SAFE public projects git trees. - safe/jmp/linux-2.6/blob - 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/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 #define CPUID_6_ECX_APERFMPERF_CAPABILITY       (0x1)
  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 *, drv_data);
  73
  74 struct acpi_msr_data {
  75         u64 saved_aperf, saved_mperf;
  76 };
  77
  78 static DEFINE_PER_CPU(struct acpi_msr_data, msr_data);
  79
  80 DEFINE_TRACE(power_mark);
  81
  82 /* acpi_perf_data is a pointer to percpu data. */
  83 static struct acpi_processor_performance *acpi_perf_data;
  84
  85 static struct cpufreq_driver acpi_cpufreq_driver;
  86
  87 static unsigned int acpi_pstate_strict;
  88
  89 static int check_est_cpu(unsigned int cpuid)
  90 {
  91         struct cpuinfo_x86 *cpu = &cpu_data(cpuid);
  92
  93         return cpu_has(cpu, X86_FEATURE_EST);
  94 }
  95
  96 static unsigned extract_io(u32 value, struct acpi_cpufreq_data *data)
  97 {
  98         struct acpi_processor_performance *perf;
  99         int i;
 100
 101         perf = data->acpi_data;
 102
 103         for (i = 0; i < perf->state_count; i++) {
 104                 if (value == perf->states[i].status)
 105                         return data->freq_table[i].frequency;
 106         }
 107         return 0;
 108 }
 109
 110 static unsigned extract_msr(u32 msr, struct acpi_cpufreq_data *data)
 111 {
 112         int i;
 113         struct acpi_processor_performance *perf;
 114
 115         msr &= INTEL_MSR_RANGE;
 116         perf = data->acpi_data;
 117
 118         for (i = 0; data->freq_table[i].frequency != CPUFREQ_TABLE_END; i++) {
 119                 if (msr == perf->states[data->freq_table[i].index].status)
 120                         return data->freq_table[i].frequency;
 121         }
 122         return data->freq_table[0].frequency;
 123 }
 124
 125 static unsigned extract_freq(u32 val, struct acpi_cpufreq_data *data)
 126 {
 127         switch (data->cpu_feature) {
 128         case SYSTEM_INTEL_MSR_CAPABLE:
 129                 return extract_msr(val, data);
 130         case SYSTEM_IO_CAPABLE:
 131                 return extract_io(val, data);
 132         default:
 133                 return 0;
 134         }
 135 }
 136
 137 struct msr_addr {
 138         u32 reg;
 139 };
 140
 141 struct io_addr {
 142         u16 port;
 143         u8 bit_width;
 144 };
 145
 146 struct drv_cmd {
 147         unsigned int type;
 148         const struct cpumask *mask;
 149         union {
 150                 struct msr_addr msr;
 151                 struct io_addr io;
 152         } addr;
 153         u32 val;
 154 };
 155
 156 /* Called via smp_call_function_single(), on the target CPU */
 157 static void do_drv_read(void *_cmd)
 158 {
 159         struct drv_cmd *cmd = _cmd;
 160         u32 h;
 161
 162         switch (cmd->type) {
 163         case SYSTEM_INTEL_MSR_CAPABLE:
 164                 rdmsr(cmd->addr.msr.reg, cmd->val, h);
 165                 break;
 166         case SYSTEM_IO_CAPABLE:
 167                 acpi_os_read_port((acpi_io_address)cmd->addr.io.port,
 168                                 &cmd->val,
 169                                 (u32)cmd->addr.io.bit_width);
 170                 break;
 171         default:
 172                 break;
 173         }
 174 }
 175
 176 /* Called via smp_call_function_many(), on the target CPUs */
 177 static void do_drv_write(void *_cmd)
 178 {
 179         struct drv_cmd *cmd = _cmd;
 180         u32 lo, hi;
 181
 182         switch (cmd->type) {
 183         case SYSTEM_INTEL_MSR_CAPABLE:
 184                 rdmsr(cmd->addr.msr.reg, lo, hi);
 185                 lo = (lo & ~INTEL_MSR_RANGE) | (cmd->val & INTEL_MSR_RANGE);
 186                 wrmsr(cmd->addr.msr.reg, lo, hi);
 187                 break;
 188         case SYSTEM_IO_CAPABLE:
 189                 acpi_os_write_port((acpi_io_address)cmd->addr.io.port,
 190                                 cmd->val,
