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