git.oblomov.eu Git - linux-2.6/blob - drivers/acpi/processor_idle.c

   1 /*
   2  * processor_idle - idle state submodule to the ACPI processor driver
   3  *
   4  *  Copyright (C) 2001, 2002 Andy Grover <andrew.grover@intel.com>
   5  *  Copyright (C) 2001, 2002 Paul Diefenbaugh <paul.s.diefenbaugh@intel.com>
   6  *  Copyright (C) 2004       Dominik Brodowski <linux@brodo.de>
   7  *  Copyright (C) 2004  Anil S Keshavamurthy <anil.s.keshavamurthy@intel.com>
   8  *                      - Added processor hotplug support
   9  *  Copyright (C) 2005  Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>
  10  *                      - Added support for C3 on SMP
  11  *
  12  * ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
  13  *
  14  *  This program is free software; you can redistribute it and/or modify
  15  *  it under the terms of the GNU General Public License as published by
  16  *  the Free Software Foundation; either version 2 of the License, or (at
  17  *  your option) any later version.
  18  *
  19  *  This program is distributed in the hope that it will be useful, but
  20  *  WITHOUT ANY WARRANTY; without even the implied warranty of
  21  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  22  *  General Public License for more details.
  23  *
  24  *  You should have received a copy of the GNU General Public License along
  25  *  with this program; if not, write to the Free Software Foundation, Inc.,
  26  *  59 Temple Place, Suite 330, Boston, MA 02111-1307 USA.
  27  *
  28  * ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
  29  */
  30
  31 #include <linux/kernel.h>
  32 #include <linux/module.h>
  33 #include <linux/init.h>
  34 #include <linux/cpufreq.h>
  35 #include <linux/proc_fs.h>
  36 #include <linux/seq_file.h>
  37 #include <linux/acpi.h>
  38 #include <linux/dmi.h>
  39 #include <linux/moduleparam.h>
  40 #include <linux/sched.h>        /* need_resched() */
  41
  42 #include <asm/io.h>
  43 #include <asm/uaccess.h>
  44
  45 #include <acpi/acpi_bus.h>
  46 #include <acpi/processor.h>
  47
  48 #define ACPI_PROCESSOR_COMPONENT        0x01000000
  49 #define ACPI_PROCESSOR_CLASS            "processor"
  50 #define ACPI_PROCESSOR_DRIVER_NAME      "ACPI Processor Driver"
  51 #define _COMPONENT              ACPI_PROCESSOR_COMPONENT
  52 ACPI_MODULE_NAME("acpi_processor")
  53 #define ACPI_PROCESSOR_FILE_POWER       "power"
  54 #define US_TO_PM_TIMER_TICKS(t)         ((t * (PM_TIMER_FREQUENCY/1000)) / 1000)
  55 #define C2_OVERHEAD                     4       /* 1us (3.579 ticks per us) */
  56 #define C3_OVERHEAD                     4       /* 1us (3.579 ticks per us) */
  57 static void (*pm_idle_save) (void) __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         clear_thread_flag(TIF_POLLING_NRFLAG);
 210         smp_mb__after_clear_bit();
 211         if (!need_resched())
 212                 safe_halt();
 213         set_thread_flag(TIF_POLLING_NRFLAG);
 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                 clear_thread_flag(TIF_POLLING_NRFLAG);
 334                 smp_mb__after_clear_bit();
 335                 if (need_resched()) {
 336                         set_thread_flag(TIF_POLLING_NRFLAG);
 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                 /* Re-enable interrupts */
 373                 local_irq_enable();
 374                 set_thread_flag(TIF_POLLING_NRFLAG);
 375                 /* Compute time (ticks) that we were actually asleep */
 376                 sleep_ticks =
 377                     ticks_elapsed(t1, t2) - cx->latency_ticks - C2_OVERHEAD;
 378                 break;
 379
 380         case ACPI_STATE_C3:
 381
 382                 if (pr->flags.bm_check) {
 383                         if (atomic_inc_return(&c3_cpu_count) ==
 384                             num_online_cpus()) {
 385                                 /*
 386                                  * All CPUs are trying to go to C3
