drivers/acpi/processor_idle.c

   1 /*
   2  * processor_idle - idle state submodule to the ACPI processor driver
   3  *
   4  *  Copyright (C) 2001, 2002 Andy Grover <andrew.grover@intel.com>
   5  *  Copyright (C) 2001, 2002 Paul Diefenbaugh <paul.s.diefenbaugh@intel.com>
   6  *  Copyright (C) 2004, 2005 Dominik Brodowski <linux@brodo.de>
   7  *  Copyright (C) 2004  Anil S Keshavamurthy <anil.s.keshavamurthy@intel.com>
   8  *                      - Added processor hotplug support
   9  *  Copyright (C) 2005  Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>
  10  *                      - Added support for C3 on SMP
  11  *
  12  * ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
  13  *
  14  *  This program is free software; you can redistribute it and/or modify
  15  *  it under the terms of the GNU General Public License as published by
  16  *  the Free Software Foundation; either version 2 of the License, or (at
  17  *  your option) any later version.
  18  *
  19  *  This program is distributed in the hope that it will be useful, but
  20  *  WITHOUT ANY WARRANTY; without even the implied warranty of
  21  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  22  *  General Public License for more details.
  23  *
  24  *  You should have received a copy of the GNU General Public License along
  25  *  with this program; if not, write to the Free Software Foundation, Inc.,
  26  *  59 Temple Place, Suite 330, Boston, MA 02111-1307 USA.
  27  *
  28  * ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
  29  */
  30
  31 #include <linux/kernel.h>
  32 #include <linux/module.h>
  33 #include <linux/init.h>
  34 #include <linux/cpufreq.h>
  35 #include <linux/proc_fs.h>
  36 #include <linux/seq_file.h>
  37 #include <linux/acpi.h>
  38 #include <linux/dmi.h>
  39 #include <linux/moduleparam.h>
  40 #include <linux/sched.h>        /* need_resched() */
  41 #include <linux/latency.h>
  42
  43 /*
  44  * Include the apic definitions for x86 to have the APIC timer related defines
  45  * available also for UP (on SMP it gets magically included via linux/smp.h).
  46  */
  47 #ifdef CONFIG_X86
  48 #include <asm/apic.h>
  49 #endif
  50
  51 #include <asm/io.h>
  52 #include <asm/uaccess.h>
  53
  54 #include <acpi/acpi_bus.h>
  55 #include <acpi/processor.h>
  56
  57 #define ACPI_PROCESSOR_COMPONENT        0x01000000
  58 #define ACPI_PROCESSOR_CLASS            "processor"
  59 #define ACPI_PROCESSOR_DRIVER_NAME      "ACPI Processor Driver"
  60 #define _COMPONENT              ACPI_PROCESSOR_COMPONENT
  61 ACPI_MODULE_NAME("acpi_processor")
  62 #define ACPI_PROCESSOR_FILE_POWER       "power"
  63 #define US_TO_PM_TIMER_TICKS(t)         ((t * (PM_TIMER_FREQUENCY/1000)) / 1000)
  64 #define C2_OVERHEAD                     4       /* 1us (3.579 ticks per us) */
  65 #define C3_OVERHEAD                     4       /* 1us (3.579 ticks per us) */
  66 static void (*pm_idle_save) (void) __read_mostly;
  67 module_param(max_cstate, uint, 0644);
  68
  69 static unsigned int nocst __read_mostly;
  70 module_param(nocst, uint, 0000);
  71
  72 /*
  73  * bm_history -- bit-mask with a bit per jiffy of bus-master activity
  74  * 1000 HZ: 0xFFFFFFFF: 32 jiffies = 32ms
  75  * 800 HZ: 0xFFFFFFFF: 32 jiffies = 40ms
  76  * 100 HZ: 0x0000000F: 4 jiffies = 40ms
  77  * reduce history for more aggressive entry into C3
  78  */
  79 static unsigned int bm_history __read_mostly =
  80     (HZ >= 800 ? 0xFFFFFFFF : ((1U << (HZ / 25)) - 1));
  81 module_param(bm_history, uint, 0644);
  82 /* --------------------------------------------------------------------------
  83                                 Power Management
  84    -------------------------------------------------------------------------- */
  85
  86 /*
  87  * IBM ThinkPad R40e crashes mysteriously when going into C2 or C3.
  88  * For now disable this. Probably a bug somewhere else.
  89  *
  90  * To skip this limit, boot/load with a large max_cstate limit.
  91  */
  92 static int set_max_cstate(struct dmi_system_id *id)
  93 {
  94         if (max_cstate > ACPI_PROCESSOR_MAX_POWER)
  95                 return 0;
  96
  97         printk(KERN_NOTICE PREFIX "%s detected - limiting to C%ld max_cstate."
