Pileus Git - ~andy/linux/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) __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 wait op - must do something useless after P_LVL2 read
 369                    because chipsets cannot guarantee that STPCLK# signal
 370                    gets asserted in time to freeze execution properly. */
 371                 t2 = inl(acpi_fadt.xpm_tmr_blk.address);
 372                 /* Get end time (ticks) */
 373                 t2 = inl(acpi_fadt.xpm_tmr_blk.address);
 374
 375 #ifdef CONFIG_GENERIC_TIME
 376                 /* TSC halts in C2, so notify users */
 377                 mark_tsc_unstable();
 378 #endif
 379                 /* Re-enable interrupts */
 380                 local_irq_enable();
 381                 current_thread_info()->status |= TS_POLLING;
 382                 /* Compute time (ticks) that we were actually asleep */
 383                 sleep_ticks =
 384                     ticks_elapsed(t1, t2) - cx->latency_ticks - C2_OVERHEAD;
 385                 break;
 386
 387         case ACPI_STATE_C3:
 388
 389                 if (pr->flags.bm_check) {
 390                         if (atomic_inc_return(&c3_cpu_count) ==
 391                             num_online_cpus()) {
 392                                 /*
 393                                  * All CPUs are trying to go to C3
 394                                  * Disable bus master arbitration
 395                                  */
 396                                 acpi_set_register(ACPI_BITREG_ARB_DISABLE, 1,
 397                                                   ACPI_MTX_DO_NOT_LOCK);
 398                         }
 399                 } else {
 400                         /* SMP with no shared cache... Invalidate cache  */
 401                         ACPI_FLUSH_CPU_CACHE();
 402                 }
 403
 404                 /* Get start time (ticks) */
 405                 t1 = inl(acpi_fadt.xpm_tmr_blk.address);
 406                 /* Invoke C3 */
 407                 inb(cx->address);
 408                 /* Dummy wait op (see above) */
 409                 t2 = inl(acpi_fadt.xpm_tmr_blk.address);
 410                 /* Get end time (ticks) */
 411                 t2 = inl(acpi_fadt.xpm_tmr_blk.address);
 412                 if (pr->flags.bm_check) {
 413                         /* Enable bus master arbitration */
 414                         atomic_dec(&c3_cpu_count);
 415                         acpi_set_register(ACPI_BITREG_ARB_DISABLE, 0,
 416                                           ACPI_MTX_DO_NOT_LOCK);
 417                 }
 418
 419 #ifdef CONFIG_GENERIC_TIME
 420                 /* TSC halts in C3, so notify users */
 421                 mark_tsc_unstable();
 422 #endif
 423                 /* Re-enable interrupts */
 424                 local_irq_enable();
 425                 current_thread_info()->status |= TS_POLLING;
 426                 /* Compute time (ticks) that we were actually asleep */
 427                 sleep_ticks =
 428                     ticks_elapsed(t1, t2) - cx->latency_ticks - C3_OVERHEAD;
 429                 break;
 430
 431         default:
 432                 local_irq_enable();
 433                 return;
 434         }
 435
 436         next_state = pr->power.state;
 437
 438 #ifdef CONFIG_HOTPLUG_CPU
 439         /* Don't do promotion/demotion */
 440         if ((cx->type == ACPI_STATE_C1) && (num_online_cpus() > 1) &&
 441             !pr->flags.has_cst && !acpi_fadt.plvl2_up) {
 442                 next_state = cx;
 443                 goto end;
 444         }
 445 #endif
 446
 447         /*
 448          * Promotion?
 449          * ----------
 450          * Track the number of longs (time asleep is greater than threshold)
 451          * and promote when the count threshold is reached.  Note that bus
 452          * mastering activity may prevent promotions.
 453          * Do not promote above max_cstate.
 454          */
 455         if (cx->promotion.state &&
 456             ((cx->promotion.state - pr->power.states) <= max_cstate)) {
 457                 if (sleep_ticks > cx->promotion.threshold.ticks) {
 458                         cx->promotion.count++;
 459                         cx->demotion.count = 0;
 460                         if (cx->promotion.count >=
 461                             cx->promotion.threshold.count) {
 462                                 if (pr->flags.bm_check) {
 463                                         if (!
