drv_cmd *cmd)
 199 {
 200         int this_cpu;
 201
 202         this_cpu = get_cpu();
 203         if (cpumask_test_cpu(this_cpu, cmd->mask))
 204                 do_drv_write(cmd);
 205         smp_call_function_many(cmd->mask, do_drv_write, cmd, 1);
 206         put_cpu();
 207 }
 208
 209 static u32 get_cur_val(const struct cpumask *mask)
 210 {
 211         struct acpi_processor_performance *perf;
 212         struct drv_cmd cmd;
 213
 214         if (unlikely(cpumask_empty(mask)))
 215                 return 0;
 216
 217         switch (per_cpu(acfreq_data, cpumask_first(mask))->cpu_feature) {
 218         case SYSTEM_INTEL_MSR_CAPABLE:
 219                 cmd.type = SYSTEM_INTEL_MSR_CAPABLE;
 220                 cmd.addr.msr.reg = MSR_IA32_PERF_STATUS;
 221                 break;
 222         case SYSTEM_IO_CAPABLE:
 223                 cmd.type = SYSTEM_IO_CAPABLE;
 224                 perf = per_cpu(acfreq_data, cpumask_first(mask))->acpi_data;
 225                 cmd.addr.io.port = perf->control_register.address;
 226                 cmd.addr.io.bit_width = perf->control_register.bit_width;
 227                 break;
 228         default:
 229                 return 0;
 230         }
 231
 232         cmd.mask = mask;
 233         drv_read(&cmd);
 234
 235         dprintk("get_cur_val = %u\n", cmd.val);
 236
 237         return cmd.val;
 238 }
 239
 240 /* Called via smp_call_function_single(), on the target CPU */
 241 static void read_measured_perf_ctrs(void *_cur)
 242 {
 243         struct aperfmperf *am = _cur;
 244
 245         get_aperfmperf(am);
 246 }
 247
 248 /*
 249  * Return the measured active (C0) frequency on this CPU since last call
 250  * to this function.
 251  * Input: cpu number
 252  * Return: Average CPU frequency in terms of max frequency (zero on error)
 253  *
 254  * We use IA32_MPERF and IA32_APERF MSRs to get the measured performance
 255  * over a period of time, while CPU is in C0 state.
 256  * IA32_MPERF counts at the rate of max advertised frequency
 257  * IA32_APERF counts at the rate of actual CPU frequency
 258  * Only IA32_APERF/IA32_MPERF ratio is architecturally defined and
 259  * no meaning should be associated with absolute values of these MSRs.
 260  */
 261 static unsigned int get_measured_perf(struct cpufreq_policy *policy,
 262                                       unsigned int cpu)
 263 {
 264         struct aperfmperf perf;
 265         unsigned long ratio;
 266         unsigned int retval;
 267
 268         if (smp_call_function_single(cpu, read_measured_perf_ctrs, &perf, 1))
 269                 return 0;
 270
 271         ratio = calc_aperfmperf_ratio(&per_cpu(acfreq_old_perf, cpu), &perf);
 272         per_cpu(acfreq_old_perf, cpu) = perf;
 273
 274         retval = (policy->cpuinfo.max_freq * ratio) >> APERFMPERF_SHIFT;
 275
 276         return retval;
 277 }
 278
 279 static unsigned int get_cur_freq_on_cpu(unsigned int cpu)
 280 {
 281         struct acpi_cpufreq_data *data = per_cpu(acfreq_data, cpu);
 282         unsigned int freq;
 283         unsigned int cached_freq;
 284
 285         dprintk("get_cur_freq_on_cpu (%d)\n", cpu);
 286
 287         if (unlikely(data == NULL ||
 288                      data->acpi_data == NULL || data->freq_table == NULL)) {
 289                 return 0;
 290         }
 291
 292         cached_freq = data->freq_table[data->acpi_data->state].frequency;
 293         freq = extract_freq(get_cur_val(cpumask_of(cpu)), data);
 294         if (freq != cached_freq) {
 295                 /*
 296                  * The dreaded BIOS frequency change behind our back.
 297                  * Force set the frequency on next target call.
