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) __read_mostly;
  58 module_param(max_cstate, uint, 0644);
  59
  60 static unsigned int nocst __read_mostly;
  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 __read_mostly =
  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 /* Actually this shouldn't be __cpuinitdata, would be better to fix the
  98    callers to only run once -AK */
  99 static struct dmi_system_id __cpuinitdata processor_power_dmi_table[] = {
 100         { set_max_cstate, "IBM ThinkPad R40e", {
 101           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 102           DMI_MATCH(DMI_BIOS_VERSION,"1SET60WW")}, (void *)1},
 103         { set_max_cstate, "IBM ThinkPad R40e", {
 104           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 105           DMI_MATCH(DMI_BIOS_VERSION,"1SET43WW") }, (void*)1},
 106         { set_max_cstate, "IBM ThinkPad R40e", {
 107           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 108           DMI_MATCH(DMI_BIOS_VERSION,"1SET45WW") }, (void*)1},
 109         { set_max_cstate, "IBM ThinkPad R40e", {
 110           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 111           DMI_MATCH(DMI_BIOS_VERSION,"1SET47WW") }, (void*)1},
 112         { set_max_cstate, "IBM ThinkPad R40e", {
 113           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 114           DMI_MATCH(DMI_BIOS_VERSION,"1SET50WW") }, (void*)1},
 115         { set_max_cstate, "IBM ThinkPad R40e", {
 116           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 117           DMI_MATCH(DMI_BIOS_VERSION,"1SET52WW") }, (void*)1},
 118         { set_max_cstate, "IBM ThinkPad R40e", {
 119           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 120           DMI_MATCH(DMI_BIOS_VERSION,"1SET55WW") }, (void*)1},
 121         { set_max_cstate, "IBM ThinkPad R40e", {
 122           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 123           DMI_MATCH(DMI_BIOS_VERSION,"1SET56WW") }, (void*)1},
 124         { set_max_cstate, "IBM ThinkPad R40e", {
 125           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 126           DMI_MATCH(DMI_BIOS_VERSION,"1SET59WW") }, (void*)1},
 127         { set_max_cstate, "IBM ThinkPad R40e", {
 128           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 129           DMI_MATCH(DMI_BIOS_VERSION,"1SET60WW") }, (void*)1},
 130         { set_max_cstate, "IBM ThinkPad R40e", {
 131           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 132           DMI_MATCH(DMI_BIOS_VERSION,"1SET61WW") }, (void*)1},
 133         { set_max_cstate, "IBM ThinkPad R40e", {
 134           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 135           DMI_MATCH(DMI_BIOS_VERSION,"1SET62WW") }, (void*)1},
 136         { set_max_cstate, "IBM ThinkPad R40e", {
 137           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 138           DMI_MATCH(DMI_BIOS_VERSION,"1SET64WW") }, (void*)1},
 139         { set_max_cstate, "IBM ThinkPad R40e", {
 140           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 141           DMI_MATCH(DMI_BIOS_VERSION,"1SET65WW") }, (void*)1},
 142         { set_max_cstate, "IBM ThinkPad R40e", {
 143           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 144           DMI_MATCH(DMI_BIOS_VERSION,"1SET68WW") }, (void*)1},
 145         { set_max_cstate, "Medion 41700", {
 146           DMI_MATCH(DMI_BIOS_VENDOR,"Phoenix Technologies LTD"),
 147           DMI_MATCH(DMI_BIOS_VERSION,"R01-A1J")}, (void *)1},
 148         { set_max_cstate, "Clevo 5600D", {
 149           DMI_MATCH(DMI_BIOS_VENDOR,"Phoenix Technologies LTD"),
 150           DMI_MATCH(DMI_BIOS_VERSION,"SHE845M0.86C.0013.D.0302131307")},
 151          (void *)2},
 152         {},
 153 };
 154
 155 static inline u32 ticks_elapsed(u32 t1, u32 t2)
 156 {
 157         if (t2 >= t1)
 158                 return (t2 - t1);
 159         else if (!acpi_fadt.tmr_val_ext)
 160                 return (((0x00FFFFFF - t1) + t2) & 0x00FFFFFF);
 161         else
 162                 return ((0xFFFFFFFF - t1) + t2);
 163 }
 164
 165 static void
 166 acpi_processor_power_activate(struct acpi_processor *pr,
 167                               struct acpi_processor_cx *new)
 168 {
 169         struct acpi_processor_cx *old;
 170
 171         if (!pr || !new)
 172                 return;
 173
 174         old = pr->power.state;
 175
 176         if (old)
 177                 old->promotion.count = 0;
 178         new->demotion.count = 0;
 179
 180         /* Cleanup from old state. */
 181         if (old) {
 182                 switch (old->type) {
 183                 case ACPI_STATE_C3:
 184                         /* Disable bus master reload */
 185                         if (new->type != ACPI_STATE_C3 && pr->flags.bm_check)
 186                                 acpi_set_register(ACPI_BITREG_BUS_MASTER_RLD, 0,
 187                                                   ACPI_MTX_DO_NOT_LOCK);
 188                         break;
 189                 }
 190         }
 191
 192         /* Prepare to use new state. */
 193         switch (new->type) {
 194         case ACPI_STATE_C3:
 195                 /* Enable bus master reload */
 196                 if (old->type != ACPI_STATE_C3 && pr->flags.bm_check)
 197                         acpi_set_register(ACPI_BITREG_BUS_MASTER_RLD, 1,
 198                                           ACPI_MTX_DO_NOT_LOCK);
 199                 break;
 200         }
 201
 202         pr->power.state = new;
 203
 204         return;
 205 }
 206
 207 static void acpi_safe_halt(void)
 208 {
 209         current_thread_info()->status &= ~TS_POLLING;
 210         smp_mb__after_clear_bit();
 211         if (!need_resched())
 212                 safe_halt();
 213         current_thread_info()->status |= TS_POLLING;
 214 }
 215
 216 static atomic_t c3_cpu_count;
 217
 218 static void acpi_processor_idle(void)
 219 {
 220         struct acpi_processor *pr = NULL;
 221         struct acpi_processor_cx *cx = NULL;
 222         struct acpi_processor_cx *next_state = NULL;
 223         int sleep_ticks = 0;
 224         u32 t1, t2 = 0;
 225
 226         pr = processors[smp_processor_id()];
 227         if (!pr)
 228                 return;
 229
 230         /*
 231          * Interrupts must be disabled during bus mastering calculations and
 232          * for C2/C3 transitions.
