drivers/acpi/processor_idle.c

   1 /*
   2  * processor_idle - idle state submodule to the ACPI processor 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) 2004       Dominik Brodowski <linux@brodo.de>
   7  *  Copyright (C) 2004  Anil S Keshavamurthy <anil.s.keshavamurthy@intel.com>
   8  *                      - Added processor hotplug support
   9  *  Copyright (C) 2005  Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>
  10  *                      - Added support for C3 on SMP
  11  *
  12  * ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
  13  *
  14  *  This program is free software; you can redistribute it and/or modify
  15  *  it under the terms of the GNU General Public License as published by
  16  *  the Free Software Foundation; either version 2 of the License, or (at
  17  *  your option) any later version.
  18  *
  19  *  This program is distributed in the hope that it will be useful, but
  20  *  WITHOUT ANY WARRANTY; without even the implied warranty of
  21  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  22  *  General Public License for more details.
  23  *
  24  *  You should have received a copy of the GNU General Public License along
  25  *  with this program; if not, write to the Free Software Foundation, Inc.,
  26  *  59 Temple Place, Suite 330, Boston, MA 02111-1307 USA.
  27  *
  28  * ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
  29  */
  30
  31 #include <linux/kernel.h>
  32 #include <linux/module.h>
  33 #include <linux/init.h>
  34 #include <linux/cpufreq.h>
  35 #include <linux/proc_fs.h>
  36 #include <linux/seq_file.h>
  37 #include <linux/acpi.h>
  38 #include <linux/dmi.h>
  39 #include <linux/moduleparam.h>
  40 #include <linux/sched.h>        /* need_resched() */
  41
  42 #include <asm/io.h>
  43 #include <asm/uaccess.h>
  44
  45 #include <acpi/acpi_bus.h>
  46 #include <acpi/processor.h>
  47
  48 #define ACPI_PROCESSOR_COMPONENT        0x01000000
  49 #define ACPI_PROCESSOR_CLASS            "processor"
  50 #define ACPI_PROCESSOR_DRIVER_NAME      "ACPI Processor Driver"
  51 #define _COMPONENT              ACPI_PROCESSOR_COMPONENT
  52 ACPI_MODULE_NAME("acpi_processor")
  53 #define ACPI_PROCESSOR_FILE_POWER       "power"
  54 #define US_TO_PM_TIMER_TICKS(t)         ((t * (PM_TIMER_FREQUENCY/1000)) / 1000)
  55 #define C2_OVERHEAD                     4       /* 1us (3.579 ticks per us) */
  56 #define C3_OVERHEAD                     4       /* 1us (3.579 ticks per us) */
  57 static void (*pm_idle_save) (void);
  58 module_param(max_cstate, uint, 0644);
  59
  60 static unsigned int nocst = 0;
  61 module_param(nocst, uint, 0000);
  62
  63 /*
  64  * bm_history -- bit-mask with a bit per jiffy of bus-master activity
  65  * 1000 HZ: 0xFFFFFFFF: 32 jiffies = 32ms
  66  * 800 HZ: 0xFFFFFFFF: 32 jiffies = 40ms
  67  * 100 HZ: 0x0000000F: 4 jiffies = 40ms
  68  * reduce history for more aggressive entry into C3
  69  */
  70 static unsigned int bm_history =
  71     (HZ >= 800 ? 0xFFFFFFFF : ((1U << (HZ / 25)) - 1));
  72 module_param(bm_history, uint, 0644);
  73 /* --------------------------------------------------------------------------
  74                                 Power Management
  75    -------------------------------------------------------------------------- */
  76
  77 /*
  78  * IBM ThinkPad R40e crashes mysteriously when going into C2 or C3.
  79  * For now disable this. Probably a bug somewhere else.
  80  *
  81  * To skip this limit, boot/load with a large max_cstate limit.
  82  */
  83 static int set_max_cstate(struct dmi_system_id *id)
  84 {
  85         if (max_cstate > ACPI_PROCESSOR_MAX_POWER)
  86                 return 0;
  87
  88         printk(KERN_NOTICE PREFIX "%s detected - limiting to C%ld max_cstate."
