SAFE public projects git trees. - safe/jmp/linux-2.6/blob - drivers/acpi/processor_idle.c

   1 /*
   2  * processor_idle - idle state submodule to the ACPI processor driver
   3  *
   4  *  Copyright (C) 2001, 2002 Andy Grover <andrew.grover@intel.com>
   5  *  Copyright (C) 2001, 2002 Paul Diefenbaugh <paul.s.diefenbaugh@intel.com>
   6  *  Copyright (C) 2004, 2005 Dominik Brodowski <linux@brodo.de>
   7  *  Copyright (C) 2004  Anil S Keshavamurthy <anil.s.keshavamurthy@intel.com>
   8  *                      - Added processor hotplug support
   9  *  Copyright (C) 2005  Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>
  10  *                      - Added support for C3 on SMP
  11  *
  12  * ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
  13  *
  14  *  This program is free software; you can redistribute it and/or modify
  15  *  it under the terms of the GNU General Public License as published by
  16  *  the Free Software Foundation; either version 2 of the License, or (at
  17  *  your option) any later version.
  18  *
  19  *  This program is distributed in the hope that it will be useful, but
  20  *  WITHOUT ANY WARRANTY; without even the implied warranty of
  21  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  22  *  General Public License for more details.
  23  *
  24  *  You should have received a copy of the GNU General Public License along
  25  *  with this program; if not, write to the Free Software Foundation, Inc.,
  26  *  59 Temple Place, Suite 330, Boston, MA 02111-1307 USA.
  27  *
  28  * ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
  29  */
  30
  31 #include <linux/kernel.h>
  32 #include <linux/module.h>
  33 #include <linux/init.h>
  34 #include <linux/cpufreq.h>
  35 #include <linux/proc_fs.h>
  36 #include <linux/seq_file.h>
  37 #include <linux/acpi.h>
  38 #include <linux/dmi.h>
  39 #include <linux/moduleparam.h>
  40 #include <linux/sched.h>        /* need_resched() */
  41 #include <linux/latency.h>
  42
  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 _COMPONENT              ACPI_PROCESSOR_COMPONENT
  52 ACPI_MODULE_NAME("processor_idle");
  53 #define ACPI_PROCESSOR_FILE_POWER       "power"
  54 #define US_TO_PM_TIMER_TICKS(t)         ((t * (PM_TIMER_FREQUENCY/1000)) / 1000)
  55 #define C2_OVERHEAD                     4       /* 1us (3.579 ticks per us) */
  56 #define C3_OVERHEAD                     4       /* 1us (3.579 ticks per us) */
  57 static void (*pm_idle_save) (void) __read_mostly;
  58 module_param(max_cstate, uint, 0644);
  59
  60 static unsigned int nocst __read_mostly;
  61 module_param(nocst, uint, 0000);
  62
  63 /*
  64  * bm_history -- bit-mask with a bit per jiffy of bus-master activity
  65  * 1000 HZ: 0xFFFFFFFF: 32 jiffies = 32ms
  66  * 800 HZ: 0xFFFFFFFF: 32 jiffies = 40ms
  67  * 100 HZ: 0x0000000F: 4 jiffies = 40ms
  68  * reduce history for more aggressive entry into C3
  69  */
  70 static unsigned int bm_history __read_mostly =
  71     (HZ >= 800 ? 0xFFFFFFFF : ((1U << (HZ / 25)) - 1));
  72 module_param(bm_history, uint, 0644);
  73 /* --------------------------------------------------------------------------
  74                                 Power Management
  75    -------------------------------------------------------------------------- */
  76
  77 /*
  78  * IBM ThinkPad R40e crashes mysteriously when going into C2 or C3.
  79  * For now disable this. Probably a bug somewhere else.
  80  *
  81  * To skip this limit, boot/load with a large max_cstate limit.
  82  */
  83 static int set_max_cstate(struct dmi_system_id *id)
  84 {
  85         if (max_cstate > ACPI_PROCESSOR_MAX_POWER)
  86                 return 0;
  87
  88         printk(KERN_NOTICE PREFIX "%s detected - limiting to C%ld max_cstate."
  89                " Override with \"processor.max_cstate=%d\"\n", id->ident,
  90                (long)id->driver_data, ACPI_PROCESSOR_MAX_POWER + 1);
  91
  92         max_cstate = (long)id->driver_data;
  93
  94         return 0;
  95 }
  96
  97 /* Actually this shouldn't be __cpuinitdata, would be better to fix the
  98    callers to only run once -AK */
  99 static struct dmi_system_id __cpuinitdata processor_power_dmi_table[] = {
 100         { set_max_cstate, "IBM ThinkPad R40e", {
 101           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 102           DMI_MATCH(DMI_BIOS_VERSION,"1SET70WW")}, (void *)1},
 103         { set_max_cstate, "IBM ThinkPad R40e", {
 104           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 105           DMI_MATCH(DMI_BIOS_VERSION,"1SET60WW")}, (void *)1},
 106         { set_max_cstate, "IBM ThinkPad R40e", {
 107           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 108           DMI_MATCH(DMI_BIOS_VERSION,"1SET43WW") }, (void*)1},
 109         { set_max_cstate, "IBM ThinkPad R40e", {
 110           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 111           DMI_MATCH(DMI_BIOS_VERSION,"1SET45WW") }, (void*)1},
 112         { set_max_cstate, "IBM ThinkPad R40e", {
 113           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 114           DMI_MATCH(DMI_BIOS_VERSION,"1SET47WW") }, (void*)1},
 115         { set_max_cstate, "IBM ThinkPad R40e", {
 116           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 117           DMI_MATCH(DMI_BIOS_VERSION,"1SET50WW") }, (void*)1},
 118         { set_max_cstate, "IBM ThinkPad R40e", {
 119           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 120           DMI_MATCH(DMI_BIOS_VERSION,"1SET52WW") }, (void*)1},