 387                                  * Disable bus master arbitration
 388                                  */
 389                                 acpi_set_register(ACPI_BITREG_ARB_DISABLE, 1,
 390                                                   ACPI_MTX_DO_NOT_LOCK);
 391                         }
 392                 } else {
 393                         /* SMP with no shared cache... Invalidate cache  */
 394                         ACPI_FLUSH_CPU_CACHE();
 395                 }
 396
 397                 /* Get start time (ticks) */
 398                 t1 = inl(acpi_fadt.xpm_tmr_blk.address);
 399                 /* Invoke C3 */
 400                 inb(cx->address);
 401                 /* Dummy op - must do something useless after P_LVL3 read */
 402                 t2 = inl(acpi_fadt.xpm_tmr_blk.address);
 403                 /* Get end time (ticks) */
 404                 t2 = inl(acpi_fadt.xpm_tmr_blk.address);
 405                 if (pr->flags.bm_check) {
 406                         /* Enable bus master arbitration */
 407                         atomic_dec(&c3_cpu_count);
 408                         acpi_set_register(ACPI_BITREG_ARB_DISABLE, 0,
 409                                           ACPI_MTX_DO_NOT_LOCK);
 410                 }
 411
 412                 /* Re-enable interrupts */
 413                 local_irq_enable();
 414                 set_thread_flag(TIF_POLLING_NRFLAG);
 415                 /* Compute time (ticks) that we were actually asleep */
 416                 sleep_ticks =
 417                     ticks_elapsed(t1, t2) - cx->latency_ticks - C3_OVERHEAD;
 418                 break;
 419
 420         default:
 421                 local_irq_enable();
 422                 return;
 423         }
 424
 425         next_state = pr->power.state;
 426
 427 #ifdef CONFIG_HOTPLUG_CPU
 428         /* Don't do promotion/demotion */
 429         if ((cx->type == ACPI_STATE_C1) && (num_online_cpus() > 1) &&
 430             !pr->flags.has_cst && !acpi_fadt.plvl2_up) {
 431                 next_state = cx;
 432                 goto end;
 433         }
 434 #endif
 435
 436         /*
 437          * Promotion?
 438          * ----------
 439          * Track the number of longs (time asleep is greater than threshold)
 440          * and promote when the count threshold is reached.  Note that bus
 441          * mastering activity may prevent promotions.
 442          * Do not promote above max_cstate.
 443          */
 444         if (cx->promotion.state &&
 445             ((cx->promotion.state - pr->power.states) <= max_cstate)) {
 446                 if (sleep_ticks > cx->promotion.threshold.ticks) {
 447                         cx->promotion.count++;
 448                         cx->demotion.count = 0;
 449                         if (cx->promotion.count >=
 450                             cx->promotion.threshold.count) {
 451                                 if (pr->flags.bm_check) {
 452                                         if (!
 453                                             (pr->power.bm_activity & cx->
 454                                              promotion.threshold.bm)) {
 455                                                 next_state =
 456                                                     cx->promotion.state;
 457                                                 goto end;
 458                                         }
 459                                 } else {
 460                                         next_state = cx->promotion.state;
 461                                         goto end;
 462                                 }
 463                         }
 464                 }
 465         }
 466
 467         /*
 468          * Demotion?
 469          * ---------
 470          * Track the number of shorts (time asleep is less than time threshold)
 471          * and demote when the usage threshold is reached.
