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
  43 /*
  44  * Include the apic definitions for x86 to have the APIC timer related defines
  45  * available also for UP (on SMP it gets magically included via linux/smp.h).
  46  * asm/acpi.h is not an option, as it would require more include magic. Also
  47  * creating an empty asm-ia64/apic.h would just trade pest vs. cholera.
  48  */
  49 #ifdef CONFIG_X86
  50 #include <asm/apic.h>
  51 #endif
  52
  53 #include <asm/io.h>
  54 #include <asm/uaccess.h>
  55
  56 #include <acpi/acpi_bus.h>
  57 #include <acpi/processor.h>
  58
  59 #define ACPI_PROCESSOR_COMPONENT        0x01000000
  60 #define ACPI_PROCESSOR_CLASS            "processor"
  61 #define ACPI_PROCESSOR_DRIVER_NAME      "ACPI Processor Driver"
  62 #define _COMPONENT              ACPI_PROCESSOR_COMPONENT
  63 ACPI_MODULE_NAME("acpi_processor")
  64 #define ACPI_PROCESSOR_FILE_POWER       "power"
  65 #define US_TO_PM_TIMER_TICKS(t)         ((t * (PM_TIMER_FREQUENCY/1000)) / 1000)
  66 #define C2_OVERHEAD                     4       /* 1us (3.579 ticks per us) */
  67 #define C3_OVERHEAD                     4       /* 1us (3.579 ticks per us) */
  68 static void (*pm_idle_save) (void) __read_mostly;
  69 module_param(max_cstate, uint, 0644);
  70
  71 static unsigned int nocst __read_mostly;
  72 module_param(nocst, uint, 0000);
  73
  74 /*
  75  * bm_history -- bit-mask with a bit per jiffy of bus-master activity
  76  * 1000 HZ: 0xFFFFFFFF: 32 jiffies = 32ms
  77  * 800 HZ: 0xFFFFFFFF: 32 jiffies = 40ms
  78  * 100 HZ: 0x0000000F: 4 jiffies = 40ms
  79  * reduce history for more aggressive entry into C3
  80  */
  81 static unsigned int bm_history __read_mostly =
  82     (HZ >= 800 ? 0xFFFFFFFF : ((1U << (HZ / 25)) - 1));
  83 module_param(bm_history, uint, 0644);
  84 /* --------------------------------------------------------------------------
  85                                 Power Management
  86    -------------------------------------------------------------------------- */
  87
  88 /*
  89  * IBM ThinkPad R40e crashes mysteriously when going into C2 or C3.
  90  * For now disable this. Probably a bug somewhere else.
  91  *
  92  * To skip this limit, boot/load with a large max_cstate limit.
  93  */
  94 static int set_max_cstate(struct dmi_system_id *id)
  95 {
  96         if (max_cstate > ACPI_PROCESSOR_MAX_POWER)
  97                 return 0;
  98
  99         printk(KERN_NOTICE PREFIX "%s detected - limiting to C%ld max_cstate."
 100                " Override with \"processor.max_cstate=%d\"\n", id->ident,
 101                (long)id->driver_data, ACPI_PROCESSOR_MAX_POWER + 1);
 102
 103         max_cstate = (long)id->driver_data;
 104
 105         return 0;
 106 }
 107
 108 /* Actually this shouldn't be __cpuinitdata, would be better to fix the
 109    callers to only run once -AK */
 110 static struct dmi_system_id __cpuinitdata processor_power_dmi_table[] = {
 111         { set_max_cstate, "IBM ThinkPad R40e", {
 112           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 113           DMI_MATCH(DMI_BIOS_VERSION,"1SET70WW")}, (void *)1},
 114         { set_max_cstate, "IBM ThinkPad R40e", {
 115           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 116           DMI_MATCH(DMI_BIOS_VERSION,"1SET60WW")}, (void *)1},
 117         { set_max_cstate, "IBM ThinkPad R40e", {
 118           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 119           DMI_MATCH(DMI_BIOS_VERSION,"1SET43WW") }, (void*)1},
 120         { set_max_cstate, "IBM ThinkPad R40e", {
 121           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 122           DMI_MATCH(DMI_BIOS_VERSION,"1SET45WW") }, (void*)1},
 123         { set_max_cstate, "IBM ThinkPad R40e", {
 124           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 125           DMI_MATCH(DMI_BIOS_VERSION,"1SET47WW") }, (void*)1},
 126         { set_max_cstate, "IBM ThinkPad R40e", {
 127           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 128           DMI_MATCH(DMI_BIOS_VERSION,"1SET50WW") }, (void*)1},
 129         { set_max_cstate, "IBM ThinkPad R40e", {
 130           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 131           DMI_MATCH(DMI_BIOS_VERSION,"1SET52WW") }, (void*)1},
 132         { set_max_cstate, "IBM ThinkPad R40e", {
 133           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 134           DMI_MATCH(DMI_BIOS_VERSION,"1SET55WW") }, (void*)1},
 135         { set_max_cstate, "IBM ThinkPad R40e", {
 136           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 137           DMI_MATCH(DMI_BIOS_VERSION,"1SET56WW") }, (void*)1},
 138         { set_max_cstate, "IBM ThinkPad R40e", {
 139           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 140           DMI_MATCH(DMI_BIOS_VERSION,"1SET59WW") }, (void*)1},
 141         { set_max_cstate, "IBM ThinkPad R40e", {
 142           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 143           DMI_MATCH(DMI_BIOS_VERSION,"1SET60WW") }, (void*)1},
 144         { set_max_cstate, "IBM ThinkPad R40e", {
 145           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 146           DMI_MATCH(DMI_BIOS_VERSION,"1SET61WW") }, (void*)1},
 147         { set_max_cstate, "IBM ThinkPad R40e", {
