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 #include <linux/latency.h>
  42 #include <linux/clockchips.h>
  43 #include <linux/cpuidle.h>
  44
  45 /*
  46  * Include the apic definitions for x86 to have the APIC timer related defines
  47  * available also for UP (on SMP it gets magically included via linux/smp.h).
  48  * asm/acpi.h is not an option, as it would require more include magic. Also
  49  * creating an empty asm-ia64/apic.h would just trade pest vs. cholera.
  50  */
  51 #ifdef CONFIG_X86
  52 #include <asm/apic.h>
  53 #endif
  54
  55 #include <asm/io.h>
  56 #include <asm/uaccess.h>
  57
  58 #include <acpi/acpi_bus.h>
  59 #include <acpi/processor.h>
  60
  61 #define ACPI_PROCESSOR_COMPONENT        0x01000000
  62 #define ACPI_PROCESSOR_CLASS            "processor"
  63 #define _COMPONENT              ACPI_PROCESSOR_COMPONENT
  64 ACPI_MODULE_NAME("processor_idle");
  65 #define ACPI_PROCESSOR_FILE_POWER       "power"
  66 #define US_TO_PM_TIMER_TICKS(t)         ((t * (PM_TIMER_FREQUENCY/1000)) / 1000)
  67 #define PM_TIMER_TICK_NS                (1000000000ULL/PM_TIMER_FREQUENCY)
  68 #ifndef CONFIG_CPU_IDLE
  69 #define C2_OVERHEAD                     4       /* 1us (3.579 ticks per us) */
  70 #define C3_OVERHEAD                     4       /* 1us (3.579 ticks per us) */
  71 static void (*pm_idle_save) (void) __read_mostly;
  72 #else
  73 #define C2_OVERHEAD                     1       /* 1us */
  74 #define C3_OVERHEAD                     1       /* 1us */
  75 #endif
  76 #define PM_TIMER_TICKS_TO_US(p)         (((p) * 1000)/(PM_TIMER_FREQUENCY/1000))
  77
  78 static unsigned int max_cstate __read_mostly = ACPI_PROCESSOR_MAX_POWER;
  79 #ifdef CONFIG_CPU_IDLE
  80 module_param(max_cstate, uint, 0000);
  81 #else
  82 module_param(max_cstate, uint, 0644);
  83 #endif
  84 static unsigned int nocst __read_mostly;
  85 module_param(nocst, uint, 0000);
  86
  87 #ifndef CONFIG_CPU_IDLE
  88 /*
  89  * bm_history -- bit-mask with a bit per jiffy of bus-master activity
  90  * 1000 HZ: 0xFFFFFFFF: 32 jiffies = 32ms
  91  * 800 HZ: 0xFFFFFFFF: 32 jiffies = 40ms
  92  * 100 HZ: 0x0000000F: 4 jiffies = 40ms
  93  * reduce history for more aggressive entry into C3
  94  */
  95 static unsigned int bm_history __read_mostly =
  96     (HZ >= 800 ? 0xFFFFFFFF : ((1U << (HZ / 25)) - 1));
  97 module_param(bm_history, uint, 0644);
  98
  99 static int acpi_processor_set_power_policy(struct acpi_processor *pr);
 100
 101 #endif
 102
 103 /*
 104  * IBM ThinkPad R40e crashes mysteriously when going into C2 or C3.
 105  * For now disable this. Probably a bug somewhere else.
 106  *
 107  * To skip this limit, boot/load with a large max_cstate limit.
 108  */
 109 static int set_max_cstate(const struct dmi_system_id *id)
 110 {
 111         if (max_cstate > ACPI_PROCESSOR_MAX_POWER)
 112                 return 0;
 113
 114         printk(KERN_NOTICE PREFIX "%s detected - limiting to C%ld max_cstate."
 115                " Override with \"processor.max_cstate=%d\"\n", id->ident,
 116                (long)id->driver_data, ACPI_PROCESSOR_MAX_POWER + 1);
 117
 118         max_cstate = (long)id->driver_data;
 119
 120         return 0;
 121 }
 122
 123 /* Actually this shouldn't be __cpuinitdata, would be better to fix the
 124    callers to only run once -AK */
 125 static struct dmi_system_id __cpuinitdata processor_power_dmi_table[] = {
 126         { set_max_cstate, "IBM ThinkPad R40e", {
 127           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 128           DMI_MATCH(DMI_BIOS_VERSION,"1SET70WW")}, (void *)1},
 129         { set_max_cstate, "IBM ThinkPad R40e", {
 130           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 131           DMI_MATCH(DMI_BIOS_VERSION,"1SET60WW")}, (void *)1},
 132         { set_max_cstate, "IBM ThinkPad R40e", {
 133           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 134           DMI_MATCH(DMI_BIOS_VERSION,"1SET43WW") }, (void*)1},
 135         { set_max_cstate, "IBM ThinkPad R40e", {
 136           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 137           DMI_MATCH(DMI_BIOS_VERSION,"1SET45WW") }, (void*)1},
 138         { set_max_cstate, "IBM ThinkPad R40e", {
 139           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 140           DMI_MATCH(DMI_BIOS_VERSION,"1SET47WW") }, (void*)1},
 141         { set_max_cstate, "IBM ThinkPad R40e", {
 142           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 143           DMI_MATCH(DMI_BIOS_VERSION,"1SET50WW") }, (void*)1},
 144         { set_max_cstate, "IBM ThinkPad R40e", {
 145           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 146           DMI_MATCH(DMI_BIOS_VERSION,"1SET52WW") }, (void*)1},
 147         { set_max_cstate, "IBM ThinkPad R40e", {
 148           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 149           DMI_MATCH(DMI_BIOS_VERSION,"1SET55WW") }, (void*)1},
 150         { set_max_cstate, "IBM ThinkPad R40e", {
 151           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 152           DMI_MATCH(DMI_BIOS_VERSION,"1SET56WW") }, (void*)1},
 153         { set_max_cstate, "IBM ThinkPad R40e", {
 154           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 155           DMI_MATCH(DMI_BIOS_VERSION,"1SET59WW") }, (void*)1},
 156         { set_max_cstate, "IBM ThinkPad R40e", {
 157           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 158           DMI_MATCH(DMI_BIOS_VERSION,"1SET60WW") }, (void*)1},
 159         { set_max_cstate, "IBM ThinkPad R40e", {
 160           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 161           DMI_MATCH(DMI_BIOS_VERSION,"1SET61WW") }, (void*)1},
 162         { set_max_cstate, "IBM ThinkPad R40e", {
 163           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 164           DMI_MATCH(DMI_BIOS_VERSION,"1SET62WW") }, (void*)1},
 165         { set_max_cstate, "IBM ThinkPad R40e", {
 166           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 167           DMI_MATCH(DMI_BIOS_VERSION,"1SET64WW") }, (void*)1},
 168         { set_max_cstate, "IBM ThinkPad R40e", {
 169           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 170           DMI_MATCH(DMI_BIOS_VERSION,"1SET65WW") }, (void*)1},
 171         { set_max_cstate, "IBM ThinkPad R40e", {
 172           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 173           DMI_MATCH(DMI_BIOS_VERSION,"1SET68WW") }, (void*)1},
 174         { set_max_cstate, "Medion 41700", {
 175           DMI_MATCH(DMI_BIOS_VENDOR,"Phoenix Technologies LTD"),
 176           DMI_MATCH(DMI_BIOS_VERSION,"R01-A1J")}, (void *)1},
 177         { set_max_cstate, "Clevo 5600D", {
 178           DMI_MATCH(DMI_BIOS_VENDOR,"Phoenix Technologies LTD"),
 179           DMI_MATCH(DMI_BIOS_VERSION,"SHE845M0.86C.0013.D.0302131307")},
 180          (void *)2},
 181         {},
 182 };
 183
 184 static inline u32 ticks_elapsed(u32 t1, u32 t2)
 185 {
 186         if (t2 >= t1)
 187                 return (t2 - t1);
 188         else if (!(acpi_gbl_FADT.flags & ACPI_FADT_32BIT_TIMER))
 189                 return (((0x00FFFFFF - t1) + t2) & 0x00FFFFFF);
 190         else
 191                 return ((0xFFFFFFFF - t1) + t2);
 192 }
 193
 194 static inline u32 ticks_elapsed_in_us(u32 t1, u32 t2)
 195 {
 196         if (t2 >= t1)
 197                 return PM_TIMER_TICKS_TO_US(t2 - t1);
 198         else if (!(acpi_gbl_FADT.flags & ACPI_FADT_32BIT_TIMER))
 199                 return PM_TIMER_TICKS_TO_US(((0x00FFFFFF - t1) + t2) & 0x00FFFFFF);
 200         else
 201                 return PM_TIMER_TICKS_TO_US((0xFFFFFFFF - t1) + t2);
 202 }
 203
 204 static void acpi_safe_halt(void)
 205 {