 233          */
 234         local_irq_disable();
 235
 236         /*
 237          * Check whether we truly need to go idle, or should
 238          * reschedule:
 239          */
 240         if (unlikely(need_resched())) {
 241                 local_irq_enable();
 242                 return;
 243         }
 244
 245         cx = pr->power.state;
 246         if (!cx) {
 247                 if (pm_idle_save)
 248                         pm_idle_save();
 249                 else
 250                         acpi_safe_halt();
 251                 return;
 252         }
 253
 254         /*
 255          * Check BM Activity
 256          * -----------------
 257          * Check for bus mastering activity (if required), record, and check
 258          * for demotion.
 259          */
 260         if (pr->flags.bm_check) {
 261                 u32 bm_status = 0;
 262                 unsigned long diff = jiffies - pr->power.bm_check_timestamp;
 263
 264                 if (diff > 32)
 265                         diff = 32;
 266
 267                 while (diff) {
 268                         /* if we didn't get called, assume there was busmaster activity */
 269                         diff--;
 270                         if (diff)
 271                                 pr->power.bm_activity |= 0x1;
 272                         pr->power.bm_activity <<= 1;
 273                 }
 274
 275                 acpi_get_register(ACPI_BITREG_BUS_MASTER_STATUS,
 276                                   &bm_status, ACPI_MTX_DO_NOT_LOCK);
 277                 if (bm_status) {
 278                         pr->power.bm_activity++;
 279                         acpi_set_register(ACPI_BITREG_BUS_MASTER_STATUS,
 280                                           1, ACPI_MTX_DO_NOT_LOCK);
 281                 }
 282                 /*
 283                  * PIIX4 Erratum #18: Note that BM_STS doesn't always reflect
 284                  * the true state of bus mastering activity; forcing us to
 285                  * manually check the BMIDEA bit of each IDE channel.
 286                  */
 287                 else if (errata.piix4.bmisx) {
 288                         if ((inb_p(errata.piix4.bmisx + 0x02) & 0x01)
 289                             || (inb_p(errata.piix4.bmisx + 0x0A) & 0x01))
 290                                 pr->power.bm_activity++;
 291                 }
 292
 293                 pr->power.bm_check_timestamp = jiffies;
 294
 295                 /*
 296                  * Apply bus mastering demotion policy.  Automatically demote
 297                  * to avoid a faulty transition.  Note that the processor
 298                  * won't enter a low-power state during this call (to this
 299                  * funciton) but should upon the next.
 300                  *
 301                  * TBD: A better policy might be to fallback to the demotion
 302                  *      state (use it for this quantum only) istead of
 303                  *      demoting -- and rely on duration as our sole demotion
 304                  *      qualification.  This may, however, introduce DMA
 305                  *      issues (e.g. floppy DMA transfer overrun/underrun).
 306                  */
 307                 if (pr->power.bm_activity & cx->demotion.threshold.bm) {
 308                         local_irq_enable();
 309                         next_state = cx->demotion.state;
 310                         goto end;
 311                 }
 312         }
 313
 314 #ifdef CONFIG_HOTPLUG_CPU
 315         /*
 316          * Check for P_LVL2_UP flag before entering C2 and above on
 317          * an SMP system. We do it here instead of doing it at _CST/P_LVL
 318          * detection phase, to work cleanly with logical CPU hotplug.
 319          */
 320         if ((cx->type != ACPI_STATE_C1) && (num_online_cpus() > 1) &&
 321             !pr->flags.has_cst && !acpi_fadt.plvl2_up)
 322                 cx = &pr->power.states[ACPI_STATE_C1];
 323 #endif
 324
 325         cx->usage++;
 326
 327         /*
 328          * Sleep:
 329          * ------
 330          * Invoke the current Cx state to put the processor to sleep.
 331          */
 332         if (cx->type == ACPI_STATE_C2 || cx->type == ACPI_STATE_C3) {
 333                 current_thread_info()->status &= ~TS_POLLING;
 334                 smp_mb__after_clear_bit();
 335                 if (need_resched()) {
 336                         current_thread_info()->status |= TS_POLLING;
 337                         local_irq_enable();
 338                         return;
 339                 }
 340         }
 341
 342         switch (cx->type) {
 343
 344         case ACPI_STATE_C1:
 345                 /*
 346                  * Invoke C1.