  98                " Override with \"processor.max_cstate=%d\"\n", id->ident,
  99                (long)id->driver_data, ACPI_PROCESSOR_MAX_POWER + 1);
 100
 101         max_cstate = (long)id->driver_data;
 102
 103         return 0;
 104 }
 105
 106 /* Actually this shouldn't be __cpuinitdata, would be better to fix the
 107    callers to only run once -AK */
 108 static struct dmi_system_id __cpuinitdata processor_power_dmi_table[] = {
 109         { set_max_cstate, "IBM ThinkPad R40e", {
 110           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 111           DMI_MATCH(DMI_BIOS_VERSION,"1SET70WW")}, (void *)1},
 112         { set_max_cstate, "IBM ThinkPad R40e", {
 113           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 114           DMI_MATCH(DMI_BIOS_VERSION,"1SET60WW")}, (void *)1},
 115         { set_max_cstate, "IBM ThinkPad R40e", {
 116           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 117           DMI_MATCH(DMI_BIOS_VERSION,"1SET43WW") }, (void*)1},
 118         { set_max_cstate, "IBM ThinkPad R40e", {
 119           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 120           DMI_MATCH(DMI_BIOS_VERSION,"1SET45WW") }, (void*)1},
 121         { set_max_cstate, "IBM ThinkPad R40e", {
 122           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 123           DMI_MATCH(DMI_BIOS_VERSION,"1SET47WW") }, (void*)1},
 124         { set_max_cstate, "IBM ThinkPad R40e", {
 125           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 126           DMI_MATCH(DMI_BIOS_VERSION,"1SET50WW") }, (void*)1},
 127         { set_max_cstate, "IBM ThinkPad R40e", {
 128           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 129           DMI_MATCH(DMI_BIOS_VERSION,"1SET52WW") }, (void*)1},
 130         { set_max_cstate, "IBM ThinkPad R40e", {
 131           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 132           DMI_MATCH(DMI_BIOS_VERSION,"1SET55WW") }, (void*)1},
 133         { set_max_cstate, "IBM ThinkPad R40e", {
 134           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 135           DMI_MATCH(DMI_BIOS_VERSION,"1SET56WW") }, (void*)1},
 136         { set_max_cstate, "IBM ThinkPad R40e", {
 137           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 138           DMI_MATCH(DMI_BIOS_VERSION,"1SET59WW") }, (void*)1},
 139         { set_max_cstate, "IBM ThinkPad R40e", {
 140           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 141           DMI_MATCH(DMI_BIOS_VERSION,"1SET60WW") }, (void*)1},
 142         { set_max_cstate, "IBM ThinkPad R40e", {
 143           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 144           DMI_MATCH(DMI_BIOS_VERSION,"1SET61WW") }, (void*)1},
 145         { set_max_cstate, "IBM ThinkPad R40e", {
 146           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 147           DMI_MATCH(DMI_BIOS_VERSION,"1SET62WW") }, (void*)1},
 148         { set_max_cstate, "IBM ThinkPad R40e", {
 149           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 150           DMI_MATCH(DMI_BIOS_VERSION,"1SET64WW") }, (void*)1},
 151         { set_max_cstate, "IBM ThinkPad R40e", {
 152           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 153           DMI_MATCH(DMI_BIOS_VERSION,"1SET65WW") }, (void*)1},
 154         { set_max_cstate, "IBM ThinkPad R40e", {
 155           DMI_MATCH(DMI_BIOS_VENDOR,"IBM"),
 156           DMI_MATCH(DMI_BIOS_VERSION,"1SET68WW") }, (void*)1},
 157         { set_max_cstate, "Medion 41700", {
 158           DMI_MATCH(DMI_BIOS_VENDOR,"Phoenix Technologies LTD"),
 159           DMI_MATCH(DMI_BIOS_VERSION,"R01-A1J")}, (void *)1},
 160         { set_max_cstate, "Clevo 5600D", {
 161           DMI_MATCH(DMI_BIOS_VENDOR,"Phoenix Technologies LTD"),
 162           DMI_MATCH(DMI_BIOS_VERSION,"SHE845M0.86C.0013.D.0302131307")},
 163          (void *)2},
 164         {},
 165 };
 166
 167 static inline u32 ticks_elapsed(u32 t1, u32 t2)
 168 {
 169         if (t2 >= t1)
 170                 return (t2 - t1);
 171         else if (!(acpi_gbl_FADT.flags & ACPI_FADT_32BIT_TIMER))
 172                 return (((0x00FFFFFF - t1) + t2) & 0x00FFFFFF);
 173         else
 174                 return ((0xFFFFFFFF - t1) + t2);
 175 }
 176
 177 static void
 178 acpi_processor_power_activate(struct acpi_processor *pr,
 179                               struct acpi_processor_cx *new)
 180 {
 181         struct acpi_processor_cx *old;
 182
 183         if (!pr || !new)
 184                 return;
 185
 186         old = pr->power.state;
 187
 188         if (old)
 189                 old->promotion.count = 0;
 190         new->demotion.count = 0;
 191
 192         /* Cleanup from old state. */
 193         if (old) {
 194                 switch (old->type) {
 195                 case ACPI_STATE_C3:
 196                         /* Disable bus master reload */
 197                         if (new->type != ACPI_STATE_C3 && pr->flags.bm_check)
 198                                 acpi_set_register(ACPI_BITREG_BUS_MASTER_RLD, 0);
 199                         break;
 200                 }
 201         }
 202
 203         /* Prepare to use new state. */
 204         switch (new->type) {
 205         case ACPI_STATE_C3:
 206                 /* Enable bus master reload */
 207                 if (old->type != ACPI_STATE_C3 && pr->flags.bm_check)
 208                         acpi_set_register(ACPI_BITREG_BUS_MASTER_RLD, 1);
 209                 break;
 210         }
 211
 212         pr->power.state = new;
 213
 214         return;
 215 }
 216
 217 static void acpi_safe_halt(void)
 218 {
 219         current_thread_info()->status &= ~TS_POLLING;
 220         /*
 221          * TS_POLLING-cleared state must be visible before we
 222          * test NEED_RESCHED:
 223          */
 224         smp_mb();
 225         if (!need_resched())
 226                 safe_halt();
 227         current_thread_info()->status |= TS_POLLING;
 228 }
 229
 230 static atomic_t c3_cpu_count;
 231
 232 /* Common C-state entry for C2, C3, .. */
 233 static void acpi_cstate_enter(struct acpi_processor_cx *cstate)
 234 {
 235         if (cstate->space_id == ACPI_CSTATE_FFH) {
 236                 /* Call into architectural FFH based C-state */
 237                 acpi_processor_ffh_cstate_enter(cstate);
 238         } else {
 239                 int unused;
 240                 /* IO port based C-state */
 241                 inb(cstate->address);
 242                 /* Dummy wait op - must do something useless after P_LVL2 read
 243                    because chipsets cannot guarantee that STPCLK# signal
 244                    gets asserted in time to freeze execution properly. */
 245                 unused = inl(acpi_gbl_FADT.xpm_timer_block.address);
 246         }
 247 }
 248
 249 static void acpi_processor_idle(void)
 250 {
 251         struct acpi_processor *pr = NULL;
 252         struct acpi_processor_cx *cx = NULL;
 253         struct acpi_processor_cx *next_state = NULL;
 254         int sleep_ticks = 0;
 255         u32 t1, t2 = 0;
 256
 257         pr = processors[smp_processor_id()];
 258         if (!pr)
 259                 return;
 260
 261         /*
 262          * Interrupts must be disabled during bus mastering calculations and
 263          * for C2/C3 transitions.