 206         current_thread_info()->status &= ~TS_POLLING;
 207         /*
 208          * TS_POLLING-cleared state must be visible before we
 209          * test NEED_RESCHED:
 210          */
 211         smp_mb();
 212         if (!need_resched())
 213                 safe_halt();
 214         current_thread_info()->status |= TS_POLLING;
 215 }
 216
 217 #ifndef CONFIG_CPU_IDLE
 218
 219 static void
 220 acpi_processor_power_activate(struct acpi_processor *pr,
 221                               struct acpi_processor_cx *new)
 222 {
 223         struct acpi_processor_cx *old;
 224
 225         if (!pr || !new)
 226                 return;
 227
 228         old = pr->power.state;
 229
 230         if (old)
 231                 old->promotion.count = 0;
 232         new->demotion.count = 0;
 233
 234         /* Cleanup from old state. */
 235         if (old) {
 236                 switch (old->type) {
 237                 case ACPI_STATE_C3:
 238                         /* Disable bus master reload */
 239                         if (new->type != ACPI_STATE_C3 && pr->flags.bm_check)
 240                                 acpi_set_register(ACPI_BITREG_BUS_MASTER_RLD, 0);
 241                         break;
 242                 }
 243         }
 244
 245         /* Prepare to use new state. */
 246         switch (new->type) {
 247         case ACPI_STATE_C3:
 248                 /* Enable bus master reload */
 249                 if (old->type != ACPI_STATE_C3 && pr->flags.bm_check)
 250                         acpi_set_register(ACPI_BITREG_BUS_MASTER_RLD, 1);
 251                 break;
 252         }
 253
 254         pr->power.state = new;
 255
 256         return;
 257 }
 258
 259 static atomic_t c3_cpu_count;
 260
 261 /* Common C-state entry for C2, C3, .. */
 262 static void acpi_cstate_enter(struct acpi_processor_cx *cstate)
 263 {
 264         if (cstate->space_id == ACPI_CSTATE_FFH) {
 265                 /* Call into architectural FFH based C-state */
 266                 acpi_processor_ffh_cstate_enter(cstate);
 267         } else {
 268                 int unused;
 269                 /* IO port based C-state */
 270                 inb(cstate->address);
 271                 /* Dummy wait op - must do something useless after P_LVL2 read
 272                    because chipsets cannot guarantee that STPCLK# signal
 273                    gets asserted in time to freeze execution properly. */
 274                 unused = inl(acpi_gbl_FADT.xpm_timer_block.address);
 275         }
 276 }
 277 #endif /* !CONFIG_CPU_IDLE */
 278
 279 #ifdef ARCH_APICTIMER_STOPS_ON_C3
 280
 281 /*
 282  * Some BIOS implementations switch to C3 in the published C2 state.
 283  * This seems to be a common problem on AMD boxen, but other vendors
 284  * are affected too. We pick the most conservative approach: we assume
 285  * that the local APIC stops in both C2 and C3.
 286  */
 287 static void acpi_timer_check_state(int state, struct acpi_processor *pr,
 288                                    struct acpi_processor_cx *cx)
 289 {
 290         struct acpi_processor_power *pwr = &pr->power;
 291         u8 type = local_apic_timer_c2_ok ? ACPI_STATE_C3 : ACPI_STATE_C2;
 292
 293         /*
 294          * Check, if one of the previous states already marked the lapic
 295          * unstable
 296          */
 297         if (pwr->timer_broadcast_on_state < state)
 298                 return;
 299
 300         if (cx->type >= type)
 301                 pr->power.timer_broadcast_on_state = state;
 302 }
 303
 304 static void acpi_propagate_timer_broadcast(struct acpi_processor *pr)
 305 {
 306         unsigned long reason;
 307
 308         reason = pr->power.timer_broadcast_on_state < INT_MAX ?
 309                 CLOCK_EVT_NOTIFY_BROADCAST_ON : CLOCK_EVT_NOTIFY_BROADCAST_OFF;
 310
 311         clockevents_notify(reason, &pr->id);
 312 }
 313
 314 /* Power(C) State timer broadcast control */
 315 static void acpi_state_timer_broadcast(struct acpi_processor *pr,
 316                                        struct acpi_processor_cx *cx,
 317                                        int broadcast)
 318 {
 319         int state = cx - pr->power.states;
 320
 321         if (state >= pr->power.timer_broadcast_on_state) {
 322                 unsigned long reason;
 323
 324                 reason = broadcast ?  CLOCK_EVT_NOTIFY_BROADCAST_ENTER :
 325                         CLOCK_EVT_NOTIFY_BROADCAST_EXIT;
 326                 clockevents_notify(reason, &pr->id);
 327         }
 328 }
 329
 330 #else
 331
 332 static void acpi_timer_check_state(int state, struct acpi_processor *pr,
 333                                    struct acpi_processor_cx *cstate) { }
 334 static void acpi_propagate_timer_broadcast(struct acpi_processor *pr) { }
 335 static void acpi_state_timer_broadcast(struct acpi_processor *pr,
 336                                        struct acpi_processor_cx *cx,
 337                                        int broadcast)
 338 {
 339 }
 340
 341 #endif
 342
 343 /*
 344  * Suspend / resume control
 345  */
 346 static int acpi_idle_suspend;
 347
 348 int acpi_processor_suspend(struct acpi_device * device, pm_message_t state)
 349 {
 350         acpi_idle_suspend = 1;
 351         return 0;
 352 }
 353
 354 int acpi_processor_resume(struct acpi_device * device)
 355 {
 356         acpi_idle_suspend = 0;
 357         return 0;
 358 }
 359
 360 #ifndef CONFIG_CPU_IDLE
 361 static void acpi_processor_idle(void)
 362 {
 363         struct acpi_processor *pr = NULL;
 364         struct acpi_processor_cx *cx = NULL;
 365         struct acpi_processor_cx *next_state = NULL;
 366         int sleep_ticks = 0;
 367         u32 t1, t2 = 0;
 368
 369         /*
 370          * Interrupts must be disabled during bus mastering calculations and
 371          * for C2/C3 transitions.
 372          */
 373         local_irq_disable();
 374
 375         pr = processors[smp_processor_id()];
 376         if (!pr) {
 377                 local_irq_enable();
 378                 return;
 379         }
 380
 381         /*
 382          * Check whether we truly need to go idle, or should
 383          * reschedule:
 384          */
 385         if (unlikely(need_resched())) {
 386                 local_irq_enable();
 387                 return;
 388         }
 389
 390         cx = pr->power.state;
 391         if (!cx || acpi_idle_suspend) {
 392                 if (pm_idle_save)
 393                         pm_idle_save();
 394                 else
 395                         acpi_safe_halt();
 396                 return;
 397         }
 398
 399         /*
 400          * Check BM Activity
 401          * -----------------
 402          * Check for bus mastering activity (if required), record, and check
 403          * for demotion.
 404          */
 405         if (pr->flags.bm_check) {
 406                 u32 bm_status = 0;
 407                 unsigned long diff = jiffies - pr->power.bm_check_timestamp;
 408
 409                 if (diff > 31)
 410                         diff = 31;
 411
 412                 pr->power.bm_activity <<= diff;
 413
 414                 acpi_get_register(ACPI_BITREG_BUS_MASTER_STATUS, &bm_status);
 415                 if (bm_status) {
 416                         pr->power.bm_activity |= 0x1;
 417                         acpi_set_register(ACPI_BITREG_BUS_MASTER_STATUS, 1);
 418                 }
 419                 /*
 420                  * PIIX4 Erratum #18: Note that BM_STS doesn't always reflect
 421                  * the true state of bus mastering activity; forcing us to
 422                  * manually check the BMIDEA bit of each IDE channel.