 298                  */
 299                 data->resume = 1;
 300         }
 301
 302         dprintk("cur freq = %u\n", freq);
 303
 304         return freq;
 305 }
 306
 307 static unsigned int check_freqs(const struct cpumask *mask, unsigned int freq,
 308                                 struct acpi_cpufreq_data *data)
 309 {
 310         unsigned int cur_freq;
 311         unsigned int i;
 312
 313         for (i = 0; i < 100; i++) {
 314                 cur_freq = extract_freq(get_cur_val(mask), data);
 315                 if (cur_freq == freq)
 316                         return 1;
 317                 udelay(10);
 318         }
 319         return 0;
 320 }
 321
 322 static int acpi_cpufreq_target(struct cpufreq_policy *policy,
 323                                unsigned int target_freq, unsigned int relation)
 324 {
 325         struct acpi_cpufreq_data *data = per_cpu(acfreq_data, policy->cpu);
 326         struct acpi_processor_performance *perf;
 327         struct cpufreq_freqs freqs;
 328         struct drv_cmd cmd;
 329         unsigned int next_state = 0; /* Index into freq_table */
 330         unsigned int next_perf_state = 0; /* Index into perf table */
 331         unsigned int i;
 332         int result = 0;
 333
 334         dprintk("acpi_cpufreq_target %d (%d)\n", target_freq, policy->cpu);
 335
 336         if (unlikely(data == NULL ||
 337              data->acpi_data == NULL || data->freq_table == NULL)) {
 338                 return -ENODEV;
 339         }
 340
 341         perf = data->acpi_data;
 342         result = cpufreq_frequency_table_target(policy,
 343                                                 data->freq_table,
 344                                                 target_freq,
 345                                                 relation, &next_state);
 346         if (unlikely(result)) {
 347                 result = -ENODEV;
 348                 goto out;
 349         }
 350
 351         next_perf_state = data->freq_table[next_state].index;
 352         if (perf->state == next_perf_state) {
 353                 if (unlikely(data->resume)) {
 354                         dprintk("Called after resume, resetting to P%d\n",
 355                                 next_perf_state);
 356                         data->resume = 0;
 357                 } else {
 358                         dprintk("Already at target state (P%d)\n",
 359                                 next_perf_state);
 360                         goto out;
 361                 }
 362         }
 363
 364         trace_power_frequency(POWER_PSTATE, data->freq_table[next_state].frequency);
 365
 366         switch (data->cpu_feature) {
 367         case SYSTEM_INTEL_MSR_CAPABLE:
 368                 cmd.type = SYSTEM_INTEL_MSR_CAPABLE;
 369                 cmd.addr.msr.reg = MSR_IA32_PERF_CTL;
 370                 cmd.val = (u32) perf->states[next_perf_state].control;
 371                 break;
 372         case SYSTEM_IO_CAPABLE:
 373                 cmd.type = SYSTEM_IO_CAPABLE;
 374                 cmd.addr.io.port = perf->control_register.address;
 375                 cmd.addr.io.bit_width = perf->control_register.bit_width;
 376                 cmd.val = (u32) perf->states[next_perf_state].control;
 377                 break;
 378         default:
 379                 result = -ENODEV;
 380                 goto out;
 381         }
 382
 383         /* cpufreq holds the hotplug lock, so we are safe from here on */
 384         if (policy->shared_type != CPUFREQ_SHARED_TYPE_ANY)
 385                 cmd.mask = policy->cpus;
 386         else
 387                 cmd.mask = cpumask_of(policy->cpu);
 388
 389         freqs.old = perf->states[perf->state].core_frequency * 1000;
 390         freqs.new = data->freq_table[next_state].frequency;
 391         for_each_cpu(i, cmd.mask) {
 392                 freqs.cpu = i;
 393                 cpufreq_notify_transition(&freqs, CPUFREQ_PRECHANGE);
 394         }
 395
 396         drv_write(&cmd);
 397
 398         if (acpi_pstate_strict) {
 399                 if (!check_freqs(cmd.mask, freqs.new, data)) {
