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 /*
  55  * Include the apic definitions for x86 to have the APIC timer related defines
  56  * available also for UP (on SMP it gets magically included via linux/smp.h).
  57  */
  58 #ifdef CONFIG_X86
  59 #include <asm/apic.h>
  60 #endif
  61
  62 #include <asm/io.h>
  63 #include <asm/uaccess.h>
  64
  65 #include <acpi/acpi_bus.h>
  66 #include <acpi/processor.h>
  67
  68 #define ACPI_PROCESSOR_COMPONENT        0x01000000
  69 #define ACPI_PROCESSOR_CLASS            "processor"
  70 #define _COMPONENT              ACPI_PROCESSOR_COMPONENT
  71 ACPI_MODULE_NAME("processor_idle");
  72 #define ACPI_PROCESSOR_FILE_POWER       "power"
  73 #define US_TO_PM_TIMER_TICKS(t)         ((t * (PM_TIMER_FREQUENCY/1000)) / 1000)
  74 #define C2_OVERHEAD                     4       /* 1us (3.579 ticks per us) */
  75 #define C3_OVERHEAD                     4       /* 1us (3.579 ticks per us) */
  76 static void (*pm_idle_save) (void) __read_mostly;
  77 module_param(max_cstate, uint, 0644);
  78
  79 static unsigned int nocst __read_mostly;
  80 module_param(nocst, uint, 0000);
  81
  82 /*
  83  * bm_history -- bit-mask with a bit per jiffy of bus-master activity
  84  * 1000 HZ: 0xFFFFFFFF: 32 jiffies = 32ms
  85  * 800 HZ: 0xFFFFFFFF: 32 jiffies = 40ms
  86  * 100 HZ: 0x0000000F: 4 jiffies = 40ms
  87  * reduce history for more aggressive entry into C3
  88  */
  89 static unsigned int bm_history __read_mostly =
  90     (HZ >= 800 ? 0xFFFFFFFF : ((1U << (HZ / 25)) - 1));
  91 module_param(bm_history, uint, 0644);
  92
  93 static unsigned use_ipi = 2;
  94 module_param(use_ipi, uint, 0644);
  95 MODULE_PARM_DESC(use_ipi, "IPI (vs. LAPIC) irqs for not waking up from C2/C3"
  96                  " machines. 0=apic, 1=ipi, 2=auto\n");
  97
  98 /* --------------------------------------------------------------------------
  99                                 Power Management
 100    -------------------------------------------------------------------------- */
 101
 102 /*
 103  * IBM ThinkPad R40e crashes mysteriously when going into C2 or C3.
 104  * For now disable this. Probably a bug somewhere else.
 105  *
 106  * To skip this limit, boot/load with a large max_cstate limit.
 107  */
 108 static int set_max_cstate(struct dmi_system_id *id)
 109 {
 110         if (max_cstate > ACPI_PROCESSOR_MAX_POWER)
 111                 return 0;
 112
 113         printk(KERN_NOTICE PREFIX "%s detected - limiting to C%ld max_cstate."
 114                " Override with \"processor.max_cstate=%d\"\n", id->ident,
 115                (long)id->driver_data, ACPI_PROCESSOR_MAX_POWER + 1);
 116
 117         max_cstate = (long)id->driver_data;
 118
 119         return 0;
 120 }
 121
 122 /* Actually this shouldn't be __cpuinitdata, would be better to fix the
 123    callers to only run once -AK */
 124 static struct dmi_system_id __cpuinitdata processor_power_dmi_table[] = {
 125         { set_max_cstate, "IBM ThinkPad R40e", {
 126           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 127           DMI_MATCH(DMI_BIOS_VERSION,"1SET70WW")}, (void *)1},
 128         { set_max_cstate, "IBM ThinkPad R40e", {
 129           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 130           DMI_MATCH(DMI_BIOS_VERSION,"1SET60WW")}, (void *)1},
 131         { set_max_cstate, "IBM ThinkPad R40e", {
 132           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 133           DMI_MATCH(DMI_BIOS_VERSION,"1SET43WW") }, (void*)1},
 134         { set_max_cstate, "IBM ThinkPad R40e", {
 135           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 136           DMI_MATCH(DMI_BIOS_VERSION,"1SET45WW") }, (void*)1},
 137         { set_max_cstate, "IBM ThinkPad R40e", {
 138           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 139           DMI_MATCH(DMI_BIOS_VERSION,"1SET47WW") }, (void*)1},
 140         { set_max_cstate, "IBM ThinkPad R40e", {
 141           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 142           DMI_MATCH(DMI_BIOS_VERSION,"1SET50WW") }, (void*)1},
 143         { set_max_cstate, "IBM ThinkPad R40e", {
 144           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 145           DMI_MATCH(DMI_BIOS_VERSION,"1SET52WW") }, (void*)1},
 146         { set_max_cstate, "IBM ThinkPad R40e", {
 147           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 148           DMI_MATCH(DMI_BIOS_VERSION,"1SET55WW") }, (void*)1},
 149         { set_max_cstate, "IBM ThinkPad R40e", {
 150           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 151           DMI_MATCH(DMI_BIOS_VERSION,"1SET56WW") }, (void*)1},
 152         { set_max_cstate, "IBM ThinkPad R40e", {
 153           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 154           DMI_MATCH(DMI_BIOS_VERSION,"1SET59WW") }, (void*)1},
 155         { set_max_cstate, "IBM ThinkPad R40e", {
 156           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 157           DMI_MATCH(DMI_BIOS_VERSION,"1SET60WW") }, (void*)1},
 158         { set_max_cstate, "IBM ThinkPad R40e", {
 159           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 160           DMI_MATCH(DMI_BIOS_VERSION,"1SET61WW") }, (void*)1},
 161         { set_max_cstate, "IBM ThinkPad R40e", {
 162           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 163           DMI_MATCH(DMI_BIOS_VERSION,"1SET62WW") }, (void*)1},