  89                " Override with \"processor.max_cstate=%d\"\n", id->ident,
  90                (long)id->driver_data, ACPI_PROCESSOR_MAX_POWER + 1);
  91
  92         max_cstate = (long)id->driver_data;
  93
  94         return 0;
  95 }
  96
  97 static struct dmi_system_id __initdata processor_power_dmi_table[] = {
  98         {set_max_cstate, "IBM ThinkPad R40e", {
  99                                                DMI_MATCH(DMI_BIOS_VENDOR,
 100                                                          "IBM"),
 101                                                DMI_MATCH(DMI_BIOS_VERSION,
 102                                                          "1SET60WW")},
 103          (void *)1},
 104         {set_max_cstate, "Medion 41700", {
 105                                           DMI_MATCH(DMI_BIOS_VENDOR,
 106                                                     "Phoenix Technologies LTD"),
 107                                           DMI_MATCH(DMI_BIOS_VERSION,
 108                                                     "R01-A1J")}, (void *)1},
 109         {set_max_cstate, "Clevo 5600D", {
 110                                          DMI_MATCH(DMI_BIOS_VENDOR,
 111                                                    "Phoenix Technologies LTD"),
 112                                          DMI_MATCH(DMI_BIOS_VERSION,
 113                                                    "SHE845M0.86C.0013.D.0302131307")},
 114          (void *)2},
 115         {},
 116 };
 117
 118 static inline u32 ticks_elapsed(u32 t1, u32 t2)
 119 {
 120         if (t2 >= t1)
 121                 return (t2 - t1);
 122         else if (!acpi_fadt.tmr_val_ext)
 123                 return (((0x00FFFFFF - t1) + t2) & 0x00FFFFFF);
 124         else
 125                 return ((0xFFFFFFFF - t1) + t2);
 126 }
 127
 128 static void
 129 acpi_processor_power_activate(struct acpi_processor *pr,
 130                               struct acpi_processor_cx *new)
 131 {
 132         struct acpi_processor_cx *old;
 133
 134         if (!pr || !new)
 135                 return;
 136
 137         old = pr->power.state;
 138
 139         if (old)
 140                 old->promotion.count = 0;
 141         new->demotion.count = 0;
 142
 143         /* Cleanup from old state. */
 144         if (old) {
 145                 switch (old->type) {
 146                 case ACPI_STATE_C3:
 147                         /* Disable bus master reload */
 148                         if (new->type != ACPI_STATE_C3 && pr->flags.bm_check)
 149                                 acpi_set_register(ACPI_BITREG_BUS_MASTER_RLD, 0,
 150                                                   ACPI_MTX_DO_NOT_LOCK);
 151                         break;
 152                 }
 153         }
 154
 155         /* Prepare to use new state. */
 156         switch (new->type) {
 157         case ACPI_STATE_C3:
 158                 /* Enable bus master reload */
 159                 if (old->type != ACPI_STATE_C3 && pr->flags.bm_check)
 160                         acpi_set_register(ACPI_BITREG_BUS_MASTER_RLD, 1,
 161                                           ACPI_MTX_DO_NOT_LOCK);
 162                 break;
 163         }
 164
 165         pr->power.state = new;
 166
 167         return;
 168 }
 169
 170 static void acpi_safe_halt(void)
 171 {
 172         int polling = test_thread_flag(TIF_POLLING_NRFLAG);
 173         if (polling) {
 174                 clear_thread_flag(TIF_POLLING_NRFLAG);
 175                 smp_mb__after_clear_bit();
 176         }
 177         if (!need_resched())
 178                 safe_halt();
 179         if (polling)
 180                 set_thread_flag(TIF_POLLING_NRFLAG);
 181 }
 182
 183 static atomic_t c3_cpu_count;
 184
 185 static void acpi_processor_idle(void)
 186 {
 187         struct acpi_processor *pr = NULL;
 188         struct acpi_processor_cx *cx = NULL;
 189         struct acpi_processor_cx *next_state = NULL;
 190         int sleep_ticks = 0;
 191         u32 t1, t2 = 0;
 192
 193         pr = processors[smp_processor_id()];
 194         if (!pr)
 195                 return;
 196
 197         /*
 198          * Interrupts must be disabled during bus mastering calculations and
 199          * for C2/C3 transitions.
 200          */
 201         local_irq_disable();
 202
 203         /*
 204          * Check whether we truly need to go idle, or should
 205          * reschedule:
 206          */
 207         if (unlikely(need_resched())) {
 208                 local_irq_enable();
 209                 return;
 210         }
 211
 212         cx = pr->power.state;
 213         if (!cx) {
 214                 if (pm_idle_save)
 215                         pm_idle_save();
 216                 else
 217                         acpi_safe_halt();
 218                 return;
 219         }
 220
 221         /*
 222          * Check BM Activity
 223          * -----------------
 224          * Check for bus mastering activity (if required), record, and check
 225          * for demotion.
 226          */
 227         if (pr->flags.bm_check) {
 228                 u32 bm_status = 0;
 229                 unsigned long diff = jiffies - pr->power.bm_check_timestamp;
 230
 231                 if (diff > 32)
 232                         diff = 32;
 233
 234                 while (diff) {
 235                         /* if we didn't get called, assume there was busmaster activity */
 236                         diff--;
 237                         if (diff)
 238                                 pr->power.bm_activity |= 0x1;
 239                         pr->power.bm_activity <<= 1;
 240                 }
 241
 242                 acpi_get_register(ACPI_BITREG_BUS_MASTER_STATUS,
 243                                   &bm_status, ACPI_MTX_DO_NOT_LOCK);
 244                 if (bm_status) {
 245                         pr->power.bm_activity++;
 246                         acpi_set_register(ACPI_BITREG_BUS_MASTER_STATUS,
 247                                           1, ACPI_MTX_DO_NOT_LOCK);
 248                 }
 249                 /*
 250                  * PIIX4 Erratum #18: Note that BM_STS doesn't always reflect
 251                  * the true state of bus mastering activity; forcing us to
 252                  * manually check the BMIDEA bit of each IDE channel.
 253                  */
 254                 else if (errata.piix4.bmisx) {
 255                         if ((inb_p(errata.piix4.bmisx + 0x02) & 0x01)
 256                             || (inb_p(errata.piix4.bmisx + 0x0A) & 0x01))
 257                                 pr->power.bm_activity++;
 258                 }
 259
 260                 pr->power.bm_check_timestamp = jiffies;
 261
 262                 /*
 263                  * Apply bus mastering demotion policy.  Automatically demote
 264                  * to avoid a faulty transition.  Note that the processor
 265                  * won't enter a low-power state during this call (to this
 266                  * funciton) but should upon the next.
