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