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
  42 #include <asm/io.h>
  43 #include <asm/uaccess.h>
  44
  45 #include <acpi/acpi_bus.h>
  46 #include <acpi/processor.h>
  47
  48 #define ACPI_PROCESSOR_COMPONENT        0x01000000
  49 #define ACPI_PROCESSOR_CLASS            "processor"
  50 #define ACPI_PROCESSOR_DRIVER_NAME      "ACPI Processor Driver"
  51 #define _COMPONENT              ACPI_PROCESSOR_COMPONENT
  52 ACPI_MODULE_NAME("acpi_processor")
  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,"1SET60WW")}, (void *)1},
 103         { set_max_cstate, "IBM ThinkPad R40e", {
 104           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 105           DMI_MATCH(DMI_BIOS_VERSION,"1SET43WW") }, (void*)1},
 106         { set_max_cstate, "IBM ThinkPad R40e", {
 107           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 108           DMI_MATCH(DMI_BIOS_VERSION,"1SET45WW") }, (void*)1},
 109         { set_max_cstate, "IBM ThinkPad R40e", {
 110           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 111           DMI_MATCH(DMI_BIOS_VERSION,"1SET47WW") }, (void*)1},
 112         { set_max_cstate, "IBM ThinkPad R40e", {
 113           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 114           DMI_MATCH(DMI_BIOS_VERSION,"1SET50WW") }, (void*)1},
 115         { set_max_cstate, "IBM ThinkPad R40e", {
 116           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 117           DMI_MATCH(DMI_BIOS_VERSION,"1SET52WW") }, (void*)1},
 118         { set_max_cstate, "IBM ThinkPad R40e", {
 119           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 120           DMI_MATCH(DMI_BIOS_VERSION,"1SET55WW") }, (void*)1},
 121         { set_max_cstate, "IBM ThinkPad R40e", {
 122           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 123           DMI_MATCH(DMI_BIOS_VERSION,"1SET56WW") }, (void*)1},
 124         { set_max_cstate, "IBM ThinkPad R40e", {
 125           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 126           DMI_MATCH(DMI_BIOS_VERSION,"1SET59WW") }, (void*)1},
 127         { set_max_cstate, "IBM ThinkPad R40e", {
 128           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 129           DMI_MATCH(DMI_BIOS_VERSION,"1SET60WW") }, (void*)1},
 130         { set_max_cstate, "IBM ThinkPad R40e", {
 131           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 132           DMI_MATCH(DMI_BIOS_VERSION,"1SET61WW") }, (void*)1},
 133         { set_max_cstate, "IBM ThinkPad R40e", {
 134           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 135           DMI_MATCH(DMI_BIOS_VERSION,"1SET62WW") }, (void*)1},
 136         { set_max_cstate, "IBM ThinkPad R40e", {
 137           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 138           DMI_MATCH(DMI_BIOS_VERSION,"1SET64WW") }, (void*)1},
 139         { set_max_cstate, "IBM ThinkPad R40e", {
 140           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 141           DMI_MATCH(DMI_BIOS_VERSION,"1SET65WW") }, (void*)1},
 142         { set_max_cstate, "IBM ThinkPad R40e", {
 143           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 144           DMI_MATCH(DMI_BIOS_VERSION,"1SET68WW") }, (void*)1},
 145         { set_max_cstate, "Medion 41700", {
 146           DMI_MATCH(DMI_BIOS_VENDOR,"Phoenix Technologies LTD"),
 147           DMI_MATCH(DMI_BIOS_VERSION,"R01-A1J")}, (void *)1},
 148         { set_max_cstate, "Clevo 5600D", {
 149           DMI_MATCH(DMI_BIOS_VENDOR,"Phoenix Technologies LTD"),
 150           DMI_MATCH(DMI_BIOS_VERSION,"SHE845M0.86C.0013.D.0302131307")},
 151          (void *)2},
 152         {},
 153 };
 154
 155 static inline u32 ticks_elapsed(u32 t1, u32 t2)
 156 {
 157         if (t2 >= t1)
 158                 return (t2 - t1);
 159         else if (!acpi_fadt.tmr_val_ext)
 160                 return (((0x00FFFFFF - t1) + t2) & 0x00FFFFFF);
 161         else
 162                 return ((0xFFFFFFFF - t1) + t2);
 163 }
 164
 165 static void
 166 acpi_processor_power_activate(struct acpi_processor *pr,
 167                               struct acpi_processor_cx *new)
 168 {
 169         struct acpi_processor_cx *old;
 170
 171         if (!pr || !new)
 172                 return;
 173
 174         old = pr->power.state;
 175
 176         if (old)
 177                 old->promotion.count = 0;
 178         new->demotion.count = 0;
 179
 180         /* Cleanup from old state. */
 181         if (old) {
 182                 switch (old->type) {
 183                 case ACPI_STATE_C3:
 184                         /* Disable bus master reload */
 185                         if (new->type != ACPI_STATE_C3 && pr->flags.bm_check)
 186                                 acpi_set_register(ACPI_BITREG_BUS_MASTER_RLD, 0,
 187                                                   ACPI_MTX_DO_NOT_LOCK);
 188                         break;
 189                 }
 190         }
 191
 192         /* Prepare to use new state. */
 193         switch (new->type) {
 194         case ACPI_STATE_C3:
 195                 /* Enable bus master reload */
 196                 if (old->type != ACPI_STATE_C3 && pr->flags.bm_check)
 197                         acpi_set_register(ACPI_BITREG_BUS_MASTER_RLD, 1,
 198                                           ACPI_MTX_DO_NOT_LOCK);
 199                 break;
 200         }
 201
 202         pr->power.state = new;
 203
 204         return;
 205 }
 206
 207 static void acpi_safe_halt(void)
 208 {
 209         current_thread_info()->status &= ~TS_POLLING;
 210         smp_mb__after_clear_bit();
 211         if (!need_resched())
 212                 safe_halt();
 213         current_thread_info()->status |= TS_POLLING;
 214 }
 215
 216 static atomic_t c3_cpu_count;
 217
 218 static void acpi_processor_idle(void)
 219 {
 220         struct acpi_processor *pr = NULL;
 221         struct acpi_processor_cx *cx = NULL;
 222         struct acpi_processor_cx *next_state = NULL;
 223         int sleep_ticks = 0;
 224         u32 t1, t2 = 0;
 225
 226         pr = processors[smp_processor_id()];
 227         if (!pr)
 228                 return;
 229
 230         /*
 231          * Interrupts must be disabled during bus mastering calculations and
 232          * for C2/C3 transitions.