 400                         dprintk("acpi_cpufreq_target failed (%d)\n",
 401                                 policy->cpu);
 402                         result = -EAGAIN;
 403                         goto out;
 404                 }
 405         }
 406
 407         for_each_cpu(i, cmd.mask) {
 408                 freqs.cpu = i;
 409                 cpufreq_notify_transition(&freqs, CPUFREQ_POSTCHANGE);
 410         }
 411         perf->state = next_perf_state;
 412
 413 out:
 414         return result;
 415 }
 416
 417 static int acpi_cpufreq_verify(struct cpufreq_policy *policy)
 418 {
 419         struct acpi_cpufreq_data *data = per_cpu(acfreq_data, policy->cpu);
 420
 421         dprintk("acpi_cpufreq_verify\n");
 422
 423         return cpufreq_frequency_table_verify(policy, data->freq_table);
 424 }
 425
 426 static unsigned long
 427 acpi_cpufreq_guess_freq(struct acpi_cpufreq_data *data, unsigned int cpu)
 428 {
 429         struct acpi_processor_performance *perf = data->acpi_data;
 430
 431         if (cpu_khz) {
 432                 /* search the closest match to cpu_khz */
 433                 unsigned int i;
 434                 unsigned long freq;
 435                 unsigned long freqn = perf->states[0].core_frequency * 1000;
 436
 437                 for (i = 0; i < (perf->state_count-1); i++) {
 438                         freq = freqn;
 439                         freqn = perf->states[i+1].core_frequency * 1000;
 440                         if ((2 * cpu_khz) > (freqn + freq)) {
 441                                 perf->state = i;
 442                                 return freq;
 443                         }
 444                 }
 445                 perf->state = perf->state_count-1;
 446                 return freqn;
 447         } else {
 448                 /* assume CPU is at P0... */
 449                 perf->state = 0;
 450                 return perf->states[0].core_frequency * 1000;
 451         }
 452 }
 453
 454 static void free_acpi_perf_data(void)
 455 {
 456         unsigned int i;
 457
 458         /* Freeing a NULL pointer is OK, and alloc_percpu zeroes. */
 459         for_each_possible_cpu(i)
 460                 free_cpumask_var(per_cpu_ptr(acpi_perf_data, i)
 461                                  ->shared_cpu_map);
 462         free_percpu(acpi_perf_data);
 463 }
 464
 465 /*
 466  * acpi_cpufreq_early_init - initialize ACPI P-States library
 467  *
 468  * Initialize the ACPI P-States library (drivers/acpi/processor_perflib.c)
 469  * in order to determine correct frequency and voltage pairings. We can
 470  * do _PDC and _PSD and find out the processor dependency for the
 471  * actual init that will happen later...
 472  */
 473 static int __init acpi_cpufreq_early_init(void)
 474 {
 475         unsigned int i;
 476         dprintk("acpi_cpufreq_early_init\n");
 477
 478         acpi_perf_data = alloc_percpu(struct acpi_processor_performance);
 479         if (!acpi_perf_data) {
 480                 dprintk("Memory allocation error for acpi_perf_data.\n");
 481                 return -ENOMEM;
 482         }
 483         for_each_possible_cpu(i) {
 484                 if (!zalloc_cpumask_var_node(
 485                         &per_cpu_ptr(acpi_perf_data, i)->shared_cpu_map,
 486                         GFP_KERNEL, cpu_to_node(i))) {
 487
 488                         /* Freeing a NULL pointer is OK: alloc_percpu zeroes. */
 489                         free_acpi_perf_data();
 490                         return -ENOMEM;
 491                 }
 492         }
 493
 494         /* Do initialization in ACPI core */
 495         acpi_processor_preregister_performance(acpi_perf_data);
 496         return 0;
 497 }
 498
 499 #ifdef CONFIG_SMP
 500 /*
 501  * Some BIOSes do SW_ANY coordination internally, either set it up in hw
 502  * or do it in BIOS firmware and won't inform about it to OS. If not
 503  * detected, this has a side effect of making CPU run at a different speed
 504  * than OS intended it to run at. Detect it and handle it cleanly.