 164         { set_max_cstate, "IBM ThinkPad R40e", {
 165           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 166           DMI_MATCH(DMI_BIOS_VERSION,"1SET64WW") }, (void*)1},
 167         { set_max_cstate, "IBM ThinkPad R40e", {
 168           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 169           DMI_MATCH(DMI_BIOS_VERSION,"1SET65WW") }, (void*)1},
 170         { set_max_cstate, "IBM ThinkPad R40e", {
 171           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 172           DMI_MATCH(DMI_BIOS_VERSION,"1SET68WW") }, (void*)1},
 173         { set_max_cstate, "Medion 41700", {
 174           DMI_MATCH(DMI_BIOS_VENDOR,"Phoenix Technologies LTD"),
 175           DMI_MATCH(DMI_BIOS_VERSION,"R01-A1J")}, (void *)1},
 176         { set_max_cstate, "Clevo 5600D", {
 177           DMI_MATCH(DMI_BIOS_VENDOR,"Phoenix Technologies LTD"),
 178           DMI_MATCH(DMI_BIOS_VERSION,"SHE845M0.86C.0013.D.0302131307")},
 179          (void *)2},
 180         {},
 181 };
 182
 183 static inline u32 ticks_elapsed(u32 t1, u32 t2)
 184 {
 185         if (t2 >= t1)
 186                 return (t2 - t1);
 187         else if (!(acpi_gbl_FADT.flags & ACPI_FADT_32BIT_TIMER))
 188                 return (((0x00FFFFFF - t1) + t2) & 0x00FFFFFF);
 189         else
 190                 return ((0xFFFFFFFF - t1) + t2);
 191 }
 192
 193 static void
 194 acpi_processor_power_activate(struct acpi_processor *pr,
 195                               struct acpi_processor_cx *new)
 196 {
 197         struct acpi_processor_cx *old;
 198
 199         if (!pr || !new)
 200                 return;
 201
 202         old = pr->power.state;
 203
 204         if (old)
 205                 old->promotion.count = 0;
 206         new->demotion.count = 0;
 207
 208         /* Cleanup from old state. */
 209         if (old) {
 210                 switch (old->type) {
 211                 case ACPI_STATE_C3:
 212                         /* Disable bus master reload */
 213                         if (new->type != ACPI_STATE_C3 && pr->flags.bm_check)
 214                                 acpi_set_register(ACPI_BITREG_BUS_MASTER_RLD, 0);
 215                         break;
 216                 }
 217         }
 218
 219         /* Prepare to use new state. */
 220         switch (new->type) {
 221         case ACPI_STATE_C3:
 222                 /* Enable bus master reload */
 223                 if (old->type != ACPI_STATE_C3 && pr->flags.bm_check)
 224                         acpi_set_register(ACPI_BITREG_BUS_MASTER_RLD, 1);
 225                 break;
 226         }
 227
 228         pr->power.state = new;
 229
 230         return;
 231 }
 232
 233 static void acpi_safe_halt(void)
 234 {
 235         current_thread_info()->status &= ~TS_POLLING;
 236         /*
 237          * TS_POLLING-cleared state must be visible before we
 238          * test NEED_RESCHED:
 239          */
 240         smp_mb();
 241         if (!need_resched())
 242                 safe_halt();
 243         current_thread_info()->status |= TS_POLLING;
 244 }
 245
 246 static atomic_t c3_cpu_count;
 247
 248 /* Common C-state entry for C2, C3, .. */
 249 static void acpi_cstate_enter(struct acpi_processor_cx *cstate)
 250 {
 251         if (cstate->space_id == ACPI_CSTATE_FFH) {
 252                 /* Call into architectural FFH based C-state */
 253                 acpi_processor_ffh_cstate_enter(cstate);
 254         } else {
 255                 int unused;
 256                 /* IO port based C-state */
 257                 inb(cstate->address);
 258                 /* Dummy wait op - must do something useless after P_LVL2 read
 259                    because chipsets cannot guarantee that STPCLK# signal
 260                    gets asserted in time to freeze execution properly. */
 261                 unused = inl(acpi_gbl_FADT.xpm_timer_block.address);
 262         }
 263 }
 264
 265 #ifdef ARCH_APICTIMER_STOPS_ON_C3
 266
 267 /*
 268  * Some BIOS implementations switch to C3 in the published C2 state.
 269  * This seems to be a common problem on AMD boxen and Intel Dothan/Banias
 270  * Pentium M machines.
 271  */
 272 static void acpi_timer_check_state(int state, struct acpi_processor *pr,
 273                                    struct acpi_processor_cx *cx)
 274 {
 275         struct acpi_processor_power *pwr = &pr->power;
 276
 277         /*
 278          * Check, if one of the previous states already marked the lapic
 279          * unstable
 280          */
 281         if (pwr->timer_broadcast_on_state < state)
 282                 return;
 283
 284         if (cx->type >= ACPI_STATE_C2) {
 285                 if (boot_cpu_data.x86_vendor == X86_VENDOR_AMD)
 286                         pr->power.timer_broadcast_on_state = state;
 287                 else if ((boot_cpu_data.x86_vendor == X86_VENDOR_INTEL &&
 288                           boot_cpu_data.x86 == 6) &&
 289                          (boot_cpu_data.x86_model == 13 ||
 290                           boot_cpu_data.x86_model == 9))
 291                 {
 292                         pr->power.timer_broadcast_on_state = state;
 293                 }
 294         }
 295 }
 296
 297 static void acpi_propagate_timer_broadcast(struct acpi_processor *pr)
 298 {
 299 #ifdef CONFIG_GENERIC_CLOCKEVENTS
 300         unsigned long reason;
 301
 302         reason = pr->power.timer_broadcast_on_state < INT_MAX ?