 472          */
 473         if (cx->demotion.state) {
 474                 if (sleep_ticks < cx->demotion.threshold.ticks) {
 475                         cx->demotion.count++;
 476                         cx->promotion.count = 0;
 477                         if (cx->demotion.count >= cx->demotion.threshold.count) {
 478                                 next_state = cx->demotion.state;
 479                                 goto end;
 480                         }
 481                 }
 482         }
 483
 484       end:
 485         /*
 486          * Demote if current state exceeds max_cstate
 487          */
 488         if ((pr->power.state - pr->power.states) > max_cstate) {
 489                 if (cx->demotion.state)
 490                         next_state = cx->demotion.state;
 491         }
 492
 493         /*
 494          * New Cx State?
 495          * -------------
 496          * If we're going to start using a new Cx state we must clean up
 497          * from the previous and prepare to use the new.
 498          */
 499         if (next_state != pr->power.state)
 500                 acpi_processor_power_activate(pr, next_state);
 501 }
 502
 503 static int acpi_processor_set_power_policy(struct acpi_processor *pr)
 504 {
 505         unsigned int i;
 506         unsigned int state_is_set = 0;
 507         struct acpi_processor_cx *lower = NULL;
 508         struct acpi_processor_cx *higher = NULL;
 509         struct acpi_processor_cx *cx;
 510
 511         ACPI_FUNCTION_TRACE("acpi_processor_set_power_policy");
 512
 513         if (!pr)
 514                 return_VALUE(-EINVAL);
 515
 516         /*
 517          * This function sets the default Cx state policy (OS idle handler).
 518          * Our scheme is to promote quickly to C2 but more conservatively
 519          * to C3.  We're favoring C2  for its characteristics of low latency
 520          * (quick response), good power savings, and ability to allow bus
 521          * mastering activity.  Note that the Cx state policy is completely
 522          * customizable and can be altered dynamically.
 523          */
 524
 525         /* startup state */
 526         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
 527                 cx = &pr->power.states[i];
 528                 if (!cx->valid)
 529                         continue;
 530
 531                 if (!state_is_set)
 532                         pr->power.state = cx;
 533                 state_is_set++;
 534                 break;
 535         }
 536
 537         if (!state_is_set)
 538                 return_VALUE(-ENODEV);
 539
 540         /* demotion */
 541         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
 542                 cx = &pr->power.states[i];
 543                 if (!cx->valid)
 544                         continue;
 545
 546                 if (lower) {
 547                         cx->demotion.state = lower;
 548                         cx->demotion.threshold.ticks = cx->latency_ticks;
 549                         cx->demotion.threshold.count = 1;
 550                         if (cx->type == ACPI_STATE_C3)
 551                                 cx->demotion.threshold.bm = bm_history;
 552                 }
 553
 554                 lower = cx;
 555         }
 556
 557         /* promotion */
 558         for (i = (ACPI_PROCESSOR_MAX_POWER - 1); i > 0; i--) {
 559                 cx = &pr->power.states[i];
 560                 if (!cx->valid)
 561                         continue;
 562
 563                 if (higher) {
 564                         cx->promotion.state = higher;
 565                         cx->promotion.threshold.ticks = cx->latency_ticks;
 566                         if (cx->type >= ACPI_STATE_C2)
 567                                 cx->promotion.threshold.count = 4;
 568                         else
 569                                 cx->promotion.threshold.count = 10;
 570                         if (higher->type == ACPI_STATE_C3)
 571                                 cx->promotion.threshold.bm = bm_history;
 572                 }
 573
 574                 higher = cx;
 575         }
 576
 577         return_VALUE(0);
 578 }
 579
 580 static int acpi_processor_get_power_info_fadt(struct acpi_processor *pr)
 581 {
 582         ACPI_FUNCTION_TRACE("acpi_processor_get_power_info_fadt");
 583
 584         if (!pr)
 585                 return_VALUE(-EINVAL);
 586
 587         if (!pr->pblk)
 588                 return_VALUE(-ENODEV);
 589
 590         /* if info is obtained from pblk/fadt, type equals state */
 591         pr->power.states[ACPI_STATE_C2].type = ACPI_STATE_C2;
 592         pr->power.states[ACPI_STATE_C3].type = ACPI_STATE_C3;
 593
 594 #ifndef CONFIG_HOTPLUG_CPU
 595         /*
 596          * Check for P_LVL2_UP flag before entering C2 and above on
 597          * an SMP system.
