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