 191                                 (u32)cmd->addr.io.bit_width);
 192                 break;
 193         default:
 194                 break;
 195         }
 196 }
 197
 198 static void drv_read(struct drv_cmd *cmd)
 199 {
 200         cmd->val = 0;
 201
 202         smp_call_function_single(cpumask_any(cmd->mask), do_drv_read, cmd, 1);
 203 }
 204
 205 static void drv_write(struct drv_cmd *cmd)
 206 {
 207         int this_cpu;
 208
 209         this_cpu = get_cpu();
 210         if (cpumask_test_cpu(this_cpu, cmd->mask))
 211                 do_drv_write(cmd);
 212         smp_call_function_many(cmd->mask, do_drv_write, cmd, 1);
 213         put_cpu();
 214 }
 215
 216 static u32 get_cur_val(const struct cpumask *mask)
 217 {
 218         struct acpi_processor_performance *perf;
 219         struct drv_cmd cmd;
 220
 221         if (unlikely(cpumask_empty(mask)))
 222                 return 0;
 223
 224         switch (per_cpu(drv_data, cpumask_first(mask))->cpu_feature) {
 225         case SYSTEM_INTEL_MSR_CAPABLE:
 226                 cmd.type = SYSTEM_INTEL_MSR_CAPABLE;
 227                 cmd.addr.msr.reg = MSR_IA32_PERF_STATUS;
 228                 break;
 229         case SYSTEM_IO_CAPABLE:
 230                 cmd.type = SYSTEM_IO_CAPABLE;
 231                 perf = per_cpu(drv_data, cpumask_first(mask))->acpi_data;
 232                 cmd.addr.io.port = perf->control_register.address;
 233                 cmd.addr.io.bit_width = perf->control_register.bit_width;
 234                 break;
 235         default:
 236                 return 0;
 237         }
 238
 239         cmd.mask = mask;
 240         drv_read(&cmd);
 241
 242         dprintk("get_cur_val = %u\n", cmd.val);
 243
 244         return cmd.val;
 245 }
 246
 247 struct perf_pair {
 248         union {
 249                 struct {
 250                         u32 lo;
 251                         u32 hi;
 252                 } split;
 253                 u64 whole;
 254         } aperf, mperf;
 255 };
 256
 257 /* Called via smp_call_function_single(), on the target CPU */
 258 static void read_measured_perf_ctrs(void *_cur)
 259 {
 260         struct perf_pair *cur = _cur;
 261
 262         rdmsr(MSR_IA32_APERF, cur->aperf.split.lo, cur->aperf.split.hi);
 263         rdmsr(MSR_IA32_MPERF, cur->mperf.split.lo, cur->mperf.split.hi);
 264 }
 265
 266 /*
 267  * Return the measured active (C0) frequency on this CPU since last call
 268  * to this function.
 269  * Input: cpu number
 270  * Return: Average CPU frequency in terms of max frequency (zero on error)
 271  *
 272  * We use IA32_MPERF and IA32_APERF MSRs to get the measured performance
 273  * over a period of time, while CPU is in C0 state.
 274  * IA32_MPERF counts at the rate of max advertised frequency
 275  * IA32_APERF counts at the rate of actual CPU frequency
 276  * Only IA32_APERF/IA32_MPERF ratio is architecturally defined and
 277  * no meaning should be associated with absolute values of these MSRs.
 278  */
 279 static unsigned int get_measured_perf(struct cpufreq_policy *policy,
 280                                       unsigned int cpu)
 281 {
 282         struct perf_pair readin, cur;
 283         unsigned int perf_percent;
 284         unsigned int retval;
 285
 286         if (smp_call_function_single(cpu, read_measured_perf_ctrs, &readin, 1))
 287                 return 0;
 288
 289         cur.aperf.whole = readin.aperf.whole -
 290                                 per_cpu(msr_data, cpu).saved_aperf;
 291         cur.mperf.whole = readin.mperf.whole -
 292                                 per_cpu(msr_data, cpu).saved_mperf;
 293         per_cpu(msr_data, cpu).saved_aperf = readin.aperf.whole;
 294         per_cpu(msr_data, cpu).saved_mperf = readin.mperf.whole;
 295
 296 #ifdef __i386__
 297         /*
 298          * We dont want to do 64 bit divide with 32 bit kernel
 299          * Get an approximate value. Return failure in case we cannot get
 300          * an approximate value.