 148           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 149           DMI_MATCH(DMI_BIOS_VERSION,"1SET62WW") }, (void*)1},
 150         { set_max_cstate, "IBM ThinkPad R40e", {
 151           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 152           DMI_MATCH(DMI_BIOS_VERSION,"1SET64WW") }, (void*)1},
 153         { set_max_cstate, "IBM ThinkPad R40e", {
 154           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 155           DMI_MATCH(DMI_BIOS_VERSION,"1SET65WW") }, (void*)1},
 156         { set_max_cstate, "IBM ThinkPad R40e", {
 157           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 158           DMI_MATCH(DMI_BIOS_VERSION,"1SET68WW") }, (void*)1},
 159         { set_max_cstate, "Medion 41700", {
 160           DMI_MATCH(DMI_BIOS_VENDOR,"Phoenix Technologies LTD"),
 161           DMI_MATCH(DMI_BIOS_VERSION,"R01-A1J")}, (void *)1},
 162         { set_max_cstate, "Clevo 5600D", {
 163           DMI_MATCH(DMI_BIOS_VENDOR,"Phoenix Technologies LTD"),
 164           DMI_MATCH(DMI_BIOS_VERSION,"SHE845M0.86C.0013.D.0302131307")},
 165          (void *)2},
 166         {},
 167 };
 168
 169 static inline u32 ticks_elapsed(u32 t1, u32 t2)
 170 {
 171         if (t2 >= t1)
 172                 return (t2 - t1);
 173         else if (!(acpi_gbl_FADT.flags & ACPI_FADT_32BIT_TIMER))
 174                 return (((0x00FFFFFF - t1) + t2) & 0x00FFFFFF);
 175         else
 176                 return ((0xFFFFFFFF - t1) + t2);
 177 }
 178
 179 static void
 180 acpi_processor_power_activate(struct acpi_processor *pr,
 181                               struct acpi_processor_cx *new)
 182 {
 183         struct acpi_processor_cx *old;
 184
 185         if (!pr || !new)
 186                 return;
 187
 188         old = pr->power.state;
 189
 190         if (old)
 191                 old->promotion.count = 0;
 192         new->demotion.count = 0;
 193
 194         /* Cleanup from old state. */
 195         if (old) {
 196                 switch (old->type) {
 197                 case ACPI_STATE_C3:
 198                         /* Disable bus master reload */
 199                         if (new->type != ACPI_STATE_C3 && pr->flags.bm_check)
 200                                 acpi_set_register(ACPI_BITREG_BUS_MASTER_RLD, 0);
 201                         break;
 202                 }
 203         }
 204
 205         /* Prepare to use new state. */
 206         switch (new->type) {
 207         case ACPI_STATE_C3:
 208                 /* Enable bus master reload */
 209                 if (old->type != ACPI_STATE_C3 && pr->flags.bm_check)
 210                         acpi_set_register(ACPI_BITREG_BUS_MASTER_RLD, 1);
 211                 break;
 212         }
 213
 214         pr->power.state = new;
 215
 216         return;
 217 }
 218
 219 static void acpi_safe_halt(void)
 220 {
 221         current_thread_info()->status &= ~TS_POLLING;
 222         /*
 223          * TS_POLLING-cleared state must be visible before we
 224          * test NEED_RESCHED:
 225          */
 226         smp_mb();
 227         if (!need_resched())
 228                 safe_halt();
 229         current_thread_info()->status |= TS_POLLING;
 230 }
 231
 232 static atomic_t c3_cpu_count;
 233
 234 /* Common C-state entry for C2, C3, .. */
 235 static void acpi_cstate_enter(struct acpi_processor_cx *cstate)
 236 {
 237         if (cstate->space_id == ACPI_CSTATE_FFH) {
 238                 /* Call into architectural FFH based C-state */
 239                 acpi_processor_ffh_cstate_enter(cstate);
 240         } else {
 241                 int unused;
 242                 /* IO port based C-state */
 243                 inb(cstate->address);
 244                 /* Dummy wait op - must do something useless after P_LVL2 read
 245                    because chipsets cannot guarantee that STPCLK# signal
 246                    gets asserted in time to freeze execution properly. */
 247                 unused = inl(acpi_gbl_FADT.xpm_timer_block.address);
 248         }
 249 }
 250
 251 static void acpi_processor_idle(void)
 252 {
 253         struct acpi_processor *pr = NULL;
 254         struct acpi_processor_cx *cx = NULL;
 255         struct acpi_processor_cx *next_state = NULL;
 256         int sleep_ticks = 0;
 257         u32 t1, t2 = 0;
 258
 259         pr = processors[smp_processor_id()];
 260         if (!pr)
 261                 return;
 262
 263         /*
 264          * Interrupts must be disabled during bus mastering calculations and
 265          * for C2/C3 transitions.
 266          */
 267         local_irq_disable();
 268
 269         /*
 270          * Check whether we truly need to go idle, or should
 271          * reschedule:
 272          */
 273         if (unlikely(need_resched())) {
 274                 local_irq_enable();
 275                 return;
 276         }
 277
 278         cx = pr->power.state;
 279         if (!cx) {
 280                 if (pm_idle_save)
 281                         pm_idle_save();
 282                 else
 283                         acpi_safe_halt();
 284                 return;
 285         }
 286
 287         /*
 288          * Check BM Activity
 289          * -----------------
 290          * Check for bus mastering activity (if required), record, and check
 291          * for demotion.
