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         clear_thread_flag(TIF_POLLING_NRFLAG);
 173         smp_mb__after_clear_bit();
 174         if (!need_resched())
 175                 safe_halt();
 176         set_thread_flag(TIF_POLLING_NRFLAG);
 177 }
 178
 179 static atomic_t c3_cpu_count;
 180
 181 static void acpi_processor_idle(void)
 182 {
 183         struct acpi_processor *pr = NULL;
 184         struct acpi_processor_cx *cx = NULL;
 185         struct acpi_processor_cx *next_state = NULL;
 186         int sleep_ticks = 0;
 187         u32 t1, t2 = 0;
 188
 189         pr = processors[smp_processor_id()];
 190         if (!pr)
 191                 return;
 192
 193         /*
 194          * Interrupts must be disabled during bus mastering calculations and
 195          * for C2/C3 transitions.
 196          */
 197         local_irq_disable();
 198
 199         /*
 200          * Check whether we truly need to go idle, or should
 201          * reschedule:
 202          */
 203         if (unlikely(need_resched())) {
 204                 local_irq_enable();
 205                 return;
 206         }
 207
 208         cx = pr->power.state;
 209         if (!cx) {
 210                 if (pm_idle_save)
 211                         pm_idle_save();
 212                 else
 213                         acpi_safe_halt();
 214                 return;
 215         }
 216
 217         /*
 218          * Check BM Activity
 219          * -----------------
 220          * Check for bus mastering activity (if required), record, and check
 221          * for demotion.
 222          */
 223         if (pr->flags.bm_check) {
 224                 u32 bm_status = 0;
 225                 unsigned long diff = jiffies - pr->power.bm_check_timestamp;
 226
 227                 if (diff > 32)
 228                         diff = 32;
 229
 230                 while (diff) {
 231                         /* if we didn't get called, assume there was busmaster activity */
 232                         diff--;
 233                         if (diff)
 234                                 pr->power.bm_activity |= 0x1;
 235                         pr->power.bm_activity <<= 1;
 236                 }
 237
 238                 acpi_get_register(ACPI_BITREG_BUS_MASTER_STATUS,
 239                                   &bm_status, ACPI_MTX_DO_NOT_LOCK);
 240                 if (bm_status) {
 241                         pr->power.bm_activity++;
 242                         acpi_set_register(ACPI_BITREG_BUS_MASTER_STATUS,
 243                                           1, ACPI_MTX_DO_NOT_LOCK);
 244                 }
 245                 /*
 246                  * PIIX4 Erratum #18: Note that BM_STS doesn't always reflect
 247                  * the true state of bus mastering activity; forcing us to
 248                  * manually check the BMIDEA bit of each IDE channel.
 249                  */
 250                 else if (errata.piix4.bmisx) {
 251                         if ((inb_p(errata.piix4.bmisx + 0x02) & 0x01)
 252                             || (inb_p(errata.piix4.bmisx + 0x0A) & 0x01))
 253                                 pr->power.bm_activity++;
 254                 }
 255
 256                 pr->power.bm_check_timestamp = jiffies;
 257
 258                 /*
 259                  * Apply bus mastering demotion policy.  Automatically demote
 260                  * to avoid a faulty transition.  Note that the processor
 261                  * won't enter a low-power state during this call (to this
 262                  * funciton) but should upon the next.
 263                  *
 264                  * TBD: A better policy might be to fallback to the demotion
 265                  *      state (use it for this quantum only) istead of
 266                  *      demoting -- and rely on duration as our sole demotion
 267                  *      qualification.  This may, however, introduce DMA
 268                  *      issues (e.g. floppy DMA transfer overrun/underrun).
 269                  */
 270                 if (pr->power.bm_activity & cx->demotion.threshold.bm) {
 271                         local_irq_enable();
 272                         next_state = cx->demotion.state;
 273                         goto end;
 274                 }
 275         }
 276
 277         cx->usage++;
 278
 279 #ifdef CONFIG_HOTPLUG_CPU
 280         /*
 281          * Check for P_LVL2_UP flag before entering C2 and above on
 282          * an SMP system. We do it here instead of doing it at _CST/P_LVL
 283          * detection phase, to work cleanly with logical CPU hotplug.
 284          */
 285         if ((cx->type != ACPI_STATE_C1) && (num_online_cpus() > 1) &&
 286             !pr->flags.has_cst && acpi_fadt.plvl2_up)
 287                 cx->type = ACPI_STATE_C1;
 288 #endif
 289         /*
 290          * Sleep:
 291          * ------
 292          * Invoke the current Cx state to put the processor to sleep.
