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 ACPI_PROCESSOR_DRIVER_NAME      "ACPI Processor Driver"
  52 #define _COMPONENT              ACPI_PROCESSOR_COMPONENT
  53 ACPI_MODULE_NAME("acpi_processor")
  54 #define ACPI_PROCESSOR_FILE_POWER       "power"
  55 #define US_TO_PM_TIMER_TICKS(t)         ((t * (PM_TIMER_FREQUENCY/1000)) / 1000)
  56 #define C2_OVERHEAD                     4       /* 1us (3.579 ticks per us) */
  57 #define C3_OVERHEAD                     4       /* 1us (3.579 ticks per us) */
  58 static void (*pm_idle_save) (void) __read_mostly;
  59 module_param(max_cstate, uint, 0644);
  60
  61 static unsigned int nocst __read_mostly;
  62 module_param(nocst, uint, 0000);
  63
  64 /*
  65  * bm_history -- bit-mask with a bit per jiffy of bus-master activity
  66  * 1000 HZ: 0xFFFFFFFF: 32 jiffies = 32ms
  67  * 800 HZ: 0xFFFFFFFF: 32 jiffies = 40ms
  68  * 100 HZ: 0x0000000F: 4 jiffies = 40ms
  69  * reduce history for more aggressive entry into C3
  70  */
  71 static unsigned int bm_history __read_mostly =
  72     (HZ >= 800 ? 0xFFFFFFFF : ((1U << (HZ / 25)) - 1));
  73 module_param(bm_history, uint, 0644);
  74 /* --------------------------------------------------------------------------
  75                                 Power Management
  76    -------------------------------------------------------------------------- */
  77
  78 /*
  79  * IBM ThinkPad R40e crashes mysteriously when going into C2 or C3.
  80  * For now disable this. Probably a bug somewhere else.
  81  *
  82  * To skip this limit, boot/load with a large max_cstate limit.
  83  */
  84 static int set_max_cstate(struct dmi_system_id *id)
  85 {
  86         if (max_cstate > ACPI_PROCESSOR_MAX_POWER)
  87                 return 0;
  88
  89         printk(KERN_NOTICE PREFIX "%s detected - limiting to C%ld max_cstate."
  90                " Override with \"processor.max_cstate=%d\"\n", id->ident,
  91                (long)id->driver_data, ACPI_PROCESSOR_MAX_POWER + 1);
  92
  93         max_cstate = (long)id->driver_data;
  94
  95         return 0;
  96 }
  97
  98 /* Actually this shouldn't be __cpuinitdata, would be better to fix the
  99    callers to only run once -AK */
 100 static struct dmi_system_id __cpuinitdata processor_power_dmi_table[] = {
 101         { set_max_cstate, "IBM ThinkPad R40e", {
 102           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 103           DMI_MATCH(DMI_BIOS_VERSION,"1SET70WW")}, (void *)1},
 104         { set_max_cstate, "IBM ThinkPad R40e", {
 105           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 106           DMI_MATCH(DMI_BIOS_VERSION,"1SET60WW")}, (void *)1},
 107         { set_max_cstate, "IBM ThinkPad R40e", {
 108           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 109           DMI_MATCH(DMI_BIOS_VERSION,"1SET43WW") }, (void*)1},
 110         { set_max_cstate, "IBM ThinkPad R40e", {
 111           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 112           DMI_MATCH(DMI_BIOS_VERSION,"1SET45WW") }, (void*)1},
 113         { set_max_cstate, "IBM ThinkPad R40e", {
 114           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 115           DMI_MATCH(DMI_BIOS_VERSION,"1SET47WW") }, (void*)1},
 116         { set_max_cstate, "IBM ThinkPad R40e", {
 117           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 118           DMI_MATCH(DMI_BIOS_VERSION,"1SET50WW") }, (void*)1},
 119         { set_max_cstate, "IBM ThinkPad R40e", {
 120           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 121           DMI_MATCH(DMI_BIOS_VERSION,"1SET52WW") }, (void*)1},
 122         { set_max_cstate, "IBM ThinkPad R40e", {
 123           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 124           DMI_MATCH(DMI_BIOS_VERSION,"1SET55WW") }, (void*)1},
 125         { set_max_cstate, "IBM ThinkPad R40e", {
 126           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 127           DMI_MATCH(DMI_BIOS_VERSION,"1SET56WW") }, (void*)1},
 128         { set_max_cstate, "IBM ThinkPad R40e", {
 129           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 130           DMI_MATCH(DMI_BIOS_VERSION,"1SET59WW") }, (void*)1},
 131         { set_max_cstate, "IBM ThinkPad R40e", {
 132           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 133           DMI_MATCH(DMI_BIOS_VERSION,"1SET60WW") }, (void*)1},
 134         { set_max_cstate, "IBM ThinkPad R40e", {
 135           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 136           DMI_MATCH(DMI_BIOS_VERSION,"1SET61WW") }, (void*)1},
 137         { set_max_cstate, "IBM ThinkPad R40e", {
 138           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 139           DMI_MATCH(DMI_BIOS_VERSION,"1SET62WW") }, (void*)1},
 140         { set_max_cstate, "IBM ThinkPad R40e", {
 141           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 142           DMI_MATCH(DMI_BIOS_VERSION,"1SET64WW") }, (void*)1},
 143         { set_max_cstate, "IBM ThinkPad R40e", {
 144           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 145           DMI_MATCH(DMI_BIOS_VERSION,"1SET65WW") }, (void*)1},
 146         { set_max_cstate, "IBM ThinkPad R40e", {
 147           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 148           DMI_MATCH(DMI_BIOS_VERSION,"1SET68WW") }, (void*)1},
 149         { set_max_cstate, "Medion 41700", {
 150           DMI_MATCH(DMI_BIOS_VENDOR,"Phoenix Technologies LTD"),