 347                  * Use the appropriate idle routine, the one that would
 348                  * be used without acpi C-states.
 349                  */
 350                 if (pm_idle_save)
 351                         pm_idle_save();
 352                 else
 353                         acpi_safe_halt();
 354
 355                 /*
 356                  * TBD: Can't get time duration while in C1, as resumes
 357                  *      go to an ISR rather than here.  Need to instrument
 358                  *      base interrupt handler.
 359                  */
 360                 sleep_ticks = 0xFFFFFFFF;
 361                 break;
 362
 363         case ACPI_STATE_C2:
 364                 /* Get start time (ticks) */
 365                 t1 = inl(acpi_fadt.xpm_tmr_blk.address);
 366                 /* Invoke C2 */
 367                 inb(cx->address);
 368                 /* Dummy op - must do something useless after P_LVL2 read */
 369                 t2 = inl(acpi_fadt.xpm_tmr_blk.address);
 370                 /* Get end time (ticks) */
 371                 t2 = inl(acpi_fadt.xpm_tmr_blk.address);
 372
 373 #ifdef CONFIG_GENERIC_TIME
 374                 /* TSC halts in C2, so notify users */
 375                 mark_tsc_unstable();
 376 #endif
 377                 /* Re-enable interrupts */
 378                 local_irq_enable();
 379                 current_thread_info()->status |= TS_POLLING;
 380                 /* Compute time (ticks) that we were actually asleep */
 381                 sleep_ticks =
 382                     ticks_elapsed(t1, t2) - cx->latency_ticks - C2_OVERHEAD;
 383                 break;
 384
 385         case ACPI_STATE_C3:
 386
 387                 if (pr->flags.bm_check) {
 388                         if (atomic_inc_return(&c3_cpu_count) ==
 389                             num_online_cpus()) {
 390                                 /*
 391                                  * All CPUs are trying to go to C3
 392                                  * Disable bus master arbitration
 393                                  */
 394                                 acpi_set_register(ACPI_BITREG_ARB_DISABLE, 1,
 395                                                   ACPI_MTX_DO_NOT_LOCK);
 396                         }
 397                 } else {
 398                         /* SMP with no shared cache... Invalidate cache  */
 399                         ACPI_FLUSH_CPU_CACHE();
 400                 }
 401
 402                 /* Get start time (ticks) */
 403                 t1 = inl(acpi_fadt.xpm_tmr_blk.address);
 404                 /* Invoke C3 */
 405                 inb(cx->address);
 406                 /* Dummy op - must do something useless after P_LVL3 read */
 407                 t2 = inl(acpi_fadt.xpm_tmr_blk.address);
 408                 /* Get end time (ticks) */
 409                 t2 = inl(acpi_fadt.xpm_tmr_blk.address);
 410                 if (pr->flags.bm_check) {
 411                         /* Enable bus master arbitration */
 412                         atomic_dec(&c3_cpu_count);
 413                         acpi_set_register(ACPI_BITREG_ARB_DISABLE, 0,
 414                                           ACPI_MTX_DO_NOT_LOCK);
 415                 }
 416
 417 #ifdef CONFIG_GENERIC_TIME
 418                 /* TSC halts in C3, so notify users */
 419                 mark_tsc_unstable();
 420 #endif
 421                 /* Re-enable interrupts */
 422                 local_irq_enable();
 423                 current_thread_info()->status |= TS_POLLING;
 424                 /* Compute time (ticks) that we were actually asleep */
 425                 sleep_ticks =
 426                     ticks_elapsed(t1, t2) - cx->latency_ticks - C3_OVERHEAD;
 427                 break;
 428
 429         default:
 430                 local_irq_enable();
 431                 return;
 432         }
 433
 434         next_state = pr->power.state;
 435
 436 #ifdef CONFIG_HOTPLUG_CPU
 437         /* Don't do promotion/demotion */
 438         if ((cx->type == ACPI_STATE_C1) && (num_online_cpus() > 1) &&
 439             !pr->flags.has_cst && !acpi_fadt.plvl2_up) {
 440                 next_state = cx;
 441                 goto end;
 442         }
 443 #endif
 444
 445         /*
 446          * Promotion?
 447          * ----------
 448          * Track the number of longs (time asleep is greater than threshold)
 449          * and promote when the count threshold is reached.  Note that bus
 450          * mastering activity may prevent promotions.
 451          * Do not promote above max_cstate.
 452          */
 453         if (cx->promotion.state &&
 454             ((cx->promotion.state - pr->power.states) <= max_cstate)) {
 455                 if (sleep_ticks > cx->promotion.threshold.ticks) {
 456                         cx->promotion.count++;
 457                         cx->demotion.count = 0;
 458                         if (cx->promotion.count >=
 459                             cx->promotion.threshold.count) {
 460                                 if (pr->flags.bm_check) {
 461                                         if (!
 462                                             (pr->power.bm_activity & cx->
 463                                              promotion.threshold.bm)) {
 464                                                 next_state =
 465                                                     cx->promotion.state;
 466                                                 goto end;
 467                                         }
 468                                 } else {
 469                                         next_state = cx->promotion.state;
 470                                         goto end;
 471                                 }
 472                         }
 473                 }
 474         }
 475
 476         /*
 477          * Demotion?