 293          */
 294         if (cx->type == ACPI_STATE_C2 || cx->type == ACPI_STATE_C3) {
 295                 clear_thread_flag(TIF_POLLING_NRFLAG);
 296                 smp_mb__after_clear_bit();
 297                 if (need_resched()) {
 298                         set_thread_flag(TIF_POLLING_NRFLAG);
 299                         local_irq_enable();
 300                         return;
 301                 }
 302         }
 303
 304         switch (cx->type) {
 305
 306         case ACPI_STATE_C1:
 307                 /*
 308                  * Invoke C1.
 309                  * Use the appropriate idle routine, the one that would
 310                  * be used without acpi C-states.
 311                  */
 312                 if (pm_idle_save)
 313                         pm_idle_save();
 314                 else
 315                         acpi_safe_halt();
 316
 317                 /*
 318                  * TBD: Can't get time duration while in C1, as resumes
 319                  *      go to an ISR rather than here.  Need to instrument
 320                  *      base interrupt handler.
 321                  */
 322                 sleep_ticks = 0xFFFFFFFF;
 323                 break;
 324
 325         case ACPI_STATE_C2:
 326                 /* Get start time (ticks) */
 327                 t1 = inl(acpi_fadt.xpm_tmr_blk.address);
 328                 /* Invoke C2 */
 329                 inb(cx->address);
 330                 /* Dummy op - must do something useless after P_LVL2 read */
 331                 t2 = inl(acpi_fadt.xpm_tmr_blk.address);
 332                 /* Get end time (ticks) */
 333                 t2 = inl(acpi_fadt.xpm_tmr_blk.address);
 334                 /* Re-enable interrupts */
 335                 local_irq_enable();
 336                 set_thread_flag(TIF_POLLING_NRFLAG);
 337                 /* Compute time (ticks) that we were actually asleep */
 338                 sleep_ticks =
 339                     ticks_elapsed(t1, t2) - cx->latency_ticks - C2_OVERHEAD;
 340                 break;
 341
 342         case ACPI_STATE_C3:
 343
 344                 if (pr->flags.bm_check) {
 345                         if (atomic_inc_return(&c3_cpu_count) ==
 346                             num_online_cpus()) {
 347                                 /*
 348                                  * All CPUs are trying to go to C3
 349                                  * Disable bus master arbitration
 350                                  */
 351                                 acpi_set_register(ACPI_BITREG_ARB_DISABLE, 1,
 352                                                   ACPI_MTX_DO_NOT_LOCK);
 353                         }
 354                 } else {
 355                         /* SMP with no shared cache... Invalidate cache  */
 356                         ACPI_FLUSH_CPU_CACHE();
 357                 }
 358
 359                 /* Get start time (ticks) */
 360                 t1 = inl(acpi_fadt.xpm_tmr_blk.address);
 361                 /* Invoke C3 */
 362                 inb(cx->address);
 363                 /* Dummy op - must do something useless after P_LVL3 read */
 364                 t2 = inl(acpi_fadt.xpm_tmr_blk.address);
 365                 /* Get end time (ticks) */
 366                 t2 = inl(acpi_fadt.xpm_tmr_blk.address);
 367                 if (pr->flags.bm_check) {
 368                         /* Enable bus master arbitration */
 369                         atomic_dec(&c3_cpu_count);
 370                         acpi_set_register(ACPI_BITREG_ARB_DISABLE, 0,
 371                                           ACPI_MTX_DO_NOT_LOCK);
 372                 }
 373
 374                 /* Re-enable interrupts */
 375                 local_irq_enable();
 376                 set_thread_flag(TIF_POLLING_NRFLAG);
 377                 /* Compute time (ticks) that we were actually asleep */
 378                 sleep_ticks =
 379                     ticks_elapsed(t1, t2) - cx->latency_ticks - C3_OVERHEAD;
 380                 break;
 381
 382         default:
 383                 local_irq_enable();
 384                 return;
 385         }
 386
 387         next_state = pr->power.state;
 388
 389         /*
 390          * Promotion?
 391          * ----------
 392          * Track the number of longs (time asleep is greater than threshold)
 393          * and promote when the count threshold is reached.  Note that bus
 394          * mastering activity may prevent promotions.
 395          * Do not promote above max_cstate.
