git.oblomov.eu Git - linux-2.6/blob - 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         /*
 284          * Sleep:
 285          * ------
 286          * Invoke the current Cx state to put the processor to sleep.
 287          */
 288         switch (cx->type) {
 289
 290         case ACPI_STATE_C1:
 291                 /*
 292                  * Invoke C1.
 293                  * Use the appropriate idle routine, the one that would
 294                  * be used without acpi C-states.
 295                  */
 296                 if (pm_idle_save)
 297                         pm_idle_save();
 298                 else
 299                         acpi_safe_halt();
 300
 301                 /*
 302                  * TBD: Can't get time duration while in C1, as resumes
 303                  *      go to an ISR rather than here.  Need to instrument
 304                  *      base interrupt handler.
 305                  */
 306                 sleep_ticks = 0xFFFFFFFF;
 307                 break;
 308
 309         case ACPI_STATE_C2:
 310                 /* Get start time (ticks) */
 311                 t1 = inl(acpi_fadt.xpm_tmr_blk.address);
 312                 /* Invoke C2 */
 313                 inb(cx->address);
 314                 /* Dummy op - must do something useless after P_LVL2 read */
 315                 t2 = inl(acpi_fadt.xpm_tmr_blk.address);
 316                 /* Get end time (ticks) */
 317                 t2 = inl(acpi_fadt.xpm_tmr_blk.address);
 318                 /* Re-enable interrupts */
 319                 local_irq_enable();
 320                 /* Compute time (ticks) that we were actually asleep */
 321                 sleep_ticks =
 322                     ticks_elapsed(t1, t2) - cx->latency_ticks - C2_OVERHEAD;
 323                 break;
 324
 325         case ACPI_STATE_C3:
 326
 327                 if (pr->flags.bm_check) {
 328                         if (atomic_inc_return(&c3_cpu_count) ==
 329                             num_online_cpus()) {
 330                                 /*
 331                                  * All CPUs are trying to go to C3
 332                                  * Disable bus master arbitration
 333                                  */
 334                                 acpi_set_register(ACPI_BITREG_ARB_DISABLE, 1,
 335                                                   ACPI_MTX_DO_NOT_LOCK);
 336                         }
 337                 } else {
 338                         /* SMP with no shared cache... Invalidate cache  */
 339                         ACPI_FLUSH_CPU_CACHE();
 340                 }
 341
 342                 /* Get start time (ticks) */
 343                 t1 = inl(acpi_fadt.xpm_tmr_blk.address);
 344                 /* Invoke C3 */
 345                 inb(cx->address);
 346                 /* Dummy op - must do something useless after P_LVL3 read */
 347                 t2 = inl(acpi_fadt.xpm_tmr_blk.address);
 348                 /* Get end time (ticks) */
 349                 t2 = inl(acpi_fadt.xpm_tmr_blk.address);
 350                 if (pr->flags.bm_check) {
 351                         /* Enable bus master arbitration */
 352                         atomic_dec(&c3_cpu_count);
 353                         acpi_set_register(ACPI_BITREG_ARB_DISABLE, 0,
 354                                           ACPI_MTX_DO_NOT_LOCK);
 355                 }
 356
 357                 /* Re-enable interrupts */
 358                 local_irq_enable();
 359                 /* Compute time (ticks) that we were actually asleep */
 360                 sleep_ticks =
 361                     ticks_elapsed(t1, t2) - cx->latency_ticks - C3_OVERHEAD;
 362                 break;
 363
 364         default:
 365                 local_irq_enable();
 366                 return;
 367         }
 368
 369         next_state = pr->power.state;
 370
 371         /*
 372          * Promotion?
 373          * ----------
 374          * Track the number of longs (time asleep is greater than threshold)
 375          * and promote when the count threshold is reached.  Note that bus
 376          * mastering activity may prevent promotions.
 377          * Do not promote above max_cstate.
 378          */
 379         if (cx->promotion.state &&
 380             ((cx->promotion.state - pr->power.states) <= max_cstate)) {
 381                 if (sleep_ticks > cx->promotion.threshold.ticks) {
 382                         cx->promotion.count++;
 383                         cx->demotion.count = 0;
 384                         if (cx->promotion.count >=
 385                             cx->promotion.threshold.count) {
 386                                 if (pr->flags.bm_check) {
 387                                         if (!