 292          */
 293         if (pr->flags.bm_check) {
 294                 u32 bm_status = 0;
 295                 unsigned long diff = jiffies - pr->power.bm_check_timestamp;
 296
 297                 if (diff > 31)
 298                         diff = 31;
 299
 300                 pr->power.bm_activity <<= diff;
 301
 302                 acpi_get_register(ACPI_BITREG_BUS_MASTER_STATUS, &bm_status);
 303                 if (bm_status) {
 304                         pr->power.bm_activity |= 0x1;
 305                         acpi_set_register(ACPI_BITREG_BUS_MASTER_STATUS, 1);
 306                 }
 307                 /*
 308                  * PIIX4 Erratum #18: Note that BM_STS doesn't always reflect
 309                  * the true state of bus mastering activity; forcing us to
 310                  * manually check the BMIDEA bit of each IDE channel.
 311                  */
 312                 else if (errata.piix4.bmisx) {
 313                         if ((inb_p(errata.piix4.bmisx + 0x02) & 0x01)
 314                             || (inb_p(errata.piix4.bmisx + 0x0A) & 0x01))
 315                                 pr->power.bm_activity |= 0x1;
 316                 }
 317
 318                 pr->power.bm_check_timestamp = jiffies;
 319
 320                 /*
 321                  * If bus mastering is or was active this jiffy, demote
 322                  * to avoid a faulty transition.  Note that the processor
 323                  * won't enter a low-power state during this call (to this
 324                  * function) but should upon the next.
 325                  *
 326                  * TBD: A better policy might be to fallback to the demotion
 327                  *      state (use it for this quantum only) istead of
 328                  *      demoting -- and rely on duration as our sole demotion
 329                  *      qualification.  This may, however, introduce DMA
 330                  *      issues (e.g. floppy DMA transfer overrun/underrun).
 331                  */
 332                 if ((pr->power.bm_activity & 0x1) &&
 333                     cx->demotion.threshold.bm) {
 334                         local_irq_enable();
 335                         next_state = cx->demotion.state;
 336                         goto end;
 337                 }
 338         }
 339
 340 #ifdef CONFIG_HOTPLUG_CPU
 341         /*
 342          * Check for P_LVL2_UP flag before entering C2 and above on
 343          * an SMP system. We do it here instead of doing it at _CST/P_LVL
 344          * detection phase, to work cleanly with logical CPU hotplug.
 345          */
 346         if ((cx->type != ACPI_STATE_C1) && (num_online_cpus() > 1) &&
 347             !pr->flags.has_cst && !(acpi_gbl_FADT.flags & ACPI_FADT_C2_MP_SUPPORTED))
 348                 cx = &pr->power.states[ACPI_STATE_C1];
 349 #endif
 350
 351         /*
 352          * Sleep:
 353          * ------
 354          * Invoke the current Cx state to put the processor to sleep.
 355          */
 356         if (cx->type == ACPI_STATE_C2 || cx->type == ACPI_STATE_C3) {
 357                 current_thread_info()->status &= ~TS_POLLING;
 358                 /*
 359                  * TS_POLLING-cleared state must be visible before we
 360                  * test NEED_RESCHED:
 361                  */
 362                 smp_mb();
 363                 if (need_resched()) {
 364                         current_thread_info()->status |= TS_POLLING;
 365                         local_irq_enable();
 366                         return;
 367                 }
 368         }
 369
 370         switch (cx->type) {
 371
 372         case ACPI_STATE_C1:
 373                 /*
 374                  * Invoke C1.
 375                  * Use the appropriate idle routine, the one that would
 376                  * be used without acpi C-states.
 377                  */
 378                 if (pm_idle_save)
 379                         pm_idle_save();
 380                 else
 381                         acpi_safe_halt();
 382
 383                 /*
 384                  * TBD: Can't get time duration while in C1, as resumes
 385                  *      go to an ISR rather than here.  Need to instrument
 386                  *      base interrupt handler.
 387                  */
 388                 sleep_ticks = 0xFFFFFFFF;
 389                 break;
 390
 391         case ACPI_STATE_C2:
 392                 /* Get start time (ticks) */
 393                 t1 = inl(acpi_gbl_FADT.xpm_timer_block.address);
 394                 /* Invoke C2 */
 395                 acpi_cstate_enter(cx);
 396                 /* Get end time (ticks) */
 397                 t2 = inl(acpi_gbl_FADT.xpm_timer_block.address);
 398
 399 #ifdef CONFIG_GENERIC_TIME
 400                 /* TSC halts in C2, so notify users */
 401                 mark_tsc_unstable();
 402 #endif
 403                 /* Re-enable interrupts */
 404                 local_irq_enable();
 405                 current_thread_info()->status |= TS_POLLING;
 406                 /* Compute time (ticks) that we were actually asleep */
 407                 sleep_ticks =
 408                     ticks_elapsed(t1, t2) - cx->latency_ticks - C2_OVERHEAD;
 409                 break;
 410
 411         case ACPI_STATE_C3:
 412
 413                 if (pr->flags.bm_check) {
 414                         if (atomic_inc_return(&c3_cpu_count) ==
 415                             num_online_cpus()) {
 416                                 /*
 417                                  * All CPUs are trying to go to C3
 418                                  * Disable bus master arbitration
 419                                  */
 420                                 acpi_set_register(ACPI_BITREG_ARB_DISABLE, 1);
 421                         }
 422                 } else {
 423                         /* SMP with no shared cache... Invalidate cache  */
 424                         ACPI_FLUSH_CPU_CACHE();
 425                 }
 426
 427                 /* Get start time (ticks) */
 428                 t1 = inl(acpi_gbl_FADT.xpm_timer_block.address);
 429                 /* Invoke C3 */
 430                 acpi_cstate_enter(cx);
 431                 /* Get end time (ticks) */
 432                 t2 = inl(acpi_gbl_FADT.xpm_timer_block.address);
 433                 if (pr->flags.bm_check) {
 434                         /* Enable bus master arbitration */
 435                         atomic_dec(&c3_cpu_count);
 436                         acpi_set_register(ACPI_BITREG_ARB_DISABLE, 0);
 437                 }
 438
 439 #ifdef CONFIG_GENERIC_TIME
 440                 /* TSC halts in C3, so notify users */
 441                 mark_tsc_unstable();
 442 #endif
 443                 /* Re-enable interrupts */
 444                 local_irq_enable();
 445                 current_thread_info()->status |= TS_POLLING;
 446                 /* Compute time (ticks) that we were actually asleep */
 447                 sleep_ticks =
 448                     ticks_elapsed(t1, t2) - cx->latency_ticks - C3_OVERHEAD;
 449                 break;
 450
 451         default:
 452                 local_irq_enable();
 453                 return;
 454         }
 455         cx->usage++;
 456         if ((cx->type != ACPI_STATE_C1) && (sleep_ticks > 0))
 457                 cx->time += sleep_ticks;
 458
 459         next_state = pr->power.state;
 460
 461 #ifdef CONFIG_HOTPLUG_CPU
 462         /* Don't do promotion/demotion */
 463         if ((cx->type == ACPI_STATE_C1) && (num_online_cpus() > 1) &&
 464             !pr->flags.has_cst && !(acpi_gbl_FADT.flags & ACPI_FADT_C2_MP_SUPPORTED)) {
 465                 next_state = cx;
 466                 goto end;
 467         }
 468 #endif
 469
 470         /*
 471          * Promotion?