 396          */
 397         if (cx->promotion.state &&
 398             ((cx->promotion.state - pr->power.states) <= max_cstate)) {
 399                 if (sleep_ticks > cx->promotion.threshold.ticks) {
 400                         cx->promotion.count++;
 401                         cx->demotion.count = 0;
 402                         if (cx->promotion.count >=
 403                             cx->promotion.threshold.count) {
 404                                 if (pr->flags.bm_check) {
 405                                         if (!
 406                                             (pr->power.bm_activity & cx->
 407                                              promotion.threshold.bm)) {
 408                                                 next_state =
 409                                                     cx->promotion.state;
 410                                                 goto end;
 411                                         }
 412                                 } else {
 413                                         next_state = cx->promotion.state;
 414                                         goto end;
 415                                 }
 416                         }
 417                 }
 418         }
 419
 420         /*
 421          * Demotion?
 422          * ---------
 423          * Track the number of shorts (time asleep is less than time threshold)
 424          * and demote when the usage threshold is reached.
 425          */
 426         if (cx->demotion.state) {
 427                 if (sleep_ticks < cx->demotion.threshold.ticks) {
 428                         cx->demotion.count++;
 429                         cx->promotion.count = 0;
 430                         if (cx->demotion.count >= cx->demotion.threshold.count) {
 431                                 next_state = cx->demotion.state;
 432                                 goto end;
 433                         }
 434                 }
 435         }
 436
 437       end:
 438         /*
 439          * Demote if current state exceeds max_cstate
 440          */
 441         if ((pr->power.state - pr->power.states) > max_cstate) {
 442                 if (cx->demotion.state)
 443                         next_state = cx->demotion.state;
 444         }
 445
 446         /*
 447          * New Cx State?
 448          * -------------
 449          * If we're going to start using a new Cx state we must clean up
 450          * from the previous and prepare to use the new.
 451          */
 452         if (next_state != pr->power.state)
 453                 acpi_processor_power_activate(pr, next_state);
 454 }
 455
 456 static int acpi_processor_set_power_policy(struct acpi_processor *pr)
 457 {
 458         unsigned int i;
 459         unsigned int state_is_set = 0;
 460         struct acpi_processor_cx *lower = NULL;
 461         struct acpi_processor_cx *higher = NULL;
 462         struct acpi_processor_cx *cx;
 463
 464         ACPI_FUNCTION_TRACE("acpi_processor_set_power_policy");
 465
 466         if (!pr)
 467                 return_VALUE(-EINVAL);
 468
 469         /*
 470          * This function sets the default Cx state policy (OS idle handler).
 471          * Our scheme is to promote quickly to C2 but more conservatively
 472          * to C3.  We're favoring C2  for its characteristics of low latency
 473          * (quick response), good power savings, and ability to allow bus
 474          * mastering activity.  Note that the Cx state policy is completely
 475          * customizable and can be altered dynamically.
 476          */
 477
 478         /* startup state */
 479         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
 480                 cx = &pr->power.states[i];
 481                 if (!cx->valid)
 482                         continue;
 483
 484                 if (!state_is_set)
 485                         pr->power.state = cx;
 486                 state_is_set++;
 487                 break;
 488         }
 489
 490         if (!state_is_set)
 491                 return_VALUE(-ENODEV);
 492
 493         /* demotion */
 494         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
 495                 cx = &pr->power.states[i];
 496                 if (!cx->valid)
 497                         continue;
 498
 499                 if (lower) {
 500                         cx->demotion.state = lower;
 501                         cx->demotion.threshold.ticks = cx->latency_ticks;
 502                         cx->demotion.threshold.count = 1;
 503                         if (cx->type == ACPI_STATE_C3)
 504                                 cx->demotion.threshold.bm = bm_history;
 505                 }
 506
 507                 lower = cx;
 508         }
 509
 510         /* promotion */
 511         for (i = (ACPI_PROCESSOR_MAX_POWER - 1); i > 0; i--) {
 512                 cx = &pr->power.states[i];
 513                 if (!cx->valid)
 514                         continue;
 515
 516                 if (higher) {
 517                         cx->promotion.state = higher;
 518                         cx->promotion.threshold.ticks = cx->latency_ticks;
 519                         if (cx->type >= ACPI_STATE_C2)
 520                                 cx->promotion.threshold.count = 4;
 521                         else
 522                                 cx->promotion.threshold.count = 10;
 523                         if (higher->type == ACPI_STATE_C3)
 524                                 cx->promotion.threshold.bm = bm_history;
 525                 }
 526
 527                 higher = cx;
 528         }
 529
 530         return_VALUE(0);
 531 }
 532
 533 static int acpi_processor_get_power_info_fadt(struct acpi_processor *pr)
 534 {
 535         ACPI_FUNCTION_TRACE("acpi_processor_get_power_info_fadt");
 536
 537         if (!pr)
 538                 return_VALUE(-EINVAL);
 539
 540         if (!pr->pblk)
 541                 return_VALUE(-ENODEV);
 542
 543         memset(pr->power.states, 0, sizeof(pr->power.states));
 544
 545         /* if info is obtained from pblk/fadt, type equals state */
 546         pr->power.states[ACPI_STATE_C1].type = ACPI_STATE_C1;
 547         pr->power.states[ACPI_STATE_C2].type = ACPI_STATE_C2;
 548         pr->power.states[ACPI_STATE_C3].type = ACPI_STATE_C3;
 549
 550         /* the C0 state only exists as a filler in our array,
 551          * and all processors need to support C1 */
 552         pr->power.states[ACPI_STATE_C0].valid = 1;
 553         pr->power.states[ACPI_STATE_C1].valid = 1;
 554
 555 #ifndef CONFIG_HOTPLUG_CPU
 556         /*
 557          * Check for P_LVL2_UP flag before entering C2 and above on
 558          * an SMP system.