 233          */
 234         local_irq_disable();
 235
 236         /*
 237          * Check whether we truly need to go idle, or should
 238          * reschedule:
 239          */
 240         if (unlikely(need_resched())) {
 241                 local_irq_enable();
 242                 return;
 243         }
 244
 245         cx = pr->power.state;
 246         if (!cx) {
 247                 if (pm_idle_save)
 248                         pm_idle_save();
 249                 else
 250                         acpi_safe_halt();
 251                 return;
 252         }
 253
 254         /*
 255          * Check BM Activity
 256          * -----------------
 257          * Check for bus mastering activity (if required), record, and check
 258          * for demotion.
 259          */
 260         if (pr->flags.bm_check) {
 261                 u32 bm_status = 0;
 262                 unsigned long diff = jiffies - pr->power.bm_check_timestamp;
 263
 264                 if (diff > 31)
 265                         diff = 31;
 266
 267                 pr->power.bm_activity <<= diff;
 268
 269                 acpi_get_register(ACPI_BITREG_BUS_MASTER_STATUS,
 270                                   &bm_status, ACPI_MTX_DO_NOT_LOCK);
 271                 if (bm_status) {
 272                         pr->power.bm_activity |= 0x1;
 273                         acpi_set_register(ACPI_BITREG_BUS_MASTER_STATUS,
 274                                           1, ACPI_MTX_DO_NOT_LOCK);
 275                 }
 276                 /*
 277                  * PIIX4 Erratum #18: Note that BM_STS doesn't always reflect
 278                  * the true state of bus mastering activity; forcing us to
 279                  * manually check the BMIDEA bit of each IDE channel.
 280                  */
 281                 else if (errata.piix4.bmisx) {
 282                         if ((inb_p(errata.piix4.bmisx + 0x02) & 0x01)
 283                             || (inb_p(errata.piix4.bmisx + 0x0A) & 0x01))
 284                                 pr->power.bm_activity |= 0x1;
 285                 }
 286
 287                 pr->power.bm_check_timestamp = jiffies;
 288
 289                 /*
 290                  * If bus mastering is or was active this jiffy, demote
 291                  * to avoid a faulty transition.  Note that the processor
 292                  * won't enter a low-power state during this call (to this
 293                  * function) but should upon the next.
 294                  *
 295                  * TBD: A better policy might be to fallback to the demotion
 296                  *      state (use it for this quantum only) istead of
 297                  *      demoting -- and rely on duration as our sole demotion
 298                  *      qualification.  This may, however, introduce DMA
 299                  *      issues (e.g. floppy DMA transfer overrun/underrun).
 300                  */
 301                 if ((pr->power.bm_activity & 0x1) &&
 302                     cx->demotion.threshold.bm) {
 303                         local_irq_enable();
 304                         next_state = cx->demotion.state;
 305                         goto end;
 306                 }
 307         }
 308
 309 #ifdef CONFIG_HOTPLUG_CPU
 310         /*
 311          * Check for P_LVL2_UP flag before entering C2 and above on
 312          * an SMP system. We do it here instead of doing it at _CST/P_LVL
 313          * detection phase, to work cleanly with logical CPU hotplug.
 314          */
 315         if ((cx->type != ACPI_STATE_C1) && (num_online_cpus() > 1) &&
 316             !pr->flags.has_cst && !acpi_fadt.plvl2_up)
 317                 cx = &pr->power.states[ACPI_STATE_C1];
 318 #endif
 319
 320         /*
 321          * Sleep:
 322          * ------
 323          * Invoke the current Cx state to put the processor to sleep.
