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