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, 2005 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 > 31)
 265                         diff = 31;
 266
 267                 pr->power.bm_activity <<= diff;
 268
 269                 acpi_get_register(ACPI_BITREG_BUS_MASTER_STATUS,
 270                                   &bm_status, ACPI_MTX_DO_NOT_LOCK);
 271                 if (bm_status) {
 272                         pr->power.bm_activity |= 0x1;
 273                         acpi_set_register(ACPI_BITREG_BUS_MASTER_STATUS,
 274                                           1, ACPI_MTX_DO_NOT_LOCK);
 275                 }
 276                 /*
 277                  * PIIX4 Erratum #18: Note that BM_STS doesn't always reflect
 278                  * the true state of bus mastering activity; forcing us to
 279                  * manually check the BMIDEA bit of each IDE channel.
 280                  */
 281                 else if (errata.piix4.bmisx) {
 282                         if ((inb_p(errata.piix4.bmisx + 0x02) & 0x01)
 283                             || (inb_p(errata.piix4.bmisx + 0x0A) & 0x01))
 284                                 pr->power.bm_activity |= 0x1;
 285                 }
 286
 287                 pr->power.bm_check_timestamp = jiffies;
 288
 289                 /*
 290                  * Apply bus mastering demotion policy.  Automatically demote
 291                  * to avoid a faulty transition.  Note that the processor
 292                  * won't enter a low-power state during this call (to this
 293                  * funciton) but should upon the next.
 294                  *
 295                  * TBD: A better policy might be to fallback to the demotion
 296                  *      state (use it for this quantum only) istead of
 297                  *      demoting -- and rely on duration as our sole demotion
 298                  *      qualification.  This may, however, introduce DMA
 299                  *      issues (e.g. floppy DMA transfer overrun/underrun).
 300                  */
 301                 if (pr->power.bm_activity & cx->demotion.threshold.bm) {
 302                         local_irq_enable();
 303                         next_state = cx->demotion.state;
 304                         goto end;
 305                 }
 306         }
 307
 308 #ifdef CONFIG_HOTPLUG_CPU
 309         /*
 310          * Check for P_LVL2_UP flag before entering C2 and above on
 311          * an SMP system. We do it here instead of doing it at _CST/P_LVL
 312          * detection phase, to work cleanly with logical CPU hotplug.
 313          */
 314         if ((cx->type != ACPI_STATE_C1) && (num_online_cpus() > 1) &&
 315             !pr->flags.has_cst && !acpi_fadt.plvl2_up)
 316                 cx = &pr->power.states[ACPI_STATE_C1];
 317 #endif
 318
 319         /*
 320          * Sleep:
 321          * ------
 322          * Invoke the current Cx state to put the processor to sleep.
 323          */
 324         if (cx->type == ACPI_STATE_C2 || cx->type == ACPI_STATE_C3) {
 325                 current_thread_info()->status &= ~TS_POLLING;
 326                 smp_mb__after_clear_bit();
 327                 if (need_resched()) {
 328                         current_thread_info()->status |= TS_POLLING;
 329                         local_irq_enable();
 330                         return;
 331                 }
 332         }
 333
 334         switch (cx->type) {
 335
 336         case ACPI_STATE_C1:
 337                 /*
 338                  * Invoke C1.
 339                  * Use the appropriate idle routine, the one that would
 340                  * be used without acpi C-states.
 341                  */
 342                 if (pm_idle_save)
 343                         pm_idle_save();
 344                 else
 345                         acpi_safe_halt();
 346
 347                 /*
 348                  * TBD: Can't get time duration while in C1, as resumes
 349                  *      go to an ISR rather than here.  Need to instrument
 350                  *      base interrupt handler.