 303                 CLOCK_EVT_NOTIFY_BROADCAST_ON : CLOCK_EVT_NOTIFY_BROADCAST_OFF;
 304
 305         clockevents_notify(reason, &pr->id);
 306 #else
 307         cpumask_t mask = cpumask_of_cpu(pr->id);
 308
 309         if (use_ipi == 0)
 310                 on_each_cpu(switch_APIC_timer_to_ipi, &mask, 1, 1);
 311         else if (use_ipi == 1)
 312                 on_each_cpu(switch_ipi_to_APIC_timer, &mask, 1, 1);
 313         else {
 314                 if (pr->power.timer_broadcast_on_state < INT_MAX)
 315                         on_each_cpu(switch_APIC_timer_to_ipi, &mask, 1, 1);
 316                 else
 317                         on_each_cpu(switch_ipi_to_APIC_timer, &mask, 1, 1);
 318         }
 319 #endif
 320 }
 321
 322 /* Power(C) State timer broadcast control */
 323 static void acpi_state_timer_broadcast(struct acpi_processor *pr,
 324                                        struct acpi_processor_cx *cx,
 325                                        int broadcast)
 326 {
 327 #ifdef CONFIG_GENERIC_CLOCKEVENTS
 328
 329         int state = cx - pr->power.states;
 330
 331         if (state >= pr->power.timer_broadcast_on_state) {
 332                 unsigned long reason;
 333
 334                 reason = broadcast ?  CLOCK_EVT_NOTIFY_BROADCAST_ENTER :
 335                         CLOCK_EVT_NOTIFY_BROADCAST_EXIT;
 336                 clockevents_notify(reason, &pr->id);
 337         }
 338 #endif
 339 }
 340
 341 #else
 342
 343 static void acpi_timer_check_state(int state, struct acpi_processor *pr,
 344                                    struct acpi_processor_cx *cstate) { }
 345 static void acpi_propagate_timer_broadcast(struct acpi_processor *pr) { }
 346 static void acpi_state_timer_broadcast(struct acpi_processor *pr,
 347                                        struct acpi_processor_cx *cx,
 348                                        int broadcast)
 349 {
 350 }
 351
 352 #endif
 353
 354 static void acpi_processor_idle(void)
 355 {
 356         struct acpi_processor *pr = NULL;
 357         struct acpi_processor_cx *cx = NULL;
 358         struct acpi_processor_cx *next_state = NULL;
 359         int sleep_ticks = 0;
 360         u32 t1, t2 = 0;
 361
 362         pr = processors[smp_processor_id()];
 363         if (!pr)
 364                 return;
 365
 366         /*
 367          * Interrupts must be disabled during bus mastering calculations and
 368          * for C2/C3 transitions.
 369          */
 370         local_irq_disable();
 371
 372         /*
 373          * Check whether we truly need to go idle, or should
 374          * reschedule:
 375          */
 376         if (unlikely(need_resched())) {
 377                 local_irq_enable();
 378                 return;
 379         }
 380
 381         cx = pr->power.state;
 382         if (!cx) {
 383                 if (pm_idle_save)
 384                         pm_idle_save();
 385                 else
 386                         acpi_safe_halt();
 387                 return;
 388         }
 389
 390         /*
 391          * Check BM Activity
 392          * -----------------
 393          * Check for bus mastering activity (if required), record, and check
 394          * for demotion.
 395          */
 396         if (pr->flags.bm_check) {
 397                 u32 bm_status = 0;
 398                 unsigned long diff = jiffies - pr->power.bm_check_timestamp;
 399
 400                 if (diff > 31)
 401                         diff = 31;
 402
 403                 pr->power.bm_activity <<= diff;
 404
 405                 acpi_get_register(ACPI_BITREG_BUS_MASTER_STATUS, &bm_status);
 406                 if (bm_status) {
 407                         pr->power.bm_activity |= 0x1;
 408                         acpi_set_register(ACPI_BITREG_BUS_MASTER_STATUS, 1);
 409                 }
 410                 /*
 411                  * PIIX4 Erratum #18: Note that BM_STS doesn't always reflect
 412                  * the true state of bus mastering activity; forcing us to
 413                  * manually check the BMIDEA bit of each IDE channel.
 414                  */
 415                 else if (errata.piix4.bmisx) {
 416                         if ((inb_p(errata.piix4.bmisx + 0x02) & 0x01)
 417                             || (inb_p(errata.piix4.bmisx + 0x0A) & 0x01))
 418                                 pr->power.bm_activity |= 0x1;
 419                 }
 420
 421                 pr->power.bm_check_timestamp = jiffies;
 422
 423                 /*
 424                  * If bus mastering is or was active this jiffy, demote
 425                  * to avoid a faulty transition.  Note that the processor
 426                  * won't enter a low-power state during this call (to this
 427                  * function) but should upon the next.
 428                  *
 429                  * TBD: A better policy might be to fallback to the demotion
 430                  *      state (use it for this quantum only) istead of
 431                  *      demoting -- and rely on duration as our sole demotion
 432                  *      qualification.  This may, however, introduce DMA
 433                  *      issues (e.g. floppy DMA transfer overrun/underrun).
 434                  */
 435                 if ((pr->power.bm_activity & 0x1) &&
 436                     cx->demotion.threshold.bm) {
 437                         local_irq_enable();
 438                         next_state = cx->demotion.state;
 439                         goto end;
 440                 }
 441         }
 442
 443 #ifdef CONFIG_HOTPLUG_CPU
 444         /*
 445          * Check for P_LVL2_UP flag before entering C2 and above on
 446          * an SMP system. We do it here instead of doing it at _CST/P_LVL
 447          * detection phase, to work cleanly with logical CPU hotplug.
 448          */
 449         if ((cx->type != ACPI_STATE_C1) && (num_online_cpus() > 1) &&
 450             !pr->flags.has_cst && !(acpi_gbl_FADT.flags & ACPI_FADT_C2_MP_SUPPORTED))
 451                 cx = &pr->power.states[ACPI_STATE_C1];
 452 #endif
 453
 454         /*
 455          * Sleep:
 456          * ------
 457          * Invoke the current Cx state to put the processor to sleep.
 458          */
 459         if (cx->type == ACPI_STATE_C2 || cx->type == ACPI_STATE_C3) {
 460                 current_thread_info()->status &= ~TS_POLLING;
 461                 /*
 462                  * TS_POLLING-cleared state must be visible before we
 463                  * test NEED_RESCHED:
 464                  */
 465                 smp_mb();
 466                 if (need_resched()) {
 467                         current_thread_info()->status |= TS_POLLING;
 468                         local_irq_enable();
 469                         return;
 470                 }
 471         }
 472
 473         switch (cx->type) {
 474
 475         case ACPI_STATE_C1:
 476                 /*
 477                  * Invoke C1.
 478                  * Use the appropriate idle routine, the one that would
 479                  * be used without acpi C-states.
