git.oblomov.eu Git - linux-2.6/blob - drivers/acpi/processor_idle.c

   1 /*
   2  * processor_idle - idle state submodule to the ACPI processor driver
   3  *
   4  *  Copyright (C) 2001, 2002 Andy Grover <andrew.grover@intel.com>
   5  *  Copyright (C) 2001, 2002 Paul Diefenbaugh <paul.s.diefenbaugh@intel.com>
   6  *  Copyright (C) 2004, 2005 Dominik Brodowski <linux@brodo.de>
   7  *  Copyright (C) 2004  Anil S Keshavamurthy <anil.s.keshavamurthy@intel.com>
   8  *                      - Added processor hotplug support
   9  *  Copyright (C) 2005  Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>
  10  *                      - Added support for C3 on SMP
  11  *
  12  * ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
  13  *
  14  *  This program is free software; you can redistribute it and/or modify
  15  *  it under the terms of the GNU General Public License as published by
  16  *  the Free Software Foundation; either version 2 of the License, or (at
  17  *  your option) any later version.
  18  *
  19  *  This program is distributed in the hope that it will be useful, but
  20  *  WITHOUT ANY WARRANTY; without even the implied warranty of
  21  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  22  *  General Public License for more details.
  23  *
  24  *  You should have received a copy of the GNU General Public License along
  25  *  with this program; if not, write to the Free Software Foundation, Inc.,
  26  *  59 Temple Place, Suite 330, Boston, MA 02111-1307 USA.
  27  *
  28  * ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
  29  */
  30
  31 #include <linux/kernel.h>
  32 #include <linux/module.h>
  33 #include <linux/init.h>
  34 #include <linux/cpufreq.h>
  35 #include <linux/proc_fs.h>
  36 #include <linux/seq_file.h>
  37 #include <linux/acpi.h>
  38 #include <linux/dmi.h>
  39 #include <linux/moduleparam.h>
  40 #include <linux/sched.h>        /* need_resched() */
  41 #include <linux/latency.h>
  42 #include <linux/clockchips.h>
  43
  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         /*
 336          * Interrupts must be disabled during bus mastering calculations and
 337          * for C2/C3 transitions.
 338          */
 339         local_irq_disable();
 340
 341         pr = processors[smp_processor_id()];
 342         if (!pr) {
 343                 local_irq_enable();
 344                 return;
 345         }
 346
 347         /*
 348          * Check whether we truly need to go idle, or should
 349          * reschedule:
 350          */
 351         if (unlikely(need_resched())) {
 352                 local_irq_enable();
 353                 return;
 354         }
 355
 356         cx = pr->power.state;
 357         if (!cx) {
 358                 if (pm_idle_save)
 359                         pm_idle_save();
 360                 else
 361                         acpi_safe_halt();
 362                 return;
 363         }
 364
 365         /*
 366          * Check BM Activity
 367          * -----------------
 368          * Check for bus mastering activity (if required), record, and check
 369          * for demotion.
 370          */
 371         if (pr->flags.bm_check) {
 372                 u32 bm_status = 0;
 373                 unsigned long diff = jiffies - pr->power.bm_check_timestamp;
 374
 375                 if (diff > 31)
 376                         diff = 31;
 377
 378                 pr->power.bm_activity <<= diff;
 379
 380                 acpi_get_register(ACPI_BITREG_BUS_MASTER_STATUS, &bm_status);
 381                 if (bm_status) {
 382                         pr->power.bm_activity |= 0x1;
 383                         acpi_set_register(ACPI_BITREG_BUS_MASTER_STATUS, 1);
 384                 }
 385                 /*
 386                  * PIIX4 Erratum #18: Note that BM_STS doesn't always reflect
 387                  * the true state of bus mastering activity; forcing us to
 388                  * manually check the BMIDEA bit of each IDE channel.
 389                  */
 390                 else if (errata.piix4.bmisx) {
 391                         if ((inb_p(errata.piix4.bmisx + 0x02) & 0x01)
 392                             || (inb_p(errata.piix4.bmisx + 0x0A) & 0x01))
 393                                 pr->power.bm_activity |= 0x1;
 394                 }
 395
 396                 pr->power.bm_check_timestamp = jiffies;
 397
 398                 /*
 399                  * If bus mastering is or was active this jiffy, demote
 400                  * to avoid a faulty transition.  Note that the processor
 401                  * won't enter a low-power state during this call (to this
 402                  * function) but should upon the next.
 403                  *
 404                  * TBD: A better policy might be to fallback to the demotion
 405                  *      state (use it for this quantum only) istead of
 406                  *      demoting -- and rely on duration as our sole demotion
 407                  *      qualification.  This may, however, introduce DMA
 408                  *      issues (e.g. floppy DMA transfer overrun/underrun).