 598          */
 599         if ((num_online_cpus() > 1) && !acpi_fadt.plvl2_up)
 600                 return_VALUE(-ENODEV);
 601 #endif
 602
 603         /* determine C2 and C3 address from pblk */
 604         pr->power.states[ACPI_STATE_C2].address = pr->pblk + 4;
 605         pr->power.states[ACPI_STATE_C3].address = pr->pblk + 5;
 606
 607         /* determine latencies from FADT */
 608         pr->power.states[ACPI_STATE_C2].latency = acpi_fadt.plvl2_lat;
 609         pr->power.states[ACPI_STATE_C3].latency = acpi_fadt.plvl3_lat;
 610
 611         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 612                           "lvl2[0x%08x] lvl3[0x%08x]\n",
 613                           pr->power.states[ACPI_STATE_C2].address,
 614                           pr->power.states[ACPI_STATE_C3].address));
 615
 616         return_VALUE(0);
 617 }
 618
 619 static int acpi_processor_get_power_info_default_c1(struct acpi_processor *pr)
 620 {
 621         ACPI_FUNCTION_TRACE("acpi_processor_get_power_info_default_c1");
 622
 623         /* Zero initialize all the C-states info. */
 624         memset(pr->power.states, 0, sizeof(pr->power.states));
 625
 626         /* set the first C-State to C1 */
 627         pr->power.states[ACPI_STATE_C1].type = ACPI_STATE_C1;
 628
 629         /* the C0 state only exists as a filler in our array,
 630          * and all processors need to support C1 */
 631         pr->power.states[ACPI_STATE_C0].valid = 1;
 632         pr->power.states[ACPI_STATE_C1].valid = 1;
 633
 634         return_VALUE(0);
 635 }
 636
 637 static int acpi_processor_get_power_info_cst(struct acpi_processor *pr)
 638 {
 639         acpi_status status = 0;
 640         acpi_integer count;
 641         int current_count;
 642         int i;
 643         struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
 644         union acpi_object *cst;
 645
 646         ACPI_FUNCTION_TRACE("acpi_processor_get_power_info_cst");
 647
 648         if (nocst)
 649                 return_VALUE(-ENODEV);
 650
 651         current_count = 1;
 652
 653         /* Zero initialize C2 onwards and prepare for fresh CST lookup */
 654         for (i = 2; i < ACPI_PROCESSOR_MAX_POWER; i++)
 655                 memset(&(pr->power.states[i]), 0,
 656                                 sizeof(struct acpi_processor_cx));
 657
 658         status = acpi_evaluate_object(pr->handle, "_CST", NULL, &buffer);
 659         if (ACPI_FAILURE(status)) {
 660                 ACPI_DEBUG_PRINT((ACPI_DB_INFO, "No _CST, giving up\n"));
 661                 return_VALUE(-ENODEV);
 662         }
 663
 664         cst = (union acpi_object *)buffer.pointer;
 665
 666         /* There must be at least 2 elements */
 667         if (!cst || (cst->type != ACPI_TYPE_PACKAGE) || cst->package.count < 2) {
 668                 ACPI_DEBUG_PRINT((ACPI_DB_ERROR,
 669                                   "not enough elements in _CST\n"));
 670                 status = -EFAULT;
 671                 goto end;
 672         }
 673
 674         count = cst->package.elements[0].integer.value;
 675
 676         /* Validate number of power states. */
 677         if (count < 1 || count != cst->package.count - 1) {
 678                 ACPI_DEBUG_PRINT((ACPI_DB_ERROR,
 679                                   "count given by _CST is not valid\n"));
 680                 status = -EFAULT;
 681                 goto end;
 682         }
 683
 684         /* Tell driver that at least _CST is supported. */
 685         pr->flags.has_cst = 1;
 686
 687         for (i = 1; i <= count; i++) {
 688                 union acpi_object *element;
 689                 union acpi_object *obj;