 301          */
 302         if (unlikely(cur.aperf.split.hi || cur.mperf.split.hi)) {
 303                 int shift_count;
 304                 u32 h;
 305
 306                 h = max_t(u32, cur.aperf.split.hi, cur.mperf.split.hi);
 307                 shift_count = fls(h);
 308
 309                 cur.aperf.whole >>= shift_count;
 310                 cur.mperf.whole >>= shift_count;
 311         }
 312
 313         if (((unsigned long)(-1) / 100) < cur.aperf.split.lo) {
 314                 int shift_count = 7;
 315                 cur.aperf.split.lo >>= shift_count;
 316                 cur.mperf.split.lo >>= shift_count;
 317         }
 318
 319         if (cur.aperf.split.lo && cur.mperf.split.lo)
 320                 perf_percent = (cur.aperf.split.lo * 100) / cur.mperf.split.lo;
 321         else
 322                 perf_percent = 0;
 323
 324 #else
 325         if (unlikely(((unsigned long)(-1) / 100) < cur.aperf.whole)) {
 326                 int shift_count = 7;
 327                 cur.aperf.whole >>= shift_count;
 328                 cur.mperf.whole >>= shift_count;
 329         }
 330
 331         if (cur.aperf.whole && cur.mperf.whole)
 332                 perf_percent = (cur.aperf.whole * 100) / cur.mperf.whole;
 333         else
 334                 perf_percent = 0;
 335
 336 #endif
 337
 338         retval = (policy->cpuinfo.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 struct cpumask *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         perf = data->acpi_data;
 407         result = cpufreq_frequency_table_target(policy,
 408                                                 data->freq_table,
 409                                                 target_freq,
 410                                                 relation, &next_state);
 411         if (unlikely(result)) {
 412                 result = -ENODEV;
 413                 goto out;
 414         }
 415
 416         next_perf_state = data->freq_table[next_state].index;
 417         if (perf->state == next_perf_state) {
 418                 if (unlikely(data->resume)) {
 419                         dprintk("Called after resume, resetting to P%d\n",
 420                                 next_perf_state);
 421                         data->resume = 0;
 422                 } else {
 423                         dprintk("Already at target state (P%d)\n",
 424                                 next_perf_state);
 425                         goto out;
 426                 }
 427         }
 428
 429         trace_power_mark(&it, POWER_PSTATE, next_perf_state);
 430
 431         switch (data->cpu_feature) {
 432         case SYSTEM_INTEL_MSR_CAPABLE:
 433                 cmd.type = SYSTEM_INTEL_MSR_CAPABLE;
 434                 cmd.addr.msr.reg = MSR_IA32_PERF_CTL;
 435                 cmd.val = (u32) perf->states[next_perf_state].control;
 436                 break;
 437         case SYSTEM_IO_CAPABLE:
 438                 cmd.type = SYSTEM_IO_CAPABLE;
 439                 cmd.addr.io.port = perf->control_register.address;
 440                 cmd.addr.io.bit_width = perf->control_register.bit_width;
 441                 cmd.val = (u32) perf->states[next_perf_state].control;
 442                 break;
 443         default:
 444                 result = -ENODEV;
 445                 goto out;
 446         }
 447
 448         /* cpufreq holds the hotplug lock, so we are safe from here on */
 449         if (policy->shared_type != CPUFREQ_SHARED_TYPE_ANY)
 450                 cmd.mask = policy->cpus;
 451         else
 452                 cmd.mask = cpumask_of(policy->cpu);
 453
 454         freqs.old = perf->states[perf->state].core_frequency * 1000;
 455         freqs.new = data->freq_table[next_state].frequency;
 456         for_each_cpu(i, cmd.mask) {
 457                 freqs.cpu = i;
 458                 cpufreq_notify_transition(&freqs, CPUFREQ_PRECHANGE);
 459         }
 460
 461         drv_write(&cmd);
 462
 463         if (acpi_pstate_strict) {
 464                 if (!check_freqs(cmd.mask, freqs.new, data)) {