 472          * ----------
 473          * Track the number of longs (time asleep is greater than threshold)
 474          * and promote when the count threshold is reached.  Note that bus
 475          * mastering activity may prevent promotions.
 476          * Do not promote above max_cstate.
 477          */
 478         if (cx->promotion.state &&
 479             ((cx->promotion.state - pr->power.states) <= max_cstate)) {
 480                 if (sleep_ticks > cx->promotion.threshold.ticks &&
 481                   cx->promotion.state->latency <= system_latency_constraint()) {
 482                         cx->promotion.count++;
 483                         cx->demotion.count = 0;
 484                         if (cx->promotion.count >=
 485                             cx->promotion.threshold.count) {
 486                                 if (pr->flags.bm_check) {
 487                                         if (!
 488                                             (pr->power.bm_activity & cx->
 489                                              promotion.threshold.bm)) {
 490                                                 next_state =
 491                                                     cx->promotion.state;
 492                                                 goto end;
 493                                         }
 494                                 } else {
 495                                         next_state = cx->promotion.state;
 496                                         goto end;
 497                                 }
 498                         }
 499                 }
 500         }
 501
 502         /*
 503          * Demotion?
 504          * ---------
 505          * Track the number of shorts (time asleep is less than time threshold)
 506          * and demote when the usage threshold is reached.
 507          */
 508         if (cx->demotion.state) {
 509                 if (sleep_ticks < cx->demotion.threshold.ticks) {
 510                         cx->demotion.count++;
 511                         cx->promotion.count = 0;
 512                         if (cx->demotion.count >= cx->demotion.threshold.count) {
 513                                 next_state = cx->demotion.state;
 514                                 goto end;
 515                         }
 516                 }
 517         }
 518
 519       end:
 520         /*
 521          * Demote if current state exceeds max_cstate
 522          * or if the latency of the current state is unacceptable
 523          */
 524         if ((pr->power.state - pr->power.states) > max_cstate ||
 525                 pr->power.state->latency > system_latency_constraint()) {
 526                 if (cx->demotion.state)
 527                         next_state = cx->demotion.state;
 528         }
 529
 530         /*
 531          * New Cx State?
 532          * -------------
 533          * If we're going to start using a new Cx state we must clean up
 534          * from the previous and prepare to use the new.
 535          */
 536         if (next_state != pr->power.state)
 537                 acpi_processor_power_activate(pr, next_state);
 538 }
 539
 540 static int acpi_processor_set_power_policy(struct acpi_processor *pr)
 541 {
 542         unsigned int i;
 543         unsigned int state_is_set = 0;
 544         struct acpi_processor_cx *lower = NULL;
 545         struct acpi_processor_cx *higher = NULL;
 546         struct acpi_processor_cx *cx;
 547
 548
 549         if (!pr)
 550                 return -EINVAL;
 551
 552         /*
 553          * This function sets the default Cx state policy (OS idle handler).
 554          * Our scheme is to promote quickly to C2 but more conservatively
 555          * to C3.  We're favoring C2  for its characteristics of low latency
 556          * (quick response), good power savings, and ability to allow bus
 557          * mastering activity.  Note that the Cx state policy is completely
 558          * customizable and can be altered dynamically.
 559          */
 560
 561         /* startup state */
 562         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
 563                 cx = &pr->power.states[i];
 564                 if (!cx->valid)
 565                         continue;
 566
 567                 if (!state_is_set)
 568                         pr->power.state = cx;
 569                 state_is_set++;
 570                 break;
 571         }
 572
 573         if (!state_is_set)
 574                 return -ENODEV;
 575
 576         /* demotion */
 577         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
 578                 cx = &pr->power.states[i];
 579                 if (!cx->valid)
 580                         continue;
 581
 582                 if (lower) {
 583                         cx->demotion.state = lower;
 584                         cx->demotion.threshold.ticks = cx->latency_ticks;
 585                         cx->demotion.threshold.count = 1;
 586                         if (cx->type == ACPI_STATE_C3)
 587                                 cx->demotion.threshold.bm = bm_history;
 588                 }
 589
 590                 lower = cx;
 591         }
 592
 593         /* promotion */
 594         for (i = (ACPI_PROCESSOR_MAX_POWER - 1); i > 0; i--) {
 595                 cx = &pr->power.states[i];
 596                 if (!cx->valid)
 597                         continue;
 598
 599                 if (higher) {
 600                         cx->promotion.state = higher;
 601                         cx->promotion.threshold.ticks = cx->latency_ticks;
 602                         if (cx->type >= ACPI_STATE_C2)
 603                                 cx->promotion.threshold.count = 4;
 604                         else
 605                                 cx->promotion.threshold.count = 10;
 606                         if (higher->type == ACPI_STATE_C3)
 607                                 cx->promotion.threshold.bm = bm_history;
 608                 }
 609
 610                 higher = cx;
 611         }
 612
 613         return 0;
 614 }
 615
 616 static int acpi_processor_get_power_info_fadt(struct acpi_processor *pr)
 617 {
 618
 619         if (!pr)
 620                 return -EINVAL;
 621
 622         if (!pr->pblk)
 623                 return -ENODEV;
 624
 625         /* if info is obtained from pblk/fadt, type equals state */
 626         pr->power.states[ACPI_STATE_C2].type = ACPI_STATE_C2;
 627         pr->power.states[ACPI_STATE_C3].type = ACPI_STATE_C3;
 628
 629 #ifndef CONFIG_HOTPLUG_CPU
 630         /*
 631          * Check for P_LVL2_UP flag before entering C2 and above on
 632          * an SMP system.