 690                 struct acpi_power_register *reg;
 691                 struct acpi_processor_cx cx;
 692
 693                 memset(&cx, 0, sizeof(cx));
 694
 695                 element = (union acpi_object *)&(cst->package.elements[i]);
 696                 if (element->type != ACPI_TYPE_PACKAGE)
 697                         continue;
 698
 699                 if (element->package.count != 4)
 700                         continue;
 701
 702                 obj = (union acpi_object *)&(element->package.elements[0]);
 703
 704                 if (obj->type != ACPI_TYPE_BUFFER)
 705                         continue;
 706
 707                 reg = (struct acpi_power_register *)obj->buffer.pointer;
 708
 709                 if (reg->space_id != ACPI_ADR_SPACE_SYSTEM_IO &&
 710                     (reg->space_id != ACPI_ADR_SPACE_FIXED_HARDWARE))
 711                         continue;
 712
 713                 cx.address = (reg->space_id == ACPI_ADR_SPACE_FIXED_HARDWARE) ?
 714                     0 : reg->address;
 715
 716                 /* There should be an easy way to extract an integer... */
 717                 obj = (union acpi_object *)&(element->package.elements[1]);
 718                 if (obj->type != ACPI_TYPE_INTEGER)
 719                         continue;
 720
 721                 cx.type = obj->integer.value;
 722
 723                 if ((cx.type != ACPI_STATE_C1) &&
 724                     (reg->space_id != ACPI_ADR_SPACE_SYSTEM_IO))
 725                         continue;
 726
 727                 if ((cx.type < ACPI_STATE_C2) || (cx.type > ACPI_STATE_C3))
 728                         continue;
 729
 730                 obj = (union acpi_object *)&(element->package.elements[2]);
 731                 if (obj->type != ACPI_TYPE_INTEGER)
 732                         continue;
 733
 734                 cx.latency = obj->integer.value;
 735
 736                 obj = (union acpi_object *)&(element->package.elements[3]);
 737                 if (obj->type != ACPI_TYPE_INTEGER)
 738                         continue;
 739
 740                 cx.power = obj->integer.value;
 741
 742                 current_count++;
 743                 memcpy(&(pr->power.states[current_count]), &cx, sizeof(cx));
 744
 745                 /*
 746                  * We support total ACPI_PROCESSOR_MAX_POWER - 1
 747                  * (From 1 through ACPI_PROCESSOR_MAX_POWER - 1)
 748                  */
 749                 if (current_count >= (ACPI_PROCESSOR_MAX_POWER - 1)) {
 750                         printk(KERN_WARNING
 751                                "Limiting number of power states to max (%d)\n",
 752                                ACPI_PROCESSOR_MAX_POWER);
 753                         printk(KERN_WARNING
 754                                "Please increase ACPI_PROCESSOR_MAX_POWER if needed.\n");
 755                         break;
 756                 }
 757         }
 758
 759         ACPI_DEBUG_PRINT((ACPI_DB_INFO, "Found %d power states\n",
 760                           current_count));
 761
 762         /* Validate number of power states discovered */
 763         if (current_count < 2)
 764                 status = -EFAULT;
 765
 766       end:
 767         acpi_os_free(buffer.pointer);
 768
 769         return_VALUE(status);
 770 }
 771
 772 static void acpi_processor_power_verify_c2(struct acpi_processor_cx *cx)
 773 {
 774         ACPI_FUNCTION_TRACE("acpi_processor_get_power_verify_c2");
 775
 776         if (!cx->address)
 777                 return_VOID;
 778
 779         /*
 780          * C2 latency must be less than or equal to 100
 781          * microseconds.
