SAFE public projects git trees. - safe/jmp/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 #ifdef CONFIG_HOTPLUG_CPU
 282         /*
 283          * Check for P_LVL2_UP flag before entering C2 and above on
 284          * an SMP system. We do it here instead of doing it at _CST/P_LVL
 285          * detection phase, to work cleanly with logical CPU hotplug.
 286          */
 287         if ((cx->type != ACPI_STATE_C1) && (num_online_cpus() > 1) &&
 288             !pr->flags.has_cst && !acpi_fadt.plvl2_up)
 289                 cx = &pr->power.states[ACPI_STATE_C1];
 290 #endif
 291
 292         cx->usage++;
 293
 294         /*
 295          * Sleep:
 296          * ------
 297          * Invoke the current Cx state to put the processor to sleep.
 298          */
 299         switch (cx->type) {
 300
 301         case ACPI_STATE_C1:
 302                 /*
 303                  * Invoke C1.
 304                  * Use the appropriate idle routine, the one that would
 305                  * be used without acpi C-states.
 306                  */
 307                 if (pm_idle_save)
 308                         pm_idle_save();
 309                 else
 310                         acpi_safe_halt();
 311
 312                 /*
 313                  * TBD: Can't get time duration while in C1, as resumes
 314                  *      go to an ISR rather than here.  Need to instrument
 315                  *      base interrupt handler.
 316                  */
 317                 sleep_ticks = 0xFFFFFFFF;
 318                 break;
 319
 320         case ACPI_STATE_C2:
 321                 /* Get start time (ticks) */
 322                 t1 = inl(acpi_fadt.xpm_tmr_blk.address);
 323                 /* Invoke C2 */
 324                 inb(cx->address);
 325                 /* Dummy op - must do something useless after P_LVL2 read */
 326                 t2 = inl(acpi_fadt.xpm_tmr_blk.address);
 327                 /* Get end time (ticks) */
 328                 t2 = inl(acpi_fadt.xpm_tmr_blk.address);
 329                 /* Re-enable interrupts */
 330                 local_irq_enable();
 331                 /* Compute time (ticks) that we were actually asleep */
 332                 sleep_ticks =
 333                     ticks_elapsed(t1, t2) - cx->latency_ticks - C2_OVERHEAD;
 334                 break;
 335
 336         case ACPI_STATE_C3:
 337
 338                 if (pr->flags.bm_check) {
 339                         if (atomic_inc_return(&c3_cpu_count) ==
 340                             num_online_cpus()) {
 341                                 /*
 342                                  * All CPUs are trying to go to C3
 343                                  * Disable bus master arbitration
 344                                  */
 345                                 acpi_set_register(ACPI_BITREG_ARB_DISABLE, 1,
 346                                                   ACPI_MTX_DO_NOT_LOCK);
 347                         }
 348                 } else {
 349                         /* SMP with no shared cache... Invalidate cache  */
 350                         ACPI_FLUSH_CPU_CACHE();
 351                 }
 352
 353                 /* Get start time (ticks) */
 354                 t1 = inl(acpi_fadt.xpm_tmr_blk.address);
 355                 /* Invoke C3 */
 356                 inb(cx->address);
 357                 /* Dummy op - must do something useless after P_LVL3 read */
 358                 t2 = inl(acpi_fadt.xpm_tmr_blk.address);
 359                 /* Get end time (ticks) */
 360                 t2 = inl(acpi_fadt.xpm_tmr_blk.address);
 361                 if (pr->flags.bm_check) {
 362                         /* Enable bus master arbitration */
 363                         atomic_dec(&c3_cpu_count);
 364                         acpi_set_register(ACPI_BITREG_ARB_DISABLE, 0,
 365                                           ACPI_MTX_DO_NOT_LOCK);
 366                 }
 367
 368                 /* Re-enable interrupts */
 369                 local_irq_enable();
 370                 /* Compute time (ticks) that we were actually asleep */
 371                 sleep_ticks =
 372                     ticks_elapsed(t1, t2) - cx->latency_ticks - C3_OVERHEAD;
 373                 break;
 374
 375         default:
 376                 local_irq_enable();
 377                 return;
 378         }
 379
 380         next_state = pr->power.state;
 381
 382 #ifdef CONFIG_HOTPLUG_CPU
 383         /* Don't do promotion/demotion */
 384         if ((cx->type == ACPI_STATE_C1) && (num_online_cpus() > 1) &&
 385             !pr->flags.has_cst && !acpi_fadt.plvl2_up) {
 386                 next_state = cx;
 387                 goto end;
 388         }
 389 #endif
 390
 391         /*
 392          * Promotion?
