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 <linux/slab.h>
  37 #include <trace/events/power.h>
  38
  39 #include <linux/acpi.h>
  40 #include <linux/io.h>
  41 #include <linux/delay.h>
  42 #include <linux/uaccess.h>
  43
  44 #include <acpi/processor.h>
  45
  46 #include <asm/msr.h>
  47 #include <asm/processor.h>
  48 #include <asm/cpufeature.h>
  49
  50 #define dprintk(msg...) cpufreq_debug_printk(CPUFREQ_DEBUG_DRIVER, \
  51                 "acpi-cpufreq", msg)
  52
  53 MODULE_AUTHOR("Paul Diefenbaugh, Dominik Brodowski");
  54 MODULE_DESCRIPTION("ACPI Processor P-States Driver");
  55 MODULE_LICENSE("GPL");
  56
  57 enum {
  58         UNDEFINED_CAPABLE = 0,
  59         SYSTEM_INTEL_MSR_CAPABLE,
  60         SYSTEM_IO_CAPABLE,
  61 };
  62
  63 #define INTEL_MSR_RANGE         (0xffff)
  64
  65 struct acpi_cpufreq_data {
  66         struct acpi_processor_performance *acpi_data;
  67         struct cpufreq_frequency_table *freq_table;
  68         unsigned int resume;
  69         unsigned int cpu_feature;
  70 };
  71
  72 static DEFINE_PER_CPU(struct acpi_cpufreq_data *, acfreq_data);
  73
  74 static DEFINE_PER_CPU(struct aperfmperf, acfreq_old_perf);
  75
  76 /* acpi_perf_data is a pointer to percpu data. */
  77 static struct acpi_processor_performance *acpi_perf_data;
  78
  79 static struct cpufreq_driver acpi_cpufreq_driver;
  80
  81 static unsigned int acpi_pstate_strict;
  82
  83 static int check_est_cpu(unsigned int cpuid)
  84 {
  85         struct cpuinfo_x86 *cpu = &cpu_data(cpuid);
  86
  87         return cpu_has(cpu, X86_FEATURE_EST);
  88 }
  89
  90 static unsigned extract_io(u32 value, struct acpi_cpufreq_data *data)
  91 {
  92         struct acpi_processor_performance *perf;
  93         int i;
  94
  95         perf = data->acpi_data;
  96
  97         for (i = 0; i < perf->state_count; i++) {
  98                 if (value == perf->states[i].status)
  99                         return data->freq_table[i].frequency;
 100         }
 101         return 0;
 102 }
 103
 104 static unsigned extract_msr(u32 msr, struct acpi_cpufreq_data *data)
 105 {
 106         int i;
 107         struct acpi_processor_performance *perf;
 108
 109         msr &= INTEL_MSR_RANGE;
 110         perf = data->acpi_data;
 111
 112         for (i = 0; data->freq_table[i].frequency != CPUFREQ_TABLE_END; i++) {
 113                 if (msr == perf->states[data->freq_table[i].index].status)
 114                         return data->freq_table[i].frequency;
 115         }
 116         return data->freq_table[0].frequency;
 117 }
 118
 119 static unsigned extract_freq(u32 val, struct acpi_cpufreq_data *data)
 120 {
 121         switch (data->cpu_feature) {
 122         case SYSTEM_INTEL_MSR_CAPABLE:
 123                 return extract_msr(val, data);
 124         case SYSTEM_IO_CAPABLE:
 125                 return extract_io(val, data);
 126         default:
 127                 return 0;
 128         }
 129 }
 130
 131 struct msr_addr {
 132         u32 reg;
 133 };
 134
 135 struct io_addr {
 136         u16 port;
 137         u8 bit_width;
 138 };
 139
 140 struct drv_cmd {
 141         unsigned int type;
 142         const struct cpumask *mask;
 143         union {
 144                 struct msr_addr msr;
 145                 struct io_addr io;
 146         } addr;
 147         u32 val;
 148 };
 149
 150 /* Called via smp_call_function_single(), on the target CPU */
 151 static void do_drv_read(void *_cmd)
 152 {
 153         struct drv_cmd *cmd = _cmd;
 154         u32 h;
 155
 156         switch (cmd->type) {
 157         case SYSTEM_INTEL_MSR_CAPABLE:
 158                 rdmsr(cmd->addr.msr.reg, cmd->val, h);
