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/pm_qos_params.h>
  42 #include <linux/clockchips.h>
  43 #include <linux/cpuidle.h>
  44
  45 /*
  46  * Include the apic definitions for x86 to have the APIC timer related defines
  47  * available also for UP (on SMP it gets magically included via linux/smp.h).
  48  * asm/acpi.h is not an option, as it would require more include magic. Also
  49  * creating an empty asm-ia64/apic.h would just trade pest vs. cholera.
  50  */
  51 #ifdef CONFIG_X86
  52 #include <asm/apic.h>
  53 #endif
  54
  55 #include <asm/io.h>
  56 #include <asm/uaccess.h>
  57
  58 #include <acpi/acpi_bus.h>
  59 #include <acpi/processor.h>
  60
  61 #define ACPI_PROCESSOR_COMPONENT        0x01000000
  62 #define ACPI_PROCESSOR_CLASS            "processor"
  63 #define _COMPONENT              ACPI_PROCESSOR_COMPONENT
  64 ACPI_MODULE_NAME("processor_idle");
  65 #define ACPI_PROCESSOR_FILE_POWER       "power"
  66 #define US_TO_PM_TIMER_TICKS(t)         ((t * (PM_TIMER_FREQUENCY/1000)) / 1000)
  67 #define PM_TIMER_TICK_NS                (1000000000ULL/PM_TIMER_FREQUENCY)
  68 #ifndef CONFIG_CPU_IDLE
  69 #define C2_OVERHEAD                     4       /* 1us (3.579 ticks per us) */
  70 #define C3_OVERHEAD                     4       /* 1us (3.579 ticks per us) */
  71 static void (*pm_idle_save) (void) __read_mostly;
  72 #else
  73 #define C2_OVERHEAD                     1       /* 1us */
  74 #define C3_OVERHEAD                     1       /* 1us */
  75 #endif
  76 #define PM_TIMER_TICKS_TO_US(p)         (((p) * 1000)/(PM_TIMER_FREQUENCY/1000))
  77
  78 static unsigned int max_cstate __read_mostly = ACPI_PROCESSOR_MAX_POWER;
  79 #ifdef CONFIG_CPU_IDLE
  80 module_param(max_cstate, uint, 0000);
  81 #else
  82 module_param(max_cstate, uint, 0644);
  83 #endif
  84 static unsigned int nocst __read_mostly;
  85 module_param(nocst, uint, 0000);
  86
  87 #ifndef CONFIG_CPU_IDLE
  88 /*
  89  * bm_history -- bit-mask with a bit per jiffy of bus-master activity
  90  * 1000 HZ: 0xFFFFFFFF: 32 jiffies = 32ms
  91  * 800 HZ: 0xFFFFFFFF: 32 jiffies = 40ms
  92  * 100 HZ: 0x0000000F: 4 jiffies = 40ms
  93  * reduce history for more aggressive entry into C3
  94  */
  95 static unsigned int bm_history __read_mostly =
  96     (HZ >= 800 ? 0xFFFFFFFF : ((1U << (HZ / 25)) - 1));
  97 module_param(bm_history, uint, 0644);
  98
  99 static int acpi_processor_set_power_policy(struct acpi_processor *pr);
 100
 101 #else   /* CONFIG_CPU_IDLE */
 102 static unsigned int latency_factor __read_mostly = 2;
 103 module_param(latency_factor, uint, 0644);
 104 #endif
 105
 106 /*
 107  * IBM ThinkPad R40e crashes mysteriously when going into C2 or C3.
 108  * For now disable this. Probably a bug somewhere else.
 109  *
 110  * To skip this limit, boot/load with a large max_cstate limit.
 111  */
 112 static int set_max_cstate(const struct dmi_system_id *id)
 113 {
 114         if (max_cstate > ACPI_PROCESSOR_MAX_POWER)
 115                 return 0;
 116
 117         printk(KERN_NOTICE PREFIX "%s detected - limiting to C%ld max_cstate."
 118                " Override with \"processor.max_cstate=%d\"\n", id->ident,
 119                (long)id->driver_data, ACPI_PROCESSOR_MAX_POWER + 1);
 120
 121         max_cstate = (long)id->driver_data;
 122
 123         return 0;
 124 }
 125
 126 /* Actually this shouldn't be __cpuinitdata, would be better to fix the
 127    callers to only run once -AK */
 128 static struct dmi_system_id __cpuinitdata processor_power_dmi_table[] = {
 129         { set_max_cstate, "IBM ThinkPad R40e", {
 130           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 131           DMI_MATCH(DMI_BIOS_VERSION,"1SET70WW")}, (void *)1},
 132         { set_max_cstate, "IBM ThinkPad R40e", {
 133           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 134           DMI_MATCH(DMI_BIOS_VERSION,"1SET60WW")}, (void *)1},
 135         { set_max_cstate, "IBM ThinkPad R40e", {
 136           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 137           DMI_MATCH(DMI_BIOS_VERSION,"1SET43WW") }, (void*)1},
 138         { set_max_cstate, "IBM ThinkPad R40e", {
 139           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 140           DMI_MATCH(DMI_BIOS_VERSION,"1SET45WW") }, (void*)1},
 141         { set_max_cstate, "IBM ThinkPad R40e", {
 142           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 143           DMI_MATCH(DMI_BIOS_VERSION,"1SET47WW") }, (void*)1},
 144         { set_max_cstate, "IBM ThinkPad R40e", {
 145           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 146           DMI_MATCH(DMI_BIOS_VERSION,"1SET50WW") }, (void*)1},
 147         { set_max_cstate, "IBM ThinkPad R40e", {
 148           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 149           DMI_MATCH(DMI_BIOS_VERSION,"1SET52WW") }, (void*)1},
 150         { set_max_cstate, "IBM ThinkPad R40e", {
 151           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 152           DMI_MATCH(DMI_BIOS_VERSION,"1SET55WW") }, (void*)1},
 153         { set_max_cstate, "IBM ThinkPad R40e", {
 154           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 155           DMI_MATCH(DMI_BIOS_VERSION,"1SET56WW") }, (void*)1},
 156         { set_max_cstate, "IBM ThinkPad R40e", {
 157           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 158           DMI_MATCH(DMI_BIOS_VERSION,"1SET59WW") }, (void*)1},
 159         { set_max_cstate, "IBM ThinkPad R40e", {
 160           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 161           DMI_MATCH(DMI_BIOS_VERSION,"1SET60WW") }, (void*)1},
 162         { set_max_cstate, "IBM ThinkPad R40e", {
 163           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 164           DMI_MATCH(DMI_BIOS_VERSION,"1SET61WW") }, (void*)1},
 165         { set_max_cstate, "IBM ThinkPad R40e", {
 166           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 167           DMI_MATCH(DMI_BIOS_VERSION,"1SET62WW") }, (void*)1},
 168         { set_max_cstate, "IBM ThinkPad R40e", {
 169           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 170           DMI_MATCH(DMI_BIOS_VERSION,"1SET64WW") }, (void*)1},
 171         { set_max_cstate, "IBM ThinkPad R40e", {
 172           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 173           DMI_MATCH(DMI_BIOS_VERSION,"1SET65WW") }, (void*)1},
 174         { set_max_cstate, "IBM ThinkPad R40e", {
 175           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 176           DMI_MATCH(DMI_BIOS_VERSION,"1SET68WW") }, (void*)1},
 177         { set_max_cstate, "Medion 41700", {
 178           DMI_MATCH(DMI_BIOS_VENDOR,"Phoenix Technologies LTD"),
 179           DMI_MATCH(DMI_BIOS_VERSION,"R01-A1J")}, (void *)1},
 180         { set_max_cstate, "Clevo 5600D", {
 181           DMI_MATCH(DMI_BIOS_VENDOR,"Phoenix Technologies LTD"),
 182           DMI_MATCH(DMI_BIOS_VERSION,"SHE845M0.86C.0013.D.0302131307")},
 183          (void *)2},
 184         {},
 185 };
 186
 187 static inline u32 ticks_elapsed(u32 t1, u32 t2)
 188 {
 189         if (t2 >= t1)
 190                 return (t2 - t1);
 191         else if (!(acpi_gbl_FADT.flags & ACPI_FADT_32BIT_TIMER))
 192                 return (((0x00FFFFFF - t1) + t2) & 0x00FFFFFF);
 193         else
 194                 return ((0xFFFFFFFF - t1) + t2);
 195 }
 196
 197 static inline u32 ticks_elapsed_in_us(u32 t1, u32 t2)
 198 {
 199         if (t2 >= t1)
 200                 return PM_TIMER_TICKS_TO_US(t2 - t1);
 201         else if (!(acpi_gbl_FADT.flags & ACPI_FADT_32BIT_TIMER))
 202                 return PM_TIMER_TICKS_TO_US(((0x00FFFFFF - t1) + t2) & 0x00FFFFFF);
 203         else
 204                 return PM_TIMER_TICKS_TO_US((0xFFFFFFFF - t1) + t2);
 205 }
 206
 207 /*
 208  * Callers should disable interrupts before the call and enable
 209  * interrupts after return.