 393          * ----------
 394          * Track the number of longs (time asleep is greater than threshold)
 395          * and promote when the count threshold is reached.  Note that bus
 396          * mastering activity may prevent promotions.
 397          * Do not promote above max_cstate.
 398          */
 399         if (cx->promotion.state &&
 400             ((cx->promotion.state - pr->power.states) <= max_cstate)) {
 401                 if (sleep_ticks > cx->promotion.threshold.ticks) {
 402                         cx->promotion.count++;
 403                         cx->demotion.count = 0;
 404                         if (cx->promotion.count >=
 405                             cx->promotion.threshold.count) {
 406                                 if (pr->flags.bm_check) {
 407                                         if (!
 408                                             (pr->power.bm_activity & cx->
 409                                              promotion.threshold.bm)) {
 410                                                 next_state =
 411                                                     cx->promotion.state;
 412                                                 goto end;
 413                                         }
 414                                 } else {
 415                                         next_state = cx->promotion.state;
 416                                         goto end;
 417                                 }
 418                         }
 419                 }
 420         }
 421
 422         /*
 423          * Demotion?
 424          * ---------
 425          * Track the number of shorts (time asleep is less than time threshold)
 426          * and demote when the usage threshold is reached.
 427          */
 428         if (cx->demotion.state) {
 429                 if (sleep_ticks < cx->demotion.threshold.ticks) {
 430                         cx->demotion.count++;
 431                         cx->promotion.count = 0;
 432                         if (cx->demotion.count >= cx->demotion.threshold.count) {
 433                                 next_state = cx->demotion.state;
 434                                 goto end;
 435                         }
 436                 }
 437         }
 438
 439       end:
 440         /*
 441          * Demote if current state exceeds max_cstate
 442          */
 443         if ((pr->power.state - pr->power.states) > max_cstate) {
 444                 if (cx->demotion.state)
 445                         next_state = cx->demotion.state;
 446         }
 447
 448         /*
 449          * New Cx State?
 450          * -------------
 451          * If we're going to start using a new Cx state we must clean up
 452          * from the previous and prepare to use the new.
 453          */
 454         if (next_state != pr->power.state)
 455                 acpi_processor_power_activate(pr, next_state);
 456 }
 457
 458 static int acpi_processor_set_power_policy(struct acpi_processor *pr)
 459 {
 460         unsigned int i;
 461         unsigned int state_is_set = 0;
 462         struct acpi_processor_cx *lower = NULL;
 463         struct acpi_processor_cx *higher = NULL;
 464         struct acpi_processor_cx *cx;
 465
 466         ACPI_FUNCTION_TRACE("acpi_processor_set_power_policy");
 467
 468         if (!pr)
 469                 return_VALUE(-EINVAL);
 470
 471         /*
 472          * This function sets the default Cx state policy (OS idle handler).
 473          * Our scheme is to promote quickly to C2 but more conservatively
 474          * to C3.  We're favoring C2  for its characteristics of low latency
 475          * (quick response), good power savings, and ability to allow bus
 476          * mastering activity.  Note that the Cx state policy is completely
 477          * customizable and can be altered dynamically.