 264          */
 265         local_irq_disable();
 266
 267         /*
 268          * Check whether we truly need to go idle, or should
 269          * reschedule:
 270          */
 271         if (unlikely(need_resched())) {
 272                 local_irq_enable();
 273                 return;
 274         }
 275
 276         cx = pr->power.state;
 277         if (!cx) {
 278                 if (pm_idle_save)
 279                         pm_idle_save();
 280                 else
 281                         acpi_safe_halt();
 282                 return;
 283         }
 284
 285         /*
 286          * Check BM Activity
 287          * -----------------
 288          * Check for bus mastering activity (if required), record, and check
 289          * for demotion.
 290          */
 291         if (pr->flags.bm_check) {
 292                 u32 bm_status = 0;
 293                 unsigned long diff = jiffies - pr->power.bm_check_timestamp;
 294
 295                 if (diff > 31)
 296                         diff = 31;
 297
 298                 pr->power.bm_activity <<= diff;
 299
 300                 acpi_get_register(ACPI_BITREG_BUS_MASTER_STATUS, &bm_status);
 301                 if (bm_status) {
 302                         pr->power.bm_activity |= 0x1;
 303                         acpi_set_register(ACPI_BITREG_BUS_MASTER_STATUS, 1);
 304                 }
 305                 /*
 306                  * PIIX4 Erratum #18: Note that BM_STS doesn't always reflect
 307                  * the true state of bus mastering activity; forcing us to
 308                  * manually check the BMIDEA bit of each IDE channel.
 309                  */
 310                 else if (errata.piix4.bmisx) {
 311                         if ((inb_p(errata.piix4.bmisx + 0x02) & 0x01)
 312                             || (inb_p(errata.piix4.bmisx + 0x0A) & 0x01))
 313                                 pr->power.bm_activity |= 0x1;
 314                 }
 315
 316                 pr->power.bm_check_timestamp = jiffies;
 317
 318                 /*
 319                  * If bus mastering is or was active this jiffy, demote
 320                  * to avoid a faulty transition.  Note that the processor
 321                  * won't enter a low-power state during this call (to this
 322                  * function) but should upon the next.
 323                  *
 324                  * TBD: A better policy might be to fallback to the demotion
 325                  *      state (use it for this quantum only) istead of
 326                  *      demoting -- and rely on duration as our sole demotion
 327                  *      qualification.  This may, however, introduce DMA
 328                  *      issues (e.g. floppy DMA transfer overrun/underrun).
 329                  */
 330                 if ((pr->power.bm_activity & 0x1) &&
 331                     cx->demotion.threshold.bm) {
 332                         local_irq_enable();
 333                         next_state = cx->demotion.state;
 334                         goto end;
 335                 }
 336         }
 337
 338 #ifdef CONFIG_HOTPLUG_CPU
 339         /*
 340          * Check for P_LVL2_UP flag before entering C2 and above on
 341          * an SMP system. We do it here instead of doing it at _CST/P_LVL
 342          * detection phase, to work cleanly with logical CPU hotplug.
 343          */
 344         if ((cx->type != ACPI_STATE_C1) && (num_online_cpus() > 1) &&
 345             !pr->flags.has_cst && !(acpi_gbl_FADT.flags & ACPI_FADT_C2_MP_SUPPORTED))
 346                 cx = &pr->power.states[ACPI_STATE_C1];
 347 #endif
 348
 349         /*
 350          * Sleep:
 351          * ------
 352          * Invoke the current Cx state to put the processor to sleep.
 353          */
 354         if (cx->type == ACPI_STATE_C2 || cx->type == ACPI_STATE_C3) {
 355                 current_thread_info()->status &= ~TS_POLLING;
 356                 /*
 357                  * TS_POLLING-cleared state must be visible before we
 358                  * test NEED_RESCHED:
 359                  */
 360                 smp_mb();
 361                 if (need_resched()) {
 362                         current_thread_info()->status |= TS_POLLING;
 363                         local_irq_enable();
 364                         return;
 365                 }
 366         }
 367
 368         switch (cx->type) {
 369
 370         case ACPI_STATE_C1:
 371                 /*
 372                  * Invoke C1.
 373                  * Use the appropriate idle routine, the one that would
 374                  * be used without acpi C-states.
 375                  */
 376                 if (pm_idle_save)
 377                         pm_idle_save();
 378                 else
 379                         acpi_safe_halt();
 380
 381                 /*
 382                  * TBD: Can't get time duration while in C1, as resumes
 383                  *      go to an ISR rather than here.  Need to instrument
 384                  *      base interrupt handler.