 559          */
 560         if ((num_online_cpus() > 1) && acpi_fadt.plvl2_up)
 561                 return_VALUE(-ENODEV);
 562 #endif
 563
 564         /* determine C2 and C3 address from pblk */
 565         pr->power.states[ACPI_STATE_C2].address = pr->pblk + 4;
 566         pr->power.states[ACPI_STATE_C3].address = pr->pblk + 5;
 567
 568         /* determine latencies from FADT */
 569         pr->power.states[ACPI_STATE_C2].latency = acpi_fadt.plvl2_lat;
 570         pr->power.states[ACPI_STATE_C3].latency = acpi_fadt.plvl3_lat;
 571
 572         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 573                           "lvl2[0x%08x] lvl3[0x%08x]\n",
 574                           pr->power.states[ACPI_STATE_C2].address,
 575                           pr->power.states[ACPI_STATE_C3].address));
 576
 577         return_VALUE(0);
 578 }
 579
 580 static int acpi_processor_get_power_info_default_c1(struct acpi_processor *pr)
 581 {
 582         ACPI_FUNCTION_TRACE("acpi_processor_get_power_info_default_c1");
 583
 584         memset(pr->power.states, 0, sizeof(pr->power.states));
 585
 586         /* if info is obtained from pblk/fadt, type equals state */
 587         pr->power.states[ACPI_STATE_C1].type = ACPI_STATE_C1;
 588         pr->power.states[ACPI_STATE_C2].type = ACPI_STATE_C2;
 589         pr->power.states[ACPI_STATE_C3].type = ACPI_STATE_C3;
 590
 591         /* the C0 state only exists as a filler in our array,
 592          * and all processors need to support C1 */
 593         pr->power.states[ACPI_STATE_C0].valid = 1;
 594         pr->power.states[ACPI_STATE_C1].valid = 1;
 595
 596         return_VALUE(0);
 597 }
 598
 599 static int acpi_processor_get_power_info_cst(struct acpi_processor *pr)
 600 {
 601         acpi_status status = 0;
 602         acpi_integer count;
 603         int i;
 604         struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
 605         union acpi_object *cst;
 606
 607         ACPI_FUNCTION_TRACE("acpi_processor_get_power_info_cst");
 608
 609         if (nocst)
 610                 return_VALUE(-ENODEV);
 611
 612         pr->power.count = 0;
 613         for (i = 0; i < ACPI_PROCESSOR_MAX_POWER; i++)
 614                 memset(&(pr->power.states[i]), 0,
 615                        sizeof(struct acpi_processor_cx));
 616
 617         status = acpi_evaluate_object(pr->handle, "_CST", NULL, &buffer);
 618         if (ACPI_FAILURE(status)) {
 619                 ACPI_DEBUG_PRINT((ACPI_DB_INFO, "No _CST, giving up\n"));
 620                 return_VALUE(-ENODEV);
 621         }
 622
 623         cst = (union acpi_object *)buffer.pointer;
 624
 625         /* There must be at least 2 elements */
 626         if (!cst || (cst->type != ACPI_TYPE_PACKAGE) || cst->package.count < 2) {
 627                 ACPI_DEBUG_PRINT((ACPI_DB_ERROR,
 628                                   "not enough elements in _CST\n"));
 629                 status = -EFAULT;
 630                 goto end;
 631         }
 632
 633         count = cst->package.elements[0].integer.value;
 634
 635         /* Validate number of power states. */
 636         if (count < 1 || count != cst->package.count - 1) {
 637                 ACPI_DEBUG_PRINT((ACPI_DB_ERROR,
 638                                   "count given by _CST is not valid\n"));
 639                 status = -EFAULT;
 640                 goto end;
 641         }
 642
 643         /* We support up to ACPI_PROCESSOR_MAX_POWER. */
 644         if (count > ACPI_PROCESSOR_MAX_POWER) {
 645                 printk(KERN_WARNING