 151           DMI_MATCH(DMI_BIOS_VERSION,"R01-A1J")}, (void *)1},
 152         { set_max_cstate, "Clevo 5600D", {
 153           DMI_MATCH(DMI_BIOS_VENDOR,"Phoenix Technologies LTD"),
 154           DMI_MATCH(DMI_BIOS_VERSION,"SHE845M0.86C.0013.D.0302131307")},
 155          (void *)2},
 156         {},
 157 };
 158
 159 static inline u32 ticks_elapsed(u32 t1, u32 t2)
 160 {
 161         if (t2 >= t1)
 162                 return (t2 - t1);
 163         else if (!acpi_fadt.tmr_val_ext)
 164                 return (((0x00FFFFFF - t1) + t2) & 0x00FFFFFF);
 165         else
 166                 return ((0xFFFFFFFF - t1) + t2);
 167 }
 168
 169 static void
 170 acpi_processor_power_activate(struct acpi_processor *pr,
 171                               struct acpi_processor_cx *new)
 172 {
 173         struct acpi_processor_cx *old;
 174
 175         if (!pr || !new)
 176                 return;
 177
 178         old = pr->power.state;
 179
 180         if (old)
 181                 old->promotion.count = 0;
 182         new->demotion.count = 0;
 183
 184         /* Cleanup from old state. */
 185         if (old) {
 186                 switch (old->type) {
 187                 case ACPI_STATE_C3:
 188                         /* Disable bus master reload */
 189                         if (new->type != ACPI_STATE_C3 && pr->flags.bm_check)
 190                                 acpi_set_register(ACPI_BITREG_BUS_MASTER_RLD, 0,
 191                                                   ACPI_MTX_DO_NOT_LOCK);
 192                         break;
 193                 }
 194         }
 195
 196         /* Prepare to use new state. */
 197         switch (new->type) {
 198         case ACPI_STATE_C3:
 199                 /* Enable bus master reload */
 200                 if (old->type != ACPI_STATE_C3 && pr->flags.bm_check)
 201                         acpi_set_register(ACPI_BITREG_BUS_MASTER_RLD, 1,
 202                                           ACPI_MTX_DO_NOT_LOCK);
 203                 break;
 204         }
 205
 206         pr->power.state = new;
 207
 208         return;
 209 }
 210
 211 static void acpi_safe_halt(void)
 212 {
 213         current_thread_info()->status &= ~TS_POLLING;
 214         smp_mb__after_clear_bit();
 215         if (!need_resched())
 216                 safe_halt();
 217         current_thread_info()->status |= TS_POLLING;
 218 }
 219
 220 static atomic_t c3_cpu_count;
 221
 222 static void acpi_processor_idle(void)
 223 {
 224         struct acpi_processor *pr = NULL;
 225         struct acpi_processor_cx *cx = NULL;
 226         struct acpi_processor_cx *next_state = NULL;
 227         int sleep_ticks = 0;
 228         u32 t1, t2 = 0;
 229
 230         pr = processors[smp_processor_id()];
 231         if (!pr)
 232                 return;
 233
 234         /*
 235          * Interrupts must be disabled during bus mastering calculations and
 236          * for C2/C3 transitions.
 237          */
 238         local_irq_disable();
 239
 240         /*
 241          * Check whether we truly need to go idle, or should
 242          * reschedule:
 243          */
 244         if (unlikely(need_resched())) {
 245                 local_irq_enable();
 246                 return;
 247         }
 248
 249         cx = pr->power.state;
 250         if (!cx) {
 251                 if (pm_idle_save)
 252                         pm_idle_save();
 253                 else
 254                         acpi_safe_halt();
 255                 return;
 256         }
 257
 258         /*
 259          * Check BM Activity
 260          * -----------------
 261          * Check for bus mastering activity (if required), record, and check
 262          * for demotion.
 263          */
 264         if (pr->flags.bm_check) {
 265                 u32 bm_status = 0;
 266                 unsigned long diff = jiffies - pr->power.bm_check_timestamp;
 267
 268                 if (diff > 31)
 269                         diff = 31;
 270
 271                 pr->power.bm_activity <<= diff;
 272
 273                 acpi_get_register(ACPI_BITREG_BUS_MASTER_STATUS,
 274                                   &bm_status, ACPI_MTX_DO_NOT_LOCK);
 275                 if (bm_status) {
 276                         pr->power.bm_activity |= 0x1;
 277                         acpi_set_register(ACPI_BITREG_BUS_MASTER_STATUS,
 278                                           1, ACPI_MTX_DO_NOT_LOCK);
 279                 }
 280                 /*
 281                  * PIIX4 Erratum #18: Note that BM_STS doesn't always reflect
 282                  * the true state of bus mastering activity; forcing us to
 283                  * manually check the BMIDEA bit of each IDE channel.
 284                  */
 285                 else if (errata.piix4.bmisx) {
 286                         if ((inb_p(errata.piix4.bmisx + 0x02) & 0x01)
 287                             || (inb_p(errata.piix4.bmisx + 0x0A) & 0x01))
 288                                 pr->power.bm_activity |= 0x1;
 289                 }
 290
 291                 pr->power.bm_check_timestamp = jiffies;
 292
 293                 /*
 294                  * If bus mastering is or was active this jiffy, demote
 295                  * to avoid a faulty transition.  Note that the processor
 296                  * won't enter a low-power state during this call (to this
 297                  * function) but should upon the next.