 388                                             (pr->power.bm_activity & cx->
 389                                              promotion.threshold.bm)) {
 390                                                 next_state =
 391                                                     cx->promotion.state;
 392                                                 goto end;
 393                                         }
 394                                 } else {
 395                                         next_state = cx->promotion.state;
 396                                         goto end;
 397                                 }
 398                         }
 399                 }
 400         }
 401
 402         /*
 403          * Demotion?
 404          * ---------
 405          * Track the number of shorts (time asleep is less than time threshold)
 406          * and demote when the usage threshold is reached.
 407          */
 408         if (cx->demotion.state) {
 409                 if (sleep_ticks < cx->demotion.threshold.ticks) {
 410                         cx->demotion.count++;
 411                         cx->promotion.count = 0;
 412                         if (cx->demotion.count >= cx->demotion.threshold.count) {
 413                                 next_state = cx->demotion.state;
 414                                 goto end;
 415                         }
 416                 }
 417         }
 418
 419       end:
 420         /*
 421          * Demote if current state exceeds max_cstate
 422          */
 423         if ((pr->power.state - pr->power.states) > max_cstate) {
 424                 if (cx->demotion.state)
 425                         next_state = cx->demotion.state;
 426         }
 427
 428         /*
 429          * New Cx State?
 430          * -------------
 431          * If we're going to start using a new Cx state we must clean up
 432          * from the previous and prepare to use the new.
 433          */
 434         if (next_state != pr->power.state)
 435                 acpi_processor_power_activate(pr, next_state);
 436 }
 437
 438 static int acpi_processor_set_power_policy(struct acpi_processor *pr)
 439 {
 440         unsigned int i;
 441         unsigned int state_is_set = 0;
 442         struct acpi_processor_cx *lower = NULL;
 443         struct acpi_processor_cx *higher = NULL;
 444         struct acpi_processor_cx *cx;
 445
 446         ACPI_FUNCTION_TRACE("acpi_processor_set_power_policy");
 447
 448         if (!pr)
 449                 return_VALUE(-EINVAL);
 450
 451         /*
 452          * This function sets the default Cx state policy (OS idle handler).
 453          * Our scheme is to promote quickly to C2 but more conservatively
 454          * to C3.  We're favoring C2  for its characteristics of low latency
 455          * (quick response), good power savings, and ability to allow bus
 456          * mastering activity.  Note that the Cx state policy is completely
 457          * customizable and can be altered dynamically.
 458          */
 459
 460         /* startup state */
 461         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
 462                 cx = &pr->power.states[i];
 463                 if (!cx->valid)
 464                         continue;
 465
 466                 if (!state_is_set)
 467                         pr->power.state = cx;
 468                 state_is_set++;
 469                 break;
 470         }
 471
 472         if (!state_is_set)
 473                 return_VALUE(-ENODEV);
 474
 475         /* demotion */
 476         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
 477                 cx = &pr->power.states[i];
 478                 if (!cx->valid)
 479                         continue;
 480
 481                 if (lower) {
 482                         cx->demotion.state = lower;
 483                         cx->demotion.threshold.ticks = cx->latency_ticks;
 484                         cx->demotion.threshold.count = 1;
 485                         if (cx->type == ACPI_STATE_C3)
 486                                 cx->demotion.threshold.bm = bm_history;
 487                 }
 488
 489                 lower = cx;
 490         }
 491
 492         /* promotion */
 493         for (i = (ACPI_PROCESSOR_MAX_POWER - 1); i > 0; i--) {
 494                 cx = &pr->power.states[i];
 495                 if (!cx->valid)
 496                         continue;
 497
 498                 if (higher) {
 499                         cx->promotion.state = higher;