 210  */
 211 static void acpi_safe_halt(void)
 212 {
 213         current_thread_info()->status &= ~TS_POLLING;
 214         /*
 215          * TS_POLLING-cleared state must be visible before we
 216          * test NEED_RESCHED:
 217          */
 218         smp_mb();
 219         if (!need_resched())
 220                 safe_halt();
 221         current_thread_info()->status |= TS_POLLING;
 222 }
 223
 224 #ifndef CONFIG_CPU_IDLE
 225
 226 static void
 227 acpi_processor_power_activate(struct acpi_processor *pr,
 228                               struct acpi_processor_cx *new)
 229 {
 230         struct acpi_processor_cx *old;
 231
 232         if (!pr || !new)
 233                 return;
 234
 235         old = pr->power.state;
 236
 237         if (old)
 238                 old->promotion.count = 0;
 239         new->demotion.count = 0;
 240
 241         /* Cleanup from old state. */
 242         if (old) {
 243                 switch (old->type) {
 244                 case ACPI_STATE_C3:
 245                         /* Disable bus master reload */
 246                         if (new->type != ACPI_STATE_C3 && pr->flags.bm_check)
 247                                 acpi_set_register(ACPI_BITREG_BUS_MASTER_RLD, 0);
 248                         break;
 249                 }
 250         }
 251
 252         /* Prepare to use new state. */
 253         switch (new->type) {
 254         case ACPI_STATE_C3:
 255                 /* Enable bus master reload */
 256                 if (old->type != ACPI_STATE_C3 && pr->flags.bm_check)
 257                         acpi_set_register(ACPI_BITREG_BUS_MASTER_RLD, 1);
 258                 break;
 259         }
 260
 261         pr->power.state = new;
 262
 263         return;
 264 }
 265
 266 static atomic_t c3_cpu_count;
 267
 268 /* Common C-state entry for C2, C3, .. */
 269 static void acpi_cstate_enter(struct acpi_processor_cx *cstate)
 270 {
 271         if (cstate->entry_method == ACPI_CSTATE_FFH) {
 272                 /* Call into architectural FFH based C-state */
 273                 acpi_processor_ffh_cstate_enter(cstate);
 274         } else {
 275                 int unused;
 276                 /* IO port based C-state */
 277                 inb(cstate->address);
 278                 /* Dummy wait op - must do something useless after P_LVL2 read
 279                    because chipsets cannot guarantee that STPCLK# signal
 280                    gets asserted in time to freeze execution properly. */
 281                 unused = inl(acpi_gbl_FADT.xpm_timer_block.address);
 282         }
 283 }
 284 #endif /* !CONFIG_CPU_IDLE */
 285
 286 #ifdef ARCH_APICTIMER_STOPS_ON_C3
 287
 288 /*
 289  * Some BIOS implementations switch to C3 in the published C2 state.
 290  * This seems to be a common problem on AMD boxen, but other vendors
 291  * are affected too. We pick the most conservative approach: we assume
 292  * that the local APIC stops in both C2 and C3.
 293  */
 294 static void acpi_timer_check_state(int state, struct acpi_processor *pr,
 295                                    struct acpi_processor_cx *cx)
 296 {
 297         struct acpi_processor_power *pwr = &pr->power;
 298         u8 type = local_apic_timer_c2_ok ? ACPI_STATE_C3 : ACPI_STATE_C2;
 299
 300         /*
 301          * Check, if one of the previous states already marked the lapic
 302          * unstable
 303          */
 304         if (pwr->timer_broadcast_on_state < state)
 305                 return;
 306
 307         if (cx->type >= type)
 308                 pr->power.timer_broadcast_on_state = state;
 309 }
 310
 311 static void acpi_propagate_timer_broadcast(struct acpi_processor *pr)
 312 {
 313         unsigned long reason;
 314
 315         reason = pr->power.timer_broadcast_on_state < INT_MAX ?
 316                 CLOCK_EVT_NOTIFY_BROADCAST_ON : CLOCK_EVT_NOTIFY_BROADCAST_OFF;
 317
 318         clockevents_notify(reason, &pr->id);
 319 }
 320
 321 /* Power(C) State timer broadcast control */
 322 static void acpi_state_timer_broadcast(struct acpi_processor *pr,
 323                                        struct acpi_processor_cx *cx,
 324                                        int broadcast)
 325 {
 326         int state = cx - pr->power.states;
 327
 328         if (state >= pr->power.timer_broadcast_on_state) {
 329                 unsigned long reason;
 330
 331                 reason = broadcast ?  CLOCK_EVT_NOTIFY_BROADCAST_ENTER :
 332                         CLOCK_EVT_NOTIFY_BROADCAST_EXIT;
 333                 clockevents_notify(reason, &pr->id);
 334         }
 335 }
 336
 337 #else
 338
 339 static void acpi_timer_check_state(int state, struct acpi_processor *pr,
 340                                    struct acpi_processor_cx *cstate) { }
 341 static void acpi_propagate_timer_broadcast(struct acpi_processor *pr) { }
 342 static void acpi_state_timer_broadcast(struct acpi_processor *pr,
 343                                        struct acpi_processor_cx *cx,
 344                                        int broadcast)
 345 {
 346 }
 347
 348 #endif
 349
 350 /*
 351  * Suspend / resume control
 352  */
 353 static int acpi_idle_suspend;
 354
 355 int acpi_processor_suspend(struct acpi_device * device, pm_message_t state)
 356 {
 357         acpi_idle_suspend = 1;
 358         return 0;
 359 }
 360
 361 int acpi_processor_resume(struct acpi_device * device)
 362 {
 363         acpi_idle_suspend = 0;
 364         return 0;
 365 }
 366
 367 #if defined (CONFIG_GENERIC_TIME) && defined (CONFIG_X86_TSC)
 368 static int tsc_halts_in_c(int state)
 369 {
 370         switch (boot_cpu_data.x86_vendor) {
 371         case X86_VENDOR_AMD:
 372                 /*
 373                  * AMD Fam10h TSC will tick in all
 374                  * C/P/S0/S1 states when this bit is set.
 375                  */
 376                 if (boot_cpu_has(X86_FEATURE_CONSTANT_TSC))
 377                         return 0;
 378                 /*FALL THROUGH*/
 379         case X86_VENDOR_INTEL:
 380                 /* Several cases known where TSC halts in C2 too */
 381         default:
 382                 return state > ACPI_STATE_C1;
 383         }
 384 }
 385 #endif
 386
 387 #ifndef CONFIG_CPU_IDLE
 388 static void acpi_processor_idle(void)
 389 {
 390         struct acpi_processor *pr = NULL;
 391         struct acpi_processor_cx *cx = NULL;
 392         struct acpi_processor_cx *next_state = NULL;
 393         int sleep_ticks = 0;
 394         u32 t1, t2 = 0;
 395
 396         /*
 397          * Interrupts must be disabled during bus mastering calculations and
 398          * for C2/C3 transitions.
 399          */
 400         local_irq_disable();
 401
 402         pr = processors[smp_processor_id()];
 403         if (!pr) {
 404                 local_irq_enable();
 405                 return;
 406         }
 407
 408         /*
 409          * Check whether we truly need to go idle, or should
 410          * reschedule:
 411          */
 412         if (unlikely(need_resched())) {
 413                 local_irq_enable();
 414                 return;
 415         }
 416
 417         cx = pr->power.state;
 418         if (!cx || acpi_idle_suspend) {
 419                 if (pm_idle_save)
 420                         pm_idle_save();
 421                 else
 422                         acpi_safe_halt();
 423
 424                 local_irq_enable();
 425                 return;
 426         }
 427
 428         /*
 429          * Check BM Activity
 430          * -----------------
 431          * Check for bus mastering activity (if required), record, and check
 432          * for demotion.
 433          */
 434         if (pr->flags.bm_check) {
 435                 u32 bm_status = 0;
 436                 unsigned long diff = jiffies - pr->power.bm_check_timestamp;
 437
 438                 if (diff > 31)
 439                         diff = 31;
 440
 441                 pr->power.bm_activity <<= diff;
 442
 443                 acpi_get_register(ACPI_BITREG_BUS_MASTER_STATUS, &bm_status);
 444                 if (bm_status) {
 445                         pr->power.bm_activity |= 0x1;
 446                         acpi_set_register(ACPI_BITREG_BUS_MASTER_STATUS, 1);
 447                 }
 448                 /*
 449                  * PIIX4 Erratum #18: Note that BM_STS doesn't always reflect
 450                  * the true state of bus mastering activity; forcing us to
 451                  * manually check the BMIDEA bit of each IDE channel.