 464                                             (pr->power.bm_activity & cx->
 465                                              promotion.threshold.bm)) {
 466                                                 next_state =
 467                                                     cx->promotion.state;
 468                                                 goto end;
 469                                         }
 470                                 } else {
 471                                         next_state = cx->promotion.state;
 472                                         goto end;
 473                                 }
 474                         }
 475                 }
 476         }
 477
 478         /*
 479          * Demotion?
 480          * ---------
 481          * Track the number of shorts (time asleep is less than time threshold)
 482          * and demote when the usage threshold is reached.
 483          */
 484         if (cx->demotion.state) {
 485                 if (sleep_ticks < cx->demotion.threshold.ticks) {
 486                         cx->demotion.count++;
 487                         cx->promotion.count = 0;
 488                         if (cx->demotion.count >= cx->demotion.threshold.count) {
 489                                 next_state = cx->demotion.state;
 490                                 goto end;
 491                         }
 492                 }
 493         }
 494
 495       end:
 496         /*
 497          * Demote if current state exceeds max_cstate
 498          */
 499         if ((pr->power.state - pr->power.states) > max_cstate) {
 500                 if (cx->demotion.state)
 501                         next_state = cx->demotion.state;
 502         }
 503
 504         /*
 505          * New Cx State?
 506          * -------------
 507          * If we're going to start using a new Cx state we must clean up
 508          * from the previous and prepare to use the new.
 509          */
 510         if (next_state != pr->power.state)
 511                 acpi_processor_power_activate(pr, next_state);
 512 }
 513
 514 static int acpi_processor_set_power_policy(struct acpi_processor *pr)
 515 {
 516         unsigned int i;
 517         unsigned int state_is_set = 0;
 518         struct acpi_processor_cx *lower = NULL;
 519         struct acpi_processor_cx *higher = NULL;
 520         struct acpi_processor_cx *cx;
 521
 522
 523         if (!pr)
 524                 return -EINVAL;
 525
 526         /*
 527          * This function sets the default Cx state policy (OS idle handler).
 528          * Our scheme is to promote quickly to C2 but more conservatively
 529          * to C3.  We're favoring C2  for its characteristics of low latency
 530          * (quick response), good power savings, and ability to allow bus
 531          * mastering activity.  Note that the Cx state policy is completely
 532          * customizable and can be altered dynamically.
 533          */
 534
 535         /* startup state */
 536         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
 537                 cx = &pr->power.states[i];
 538                 if (!cx->valid)
 539                         continue;
 540
 541                 if (!state_is_set)
 542                         pr->power.state = cx;
 543                 state_is_set++;
 544                 break;
 545         }
 546
 547         if (!state_is_set)
 548                 return -ENODEV;
 549
 550         /* demotion */
 551         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
 552                 cx = &pr->power.states[i];
 553                 if (!cx->valid)
 554                         continue;
 555
 556                 if (lower) {
 557                         cx->demotion.state = lower;
 558                         cx->demotion.threshold.ticks = cx->latency_ticks;
 559                         cx->demotion.threshold.count = 1;
 560                         if (cx->type == ACPI_STATE_C3)
 561                                 cx->demotion.threshold.bm = bm_history;
 562                 }
 563
 564                 lower = cx;
 565         }
 566
 567         /* promotion */
 568         for (i = (ACPI_PROCESSOR_MAX_POWER - 1); i > 0; i--) {
 569                 cx = &pr->power.states[i];
 570                 if (!cx->valid)
 571                         continue;
 572
 573                 if (higher) {
 574                         cx->promotion.state = higher;