 159                 break;
 160         case SYSTEM_IO_CAPABLE:
 161                 acpi_os_read_port((acpi_io_address)cmd->addr.io.port,
 162                                 &cmd->val,
 163                                 (u32)cmd->addr.io.bit_width);
 164                 break;
 165         default:
 166                 break;
 167         }
 168 }
 169
 170 /* Called via smp_call_function_many(), on the target CPUs */
 171 static void do_drv_write(void *_cmd)
 172 {
 173         struct drv_cmd *cmd = _cmd;
 174         u32 lo, hi;
 175
 176         switch (cmd->type) {
 177         case SYSTEM_INTEL_MSR_CAPABLE:
 178                 rdmsr(cmd->addr.msr.reg, lo, hi);
 179                 lo = (lo & ~INTEL_MSR_RANGE) | (cmd->val & INTEL_MSR_RANGE);
 180                 wrmsr(cmd->addr.msr.reg, lo, hi);
 181                 break;
 182         case SYSTEM_IO_CAPABLE:
 183                 acpi_os_write_port((acpi_io_address)cmd->addr.io.port,
 184                                 cmd->val,
 185                                 (u32)cmd->addr.io.bit_width);
 186                 break;
 187         default:
 188                 break;
 189         }
 190 }
 191
 192 static void drv_read(struct drv_cmd *cmd)
 193 {
 194         int err;
 195         cmd->val = 0;
 196
 197         err = smp_call_function_any(cmd->mask, do_drv_read, cmd, 1);
 198         WARN_ON_ONCE(err);      /* smp_call_function_any() was buggy? */
 199 }
 200
 201 static void drv_write(struct drv_cmd *cmd)
 202 {
 203         int this_cpu;
 204
 205         this_cpu = get_cpu();
 206         if (cpumask_test_cpu(this_cpu, cmd->mask))
 207                 do_drv_write(cmd);
 208         smp_call_function_many(cmd->mask, do_drv_write, cmd, 1);
 209         put_cpu();
 210 }
 211
 212 static u32 get_cur_val(const struct cpumask *mask)
 213 {
 214         struct acpi_processor_performance *perf;
 215         struct drv_cmd cmd;
 216
 217         if (unlikely(cpumask_empty(mask)))
 218                 return 0;
 219
 220         switch (per_cpu(acfreq_data, cpumask_first(mask))->cpu_feature) {
 221         case SYSTEM_INTEL_MSR_CAPABLE:
 222                 cmd.type = SYSTEM_INTEL_MSR_CAPABLE;
 223                 cmd.addr.msr.reg = MSR_IA32_PERF_STATUS;
 224                 break;
 225         case SYSTEM_IO_CAPABLE:
 226                 cmd.type = SYSTEM_IO_CAPABLE;
 227                 perf = per_cpu(acfreq_data, cpumask_first(mask))->acpi_data;
 228                 cmd.addr.io.port = perf->control_register.address;
 229                 cmd.addr.io.bit_width = perf->control_register.bit_width;
 230                 break;
 231         default:
 232                 return 0;
 233         }
 234
 235         cmd.mask = mask;
 236         drv_read(&cmd);
 237
 238         dprintk("get_cur_val = %u\n", cmd.val);
 239
 240         return cmd.val;
 241 }
 242
 243 /* Called via smp_call_function_single(), on the target CPU */
 244 static void read_measured_perf_ctrs(void *_cur)
 245 {
 246         struct aperfmperf *am = _cur;
 247
 248         get_aperfmperf(am);
 249 }
 250
 251 /*
 252  * Return the measured active (C0) frequency on this CPU since last call
 253  * to this function.
 254  * Input: cpu number
 255  * Return: Average CPU frequency in terms of max frequency (zero on error)
 256  *
 257  * We use IA32_MPERF and IA32_APERF MSRs to get the measured performance
 258  * over a period of time, while CPU is in C0 state.
 259  * IA32_MPERF counts at the rate of max advertised frequency
 260  * IA32_APERF counts at the rate of actual CPU frequency
 261  * Only IA32_APERF/IA32_MPERF ratio is architecturally defined and
 262  * no meaning should be associated with absolute values of these MSRs.