 452                  */
 453                 else if (errata.piix4.bmisx) {
 454                         if ((inb_p(errata.piix4.bmisx + 0x02) & 0x01)
 455                             || (inb_p(errata.piix4.bmisx + 0x0A) & 0x01))
 456                                 pr->power.bm_activity |= 0x1;
 457                 }
 458
 459                 pr->power.bm_check_timestamp = jiffies;
 460
 461                 /*
 462                  * If bus mastering is or was active this jiffy, demote
 463                  * to avoid a faulty transition.  Note that the processor
 464                  * won't enter a low-power state during this call (to this
 465                  * function) but should upon the next.
 466                  *
 467                  * TBD: A better policy might be to fallback to the demotion
 468                  *      state (use it for this quantum only) istead of
 469                  *      demoting -- and rely on duration as our sole demotion
 470                  *      qualification.  This may, however, introduce DMA
 471                  *      issues (e.g. floppy DMA transfer overrun/underrun).
 472                  */
 473                 if ((pr->power.bm_activity & 0x1) &&
 474                     cx->demotion.threshold.bm) {
 475                         local_irq_enable();
 476                         next_state = cx->demotion.state;
 477                         goto end;
 478                 }
 479         }
 480
 481 #ifdef CONFIG_HOTPLUG_CPU
 482         /*
 483          * Check for P_LVL2_UP flag before entering C2 and above on
 484          * an SMP system. We do it here instead of doing it at _CST/P_LVL
 485          * detection phase, to work cleanly with logical CPU hotplug.
 486          */
 487         if ((cx->type != ACPI_STATE_C1) && (num_online_cpus() > 1) &&
 488             !pr->flags.has_cst && !(acpi_gbl_FADT.flags & ACPI_FADT_C2_MP_SUPPORTED))
 489                 cx = &pr->power.states[ACPI_STATE_C1];
 490 #endif
 491
 492         /*
 493          * Sleep:
 494          * ------
 495          * Invoke the current Cx state to put the processor to sleep.
 496          */
 497         if (cx->type == ACPI_STATE_C2 || cx->type == ACPI_STATE_C3) {
 498                 current_thread_info()->status &= ~TS_POLLING;
 499                 /*
 500                  * TS_POLLING-cleared state must be visible before we
 501                  * test NEED_RESCHED:
 502                  */
 503                 smp_mb();
 504                 if (need_resched()) {
 505                         current_thread_info()->status |= TS_POLLING;
 506                         local_irq_enable();
 507                         return;
 508                 }
 509         }
 510
 511         switch (cx->type) {
 512
 513         case ACPI_STATE_C1:
 514                 /*
 515                  * Invoke C1.
 516                  * Use the appropriate idle routine, the one that would
 517                  * be used without acpi C-states.
 518                  */
 519                 if (pm_idle_save)
 520                         pm_idle_save();
 521                 else
 522                         acpi_safe_halt();
 523
 524                 /*
 525                  * TBD: Can't get time duration while in C1, as resumes
 526                  *      go to an ISR rather than here.  Need to instrument
 527                  *      base interrupt handler.
 528                  *
 529                  * Note: the TSC better not stop in C1, sched_clock() will
 530                  *       skew otherwise.
 531                  */
 532                 sleep_ticks = 0xFFFFFFFF;
 533                 local_irq_enable();
 534                 break;
 535
 536         case ACPI_STATE_C2:
 537                 /* Get start time (ticks) */
 538                 t1 = inl(acpi_gbl_FADT.xpm_timer_block.address);
 539                 /* Tell the scheduler that we are going deep-idle: */
 540                 sched_clock_idle_sleep_event();
 541                 /* Invoke C2 */
 542                 acpi_state_timer_broadcast(pr, cx, 1);
 543                 acpi_cstate_enter(cx);
 544                 /* Get end time (ticks) */
 545                 t2 = inl(acpi_gbl_FADT.xpm_timer_block.address);
 546
 547 #if defined (CONFIG_GENERIC_TIME) && defined (CONFIG_X86_TSC)
 548                 /* TSC halts in C2, so notify users */
 549                 if (tsc_halts_in_c(ACPI_STATE_C2))
 550                         mark_tsc_unstable("possible TSC halt in C2");
 551 #endif
 552                 /* Compute time (ticks) that we were actually asleep */
 553                 sleep_ticks = ticks_elapsed(t1, t2);
 554
 555                 /* Tell the scheduler how much we idled: */
 556                 sched_clock_idle_wakeup_event(sleep_ticks*PM_TIMER_TICK_NS);
 557
 558                 /* Re-enable interrupts */
 559                 local_irq_enable();
 560                 /* Do not account our idle-switching overhead: */
 561                 sleep_ticks -= cx->latency_ticks + C2_OVERHEAD;
 562
 563                 current_thread_info()->status |= TS_POLLING;
 564                 acpi_state_timer_broadcast(pr, cx, 0);
 565                 break;
 566
 567         case ACPI_STATE_C3:
 568                 acpi_unlazy_tlb(smp_processor_id());
 569                 /*
 570                  * Must be done before busmaster disable as we might
 571                  * need to access HPET !
 572                  */
 573                 acpi_state_timer_broadcast(pr, cx, 1);
 574                 /*
 575                  * disable bus master
 576                  * bm_check implies we need ARB_DIS
 577                  * !bm_check implies we need cache flush
 578                  * bm_control implies whether we can do ARB_DIS
 579                  *
 580                  * That leaves a case where bm_check is set and bm_control is
 581                  * not set. In that case we cannot do much, we enter C3
 582                  * without doing anything.
 583                  */
 584                 if (pr->flags.bm_check && pr->flags.bm_control) {
 585                         if (atomic_inc_return(&c3_cpu_count) ==
 586                             num_online_cpus()) {
 587                                 /*
 588                                  * All CPUs are trying to go to C3
 589                                  * Disable bus master arbitration
 590                                  */
 591                                 acpi_set_register(ACPI_BITREG_ARB_DISABLE, 1);
 592                         }
 593                 } else if (!pr->flags.bm_check) {
 594                         /* SMP with no shared cache... Invalidate cache  */
 595                         ACPI_FLUSH_CPU_CACHE();
 596                 }
 597
 598                 /* Get start time (ticks) */
 599                 t1 = inl(acpi_gbl_FADT.xpm_timer_block.address);
 600                 /* Invoke C3 */
 601                 /* Tell the scheduler that we are going deep-idle: */
 602                 sched_clock_idle_sleep_event();
 603                 acpi_cstate_enter(cx);
 604                 /* Get end time (ticks) */
 605                 t2 = inl(acpi_gbl_FADT.xpm_timer_block.address);
 606                 if (pr->flags.bm_check && pr->flags.bm_control) {
 607                         /* Enable bus master arbitration */
 608                         atomic_dec(&c3_cpu_count);
 609                         acpi_set_register(ACPI_BITREG_ARB_DISABLE, 0);
 610                 }
 611
 612 #if defined (CONFIG_GENERIC_TIME) && defined (CONFIG_X86_TSC)
 613                 /* TSC halts in C3, so notify users */
 614                 if (tsc_halts_in_c(ACPI_STATE_C3))
 615                         mark_tsc_unstable("TSC halts in C3");
 616 #endif
 617                 /* Compute time (ticks) that we were actually asleep */
 618                 sleep_ticks = ticks_elapsed(t1, t2);
 619                 /* Tell the scheduler how much we idled: */
 620                 sched_clock_idle_wakeup_event(sleep_ticks*PM_TIMER_TICK_NS);
 621
 622                 /* Re-enable interrupts */
 623                 local_irq_enable();
 624                 /* Do not account our idle-switching overhead: */
 625                 sleep_ticks -= cx->latency_ticks + C3_OVERHEAD;
 626
 627                 current_thread_info()->status |= TS_POLLING;
 628                 acpi_state_timer_broadcast(pr, cx, 0);
 629                 break;
 630
 631         default:
 632                 local_irq_enable();
 633                 return;
 634         }
 635         cx->usage++;
 636         if ((cx->type != ACPI_STATE_C1) && (sleep_ticks > 0))
 637                 cx->time += sleep_ticks;
 638
 639         next_state = pr->power.state;
 640
 641 #ifdef CONFIG_HOTPLUG_CPU
 642         /* Don't do promotion/demotion */
 643         if ((cx->type == ACPI_STATE_C1) && (num_online_cpus() > 1) &&
 644             !pr->flags.has_cst && !(acpi_gbl_FADT.flags & ACPI_FADT_C2_MP_SUPPORTED)) {
 645                 next_state = cx;
 646                 goto end;
 647         }
 648 #endif
 649
 650         /*
 651          * Promotion?