 267                  *
 268                  * TBD: A better policy might be to fallback to the demotion
 269                  *      state (use it for this quantum only) istead of
 270                  *      demoting -- and rely on duration as our sole demotion
 271                  *      qualification.  This may, however, introduce DMA
 272                  *      issues (e.g. floppy DMA transfer overrun/underrun).
 273                  */
 274                 if (pr->power.bm_activity & cx->demotion.threshold.bm) {
 275                         local_irq_enable();
 276                         next_state = cx->demotion.state;
 277                         goto end;
 278                 }
 279         }
 280
 281         cx->usage++;
 282
 283 #ifdef CONFIG_HOTPLUG_CPU
 284         /*
 285          * Check for P_LVL2_UP flag before entering C2 and above on
 286          * an SMP system. We do it here instead of doing it at _CST/P_LVL
 287          * detection phase, to work cleanly with logical CPU hotplug.
 288          */
 289         if ((cx->type != ACPI_STATE_C1) && (num_online_cpus() > 1) &&
 290             !pr->flags.has_cst && acpi_fadt.plvl2_up)
 291                 cx->type = ACPI_STATE_C1;
 292 #endif
 293         /*
 294          * Sleep:
 295          * ------
 296          * Invoke the current Cx state to put the processor to sleep.
 297          */
 298         switch (cx->type) {
 299
 300         case ACPI_STATE_C1:
 301                 /*
 302                  * Invoke C1.
 303                  * Use the appropriate idle routine, the one that would
 304                  * be used without acpi C-states.
 305                  */
 306                 if (pm_idle_save)
 307                         pm_idle_save();
 308                 else
 309                         acpi_safe_halt();
 310
 311                 /*
 312                  * TBD: Can't get time duration while in C1, as resumes
 313                  *      go to an ISR rather than here.  Need to instrument
 314                  *      base interrupt handler.
 315                  */
 316                 sleep_ticks = 0xFFFFFFFF;
 317                 break;
 318
 319         case ACPI_STATE_C2:
 320                 /* Get start time (ticks) */
 321                 t1 = inl(acpi_fadt.xpm_tmr_blk.address);
 322                 /* Invoke C2 */
 323                 inb(cx->address);
 324                 /* Dummy op - must do something useless after P_LVL2 read */
 325                 t2 = inl(acpi_fadt.xpm_tmr_blk.address);
 326                 /* Get end time (ticks) */
 327                 t2 = inl(acpi_fadt.xpm_tmr_blk.address);
 328                 /* Re-enable interrupts */
 329                 local_irq_enable();
 330                 /* Compute time (ticks) that we were actually asleep */
 331                 sleep_ticks =
 332                     ticks_elapsed(t1, t2) - cx->latency_ticks - C2_OVERHEAD;
 333                 break;
 334
 335         case ACPI_STATE_C3:
 336
 337                 if (pr->flags.bm_check) {
 338                         if (atomic_inc_return(&c3_cpu_count) ==
 339                             num_online_cpus()) {
 340                                 /*
 341                                  * All CPUs are trying to go to C3
 342                                  * Disable bus master arbitration
 343                                  */
 344                                 acpi_set_register(ACPI_BITREG_ARB_DISABLE, 1,
 345                                                   ACPI_MTX_DO_NOT_LOCK);
 346                         }
 347                 } else {
 348                         /* SMP with no shared cache... Invalidate cache  */
 349                         ACPI_FLUSH_CPU_CACHE();
 350                 }
 351
 352                 /* Get start time (ticks) */
 353                 t1 = inl(acpi_fadt.xpm_tmr_blk.address);
 354                 /* Invoke C3 */
 355                 inb(cx->address);
 356                 /* Dummy op - must do something useless after P_LVL3 read */
 357                 t2 = inl(acpi_fadt.xpm_tmr_blk.address);
 358                 /* Get end time (ticks) */
 359                 t2 = inl(acpi_fadt.xpm_tmr_blk.address);
 360                 if (pr->flags.bm_check) {
 361                         /* Enable bus master arbitration */
 362                         atomic_dec(&c3_cpu_count);
 363                         acpi_set_register(ACPI_BITREG_ARB_DISABLE, 0,
 364                                           ACPI_MTX_DO_NOT_LOCK);
 365                 }
 366
 367                 /* Re-enable interrupts */
 368                 local_irq_enable();
 369                 /* Compute time (ticks) that we were actually asleep */
 370                 sleep_ticks =
 371                     ticks_elapsed(t1, t2) - cx->latency_ticks - C3_OVERHEAD;
 372                 break;
 373
 374         default:
 375                 local_irq_enable();
 376                 return;
 377         }
 378
 379         next_state = pr->power.state;
 380
 381         /*
 382          * Promotion?
 383          * ----------
 384          * Track the number of longs (time asleep is greater than threshold)
 385          * and promote when the count threshold is reached.  Note that bus
 386          * mastering activity may prevent promotions.
 387          * Do not promote above max_cstate.