 575                         cx->promotion.threshold.ticks = cx->latency_ticks;
 576                         if (cx->type >= ACPI_STATE_C2)
 577                                 cx->promotion.threshold.count = 4;
 578                         else
 579                                 cx->promotion.threshold.count = 10;
 580                         if (higher->type == ACPI_STATE_C3)
 581                                 cx->promotion.threshold.bm = bm_history;
 582                 }
 583
 584                 higher = cx;
 585         }
 586
 587         return 0;
 588 }
 589
 590 static int acpi_processor_get_power_info_fadt(struct acpi_processor *pr)
 591 {
 592
 593         if (!pr)
 594                 return -EINVAL;
 595
 596         if (!pr->pblk)
 597                 return -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 -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 0;
 626 }
 627
 628 static int acpi_processor_get_power_info_default_c1(struct acpi_processor *pr)
 629 {
 630
 631         /* Zero initialize all the C-states info. */
 632         memset(pr->power.states, 0, sizeof(pr->power.states));
 633
 634         /* set the first C-State to C1 */
 635         pr->power.states[ACPI_STATE_C1].type = ACPI_STATE_C1;
 636
 637         /* the C0 state only exists as a filler in our array,
 638          * and all processors need to support C1 */
 639         pr->power.states[ACPI_STATE_C0].valid = 1;
 640         pr->power.states[ACPI_STATE_C1].valid = 1;
 641
 642         return 0;
 643 }
 644
 645 static int acpi_processor_get_power_info_cst(struct acpi_processor *pr)
 646 {
 647         acpi_status status = 0;
 648         acpi_integer count;
 649         int current_count;
 650         int i;
 651         struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
 652         union acpi_object *cst;
 653
 654
 655         if (nocst)
 656                 return -ENODEV;
 657
 658         current_count = 1;
 659
 660         /* Zero initialize C2 onwards and prepare for fresh CST lookup */
 661         for (i = 2; i < ACPI_PROCESSOR_MAX_POWER; i++)
 662                 memset(&(pr->power.states[i]), 0,
 663                                 sizeof(struct acpi_processor_cx));
 664
 665         status = acpi_evaluate_object(pr->handle, "_CST", NULL, &buffer);
 666         if (ACPI_FAILURE(status)) {
 667                 ACPI_DEBUG_PRINT((ACPI_DB_INFO, "No _CST, giving up\n"));
 668                 return -ENODEV;
 669         }
 670
 671         cst = (union acpi_object *)buffer.pointer;
 672
 673         /* There must be at least 2 elements */
 674         if (!cst || (cst->type != ACPI_TYPE_PACKAGE) || cst->package.count < 2) {
 675                 printk(KERN_ERR PREFIX "not enough elements in _CST\n");
 676                 status = -EFAULT;
 677                 goto end;
 678         }
 679
 680         count = cst->package.elements[0].integer.value;
 681
 682         /* Validate number of power states. */
 683         if (count < 1 || count != cst->package.count - 1) {
 684                 printk(KERN_ERR PREFIX "count given by _CST is not valid\n");
 685                 status = -EFAULT;
 686                 goto end;
 687         }
 688
 689         /* Tell driver that at least _CST is supported. */
 690         pr->flags.has_cst = 1;
 691
 692         for (i = 1; i <= count; i++) {
 693                 union acpi_object *element;
 694                 union acpi_object *obj;
 695                 struct acpi_power_register *reg;
 696                 struct acpi_processor_cx cx;
 697
 698                 memset(&cx, 0, sizeof(cx));
 699
 700                 element = (union acpi_object *)&(cst->package.elements[i]);
 701                 if (element->type != ACPI_TYPE_PACKAGE)
 702                         continue;
 703
 704                 if (element->package.count != 4)
 705                         continue;
 706
 707                 obj = (union acpi_object *)&(element->package.elements[0]);
 708
 709                 if (obj->type != ACPI_TYPE_BUFFER)
 710                         continue;
 711
 712                 reg = (struct acpi_power_register *)obj->buffer.pointer;
 713
 714                 if (reg->space_id != ACPI_ADR_SPACE_SYSTEM_IO &&
 715                     (reg->space_id != ACPI_ADR_SPACE_FIXED_HARDWARE))
 716                         continue;
 717
 718                 cx.address = (reg->space_id == ACPI_ADR_SPACE_FIXED_HARDWARE) ?