 423                  */
 424                 else if (errata.piix4.bmisx) {
 425                         if ((inb_p(errata.piix4.bmisx + 0x02) & 0x01)
 426                             || (inb_p(errata.piix4.bmisx + 0x0A) & 0x01))
 427                                 pr->power.bm_activity |= 0x1;
 428                 }
 429
 430                 pr->power.bm_check_timestamp = jiffies;
 431
 432                 /*
 433                  * If bus mastering is or was active this jiffy, demote
 434                  * to avoid a faulty transition.  Note that the processor
 435                  * won't enter a low-power state during this call (to this
 436                  * function) but should upon the next.
 437                  *
 438                  * TBD: A better policy might be to fallback to the demotion
 439                  *      state (use it for this quantum only) istead of
 440                  *      demoting -- and rely on duration as our sole demotion
 441                  *      qualification.  This may, however, introduce DMA
 442                  *      issues (e.g. floppy DMA transfer overrun/underrun).
 443                  */
 444                 if ((pr->power.bm_activity & 0x1) &&
 445                     cx->demotion.threshold.bm) {
 446                         local_irq_enable();
 447                         next_state = cx->demotion.state;
 448                         goto end;
 449                 }
 450         }
 451
 452 #ifdef CONFIG_HOTPLUG_CPU
 453         /*
 454          * Check for P_LVL2_UP flag before entering C2 and above on
 455          * an SMP system. We do it here instead of doing it at _CST/P_LVL
 456          * detection phase, to work cleanly with logical CPU hotplug.
 457          */
 458         if ((cx->type != ACPI_STATE_C1) && (num_online_cpus() > 1) &&
 459             !pr->flags.has_cst && !(acpi_gbl_FADT.flags & ACPI_FADT_C2_MP_SUPPORTED))
 460                 cx = &pr->power.states[ACPI_STATE_C1];
 461 #endif
 462
 463         /*
 464          * Sleep:
 465          * ------
 466          * Invoke the current Cx state to put the processor to sleep.
 467          */
 468         if (cx->type == ACPI_STATE_C2 || cx->type == ACPI_STATE_C3) {
 469                 current_thread_info()->status &= ~TS_POLLING;
 470                 /*
 471                  * TS_POLLING-cleared state must be visible before we
 472                  * test NEED_RESCHED:
 473                  */
 474                 smp_mb();
 475                 if (need_resched()) {
 476                         current_thread_info()->status |= TS_POLLING;
 477                         local_irq_enable();
 478                         return;
 479                 }
 480         }
 481
 482         switch (cx->type) {
 483
 484         case ACPI_STATE_C1:
 485                 /*
 486                  * Invoke C1.
 487                  * Use the appropriate idle routine, the one that would
 488                  * be used without acpi C-states.
 489                  */
 490                 if (pm_idle_save)
 491                         pm_idle_save();
 492                 else
 493                         acpi_safe_halt();
 494
 495                 /*
 496                  * TBD: Can't get time duration while in C1, as resumes
 497                  *      go to an ISR rather than here.  Need to instrument
 498                  *      base interrupt handler.
 499                  *
 500                  * Note: the TSC better not stop in C1, sched_clock() will
 501                  *       skew otherwise.
 502                  */
 503                 sleep_ticks = 0xFFFFFFFF;
 504                 break;
 505
 506         case ACPI_STATE_C2:
 507                 /* Get start time (ticks) */
 508                 t1 = inl(acpi_gbl_FADT.xpm_timer_block.address);
 509                 /* Tell the scheduler that we are going deep-idle: */
 510                 sched_clock_idle_sleep_event();
 511                 /* Invoke C2 */
 512                 acpi_state_timer_broadcast(pr, cx, 1);
 513                 acpi_cstate_enter(cx);
 514                 /* Get end time (ticks) */
 515                 t2 = inl(acpi_gbl_FADT.xpm_timer_block.address);
 516
 517 #if defined (CONFIG_GENERIC_TIME) && defined (CONFIG_X86_TSC)
 518                 /* TSC halts in C2, so notify users */
 519                 mark_tsc_unstable("possible TSC halt in C2");
 520 #endif
 521                 /* Compute time (ticks) that we were actually asleep */
 522                 sleep_ticks = ticks_elapsed(t1, t2);
 523
 524                 /* Tell the scheduler how much we idled: */
 525                 sched_clock_idle_wakeup_event(sleep_ticks*PM_TIMER_TICK_NS);
 526
 527                 /* Re-enable interrupts */
 528                 local_irq_enable();
 529                 /* Do not account our idle-switching overhead: */
 530                 sleep_ticks -= cx->latency_ticks + C2_OVERHEAD;
 531
 532                 current_thread_info()->status |= TS_POLLING;
 533                 acpi_state_timer_broadcast(pr, cx, 0);
 534                 break;
 535
 536         case ACPI_STATE_C3:
 537                 /*
 538                  * Must be done before busmaster disable as we might
 539                  * need to access HPET !
 540                  */
 541                 acpi_state_timer_broadcast(pr, cx, 1);
 542                 /*
 543                  * disable bus master
 544                  * bm_check implies we need ARB_DIS
 545                  * !bm_check implies we need cache flush
 546                  * bm_control implies whether we can do ARB_DIS
 547                  *
 548                  * That leaves a case where bm_check is set and bm_control is
 549                  * not set. In that case we cannot do much, we enter C3
 550                  * without doing anything.
 551                  */
 552                 if (pr->flags.bm_check && pr->flags.bm_control) {
 553                         if (atomic_inc_return(&c3_cpu_count) ==
 554                             num_online_cpus()) {
 555                                 /*
 556                                  * All CPUs are trying to go to C3
 557                                  * Disable bus master arbitration
 558                                  */
 559                                 acpi_set_register(ACPI_BITREG_ARB_DISABLE, 1);
 560                         }
 561                 } else if (!pr->flags.bm_check) {
 562                         /* SMP with no shared cache... Invalidate cache  */
 563                         ACPI_FLUSH_CPU_CACHE();
 564                 }
 565
 566                 /* Get start time (ticks) */
 567                 t1 = inl(acpi_gbl_FADT.xpm_timer_block.address);
 568                 /* Invoke C3 */
 569                 /* Tell the scheduler that we are going deep-idle: */
 570                 sched_clock_idle_sleep_event();
 571                 acpi_cstate_enter(cx);
 572                 /* Get end time (ticks) */
 573                 t2 = inl(acpi_gbl_FADT.xpm_timer_block.address);
 574                 if (pr->flags.bm_check && pr->flags.bm_control) {
 575                         /* Enable bus master arbitration */
 576                         atomic_dec(&c3_cpu_count);
 577                         acpi_set_register(ACPI_BITREG_ARB_DISABLE, 0);
 578                 }
 579
 580 #if defined (CONFIG_GENERIC_TIME) && defined (CONFIG_X86_TSC)
 581                 /* TSC halts in C3, so notify users */
 582                 mark_tsc_unstable("TSC halts in C3");
 583 #endif
 584                 /* Compute time (ticks) that we were actually asleep */
 585                 sleep_ticks = ticks_elapsed(t1, t2);
 586                 /* Tell the scheduler how much we idled: */
 587                 sched_clock_idle_wakeup_event(sleep_ticks*PM_TIMER_TICK_NS);
 588
 589                 /* Re-enable interrupts */
 590                 local_irq_enable();
 591                 /* Do not account our idle-switching overhead: */
 592                 sleep_ticks -= cx->latency_ticks + C3_OVERHEAD;
 593
 594                 current_thread_info()->status |= TS_POLLING;
 595                 acpi_state_timer_broadcast(pr, cx, 0);
 596                 break;
 597
 598         default:
 599                 local_irq_enable();
 600                 return;
 601         }
 602         cx->usage++;
 603         if ((cx->type != ACPI_STATE_C1) && (sleep_ticks > 0))
 604                 cx->time += sleep_ticks;
 605
 606         next_state = pr->power.state;
 607
 608 #ifdef CONFIG_HOTPLUG_CPU
 609         /* Don't do promotion/demotion */
 610         if ((cx->type == ACPI_STATE_C1) && (num_online_cpus() > 1) &&
 611             !pr->flags.has_cst && !(acpi_gbl_FADT.flags & ACPI_FADT_C2_MP_SUPPORTED)) {
 612                 next_state = cx;
 613                 goto end;
 614         }
 615 #endif
 616
 617         /*
 618          * Promotion?
 619          * ----------
 620          * Track the number of longs (time asleep is greater than threshold)
 621          * and promote when the count threshold is reached.  Note that bus
 622          * mastering activity may prevent promotions.
 623          * Do not promote above max_cstate.