 388          */
 389         if (cx->promotion.state &&
 390             ((cx->promotion.state - pr->power.states) <= max_cstate)) {
 391                 if (sleep_ticks > cx->promotion.threshold.ticks) {
 392                         cx->promotion.count++;
 393                         cx->demotion.count = 0;
 394                         if (cx->promotion.count >=
 395                             cx->promotion.threshold.count) {
 396                                 if (pr->flags.bm_check) {
 397                                         if (!
 398                                             (pr->power.bm_activity & cx->
 399                                              promotion.threshold.bm)) {
 400                                                 next_state =
 401                                                     cx->promotion.state;
 402                                                 goto end;
 403                                         }
 404                                 } else {
 405                                         next_state = cx->promotion.state;
 406                                         goto end;
 407                                 }
 408                         }
 409                 }
 410         }
 411
 412         /*
 413          * Demotion?
 414          * ---------
 415          * Track the number of shorts (time asleep is less than time threshold)
 416          * and demote when the usage threshold is reached.
 417          */
 418         if (cx->demotion.state) {
 419                 if (sleep_ticks < cx->demotion.threshold.ticks) {
 420                         cx->demotion.count++;
 421                         cx->promotion.count = 0;
 422                         if (cx->demotion.count >= cx->demotion.threshold.count) {
 423                                 next_state = cx->demotion.state;
 424                                 goto end;
 425                         }
 426                 }
 427         }
 428
 429       end:
 430         /*
 431          * Demote if current state exceeds max_cstate
 432          */
 433         if ((pr->power.state - pr->power.states) > max_cstate) {
 434                 if (cx->demotion.state)
 435                         next_state = cx->demotion.state;
 436         }
 437
 438         /*
 439          * New Cx State?
 440          * -------------
 441          * If we're going to start using a new Cx state we must clean up
 442          * from the previous and prepare to use the new.
 443          */
 444         if (next_state != pr->power.state)
 445                 acpi_processor_power_activate(pr, next_state);
 446 }
 447
 448 static int acpi_processor_set_power_policy(struct acpi_processor *pr)
 449 {
 450         unsigned int i;
 451         unsigned int state_is_set = 0;
 452         struct acpi_processor_cx *lower = NULL;
 453         struct acpi_processor_cx *higher = NULL;
 454         struct acpi_processor_cx *cx;
 455
 456         ACPI_FUNCTION_TRACE("acpi_processor_set_power_policy");
 457
 458         if (!pr)
 459                 return_VALUE(-EINVAL);
 460
 461         /*
 462          * This function sets the default Cx state policy (OS idle handler).
 463          * Our scheme is to promote quickly to C2 but more conservatively
 464          * to C3.  We're favoring C2  for its characteristics of low latency
 465          * (quick response), good power savings, and ability to allow bus
 466          * mastering activity.  Note that the Cx state policy is completely
 467          * customizable and can be altered dynamically.
 468          */
 469
 470         /* startup state */
 471         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
 472                 cx = &pr->power.states[i];
 473                 if (!cx->valid)
 474                         continue;
 475
 476                 if (!state_is_set)
 477                         pr->power.state = cx;
 478                 state_is_set++;
 479                 break;
 480         }
 481
 482         if (!state_is_set)
 483                 return_VALUE(-ENODEV);
 484
 485         /* demotion */
 486         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
 487                 cx = &pr->power.states[i];
 488                 if (!cx->valid)
 489                         continue;
 490
 491                 if (lower) {
 492                         cx->demotion.state = lower;
 493                         cx->demotion.threshold.ticks = cx->latency_ticks;
 494                         cx->demotion.threshold.count = 1;
 495                         if (cx->type == ACPI_STATE_C3)
 496                                 cx->demotion.threshold.bm = bm_history;
 497                 }
 498
 499                 lower = cx;
 500         }
 501
 502         /* promotion */
 503         for (i = (ACPI_PROCESSOR_MAX_POWER - 1); i > 0; i--) {
 504                 cx = &pr->power.states[i];
 505                 if (!cx->valid)
 506                         continue;
 507
 508                 if (higher) {
 509                         cx->promotion.state = higher;
 510                         cx->promotion.threshold.ticks = cx->latency_ticks;
 511                         if (cx->type >= ACPI_STATE_C2)
 512                                 cx->promotion.threshold.count = 4;
 513                         else
 514                                 cx->promotion.threshold.count = 10;
 515                         if (higher->type == ACPI_STATE_C3)
 516                                 cx->promotion.threshold.bm = bm_history;
 517                 }
 518
 519                 higher = cx;
 520         }
 521
 522         return_VALUE(0);
 523 }
 524
 525 static int acpi_processor_get_power_info_fadt(struct acpi_processor *pr)
 526 {
 527         ACPI_FUNCTION_TRACE("acpi_processor_get_power_info_fadt");
 528
 529         if (!pr)
 530                 return_VALUE(-EINVAL);
 531
 532         if (!pr->pblk)
 533                 return_VALUE(-ENODEV);
 534
 535         memset(pr->power.states, 0, sizeof(pr->power.states));
 536
 537         /* if info is obtained from pblk/fadt, type equals state */
 538         pr->power.states[ACPI_STATE_C1].type = ACPI_STATE_C1;
 539         pr->power.states[ACPI_STATE_C2].type = ACPI_STATE_C2;
 540         pr->power.states[ACPI_STATE_C3].type = ACPI_STATE_C3;
 541
 542         /* the C0 state only exists as a filler in our array,
 543          * and all processors need to support C1 */
 544         pr->power.states[ACPI_STATE_C0].valid = 1;
 545         pr->power.states[ACPI_STATE_C1].valid = 1;
 546
 547 #ifndef CONFIG_HOTPLUG_CPU
 548         /*
 549          * Check for P_LVL2_UP flag before entering C2 and above on
 550          * an SMP system.