 324          */
 325         if (cx->type == ACPI_STATE_C2 || cx->type == ACPI_STATE_C3) {
 326                 current_thread_info()->status &= ~TS_POLLING;
 327                 smp_mb__after_clear_bit();
 328                 if (need_resched()) {
 329                         current_thread_info()->status |= TS_POLLING;
 330                         local_irq_enable();
 331                         return;
 332                 }
 333         }
 334
 335         switch (cx->type) {
 336
 337         case ACPI_STATE_C1:
 338                 /*
 339                  * Invoke C1.
 340                  * Use the appropriate idle routine, the one that would
 341                  * be used without acpi C-states.
 342                  */
 343                 if (pm_idle_save)
 344                         pm_idle_save();
 345                 else
 346                         acpi_safe_halt();
 347
 348                 /*
 349                  * TBD: Can't get time duration while in C1, as resumes
 350                  *      go to an ISR rather than here.  Need to instrument
 351                  *      base interrupt handler.
 352                  */
 353                 sleep_ticks = 0xFFFFFFFF;
 354                 break;
 355
 356         case ACPI_STATE_C2:
 357                 /* Get start time (ticks) */
 358                 t1 = inl(acpi_fadt.xpm_tmr_blk.address);
 359                 /* Invoke C2 */
 360                 inb(cx->address);
 361                 /* Dummy op - must do something useless after P_LVL2 read */
 362                 t2 = inl(acpi_fadt.xpm_tmr_blk.address);
 363                 /* Get end time (ticks) */
 364                 t2 = inl(acpi_fadt.xpm_tmr_blk.address);
 365
 366 #ifdef CONFIG_GENERIC_TIME
 367                 /* TSC halts in C2, so notify users */
 368                 mark_tsc_unstable();
 369 #endif
 370                 /* Re-enable interrupts */
 371                 local_irq_enable();
 372                 current_thread_info()->status |= TS_POLLING;
 373                 /* Compute time (ticks) that we were actually asleep */
 374                 sleep_ticks =
 375                     ticks_elapsed(t1, t2) - cx->latency_ticks - C2_OVERHEAD;
 376                 break;
 377
 378         case ACPI_STATE_C3:
 379
 380                 if (pr->flags.bm_check) {
 381                         if (atomic_inc_return(&c3_cpu_count) ==
 382                             num_online_cpus()) {
 383                                 /*
 384                                  * All CPUs are trying to go to C3
 385                                  * Disable bus master arbitration
 386                                  */
 387                                 acpi_set_register(ACPI_BITREG_ARB_DISABLE, 1,
 388                                                   ACPI_MTX_DO_NOT_LOCK);
 389                         }
 390                 } else {
 391                         /* SMP with no shared cache... Invalidate cache  */
 392                         ACPI_FLUSH_CPU_CACHE();
 393                 }
 394
 395                 /* Get start time (ticks) */
 396                 t1 = inl(acpi_fadt.xpm_tmr_blk.address);
 397                 /* Invoke C3 */
 398                 inb(cx->address);
 399                 /* Dummy op - must do something useless after P_LVL3 read */
 400                 t2 = inl(acpi_fadt.xpm_tmr_blk.address);
 401                 /* Get end time (ticks) */
 402                 t2 = inl(acpi_fadt.xpm_tmr_blk.address);
 403                 if (pr->flags.bm_check) {
 404                         /* Enable bus master arbitration */
 405                         atomic_dec(&c3_cpu_count);
 406                         acpi_set_register(ACPI_BITREG_ARB_DISABLE, 0,
 407                                           ACPI_MTX_DO_NOT_LOCK);
 408                 }
 409
 410 #ifdef CONFIG_GENERIC_TIME
 411                 /* TSC halts in C3, so notify users */
 412                 mark_tsc_unstable();
 413 #endif
 414                 /* Re-enable interrupts */
 415                 local_irq_enable();
 416                 current_thread_info()->status |= TS_POLLING;
 417                 /* Compute time (ticks) that we were actually asleep */
 418                 sleep_ticks =
 419                     ticks_elapsed(t1, t2) - cx->latency_ticks - C3_OVERHEAD;
 420                 break;
 421
 422         default:
 423                 local_irq_enable();
 424                 return;
 425         }
 426         cx->usage++;
 427         if ((cx->type != ACPI_STATE_C1) && (sleep_ticks > 0))
 428                 cx->time += sleep_ticks;
 429
 430         next_state = pr->power.state;
 431
 432 #ifdef CONFIG_HOTPLUG_CPU
 433         /* Don't do promotion/demotion */
 434         if ((cx->type == ACPI_STATE_C1) && (num_online_cpus() > 1) &&
 435             !pr->flags.has_cst && !acpi_fadt.plvl2_up) {
 436                 next_state = cx;
 437                 goto end;
 438         }
 439 #endif
 440
 441         /*
 442          * Promotion?
 443          * ----------
 444          * Track the number of longs (time asleep is greater than threshold)
 445          * and promote when the count threshold is reached.  Note that bus
 446          * mastering activity may prevent promotions.
 447          * Do not promote above max_cstate.