 646                        "Limiting number of power states to max (%d)\n",
 647                        ACPI_PROCESSOR_MAX_POWER);
 648                 printk(KERN_WARNING
 649                        "Please increase ACPI_PROCESSOR_MAX_POWER if needed.\n");
 650                 count = ACPI_PROCESSOR_MAX_POWER;
 651         }
 652
 653         /* Tell driver that at least _CST is supported. */
 654         pr->flags.has_cst = 1;
 655
 656         for (i = 1; i <= count; i++) {
 657                 union acpi_object *element;
 658                 union acpi_object *obj;
 659                 struct acpi_power_register *reg;
 660                 struct acpi_processor_cx cx;
 661
 662                 memset(&cx, 0, sizeof(cx));
 663
 664                 element = (union acpi_object *)&(cst->package.elements[i]);
 665                 if (element->type != ACPI_TYPE_PACKAGE)
 666                         continue;
 667
 668                 if (element->package.count != 4)
 669                         continue;
 670
 671                 obj = (union acpi_object *)&(element->package.elements[0]);
 672
 673                 if (obj->type != ACPI_TYPE_BUFFER)
 674                         continue;
 675
 676                 reg = (struct acpi_power_register *)obj->buffer.pointer;
 677
 678                 if (reg->space_id != ACPI_ADR_SPACE_SYSTEM_IO &&
 679                     (reg->space_id != ACPI_ADR_SPACE_FIXED_HARDWARE))
 680                         continue;
 681
 682                 cx.address = (reg->space_id == ACPI_ADR_SPACE_FIXED_HARDWARE) ?
 683                     0 : reg->address;
 684
 685                 /* There should be an easy way to extract an integer... */
 686                 obj = (union acpi_object *)&(element->package.elements[1]);
 687                 if (obj->type != ACPI_TYPE_INTEGER)
 688                         continue;
 689
 690                 cx.type = obj->integer.value;
 691
 692                 if ((cx.type != ACPI_STATE_C1) &&
 693                     (reg->space_id != ACPI_ADR_SPACE_SYSTEM_IO))
 694                         continue;
 695
 696                 if ((cx.type < ACPI_STATE_C1) || (cx.type > ACPI_STATE_C3))
 697                         continue;
 698
 699                 obj = (union acpi_object *)&(element->package.elements[2]);
 700                 if (obj->type != ACPI_TYPE_INTEGER)
 701                         continue;
 702
 703                 cx.latency = obj->integer.value;
 704
 705                 obj = (union acpi_object *)&(element->package.elements[3]);
 706                 if (obj->type != ACPI_TYPE_INTEGER)
 707                         continue;
 708
 709                 cx.power = obj->integer.value;
 710
 711                 (pr->power.count)++;
 712                 memcpy(&(pr->power.states[pr->power.count]), &cx, sizeof(cx));
 713         }
 714
 715         ACPI_DEBUG_PRINT((ACPI_DB_INFO, "Found %d power states\n",
 716                           pr->power.count));
 717
 718         /* Validate number of power states discovered */
 719         if (pr->power.count < 2)
 720                 status = -EFAULT;
 721
 722       end:
 723         acpi_os_free(buffer.pointer);
 724
 725         return_VALUE(status);
 726 }
 727
 728 static void acpi_processor_power_verify_c2(struct acpi_processor_cx *cx)
 729 {
 730         ACPI_FUNCTION_TRACE("acpi_processor_get_power_verify_c2");
 731
 732         if (!cx->address)
 733                 return_VOID;
 734
 735         /*
 736          * C2 latency must be less than or equal to 100
 737          * microseconds.