 121         { set_max_cstate, "IBM ThinkPad R40e", {
 122           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 123           DMI_MATCH(DMI_BIOS_VERSION,"1SET55WW") }, (void*)1},
 124         { set_max_cstate, "IBM ThinkPad R40e", {
 125           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 126           DMI_MATCH(DMI_BIOS_VERSION,"1SET56WW") }, (void*)1},
 127         { set_max_cstate, "IBM ThinkPad R40e", {
 128           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 129           DMI_MATCH(DMI_BIOS_VERSION,"1SET59WW") }, (void*)1},
 130         { set_max_cstate, "IBM ThinkPad R40e", {
 131           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 132           DMI_MATCH(DMI_BIOS_VERSION,"1SET60WW") }, (void*)1},
 133         { set_max_cstate, "IBM ThinkPad R40e", {
 134           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 135           DMI_MATCH(DMI_BIOS_VERSION,"1SET61WW") }, (void*)1},
 136         { set_max_cstate, "IBM ThinkPad R40e", {
 137           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 138           DMI_MATCH(DMI_BIOS_VERSION,"1SET62WW") }, (void*)1},
 139         { set_max_cstate, "IBM ThinkPad R40e", {
 140           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 141           DMI_MATCH(DMI_BIOS_VERSION,"1SET64WW") }, (void*)1},
 142         { set_max_cstate, "IBM ThinkPad R40e", {
 143           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 144           DMI_MATCH(DMI_BIOS_VERSION,"1SET65WW") }, (void*)1},
 145         { set_max_cstate, "IBM ThinkPad R40e", {
 146           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 147           DMI_MATCH(DMI_BIOS_VERSION,"1SET68WW") }, (void*)1},
 148         { set_max_cstate, "Medion 41700", {
 149           DMI_MATCH(DMI_BIOS_VENDOR,"Phoenix Technologies LTD"),
 150           DMI_MATCH(DMI_BIOS_VERSION,"R01-A1J")}, (void *)1},
 151         { set_max_cstate, "Clevo 5600D", {
 152           DMI_MATCH(DMI_BIOS_VENDOR,"Phoenix Technologies LTD"),
 153           DMI_MATCH(DMI_BIOS_VERSION,"SHE845M0.86C.0013.D.0302131307")},
 154          (void *)2},
 155         {},
 156 };
 157
 158 static inline u32 ticks_elapsed(u32 t1, u32 t2)
 159 {
 160         if (t2 >= t1)
 161                 return (t2 - t1);
 162         else if (!(acpi_gbl_FADT.flags & ACPI_FADT_32BIT_TIMER))
 163                 return (((0x00FFFFFF - t1) + t2) & 0x00FFFFFF);
 164         else
 165                 return ((0xFFFFFFFF - t1) + t2);
 166 }
 167
 168 static void
 169 acpi_processor_power_activate(struct acpi_processor *pr,
 170                               struct acpi_processor_cx *new)
 171 {
 172         struct acpi_processor_cx *old;
 173
 174         if (!pr || !new)
 175                 return;
 176
 177         old = pr->power.state;
 178
 179         if (old)
 180                 old->promotion.count = 0;
 181         new->demotion.count = 0;
 182
 183         /* Cleanup from old state. */
 184         if (old) {
 185                 switch (old->type) {
 186                 case ACPI_STATE_C3:
 187                         /* Disable bus master reload */
 188                         if (new->type != ACPI_STATE_C3 && pr->flags.bm_check)
 189                                 acpi_set_register(ACPI_BITREG_BUS_MASTER_RLD, 0);
 190                         break;
 191                 }
 192         }
 193
 194         /* Prepare to use new state. */
 195         switch (new->type) {
 196         case ACPI_STATE_C3:
 197                 /* Enable bus master reload */
 198                 if (old->type != ACPI_STATE_C3 && pr->flags.bm_check)
 199                         acpi_set_register(ACPI_BITREG_BUS_MASTER_RLD, 1);
 200                 break;
 201         }
 202
 203         pr->power.state = new;
 204
 205         return;
 206 }
 207
 208 static void acpi_safe_halt(void)
 209 {
 210         current_thread_info()->status &= ~TS_POLLING;
 211         /*
 212          * TS_POLLING-cleared state must be visible before we
 213          * test NEED_RESCHED:
 214          */
 215         smp_mb();
 216         if (!need_resched())
 217                 safe_halt();
 218         current_thread_info()->status |= TS_POLLING;
 219 }
 220
 221 static atomic_t c3_cpu_count;
 222
 223 /* Common C-state entry for C2, C3, .. */
 224 static void acpi_cstate_enter(struct acpi_processor_cx *cstate)
 225 {
 226         if (cstate->space_id == ACPI_CSTATE_FFH) {
 227                 /* Call into architectural FFH based C-state */
 228                 acpi_processor_ffh_cstate_enter(cstate);
 229         } else {
 230                 int unused;
 231                 /* IO port based C-state */
 232                 inb(cstate->address);
 233                 /* Dummy wait op - must do something useless after P_LVL2 read
 234                    because chipsets cannot guarantee that STPCLK# signal
 235                    gets asserted in time to freeze execution properly. */
 236                 unused = inl(acpi_gbl_FADT.xpm_timer_block.address);
 237         }
 238 }
 239
 240 static void acpi_processor_idle(void)
 241 {
 242         struct acpi_processor *pr = NULL;
 243         struct acpi_processor_cx *cx = NULL;
 244         struct acpi_processor_cx *next_state = NULL;
 245         int sleep_ticks = 0;
 246         u32 t1, t2 = 0;
 247
 248         pr = processors[smp_processor_id()];
 249         if (!pr)
 250                 return;
 251
 252         /*
 253          * Interrupts must be disabled during bus mastering calculations and
 254          * for C2/C3 transitions.