 263  */
 264 static unsigned int get_measured_perf(struct cpufreq_policy *policy,
 265                                       unsigned int cpu)
 266 {
 267         struct aperfmperf perf;
 268         unsigned long ratio;
 269         unsigned int retval;
 270
 271         if (smp_call_function_single(cpu, read_measured_perf_ctrs, &perf, 1))
 272                 return 0;
 273
 274         ratio = calc_aperfmperf_ratio(&per_cpu(acfreq_old_perf, cpu), &perf);
 275         per_cpu(acfreq_old_perf, cpu) = perf;
 276
 277         retval = (policy->cpuinfo.max_freq * ratio) >> APERFMPERF_SHIFT;
 278
 279         return retval;
 280 }
 281
 282 static unsigned int get_cur_freq_on_cpu(unsigned int cpu)
 283 {
 284         struct acpi_cpufreq_data *data = per_cpu(acfreq_data, cpu);
 285         unsigned int freq;
 286         unsigned int cached_freq;
 287
 288         dprintk("get_cur_freq_on_cpu (%d)\n", cpu);
 289
 290         if (unlikely(data == NULL ||
 291                      data->acpi_data == NULL || data->freq_table == NULL)) {
 292                 return 0;
 293         }
 294
 295         cached_freq = data->freq_table[data->acpi_data->state].frequency;
 296         freq = extract_freq(get_cur_val(cpumask_of(cpu)), data);
 297         if (freq != cached_freq) {
 298                 /*
 299                  * The dreaded BIOS frequency change behind our back.
 300                  * Force set the frequency on next target call.
 301                  */
 302                 data->resume = 1;
 303         }
 304
 305         dprintk("cur freq = %u\n", freq);
 306
 307         return freq;
 308 }
 309
 310 static unsigned int check_freqs(const struct cpumask *mask, unsigned int freq,
 311                                 struct acpi_cpufreq_data *data)
 312 {
 313         unsigned int cur_freq;
 314         unsigned int i;
 315
 316         for (i = 0; i < 100; i++) {
 317                 cur_freq = extract_freq(get_cur_val(mask), data);
 318                 if (cur_freq == freq)
 319                         return 1;
 320                 udelay(10);
 321         }
 322         return 0;
 323 }
 324
 325 static int acpi_cpufreq_target(struct cpufreq_policy *policy,
 326                                unsigned int target_freq, unsigned int relation)
 327 {
 328         struct acpi_cpufreq_data *data = per_cpu(acfreq_data, policy->cpu);
 329         struct acpi_processor_performance *perf;
 330         struct cpufreq_freqs freqs;
 331         struct drv_cmd cmd;
 332         unsigned int next_state = 0; /* Index into freq_table */
 333         unsigned int next_perf_state = 0; /* Index into perf table */
 334         unsigned int i;
 335         int result = 0;
 336
 337         dprintk("acpi_cpufreq_target %d (%d)\n", target_freq, policy->cpu);
 338
 339         if (unlikely(data == NULL ||
 340              data->acpi_data == NULL || data->freq_table == NULL)) {
 341                 return -ENODEV;
 342         }
 343
 344         perf = data->acpi_data;
 345         result = cpufreq_frequency_table_target(policy,
 346                                                 data->freq_table,
 347                                                 target_freq,
 348                                                 relation, &next_state);
 349         if (unlikely(result)) {
 350                 result = -ENODEV;
 351                 goto out;
 352         }
 353
 354         next_perf_state = data->freq_table[next_state].index;
 355         if (perf->state == next_perf_state) {
 356                 if (unlikely(data->resume)) {
 357                         dprintk("Called after resume, resetting to P%d\n",
 358                                 next_perf_state);
 359                         data->resume = 0;
 360                 } else {
 361                         dprintk("Already at target state (P%d)\n",
 362                                 next_perf_state);
 363                         goto out;
 364                 }
 365         }
 366
 367         trace_power_frequency(POWER_PSTATE, data->freq_table[next_state].frequency);
 368
 369         switch (data->cpu_feature) {
 370         case SYSTEM_INTEL_MSR_CAPABLE:
 371                 cmd.type = SYSTEM_INTEL_MSR_CAPABLE;
 372                 cmd.addr.msr.reg = MSR_IA32_PERF_CTL;
 373                 cmd.val = (u32) perf->states[next_perf_state].control;
 374                 break;
 375         case SYSTEM_IO_CAPABLE:
 376                 cmd.type = SYSTEM_IO_CAPABLE;
 377                 cmd.addr.io.port = perf->control_register.address;
 378                 cmd.addr.io.bit_width = perf->control_register.bit_width;
 379                 cmd.val = (u32) perf->states[next_perf_state].control;
 380                 break;
 381         default:
 382                 result = -ENODEV;