 480                  */
 481                 if (pm_idle_save)
 482                         pm_idle_save();
 483                 else
 484                         acpi_safe_halt();
 485
 486                 /*
 487                  * TBD: Can't get time duration while in C1, as resumes
 488                  *      go to an ISR rather than here.  Need to instrument
 489                  *      base interrupt handler.
 490                  */
 491                 sleep_ticks = 0xFFFFFFFF;
 492                 break;
 493
 494         case ACPI_STATE_C2:
 495                 /* Get start time (ticks) */
 496                 t1 = inl(acpi_gbl_FADT.xpm_timer_block.address);
 497                 /* Invoke C2 */
 498                 acpi_state_timer_broadcast(pr, cx, 1);
 499                 acpi_cstate_enter(cx);
 500                 /* Get end time (ticks) */
 501                 t2 = inl(acpi_gbl_FADT.xpm_timer_block.address);
 502
 503 #ifdef CONFIG_GENERIC_TIME
 504                 /* TSC halts in C2, so notify users */
 505                 mark_tsc_unstable();
 506 #endif
 507                 /* Re-enable interrupts */
 508                 local_irq_enable();
 509                 current_thread_info()->status |= TS_POLLING;
 510                 /* Compute time (ticks) that we were actually asleep */
 511                 sleep_ticks =
 512                     ticks_elapsed(t1, t2) - cx->latency_ticks - C2_OVERHEAD;
 513                 acpi_state_timer_broadcast(pr, cx, 0);
 514                 break;
 515
 516         case ACPI_STATE_C3:
 517
 518                 if (pr->flags.bm_check) {
 519                         if (atomic_inc_return(&c3_cpu_count) ==
 520                             num_online_cpus()) {
 521                                 /*
 522                                  * All CPUs are trying to go to C3
 523                                  * Disable bus master arbitration
 524                                  */
 525                                 acpi_set_register(ACPI_BITREG_ARB_DISABLE, 1);
 526                         }
 527                 } else {
 528                         /* SMP with no shared cache... Invalidate cache  */
 529                         ACPI_FLUSH_CPU_CACHE();
 530                 }
 531
 532                 /* Get start time (ticks) */
 533                 t1 = inl(acpi_gbl_FADT.xpm_timer_block.address);
 534                 /* Invoke C3 */
 535                 acpi_state_timer_broadcast(pr, cx, 1);
 536                 acpi_cstate_enter(cx);
 537                 /* Get end time (ticks) */
 538                 t2 = inl(acpi_gbl_FADT.xpm_timer_block.address);
 539                 if (pr->flags.bm_check) {
 540                         /* Enable bus master arbitration */
 541                         atomic_dec(&c3_cpu_count);
 542                         acpi_set_register(ACPI_BITREG_ARB_DISABLE, 0);
 543                 }
 544
 545 #ifdef CONFIG_GENERIC_TIME
 546                 /* TSC halts in C3, so notify users */
 547                 mark_tsc_unstable();
 548 #endif
 549                 /* Re-enable interrupts */
 550                 local_irq_enable();
 551                 current_thread_info()->status |= TS_POLLING;
 552                 /* Compute time (ticks) that we were actually asleep */
 553                 sleep_ticks =
 554                     ticks_elapsed(t1, t2) - cx->latency_ticks - C3_OVERHEAD;
 555                 acpi_state_timer_broadcast(pr, cx, 0);
 556                 break;
 557
 558         default:
 559                 local_irq_enable();
 560                 return;
 561         }
 562         cx->usage++;
 563         if ((cx->type != ACPI_STATE_C1) && (sleep_ticks > 0))
 564                 cx->time += sleep_ticks;
 565
 566         next_state = pr->power.state;
 567
 568 #ifdef CONFIG_HOTPLUG_CPU
 569         /* Don't do promotion/demotion */
 570         if ((cx->type == ACPI_STATE_C1) && (num_online_cpus() > 1) &&
 571             !pr->flags.has_cst && !(acpi_gbl_FADT.flags & ACPI_FADT_C2_MP_SUPPORTED)) {
 572                 next_state = cx;
 573                 goto end;
 574         }
 575 #endif
 576
 577         /*
 578          * Promotion?
 579          * ----------
 580          * Track the number of longs (time asleep is greater than threshold)
 581          * and promote when the count threshold is reached.  Note that bus
 582          * mastering activity may prevent promotions.
 583          * Do not promote above max_cstate.
 584          */
 585         if (cx->promotion.state &&
 586             ((cx->promotion.state - pr->power.states) <= max_cstate)) {
 587                 if (sleep_ticks > cx->promotion.threshold.ticks &&
 588                   cx->promotion.state->latency <= system_latency_constraint()) {
 589                         cx->promotion.count++;
 590                         cx->demotion.count = 0;
 591                         if (cx->promotion.count >=
 592                             cx->promotion.threshold.count) {
 593                                 if (pr->flags.bm_check) {
 594                                         if (!
 595                                             (pr->power.bm_activity & cx->
 596                                              promotion.threshold.bm)) {
 597                                                 next_state =
 598                                                     cx->promotion.state;
 599                                                 goto end;
 600                                         }
 601                                 } else {
 602                                         next_state = cx->promotion.state;
 603                                         goto end;
 604                                 }
 605                         }
 606                 }
 607         }
 608
 609         /*
 610          * Demotion?
 611          * ---------
 612          * Track the number of shorts (time asleep is less than time threshold)
 613          * and demote when the usage threshold is reached.
 614          */
 615         if (cx->demotion.state) {
 616                 if (sleep_ticks < cx->demotion.threshold.ticks) {
 617                         cx->demotion.count++;
 618                         cx->promotion.count = 0;
 619                         if (cx->demotion.count >= cx->demotion.threshold.count) {
 620                                 next_state = cx->demotion.state;
 621                                 goto end;
 622                         }
 623                 }
 624         }
 625
 626       end:
 627         /*
 628          * Demote if current state exceeds max_cstate
 629          * or if the latency of the current state is unacceptable
 630          */
 631         if ((pr->power.state - pr->power.states) > max_cstate ||
 632                 pr->power.state->latency > system_latency_constraint()) {
 633                 if (cx->demotion.state)
 634                         next_state = cx->demotion.state;
 635         }
 636
 637         /*
 638          * New Cx State?