 478          */
 479
 480         /* startup state */
 481         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
 482                 cx = &pr->power.states[i];
 483                 if (!cx->valid)
 484                         continue;
 485
 486                 if (!state_is_set)
 487                         pr->power.state = cx;
 488                 state_is_set++;
 489                 break;
 490         }
 491
 492         if (!state_is_set)
 493                 return_VALUE(-ENODEV);
 494
 495         /* demotion */
 496         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
 497                 cx = &pr->power.states[i];
 498                 if (!cx->valid)
 499                         continue;
 500
 501                 if (lower) {
 502                         cx->demotion.state = lower;
 503                         cx->demotion.threshold.ticks = cx->latency_ticks;
 504                         cx->demotion.threshold.count = 1;
 505                         if (cx->type == ACPI_STATE_C3)
 506                                 cx->demotion.threshold.bm = bm_history;
 507                 }
 508
 509                 lower = cx;
 510         }
 511
 512         /* promotion */
 513         for (i = (ACPI_PROCESSOR_MAX_POWER - 1); i > 0; i--) {
 514                 cx = &pr->power.states[i];
 515                 if (!cx->valid)
 516                         continue;
 517
 518                 if (higher) {
 519                         cx->promotion.state = higher;
 520                         cx->promotion.threshold.ticks = cx->latency_ticks;
 521                         if (cx->type >= ACPI_STATE_C2)
 522                                 cx->promotion.threshold.count = 4;
 523                         else
 524                                 cx->promotion.threshold.count = 10;
 525                         if (higher->type == ACPI_STATE_C3)
 526                                 cx->promotion.threshold.bm = bm_history;
 527                 }
 528
 529                 higher = cx;
 530         }
 531
 532         return_VALUE(0);
 533 }
 534
 535 static int acpi_processor_get_power_info_fadt(struct acpi_processor *pr)
 536 {
 537         ACPI_FUNCTION_TRACE("acpi_processor_get_power_info_fadt");
 538
 539         if (!pr)
 540                 return_VALUE(-EINVAL);
 541
 542         if (!pr->pblk)
 543                 return_VALUE(-ENODEV);
 544
 545         memset(pr->power.states, 0, sizeof(pr->power.states));
 546
 547         /* if info is obtained from pblk/fadt, type equals state */
 548         pr->power.states[ACPI_STATE_C1].type = ACPI_STATE_C1;
 549         pr->power.states[ACPI_STATE_C2].type = ACPI_STATE_C2;
 550         pr->power.states[ACPI_STATE_C3].type = ACPI_STATE_C3;
 551
 552         /* the C0 state only exists as a filler in our array,
 553          * and all processors need to support C1 */
 554         pr->power.states[ACPI_STATE_C0].valid = 1;
 555         pr->power.states[ACPI_STATE_C1].valid = 1;
 556
 557 #ifndef CONFIG_HOTPLUG_CPU
 558         /*
 559          * Check for P_LVL2_UP flag before entering C2 and above on
 560          * an SMP system.