 409                  */
 410                 if ((pr->power.bm_activity & 0x1) &&
 411                     cx->demotion.threshold.bm) {
 412                         local_irq_enable();
 413                         next_state = cx->demotion.state;
 414                         goto end;
 415                 }
 416         }
 417
 418 #ifdef CONFIG_HOTPLUG_CPU
 419         /*
 420          * Check for P_LVL2_UP flag before entering C2 and above on
 421          * an SMP system. We do it here instead of doing it at _CST/P_LVL
 422          * detection phase, to work cleanly with logical CPU hotplug.
 423          */
 424         if ((cx->type != ACPI_STATE_C1) && (num_online_cpus() > 1) &&
 425             !pr->flags.has_cst && !(acpi_gbl_FADT.flags & ACPI_FADT_C2_MP_SUPPORTED))
 426                 cx = &pr->power.states[ACPI_STATE_C1];
 427 #endif
 428
 429         /*
 430          * Sleep:
 431          * ------
 432          * Invoke the current Cx state to put the processor to sleep.
 433          */
 434         if (cx->type == ACPI_STATE_C2 || cx->type == ACPI_STATE_C3) {
 435                 current_thread_info()->status &= ~TS_POLLING;
 436                 /*
 437                  * TS_POLLING-cleared state must be visible before we
 438                  * test NEED_RESCHED:
 439                  */
 440                 smp_mb();
 441                 if (need_resched()) {
 442                         current_thread_info()->status |= TS_POLLING;
 443                         local_irq_enable();
 444                         return;
 445                 }
 446         }
 447
 448         switch (cx->type) {
 449
 450         case ACPI_STATE_C1:
 451                 /*
 452                  * Invoke C1.
 453                  * Use the appropriate idle routine, the one that would
 454                  * be used without acpi C-states.
 455                  */
 456                 if (pm_idle_save)
 457                         pm_idle_save();
 458                 else
 459                         acpi_safe_halt();
 460
 461                 /*
 462                  * TBD: Can't get time duration while in C1, as resumes
 463                  *      go to an ISR rather than here.  Need to instrument
 464                  *      base interrupt handler.
 465                  */
 466                 sleep_ticks = 0xFFFFFFFF;
 467                 break;
 468
 469         case ACPI_STATE_C2:
 470                 /* Get start time (ticks) */
 471                 t1 = inl(acpi_gbl_FADT.xpm_timer_block.address);
 472                 /* Invoke C2 */
 473                 acpi_state_timer_broadcast(pr, cx, 1);
 474                 acpi_cstate_enter(cx);
 475                 /* Get end time (ticks) */
 476                 t2 = inl(acpi_gbl_FADT.xpm_timer_block.address);
 477
 478 #if defined (CONFIG_GENERIC_TIME) && defined (CONFIG_X86_TSC)
 479                 /* TSC halts in C2, so notify users */
 480                 mark_tsc_unstable("possible TSC halt in C2");
 481 #endif
 482                 /* Re-enable interrupts */
 483                 local_irq_enable();
 484                 current_thread_info()->status |= TS_POLLING;
 485                 /* Compute time (ticks) that we were actually asleep */
 486                 sleep_ticks =
 487                     ticks_elapsed(t1, t2) - cx->latency_ticks - C2_OVERHEAD;
 488                 acpi_state_timer_broadcast(pr, cx, 0);
 489                 break;
 490
 491         case ACPI_STATE_C3:
 492
 493                 /*
 494                  * disable bus master
 495                  * bm_check implies we need ARB_DIS
 496                  * !bm_check implies we need cache flush
 497                  * bm_control implies whether we can do ARB_DIS
 498                  *
 499                  * That leaves a case where bm_check is set and bm_control is
 500                  * not set. In that case we cannot do much, we enter C3
 501                  * without doing anything.