 351                  */
 352                 sleep_ticks = 0xFFFFFFFF;
 353                 break;
 354
 355         case ACPI_STATE_C2:
 356                 /* Get start time (ticks) */
 357                 t1 = inl(acpi_fadt.xpm_tmr_blk.address);
 358                 /* Invoke C2 */
 359                 inb(cx->address);
 360                 /* Dummy op - must do something useless after P_LVL2 read */
 361                 t2 = inl(acpi_fadt.xpm_tmr_blk.address);
 362                 /* Get end time (ticks) */
 363                 t2 = inl(acpi_fadt.xpm_tmr_blk.address);
 364
 365 #ifdef CONFIG_GENERIC_TIME
 366                 /* TSC halts in C2, so notify users */
 367                 mark_tsc_unstable();
 368 #endif
 369                 /* Re-enable interrupts */
 370                 local_irq_enable();
 371                 current_thread_info()->status |= TS_POLLING;
 372                 /* Compute time (ticks) that we were actually asleep */
 373                 sleep_ticks =
 374                     ticks_elapsed(t1, t2) - cx->latency_ticks - C2_OVERHEAD;
 375                 break;
 376
 377         case ACPI_STATE_C3:
 378
 379                 if (pr->flags.bm_check) {
 380                         if (atomic_inc_return(&c3_cpu_count) ==
 381                             num_online_cpus()) {
 382                                 /*
 383                                  * All CPUs are trying to go to C3
 384                                  * Disable bus master arbitration
 385                                  */
 386                                 acpi_set_register(ACPI_BITREG_ARB_DISABLE, 1,
 387                                                   ACPI_MTX_DO_NOT_LOCK);
 388                         }
 389                 } else {
 390                         /* SMP with no shared cache... Invalidate cache  */
 391                         ACPI_FLUSH_CPU_CACHE();
 392                 }
 393
 394                 /* Get start time (ticks) */
 395                 t1 = inl(acpi_fadt.xpm_tmr_blk.address);
 396                 /* Invoke C3 */
 397                 inb(cx->address);
 398                 /* Dummy op - must do something useless after P_LVL3 read */
 399                 t2 = inl(acpi_fadt.xpm_tmr_blk.address);
 400                 /* Get end time (ticks) */
 401                 t2 = inl(acpi_fadt.xpm_tmr_blk.address);
 402                 if (pr->flags.bm_check) {
 403                         /* Enable bus master arbitration */
 404                         atomic_dec(&c3_cpu_count);
 405                         acpi_set_register(ACPI_BITREG_ARB_DISABLE, 0,
 406                                           ACPI_MTX_DO_NOT_LOCK);
 407                 }
 408
 409 #ifdef CONFIG_GENERIC_TIME
 410                 /* TSC halts in C3, so notify users */
 411                 mark_tsc_unstable();
 412 #endif
 413                 /* Re-enable interrupts */
 414                 local_irq_enable();
 415                 current_thread_info()->status |= TS_POLLING;
 416                 /* Compute time (ticks) that we were actually asleep */
 417                 sleep_ticks =
 418                     ticks_elapsed(t1, t2) - cx->latency_ticks - C3_OVERHEAD;
 419                 break;
 420
 421         default:
 422                 local_irq_enable();
 423                 return;
 424         }
 425         cx->usage++;
 426         if ((cx->type != ACPI_STATE_C1) && (sleep_ticks > 0))
 427                 cx->time += sleep_ticks;
 428
 429         next_state = pr->power.state;
 430
 431 #ifdef CONFIG_HOTPLUG_CPU
 432         /* Don't do promotion/demotion */
 433         if ((cx->type == ACPI_STATE_C1) && (num_online_cpus() > 1) &&
 434             !pr->flags.has_cst && !acpi_fadt.plvl2_up) {
 435                 next_state = cx;
 436                 goto end;
 437         }
 438 #endif
 439
 440         /*
 441          * Promotion?
 442          * ----------
 443          * Track the number of longs (time asleep is greater than threshold)
 444          * and promote when the count threshold is reached.  Note that bus
 445          * mastering activity may prevent promotions.
 446          * Do not promote above max_cstate.
 447          */
 448         if (cx->promotion.state &&
 449             ((cx->promotion.state - pr->power.states) <= max_cstate)) {
 450                 if (sleep_ticks > cx->promotion.threshold.ticks) {
 451                         cx->promotion.count++;
 452                         cx->demotion.count = 0;
 453                         if (cx->promotion.count >=
 454                             cx->promotion.threshold.count) {
 455                                 if (pr->flags.bm_check) {
 456                                         if (!