 624          */
 625         if (cx->promotion.state &&
 626             ((cx->promotion.state - pr->power.states) <= max_cstate)) {
 627                 if (sleep_ticks > cx->promotion.threshold.ticks &&
 628                   cx->promotion.state->latency <= system_latency_constraint()) {
 629                         cx->promotion.count++;
 630                         cx->demotion.count = 0;
 631                         if (cx->promotion.count >=
 632                             cx->promotion.threshold.count) {
 633                                 if (pr->flags.bm_check) {
 634                                         if (!
 635                                             (pr->power.bm_activity & cx->
 636                                              promotion.threshold.bm)) {
 637                                                 next_state =
 638                                                     cx->promotion.state;
 639                                                 goto end;
 640                                         }
 641                                 } else {
 642                                         next_state = cx->promotion.state;
 643                                         goto end;
 644                                 }
 645                         }
 646                 }
 647         }
 648
 649         /*
 650          * Demotion?
 651          * ---------
 652          * Track the number of shorts (time asleep is less than time threshold)
 653          * and demote when the usage threshold is reached.
 654          */
 655         if (cx->demotion.state) {
 656                 if (sleep_ticks < cx->demotion.threshold.ticks) {
 657                         cx->demotion.count++;
 658                         cx->promotion.count = 0;
 659                         if (cx->demotion.count >= cx->demotion.threshold.count) {
 660                                 next_state = cx->demotion.state;
 661                                 goto end;
 662                         }
 663                 }
 664         }
 665
 666       end:
 667         /*
 668          * Demote if current state exceeds max_cstate
 669          * or if the latency of the current state is unacceptable
 670          */
 671         if ((pr->power.state - pr->power.states) > max_cstate ||
 672                 pr->power.state->latency > system_latency_constraint()) {
 673                 if (cx->demotion.state)
 674                         next_state = cx->demotion.state;
 675         }
 676
 677         /*
 678          * New Cx State?
 679          * -------------
 680          * If we're going to start using a new Cx state we must clean up
 681          * from the previous and prepare to use the new.
 682          */
 683         if (next_state != pr->power.state)
 684                 acpi_processor_power_activate(pr, next_state);
 685 }
 686
 687 static int acpi_processor_set_power_policy(struct acpi_processor *pr)
 688 {
 689         unsigned int i;
 690         unsigned int state_is_set = 0;
 691         struct acpi_processor_cx *lower = NULL;
 692         struct acpi_processor_cx *higher = NULL;
 693         struct acpi_processor_cx *cx;
 694
 695
 696         if (!pr)
 697                 return -EINVAL;
 698
 699         /*
 700          * This function sets the default Cx state policy (OS idle handler).
 701          * Our scheme is to promote quickly to C2 but more conservatively
 702          * to C3.  We're favoring C2  for its characteristics of low latency
 703          * (quick response), good power savings, and ability to allow bus
 704          * mastering activity.  Note that the Cx state policy is completely
 705          * customizable and can be altered dynamically.
 706          */
 707
 708         /* startup state */
 709         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
 710                 cx = &pr->power.states[i];
 711                 if (!cx->valid)
 712                         continue;
 713
 714                 if (!state_is_set)
 715                         pr->power.state = cx;
 716                 state_is_set++;
 717                 break;
 718         }
 719
 720         if (!state_is_set)
 721                 return -ENODEV;
 722
 723         /* demotion */
 724         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
 725                 cx = &pr->power.states[i];
 726                 if (!cx->valid)
 727                         continue;
 728
 729                 if (lower) {
 730                         cx->demotion.state = lower;
 731                         cx->demotion.threshold.ticks = cx->latency_ticks;
 732                         cx->demotion.threshold.count = 1;
 733                         if (cx->type == ACPI_STATE_C3)
 734                                 cx->demotion.threshold.bm = bm_history;
 735                 }
 736
 737                 lower = cx;
 738         }
 739
 740         /* promotion */
 741         for (i = (ACPI_PROCESSOR_MAX_POWER - 1); i > 0; i--) {
 742                 cx = &pr->power.states[i];
 743                 if (!cx->valid)
 744                         continue;
 745
 746                 if (higher) {
 747                         cx->promotion.state = higher;
 748                         cx->promotion.threshold.ticks = cx->latency_ticks;
 749                         if (cx->type >= ACPI_STATE_C2)
 750                                 cx->promotion.threshold.count = 4;
 751                         else
 752                                 cx->promotion.threshold.count = 10;
 753                         if (higher->type == ACPI_STATE_C3)
 754                                 cx->promotion.threshold.bm = bm_history;
 755                 }
 756
 757                 higher = cx;
 758         }
 759
 760         return 0;
 761 }
 762 #endif /* !CONFIG_CPU_IDLE */
 763
 764 static int acpi_processor_get_power_info_fadt(struct acpi_processor *pr)
 765 {
 766
 767         if (!pr)
 768                 return -EINVAL;
 769
 770         if (!pr->pblk)
 771                 return -ENODEV;
 772
 773         /* if info is obtained from pblk/fadt, type equals state */
 774         pr->power.states[ACPI_STATE_C2].type = ACPI_STATE_C2;
 775         pr->power.states[ACPI_STATE_C3].type = ACPI_STATE_C3;
 776
 777 #ifndef CONFIG_HOTPLUG_CPU
 778         /*
 779          * Check for P_LVL2_UP flag before entering C2 and above on
 780          * an SMP system.
 781          */
 782         if ((num_online_cpus() > 1) &&
 783             !(acpi_gbl_FADT.flags & ACPI_FADT_C2_MP_SUPPORTED))
 784                 return -ENODEV;
 785 #endif
 786
 787         /* determine C2 and C3 address from pblk */
 788         pr->power.states[ACPI_STATE_C2].address = pr->pblk + 4;
 789         pr->power.states[ACPI_STATE_C3].address = pr->pblk + 5;
 790
 791         /* determine latencies from FADT */
 792         pr->power.states[ACPI_STATE_C2].latency = acpi_gbl_FADT.C2latency;
 793         pr->power.states[ACPI_STATE_C3].latency = acpi_gbl_FADT.C3latency;
 794
 795         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 796                           "lvl2[0x%08x] lvl3[0x%08x]\n",
 797                           pr->power.states[ACPI_STATE_C2].address,
 798                           pr->power.states[ACPI_STATE_C3].address));
 799
 800         return 0;
 801 }
 802
 803 static int acpi_processor_get_power_info_default(struct acpi_processor *pr)
 804 {
 805         if (!pr->power.states[ACPI_STATE_C1].valid) {
 806                 /* set the first C-State to C1 */
 807                 /* all processors need to support C1 */
 808                 pr->power.states[ACPI_STATE_C1].type = ACPI_STATE_C1;
 809                 pr->power.states[ACPI_STATE_C1].valid = 1;
 810         }
 811         /* the C0 state only exists as a filler in our array */
 812         pr->power.states[ACPI_STATE_C0].valid = 1;
 813         return 0;
 814 }
 815
 816 static int acpi_processor_get_power_info_cst(struct acpi_processor *pr)
 817 {
 818         acpi_status status = 0;
 819         acpi_integer count;
 820         int current_count;
 821         int i;
 822         struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
 823         union acpi_object *cst;
 824
 825
 826         if (nocst)
 827                 return -ENODEV;
 828
 829         current_count = 0;
 830
 831         status = acpi_evaluate_object(pr->handle, "_CST", NULL, &buffer);
 832         if (ACPI_FAILURE(status)) {
 833                 ACPI_DEBUG_PRINT((ACPI_DB_INFO, "No _CST, giving up\n"));
 834                 return -ENODEV;
 835         }
 836
 837         cst = buffer.pointer;
 838
 839         /* There must be at least 2 elements */
 840         if (!cst || (cst->type != ACPI_TYPE_PACKAGE) || cst->package.count < 2) {
 841                 printk(KERN_ERR PREFIX "not enough elements in _CST\n");
 842                 status = -EFAULT;
 843                 goto end;
 844         }
 845
 846         count = cst->package.elements[0].integer.value;
 847
 848         /* Validate number of power states. */
 849         if (count < 1 || count != cst->package.count - 1) {
 850                 printk(KERN_ERR PREFIX "count given by _CST is not valid\n");
 851                 status = -EFAULT;
 852                 goto end;
 853         }
 854
 855         /* Tell driver that at least _CST is supported. */
 856         pr->flags.has_cst = 1;
 857
 858         for (i = 1; i <= count; i++) {
 859                 union acpi_object *element;
 860                 union acpi_object *obj;
 861                 struct acpi_power_register *reg;
 862                 struct acpi_processor_cx cx;
 863
 864                 memset(&cx, 0, sizeof(cx));
 865
 866                 element = &(cst->package.elements[i]);
 867                 if (element->type != ACPI_TYPE_PACKAGE)
 868                         continue;
 869
 870                 if (element->package.count != 4)
 871                         continue;
 872
 873                 obj = &(element->package.elements[0]);
 874
 875                 if (obj->type != ACPI_TYPE_BUFFER)
 876                         continue;
 877
 878                 reg = (struct acpi_power_register *)obj->buffer.pointer;
 879
 880                 if (reg->space_id != ACPI_ADR_SPACE_SYSTEM_IO &&
 881                     (reg->space_id != ACPI_ADR_SPACE_FIXED_HARDWARE))
 882                         continue;
 883
 884                 /* There should be an easy way to extract an integer... */
 885                 obj = &(element->package.elements[1]);
 886                 if (obj->type != ACPI_TYPE_INTEGER)
 887                         continue;
 888
 889                 cx.type = obj->integer.value;
 890                 /*
 891                  * Some buggy BIOSes won't list C1 in _CST -
 892                  * Let acpi_processor_get_power_info_default() handle them later
 893                  */
 894                 if (i == 1 && cx.type != ACPI_STATE_C1)
 895                         current_count++;
 896
 897                 cx.address = reg->address;
 898                 cx.index = current_count + 1;
 899
 900                 cx.space_id = ACPI_CSTATE_SYSTEMIO;
 901                 if (reg->space_id == ACPI_ADR_SPACE_FIXED_HARDWARE) {
 902                         if (acpi_processor_ffh_cstate_probe
 903                                         (pr->id, &cx, reg) == 0) {
 904                                 cx.space_id = ACPI_CSTATE_FFH;
 905                         } else if (cx.type != ACPI_STATE_C1) {
 906                                 /*
 907                                  * C1 is a special case where FIXED_HARDWARE
 908                                  * can be handled in non-MWAIT way as well.