 782          */
 783         else if (cx->latency > ACPI_PROCESSOR_MAX_C2_LATENCY) {
 784                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 785                                   "latency too large [%d]\n", cx->latency));
 786                 return_VOID;
 787         }
 788
 789         /*
 790          * Otherwise we've met all of our C2 requirements.
 791          * Normalize the C2 latency to expidite policy
 792          */
 793         cx->valid = 1;
 794         cx->latency_ticks = US_TO_PM_TIMER_TICKS(cx->latency);
 795
 796         return_VOID;
 797 }
 798
 799 static void acpi_processor_power_verify_c3(struct acpi_processor *pr,
 800                                            struct acpi_processor_cx *cx)
 801 {
 802         static int bm_check_flag;
 803
 804         ACPI_FUNCTION_TRACE("acpi_processor_get_power_verify_c3");
 805
 806         if (!cx->address)
 807                 return_VOID;
 808
 809         /*
 810          * C3 latency must be less than or equal to 1000
 811          * microseconds.
 812          */
 813         else if (cx->latency > ACPI_PROCESSOR_MAX_C3_LATENCY) {
 814                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 815                                   "latency too large [%d]\n", cx->latency));
 816                 return_VOID;
 817         }
 818
 819         /*
 820          * PIIX4 Erratum #18: We don't support C3 when Type-F (fast)
 821          * DMA transfers are used by any ISA device to avoid livelock.
 822          * Note that we could disable Type-F DMA (as recommended by
 823          * the erratum), but this is known to disrupt certain ISA
 824          * devices thus we take the conservative approach.
 825          */
 826         else if (errata.piix4.fdma) {
 827                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 828                                   "C3 not supported on PIIX4 with Type-F DMA\n"));
 829                 return_VOID;
 830         }
 831
 832         /* All the logic here assumes flags.bm_check is same across all CPUs */
 833         if (!bm_check_flag) {
 834                 /* Determine whether bm_check is needed based on CPU  */
 835                 acpi_processor_power_init_bm_check(&(pr->flags), pr->id);
 836                 bm_check_flag = pr->flags.bm_check;
 837         } else {
 838                 pr->flags.bm_check = bm_check_flag;
 839         }
 840
 841         if (pr->flags.bm_check) {
 842                 /* bus mastering control is necessary */
 843                 if (!pr->flags.bm_control) {
 844                         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 845                                           "C3 support requires bus mastering control\n"));
 846                         return_VOID;
 847                 }
 848         } else {
 849                 /*
 850                  * WBINVD should be set in fadt, for C3 state to be
 851                  * supported on when bm_check is not required.
 852                  */
 853                 if (acpi_fadt.wb_invd != 1) {
 854                         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 855                                           "Cache invalidation should work properly"
 856                                           " for C3 to be enabled on SMP systems\n"));
 857                         return_VOID;
 858                 }
 859                 acpi_set_register(ACPI_BITREG_BUS_MASTER_RLD,
 860                                   0, ACPI_MTX_DO_NOT_LOCK);
 861         }
 862
 863         /*
 864          * Otherwise we've met all of our C3 requirements.