 457                                             (pr->power.bm_activity & cx->
 458                                              promotion.threshold.bm)) {
 459                                                 next_state =
 460                                                     cx->promotion.state;
 461                                                 goto end;
 462                                         }
 463                                 } else {
 464                                         next_state = cx->promotion.state;
 465                                         goto end;
 466                                 }
 467                         }
 468                 }
 469         }
 470
 471         /*
 472          * Demotion?
 473          * ---------
 474          * Track the number of shorts (time asleep is less than time threshold)
 475          * and demote when the usage threshold is reached.
 476          */
 477         if (cx->demotion.state) {
 478                 if (sleep_ticks < cx->demotion.threshold.ticks) {
 479                         cx->demotion.count++;
 480                         cx->promotion.count = 0;
 481                         if (cx->demotion.count >= cx->demotion.threshold.count) {
 482                                 next_state = cx->demotion.state;
 483                                 goto end;
 484                         }
 485                 }
 486         }
 487
 488       end:
 489         /*
 490          * Demote if current state exceeds max_cstate
 491          */
 492         if ((pr->power.state - pr->power.states) > max_cstate) {
 493                 if (cx->demotion.state)
 494                         next_state = cx->demotion.state;
 495         }
 496
 497         /*
 498          * New Cx State?
 499          * -------------
 500          * If we're going to start using a new Cx state we must clean up
 501          * from the previous and prepare to use the new.
 502          */
 503         if (next_state != pr->power.state)
 504                 acpi_processor_power_activate(pr, next_state);
 505 }
 506
 507 static int acpi_processor_set_power_policy(struct acpi_processor *pr)
 508 {
 509         unsigned int i;
 510         unsigned int state_is_set = 0;
 511         struct acpi_processor_cx *lower = NULL;
 512         struct acpi_processor_cx *higher = NULL;
 513         struct acpi_processor_cx *cx;
 514
 515
 516         if (!pr)
 517                 return -EINVAL;
 518
 519         /*
 520          * This function sets the default Cx state policy (OS idle handler).
 521          * Our scheme is to promote quickly to C2 but more conservatively
 522          * to C3.  We're favoring C2  for its characteristics of low latency
 523          * (quick response), good power savings, and ability to allow bus
 524          * mastering activity.  Note that the Cx state policy is completely
 525          * customizable and can be altered dynamically.
 526          */
 527
 528         /* startup state */
 529         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
 530                 cx = &pr->power.states[i];
 531                 if (!cx->valid)
 532                         continue;
 533
 534                 if (!state_is_set)
 535                         pr->power.state = cx;
 536                 state_is_set++;
 537                 break;
 538         }
 539
 540         if (!state_is_set)
 541                 return -ENODEV;
 542
 543         /* demotion */
 544         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
 545                 cx = &pr->power.states[i];
 546                 if (!cx->valid)
 547                         continue;
 548
 549                 if (lower) {
 550                         cx->demotion.state = lower;
 551                         cx->demotion.threshold.ticks = cx->latency_ticks;
 552                         cx->demotion.threshold.count = 1;
 553                         if (cx->type == ACPI_STATE_C3)
 554                                 cx->demotion.threshold.bm = bm_history;
 555                 }
 556
 557                 lower = cx;
 558         }
 559
 560         /* promotion */
 561         for (i = (ACPI_PROCESSOR_MAX_POWER - 1); i > 0; i--) {
 562                 cx = &pr->power.states[i];
 563                 if (!cx->valid)
 564                         continue;
 565
 566                 if (higher) {
 567                         cx->promotion.state = higher;
 568                         cx->promotion.threshold.ticks = cx->latency_ticks;
 569                         if (cx->type >= ACPI_STATE_C2)
 570                                 cx->promotion.threshold.count = 4;
 571                         else
 572                                 cx->promotion.threshold.count = 10;
 573                         if (higher->type == ACPI_STATE_C3)
 574                                 cx->promotion.threshold.bm = bm_history;
 575                 }
 576
 577                 higher = cx;
 578         }
 579
 580         return 0;
 581 }
 582
 583 static int acpi_processor_get_power_info_fadt(struct acpi_processor *pr)
 584 {
 585
 586         if (!pr)
 587                 return -EINVAL;
 588
 589         if (!pr->pblk)
 590                 return -ENODEV;
 591
 592         /* if info is obtained from pblk/fadt, type equals state */
 593         pr->power.states[ACPI_STATE_C2].type = ACPI_STATE_C2;
 594         pr->power.states[ACPI_STATE_C3].type = ACPI_STATE_C3;
 595
 596 #ifndef CONFIG_HOTPLUG_CPU
 597         /*
 598          * Check for P_LVL2_UP flag before entering C2 and above on
 599          * an SMP system.