 383                 goto out;
 384         }
 385
 386         /* cpufreq holds the hotplug lock, so we are safe from here on */
 387         if (policy->shared_type != CPUFREQ_SHARED_TYPE_ANY)
 388                 cmd.mask = policy->cpus;
 389         else
 390                 cmd.mask = cpumask_of(policy->cpu);
 391
 392         freqs.old = perf->states[perf->state].core_frequency * 1000;
 393         freqs.new = data->freq_table[next_state].frequency;
 394         for_each_cpu(i, cmd.mask) {
 395                 freqs.cpu = i;
 396                 cpufreq_notify_transition(&freqs, CPUFREQ_PRECHANGE);
 397         }
 398
 399         drv_write(&cmd);
 400
 401         if (acpi_pstate_strict) {
 402                 if (!check_freqs(cmd.mask, freqs.new, data)) {
 403                         dprintk("acpi_cpufreq_target failed (%d)\n",
 404                                 policy->cpu);
 405                         result = -EAGAIN;
 406                         goto out;
 407                 }
 408         }
 409
 410         for_each_cpu(i, cmd.mask) {
 411                 freqs.cpu = i;
 412                 cpufreq_notify_transition(&freqs, CPUFREQ_POSTCHANGE);
 413         }
 414         perf->state = next_perf_state;
 415
 416 out:
 417         return result;
 418 }
 419
 420 static int acpi_cpufreq_verify(struct cpufreq_policy *policy)
 421 {
 422         struct acpi_cpufreq_data *data = per_cpu(acfreq_data, policy->cpu);
 423
 424         dprintk("acpi_cpufreq_verify\n");
 425
 426         return cpufreq_frequency_table_verify(policy, data->freq_table);
 427 }
 428
 429 static unsigned long
 430 acpi_cpufreq_guess_freq(struct acpi_cpufreq_data *data, unsigned int cpu)
 431 {
 432         struct acpi_processor_performance *perf = data->acpi_data;
 433
 434         if (cpu_khz) {
 435                 /* search the closest match to cpu_khz */
 436                 unsigned int i;
 437                 unsigned long freq;
 438                 unsigned long freqn = perf->states[0].core_frequency * 1000;
 439
 440                 for (i = 0; i < (perf->state_count-1); i++) {
 441                         freq = freqn;
 442                         freqn = perf->states[i+1].core_frequency * 1000;
 443                         if ((2 * cpu_khz) > (freqn + freq)) {
 444                                 perf->state = i;
 445                                 return freq;
 446                         }
 447                 }
 448                 perf->state = perf->state_count-1;
 449                 return freqn;
 450         } else {
 451                 /* assume CPU is at P0... */
 452                 perf->state = 0;
 453                 return perf->states[0].core_frequency * 1000;
 454         }
 455 }
 456
 457 static void free_acpi_perf_data(void)
 458 {
 459         unsigned int i;
 460
 461         /* Freeing a NULL pointer is OK, and alloc_percpu zeroes. */
 462         for_each_possible_cpu(i)
 463                 free_cpumask_var(per_cpu_ptr(acpi_perf_data, i)
 464                                  ->shared_cpu_map);
 465         free_percpu(acpi_perf_data);
 466 }
 467
 468 /*
 469  * acpi_cpufreq_early_init - initialize ACPI P-States library
 470  *
 471  * Initialize the ACPI P-States library (drivers/acpi/processor_perflib.c)
 472  * in order to determine correct frequency and voltage pairings. We can
 473  * do _PDC and _PSD and find out the processor dependency for the
 474  * actual init that will happen later...
 475  */
 476 static int __init acpi_cpufreq_early_init(void)
 477 {
 478         unsigned int i;
 479         dprintk("acpi_cpufreq_early_init\n");
 480
 481         acpi_perf_data = alloc_percpu(struct acpi_processor_performance);
 482         if (!acpi_perf_data) {
 483                 dprintk("Memory allocation error for acpi_perf_data.\n");
 484                 return -ENOMEM;
 485         }
 486         for_each_possible_cpu(i) {
 487                 if (!zalloc_cpumask_var_node(
 488                         &per_cpu_ptr(acpi_perf_data, i)->shared_cpu_map,
 489                         GFP_KERNEL, cpu_to_node(i))) {
 490
 491                         /* Freeing a NULL pointer is OK: alloc_percpu zeroes. */
 492                         free_acpi_perf_data();
 493                         return -ENOMEM;
 494                 }
 495         }
 496
 497         /* Do initialization in ACPI core */
 498         acpi_processor_preregister_performance(acpi_perf_data);
 499         return 0;
 500 }
 501
 502 #ifdef CONFIG_SMP
 503 /*
 504  * Some BIOSes do SW_ANY coordination internally, either set it up in hw
 505  * or do it in BIOS firmware and won't inform about it to OS. If not
 506  * detected, this has a side effect of making CPU run at a different speed
 507  * than OS intended it to run at. Detect it and handle it cleanly.