 505  */
 506 static int bios_with_sw_any_bug;
 507
 508 static int sw_any_bug_found(const struct dmi_system_id *d)
 509 {
 510         bios_with_sw_any_bug = 1;
 511         return 0;
 512 }
 513
 514 static const struct dmi_system_id sw_any_bug_dmi_table[] = {
 515         {
 516                 .callback = sw_any_bug_found,
 517                 .ident = "Supermicro Server X6DLP",
 518                 .matches = {
 519                         DMI_MATCH(DMI_SYS_VENDOR, "Supermicro"),
 520                         DMI_MATCH(DMI_BIOS_VERSION, "080010"),
 521                         DMI_MATCH(DMI_PRODUCT_NAME, "X6DLP"),
 522                 },
 523         },
 524         { }
 525 };
 526
 527 static int acpi_cpufreq_blacklist(struct cpuinfo_x86 *c)
 528 {
 529         /* Intel Xeon Processor 7100 Series Specification Update
 530          * http://www.intel.com/Assets/PDF/specupdate/314554.pdf
 531          * AL30: A Machine Check Exception (MCE) Occurring during an
 532          * Enhanced Intel SpeedStep Technology Ratio Change May Cause
 533          * Both Processor Cores to Lock Up. */
 534         if (c->x86_vendor == X86_VENDOR_INTEL) {
 535                 if ((c->x86 == 15) &&
 536                     (c->x86_model == 6) &&
 537                     (c->x86_mask == 8)) {
 538                         printk(KERN_INFO "acpi-cpufreq: Intel(R) "
 539                             "Xeon(R) 7100 Errata AL30, processors may "
 540                             "lock up on frequency changes: disabling "
 541                             "acpi-cpufreq.\n");
 542                         return -ENODEV;
 543                     }
 544                 }
 545         return 0;
 546 }
 547 #endif
 548
 549 static int acpi_cpufreq_cpu_init(struct cpufreq_policy *policy)
 550 {
 551         unsigned int i;
 552         unsigned int valid_states = 0;
 553         unsigned int cpu = policy->cpu;
 554         struct acpi_cpufreq_data *data;
 555         unsigned int result = 0;
 556         struct cpuinfo_x86 *c = &cpu_data(policy->cpu);
 557         struct acpi_processor_performance *perf;
 558 #ifdef CONFIG_SMP
 559         static int blacklisted;
 560 #endif
 561
 562         dprintk("acpi_cpufreq_cpu_init\n");
 563
 564 #ifdef CONFIG_SMP
 565         if (blacklisted)
 566                 return blacklisted;
 567         blacklisted = acpi_cpufreq_blacklist(c);
 568         if (blacklisted)
 569                 return blacklisted;
 570 #endif
 571
 572         data = kzalloc(sizeof(struct acpi_cpufreq_data), GFP_KERNEL);
 573         if (!data)
 574                 return -ENOMEM;
 575
 576         data->acpi_data = per_cpu_ptr(acpi_perf_data, cpu);
 577         per_cpu(acfreq_data, cpu) = data;
 578
 579         if (cpu_has(c, X86_FEATURE_CONSTANT_TSC))
 580                 acpi_cpufreq_driver.flags |= CPUFREQ_CONST_LOOPS;
 581
 582         result = acpi_processor_register_performance(data->acpi_data, cpu);
 583         if (result)
 584                 goto err_free;
 585
 586         perf = data->acpi_data;
 587         policy->shared_type = perf->shared_type;
 588
 589         /*
 590          * Will let policy->cpus know about dependency only when software
 591          * coordination is required.