 500                         cx->promotion.threshold.ticks = cx->latency_ticks;
 501                         if (cx->type >= ACPI_STATE_C2)
 502                                 cx->promotion.threshold.count = 4;
 503                         else
 504                                 cx->promotion.threshold.count = 10;
 505                         if (higher->type == ACPI_STATE_C3)
 506                                 cx->promotion.threshold.bm = bm_history;
 507                 }
 508
 509                 higher = cx;
 510         }
 511
 512         return_VALUE(0);
 513 }
 514
 515 static int acpi_processor_get_power_info_fadt(struct acpi_processor *pr)
 516 {
 517         int i;
 518
 519         ACPI_FUNCTION_TRACE("acpi_processor_get_power_info_fadt");
 520
 521         if (!pr)
 522                 return_VALUE(-EINVAL);
 523
 524         if (!pr->pblk)
 525                 return_VALUE(-ENODEV);
 526
 527         for (i = 0; i < ACPI_PROCESSOR_MAX_POWER; i++)
 528                 memset(pr->power.states, 0, sizeof(struct acpi_processor_cx));
 529
 530         /* if info is obtained from pblk/fadt, type equals state */
 531         pr->power.states[ACPI_STATE_C1].type = ACPI_STATE_C1;
 532         pr->power.states[ACPI_STATE_C2].type = ACPI_STATE_C2;
 533         pr->power.states[ACPI_STATE_C3].type = ACPI_STATE_C3;
 534
 535         /* the C0 state only exists as a filler in our array,
 536          * and all processors need to support C1 */
 537         pr->power.states[ACPI_STATE_C0].valid = 1;
 538         pr->power.states[ACPI_STATE_C1].valid = 1;
 539
 540         /* determine C2 and C3 address from pblk */
 541         pr->power.states[ACPI_STATE_C2].address = pr->pblk + 4;
 542         pr->power.states[ACPI_STATE_C3].address = pr->pblk + 5;
 543
 544         /* determine latencies from FADT */
 545         pr->power.states[ACPI_STATE_C2].latency = acpi_fadt.plvl2_lat;
 546         pr->power.states[ACPI_STATE_C3].latency = acpi_fadt.plvl3_lat;
 547
 548         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 549                           "lvl2[0x%08x] lvl3[0x%08x]\n",
 550                           pr->power.states[ACPI_STATE_C2].address,
 551                           pr->power.states[ACPI_STATE_C3].address));
 552
 553         return_VALUE(0);
 554 }
 555
 556 static int acpi_processor_get_power_info_default_c1(struct acpi_processor *pr)
 557 {
 558         int i;
 559
 560         ACPI_FUNCTION_TRACE("acpi_processor_get_power_info_default_c1");
 561
 562         for (i = 0; i < ACPI_PROCESSOR_MAX_POWER; i++)
 563                 memset(&(pr->power.states[i]), 0,
 564                        sizeof(struct acpi_processor_cx));
 565
 566         /* if info is obtained from pblk/fadt, type equals state */
 567         pr->power.states[ACPI_STATE_C1].type = ACPI_STATE_C1;
 568         pr->power.states[ACPI_STATE_C2].type = ACPI_STATE_C2;
 569         pr->power.states[ACPI_STATE_C3].type = ACPI_STATE_C3;
 570
 571         /* the C0 state only exists as a filler in our array,
 572          * and all processors need to support C1 */
 573         pr->power.states[ACPI_STATE_C0].valid = 1;
 574         pr->power.states[ACPI_STATE_C1].valid = 1;
 575
 576         return_VALUE(0);
 577 }
 578
 579 static int acpi_processor_get_power_info_cst(struct acpi_processor *pr)
 580 {
 581         acpi_status status = 0;
 582         acpi_integer count;
 583         int i;
 584         struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
 585         union acpi_object *cst;
 586
 587         ACPI_FUNCTION_TRACE("acpi_processor_get_power_info_cst");
 588
 589         if (nocst)
 590                 return_VALUE(-ENODEV);
 591
 592         pr->power.count = 0;
 593         for (i = 0; i < ACPI_PROCESSOR_MAX_POWER; i++)
 594                 memset(&(pr->power.states[i]), 0,
 595                        sizeof(struct acpi_processor_cx));
 596
 597         status = acpi_evaluate_object(pr->handle, "_CST", NULL, &buffer);
 598         if (ACPI_FAILURE(status)) {
 599                 ACPI_DEBUG_PRINT((ACPI_DB_INFO, "No _CST, giving up\n"));
 600                 return_VALUE(-ENODEV);
 601         }
 602
 603         cst = (union acpi_object *)buffer.pointer;
 604
 605         /* There must be at least 2 elements */
 606         if (!cst || (cst->type != ACPI_TYPE_PACKAGE) || cst->package.count < 2) {
 607                 ACPI_DEBUG_PRINT((ACPI_DB_ERROR,
 608                                   "not enough elements in _CST\n"));
 609                 status = -EFAULT;