 551          */
 552         if ((num_online_cpus() > 1) && acpi_fadt.plvl2_up)
 553                 return_VALUE(-ENODEV);
 554 #endif
 555
 556         /* determine C2 and C3 address from pblk */
 557         pr->power.states[ACPI_STATE_C2].address = pr->pblk + 4;
 558         pr->power.states[ACPI_STATE_C3].address = pr->pblk + 5;
 559
 560         /* determine latencies from FADT */
 561         pr->power.states[ACPI_STATE_C2].latency = acpi_fadt.plvl2_lat;
 562         pr->power.states[ACPI_STATE_C3].latency = acpi_fadt.plvl3_lat;
 563
 564         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 565                           "lvl2[0x%08x] lvl3[0x%08x]\n",
 566                           pr->power.states[ACPI_STATE_C2].address,
 567                           pr->power.states[ACPI_STATE_C3].address));
 568
 569         return_VALUE(0);
 570 }
 571
 572 static int acpi_processor_get_power_info_default_c1(struct acpi_processor *pr)
 573 {
 574         ACPI_FUNCTION_TRACE("acpi_processor_get_power_info_default_c1");
 575
 576         memset(pr->power.states, 0, sizeof(pr->power.states));
 577
 578         /* if info is obtained from pblk/fadt, type equals state */
 579         pr->power.states[ACPI_STATE_C1].type = ACPI_STATE_C1;
 580         pr->power.states[ACPI_STATE_C2].type = ACPI_STATE_C2;
 581         pr->power.states[ACPI_STATE_C3].type = ACPI_STATE_C3;
 582
 583         /* the C0 state only exists as a filler in our array,
 584          * and all processors need to support C1 */
 585         pr->power.states[ACPI_STATE_C0].valid = 1;
 586         pr->power.states[ACPI_STATE_C1].valid = 1;
 587
 588         return_VALUE(0);
 589 }
 590
 591 static int acpi_processor_get_power_info_cst(struct acpi_processor *pr)
 592 {
 593         acpi_status status = 0;
 594         acpi_integer count;
 595         int i;
 596         struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
 597         union acpi_object *cst;
 598
 599         ACPI_FUNCTION_TRACE("acpi_processor_get_power_info_cst");
 600
 601         if (nocst)
 602                 return_VALUE(-ENODEV);
 603
 604         pr->power.count = 0;
 605         for (i = 0; i < ACPI_PROCESSOR_MAX_POWER; i++)
 606                 memset(&(pr->power.states[i]), 0,
 607                        sizeof(struct acpi_processor_cx));
 608
 609         status = acpi_evaluate_object(pr->handle, "_CST", NULL, &buffer);
 610         if (ACPI_FAILURE(status)) {
 611                 ACPI_DEBUG_PRINT((ACPI_DB_INFO, "No _CST, giving up\n"));
 612                 return_VALUE(-ENODEV);
 613         }
 614
 615         cst = (union acpi_object *)buffer.pointer;
 616
 617         /* There must be at least 2 elements */
 618         if (!cst || (cst->type != ACPI_TYPE_PACKAGE) || cst->package.count < 2) {
 619                 ACPI_DEBUG_PRINT((ACPI_DB_ERROR,
 620                                   "not enough elements in _CST\n"));
 621                 status = -EFAULT;
 622                 goto end;
 623         }
 624
 625         count = cst->package.elements[0].integer.value;
 626
 627         /* Validate number of power states. */
 628         if (count < 1 || count != cst->package.count - 1) {
 629                 ACPI_DEBUG_PRINT((ACPI_DB_ERROR,
 630                                   "count given by _CST is not valid\n"));
 631                 status = -EFAULT;
 632                 goto end;
 633         }
 634
 635         /* We support up to ACPI_PROCESSOR_MAX_POWER. */
 636         if (count > ACPI_PROCESSOR_MAX_POWER) {
 637                 printk(KERN_WARNING
 638                        "Limiting number of power states to max (%d)\n",
 639                        ACPI_PROCESSOR_MAX_POWER);
 640                 printk(KERN_WARNING
 641                        "Please increase ACPI_PROCESSOR_MAX_POWER if needed.\n");
 642                 count = ACPI_PROCESSOR_MAX_POWER;
 643         }
 644
 645         /* Tell driver that at least _CST is supported. */
 646         pr->flags.has_cst = 1;
 647
 648         for (i = 1; i <= count; i++) {
 649                 union acpi_object *element;
 650                 union acpi_object *obj;
 651                 struct acpi_power_register *reg;
 652                 struct acpi_processor_cx cx;
 653
 654                 memset(&cx, 0, sizeof(cx));
 655
 656                 element = (union acpi_object *)&(cst->package.elements[i]);
 657                 if (element->type != ACPI_TYPE_PACKAGE)
 658                         continue;
 659
 660                 if (element->package.count != 4)
 661                         continue;
 662
 663                 obj = (union acpi_object *)&(element->package.elements[0]);
 664
 665                 if (obj->type != ACPI_TYPE_BUFFER)
 666                         continue;
 667
 668                 reg = (struct acpi_power_register *)obj->buffer.pointer;
 669
 670                 if (reg->space_id != ACPI_ADR_SPACE_SYSTEM_IO &&
 671                     (reg->space_id != ACPI_ADR_SPACE_FIXED_HARDWARE))
 672                         continue;
 673
 674                 cx.address = (reg->space_id == ACPI_ADR_SPACE_FIXED_HARDWARE) ?