 592          */
 593         if (policy->shared_type == CPUFREQ_SHARED_TYPE_ALL ||
 594             policy->shared_type == CPUFREQ_SHARED_TYPE_ANY) {
 595                 cpumask_copy(policy->cpus, perf->shared_cpu_map);
 596         }
 597         cpumask_copy(policy->related_cpus, perf->shared_cpu_map);
 598
 599 #ifdef CONFIG_SMP
 600         dmi_check_system(sw_any_bug_dmi_table);
 601         if (bios_with_sw_any_bug && cpumask_weight(policy->cpus) == 1) {
 602                 policy->shared_type = CPUFREQ_SHARED_TYPE_ALL;
 603                 cpumask_copy(policy->cpus, cpu_core_mask(cpu));
 604         }
 605 #endif
 606
 607         /* capability check */
 608         if (perf->state_count <= 1) {
 609                 dprintk("No P-States\n");
 610                 result = -ENODEV;
 611                 goto err_unreg;
 612         }
 613
 614         if (perf->control_register.space_id != perf->status_register.space_id) {
 615                 result = -ENODEV;
 616                 goto err_unreg;
 617         }
 618
 619         switch (perf->control_register.space_id) {
 620         case ACPI_ADR_SPACE_SYSTEM_IO:
 621                 dprintk("SYSTEM IO addr space\n");
 622                 data->cpu_feature = SYSTEM_IO_CAPABLE;
 623                 break;
 624         case ACPI_ADR_SPACE_FIXED_HARDWARE:
 625                 dprintk("HARDWARE addr space\n");
 626                 if (!check_est_cpu(cpu)) {
 627                         result = -ENODEV;
 628                         goto err_unreg;
 629                 }
 630                 data->cpu_feature = SYSTEM_INTEL_MSR_CAPABLE;
 631                 break;
 632         default:
 633                 dprintk("Unknown addr space %d\n",
 634                         (u32) (perf->control_register.space_id));
 635                 result = -ENODEV;
 636                 goto err_unreg;
 637         }
 638
 639         data->freq_table = kmalloc(sizeof(struct cpufreq_frequency_table) *
 640                     (perf->state_count+1), GFP_KERNEL);
 641         if (!data->freq_table) {
 642                 result = -ENOMEM;
 643                 goto err_unreg;
 644         }
 645
 646         /* detect transition latency */
 647         policy->cpuinfo.transition_latency = 0;
 648         for (i = 0; i < perf->state_count; i++) {
 649                 if ((perf->states[i].transition_latency * 1000) >
 650                     policy->cpuinfo.transition_latency)
 651                         policy->cpuinfo.transition_latency =
 652                             perf->states[i].transition_latency * 1000;
 653         }
 654
 655         /* Check for high latency (>20uS) from buggy BIOSes, like on T42 */
 656         if (perf->control_register.space_id == ACPI_ADR_SPACE_FIXED_HARDWARE &&
 657             policy->cpuinfo.transition_latency > 20 * 1000) {
 658                 policy->cpuinfo.transition_latency = 20 * 1000;
 659                 printk_once(KERN_INFO
 660                             "P-state transition latency capped at 20 uS\n");
 661         }
 662
 663         /* table init */
 664         for (i = 0; i < perf->state_count; i++) {
 665                 if (i > 0 && perf->states[i].core_frequency >=
 666                     data->freq_table[valid_states-1].frequency / 1000)
 667                         continue;
 668
 669                 data->freq_table[valid_states].index = i;
 670                 data->freq_table[valid_states].frequency =
 671                     perf->states[i].core_frequency * 1000;
 672                 valid_states++;
 673         }
 674         data->freq_table[valid_states].frequency = CPUFREQ_TABLE_END;
 675         perf->state = 0;
 676
 677         result = cpufreq_frequency_table_cpuinfo(policy, data->freq_table);
 678         if (result)
 679                 goto err_freqfree;
 680
 681         if (perf->states[0].core_frequency * 1000 != policy->cpuinfo.max_freq)
 682                 printk(KERN_WARNING FW_WARN "P-state 0 is not max freq\n");
 683
 684         switch (perf->control_register.space_id) {
 685         case ACPI_ADR_SPACE_SYSTEM_IO:
 686                 /* Current speed is unknown and not detectable by IO port */
 687                 policy->cur = acpi_cpufreq_guess_freq(data, policy->cpu);
 688                 break;
 689         case ACPI_ADR_SPACE_FIXED_HARDWARE:
 690                 acpi_cpufreq_driver.get = get_cur_freq_on_cpu;
 691                 policy->cur = get_cur_freq_on_cpu(cpu);
 692                 break;
 693         default:
 694                 break;
 695         }
 696
 697         /* notify BIOS that we exist */
 698         acpi_processor_notify_smm(THIS_MODULE);
 699
 700         /* Check for APERF/MPERF support in hardware */
 701         if (cpu_has(c, X86_FEATURE_APERFMPERF))
 702                 acpi_cpufreq_driver.getavg = get_measured_perf;
 703
 704         dprintk("CPU%u - ACPI performance management activated.\n", cpu);
 705         for (i = 0; i < perf->state_count; i++)
 706                 dprintk("     %cP%d: %d MHz, %d mW, %d uS\n",
 707                         (i == perf->state ? '*' : ' '), i,
 708                         (u32) perf->states[i].core_frequency,
 709                         (u32) perf->states[i].power,
 710                         (u32) perf->states[i].transition_latency);
 711
 712         cpufreq_frequency_table_get_attr(data->freq_table, policy->cpu);
 713
 714         /*
 715          * the first call to ->target() should result in us actually
 716          * writing something to the appropriate registers.