 719                     0 : reg->address;
 720
 721                 /* There should be an easy way to extract an integer... */
 722                 obj = (union acpi_object *)&(element->package.elements[1]);
 723                 if (obj->type != ACPI_TYPE_INTEGER)
 724                         continue;
 725
 726                 cx.type = obj->integer.value;
 727
 728                 if ((cx.type != ACPI_STATE_C1) &&
 729                     (reg->space_id != ACPI_ADR_SPACE_SYSTEM_IO))
 730                         continue;
 731
 732                 if ((cx.type < ACPI_STATE_C2) || (cx.type > ACPI_STATE_C3))
 733                         continue;
 734
 735                 obj = (union acpi_object *)&(element->package.elements[2]);
 736                 if (obj->type != ACPI_TYPE_INTEGER)
 737                         continue;
 738
 739                 cx.latency = obj->integer.value;
 740
 741                 obj = (union acpi_object *)&(element->package.elements[3]);
 742                 if (obj->type != ACPI_TYPE_INTEGER)
 743                         continue;
 744
 745                 cx.power = obj->integer.value;
 746
 747                 current_count++;
 748                 memcpy(&(pr->power.states[current_count]), &cx, sizeof(cx));
 749
 750                 /*
 751                  * We support total ACPI_PROCESSOR_MAX_POWER - 1
 752                  * (From 1 through ACPI_PROCESSOR_MAX_POWER - 1)
 753                  */
 754                 if (current_count >= (ACPI_PROCESSOR_MAX_POWER - 1)) {
 755                         printk(KERN_WARNING
 756                                "Limiting number of power states to max (%d)\n",
 757                                ACPI_PROCESSOR_MAX_POWER);
 758                         printk(KERN_WARNING
 759                                "Please increase ACPI_PROCESSOR_MAX_POWER if needed.\n");
 760                         break;
 761                 }
 762         }
 763
 764         ACPI_DEBUG_PRINT((ACPI_DB_INFO, "Found %d power states\n",
 765                           current_count));
 766
 767         /* Validate number of power states discovered */
 768         if (current_count < 2)
 769                 status = -EFAULT;
 770
 771       end:
 772         acpi_os_free(buffer.pointer);
 773
 774         return status;
 775 }
 776
 777 static void acpi_processor_power_verify_c2(struct acpi_processor_cx *cx)
 778 {
 779
 780         if (!cx->address)
 781                 return;
 782
 783         /*
 784          * C2 latency must be less than or equal to 100
 785          * microseconds.
 786          */
 787         else if (cx->latency > ACPI_PROCESSOR_MAX_C2_LATENCY) {
 788                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 789                                   "latency too large [%d]\n", cx->latency));
 790                 return;
 791         }
 792
 793         /*
 794          * Otherwise we've met all of our C2 requirements.
 795          * Normalize the C2 latency to expidite policy
 796          */
 797         cx->valid = 1;
 798         cx->latency_ticks = US_TO_PM_TIMER_TICKS(cx->latency);
 799
 800         return;
 801 }
 802
 803 static void acpi_processor_power_verify_c3(struct acpi_processor *pr,
 804                                            struct acpi_processor_cx *cx)
 805 {
 806         static int bm_check_flag;
 807
 808
 809         if (!cx->address)
 810                 return;
 811
 812         /*
 813          * C3 latency must be less than or equal to 1000
 814          * microseconds.
 815          */
 816         else if (cx->latency > ACPI_PROCESSOR_MAX_C3_LATENCY) {
 817                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 818                                   "latency too large [%d]\n", cx->latency));
 819                 return;
 820         }
 821
 822         /*
 823          * PIIX4 Erratum #18: We don't support C3 when Type-F (fast)
 824          * DMA transfers are used by any ISA device to avoid livelock.
 825          * Note that we could disable Type-F DMA (as recommended by
 826          * the erratum), but this is known to disrupt certain ISA
 827          * devices thus we take the conservative approach.
