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