drivers/acpi/processor_idle.c

   1 /*
   2  * processor_idle - idle state submodule to the ACPI processor driver
   3  *
   4  *  Copyright (C) 2001, 2002 Andy Grover <andrew.grover@intel.com>
   5  *  Copyright (C) 2001, 2002 Paul Diefenbaugh <paul.s.diefenbaugh@intel.com>
   6  *  Copyright (C) 2004, 2005 Dominik Brodowski <linux@brodo.de>
   7  *  Copyright (C) 2004  Anil S Keshavamurthy <anil.s.keshavamurthy@intel.com>
   8  *                      - Added processor hotplug support
   9  *  Copyright (C) 2005  Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>
  10  *                      - Added support for C3 on SMP
  11  *
  12  * ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
  13  *
  14  *  This program is free software; you can redistribute it and/or modify
  15  *  it under the terms of the GNU General Public License as published by
  16  *  the Free Software Foundation; either version 2 of the License, or (at
  17  *  your option) any later version.
  18  *
  19  *  This program is distributed in the hope that it will be useful, but
  20  *  WITHOUT ANY WARRANTY; without even the implied warranty of
  21  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  22  *  General Public License for more details.
  23  *
  24  *  You should have received a copy of the GNU General Public License along
  25  *  with this program; if not, write to the Free Software Foundation, Inc.,
  26  *  59 Temple Place, Suite 330, Boston, MA 02111-1307 USA.
  27  *
  28  * ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
  29  */
  30
  31 #include <linux/module.h>
  32 #include <linux/acpi.h>
  33 #include <linux/dmi.h>
  34 #include <linux/sched.h>       /* need_resched() */
  35 #include <linux/clockchips.h>
  36 #include <linux/cpuidle.h>
  37 #include <linux/syscore_ops.h>
  38
  39 /*
  40  * Include the apic definitions for x86 to have the APIC timer related defines
  41  * available also for UP (on SMP it gets magically included via linux/smp.h).
  42  * asm/acpi.h is not an option, as it would require more include magic. Also
  43  * creating an empty asm-ia64/apic.h would just trade pest vs. cholera.
  44  */
  45 #ifdef CONFIG_X86
  46 #include <asm/apic.h>
  47 #endif
  48
  49 #include <acpi/acpi_bus.h>
  50 #include <acpi/processor.h>
  51
  52 #define PREFIX "ACPI: "
  53
  54 #define ACPI_PROCESSOR_CLASS            "processor"
  55 #define _COMPONENT              ACPI_PROCESSOR_COMPONENT
  56 ACPI_MODULE_NAME("processor_idle");
  57
  58 static unsigned int max_cstate __read_mostly = ACPI_PROCESSOR_MAX_POWER;
  59 module_param(max_cstate, uint, 0000);
  60 static unsigned int nocst __read_mostly;
  61 module_param(nocst, uint, 0000);
  62 static int bm_check_disable __read_mostly;
  63 module_param(bm_check_disable, uint, 0000);
  64
  65 static unsigned int latency_factor __read_mostly = 2;
  66 module_param(latency_factor, uint, 0644);
  67
  68 static DEFINE_PER_CPU(struct cpuidle_device *, acpi_cpuidle_device);
  69
  70 static DEFINE_PER_CPU(struct acpi_processor_cx * [CPUIDLE_STATE_MAX],
  71                                                                 acpi_cstate);
  72
  73 static int disabled_by_idle_boot_param(void)
  74 {
  75         return boot_option_idle_override == IDLE_POLL ||
  76                 boot_option_idle_override == IDLE_HALT;
  77 }
  78
  79 /*
  80  * IBM ThinkPad R40e crashes mysteriously when going into C2 or C3.
  81  * For now disable this. Probably a bug somewhere else.
  82  *
  83  * To skip this limit, boot/load with a large max_cstate limit.
  84  */
  85 static int set_max_cstate(const struct dmi_system_id *id)
  86 {
  87         if (max_cstate > ACPI_PROCESSOR_MAX_POWER)
  88                 return 0;
  89
  90         printk(KERN_NOTICE PREFIX "%s detected - limiting to C%ld max_cstate."
  91                " Override with \"processor.max_cstate=%d\"\n", id->ident,
  92                (long)id->driver_data, ACPI_PROCESSOR_MAX_POWER + 1);
  93
  94         max_cstate = (long)id->driver_data;
  95
  96         return 0;
  97 }
  98
  99 static struct dmi_system_id processor_power_dmi_table[] = {
 100         { set_max_cstate, "Clevo 5600D", {
 101           DMI_MATCH(DMI_BIOS_VENDOR,"Phoenix Technologies LTD"),
 102           DMI_MATCH(DMI_BIOS_VERSION,"SHE845M0.86C.0013.D.0302131307")},
 103          (void *)2},
 104         { set_max_cstate, "Pavilion zv5000", {
 105           DMI_MATCH(DMI_SYS_VENDOR, "Hewlett-Packard"),
 106           DMI_MATCH(DMI_PRODUCT_NAME,"Pavilion zv5000 (DS502A#ABA)")},
 107          (void *)1},
 108         { set_max_cstate, "Asus L8400B", {
 109           DMI_MATCH(DMI_SYS_VENDOR, "ASUSTeK Computer Inc."),
 110           DMI_MATCH(DMI_PRODUCT_NAME,"L8400B series Notebook PC")},
 111          (void *)1},
 112         {},
 113 };
 114
 115
 116 /*
 117  * Callers should disable interrupts before the call and enable
 118  * interrupts after return.
