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