 600          */
 601         if ((num_online_cpus() > 1) && !acpi_fadt.plvl2_up)
 602                 return -ENODEV;
 603 #endif
 604
 605         /* determine C2 and C3 address from pblk */
 606         pr->power.states[ACPI_STATE_C2].address = pr->pblk + 4;
 607         pr->power.states[ACPI_STATE_C3].address = pr->pblk + 5;
 608
 609         /* determine latencies from FADT */
 610         pr->power.states[ACPI_STATE_C2].latency = acpi_fadt.plvl2_lat;
 611         pr->power.states[ACPI_STATE_C3].latency = acpi_fadt.plvl3_lat;
 612
 613         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 614                           "lvl2[0x%08x] lvl3[0x%08x]\n",
 615                           pr->power.states[ACPI_STATE_C2].address,
 616                           pr->power.states[ACPI_STATE_C3].address));
 617
 618         return 0;
 619 }
 620
 621 static int acpi_processor_get_power_info_default_c1(struct acpi_processor *pr)
 622 {
 623
 624         /* Zero initialize all the C-states info. */
 625         memset(pr->power.states, 0, sizeof(pr->power.states));
 626
 627         /* set the first C-State to C1 */
 628         pr->power.states[ACPI_STATE_C1].type = ACPI_STATE_C1;
 629
 630         /* the C0 state only exists as a filler in our array,
 631          * and all processors need to support C1 */
 632         pr->power.states[ACPI_STATE_C0].valid = 1;
 633         pr->power.states[ACPI_STATE_C1].valid = 1;
 634
 635         return 0;
 636 }
 637
 638 static int acpi_processor_get_power_info_cst(struct acpi_processor *pr)
 639 {
 640         acpi_status status = 0;
 641         acpi_integer count;
 642         int current_count;
 643         int i;
 644         struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
 645         union acpi_object *cst;
 646
 647
 648         if (nocst)
 649                 return -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 -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                 printk(KERN_ERR PREFIX "not enough elements in _CST\n");
 669                 status = -EFAULT;
 670                 goto end;
 671         }
 672
 673         count = cst->package.elements[0].integer.value;
 674
 675         /* Validate number of power states. */
 676         if (count < 1 || count != cst->package.count - 1) {
 677                 printk(KERN_ERR PREFIX "count given by _CST is not valid\n");
 678                 status = -EFAULT;
 679                 goto end;
 680         }
 681
 682         /* Tell driver that at least _CST is supported. */
 683         pr->flags.has_cst = 1;
 684
 685         for (i = 1; i <= count; i++) {
 686                 union acpi_object *element;
 687                 union acpi_object *obj;
 688                 struct acpi_power_register *reg;
 689                 struct acpi_processor_cx cx;
 690
 691                 memset(&cx, 0, sizeof(cx));
 692
 693                 element = (union acpi_object *)&(cst->package.elements[i]);
 694                 if (element->type != ACPI_TYPE_PACKAGE)
 695                         continue;
 696
 697                 if (element->package.count != 4)
 698                         continue;
 699
 700                 obj = (union acpi_object *)&(element->package.elements[0]);
 701
 702                 if (obj->type != ACPI_TYPE_BUFFER)
 703                         continue;
 704
 705                 reg = (struct acpi_power_register *)obj->buffer.pointer;
 706
 707                 if (reg->space_id != ACPI_ADR_SPACE_SYSTEM_IO &&
 708                     (reg->space_id != ACPI_ADR_SPACE_FIXED_HARDWARE))
 709                         continue;
 710
 711                 cx.address = (reg->space_id == ACPI_ADR_SPACE_FIXED_HARDWARE) ?