 448          */
 449         if (cx->promotion.state &&
 450             ((cx->promotion.state - pr->power.states) <= max_cstate)) {
 451                 if (sleep_ticks > cx->promotion.threshold.ticks) {
 452                         cx->promotion.count++;
 453                         cx->demotion.count = 0;
 454                         if (cx->promotion.count >=
 455                             cx->promotion.threshold.count) {
 456                                 if (pr->flags.bm_check) {
 457                                         if (!
 458                                             (pr->power.bm_activity & cx->
 459                                              promotion.threshold.bm)) {
 460                                                 next_state =
 461                                                     cx->promotion.state;
 462                                                 goto end;
 463                                         }
 464                                 } else {
 465                                         next_state = cx->promotion.state;
 466                                         goto end;
 467                                 }
 468                         }
 469                 }
 470         }
 471
 472         /*
 473          * Demotion?
 474          * ---------
 475          * Track the number of shorts (time asleep is less than time threshold)
 476          * and demote when the usage threshold is reached.
 477          */
 478         if (cx->demotion.state) {
 479                 if (sleep_ticks < cx->demotion.threshold.ticks) {
 480                         cx->demotion.count++;
 481                         cx->promotion.count = 0;
 482                         if (cx->demotion.count >= cx->demotion.threshold.count) {
 483                                 next_state = cx->demotion.state;
 484                                 goto end;
 485                         }
 486                 }
 487         }
 488
 489       end:
 490         /*
 491          * Demote if current state exceeds max_cstate
 492          */
 493         if ((pr->power.state - pr->power.states) > max_cstate) {
 494                 if (cx->demotion.state)
 495                         next_state = cx->demotion.state;
 496         }
 497
 498         /*
 499          * New Cx State?
 500          * -------------
 501          * If we're going to start using a new Cx state we must clean up
 502          * from the previous and prepare to use the new.
 503          */
 504         if (next_state != pr->power.state)
 505                 acpi_processor_power_activate(pr, next_state);
 506 }
 507
 508 static int acpi_processor_set_power_policy(struct acpi_processor *pr)
 509 {
 510         unsigned int i;
 511         unsigned int state_is_set = 0;
 512         struct acpi_processor_cx *lower = NULL;
 513         struct acpi_processor_cx *higher = NULL;
 514         struct acpi_processor_cx *cx;
 515
 516
 517         if (!pr)
 518                 return -EINVAL;
 519
 520         /*
 521          * This function sets the default Cx state policy (OS idle handler).
 522          * Our scheme is to promote quickly to C2 but more conservatively
 523          * to C3.  We're favoring C2  for its characteristics of low latency
 524          * (quick response), good power savings, and ability to allow bus
 525          * mastering activity.  Note that the Cx state policy is completely
 526          * customizable and can be altered dynamically.
 527          */
 528
 529         /* startup state */
 530         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
 531                 cx = &pr->power.states[i];
 532                 if (!cx->valid)
 533                         continue;
 534
 535                 if (!state_is_set)
 536                         pr->power.state = cx;
 537                 state_is_set++;
 538                 break;
 539         }
 540
 541         if (!state_is_set)
 542                 return -ENODEV;
 543
 544         /* demotion */
 545         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
 546                 cx = &pr->power.states[i];
 547                 if (!cx->valid)
 548                         continue;
 549
 550                 if (lower) {
 551                         cx->demotion.state = lower;
 552                         cx->demotion.threshold.ticks = cx->latency_ticks;
 553                         cx->demotion.threshold.count = 1;
 554                         if (cx->type == ACPI_STATE_C3)
 555                                 cx->demotion.threshold.bm = bm_history;
 556                 }
 557
 558                 lower = cx;
 559         }
 560
 561         /* promotion */
 562         for (i = (ACPI_PROCESSOR_MAX_POWER - 1); i > 0; i--) {
 563                 cx = &pr->power.states[i];
 564                 if (!cx->valid)
 565                         continue;
 566
 567                 if (higher) {
 568                         cx->promotion.state = higher;
 569                         cx->promotion.threshold.ticks = cx->latency_ticks;
 570                         if (cx->type >= ACPI_STATE_C2)
 571                                 cx->promotion.threshold.count = 4;
 572                         else
 573                                 cx->promotion.threshold.count = 10;
 574                         if (higher->type == ACPI_STATE_C3)
 575                                 cx->promotion.threshold.bm = bm_history;
 576                 }
 577
 578                 higher = cx;
 579         }
 580
 581         return 0;
 582 }
 583
 584 static int acpi_processor_get_power_info_fadt(struct acpi_processor *pr)
 585 {
 586
 587         if (!pr)
 588                 return -EINVAL;
 589
 590         if (!pr->pblk)
 591                 return -ENODEV;
 592
 593         /* if info is obtained from pblk/fadt, type equals state */
 594         pr->power.states[ACPI_STATE_C2].type = ACPI_STATE_C2;
 595         pr->power.states[ACPI_STATE_C3].type = ACPI_STATE_C3;
 596
 597 #ifndef CONFIG_HOTPLUG_CPU
 598         /*
 599          * Check for P_LVL2_UP flag before entering C2 and above on
 600          * an SMP system.