 561          */
 562         if ((num_online_cpus() > 1) && !acpi_fadt.plvl2_up)
 563                 return_VALUE(-ENODEV);
 564 #endif
 565
 566         /* determine C2 and C3 address from pblk */
 567         pr->power.states[ACPI_STATE_C2].address = pr->pblk + 4;
 568         pr->power.states[ACPI_STATE_C3].address = pr->pblk + 5;
 569
 570         /* determine latencies from FADT */
 571         pr->power.states[ACPI_STATE_C2].latency = acpi_fadt.plvl2_lat;
 572         pr->power.states[ACPI_STATE_C3].latency = acpi_fadt.plvl3_lat;
 573
 574         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 575                           "lvl2[0x%08x] lvl3[0x%08x]\n",
 576                           pr->power.states[ACPI_STATE_C2].address,
 577                           pr->power.states[ACPI_STATE_C3].address));
 578
 579         return_VALUE(0);
 580 }
 581
 582 static int acpi_processor_get_power_info_default_c1(struct acpi_processor *pr)
 583 {
 584         ACPI_FUNCTION_TRACE("acpi_processor_get_power_info_default_c1");
 585
 586         memset(pr->power.states, 0, sizeof(pr->power.states));
 587
 588         /* if info is obtained from pblk/fadt, type equals state */
 589         pr->power.states[ACPI_STATE_C1].type = ACPI_STATE_C1;
 590         pr->power.states[ACPI_STATE_C2].type = ACPI_STATE_C2;
 591         pr->power.states[ACPI_STATE_C3].type = ACPI_STATE_C3;
 592
 593         /* the C0 state only exists as a filler in our array,
 594          * and all processors need to support C1 */
 595         pr->power.states[ACPI_STATE_C0].valid = 1;
 596         pr->power.states[ACPI_STATE_C1].valid = 1;
 597
 598         return_VALUE(0);
 599 }
 600
 601 static int acpi_processor_get_power_info_cst(struct acpi_processor *pr)
 602 {
 603         acpi_status status = 0;
 604         acpi_integer count;
 605         int i;
 606         struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
 607         union acpi_object *cst;
 608
 609         ACPI_FUNCTION_TRACE("acpi_processor_get_power_info_cst");
 610
 611         if (nocst)
 612                 return_VALUE(-ENODEV);
 613
 614         pr->power.count = 0;
 615         for (i = 0; i < ACPI_PROCESSOR_MAX_POWER; i++)
 616                 memset(&(pr->power.states[i]), 0,
 617                        sizeof(struct acpi_processor_cx));
 618
 619         status = acpi_evaluate_object(pr->handle, "_CST", NULL, &buffer);
 620         if (ACPI_FAILURE(status)) {
 621                 ACPI_DEBUG_PRINT((ACPI_DB_INFO, "No _CST, giving up\n"));
 622                 return_VALUE(-ENODEV);
 623         }
 624
 625         cst = (union acpi_object *)buffer.pointer;
 626
 627         /* There must be at least 2 elements */
 628         if (!cst || (cst->type != ACPI_TYPE_PACKAGE) || cst->package.count < 2) {
 629                 ACPI_DEBUG_PRINT((ACPI_DB_ERROR,
 630                                   "not enough elements in _CST\n"));
 631                 status = -EFAULT;
 632                 goto end;
 633         }
 634
 635         count = cst->package.elements[0].integer.value;
 636
 637         /* Validate number of power states. */
 638         if (count < 1 || count != cst->package.count - 1) {
 639                 ACPI_DEBUG_PRINT((ACPI_DB_ERROR,
 640                                   "count given by _CST is not valid\n"));
 641                 status = -EFAULT;
 642                 goto end;
 643         }
 644
 645         /* We support up to ACPI_PROCESSOR_MAX_POWER. */
 646         if (count > ACPI_PROCESSOR_MAX_POWER) {
 647                 printk(KERN_WARNING
 648                        "Limiting number of power states to max (%d)\n",
 649                        ACPI_PROCESSOR_MAX_POWER);
 650                 printk(KERN_WARNING
 651                        "Please increase ACPI_PROCESSOR_MAX_POWER if needed.\n");
 652                 count = ACPI_PROCESSOR_MAX_POWER;
 653         }
 654
 655         /* Tell driver that at least _CST is supported. */
 656         pr->flags.has_cst = 1;
 657
 658         for (i = 1; i <= count; i++) {
 659                 union acpi_object *element;
 660                 union acpi_object *obj;
 661                 struct acpi_power_register *reg;
 662                 struct acpi_processor_cx cx;
 663
 664                 memset(&cx, 0, sizeof(cx));
 665
 666                 element = (union acpi_object *)&(cst->package.elements[i]);
 667                 if (element->type != ACPI_TYPE_PACKAGE)
 668                         continue;
 669
 670                 if (element->package.count != 4)
 671                         continue;
 672
 673                 obj = (union acpi_object *)&(element->package.elements[0]);
 674
 675                 if (obj->type != ACPI_TYPE_BUFFER)
 676                         continue;
 677
 678                 reg = (struct acpi_power_register *)obj->buffer.pointer;
 679
 680                 if (reg->space_id != ACPI_ADR_SPACE_SYSTEM_IO &&
 681                     (reg->space_id != ACPI_ADR_SPACE_FIXED_HARDWARE))
 682                         continue;
 683
 684                 cx.address = (reg->space_id == ACPI_ADR_SPACE_FIXED_HARDWARE) ?