 712                     0 : reg->address;
 713
 714                 /* There should be an easy way to extract an integer... */
 715                 obj = (union acpi_object *)&(element->package.elements[1]);
 716                 if (obj->type != ACPI_TYPE_INTEGER)
 717                         continue;
 718
 719                 cx.type = obj->integer.value;
 720
 721                 if ((cx.type != ACPI_STATE_C1) &&
 722                     (reg->space_id != ACPI_ADR_SPACE_SYSTEM_IO))
 723                         continue;
 724
 725                 if ((cx.type < ACPI_STATE_C2) || (cx.type > ACPI_STATE_C3))
 726                         continue;
 727
 728                 obj = (union acpi_object *)&(element->package.elements[2]);
 729                 if (obj->type != ACPI_TYPE_INTEGER)
 730                         continue;
 731
 732                 cx.latency = obj->integer.value;
 733
 734                 obj = (union acpi_object *)&(element->package.elements[3]);
 735                 if (obj->type != ACPI_TYPE_INTEGER)
 736                         continue;
 737
 738                 cx.power = obj->integer.value;
 739
 740                 current_count++;
 741                 memcpy(&(pr->power.states[current_count]), &cx, sizeof(cx));
 742
 743                 /*
 744                  * We support total ACPI_PROCESSOR_MAX_POWER - 1
 745                  * (From 1 through ACPI_PROCESSOR_MAX_POWER - 1)
 746                  */
 747                 if (current_count >= (ACPI_PROCESSOR_MAX_POWER - 1)) {
 748                         printk(KERN_WARNING
 749                                "Limiting number of power states to max (%d)\n",
 750                                ACPI_PROCESSOR_MAX_POWER);
 751                         printk(KERN_WARNING
 752                                "Please increase ACPI_PROCESSOR_MAX_POWER if needed.\n");
 753                         break;
 754                 }
 755         }
 756
 757         ACPI_DEBUG_PRINT((ACPI_DB_INFO, "Found %d power states\n",
 758                           current_count));
 759
 760         /* Validate number of power states discovered */
 761         if (current_count < 2)
 762                 status = -EFAULT;
 763
 764       end:
 765         acpi_os_free(buffer.pointer);
 766
 767         return status;
 768 }
 769
 770 static void acpi_processor_power_verify_c2(struct acpi_processor_cx *cx)
 771 {
 772
 773         if (!cx->address)
 774                 return;
 775
 776         /*
 777          * C2 latency must be less than or equal to 100
 778          * microseconds.
 779          */
 780         else if (cx->latency > ACPI_PROCESSOR_MAX_C2_LATENCY) {
 781                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 782                                   "latency too large [%d]\n", cx->latency));
 783                 return;
 784         }
 785
 786         /*
 787          * Otherwise we've met all of our C2 requirements.
 788          * Normalize the C2 latency to expidite policy
 789          */
 790         cx->valid = 1;
 791         cx->latency_ticks = US_TO_PM_TIMER_TICKS(cx->latency);
 792
 793         return;
 794 }
 795
 796 static void acpi_processor_power_verify_c3(struct acpi_processor *pr,
 797                                            struct acpi_processor_cx *cx)
 798 {
 799         static int bm_check_flag;
 800
 801
 802         if (!cx->address)
 803                 return;
 804
 805         /*
 806          * C3 latency must be less than or equal to 1000
 807          * microseconds.
 808          */
 809         else if (cx->latency > ACPI_PROCESSOR_MAX_C3_LATENCY) {
 810                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 811                                   "latency too large [%d]\n", cx->latency));
 812                 return;
 813         }
 814
 815         /*
 816          * PIIX4 Erratum #18: We don't support C3 when Type-F (fast)
 817          * DMA transfers are used by any ISA device to avoid livelock.