 385                  */
 386                 sleep_ticks = 0xFFFFFFFF;
 387                 break;
 388
 389         case ACPI_STATE_C2:
 390                 /* Get start time (ticks) */
 391                 t1 = inl(acpi_gbl_FADT.xpm_timer_block.address);
 392                 /* Invoke C2 */
 393                 acpi_cstate_enter(cx);
 394                 /* Get end time (ticks) */
 395                 t2 = inl(acpi_gbl_FADT.xpm_timer_block.address);
 396
 397 #ifdef CONFIG_GENERIC_TIME
 398                 /* TSC halts in C2, so notify users */
 399                 mark_tsc_unstable();
 400 #endif
 401                 /* Re-enable interrupts */
 402                 local_irq_enable();
 403                 current_thread_info()->status |= TS_POLLING;
 404                 /* Compute time (ticks) that we were actually asleep */
 405                 sleep_ticks =
 406                     ticks_elapsed(t1, t2) - cx->latency_ticks - C2_OVERHEAD;
 407                 break;
 408
 409         case ACPI_STATE_C3:
 410
 411                 if (pr->flags.bm_check) {
 412                         if (atomic_inc_return(&c3_cpu_count) ==
 413                             num_online_cpus()) {
 414                                 /*
 415                                  * All CPUs are trying to go to C3
 416                                  * Disable bus master arbitration
 417                                  */
 418                                 acpi_set_register(ACPI_BITREG_ARB_DISABLE, 1);
 419                         }
 420                 } else {
 421                         /* SMP with no shared cache... Invalidate cache  */
 422                         ACPI_FLUSH_CPU_CACHE();
 423                 }
 424
 425                 /* Get start time (ticks) */
 426                 t1 = inl(acpi_gbl_FADT.xpm_timer_block.address);
 427                 /* Invoke C3 */
 428                 acpi_cstate_enter(cx);
 429                 /* Get end time (ticks) */
 430                 t2 = inl(acpi_gbl_FADT.xpm_timer_block.address);
 431                 if (pr->flags.bm_check) {
 432                         /* Enable bus master arbitration */
 433                         atomic_dec(&c3_cpu_count);
 434                         acpi_set_register(ACPI_BITREG_ARB_DISABLE, 0);
 435                 }
 436
 437 #ifdef CONFIG_GENERIC_TIME
 438                 /* TSC halts in C3, so notify users */
 439                 mark_tsc_unstable();
 440 #endif
 441                 /* Re-enable interrupts */
 442                 local_irq_enable();
 443                 current_thread_info()->status |= TS_POLLING;
 444                 /* Compute time (ticks) that we were actually asleep */
 445                 sleep_ticks =
 446                     ticks_elapsed(t1, t2) - cx->latency_ticks - C3_OVERHEAD;
 447                 break;
 448
 449         default:
 450                 local_irq_enable();
 451                 return;
 452         }
 453         cx->usage++;
 454         if ((cx->type != ACPI_STATE_C1) && (sleep_ticks > 0))
 455                 cx->time += sleep_ticks;
 456
 457         next_state = pr->power.state;
 458
 459 #ifdef CONFIG_HOTPLUG_CPU
 460         /* Don't do promotion/demotion */
 461         if ((cx->type == ACPI_STATE_C1) && (num_online_cpus() > 1) &&
 462             !pr->flags.has_cst && !(acpi_gbl_FADT.flags & ACPI_FADT_C2_MP_SUPPORTED)) {
 463                 next_state = cx;
 464                 goto end;
 465         }
 466 #endif
 467
 468         /*
 469          * Promotion?
 470          * ----------
 471          * Track the number of longs (time asleep is greater than threshold)
 472          * and promote when the count threshold is reached.  Note that bus
 473          * mastering activity may prevent promotions.
 474          * Do not promote above max_cstate.
 475          */
 476         if (cx->promotion.state &&
 477             ((cx->promotion.state - pr->power.states) <= max_cstate)) {
 478                 if (sleep_ticks > cx->promotion.threshold.ticks &&
 479                   cx->promotion.state->latency <= system_latency_constraint()) {
 480                         cx->promotion.count++;
 481                         cx->demotion.count = 0;
 482                         if (cx->promotion.count >=
 483                             cx->promotion.threshold.count) {
 484                                 if (pr->flags.bm_check) {
 485                                         if (!
 486                                             (pr->power.bm_activity & cx->
 487                                              promotion.threshold.bm)) {
 488                                                 next_state =
 489                                                     cx->promotion.state;
 490                                                 goto end;
 491                                         }
 492                                 } else {
 493                                         next_state = cx->promotion.state;
 494                                         goto end;
 495                                 }
 496                         }
 497                 }
 498         }
 499
 500         /*
 501          * Demotion?
 502          * ---------
 503          * Track the number of shorts (time asleep is less than time threshold)
 504          * and demote when the usage threshold is reached.
 505          */
 506         if (cx->demotion.state) {
 507                 if (sleep_ticks < cx->demotion.threshold.ticks) {
 508                         cx->demotion.count++;
 509                         cx->promotion.count = 0;
 510                         if (cx->demotion.count >= cx->demotion.threshold.count) {
 511                                 next_state = cx->demotion.state;
 512                                 goto end;
 513                         }
 514                 }
 515         }
 516
 517       end:
 518         /*
 519          * Demote if current state exceeds max_cstate
 520          * or if the latency of the current state is unacceptable
 521          */
 522         if ((pr->power.state - pr->power.states) > max_cstate ||
 523                 pr->power.state->latency > system_latency_constraint()) {
 524                 if (cx->demotion.state)
 525                         next_state = cx->demotion.state;
 526         }
 527
 528         /*
 529          * New Cx State?
 530          * -------------
 531          * If we're going to start using a new Cx state we must clean up
 532          * from the previous and prepare to use the new.
 533          */
 534         if (next_state != pr->power.state)
 535                 acpi_processor_power_activate(pr, next_state);
 536 }
 537
 538 static int acpi_processor_set_power_policy(struct acpi_processor *pr)
 539 {
 540         unsigned int i;
 541         unsigned int state_is_set = 0;
 542         struct acpi_processor_cx *lower = NULL;
 543         struct acpi_processor_cx *higher = NULL;
 544         struct acpi_processor_cx *cx;
 545
 546
 547         if (!pr)
 548                 return -EINVAL;
 549
 550         /*
 551          * This function sets the default Cx state policy (OS idle handler).
 552          * Our scheme is to promote quickly to C2 but more conservatively
 553          * to C3.  We're favoring C2  for its characteristics of low latency
 554          * (quick response), good power savings, and ability to allow bus
 555          * mastering activity.  Note that the Cx state policy is completely
 556          * customizable and can be altered dynamically.