 610                 goto end;
 611         }
 612
 613         count = cst->package.elements[0].integer.value;
 614
 615         /* Validate number of power states. */
 616         if (count < 1 || count != cst->package.count - 1) {
 617                 ACPI_DEBUG_PRINT((ACPI_DB_ERROR,
 618                                   "count given by _CST is not valid\n"));
 619                 status = -EFAULT;
 620                 goto end;
 621         }
 622
 623         /* We support up to ACPI_PROCESSOR_MAX_POWER. */
 624         if (count > ACPI_PROCESSOR_MAX_POWER) {
 625                 printk(KERN_WARNING
 626                        "Limiting number of power states to max (%d)\n",
 627                        ACPI_PROCESSOR_MAX_POWER);
 628                 printk(KERN_WARNING
 629                        "Please increase ACPI_PROCESSOR_MAX_POWER if needed.\n");
 630                 count = ACPI_PROCESSOR_MAX_POWER;
 631         }
 632
 633         /* Tell driver that at least _CST is supported. */
 634         pr->flags.has_cst = 1;
 635
 636         for (i = 1; i <= count; i++) {
 637                 union acpi_object *element;
 638                 union acpi_object *obj;
 639                 struct acpi_power_register *reg;
 640                 struct acpi_processor_cx cx;
 641
 642                 memset(&cx, 0, sizeof(cx));
 643
 644                 element = (union acpi_object *)&(cst->package.elements[i]);
 645                 if (element->type != ACPI_TYPE_PACKAGE)
 646                         continue;
 647
 648                 if (element->package.count != 4)
 649                         continue;
 650
 651                 obj = (union acpi_object *)&(element->package.elements[0]);
 652
 653                 if (obj->type != ACPI_TYPE_BUFFER)
 654                         continue;
 655
 656                 reg = (struct acpi_power_register *)obj->buffer.pointer;
 657
 658                 if (reg->space_id != ACPI_ADR_SPACE_SYSTEM_IO &&
 659                     (reg->space_id != ACPI_ADR_SPACE_FIXED_HARDWARE))
 660                         continue;
 661
 662                 cx.address = (reg->space_id == ACPI_ADR_SPACE_FIXED_HARDWARE) ?
 663                     0 : reg->address;
 664
 665                 /* There should be an easy way to extract an integer... */
 666                 obj = (union acpi_object *)&(element->package.elements[1]);
 667                 if (obj->type != ACPI_TYPE_INTEGER)
 668                         continue;
 669
 670                 cx.type = obj->integer.value;
 671
 672                 if ((cx.type != ACPI_STATE_C1) &&
 673                     (reg->space_id != ACPI_ADR_SPACE_SYSTEM_IO))
 674                         continue;
 675
 676                 if ((cx.type < ACPI_STATE_C1) || (cx.type > ACPI_STATE_C3))
 677                         continue;
 678
 679                 obj = (union acpi_object *)&(element->package.elements[2]);
 680                 if (obj->type != ACPI_TYPE_INTEGER)
 681                         continue;
 682
 683                 cx.latency = obj->integer.value;
 684
 685                 obj = (union acpi_object *)&(element->package.elements[3]);
 686                 if (obj->type != ACPI_TYPE_INTEGER)
 687                         continue;
 688
 689                 cx.power = obj->integer.value;
 690
 691                 (pr->power.count)++;
 692                 memcpy(&(pr->power.states[pr->power.count]), &cx, sizeof(cx));
 693         }
 694
 695         ACPI_DEBUG_PRINT((ACPI_DB_INFO, "Found %d power states\n",
 696                           pr->power.count));
 697
 698         /* Validate number of power states discovered */
 699         if (pr->power.count < 2)
 700                 status = -ENODEV;
 701
 702       end:
 703         acpi_os_free(buffer.pointer);
 704
 705         return_VALUE(status);
 706 }
 707
 708 static void acpi_processor_power_verify_c2(struct acpi_processor_cx *cx)
 709 {
 710         ACPI_FUNCTION_TRACE("acpi_processor_get_power_verify_c2");
 711
 712         if (!cx->address)
 713                 return_VOID;
 714
 715         /*
 716          * C2 latency must be less than or equal to 100
 717          * microseconds.
 718          */
 719         else if (cx->latency > ACPI_PROCESSOR_MAX_C2_LATENCY) {
 720                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 721                                   "latency too large [%d]\n", cx->latency));
 722                 return_VOID;
 723         }
 724
 725         /*
 726          * Otherwise we've met all of our C2 requirements.