 465                         dprintk("acpi_cpufreq_target failed (%d)\n",
 466                                 policy->cpu);
 467                         result = -EAGAIN;
 468                         goto out;
 469                 }
 470         }
 471
 472         for_each_cpu(i, cmd.mask) {
 473                 freqs.cpu = i;
 474                 cpufreq_notify_transition(&freqs, CPUFREQ_POSTCHANGE);
 475         }
 476         perf->state = next_perf_state;
 477
 478 out:
 479         return result;
 480 }
 481
 482 static int acpi_cpufreq_verify(struct cpufreq_policy *policy)
 483 {
 484         struct acpi_cpufreq_data *data = per_cpu(drv_data, policy->cpu);
 485
 486         dprintk("acpi_cpufreq_verify\n");
 487
 488         return cpufreq_frequency_table_verify(policy, data->freq_table);
 489 }
 490
 491 static unsigned long
 492 acpi_cpufreq_guess_freq(struct acpi_cpufreq_data *data, unsigned int cpu)
 493 {
 494         struct acpi_processor_performance *perf = data->acpi_data;
 495
 496         if (cpu_khz) {
 497                 /* search the closest match to cpu_khz */
 498                 unsigned int i;
 499                 unsigned long freq;
 500                 unsigned long freqn = perf->states[0].core_frequency * 1000;
 501
 502                 for (i = 0; i < (perf->state_count-1); i++) {
 503                         freq = freqn;
 504                         freqn = perf->states[i+1].core_frequency * 1000;
 505                         if ((2 * cpu_khz) > (freqn + freq)) {
 506                                 perf->state = i;
 507                                 return freq;
 508                         }
 509                 }
 510                 perf->state = perf->state_count-1;
 511                 return freqn;
 512         } else {
 513                 /* assume CPU is at P0... */
 514                 perf->state = 0;
 515                 return perf->states[0].core_frequency * 1000;
 516         }
 517 }
 518
 519 static void free_acpi_perf_data(void)
 520 {
 521         unsigned int i;
 522
 523         /* Freeing a NULL pointer is OK, and alloc_percpu zeroes. */
 524         for_each_possible_cpu(i)
 525                 free_cpumask_var(per_cpu_ptr(acpi_perf_data, i)
 526                                  ->shared_cpu_map);
 527         free_percpu(acpi_perf_data);
 528 }
 529
 530 /*
 531  * acpi_cpufreq_early_init - initialize ACPI P-States library
 532  *
 533  * Initialize the ACPI P-States library (drivers/acpi/processor_perflib.c)
 534  * in order to determine correct frequency and voltage pairings. We can
 535  * do _PDC and _PSD and find out the processor dependency for the
 536  * actual init that will happen later...
 537  */
 538 static int __init acpi_cpufreq_early_init(void)
 539 {
 540         unsigned int i;
 541         dprintk("acpi_cpufreq_early_init\n");
 542
 543         acpi_perf_data = alloc_percpu(struct acpi_processor_performance);
 544         if (!acpi_perf_data) {
 545                 dprintk("Memory allocation error for acpi_perf_data.\n");
 546                 return -ENOMEM;
 547         }
 548         for_each_possible_cpu(i) {
 549                 if (!zalloc_cpumask_var_node(
 550                         &per_cpu_ptr(acpi_perf_data, i)->shared_cpu_map,
 551                         GFP_KERNEL, cpu_to_node(i))) {
 552
 553                         /* Freeing a NULL pointer is OK: alloc_percpu zeroes. */
 554                         free_acpi_perf_data();
 555                         return -ENOMEM;
 556                 }
 557         }
 558
 559         /* Do initialization in ACPI core */
 560         acpi_processor_preregister_performance(acpi_perf_data);
 561         return 0;
 562 }
 563
 564 #ifdef CONFIG_SMP
 565 /*
 566  * Some BIOSes do SW_ANY coordination internally, either set it up in hw
 567  * or do it in BIOS firmware and won't inform about it to OS. If not
 568  * detected, this has a side effect of making CPU run at a different speed
 569  * than OS intended it to run at. Detect it and handle it cleanly.