 478          * ---------
 479          * Track the number of shorts (time asleep is less than time threshold)
 480          * and demote when the usage threshold is reached.
 481          */
 482         if (cx->demotion.state) {
 483                 if (sleep_ticks < cx->demotion.threshold.ticks) {
 484                         cx->demotion.count++;
 485                         cx->promotion.count = 0;
 486                         if (cx->demotion.count >= cx->demotion.threshold.count) {
 487                                 next_state = cx->demotion.state;
 488                                 goto end;
 489                         }
 490                 }
 491         }
 492
 493       end:
 494         /*
 495          * Demote if current state exceeds max_cstate
 496          */
 497         if ((pr->power.state - pr->power.states) > max_cstate) {
 498                 if (cx->demotion.state)
 499                         next_state = cx->demotion.state;
 500         }
 501
 502         /*
 503          * New Cx State?
 504          * -------------
 505          * If we're going to start using a new Cx state we must clean up
 506          * from the previous and prepare to use the new.
 507          */
 508         if (next_state != pr->power.state)
 509                 acpi_processor_power_activate(pr, next_state);
 510 }
 511
 512 static int acpi_processor_set_power_policy(struct acpi_processor *pr)
 513 {
 514         unsigned int i;
 515         unsigned int state_is_set = 0;
 516         struct acpi_processor_cx *lower = NULL;
 517         struct acpi_processor_cx *higher = NULL;
 518         struct acpi_processor_cx *cx;
 519
 520         ACPI_FUNCTION_TRACE("acpi_processor_set_power_policy");
 521
 522         if (!pr)
 523                 return_VALUE(-EINVAL);
 524
 525         /*
 526          * This function sets the default Cx state policy (OS idle handler).
 527          * Our scheme is to promote quickly to C2 but more conservatively
 528          * to C3.  We're favoring C2  for its characteristics of low latency
 529          * (quick response), good power savings, and ability to allow bus
 530          * mastering activity.  Note that the Cx state policy is completely
 531          * customizable and can be altered dynamically.
 532          */
 533
 534         /* startup state */
 535         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
 536                 cx = &pr->power.states[i];
 537                 if (!cx->valid)
 538                         continue;
 539
 540                 if (!state_is_set)
 541                         pr->power.state = cx;
 542                 state_is_set++;
 543                 break;
 544         }
 545
 546         if (!state_is_set)
 547                 return_VALUE(-ENODEV);
 548
 549         /* demotion */
 550         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
 551                 cx = &pr->power.states[i];
 552                 if (!cx->valid)
 553                         continue;
 554
 555                 if (lower) {
 556                         cx->demotion.state = lower;
 557                         cx->demotion.threshold.ticks = cx->latency_ticks;
 558                         cx->demotion.threshold.count = 1;
 559                         if (cx->type == ACPI_STATE_C3)
 560                                 cx->demotion.threshold.bm = bm_history;
 561                 }
 562
 563                 lower = cx;
 564         }
 565
 566         /* promotion */
 567         for (i = (ACPI_PROCESSOR_MAX_POWER - 1); i > 0; i--) {
 568                 cx = &pr->power.states[i];
 569                 if (!cx->valid)
 570                         continue;
 571
 572                 if (higher) {
 573                         cx->promotion.state = higher;
 574                         cx->promotion.threshold.ticks = cx->latency_ticks;
 575                         if (cx->type >= ACPI_STATE_C2)
 576                                 cx->promotion.threshold.count = 4;
 577                         else
 578                                 cx->promotion.threshold.count = 10;
 579                         if (higher->type == ACPI_STATE_C3)
 580                                 cx->promotion.threshold.bm = bm_history;
 581                 }
 582
 583                 higher = cx;
 584         }
 585
 586         return_VALUE(0);
 587 }
 588
 589 static int acpi_processor_get_power_info_fadt(struct acpi_processor *pr)
 590 {
 591         ACPI_FUNCTION_TRACE("acpi_processor_get_power_info_fadt");
 592
 593         if (!pr)
 594                 return_VALUE(-EINVAL);
 595
 596         if (!pr->pblk)
 597                 return_VALUE(-ENODEV);
 598
 599         /* if info is obtained from pblk/fadt, type equals state */
 600         pr->power.states[ACPI_STATE_C2].type = ACPI_STATE_C2;
 601         pr->power.states[ACPI_STATE_C3].type = ACPI_STATE_C3;
 602
 603 #ifndef CONFIG_HOTPLUG_CPU
 604         /*
 605          * Check for P_LVL2_UP flag before entering C2 and above on
 606          * an SMP system.