 298                  *
 299                  * TBD: A better policy might be to fallback to the demotion
 300                  *      state (use it for this quantum only) istead of
 301                  *      demoting -- and rely on duration as our sole demotion
 302                  *      qualification.  This may, however, introduce DMA
 303                  *      issues (e.g. floppy DMA transfer overrun/underrun).
 304                  */
 305                 if ((pr->power.bm_activity & 0x1) &&
 306                     cx->demotion.threshold.bm) {
 307                         local_irq_enable();
 308                         next_state = cx->demotion.state;
 309                         goto end;
 310                 }
 311         }
 312
 313 #ifdef CONFIG_HOTPLUG_CPU
 314         /*
 315          * Check for P_LVL2_UP flag before entering C2 and above on
 316          * an SMP system. We do it here instead of doing it at _CST/P_LVL
 317          * detection phase, to work cleanly with logical CPU hotplug.
 318          */
 319         if ((cx->type != ACPI_STATE_C1) && (num_online_cpus() > 1) &&
 320             !pr->flags.has_cst && !acpi_fadt.plvl2_up)
 321                 cx = &pr->power.states[ACPI_STATE_C1];
 322 #endif
 323
 324         /*
 325          * Sleep:
 326          * ------
 327          * Invoke the current Cx state to put the processor to sleep.
 328          */
 329         if (cx->type == ACPI_STATE_C2 || cx->type == ACPI_STATE_C3) {
 330                 current_thread_info()->status &= ~TS_POLLING;
 331                 smp_mb__after_clear_bit();
 332                 if (need_resched()) {
 333                         current_thread_info()->status |= TS_POLLING;
 334                         local_irq_enable();
 335                         return;
 336                 }
 337         }
 338
 339         switch (cx->type) {
 340
 341         case ACPI_STATE_C1:
 342                 /*
 343                  * Invoke C1.
 344                  * Use the appropriate idle routine, the one that would
 345                  * be used without acpi C-states.
 346                  */
 347                 if (pm_idle_save)
 348                         pm_idle_save();
 349                 else
 350                         acpi_safe_halt();
 351
 352                 /*
 353                  * TBD: Can't get time duration while in C1, as resumes
 354                  *      go to an ISR rather than here.  Need to instrument
 355                  *      base interrupt handler.
 356                  */
 357                 sleep_ticks = 0xFFFFFFFF;
 358                 break;
 359
 360         case ACPI_STATE_C2:
 361                 /* Get start time (ticks) */
 362                 t1 = inl(acpi_fadt.xpm_tmr_blk.address);
 363                 /* Invoke C2 */
 364                 inb(cx->address);
 365                 /* Dummy wait op - must do something useless after P_LVL2 read
 366                    because chipsets cannot guarantee that STPCLK# signal
 367                    gets asserted in time to freeze execution properly. */
 368                 t2 = inl(acpi_fadt.xpm_tmr_blk.address);
 369                 /* Get end time (ticks) */
 370                 t2 = inl(acpi_fadt.xpm_tmr_blk.address);
 371
 372 #ifdef CONFIG_GENERIC_TIME
 373                 /* TSC halts in C2, so notify users */
 374                 mark_tsc_unstable();
 375 #endif
 376                 /* Re-enable interrupts */
 377                 local_irq_enable();
 378                 current_thread_info()->status |= TS_POLLING;
 379                 /* Compute time (ticks) that we were actually asleep */
 380                 sleep_ticks =
 381                     ticks_elapsed(t1, t2) - cx->latency_ticks - C2_OVERHEAD;
 382                 break;
 383
 384         case ACPI_STATE_C3:
 385
 386                 if (pr->flags.bm_check) {
 387                         if (atomic_inc_return(&c3_cpu_count) ==
 388                             num_online_cpus()) {
 389                                 /*
 390                                  * All CPUs are trying to go to C3
 391                                  * Disable bus master arbitration
 392                                  */
 393                                 acpi_set_register(ACPI_BITREG_ARB_DISABLE, 1,
 394                                                   ACPI_MTX_DO_NOT_LOCK);
 395                         }
 396                 } else {
 397                         /* SMP with no shared cache... Invalidate cache  */
 398                         ACPI_FLUSH_CPU_CACHE();
 399                 }
 400
 401                 /* Get start time (ticks) */
 402                 t1 = inl(acpi_fadt.xpm_tmr_blk.address);
 403                 /* Invoke C3 */
 404                 inb(cx->address);
 405                 /* Dummy wait op (see above) */
 406                 t2 = inl(acpi_fadt.xpm_tmr_blk.address);
 407                 /* Get end time (ticks) */
 408                 t2 = inl(acpi_fadt.xpm_tmr_blk.address);
 409                 if (pr->flags.bm_check) {
 410                         /* Enable bus master arbitration */
 411                         atomic_dec(&c3_cpu_count);
 412                         acpi_set_register(ACPI_BITREG_ARB_DISABLE, 0,
 413                                           ACPI_MTX_DO_NOT_LOCK);
 414                 }
 415
 416 #ifdef CONFIG_GENERIC_TIME
 417                 /* TSC halts in C3, so notify users */
 418                 mark_tsc_unstable();
 419 #endif
 420                 /* Re-enable interrupts */
 421                 local_irq_enable();
 422                 current_thread_info()->status |= TS_POLLING;
 423                 /* Compute time (ticks) that we were actually asleep */
 424                 sleep_ticks =
 425                     ticks_elapsed(t1, t2) - cx->latency_ticks - C3_OVERHEAD;
 426                 break;
 427
 428         default:
 429                 local_irq_enable();
 430                 return;
 431         }
 432         cx->usage++;
 433         if ((cx->type != ACPI_STATE_C1) && (sleep_ticks > 0))
 434                 cx->time += sleep_ticks;
 435
 436         next_state = pr->power.state;
 437
 438 #ifdef CONFIG_HOTPLUG_CPU
 439         /* Don't do promotion/demotion */
 440         if ((cx->type == ACPI_STATE_C1) && (num_online_cpus() > 1) &&
 441             !pr->flags.has_cst && !acpi_fadt.plvl2_up) {
 442                 next_state = cx;
 443                 goto end;
 444         }
 445 #endif
 446
 447         /*
 448          * Promotion?