 828          */
 829         else if (errata.piix4.fdma) {
 830                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 831                                   "C3 not supported on PIIX4 with Type-F DMA\n"));
 832                 return;
 833         }
 834
 835         /* All the logic here assumes flags.bm_check is same across all CPUs */
 836         if (!bm_check_flag) {
 837                 /* Determine whether bm_check is needed based on CPU  */
 838                 acpi_processor_power_init_bm_check(&(pr->flags), pr->id);
 839                 bm_check_flag = pr->flags.bm_check;
 840         } else {
 841                 pr->flags.bm_check = bm_check_flag;
 842         }
 843
 844         if (pr->flags.bm_check) {
 845                 /* bus mastering control is necessary */
 846                 if (!pr->flags.bm_control) {
 847                         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 848                                           "C3 support requires bus mastering control\n"));
 849                         return;
 850                 }
 851         } else {
 852                 /*
 853                  * WBINVD should be set in fadt, for C3 state to be
 854                  * supported on when bm_check is not required.
 855                  */
 856                 if (acpi_fadt.wb_invd != 1) {
 857                         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 858                                           "Cache invalidation should work properly"
 859                                           " for C3 to be enabled on SMP systems\n"));
 860                         return;
 861                 }
 862                 acpi_set_register(ACPI_BITREG_BUS_MASTER_RLD,
 863                                   0, ACPI_MTX_DO_NOT_LOCK);
 864         }
 865
 866         /*
 867          * Otherwise we've met all of our C3 requirements.
 868          * Normalize the C3 latency to expidite policy.  Enable
 869          * checking of bus mastering status (bm_check) so we can
 870          * use this in our C3 policy
 871          */
 872         cx->valid = 1;
 873         cx->latency_ticks = US_TO_PM_TIMER_TICKS(cx->latency);
 874
 875         return;
 876 }
 877
 878 static int acpi_processor_power_verify(struct acpi_processor *pr)
 879 {
 880         unsigned int i;
 881         unsigned int working = 0;
 882
 883 #ifdef ARCH_APICTIMER_STOPS_ON_C3
 884         int timer_broadcast = 0;
 885         cpumask_t mask = cpumask_of_cpu(pr->id);
 886         on_each_cpu(switch_ipi_to_APIC_timer, &mask, 1, 1);
 887 #endif
 888
 889         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
 890                 struct acpi_processor_cx *cx = &pr->power.states[i];
 891
 892                 switch (cx->type) {
 893                 case ACPI_STATE_C1:
 894                         cx->valid = 1;
 895                         break;
 896
 897                 case ACPI_STATE_C2:
 898                         acpi_processor_power_verify_c2(cx);
 899 #ifdef ARCH_APICTIMER_STOPS_ON_C3
 900                         /* Some AMD systems fake C3 as C2, but still
 901                            have timer troubles */
 902                         if (cx->valid &&
 903                                 boot_cpu_data.x86_vendor == X86_VENDOR_AMD)
 904                                 timer_broadcast++;
 905 #endif
 906                         break;
 907
 908                 case ACPI_STATE_C3:
 909                         acpi_processor_power_verify_c3(pr, cx);
 910 #ifdef ARCH_APICTIMER_STOPS_ON_C3
 911                         if (cx->valid)
 912                                 timer_broadcast++;
 913 #endif
 914                         break;
 915                 }
 916
 917                 if (cx->valid)
 918                         working++;
 919         }
 920
 921 #ifdef ARCH_APICTIMER_STOPS_ON_C3
 922         if (timer_broadcast)
 923                 on_each_cpu(switch_APIC_timer_to_ipi, &mask, 1, 1);
 924 #endif
 925
 926         return (working);
 927 }
 928
 929 static int acpi_processor_get_power_info(struct acpi_processor *pr)
 930 {
 931         unsigned int i;
 932         int result;
 933
 934
 935         /* NOTE: the idle thread may not be running while calling
 936          * this function */
 937
 938         /* Adding C1 state */
 939         acpi_processor_get_power_info_default_c1(pr);
 940         result = acpi_processor_get_power_info_cst(pr);
 941         if (result == -ENODEV)
 942                 acpi_processor_get_power_info_fadt(pr);
 943
 944         pr->power.count = acpi_processor_power_verify(pr);
 945
 946         /*
 947          * Set Default Policy
 948          * ------------------
 949          * Now that we know which states are supported, set the default
 950          * policy.  Note that this policy can be changed dynamically
 951          * (e.g. encourage deeper sleeps to conserve battery life when
 952          * not on AC).
