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/module.h>
  32 #include <linux/acpi.h>
  33 #include <linux/dmi.h>
  34 #include <linux/sched.h>       /* need_resched() */
  35 #include <linux/clockchips.h>
  36 #include <linux/cpuidle.h>
  37 #include <linux/syscore_ops.h>
  38
  39 /*
  40  * Include the apic definitions for x86 to have the APIC timer related defines
  41  * available also for UP (on SMP it gets magically included via linux/smp.h).
  42  * asm/acpi.h is not an option, as it would require more include magic. Also
  43  * creating an empty asm-ia64/apic.h would just trade pest vs. cholera.
  44  */
  45 #ifdef CONFIG_X86
  46 #include <asm/apic.h>
  47 #endif
  48
  49 #include <acpi/acpi_bus.h>
  50 #include <acpi/processor.h>
  51
  52 #define PREFIX "ACPI: "
  53
  54 #define ACPI_PROCESSOR_CLASS            "processor"
  55 #define _COMPONENT              ACPI_PROCESSOR_COMPONENT
  56 ACPI_MODULE_NAME("processor_idle");
  57
  58 static unsigned int max_cstate __read_mostly = ACPI_PROCESSOR_MAX_POWER;
  59 module_param(max_cstate, uint, 0000);
  60 static unsigned int nocst __read_mostly;
  61 module_param(nocst, uint, 0000);
  62 static int bm_check_disable __read_mostly;
  63 module_param(bm_check_disable, uint, 0000);
  64
  65 static unsigned int latency_factor __read_mostly = 2;
  66 module_param(latency_factor, uint, 0644);
  67
  68 static DEFINE_PER_CPU(struct cpuidle_device *, acpi_cpuidle_device);
  69
  70 static DEFINE_PER_CPU(struct acpi_processor_cx * [CPUIDLE_STATE_MAX],
  71                                                                 acpi_cstate);
  72
  73 static int disabled_by_idle_boot_param(void)
  74 {
  75         return boot_option_idle_override == IDLE_POLL ||
  76                 boot_option_idle_override == IDLE_HALT;
  77 }
  78
  79 /*
  80  * IBM ThinkPad R40e crashes mysteriously when going into C2 or C3.
  81  * For now disable this. Probably a bug somewhere else.
  82  *
  83  * To skip this limit, boot/load with a large max_cstate limit.
  84  */
  85 static int set_max_cstate(const struct dmi_system_id *id)
  86 {
  87         if (max_cstate > ACPI_PROCESSOR_MAX_POWER)
  88                 return 0;
  89
  90         printk(KERN_NOTICE PREFIX "%s detected - limiting to C%ld max_cstate."
  91                " Override with \"processor.max_cstate=%d\"\n", id->ident,
  92                (long)id->driver_data, ACPI_PROCESSOR_MAX_POWER + 1);
  93
  94         max_cstate = (long)id->driver_data;
  95
  96         return 0;
  97 }
  98
  99 static struct dmi_system_id processor_power_dmi_table[] = {
 100         { set_max_cstate, "Clevo 5600D", {
 101           DMI_MATCH(DMI_BIOS_VENDOR,"Phoenix Technologies LTD"),
 102           DMI_MATCH(DMI_BIOS_VERSION,"SHE845M0.86C.0013.D.0302131307")},
 103          (void *)2},
 104         { set_max_cstate, "Pavilion zv5000", {
 105           DMI_MATCH(DMI_SYS_VENDOR, "Hewlett-Packard"),
 106           DMI_MATCH(DMI_PRODUCT_NAME,"Pavilion zv5000 (DS502A#ABA)")},
 107          (void *)1},
 108         { set_max_cstate, "Asus L8400B", {
 109           DMI_MATCH(DMI_SYS_VENDOR, "ASUSTeK Computer Inc."),
 110           DMI_MATCH(DMI_PRODUCT_NAME,"L8400B series Notebook PC")},
 111          (void *)1},
 112         {},
 113 };
 114
 115
 116 /*
 117  * Callers should disable interrupts before the call and enable
 118  * interrupts after return.
 119  */
 120 static void acpi_safe_halt(void)
 121 {
 122         current_thread_info()->status &= ~TS_POLLING;
 123         /*
 124          * TS_POLLING-cleared state must be visible before we
 125          * test NEED_RESCHED:
 126          */
 127         smp_mb();
 128         if (!need_resched()) {
 129                 safe_halt();
 130                 local_irq_disable();
 131         }
 132         current_thread_info()->status |= TS_POLLING;
 133 }
 134
 135 #ifdef ARCH_APICTIMER_STOPS_ON_C3
 136
 137 /*
 138  * Some BIOS implementations switch to C3 in the published C2 state.
 139  * This seems to be a common problem on AMD boxen, but other vendors
 140  * are affected too. We pick the most conservative approach: we assume
 141  * that the local APIC stops in both C2 and C3.
 142  */
 143 static void lapic_timer_check_state(int state, struct acpi_processor *pr,
 144                                    struct acpi_processor_cx *cx)
 145 {
 146         struct acpi_processor_power *pwr = &pr->power;
 147         u8 type = local_apic_timer_c2_ok ? ACPI_STATE_C3 : ACPI_STATE_C2;
 148
 149         if (cpu_has(&cpu_data(pr->id), X86_FEATURE_ARAT))
 150                 return;
 151
 152         if (amd_e400_c1e_detected)
 153                 type = ACPI_STATE_C1;
 154
 155         /*
 156          * Check, if one of the previous states already marked the lapic
 157          * unstable
 158          */
 159         if (pwr->timer_broadcast_on_state < state)
 160                 return;
 161
 162         if (cx->type >= type)
 163                 pr->power.timer_broadcast_on_state = state;
 164 }
 165
 166 static void __lapic_timer_propagate_broadcast(void *arg)
 167 {
 168         struct acpi_processor *pr = (struct acpi_processor *) arg;
 169         unsigned long reason;
 170
 171         reason = pr->power.timer_broadcast_on_state < INT_MAX ?