 502                  */
 503                 if (pr->flags.bm_check && pr->flags.bm_control) {
 504                         if (atomic_inc_return(&c3_cpu_count) ==
 505                             num_online_cpus()) {
 506                                 /*
 507                                  * All CPUs are trying to go to C3
 508                                  * Disable bus master arbitration
 509                                  */
 510                                 acpi_set_register(ACPI_BITREG_ARB_DISABLE, 1);
 511                         }
 512                 } else if (!pr->flags.bm_check) {
 513                         /* SMP with no shared cache... Invalidate cache  */
 514                         ACPI_FLUSH_CPU_CACHE();
 515                 }
 516
 517                 /* Get start time (ticks) */
 518                 t1 = inl(acpi_gbl_FADT.xpm_timer_block.address);
 519                 /* Invoke C3 */
 520                 acpi_state_timer_broadcast(pr, cx, 1);
 521                 acpi_cstate_enter(cx);
 522                 /* Get end time (ticks) */
 523                 t2 = inl(acpi_gbl_FADT.xpm_timer_block.address);
 524                 if (pr->flags.bm_check && pr->flags.bm_control) {
 525                         /* Enable bus master arbitration */
 526                         atomic_dec(&c3_cpu_count);
 527                         acpi_set_register(ACPI_BITREG_ARB_DISABLE, 0);
 528                 }
 529
 530 #if defined (CONFIG_GENERIC_TIME) && defined (CONFIG_X86_TSC)
 531                 /* TSC halts in C3, so notify users */
 532                 mark_tsc_unstable("TSC halts in C3");
 533 #endif
 534                 /* Re-enable interrupts */
 535                 local_irq_enable();
 536                 current_thread_info()->status |= TS_POLLING;
 537                 /* Compute time (ticks) that we were actually asleep */
 538                 sleep_ticks =
 539                     ticks_elapsed(t1, t2) - cx->latency_ticks - C3_OVERHEAD;
 540                 acpi_state_timer_broadcast(pr, cx, 0);
 541                 break;
 542
 543         default:
 544                 local_irq_enable();
 545                 return;
 546         }
 547         cx->usage++;
 548         if ((cx->type != ACPI_STATE_C1) && (sleep_ticks > 0))
 549                 cx->time += sleep_ticks;
 550
 551         next_state = pr->power.state;
 552
 553 #ifdef CONFIG_HOTPLUG_CPU
 554         /* Don't do promotion/demotion */
 555         if ((cx->type == ACPI_STATE_C1) && (num_online_cpus() > 1) &&
 556             !pr->flags.has_cst && !(acpi_gbl_FADT.flags & ACPI_FADT_C2_MP_SUPPORTED)) {
 557                 next_state = cx;
 558                 goto end;
 559         }
 560 #endif
 561
 562         /*
 563          * Promotion?
 564          * ----------
 565          * Track the number of longs (time asleep is greater than threshold)
 566          * and promote when the count threshold is reached.  Note that bus
 567          * mastering activity may prevent promotions.
 568          * Do not promote above max_cstate.
 569          */
 570         if (cx->promotion.state &&
 571             ((cx->promotion.state - pr->power.states) <= max_cstate)) {
 572                 if (sleep_ticks > cx->promotion.threshold.ticks &&
 573                   cx->promotion.state->latency <= system_latency_constraint()) {
 574                         cx->promotion.count++;
 575                         cx->demotion.count = 0;
 576                         if (cx->promotion.count >=
 577                             cx->promotion.threshold.count) {
 578                                 if (pr->flags.bm_check) {
 579                                         if (!
 580                                             (pr->power.bm_activity & cx->
 581                                              promotion.threshold.bm)) {
 582                                                 next_state =
 583                                                     cx->promotion.state;
 584                                                 goto end;
 585                                         }
 586                                 } else {
 587                                         next_state = cx->promotion.state;
 588                                         goto end;
 589                                 }
 590                         }
 591                 }
 592         }
 593
 594         /*
 595          * Demotion?
 596          * ---------
 597          * Track the number of shorts (time asleep is less than time threshold)
 598          * and demote when the usage threshold is reached.
 599          */
 600         if (cx->demotion.state) {
 601                 if (sleep_ticks < cx->demotion.threshold.ticks) {
 602                         cx->demotion.count++;
 603                         cx->promotion.count = 0;
 604                         if (cx->demotion.count >= cx->demotion.threshold.count) {
 605                                 next_state = cx->demotion.state;
 606                                 goto end;
 607                         }
 608                 }
 609         }
 610
 611       end:
 612         /*
 613          * Demote if current state exceeds max_cstate
 614          * or if the latency of the current state is unacceptable
 615          */
 616         if ((pr->power.state - pr->power.states) > max_cstate ||
 617                 pr->power.state->latency > system_latency_constraint()) {
 618                 if (cx->demotion.state)
 619                         next_state = cx->demotion.state;
 620         }
 621
 622         /*
 623          * New Cx State?
 624          * -------------
 625          * If we're going to start using a new Cx state we must clean up
 626          * from the previous and prepare to use the new.
 627          */
 628         if (next_state != pr->power.state)
 629                 acpi_processor_power_activate(pr, next_state);
 630 }
 631
 632 static int acpi_processor_set_power_policy(struct acpi_processor *pr)
 633 {
 634         unsigned int i;
 635         unsigned int state_is_set = 0;
 636         struct acpi_processor_cx *lower = NULL;
 637         struct acpi_processor_cx *higher = NULL;
 638         struct acpi_processor_cx *cx;
 639
 640
 641         if (!pr)
 642                 return -EINVAL;
 643
 644         /*
 645          * This function sets the default Cx state policy (OS idle handler).
 646          * Our scheme is to promote quickly to C2 but more conservatively
 647          * to C3.  We're favoring C2  for its characteristics of low latency
 648          * (quick response), good power savings, and ability to allow bus
 649          * mastering activity.  Note that the Cx state policy is completely
 650          * customizable and can be altered dynamically.