 953          */
 954         result = acpi_processor_set_power_policy(pr);
 955         if (result)
 956                 return result;
 957
 958         /*
 959          * if one state of type C2 or C3 is available, mark this
 960          * CPU as being "idle manageable"
 961          */
 962         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
 963                 if (pr->power.states[i].valid) {
 964                         pr->power.count = i;
 965                         if (pr->power.states[i].type >= ACPI_STATE_C2)
 966                                 pr->flags.power = 1;
 967                 }
 968         }
 969
 970         return 0;
 971 }
 972
 973 int acpi_processor_cst_has_changed(struct acpi_processor *pr)
 974 {
 975         int result = 0;
 976
 977
 978         if (!pr)
 979                 return -EINVAL;
 980
 981         if (nocst) {
 982                 return -ENODEV;
 983         }
 984
 985         if (!pr->flags.power_setup_done)
 986                 return -ENODEV;
 987
 988         /* Fall back to the default idle loop */
 989         pm_idle = pm_idle_save;
 990         synchronize_sched();    /* Relies on interrupts forcing exit from idle. */
 991
 992         pr->flags.power = 0;
 993         result = acpi_processor_get_power_info(pr);
 994         if ((pr->flags.power == 1) && (pr->flags.power_setup_done))
 995                 pm_idle = acpi_processor_idle;
 996
 997         return result;
 998 }
 999
1000 /* proc interface */
1001
1002 static int acpi_processor_power_seq_show(struct seq_file *seq, void *offset)
1003 {
1004         struct acpi_processor *pr = (struct acpi_processor *)seq->private;
1005         unsigned int i;
1006
1007
1008         if (!pr)
1009                 goto end;
1010
1011         seq_printf(seq, "active state:            C%zd\n"
1012                    "max_cstate:              C%d\n"
1013                    "bus master activity:     %08x\n",
1014                    pr->power.state ? pr->power.state - pr->power.states : 0,
1015                    max_cstate, (unsigned)pr->power.bm_activity);
1016
1017         seq_puts(seq, "states:\n");
1018
1019         for (i = 1; i <= pr->power.count; i++) {
1020                 seq_printf(seq, "   %cC%d:                  ",
1021                            (&pr->power.states[i] ==
1022                             pr->power.state ? '*' : ' '), i);
1023
1024                 if (!pr->power.states[i].valid) {
1025                         seq_puts(seq, "<not supported>\n");
1026                         continue;
1027                 }
1028
1029                 switch (pr->power.states[i].type) {
1030                 case ACPI_STATE_C1:
1031                         seq_printf(seq, "type[C1] ");
1032                         break;
1033                 case ACPI_STATE_C2:
1034                         seq_printf(seq, "type[C2] ");
1035                         break;
1036                 case ACPI_STATE_C3:
1037                         seq_printf(seq, "type[C3] ");
1038                         break;
1039                 default:
1040                         seq_printf(seq, "type[--] ");
1041                         break;
1042                 }
1043
1044                 if (pr->power.states[i].promotion.state)
1045                         seq_printf(seq, "promotion[C%zd] ",
1046                                    (pr->power.states[i].promotion.state -
1047                                     pr->power.states));
1048                 else
1049                         seq_puts(seq, "promotion[--] ");
1050
1051                 if (pr->power.states[i].demotion.state)
1052                         seq_printf(seq, "demotion[C%zd] ",
1053                                    (pr->power.states[i].demotion.state -
1054                                     pr->power.states));
1055                 else
1056                         seq_puts(seq, "demotion[--] ");
1057
1058                 seq_printf(seq, "latency[%03d] usage[%08d]\n",
1059                            pr->power.states[i].latency,