 255          */
 256         local_irq_disable();
 257
 258         /*
 259          * Check whether we truly need to go idle, or should
 260          * reschedule:
 261          */
 262         if (unlikely(need_resched())) {
 263                 local_irq_enable();
 264                 return;
 265         }
 266
 267         cx = pr->power.state;
 268         if (!cx) {
 269                 if (pm_idle_save)
 270                         pm_idle_save();
 271                 else
 272                         acpi_safe_halt();
 273                 return;
 274         }
 275
 276         /*
 277          * Check BM Activity
 278          * -----------------
 279          * Check for bus mastering activity (if required), record, and check
 280          * for demotion.
 281          */
 282         if (pr->flags.bm_check) {
 283                 u32 bm_status = 0;
 284                 unsigned long diff = jiffies - pr->power.bm_check_timestamp;
 285
 286                 if (diff > 31)
 287                         diff = 31;
 288
 289                 pr->power.bm_activity <<= diff;
 290
 291                 acpi_get_register(ACPI_BITREG_BUS_MASTER_STATUS, &bm_status);
 292                 if (bm_status) {
 293                         pr->power.bm_activity |= 0x1;
 294                         acpi_set_register(ACPI_BITREG_BUS_MASTER_STATUS, 1);
 295                 }
 296                 /*
 297                  * PIIX4 Erratum #18: Note that BM_STS doesn't always reflect
 298                  * the true state of bus mastering activity; forcing us to
 299                  * manually check the BMIDEA bit of each IDE channel.
 300                  */
 301                 else if (errata.piix4.bmisx) {
 302                         if ((inb_p(errata.piix4.bmisx + 0x02) & 0x01)
 303                             || (inb_p(errata.piix4.bmisx + 0x0A) & 0x01))
 304                                 pr->power.bm_activity |= 0x1;
 305                 }
 306
 307                 pr->power.bm_check_timestamp = jiffies;
 308
 309                 /*
 310                  * If bus mastering is or was active this jiffy, demote
 311                  * to avoid a faulty transition.  Note that the processor
 312                  * won't enter a low-power state during this call (to this
 313                  * function) but should upon the next.
 314                  *
 315                  * TBD: A better policy might be to fallback to the demotion
 316                  *      state (use it for this quantum only) istead of
 317                  *      demoting -- and rely on duration as our sole demotion
 318                  *      qualification.  This may, however, introduce DMA
 319                  *      issues (e.g. floppy DMA transfer overrun/underrun).
 320                  */
 321                 if ((pr->power.bm_activity & 0x1) &&
 322                     cx->demotion.threshold.bm) {
 323                         local_irq_enable();
 324                         next_state = cx->demotion.state;
 325                         goto end;
 326                 }
 327         }
 328
 329 #ifdef CONFIG_HOTPLUG_CPU
 330         /*
 331          * Check for P_LVL2_UP flag before entering C2 and above on
 332          * an SMP system. We do it here instead of doing it at _CST/P_LVL
 333          * detection phase, to work cleanly with logical CPU hotplug.
 334          */
 335         if ((cx->type != ACPI_STATE_C1) && (num_online_cpus() > 1) &&
 336             !pr->flags.has_cst && !(acpi_gbl_FADT.flags & ACPI_FADT_C2_MP_SUPPORTED))
 337                 cx = &pr->power.states[ACPI_STATE_C1];
 338 #endif
 339
 340         /*
 341          * Sleep:
 342          * ------
 343          * Invoke the current Cx state to put the processor to sleep.
 344          */
 345         if (cx->type == ACPI_STATE_C2 || cx->type == ACPI_STATE_C3) {
 346                 current_thread_info()->status &= ~TS_POLLING;
 347                 /*
 348                  * TS_POLLING-cleared state must be visible before we
 349                  * test NEED_RESCHED:
 350                  */
 351                 smp_mb();
 352                 if (need_resched()) {
 353                         current_thread_info()->status |= TS_POLLING;
 354                         local_irq_enable();
 355                         return;
 356                 }
 357         }
 358
 359         switch (cx->type) {
 360
 361         case ACPI_STATE_C1:
 362                 /*
 363                  * Invoke C1.
 364                  * Use the appropriate idle routine, the one that would
 365                  * be used without acpi C-states.
 366                  */
 367                 if (pm_idle_save)
 368                         pm_idle_save();
 369                 else
 370                         acpi_safe_halt();
 371
 372                 /*
 373                  * TBD: Can't get time duration while in C1, as resumes
 374                  *      go to an ISR rather than here.  Need to instrument
 375                  *      base interrupt handler.