 601          */
 602         if ((num_online_cpus() > 1) && !acpi_fadt.plvl2_up)
 603                 return -ENODEV;
 604 #endif
 605
 606         /* determine C2 and C3 address from pblk */
 607         pr->power.states[ACPI_STATE_C2].address = pr->pblk + 4;
 608         pr->power.states[ACPI_STATE_C3].address = pr->pblk + 5;
 609
 610         /* determine latencies from FADT */
 611         pr->power.states[ACPI_STATE_C2].latency = acpi_fadt.plvl2_lat;
 612         pr->power.states[ACPI_STATE_C3].latency = acpi_fadt.plvl3_lat;
 613
 614         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 615                           "lvl2[0x%08x] lvl3[0x%08x]\n",
 616                           pr->power.states[ACPI_STATE_C2].address,
 617                           pr->power.states[ACPI_STATE_C3].address));
 618
 619         return 0;
 620 }
 621
 622 static int acpi_processor_get_power_info_default_c1(struct acpi_processor *pr)
 623 {
 624
 625         /* Zero initialize all the C-states info. */
 626         memset(pr->power.states, 0, sizeof(pr->power.states));
 627
 628         /* set the first C-State to C1 */
 629         pr->power.states[ACPI_STATE_C1].type = ACPI_STATE_C1;
 630
 631         /* the C0 state only exists as a filler in our array,
 632          * and all processors need to support C1 */
 633         pr->power.states[ACPI_STATE_C0].valid = 1;
 634         pr->power.states[ACPI_STATE_C1].valid = 1;
 635
 636         return 0;
 637 }
 638
 639 static int acpi_processor_get_power_info_cst(struct acpi_processor *pr)
 640 {
 641         acpi_status status = 0;
 642         acpi_integer count;
 643         int current_count;
 644         int i;
 645         struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
 646         union acpi_object *cst;
 647
 648
 649         if (nocst)
 650                 return -ENODEV;
 651
 652         current_count = 1;
 653
 654         /* Zero initialize C2 onwards and prepare for fresh CST lookup */
 655         for (i = 2; i < ACPI_PROCESSOR_MAX_POWER; i++)
 656                 memset(&(pr->power.states[i]), 0,
 657                                 sizeof(struct acpi_processor_cx));
 658
 659         status = acpi_evaluate_object(pr->handle, "_CST", NULL, &buffer);
 660         if (ACPI_FAILURE(status)) {
 661                 ACPI_DEBUG_PRINT((ACPI_DB_INFO, "No _CST, giving up\n"));
 662                 return -ENODEV;
 663         }
 664
 665         cst = (union acpi_object *)buffer.pointer;
 666
 667         /* There must be at least 2 elements */
 668         if (!cst || (cst->type != ACPI_TYPE_PACKAGE) || cst->package.count < 2) {
 669                 printk(KERN_ERR PREFIX "not enough elements in _CST\n");
 670                 status = -EFAULT;
 671                 goto end;
 672         }
 673
 674         count = cst->package.elements[0].integer.value;
 675
 676         /* Validate number of power states. */
 677         if (count < 1 || count != cst->package.count - 1) {
 678                 printk(KERN_ERR PREFIX "count given by _CST is not valid\n");
 679                 status = -EFAULT;
 680                 goto end;
 681         }
 682
 683         /* Tell driver that at least _CST is supported. */
 684         pr->flags.has_cst = 1;
 685
 686         for (i = 1; i <= count; i++) {
 687                 union acpi_object *element;
 688                 union acpi_object *obj;
 689                 struct acpi_power_register *reg;
 690                 struct acpi_processor_cx cx;
 691
 692                 memset(&cx, 0, sizeof(cx));
 693
 694                 element = (union acpi_object *)&(cst->package.elements[i]);
 695                 if (element->type != ACPI_TYPE_PACKAGE)
 696                         continue;
 697
 698                 if (element->package.count != 4)
 699                         continue;
 700
 701                 obj = (union acpi_object *)&(element->package.elements[0]);
 702
 703                 if (obj->type != ACPI_TYPE_BUFFER)
 704                         continue;
 705
 706                 reg = (struct acpi_power_register *)obj->buffer.pointer;
 707
 708                 if (reg->space_id != ACPI_ADR_SPACE_SYSTEM_IO &&
 709                     (reg->space_id != ACPI_ADR_SPACE_FIXED_HARDWARE))
 710                         continue;
 711
 712                 cx.address = (reg->space_id == ACPI_ADR_SPACE_FIXED_HARDWARE) ?