 172                 CLOCK_EVT_NOTIFY_BROADCAST_ON : CLOCK_EVT_NOTIFY_BROADCAST_OFF;
 173
 174         clockevents_notify(reason, &pr->id);
 175 }
 176
 177 static void lapic_timer_propagate_broadcast(struct acpi_processor *pr)
 178 {
 179         smp_call_function_single(pr->id, __lapic_timer_propagate_broadcast,
 180                                  (void *)pr, 1);
 181 }
 182
 183 /* Power(C) State timer broadcast control */
 184 static void lapic_timer_state_broadcast(struct acpi_processor *pr,
 185                                        struct acpi_processor_cx *cx,
 186                                        int broadcast)
 187 {
 188         int state = cx - pr->power.states;
 189
 190         if (state >= pr->power.timer_broadcast_on_state) {
 191                 unsigned long reason;
 192
 193                 reason = broadcast ?  CLOCK_EVT_NOTIFY_BROADCAST_ENTER :
 194                         CLOCK_EVT_NOTIFY_BROADCAST_EXIT;
 195                 clockevents_notify(reason, &pr->id);
 196         }
 197 }
 198
 199 #else
 200
 201 static void lapic_timer_check_state(int state, struct acpi_processor *pr,
 202                                    struct acpi_processor_cx *cstate) { }
 203 static void lapic_timer_propagate_broadcast(struct acpi_processor *pr) { }
 204 static void lapic_timer_state_broadcast(struct acpi_processor *pr,
 205                                        struct acpi_processor_cx *cx,
 206                                        int broadcast)
 207 {
 208 }
 209
 210 #endif
 211
 212 #ifdef CONFIG_PM_SLEEP
 213 static u32 saved_bm_rld;
 214
 215 static int acpi_processor_suspend(void)
 216 {
 217         acpi_read_bit_register(ACPI_BITREG_BUS_MASTER_RLD, &saved_bm_rld);
 218         return 0;
 219 }
 220
 221 static void acpi_processor_resume(void)
 222 {
 223         u32 resumed_bm_rld;
 224
 225         acpi_read_bit_register(ACPI_BITREG_BUS_MASTER_RLD, &resumed_bm_rld);
 226         if (resumed_bm_rld == saved_bm_rld)
 227                 return;
 228
 229         acpi_write_bit_register(ACPI_BITREG_BUS_MASTER_RLD, saved_bm_rld);
 230 }
 231
 232 static struct syscore_ops acpi_processor_syscore_ops = {
 233         .suspend = acpi_processor_suspend,
 234         .resume = acpi_processor_resume,
 235 };
 236
 237 void acpi_processor_syscore_init(void)
 238 {
 239         register_syscore_ops(&acpi_processor_syscore_ops);
 240 }
 241
 242 void acpi_processor_syscore_exit(void)
 243 {
 244         unregister_syscore_ops(&acpi_processor_syscore_ops);
 245 }
 246 #endif /* CONFIG_PM_SLEEP */
 247
 248 #if defined(CONFIG_X86)
 249 static void tsc_check_state(int state)
 250 {
 251         switch (boot_cpu_data.x86_vendor) {
 252         case X86_VENDOR_AMD:
 253         case X86_VENDOR_INTEL:
 254                 /*
 255                  * AMD Fam10h TSC will tick in all
 256                  * C/P/S0/S1 states when this bit is set.
 257                  */
 258                 if (boot_cpu_has(X86_FEATURE_NONSTOP_TSC))
 259                         return;
 260
 261                 /*FALL THROUGH*/
 262         default:
 263                 /* TSC could halt in idle, so notify users */
 264                 if (state > ACPI_STATE_C1)
 265                         mark_tsc_unstable("TSC halts in idle");
 266         }
 267 }
 268 #else
 269 static void tsc_check_state(int state) { return; }
 270 #endif
 271
 272 static int acpi_processor_get_power_info_fadt(struct acpi_processor *pr)
 273 {
 274
 275         if (!pr)
 276                 return -EINVAL;
 277
 278         if (!pr->pblk)
 279                 return -ENODEV;
 280
 281         /* if info is obtained from pblk/fadt, type equals state */
 282         pr->power.states[ACPI_STATE_C2].type = ACPI_STATE_C2;
 283         pr->power.states[ACPI_STATE_C3].type = ACPI_STATE_C3;
 284
 285 #ifndef CONFIG_HOTPLUG_CPU
 286         /*
 287          * Check for P_LVL2_UP flag before entering C2 and above on
 288          * an SMP system.
 289          */
 290         if ((num_online_cpus() > 1) &&
 291             !(acpi_gbl_FADT.flags & ACPI_FADT_C2_MP_SUPPORTED))
 292                 return -ENODEV;
 293 #endif
 294
 295         /* determine C2 and C3 address from pblk */
 296         pr->power.states[ACPI_STATE_C2].address = pr->pblk + 4;
 297         pr->power.states[ACPI_STATE_C3].address = pr->pblk + 5;
 298
 299         /* determine latencies from FADT */
 300         pr->power.states[ACPI_STATE_C2].latency = acpi_gbl_FADT.c2_latency;
 301         pr->power.states[ACPI_STATE_C3].latency = acpi_gbl_FADT.c3_latency;
 302
 303         /*
 304          * FADT specified C2 latency must be less than or equal to
 305          * 100 microseconds.
 306          */
 307         if (acpi_gbl_FADT.c2_latency > ACPI_PROCESSOR_MAX_C2_LATENCY) {
 308                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 309                         "C2 latency too large [%d]\n", acpi_gbl_FADT.c2_latency));
 310                 /* invalidate C2 */
 311                 pr->power.states[ACPI_STATE_C2].address = 0;
 312         }
 313
 314         /*
 315          * FADT supplied C3 latency must be less than or equal to
 316          * 1000 microseconds.