 376                  */
 377                 sleep_ticks = 0xFFFFFFFF;
 378                 break;
 379
 380         case ACPI_STATE_C2:
 381                 /* Get start time (ticks) */
 382                 t1 = inl(acpi_gbl_FADT.xpm_timer_block.address);
 383                 /* Invoke C2 */
 384                 acpi_cstate_enter(cx);
 385                 /* Get end time (ticks) */
 386                 t2 = inl(acpi_gbl_FADT.xpm_timer_block.address);
 387
 388 #ifdef CONFIG_GENERIC_TIME
 389                 /* TSC halts in C2, so notify users */
 390                 mark_tsc_unstable();
 391 #endif
 392                 /* Re-enable interrupts */
 393                 local_irq_enable();
 394                 current_thread_info()->status |= TS_POLLING;
 395                 /* Compute time (ticks) that we were actually asleep */
 396                 sleep_ticks =
 397                     ticks_elapsed(t1, t2) - cx->latency_ticks - C2_OVERHEAD;
 398                 break;
 399
 400         case ACPI_STATE_C3:
 401
 402                 if (pr->flags.bm_check) {
 403                         if (atomic_inc_return(&c3_cpu_count) ==
 404                             num_online_cpus()) {
 405                                 /*
 406                                  * All CPUs are trying to go to C3
 407                                  * Disable bus master arbitration
 408                                  */
 409                                 acpi_set_register(ACPI_BITREG_ARB_DISABLE, 1);
 410                         }
 411                 } else {
 412                         /* SMP with no shared cache... Invalidate cache  */
 413                         ACPI_FLUSH_CPU_CACHE();
 414                 }
 415
 416                 /* Get start time (ticks) */
 417                 t1 = inl(acpi_gbl_FADT.xpm_timer_block.address);
 418                 /* Invoke C3 */
 419                 acpi_cstate_enter(cx);
 420                 /* Get end time (ticks) */
 421                 t2 = inl(acpi_gbl_FADT.xpm_timer_block.address);
 422                 if (pr->flags.bm_check) {
 423                         /* Enable bus master arbitration */
 424                         atomic_dec(&c3_cpu_count);
 425                         acpi_set_register(ACPI_BITREG_ARB_DISABLE, 0);
 426                 }
 427
 428 #ifdef CONFIG_GENERIC_TIME
 429                 /* TSC halts in C3, so notify users */
 430                 mark_tsc_unstable();
 431 #endif
 432                 /* Re-enable interrupts */
 433                 local_irq_enable();
 434                 current_thread_info()->status |= TS_POLLING;
 435                 /* Compute time (ticks) that we were actually asleep */
 436                 sleep_ticks =
 437                     ticks_elapsed(t1, t2) - cx->latency_ticks - C3_OVERHEAD;
 438                 break;
 439
 440         default:
 441                 local_irq_enable();
 442                 return;
 443         }
 444         cx->usage++;
 445         if ((cx->type != ACPI_STATE_C1) && (sleep_ticks > 0))
 446                 cx->time += sleep_ticks;
 447
 448         next_state = pr->power.state;
 449
 450 #ifdef CONFIG_HOTPLUG_CPU
 451         /* Don't do promotion/demotion */
 452         if ((cx->type == ACPI_STATE_C1) && (num_online_cpus() > 1) &&
 453             !pr->flags.has_cst && !(acpi_gbl_FADT.flags & ACPI_FADT_C2_MP_SUPPORTED)) {
 454                 next_state = cx;
 455                 goto end;
 456         }
 457 #endif
 458
 459         /*
 460          * Promotion?
 461          * ----------
 462          * Track the number of longs (time asleep is greater than threshold)
 463          * and promote when the count threshold is reached.  Note that bus
 464          * mastering activity may prevent promotions.
 465          * Do not promote above max_cstate.
 466          */
 467         if (cx->promotion.state &&
 468             ((cx->promotion.state - pr->power.states) <= max_cstate)) {
 469                 if (sleep_ticks > cx->promotion.threshold.ticks &&
 470                   cx->promotion.state->latency <= system_latency_constraint()) {
 471                         cx->promotion.count++;
 472                         cx->demotion.count = 0;
 473                         if (cx->promotion.count >=
 474                             cx->promotion.threshold.count) {
 475                                 if (pr->flags.bm_check) {
 476                                         if (!
 477                                             (pr->power.bm_activity & cx->
 478                                              promotion.threshold.bm)) {
 479                                                 next_state =
 480                                                     cx->promotion.state;
 481                                                 goto end;
 482                                         }
 483                                 } else {
 484                                         next_state = cx->promotion.state;
 485                                         goto end;
 486                                 }
 487                         }
 488                 }
 489         }
 490
 491         /*
 492          * Demotion?
 493          * ---------
 494          * Track the number of shorts (time asleep is less than time threshold)
 495          * and demote when the usage threshold is reached.
 496          */
 497         if (cx->demotion.state) {
 498                 if (sleep_ticks < cx->demotion.threshold.ticks) {
 499                         cx->demotion.count++;
 500                         cx->promotion.count = 0;
 501                         if (cx->demotion.count >= cx->demotion.threshold.count) {
 502                                 next_state = cx->demotion.state;
 503                                 goto end;
 504                         }
 505                 }
 506         }
 507
 508       end:
 509         /*
 510          * Demote if current state exceeds max_cstate
 511          * or if the latency of the current state is unacceptable
 512          */
 513         if ((pr->power.state - pr->power.states) > max_cstate ||
 514                 pr->power.state->latency > system_latency_constraint()) {
 515                 if (cx->demotion.state)
 516                         next_state = cx->demotion.state;
 517         }
 518
 519         /*
 520          * New Cx State?
 521          * -------------
 522          * If we're going to start using a new Cx state we must clean up
 523          * from the previous and prepare to use the new.
 524          */
 525         if (next_state != pr->power.state)
 526                 acpi_processor_power_activate(pr, next_state);
 527 }
 528
 529 static int acpi_processor_set_power_policy(struct acpi_processor *pr)
 530 {
 531         unsigned int i;
 532         unsigned int state_is_set = 0;
 533         struct acpi_processor_cx *lower = NULL;
 534         struct acpi_processor_cx *higher = NULL;
 535         struct acpi_processor_cx *cx;
 536
 537
 538         if (!pr)
 539                 return -EINVAL;
 540
 541         /*
 542          * This function sets the default Cx state policy (OS idle handler).
 543          * Our scheme is to promote quickly to C2 but more conservatively
 544          * to C3.  We're favoring C2  for its characteristics of low latency
 545          * (quick response), good power savings, and ability to allow bus
 546          * mastering activity.  Note that the Cx state policy is completely
 547          * customizable and can be altered dynamically.
