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