 607          */
 608         if ((num_online_cpus() > 1) && !acpi_fadt.plvl2_up)
 609                 return_VALUE(-ENODEV);
 610 #endif
 611
 612         /* determine C2 and C3 address from pblk */
 613         pr->power.states[ACPI_STATE_C2].address = pr->pblk + 4;
 614         pr->power.states[ACPI_STATE_C3].address = pr->pblk + 5;
 615
 616         /* determine latencies from FADT */
 617         pr->power.states[ACPI_STATE_C2].latency = acpi_fadt.plvl2_lat;
 618         pr->power.states[ACPI_STATE_C3].latency = acpi_fadt.plvl3_lat;
 619
 620         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 621                           "lvl2[0x%08x] lvl3[0x%08x]\n",
 622                           pr->power.states[ACPI_STATE_C2].address,
 623                           pr->power.states[ACPI_STATE_C3].address));
 624
 625         return_VALUE(0);
 626 }
 627
 628 static int acpi_processor_get_power_info_default_c1(struct acpi_processor *pr)
 629 {
 630         ACPI_FUNCTION_TRACE("acpi_processor_get_power_info_default_c1");
 631
 632         /* Zero initialize all the C-states info. */
 633         memset(pr->power.states, 0, sizeof(pr->power.states));
 634
 635         /* set the first C-State to C1 */
 636         pr->power.states[ACPI_STATE_C1].type = ACPI_STATE_C1;
 637
 638         /* the C0 state only exists as a filler in our array,
 639          * and all processors need to support C1 */
 640         pr->power.states[ACPI_STATE_C0].valid = 1;
 641         pr->power.states[ACPI_STATE_C1].valid = 1;
 642
 643         return_VALUE(0);
 644 }
 645
 646 static int acpi_processor_get_power_info_cst(struct acpi_processor *pr)
 647 {
 648         acpi_status status = 0;
 649         acpi_integer count;
 650         int current_count;
 651         int i;
 652         struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
 653         union acpi_object *cst;
 654
 655         ACPI_FUNCTION_TRACE("acpi_processor_get_power_info_cst");
 656
 657         if (nocst)
 658                 return_VALUE(-ENODEV);
 659
 660         current_count = 1;
 661
 662         /* Zero initialize C2 onwards and prepare for fresh CST lookup */
 663         for (i = 2; i < ACPI_PROCESSOR_MAX_POWER; i++)
 664                 memset(&(pr->power.states[i]), 0,
 665                                 sizeof(struct acpi_processor_cx));
 666
 667         status = acpi_evaluate_object(pr->handle, "_CST", NULL, &buffer);
 668         if (ACPI_FAILURE(status)) {
 669                 ACPI_DEBUG_PRINT((ACPI_DB_INFO, "No _CST, giving up\n"));
 670                 return_VALUE(-ENODEV);
 671         }
 672
 673         cst = (union acpi_object *)buffer.pointer;
 674
 675         /* There must be at least 2 elements */
 676         if (!cst || (cst->type != ACPI_TYPE_PACKAGE) || cst->package.count < 2) {
 677                 ACPI_DEBUG_PRINT((ACPI_DB_ERROR,
 678                                   "not enough elements in _CST\n"));
 679                 status = -EFAULT;
 680                 goto end;
 681         }
 682
 683         count = cst->package.elements[0].integer.value;
 684
 685         /* Validate number of power states. */
 686         if (count < 1 || count != cst->package.count - 1) {
 687                 ACPI_DEBUG_PRINT((ACPI_DB_ERROR,
 688                                   "count given by _CST is not valid\n"));
 689                 status = -EFAULT;
 690                 goto end;
 691         }
 692
 693         /* Tell driver that at least _CST is supported. */
 694         pr->flags.has_cst = 1;
 695
 696         for (i = 1; i <= count; i++) {
 697                 union acpi_object *element;
 698                 union acpi_object *obj;
 699                 struct acpi_power_register *reg;
 700                 struct acpi_processor_cx cx;
 701
 702                 memset(&cx, 0, sizeof(cx));
 703
 704                 element = (union acpi_object *)&(cst->package.elements[i]);
 705                 if (element->type != ACPI_TYPE_PACKAGE)
 706                         continue;
 707
 708                 if (element->package.count != 4)
 709                         continue;
 710
 711                 obj = (union acpi_object *)&(element->package.elements[0]);
 712
 713                 if (obj->type != ACPI_TYPE_BUFFER)
 714                         continue;
 715
 716                 reg = (struct acpi_power_register *)obj->buffer.pointer;
 717
 718                 if (reg->space_id != ACPI_ADR_SPACE_SYSTEM_IO &&
 719                     (reg->space_id != ACPI_ADR_SPACE_FIXED_HARDWARE))
 720                         continue;
 721
 722                 cx.address = (reg->space_id == ACPI_ADR_SPACE_FIXED_HARDWARE) ?