 738          */
 739         else if (cx->latency > ACPI_PROCESSOR_MAX_C2_LATENCY) {
 740                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 741                                   "latency too large [%d]\n", cx->latency));
 742                 return_VOID;
 743         }
 744
 745         /*
 746          * Otherwise we've met all of our C2 requirements.
 747          * Normalize the C2 latency to expidite policy
 748          */
 749         cx->valid = 1;
 750         cx->latency_ticks = US_TO_PM_TIMER_TICKS(cx->latency);
 751
 752         return_VOID;
 753 }
 754
 755 static void acpi_processor_power_verify_c3(struct acpi_processor *pr,
 756                                            struct acpi_processor_cx *cx)
 757 {
 758         static int bm_check_flag;
 759
 760         ACPI_FUNCTION_TRACE("acpi_processor_get_power_verify_c3");
 761
 762         if (!cx->address)
 763                 return_VOID;
 764
 765         /*
 766          * C3 latency must be less than or equal to 1000
 767          * microseconds.
 768          */
 769         else if (cx->latency > ACPI_PROCESSOR_MAX_C3_LATENCY) {
 770                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 771                                   "latency too large [%d]\n", cx->latency));
 772                 return_VOID;
 773         }
 774
 775         /*
 776          * PIIX4 Erratum #18: We don't support C3 when Type-F (fast)
 777          * DMA transfers are used by any ISA device to avoid livelock.
 778          * Note that we could disable Type-F DMA (as recommended by
 779          * the erratum), but this is known to disrupt certain ISA
 780          * devices thus we take the conservative approach.
 781          */
 782         else if (errata.piix4.fdma) {
 783                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 784                                   "C3 not supported on PIIX4 with Type-F DMA\n"));
 785                 return_VOID;
 786         }
 787
 788         /* All the logic here assumes flags.bm_check is same across all CPUs */
 789         if (!bm_check_flag) {
 790                 /* Determine whether bm_check is needed based on CPU  */
 791                 acpi_processor_power_init_bm_check(&(pr->flags), pr->id);
 792                 bm_check_flag = pr->flags.bm_check;
 793         } else {
 794                 pr->flags.bm_check = bm_check_flag;
 795         }
 796
 797         if (pr->flags.bm_check) {
 798                 /* bus mastering control is necessary */
 799                 if (!pr->flags.bm_control) {
 800                         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 801                                           "C3 support requires bus mastering control\n"));
 802                         return_VOID;
 803                 }
 804         } else {
 805                 /*
 806                  * WBINVD should be set in fadt, for C3 state to be
 807                  * supported on when bm_check is not required.
 808                  */
 809                 if (acpi_fadt.wb_invd != 1) {
 810                         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 811                                           "Cache invalidation should work properly"
 812                                           " for C3 to be enabled on SMP systems\n"));
 813                         return_VOID;
 814                 }
 815                 acpi_set_register(ACPI_BITREG_BUS_MASTER_RLD,
 816                                   0, ACPI_MTX_DO_NOT_LOCK);
 817         }
 818
 819         /*
 820          * Otherwise we've met all of our C3 requirements.
 821          * Normalize the C3 latency to expidite policy.  Enable
 822          * checking of bus mastering status (bm_check) so we can
 823          * use this in our C3 policy
 824          */
 825         cx->valid = 1;
 826         cx->latency_ticks = US_TO_PM_TIMER_TICKS(cx->latency);
 827
 828         return_VOID;
 829 }
 830
 831 static int acpi_processor_power_verify(struct acpi_processor *pr)
 832 {
 833         unsigned int i;
 834         unsigned int working = 0;
 835
 836         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
 837                 struct acpi_processor_cx *cx = &pr->power.states[i];
 838
 839                 switch (cx->type) {
 840                 case ACPI_STATE_C1:
 841                         cx->valid = 1;
 842                         break;
 843
 844                 case ACPI_STATE_C2:
 845                         acpi_processor_power_verify_c2(cx);
 846                         break;
 847
 848                 case ACPI_STATE_C3:
 849                         acpi_processor_power_verify_c3(pr, cx);
 850                         break;
 851                 }
 852
 853                 if (cx->valid)
 854                         working++;
 855         }
 856
 857         return (working);
 858 }
 859
 860 static int acpi_processor_get_power_info(struct acpi_processor *pr)
 861 {
 862         unsigned int i;
 863         int result;
 864
 865         ACPI_FUNCTION_TRACE("acpi_processor_get_power_info");
 866
 867         /* NOTE: the idle thread may not be running while calling
 868          * this function */
 869
 870         result = acpi_processor_get_power_info_cst(pr);
 871         if (result == -ENODEV)
 872                 result = acpi_processor_get_power_info_fadt(pr);
 873
 874         if ((result) || (acpi_processor_power_verify(pr) < 2))
 875                 result = acpi_processor_get_power_info_default_c1(pr);
 876
 877         /*
 878          * Set Default Policy
 879          * ------------------
 880          * Now that we know which states are supported, set the default
 881          * policy.  Note that this policy can be changed dynamically
 882          * (e.g. encourage deeper sleeps to conserve battery life when
 883          * not on AC).