 685                     0 : reg->address;
 686
 687                 /* There should be an easy way to extract an integer... */
 688                 obj = (union acpi_object *)&(element->package.elements[1]);
 689                 if (obj->type != ACPI_TYPE_INTEGER)
 690                         continue;
 691
 692                 cx.type = obj->integer.value;
 693
 694                 if ((cx.type != ACPI_STATE_C1) &&
 695                     (reg->space_id != ACPI_ADR_SPACE_SYSTEM_IO))
 696                         continue;
 697
 698                 if ((cx.type < ACPI_STATE_C1) || (cx.type > ACPI_STATE_C3))
 699                         continue;
 700
 701                 obj = (union acpi_object *)&(element->package.elements[2]);
 702                 if (obj->type != ACPI_TYPE_INTEGER)
 703                         continue;
 704
 705                 cx.latency = obj->integer.value;
 706
 707                 obj = (union acpi_object *)&(element->package.elements[3]);
 708                 if (obj->type != ACPI_TYPE_INTEGER)
 709                         continue;
 710
 711                 cx.power = obj->integer.value;
 712
 713                 (pr->power.count)++;
 714                 memcpy(&(pr->power.states[pr->power.count]), &cx, sizeof(cx));
 715         }
 716
 717         ACPI_DEBUG_PRINT((ACPI_DB_INFO, "Found %d power states\n",
 718                           pr->power.count));
 719
 720         /* Validate number of power states discovered */
 721         if (pr->power.count < 2)
 722                 status = -EFAULT;
 723
 724       end:
 725         acpi_os_free(buffer.pointer);
 726
 727         return_VALUE(status);
 728 }
 729
 730 static void acpi_processor_power_verify_c2(struct acpi_processor_cx *cx)
 731 {
 732         ACPI_FUNCTION_TRACE("acpi_processor_get_power_verify_c2");
 733
 734         if (!cx->address)
 735                 return_VOID;
 736
 737         /*
 738          * C2 latency must be less than or equal to 100
 739          * microseconds.
 740          */
 741         else if (cx->latency > ACPI_PROCESSOR_MAX_C2_LATENCY) {
 742                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 743                                   "latency too large [%d]\n", cx->latency));
 744                 return_VOID;
 745         }
 746
 747         /*
 748          * Otherwise we've met all of our C2 requirements.
 749          * Normalize the C2 latency to expidite policy
 750          */
 751         cx->valid = 1;
 752         cx->latency_ticks = US_TO_PM_TIMER_TICKS(cx->latency);
 753
 754         return_VOID;
 755 }
 756
 757 static void acpi_processor_power_verify_c3(struct acpi_processor *pr,
 758                                            struct acpi_processor_cx *cx)
 759 {
 760         static int bm_check_flag;
 761
 762         ACPI_FUNCTION_TRACE("acpi_processor_get_power_verify_c3");
 763
 764         if (!cx->address)
 765                 return_VOID;
 766
 767         /*
 768          * C3 latency must be less than or equal to 1000
 769          * microseconds.
 770          */
 771         else if (cx->latency > ACPI_PROCESSOR_MAX_C3_LATENCY) {
 772                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 773                                   "latency too large [%d]\n", cx->latency));
 774                 return_VOID;
 775         }
 776
 777         /*
 778          * PIIX4 Erratum #18: We don't support C3 when Type-F (fast)
 779          * DMA transfers are used by any ISA device to avoid livelock.
 780          * Note that we could disable Type-F DMA (as recommended by
 781          * the erratum), but this is known to disrupt certain ISA
 782          * devices thus we take the conservative approach.