 449          * ----------
 450          * Track the number of longs (time asleep is greater than threshold)
 451          * and promote when the count threshold is reached.  Note that bus
 452          * mastering activity may prevent promotions.
 453          * Do not promote above max_cstate.
 454          */
 455         if (cx->promotion.state &&
 456             ((cx->promotion.state - pr->power.states) <= max_cstate)) {
 457                 if (sleep_ticks > cx->promotion.threshold.ticks &&
 458                   cx->promotion.state->latency <= system_latency_constraint()) {
 459                         cx->promotion.count++;
 460                         cx->demotion.count = 0;
 461                         if (cx->promotion.count >=
 462                             cx->promotion.threshold.count) {
 463                                 if (pr->flags.bm_check) {
 464                                         if (!
 465                                             (pr->power.bm_activity & cx->
 466                                              promotion.threshold.bm)) {
 467                                                 next_state =
 468                                                     cx->promotion.state;
 469                                                 goto end;
 470                                         }
 471                                 } else {
 472                                         next_state = cx->promotion.state;
 473                                         goto end;
 474                                 }
 475                         }
 476                 }
 477         }
 478
 479         /*
 480          * Demotion?
 481          * ---------
 482          * Track the number of shorts (time asleep is less than time threshold)
 483          * and demote when the usage threshold is reached.
 484          */
 485         if (cx->demotion.state) {
 486                 if (sleep_ticks < cx->demotion.threshold.ticks) {
 487                         cx->demotion.count++;
 488                         cx->promotion.count = 0;
 489                         if (cx->demotion.count >= cx->demotion.threshold.count) {
 490                                 next_state = cx->demotion.state;
 491                                 goto end;
 492                         }
 493                 }
 494         }
 495
 496       end:
 497         /*
 498          * Demote if current state exceeds max_cstate
 499          * or if the latency of the current state is unacceptable
 500          */
 501         if ((pr->power.state - pr->power.states) > max_cstate ||
 502                 pr->power.state->latency > system_latency_constraint()) {
 503                 if (cx->demotion.state)
 504                         next_state = cx->demotion.state;
 505         }
 506
 507         /*
 508          * New Cx State?
 509          * -------------
 510          * If we're going to start using a new Cx state we must clean up
 511          * from the previous and prepare to use the new.
 512          */
 513         if (next_state != pr->power.state)
 514                 acpi_processor_power_activate(pr, next_state);
 515 }
 516
 517 static int acpi_processor_set_power_policy(struct acpi_processor *pr)
 518 {
 519         unsigned int i;
 520         unsigned int state_is_set = 0;
 521         struct acpi_processor_cx *lower = NULL;
 522         struct acpi_processor_cx *higher = NULL;
 523         struct acpi_processor_cx *cx;
 524
 525
 526         if (!pr)
 527                 return -EINVAL;
 528
 529         /*
 530          * This function sets the default Cx state policy (OS idle handler).
 531          * Our scheme is to promote quickly to C2 but more conservatively
 532          * to C3.  We're favoring C2  for its characteristics of low latency
 533          * (quick response), good power savings, and ability to allow bus
 534          * mastering activity.  Note that the Cx state policy is completely
 535          * customizable and can be altered dynamically.
 536          */
 537
 538         /* startup state */
 539         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
 540                 cx = &pr->power.states[i];
 541                 if (!cx->valid)
 542                         continue;
 543
 544                 if (!state_is_set)
 545                         pr->power.state = cx;
 546                 state_is_set++;
 547                 break;
 548         }
 549
 550         if (!state_is_set)
 551                 return -ENODEV;
 552
 553         /* demotion */
 554         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
 555                 cx = &pr->power.states[i];
 556                 if (!cx->valid)
 557                         continue;
 558
 559                 if (lower) {
 560                         cx->demotion.state = lower;
 561                         cx->demotion.threshold.ticks = cx->latency_ticks;
 562                         cx->demotion.threshold.count = 1;
 563                         if (cx->type == ACPI_STATE_C3)
 564                                 cx->demotion.threshold.bm = bm_history;
 565                 }
 566
 567                 lower = cx;
 568         }
 569
 570         /* promotion */
 571         for (i = (ACPI_PROCESSOR_MAX_POWER - 1); i > 0; i--) {
 572                 cx = &pr->power.states[i];
 573                 if (!cx->valid)
 574                         continue;
 575
 576                 if (higher) {
 577                         cx->promotion.state = higher;
 578                         cx->promotion.threshold.ticks = cx->latency_ticks;
 579                         if (cx->type >= ACPI_STATE_C2)
 580                                 cx->promotion.threshold.count = 4;
 581                         else
 582                                 cx->promotion.threshold.count = 10;
 583                         if (higher->type == ACPI_STATE_C3)
 584                                 cx->promotion.threshold.bm = bm_history;
 585                 }
 586
 587                 higher = cx;
 588         }
 589
 590         return 0;
 591 }
 592
 593 static int acpi_processor_get_power_info_fadt(struct acpi_processor *pr)
 594 {
 595
 596         if (!pr)
 597                 return -EINVAL;
 598
 599         if (!pr->pblk)
 600                 return -ENODEV;
 601
 602         /* if info is obtained from pblk/fadt, type equals state */
 603         pr->power.states[ACPI_STATE_C2].type = ACPI_STATE_C2;
 604         pr->power.states[ACPI_STATE_C3].type = ACPI_STATE_C3;
 605
 606 #ifndef CONFIG_HOTPLUG_CPU
 607         /*
 608          * Check for P_LVL2_UP flag before entering C2 and above on
 609          * an SMP system.