 884          */
 885         result = acpi_processor_set_power_policy(pr);
 886         if (result)
 887                 return_VALUE(result);
 888
 889         /*
 890          * if one state of type C2 or C3 is available, mark this
 891          * CPU as being "idle manageable"
 892          */
 893         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
 894                 if (pr->power.states[i].valid) {
 895                         pr->power.count = i;
 896                         if (pr->power.states[i].type >= ACPI_STATE_C2)
 897                                 pr->flags.power = 1;
 898                 }
 899         }
 900
 901         return_VALUE(0);
 902 }
 903
 904 int acpi_processor_cst_has_changed(struct acpi_processor *pr)
 905 {
 906         int result = 0;
 907
 908         ACPI_FUNCTION_TRACE("acpi_processor_cst_has_changed");
 909
 910         if (!pr)
 911                 return_VALUE(-EINVAL);
 912
 913         if (nocst) {
 914                 return_VALUE(-ENODEV);
 915         }
 916
 917         if (!pr->flags.power_setup_done)
 918                 return_VALUE(-ENODEV);
 919
 920         /* Fall back to the default idle loop */
 921         pm_idle = pm_idle_save;
 922         synchronize_sched();    /* Relies on interrupts forcing exit from idle. */
 923
 924         pr->flags.power = 0;
 925         result = acpi_processor_get_power_info(pr);
 926         if ((pr->flags.power == 1) && (pr->flags.power_setup_done))
 927                 pm_idle = acpi_processor_idle;
 928
 929         return_VALUE(result);
 930 }
 931
 932 /* proc interface */
 933
 934 static int acpi_processor_power_seq_show(struct seq_file *seq, void *offset)
 935 {
 936         struct acpi_processor *pr = (struct acpi_processor *)seq->private;
 937         unsigned int i;
 938
 939         ACPI_FUNCTION_TRACE("acpi_processor_power_seq_show");
 940
 941         if (!pr)
 942                 goto end;
 943
 944         seq_printf(seq, "active state:            C%zd\n"
 945                    "max_cstate:              C%d\n"
 946                    "bus master activity:     %08x\n",
 947                    pr->power.state ? pr->power.state - pr->power.states : 0,
 948                    max_cstate, (unsigned)pr->power.bm_activity);
 949
 950         seq_puts(seq, "states:\n");
 951
 952         for (i = 1; i <= pr->power.count; i++) {
 953                 seq_printf(seq, "   %cC%d:                  ",
 954                            (&pr->power.states[i] ==
 955                             pr->power.state ? '*' : ' '), i);
 956
 957                 if (!pr->power.states[i].valid) {
 958                         seq_puts(seq, "<not supported>\n");
 959                         continue;
 960                 }
 961
 962                 switch (pr->power.states[i].type) {
 963                 case ACPI_STATE_C1:
 964                         seq_printf(seq, "type[C1] ");
 965                         break;
 966                 case ACPI_STATE_C2:
 967                         seq_printf(seq, "type[C2] ");
 968                         break;
 969                 case ACPI_STATE_C3:
 970                         seq_printf(seq, "type[C3] ");
 971                         break;
 972                 default:
 973                         seq_printf(seq, "type[--] ");
 974                         break;
 975                 }
 976
 977                 if (pr->power.states[i].promotion.state)
 978                         seq_printf(seq, "promotion[C%zd] ",
 979                                    (pr->power.states[i].promotion.state -
 980                                     pr->power.states));
 981                 else
 982                         seq_puts(seq, "promotion[--] ");
 983
 984                 if (pr->power.states[i].demotion.state)
 985                         seq_printf(seq, "demotion[C%zd] ",
 986                                    (pr->power.states[i].demotion.state -
 987                                     pr->power.states));
 988                 else
 989                         seq_puts(seq, "demotion[--] ");
 990
 991                 seq_printf(seq, "latency[%03d] usage[%08d]\n",
 992                            pr->power.states[i].latency,
 993                            pr->power.states[i].usage);
 994         }
 995
 996       end:
 997         return_VALUE(0);
 998 }
 999