 317          */
 318         if (acpi_gbl_FADT.c3_latency > ACPI_PROCESSOR_MAX_C3_LATENCY) {
 319                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 320                         "C3 latency too large [%d]\n", acpi_gbl_FADT.c3_latency));
 321                 /* invalidate C3 */
 322                 pr->power.states[ACPI_STATE_C3].address = 0;
 323         }
 324
 325         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 326                           "lvl2[0x%08x] lvl3[0x%08x]\n",
 327                           pr->power.states[ACPI_STATE_C2].address,
 328                           pr->power.states[ACPI_STATE_C3].address));
 329
 330         return 0;
 331 }
 332
 333 static int acpi_processor_get_power_info_default(struct acpi_processor *pr)
 334 {
 335         if (!pr->power.states[ACPI_STATE_C1].valid) {
 336                 /* set the first C-State to C1 */
 337                 /* all processors need to support C1 */
 338                 pr->power.states[ACPI_STATE_C1].type = ACPI_STATE_C1;
 339                 pr->power.states[ACPI_STATE_C1].valid = 1;
 340                 pr->power.states[ACPI_STATE_C1].entry_method = ACPI_CSTATE_HALT;
 341         }
 342         /* the C0 state only exists as a filler in our array */
 343         pr->power.states[ACPI_STATE_C0].valid = 1;
 344         return 0;
 345 }
 346
 347 static int acpi_processor_get_power_info_cst(struct acpi_processor *pr)
 348 {
 349         acpi_status status = 0;
 350         u64 count;
 351         int current_count;
 352         int i;
 353         struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
 354         union acpi_object *cst;
 355
 356
 357         if (nocst)
 358                 return -ENODEV;
 359
 360         current_count = 0;
 361
 362         status = acpi_evaluate_object(pr->handle, "_CST", NULL, &buffer);
 363         if (ACPI_FAILURE(status)) {
 364                 ACPI_DEBUG_PRINT((ACPI_DB_INFO, "No _CST, giving up\n"));
 365                 return -ENODEV;
 366         }
 367
 368         cst = buffer.pointer;
 369
 370         /* There must be at least 2 elements */
 371         if (!cst || (cst->type != ACPI_TYPE_PACKAGE) || cst->package.count < 2) {
 372                 printk(KERN_ERR PREFIX "not enough elements in _CST\n");
 373                 status = -EFAULT;
 374                 goto end;
 375         }
 376
 377         count = cst->package.elements[0].integer.value;
 378
 379         /* Validate number of power states. */
 380         if (count < 1 || count != cst->package.count - 1) {
 381                 printk(KERN_ERR PREFIX "count given by _CST is not valid\n");
 382                 status = -EFAULT;
 383                 goto end;
 384         }
 385
 386         /* Tell driver that at least _CST is supported. */
 387         pr->flags.has_cst = 1;
 388
 389         for (i = 1; i <= count; i++) {
 390                 union acpi_object *element;
 391                 union acpi_object *obj;
 392                 struct acpi_power_register *reg;
 393                 struct acpi_processor_cx cx;
 394
 395                 memset(&cx, 0, sizeof(cx));
 396
 397                 element = &(cst->package.elements[i]);
 398                 if (element->type != ACPI_TYPE_PACKAGE)
 399                         continue;
 400
 401                 if (element->package.count != 4)
 402                         continue;
 403
 404                 obj = &(element->package.elements[0]);
 405
 406                 if (obj->type != ACPI_TYPE_BUFFER)
 407                         continue;
 408
 409                 reg = (struct acpi_power_register *)obj->buffer.pointer;
 410
 411                 if (reg->space_id != ACPI_ADR_SPACE_SYSTEM_IO &&
 412                     (reg->space_id != ACPI_ADR_SPACE_FIXED_HARDWARE))
 413                         continue;
 414
 415                 /* There should be an easy way to extract an integer... */
 416                 obj = &(element->package.elements[1]);
 417                 if (obj->type != ACPI_TYPE_INTEGER)
 418                         continue;
 419
 420                 cx.type = obj->integer.value;
 421                 /*
 422                  * Some buggy BIOSes won't list C1 in _CST -
 423                  * Let acpi_processor_get_power_info_default() handle them later
 424                  */
 425                 if (i == 1 && cx.type != ACPI_STATE_C1)
 426                         current_count++;
 427
 428                 cx.address = reg->address;
 429                 cx.index = current_count + 1;
 430
 431                 cx.entry_method = ACPI_CSTATE_SYSTEMIO;
 432                 if (reg->space_id == ACPI_ADR_SPACE_FIXED_HARDWARE) {
 433                         if (acpi_processor_ffh_cstate_probe
 434                                         (pr->id, &cx, reg) == 0) {
 435                                 cx.entry_method = ACPI_CSTATE_FFH;
 436                         } else if (cx.type == ACPI_STATE_C1) {
 437                                 /*
 438                                  * C1 is a special case where FIXED_HARDWARE
 439                                  * can be handled in non-MWAIT way as well.
 440                                  * In that case, save this _CST entry info.
 441                                  * Otherwise, ignore this info and continue.
 442                                  */
 443                                 cx.entry_method = ACPI_CSTATE_HALT;
 444                                 snprintf(cx.desc, ACPI_CX_DESC_LEN, "ACPI HLT");
 445                         } else {
 446                                 continue;
 447                         }
 448                         if (cx.type == ACPI_STATE_C1 &&
 449                             (boot_option_idle_override == IDLE_NOMWAIT)) {
 450                                 /*
 451                                  * In most cases the C1 space_id obtained from
 452                                  * _CST object is FIXED_HARDWARE access mode.
 453                                  * But when the option of idle=halt is added,
 454                                  * the entry_method type should be changed from
 455                                  * CSTATE_FFH to CSTATE_HALT.