 818          * Note that we could disable Type-F DMA (as recommended by
 819          * the erratum), but this is known to disrupt certain ISA
 820          * devices thus we take the conservative approach.
 821          */
 822         else if (errata.piix4.fdma) {
 823                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 824                                   "C3 not supported on PIIX4 with Type-F DMA\n"));
 825                 return;
 826         }
 827
 828         /* All the logic here assumes flags.bm_check is same across all CPUs */
 829         if (!bm_check_flag) {
 830                 /* Determine whether bm_check is needed based on CPU  */
 831                 acpi_processor_power_init_bm_check(&(pr->flags), pr->id);
 832                 bm_check_flag = pr->flags.bm_check;
 833         } else {
 834                 pr->flags.bm_check = bm_check_flag;
 835         }
 836
 837         if (pr->flags.bm_check) {
 838                 /* bus mastering control is necessary */
 839                 if (!pr->flags.bm_control) {
 840                         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 841                                           "C3 support requires bus mastering control\n"));
 842                         return;
 843                 }
 844         } else {
 845                 /*
 846                  * WBINVD should be set in fadt, for C3 state to be
 847                  * supported on when bm_check is not required.
 848                  */
 849                 if (acpi_fadt.wb_invd != 1) {
 850                         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 851                                           "Cache invalidation should work properly"
 852                                           " for C3 to be enabled on SMP systems\n"));
 853                         return;
 854                 }
 855                 acpi_set_register(ACPI_BITREG_BUS_MASTER_RLD,
 856                                   0, ACPI_MTX_DO_NOT_LOCK);
 857         }
 858
 859         /*
 860          * Otherwise we've met all of our C3 requirements.
 861          * Normalize the C3 latency to expidite policy.  Enable
 862          * checking of bus mastering status (bm_check) so we can
 863          * use this in our C3 policy
 864          */
 865         cx->valid = 1;
 866         cx->latency_ticks = US_TO_PM_TIMER_TICKS(cx->latency);
 867
 868         return;
 869 }
 870
 871 static int acpi_processor_power_verify(struct acpi_processor *pr)
 872 {
 873         unsigned int i;
 874         unsigned int working = 0;
 875
 876 #ifdef ARCH_APICTIMER_STOPS_ON_C3
 877         int timer_broadcast = 0;
 878         cpumask_t mask = cpumask_of_cpu(pr->id);
 879         on_each_cpu(switch_ipi_to_APIC_timer, &mask, 1, 1);
 880 #endif
 881
 882         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
 883                 struct acpi_processor_cx *cx = &pr->power.states[i];
 884
 885                 switch (cx->type) {
 886                 case ACPI_STATE_C1:
 887                         cx->valid = 1;
 888                         break;
 889
 890                 case ACPI_STATE_C2:
 891                         acpi_processor_power_verify_c2(cx);
 892 #ifdef ARCH_APICTIMER_STOPS_ON_C3
 893                         /* Some AMD systems fake C3 as C2, but still
 894                            have timer troubles */
 895                         if (cx->valid &&
 896                                 boot_cpu_data.x86_vendor == X86_VENDOR_AMD)
 897                                 timer_broadcast++;
 898 #endif
 899                         break;
 900
 901                 case ACPI_STATE_C3:
 902                         acpi_processor_power_verify_c3(pr, cx);
 903 #ifdef ARCH_APICTIMER_STOPS_ON_C3
 904                         if (cx->valid)
 905                                 timer_broadcast++;
 906 #endif
 907                         break;
 908                 }
 909
 910                 if (cx->valid)
 911                         working++;
 912         }
 913
 914 #ifdef ARCH_APICTIMER_STOPS_ON_C3
 915         if (timer_broadcast)
 916                 on_each_cpu(switch_APIC_timer_to_ipi, &mask, 1, 1);
 917 #endif
 918
 919         return (working);
 920 }
 921
 922 static int acpi_processor_get_power_info(struct acpi_processor *pr)
 923 {
 924         unsigned int i;
 925         int result;
 926
 927
 928         /* NOTE: the idle thread may not be running while calling
 929          * this function */
 930
 931         /* Adding C1 state */
 932         acpi_processor_get_power_info_default_c1(pr);
 933         result = acpi_processor_get_power_info_cst(pr);
 934         if (result == -ENODEV)
 935                 acpi_processor_get_power_info_fadt(pr);
 936
 937         pr->power.count = acpi_processor_power_verify(pr);
 938
 939         /*
 940          * Set Default Policy
 941          * ------------------
 942          * Now that we know which states are supported, set the default
 943          * policy.  Note that this policy can be changed dynamically
 944          * (e.g. encourage deeper sleeps to conserve battery life when
 945          * not on AC).