 633          */
 634         if ((num_online_cpus() > 1) &&
 635             !(acpi_gbl_FADT.flags & ACPI_FADT_C2_MP_SUPPORTED))
 636                 return -ENODEV;
 637 #endif
 638
 639         /* determine C2 and C3 address from pblk */
 640         pr->power.states[ACPI_STATE_C2].address = pr->pblk + 4;
 641         pr->power.states[ACPI_STATE_C3].address = pr->pblk + 5;
 642
 643         /* determine latencies from FADT */
 644         pr->power.states[ACPI_STATE_C2].latency = acpi_gbl_FADT.C2latency;
 645         pr->power.states[ACPI_STATE_C3].latency = acpi_gbl_FADT.C3latency;
 646
 647         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 648                           "lvl2[0x%08x] lvl3[0x%08x]\n",
 649                           pr->power.states[ACPI_STATE_C2].address,
 650                           pr->power.states[ACPI_STATE_C3].address));
 651
 652         return 0;
 653 }
 654
 655 static int acpi_processor_get_power_info_default(struct acpi_processor *pr)
 656 {
 657         if (!pr->power.states[ACPI_STATE_C1].valid) {
 658                 /* set the first C-State to C1 */
 659                 /* all processors need to support C1 */
 660                 pr->power.states[ACPI_STATE_C1].type = ACPI_STATE_C1;
 661                 pr->power.states[ACPI_STATE_C1].valid = 1;
 662         }
 663         /* the C0 state only exists as a filler in our array */
 664         pr->power.states[ACPI_STATE_C0].valid = 1;
 665         return 0;
 666 }
 667
 668 static int acpi_processor_get_power_info_cst(struct acpi_processor *pr)
 669 {
 670         acpi_status status = 0;
 671         acpi_integer count;
 672         int current_count;
 673         int i;
 674         struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
 675         union acpi_object *cst;
 676
 677
 678         if (nocst)
 679                 return -ENODEV;
 680
 681         current_count = 0;
 682
 683         status = acpi_evaluate_object(pr->handle, "_CST", NULL, &buffer);
 684         if (ACPI_FAILURE(status)) {
 685                 ACPI_DEBUG_PRINT((ACPI_DB_INFO, "No _CST, giving up\n"));
 686                 return -ENODEV;
 687         }
 688
 689         cst = buffer.pointer;
 690
 691         /* There must be at least 2 elements */
 692         if (!cst || (cst->type != ACPI_TYPE_PACKAGE) || cst->package.count < 2) {
 693                 printk(KERN_ERR PREFIX "not enough elements in _CST\n");
 694                 status = -EFAULT;
 695                 goto end;
 696         }
 697
 698         count = cst->package.elements[0].integer.value;
 699
 700         /* Validate number of power states. */
 701         if (count < 1 || count != cst->package.count - 1) {
 702                 printk(KERN_ERR PREFIX "count given by _CST is not valid\n");
 703                 status = -EFAULT;
 704                 goto end;
 705         }
 706
 707         /* Tell driver that at least _CST is supported. */
 708         pr->flags.has_cst = 1;
 709
 710         for (i = 1; i <= count; i++) {
 711                 union acpi_object *element;
 712                 union acpi_object *obj;
 713                 struct acpi_power_register *reg;
 714                 struct acpi_processor_cx cx;
 715
 716                 memset(&cx, 0, sizeof(cx));
 717
 718                 element = &(cst->package.elements[i]);
 719                 if (element->type != ACPI_TYPE_PACKAGE)
 720                         continue;
 721
 722                 if (element->package.count != 4)
 723                         continue;
 724
 725                 obj = &(element->package.elements[0]);
 726
 727                 if (obj->type != ACPI_TYPE_BUFFER)
 728                         continue;
 729
 730                 reg = (struct acpi_power_register *)obj->buffer.pointer;
 731
 732                 if (reg->space_id != ACPI_ADR_SPACE_SYSTEM_IO &&
 733                     (reg->space_id != ACPI_ADR_SPACE_FIXED_HARDWARE))
 734                         continue;
 735
 736                 /* There should be an easy way to extract an integer... */
 737                 obj = &(element->package.elements[1]);
 738                 if (obj->type != ACPI_TYPE_INTEGER)
 739                         continue;
 740
 741                 cx.type = obj->integer.value;
 742                 /*
 743                  * Some buggy BIOSes won't list C1 in _CST -
 744                  * Let acpi_processor_get_power_info_default() handle them later
 745                  */
 746                 if (i == 1 && cx.type != ACPI_STATE_C1)
 747                         current_count++;
 748
 749                 cx.address = reg->address;
 750                 cx.index = current_count + 1;
 751
 752                 cx.space_id = ACPI_CSTATE_SYSTEMIO;
 753                 if (reg->space_id == ACPI_ADR_SPACE_FIXED_HARDWARE) {
 754                         if (acpi_processor_ffh_cstate_probe
 755                                         (pr->id, &cx, reg) == 0) {
 756                                 cx.space_id = ACPI_CSTATE_FFH;
 757                         } else if (cx.type != ACPI_STATE_C1) {
 758                                 /*
 759                                  * C1 is a special case where FIXED_HARDWARE
 760                                  * can be handled in non-MWAIT way as well.