 508  */
 509 static int bios_with_sw_any_bug;
 510
 511 static int sw_any_bug_found(const struct dmi_system_id *d)
 512 {
 513         bios_with_sw_any_bug = 1;
 514         return 0;
 515 }
 516
 517 static const struct dmi_system_id sw_any_bug_dmi_table[] = {
 518         {
 519                 .callback = sw_any_bug_found,
 520                 .ident = "Supermicro Server X6DLP",
 521                 .matches = {
 522                         DMI_MATCH(DMI_SYS_VENDOR, "Supermicro"),
 523                         DMI_MATCH(DMI_BIOS_VERSION, "080010"),
 524                         DMI_MATCH(DMI_PRODUCT_NAME, "X6DLP"),
 525                 },
 526         },
 527         { }
 528 };
 529
 530 static int acpi_cpufreq_blacklist(struct cpuinfo_x86 *c)
 531 {
 532         /* Intel Xeon Processor 7100 Series Specification Update
 533          * http://www.intel.com/Assets/PDF/specupdate/314554.pdf
 534          * AL30: A Machine Check Exception (MCE) Occurring during an
 535          * Enhanced Intel SpeedStep Technology Ratio Change May Cause
 536          * Both Processor Cores to Lock Up. */
 537         if (c->x86_vendor == X86_VENDOR_INTEL) {
 538                 if ((c->x86 == 15) &&
 539                     (c->x86_model == 6) &&
 540                     (c->x86_mask == 8)) {
 541                         printk(KERN_INFO "acpi-cpufreq: Intel(R) "
 542                             "Xeon(R) 7100 Errata AL30, processors may "
 543                             "lock up on frequency changes: disabling "
 544                             "acpi-cpufreq.\n");
 545                         return -ENODEV;
 546                     }
 547                 }
 548         return 0;
 549 }
 550 #endif
 551
 552 static int acpi_cpufreq_cpu_init(struct cpufreq_policy *policy)
 553 {
 554         unsigned int i;
 555         unsigned int valid_states = 0;
 556         unsigned int cpu = policy->cpu;
 557         struct acpi_cpufreq_data *data;
 558         unsigned int result = 0;
 559         struct cpuinfo_x86 *c = &cpu_data(policy->cpu);
 560         struct acpi_processor_performance *perf;
 561 #ifdef CONFIG_SMP
 562         static int blacklisted;
 563 #endif
 564
 565         dprintk("acpi_cpufreq_cpu_init\n");
 566
 567 #ifdef CONFIG_SMP
 568         if (blacklisted)
 569                 return blacklisted;
 570         blacklisted = acpi_cpufreq_blacklist(c);
 571         if (blacklisted)
 572                 return blacklisted;
 573 #endif
 574
 575         data = kzalloc(sizeof(struct acpi_cpufreq_data), GFP_KERNEL);
 576         if (!data)
 577                 return -ENOMEM;
 578
 579         data->acpi_data = per_cpu_ptr(acpi_perf_data, cpu);
 580         per_cpu(acfreq_data, cpu) = data;
 581
 582         if (cpu_has(c, X86_FEATURE_CONSTANT_TSC))
 583                 acpi_cpufreq_driver.flags |= CPUFREQ_CONST_LOOPS;
 584
 585         result = acpi_processor_register_performance(data->acpi_data, cpu);
 586         if (result)
 587                 goto err_free;
 588
 589         perf = data->acpi_data;
 590         policy->shared_type = perf->shared_type;
 591
 592         /*
 593          * Will let policy->cpus know about dependency only when software
 594          * coordination is required.