 946          */
 947         result = acpi_processor_set_power_policy(pr);
 948         if (result)
 949                 return result;
 950
 951         /*
 952          * if one state of type C2 or C3 is available, mark this
 953          * CPU as being "idle manageable"
 954          */
 955         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
 956                 if (pr->power.states[i].valid) {
 957                         pr->power.count = i;
 958                         if (pr->power.states[i].type >= ACPI_STATE_C2)
 959                                 pr->flags.power = 1;
 960                 }
 961         }
 962
 963         return 0;
 964 }
 965
 966 int acpi_processor_cst_has_changed(struct acpi_processor *pr)
 967 {
 968         int result = 0;
 969
 970
 971         if (!pr)
 972                 return -EINVAL;
 973
 974         if (nocst) {
 975                 return -ENODEV;
 976         }
 977
 978         if (!pr->flags.power_setup_done)
 979                 return -ENODEV;
 980
 981         /* Fall back to the default idle loop */
 982         pm_idle = pm_idle_save;
 983         synchronize_sched();    /* Relies on interrupts forcing exit from idle. */
 984
 985         pr->flags.power = 0;
 986         result = acpi_processor_get_power_info(pr);
 987         if ((pr->flags.power == 1) && (pr->flags.power_setup_done))
 988                 pm_idle = acpi_processor_idle;
 989
 990         return result;
 991 }
 992
 993 /* proc interface */
 994
 995 static int acpi_processor_power_seq_show(struct seq_file *seq, void *offset)
 996 {
 997         struct acpi_processor *pr = (struct acpi_processor *)seq->private;
 998         unsigned int i;
 999
1000
1001         if (!pr)
1002                 goto end;
1003
1004         seq_printf(seq, "active state:            C%zd\n"
1005                    "max_cstate:              C%d\n"
1006                    "bus master activity:     %08x\n",
1007                    pr->power.state ? pr->power.state - pr->power.states : 0,
1008                    max_cstate, (unsigned)pr->power.bm_activity);
1009
1010         seq_puts(seq, "states:\n");
1011
1012         for (i = 1; i <= pr->power.count; i++) {
1013                 seq_printf(seq, "   %cC%d:                  ",
1014                            (&pr->power.states[i] ==
1015                             pr->power.state ? '*' : ' '), i);
1016
1017                 if (!pr->power.states[i].valid) {
1018                         seq_puts(seq, "<not supported>\n");
1019                         continue;
1020                 }
1021
1022                 switch (pr->power.states[i].type) {
1023                 case ACPI_STATE_C1:
1024                         seq_printf(seq, "type[C1] ");
1025                         break;
1026                 case ACPI_STATE_C2:
1027                         seq_printf(seq, "type[C2] ");
1028                         break;
1029                 case ACPI_STATE_C3:
1030                         seq_printf(seq, "type[C3] ");
1031                         break;
1032                 default:
1033                         seq_printf(seq, "type[--] ");
1034                         break;
1035                 }
1036
1037                 if (pr->power.states[i].promotion.state)
1038                         seq_printf(seq, "promotion[C%zd] ",
1039                                    (pr->power.states[i].promotion.state -
1040                                     pr->power.states));
1041                 else
1042                         seq_puts(seq, "promotion[--] ");
1043
1044                 if (pr->power.states[i].demotion.state)
1045                         seq_printf(seq, "demotion[C%zd] ",
1046                                    (pr->power.states[i].demotion.state -
1047                                     pr->power.states));
1048                 else
1049                         seq_puts(seq, "demotion[--] ");
1050
1051                 seq_printf(seq, "latency[%03d] usage[%08d] duration[%020llu]\n",
1052                            pr->power.states[i].latency,
1053                            pr->power.states[i].usage,