 610          */
 611         if ((num_online_cpus() > 1) && !acpi_fadt.plvl2_up)
 612                 return -ENODEV;
 613 #endif
 614
 615         /* determine C2 and C3 address from pblk */
 616         pr->power.states[ACPI_STATE_C2].address = pr->pblk + 4;
 617         pr->power.states[ACPI_STATE_C3].address = pr->pblk + 5;
 618
 619         /* determine latencies from FADT */
 620         pr->power.states[ACPI_STATE_C2].latency = acpi_fadt.plvl2_lat;
 621         pr->power.states[ACPI_STATE_C3].latency = acpi_fadt.plvl3_lat;
 622
 623         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 624                           "lvl2[0x%08x] lvl3[0x%08x]\n",
 625                           pr->power.states[ACPI_STATE_C2].address,
 626                           pr->power.states[ACPI_STATE_C3].address));
 627
 628         return 0;
 629 }
 630
 631 static int acpi_processor_get_power_info_default_c1(struct acpi_processor *pr)
 632 {
 633
 634         /* Zero initialize all the C-states info. */
 635         memset(pr->power.states, 0, sizeof(pr->power.states));
 636
 637         /* set the first C-State to C1 */
 638         pr->power.states[ACPI_STATE_C1].type = ACPI_STATE_C1;
 639
 640         /* the C0 state only exists as a filler in our array,
 641          * and all processors need to support C1 */
 642         pr->power.states[ACPI_STATE_C0].valid = 1;
 643         pr->power.states[ACPI_STATE_C1].valid = 1;
 644
 645         return 0;
 646 }
 647
 648 static int acpi_processor_get_power_info_cst(struct acpi_processor *pr)
 649 {
 650         acpi_status status = 0;
 651         acpi_integer count;
 652         int current_count;
 653         int i;
 654         struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
 655         union acpi_object *cst;
 656
 657
 658         if (nocst)
 659                 return -ENODEV;
 660
 661         current_count = 1;
 662
 663         /* Zero initialize C2 onwards and prepare for fresh CST lookup */
 664         for (i = 2; i < ACPI_PROCESSOR_MAX_POWER; i++)
 665                 memset(&(pr->power.states[i]), 0,
 666                                 sizeof(struct acpi_processor_cx));
 667
 668         status = acpi_evaluate_object(pr->handle, "_CST", NULL, &buffer);
 669         if (ACPI_FAILURE(status)) {
 670                 ACPI_DEBUG_PRINT((ACPI_DB_INFO, "No _CST, giving up\n"));
 671                 return -ENODEV;
 672         }
 673
 674         cst = buffer.pointer;
 675
 676         /* There must be at least 2 elements */
 677         if (!cst || (cst->type != ACPI_TYPE_PACKAGE) || cst->package.count < 2) {
 678                 printk(KERN_ERR PREFIX "not enough elements in _CST\n");
 679                 status = -EFAULT;
 680                 goto end;
 681         }
 682
 683         count = cst->package.elements[0].integer.value;
 684
 685         /* Validate number of power states. */
 686         if (count < 1 || count != cst->package.count - 1) {
 687                 printk(KERN_ERR PREFIX "count given by _CST is not valid\n");
 688                 status = -EFAULT;
 689                 goto end;
 690         }
 691
 692         /* Tell driver that at least _CST is supported. */
 693         pr->flags.has_cst = 1;
 694
 695         for (i = 1; i <= count; i++) {
 696                 union acpi_object *element;
 697                 union acpi_object *obj;
 698                 struct acpi_power_register *reg;
 699                 struct acpi_processor_cx cx;
 700
 701                 memset(&cx, 0, sizeof(cx));
 702
 703                 element = &(cst->package.elements[i]);
 704                 if (element->type != ACPI_TYPE_PACKAGE)
 705                         continue;
 706
 707                 if (element->package.count != 4)
 708                         continue;
 709
 710                 obj = &(element->package.elements[0]);
 711
 712                 if (obj->type != ACPI_TYPE_BUFFER)
 713                         continue;
 714
 715                 reg = (struct acpi_power_register *)obj->buffer.pointer;
 716
 717                 if (reg->space_id != ACPI_ADR_SPACE_SYSTEM_IO &&
 718                     (reg->space_id != ACPI_ADR_SPACE_FIXED_HARDWARE))
 719                         continue;
 720
 721                 cx.address = (reg->space_id == ACPI_ADR_SPACE_FIXED_HARDWARE) ?