 548          */
 549
 550         /* startup state */
 551         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
 552                 cx = &pr->power.states[i];
 553                 if (!cx->valid)
 554                         continue;
 555
 556                 if (!state_is_set)
 557                         pr->power.state = cx;
 558                 state_is_set++;
 559                 break;
 560         }
 561
 562         if (!state_is_set)
 563                 return -ENODEV;
 564
 565         /* demotion */
 566         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
 567                 cx = &pr->power.states[i];
 568                 if (!cx->valid)
 569                         continue;
 570
 571                 if (lower) {
 572                         cx->demotion.state = lower;
 573                         cx->demotion.threshold.ticks = cx->latency_ticks;
 574                         cx->demotion.threshold.count = 1;
 575                         if (cx->type == ACPI_STATE_C3)
 576                                 cx->demotion.threshold.bm = bm_history;
 577                 }
 578
 579                 lower = cx;
 580         }
 581
 582         /* promotion */
 583         for (i = (ACPI_PROCESSOR_MAX_POWER - 1); i > 0; i--) {
 584                 cx = &pr->power.states[i];
 585                 if (!cx->valid)
 586                         continue;
 587
 588                 if (higher) {
 589                         cx->promotion.state = higher;
 590                         cx->promotion.threshold.ticks = cx->latency_ticks;
 591                         if (cx->type >= ACPI_STATE_C2)
 592                                 cx->promotion.threshold.count = 4;
 593                         else
 594                                 cx->promotion.threshold.count = 10;
 595                         if (higher->type == ACPI_STATE_C3)
 596                                 cx->promotion.threshold.bm = bm_history;
 597                 }
 598
 599                 higher = cx;
 600         }
 601
 602         return 0;
 603 }
 604
 605 static int acpi_processor_get_power_info_fadt(struct acpi_processor *pr)
 606 {
 607
 608         if (!pr)
 609                 return -EINVAL;
 610
 611         if (!pr->pblk)
 612                 return -ENODEV;
 613
 614         /* if info is obtained from pblk/fadt, type equals state */
 615         pr->power.states[ACPI_STATE_C2].type = ACPI_STATE_C2;
 616         pr->power.states[ACPI_STATE_C3].type = ACPI_STATE_C3;
 617
 618 #ifndef CONFIG_HOTPLUG_CPU
 619         /*
 620          * Check for P_LVL2_UP flag before entering C2 and above on
 621          * an SMP system.
 622          */
 623         if ((num_online_cpus() > 1) &&
 624             !(acpi_gbl_FADT.flags & ACPI_FADT_C2_MP_SUPPORTED))
 625                 return -ENODEV;
 626 #endif
 627
 628         /* determine C2 and C3 address from pblk */
 629         pr->power.states[ACPI_STATE_C2].address = pr->pblk + 4;
 630         pr->power.states[ACPI_STATE_C3].address = pr->pblk + 5;
 631
 632         /* determine latencies from FADT */
 633         pr->power.states[ACPI_STATE_C2].latency = acpi_gbl_FADT.C2latency;
 634         pr->power.states[ACPI_STATE_C3].latency = acpi_gbl_FADT.C3latency;
 635
 636         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 637                           "lvl2[0x%08x] lvl3[0x%08x]\n",
 638                           pr->power.states[ACPI_STATE_C2].address,
 639                           pr->power.states[ACPI_STATE_C3].address));
 640
 641         return 0;
 642 }
 643
 644 static int acpi_processor_get_power_info_default(struct acpi_processor *pr)
 645 {
 646         if (!pr->power.states[ACPI_STATE_C1].valid) {
 647                 /* set the first C-State to C1 */
 648                 /* all processors need to support C1 */
 649                 pr->power.states[ACPI_STATE_C1].type = ACPI_STATE_C1;
 650                 pr->power.states[ACPI_STATE_C1].valid = 1;
 651         }
 652         /* the C0 state only exists as a filler in our array */
 653         pr->power.states[ACPI_STATE_C0].valid = 1;
 654         return 0;
 655 }
 656
 657 static int acpi_processor_get_power_info_cst(struct acpi_processor *pr)
 658 {
 659         acpi_status status = 0;
 660         acpi_integer count;
 661         int current_count;
 662         int i;
 663         struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
 664         union acpi_object *cst;
 665
 666
 667         if (nocst)
 668                 return -ENODEV;
 669
 670         current_count = 0;
 671
 672         status = acpi_evaluate_object(pr->handle, "_CST", NULL, &buffer);
 673         if (ACPI_FAILURE(status)) {
 674                 ACPI_DEBUG_PRINT((ACPI_DB_INFO, "No _CST, giving up\n"));
 675                 return -ENODEV;
 676         }
 677
 678         cst = buffer.pointer;
 679
 680         /* There must be at least 2 elements */
 681         if (!cst || (cst->type != ACPI_TYPE_PACKAGE) || cst->package.count < 2) {
 682                 printk(KERN_ERR PREFIX "not enough elements in _CST\n");
 683                 status = -EFAULT;
 684                 goto end;
 685         }
 686
 687         count = cst->package.elements[0].integer.value;
 688
 689         /* Validate number of power states. */
 690         if (count < 1 || count != cst->package.count - 1) {
 691                 printk(KERN_ERR PREFIX "count given by _CST is not valid\n");
 692                 status = -EFAULT;
 693                 goto end;
 694         }
 695
 696         /* Tell driver that at least _CST is supported. */
 697         pr->flags.has_cst = 1;
 698
 699         for (i = 1; i <= count; i++) {
 700                 union acpi_object *element;
 701                 union acpi_object *obj;
 702                 struct acpi_power_register *reg;
 703                 struct acpi_processor_cx cx;
 704
 705                 memset(&cx, 0, sizeof(cx));
 706
 707                 element = &(cst->package.elements[i]);
 708                 if (element->type != ACPI_TYPE_PACKAGE)
 709                         continue;
 710
 711                 if (element->package.count != 4)
 712                         continue;
 713
 714                 obj = &(element->package.elements[0]);
 715
 716                 if (obj->type != ACPI_TYPE_BUFFER)
 717                         continue;
 718
 719                 reg = (struct acpi_power_register *)obj->buffer.pointer;
 720
 721                 if (reg->space_id != ACPI_ADR_SPACE_SYSTEM_IO &&
 722                     (reg->space_id != ACPI_ADR_SPACE_FIXED_HARDWARE))
 723                         continue;
 724
 725                 /* There should be an easy way to extract an integer... */
 726                 obj = &(element->package.elements[1]);
 727                 if (obj->type != ACPI_TYPE_INTEGER)
 728                         continue;
 729
 730                 cx.type = obj->integer.value;
 731                 /*
 732                  * Some buggy BIOSes won't list C1 in _CST -
 733                  * Let acpi_processor_get_power_info_default() handle them later
 734                  */
 735                 if (i == 1 && cx.type != ACPI_STATE_C1)
 736                         current_count++;
 737
 738                 cx.address = reg->address;
 739                 cx.index = current_count + 1;
 740
 741                 cx.space_id = ACPI_CSTATE_SYSTEMIO;
 742                 if (reg->space_id == ACPI_ADR_SPACE_FIXED_HARDWARE) {
 743                         if (acpi_processor_ffh_cstate_probe
 744                                         (pr->id, &cx, reg) == 0) {
 745                                 cx.space_id = ACPI_CSTATE_FFH;
 746                         } else if (cx.type != ACPI_STATE_C1) {
 747                                 /*
 748                                  * C1 is a special case where FIXED_HARDWARE
 749                                  * can be handled in non-MWAIT way as well.