 783          */
 784         else if (errata.piix4.fdma) {
 785                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 786                                   "C3 not supported on PIIX4 with Type-F DMA\n"));
 787                 return_VOID;
 788         }
 789
 790         /* All the logic here assumes flags.bm_check is same across all CPUs */
 791         if (!bm_check_flag) {
 792                 /* Determine whether bm_check is needed based on CPU  */
 793                 acpi_processor_power_init_bm_check(&(pr->flags), pr->id);
 794                 bm_check_flag = pr->flags.bm_check;
 795         } else {
 796                 pr->flags.bm_check = bm_check_flag;
 797         }
 798
 799         if (pr->flags.bm_check) {
 800                 /* bus mastering control is necessary */
 801                 if (!pr->flags.bm_control) {
 802                         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 803                                           "C3 support requires bus mastering control\n"));
 804                         return_VOID;
 805                 }
 806         } else {
 807                 /*
 808                  * WBINVD should be set in fadt, for C3 state to be
 809                  * supported on when bm_check is not required.
 810                  */
 811                 if (acpi_fadt.wb_invd != 1) {
 812                         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 813                                           "Cache invalidation should work properly"
 814                                           " for C3 to be enabled on SMP systems\n"));
 815                         return_VOID;
 816                 }
 817                 acpi_set_register(ACPI_BITREG_BUS_MASTER_RLD,
 818                                   0, ACPI_MTX_DO_NOT_LOCK);
 819         }
 820
 821         /*
 822          * Otherwise we've met all of our C3 requirements.
 823          * Normalize the C3 latency to expidite policy.  Enable
 824          * checking of bus mastering status (bm_check) so we can
 825          * use this in our C3 policy
 826          */
 827         cx->valid = 1;
 828         cx->latency_ticks = US_TO_PM_TIMER_TICKS(cx->latency);
 829
 830         return_VOID;
 831 }
 832
 833 static int acpi_processor_power_verify(struct acpi_processor *pr)
 834 {
 835         unsigned int i;
 836         unsigned int working = 0;
 837
 838         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
 839                 struct acpi_processor_cx *cx = &pr->power.states[i];
 840
 841                 switch (cx->type) {
 842                 case ACPI_STATE_C1:
 843                         cx->valid = 1;
 844                         break;
 845
 846                 case ACPI_STATE_C2:
 847                         acpi_processor_power_verify_c2(cx);
 848                         break;
 849
 850                 case ACPI_STATE_C3:
 851                         acpi_processor_power_verify_c3(pr, cx);
 852                         break;
 853                 }
 854
 855                 if (cx->valid)
 856                         working++;
 857         }
 858
 859         return (working);
 860 }
 861
 862 static int acpi_processor_get_power_info(struct acpi_processor *pr)
 863 {
 864         unsigned int i;
 865         int result;
 866
 867         ACPI_FUNCTION_TRACE("acpi_processor_get_power_info");
 868
 869         /* NOTE: the idle thread may not be running while calling
 870          * this function */
 871
 872         result = acpi_processor_get_power_info_cst(pr);
 873         if (result == -ENODEV)
 874                 result = acpi_processor_get_power_info_fadt(pr);
 875
 876         if ((result) || (acpi_processor_power_verify(pr) < 2))
 877                 result = acpi_processor_get_power_info_default_c1(pr);
 878
 879         /*
 880          * Set Default Policy
 881          * ------------------
 882          * Now that we know which states are supported, set the default
 883          * policy.  Note that this policy can be changed dynamically
 884          * (e.g. encourage deeper sleeps to conserve battery life when
 885          * not on AC).