 639          * -------------
 640          * If we're going to start using a new Cx state we must clean up
 641          * from the previous and prepare to use the new.
 642          */
 643         if (next_state != pr->power.state)
 644                 acpi_processor_power_activate(pr, next_state);
 645 }
 646
 647 static int acpi_processor_set_power_policy(struct acpi_processor *pr)
 648 {
 649         unsigned int i;
 650         unsigned int state_is_set = 0;
 651         struct acpi_processor_cx *lower = NULL;
 652         struct acpi_processor_cx *higher = NULL;
 653         struct acpi_processor_cx *cx;
 654
 655
 656         if (!pr)
 657                 return -EINVAL;
 658
 659         /*
 660          * This function sets the default Cx state policy (OS idle handler).
 661          * Our scheme is to promote quickly to C2 but more conservatively
 662          * to C3.  We're favoring C2  for its characteristics of low latency
 663          * (quick response), good power savings, and ability to allow bus
 664          * mastering activity.  Note that the Cx state policy is completely
 665          * customizable and can be altered dynamically.
 666          */
 667
 668         /* startup state */
 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 (!state_is_set)
 675                         pr->power.state = cx;
 676                 state_is_set++;
 677                 break;
 678         }
 679
 680         if (!state_is_set)
 681                 return -ENODEV;
 682
 683         /* demotion */
 684         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
 685                 cx = &pr->power.states[i];
 686                 if (!cx->valid)
 687                         continue;
 688
 689                 if (lower) {
 690                         cx->demotion.state = lower;
 691                         cx->demotion.threshold.ticks = cx->latency_ticks;
 692                         cx->demotion.threshold.count = 1;
 693                         if (cx->type == ACPI_STATE_C3)
 694                                 cx->demotion.threshold.bm = bm_history;
 695                 }
 696
 697                 lower = cx;
 698         }
 699
 700         /* promotion */
 701         for (i = (ACPI_PROCESSOR_MAX_POWER - 1); i > 0; i--) {
 702                 cx = &pr->power.states[i];
 703                 if (!cx->valid)
 704                         continue;
 705
 706                 if (higher) {
 707                         cx->promotion.state = higher;
 708                         cx->promotion.threshold.ticks = cx->latency_ticks;
 709                         if (cx->type >= ACPI_STATE_C2)
 710                                 cx->promotion.threshold.count = 4;
 711                         else
 712                                 cx->promotion.threshold.count = 10;
 713                         if (higher->type == ACPI_STATE_C3)
 714                                 cx->promotion.threshold.bm = bm_history;
 715                 }
 716
 717                 higher = cx;
 718         }
 719
 720         return 0;
 721 }
 722
 723 static int acpi_processor_get_power_info_fadt(struct acpi_processor *pr)
 724 {
 725
 726         if (!pr)
 727                 return -EINVAL;
 728
 729         if (!pr->pblk)
 730                 return -ENODEV;
 731
 732         /* if info is obtained from pblk/fadt, type equals state */
 733         pr->power.states[ACPI_STATE_C2].type = ACPI_STATE_C2;
 734         pr->power.states[ACPI_STATE_C3].type = ACPI_STATE_C3;
 735
 736 #ifndef CONFIG_HOTPLUG_CPU
 737         /*
 738          * Check for P_LVL2_UP flag before entering C2 and above on
 739          * an SMP system.
 740          */
 741         if ((num_online_cpus() > 1) &&
 742             !(acpi_gbl_FADT.flags & ACPI_FADT_C2_MP_SUPPORTED))
 743                 return -ENODEV;
 744 #endif
 745
 746         /* determine C2 and C3 address from pblk */
 747         pr->power.states[ACPI_STATE_C2].address = pr->pblk + 4;
 748         pr->power.states[ACPI_STATE_C3].address = pr->pblk + 5;
 749
 750         /* determine latencies from FADT */
 751         pr->power.states[ACPI_STATE_C2].latency = acpi_gbl_FADT.C2latency;
 752         pr->power.states[ACPI_STATE_C3].latency = acpi_gbl_FADT.C3latency;
 753
 754         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 755                           "lvl2[0x%08x] lvl3[0x%08x]\n",
 756                           pr->power.states[ACPI_STATE_C2].address,
 757                           pr->power.states[ACPI_STATE_C3].address));
 758
 759         return 0;
 760 }
 761
 762 static int acpi_processor_get_power_info_default(struct acpi_processor *pr)
 763 {
 764         if (!pr->power.states[ACPI_STATE_C1].valid) {
 765                 /* set the first C-State to C1 */
 766                 /* all processors need to support C1 */
 767                 pr->power.states[ACPI_STATE_C1].type = ACPI_STATE_C1;
 768                 pr->power.states[ACPI_STATE_C1].valid = 1;
 769         }
 770         /* the C0 state only exists as a filler in our array */
 771         pr->power.states[ACPI_STATE_C0].valid = 1;
 772         return 0;
 773 }
 774
 775 static int acpi_processor_get_power_info_cst(struct acpi_processor *pr)
 776 {
 777         acpi_status status = 0;
 778         acpi_integer count;
 779         int current_count;
 780         int i;
 781         struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
 782         union acpi_object *cst;
 783
 784
 785         if (nocst)
 786                 return -ENODEV;
 787
 788         current_count = 0;
 789
 790         status = acpi_evaluate_object(pr->handle, "_CST", NULL, &buffer);
 791         if (ACPI_FAILURE(status)) {
 792                 ACPI_DEBUG_PRINT((ACPI_DB_INFO, "No _CST, giving up\n"));
 793                 return -ENODEV;
 794         }
 795
 796         cst = buffer.pointer;
 797
 798         /* There must be at least 2 elements */
 799         if (!cst || (cst->type != ACPI_TYPE_PACKAGE) || cst->package.count < 2) {
 800                 printk(KERN_ERR PREFIX "not enough elements in _CST\n");
 801                 status = -EFAULT;
 802                 goto end;
 803         }
 804
 805         count = cst->package.elements[0].integer.value;
 806
 807         /* Validate number of power states. */
 808         if (count < 1 || count != cst->package.count - 1) {
 809                 printk(KERN_ERR PREFIX "count given by _CST is not valid\n");