 722                     0 : reg->address;
 723
 724                 /* There should be an easy way to extract an integer... */
 725                 obj = &(element->package.elements[1]);
 726                 if (obj->type != ACPI_TYPE_INTEGER)
 727                         continue;
 728
 729                 cx.type = obj->integer.value;
 730
 731                 if ((cx.type != ACPI_STATE_C1) &&
 732                     (reg->space_id != ACPI_ADR_SPACE_SYSTEM_IO))
 733                         continue;
 734
 735                 if ((cx.type < ACPI_STATE_C2) || (cx.type > ACPI_STATE_C3))
 736                         continue;
 737
 738                 obj = &(element->package.elements[2]);
 739                 if (obj->type != ACPI_TYPE_INTEGER)
 740                         continue;
 741
 742                 cx.latency = obj->integer.value;
 743
 744                 obj = &(element->package.elements[3]);
 745                 if (obj->type != ACPI_TYPE_INTEGER)
 746                         continue;
 747
 748                 cx.power = obj->integer.value;
 749
 750                 current_count++;
 751                 memcpy(&(pr->power.states[current_count]), &cx, sizeof(cx));
 752
 753                 /*
 754                  * We support total ACPI_PROCESSOR_MAX_POWER - 1
 755                  * (From 1 through ACPI_PROCESSOR_MAX_POWER - 1)
 756                  */
 757                 if (current_count >= (ACPI_PROCESSOR_MAX_POWER - 1)) {
 758                         printk(KERN_WARNING
 759                                "Limiting number of power states to max (%d)\n",
 760                                ACPI_PROCESSOR_MAX_POWER);
 761                         printk(KERN_WARNING
 762                                "Please increase ACPI_PROCESSOR_MAX_POWER if needed.\n");
 763                         break;
 764                 }
 765         }
 766
 767         ACPI_DEBUG_PRINT((ACPI_DB_INFO, "Found %d power states\n",
 768                           current_count));
 769
 770         /* Validate number of power states discovered */
 771         if (current_count < 2)
 772                 status = -EFAULT;
 773
 774       end:
 775         kfree(buffer.pointer);
 776
 777         return status;
 778 }
 779
 780 static void acpi_processor_power_verify_c2(struct acpi_processor_cx *cx)
 781 {
 782
 783         if (!cx->address)
 784                 return;
 785
 786         /*
 787          * C2 latency must be less than or equal to 100
 788          * microseconds.
 789          */
 790         else if (cx->latency > ACPI_PROCESSOR_MAX_C2_LATENCY) {
 791                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 792                                   "latency too large [%d]\n", cx->latency));
 793                 return;
 794         }
 795
 796         /*
 797          * Otherwise we've met all of our C2 requirements.
 798          * Normalize the C2 latency to expidite policy
 799          */
 800         cx->valid = 1;
 801         cx->latency_ticks = US_TO_PM_TIMER_TICKS(cx->latency);
 802
 803         return;
 804 }
 805
 806 static void acpi_processor_power_verify_c3(struct acpi_processor *pr,
 807                                            struct acpi_processor_cx *cx)
 808 {
 809         static int bm_check_flag;
 810
 811
 812         if (!cx->address)
 813                 return;
 814
 815         /*
 816          * C3 latency must be less than or equal to 1000
 817          * microseconds.
 818          */
 819         else if (cx->latency > ACPI_PROCESSOR_MAX_C3_LATENCY) {
 820                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 821                                   "latency too large [%d]\n", cx->latency));
 822                 return;
 823         }
 824
 825         /*
 826          * PIIX4 Erratum #18: We don't support C3 when Type-F (fast)
 827          * DMA transfers are used by any ISA device to avoid livelock.
 828          * Note that we could disable Type-F DMA (as recommended by
 829          * the erratum), but this is known to disrupt certain ISA
 830          * devices thus we take the conservative approach.
 831          */
 832         else if (errata.piix4.fdma) {
 833                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 834                                   "C3 not supported on PIIX4 with Type-F DMA\n"));
 835                 return;
 836         }
 837
 838         /* All the logic here assumes flags.bm_check is same across all CPUs */
 839         if (!bm_check_flag) {
 840                 /* Determine whether bm_check is needed based on CPU  */
 841                 acpi_processor_power_init_bm_check(&(pr->flags), pr->id);
 842                 bm_check_flag = pr->flags.bm_check;
 843         } else {
 844                 pr->flags.bm_check = bm_check_flag;
 845         }
 846
 847         if (pr->flags.bm_check) {
 848                 /* bus mastering control is necessary */
 849                 if (!pr->flags.bm_control) {
 850                         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 851                                           "C3 support requires bus mastering control\n"));
 852                         return;
 853                 }
 854         } else {
 855                 /*
 856                  * WBINVD should be set in fadt, for C3 state to be
 857                  * supported on when bm_check is not required.
 858                  */
 859                 if (acpi_fadt.wb_invd != 1) {
 860                         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 861                                           "Cache invalidation should work properly"
 862                                           " for C3 to be enabled on SMP systems\n"));
 863                         return;
 864                 }
 865                 acpi_set_register(ACPI_BITREG_BUS_MASTER_RLD,
 866                                   0, ACPI_MTX_DO_NOT_LOCK);
 867         }
 868
 869         /*
 870          * Otherwise we've met all of our C3 requirements.