 886          */
 887         result = acpi_processor_set_power_policy(pr);
 888         if (result)
 889                 return_VALUE(result);
 890
 891         /*
 892          * if one state of type C2 or C3 is available, mark this
 893          * CPU as being "idle manageable"
 894          */
 895         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
 896                 if (pr->power.states[i].valid) {
 897                         pr->power.count = i;
 898                         if (pr->power.states[i].type >= ACPI_STATE_C2)
 899                                 pr->flags.power = 1;
 900                 }
 901         }
 902
 903         return_VALUE(0);
 904 }
 905
 906 int acpi_processor_cst_has_changed(struct acpi_processor *pr)
 907 {
 908         int result = 0;
 909
 910         ACPI_FUNCTION_TRACE("acpi_processor_cst_has_changed");
 911
 912         if (!pr)
 913                 return_VALUE(-EINVAL);
 914
 915         if (nocst) {
 916                 return_VALUE(-ENODEV);
 917         }
 918
 919         if (!pr->flags.power_setup_done)
 920                 return_VALUE(-ENODEV);
 921
 922         /* Fall back to the default idle loop */
 923         pm_idle = pm_idle_save;
 924         synchronize_sched();    /* Relies on interrupts forcing exit from idle. */
 925
 926         pr->flags.power = 0;
 927         result = acpi_processor_get_power_info(pr);
 928         if ((pr->flags.power == 1) && (pr->flags.power_setup_done))
 929                 pm_idle = acpi_processor_idle;
 930
 931         return_VALUE(result);
 932 }
 933
 934 /* proc interface */
 935
 936 static int acpi_processor_power_seq_show(struct seq_file *seq, void *offset)
 937 {
 938         struct acpi_processor *pr = (struct acpi_processor *)seq->private;
 939         unsigned int i;
 940
 941         ACPI_FUNCTION_TRACE("acpi_processor_power_seq_show");
 942
 943         if (!pr)
 944                 goto end;
 945
 946         seq_printf(seq, "active state:            C%zd\n"
 947                    "max_cstate:              C%d\n"
 948                    "bus master activity:     %08x\n",
 949                    pr->power.state ? pr->power.state - pr->power.states : 0,
 950                    max_cstate, (unsigned)pr->power.bm_activity);
 951
 952         seq_puts(seq, "states:\n");
 953
 954         for (i = 1; i <= pr->power.count; i++) {
 955                 seq_printf(seq, "   %cC%d:                  ",
 956                            (&pr->power.states[i] ==
 957                             pr->power.state ? '*' : ' '), i);
 958
 959                 if (!pr->power.states[i].valid) {
 960                         seq_puts(seq, "<not supported>\n");
 961                         continue;
 962                 }
 963
 964                 switch (pr->power.states[i].type) {
 965                 case ACPI_STATE_C1:
 966                         seq_printf(seq, "type[C1] ");
 967                         break;
 968                 case ACPI_STATE_C2:
 969                         seq_printf(seq, "type[C2] ");
 970                         break;
 971                 case ACPI_STATE_C3:
 972                         seq_printf(seq, "type[C3] ");
 973                         break;
 974                 default:
 975                         seq_printf(seq, "type[--] ");
 976                         break;
 977                 }
 978
 979                 if (pr->power.states[i].promotion.state)
 980                         seq_printf(seq, "promotion[C%zd] ",
 981                                    (pr->power.states[i].promotion.state -
 982                                     pr->power.states));
 983                 else
 984                         seq_puts(seq, "promotion[--] ");
 985
 986                 if (pr->power.states[i].demotion.state)
 987                         seq_printf(seq, "demotion[C%zd] ",
 988                                    (pr->power.states[i].demotion.state -
 989                                     pr->power.states));
 990                 else
 991                         seq_puts(seq, "demotion[--] ");
 992
 993                 seq_printf(seq, "latency[%03d] usage[%08d]\n",
 994                            pr->power.states[i].latency,
 995                            pr->power.states[i].usage);
 996         }
 997
 998       end:
 999         return_VALUE(0);
1000 }
1001