 570  */
 571 static int bios_with_sw_any_bug;
 572
 573 static int sw_any_bug_found(const struct dmi_system_id *d)
 574 {
 575         bios_with_sw_any_bug = 1;
 576         return 0;
 577 }
 578
 579 static const struct dmi_system_id sw_any_bug_dmi_table[] = {
 580         {
 581                 .callback = sw_any_bug_found,
 582                 .ident = "Supermicro Server X6DLP",
 583                 .matches = {
 584                         DMI_MATCH(DMI_SYS_VENDOR, "Supermicro"),
 585                         DMI_MATCH(DMI_BIOS_VERSION, "080010"),
 586                         DMI_MATCH(DMI_PRODUCT_NAME, "X6DLP"),
 587                 },
 588         },
 589         { }
 590 };
 591
 592 static int acpi_cpufreq_blacklist(struct cpuinfo_x86 *c)
 593 {
 594         /* http://www.intel.com/Assets/PDF/specupdate/314554.pdf
 595          * AL30: A Machine Check Exception (MCE) Occurring during an
 596          * Enhanced Intel SpeedStep Technology Ratio Change May Cause
 597          * Both Processor Cores to Lock Up when HT is enabled*/
 598         if (c->x86_vendor == X86_VENDOR_INTEL) {
 599                 if ((c->x86 == 15) &&
 600                     (c->x86_model == 6) &&
 601                     (c->x86_mask == 8) && smt_capable())
 602                         return -ENODEV;
 603                 }
 604         return 0;
 605 }
 606 #endif
 607
 608 static int acpi_cpufreq_cpu_init(struct cpufreq_policy *policy)
 609 {
 610         unsigned int i;
 611         unsigned int valid_states = 0;
 612         unsigned int cpu = policy->cpu;
 613         struct acpi_cpufreq_data *data;
 614         unsigned int result = 0;
 615         struct cpuinfo_x86 *c = &cpu_data(policy->cpu);
 616         struct acpi_processor_performance *perf;
 617
 618         dprintk("acpi_cpufreq_cpu_init\n");
 619
 620 #ifdef CONFIG_SMP
 621         result = acpi_cpufreq_blacklist(c);
 622         if (result)
 623                 return result;
 624 #endif
 625
 626         data = kzalloc(sizeof(struct acpi_cpufreq_data), GFP_KERNEL);
 627         if (!data)
 628                 return -ENOMEM;
 629
 630         data->acpi_data = per_cpu_ptr(acpi_perf_data, cpu);
 631         per_cpu(drv_data, cpu) = data;
 632
 633         if (cpu_has(c, X86_FEATURE_CONSTANT_TSC))
 634                 acpi_cpufreq_driver.flags |= CPUFREQ_CONST_LOOPS;
 635
 636         result = acpi_processor_register_performance(data->acpi_data, cpu);
 637         if (result)
 638                 goto err_free;
 639
 640         perf = data->acpi_data;
 641         policy->shared_type = perf->shared_type;
 642
 643         /*
 644          * Will let policy->cpus know about dependency only when software
 645          * coordination is required.