 761                                  * In that case, save this _CST entry info.
 762                                  * That is, we retain space_id of SYSTEM_IO for
 763                                  * halt based C1.
 764                                  * Otherwise, ignore this info and continue.
 765                                  */
 766                                 continue;
 767                         }
 768                 }
 769
 770                 obj = &(element->package.elements[2]);
 771                 if (obj->type != ACPI_TYPE_INTEGER)
 772                         continue;
 773
 774                 cx.latency = obj->integer.value;
 775
 776                 obj = &(element->package.elements[3]);
 777                 if (obj->type != ACPI_TYPE_INTEGER)
 778                         continue;
 779
 780                 cx.power = obj->integer.value;
 781
 782                 current_count++;
 783                 memcpy(&(pr->power.states[current_count]), &cx, sizeof(cx));
 784
 785                 /*
 786                  * We support total ACPI_PROCESSOR_MAX_POWER - 1
 787                  * (From 1 through ACPI_PROCESSOR_MAX_POWER - 1)
 788                  */
 789                 if (current_count >= (ACPI_PROCESSOR_MAX_POWER - 1)) {
 790                         printk(KERN_WARNING
 791                                "Limiting number of power states to max (%d)\n",
 792                                ACPI_PROCESSOR_MAX_POWER);
 793                         printk(KERN_WARNING
 794                                "Please increase ACPI_PROCESSOR_MAX_POWER if needed.\n");
 795                         break;
 796                 }
 797         }
 798
 799         ACPI_DEBUG_PRINT((ACPI_DB_INFO, "Found %d power states\n",
 800                           current_count));
 801
 802         /* Validate number of power states discovered */
 803         if (current_count < 2)
 804                 status = -EFAULT;
 805
 806       end:
 807         kfree(buffer.pointer);
 808
 809         return status;
 810 }
 811
 812 static void acpi_processor_power_verify_c2(struct acpi_processor_cx *cx)
 813 {
 814
 815         if (!cx->address)
 816                 return;
 817
 818         /*
 819          * C2 latency must be less than or equal to 100
 820          * microseconds.
 821          */
 822         else if (cx->latency > ACPI_PROCESSOR_MAX_C2_LATENCY) {
 823                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 824                                   "latency too large [%d]\n", cx->latency));
 825                 return;
 826         }
 827
 828         /*
 829          * Otherwise we've met all of our C2 requirements.
 830          * Normalize the C2 latency to expidite policy
 831          */
 832         cx->valid = 1;
 833         cx->latency_ticks = US_TO_PM_TIMER_TICKS(cx->latency);
 834
 835         return;
 836 }
 837
 838 static void acpi_processor_power_verify_c3(struct acpi_processor *pr,
 839                                            struct acpi_processor_cx *cx)
 840 {
 841         static int bm_check_flag;
 842
 843
 844         if (!cx->address)
 845                 return;
 846
 847         /*
 848          * C3 latency must be less than or equal to 1000
 849          * microseconds.
 850          */
 851         else if (cx->latency > ACPI_PROCESSOR_MAX_C3_LATENCY) {
 852                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 853                                   "latency too large [%d]\n", cx->latency));
 854                 return;
 855         }
 856
 857         /*
 858          * PIIX4 Erratum #18: We don't support C3 when Type-F (fast)
 859          * DMA transfers are used by any ISA device to avoid livelock.
 860          * Note that we could disable Type-F DMA (as recommended by
 861          * the erratum), but this is known to disrupt certain ISA
 862          * devices thus we take the conservative approach.
 863          */
 864         else if (errata.piix4.fdma) {
 865                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 866                                   "C3 not supported on PIIX4 with Type-F DMA\n"));
 867                 return;
 868         }
 869
 870         /* All the logic here assumes flags.bm_check is same across all CPUs */
 871         if (!bm_check_flag) {
 872                 /* Determine whether bm_check is needed based on CPU  */
 873                 acpi_processor_power_init_bm_check(&(pr->flags), pr->id);
 874                 bm_check_flag = pr->flags.bm_check;
 875         } else {
 876                 pr->flags.bm_check = bm_check_flag;
 877         }
 878
 879         if (pr->flags.bm_check) {
 880                 /* bus mastering control is necessary */
 881                 if (!pr->flags.bm_control) {
 882                         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 883                                           "C3 support requires bus mastering control\n"));
 884                         return;
 885                 }
 886         } else {
 887                 /*
 888                  * WBINVD should be set in fadt, for C3 state to be
 889                  * supported on when bm_check is not required.
 890                  */
 891                 if (!(acpi_gbl_FADT.flags & ACPI_FADT_WBINVD)) {
 892                         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 893                                           "Cache invalidation should work properly"
 894                                           " for C3 to be enabled on SMP systems\n"));
 895                         return;
 896                 }
 897                 acpi_set_register(ACPI_BITREG_BUS_MASTER_RLD, 0);
 898         }
 899
 900         /*
 901          * Otherwise we've met all of our C3 requirements.