1000 static int acpi_processor_power_open_fs(struct inode *inode, struct file *file)
1001 {
1002         return single_open(file, acpi_processor_power_seq_show,
1003                            PDE(inode)->data);
1004 }
1005
1006 static struct file_operations acpi_processor_power_fops = {
1007         .open = acpi_processor_power_open_fs,
1008         .read = seq_read,
1009         .llseek = seq_lseek,
1010         .release = single_release,
1011 };
1012
1013 int acpi_processor_power_init(struct acpi_processor *pr,
1014                               struct acpi_device *device)
1015 {
1016         acpi_status status = 0;
1017         static int first_run = 0;
1018         struct proc_dir_entry *entry = NULL;
1019         unsigned int i;
1020
1021         ACPI_FUNCTION_TRACE("acpi_processor_power_init");
1022
1023         if (!first_run) {
1024                 dmi_check_system(processor_power_dmi_table);
1025                 if (max_cstate < ACPI_C_STATES_MAX)
1026                         printk(KERN_NOTICE
1027                                "ACPI: processor limited to max C-state %d\n",
1028                                max_cstate);
1029                 first_run++;
1030         }
1031
1032         if (!pr)
1033                 return_VALUE(-EINVAL);
1034
1035         if (acpi_fadt.cst_cnt && !nocst) {
1036                 status =
1037                     acpi_os_write_port(acpi_fadt.smi_cmd, acpi_fadt.cst_cnt, 8);
1038                 if (ACPI_FAILURE(status)) {
1039                         ACPI_DEBUG_PRINT((ACPI_DB_ERROR,
1040                                           "Notifying BIOS of _CST ability failed\n"));
1041                 }
1042         }
1043
1044         acpi_processor_power_init_pdc(&(pr->power), pr->id);
1045         acpi_processor_set_pdc(pr, pr->power.pdc);
1046         acpi_processor_get_power_info(pr);
1047
1048         /*
1049          * Install the idle handler if processor power management is supported.
1050          * Note that we use previously set idle handler will be used on
1051          * platforms that only support C1.
1052          */
1053         if ((pr->flags.power) && (!boot_option_idle_override)) {
1054                 printk(KERN_INFO PREFIX "CPU%d (power states:", pr->id);
1055                 for (i = 1; i <= pr->power.count; i++)
1056                         if (pr->power.states[i].valid)
1057                                 printk(" C%d[C%d]", i,
1058                                        pr->power.states[i].type);
1059                 printk(")\n");
1060
1061                 if (pr->id == 0) {
1062                         pm_idle_save = pm_idle;
1063                         pm_idle = acpi_processor_idle;
1064                 }
1065         }
1066
1067         /* 'power' [R] */
1068         entry = create_proc_entry(ACPI_PROCESSOR_FILE_POWER,
1069                                   S_IRUGO, acpi_device_dir(device));
1070         if (!entry)
1071                 ACPI_DEBUG_PRINT((ACPI_DB_ERROR,
1072                                   "Unable to create '%s' fs entry\n",
1073                                   ACPI_PROCESSOR_FILE_POWER));
1074         else {
1075                 entry->proc_fops = &acpi_processor_power_fops;
1076                 entry->data = acpi_driver_data(device);
1077                 entry->owner = THIS_MODULE;
1078         }
1079
1080         pr->flags.power_setup_done = 1;
1081
1082         return_VALUE(0);
1083 }
1084
1085 int acpi_processor_power_exit(struct acpi_processor *pr,
1086                               struct acpi_device *device)
1087 {
1088         ACPI_FUNCTION_TRACE("acpi_processor_power_exit");
1089
1090         pr->flags.power_setup_done = 0;
1091
1092         if (acpi_device_dir(device))
1093                 remove_proc_entry(ACPI_PROCESSOR_FILE_POWER,
1094                                   acpi_device_dir(device));
1095
1096         /* Unregister the idle handler when processor #0 is removed. */
1097         if (pr->id == 0) {
1098                 pm_idle = pm_idle_save;
1099
1100                 /*
1101                  * We are about to unload the current idle thread pm callback
1102                  * (pm_idle), Wait for all processors to update cached/local
1103                  * copies of pm_idle before proceeding.
1104                  */
1105                 cpu_idle_wait();
1106         }
1107
1108         return_VALUE(0);
1109 }