 456                                  * When the option of idle=nomwait is added,
 457                                  * the C1 entry_method type should be
 458                                  * CSTATE_HALT.
 459                                  */
 460                                 cx.entry_method = ACPI_CSTATE_HALT;
 461                                 snprintf(cx.desc, ACPI_CX_DESC_LEN, "ACPI HLT");
 462                         }
 463                 } else {
 464                         snprintf(cx.desc, ACPI_CX_DESC_LEN, "ACPI IOPORT 0x%x",
 465                                  cx.address);
 466                 }
 467
 468                 if (cx.type == ACPI_STATE_C1) {
 469                         cx.valid = 1;
 470                 }
 471
 472                 obj = &(element->package.elements[2]);
 473                 if (obj->type != ACPI_TYPE_INTEGER)
 474                         continue;
 475
 476                 cx.latency = obj->integer.value;
 477
 478                 obj = &(element->package.elements[3]);
 479                 if (obj->type != ACPI_TYPE_INTEGER)
 480                         continue;
 481
 482                 current_count++;
 483                 memcpy(&(pr->power.states[current_count]), &cx, sizeof(cx));
 484
 485                 /*
 486                  * We support total ACPI_PROCESSOR_MAX_POWER - 1
 487                  * (From 1 through ACPI_PROCESSOR_MAX_POWER - 1)
 488                  */
 489                 if (current_count >= (ACPI_PROCESSOR_MAX_POWER - 1)) {
 490                         printk(KERN_WARNING
 491                                "Limiting number of power states to max (%d)\n",
 492                                ACPI_PROCESSOR_MAX_POWER);
 493                         printk(KERN_WARNING
 494                                "Please increase ACPI_PROCESSOR_MAX_POWER if needed.\n");
 495                         break;
 496                 }
 497         }
 498
 499         ACPI_DEBUG_PRINT((ACPI_DB_INFO, "Found %d power states\n",
 500                           current_count));
 501
 502         /* Validate number of power states discovered */
 503         if (current_count < 2)
 504                 status = -EFAULT;
 505
 506       end:
 507         kfree(buffer.pointer);
 508
 509         return status;
 510 }
 511
 512 static void acpi_processor_power_verify_c3(struct acpi_processor *pr,
 513                                            struct acpi_processor_cx *cx)
 514 {
 515         static int bm_check_flag = -1;
 516         static int bm_control_flag = -1;
 517
 518
 519         if (!cx->address)
 520                 return;
 521
 522         /*
 523          * PIIX4 Erratum #18: We don't support C3 when Type-F (fast)
 524          * DMA transfers are used by any ISA device to avoid livelock.
 525          * Note that we could disable Type-F DMA (as recommended by
 526          * the erratum), but this is known to disrupt certain ISA
 527          * devices thus we take the conservative approach.
 528          */
 529         else if (errata.piix4.fdma) {
 530                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 531                                   "C3 not supported on PIIX4 with Type-F DMA\n"));
 532                 return;
 533         }
 534
 535         /* All the logic here assumes flags.bm_check is same across all CPUs */
 536         if (bm_check_flag == -1) {
 537                 /* Determine whether bm_check is needed based on CPU  */
 538                 acpi_processor_power_init_bm_check(&(pr->flags), pr->id);
 539                 bm_check_flag = pr->flags.bm_check;
 540                 bm_control_flag = pr->flags.bm_control;
 541         } else {
 542                 pr->flags.bm_check = bm_check_flag;
 543                 pr->flags.bm_control = bm_control_flag;
 544         }
 545
 546         if (pr->flags.bm_check) {
 547                 if (!pr->flags.bm_control) {
 548                         if (pr->flags.has_cst != 1) {
 549                                 /* bus mastering control is necessary */
 550                                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 551                                         "C3 support requires BM control\n"));
 552                                 return;
 553                         } else {
 554                                 /* Here we enter C3 without bus mastering */
 555                                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 556                                         "C3 support without BM control\n"));
 557                         }
 558                 }
 559         } else {
 560                 /*
 561                  * WBINVD should be set in fadt, for C3 state to be
 562                  * supported on when bm_check is not required.
 563                  */
 564                 if (!(acpi_gbl_FADT.flags & ACPI_FADT_WBINVD)) {
 565                         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 566                                           "Cache invalidation should work properly"
 567                                           " for C3 to be enabled on SMP systems\n"));
 568                         return;
 569                 }
 570         }
 571
 572         /*
 573          * Otherwise we've met all of our C3 requirements.
 574          * Normalize the C3 latency to expidite policy.  Enable
 575          * checking of bus mastering status (bm_check) so we can
 576          * use this in our C3 policy
 577          */
 578         cx->valid = 1;
 579
 580         /*
 581          * On older chipsets, BM_RLD needs to be set
 582          * in order for Bus Master activity to wake the
 583          * system from C3.  Newer chipsets handle DMA
 584          * during C3 automatically and BM_RLD is a NOP.
 585          * In either case, the proper way to
 586          * handle BM_RLD is to set it and leave it set.