 651          */
 652
 653         /* startup state */
 654         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
 655                 cx = &pr->power.states[i];
 656                 if (!cx->valid)
 657                         continue;
 658
 659                 if (!state_is_set)
 660                         pr->power.state = cx;
 661                 state_is_set++;
 662                 break;
 663         }
 664
 665         if (!state_is_set)
 666                 return -ENODEV;
 667
 668         /* demotion */
 669         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
 670                 cx = &pr->power.states[i];
 671                 if (!cx->valid)
 672                         continue;
 673
 674                 if (lower) {
 675                         cx->demotion.state = lower;
 676                         cx->demotion.threshold.ticks = cx->latency_ticks;
 677                         cx->demotion.threshold.count = 1;
 678                         if (cx->type == ACPI_STATE_C3)
 679                                 cx->demotion.threshold.bm = bm_history;
 680                 }
 681
 682                 lower = cx;
 683         }
 684
 685         /* promotion */
 686         for (i = (ACPI_PROCESSOR_MAX_POWER - 1); i > 0; i--) {
 687                 cx = &pr->power.states[i];
 688                 if (!cx->valid)
 689                         continue;
 690
 691                 if (higher) {
 692                         cx->promotion.state = higher;
 693                         cx->promotion.threshold.ticks = cx->latency_ticks;
 694                         if (cx->type >= ACPI_STATE_C2)
 695                                 cx->promotion.threshold.count = 4;
 696                         else
 697                                 cx->promotion.threshold.count = 10;
 698                         if (higher->type == ACPI_STATE_C3)
 699                                 cx->promotion.threshold.bm = bm_history;
 700                 }
 701
 702                 higher = cx;
 703         }
 704
 705         return 0;
 706 }
 707
 708 static int acpi_processor_get_power_info_fadt(struct acpi_processor *pr)
 709 {
 710
 711         if (!pr)
 712                 return -EINVAL;
 713
 714         if (!pr->pblk)
 715                 return -ENODEV;
 716
 717         /* if info is obtained from pblk/fadt, type equals state */
 718         pr->power.states[ACPI_STATE_C2].type = ACPI_STATE_C2;
 719         pr->power.states[ACPI_STATE_C3].type = ACPI_STATE_C3;
 720
 721 #ifndef CONFIG_HOTPLUG_CPU
 722         /*
 723          * Check for P_LVL2_UP flag before entering C2 and above on
 724          * an SMP system.
 725          */
 726         if ((num_online_cpus() > 1) &&
 727             !(acpi_gbl_FADT.flags & ACPI_FADT_C2_MP_SUPPORTED))
 728                 return -ENODEV;
 729 #endif
 730
 731         /* determine C2 and C3 address from pblk */
 732         pr->power.states[ACPI_STATE_C2].address = pr->pblk + 4;
 733         pr->power.states[ACPI_STATE_C3].address = pr->pblk + 5;
 734
 735         /* determine latencies from FADT */
 736         pr->power.states[ACPI_STATE_C2].latency = acpi_gbl_FADT.C2latency;
 737         pr->power.states[ACPI_STATE_C3].latency = acpi_gbl_FADT.C3latency;
 738
 739         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 740                           "lvl2[0x%08x] lvl3[0x%08x]\n",
 741                           pr->power.states[ACPI_STATE_C2].address,
 742                           pr->power.states[ACPI_STATE_C3].address));
 743
 744         return 0;
 745 }
 746
 747 static int acpi_processor_get_power_info_default(struct acpi_processor *pr)
 748 {
 749         if (!pr->power.states[ACPI_STATE_C1].valid) {
 750                 /* set the first C-State to C1 */
 751                 /* all processors need to support C1 */
 752                 pr->power.states[ACPI_STATE_C1].type = ACPI_STATE_C1;
 753                 pr->power.states[ACPI_STATE_C1].valid = 1;
 754         }
 755         /* the C0 state only exists as a filler in our array */
 756         pr->power.states[ACPI_STATE_C0].valid = 1;
 757         return 0;
 758 }
 759
 760 static int acpi_processor_get_power_info_cst(struct acpi_processor *pr)
 761 {
 762         acpi_status status = 0;
 763         acpi_integer count;
 764         int current_count;
 765         int i;
 766         struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
 767         union acpi_object *cst;
 768
 769
 770         if (nocst)
 771                 return -ENODEV;
 772
 773         current_count = 0;
 774
 775         status = acpi_evaluate_object(pr->handle, "_CST", NULL, &buffer);
 776         if (ACPI_FAILURE(status)) {
 777                 ACPI_DEBUG_PRINT((ACPI_DB_INFO, "No _CST, giving up\n"));
 778                 return -ENODEV;
 779         }
 780
 781         cst = buffer.pointer;
 782
 783         /* There must be at least 2 elements */
 784         if (!cst || (cst->type != ACPI_TYPE_PACKAGE) || cst->package.count < 2) {
 785                 printk(KERN_ERR PREFIX "not enough elements in _CST\n");
 786                 status = -EFAULT;