 871          * Normalize the C3 latency to expidite policy.  Enable
 872          * checking of bus mastering status (bm_check) so we can
 873          * use this in our C3 policy
 874          */
 875         cx->valid = 1;
 876         cx->latency_ticks = US_TO_PM_TIMER_TICKS(cx->latency);
 877
 878         return;
 879 }
 880
 881 static int acpi_processor_power_verify(struct acpi_processor *pr)
 882 {
 883         unsigned int i;
 884         unsigned int working = 0;
 885
 886 #ifdef ARCH_APICTIMER_STOPS_ON_C3
 887         int timer_broadcast = 0;
 888         cpumask_t mask = cpumask_of_cpu(pr->id);
 889         on_each_cpu(switch_ipi_to_APIC_timer, &mask, 1, 1);
 890 #endif
 891
 892         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
 893                 struct acpi_processor_cx *cx = &pr->power.states[i];
 894
 895                 switch (cx->type) {
 896                 case ACPI_STATE_C1:
 897                         cx->valid = 1;
 898                         break;
 899
 900                 case ACPI_STATE_C2:
 901                         acpi_processor_power_verify_c2(cx);
 902 #ifdef ARCH_APICTIMER_STOPS_ON_C3
 903                         /* Some AMD systems fake C3 as C2, but still
 904                            have timer troubles */
 905                         if (cx->valid &&
 906                                 boot_cpu_data.x86_vendor == X86_VENDOR_AMD)
 907                                 timer_broadcast++;
 908 #endif
 909                         break;
 910
 911                 case ACPI_STATE_C3:
 912                         acpi_processor_power_verify_c3(pr, cx);
 913 #ifdef ARCH_APICTIMER_STOPS_ON_C3
 914                         if (cx->valid)
 915                                 timer_broadcast++;
 916 #endif
 917                         break;
 918                 }
 919
 920                 if (cx->valid)
 921                         working++;
 922         }
 923
 924 #ifdef ARCH_APICTIMER_STOPS_ON_C3
 925         if (timer_broadcast)
 926                 on_each_cpu(switch_APIC_timer_to_ipi, &mask, 1, 1);
 927 #endif
 928
 929         return (working);
 930 }
 931
 932 static int acpi_processor_get_power_info(struct acpi_processor *pr)
 933 {
 934         unsigned int i;
 935         int result;
 936
 937
 938         /* NOTE: the idle thread may not be running while calling
 939          * this function */
 940
 941         /* Adding C1 state */
 942         acpi_processor_get_power_info_default_c1(pr);
 943         result = acpi_processor_get_power_info_cst(pr);
 944         if (result == -ENODEV)
 945                 acpi_processor_get_power_info_fadt(pr);
 946
 947         pr->power.count = acpi_processor_power_verify(pr);
 948
 949         /*
 950          * Set Default Policy
 951          * ------------------
 952          * Now that we know which states are supported, set the default
 953          * policy.  Note that this policy can be changed dynamically
 954          * (e.g. encourage deeper sleeps to conserve battery life when
 955          * not on AC).
 956          */
 957         result = acpi_processor_set_power_policy(pr);
 958         if (result)
 959                 return result;
 960
 961         /*
 962          * if one state of type C2 or C3 is available, mark this
 963          * CPU as being "idle manageable"
 964          */
 965         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
 966                 if (pr->power.states[i].valid) {
 967                         pr->power.count = i;
 968                         if (pr->power.states[i].type >= ACPI_STATE_C2)
 969                                 pr->flags.power = 1;
 970                 }
 971         }
 972
 973         return 0;
 974 }
 975
 976 int acpi_processor_cst_has_changed(struct acpi_processor *pr)
 977 {
 978         int result = 0;
 979
 980
 981         if (!pr)
 982                 return -EINVAL;
 983
 984         if (nocst) {
 985                 return -ENODEV;
 986         }
 987
 988         if (!pr->flags.power_setup_done)
 989                 return -ENODEV;
 990
 991         /* Fall back to the default idle loop */
 992         pm_idle = pm_idle_save;
 993         synchronize_sched();    /* Relies on interrupts forcing exit from idle. */
 994
 995         pr->flags.power = 0;
 996         result = acpi_processor_get_power_info(pr);
 997         if ((pr->flags.power == 1) && (pr->flags.power_setup_done))
 998                 pm_idle = acpi_processor_idle;
 999
1000         return result;
1001 }
1002
1003 /* proc interface */
1004
1005 static int acpi_processor_power_seq_show(struct seq_file *seq, void *offset)
1006 {
1007         struct acpi_processor *pr = seq->private;
1008         unsigned int i;
1009
1010
1011         if (!pr)
1012                 goto end;
1013
1014         seq_printf(seq, "active state:            C%zd\n"
1015                    "max_cstate:              C%d\n"
1016                    "bus master activity:     %08x\n"
1017                    "maximum allowed latency: %d usec\n",
1018                    pr->power.state ? pr->power.state - pr->power.states : 0,
1019                    max_cstate, (unsigned)pr->power.bm_activity,
1020                    system_latency_constraint());
1021
1022         seq_puts(seq, "states:\n");
1023
1024         for (i = 1; i <= pr->power.count; i++) {
1025                 seq_printf(seq, "   %cC%d:                  ",
1026                            (&pr->power.states[i] ==
1027                             pr->power.state ? '*' : ' '), i);
1028
1029                 if (!pr->power.states[i].valid) {
1030                         seq_puts(seq, "<not supported>\n");
1031                         continue;
1032                 }
1033
1034                 switch (pr->power.states[i].type) {
1035                 case ACPI_STATE_C1:
1036                         seq_printf(seq, "type[C1] ");
1037                         break;
1038                 case ACPI_STATE_C2:
1039                         seq_printf(seq, "type[C2] ");
1040                         break;
1041                 case ACPI_STATE_C3:
1042                         seq_printf(seq, "type[C3] ");