 675                     0 : reg->address;
 676
 677                 /* There should be an easy way to extract an integer... */
 678                 obj = (union acpi_object *)&(element->package.elements[1]);
 679                 if (obj->type != ACPI_TYPE_INTEGER)
 680                         continue;
 681
 682                 cx.type = obj->integer.value;
 683
 684                 if ((cx.type != ACPI_STATE_C1) &&
 685                     (reg->space_id != ACPI_ADR_SPACE_SYSTEM_IO))
 686                         continue;
 687
 688                 if ((cx.type < ACPI_STATE_C1) || (cx.type > ACPI_STATE_C3))
 689                         continue;
 690
 691                 obj = (union acpi_object *)&(element->package.elements[2]);
 692                 if (obj->type != ACPI_TYPE_INTEGER)
 693                         continue;
 694
 695                 cx.latency = obj->integer.value;
 696
 697                 obj = (union acpi_object *)&(element->package.elements[3]);
 698                 if (obj->type != ACPI_TYPE_INTEGER)
 699                         continue;
 700
 701                 cx.power = obj->integer.value;
 702
 703                 (pr->power.count)++;
 704                 memcpy(&(pr->power.states[pr->power.count]), &cx, sizeof(cx));
 705         }
 706
 707         ACPI_DEBUG_PRINT((ACPI_DB_INFO, "Found %d power states\n",
 708                           pr->power.count));
 709
 710         /* Validate number of power states discovered */
 711         if (pr->power.count < 2)
 712                 status = -EFAULT;
 713
 714       end:
 715         acpi_os_free(buffer.pointer);
 716
 717         return_VALUE(status);
 718 }
 719
 720 static void acpi_processor_power_verify_c2(struct acpi_processor_cx *cx)
 721 {
 722         ACPI_FUNCTION_TRACE("acpi_processor_get_power_verify_c2");
 723
 724         if (!cx->address)
 725                 return_VOID;
 726
 727         /*
 728          * C2 latency must be less than or equal to 100
 729          * microseconds.
 730          */
 731         else if (cx->latency > ACPI_PROCESSOR_MAX_C2_LATENCY) {
 732                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 733                                   "latency too large [%d]\n", cx->latency));
 734                 return_VOID;
 735         }
 736
 737         /*
 738          * Otherwise we've met all of our C2 requirements.
 739          * Normalize the C2 latency to expidite policy
 740          */
 741         cx->valid = 1;
 742         cx->latency_ticks = US_TO_PM_TIMER_TICKS(cx->latency);
 743
 744         return_VOID;
 745 }
 746
 747 static void acpi_processor_power_verify_c3(struct acpi_processor *pr,
 748                                            struct acpi_processor_cx *cx)
 749 {
 750         static int bm_check_flag;
 751
 752         ACPI_FUNCTION_TRACE("acpi_processor_get_power_verify_c3");
 753
 754         if (!cx->address)
 755                 return_VOID;
 756
 757         /*
 758          * C3 latency must be less than or equal to 1000
 759          * microseconds.
 760          */
 761         else if (cx->latency > ACPI_PROCESSOR_MAX_C3_LATENCY) {
 762                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 763                                   "latency too large [%d]\n", cx->latency));
 764                 return_VOID;
 765         }
 766
 767         /*
 768          * PIIX4 Erratum #18: We don't support C3 when Type-F (fast)
 769          * DMA transfers are used by any ISA device to avoid livelock.
 770          * Note that we could disable Type-F DMA (as recommended by
 771          * the erratum), but this is known to disrupt certain ISA
 772          * devices thus we take the conservative approach.
 773          */
 774         else if (errata.piix4.fdma) {
 775                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 776                                   "C3 not supported on PIIX4 with Type-F DMA\n"));
 777                 return_VOID;
 778         }
 779
 780         /* All the logic here assumes flags.bm_check is same across all CPUs */
 781         if (!bm_check_flag) {
 782                 /* Determine whether bm_check is needed based on CPU  */
 783                 acpi_processor_power_init_bm_check(&(pr->flags), pr->id);
 784                 bm_check_flag = pr->flags.bm_check;
 785         } else {
 786                 pr->flags.bm_check = bm_check_flag;
 787         }
 788
 789         if (pr->flags.bm_check) {
 790                 /* bus mastering control is necessary */
 791                 if (!pr->flags.bm_control) {
 792                         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 793                                           "C3 support requires bus mastering control\n"));
 794                         return_VOID;
 795                 }
 796         } else {
 797                 /*
 798                  * WBINVD should be set in fadt, for C3 state to be
 799                  * supported on when bm_check is not required.