 717          */
 718         data->resume = 1;
 719
 720         return result;
 721
 722 err_freqfree:
 723         kfree(data->freq_table);
 724 err_unreg:
 725         acpi_processor_unregister_performance(perf, cpu);
 726 err_free:
 727         kfree(data);
 728         per_cpu(acfreq_data, cpu) = NULL;
 729
 730         return result;
 731 }
 732
 733 static int acpi_cpufreq_cpu_exit(struct cpufreq_policy *policy)
 734 {
 735         struct acpi_cpufreq_data *data = per_cpu(acfreq_data, policy->cpu);
 736
 737         dprintk("acpi_cpufreq_cpu_exit\n");
 738
 739         if (data) {
 740                 cpufreq_frequency_table_put_attr(policy->cpu);
 741                 per_cpu(acfreq_data, policy->cpu) = NULL;
 742                 acpi_processor_unregister_performance(data->acpi_data,
 743                                                       policy->cpu);
 744                 kfree(data);
 745         }
 746
 747         return 0;
 748 }
 749
 750 static int acpi_cpufreq_resume(struct cpufreq_policy *policy)
 751 {
 752         struct acpi_cpufreq_data *data = per_cpu(acfreq_data, policy->cpu);
 753
 754         dprintk("acpi_cpufreq_resume\n");
 755
 756         data->resume = 1;
 757
 758         return 0;
 759 }
 760
 761 static struct freq_attr *acpi_cpufreq_attr[] = {
 762         &cpufreq_freq_attr_scaling_available_freqs,
 763         NULL,
 764 };
 765
 766 static struct cpufreq_driver acpi_cpufreq_driver = {
 767         .verify         = acpi_cpufreq_verify,
 768         .target         = acpi_cpufreq_target,
 769         .bios_limit     = acpi_processor_get_bios_limit,
 770         .init           = acpi_cpufreq_cpu_init,
 771         .exit           = acpi_cpufreq_cpu_exit,
 772         .resume         = acpi_cpufreq_resume,
 773         .name           = "acpi-cpufreq",
 774         .owner          = THIS_MODULE,
 775         .attr           = acpi_cpufreq_attr,
 776 };
 777
 778 static int __init acpi_cpufreq_init(void)
 779 {
 780         int ret;
 781
 782         if (acpi_disabled)
 783                 return 0;
 784
 785         dprintk("acpi_cpufreq_init\n");
 786
 787         ret = acpi_cpufreq_early_init();
 788         if (ret)
 789                 return ret;
 790
 791         ret = cpufreq_register_driver(&acpi_cpufreq_driver);
 792         if (ret)
 793                 free_acpi_perf_data();
 794
 795         return ret;
 796 }
 797
 798 static void __exit acpi_cpufreq_exit(void)
 799 {
 800         dprintk("acpi_cpufreq_exit\n");
 801
 802         cpufreq_unregister_driver(&acpi_cpufreq_driver);
 803
 804         free_percpu(acpi_perf_data);
 805 }
 806
 807 module_param(acpi_pstate_strict, uint, 0644);
 808 MODULE_PARM_DESC(acpi_pstate_strict,
 809         "value 0 or non-zero. non-zero -> strict ACPI checks are "
 810         "performed during frequency changes.");
 811
 812 late_initcall(acpi_cpufreq_init);
 813 module_exit(acpi_cpufreq_exit);
 814
 815 MODULE_ALIAS("acpi");