 723                     0 : reg->address;
 724
 725                 /* There should be an easy way to extract an integer... */
 726                 obj = (union acpi_object *)&(element->package.elements[1]);
 727                 if (obj->type != ACPI_TYPE_INTEGER)
 728                         continue;
 729
 730                 cx.type = obj->integer.value;
 731
 732                 if ((cx.type != ACPI_STATE_C1) &&
 733                     (reg->space_id != ACPI_ADR_SPACE_SYSTEM_IO))
 734                         continue;
 735
 736                 if ((cx.type < ACPI_STATE_C2) || (cx.type > ACPI_STATE_C3))
 737                         continue;
 738
 739                 obj = (union acpi_object *)&(element->package.elements[2]);
 740                 if (obj->type != ACPI_TYPE_INTEGER)
 741                         continue;
 742
 743                 cx.latency = obj->integer.value;
 744
 745                 obj = (union acpi_object *)&(element->package.elements[3]);
 746                 if (obj->type != ACPI_TYPE_INTEGER)
 747                         continue;
 748
 749                 cx.power = obj->integer.value;
 750
 751                 current_count++;
 752                 memcpy(&(pr->power.states[current_count]), &cx, sizeof(cx));
 753
 754                 /*
 755                  * We support total ACPI_PROCESSOR_MAX_POWER - 1
 756                  * (From 1 through ACPI_PROCESSOR_MAX_POWER - 1)
 757                  */
 758                 if (current_count >= (ACPI_PROCESSOR_MAX_POWER - 1)) {
 759                         printk(KERN_WARNING
 760                                "Limiting number of power states to max (%d)\n",
 761                                ACPI_PROCESSOR_MAX_POWER);
 762                         printk(KERN_WARNING
 763                                "Please increase ACPI_PROCESSOR_MAX_POWER if needed.\n");
 764                         break;
 765                 }
 766         }
 767
 768         ACPI_DEBUG_PRINT((ACPI_DB_INFO, "Found %d power states\n",
 769                           current_count));
 770
 771         /* Validate number of power states discovered */
 772         if (current_count < 2)
 773                 status = -EFAULT;
 774
 775       end:
 776         acpi_os_free(buffer.pointer);
 777
 778         return_VALUE(status);
 779 }
 780
 781 static void acpi_processor_power_verify_c2(struct acpi_processor_cx *cx)
 782 {
 783         ACPI_FUNCTION_TRACE("acpi_processor_get_power_verify_c2");
 784
 785         if (!cx->address)
 786                 return_VOID;
 787
 788         /*
 789          * C2 latency must be less than or equal to 100
 790          * microseconds.
 791          */
 792         else if (cx->latency > ACPI_PROCESSOR_MAX_C2_LATENCY) {
 793                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 794                                   "latency too large [%d]\n", cx->latency));
 795                 return_VOID;
 796         }
 797
 798         /*
 799          * Otherwise we've met all of our C2 requirements.
 800          * Normalize the C2 latency to expidite policy
 801          */
 802         cx->valid = 1;
 803         cx->latency_ticks = US_TO_PM_TIMER_TICKS(cx->latency);
 804
 805         return_VOID;
 806 }
 807
 808 static void acpi_processor_power_verify_c3(struct acpi_processor *pr,
 809                                            struct acpi_processor_cx *cx)
 810 {
 811         static int bm_check_flag;
 812
 813         ACPI_FUNCTION_TRACE("acpi_processor_get_power_verify_c3");
 814
 815         if (!cx->address)
 816                 return_VOID;
 817
 818         /*
 819          * C3 latency must be less than or equal to 1000
 820          * microseconds.
 821          */
 822         else if (cx->latency > ACPI_PROCESSOR_MAX_C3_LATENCY) {
 823                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 824                                   "latency too large [%d]\n", cx->latency));
 825                 return_VOID;
 826         }
 827
 828         /*
 829          * PIIX4 Erratum #18: We don't support C3 when Type-F (fast)
 830          * DMA transfers are used by any ISA device to avoid livelock.
 831          * Note that we could disable Type-F DMA (as recommended by
 832          * the erratum), but this is known to disrupt certain ISA
 833          * devices thus we take the conservative approach.
 834          */
 835         else if (errata.piix4.fdma) {
 836                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 837                                   "C3 not supported on PIIX4 with Type-F DMA\n"));
 838                 return_VOID;
 839         }
 840
 841         /* All the logic here assumes flags.bm_check is same across all CPUs */
 842         if (!bm_check_flag) {
 843                 /* Determine whether bm_check is needed based on CPU  */
 844                 acpi_processor_power_init_bm_check(&(pr->flags), pr->id);
 845                 bm_check_flag = pr->flags.bm_check;
 846         } else {
 847                 pr->flags.bm_check = bm_check_flag;
 848         }
 849
 850         if (pr->flags.bm_check) {
 851                 /* bus mastering control is necessary */
 852                 if (!pr->flags.bm_control) {
 853                         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 854                                           "C3 support requires bus mastering control\n"));
 855                         return_VOID;
 856                 }
 857         } else {
 858                 /*
 859                  * WBINVD should be set in fadt, for C3 state to be
 860                  * supported on when bm_check is not required.
 861                  */
 862                 if (acpi_fadt.wb_invd != 1) {
 863                         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 864                                           "Cache invalidation should work properly"
 865                                           " for C3 to be enabled on SMP systems\n"));
 866                         return_VOID;
 867                 }
 868                 acpi_set_register(ACPI_BITREG_BUS_MASTER_RLD,
 869                                   0, ACPI_MTX_DO_NOT_LOCK);
 870         }
 871
 872         /*
 873          * Otherwise we've met all of our C3 requirements.