 652          * ----------
 653          * Track the number of longs (time asleep is greater than threshold)
 654          * and promote when the count threshold is reached.  Note that bus
 655          * mastering activity may prevent promotions.
 656          * Do not promote above max_cstate.
 657          */
 658         if (cx->promotion.state &&
 659             ((cx->promotion.state - pr->power.states) <= max_cstate)) {
 660                 if (sleep_ticks > cx->promotion.threshold.ticks &&
 661                   cx->promotion.state->latency <=
 662                                 pm_qos_requirement(PM_QOS_CPU_DMA_LATENCY)) {
 663                         cx->promotion.count++;
 664                         cx->demotion.count = 0;
 665                         if (cx->promotion.count >=
 666                             cx->promotion.threshold.count) {
 667                                 if (pr->flags.bm_check) {
 668                                         if (!
 669                                             (pr->power.bm_activity & cx->
 670                                              promotion.threshold.bm)) {
 671                                                 next_state =
 672                                                     cx->promotion.state;
 673                                                 goto end;
 674                                         }
 675                                 } else {
 676                                         next_state = cx->promotion.state;
 677                                         goto end;
 678                                 }
 679                         }
 680                 }
 681         }
 682
 683         /*
 684          * Demotion?
 685          * ---------
 686          * Track the number of shorts (time asleep is less than time threshold)
 687          * and demote when the usage threshold is reached.
 688          */
 689         if (cx->demotion.state) {
 690                 if (sleep_ticks < cx->demotion.threshold.ticks) {
 691                         cx->demotion.count++;
 692                         cx->promotion.count = 0;
 693                         if (cx->demotion.count >= cx->demotion.threshold.count) {
 694                                 next_state = cx->demotion.state;
 695                                 goto end;
 696                         }
 697                 }
 698         }
 699
 700       end:
 701         /*
 702          * Demote if current state exceeds max_cstate
 703          * or if the latency of the current state is unacceptable
 704          */
 705         if ((pr->power.state - pr->power.states) > max_cstate ||
 706                 pr->power.state->latency >
 707                                 pm_qos_requirement(PM_QOS_CPU_DMA_LATENCY)) {
 708                 if (cx->demotion.state)
 709                         next_state = cx->demotion.state;
 710         }
 711
 712         /*
 713          * New Cx State?
 714          * -------------
 715          * If we're going to start using a new Cx state we must clean up
 716          * from the previous and prepare to use the new.
 717          */
 718         if (next_state != pr->power.state)
 719                 acpi_processor_power_activate(pr, next_state);
 720 }
 721
 722 static int acpi_processor_set_power_policy(struct acpi_processor *pr)
 723 {
 724         unsigned int i;
 725         unsigned int state_is_set = 0;
 726         struct acpi_processor_cx *lower = NULL;
 727         struct acpi_processor_cx *higher = NULL;
 728         struct acpi_processor_cx *cx;
 729
 730
 731         if (!pr)
 732                 return -EINVAL;
 733
 734         /*
 735          * This function sets the default Cx state policy (OS idle handler).
 736          * Our scheme is to promote quickly to C2 but more conservatively
 737          * to C3.  We're favoring C2  for its characteristics of low latency
 738          * (quick response), good power savings, and ability to allow bus
 739          * mastering activity.  Note that the Cx state policy is completely
 740          * customizable and can be altered dynamically.
 741          */
 742
 743         /* startup state */
 744         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
 745                 cx = &pr->power.states[i];
 746                 if (!cx->valid)
 747                         continue;
 748
 749                 if (!state_is_set)
 750                         pr->power.state = cx;
 751                 state_is_set++;
 752                 break;
 753         }
 754
 755         if (!state_is_set)
 756                 return -ENODEV;
 757
 758         /* demotion */
 759         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
 760                 cx = &pr->power.states[i];
 761                 if (!cx->valid)
 762                         continue;
 763
 764                 if (lower) {
 765                         cx->demotion.state = lower;
 766                         cx->demotion.threshold.ticks = cx->latency_ticks;
 767                         cx->demotion.threshold.count = 1;
 768                         if (cx->type == ACPI_STATE_C3)
 769                                 cx->demotion.threshold.bm = bm_history;
 770                 }
 771
 772                 lower = cx;
 773         }
 774
 775         /* promotion */
 776         for (i = (ACPI_PROCESSOR_MAX_POWER - 1); i > 0; i--) {
 777                 cx = &pr->power.states[i];
 778                 if (!cx->valid)
 779                         continue;
 780
 781                 if (higher) {
 782                         cx->promotion.state = higher;
 783                         cx->promotion.threshold.ticks = cx->latency_ticks;
 784                         if (cx->type >= ACPI_STATE_C2)
 785                                 cx->promotion.threshold.count = 4;
 786                         else
 787                                 cx->promotion.threshold.count = 10;
 788                         if (higher->type == ACPI_STATE_C3)
 789                                 cx->promotion.threshold.bm = bm_history;
 790                 }
 791
 792                 higher = cx;
 793         }
 794
 795         return 0;
 796 }
 797 #endif /* !CONFIG_CPU_IDLE */
 798
 799 static int acpi_processor_get_power_info_fadt(struct acpi_processor *pr)
 800 {
 801
 802         if (!pr)
 803                 return -EINVAL;
 804
 805         if (!pr->pblk)
 806                 return -ENODEV;
 807
 808         /* if info is obtained from pblk/fadt, type equals state */
 809         pr->power.states[ACPI_STATE_C2].type = ACPI_STATE_C2;
 810         pr->power.states[ACPI_STATE_C3].type = ACPI_STATE_C3;
 811
 812 #ifndef CONFIG_HOTPLUG_CPU
 813         /*
 814          * Check for P_LVL2_UP flag before entering C2 and above on
 815          * an SMP system.
 816          */
 817         if ((num_online_cpus() > 1) &&
 818             !(acpi_gbl_FADT.flags & ACPI_FADT_C2_MP_SUPPORTED))
 819                 return -ENODEV;
 820 #endif
 821
 822         /* determine C2 and C3 address from pblk */
 823         pr->power.states[ACPI_STATE_C2].address = pr->pblk + 4;
 824         pr->power.states[ACPI_STATE_C3].address = pr->pblk + 5;
 825
 826         /* determine latencies from FADT */
 827         pr->power.states[ACPI_STATE_C2].latency = acpi_gbl_FADT.C2latency;
 828         pr->power.states[ACPI_STATE_C3].latency = acpi_gbl_FADT.C3latency;
 829
 830         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 831                           "lvl2[0x%08x] lvl3[0x%08x]\n",
 832                           pr->power.states[ACPI_STATE_C2].address,
 833                           pr->power.states[ACPI_STATE_C3].address));
 834
 835         return 0;
 836 }
 837
 838 static int acpi_processor_get_power_info_default(struct acpi_processor *pr)
 839 {
 840         if (!pr->power.states[ACPI_STATE_C1].valid) {
 841                 /* set the first C-State to C1 */
 842                 /* all processors need to support C1 */
 843                 pr->power.states[ACPI_STATE_C1].type = ACPI_STATE_C1;
 844                 pr->power.states[ACPI_STATE_C1].valid = 1;
 845         }
 846         /* the C0 state only exists as a filler in our array */
 847         pr->power.states[ACPI_STATE_C0].valid = 1;
 848         return 0;
 849 }
 850
 851 static int acpi_processor_get_power_info_cst(struct acpi_processor *pr)
 852 {
 853         acpi_status status = 0;
 854         acpi_integer count;
 855         int current_count;
 856         int i;
 857         struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
 858         union acpi_object *cst;
 859
 860
 861         if (nocst)
 862                 return -ENODEV;
 863
 864         current_count = 0;
 865
 866         status = acpi_evaluate_object(pr->handle, "_CST", NULL, &buffer);
 867         if (ACPI_FAILURE(status)) {
 868                 ACPI_DEBUG_PRINT((ACPI_DB_INFO, "No _CST, giving up\n"));
 869                 return -ENODEV;
 870         }
 871
 872         cst = buffer.pointer;
 873
 874         /* There must be at least 2 elements */
 875         if (!cst || (cst->type != ACPI_TYPE_PACKAGE) || cst->package.count < 2) {
 876                 printk(KERN_ERR PREFIX "not enough elements in _CST\n");
 877                 status = -EFAULT;
 878                 goto end;
 879         }
 880
 881         count = cst->package.elements[0].integer.value;
 882
 883         /* Validate number of power states. */
 884         if (count < 1 || count != cst->package.count - 1) {
 885                 printk(KERN_ERR PREFIX "count given by _CST is not valid\n");
 886                 status = -EFAULT;
 887                 goto end;
 888         }
 889
 890         /* Tell driver that at least _CST is supported. */
 891         pr->flags.has_cst = 1;
 892
 893         for (i = 1; i <= count; i++) {
 894                 union acpi_object *element;
 895                 union acpi_object *obj;
 896                 struct acpi_power_register *reg;
 897                 struct acpi_processor_cx cx;
 898
 899                 memset(&cx, 0, sizeof(cx));
 900
 901                 element = &(cst->package.elements[i]);
 902                 if (element->type != ACPI_TYPE_PACKAGE)
 903                         continue;
 904
 905                 if (element->package.count != 4)
 906                         continue;
 907
 908                 obj = &(element->package.elements[0]);
 909
 910                 if (obj->type != ACPI_TYPE_BUFFER)
 911                         continue;
 912
 913                 reg = (struct acpi_power_register *)obj->buffer.pointer;
 914
 915                 if (reg->space_id != ACPI_ADR_SPACE_SYSTEM_IO &&
 916                     (reg->space_id != ACPI_ADR_SPACE_FIXED_HARDWARE))
 917                         continue;
 918
 919                 /* There should be an easy way to extract an integer... */
 920                 obj = &(element->package.elements[1]);
 921                 if (obj->type != ACPI_TYPE_INTEGER)
 922                         continue;
 923
 924                 cx.type = obj->integer.value;
 925                 /*
 926                  * Some buggy BIOSes won't list C1 in _CST -
 927                  * Let acpi_processor_get_power_info_default() handle them later
 928                  */
 929                 if (i == 1 && cx.type != ACPI_STATE_C1)
 930                         current_count++;
 931
 932                 cx.address = reg->address;
 933                 cx.index = current_count + 1;
 934
 935                 cx.entry_method = ACPI_CSTATE_SYSTEMIO;
 936                 if (reg->space_id == ACPI_ADR_SPACE_FIXED_HARDWARE) {
 937                         if (acpi_processor_ffh_cstate_probe
 938                                         (pr->id, &cx, reg) == 0) {
 939                                 cx.entry_method = ACPI_CSTATE_FFH;
 940                         } else if (cx.type == ACPI_STATE_C1) {
 941                                 /*
 942                                  * C1 is a special case where FIXED_HARDWARE
 943                                  * can be handled in non-MWAIT way as well.