 557          */
 558
 559         /* startup state */
 560         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
 561                 cx = &pr->power.states[i];
 562                 if (!cx->valid)
 563                         continue;
 564
 565                 if (!state_is_set)
 566                         pr->power.state = cx;
 567                 state_is_set++;
 568                 break;
 569         }
 570
 571         if (!state_is_set)
 572                 return -ENODEV;
 573
 574         /* demotion */
 575         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
 576                 cx = &pr->power.states[i];
 577                 if (!cx->valid)
 578                         continue;
 579
 580                 if (lower) {
 581                         cx->demotion.state = lower;
 582                         cx->demotion.threshold.ticks = cx->latency_ticks;
 583                         cx->demotion.threshold.count = 1;
 584                         if (cx->type == ACPI_STATE_C3)
 585                                 cx->demotion.threshold.bm = bm_history;
 586                 }
 587
 588                 lower = cx;
 589         }
 590
 591         /* promotion */
 592         for (i = (ACPI_PROCESSOR_MAX_POWER - 1); i > 0; i--) {
 593                 cx = &pr->power.states[i];
 594                 if (!cx->valid)
 595                         continue;
 596
 597                 if (higher) {
 598                         cx->promotion.state = higher;
 599                         cx->promotion.threshold.ticks = cx->latency_ticks;
 600                         if (cx->type >= ACPI_STATE_C2)
 601                                 cx->promotion.threshold.count = 4;
 602                         else
 603                                 cx->promotion.threshold.count = 10;
 604                         if (higher->type == ACPI_STATE_C3)
 605                                 cx->promotion.threshold.bm = bm_history;
 606                 }
 607
 608                 higher = cx;
 609         }
 610
 611         return 0;
 612 }
 613
 614 static int acpi_processor_get_power_info_fadt(struct acpi_processor *pr)
 615 {
 616
 617         if (!pr)
 618                 return -EINVAL;
 619
 620         if (!pr->pblk)
 621                 return -ENODEV;
 622
 623         /* if info is obtained from pblk/fadt, type equals state */
 624         pr->power.states[ACPI_STATE_C2].type = ACPI_STATE_C2;
 625         pr->power.states[ACPI_STATE_C3].type = ACPI_STATE_C3;
 626
 627 #ifndef CONFIG_HOTPLUG_CPU
 628         /*
 629          * Check for P_LVL2_UP flag before entering C2 and above on
 630          * an SMP system.
 631          */
 632         if ((num_online_cpus() > 1) &&
 633             !(acpi_gbl_FADT.flags & ACPI_FADT_C2_MP_SUPPORTED))
 634                 return -ENODEV;
 635 #endif
 636
 637         /* determine C2 and C3 address from pblk */
 638         pr->power.states[ACPI_STATE_C2].address = pr->pblk + 4;
 639         pr->power.states[ACPI_STATE_C3].address = pr->pblk + 5;
 640
 641         /* determine latencies from FADT */
 642         pr->power.states[ACPI_STATE_C2].latency = acpi_gbl_FADT.C2latency;
 643         pr->power.states[ACPI_STATE_C3].latency = acpi_gbl_FADT.C3latency;
 644
 645         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 646                           "lvl2[0x%08x] lvl3[0x%08x]\n",
 647                           pr->power.states[ACPI_STATE_C2].address,
 648                           pr->power.states[ACPI_STATE_C3].address));
 649
 650         return 0;
 651 }
 652
 653 static int acpi_processor_get_power_info_default(struct acpi_processor *pr)
 654 {
 655         if (!pr->power.states[ACPI_STATE_C1].valid) {
 656                 /* set the first C-State to C1 */
 657                 /* all processors need to support C1 */
 658                 pr->power.states[ACPI_STATE_C1].type = ACPI_STATE_C1;
 659                 pr->power.states[ACPI_STATE_C1].valid = 1;
 660         }
 661         /* the C0 state only exists as a filler in our array */
 662         pr->power.states[ACPI_STATE_C0].valid = 1;
 663         return 0;
 664 }
 665
 666 static int acpi_processor_get_power_info_cst(struct acpi_processor *pr)
 667 {
 668         acpi_status status = 0;
 669         acpi_integer count;
 670         int current_count;
 671         int i;
 672         struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
 673         union acpi_object *cst;
 674
 675
 676         if (nocst)
 677                 return -ENODEV;
 678
 679         current_count = 0;
 680
 681         status = acpi_evaluate_object(pr->handle, "_CST", NULL, &buffer);
 682         if (ACPI_FAILURE(status)) {
 683                 ACPI_DEBUG_PRINT((ACPI_DB_INFO, "No _CST, giving up\n"));
 684                 return -ENODEV;
 685         }
 686
 687         cst = buffer.pointer;
 688
 689         /* There must be at least 2 elements */
 690         if (!cst || (cst->type != ACPI_TYPE_PACKAGE) || cst->package.count < 2) {
 691                 printk(KERN_ERR PREFIX "not enough elements in _CST\n");
 692                 status = -EFAULT;
 693                 goto end;
 694         }
 695
 696         count = cst->package.elements[0].integer.value;
 697
 698         /* Validate number of power states. */
 699         if (count < 1 || count != cst->package.count - 1) {
 700                 printk(KERN_ERR PREFIX "count given by _CST is not valid\n");
 701                 status = -EFAULT;
 702                 goto end;
 703         }
 704
 705         /* Tell driver that at least _CST is supported. */
 706         pr->flags.has_cst = 1;
 707
 708         for (i = 1; i <= count; i++) {
 709                 union acpi_object *element;
 710                 union acpi_object *obj;
 711                 struct acpi_power_register *reg;
 712                 struct acpi_processor_cx cx;
 713
 714                 memset(&cx, 0, sizeof(cx));
 715
 716                 element = &(cst->package.elements[i]);
 717                 if (element->type != ACPI_TYPE_PACKAGE)
 718                         continue;
 719
 720                 if (element->package.count != 4)
 721                         continue;
 722
 723                 obj = &(element->package.elements[0]);
 724
 725                 if (obj->type != ACPI_TYPE_BUFFER)
 726                         continue;
 727
 728                 reg = (struct acpi_power_register *)obj->buffer.pointer;
 729
 730                 if (reg->space_id != ACPI_ADR_SPACE_SYSTEM_IO &&
 731                     (reg->space_id != ACPI_ADR_SPACE_FIXED_HARDWARE))
 732                         continue;
 733
 734                 /* There should be an easy way to extract an integer... */
 735                 obj = &(element->package.elements[1]);
 736                 if (obj->type != ACPI_TYPE_INTEGER)
 737                         continue;
 738
 739                 cx.type = obj->integer.value;
 740                 /*
 741                  * Some buggy BIOSes won't list C1 in _CST -
 742                  * Let acpi_processor_get_power_info_default() handle them later
 743                  */
 744                 if (i == 1 && cx.type != ACPI_STATE_C1)
 745                         current_count++;
 746
 747                 cx.address = reg->address;
 748                 cx.index = current_count + 1;
 749
 750                 cx.space_id = ACPI_CSTATE_SYSTEMIO;
 751                 if (reg->space_id == ACPI_ADR_SPACE_FIXED_HARDWARE) {
 752                         if (acpi_processor_ffh_cstate_probe
 753                                         (pr->id, &cx, reg) == 0) {
 754                                 cx.space_id = ACPI_CSTATE_FFH;
 755                         } else if (cx.type != ACPI_STATE_C1) {
 756                                 /*
 757                                  * C1 is a special case where FIXED_HARDWARE
 758                                  * can be handled in non-MWAIT way as well.