 713                     0 : reg->address;
 714
 715                 /* There should be an easy way to extract an integer... */
 716                 obj = (union acpi_object *)&(element->package.elements[1]);
 717                 if (obj->type != ACPI_TYPE_INTEGER)
 718                         continue;
 719
 720                 cx.type = obj->integer.value;
 721
 722                 if ((cx.type != ACPI_STATE_C1) &&
 723                     (reg->space_id != ACPI_ADR_SPACE_SYSTEM_IO))
 724                         continue;
 725
 726                 if ((cx.type < ACPI_STATE_C2) || (cx.type > ACPI_STATE_C3))
 727                         continue;
 728
 729                 obj = (union acpi_object *)&(element->package.elements[2]);
 730                 if (obj->type != ACPI_TYPE_INTEGER)
 731                         continue;
 732
 733                 cx.latency = obj->integer.value;
 734
 735                 obj = (union acpi_object *)&(element->package.elements[3]);
 736                 if (obj->type != ACPI_TYPE_INTEGER)
 737                         continue;
 738
 739                 cx.power = obj->integer.value;
 740
 741                 current_count++;
 742                 memcpy(&(pr->power.states[current_count]), &cx, sizeof(cx));
 743
 744                 /*
 745                  * We support total ACPI_PROCESSOR_MAX_POWER - 1
 746                  * (From 1 through ACPI_PROCESSOR_MAX_POWER - 1)
 747                  */
 748                 if (current_count >= (ACPI_PROCESSOR_MAX_POWER - 1)) {
 749                         printk(KERN_WARNING
 750                                "Limiting number of power states to max (%d)\n",
 751                                ACPI_PROCESSOR_MAX_POWER);
 752                         printk(KERN_WARNING
 753                                "Please increase ACPI_PROCESSOR_MAX_POWER if needed.\n");
 754                         break;
 755                 }
 756         }
 757
 758         ACPI_DEBUG_PRINT((ACPI_DB_INFO, "Found %d power states\n",
 759                           current_count));
 760
 761         /* Validate number of power states discovered */
 762         if (current_count < 2)
 763                 status = -EFAULT;
 764
 765       end:
 766         acpi_os_free(buffer.pointer);
 767
 768         return status;
 769 }
 770
 771 static void acpi_processor_power_verify_c2(struct acpi_processor_cx *cx)
 772 {
 773
 774         if (!cx->address)
 775                 return;
 776
 777         /*
 778          * C2 latency must be less than or equal to 100
 779          * microseconds.
 780          */
 781         else if (cx->latency > ACPI_PROCESSOR_MAX_C2_LATENCY) {
 782                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 783                                   "latency too large [%d]\n", cx->latency));
 784                 return;
 785         }
 786
 787         /*
 788          * Otherwise we've met all of our C2 requirements.
 789          * Normalize the C2 latency to expidite policy
 790          */
 791         cx->valid = 1;
 792         cx->latency_ticks = US_TO_PM_TIMER_TICKS(cx->latency);
 793
 794         return;
 795 }
 796
 797 static void acpi_processor_power_verify_c3(struct acpi_processor *pr,
 798                                            struct acpi_processor_cx *cx)
 799 {
 800         static int bm_check_flag;
 801
 802
 803         if (!cx->address)
 804                 return;
 805
 806         /*
 807          * C3 latency must be less than or equal to 1000
 808          * microseconds.
 809          */
 810         else if (cx->latency > ACPI_PROCESSOR_MAX_C3_LATENCY) {
 811                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 812                                   "latency too large [%d]\n", cx->latency));
 813                 return;
 814         }
 815
 816         /*
 817          * PIIX4 Erratum #18: We don't support C3 when Type-F (fast)
 818          * DMA transfers are used by any ISA device to avoid livelock.
 819          * Note that we could disable Type-F DMA (as recommended by
 820          * the erratum), but this is known to disrupt certain ISA
 821          * devices thus we take the conservative approach.
 822          */
 823         else if (errata.piix4.fdma) {
 824                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 825                                   "C3 not supported on PIIX4 with Type-F DMA\n"));
 826                 return;
 827         }
 828
 829         /* All the logic here assumes flags.bm_check is same across all CPUs */
 830         if (!bm_check_flag) {
 831                 /* Determine whether bm_check is needed based on CPU  */
 832                 acpi_processor_power_init_bm_check(&(pr->flags), pr->id);
 833                 bm_check_flag = pr->flags.bm_check;
 834         } else {
 835                 pr->flags.bm_check = bm_check_flag;
 836         }
 837
 838         if (pr->flags.bm_check) {
 839                 /* bus mastering control is necessary */
 840                 if (!pr->flags.bm_control) {
 841                         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 842                                           "C3 support requires bus mastering control\n"));
 843                         return;
 844                 }
 845         } else {
 846                 /*
 847                  * WBINVD should be set in fadt, for C3 state to be
 848                  * supported on when bm_check is not required.
 849                  */
 850                 if (acpi_fadt.wb_invd != 1) {
 851                         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 852                                           "Cache invalidation should work properly"
 853                                           " for C3 to be enabled on SMP systems\n"));
 854                         return;
 855                 }
 856                 acpi_set_register(ACPI_BITREG_BUS_MASTER_RLD,
 857                                   0, ACPI_MTX_DO_NOT_LOCK);
 858         }
 859
 860         /*
 861          * Otherwise we've met all of our C3 requirements.