 944                                  * In that case, save this _CST entry info.
 945                                  * Otherwise, ignore this info and continue.
 946                                  */
 947                                 cx.entry_method = ACPI_CSTATE_HALT;
 948                         } else {
 949                                 continue;
 950                         }
 951                 }
 952
 953                 obj = &(element->package.elements[2]);
 954                 if (obj->type != ACPI_TYPE_INTEGER)
 955                         continue;
 956
 957                 cx.latency = obj->integer.value;
 958
 959                 obj = &(element->package.elements[3]);
 960                 if (obj->type != ACPI_TYPE_INTEGER)
 961                         continue;
 962
 963                 cx.power = obj->integer.value;
 964
 965                 current_count++;
 966                 memcpy(&(pr->power.states[current_count]), &cx, sizeof(cx));
 967
 968                 /*
 969                  * We support total ACPI_PROCESSOR_MAX_POWER - 1
 970                  * (From 1 through ACPI_PROCESSOR_MAX_POWER - 1)
 971                  */
 972                 if (current_count >= (ACPI_PROCESSOR_MAX_POWER - 1)) {
 973                         printk(KERN_WARNING
 974                                "Limiting number of power states to max (%d)\n",
 975                                ACPI_PROCESSOR_MAX_POWER);
 976                         printk(KERN_WARNING
 977                                "Please increase ACPI_PROCESSOR_MAX_POWER if needed.\n");
 978                         break;
 979                 }
 980         }
 981
 982         ACPI_DEBUG_PRINT((ACPI_DB_INFO, "Found %d power states\n",
 983                           current_count));
 984
 985         /* Validate number of power states discovered */
 986         if (current_count < 2)
 987                 status = -EFAULT;
 988
 989       end:
 990         kfree(buffer.pointer);
 991
 992         return status;
 993 }
 994
 995 static void acpi_processor_power_verify_c2(struct acpi_processor_cx *cx)
 996 {
 997
 998         if (!cx->address)
 999                 return;
1000
1001         /*
1002          * C2 latency must be less than or equal to 100
1003          * microseconds.
1004          */
1005         else if (cx->latency > ACPI_PROCESSOR_MAX_C2_LATENCY) {
1006                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
1007                                   "latency too large [%d]\n", cx->latency));
1008                 return;
1009         }
1010
1011         /*
1012          * Otherwise we've met all of our C2 requirements.
1013          * Normalize the C2 latency to expidite policy
1014          */
1015         cx->valid = 1;
1016
1017 #ifndef CONFIG_CPU_IDLE
1018         cx->latency_ticks = US_TO_PM_TIMER_TICKS(cx->latency);
1019 #else
1020         cx->latency_ticks = cx->latency;
1021 #endif
1022
1023         return;
1024 }
1025
1026 static void acpi_processor_power_verify_c3(struct acpi_processor *pr,
1027                                            struct acpi_processor_cx *cx)
1028 {
1029         static int bm_check_flag;
1030
1031
1032         if (!cx->address)
1033                 return;
1034
1035         /*
1036          * C3 latency must be less than or equal to 1000
1037          * microseconds.
1038          */
1039         else if (cx->latency > ACPI_PROCESSOR_MAX_C3_LATENCY) {
1040                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
1041                                   "latency too large [%d]\n", cx->latency));
1042                 return;
1043         }
1044
1045         /*
1046          * PIIX4 Erratum #18: We don't support C3 when Type-F (fast)
1047          * DMA transfers are used by any ISA device to avoid livelock.
1048          * Note that we could disable Type-F DMA (as recommended by
1049          * the erratum), but this is known to disrupt certain ISA
1050          * devices thus we take the conservative approach.
1051          */
1052         else if (errata.piix4.fdma) {
1053                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
1054                                   "C3 not supported on PIIX4 with Type-F DMA\n"));
1055                 return;
1056         }
1057
1058         /* All the logic here assumes flags.bm_check is same across all CPUs */
1059         if (!bm_check_flag) {
1060                 /* Determine whether bm_check is needed based on CPU  */
1061                 acpi_processor_power_init_bm_check(&(pr->flags), pr->id);
1062                 bm_check_flag = pr->flags.bm_check;
1063         } else {
1064                 pr->flags.bm_check = bm_check_flag;
1065         }
1066
1067         if (pr->flags.bm_check) {
1068                 if (!pr->flags.bm_control) {
1069                         if (pr->flags.has_cst != 1) {
1070                                 /* bus mastering control is necessary */
1071                                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
1072                                         "C3 support requires BM control\n"));
1073                                 return;
1074                         } else {
1075                                 /* Here we enter C3 without bus mastering */
1076                                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
1077                                         "C3 support without BM control\n"));
1078                         }
1079                 }
1080         } else {
1081                 /*
1082                  * WBINVD should be set in fadt, for C3 state to be
1083                  * supported on when bm_check is not required.
1084                  */
1085                 if (!(acpi_gbl_FADT.flags & ACPI_FADT_WBINVD)) {
1086                         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
1087                                           "Cache invalidation should work properly"
1088                                           " for C3 to be enabled on SMP systems\n"));
1089                         return;
1090                 }
1091                 acpi_set_register(ACPI_BITREG_BUS_MASTER_RLD, 0);
1092         }
1093
1094         /*
1095          * Otherwise we've met all of our C3 requirements.