 800                  */
 801                 if (acpi_fadt.wb_invd != 1) {
 802                         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 803                                           "Cache invalidation should work properly"
 804                                           " for C3 to be enabled on SMP systems\n"));
 805                         return_VOID;
 806                 }
 807                 acpi_set_register(ACPI_BITREG_BUS_MASTER_RLD,
 808                                   0, ACPI_MTX_DO_NOT_LOCK);
 809         }
 810
 811         /*
 812          * Otherwise we've met all of our C3 requirements.
 813          * Normalize the C3 latency to expidite policy.  Enable
 814          * checking of bus mastering status (bm_check) so we can
 815          * use this in our C3 policy
 816          */
 817         cx->valid = 1;
 818         cx->latency_ticks = US_TO_PM_TIMER_TICKS(cx->latency);
 819
 820         return_VOID;
 821 }
 822
 823 static int acpi_processor_power_verify(struct acpi_processor *pr)
 824 {
 825         unsigned int i;
 826         unsigned int working = 0;
 827
 828         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
 829                 struct acpi_processor_cx *cx = &pr->power.states[i];
 830
 831                 switch (cx->type) {
 832                 case ACPI_STATE_C1:
 833                         cx->valid = 1;
 834                         break;
 835
 836                 case ACPI_STATE_C2:
 837                         acpi_processor_power_verify_c2(cx);
 838                         break;
 839
 840                 case ACPI_STATE_C3:
 841                         acpi_processor_power_verify_c3(pr, cx);
 842                         break;
 843                 }
 844
 845                 if (cx->valid)
 846                         working++;
 847         }
 848
 849         return (working);
 850 }
 851
 852 static int acpi_processor_get_power_info(struct acpi_processor *pr)
 853 {
 854         unsigned int i;
 855         int result;
 856
 857         ACPI_FUNCTION_TRACE("acpi_processor_get_power_info");
 858
 859         /* NOTE: the idle thread may not be running while calling
 860          * this function */
 861
 862         result = acpi_processor_get_power_info_cst(pr);
 863         if (result == -ENODEV)
 864                 result = acpi_processor_get_power_info_fadt(pr);
 865
 866         if ((result) || (acpi_processor_power_verify(pr) < 2))
 867                 result = acpi_processor_get_power_info_default_c1(pr);
 868
 869         /*
 870          * Set Default Policy
 871          * ------------------
 872          * Now that we know which states are supported, set the default
 873          * policy.  Note that this policy can be changed dynamically
 874          * (e.g. encourage deeper sleeps to conserve battery life when
 875          * not on AC).
 876          */
 877         result = acpi_processor_set_power_policy(pr);
 878         if (result)
 879                 return_VALUE(result);
 880
 881         /*
 882          * if one state of type C2 or C3 is available, mark this
 883          * CPU as being "idle manageable"
 884          */
 885         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
 886                 if (pr->power.states[i].valid) {
 887                         pr->power.count = i;
 888                         if (pr->power.states[i].type >= ACPI_STATE_C2)
 889                                 pr->flags.power = 1;
 890                 }
 891         }
 892
 893         return_VALUE(0);
 894 }
 895
 896 int acpi_processor_cst_has_changed(struct acpi_processor *pr)
 897 {
 898         int result = 0;
 899
 900         ACPI_FUNCTION_TRACE("acpi_processor_cst_has_changed");
 901
 902         if (!pr)
 903                 return_VALUE(-EINVAL);
 904
 905         if (nocst) {
 906                 return_VALUE(-ENODEV);
 907         }
 908
 909         if (!pr->flags.power_setup_done)
 910                 return_VALUE(-ENODEV);
 911
 912         /* Fall back to the default idle loop */
 913         pm_idle = pm_idle_save;
 914         synchronize_sched();    /* Relies on interrupts forcing exit from idle. */
 915
 916         pr->flags.power = 0;
 917         result = acpi_processor_get_power_info(pr);
 918         if ((pr->flags.power == 1) && (pr->flags.power_setup_done))
 919                 pm_idle = acpi_processor_idle;
 920
 921         return_VALUE(result);
 922 }
 923
 924 /* proc interface */
 925
 926 static int acpi_processor_power_seq_show(struct seq_file *seq, void *offset)
 927 {
 928         struct acpi_processor *pr = (struct acpi_processor *)seq->private;
 929         unsigned int i;
 930
 931         ACPI_FUNCTION_TRACE("acpi_processor_power_seq_show");
 932
 933         if (!pr)
 934                 goto end;
 935
 936         seq_printf(seq, "active state:            C%zd\n"
 937                    "max_cstate:              C%d\n"
 938                    "bus master activity:     %08x\n",
 939                    pr->power.state ? pr->power.state - pr->power.states : 0,
 940                    max_cstate, (unsigned)pr->power.bm_activity);
 941
 942         seq_puts(seq, "states:\n");
 943
 944         for (i = 1; i <= pr->power.count; i++) {
 945                 seq_printf(seq, "   %cC%d:                  ",
 946                            (&pr->power.states[i] ==
 947                             pr->power.state ? '*' : ' '), i);
 948
 949                 if (!pr->power.states[i].valid) {
 950                         seq_puts(seq, "<not supported>\n");
 951                         continue;
 952                 }
 953
 954                 switch (pr->power.states[i].type) {
 955                 case ACPI_STATE_C1:
 956                         seq_printf(seq, "type[C1] ");
 957                         break;