 874          * Normalize the C3 latency to expidite policy.  Enable
 875          * checking of bus mastering status (bm_check) so we can
 876          * use this in our C3 policy
 877          */
 878         cx->valid = 1;
 879         cx->latency_ticks = US_TO_PM_TIMER_TICKS(cx->latency);
 880
 881         return_VOID;
 882 }
 883
 884 static int acpi_processor_power_verify(struct acpi_processor *pr)
 885 {
 886         unsigned int i;
 887         unsigned int working = 0;
 888
 889 #ifdef ARCH_APICTIMER_STOPS_ON_C3
 890         int timer_broadcast = 0;
 891         cpumask_t mask = cpumask_of_cpu(pr->id);
 892         on_each_cpu(switch_ipi_to_APIC_timer, &mask, 1, 1);
 893 #endif
 894
 895         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
 896                 struct acpi_processor_cx *cx = &pr->power.states[i];
 897
 898                 switch (cx->type) {
 899                 case ACPI_STATE_C1:
 900                         cx->valid = 1;
 901                         break;
 902
 903                 case ACPI_STATE_C2:
 904                         acpi_processor_power_verify_c2(cx);
 905 #ifdef ARCH_APICTIMER_STOPS_ON_C3
 906                         /* Some AMD systems fake C3 as C2, but still
 907                            have timer troubles */
 908                         if (cx->valid &&
 909                                 boot_cpu_data.x86_vendor == X86_VENDOR_AMD)
 910                                 timer_broadcast++;
 911 #endif
 912                         break;
 913
 914                 case ACPI_STATE_C3:
 915                         acpi_processor_power_verify_c3(pr, cx);
 916 #ifdef ARCH_APICTIMER_STOPS_ON_C3
 917                         if (cx->valid)
 918                                 timer_broadcast++;
 919 #endif
 920                         break;
 921                 }
 922
 923                 if (cx->valid)
 924                         working++;
 925         }
 926
 927 #ifdef ARCH_APICTIMER_STOPS_ON_C3
 928         if (timer_broadcast)
 929                 on_each_cpu(switch_APIC_timer_to_ipi, &mask, 1, 1);
 930 #endif
 931
 932         return (working);
 933 }
 934
 935 static int acpi_processor_get_power_info(struct acpi_processor *pr)
 936 {
 937         unsigned int i;
 938         int result;
 939
 940         ACPI_FUNCTION_TRACE("acpi_processor_get_power_info");
 941
 942         /* NOTE: the idle thread may not be running while calling
 943          * this function */
 944
 945         /* Adding C1 state */
 946         acpi_processor_get_power_info_default_c1(pr);
 947         result = acpi_processor_get_power_info_cst(pr);
 948         if (result == -ENODEV)
 949                 acpi_processor_get_power_info_fadt(pr);
 950
 951         pr->power.count = acpi_processor_power_verify(pr);
 952
 953         /*
 954          * Set Default Policy
 955          * ------------------
 956          * Now that we know which states are supported, set the default
 957          * policy.  Note that this policy can be changed dynamically
 958          * (e.g. encourage deeper sleeps to conserve battery life when
 959          * not on AC).
 960          */
 961         result = acpi_processor_set_power_policy(pr);
 962         if (result)
 963                 return_VALUE(result);
 964
 965         /*
 966          * if one state of type C2 or C3 is available, mark this
 967          * CPU as being "idle manageable"
 968          */
 969         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
 970                 if (pr->power.states[i].valid) {
 971                         pr->power.count = i;
 972                         if (pr->power.states[i].type >= ACPI_STATE_C2)
 973                                 pr->flags.power = 1;
 974                 }
 975         }
 976
 977         return_VALUE(0);
 978 }
 979
 980 int acpi_processor_cst_has_changed(struct acpi_processor *pr)
 981 {
 982         int result = 0;
 983
 984         ACPI_FUNCTION_TRACE("acpi_processor_cst_has_changed");
 985
 986         if (!pr)
 987                 return_VALUE(-EINVAL);
 988
 989         if (nocst) {
 990                 return_VALUE(-ENODEV);
 991         }
 992
 993         if (!pr->flags.power_setup_done)
 994                 return_VALUE(-ENODEV);
 995
 996         /* Fall back to the default idle loop */
 997         pm_idle = pm_idle_save;
 998         synchronize_sched();    /* Relies on interrupts forcing exit from idle. */
 999
1000         pr->flags.power = 0;
1001         result = acpi_processor_get_power_info(pr);
1002         if ((pr->flags.power == 1) && (pr->flags.power_setup_done))
1003                 pm_idle = acpi_processor_idle;
1004
1005         return_VALUE(result);
1006 }
1007
1008 /* proc interface */
1009
1010 static int acpi_processor_power_seq_show(struct seq_file *seq, void *offset)
1011 {
1012         struct acpi_processor *pr = (struct acpi_processor *)seq->private;
1013         unsigned int i;
1014
1015         ACPI_FUNCTION_TRACE("acpi_processor_power_seq_show");
1016
1017         if (!pr)
1018                 goto end;
1019
1020         seq_printf(seq, "active state:            C%zd\n"
1021                    "max_cstate:              C%d\n"
1022                    "bus master activity:     %08x\n",
1023                    pr->power.state ? pr->power.state - pr->power.states : 0,
1024                    max_cstate, (unsigned)pr->power.bm_activity);
1025
1026         seq_puts(seq, "states:\n");
1027
1028         for (i = 1; i <= pr->power.count; i++) {
1029                 seq_printf(seq, "   %cC%d:                  ",
1030                            (&pr->power.states[i] ==
1031                             pr->power.state ? '*' : ' '), i);
1032
1033                 if (!pr->power.states[i].valid) {
1034                         seq_puts(seq, "<not supported>\n");
1035                         continue;
1036                 }