 810                 status = -EFAULT;
 811                 goto end;
 812         }
 813
 814         /* Tell driver that at least _CST is supported. */
 815         pr->flags.has_cst = 1;
 816
 817         for (i = 1; i <= count; i++) {
 818                 union acpi_object *element;
 819                 union acpi_object *obj;
 820                 struct acpi_power_register *reg;
 821                 struct acpi_processor_cx cx;
 822
 823                 memset(&cx, 0, sizeof(cx));
 824
 825                 element = &(cst->package.elements[i]);
 826                 if (element->type != ACPI_TYPE_PACKAGE)
 827                         continue;
 828
 829                 if (element->package.count != 4)
 830                         continue;
 831
 832                 obj = &(element->package.elements[0]);
 833
 834                 if (obj->type != ACPI_TYPE_BUFFER)
 835                         continue;
 836
 837                 reg = (struct acpi_power_register *)obj->buffer.pointer;
 838
 839                 if (reg->space_id != ACPI_ADR_SPACE_SYSTEM_IO &&
 840                     (reg->space_id != ACPI_ADR_SPACE_FIXED_HARDWARE))
 841                         continue;
 842
 843                 /* There should be an easy way to extract an integer... */
 844                 obj = &(element->package.elements[1]);
 845                 if (obj->type != ACPI_TYPE_INTEGER)
 846                         continue;
 847
 848                 cx.type = obj->integer.value;
 849                 /*
 850                  * Some buggy BIOSes won't list C1 in _CST -
 851                  * Let acpi_processor_get_power_info_default() handle them later
 852                  */
 853                 if (i == 1 && cx.type != ACPI_STATE_C1)
 854                         current_count++;
 855
 856                 cx.address = reg->address;
 857                 cx.index = current_count + 1;
 858
 859                 cx.space_id = ACPI_CSTATE_SYSTEMIO;
 860                 if (reg->space_id == ACPI_ADR_SPACE_FIXED_HARDWARE) {
 861                         if (acpi_processor_ffh_cstate_probe
 862                                         (pr->id, &cx, reg) == 0) {
 863                                 cx.space_id = ACPI_CSTATE_FFH;
 864                         } else if (cx.type != ACPI_STATE_C1) {
 865                                 /*
 866                                  * C1 is a special case where FIXED_HARDWARE
 867                                  * can be handled in non-MWAIT way as well.
 868                                  * In that case, save this _CST entry info.
 869                                  * That is, we retain space_id of SYSTEM_IO for
 870                                  * halt based C1.
 871                                  * Otherwise, ignore this info and continue.
 872                                  */
 873                                 continue;
 874                         }
 875                 }
 876
 877                 obj = &(element->package.elements[2]);
 878                 if (obj->type != ACPI_TYPE_INTEGER)
 879                         continue;
 880
 881                 cx.latency = obj->integer.value;
 882
 883                 obj = &(element->package.elements[3]);
 884                 if (obj->type != ACPI_TYPE_INTEGER)
 885                         continue;
 886
 887                 cx.power = obj->integer.value;
 888
 889                 current_count++;
 890                 memcpy(&(pr->power.states[current_count]), &cx, sizeof(cx));
 891
 892                 /*
 893                  * We support total ACPI_PROCESSOR_MAX_POWER - 1
 894                  * (From 1 through ACPI_PROCESSOR_MAX_POWER - 1)
 895                  */
 896                 if (current_count >= (ACPI_PROCESSOR_MAX_POWER - 1)) {
 897                         printk(KERN_WARNING
 898                                "Limiting number of power states to max (%d)\n",
 899                                ACPI_PROCESSOR_MAX_POWER);
 900                         printk(KERN_WARNING
 901                                "Please increase ACPI_PROCESSOR_MAX_POWER if needed.\n");
 902                         break;
 903                 }
 904         }
 905
 906         ACPI_DEBUG_PRINT((ACPI_DB_INFO, "Found %d power states\n",
 907                           current_count));
 908
 909         /* Validate number of power states discovered */
 910         if (current_count < 2)
 911                 status = -EFAULT;
 912
 913       end:
 914         kfree(buffer.pointer);
 915
 916         return status;
 917 }
 918
 919 static void acpi_processor_power_verify_c2(struct acpi_processor_cx *cx)
 920 {
 921
 922         if (!cx->address)
 923                 return;
 924
 925         /*
 926          * C2 latency must be less than or equal to 100
 927          * microseconds.
 928          */
 929         else if (cx->latency > ACPI_PROCESSOR_MAX_C2_LATENCY) {
 930                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 931                                   "latency too large [%d]\n", cx->latency));
 932                 return;
 933         }
 934
 935         /*
 936          * Otherwise we've met all of our C2 requirements.
 937          * Normalize the C2 latency to expidite policy
 938          */
 939         cx->valid = 1;
 940         cx->latency_ticks = US_TO_PM_TIMER_TICKS(cx->latency);
 941
 942         return;
 943 }
 944
 945 static void acpi_processor_power_verify_c3(struct acpi_processor *pr,
 946                                            struct acpi_processor_cx *cx)
 947 {
 948         static int bm_check_flag;
 949
 950
 951         if (!cx->address)
 952                 return;
 953
 954         /*
 955          * C3 latency must be less than or equal to 1000
 956          * microseconds.
 957          */
 958         else if (cx->latency > ACPI_PROCESSOR_MAX_C3_LATENCY) {
 959                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 960                                   "latency too large [%d]\n", cx->latency));
 961                 return;
 962         }
 963
 964         /*
 965          * PIIX4 Erratum #18: We don't support C3 when Type-F (fast)
 966          * DMA transfers are used by any ISA device to avoid livelock.
 967          * Note that we could disable Type-F DMA (as recommended by
 968          * the erratum), but this is known to disrupt certain ISA
 969          * devices thus we take the conservative approach.