 787                 goto end;
 788         }
 789
 790         count = cst->package.elements[0].integer.value;
 791
 792         /* Validate number of power states. */
 793         if (count < 1 || count != cst->package.count - 1) {
 794                 printk(KERN_ERR PREFIX "count given by _CST is not valid\n");
 795                 status = -EFAULT;
 796                 goto end;
 797         }
 798
 799         /* Tell driver that at least _CST is supported. */
 800         pr->flags.has_cst = 1;
 801
 802         for (i = 1; i <= count; i++) {
 803                 union acpi_object *element;
 804                 union acpi_object *obj;
 805                 struct acpi_power_register *reg;
 806                 struct acpi_processor_cx cx;
 807
 808                 memset(&cx, 0, sizeof(cx));
 809
 810                 element = &(cst->package.elements[i]);
 811                 if (element->type != ACPI_TYPE_PACKAGE)
 812                         continue;
 813
 814                 if (element->package.count != 4)
 815                         continue;
 816
 817                 obj = &(element->package.elements[0]);
 818
 819                 if (obj->type != ACPI_TYPE_BUFFER)
 820                         continue;
 821
 822                 reg = (struct acpi_power_register *)obj->buffer.pointer;
 823
 824                 if (reg->space_id != ACPI_ADR_SPACE_SYSTEM_IO &&
 825                     (reg->space_id != ACPI_ADR_SPACE_FIXED_HARDWARE))
 826                         continue;
 827
 828                 /* There should be an easy way to extract an integer... */
 829                 obj = &(element->package.elements[1]);
 830                 if (obj->type != ACPI_TYPE_INTEGER)
 831                         continue;
 832
 833                 cx.type = obj->integer.value;
 834                 /*
 835                  * Some buggy BIOSes won't list C1 in _CST -
 836                  * Let acpi_processor_get_power_info_default() handle them later
 837                  */
 838                 if (i == 1 && cx.type != ACPI_STATE_C1)
 839                         current_count++;
 840
 841                 cx.address = reg->address;
 842                 cx.index = current_count + 1;
 843
 844                 cx.space_id = ACPI_CSTATE_SYSTEMIO;
 845                 if (reg->space_id == ACPI_ADR_SPACE_FIXED_HARDWARE) {
 846                         if (acpi_processor_ffh_cstate_probe
 847                                         (pr->id, &cx, reg) == 0) {
 848                                 cx.space_id = ACPI_CSTATE_FFH;
 849                         } else if (cx.type != ACPI_STATE_C1) {
 850                                 /*
 851                                  * C1 is a special case where FIXED_HARDWARE
 852                                  * can be handled in non-MWAIT way as well.
 853                                  * In that case, save this _CST entry info.
 854                                  * That is, we retain space_id of SYSTEM_IO for
 855                                  * halt based C1.
 856                                  * Otherwise, ignore this info and continue.
 857                                  */
 858                                 continue;
 859                         }
 860                 }
 861
 862                 obj = &(element->package.elements[2]);
 863                 if (obj->type != ACPI_TYPE_INTEGER)
 864                         continue;
 865
 866                 cx.latency = obj->integer.value;
 867
 868                 obj = &(element->package.elements[3]);
 869                 if (obj->type != ACPI_TYPE_INTEGER)
 870                         continue;
 871
 872                 cx.power = obj->integer.value;
 873
 874                 current_count++;
 875                 memcpy(&(pr->power.states[current_count]), &cx, sizeof(cx));
 876
 877                 /*
 878                  * We support total ACPI_PROCESSOR_MAX_POWER - 1
 879                  * (From 1 through ACPI_PROCESSOR_MAX_POWER - 1)
 880                  */
 881                 if (current_count >= (ACPI_PROCESSOR_MAX_POWER - 1)) {
 882                         printk(KERN_WARNING
 883                                "Limiting number of power states to max (%d)\n",
 884                                ACPI_PROCESSOR_MAX_POWER);
 885                         printk(KERN_WARNING
 886                                "Please increase ACPI_PROCESSOR_MAX_POWER if needed.\n");
 887                         break;
 888                 }
 889         }
 890
 891         ACPI_DEBUG_PRINT((ACPI_DB_INFO, "Found %d power states\n",
 892                           current_count));
 893
 894         /* Validate number of power states discovered */
 895         if (current_count < 2)
 896                 status = -EFAULT;
 897
 898       end:
 899         kfree(buffer.pointer);
 900
 901         return status;
 902 }
 903
 904 static void acpi_processor_power_verify_c2(struct acpi_processor_cx *cx)
 905 {
 906
 907         if (!cx->address)
 908                 return;
 909
 910         /*
 911          * C2 latency must be less than or equal to 100
 912          * microseconds.