1002 static int acpi_processor_power_open_fs(struct inode *inode, struct file *file)
1003 {
1004         return single_open(file, acpi_processor_power_seq_show,
1005                            PDE(inode)->data);
1006 }
1007
1008 static struct file_operations acpi_processor_power_fops = {
1009         .open = acpi_processor_power_open_fs,
1010         .read = seq_read,
1011         .llseek = seq_lseek,
1012         .release = single_release,
1013 };
1014
1015 int acpi_processor_power_init(struct acpi_processor *pr,
1016                               struct acpi_device *device)
1017 {
1018         acpi_status status = 0;
1019         static int first_run = 0;
1020         struct proc_dir_entry *entry = NULL;
1021         unsigned int i;
1022
1023         ACPI_FUNCTION_TRACE("acpi_processor_power_init");
1024
1025         if (!first_run) {
1026                 dmi_check_system(processor_power_dmi_table);
1027                 if (max_cstate < ACPI_C_STATES_MAX)
1028                         printk(KERN_NOTICE
1029                                "ACPI: processor limited to max C-state %d\n",
1030                                max_cstate);
1031                 first_run++;
1032         }
1033
1034         if (!pr)
1035                 return_VALUE(-EINVAL);
1036
1037         if (acpi_fadt.cst_cnt && !nocst) {
1038                 status =
1039                     acpi_os_write_port(acpi_fadt.smi_cmd, acpi_fadt.cst_cnt, 8);
1040                 if (ACPI_FAILURE(status)) {
1041                         ACPI_DEBUG_PRINT((ACPI_DB_ERROR,
1042                                           "Notifying BIOS of _CST ability failed\n"));
1043                 }
1044         }
1045
1046         acpi_processor_power_init_pdc(&(pr->power), pr->id);
1047         acpi_processor_set_pdc(pr, pr->power.pdc);
1048         acpi_processor_get_power_info(pr);
1049
1050         /*
1051          * Install the idle handler if processor power management is supported.
1052          * Note that we use previously set idle handler will be used on
1053          * platforms that only support C1.
1054          */
1055         if ((pr->flags.power) && (!boot_option_idle_override)) {
1056                 printk(KERN_INFO PREFIX "CPU%d (power states:", pr->id);
1057                 for (i = 1; i <= pr->power.count; i++)
1058                         if (pr->power.states[i].valid)
1059                                 printk(" C%d[C%d]", i,
1060                                        pr->power.states[i].type);
1061                 printk(")\n");
1062
1063                 if (pr->id == 0) {
1064                         pm_idle_save = pm_idle;
1065                         pm_idle = acpi_processor_idle;
1066                 }
1067         }
1068
1069         /* 'power' [R] */
1070         entry = create_proc_entry(ACPI_PROCESSOR_FILE_POWER,
1071                                   S_IRUGO, acpi_device_dir(device));
1072         if (!entry)
1073                 ACPI_DEBUG_PRINT((ACPI_DB_ERROR,
1074                                   "Unable to create '%s' fs entry\n",
1075                                   ACPI_PROCESSOR_FILE_POWER));
1076         else {
1077                 entry->proc_fops = &acpi_processor_power_fops;
1078                 entry->data = acpi_driver_data(device);
1079                 entry->owner = THIS_MODULE;
1080         }
1081
1082         pr->flags.power_setup_done = 1;
1083
1084         return_VALUE(0);
1085 }
1086
1087 int acpi_processor_power_exit(struct acpi_processor *pr,
1088                               struct acpi_device *device)
1089 {
1090         ACPI_FUNCTION_TRACE("acpi_processor_power_exit");
1091
1092         pr->flags.power_setup_done = 0;
1093
1094         if (acpi_device_dir(device))
1095                 remove_proc_entry(ACPI_PROCESSOR_FILE_POWER,
1096                                   acpi_device_dir(device));
1097
1098         /* Unregister the idle handler when processor #0 is removed. */
1099         if (pr->id == 0) {
1100                 pm_idle = pm_idle_save;
1101
1102                 /*
1103                  * We are about to unload the current idle thread pm callback
1104                  * (pm_idle), Wait for all processors to update cached/local
1105                  * copies of pm_idle before proceeding.
1106                  */
1107                 cpu_idle_wait();
1108         }
1109
1110         return_VALUE(0);
1111 }