1054                            pr->power.states[i].time);
1055         }
1056
1057       end:
1058         return 0;
1059 }
1060
1061 static int acpi_processor_power_open_fs(struct inode *inode, struct file *file)
1062 {
1063         return single_open(file, acpi_processor_power_seq_show,
1064                            PDE(inode)->data);
1065 }
1066
1067 static struct file_operations acpi_processor_power_fops = {
1068         .open = acpi_processor_power_open_fs,
1069         .read = seq_read,
1070         .llseek = seq_lseek,
1071         .release = single_release,
1072 };
1073
1074 int acpi_processor_power_init(struct acpi_processor *pr,
1075                               struct acpi_device *device)
1076 {
1077         acpi_status status = 0;
1078         static int first_run;
1079         struct proc_dir_entry *entry = NULL;
1080         unsigned int i;
1081
1082
1083         if (!first_run) {
1084                 dmi_check_system(processor_power_dmi_table);
1085                 if (max_cstate < ACPI_C_STATES_MAX)
1086                         printk(KERN_NOTICE
1087                                "ACPI: processor limited to max C-state %d\n",
1088                                max_cstate);
1089                 first_run++;
1090         }
1091
1092         if (!pr)
1093                 return -EINVAL;
1094
1095         if (acpi_fadt.cst_cnt && !nocst) {
1096                 status =
1097                     acpi_os_write_port(acpi_fadt.smi_cmd, acpi_fadt.cst_cnt, 8);
1098                 if (ACPI_FAILURE(status)) {
1099                         ACPI_EXCEPTION((AE_INFO, status,
1100                                         "Notifying BIOS of _CST ability failed"));
1101                 }
1102         }
1103
1104         acpi_processor_get_power_info(pr);
1105
1106         /*
1107          * Install the idle handler if processor power management is supported.
1108          * Note that we use previously set idle handler will be used on
1109          * platforms that only support C1.
1110          */
1111         if ((pr->flags.power) && (!boot_option_idle_override)) {
1112                 printk(KERN_INFO PREFIX "CPU%d (power states:", pr->id);
1113                 for (i = 1; i <= pr->power.count; i++)
1114                         if (pr->power.states[i].valid)
1115                                 printk(" C%d[C%d]", i,
1116                                        pr->power.states[i].type);
1117                 printk(")\n");
1118
1119                 if (pr->id == 0) {
1120                         pm_idle_save = pm_idle;
1121                         pm_idle = acpi_processor_idle;
1122                 }
1123         }
1124
1125         /* 'power' [R] */
1126         entry = create_proc_entry(ACPI_PROCESSOR_FILE_POWER,
1127                                   S_IRUGO, acpi_device_dir(device));
1128         if (!entry)
1129                 return -EIO;
1130         else {
1131                 entry->proc_fops = &acpi_processor_power_fops;
1132                 entry->data = acpi_driver_data(device);
1133                 entry->owner = THIS_MODULE;
1134         }
1135
1136         pr->flags.power_setup_done = 1;
1137
1138         return 0;
1139 }
1140
1141 int acpi_processor_power_exit(struct acpi_processor *pr,
1142                               struct acpi_device *device)
1143 {
1144
1145         pr->flags.power_setup_done = 0;
1146
1147         if (acpi_device_dir(device))
1148                 remove_proc_entry(ACPI_PROCESSOR_FILE_POWER,
1149                                   acpi_device_dir(device));
1150
1151         /* Unregister the idle handler when processor #0 is removed. */
1152         if (pr->id == 0) {
1153                 pm_idle = pm_idle_save;
1154
1155                 /*
1156                  * We are about to unload the current idle thread pm callback
1157                  * (pm_idle), Wait for all processors to update cached/local
1158                  * copies of pm_idle before proceeding.
1159                  */
1160                 cpu_idle_wait();
1161         }
1162
1163         return 0;
1164 }