 727          * Normalize the C2 latency to expidite policy
 728          */
 729         cx->valid = 1;
 730         cx->latency_ticks = US_TO_PM_TIMER_TICKS(cx->latency);
 731
 732         return_VOID;
 733 }
 734
 735 static void acpi_processor_power_verify_c3(struct acpi_processor *pr,
 736                                            struct acpi_processor_cx *cx)
 737 {
 738         static int bm_check_flag;
 739
 740         ACPI_FUNCTION_TRACE("acpi_processor_get_power_verify_c3");
 741
 742         if (!cx->address)
 743                 return_VOID;
 744
 745         /*
 746          * C3 latency must be less than or equal to 1000
 747          * microseconds.
 748          */
 749         else if (cx->latency > ACPI_PROCESSOR_MAX_C3_LATENCY) {
 750                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 751                                   "latency too large [%d]\n", cx->latency));
 752                 return_VOID;
 753         }
 754
 755         /*
 756          * PIIX4 Erratum #18: We don't support C3 when Type-F (fast)
 757          * DMA transfers are used by any ISA device to avoid livelock.
 758          * Note that we could disable Type-F DMA (as recommended by
 759          * the erratum), but this is known to disrupt certain ISA
 760          * devices thus we take the conservative approach.
 761          */
 762         else if (errata.piix4.fdma) {
 763                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 764                                   "C3 not supported on PIIX4 with Type-F DMA\n"));
 765                 return_VOID;
 766         }
 767
 768         /* All the logic here assumes flags.bm_check is same across all CPUs */
 769         if (!bm_check_flag) {
 770                 /* Determine whether bm_check is needed based on CPU  */
 771                 acpi_processor_power_init_bm_check(&(pr->flags), pr->id);
 772                 bm_check_flag = pr->flags.bm_check;
 773         } else {
 774                 pr->flags.bm_check = bm_check_flag;
 775         }
 776
 777         if (pr->flags.bm_check) {
 778                 /* bus mastering control is necessary */
 779                 if (!pr->flags.bm_control) {
 780                         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 781                                           "C3 support requires bus mastering control\n"));
 782                         return_VOID;
 783                 }
 784         } else {
 785                 /*
 786                  * WBINVD should be set in fadt, for C3 state to be
 787                  * supported on when bm_check is not required.
 788                  */
 789                 if (acpi_fadt.wb_invd != 1) {
 790                         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 791                                           "Cache invalidation should work properly"
 792                                           " for C3 to be enabled on SMP systems\n"));
 793                         return_VOID;
 794                 }
 795                 acpi_set_register(ACPI_BITREG_BUS_MASTER_RLD,
 796                                   0, ACPI_MTX_DO_NOT_LOCK);
 797         }
 798
 799         /*
 800          * Otherwise we've met all of our C3 requirements.
 801          * Normalize the C3 latency to expidite policy.  Enable
 802          * checking of bus mastering status (bm_check) so we can
 803          * use this in our C3 policy
 804          */
 805         cx->valid = 1;
 806         cx->latency_ticks = US_TO_PM_TIMER_TICKS(cx->latency);
 807
 808         return_VOID;
 809 }
 810
 811 static int acpi_processor_power_verify(struct acpi_processor *pr)
 812 {
 813         unsigned int i;
 814         unsigned int working = 0;
 815
 816         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
 817                 struct acpi_processor_cx *cx = &pr->power.states[i];
 818
 819                 switch (cx->type) {
 820                 case ACPI_STATE_C1:
 821                         cx->valid = 1;
 822                         break;
 823
 824                 case ACPI_STATE_C2:
 825                         acpi_processor_power_verify_c2(cx);
 826                         break;
 827
 828                 case ACPI_STATE_C3:
 829                         acpi_processor_power_verify_c3(pr, cx);
 830                         break;
 831                 }
 832
 833                 if (cx->valid)
 834                         working++;
 835         }
 836
 837         return (working);
 838 }
 839
 840 static int acpi_processor_get_power_info(struct acpi_processor *pr)
 841 {
 842         unsigned int i;
 843         int result;
 844
 845         ACPI_FUNCTION_TRACE("acpi_processor_get_power_info");
 846
 847         /* NOTE: the idle thread may not be running while calling
 848          * this function */
 849
 850         result = acpi_processor_get_power_info_cst(pr);
 851         if ((result) || (acpi_processor_power_verify(pr) < 2)) {
 852                 result = acpi_processor_get_power_info_fadt(pr);
 853                 if ((result) || (acpi_processor_power_verify(pr) < 2))
 854                         result = acpi_processor_get_power_info_default_c1(pr);
 855         }
 856
 857         /*
 858          * Set Default Policy
 859          * ------------------
 860          * Now that we know which states are supported, set the default
 861          * policy.  Note that this policy can be changed dynamically
 862          * (e.g. encourage deeper sleeps to conserve battery life when
 863          * not on AC).