 909                                  * In that case, save this _CST entry info.
 910                                  * That is, we retain space_id of SYSTEM_IO for
 911                                  * halt based C1.
 912                                  * Otherwise, ignore this info and continue.
 913                                  */
 914                                 continue;
 915                         }
 916                 }
 917
 918                 obj = &(element->package.elements[2]);
 919                 if (obj->type != ACPI_TYPE_INTEGER)
 920                         continue;
 921
 922                 cx.latency = obj->integer.value;
 923
 924                 obj = &(element->package.elements[3]);
 925                 if (obj->type != ACPI_TYPE_INTEGER)
 926                         continue;
 927
 928                 cx.power = obj->integer.value;
 929
 930                 current_count++;
 931                 memcpy(&(pr->power.states[current_count]), &cx, sizeof(cx));
 932
 933                 /*
 934                  * We support total ACPI_PROCESSOR_MAX_POWER - 1
 935                  * (From 1 through ACPI_PROCESSOR_MAX_POWER - 1)
 936                  */
 937                 if (current_count >= (ACPI_PROCESSOR_MAX_POWER - 1)) {
 938                         printk(KERN_WARNING
 939                                "Limiting number of power states to max (%d)\n",
 940                                ACPI_PROCESSOR_MAX_POWER);
 941                         printk(KERN_WARNING
 942                                "Please increase ACPI_PROCESSOR_MAX_POWER if needed.\n");
 943                         break;
 944                 }
 945         }
 946
 947         ACPI_DEBUG_PRINT((ACPI_DB_INFO, "Found %d power states\n",
 948                           current_count));
 949
 950         /* Validate number of power states discovered */
 951         if (current_count < 2)
 952                 status = -EFAULT;
 953
 954       end:
 955         kfree(buffer.pointer);
 956
 957         return status;
 958 }
 959
 960 static void acpi_processor_power_verify_c2(struct acpi_processor_cx *cx)
 961 {
 962
 963         if (!cx->address)
 964                 return;
 965
 966         /*
 967          * C2 latency must be less than or equal to 100
 968          * microseconds.
 969          */
 970         else if (cx->latency > ACPI_PROCESSOR_MAX_C2_LATENCY) {
 971                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 972                                   "latency too large [%d]\n", cx->latency));
 973                 return;
 974         }
 975
 976         /*
 977          * Otherwise we've met all of our C2 requirements.
 978          * Normalize the C2 latency to expidite policy
 979          */
 980         cx->valid = 1;
 981
 982 #ifndef CONFIG_CPU_IDLE
 983         cx->latency_ticks = US_TO_PM_TIMER_TICKS(cx->latency);
 984 #else
 985         cx->latency_ticks = cx->latency;
 986 #endif
 987
 988         return;
 989 }
 990
 991 static void acpi_processor_power_verify_c3(struct acpi_processor *pr,
 992                                            struct acpi_processor_cx *cx)
 993 {
 994         static int bm_check_flag;
 995
 996
 997         if (!cx->address)
 998                 return;
 999
1000         /*
1001          * C3 latency must be less than or equal to 1000
1002          * microseconds.
1003          */
1004         else if (cx->latency > ACPI_PROCESSOR_MAX_C3_LATENCY) {
1005                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
1006                                   "latency too large [%d]\n", cx->latency));
1007                 return;
1008         }
1009
1010         /*
1011          * PIIX4 Erratum #18: We don't support C3 when Type-F (fast)
1012          * DMA transfers are used by any ISA device to avoid livelock.
1013          * Note that we could disable Type-F DMA (as recommended by
1014          * the erratum), but this is known to disrupt certain ISA
1015          * devices thus we take the conservative approach.
1016          */
1017         else if (errata.piix4.fdma) {
1018                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
1019                                   "C3 not supported on PIIX4 with Type-F DMA\n"));
1020                 return;
1021         }
1022
1023         /* All the logic here assumes flags.bm_check is same across all CPUs */
1024         if (!bm_check_flag) {
1025                 /* Determine whether bm_check is needed based on CPU  */
1026                 acpi_processor_power_init_bm_check(&(pr->flags), pr->id);
1027                 bm_check_flag = pr->flags.bm_check;
1028         } else {
1029                 pr->flags.bm_check = bm_check_flag;
1030         }
1031
1032         if (pr->flags.bm_check) {
1033                 if (!pr->flags.bm_control) {
1034                         if (pr->flags.has_cst != 1) {
1035                                 /* bus mastering control is necessary */
1036                                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
1037                                         "C3 support requires BM control\n"));
1038                                 return;
1039                         } else {
1040                                 /* Here we enter C3 without bus mastering */
1041                                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
1042                                         "C3 support without BM control\n"));
1043                         }
1044                 }
1045         } else {
1046                 /*
1047                  * WBINVD should be set in fadt, for C3 state to be
1048                  * supported on when bm_check is not required.
1049                  */
1050                 if (!(acpi_gbl_FADT.flags & ACPI_FADT_WBINVD)) {
1051                         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
1052                                           "Cache invalidation should work properly"
1053                                           " for C3 to be enabled on SMP systems\n"));
1054                         return;
1055                 }
1056                 acpi_set_register(ACPI_BITREG_BUS_MASTER_RLD, 0);
1057         }
1058
1059         /*
1060          * Otherwise we've met all of our C3 requirements.