 595          */
 596         if (policy->shared_type == CPUFREQ_SHARED_TYPE_ALL ||
 597             policy->shared_type == CPUFREQ_SHARED_TYPE_ANY) {
 598                 cpumask_copy(policy->cpus, perf->shared_cpu_map);
 599         }
 600         cpumask_copy(policy->related_cpus, perf->shared_cpu_map);
 601
 602 #ifdef CONFIG_SMP
 603         dmi_check_system(sw_any_bug_dmi_table);
 604         if (bios_with_sw_any_bug && cpumask_weight(policy->cpus) == 1) {
 605                 policy->shared_type = CPUFREQ_SHARED_TYPE_ALL;
 606                 cpumask_copy(policy->cpus, cpu_core_mask(cpu));
 607         }
 608 #endif
 609
 610         /* capability check */
 611         if (perf->state_count <= 1) {
 612                 dprintk("No P-States\n");
 613                 result = -ENODEV;
 614                 goto err_unreg;
 615         }
 616
 617         if (perf->control_register.space_id != perf->status_register.space_id) {
 618                 result = -ENODEV;
 619                 goto err_unreg;
 620         }
 621
 622         switch (perf->control_register.space_id) {
 623         case ACPI_ADR_SPACE_SYSTEM_IO:
 624                 dprintk("SYSTEM IO addr space\n");
 625                 data->cpu_feature = SYSTEM_IO_CAPABLE;
 626                 break;
 627         case ACPI_ADR_SPACE_FIXED_HARDWARE:
 628                 dprintk("HARDWARE addr space\n");
 629                 if (!check_est_cpu(cpu)) {
 630                         result = -ENODEV;
 631                         goto err_unreg;
 632                 }
 633                 data->cpu_feature = SYSTEM_INTEL_MSR_CAPABLE;
 634                 break;
 635         default:
 636                 dprintk("Unknown addr space %d\n",
 637                         (u32) (perf->control_register.space_id));
 638                 result = -ENODEV;
 639                 goto err_unreg;
 640         }
 641
 642         data->freq_table = kmalloc(sizeof(struct cpufreq_frequency_table) *
 643                     (perf->state_count+1), GFP_KERNEL);
 644         if (!data->freq_table) {
 645                 result = -ENOMEM;
 646                 goto err_unreg;
 647         }
 648
 649         /* detect transition latency */
 650         policy->cpuinfo.transition_latency = 0;
 651         for (i = 0; i < perf->state_count; i++) {
 652                 if ((perf->states[i].transition_latency * 1000) >
 653                     policy->cpuinfo.transition_latency)
 654                         policy->cpuinfo.transition_latency =
 655                             perf->states[i].transition_latency * 1000;
 656         }
 657
 658         /* Check for high latency (>20uS) from buggy BIOSes, like on T42 */
 659         if (perf->control_register.space_id == ACPI_ADR_SPACE_FIXED_HARDWARE &&
 660             policy->cpuinfo.transition_latency > 20 * 1000) {
 661                 policy->cpuinfo.transition_latency = 20 * 1000;
 662                 printk_once(KERN_INFO
 663                             "P-state transition latency capped at 20 uS\n");
 664         }
 665
 666         /* table init */
 667         for (i = 0; i < perf->state_count; i++) {
 668                 if (i > 0 && perf->states[i].core_frequency >=
 669                     data->freq_table[valid_states-1].frequency / 1000)
 670                         continue;
 671
 672                 data->freq_table[valid_states].index = i;
 673                 data->freq_table[valid_states].frequency =
 674                     perf->states[i].core_frequency * 1000;
 675                 valid_states++;
 676         }
 677         data->freq_table[valid_states].frequency = CPUFREQ_TABLE_END;
 678         perf->state = 0;
 679
 680         result = cpufreq_frequency_table_cpuinfo(policy, data->freq_table);
 681         if (result)
 682                 goto err_freqfree;
 683
 684         if (perf->states[0].core_frequency * 1000 != policy->cpuinfo.max_freq)
 685                 printk(KERN_WARNING FW_WARN "P-state 0 is not max freq\n");
 686
 687         switch (perf->control_register.space_id) {
 688         case ACPI_ADR_SPACE_SYSTEM_IO:
 689                 /* Current speed is unknown and not detectable by IO port */
 690                 policy->cur = acpi_cpufreq_guess_freq(data, policy->cpu);
 691                 break;
 692         case ACPI_ADR_SPACE_FIXED_HARDWARE:
 693                 acpi_cpufreq_driver.get = get_cur_freq_on_cpu;
 694                 policy->cur = get_cur_freq_on_cpu(cpu);
 695                 break;
 696         default:
 697                 break;
 698         }
 699
 700         /* notify BIOS that we exist */
 701         acpi_processor_notify_smm(THIS_MODULE);
 702
 703         /* Check for APERF/MPERF support in hardware */
 704         if (cpu_has(c, X86_FEATURE_APERFMPERF))
 705                 acpi_cpufreq_driver.getavg = get_measured_perf;
 706
 707         dprintk("CPU%u - ACPI performance management activated.\n", cpu);
 708         for (i = 0; i < perf->state_count; i++)
 709                 dprintk("     %cP%d: %d MHz, %d mW, %d uS\n",
 710                         (i == perf->state ? '*' : ' '), i,
 711                         (u32) perf->states[i].core_frequency,
 712                         (u32) perf->states[i].power,
 713                         (u32) perf->states[i].transition_latency);
 714
 715         cpufreq_frequency_table_get_attr(data->freq_table, policy->cpu);
 716
 717         /*
 718          * the first call to ->target() should result in us actually
 719          * writing something to the appropriate registers.