 902          * Normalize the C3 latency to expidite policy.  Enable
 903          * checking of bus mastering status (bm_check) so we can
 904          * use this in our C3 policy
 905          */
 906         cx->valid = 1;
 907         cx->latency_ticks = US_TO_PM_TIMER_TICKS(cx->latency);
 908
 909         return;
 910 }
 911
 912 static int acpi_processor_power_verify(struct acpi_processor *pr)
 913 {
 914         unsigned int i;
 915         unsigned int working = 0;
 916
 917 #ifdef ARCH_APICTIMER_STOPS_ON_C3
 918         int timer_broadcast = 0;
 919         cpumask_t mask = cpumask_of_cpu(pr->id);
 920         on_each_cpu(switch_ipi_to_APIC_timer, &mask, 1, 1);
 921 #endif
 922
 923         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
 924                 struct acpi_processor_cx *cx = &pr->power.states[i];
 925
 926                 switch (cx->type) {
 927                 case ACPI_STATE_C1:
 928                         cx->valid = 1;
 929                         break;
 930
 931                 case ACPI_STATE_C2:
 932                         acpi_processor_power_verify_c2(cx);
 933 #ifdef ARCH_APICTIMER_STOPS_ON_C3
 934                         /* Some AMD systems fake C3 as C2, but still
 935                            have timer troubles */
 936                         if (cx->valid &&
 937                                 boot_cpu_data.x86_vendor == X86_VENDOR_AMD)
 938                                 timer_broadcast++;
 939 #endif
 940                         break;
 941
 942                 case ACPI_STATE_C3:
 943                         acpi_processor_power_verify_c3(pr, cx);
 944 #ifdef ARCH_APICTIMER_STOPS_ON_C3
 945                         if (cx->valid)
 946                                 timer_broadcast++;
 947 #endif
 948                         break;
 949                 }
 950
 951                 if (cx->valid)
 952                         working++;
 953         }
 954
 955 #ifdef ARCH_APICTIMER_STOPS_ON_C3
 956         if (timer_broadcast)
 957                 on_each_cpu(switch_APIC_timer_to_ipi, &mask, 1, 1);
 958 #endif
 959
 960         return (working);
 961 }
 962
 963 static int acpi_processor_get_power_info(struct acpi_processor *pr)
 964 {
 965         unsigned int i;
 966         int result;
 967
 968
 969         /* NOTE: the idle thread may not be running while calling
 970          * this function */
 971
 972         /* Zero initialize all the C-states info. */
 973         memset(pr->power.states, 0, sizeof(pr->power.states));
 974
 975         result = acpi_processor_get_power_info_cst(pr);
 976         if (result == -ENODEV)
 977                 result = acpi_processor_get_power_info_fadt(pr);
 978
 979         if (result)
 980                 return result;
 981
 982         acpi_processor_get_power_info_default(pr);
 983
 984         pr->power.count = acpi_processor_power_verify(pr);
 985
 986         /*
 987          * Set Default Policy
 988          * ------------------
 989          * Now that we know which states are supported, set the default
 990          * policy.  Note that this policy can be changed dynamically
 991          * (e.g. encourage deeper sleeps to conserve battery life when
 992          * not on AC).
 993          */
 994         result = acpi_processor_set_power_policy(pr);
 995         if (result)
 996                 return result;
 997
 998         /*
 999          * if one state of type C2 or C3 is available, mark this
1000          * CPU as being "idle manageable"
1001          */
1002         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
1003                 if (pr->power.states[i].valid) {
1004                         pr->power.count = i;
1005                         if (pr->power.states[i].type >= ACPI_STATE_C2)
1006                                 pr->flags.power = 1;
1007                 }
1008         }
1009
1010         return 0;
1011 }
1012
1013 int acpi_processor_cst_has_changed(struct acpi_processor *pr)
1014 {
1015         int result = 0;
1016
1017
1018         if (!pr)
1019                 return -EINVAL;
1020
1021         if (nocst) {
1022                 return -ENODEV;
1023         }
1024
1025         if (!pr->flags.power_setup_done)
1026                 return -ENODEV;
1027
1028         /* Fall back to the default idle loop */
1029         pm_idle = pm_idle_save;
1030         synchronize_sched();    /* Relies on interrupts forcing exit from idle. */
1031
1032         pr->flags.power = 0;
1033         result = acpi_processor_get_power_info(pr);
1034         if ((pr->flags.power == 1) && (pr->flags.power_setup_done))
1035                 pm_idle = acpi_processor_idle;
1036
1037         return result;
1038 }
1039
1040 /* proc interface */
1041
1042 static int acpi_processor_power_seq_show(struct seq_file *seq, void *offset)
1043 {
1044         struct acpi_processor *pr = seq->private;
1045         unsigned int i;
1046
1047
1048         if (!pr)
1049                 goto end;
1050
1051         seq_printf(seq, "active state:            C%zd\n"
1052                    "max_cstate:              C%d\n"
1053                    "bus master activity:     %08x\n"
1054                    "maximum allowed latency: %d usec\n",
1055                    pr->power.state ? pr->power.state - pr->power.states : 0,
1056                    max_cstate, (unsigned)pr->power.bm_activity,
1057                    system_latency_constraint());
1058
1059         seq_puts(seq, "states:\n");
1060
1061         for (i = 1; i <= pr->power.count; i++) {
1062                 seq_printf(seq, "   %cC%d:                  ",
1063                            (&pr->power.states[i] ==
1064                             pr->power.state ? '*' : ' '), i);
1065
1066                 if (!pr->power.states[i].valid) {
1067                         seq_puts(seq, "<not supported>\n");
1068                         continue;
1069                 }
1070
1071                 switch (pr->power.states[i].type) {
1072                 case ACPI_STATE_C1:
1073                         seq_printf(seq, "type[C1] ");
1074                         break;
1075                 case ACPI_STATE_C2:
1076                         seq_printf(seq, "type[C2] ");
1077                         break;
1078                 case ACPI_STATE_C3:
1079                         seq_printf(seq, "type[C3] ");