 970          */
 971         else if (errata.piix4.fdma) {
 972                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 973                                   "C3 not supported on PIIX4 with Type-F DMA\n"));
 974                 return;
 975         }
 976
 977         /* All the logic here assumes flags.bm_check is same across all CPUs */
 978         if (!bm_check_flag) {
 979                 /* Determine whether bm_check is needed based on CPU  */
 980                 acpi_processor_power_init_bm_check(&(pr->flags), pr->id);
 981                 bm_check_flag = pr->flags.bm_check;
 982         } else {
 983                 pr->flags.bm_check = bm_check_flag;
 984         }
 985
 986         if (pr->flags.bm_check) {
 987                 /* bus mastering control is necessary */
 988                 if (!pr->flags.bm_control) {
 989                         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 990                                           "C3 support requires bus mastering control\n"));
 991                         return;
 992                 }
 993         } else {
 994                 /*
 995                  * WBINVD should be set in fadt, for C3 state to be
 996                  * supported on when bm_check is not required.
 997                  */
 998                 if (!(acpi_gbl_FADT.flags & ACPI_FADT_WBINVD)) {
 999                         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
1000                                           "Cache invalidation should work properly"
1001                                           " for C3 to be enabled on SMP systems\n"));
1002                         return;
1003                 }
1004                 acpi_set_register(ACPI_BITREG_BUS_MASTER_RLD, 0);
1005         }
1006
1007         /*
1008          * Otherwise we've met all of our C3 requirements.
1009          * Normalize the C3 latency to expidite policy.  Enable
1010          * checking of bus mastering status (bm_check) so we can
1011          * use this in our C3 policy
1012          */
1013         cx->valid = 1;
1014         cx->latency_ticks = US_TO_PM_TIMER_TICKS(cx->latency);
1015
1016         return;
1017 }
1018
1019 static int acpi_processor_power_verify(struct acpi_processor *pr)
1020 {
1021         unsigned int i;
1022         unsigned int working = 0;
1023
1024         pr->power.timer_broadcast_on_state = INT_MAX;
1025
1026         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
1027                 struct acpi_processor_cx *cx = &pr->power.states[i];
1028
1029                 switch (cx->type) {
1030                 case ACPI_STATE_C1:
1031                         cx->valid = 1;
1032                         break;
1033
1034                 case ACPI_STATE_C2:
1035                         acpi_processor_power_verify_c2(cx);
1036                         if (cx->valid && use_ipi != 0 && use_ipi != 1)
1037                                 acpi_timer_check_state(i, pr, cx);
1038                         break;
1039
1040                 case ACPI_STATE_C3:
1041                         acpi_processor_power_verify_c3(pr, cx);
1042                         if (cx->valid && use_ipi != 0 && use_ipi != 1)
1043                                 acpi_timer_check_state(i, pr, cx);
1044                         break;
1045                 }
1046
1047                 if (cx->valid)
1048                         working++;
1049         }
1050
1051         acpi_propagate_timer_broadcast(pr);
1052
1053         return (working);
1054 }
1055
1056 static int acpi_processor_get_power_info(struct acpi_processor *pr)
1057 {
1058         unsigned int i;
1059         int result;
1060
1061
1062         /* NOTE: the idle thread may not be running while calling
1063          * this function */
1064
1065         /* Zero initialize all the C-states info. */
1066         memset(pr->power.states, 0, sizeof(pr->power.states));
1067
1068         result = acpi_processor_get_power_info_cst(pr);
1069         if (result == -ENODEV)
1070                 result = acpi_processor_get_power_info_fadt(pr);
1071
1072         if (result)
1073                 return result;
1074
1075         acpi_processor_get_power_info_default(pr);
1076
1077         pr->power.count = acpi_processor_power_verify(pr);
1078
1079         /*
1080          * Set Default Policy
1081          * ------------------
1082          * Now that we know which states are supported, set the default
1083          * policy.  Note that this policy can be changed dynamically
1084          * (e.g. encourage deeper sleeps to conserve battery life when
1085          * not on AC).
1086          */
1087         result = acpi_processor_set_power_policy(pr);
1088         if (result)
1089                 return result;
1090
1091         /*
1092          * if one state of type C2 or C3 is available, mark this
1093          * CPU as being "idle manageable"
1094          */
1095         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
1096                 if (pr->power.states[i].valid) {
1097                         pr->power.count = i;
1098                         if (pr->power.states[i].type >= ACPI_STATE_C2)
1099                                 pr->flags.power = 1;
1100                 }
1101         }
1102
1103         return 0;
1104 }
1105
1106 int acpi_processor_cst_has_changed(struct acpi_processor *pr)
1107 {
1108         int result = 0;
1109
1110
1111         if (!pr)
1112                 return -EINVAL;
1113
1114         if (nocst) {
1115                 return -ENODEV;
1116         }
1117
1118         if (!pr->flags.power_setup_done)
1119                 return -ENODEV;
1120
1121         /* Fall back to the default idle loop */
1122         pm_idle = pm_idle_save;
1123         synchronize_sched();    /* Relies on interrupts forcing exit from idle. */
1124
1125         pr->flags.power = 0;
1126         result = acpi_processor_get_power_info(pr);
1127         if ((pr->flags.power == 1) && (pr->flags.power_setup_done))
1128                 pm_idle = acpi_processor_idle;
1129
1130         return result;
1131 }
1132
1133 /* proc interface */
1134
1135 static int acpi_processor_power_seq_show(struct seq_file *seq, void *offset)
1136 {
1137         struct acpi_processor *pr = seq->private;
1138         unsigned int i;
1139
1140
1141         if (!pr)
1142                 goto end;
1143
1144         seq_printf(seq, "active state:            C%zd\n"
1145                    "max_cstate:              C%d\n"
1146                    "bus master activity:     %08x\n"
1147                    "maximum allowed latency: %d usec\n",
1148                    pr->power.state ? pr->power.state - pr->power.states : 0,
1149                    max_cstate, (unsigned)pr->power.bm_activity,
1150                    system_latency_constraint());
1151
1152         seq_puts(seq, "states:\n");
1153
1154         for (i = 1; i <= pr->power.count; i++) {
1155                 seq_printf(seq, "   %cC%d:                  ",
1156                            (&pr->power.states[i] ==