1061          * Normalize the C3 latency to expidite policy.  Enable
1062          * checking of bus mastering status (bm_check) so we can
1063          * use this in our C3 policy
1064          */
1065         cx->valid = 1;
1066
1067 #ifndef CONFIG_CPU_IDLE
1068         cx->latency_ticks = US_TO_PM_TIMER_TICKS(cx->latency);
1069 #else
1070         cx->latency_ticks = cx->latency;
1071 #endif
1072
1073         return;
1074 }
1075
1076 static int acpi_processor_power_verify(struct acpi_processor *pr)
1077 {
1078         unsigned int i;
1079         unsigned int working = 0;
1080
1081         pr->power.timer_broadcast_on_state = INT_MAX;
1082
1083         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
1084                 struct acpi_processor_cx *cx = &pr->power.states[i];
1085
1086                 switch (cx->type) {
1087                 case ACPI_STATE_C1:
1088                         cx->valid = 1;
1089                         break;
1090
1091                 case ACPI_STATE_C2:
1092                         acpi_processor_power_verify_c2(cx);
1093                         if (cx->valid)
1094                                 acpi_timer_check_state(i, pr, cx);
1095                         break;
1096
1097                 case ACPI_STATE_C3:
1098                         acpi_processor_power_verify_c3(pr, cx);
1099                         if (cx->valid)
1100                                 acpi_timer_check_state(i, pr, cx);
1101                         break;
1102                 }
1103
1104                 if (cx->valid)
1105                         working++;
1106         }
1107
1108         acpi_propagate_timer_broadcast(pr);
1109
1110         return (working);
1111 }
1112
1113 static int acpi_processor_get_power_info(struct acpi_processor *pr)
1114 {
1115         unsigned int i;
1116         int result;
1117
1118
1119         /* NOTE: the idle thread may not be running while calling
1120          * this function */
1121
1122         /* Zero initialize all the C-states info. */
1123         memset(pr->power.states, 0, sizeof(pr->power.states));
1124
1125         result = acpi_processor_get_power_info_cst(pr);
1126         if (result == -ENODEV)
1127                 result = acpi_processor_get_power_info_fadt(pr);
1128
1129         if (result)
1130                 return result;
1131
1132         acpi_processor_get_power_info_default(pr);
1133
1134         pr->power.count = acpi_processor_power_verify(pr);
1135
1136 #ifndef CONFIG_CPU_IDLE
1137         /*
1138          * Set Default Policy
1139          * ------------------
1140          * Now that we know which states are supported, set the default
1141          * policy.  Note that this policy can be changed dynamically
1142          * (e.g. encourage deeper sleeps to conserve battery life when
1143          * not on AC).
1144          */
1145         result = acpi_processor_set_power_policy(pr);
1146         if (result)
1147                 return result;
1148 #endif
1149
1150         /*
1151          * if one state of type C2 or C3 is available, mark this
1152          * CPU as being "idle manageable"
1153          */
1154         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
1155                 if (pr->power.states[i].valid) {
1156                         pr->power.count = i;
1157                         if (pr->power.states[i].type >= ACPI_STATE_C2)
1158                                 pr->flags.power = 1;
1159                 }
1160         }
1161
1162         return 0;
1163 }
1164
1165 static int acpi_processor_power_seq_show(struct seq_file *seq, void *offset)
1166 {
1167         struct acpi_processor *pr = seq->private;
1168         unsigned int i;
1169
1170
1171         if (!pr)
1172                 goto end;
1173
1174         seq_printf(seq, "active state:            C%zd\n"
1175                    "max_cstate:              C%d\n"
1176                    "bus master activity:     %08x\n"
1177                    "maximum allowed latency: %d usec\n",
1178                    pr->power.state ? pr->power.state - pr->power.states : 0,
1179                    max_cstate, (unsigned)pr->power.bm_activity,
1180                    system_latency_constraint());
1181
1182         seq_puts(seq, "states:\n");
1183
1184         for (i = 1; i <= pr->power.count; i++) {
1185                 seq_printf(seq, "   %cC%d:                  ",
1186                            (&pr->power.states[i] ==
1187                             pr->power.state ? '*' : ' '), i);
1188
1189                 if (!pr->power.states[i].valid) {
1190                         seq_puts(seq, "<not supported>\n");
1191                         continue;
1192                 }
1193
1194                 switch (pr->power.states[i].type) {
1195                 case ACPI_STATE_C1:
1196                         seq_printf(seq, "type[C1] ");
1197                         break;
1198                 case ACPI_STATE_C2:
1199                         seq_printf(seq, "type[C2] ");
1200                         break;
1201                 case ACPI_STATE_C3:
1202                         seq_printf(seq, "type[C3] ");
1203                         break;
1204                 default:
1205                         seq_printf(seq, "type[--] ");
1206                         break;
1207                 }
1208
1209                 if (pr->power.states[i].promotion.state)
1210                         seq_printf(seq, "promotion[C%zd] ",
1211                                    (pr->power.states[i].promotion.state -
1212                                     pr->power.states));
1213                 else
1214                         seq_puts(seq, "promotion[--] ");
1215
1216                 if (pr->power.states[i].demotion.state)
1217                         seq_printf(seq, "demotion[C%zd] ",
1218                                    (pr->power.states[i].demotion.state -
1219                                     pr->power.states));
1220                 else
1221                         seq_puts(seq, "demotion[--] ");
1222
1223                 seq_printf(seq, "latency[%03d] usage[%08d] duration[%020llu]\n",
1224                            pr->power.states[i].latency,
1225                            pr->power.states[i].usage,
1226                            (unsigned long long)pr->power.states[i].time);
1227         }
1228
1229       end:
1230         return 0;
1231 }
1232
1233 static int acpi_processor_power_open_fs(struct inode *inode, struct file *file)
1234 {
1235         return single_open(file, acpi_processor_power_seq_show,
1236                            PDE(inode)->data);
1237 }
1238
1239 static const struct file_operations acpi_processor_power_fops = {
1240         .open = acpi_processor_power_open_fs,
1241         .read = seq_read,
1242         .llseek = seq_lseek,
1243         .release = single_release,
1244 };
1245
1246 #ifndef CONFIG_CPU_IDLE
1247
1248 int acpi_processor_cst_has_changed(struct acpi_processor *pr)
1249 {
1250         int result = 0;
1251
1252
1253         if (!pr)
1254                 return -EINVAL;
1255
1256         if (nocst) {
1257                 return -ENODEV;
1258         }
1259
1260         if (!pr->flags.power_setup_done)
1261                 return -ENODEV;
1262
1263         /* Fall back to the default idle loop */
1264         pm_idle = pm_idle_save;
1265         synchronize_sched();    /* Relies on interrupts forcing exit from idle. */
1266
1267         pr->flags.power = 0;
1268         result = acpi_processor_get_power_info(pr);
1269         if ((pr->flags.power == 1) && (pr->flags.power_setup_done))
1270                 pm_idle = acpi_processor_idle;
1271
1272         return result;
1273 }
1274
1275 #ifdef CONFIG_SMP
1276 static void smp_callback(void *v)
1277 {
1278         /* we already woke the CPU up, nothing more to do */
1279 }
1280
1281 /*
1282  * This function gets called when a part of the kernel has a new latency
1283  * requirement.  This means we need to get all processors out of their C-state,
1284  * and then recalculate a new suitable C-state. Just do a cross-cpu IPI; that
1285  * wakes them all right up.
1286  */
1287 static int acpi_processor_latency_notify(struct notifier_block *b,
1288                 unsigned long l, void *v)
1289 {
1290         smp_call_function(smp_callback, NULL, 0, 1);
1291         return NOTIFY_OK;
1292 }
1293
1294 static struct notifier_block acpi_processor_latency_notifier = {
1295         .notifier_call = acpi_processor_latency_notify,
1296 };
1297
1298 #endif
1299
1300 #else /* CONFIG_CPU_IDLE */
1301
1302 /**
1303  * acpi_idle_bm_check - checks if bus master activity was detected
1304  */
1305 static int acpi_idle_bm_check(void)
1306 {
1307         u32 bm_status = 0;
1308
1309         acpi_get_register(ACPI_BITREG_BUS_MASTER_STATUS, &bm_status);
1310         if (bm_status)
1311                 acpi_set_register(ACPI_BITREG_BUS_MASTER_STATUS, 1);
1312         /*
1313          * PIIX4 Erratum #18: Note that BM_STS doesn't always reflect
1314          * the true state of bus mastering activity; forcing us to
1315          * manually check the BMIDEA bit of each IDE channel.