 759                                  * In that case, save this _CST entry info.
 760                                  * That is, we retain space_id of SYSTEM_IO for
 761                                  * halt based C1.
 762                                  * Otherwise, ignore this info and continue.
 763                                  */
 764                                 continue;
 765                         }
 766                 }
 767
 768                 obj = &(element->package.elements[2]);
 769                 if (obj->type != ACPI_TYPE_INTEGER)
 770                         continue;
 771
 772                 cx.latency = obj->integer.value;
 773
 774                 obj = &(element->package.elements[3]);
 775                 if (obj->type != ACPI_TYPE_INTEGER)
 776                         continue;
 777
 778                 cx.power = obj->integer.value;
 779
 780                 current_count++;
 781                 memcpy(&(pr->power.states[current_count]), &cx, sizeof(cx));
 782
 783                 /*
 784                  * We support total ACPI_PROCESSOR_MAX_POWER - 1
 785                  * (From 1 through ACPI_PROCESSOR_MAX_POWER - 1)
 786                  */
 787                 if (current_count >= (ACPI_PROCESSOR_MAX_POWER - 1)) {
 788                         printk(KERN_WARNING
 789                                "Limiting number of power states to max (%d)\n",
 790                                ACPI_PROCESSOR_MAX_POWER);
 791                         printk(KERN_WARNING
 792                                "Please increase ACPI_PROCESSOR_MAX_POWER if needed.\n");
 793                         break;
 794                 }
 795         }
 796
 797         ACPI_DEBUG_PRINT((ACPI_DB_INFO, "Found %d power states\n",
 798                           current_count));
 799
 800         /* Validate number of power states discovered */
 801         if (current_count < 2)
 802                 status = -EFAULT;
 803
 804       end:
 805         kfree(buffer.pointer);
 806
 807         return status;
 808 }
 809
 810 static void acpi_processor_power_verify_c2(struct acpi_processor_cx *cx)
 811 {
 812
 813         if (!cx->address)
 814                 return;
 815
 816         /*
 817          * C2 latency must be less than or equal to 100
 818          * microseconds.
 819          */
 820         else if (cx->latency > ACPI_PROCESSOR_MAX_C2_LATENCY) {
 821                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 822                                   "latency too large [%d]\n", cx->latency));
 823                 return;
 824         }
 825
 826         /*
 827          * Otherwise we've met all of our C2 requirements.
 828          * Normalize the C2 latency to expidite policy
 829          */
 830         cx->valid = 1;
 831         cx->latency_ticks = US_TO_PM_TIMER_TICKS(cx->latency);
 832
 833         return;
 834 }
 835
 836 static void acpi_processor_power_verify_c3(struct acpi_processor *pr,
 837                                            struct acpi_processor_cx *cx)
 838 {
 839         static int bm_check_flag;
 840
 841
 842         if (!cx->address)
 843                 return;
 844
 845         /*
 846          * C3 latency must be less than or equal to 1000
 847          * microseconds.
 848          */
 849         else if (cx->latency > ACPI_PROCESSOR_MAX_C3_LATENCY) {
 850                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 851                                   "latency too large [%d]\n", cx->latency));
 852                 return;
 853         }
 854
 855         /*
 856          * PIIX4 Erratum #18: We don't support C3 when Type-F (fast)
 857          * DMA transfers are used by any ISA device to avoid livelock.
 858          * Note that we could disable Type-F DMA (as recommended by
 859          * the erratum), but this is known to disrupt certain ISA
 860          * devices thus we take the conservative approach.
 861          */
 862         else if (errata.piix4.fdma) {
 863                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 864                                   "C3 not supported on PIIX4 with Type-F DMA\n"));
 865                 return;
 866         }
 867
 868         /* All the logic here assumes flags.bm_check is same across all CPUs */
 869         if (!bm_check_flag) {
 870                 /* Determine whether bm_check is needed based on CPU  */
 871                 acpi_processor_power_init_bm_check(&(pr->flags), pr->id);
 872                 bm_check_flag = pr->flags.bm_check;
 873         } else {
 874                 pr->flags.bm_check = bm_check_flag;
 875         }
 876
 877         if (pr->flags.bm_check) {
 878                 /* bus mastering control is necessary */
 879                 if (!pr->flags.bm_control) {
 880                         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 881                                           "C3 support requires bus mastering control\n"));
 882                         return;
 883                 }
 884         } else {
 885                 /*
 886                  * WBINVD should be set in fadt, for C3 state to be
 887                  * supported on when bm_check is not required.
 888                  */
 889                 if (!(acpi_gbl_FADT.flags & ACPI_FADT_WBINVD)) {
 890                         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 891                                           "Cache invalidation should work properly"
 892                                           " for C3 to be enabled on SMP systems\n"));
 893                         return;
 894                 }
 895                 acpi_set_register(ACPI_BITREG_BUS_MASTER_RLD, 0);
 896         }
 897
 898         /*
 899          * Otherwise we've met all of our C3 requirements.