 862          * Normalize the C3 latency to expidite policy.  Enable
 863          * checking of bus mastering status (bm_check) so we can
 864          * use this in our C3 policy
 865          */
 866         cx->valid = 1;
 867         cx->latency_ticks = US_TO_PM_TIMER_TICKS(cx->latency);
 868
 869         return;
 870 }
 871
 872 static int acpi_processor_power_verify(struct acpi_processor *pr)
 873 {
 874         unsigned int i;
 875         unsigned int working = 0;
 876
 877 #ifdef ARCH_APICTIMER_STOPS_ON_C3
 878         int timer_broadcast = 0;
 879         cpumask_t mask = cpumask_of_cpu(pr->id);
 880         on_each_cpu(switch_ipi_to_APIC_timer, &mask, 1, 1);
 881 #endif
 882
 883         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
 884                 struct acpi_processor_cx *cx = &pr->power.states[i];
 885
 886                 switch (cx->type) {
 887                 case ACPI_STATE_C1:
 888                         cx->valid = 1;
 889                         break;
 890
 891                 case ACPI_STATE_C2:
 892                         acpi_processor_power_verify_c2(cx);
 893 #ifdef ARCH_APICTIMER_STOPS_ON_C3
 894                         /* Some AMD systems fake C3 as C2, but still
 895                            have timer troubles */
 896                         if (cx->valid &&
 897                                 boot_cpu_data.x86_vendor == X86_VENDOR_AMD)
 898                                 timer_broadcast++;
 899 #endif
 900                         break;
 901
 902                 case ACPI_STATE_C3:
 903                         acpi_processor_power_verify_c3(pr, cx);
 904 #ifdef ARCH_APICTIMER_STOPS_ON_C3
 905                         if (cx->valid)
 906                                 timer_broadcast++;
 907 #endif
 908                         break;
 909                 }
 910
 911                 if (cx->valid)
 912                         working++;
 913         }
 914
 915 #ifdef ARCH_APICTIMER_STOPS_ON_C3
 916         if (timer_broadcast)
 917                 on_each_cpu(switch_APIC_timer_to_ipi, &mask, 1, 1);
 918 #endif
 919
 920         return (working);
 921 }
 922
 923 static int acpi_processor_get_power_info(struct acpi_processor *pr)
 924 {
 925         unsigned int i;
 926         int result;
 927
 928
 929         /* NOTE: the idle thread may not be running while calling
 930          * this function */
 931
 932         /* Adding C1 state */
 933         acpi_processor_get_power_info_default_c1(pr);
 934         result = acpi_processor_get_power_info_cst(pr);
 935         if (result == -ENODEV)
 936                 acpi_processor_get_power_info_fadt(pr);
 937
 938         pr->power.count = acpi_processor_power_verify(pr);
 939
 940         /*
 941          * Set Default Policy
 942          * ------------------
 943          * Now that we know which states are supported, set the default
 944          * policy.  Note that this policy can be changed dynamically
 945          * (e.g. encourage deeper sleeps to conserve battery life when
 946          * not on AC).
 947          */
 948         result = acpi_processor_set_power_policy(pr);
 949         if (result)
 950                 return result;
 951
 952         /*
 953          * if one state of type C2 or C3 is available, mark this
 954          * CPU as being "idle manageable"
 955          */
 956         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
 957                 if (pr->power.states[i].valid) {
 958                         pr->power.count = i;
 959                         if (pr->power.states[i].type >= ACPI_STATE_C2)
 960                                 pr->flags.power = 1;
 961                 }
 962         }
 963
 964         return 0;
 965 }
 966
 967 int acpi_processor_cst_has_changed(struct acpi_processor *pr)
 968 {
 969         int result = 0;
 970
 971
 972         if (!pr)
 973                 return -EINVAL;
 974
 975         if (nocst) {
 976                 return -ENODEV;
 977         }
 978
 979         if (!pr->flags.power_setup_done)
 980                 return -ENODEV;
 981
 982         /* Fall back to the default idle loop */
 983         pm_idle = pm_idle_save;
 984         synchronize_sched();    /* Relies on interrupts forcing exit from idle. */
 985
 986         pr->flags.power = 0;
 987         result = acpi_processor_get_power_info(pr);
 988         if ((pr->flags.power == 1) && (pr->flags.power_setup_done))
 989                 pm_idle = acpi_processor_idle;
 990
 991         return result;
 992 }
 993
 994 /* proc interface */
 995
 996 static int acpi_processor_power_seq_show(struct seq_file *seq, void *offset)
 997 {
 998         struct acpi_processor *pr = (struct acpi_processor *)seq->private;
 999         unsigned int i;
1000
1001
1002         if (!pr)
1003                 goto end;
1004
1005         seq_printf(seq, "active state:            C%zd\n"
1006                    "max_cstate:              C%d\n"
1007                    "bus master activity:     %08x\n",
1008                    pr->power.state ? pr->power.state - pr->power.states : 0,
1009                    max_cstate, (unsigned)pr->power.bm_activity);
1010
1011         seq_puts(seq, "states:\n");
1012
1013         for (i = 1; i <= pr->power.count; i++) {
1014                 seq_printf(seq, "   %cC%d:                  ",
1015                            (&pr->power.states[i] ==
1016                             pr->power.state ? '*' : ' '), i);
1017
1018                 if (!pr->power.states[i].valid) {
1019                         seq_puts(seq, "<not supported>\n");
1020                         continue;
1021                 }