 646          */
 647         if (policy->shared_type == CPUFREQ_SHARED_TYPE_ALL ||
 648             policy->shared_type == CPUFREQ_SHARED_TYPE_ANY) {
 649                 cpumask_copy(policy->cpus, perf->shared_cpu_map);
 650         }
 651         cpumask_copy(policy->related_cpus, perf->shared_cpu_map);
 652
 653 #ifdef CONFIG_SMP
 654         dmi_check_system(sw_any_bug_dmi_table);
 655         if (bios_with_sw_any_bug && cpumask_weight(policy->cpus) == 1) {
 656                 policy->shared_type = CPUFREQ_SHARED_TYPE_ALL;
 657                 cpumask_copy(policy->cpus, cpu_core_mask(cpu));
 658         }
 659 #endif
 660
 661         /* capability check */
 662         if (perf->state_count <= 1) {
 663                 dprintk("No P-States\n");
 664                 result = -ENODEV;
 665                 goto err_unreg;
 666         }
 667
 668         if (perf->control_register.space_id != perf->status_register.space_id) {
 669                 result = -ENODEV;
 670                 goto err_unreg;
 671         }
 672
 673         switch (perf->control_register.space_id) {
 674         case ACPI_ADR_SPACE_SYSTEM_IO:
 675                 dprintk("SYSTEM IO addr space\n");
 676                 data->cpu_feature = SYSTEM_IO_CAPABLE;
 677                 break;
 678         case ACPI_ADR_SPACE_FIXED_HARDWARE:
 679                 dprintk("HARDWARE addr space\n");
 680                 if (!check_est_cpu(cpu)) {
 681                         result = -ENODEV;
 682                         goto err_unreg;
 683                 }
 684                 data->cpu_feature = SYSTEM_INTEL_MSR_CAPABLE;
 685                 break;
 686         default:
 687                 dprintk("Unknown addr space %d\n",
 688                         (u32) (perf->control_register.space_id));
 689                 result = -ENODEV;
 690                 goto err_unreg;
 691         }
 692
 693         data->freq_table = kmalloc(sizeof(struct cpufreq_frequency_table) *
 694                     (perf->state_count+1), GFP_KERNEL);
 695         if (!data->freq_table) {
 696                 result = -ENOMEM;
 697                 goto err_unreg;
 698         }
 699
 700         /* detect transition latency */
 701         policy->cpuinfo.transition_latency = 0;
 702         for (i = 0; i < perf->state_count; i++) {
 703                 if ((perf->states[i].transition_latency * 1000) >
 704                     policy->cpuinfo.transition_latency)
 705                         policy->cpuinfo.transition_latency =
 706                             perf->states[i].transition_latency * 1000;
 707         }
 708
 709         /* Check for high latency (>20uS) from buggy BIOSes, like on T42 */
 710         if (perf->control_register.space_id == ACPI_ADR_SPACE_FIXED_HARDWARE &&
 711             policy->cpuinfo.transition_latency > 20 * 1000) {
 712                 policy->cpuinfo.transition_latency = 20 * 1000;
 713                 printk_once(KERN_INFO
 714                             "P-state transition latency capped at 20 uS\n");
 715         }
 716
 717         /* table init */
 718         for (i = 0; i < perf->state_count; i++) {
 719                 if (i > 0 && perf->states[i].core_frequency >=
 720                     data->freq_table[valid_states-1].frequency / 1000)
 721                         continue;
 722
 723                 data->freq_table[valid_states].index = i;
 724                 data->freq_table[valid_states].frequency =
 725                     perf->states[i].core_frequency * 1000;
 726                 valid_states++;
 727         }
 728         data->freq_table[valid_states].frequency = CPUFREQ_TABLE_END;
 729         perf->state = 0;
 730
 731         result = cpufreq_frequency_table_cpuinfo(policy, data->freq_table);
 732         if (result)
 733                 goto err_freqfree;
 734
 735         if (perf->states[0].core_frequency * 1000 != policy->cpuinfo.max_freq)
 736                 printk(KERN_WARNING FW_WARN "P-state 0 is not max freq\n");
 737
 738         switch (perf->control_register.space_id) {
 739         case ACPI_ADR_SPACE_SYSTEM_IO:
 740                 /* Current speed is unknown and not detectable by IO port */
 741                 policy->cur = acpi_cpufreq_guess_freq(data, policy->cpu);
 742                 break;
 743         case ACPI_ADR_SPACE_FIXED_HARDWARE:
 744                 acpi_cpufreq_driver.get = get_cur_freq_on_cpu;
 745                 policy->cur = get_cur_freq_on_cpu(cpu);
 746                 break;
 747         default:
 748                 break;
 749         }
 750
 751         /* notify BIOS that we exist */
 752         acpi_processor_notify_smm(THIS_MODULE);
 753
 754         /* Check for APERF/MPERF support in hardware */
 755         if (c->x86_vendor == X86_VENDOR_INTEL && c->cpuid_level >= 6) {
 756                 unsigned int ecx;
 757                 ecx = cpuid_ecx(6);
 758                 if (ecx & CPUID_6_ECX_APERFMPERF_CAPABILITY)
 759                         acpi_cpufreq_driver.getavg = get_measured_perf;
 760         }
 761
 762         dprintk("CPU%u - ACPI performance management activated.\n", cpu);
 763         for (i = 0; i < perf->state_count; i++)
 764                 dprintk("     %cP%d: %d MHz, %d mW, %d uS\n",
 765                         (i == perf->state ? '*' : ' '), i,
 766                         (u32) perf->states[i].core_frequency,
 767                         (u32) perf->states[i].power,
 768                         (u32) perf->states[i].transition_latency);
 769
 770         cpufreq_frequency_table_get_attr(data->freq_table, policy->cpu);
 771
 772         /*
 773          * the first call to ->target() should result in us actually
 774          * writing something to the appropriate registers.