 900          * Normalize the C3 latency to expidite policy.  Enable
 901          * checking of bus mastering status (bm_check) so we can
 902          * use this in our C3 policy
 903          */
 904         cx->valid = 1;
 905         cx->latency_ticks = US_TO_PM_TIMER_TICKS(cx->latency);
 906
 907         return;
 908 }
 909
 910 static int acpi_processor_power_verify(struct acpi_processor *pr)
 911 {
 912         unsigned int i;
 913         unsigned int working = 0;
 914
 915 #ifdef ARCH_APICTIMER_STOPS_ON_C3
 916         int timer_broadcast = 0;
 917         cpumask_t mask = cpumask_of_cpu(pr->id);
 918         on_each_cpu(switch_ipi_to_APIC_timer, &mask, 1, 1);
 919 #endif
 920
 921         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
 922                 struct acpi_processor_cx *cx = &pr->power.states[i];
 923
 924                 switch (cx->type) {
 925                 case ACPI_STATE_C1:
 926                         cx->valid = 1;
 927                         break;
 928
 929                 case ACPI_STATE_C2:
 930                         acpi_processor_power_verify_c2(cx);
 931 #ifdef ARCH_APICTIMER_STOPS_ON_C3
 932                         /* Some AMD systems fake C3 as C2, but still
 933                            have timer troubles */
 934                         if (cx->valid &&
 935                                 boot_cpu_data.x86_vendor == X86_VENDOR_AMD)
 936                                 timer_broadcast++;
 937 #endif
 938                         break;
 939
 940                 case ACPI_STATE_C3:
 941                         acpi_processor_power_verify_c3(pr, cx);
 942 #ifdef ARCH_APICTIMER_STOPS_ON_C3
 943                         if (cx->valid)
 944                                 timer_broadcast++;
 945 #endif
 946                         break;
 947                 }
 948
 949                 if (cx->valid)
 950                         working++;
 951         }
 952
 953 #ifdef ARCH_APICTIMER_STOPS_ON_C3
 954         if (timer_broadcast)
 955                 on_each_cpu(switch_APIC_timer_to_ipi, &mask, 1, 1);
 956 #endif
 957
 958         return (working);
 959 }
 960
 961 static int acpi_processor_get_power_info(struct acpi_processor *pr)
 962 {
 963         unsigned int i;
 964         int result;
 965
 966
 967         /* NOTE: the idle thread may not be running while calling
 968          * this function */
 969
 970         /* Zero initialize all the C-states info. */
 971         memset(pr->power.states, 0, sizeof(pr->power.states));
 972
 973         result = acpi_processor_get_power_info_cst(pr);
 974         if (result == -ENODEV)
 975                 result = acpi_processor_get_power_info_fadt(pr);
 976
 977         if (result)
 978                 return result;
 979
 980         acpi_processor_get_power_info_default(pr);
 981
 982         pr->power.count = acpi_processor_power_verify(pr);
 983
 984         /*
 985          * Set Default Policy
 986          * ------------------
 987          * Now that we know which states are supported, set the default
 988          * policy.  Note that this policy can be changed dynamically
 989          * (e.g. encourage deeper sleeps to conserve battery life when
 990          * not on AC).
 991          */
 992         result = acpi_processor_set_power_policy(pr);
 993         if (result)
 994                 return result;
 995
 996         /*
 997          * if one state of type C2 or C3 is available, mark this
 998          * CPU as being "idle manageable"
 999          */
1000         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
1001                 if (pr->power.states[i].valid) {
1002                         pr->power.count = i;
1003                         if (pr->power.states[i].type >= ACPI_STATE_C2)
1004                                 pr->flags.power = 1;
1005                 }
1006         }
1007
1008         return 0;
1009 }
1010
1011 int acpi_processor_cst_has_changed(struct acpi_processor *pr)
1012 {
1013         int result = 0;
1014
1015
1016         if (!pr)
1017                 return -EINVAL;
1018
1019         if (nocst) {
1020                 return -ENODEV;
1021         }
1022
1023         if (!pr->flags.power_setup_done)
1024                 return -ENODEV;
1025
1026         /* Fall back to the default idle loop */
1027         pm_idle = pm_idle_save;
1028         synchronize_sched();    /* Relies on interrupts forcing exit from idle. */
1029
1030         pr->flags.power = 0;
1031         result = acpi_processor_get_power_info(pr);
1032         if ((pr->flags.power == 1) && (pr->flags.power_setup_done))
1033                 pm_idle = acpi_processor_idle;
1034
1035         return result;
1036 }
1037
1038 /* proc interface */
1039
1040 static int acpi_processor_power_seq_show(struct seq_file *seq, void *offset)
1041 {
1042         struct acpi_processor *pr = seq->private;
1043         unsigned int i;
1044
1045
1046         if (!pr)
1047                 goto end;
1048
1049         seq_printf(seq, "active state:            C%zd\n"
1050                    "max_cstate:              C%d\n"
1051                    "bus master activity:     %08x\n"
1052                    "maximum allowed latency: %d usec\n",
1053                    pr->power.state ? pr->power.state - pr->power.states : 0,
1054                    max_cstate, (unsigned)pr->power.bm_activity,
1055                    system_latency_constraint());
1056
1057         seq_puts(seq, "states:\n");
1058
1059         for (i = 1; i <= pr->power.count; i++) {
1060                 seq_printf(seq, "   %cC%d:                  ",
1061                            (&pr->power.states[i] ==
1062                             pr->power.state ? '*' : ' '), i);
1063
1064                 if (!pr->power.states[i].valid) {
1065                         seq_puts(seq, "<not supported>\n");
1066                         continue;
1067                 }
1068
1069                 switch (pr->power.states[i].type) {
1070                 case ACPI_STATE_C1:
1071                         seq_printf(seq, "type[C1] ");
1072                         break;
1073                 case ACPI_STATE_C2:
1074                         seq_printf(seq, "type[C2] ");
1075                         break;
1076                 case ACPI_STATE_C3:
1077                         seq_printf(seq, "type[C3] ");