 587          */
 588         acpi_write_bit_register(ACPI_BITREG_BUS_MASTER_RLD, 1);
 589
 590         return;
 591 }
 592
 593 static int acpi_processor_power_verify(struct acpi_processor *pr)
 594 {
 595         unsigned int i;
 596         unsigned int working = 0;
 597
 598         pr->power.timer_broadcast_on_state = INT_MAX;
 599
 600         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER && i <= max_cstate; i++) {
 601                 struct acpi_processor_cx *cx = &pr->power.states[i];
 602
 603                 switch (cx->type) {
 604                 case ACPI_STATE_C1:
 605                         cx->valid = 1;
 606                         break;
 607
 608                 case ACPI_STATE_C2:
 609                         if (!cx->address)
 610                                 break;
 611                         cx->valid = 1;
 612                         break;
 613
 614                 case ACPI_STATE_C3:
 615                         acpi_processor_power_verify_c3(pr, cx);
 616                         break;
 617                 }
 618                 if (!cx->valid)
 619                         continue;
 620
 621                 lapic_timer_check_state(i, pr, cx);
 622                 tsc_check_state(cx->type);
 623                 working++;
 624         }
 625
 626         lapic_timer_propagate_broadcast(pr);
 627
 628         return (working);
 629 }
 630
 631 static int acpi_processor_get_power_info(struct acpi_processor *pr)
 632 {
 633         unsigned int i;
 634         int result;
 635
 636
 637         /* NOTE: the idle thread may not be running while calling
 638          * this function */
 639
 640         /* Zero initialize all the C-states info. */
 641         memset(pr->power.states, 0, sizeof(pr->power.states));
 642
 643         result = acpi_processor_get_power_info_cst(pr);
 644         if (result == -ENODEV)
 645                 result = acpi_processor_get_power_info_fadt(pr);
 646
 647         if (result)
 648                 return result;
 649
 650         acpi_processor_get_power_info_default(pr);
 651
 652         pr->power.count = acpi_processor_power_verify(pr);
 653
 654         /*
 655          * if one state of type C2 or C3 is available, mark this
 656          * CPU as being "idle manageable"
 657          */
 658         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
 659                 if (pr->power.states[i].valid) {
 660                         pr->power.count = i;
 661                         if (pr->power.states[i].type >= ACPI_STATE_C2)
 662                                 pr->flags.power = 1;
 663                 }
 664         }
 665
 666         return 0;
 667 }
 668
 669 /**
 670  * acpi_idle_bm_check - checks if bus master activity was detected
 671  */
 672 static int acpi_idle_bm_check(void)
 673 {
 674         u32 bm_status = 0;
 675
 676         if (bm_check_disable)
 677                 return 0;
 678
 679         acpi_read_bit_register(ACPI_BITREG_BUS_MASTER_STATUS, &bm_status);
 680         if (bm_status)
 681                 acpi_write_bit_register(ACPI_BITREG_BUS_MASTER_STATUS, 1);
 682         /*
 683          * PIIX4 Erratum #18: Note that BM_STS doesn't always reflect
 684          * the true state of bus mastering activity; forcing us to
 685          * manually check the BMIDEA bit of each IDE channel.
 686          */
 687         else if (errata.piix4.bmisx) {
 688                 if ((inb_p(errata.piix4.bmisx + 0x02) & 0x01)
 689                     || (inb_p(errata.piix4.bmisx + 0x0A) & 0x01))
 690                         bm_status = 1;
 691         }
 692         return bm_status;
 693 }
 694
 695 /**
 696  * acpi_idle_do_entry - a helper function that does C2 and C3 type entry
 697  * @cx: cstate data
 698  *
 699  * Caller disables interrupt before call and enables interrupt after return.
 700  */
 701 static inline void acpi_idle_do_entry(struct acpi_processor_cx *cx)
 702 {
 703         /* Don't trace irqs off for idle */
 704         stop_critical_timings();
 705         if (cx->entry_method == ACPI_CSTATE_FFH) {
 706                 /* Call into architectural FFH based C-state */
 707                 acpi_processor_ffh_cstate_enter(cx);
 708         } else if (cx->entry_method == ACPI_CSTATE_HALT) {
 709                 acpi_safe_halt();
 710         } else {
 711                 /* IO port based C-state */
 712                 inb(cx->address);
 713                 /* Dummy wait op - must do something useless after P_LVL2 read
 714                    because chipsets cannot guarantee that STPCLK# signal
 715                    gets asserted in time to freeze execution properly. */
 716                 inl(acpi_gbl_FADT.xpm_timer_block.address);
 717         }
 718         start_critical_timings();
 719 }
 720
 721 /**
 722  * acpi_idle_enter_c1 - enters an ACPI C1 state-type
 723  * @dev: the target CPU
 724  * @drv: cpuidle driver containing cpuidle state info
 725  * @index: index of target state
 726  *
 727  * This is equivalent to the HALT instruction.
 728  */
 729 static int acpi_idle_enter_c1(struct cpuidle_device *dev,
 730                 struct cpuidle_driver *drv, int index)
 731 {
 732         struct acpi_processor *pr;
 733         struct acpi_processor_cx *cx = per_cpu(acpi_cstate[index], dev->cpu);
 734
 735         pr = __this_cpu_read(processors);
 736
 737         if (unlikely(!pr))
 738                 return -EINVAL;
 739
 740         lapic_timer_state_broadcast(pr, cx, 1);
 741         acpi_idle_do_entry(cx);
 742
 743         lapic_timer_state_broadcast(pr, cx, 0);
 744
 745         return index;
 746 }
 747
 748
 749 /**
 750  * acpi_idle_play_dead - enters an ACPI state for long-term idle (i.e. off-lining)
 751  * @dev: the target CPU
 752  * @index: the index of suggested state
 753  */
 754 static int acpi_idle_play_dead(struct cpuidle_device *dev, int index)
 755 {
 756         struct acpi_processor_cx *cx = per_cpu(acpi_cstate[index], dev->cpu);
 757
 758         ACPI_FLUSH_CPU_CACHE();
 759
 760         while (1) {
 761
 762                 if (cx->entry_method == ACPI_CSTATE_HALT)
 763                         safe_halt();
 764                 else if (cx->entry_method == ACPI_CSTATE_SYSTEMIO) {
 765                         inb(cx->address);
 766                         /* See comment in acpi_idle_do_entry() */
 767                         inl(acpi_gbl_FADT.xpm_timer_block.address);
 768                 } else
 769                         return -ENODEV;
 770         }
 771
 772         /* Never reached */
 773         return 0;
 774 }
 775
 776 /**
 777  * acpi_idle_enter_simple - enters an ACPI state without BM handling
 778  * @dev: the target CPU
 779  * @drv: cpuidle driver with cpuidle state information
 780  * @index: the index of suggested state
 781  */
 782 static int acpi_idle_enter_simple(struct cpuidle_device *dev,
 783                 struct cpuidle_driver *drv, int index)
 784 {
 785         struct acpi_processor *pr;
 786         struct acpi_processor_cx *cx = per_cpu(acpi_cstate[index], dev->cpu);
 787
 788         pr = __this_cpu_read(processors);
 789
 790         if (unlikely(!pr))
 791                 return -EINVAL;
 792
 793         if (cx->entry_method != ACPI_CSTATE_FFH) {
 794                 current_thread_info()->status &= ~TS_POLLING;
 795                 /*
 796                  * TS_POLLING-cleared state must be visible before we test
 797                  * NEED_RESCHED:
 798                  */
 799                 smp_mb();
 800
 801                 if (unlikely(need_resched())) {
 802                         current_thread_info()->status |= TS_POLLING;
 803                         return -EINVAL;
 804                 }
 805         }
 806
 807         /*
 808          * Must be done before busmaster disable as we might need to
 809          * access HPET !