 864          */
 865         result = acpi_processor_set_power_policy(pr);
 866         if (result)
 867                 return_VALUE(result);
 868
 869         /*
 870          * if one state of type C2 or C3 is available, mark this
 871          * CPU as being "idle manageable"
 872          */
 873         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
 874                 if (pr->power.states[i].valid) {
 875                         pr->power.count = i;
 876                         pr->flags.power = 1;
 877                 }
 878         }
 879
 880         return_VALUE(0);
 881 }
 882
 883 int acpi_processor_cst_has_changed(struct acpi_processor *pr)
 884 {
 885         int result = 0;
 886
 887         ACPI_FUNCTION_TRACE("acpi_processor_cst_has_changed");
 888
 889         if (!pr)
 890                 return_VALUE(-EINVAL);
 891
 892         if (nocst) {
 893                 return_VALUE(-ENODEV);
 894         }
 895
 896         if (!pr->flags.power_setup_done)
 897                 return_VALUE(-ENODEV);
 898
 899         /* Fall back to the default idle loop */
 900         pm_idle = pm_idle_save;
 901         synchronize_sched();    /* Relies on interrupts forcing exit from idle. */
 902
 903         pr->flags.power = 0;
 904         result = acpi_processor_get_power_info(pr);
 905         if ((pr->flags.power == 1) && (pr->flags.power_setup_done))
 906                 pm_idle = acpi_processor_idle;
 907
 908         return_VALUE(result);
 909 }
 910
 911 /* proc interface */
 912
 913 static int acpi_processor_power_seq_show(struct seq_file *seq, void *offset)
 914 {
 915         struct acpi_processor *pr = (struct acpi_processor *)seq->private;
 916         unsigned int i;
 917
 918         ACPI_FUNCTION_TRACE("acpi_processor_power_seq_show");
 919
 920         if (!pr)
 921                 goto end;
 922
 923         seq_printf(seq, "active state:            C%zd\n"
 924                    "max_cstate:              C%d\n"
 925                    "bus master activity:     %08x\n",
 926                    pr->power.state ? pr->power.state - pr->power.states : 0,
 927                    max_cstate, (unsigned)pr->power.bm_activity);
 928
 929         seq_puts(seq, "states:\n");
 930
 931         for (i = 1; i <= pr->power.count; i++) {
 932                 seq_printf(seq, "   %cC%d:                  ",
 933                            (&pr->power.states[i] ==
 934                             pr->power.state ? '*' : ' '), i);
 935
 936                 if (!pr->power.states[i].valid) {
 937                         seq_puts(seq, "<not supported>\n");
 938                         continue;
 939                 }
 940
 941                 switch (pr->power.states[i].type) {
 942                 case ACPI_STATE_C1:
 943                         seq_printf(seq, "type[C1] ");
 944                         break;
 945                 case ACPI_STATE_C2:
 946                         seq_printf(seq, "type[C2] ");
 947                         break;
 948                 case ACPI_STATE_C3:
 949                         seq_printf(seq, "type[C3] ");
 950                         break;
 951                 default:
 952                         seq_printf(seq, "type[--] ");
 953                         break;
 954                 }
 955
 956                 if (pr->power.states[i].promotion.state)
 957                         seq_printf(seq, "promotion[C%zd] ",
 958                                    (pr->power.states[i].promotion.state -
 959                                     pr->power.states));
 960                 else
 961                         seq_puts(seq, "promotion[--] ");
 962
 963                 if (pr->power.states[i].demotion.state)
 964                         seq_printf(seq, "demotion[C%zd] ",
 965                                    (pr->power.states[i].demotion.state -
 966                                     pr->power.states));
 967                 else
 968                         seq_puts(seq, "demotion[--] ");
 969
 970                 seq_printf(seq, "latency[%03d] usage[%08d]\n",
 971                            pr->power.states[i].latency,
 972                            pr->power.states[i].usage);
 973         }
 974
 975       end:
 976         return_VALUE(0);
 977 }
 978
 979 static int acpi_processor_power_open_fs(struct inode *inode, struct file *file)