1022
1023                 switch (pr->power.states[i].type) {
1024                 case ACPI_STATE_C1:
1025                         seq_printf(seq, "type[C1] ");
1026                         break;
1027                 case ACPI_STATE_C2:
1028                         seq_printf(seq, "type[C2] ");
1029                         break;
1030                 case ACPI_STATE_C3:
1031                         seq_printf(seq, "type[C3] ");
1032                         break;
1033                 default:
1034                         seq_printf(seq, "type[--] ");
1035                         break;
1036                 }
1037
1038                 if (pr->power.states[i].promotion.state)
1039                         seq_printf(seq, "promotion[C%zd] ",
1040                                    (pr->power.states[i].promotion.state -
1041                                     pr->power.states));
1042                 else
1043                         seq_puts(seq, "promotion[--] ");
1044
1045                 if (pr->power.states[i].demotion.state)
1046                         seq_printf(seq, "demotion[C%zd] ",
1047                                    (pr->power.states[i].demotion.state -
1048                                     pr->power.states));
1049                 else
1050                         seq_puts(seq, "demotion[--] ");
1051
1052                 seq_printf(seq, "latency[%03d] usage[%08d] duration[%020llu]\n",
1053                            pr->power.states[i].latency,
1054                            pr->power.states[i].usage,
1055                            pr->power.states[i].time);
1056         }
1057
1058       end:
1059         return 0;
1060 }
1061
1062 static int acpi_processor_power_open_fs(struct inode *inode, struct file *file)
1063 {
1064         return single_open(file, acpi_processor_power_seq_show,
1065                            PDE(inode)->data);
1066 }
1067
1068 static struct file_operations acpi_processor_power_fops = {
1069         .open = acpi_processor_power_open_fs,
1070         .read = seq_read,
1071         .llseek = seq_lseek,
1072         .release = single_release,
1073 };
1074
1075 int acpi_processor_power_init(struct acpi_processor *pr,
1076                               struct acpi_device *device)
1077 {
1078         acpi_status status = 0;
1079         static int first_run;
1080         struct proc_dir_entry *entry = NULL;
1081         unsigned int i;
1082
1083
1084         if (!first_run) {
1085                 dmi_check_system(processor_power_dmi_table);
1086                 if (max_cstate < ACPI_C_STATES_MAX)
1087                         printk(KERN_NOTICE
1088                                "ACPI: processor limited to max C-state %d\n",
1089                                max_cstate);
1090                 first_run++;
1091         }
1092
1093         if (!pr)
1094                 return -EINVAL;
1095
1096         if (acpi_fadt.cst_cnt && !nocst) {
1097                 status =
1098                     acpi_os_write_port(acpi_fadt.smi_cmd, acpi_fadt.cst_cnt, 8);
1099                 if (ACPI_FAILURE(status)) {
1100                         ACPI_EXCEPTION((AE_INFO, status,
1101                                         "Notifying BIOS of _CST ability failed"));
1102                 }
1103         }
1104
1105         acpi_processor_get_power_info(pr);
1106
1107         /*
1108          * Install the idle handler if processor power management is supported.
1109          * Note that we use previously set idle handler will be used on
1110          * platforms that only support C1.
1111          */
1112         if ((pr->flags.power) && (!boot_option_idle_override)) {
1113                 printk(KERN_INFO PREFIX "CPU%d (power states:", pr->id);
1114                 for (i = 1; i <= pr->power.count; i++)
1115                         if (pr->power.states[i].valid)
1116                                 printk(" C%d[C%d]", i,
1117                                        pr->power.states[i].type);
1118                 printk(")\n");
1119
1120                 if (pr->id == 0) {
1121                         pm_idle_save = pm_idle;
1122                         pm_idle = acpi_processor_idle;
1123                 }
1124         }
1125
1126         /* 'power' [R] */
1127         entry = create_proc_entry(ACPI_PROCESSOR_FILE_POWER,
1128                                   S_IRUGO, acpi_device_dir(device));
1129         if (!entry)
1130                 return -EIO;
1131         else {
1132                 entry->proc_fops = &acpi_processor_power_fops;
1133                 entry->data = acpi_driver_data(device);
1134                 entry->owner = THIS_MODULE;
1135         }
1136
1137         pr->flags.power_setup_done = 1;
1138
1139         return 0;
1140 }
1141
1142 int acpi_processor_power_exit(struct acpi_processor *pr,
1143                               struct acpi_device *device)
1144 {
1145
1146         pr->flags.power_setup_done = 0;
1147
1148         if (acpi_device_dir(device))
1149                 remove_proc_entry(ACPI_PROCESSOR_FILE_POWER,
1150                                   acpi_device_dir(device));
1151
1152         /* Unregister the idle handler when processor #0 is removed. */
1153         if (pr->id == 0) {
1154                 pm_idle = pm_idle_save;
1155
1156                 /*
1157                  * We are about to unload the current idle thread pm callback
1158                  * (pm_idle), Wait for all processors to update cached/local
1159                  * copies of pm_idle before proceeding.
1160                  */
1161                 cpu_idle_wait();
1162         }
1163
1164         return 0;
1165 }