 865          * Normalize the C3 latency to expidite policy.  Enable
 866          * checking of bus mastering status (bm_check) so we can
 867          * use this in our C3 policy
 868          */
 869         cx->valid = 1;
 870         cx->latency_ticks = US_TO_PM_TIMER_TICKS(cx->latency);
 871
 872         return_VOID;
 873 }
 874
 875 static int acpi_processor_power_verify(struct acpi_processor *pr)
 876 {
 877         unsigned int i;
 878         unsigned int working = 0;
 879
 880 #ifdef ARCH_APICTIMER_STOPS_ON_C3
 881         int timer_broadcast = 0;
 882         cpumask_t mask = cpumask_of_cpu(pr->id);
 883         on_each_cpu(switch_ipi_to_APIC_timer, &mask, 1, 1);
 884 #endif
 885
 886         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
 887                 struct acpi_processor_cx *cx = &pr->power.states[i];
 888
 889                 switch (cx->type) {
 890                 case ACPI_STATE_C1:
 891                         cx->valid = 1;
 892                         break;
 893
 894                 case ACPI_STATE_C2:
 895                         acpi_processor_power_verify_c2(cx);
 896 #ifdef ARCH_APICTIMER_STOPS_ON_C3
 897                         /* Some AMD systems fake C3 as C2, but still
 898                            have timer troubles */
 899                         if (cx->valid &&
 900                                 boot_cpu_data.x86_vendor == X86_VENDOR_AMD)
 901                                 timer_broadcast++;
 902 #endif
 903                         break;
 904
 905                 case ACPI_STATE_C3:
 906                         acpi_processor_power_verify_c3(pr, cx);
 907 #ifdef ARCH_APICTIMER_STOPS_ON_C3
 908                         if (cx->valid)
 909                                 timer_broadcast++;
 910 #endif
 911                         break;
 912                 }
 913
 914                 if (cx->valid)
 915                         working++;
 916         }
 917
 918 #ifdef ARCH_APICTIMER_STOPS_ON_C3
 919         if (timer_broadcast)
 920                 on_each_cpu(switch_APIC_timer_to_ipi, &mask, 1, 1);
 921 #endif
 922
 923         return (working);
 924 }
 925
 926 static int acpi_processor_get_power_info(struct acpi_processor *pr)
 927 {
 928         unsigned int i;
 929         int result;
 930
 931         ACPI_FUNCTION_TRACE("acpi_processor_get_power_info");
 932
 933         /* NOTE: the idle thread may not be running while calling
 934          * this function */
 935
 936         /* Adding C1 state */
 937         acpi_processor_get_power_info_default_c1(pr);
 938         result = acpi_processor_get_power_info_cst(pr);
 939         if (result == -ENODEV)
 940                 acpi_processor_get_power_info_fadt(pr);
 941
 942         pr->power.count = acpi_processor_power_verify(pr);
 943
 944         /*
 945          * Set Default Policy
 946          * ------------------
 947          * Now that we know which states are supported, set the default
 948          * policy.  Note that this policy can be changed dynamically
 949          * (e.g. encourage deeper sleeps to conserve battery life when
 950          * not on AC).
 951          */
 952         result = acpi_processor_set_power_policy(pr);
 953         if (result)
 954                 return_VALUE(result);
 955
 956         /*
 957          * if one state of type C2 or C3 is available, mark this
 958          * CPU as being "idle manageable"
 959          */
 960         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
 961                 if (pr->power.states[i].valid) {
 962                         pr->power.count = i;
 963                         if (pr->power.states[i].type >= ACPI_STATE_C2)
 964                                 pr->flags.power = 1;
 965                 }
 966         }
 967
 968         return_VALUE(0);
 969 }
 970
 971 int acpi_processor_cst_has_changed(struct acpi_processor *pr)
 972 {
 973         int result = 0;
 974
 975         ACPI_FUNCTION_TRACE("acpi_processor_cst_has_changed");
 976
 977         if (!pr)
 978                 return_VALUE(-EINVAL);
 979
 980         if (nocst) {
 981                 return_VALUE(-ENODEV);
 982         }
 983
 984         if (!pr->flags.power_setup_done)
 985                 return_VALUE(-ENODEV);
 986
 987         /* Fall back to the default idle loop */