1080                         break;
1081                 default:
1082                         seq_printf(seq, "type[--] ");
1083                         break;
1084                 }
1085
1086                 if (pr->power.states[i].promotion.state)
1087                         seq_printf(seq, "promotion[C%zd] ",
1088                                    (pr->power.states[i].promotion.state -
1089                                     pr->power.states));
1090                 else
1091                         seq_puts(seq, "promotion[--] ");
1092
1093                 if (pr->power.states[i].demotion.state)
1094                         seq_printf(seq, "demotion[C%zd] ",
1095                                    (pr->power.states[i].demotion.state -
1096                                     pr->power.states));
1097                 else
1098                         seq_puts(seq, "demotion[--] ");
1099
1100                 seq_printf(seq, "latency[%03d] usage[%08d] duration[%020llu]\n",
1101                            pr->power.states[i].latency,
1102                            pr->power.states[i].usage,
1103                            (unsigned long long)pr->power.states[i].time);
1104         }
1105
1106       end:
1107         return 0;
1108 }
1109
1110 static int acpi_processor_power_open_fs(struct inode *inode, struct file *file)
1111 {
1112         return single_open(file, acpi_processor_power_seq_show,
1113                            PDE(inode)->data);
1114 }
1115
1116 static const struct file_operations acpi_processor_power_fops = {
1117         .open = acpi_processor_power_open_fs,
1118         .read = seq_read,
1119         .llseek = seq_lseek,
1120         .release = single_release,
1121 };
1122
1123 #ifdef CONFIG_SMP
1124 static void smp_callback(void *v)
1125 {
1126         /* we already woke the CPU up, nothing more to do */
1127 }
1128
1129 /*
1130  * This function gets called when a part of the kernel has a new latency
1131  * requirement.  This means we need to get all processors out of their C-state,
1132  * and then recalculate a new suitable C-state. Just do a cross-cpu IPI; that
1133  * wakes them all right up.
1134  */
1135 static int acpi_processor_latency_notify(struct notifier_block *b,
1136                 unsigned long l, void *v)
1137 {
1138         smp_call_function(smp_callback, NULL, 0, 1);
1139         return NOTIFY_OK;
1140 }
1141
1142 static struct notifier_block acpi_processor_latency_notifier = {
1143         .notifier_call = acpi_processor_latency_notify,
1144 };
1145 #endif
1146
1147 int __cpuinit acpi_processor_power_init(struct acpi_processor *pr,
1148                               struct acpi_device *device)
1149 {
1150         acpi_status status = 0;
1151         static int first_run;
1152         struct proc_dir_entry *entry = NULL;
1153         unsigned int i;
1154
1155
1156         if (!first_run) {
1157                 dmi_check_system(processor_power_dmi_table);
1158                 if (max_cstate < ACPI_C_STATES_MAX)
1159                         printk(KERN_NOTICE
1160                                "ACPI: processor limited to max C-state %d\n",
1161                                max_cstate);
1162                 first_run++;
1163 #ifdef CONFIG_SMP
1164                 register_latency_notifier(&acpi_processor_latency_notifier);
1165 #endif
1166         }
1167
1168         if (!pr)
1169                 return -EINVAL;
1170
1171         if (acpi_gbl_FADT.cst_control && !nocst) {
1172                 status =
1173                     acpi_os_write_port(acpi_gbl_FADT.smi_command, acpi_gbl_FADT.cst_control, 8);
1174                 if (ACPI_FAILURE(status)) {
1175                         ACPI_EXCEPTION((AE_INFO, status,
1176                                         "Notifying BIOS of _CST ability failed"));
1177                 }
1178         }
1179
1180         acpi_processor_get_power_info(pr);
1181
1182         /*
1183          * Install the idle handler if processor power management is supported.
1184          * Note that we use previously set idle handler will be used on
1185          * platforms that only support C1.
1186          */
1187         if ((pr->flags.power) && (!boot_option_idle_override)) {
1188                 printk(KERN_INFO PREFIX "CPU%d (power states:", pr->id);
1189                 for (i = 1; i <= pr->power.count; i++)
1190                         if (pr->power.states[i].valid)
1191                                 printk(" C%d[C%d]", i,
1192                                        pr->power.states[i].type);
1193                 printk(")\n");
1194
1195                 if (pr->id == 0) {
1196                         pm_idle_save = pm_idle;
1197                         pm_idle = acpi_processor_idle;
1198                 }
1199         }
1200
1201         /* 'power' [R] */
1202         entry = create_proc_entry(ACPI_PROCESSOR_FILE_POWER,
1203                                   S_IRUGO, acpi_device_dir(device));
1204         if (!entry)
1205                 return -EIO;
1206         else {
1207                 entry->proc_fops = &acpi_processor_power_fops;
1208                 entry->data = acpi_driver_data(device);
1209                 entry->owner = THIS_MODULE;
1210         }
1211
1212         pr->flags.power_setup_done = 1;
1213
1214         return 0;
1215 }
1216
1217 int acpi_processor_power_exit(struct acpi_processor *pr,
1218                               struct acpi_device *device)
1219 {
1220
1221         pr->flags.power_setup_done = 0;
1222
1223         if (acpi_device_dir(device))
1224                 remove_proc_entry(ACPI_PROCESSOR_FILE_POWER,
1225                                   acpi_device_dir(device));
1226
1227         /* Unregister the idle handler when processor #0 is removed. */
1228         if (pr->id == 0) {
1229                 pm_idle = pm_idle_save;
1230
1231                 /*
1232                  * We are about to unload the current idle thread pm callback
1233                  * (pm_idle), Wait for all processors to update cached/local
1234                  * copies of pm_idle before proceeding.
1235                  */
1236                 cpu_idle_wait();
1237 #ifdef CONFIG_SMP
1238                 unregister_latency_notifier(&acpi_processor_latency_notifier);
1239 #endif
1240         }
1241
1242         return 0;
1243 }