1060                            pr->power.states[i].usage);
1061         }
1062
1063       end:
1064         return 0;
1065 }
1066
1067 static int acpi_processor_power_open_fs(struct inode *inode, struct file *file)
1068 {
1069         return single_open(file, acpi_processor_power_seq_show,
1070                            PDE(inode)->data);
1071 }
1072
1073 static struct file_operations acpi_processor_power_fops = {
1074         .open = acpi_processor_power_open_fs,
1075         .read = seq_read,
1076         .llseek = seq_lseek,
1077         .release = single_release,
1078 };
1079
1080 int acpi_processor_power_init(struct acpi_processor *pr,
1081                               struct acpi_device *device)
1082 {
1083         acpi_status status = 0;
1084         static int first_run;
1085         struct proc_dir_entry *entry = NULL;
1086         unsigned int i;
1087
1088
1089         if (!first_run) {
1090                 dmi_check_system(processor_power_dmi_table);
1091                 if (max_cstate < ACPI_C_STATES_MAX)
1092                         printk(KERN_NOTICE
1093                                "ACPI: processor limited to max C-state %d\n",
1094                                max_cstate);
1095                 first_run++;
1096         }
1097
1098         if (!pr)
1099                 return -EINVAL;
1100
1101         if (acpi_fadt.cst_cnt && !nocst) {
1102                 status =
1103                     acpi_os_write_port(acpi_fadt.smi_cmd, acpi_fadt.cst_cnt, 8);
1104                 if (ACPI_FAILURE(status)) {
1105                         ACPI_EXCEPTION((AE_INFO, status,
1106                                         "Notifying BIOS of _CST ability failed"));
1107                 }
1108         }
1109
1110         acpi_processor_get_power_info(pr);
1111
1112         /*
1113          * Install the idle handler if processor power management is supported.
1114          * Note that we use previously set idle handler will be used on
1115          * platforms that only support C1.
1116          */
1117         if ((pr->flags.power) && (!boot_option_idle_override)) {
1118                 printk(KERN_INFO PREFIX "CPU%d (power states:", pr->id);
1119                 for (i = 1; i <= pr->power.count; i++)
1120                         if (pr->power.states[i].valid)
1121                                 printk(" C%d[C%d]", i,
1122                                        pr->power.states[i].type);
1123                 printk(")\n");
1124
1125                 if (pr->id == 0) {
1126                         pm_idle_save = pm_idle;
1127                         pm_idle = acpi_processor_idle;
1128                 }
1129         }
1130
1131         /* 'power' [R] */
1132         entry = create_proc_entry(ACPI_PROCESSOR_FILE_POWER,
1133                                   S_IRUGO, acpi_device_dir(device));
1134         if (!entry)
1135                 return -EIO;
1136         else {
1137                 entry->proc_fops = &acpi_processor_power_fops;
1138                 entry->data = acpi_driver_data(device);
1139                 entry->owner = THIS_MODULE;
1140         }
1141
1142         pr->flags.power_setup_done = 1;
1143
1144         return 0;
1145 }
1146
1147 int acpi_processor_power_exit(struct acpi_processor *pr,
1148                               struct acpi_device *device)
1149 {
1150
1151         pr->flags.power_setup_done = 0;
1152
1153         if (acpi_device_dir(device))
1154                 remove_proc_entry(ACPI_PROCESSOR_FILE_POWER,
1155                                   acpi_device_dir(device));
1156
1157         /* Unregister the idle handler when processor #0 is removed. */
1158         if (pr->id == 0) {
1159                 pm_idle = pm_idle_save;
1160
1161                 /*
1162                  * We are about to unload the current idle thread pm callback
1163                  * (pm_idle), Wait for all processors to update cached/local
1164                  * copies of pm_idle before proceeding.
1165                  */
1166                 cpu_idle_wait();
1167         }
1168
1169         return 0;
1170 }