1037
1038                 switch (pr->power.states[i].type) {
1039                 case ACPI_STATE_C1:
1040                         seq_printf(seq, "type[C1] ");
1041                         break;
1042                 case ACPI_STATE_C2:
1043                         seq_printf(seq, "type[C2] ");
1044                         break;
1045                 case ACPI_STATE_C3:
1046                         seq_printf(seq, "type[C3] ");
1047                         break;
1048                 default:
1049                         seq_printf(seq, "type[--] ");
1050                         break;
1051                 }
1052
1053                 if (pr->power.states[i].promotion.state)
1054                         seq_printf(seq, "promotion[C%zd] ",
1055                                    (pr->power.states[i].promotion.state -
1056                                     pr->power.states));
1057                 else
1058                         seq_puts(seq, "promotion[--] ");
1059
1060                 if (pr->power.states[i].demotion.state)
1061                         seq_printf(seq, "demotion[C%zd] ",
1062                                    (pr->power.states[i].demotion.state -
1063                                     pr->power.states));
1064                 else
1065                         seq_puts(seq, "demotion[--] ");
1066
1067                 seq_printf(seq, "latency[%03d] usage[%08d]\n",
1068                            pr->power.states[i].latency,
1069                            pr->power.states[i].usage);
1070         }
1071
1072       end:
1073         return_VALUE(0);
1074 }
1075
1076 static int acpi_processor_power_open_fs(struct inode *inode, struct file *file)
1077 {
1078         return single_open(file, acpi_processor_power_seq_show,
1079                            PDE(inode)->data);
1080 }
1081
1082 static struct file_operations acpi_processor_power_fops = {
1083         .open = acpi_processor_power_open_fs,
1084         .read = seq_read,
1085         .llseek = seq_lseek,
1086         .release = single_release,
1087 };
1088
1089 int acpi_processor_power_init(struct acpi_processor *pr,
1090                               struct acpi_device *device)
1091 {
1092         acpi_status status = 0;
1093         static int first_run;
1094         struct proc_dir_entry *entry = NULL;
1095         unsigned int i;
1096
1097         ACPI_FUNCTION_TRACE("acpi_processor_power_init");
1098
1099         if (!first_run) {
1100                 dmi_check_system(processor_power_dmi_table);
1101                 if (max_cstate < ACPI_C_STATES_MAX)
1102                         printk(KERN_NOTICE
1103                                "ACPI: processor limited to max C-state %d\n",
1104                                max_cstate);
1105                 first_run++;
1106         }
1107
1108         if (!pr)
1109                 return_VALUE(-EINVAL);
1110
1111         if (acpi_fadt.cst_cnt && !nocst) {
1112                 status =
1113                     acpi_os_write_port(acpi_fadt.smi_cmd, acpi_fadt.cst_cnt, 8);
1114                 if (ACPI_FAILURE(status)) {
1115                         ACPI_DEBUG_PRINT((ACPI_DB_ERROR,
1116                                           "Notifying BIOS of _CST ability failed\n"));
1117                 }
1118         }
1119
1120         acpi_processor_get_power_info(pr);
1121
1122         /*
1123          * Install the idle handler if processor power management is supported.
1124          * Note that we use previously set idle handler will be used on
1125          * platforms that only support C1.
1126          */
1127         if ((pr->flags.power) && (!boot_option_idle_override)) {
1128                 printk(KERN_INFO PREFIX "CPU%d (power states:", pr->id);
1129                 for (i = 1; i <= pr->power.count; i++)
1130                         if (pr->power.states[i].valid)
1131                                 printk(" C%d[C%d]", i,
1132                                        pr->power.states[i].type);
1133                 printk(")\n");
1134
1135                 if (pr->id == 0) {
1136                         pm_idle_save = pm_idle;
1137                         pm_idle = acpi_processor_idle;
1138                 }
1139         }
1140
1141         /* 'power' [R] */
1142         entry = create_proc_entry(ACPI_PROCESSOR_FILE_POWER,
1143                                   S_IRUGO, acpi_device_dir(device));
1144         if (!entry)
1145                 ACPI_DEBUG_PRINT((ACPI_DB_ERROR,
1146                                   "Unable to create '%s' fs entry\n",
1147                                   ACPI_PROCESSOR_FILE_POWER));
1148         else {
1149                 entry->proc_fops = &acpi_processor_power_fops;
1150                 entry->data = acpi_driver_data(device);
1151                 entry->owner = THIS_MODULE;
1152         }
1153
1154         pr->flags.power_setup_done = 1;
1155
1156         return_VALUE(0);
1157 }
1158
1159 int acpi_processor_power_exit(struct acpi_processor *pr,
1160                               struct acpi_device *device)
1161 {
1162         ACPI_FUNCTION_TRACE("acpi_processor_power_exit");
1163
1164         pr->flags.power_setup_done = 0;
1165
1166         if (acpi_device_dir(device))
1167                 remove_proc_entry(ACPI_PROCESSOR_FILE_POWER,
1168                                   acpi_device_dir(device));
1169
1170         /* Unregister the idle handler when processor #0 is removed. */
1171         if (pr->id == 0) {
1172                 pm_idle = pm_idle_save;
1173
1174                 /*
1175                  * We are about to unload the current idle thread pm callback
1176                  * (pm_idle), Wait for all processors to update cached/local
1177                  * copies of pm_idle before proceeding.
1178                  */
1179                 cpu_idle_wait();
1180         }
1181
1182         return_VALUE(0);
1183 }