1316          */
1317         else if (errata.piix4.bmisx) {
1318                 if ((inb_p(errata.piix4.bmisx + 0x02) & 0x01)
1319                     || (inb_p(errata.piix4.bmisx + 0x0A) & 0x01))
1320                         bm_status = 1;
1321         }
1322         return bm_status;
1323 }
1324
1325 /**
1326  * acpi_idle_update_bm_rld - updates the BM_RLD bit depending on target state
1327  * @pr: the processor
1328  * @target: the new target state
1329  */
1330 static inline void acpi_idle_update_bm_rld(struct acpi_processor *pr,
1331                                            struct acpi_processor_cx *target)
1332 {
1333         if (pr->flags.bm_rld_set && target->type != ACPI_STATE_C3) {
1334                 acpi_set_register(ACPI_BITREG_BUS_MASTER_RLD, 0);
1335                 pr->flags.bm_rld_set = 0;
1336         }
1337
1338         if (!pr->flags.bm_rld_set && target->type == ACPI_STATE_C3) {
1339                 acpi_set_register(ACPI_BITREG_BUS_MASTER_RLD, 1);
1340                 pr->flags.bm_rld_set = 1;
1341         }
1342 }
1343
1344 /**
1345  * acpi_idle_do_entry - a helper function that does C2 and C3 type entry
1346  * @cx: cstate data
1347  */
1348 static inline void acpi_idle_do_entry(struct acpi_processor_cx *cx)
1349 {
1350         if (cx->space_id == ACPI_CSTATE_FFH) {
1351                 /* Call into architectural FFH based C-state */
1352                 acpi_processor_ffh_cstate_enter(cx);
1353         } else {
1354                 int unused;
1355                 /* IO port based C-state */
1356                 inb(cx->address);
1357                 /* Dummy wait op - must do something useless after P_LVL2 read
1358                    because chipsets cannot guarantee that STPCLK# signal
1359                    gets asserted in time to freeze execution properly. */
1360                 unused = inl(acpi_gbl_FADT.xpm_timer_block.address);
1361         }
1362 }
1363
1364 /**
1365  * acpi_idle_enter_c1 - enters an ACPI C1 state-type
1366  * @dev: the target CPU
1367  * @state: the state data
1368  *
1369  * This is equivalent to the HALT instruction.
1370  */
1371 static int acpi_idle_enter_c1(struct cpuidle_device *dev,
1372                               struct cpuidle_state *state)
1373 {
1374         struct acpi_processor *pr;
1375         struct acpi_processor_cx *cx = cpuidle_get_statedata(state);
1376         pr = processors[smp_processor_id()];
1377
1378         if (unlikely(!pr))
1379                 return 0;
1380
1381         if (pr->flags.bm_check)
1382                 acpi_idle_update_bm_rld(pr, cx);
1383
1384         acpi_safe_halt();
1385
1386         cx->usage++;
1387
1388         return 0;
1389 }
1390
1391 /**
1392  * acpi_idle_enter_simple - enters an ACPI state without BM handling
1393  * @dev: the target CPU
1394  * @state: the state data
1395  */
1396 static int acpi_idle_enter_simple(struct cpuidle_device *dev,
1397                                   struct cpuidle_state *state)
1398 {
1399         struct acpi_processor *pr;
1400         struct acpi_processor_cx *cx = cpuidle_get_statedata(state);
1401         u32 t1, t2;
1402         int sleep_ticks = 0;
1403
1404         pr = processors[smp_processor_id()];
1405
1406         if (unlikely(!pr))
1407                 return 0;
1408
1409         if (acpi_idle_suspend)
1410                 return(acpi_idle_enter_c1(dev, state));
1411
1412         local_irq_disable();
1413         current_thread_info()->status &= ~TS_POLLING;
1414         /*
1415          * TS_POLLING-cleared state must be visible before we test
1416          * NEED_RESCHED:
1417          */
1418         smp_mb();
1419
1420         if (unlikely(need_resched())) {
1421                 current_thread_info()->status |= TS_POLLING;
1422                 local_irq_enable();
1423                 return 0;
1424         }
1425
1426         /*
1427          * Must be done before busmaster disable as we might need to
1428          * access HPET !
1429          */
1430         acpi_state_timer_broadcast(pr, cx, 1);
1431
1432         if (pr->flags.bm_check)
1433                 acpi_idle_update_bm_rld(pr, cx);
1434
1435         if (cx->type == ACPI_STATE_C3)
1436                 ACPI_FLUSH_CPU_CACHE();
1437
1438         t1 = inl(acpi_gbl_FADT.xpm_timer_block.address);
1439         /* Tell the scheduler that we are going deep-idle: */
1440         sched_clock_idle_sleep_event();
1441         acpi_idle_do_entry(cx);
1442         t2 = inl(acpi_gbl_FADT.xpm_timer_block.address);
1443
1444 #if defined (CONFIG_GENERIC_TIME) && defined (CONFIG_X86_TSC)
1445         /* TSC could halt in idle, so notify users */
1446         mark_tsc_unstable("TSC halts in idle");;
1447 #endif
1448         sleep_ticks = ticks_elapsed(t1, t2);
1449
1450         /* Tell the scheduler how much we idled: */
1451         sched_clock_idle_wakeup_event(sleep_ticks*PM_TIMER_TICK_NS);
1452
1453         local_irq_enable();
1454         current_thread_info()->status |= TS_POLLING;
1455
1456         cx->usage++;
1457
1458         acpi_state_timer_broadcast(pr, cx, 0);
1459         cx->time += sleep_ticks;
1460         return ticks_elapsed_in_us(t1, t2);
1461 }
1462
1463 static int c3_cpu_count;
1464 static DEFINE_SPINLOCK(c3_lock);
1465
1466 /**
1467  * acpi_idle_enter_bm - enters C3 with proper BM handling
1468  * @dev: the target CPU
1469  * @state: the state data
1470  *
1471  * If BM is detected, the deepest non-C3 idle state is entered instead.
1472  */
1473 static int acpi_idle_enter_bm(struct cpuidle_device *dev,
1474                               struct cpuidle_state *state)
1475 {
1476         struct acpi_processor *pr;
1477         struct acpi_processor_cx *cx = cpuidle_get_statedata(state);
1478         u32 t1, t2;
1479         int sleep_ticks = 0;
1480
1481         pr = processors[smp_processor_id()];
1482
1483         if (unlikely(!pr))
1484                 return 0;
1485
1486         if (acpi_idle_suspend)
1487                 return(acpi_idle_enter_c1(dev, state));
1488
1489         if (acpi_idle_bm_check()) {
1490                 if (dev->safe_state) {
1491                         return dev->safe_state->enter(dev, dev->safe_state);
1492                 } else {
1493                         acpi_safe_halt();
1494                         return 0;
1495                 }
1496         }
1497
1498         local_irq_disable();
1499         current_thread_info()->status &= ~TS_POLLING;
1500         /*
1501          * TS_POLLING-cleared state must be visible before we test
1502          * NEED_RESCHED:
1503          */
1504         smp_mb();
1505
1506         if (unlikely(need_resched())) {
1507                 current_thread_info()->status |= TS_POLLING;
1508                 local_irq_enable();
1509                 return 0;
1510         }
1511
1512         /* Tell the scheduler that we are going deep-idle: */
1513         sched_clock_idle_sleep_event();
1514         /*
1515          * Must be done before busmaster disable as we might need to
1516          * access HPET !
1517          */
1518         acpi_state_timer_broadcast(pr, cx, 1);
1519
1520         acpi_idle_update_bm_rld(pr, cx);
1521
1522         /*
1523          * disable bus master
1524          * bm_check implies we need ARB_DIS
1525          * !bm_check implies we need cache flush
1526          * bm_control implies whether we can do ARB_DIS
1527          *
1528          * That leaves a case where bm_check is set and bm_control is
1529          * not set. In that case we cannot do much, we enter C3
1530          * without doing anything.