1096          * Normalize the C3 latency to expidite policy.  Enable
1097          * checking of bus mastering status (bm_check) so we can
1098          * use this in our C3 policy
1099          */
1100         cx->valid = 1;
1101
1102 #ifndef CONFIG_CPU_IDLE
1103         cx->latency_ticks = US_TO_PM_TIMER_TICKS(cx->latency);
1104 #else
1105         cx->latency_ticks = cx->latency;
1106 #endif
1107
1108         return;
1109 }
1110
1111 static int acpi_processor_power_verify(struct acpi_processor *pr)
1112 {
1113         unsigned int i;
1114         unsigned int working = 0;
1115
1116         pr->power.timer_broadcast_on_state = INT_MAX;
1117
1118         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
1119                 struct acpi_processor_cx *cx = &pr->power.states[i];
1120
1121                 switch (cx->type) {
1122                 case ACPI_STATE_C1:
1123                         cx->valid = 1;
1124                         break;
1125
1126                 case ACPI_STATE_C2:
1127                         acpi_processor_power_verify_c2(cx);
1128                         if (cx->valid)
1129                                 acpi_timer_check_state(i, pr, cx);
1130                         break;
1131
1132                 case ACPI_STATE_C3:
1133                         acpi_processor_power_verify_c3(pr, cx);
1134                         if (cx->valid)
1135                                 acpi_timer_check_state(i, pr, cx);
1136                         break;
1137                 }
1138
1139                 if (cx->valid)
1140                         working++;
1141         }
1142
1143         acpi_propagate_timer_broadcast(pr);
1144
1145         return (working);
1146 }
1147
1148 static int acpi_processor_get_power_info(struct acpi_processor *pr)
1149 {
1150         unsigned int i;
1151         int result;
1152
1153
1154         /* NOTE: the idle thread may not be running while calling
1155          * this function */
1156
1157         /* Zero initialize all the C-states info. */
1158         memset(pr->power.states, 0, sizeof(pr->power.states));
1159
1160         result = acpi_processor_get_power_info_cst(pr);
1161         if (result == -ENODEV)
1162                 result = acpi_processor_get_power_info_fadt(pr);
1163
1164         if (result)
1165                 return result;
1166
1167         acpi_processor_get_power_info_default(pr);
1168
1169         pr->power.count = acpi_processor_power_verify(pr);
1170
1171 #ifndef CONFIG_CPU_IDLE
1172         /*
1173          * Set Default Policy
1174          * ------------------
1175          * Now that we know which states are supported, set the default
1176          * policy.  Note that this policy can be changed dynamically
1177          * (e.g. encourage deeper sleeps to conserve battery life when
1178          * not on AC).
1179          */
1180         result = acpi_processor_set_power_policy(pr);
1181         if (result)
1182                 return result;
1183 #endif
1184
1185         /*
1186          * if one state of type C2 or C3 is available, mark this
1187          * CPU as being "idle manageable"
1188          */
1189         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
1190                 if (pr->power.states[i].valid) {
1191                         pr->power.count = i;
1192                         if (pr->power.states[i].type >= ACPI_STATE_C2)
1193                                 pr->flags.power = 1;
1194                 }
1195         }
1196
1197         return 0;
1198 }
1199
1200 static int acpi_processor_power_seq_show(struct seq_file *seq, void *offset)
1201 {
1202         struct acpi_processor *pr = seq->private;
1203         unsigned int i;
1204
1205
1206         if (!pr)
1207                 goto end;
1208
1209         seq_printf(seq, "active state:            C%zd\n"
1210                    "max_cstate:              C%d\n"
1211                    "bus master activity:     %08x\n"
1212                    "maximum allowed latency: %d usec\n",
1213                    pr->power.state ? pr->power.state - pr->power.states : 0,
1214                    max_cstate, (unsigned)pr->power.bm_activity,
1215                    pm_qos_requirement(PM_QOS_CPU_DMA_LATENCY));
1216
1217         seq_puts(seq, "states:\n");
1218
1219         for (i = 1; i <= pr->power.count; i++) {
1220                 seq_printf(seq, "   %cC%d:                  ",
1221                            (&pr->power.states[i] ==
1222                             pr->power.state ? '*' : ' '), i);
1223
1224                 if (!pr->power.states[i].valid) {
1225                         seq_puts(seq, "<not supported>\n");
1226                         continue;
1227                 }
1228
1229                 switch (pr->power.states[i].type) {
1230                 case ACPI_STATE_C1:
1231                         seq_printf(seq, "type[C1] ");
1232                         break;
1233                 case ACPI_STATE_C2:
1234                         seq_printf(seq, "type[C2] ");
1235                         break;
1236                 case ACPI_STATE_C3:
1237                         seq_printf(seq, "type[C3] ");
1238                         break;
1239                 default:
1240                         seq_printf(seq, "type[--] ");
1241                         break;
1242                 }
1243
1244                 if (pr->power.states[i].promotion.state)
1245                         seq_printf(seq, "promotion[C%zd] ",
1246                                    (pr->power.states[i].promotion.state -
1247                                     pr->power.states));
1248                 else
1249                         seq_puts(seq, "promotion[--] ");
1250
1251                 if (pr->power.states[i].demotion.state)
1252                         seq_printf(seq, "demotion[C%zd] ",
1253                                    (pr->power.states[i].demotion.state -
1254                                     pr->power.states));
1255                 else
1256                         seq_puts(seq, "demotion[--] ");
1257
1258                 seq_printf(seq, "latency[%03d] usage[%08d] duration[%020llu]\n",
1259                            pr->power.states[i].latency,
1260                            pr->power.states[i].usage,
1261                            (unsigned long long)pr->power.states[i].time);
1262         }
1263
1264       end:
1265         return 0;
1266 }
1267
1268 static int acpi_processor_power_open_fs(struct inode *inode, struct file *file)
1269 {
1270         return single_open(file, acpi_processor_power_seq_show,
1271                            PDE(inode)->data);
1272 }
1273
1274 static const struct file_operations acpi_processor_power_fops = {
1275         .open = acpi_processor_power_open_fs,
1276         .read = seq_read,
1277         .llseek = seq_lseek,
1278         .release = single_release,
1279 };
1280
1281 #ifndef CONFIG_CPU_IDLE
1282
1283 int acpi_processor_cst_has_changed(struct acpi_processor *pr)
1284 {
1285         int result = 0;
1286
1287
1288         if (!pr)
1289                 return -EINVAL;
1290
1291         if (nocst) {
1292                 return -ENODEV;
1293         }
1294
1295         if (!pr->flags.power_setup_done)
1296                 return -ENODEV;
1297
1298         /* Fall back to the default idle loop */
1299         pm_idle = pm_idle_save;
1300         synchronize_sched();    /* Relies on interrupts forcing exit from idle. */
1301
1302         pr->flags.power = 0;
1303         result = acpi_processor_get_power_info(pr);
1304         if ((pr->flags.power == 1) && (pr->flags.power_setup_done))
1305                 pm_idle = acpi_processor_idle;
1306
1307         return result;
1308 }
1309
1310 #ifdef CONFIG_SMP
1311 static void smp_callback(void *v)
1312 {
1313         /* we already woke the CPU up, nothing more to do */
1314 }
1315
1316 /*
1317  * This function gets called when a part of the kernel has a new latency
1318  * requirement.  This means we need to get all processors out of their C-state,
1319  * and then recalculate a new suitable C-state. Just do a cross-cpu IPI; that
1320  * wakes them all right up.
1321  */
1322 static int acpi_processor_latency_notify(struct notifier_block *b,
1323                 unsigned long l, void *v)
1324 {
1325         smp_call_function(smp_callback, NULL, 0, 1);
1326         return NOTIFY_OK;
1327 }
1328
1329 static struct notifier_block acpi_processor_latency_notifier = {
1330         .notifier_call = acpi_processor_latency_notify,
1331 };
1332
1333 #endif
1334
1335 #else /* CONFIG_CPU_IDLE */
1336
1337 /**
1338  * acpi_idle_bm_check - checks if bus master activity was detected
1339  */
1340 static int acpi_idle_bm_check(void)
1341 {
1342         u32 bm_status = 0;
1343
1344         acpi_get_register(ACPI_BITREG_BUS_MASTER_STATUS, &bm_status);
1345         if (bm_status)
1346                 acpi_set_register(ACPI_BITREG_BUS_MASTER_STATUS, 1);
1347         /*
1348          * PIIX4 Erratum #18: Note that BM_STS doesn't always reflect
1349          * the true state of bus mastering activity; forcing us to
1350          * manually check the BMIDEA bit of each IDE channel.
1351          */
1352         else if (errata.piix4.bmisx) {
1353                 if ((inb_p(errata.piix4.bmisx + 0x02) & 0x01)
1354                     || (inb_p(errata.piix4.bmisx + 0x0A) & 0x01))
1355                         bm_status = 1;
1356         }
1357         return bm_status;
1358 }
1359
1360 /**
1361  * acpi_idle_update_bm_rld - updates the BM_RLD bit depending on target state
1362  * @pr: the processor
1363  * @target: the new target state
1364  */
1365 static inline void acpi_idle_update_bm_rld(struct acpi_processor *pr,
1366                                            struct acpi_processor_cx *target)
1367 {
1368         if (pr->flags.bm_rld_set && target->type != ACPI_STATE_C3) {
1369                 acpi_set_register(ACPI_BITREG_BUS_MASTER_RLD, 0);
1370                 pr->flags.bm_rld_set = 0;
1371         }
1372
1373         if (!pr->flags.bm_rld_set && target->type == ACPI_STATE_C3) {
1374                 acpi_set_register(ACPI_BITREG_BUS_MASTER_RLD, 1);
1375                 pr->flags.bm_rld_set = 1;
1376         }
1377 }
1378
1379 /**
1380  * acpi_idle_do_entry - a helper function that does C2 and C3 type entry
1381  * @cx: cstate data
1382  *
1383  * Caller disables interrupt before call and enables interrupt after return.