 750                                  * In that case, save this _CST entry info.
 751                                  * That is, we retain space_id of SYSTEM_IO for
 752                                  * halt based C1.
 753                                  * Otherwise, ignore this info and continue.
 754                                  */
 755                                 continue;
 756                         }
 757                 }
 758
 759                 obj = &(element->package.elements[2]);
 760                 if (obj->type != ACPI_TYPE_INTEGER)
 761                         continue;
 762
 763                 cx.latency = obj->integer.value;
 764
 765                 obj = &(element->package.elements[3]);
 766                 if (obj->type != ACPI_TYPE_INTEGER)
 767                         continue;
 768
 769                 cx.power = obj->integer.value;
 770
 771                 current_count++;
 772                 memcpy(&(pr->power.states[current_count]), &cx, sizeof(cx));
 773
 774                 /*
 775                  * We support total ACPI_PROCESSOR_MAX_POWER - 1
 776                  * (From 1 through ACPI_PROCESSOR_MAX_POWER - 1)
 777                  */
 778                 if (current_count >= (ACPI_PROCESSOR_MAX_POWER - 1)) {
 779                         printk(KERN_WARNING
 780                                "Limiting number of power states to max (%d)\n",
 781                                ACPI_PROCESSOR_MAX_POWER);
 782                         printk(KERN_WARNING
 783                                "Please increase ACPI_PROCESSOR_MAX_POWER if needed.\n");
 784                         break;
 785                 }
 786         }
 787
 788         ACPI_DEBUG_PRINT((ACPI_DB_INFO, "Found %d power states\n",
 789                           current_count));
 790
 791         /* Validate number of power states discovered */
 792         if (current_count < 2)
 793                 status = -EFAULT;
 794
 795       end:
 796         kfree(buffer.pointer);
 797
 798         return status;
 799 }
 800
 801 static void acpi_processor_power_verify_c2(struct acpi_processor_cx *cx)
 802 {
 803
 804         if (!cx->address)
 805                 return;
 806
 807         /*
 808          * C2 latency must be less than or equal to 100
 809          * microseconds.
 810          */
 811         else if (cx->latency > ACPI_PROCESSOR_MAX_C2_LATENCY) {
 812                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 813                                   "latency too large [%d]\n", cx->latency));
 814                 return;
 815         }
 816
 817         /*
 818          * Otherwise we've met all of our C2 requirements.
 819          * Normalize the C2 latency to expidite policy
 820          */
 821         cx->valid = 1;
 822         cx->latency_ticks = US_TO_PM_TIMER_TICKS(cx->latency);
 823
 824         return;
 825 }
 826
 827 static void acpi_processor_power_verify_c3(struct acpi_processor *pr,
 828                                            struct acpi_processor_cx *cx)
 829 {
 830         static int bm_check_flag;
 831
 832
 833         if (!cx->address)
 834                 return;
 835
 836         /*
 837          * C3 latency must be less than or equal to 1000
 838          * microseconds.
 839          */
 840         else if (cx->latency > ACPI_PROCESSOR_MAX_C3_LATENCY) {
 841                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 842                                   "latency too large [%d]\n", cx->latency));
 843                 return;
 844         }
 845
 846         /*
 847          * PIIX4 Erratum #18: We don't support C3 when Type-F (fast)
 848          * DMA transfers are used by any ISA device to avoid livelock.
 849          * Note that we could disable Type-F DMA (as recommended by
 850          * the erratum), but this is known to disrupt certain ISA
 851          * devices thus we take the conservative approach.
 852          */
 853         else if (errata.piix4.fdma) {
 854                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 855                                   "C3 not supported on PIIX4 with Type-F DMA\n"));
 856                 return;
 857         }
 858
 859         /* All the logic here assumes flags.bm_check is same across all CPUs */
 860         if (!bm_check_flag) {
 861                 /* Determine whether bm_check is needed based on CPU  */
 862                 acpi_processor_power_init_bm_check(&(pr->flags), pr->id);
 863                 bm_check_flag = pr->flags.bm_check;
 864         } else {
 865                 pr->flags.bm_check = bm_check_flag;
 866         }
 867
 868         if (pr->flags.bm_check) {
 869                 /* bus mastering control is necessary */
 870                 if (!pr->flags.bm_control) {
 871                         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 872                                           "C3 support requires bus mastering control\n"));
 873                         return;
 874                 }
 875         } else {
 876                 /*
 877                  * WBINVD should be set in fadt, for C3 state to be
 878                  * supported on when bm_check is not required.
 879                  */
 880                 if (!(acpi_gbl_FADT.flags & ACPI_FADT_WBINVD)) {
 881                         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 882                                           "Cache invalidation should work properly"
 883                                           " for C3 to be enabled on SMP systems\n"));
 884                         return;
 885                 }
 886                 acpi_set_register(ACPI_BITREG_BUS_MASTER_RLD, 0);
 887         }
 888
 889         /*
 890          * Otherwise we've met all of our C3 requirements.