1078                         break;
1079                 default:
1080                         seq_printf(seq, "type[--] ");
1081                         break;
1082                 }
1083
1084                 if (pr->power.states[i].promotion.state)
1085                         seq_printf(seq, "promotion[C%zd] ",
1086                                    (pr->power.states[i].promotion.state -
1087                                     pr->power.states));
1088                 else
1089                         seq_puts(seq, "promotion[--] ");
1090
1091                 if (pr->power.states[i].demotion.state)
1092                         seq_printf(seq, "demotion[C%zd] ",
1093                                    (pr->power.states[i].demotion.state -
1094                                     pr->power.states));
1095                 else
1096                         seq_puts(seq, "demotion[--] ");
1097
1098                 seq_printf(seq, "latency[%03d] usage[%08d] duration[%020llu]\n",
1099                            pr->power.states[i].latency,
1100                            pr->power.states[i].usage,
1101                            (unsigned long long)pr->power.states[i].time);
1102         }
1103
1104       end:
1105         return 0;
1106 }
1107
1108 static int acpi_processor_power_open_fs(struct inode *inode, struct file *file)
1109 {
1110         return single_open(file, acpi_processor_power_seq_show,
1111                            PDE(inode)->data);
1112 }
1113
1114 static const struct file_operations acpi_processor_power_fops = {
1115         .open = acpi_processor_power_open_fs,
1116         .read = seq_read,
1117         .llseek = seq_lseek,
1118         .release = single_release,
1119 };
1120
1121 #ifdef CONFIG_SMP
1122 static void smp_callback(void *v)
1123 {
1124         /* we already woke the CPU up, nothing more to do */
1125 }
1126
1127 /*
1128  * This function gets called when a part of the kernel has a new latency
1129  * requirement.  This means we need to get all processors out of their C-state,
1130  * and then recalculate a new suitable C-state. Just do a cross-cpu IPI; that
1131  * wakes them all right up.
1132  */
1133 static int acpi_processor_latency_notify(struct notifier_block *b,
1134                 unsigned long l, void *v)
1135 {
1136         smp_call_function(smp_callback, NULL, 0, 1);
1137         return NOTIFY_OK;
1138 }
1139
1140 static struct notifier_block acpi_processor_latency_notifier = {
1141         .notifier_call = acpi_processor_latency_notify,
1142 };
1143 #endif
1144
1145 int __cpuinit acpi_processor_power_init(struct acpi_processor *pr,
1146                               struct acpi_device *device)
1147 {
1148         acpi_status status = 0;
1149         static int first_run;
1150         struct proc_dir_entry *entry = NULL;
1151         unsigned int i;
1152
1153
1154         if (!first_run) {
1155                 dmi_check_system(processor_power_dmi_table);
1156                 if (max_cstate < ACPI_C_STATES_MAX)
1157                         printk(KERN_NOTICE
1158                                "ACPI: processor limited to max C-state %d\n",
1159                                max_cstate);
1160                 first_run++;
1161 #ifdef CONFIG_SMP
1162                 register_latency_notifier(&acpi_processor_latency_notifier);
1163 #endif
1164         }
1165
1166         if (!pr)
1167                 return -EINVAL;
1168
1169         if (acpi_gbl_FADT.cst_control && !nocst) {
1170                 status =
1171                     acpi_os_write_port(acpi_gbl_FADT.smi_command, acpi_gbl_FADT.cst_control, 8);
1172                 if (ACPI_FAILURE(status)) {
1173                         ACPI_EXCEPTION((AE_INFO, status,
1174                                         "Notifying BIOS of _CST ability failed"));
1175                 }
1176         }
1177
1178         acpi_processor_get_power_info(pr);
1179
1180         /*
1181          * Install the idle handler if processor power management is supported.
1182          * Note that we use previously set idle handler will be used on
1183          * platforms that only support C1.
1184          */
1185         if ((pr->flags.power) && (!boot_option_idle_override)) {
1186                 printk(KERN_INFO PREFIX "CPU%d (power states:", pr->id);
1187                 for (i = 1; i <= pr->power.count; i++)
1188                         if (pr->power.states[i].valid)
1189                                 printk(" C%d[C%d]", i,
1190                                        pr->power.states[i].type);
1191                 printk(")\n");
1192
1193                 if (pr->id == 0) {
1194                         pm_idle_save = pm_idle;
1195                         pm_idle = acpi_processor_idle;
1196                 }
1197         }
1198
1199         /* 'power' [R] */
1200         entry = create_proc_entry(ACPI_PROCESSOR_FILE_POWER,
1201                                   S_IRUGO, acpi_device_dir(device));
1202         if (!entry)
1203                 return -EIO;
1204         else {
1205                 entry->proc_fops = &acpi_processor_power_fops;
1206                 entry->data = acpi_driver_data(device);
1207                 entry->owner = THIS_MODULE;
1208         }
1209
1210         pr->flags.power_setup_done = 1;
1211
1212         return 0;
1213 }
1214
1215 int acpi_processor_power_exit(struct acpi_processor *pr,
1216                               struct acpi_device *device)
1217 {
1218
1219         pr->flags.power_setup_done = 0;
1220
1221         if (acpi_device_dir(device))
1222                 remove_proc_entry(ACPI_PROCESSOR_FILE_POWER,
1223                                   acpi_device_dir(device));
1224
1225         /* Unregister the idle handler when processor #0 is removed. */
1226         if (pr->id == 0) {
1227                 pm_idle = pm_idle_save;
1228
1229                 /*
1230                  * We are about to unload the current idle thread pm callback
1231                  * (pm_idle), Wait for all processors to update cached/local
1232                  * copies of pm_idle before proceeding.
1233                  */
1234                 cpu_idle_wait();
1235 #ifdef CONFIG_SMP
1236                 unregister_latency_notifier(&acpi_processor_latency_notifier);
1237 #endif
1238         }
1239
1240         return 0;
1241 }