1157                             pr->power.state ? '*' : ' '), i);
1158
1159                 if (!pr->power.states[i].valid) {
1160                         seq_puts(seq, "<not supported>\n");
1161                         continue;
1162                 }
1163
1164                 switch (pr->power.states[i].type) {
1165                 case ACPI_STATE_C1:
1166                         seq_printf(seq, "type[C1] ");
1167                         break;
1168                 case ACPI_STATE_C2:
1169                         seq_printf(seq, "type[C2] ");
1170                         break;
1171                 case ACPI_STATE_C3:
1172                         seq_printf(seq, "type[C3] ");
1173                         break;
1174                 default:
1175                         seq_printf(seq, "type[--] ");
1176                         break;
1177                 }
1178
1179                 if (pr->power.states[i].promotion.state)
1180                         seq_printf(seq, "promotion[C%zd] ",
1181                                    (pr->power.states[i].promotion.state -
1182                                     pr->power.states));
1183                 else
1184                         seq_puts(seq, "promotion[--] ");
1185
1186                 if (pr->power.states[i].demotion.state)
1187                         seq_printf(seq, "demotion[C%zd] ",
1188                                    (pr->power.states[i].demotion.state -
1189                                     pr->power.states));
1190                 else
1191                         seq_puts(seq, "demotion[--] ");
1192
1193                 seq_printf(seq, "latency[%03d] usage[%08d] duration[%020llu]\n",
1194                            pr->power.states[i].latency,
1195                            pr->power.states[i].usage,
1196                            (unsigned long long)pr->power.states[i].time);
1197         }
1198
1199       end:
1200         return 0;
1201 }
1202
1203 static int acpi_processor_power_open_fs(struct inode *inode, struct file *file)
1204 {
1205         return single_open(file, acpi_processor_power_seq_show,
1206                            PDE(inode)->data);
1207 }
1208
1209 static const struct file_operations acpi_processor_power_fops = {
1210         .open = acpi_processor_power_open_fs,
1211         .read = seq_read,
1212         .llseek = seq_lseek,
1213         .release = single_release,
1214 };
1215
1216 #ifdef CONFIG_SMP
1217 static void smp_callback(void *v)
1218 {
1219         /* we already woke the CPU up, nothing more to do */
1220 }
1221
1222 /*
1223  * This function gets called when a part of the kernel has a new latency
1224  * requirement.  This means we need to get all processors out of their C-state,
1225  * and then recalculate a new suitable C-state. Just do a cross-cpu IPI; that
1226  * wakes them all right up.
1227  */
1228 static int acpi_processor_latency_notify(struct notifier_block *b,
1229                 unsigned long l, void *v)
1230 {
1231         smp_call_function(smp_callback, NULL, 0, 1);
1232         return NOTIFY_OK;
1233 }
1234
1235 static struct notifier_block acpi_processor_latency_notifier = {
1236         .notifier_call = acpi_processor_latency_notify,
1237 };
1238 #endif
1239
1240 int __cpuinit acpi_processor_power_init(struct acpi_processor *pr,
1241                               struct acpi_device *device)
1242 {
1243         acpi_status status = 0;
1244         static int first_run;
1245         struct proc_dir_entry *entry = NULL;
1246         unsigned int i;
1247
1248
1249         if (!first_run) {
1250                 dmi_check_system(processor_power_dmi_table);
1251                 if (max_cstate < ACPI_C_STATES_MAX)
1252                         printk(KERN_NOTICE
1253                                "ACPI: processor limited to max C-state %d\n",
1254                                max_cstate);
1255                 first_run++;
1256 #ifdef CONFIG_SMP
1257                 register_latency_notifier(&acpi_processor_latency_notifier);
1258 #endif
1259         }
1260
1261         if (!pr)
1262                 return -EINVAL;
1263
1264         if (acpi_gbl_FADT.cst_control && !nocst) {
1265                 status =
1266                     acpi_os_write_port(acpi_gbl_FADT.smi_command, acpi_gbl_FADT.cst_control, 8);
1267                 if (ACPI_FAILURE(status)) {
1268                         ACPI_EXCEPTION((AE_INFO, status,
1269                                         "Notifying BIOS of _CST ability failed"));
1270                 }
1271         }
1272
1273         acpi_processor_get_power_info(pr);
1274
1275         /*
1276          * Install the idle handler if processor power management is supported.
1277          * Note that we use previously set idle handler will be used on
1278          * platforms that only support C1.
1279          */
1280         if ((pr->flags.power) && (!boot_option_idle_override)) {
1281                 printk(KERN_INFO PREFIX "CPU%d (power states:", pr->id);
1282                 for (i = 1; i <= pr->power.count; i++)
1283                         if (pr->power.states[i].valid)
1284                                 printk(" C%d[C%d]", i,
1285                                        pr->power.states[i].type);
1286                 printk(")\n");
1287
1288                 if (pr->id == 0) {
1289                         pm_idle_save = pm_idle;
1290                         pm_idle = acpi_processor_idle;
1291                 }
1292         }
1293
1294         /* 'power' [R] */
1295         entry = create_proc_entry(ACPI_PROCESSOR_FILE_POWER,
1296                                   S_IRUGO, acpi_device_dir(device));
1297         if (!entry)
1298                 return -EIO;
1299         else {
1300                 entry->proc_fops = &acpi_processor_power_fops;
1301                 entry->data = acpi_driver_data(device);
1302                 entry->owner = THIS_MODULE;
1303         }
1304
1305         pr->flags.power_setup_done = 1;
1306
1307         return 0;
1308 }
1309
1310 int acpi_processor_power_exit(struct acpi_processor *pr,
1311                               struct acpi_device *device)
1312 {
1313
1314         pr->flags.power_setup_done = 0;
1315
1316         if (acpi_device_dir(device))
1317                 remove_proc_entry(ACPI_PROCESSOR_FILE_POWER,
1318                                   acpi_device_dir(device));
1319
1320         /* Unregister the idle handler when processor #0 is removed. */
1321         if (pr->id == 0) {
1322                 pm_idle = pm_idle_save;
1323
1324                 /*
1325                  * We are about to unload the current idle thread pm callback
1326                  * (pm_idle), Wait for all processors to update cached/local
1327                  * copies of pm_idle before proceeding.
1328                  */
1329                 cpu_idle_wait();
1330 #ifdef CONFIG_SMP
1331                 unregister_latency_notifier(&acpi_processor_latency_notifier);
1332 #endif
1333         }
1334
1335         return 0;
1336 }