 891          * Normalize the C3 latency to expidite policy.  Enable
 892          * checking of bus mastering status (bm_check) so we can
 893          * use this in our C3 policy
 894          */
 895         cx->valid = 1;
 896         cx->latency_ticks = US_TO_PM_TIMER_TICKS(cx->latency);
 897
 898         return;
 899 }
 900
 901 static int acpi_processor_power_verify(struct acpi_processor *pr)
 902 {
 903         unsigned int i;
 904         unsigned int working = 0;
 905
 906 #ifdef ARCH_APICTIMER_STOPS_ON_C3
 907         int timer_broadcast = 0;
 908         cpumask_t mask = cpumask_of_cpu(pr->id);
 909         on_each_cpu(switch_ipi_to_APIC_timer, &mask, 1, 1);
 910 #endif
 911
 912         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
 913                 struct acpi_processor_cx *cx = &pr->power.states[i];
 914
 915                 switch (cx->type) {
 916                 case ACPI_STATE_C1:
 917                         cx->valid = 1;
 918                         break;
 919
 920                 case ACPI_STATE_C2:
 921                         acpi_processor_power_verify_c2(cx);
 922 #ifdef ARCH_APICTIMER_STOPS_ON_C3
 923                         /* Some AMD systems fake C3 as C2, but still
 924                            have timer troubles */
 925                         if (cx->valid &&
 926                                 boot_cpu_data.x86_vendor == X86_VENDOR_AMD)
 927                                 timer_broadcast++;
 928 #endif
 929                         break;
 930
 931                 case ACPI_STATE_C3:
 932                         acpi_processor_power_verify_c3(pr, cx);
 933 #ifdef ARCH_APICTIMER_STOPS_ON_C3
 934                         if (cx->valid)
 935                                 timer_broadcast++;
 936 #endif
 937                         break;
 938                 }
 939
 940                 if (cx->valid)
 941                         working++;
 942         }
 943
 944 #ifdef ARCH_APICTIMER_STOPS_ON_C3
 945         if (timer_broadcast)
 946                 on_each_cpu(switch_APIC_timer_to_ipi, &mask, 1, 1);
 947 #endif
 948
 949         return (working);
 950 }
 951
 952 static int acpi_processor_get_power_info(struct acpi_processor *pr)
 953 {
 954         unsigned int i;
 955         int result;
 956
 957
 958         /* NOTE: the idle thread may not be running while calling
 959          * this function */
 960
 961         /* Zero initialize all the C-states info. */
 962         memset(pr->power.states, 0, sizeof(pr->power.states));
 963
 964         result = acpi_processor_get_power_info_cst(pr);
 965         if (result == -ENODEV)
 966                 result = acpi_processor_get_power_info_fadt(pr);
 967
 968         if (result)
 969                 return result;
 970
 971         acpi_processor_get_power_info_default(pr);
 972
 973         pr->power.count = acpi_processor_power_verify(pr);
 974
 975         /*
 976          * Set Default Policy
 977          * ------------------
 978          * Now that we know which states are supported, set the default
 979          * policy.  Note that this policy can be changed dynamically
 980          * (e.g. encourage deeper sleeps to conserve battery life when
 981          * not on AC).
 982          */
 983         result = acpi_processor_set_power_policy(pr);
 984         if (result)
 985                 return result;
 986
 987         /*
 988          * if one state of type C2 or C3 is available, mark this
 989          * CPU as being "idle manageable"
 990          */
 991         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
 992                 if (pr->power.states[i].valid) {
 993                         pr->power.count = i;
 994                         if (pr->power.states[i].type >= ACPI_STATE_C2)
 995                                 pr->flags.power = 1;
 996                 }
 997         }
 998
 999         return 0;
1000 }
1001
1002 int acpi_processor_cst_has_changed(struct acpi_processor *pr)
1003 {
1004         int result = 0;
1005
1006
1007         if (!pr)
1008                 return -EINVAL;
1009
1010         if (nocst) {
1011                 return -ENODEV;
1012         }
1013
1014         if (!pr->flags.power_setup_done)
1015                 return -ENODEV;
1016
1017         /* Fall back to the default idle loop */
1018         pm_idle = pm_idle_save;
1019         synchronize_sched();    /* Relies on interrupts forcing exit from idle. */
1020
1021         pr->flags.power = 0;
1022         result = acpi_processor_get_power_info(pr);
1023         if ((pr->flags.power == 1) && (pr->flags.power_setup_done))
1024                 pm_idle = acpi_processor_idle;
1025
1026         return result;
1027 }
1028
1029 /* proc interface */
1030
1031 static int acpi_processor_power_seq_show(struct seq_file *seq, void *offset)
1032 {
1033         struct acpi_processor *pr = seq->private;
1034         unsigned int i;
1035
1036
1037         if (!pr)
1038                 goto end;
1039
1040         seq_printf(seq, "active state:            C%zd\n"
1041                    "max_cstate:              C%d\n"
1042                    "bus master activity:     %08x\n"
1043                    "maximum allowed latency: %d usec\n",
1044                    pr->power.state ? pr->power.state - pr->power.states : 0,
1045                    max_cstate, (unsigned)pr->power.bm_activity,
1046                    system_latency_constraint());
1047
1048         seq_puts(seq, "states:\n");
1049
1050         for (i = 1; i <= pr->power.count; i++) {
1051                 seq_printf(seq, "   %cC%d:                  ",
1052                            (&pr->power.states[i] ==
1053                             pr->power.state ? '*' : ' '), i);
1054
1055                 if (!pr->power.states[i].valid) {
1056                         seq_puts(seq, "<not supported>\n");
1057                         continue;
1058                 }
1059
1060                 switch (pr->power.states[i].type) {
1061                 case ACPI_STATE_C1:
1062                         seq_printf(seq, "type[C1] ");
1063                         break;
1064                 case ACPI_STATE_C2:
1065                         seq_printf(seq, "type[C2] ");
1066                         break;
1067                 case ACPI_STATE_C3:
1068                         seq_printf(seq, "type[C3] ");
