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