1043                         break;
1044                 default:
1045                         seq_printf(seq, "type[--] ");
1046                         break;
1047                 }
1048
1049                 if (pr->power.states[i].promotion.state)
1050                         seq_printf(seq, "promotion[C%zd] ",
1051                                    (pr->power.states[i].promotion.state -
1052                                     pr->power.states));
1053                 else
1054                         seq_puts(seq, "promotion[--] ");
1055
1056                 if (pr->power.states[i].demotion.state)
1057                         seq_printf(seq, "demotion[C%zd] ",
1058                                    (pr->power.states[i].demotion.state -
1059                                     pr->power.states));
1060                 else
1061                         seq_puts(seq, "demotion[--] ");
1062
1063                 seq_printf(seq, "latency[%03d] usage[%08d] duration[%020llu]\n",
1064                            pr->power.states[i].latency,
1065                            pr->power.states[i].usage,
1066                            pr->power.states[i].time);
1067         }
1068
1069       end:
1070         return 0;
1071 }
1072
1073 static int acpi_processor_power_open_fs(struct inode *inode, struct file *file)
1074 {
1075         return single_open(file, acpi_processor_power_seq_show,
1076                            PDE(inode)->data);
1077 }
1078
1079 static const struct file_operations acpi_processor_power_fops = {
1080         .open = acpi_processor_power_open_fs,
1081         .read = seq_read,
1082         .llseek = seq_lseek,
1083         .release = single_release,
1084 };
1085
1086 static void smp_callback(void *v)
1087 {
1088         /* we already woke the CPU up, nothing more to do */
1089 }
1090
1091 /*
1092  * This function gets called when a part of the kernel has a new latency
1093  * requirement.  This means we need to get all processors out of their C-state,
1094  * and then recalculate a new suitable C-state. Just do a cross-cpu IPI; that
1095  * wakes them all right up.
1096  */
1097 static int acpi_processor_latency_notify(struct notifier_block *b,
1098                 unsigned long l, void *v)
1099 {
1100         smp_call_function(smp_callback, NULL, 0, 1);
1101         return NOTIFY_OK;
1102 }
1103
1104 static struct notifier_block acpi_processor_latency_notifier = {
1105         .notifier_call = acpi_processor_latency_notify,
1106 };
1107
1108 int acpi_processor_power_init(struct acpi_processor *pr,
1109                               struct acpi_device *device)
1110 {
1111         acpi_status status = 0;
1112         static int first_run;
1113         struct proc_dir_entry *entry = NULL;
1114         unsigned int i;
1115
1116
1117         if (!first_run) {
1118                 dmi_check_system(processor_power_dmi_table);
1119                 if (max_cstate < ACPI_C_STATES_MAX)
1120                         printk(KERN_NOTICE
1121                                "ACPI: processor limited to max C-state %d\n",
1122                                max_cstate);
1123                 first_run++;
1124                 register_latency_notifier(&acpi_processor_latency_notifier);
1125         }
1126
1127         if (!pr)
1128                 return -EINVAL;
1129
1130         if (acpi_fadt.cst_cnt && !nocst) {
1131                 status =
1132                     acpi_os_write_port(acpi_fadt.smi_cmd, acpi_fadt.cst_cnt, 8);
1133                 if (ACPI_FAILURE(status)) {
1134                         ACPI_EXCEPTION((AE_INFO, status,
1135                                         "Notifying BIOS of _CST ability failed"));
1136                 }
1137         }
1138
1139         acpi_processor_get_power_info(pr);
1140
1141         /*
1142          * Install the idle handler if processor power management is supported.
1143          * Note that we use previously set idle handler will be used on
1144          * platforms that only support C1.
1145          */
1146         if ((pr->flags.power) && (!boot_option_idle_override)) {
1147                 printk(KERN_INFO PREFIX "CPU%d (power states:", pr->id);
1148                 for (i = 1; i <= pr->power.count; i++)
1149                         if (pr->power.states[i].valid)
1150                                 printk(" C%d[C%d]", i,
1151                                        pr->power.states[i].type);
1152                 printk(")\n");
1153
1154                 if (pr->id == 0) {
1155                         pm_idle_save = pm_idle;
1156                         pm_idle = acpi_processor_idle;
1157                 }
1158         }
1159
1160         /* 'power' [R] */
1161         entry = create_proc_entry(ACPI_PROCESSOR_FILE_POWER,
1162                                   S_IRUGO, acpi_device_dir(device));
1163         if (!entry)
1164                 return -EIO;
1165         else {
1166                 entry->proc_fops = &acpi_processor_power_fops;
1167                 entry->data = acpi_driver_data(device);
1168                 entry->owner = THIS_MODULE;
1169         }
1170
1171         pr->flags.power_setup_done = 1;
1172
1173         return 0;
1174 }
1175
1176 int acpi_processor_power_exit(struct acpi_processor *pr,
1177                               struct acpi_device *device)
1178 {
1179
1180         pr->flags.power_setup_done = 0;
1181
1182         if (acpi_device_dir(device))
1183                 remove_proc_entry(ACPI_PROCESSOR_FILE_POWER,
1184                                   acpi_device_dir(device));
1185
1186         /* Unregister the idle handler when processor #0 is removed. */
1187         if (pr->id == 0) {
1188                 pm_idle = pm_idle_save;
1189
1190                 /*
1191                  * We are about to unload the current idle thread pm callback
1192                  * (pm_idle), Wait for all processors to update cached/local
1193                  * copies of pm_idle before proceeding.
1194                  */
1195                 cpu_idle_wait();
1196                 unregister_latency_notifier(&acpi_processor_latency_notifier);
1197         }
1198
1199         return 0;
1200 }