1069                         break;
1070                 default:
1071                         seq_printf(seq, "type[--] ");
1072                         break;
1073                 }
1074
1075                 if (pr->power.states[i].promotion.state)
1076                         seq_printf(seq, "promotion[C%zd] ",
1077                                    (pr->power.states[i].promotion.state -
1078                                     pr->power.states));
1079                 else
1080                         seq_puts(seq, "promotion[--] ");
1081
1082                 if (pr->power.states[i].demotion.state)
1083                         seq_printf(seq, "demotion[C%zd] ",
1084                                    (pr->power.states[i].demotion.state -
1085                                     pr->power.states));
1086                 else
1087                         seq_puts(seq, "demotion[--] ");
1088
1089                 seq_printf(seq, "latency[%03d] usage[%08d] duration[%020llu]\n",
1090                            pr->power.states[i].latency,
1091                            pr->power.states[i].usage,
1092                            (unsigned long long)pr->power.states[i].time);
1093         }
1094
1095       end:
1096         return 0;
1097 }
1098
1099 static int acpi_processor_power_open_fs(struct inode *inode, struct file *file)
1100 {
1101         return single_open(file, acpi_processor_power_seq_show,
1102                            PDE(inode)->data);
1103 }
1104
1105 static const struct file_operations acpi_processor_power_fops = {
1106         .open = acpi_processor_power_open_fs,
1107         .read = seq_read,
1108         .llseek = seq_lseek,
1109         .release = single_release,
1110 };
1111
1112 #ifdef CONFIG_SMP
1113 static void smp_callback(void *v)
1114 {
1115         /* we already woke the CPU up, nothing more to do */
1116 }
1117
1118 /*
1119  * This function gets called when a part of the kernel has a new latency
1120  * requirement.  This means we need to get all processors out of their C-state,
1121  * and then recalculate a new suitable C-state. Just do a cross-cpu IPI; that
1122  * wakes them all right up.
1123  */
1124 static int acpi_processor_latency_notify(struct notifier_block *b,
1125                 unsigned long l, void *v)
1126 {
1127         smp_call_function(smp_callback, NULL, 0, 1);
1128         return NOTIFY_OK;
1129 }
1130
1131 static struct notifier_block acpi_processor_latency_notifier = {
1132         .notifier_call = acpi_processor_latency_notify,
1133 };
1134 #endif
1135
1136 int __cpuinit acpi_processor_power_init(struct acpi_processor *pr,
1137                               struct acpi_device *device)
1138 {
1139         acpi_status status = 0;
1140         static int first_run;
1141         struct proc_dir_entry *entry = NULL;
1142         unsigned int i;
1143
1144
1145         if (!first_run) {
1146                 dmi_check_system(processor_power_dmi_table);
1147                 if (max_cstate < ACPI_C_STATES_MAX)
1148                         printk(KERN_NOTICE
1149                                "ACPI: processor limited to max C-state %d\n",
1150                                max_cstate);
1151                 first_run++;
1152 #ifdef CONFIG_SMP
1153                 register_latency_notifier(&acpi_processor_latency_notifier);
1154 #endif
1155         }
1156
1157         if (!pr)
1158                 return -EINVAL;
1159
1160         if (acpi_gbl_FADT.cst_control && !nocst) {
1161                 status =
1162                     acpi_os_write_port(acpi_gbl_FADT.smi_command, acpi_gbl_FADT.cst_control, 8);
1163                 if (ACPI_FAILURE(status)) {
1164                         ACPI_EXCEPTION((AE_INFO, status,
1165                                         "Notifying BIOS of _CST ability failed"));
1166                 }
1167         }
1168
1169         acpi_processor_get_power_info(pr);
1170
1171         /*
1172          * Install the idle handler if processor power management is supported.
1173          * Note that we use previously set idle handler will be used on
1174          * platforms that only support C1.
1175          */
1176         if ((pr->flags.power) && (!boot_option_idle_override)) {
1177                 printk(KERN_INFO PREFIX "CPU%d (power states:", pr->id);
1178                 for (i = 1; i <= pr->power.count; i++)
1179                         if (pr->power.states[i].valid)
1180                                 printk(" C%d[C%d]", i,
1181                                        pr->power.states[i].type);
1182                 printk(")\n");
1183
1184                 if (pr->id == 0) {
1185                         pm_idle_save = pm_idle;
1186                         pm_idle = acpi_processor_idle;
1187                 }
1188         }
1189
1190         /* 'power' [R] */
1191         entry = create_proc_entry(ACPI_PROCESSOR_FILE_POWER,
1192                                   S_IRUGO, acpi_device_dir(device));
1193         if (!entry)
1194                 return -EIO;
1195         else {
1196                 entry->proc_fops = &acpi_processor_power_fops;
1197                 entry->data = acpi_driver_data(device);
1198                 entry->owner = THIS_MODULE;
1199         }
1200
1201         pr->flags.power_setup_done = 1;
1202
1203         return 0;
1204 }
1205
1206 int acpi_processor_power_exit(struct acpi_processor *pr,
1207                               struct acpi_device *device)
1208 {
1209
1210         pr->flags.power_setup_done = 0;
1211
1212         if (acpi_device_dir(device))
1213                 remove_proc_entry(ACPI_PROCESSOR_FILE_POWER,
1214                                   acpi_device_dir(device));
1215
1216         /* Unregister the idle handler when processor #0 is removed. */
1217         if (pr->id == 0) {
1218                 pm_idle = pm_idle_save;
1219
1220                 /*
1221                  * We are about to unload the current idle thread pm callback
1222                  * (pm_idle), Wait for all processors to update cached/local
1223                  * copies of pm_idle before proceeding.
1224                  */
1225                 cpu_idle_wait();
1226 #ifdef CONFIG_SMP
1227                 unregister_latency_notifier(&acpi_processor_latency_notifier);
1228 #endif
1229         }
1230
1231         return 0;
1232 }