 775          */
 776         data->resume = 1;
 777
 778         return result;
 779
 780 err_freqfree:
 781         kfree(data->freq_table);
 782 err_unreg:
 783         acpi_processor_unregister_performance(perf, cpu);
 784 err_free:
 785         kfree(data);
 786         per_cpu(drv_data, cpu) = NULL;
 787
 788         return result;
 789 }
 790
 791 static int acpi_cpufreq_cpu_exit(struct cpufreq_policy *policy)
 792 {
 793         struct acpi_cpufreq_data *data = per_cpu(drv_data, policy->cpu);
 794
 795         dprintk("acpi_cpufreq_cpu_exit\n");
 796
 797         if (data) {
 798                 cpufreq_frequency_table_put_attr(policy->cpu);
 799                 per_cpu(drv_data, policy->cpu) = NULL;
 800                 acpi_processor_unregister_performance(data->acpi_data,
 801                                                       policy->cpu);
 802                 kfree(data);
 803         }
 804
 805         return 0;
 806 }
 807
 808 static int acpi_cpufreq_resume(struct cpufreq_policy *policy)
 809 {
 810         struct acpi_cpufreq_data *data = per_cpu(drv_data, policy->cpu);
 811
 812         dprintk("acpi_cpufreq_resume\n");
 813
 814         data->resume = 1;
 815
 816         return 0;
 817 }
 818
 819 static struct freq_attr *acpi_cpufreq_attr[] = {
 820         &cpufreq_freq_attr_scaling_available_freqs,
 821         NULL,
 822 };
 823
 824 static struct cpufreq_driver acpi_cpufreq_driver = {
 825         .verify = acpi_cpufreq_verify,
 826         .target = acpi_cpufreq_target,
 827         .init = acpi_cpufreq_cpu_init,
 828         .exit = acpi_cpufreq_cpu_exit,
 829         .resume = acpi_cpufreq_resume,
 830         .name = "acpi-cpufreq",
 831         .owner = THIS_MODULE,
 832         .attr = acpi_cpufreq_attr,
 833 };
 834
 835 static int __init acpi_cpufreq_init(void)
 836 {
 837         int ret;
 838
 839         if (acpi_disabled)
 840                 return 0;
 841
 842         dprintk("acpi_cpufreq_init\n");
 843
 844         ret = acpi_cpufreq_early_init();
 845         if (ret)
 846                 return ret;
 847
 848         ret = cpufreq_register_driver(&acpi_cpufreq_driver);
 849         if (ret)
 850                 free_acpi_perf_data();
 851
 852         return ret;
 853 }
 854
 855 static void __exit acpi_cpufreq_exit(void)
 856 {
 857         dprintk("acpi_cpufreq_exit\n");
 858
 859         cpufreq_unregister_driver(&acpi_cpufreq_driver);
 860
 861         free_percpu(acpi_perf_data);
 862 }
 863
 864 module_param(acpi_pstate_strict, uint, 0644);
 865 MODULE_PARM_DESC(acpi_pstate_strict,
 866         "value 0 or non-zero. non-zero -> strict ACPI checks are "
 867         "performed during frequency changes.");
 868
 869 late_initcall(acpi_cpufreq_init);
 870 module_exit(acpi_cpufreq_exit);
 871
 872 MODULE_ALIAS("acpi");