 913          */
 914         else if (cx->latency > ACPI_PROCESSOR_MAX_C2_LATENCY) {
 915                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 916                                   "latency too large [%d]\n", cx->latency));
 917                 return;
 918         }
 919
 920         /*
 921          * Otherwise we've met all of our C2 requirements.
 922          * Normalize the C2 latency to expidite policy
 923          */
 924         cx->valid = 1;
 925         cx->latency_ticks = US_TO_PM_TIMER_TICKS(cx->latency);
 926
 927         return;
 928 }
 929
 930 static void acpi_processor_power_verify_c3(struct acpi_processor *pr,
 931                                            struct acpi_processor_cx *cx)
 932 {
 933         static int bm_check_flag;
 934
 935
 936         if (!cx->address)
 937                 return;
 938
 939         /*
 940          * C3 latency must be less than or equal to 1000
 941          * microseconds.
 942          */
 943         else if (cx->latency > ACPI_PROCESSOR_MAX_C3_LATENCY) {
 944                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 945                                   "latency too large [%d]\n", cx->latency));
 946                 return;
 947         }
 948
 949         /*
 950          * PIIX4 Erratum #18: We don't support C3 when Type-F (fast)
 951          * DMA transfers are used by any ISA device to avoid livelock.
 952          * Note that we could disable Type-F DMA (as recommended by
 953          * the erratum), but this is known to disrupt certain ISA
 954          * devices thus we take the conservative approach.
 955          */
 956         else if (errata.piix4.fdma) {
 957                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 958                                   "C3 not supported on PIIX4 with Type-F DMA\n"));
 959                 return;
 960         }
 961
 962         /* All the logic here assumes flags.bm_check is same across all CPUs */
 963         if (!bm_check_flag) {
 964                 /* Determine whether bm_check is needed based on CPU  */
 965                 acpi_processor_power_init_bm_check(&(pr->flags), pr->id);
 966                 bm_check_flag = pr->flags.bm_check;
 967         } else {
 968                 pr->flags.bm_check = bm_check_flag;
 969         }
 970
 971         if (pr->flags.bm_check) {
 972                 /* bus mastering control is necessary */
 973                 if (!pr->flags.bm_control) {
 974                         /* In this case we enter C3 without bus mastering */
 975                         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 976                                 "C3 support without bus mastering control\n"));
 977                 }
 978         } else {
 979                 /*
 980                  * WBINVD should be set in fadt, for C3 state to be
 981                  * supported on when bm_check is not required.
 982                  */
 983                 if (!(acpi_gbl_FADT.flags & ACPI_FADT_WBINVD)) {
 984                         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 985                                           "Cache invalidation should work properly"
 986                                           " for C3 to be enabled on SMP systems\n"));
 987                         return;
 988                 }
 989                 acpi_set_register(ACPI_BITREG_BUS_MASTER_RLD, 0);
 990         }
 991
 992         /*
 993          * Otherwise we've met all of our C3 requirements.
 994          * Normalize the C3 latency to expidite policy.  Enable
 995          * checking of bus mastering status (bm_check) so we can
 996          * use this in our C3 policy
 997          */
 998         cx->valid = 1;
 999         cx->latency_ticks = US_TO_PM_TIMER_TICKS(cx->latency);
1000
1001         return;
1002 }
1003
1004 static int acpi_processor_power_verify(struct acpi_processor *pr)
1005 {
1006         unsigned int i;
1007         unsigned int working = 0;
1008
1009         pr->power.timer_broadcast_on_state = INT_MAX;
1010
1011         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
1012                 struct acpi_processor_cx *cx = &pr->power.states[i];
1013
1014                 switch (cx->type) {
1015                 case ACPI_STATE_C1:
1016                         cx->valid = 1;
1017                         break;
1018
1019                 case ACPI_STATE_C2:
1020                         acpi_processor_power_verify_c2(cx);
1021                         if (cx->valid)
1022                                 acpi_timer_check_state(i, pr, cx);
1023                         break;
1024
1025                 case ACPI_STATE_C3:
1026                         acpi_processor_power_verify_c3(pr, cx);
1027                         if (cx->valid)
1028                                 acpi_timer_check_state(i, pr, cx);
1029                         break;
1030                 }
1031
1032                 if (cx->valid)
1033                         working++;
1034         }
1035
1036         acpi_propagate_timer_broadcast(pr);
1037
1038         return (working);
1039 }
1040
1041 static int acpi_processor_get_power_info(struct acpi_processor *pr)
1042 {
1043         unsigned int i;
1044         int result;
1045
1046
1047         /* NOTE: the idle thread may not be running while calling
1048          * this function */
1049
1050         /* Zero initialize all the C-states info. */
1051         memset(pr->power.states, 0, sizeof(pr->power.states));
1052
1053         result = acpi_processor_get_power_info_cst(pr);
1054         if (result == -ENODEV)
1055                 result = acpi_processor_get_power_info_fadt(pr);
1056
1057         if (result)
1058                 return result;
1059
1060         acpi_processor_get_power_info_default(pr);
1061
1062         pr->power.count = acpi_processor_power_verify(pr);
1063
1064         /*
1065          * Set Default Policy
1066          * ------------------
1067          * Now that we know which states are supported, set the default
1068          * policy.  Note that this policy can be changed dynamically
1069          * (e.g. encourage deeper sleeps to conserve battery life when
1070          * not on AC).