 810          */
 811         lapic_timer_state_broadcast(pr, cx, 1);
 812
 813         if (cx->type == ACPI_STATE_C3)
 814                 ACPI_FLUSH_CPU_CACHE();
 815
 816         /* Tell the scheduler that we are going deep-idle: */
 817         sched_clock_idle_sleep_event();
 818         acpi_idle_do_entry(cx);
 819
 820         sched_clock_idle_wakeup_event(0);
 821
 822         if (cx->entry_method != ACPI_CSTATE_FFH)
 823                 current_thread_info()->status |= TS_POLLING;
 824
 825         lapic_timer_state_broadcast(pr, cx, 0);
 826         return index;
 827 }
 828
 829 static int c3_cpu_count;
 830 static DEFINE_RAW_SPINLOCK(c3_lock);
 831
 832 /**
 833  * acpi_idle_enter_bm - enters C3 with proper BM handling
 834  * @dev: the target CPU
 835  * @drv: cpuidle driver containing state data
 836  * @index: the index of suggested state
 837  *
 838  * If BM is detected, the deepest non-C3 idle state is entered instead.
 839  */
 840 static int acpi_idle_enter_bm(struct cpuidle_device *dev,
 841                 struct cpuidle_driver *drv, int index)
 842 {
 843         struct acpi_processor *pr;
 844         struct acpi_processor_cx *cx = per_cpu(acpi_cstate[index], dev->cpu);
 845
 846         pr = __this_cpu_read(processors);
 847
 848         if (unlikely(!pr))
 849                 return -EINVAL;
 850
 851         if (!cx->bm_sts_skip && acpi_idle_bm_check()) {
 852                 if (drv->safe_state_index >= 0) {
 853                         return drv->states[drv->safe_state_index].enter(dev,
 854                                                 drv, drv->safe_state_index);
 855                 } else {
 856                         acpi_safe_halt();
 857                         return -EBUSY;
 858                 }
 859         }
 860
 861         if (cx->entry_method != ACPI_CSTATE_FFH) {
 862                 current_thread_info()->status &= ~TS_POLLING;
 863                 /*
 864                  * TS_POLLING-cleared state must be visible before we test
 865                  * NEED_RESCHED:
 866                  */
 867                 smp_mb();
 868
 869                 if (unlikely(need_resched())) {
 870                         current_thread_info()->status |= TS_POLLING;
 871                         return -EINVAL;
 872                 }
 873         }
 874
 875         acpi_unlazy_tlb(smp_processor_id());
 876
 877         /* Tell the scheduler that we are going deep-idle: */
 878         sched_clock_idle_sleep_event();
 879         /*
 880          * Must be done before busmaster disable as we might need to
 881          * access HPET !
 882          */
 883         lapic_timer_state_broadcast(pr, cx, 1);
 884
 885         /*
 886          * disable bus master
 887          * bm_check implies we need ARB_DIS
 888          * !bm_check implies we need cache flush
 889          * bm_control implies whether we can do ARB_DIS
 890          *
 891          * That leaves a case where bm_check is set and bm_control is
 892          * not set. In that case we cannot do much, we enter C3
 893          * without doing anything.