1384  */
1385 static inline void acpi_idle_do_entry(struct acpi_processor_cx *cx)
1386 {
1387         if (cx->entry_method == ACPI_CSTATE_FFH) {
1388                 /* Call into architectural FFH based C-state */
1389                 acpi_processor_ffh_cstate_enter(cx);
1390         } else if (cx->entry_method == ACPI_CSTATE_HALT) {
1391                 acpi_safe_halt();
1392         } else {
1393                 int unused;
1394                 /* IO port based C-state */
1395                 inb(cx->address);
1396                 /* Dummy wait op - must do something useless after P_LVL2 read
1397                    because chipsets cannot guarantee that STPCLK# signal
1398                    gets asserted in time to freeze execution properly. */
1399                 unused = inl(acpi_gbl_FADT.xpm_timer_block.address);
1400         }
1401 }
1402
1403 /**
1404  * acpi_idle_enter_c1 - enters an ACPI C1 state-type
1405  * @dev: the target CPU
1406  * @state: the state data
1407  *
1408  * This is equivalent to the HALT instruction.
1409  */
1410 static int acpi_idle_enter_c1(struct cpuidle_device *dev,
1411                               struct cpuidle_state *state)
1412 {
1413         u32 t1, t2;
1414         struct acpi_processor *pr;
1415         struct acpi_processor_cx *cx = cpuidle_get_statedata(state);
1416
1417         pr = processors[smp_processor_id()];
1418
1419         if (unlikely(!pr))
1420                 return 0;
1421
1422         local_irq_disable();
1423         if (pr->flags.bm_check)
1424                 acpi_idle_update_bm_rld(pr, cx);
1425
1426         t1 = inl(acpi_gbl_FADT.xpm_timer_block.address);
1427         acpi_idle_do_entry(cx);
1428         t2 = inl(acpi_gbl_FADT.xpm_timer_block.address);
1429
1430         local_irq_enable();
1431         cx->usage++;
1432
1433         return ticks_elapsed_in_us(t1, t2);
1434 }
1435
1436 /**
1437  * acpi_idle_enter_simple - enters an ACPI state without BM handling
1438  * @dev: the target CPU
1439  * @state: the state data
1440  */
1441 static int acpi_idle_enter_simple(struct cpuidle_device *dev,
1442                                   struct cpuidle_state *state)
1443 {
1444         struct acpi_processor *pr;
1445         struct acpi_processor_cx *cx = cpuidle_get_statedata(state);
1446         u32 t1, t2;
1447         int sleep_ticks = 0;
1448
1449         pr = processors[smp_processor_id()];
1450
1451         if (unlikely(!pr))
1452                 return 0;
1453
1454         if (acpi_idle_suspend)
1455                 return(acpi_idle_enter_c1(dev, state));
1456
1457         local_irq_disable();
1458         current_thread_info()->status &= ~TS_POLLING;
1459         /*
1460          * TS_POLLING-cleared state must be visible before we test
1461          * NEED_RESCHED:
1462          */
1463         smp_mb();
1464
1465         if (unlikely(need_resched())) {
1466                 current_thread_info()->status |= TS_POLLING;
1467                 local_irq_enable();
1468                 return 0;
1469         }
1470
1471         acpi_unlazy_tlb(smp_processor_id());
1472         /*
1473          * Must be done before busmaster disable as we might need to
1474          * access HPET !
1475          */
1476         acpi_state_timer_broadcast(pr, cx, 1);
1477
1478         if (pr->flags.bm_check)
1479                 acpi_idle_update_bm_rld(pr, cx);
1480
1481         if (cx->type == ACPI_STATE_C3)
1482                 ACPI_FLUSH_CPU_CACHE();
1483
1484         t1 = inl(acpi_gbl_FADT.xpm_timer_block.address);
1485         /* Tell the scheduler that we are going deep-idle: */
1486         sched_clock_idle_sleep_event();
1487         acpi_idle_do_entry(cx);
1488         t2 = inl(acpi_gbl_FADT.xpm_timer_block.address);
1489
1490 #if defined (CONFIG_GENERIC_TIME) && defined (CONFIG_X86_TSC)
1491         /* TSC could halt in idle, so notify users */
1492         if (tsc_halts_in_c(cx->type))
1493                 mark_tsc_unstable("TSC halts in idle");;
1494 #endif
1495         sleep_ticks = ticks_elapsed(t1, t2);
1496
1497         /* Tell the scheduler how much we idled: */
1498         sched_clock_idle_wakeup_event(sleep_ticks*PM_TIMER_TICK_NS);
1499
1500         local_irq_enable();
1501         current_thread_info()->status |= TS_POLLING;
1502
1503         cx->usage++;
1504
1505         acpi_state_timer_broadcast(pr, cx, 0);
1506         cx->time += sleep_ticks;
1507         return ticks_elapsed_in_us(t1, t2);
1508 }
1509
1510 static int c3_cpu_count;
1511 static DEFINE_SPINLOCK(c3_lock);
1512
1513 /**
1514  * acpi_idle_enter_bm - enters C3 with proper BM handling
1515  * @dev: the target CPU
1516  * @state: the state data
1517  *
1518  * If BM is detected, the deepest non-C3 idle state is entered instead.
1519  */
1520 static int acpi_idle_enter_bm(struct cpuidle_device *dev,
1521                               struct cpuidle_state *state)
1522 {
1523         struct acpi_processor *pr;
1524         struct acpi_processor_cx *cx = cpuidle_get_statedata(state);
1525         u32 t1, t2;
1526         int sleep_ticks = 0;
1527
1528         pr = processors[smp_processor_id()];
1529
1530         if (unlikely(!pr))
1531                 return 0;
1532
1533         if (acpi_idle_suspend)
1534                 return(acpi_idle_enter_c1(dev, state));
1535
1536         if (acpi_idle_bm_check()) {
1537                 if (dev->safe_state) {
1538                         return dev->safe_state->enter(dev, dev->safe_state);
1539                 } else {
1540                         local_irq_disable();
1541                         acpi_safe_halt();
1542                         local_irq_enable();
1543                         return 0;
1544                 }
1545         }
1546
1547         local_irq_disable();
1548         current_thread_info()->status &= ~TS_POLLING;
1549         /*
1550          * TS_POLLING-cleared state must be visible before we test
1551          * NEED_RESCHED:
1552          */
1553         smp_mb();
1554
1555         if (unlikely(need_resched())) {
1556                 current_thread_info()->status |= TS_POLLING;
1557                 local_irq_enable();
1558                 return 0;
1559         }
1560
1561         /* Tell the scheduler that we are going deep-idle: */
1562         sched_clock_idle_sleep_event();
1563         /*
1564          * Must be done before busmaster disable as we might need to
1565          * access HPET !
1566          */
1567         acpi_state_timer_broadcast(pr, cx, 1);
1568
1569         acpi_idle_update_bm_rld(pr, cx);
1570
1571         /*
1572          * disable bus master
1573          * bm_check implies we need ARB_DIS
1574          * !bm_check implies we need cache flush
1575          * bm_control implies whether we can do ARB_DIS
1576          *
1577          * That leaves a case where bm_check is set and bm_control is
1578          * not set. In that case we cannot do much, we enter C3
1579          * without doing anything.