1071          */
1072         result = acpi_processor_set_power_policy(pr);
1073         if (result)
1074                 return result;
1075
1076         /*
1077          * if one state of type C2 or C3 is available, mark this
1078          * CPU as being "idle manageable"
1079          */
1080         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
1081                 if (pr->power.states[i].valid) {
1082                         pr->power.count = i;
1083                         if (pr->power.states[i].type >= ACPI_STATE_C2)
1084                                 pr->flags.power = 1;
1085                 }
1086         }
1087
1088         return 0;
1089 }
1090
1091 int acpi_processor_cst_has_changed(struct acpi_processor *pr)
1092 {
1093         int result = 0;
1094
1095
1096         if (!pr)
1097                 return -EINVAL;
1098
1099         if (nocst) {
1100                 return -ENODEV;
1101         }
1102
1103         if (!pr->flags.power_setup_done)
1104                 return -ENODEV;
1105
1106         /* Fall back to the default idle loop */
1107         pm_idle = pm_idle_save;
1108         synchronize_sched();    /* Relies on interrupts forcing exit from idle. */
1109
1110         pr->flags.power = 0;
1111         result = acpi_processor_get_power_info(pr);
1112         if ((pr->flags.power == 1) && (pr->flags.power_setup_done))
1113                 pm_idle = acpi_processor_idle;
1114
1115         return result;
1116 }
1117
1118 /* proc interface */
1119
1120 static int acpi_processor_power_seq_show(struct seq_file *seq, void *offset)
1121 {
1122         struct acpi_processor *pr = seq->private;
1123         unsigned int i;
1124
1125
1126         if (!pr)
1127                 goto end;
1128
1129         seq_printf(seq, "active state:            C%zd\n"
1130                    "max_cstate:              C%d\n"
1131                    "bus master activity:     %08x\n"
1132                    "maximum allowed latency: %d usec\n",
1133                    pr->power.state ? pr->power.state - pr->power.states : 0,
1134                    max_cstate, (unsigned)pr->power.bm_activity,
1135                    system_latency_constraint());
1136
1137         seq_puts(seq, "states:\n");
1138
1139         for (i = 1; i <= pr->power.count; i++) {
1140                 seq_printf(seq, "   %cC%d:                  ",
1141                            (&pr->power.states[i] ==
1142                             pr->power.state ? '*' : ' '), i);
1143
1144                 if (!pr->power.states[i].valid) {
1145                         seq_puts(seq, "<not supported>\n");
1146                         continue;
1147                 }
1148
1149                 switch (pr->power.states[i].type) {
1150                 case ACPI_STATE_C1:
1151                         seq_printf(seq, "type[C1] ");
1152                         break;
1153                 case ACPI_STATE_C2:
1154                         seq_printf(seq, "type[C2] ");
1155                         break;
1156                 case ACPI_STATE_C3:
1157                         seq_printf(seq, "type[C3] ");
1158                         break;
1159                 default:
1160                         seq_printf(seq, "type[--] ");
1161                         break;
1162                 }
1163
1164                 if (pr->power.states[i].promotion.state)
1165                         seq_printf(seq, "promotion[C%zd] ",
1166                                    (pr->power.states[i].promotion.state -
1167                                     pr->power.states));
1168                 else
1169                         seq_puts(seq, "promotion[--] ");
1170
1171                 if (pr->power.states[i].demotion.state)
1172                         seq_printf(seq, "demotion[C%zd] ",
1173                                    (pr->power.states[i].demotion.state -
1174                                     pr->power.states));
1175                 else
1176                         seq_puts(seq, "demotion[--] ");
1177
1178                 seq_printf(seq, "latency[%03d] usage[%08d] duration[%020llu]\n",
1179                            pr->power.states[i].latency,
1180                            pr->power.states[i].usage,
1181                            (unsigned long long)pr->power.states[i].time);
1182         }
1183
1184       end:
1185         return 0;
1186 }
1187
1188 static int acpi_processor_power_open_fs(struct inode *inode, struct file *file)
1189 {
1190         return single_open(file, acpi_processor_power_seq_show,
1191                            PDE(inode)->data);
1192 }
1193
1194 static const struct file_operations acpi_processor_power_fops = {
1195         .open = acpi_processor_power_open_fs,
1196         .read = seq_read,
1197         .llseek = seq_lseek,
1198         .release = single_release,
1199 };
1200
1201 #ifdef CONFIG_SMP
1202 static void smp_callback(void *v)
1203 {
1204         /* we already woke the CPU up, nothing more to do */
1205 }
1206
1207 /*
1208  * This function gets called when a part of the kernel has a new latency
1209  * requirement.  This means we need to get all processors out of their C-state,
1210  * and then recalculate a new suitable C-state. Just do a cross-cpu IPI; that
1211  * wakes them all right up.