 980 {
 981         return single_open(file, acpi_processor_power_seq_show,
 982                            PDE(inode)->data);
 983 }
 984
 985 static struct file_operations acpi_processor_power_fops = {
 986         .open = acpi_processor_power_open_fs,
 987         .read = seq_read,
 988         .llseek = seq_lseek,
 989         .release = single_release,
 990 };
 991
 992 int acpi_processor_power_init(struct acpi_processor *pr,
 993                               struct acpi_device *device)
 994 {
 995         acpi_status status = 0;
 996         static int first_run = 0;
 997         struct proc_dir_entry *entry = NULL;
 998         unsigned int i;
 999
1000         ACPI_FUNCTION_TRACE("acpi_processor_power_init");
1001
1002         if (!first_run) {
1003                 dmi_check_system(processor_power_dmi_table);
1004                 if (max_cstate < ACPI_C_STATES_MAX)
1005                         printk(KERN_NOTICE
1006                                "ACPI: processor limited to max C-state %d\n",
1007                                max_cstate);
1008                 first_run++;
1009         }
1010
1011         if (!pr)
1012                 return_VALUE(-EINVAL);
1013
1014         if (acpi_fadt.cst_cnt && !nocst) {
1015                 status =
1016                     acpi_os_write_port(acpi_fadt.smi_cmd, acpi_fadt.cst_cnt, 8);
1017                 if (ACPI_FAILURE(status)) {
1018                         ACPI_DEBUG_PRINT((ACPI_DB_ERROR,
1019                                           "Notifying BIOS of _CST ability failed\n"));
1020                 }
1021         }
1022
1023         acpi_processor_power_init_pdc(&(pr->power), pr->id);
1024         acpi_processor_set_pdc(pr, pr->power.pdc);
1025         acpi_processor_get_power_info(pr);
1026
1027         /*
1028          * Install the idle handler if processor power management is supported.
1029          * Note that we use previously set idle handler will be used on
1030          * platforms that only support C1.
1031          */
1032         if ((pr->flags.power) && (!boot_option_idle_override)) {
1033                 printk(KERN_INFO PREFIX "CPU%d (power states:", pr->id);
1034                 for (i = 1; i <= pr->power.count; i++)
1035                         if (pr->power.states[i].valid)
1036                                 printk(" C%d[C%d]", i,
1037                                        pr->power.states[i].type);
1038                 printk(")\n");
1039
1040                 if (pr->id == 0) {
1041                         pm_idle_save = pm_idle;
1042                         pm_idle = acpi_processor_idle;
1043                 }
1044         }
1045
1046         /* 'power' [R] */
1047         entry = create_proc_entry(ACPI_PROCESSOR_FILE_POWER,
1048                                   S_IRUGO, acpi_device_dir(device));
1049         if (!entry)
1050                 ACPI_DEBUG_PRINT((ACPI_DB_ERROR,
1051                                   "Unable to create '%s' fs entry\n",
1052                                   ACPI_PROCESSOR_FILE_POWER));
1053         else {
1054                 entry->proc_fops = &acpi_processor_power_fops;
1055                 entry->data = acpi_driver_data(device);
1056                 entry->owner = THIS_MODULE;
1057         }
1058
1059         pr->flags.power_setup_done = 1;
1060
1061         return_VALUE(0);
1062 }
1063
1064 int acpi_processor_power_exit(struct acpi_processor *pr,
1065                               struct acpi_device *device)
1066 {
1067         ACPI_FUNCTION_TRACE("acpi_processor_power_exit");
1068
1069         pr->flags.power_setup_done = 0;
1070
1071         if (acpi_device_dir(device))
1072                 remove_proc_entry(ACPI_PROCESSOR_FILE_POWER,
1073                                   acpi_device_dir(device));
1074
1075         /* Unregister the idle handler when processor #0 is removed. */
1076         if (pr->id == 0) {
1077                 pm_idle = pm_idle_save;
1078
1079                 /*
1080                  * We are about to unload the current idle thread pm callback
1081                  * (pm_idle), Wait for all processors to update cached/local
1082                  * copies of pm_idle before proceeding.
1083                  */
1084                 cpu_idle_wait();
1085         }
1086
1087         return_VALUE(0);
1088 }