1531          */
1532         if (pr->flags.bm_check && pr->flags.bm_control) {
1533                 spin_lock(&c3_lock);
1534                 c3_cpu_count++;
1535                 /* Disable bus master arbitration when all CPUs are in C3 */
1536                 if (c3_cpu_count == num_online_cpus())
1537                         acpi_set_register(ACPI_BITREG_ARB_DISABLE, 1);
1538                 spin_unlock(&c3_lock);
1539         } else if (!pr->flags.bm_check) {
1540                 ACPI_FLUSH_CPU_CACHE();
1541         }
1542
1543         t1 = inl(acpi_gbl_FADT.xpm_timer_block.address);
1544         acpi_idle_do_entry(cx);
1545         t2 = inl(acpi_gbl_FADT.xpm_timer_block.address);
1546
1547         /* Re-enable bus master arbitration */
1548         if (pr->flags.bm_check && pr->flags.bm_control) {
1549                 spin_lock(&c3_lock);
1550                 acpi_set_register(ACPI_BITREG_ARB_DISABLE, 0);
1551                 c3_cpu_count--;
1552                 spin_unlock(&c3_lock);
1553         }
1554
1555 #if defined (CONFIG_GENERIC_TIME) && defined (CONFIG_X86_TSC)
1556         /* TSC could halt in idle, so notify users */
1557         mark_tsc_unstable("TSC halts in idle");
1558 #endif
1559         sleep_ticks = ticks_elapsed(t1, t2);
1560         /* Tell the scheduler how much we idled: */
1561         sched_clock_idle_wakeup_event(sleep_ticks*PM_TIMER_TICK_NS);
1562
1563         local_irq_enable();
1564         current_thread_info()->status |= TS_POLLING;
1565
1566         cx->usage++;
1567
1568         acpi_state_timer_broadcast(pr, cx, 0);
1569         cx->time += sleep_ticks;
1570         return ticks_elapsed_in_us(t1, t2);
1571 }
1572
1573 struct cpuidle_driver acpi_idle_driver = {
1574         .name =         "acpi_idle",
1575         .owner =        THIS_MODULE,
1576 };
1577
1578 /**
1579  * acpi_processor_setup_cpuidle - prepares and configures CPUIDLE
1580  * @pr: the ACPI processor
1581  */
1582 static int acpi_processor_setup_cpuidle(struct acpi_processor *pr)
1583 {
1584         int i, count = 0;
1585         struct acpi_processor_cx *cx;
1586         struct cpuidle_state *state;
1587         struct cpuidle_device *dev = &pr->power.dev;
1588
1589         if (!pr->flags.power_setup_done)
1590                 return -EINVAL;
1591
1592         if (pr->flags.power == 0) {
1593                 return -EINVAL;
1594         }
1595
1596         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER && i <= max_cstate; i++) {
1597                 cx = &pr->power.states[i];
1598                 state = &dev->states[count];
1599
1600                 if (!cx->valid)
1601                         continue;
1602
1603 #ifdef CONFIG_HOTPLUG_CPU
1604                 if ((cx->type != ACPI_STATE_C1) && (num_online_cpus() > 1) &&
1605                     !pr->flags.has_cst &&
1606                     !(acpi_gbl_FADT.flags & ACPI_FADT_C2_MP_SUPPORTED))
1607                         continue;
1608 #endif
1609                 cpuidle_set_statedata(state, cx);
1610
1611                 snprintf(state->name, CPUIDLE_NAME_LEN, "C%d", i);
1612                 state->exit_latency = cx->latency;
1613                 state->target_residency = cx->latency * 6;
1614                 state->power_usage = cx->power;
1615
1616                 state->flags = 0;
1617                 switch (cx->type) {
1618                         case ACPI_STATE_C1:
1619                         state->flags |= CPUIDLE_FLAG_SHALLOW;
1620                         state->enter = acpi_idle_enter_c1;
1621                         dev->safe_state = state;
1622                         break;
1623
1624                         case ACPI_STATE_C2:
1625                         state->flags |= CPUIDLE_FLAG_BALANCED;
1626                         state->flags |= CPUIDLE_FLAG_TIME_VALID;
1627                         state->enter = acpi_idle_enter_simple;
1628                         dev->safe_state = state;
1629                         break;
1630
1631                         case ACPI_STATE_C3:
1632                         state->flags |= CPUIDLE_FLAG_DEEP;
1633                         state->flags |= CPUIDLE_FLAG_TIME_VALID;
1634                         state->flags |= CPUIDLE_FLAG_CHECK_BM;
1635                         state->enter = pr->flags.bm_check ?
1636                                         acpi_idle_enter_bm :
1637                                         acpi_idle_enter_simple;
1638                         break;
1639                 }
1640
1641                 count++;
1642         }
1643
1644         dev->state_count = count;
1645
1646         if (!count)
1647                 return -EINVAL;
1648
1649         return 0;
1650 }
1651
1652 int acpi_processor_cst_has_changed(struct acpi_processor *pr)
1653 {
1654         int ret;
1655
1656         if (!pr)
1657                 return -EINVAL;
1658
1659         if (nocst) {
1660                 return -ENODEV;
1661         }
1662
1663         if (!pr->flags.power_setup_done)
1664                 return -ENODEV;
1665
1666         cpuidle_pause_and_lock();
1667         cpuidle_disable_device(&pr->power.dev);
1668         acpi_processor_get_power_info(pr);
1669         acpi_processor_setup_cpuidle(pr);
1670         ret = cpuidle_enable_device(&pr->power.dev);
1671         cpuidle_resume_and_unlock();
1672
1673         return ret;
1674 }
1675
1676 #endif /* CONFIG_CPU_IDLE */
1677
1678 int __cpuinit acpi_processor_power_init(struct acpi_processor *pr,
1679                               struct acpi_device *device)
1680 {
1681         acpi_status status = 0;
1682         static int first_run;
1683         struct proc_dir_entry *entry = NULL;
1684         unsigned int i;
1685
1686
1687         if (!first_run) {
1688                 dmi_check_system(processor_power_dmi_table);
1689                 max_cstate = acpi_processor_cstate_check(max_cstate);
1690                 if (max_cstate < ACPI_C_STATES_MAX)
1691                         printk(KERN_NOTICE
1692                                "ACPI: processor limited to max C-state %d\n",
1693                                max_cstate);
1694                 first_run++;
1695 #if !defined (CONFIG_CPU_IDLE) && defined (CONFIG_SMP)
1696                 register_latency_notifier(&acpi_processor_latency_notifier);
1697 #endif
1698         }
1699
1700         if (!pr)
1701                 return -EINVAL;
1702
1703         if (acpi_gbl_FADT.cst_control && !nocst) {
1704                 status =
1705                     acpi_os_write_port(acpi_gbl_FADT.smi_command, acpi_gbl_FADT.cst_control, 8);
1706                 if (ACPI_FAILURE(status)) {
1707                         ACPI_EXCEPTION((AE_INFO, status,
1708                                         "Notifying BIOS of _CST ability failed"));
1709                 }
1710         }
1711
1712         acpi_processor_get_power_info(pr);
1713         pr->flags.power_setup_done = 1;
1714
1715         /*
1716          * Install the idle handler if processor power management is supported.
1717          * Note that we use previously set idle handler will be used on
1718          * platforms that only support C1.
1719          */
1720         if ((pr->flags.power) && (!boot_option_idle_override)) {
1721 #ifdef CONFIG_CPU_IDLE
1722                 acpi_processor_setup_cpuidle(pr);
1723                 pr->power.dev.cpu = pr->id;
1724                 if (cpuidle_register_device(&pr->power.dev))
1725                         return -EIO;
1726 #endif
1727
1728                 printk(KERN_INFO PREFIX "CPU%d (power states:", pr->id);
1729                 for (i = 1; i <= pr->power.count; i++)
1730                         if (pr->power.states[i].valid)
1731                                 printk(" C%d[C%d]", i,
1732                                        pr->power.states[i].type);
1733                 printk(")\n");
1734
1735 #ifndef CONFIG_CPU_IDLE
1736                 if (pr->id == 0) {
1737                         pm_idle_save = pm_idle;
1738                         pm_idle = acpi_processor_idle;
1739                 }
1740 #endif
1741         }
1742
1743         /* 'power' [R] */
1744         entry = create_proc_entry(ACPI_PROCESSOR_FILE_POWER,
1745                                   S_IRUGO, acpi_device_dir(device));
1746         if (!entry)
1747                 return -EIO;
1748         else {
1749                 entry->proc_fops = &acpi_processor_power_fops;
1750                 entry->data = acpi_driver_data(device);
1751                 entry->owner = THIS_MODULE;
1752         }
1753
1754         return 0;
1755 }
1756
1757 int acpi_processor_power_exit(struct acpi_processor *pr,
1758                               struct acpi_device *device)
1759 {
1760 #ifdef CONFIG_CPU_IDLE
1761         if ((pr->flags.power) && (!boot_option_idle_override))
1762                 cpuidle_unregister_device(&pr->power.dev);
1763 #endif
1764         pr->flags.power_setup_done = 0;
1765
1766         if (acpi_device_dir(device))
1767                 remove_proc_entry(ACPI_PROCESSOR_FILE_POWER,
1768                                   acpi_device_dir(device));
1769
1770 #ifndef CONFIG_CPU_IDLE
1771
1772         /* Unregister the idle handler when processor #0 is removed. */
1773         if (pr->id == 0) {
1774                 pm_idle = pm_idle_save;
1775
1776                 /*
1777                  * We are about to unload the current idle thread pm callback
1778                  * (pm_idle), Wait for all processors to update cached/local
1779                  * copies of pm_idle before proceeding.
1780                  */
1781                 cpu_idle_wait();
1782 #ifdef CONFIG_SMP
1783                 unregister_latency_notifier(&acpi_processor_latency_notifier);
1784 #endif
1785         }
1786 #endif
1787
1788         return 0;
1789 }