1212  */
1213 static int acpi_processor_latency_notify(struct notifier_block *b,
1214                 unsigned long l, void *v)
1215 {
1216         smp_call_function(smp_callback, NULL, 0, 1);
1217         return NOTIFY_OK;
1218 }
1219
1220 static struct notifier_block acpi_processor_latency_notifier = {
1221         .notifier_call = acpi_processor_latency_notify,
1222 };
1223 #endif
1224
1225 int __cpuinit acpi_processor_power_init(struct acpi_processor *pr,
1226                               struct acpi_device *device)
1227 {
1228         acpi_status status = 0;
1229         static int first_run;
1230         struct proc_dir_entry *entry = NULL;
1231         unsigned int i;
1232
1233
1234         if (!first_run) {
1235                 dmi_check_system(processor_power_dmi_table);
1236                 if (max_cstate < ACPI_C_STATES_MAX)
1237                         printk(KERN_NOTICE
1238                                "ACPI: processor limited to max C-state %d\n",
1239                                max_cstate);
1240                 first_run++;
1241 #ifdef CONFIG_SMP
1242                 register_latency_notifier(&acpi_processor_latency_notifier);
1243 #endif
1244         }
1245
1246         if (!pr)
1247                 return -EINVAL;
1248
1249         if (acpi_gbl_FADT.cst_control && !nocst) {
1250                 status =
1251                     acpi_os_write_port(acpi_gbl_FADT.smi_command, acpi_gbl_FADT.cst_control, 8);
1252                 if (ACPI_FAILURE(status)) {
1253                         ACPI_EXCEPTION((AE_INFO, status,
1254                                         "Notifying BIOS of _CST ability failed"));
1255                 }
1256         }
1257
1258         acpi_processor_get_power_info(pr);
1259
1260         /*
1261          * Install the idle handler if processor power management is supported.
1262          * Note that we use previously set idle handler will be used on
1263          * platforms that only support C1.
1264          */
1265         if ((pr->flags.power) && (!boot_option_idle_override)) {
1266                 printk(KERN_INFO PREFIX "CPU%d (power states:", pr->id);
1267                 for (i = 1; i <= pr->power.count; i++)
1268                         if (pr->power.states[i].valid)
1269                                 printk(" C%d[C%d]", i,
1270                                        pr->power.states[i].type);
1271                 printk(")\n");
1272
1273                 if (pr->id == 0) {
1274                         pm_idle_save = pm_idle;
1275                         pm_idle = acpi_processor_idle;
1276                 }
1277         }
1278
1279         /* 'power' [R] */
1280         entry = create_proc_entry(ACPI_PROCESSOR_FILE_POWER,
1281                                   S_IRUGO, acpi_device_dir(device));
1282         if (!entry)
1283                 return -EIO;
1284         else {
1285                 entry->proc_fops = &acpi_processor_power_fops;
1286                 entry->data = acpi_driver_data(device);
1287                 entry->owner = THIS_MODULE;
1288         }
1289
1290         pr->flags.power_setup_done = 1;
1291
1292         return 0;
1293 }
1294
1295 int acpi_processor_power_exit(struct acpi_processor *pr,
1296                               struct acpi_device *device)
1297 {
1298
1299         pr->flags.power_setup_done = 0;
1300
1301         if (acpi_device_dir(device))
1302                 remove_proc_entry(ACPI_PROCESSOR_FILE_POWER,
1303                                   acpi_device_dir(device));
1304
1305         /* Unregister the idle handler when processor #0 is removed. */
1306         if (pr->id == 0) {
1307                 pm_idle = pm_idle_save;
1308
1309                 /*
1310                  * We are about to unload the current idle thread pm callback
1311                  * (pm_idle), Wait for all processors to update cached/local
1312                  * copies of pm_idle before proceeding.
1313                  */
1314                 cpu_idle_wait();
1315 #ifdef CONFIG_SMP
1316                 unregister_latency_notifier(&acpi_processor_latency_notifier);
1317 #endif
1318         }
1319
1320         return 0;
1321 }