 958                 case ACPI_STATE_C2:
 959                         seq_printf(seq, "type[C2] ");
 960                         break;
 961                 case ACPI_STATE_C3:
 962                         seq_printf(seq, "type[C3] ");
 963                         break;
 964                 default:
 965                         seq_printf(seq, "type[--] ");
 966                         break;
 967                 }
 968
 969                 if (pr->power.states[i].promotion.state)
 970                         seq_printf(seq, "promotion[C%zd] ",
 971                                    (pr->power.states[i].promotion.state -
 972                                     pr->power.states));
 973                 else
 974                         seq_puts(seq, "promotion[--] ");
 975
 976                 if (pr->power.states[i].demotion.state)
 977                         seq_printf(seq, "demotion[C%zd] ",
 978                                    (pr->power.states[i].demotion.state -
 979                                     pr->power.states));
 980                 else
 981                         seq_puts(seq, "demotion[--] ");
 982
 983                 seq_printf(seq, "latency[%03d] usage[%08d]\n",
 984                            pr->power.states[i].latency,
 985                            pr->power.states[i].usage);
 986         }
 987
 988       end:
 989         return_VALUE(0);
 990 }
 991
 992 static int acpi_processor_power_open_fs(struct inode *inode, struct file *file)
 993 {
 994         return single_open(file, acpi_processor_power_seq_show,
 995                            PDE(inode)->data);
 996 }
 997
 998 static struct file_operations acpi_processor_power_fops = {
 999         .open = acpi_processor_power_open_fs,
1000         .read = seq_read,
1001         .llseek = seq_lseek,
1002         .release = single_release,
1003 };
1004
1005 int acpi_processor_power_init(struct acpi_processor *pr,
1006                               struct acpi_device *device)
1007 {
1008         acpi_status status = 0;
1009         static int first_run = 0;
1010         struct proc_dir_entry *entry = NULL;
1011         unsigned int i;
1012
1013         ACPI_FUNCTION_TRACE("acpi_processor_power_init");
1014
1015         if (!first_run) {
1016                 dmi_check_system(processor_power_dmi_table);
1017                 if (max_cstate < ACPI_C_STATES_MAX)
1018                         printk(KERN_NOTICE
1019                                "ACPI: processor limited to max C-state %d\n",
1020                                max_cstate);
1021                 first_run++;
1022         }
1023
1024         if (!pr)
1025                 return_VALUE(-EINVAL);
1026
1027         if (acpi_fadt.cst_cnt && !nocst) {
1028                 status =
1029                     acpi_os_write_port(acpi_fadt.smi_cmd, acpi_fadt.cst_cnt, 8);
1030                 if (ACPI_FAILURE(status)) {
1031                         ACPI_DEBUG_PRINT((ACPI_DB_ERROR,
1032                                           "Notifying BIOS of _CST ability failed\n"));
1033                 }
1034         }
1035
1036         acpi_processor_power_init_pdc(&(pr->power), pr->id);
1037         acpi_processor_set_pdc(pr, pr->power.pdc);
1038         acpi_processor_get_power_info(pr);
1039
1040         /*
1041          * Install the idle handler if processor power management is supported.
1042          * Note that we use previously set idle handler will be used on
1043          * platforms that only support C1.
1044          */
1045         if ((pr->flags.power) && (!boot_option_idle_override)) {
1046                 printk(KERN_INFO PREFIX "CPU%d (power states:", pr->id);
1047                 for (i = 1; i <= pr->power.count; i++)
1048                         if (pr->power.states[i].valid)
1049                                 printk(" C%d[C%d]", i,
1050                                        pr->power.states[i].type);
1051                 printk(")\n");
1052
1053                 if (pr->id == 0) {
1054                         pm_idle_save = pm_idle;
1055                         pm_idle = acpi_processor_idle;
1056                 }
1057         }
1058
1059         /* 'power' [R] */
1060         entry = create_proc_entry(ACPI_PROCESSOR_FILE_POWER,
1061                                   S_IRUGO, acpi_device_dir(device));
1062         if (!entry)
1063                 ACPI_DEBUG_PRINT((ACPI_DB_ERROR,
1064                                   "Unable to create '%s' fs entry\n",
1065                                   ACPI_PROCESSOR_FILE_POWER));
1066         else {
1067                 entry->proc_fops = &acpi_processor_power_fops;
1068                 entry->data = acpi_driver_data(device);
1069                 entry->owner = THIS_MODULE;
1070         }
1071
1072         pr->flags.power_setup_done = 1;
1073
1074         return_VALUE(0);
1075 }
1076
1077 int acpi_processor_power_exit(struct acpi_processor *pr,
1078                               struct acpi_device *device)
1079 {
1080         ACPI_FUNCTION_TRACE("acpi_processor_power_exit");
1081
1082         pr->flags.power_setup_done = 0;
1083
1084         if (acpi_device_dir(device))
1085                 remove_proc_entry(ACPI_PROCESSOR_FILE_POWER,
1086                                   acpi_device_dir(device));
1087
1088         /* Unregister the idle handler when processor #0 is removed. */
1089         if (pr->id == 0) {
1090                 pm_idle = pm_idle_save;
1091
1092                 /*
1093                  * We are about to unload the current idle thread pm callback
1094                  * (pm_idle), Wait for all processors to update cached/local
1095                  * copies of pm_idle before proceeding.
1096                  */
1097                 cpu_idle_wait();
1098         }
1099
1100         return_VALUE(0);
1101 }