 894          */
 895         if (pr->flags.bm_check && pr->flags.bm_control) {
 896                 raw_spin_lock(&c3_lock);
 897                 c3_cpu_count++;
 898                 /* Disable bus master arbitration when all CPUs are in C3 */
 899                 if (c3_cpu_count == num_online_cpus())
 900                         acpi_write_bit_register(ACPI_BITREG_ARB_DISABLE, 1);
 901                 raw_spin_unlock(&c3_lock);
 902         } else if (!pr->flags.bm_check) {
 903                 ACPI_FLUSH_CPU_CACHE();
 904         }
 905
 906         acpi_idle_do_entry(cx);
 907
 908         /* Re-enable bus master arbitration */
 909         if (pr->flags.bm_check && pr->flags.bm_control) {
 910                 raw_spin_lock(&c3_lock);
 911                 acpi_write_bit_register(ACPI_BITREG_ARB_DISABLE, 0);
 912                 c3_cpu_count--;
 913                 raw_spin_unlock(&c3_lock);
 914         }
 915
 916         sched_clock_idle_wakeup_event(0);
 917
 918         if (cx->entry_method != ACPI_CSTATE_FFH)
 919                 current_thread_info()->status |= TS_POLLING;
 920
 921         lapic_timer_state_broadcast(pr, cx, 0);
 922         return index;
 923 }
 924
 925 struct cpuidle_driver acpi_idle_driver = {
 926         .name =         "acpi_idle",
 927         .owner =        THIS_MODULE,
 928 };
 929
 930 /**
 931  * acpi_processor_setup_cpuidle_cx - prepares and configures CPUIDLE
 932  * device i.e. per-cpu data
 933  *
 934  * @pr: the ACPI processor
 935  * @dev : the cpuidle device
 936  */
 937 static int acpi_processor_setup_cpuidle_cx(struct acpi_processor *pr,
 938                                            struct cpuidle_device *dev)
 939 {
 940         int i, count = CPUIDLE_DRIVER_STATE_START;
 941         struct acpi_processor_cx *cx;
 942
 943         if (!pr->flags.power_setup_done)
 944                 return -EINVAL;
 945
 946         if (pr->flags.power == 0) {
 947                 return -EINVAL;
 948         }
 949
 950         if (!dev)
 951                 return -EINVAL;
 952
 953         dev->cpu = pr->id;
 954
 955         if (max_cstate == 0)
 956                 max_cstate = 1;
 957
 958         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER && i <= max_cstate; i++) {
 959                 cx = &pr->power.states[i];
 960
 961                 if (!cx->valid)
 962                         continue;
 963
 964 #ifdef CONFIG_HOTPLUG_CPU
 965                 if ((cx->type != ACPI_STATE_C1) && (num_online_cpus() > 1) &&
 966                     !pr->flags.has_cst &&
 967                     !(acpi_gbl_FADT.flags & ACPI_FADT_C2_MP_SUPPORTED))
 968                         continue;
 969 #endif
 970                 per_cpu(acpi_cstate[count], dev->cpu) = cx;
 971
 972                 count++;
 973                 if (count == CPUIDLE_STATE_MAX)
 974                         break;
 975         }
 976
 977         dev->state_count = count;
 978
 979         if (!count)
 980                 return -EINVAL;
 981
 982         return 0;
 983 }
 984
 985 /**
 986  * acpi_processor_setup_cpuidle states- prepares and configures cpuidle
 987  * global state data i.e. idle routines
 988  *
 989  * @pr: the ACPI processor
 990  */
 991 static int acpi_processor_setup_cpuidle_states(struct acpi_processor *pr)
 992 {
 993         int i, count = CPUIDLE_DRIVER_STATE_START;
 994         struct acpi_processor_cx *cx;
 995         struct cpuidle_state *state;
 996         struct cpuidle_driver *drv = &acpi_idle_driver;
 997
 998         if (!pr->flags.power_setup_done)
 999                 return -EINVAL;
1000
1001         if (pr->flags.power == 0)
1002                 return -EINVAL;
1003
1004         drv->safe_state_index = -1;
1005         for (i = 0; i < CPUIDLE_STATE_MAX; i++) {
1006                 drv->states[i].name[0] = '\0';
1007                 drv->states[i].desc[0] = '\0';
1008         }
1009
1010         if (max_cstate == 0)
1011                 max_cstate = 1;
1012
1013         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER && i <= max_cstate; i++) {
1014                 cx = &pr->power.states[i];
1015
1016                 if (!cx->valid)
1017                         continue;
1018
1019 #ifdef CONFIG_HOTPLUG_CPU
1020                 if ((cx->type != ACPI_STATE_C1) && (num_online_cpus() > 1) &&
1021                     !pr->flags.has_cst &&
1022                     !(acpi_gbl_FADT.flags & ACPI_FADT_C2_MP_SUPPORTED))
1023                         continue;
1024 #endif
1025
1026                 state = &drv->states[count];
1027                 snprintf(state->name, CPUIDLE_NAME_LEN, "C%d", i);
1028                 strncpy(state->desc, cx->desc, CPUIDLE_DESC_LEN);
1029                 state->exit_latency = cx->latency;
1030                 state->target_residency = cx->latency * latency_factor;
1031
1032                 state->flags = 0;
1033                 switch (cx->type) {
1034                         case ACPI_STATE_C1:
1035                         if (cx->entry_method == ACPI_CSTATE_FFH)
1036                                 state->flags |= CPUIDLE_FLAG_TIME_VALID;
1037
1038                         state->enter = acpi_idle_enter_c1;
1039                         state->enter_dead = acpi_idle_play_dead;
1040                         drv->safe_state_index = count;
1041                         break;
1042
1043                         case ACPI_STATE_C2:
1044                         state->flags |= CPUIDLE_FLAG_TIME_VALID;
1045                         state->enter = acpi_idle_enter_simple;
1046                         state->enter_dead = acpi_idle_play_dead;
1047                         drv->safe_state_index = count;
1048                         break;
1049
1050                         case ACPI_STATE_C3:
1051                         state->flags |= CPUIDLE_FLAG_TIME_VALID;
1052                         state->enter = pr->flags.bm_check ?
1053                                         acpi_idle_enter_bm :
1054                                         acpi_idle_enter_simple;
1055                         break;
1056                 }
1057
1058                 count++;
1059                 if (count == CPUIDLE_STATE_MAX)
1060                         break;
1061         }
1062
1063         drv->state_count = count;
1064
1065         if (!count)
1066                 return -EINVAL;
1067
1068         return 0;
1069 }
1070
1071 int acpi_processor_hotplug(struct acpi_processor *pr)
1072 {
1073         int ret = 0;
1074         struct cpuidle_device *dev;
1075
1076         if (disabled_by_idle_boot_param())
1077                 return 0;
1078
1079         if (!pr)
1080                 return -EINVAL;
1081
1082         if (nocst) {
1083                 return -ENODEV;
1084         }
1085
1086         if (!pr->flags.power_setup_done)
1087                 return -ENODEV;
1088
1089         dev = per_cpu(acpi_cpuidle_device, pr->id);
1090         cpuidle_pause_and_lock();
1091         cpuidle_disable_device(dev);
1092         acpi_processor_get_power_info(pr);
1093         if (pr->flags.power) {
1094                 acpi_processor_setup_cpuidle_cx(pr, dev);
1095                 ret = cpuidle_enable_device(dev);
1096         }
1097         cpuidle_resume_and_unlock();
1098
1099         return ret;
1100 }
1101
1102 int acpi_processor_cst_has_changed(struct acpi_processor *pr)
1103 {
1104         int cpu;
1105         struct acpi_processor *_pr;
1106         struct cpuidle_device *dev;
1107
1108         if (disabled_by_idle_boot_param())
1109                 return 0;
1110
1111         if (!pr)
1112                 return -EINVAL;
1113
1114         if (nocst)
1115                 return -ENODEV;
1116
1117         if (!pr->flags.power_setup_done)
1118                 return -ENODEV;
1119
1120         /*
1121          * FIXME:  Design the ACPI notification to make it once per
1122          * system instead of once per-cpu.  This condition is a hack
1123          * to make the code that updates C-States be called once.