1580          */
1581         if (pr->flags.bm_check && pr->flags.bm_control) {
1582                 spin_lock(&c3_lock);
1583                 c3_cpu_count++;
1584                 /* Disable bus master arbitration when all CPUs are in C3 */
1585                 if (c3_cpu_count == num_online_cpus())
1586                         acpi_set_register(ACPI_BITREG_ARB_DISABLE, 1);
1587                 spin_unlock(&c3_lock);
1588         } else if (!pr->flags.bm_check) {
1589                 ACPI_FLUSH_CPU_CACHE();
1590         }
1591
1592         t1 = inl(acpi_gbl_FADT.xpm_timer_block.address);
1593         acpi_idle_do_entry(cx);
1594         t2 = inl(acpi_gbl_FADT.xpm_timer_block.address);
1595
1596         /* Re-enable bus master arbitration */
1597         if (pr->flags.bm_check && pr->flags.bm_control) {
1598                 spin_lock(&c3_lock);
1599                 acpi_set_register(ACPI_BITREG_ARB_DISABLE, 0);
1600                 c3_cpu_count--;
1601                 spin_unlock(&c3_lock);
1602         }
1603
1604 #if defined (CONFIG_GENERIC_TIME) && defined (CONFIG_X86_TSC)
1605         /* TSC could halt in idle, so notify users */
1606         if (tsc_halts_in_c(ACPI_STATE_C3))
1607                 mark_tsc_unstable("TSC halts in idle");
1608 #endif
1609         sleep_ticks = ticks_elapsed(t1, t2);
1610         /* Tell the scheduler how much we idled: */
1611         sched_clock_idle_wakeup_event(sleep_ticks*PM_TIMER_TICK_NS);
1612
1613         local_irq_enable();
1614         current_thread_info()->status |= TS_POLLING;
1615
1616         cx->usage++;
1617
1618         acpi_state_timer_broadcast(pr, cx, 0);
1619         cx->time += sleep_ticks;
1620         return ticks_elapsed_in_us(t1, t2);
1621 }
1622
1623 struct cpuidle_driver acpi_idle_driver = {
1624         .name =         "acpi_idle",
1625         .owner =        THIS_MODULE,
1626 };
1627
1628 /**
1629  * acpi_processor_setup_cpuidle - prepares and configures CPUIDLE
1630  * @pr: the ACPI processor
1631  */
1632 static int acpi_processor_setup_cpuidle(struct acpi_processor *pr)
1633 {
1634         int i, count = CPUIDLE_DRIVER_STATE_START;
1635         struct acpi_processor_cx *cx;
1636         struct cpuidle_state *state;
1637         struct cpuidle_device *dev = &pr->power.dev;
1638
1639         if (!pr->flags.power_setup_done)
1640                 return -EINVAL;
1641
1642         if (pr->flags.power == 0) {
1643                 return -EINVAL;
1644         }
1645
1646         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER && i <= max_cstate; i++) {
1647                 cx = &pr->power.states[i];
1648                 state = &dev->states[count];
1649
1650                 if (!cx->valid)
1651                         continue;
1652
1653 #ifdef CONFIG_HOTPLUG_CPU
1654                 if ((cx->type != ACPI_STATE_C1) && (num_online_cpus() > 1) &&
1655                     !pr->flags.has_cst &&
1656                     !(acpi_gbl_FADT.flags & ACPI_FADT_C2_MP_SUPPORTED))
1657                         continue;
1658 #endif
1659                 cpuidle_set_statedata(state, cx);
1660
1661                 snprintf(state->name, CPUIDLE_NAME_LEN, "C%d", i);
1662                 state->exit_latency = cx->latency;
1663                 state->target_residency = cx->latency * latency_factor;
1664                 state->power_usage = cx->power;
1665
1666                 state->flags = 0;
1667                 switch (cx->type) {
1668                         case ACPI_STATE_C1:
1669                         state->flags |= CPUIDLE_FLAG_SHALLOW;
1670                         state->flags |= CPUIDLE_FLAG_TIME_VALID;
1671                         state->enter = acpi_idle_enter_c1;
1672                         dev->safe_state = state;
1673                         break;
1674
1675                         case ACPI_STATE_C2:
1676                         state->flags |= CPUIDLE_FLAG_BALANCED;
1677                         state->flags |= CPUIDLE_FLAG_TIME_VALID;
1678                         state->enter = acpi_idle_enter_simple;
1679                         dev->safe_state = state;
1680                         break;
1681
1682                         case ACPI_STATE_C3:
1683                         state->flags |= CPUIDLE_FLAG_DEEP;
1684                         state->flags |= CPUIDLE_FLAG_TIME_VALID;
1685                         state->flags |= CPUIDLE_FLAG_CHECK_BM;
1686                         state->enter = pr->flags.bm_check ?
1687                                         acpi_idle_enter_bm :
1688                                         acpi_idle_enter_simple;
1689                         break;
1690                 }
1691
1692                 count++;
1693                 if (count == CPUIDLE_STATE_MAX)
1694                         break;
1695         }
1696
1697         dev->state_count = count;
1698
1699         if (!count)
1700                 return -EINVAL;
1701
1702         return 0;
1703 }
1704
1705 int acpi_processor_cst_has_changed(struct acpi_processor *pr)
1706 {
1707         int ret;
1708
1709         if (!pr)
1710                 return -EINVAL;
1711
1712         if (nocst) {
1713                 return -ENODEV;
1714         }
1715
1716         if (!pr->flags.power_setup_done)
1717                 return -ENODEV;
1718
1719         cpuidle_pause_and_lock();
1720         cpuidle_disable_device(&pr->power.dev);
1721         acpi_processor_get_power_info(pr);
1722         acpi_processor_setup_cpuidle(pr);
1723         ret = cpuidle_enable_device(&pr->power.dev);
1724         cpuidle_resume_and_unlock();
1725
1726         return ret;
1727 }
1728
1729 #endif /* CONFIG_CPU_IDLE */
1730
1731 int __cpuinit acpi_processor_power_init(struct acpi_processor *pr,
1732                               struct acpi_device *device)
1733 {
1734         acpi_status status = 0;
1735         static int first_run;
1736         struct proc_dir_entry *entry = NULL;
1737         unsigned int i;
1738
1739
1740         if (!first_run) {
1741                 dmi_check_system(processor_power_dmi_table);
1742                 max_cstate = acpi_processor_cstate_check(max_cstate);
1743                 if (max_cstate < ACPI_C_STATES_MAX)
1744                         printk(KERN_NOTICE
1745                                "ACPI: processor limited to max C-state %d\n",
1746                                max_cstate);
1747                 first_run++;
1748 #if !defined(CONFIG_CPU_IDLE) && defined(CONFIG_SMP)
1749                 pm_qos_add_notifier(PM_QOS_CPU_DMA_LATENCY,
1750                                 &acpi_processor_latency_notifier);
1751 #endif
1752         }
1753
1754         if (!pr)
1755                 return -EINVAL;
1756
1757         if (acpi_gbl_FADT.cst_control && !nocst) {
1758                 status =
1759                     acpi_os_write_port(acpi_gbl_FADT.smi_command, acpi_gbl_FADT.cst_control, 8);
1760                 if (ACPI_FAILURE(status)) {
1761                         ACPI_EXCEPTION((AE_INFO, status,
1762                                         "Notifying BIOS of _CST ability failed"));
1763                 }
1764         }
1765
1766         acpi_processor_get_power_info(pr);
1767         pr->flags.power_setup_done = 1;
1768
1769         /*
1770          * Install the idle handler if processor power management is supported.
1771          * Note that we use previously set idle handler will be used on
1772          * platforms that only support C1.
1773          */
1774         if ((pr->flags.power) && (!boot_option_idle_override)) {
1775 #ifdef CONFIG_CPU_IDLE
1776                 acpi_processor_setup_cpuidle(pr);
1777                 pr->power.dev.cpu = pr->id;
1778                 if (cpuidle_register_device(&pr->power.dev))
1779                         return -EIO;
1780 #endif
1781
1782                 printk(KERN_INFO PREFIX "CPU%d (power states:", pr->id);
1783                 for (i = 1; i <= pr->power.count; i++)
1784                         if (pr->power.states[i].valid)
1785                                 printk(" C%d[C%d]", i,
1786                                        pr->power.states[i].type);
1787                 printk(")\n");
1788
1789 #ifndef CONFIG_CPU_IDLE
1790                 if (pr->id == 0) {
1791                         pm_idle_save = pm_idle;
1792                         pm_idle = acpi_processor_idle;
1793                 }
1794 #endif
1795         }
1796
1797         /* 'power' [R] */
1798         entry = create_proc_entry(ACPI_PROCESSOR_FILE_POWER,
1799                                   S_IRUGO, acpi_device_dir(device));
1800         if (!entry)
1801                 return -EIO;
1802         else {
1803                 entry->proc_fops = &acpi_processor_power_fops;
1804                 entry->data = acpi_driver_data(device);
1805                 entry->owner = THIS_MODULE;
1806         }
1807
1808         return 0;
1809 }
1810
1811 int acpi_processor_power_exit(struct acpi_processor *pr,
1812                               struct acpi_device *device)
1813 {
1814 #ifdef CONFIG_CPU_IDLE
1815         if ((pr->flags.power) && (!boot_option_idle_override))
1816                 cpuidle_unregister_device(&pr->power.dev);
1817 #endif
1818         pr->flags.power_setup_done = 0;
1819
1820         if (acpi_device_dir(device))
1821                 remove_proc_entry(ACPI_PROCESSOR_FILE_POWER,
1822                                   acpi_device_dir(device));
1823
1824 #ifndef CONFIG_CPU_IDLE
1825
1826         /* Unregister the idle handler when processor #0 is removed. */
1827         if (pr->id == 0) {
1828                 pm_idle = pm_idle_save;
1829
1830                 /*
1831                  * We are about to unload the current idle thread pm callback
1832                  * (pm_idle), Wait for all processors to update cached/local
1833                  * copies of pm_idle before proceeding.
1834                  */
1835                 cpu_idle_wait();
1836 #ifdef CONFIG_SMP
1837                 pm_qos_remove_notifier(PM_QOS_CPU_DMA_LATENCY,
1838                                 &acpi_processor_latency_notifier);
1839 #endif
1840         }
1841 #endif
1842
1843         return 0;
1844 }