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