 720          */
 721         data->resume = 1;
 722
 723         return result;
 724
 725 err_freqfree:
 726         kfree(data->freq_table);
 727 err_unreg:
 728         acpi_processor_unregister_performance(perf, cpu);
 729 err_free:
 730         kfree(data);
 731         per_cpu(acfreq_data, cpu) = NULL;
 732
 733         return result;
 734 }
 735
 736 static int acpi_cpufreq_cpu_exit(struct cpufreq_policy *policy)
 737 {
 738         struct acpi_cpufreq_data *data = per_cpu(acfreq_data, policy->cpu);
 739
 740         dprintk("acpi_cpufreq_cpu_exit\n");
 741
 742         if (data) {
 743                 cpufreq_frequency_table_put_attr(policy->cpu);
 744                 per_cpu(acfreq_data, policy->cpu) = NULL;
 745                 acpi_processor_unregister_performance(data->acpi_data,
 746                                                       policy->cpu);
 747                 kfree(data);
 748         }
 749
 750         return 0;
 751 }
 752
 753 static int acpi_cpufreq_resume(struct cpufreq_policy *policy)
 754 {
 755         struct acpi_cpufreq_data *data = per_cpu(acfreq_data, policy->cpu);
 756
 757         dprintk("acpi_cpufreq_resume\n");
 758
 759         data->resume = 1;
 760
 761         return 0;
 762 }
 763
 764 static struct freq_attr *acpi_cpufreq_attr[] = {
 765         &cpufreq_freq_attr_scaling_available_freqs,
 766         NULL,
 767 };
 768
 769 static struct cpufreq_driver acpi_cpufreq_driver = {
 770         .verify         = acpi_cpufreq_verify,
 771         .target         = acpi_cpufreq_target,
 772         .bios_limit     = acpi_processor_get_bios_limit,
 773         .init           = acpi_cpufreq_cpu_init,
 774         .exit           = acpi_cpufreq_cpu_exit,
 775         .resume         = acpi_cpufreq_resume,
 776         .name           = "acpi-cpufreq",
 777         .owner          = THIS_MODULE,
 778         .attr           = acpi_cpufreq_attr,
 779 };
 780
 781 static int __init acpi_cpufreq_init(void)
 782 {
 783         int ret;
 784
 785         if (acpi_disabled)
 786                 return 0;
 787
 788         dprintk("acpi_cpufreq_init\n");
 789
 790         ret = acpi_cpufreq_early_init();
 791         if (ret)
 792                 return ret;
 793
 794         ret = cpufreq_register_driver(&acpi_cpufreq_driver);
 795         if (ret)
 796                 free_acpi_perf_data();
 797
 798         return ret;
 799 }
 800
 801 static void __exit acpi_cpufreq_exit(void)
 802 {
 803         dprintk("acpi_cpufreq_exit\n");
 804
 805         cpufreq_unregister_driver(&acpi_cpufreq_driver);
 806
 807         free_percpu(acpi_perf_data);
 808 }
 809
 810 module_param(acpi_pstate_strict, uint, 0644);
 811 MODULE_PARM_DESC(acpi_pstate_strict,
 812         "value 0 or non-zero. non-zero -> strict ACPI checks are "
 813         "performed during frequency changes.");
 814
 815 late_initcall(acpi_cpufreq_init);
 816 module_exit(acpi_cpufreq_exit);
 817
 818 MODULE_ALIAS("acpi");