1124          */
1125
1126         if (pr->id == 0 && cpuidle_get_driver() == &acpi_idle_driver) {
1127
1128                 cpuidle_pause_and_lock();
1129                 /* Protect against cpu-hotplug */
1130                 get_online_cpus();
1131
1132                 /* Disable all cpuidle devices */
1133                 for_each_online_cpu(cpu) {
1134                         _pr = per_cpu(processors, cpu);
1135                         if (!_pr || !_pr->flags.power_setup_done)
1136                                 continue;
1137                         dev = per_cpu(acpi_cpuidle_device, cpu);
1138                         cpuidle_disable_device(dev);
1139                 }
1140
1141                 /* Populate Updated C-state information */
1142                 acpi_processor_get_power_info(pr);
1143                 acpi_processor_setup_cpuidle_states(pr);
1144
1145                 /* Enable all cpuidle devices */
1146                 for_each_online_cpu(cpu) {
1147                         _pr = per_cpu(processors, cpu);
1148                         if (!_pr || !_pr->flags.power_setup_done)
1149                                 continue;
1150                         acpi_processor_get_power_info(_pr);
1151                         if (_pr->flags.power) {
1152                                 dev = per_cpu(acpi_cpuidle_device, cpu);
1153                                 acpi_processor_setup_cpuidle_cx(_pr, dev);
1154                                 cpuidle_enable_device(dev);
1155                         }
1156                 }
1157                 put_online_cpus();
1158                 cpuidle_resume_and_unlock();
1159         }
1160
1161         return 0;
1162 }
1163
1164 static int acpi_processor_registered;
1165
1166 int acpi_processor_power_init(struct acpi_processor *pr)
1167 {
1168         acpi_status status = 0;
1169         int retval;
1170         struct cpuidle_device *dev;
1171         static int first_run;
1172
1173         if (disabled_by_idle_boot_param())
1174                 return 0;
1175
1176         if (!first_run) {
1177                 dmi_check_system(processor_power_dmi_table);
1178                 max_cstate = acpi_processor_cstate_check(max_cstate);
1179                 if (max_cstate < ACPI_C_STATES_MAX)
1180                         printk(KERN_NOTICE
1181                                "ACPI: processor limited to max C-state %d\n",
1182                                max_cstate);
1183                 first_run++;
1184         }
1185
1186         if (!pr)
1187                 return -EINVAL;
1188
1189         if (acpi_gbl_FADT.cst_control && !nocst) {
1190                 status =
1191                     acpi_os_write_port(acpi_gbl_FADT.smi_command, acpi_gbl_FADT.cst_control, 8);
1192                 if (ACPI_FAILURE(status)) {
1193                         ACPI_EXCEPTION((AE_INFO, status,
1194                                         "Notifying BIOS of _CST ability failed"));
1195                 }
1196         }
1197
1198         acpi_processor_get_power_info(pr);
1199         pr->flags.power_setup_done = 1;
1200
1201         /*
1202          * Install the idle handler if processor power management is supported.
1203          * Note that we use previously set idle handler will be used on
1204          * platforms that only support C1.
1205          */
1206         if (pr->flags.power) {
1207                 /* Register acpi_idle_driver if not already registered */
1208                 if (!acpi_processor_registered) {
1209                         acpi_processor_setup_cpuidle_states(pr);
1210                         retval = cpuidle_register_driver(&acpi_idle_driver);
1211                         if (retval)
1212                                 return retval;
1213                         printk(KERN_DEBUG "ACPI: %s registered with cpuidle\n",
1214                                         acpi_idle_driver.name);
1215                 }
1216
1217                 dev = kzalloc(sizeof(*dev), GFP_KERNEL);
1218                 if (!dev)
1219                         return -ENOMEM;
1220                 per_cpu(acpi_cpuidle_device, pr->id) = dev;
1221
1222                 acpi_processor_setup_cpuidle_cx(pr, dev);
1223
1224                 /* Register per-cpu cpuidle_device. Cpuidle driver
1225                  * must already be registered before registering device
1226                  */
1227                 retval = cpuidle_register_device(dev);
1228                 if (retval) {
1229                         if (acpi_processor_registered == 0)
1230                                 cpuidle_unregister_driver(&acpi_idle_driver);
1231                         return retval;
1232                 }
1233                 acpi_processor_registered++;
1234         }
1235         return 0;
1236 }
1237
1238 int acpi_processor_power_exit(struct acpi_processor *pr)
1239 {
1240         struct cpuidle_device *dev = per_cpu(acpi_cpuidle_device, pr->id);
1241
1242         if (disabled_by_idle_boot_param())
1243                 return 0;
1244
1245         if (pr->flags.power) {
1246                 cpuidle_unregister_device(dev);
1247                 acpi_processor_registered--;
1248                 if (acpi_processor_registered == 0)
1249                         cpuidle_unregister_driver(&acpi_idle_driver);
1250         }
1251
1252         pr->flags.power_setup_done = 0;
1253         return 0;
1254 }