 988         pm_idle = pm_idle_save;
 989         synchronize_sched();    /* Relies on interrupts forcing exit from idle. */
 990
 991         pr->flags.power = 0;
 992         result = acpi_processor_get_power_info(pr);
 993         if ((pr->flags.power == 1) && (pr->flags.power_setup_done))
 994                 pm_idle = acpi_processor_idle;
 995
 996         return_VALUE(result);
 997 }
 998
 999 /* proc interface */
1000
1001 static int acpi_processor_power_seq_show(struct seq_file *seq, void *offset)
1002 {
1003         struct acpi_processor *pr = (struct acpi_processor *)seq->private;
1004         unsigned int i;
1005
1006         ACPI_FUNCTION_TRACE("acpi_processor_power_seq_show");
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_VALUE(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         ACPI_FUNCTION_TRACE("acpi_processor_power_init");
1089
1090         if (!first_run) {
1091                 dmi_check_system(processor_power_dmi_table);
1092                 if (max_cstate < ACPI_C_STATES_MAX)
1093                         printk(KERN_NOTICE
1094                                "ACPI: processor limited to max C-state %d\n",
1095                                max_cstate);
1096                 first_run++;
1097         }
1098
1099         if (!pr)
1100                 return_VALUE(-EINVAL);
1101
1102         if (acpi_fadt.cst_cnt && !nocst) {
1103                 status =
1104                     acpi_os_write_port(acpi_fadt.smi_cmd, acpi_fadt.cst_cnt, 8);
1105                 if (ACPI_FAILURE(status)) {
1106                         ACPI_DEBUG_PRINT((ACPI_DB_ERROR,
1107                                           "Notifying BIOS of _CST ability failed\n"));
1108                 }
1109         }
1110
1111         acpi_processor_get_power_info(pr);
1112
1113         /*
1114          * Install the idle handler if processor power management is supported.
1115          * Note that we use previously set idle handler will be used on
1116          * platforms that only support C1.
1117          */
1118         if ((pr->flags.power) && (!boot_option_idle_override)) {
1119                 printk(KERN_INFO PREFIX "CPU%d (power states:", pr->id);
1120                 for (i = 1; i <= pr->power.count; i++)
1121                         if (pr->power.states[i].valid)
1122                                 printk(" C%d[C%d]", i,
1123                                        pr->power.states[i].type);
1124                 printk(")\n");
1125
1126                 if (pr->id == 0) {
1127                         pm_idle_save = pm_idle;
1128                         pm_idle = acpi_processor_idle;
1129                 }
1130         }
1131
1132         /* 'power' [R] */
1133         entry = create_proc_entry(ACPI_PROCESSOR_FILE_POWER,
1134                                   S_IRUGO, acpi_device_dir(device));
1135         if (!entry)
1136                 ACPI_DEBUG_PRINT((ACPI_DB_ERROR,
1137                                   "Unable to create '%s' fs entry\n",
1138                                   ACPI_PROCESSOR_FILE_POWER));
1139         else {
1140                 entry->proc_fops = &acpi_processor_power_fops;
1141                 entry->data = acpi_driver_data(device);
1142                 entry->owner = THIS_MODULE;
1143         }
1144
1145         pr->flags.power_setup_done = 1;
1146
1147         return_VALUE(0);
1148 }
1149
1150 int acpi_processor_power_exit(struct acpi_processor *pr,
1151                               struct acpi_device *device)
1152 {
1153         ACPI_FUNCTION_TRACE("acpi_processor_power_exit");
1154
1155         pr->flags.power_setup_done = 0;
1156
1157         if (acpi_device_dir(device))
1158                 remove_proc_entry(ACPI_PROCESSOR_FILE_POWER,
1159                                   acpi_device_dir(device));
1160
1161         /* Unregister the idle handler when processor #0 is removed. */
1162         if (pr->id == 0) {
1163                 pm_idle = pm_idle_save;
1164
1165                 /*
1166                  * We are about to unload the current idle thread pm callback
1167                  * (pm_idle), Wait for all processors to update cached/local
1168                  * copies of pm_idle before proceeding.
1169                  */
1170                 cpu_idle_wait();
1171         }
1172
1173         return_VALUE(0);
1174 }