 119  */
 120 static void acpi_safe_halt(void)
 121 {
 122         current_thread_info()->status &= ~TS_POLLING;
 123         /*
 124          * TS_POLLING-cleared state must be visible before we
 125          * test NEED_RESCHED:
 126          */
 127         smp_mb();
 128         if (!need_resched()) {
 129                 safe_halt();
 130                 local_irq_disable();
 131         }
 132         current_thread_info()->status |= TS_POLLING;
 133 }
 134
 135 #ifdef ARCH_APICTIMER_STOPS_ON_C3
 136
 137 /*
 138  * Some BIOS implementations switch to C3 in the published C2 state.
 139  * This seems to be a common problem on AMD boxen, but other vendors
 140  * are affected too. We pick the most conservative approach: we assume
 141  * that the local APIC stops in both C2 and C3.
 142  */
 143 static void lapic_timer_check_state(int state, struct acpi_processor *pr,
 144                                    struct acpi_processor_cx *cx)
 145 {
 146         struct acpi_processor_power *pwr = &pr->power;
 147         u8 type = local_apic_timer_c2_ok ? ACPI_STATE_C3 : ACPI_STATE_C2;
 148
 149         if (cpu_has(&cpu_data(pr->id), X86_FEATURE_ARAT))
 150                 return;
 151
 152         if (amd_e400_c1e_detected)
 153                 type = ACPI_STATE_C1;
 154
 155         /*
 156          * Check, if one of the previous states already marked the lapic
 157          * unstable
 158          */
 159         if (pwr->timer_broadcast_on_state < state)
 160                 return;
 161
 162         if (cx->type >= type)
 163                 pr->power.timer_broadcast_on_state = state;
 164 }
 165
 166 static void __lapic_timer_propagate_broadcast(void *arg)
 167 {
 168         struct acpi_processor *pr = (struct acpi_processor *) arg;
 169         unsigned long reason;
 170
 171         reason = pr->power.timer_broadcast_on_state < INT_MAX ?
 172                 CLOCK_EVT_NOTIFY_BROADCAST_ON : CLOCK_EVT_NOTIFY_BROADCAST_OFF;
 173
 174         clockevents_notify(reason, &pr->id);
 175 }
 176
 177 static void lapic_timer_propagate_broadcast(struct acpi_processor *pr)
 178 {
 179         smp_call_function_single(pr->id, __lapic_timer_propagate_broadcast,
 180                                  (void *)pr, 1);
 181 }
 182
 183 /* Power(C) State timer broadcast control */
 184 static void lapic_timer_state_broadcast(struct acpi_processor *pr,
 185                                        struct acpi_processor_cx *cx,
 186                                        int broadcast)
 187 {
 188         int state = cx - pr->power.states;
 189
 190         if (state >= pr->power.timer_broadcast_on_state) {
 191                 unsigned long reason;
 192
 193                 reason = broadcast ?  CLOCK_EVT_NOTIFY_BROADCAST_ENTER :
 194                         CLOCK_EVT_NOTIFY_BROADCAST_EXIT;
 195                 clockevents_notify(reason, &pr->id);
 196         }
 197 }
 198
 199 #else
 200
 201 static void lapic_timer_check_state(int state, struct acpi_processor *pr,
 202                                    struct acpi_processor_cx *cstate) { }
 203 static void lapic_timer_propagate_broadcast(struct acpi_processor *pr) { }
 204 static void lapic_timer_state_broadcast(struct acpi_processor *pr,
 205                                        struct acpi_processor_cx *cx,
 206                                        int broadcast)
 207 {
 208 }
 209
 210 #endif
 211
 212 #ifdef CONFIG_PM_SLEEP
 213 static u32 saved_bm_rld;
 214
 215 static int acpi_processor_suspend(void)
 216 {
 217         acpi_read_bit_register(ACPI_BITREG_BUS_MASTER_RLD, &saved_bm_rld);
 218         return 0;
 219 }
 220
 221 static void acpi_processor_resume(void)
 222 {
 223         u32 resumed_bm_rld;
 224
 225         acpi_read_bit_register(ACPI_BITREG_BUS_MASTER_RLD, &resumed_bm_rld);
 226         if (resumed_bm_rld == saved_bm_rld)
 227                 return;
 228
 229         acpi_write_bit_register(ACPI_BITREG_BUS_MASTER_RLD, saved_bm_rld);
 230 }
 231
 232 static struct syscore_ops acpi_processor_syscore_ops = {
 233         .suspend = acpi_processor_suspend,
 234         .resume = acpi_processor_resume,
 235 };
 236
 237 void acpi_processor_syscore_init(void)
 238 {
 239         register_syscore_ops(&acpi_processor_syscore_ops);
 240 }
 241
 242 void acpi_processor_syscore_exit(void)
 243 {
 244         unregister_syscore_ops(&acpi_processor_syscore_ops);
 245 }
 246 #endif /* CONFIG_PM_SLEEP */
 247
 248 #if defined(CONFIG_X86)
 249 static void tsc_check_state(int state)
 250 {
 251         switch (boot_cpu_data.x86_vendor) {
 252         case X86_VENDOR_AMD:
 253         case X86_VENDOR_INTEL:
 254                 /*
 255                  * AMD Fam10h TSC will tick in all
 256                  * C/P/S0/S1 states when this bit is set.
 257                  */
 258                 if (boot_cpu_has(X86_FEATURE_NONSTOP_TSC))
 259                         return;
 260
 261                 /*FALL THROUGH*/
 262         default:
 263                 /* TSC could halt in idle, so notify users */
 264                 if (state > ACPI_STATE_C1)
 265                         mark_tsc_unstable("TSC halts in idle");
 266         }
 267 }
 268 #else
 269 static void tsc_check_state(int state) { return; }
 270 #endif
 271
 272 static int acpi_processor_get_power_info_fadt(struct acpi_processor *pr)
 273 {
 274
 275         if (!pr->pblk)
 276                 return -ENODEV;
 277
 278         /* if info is obtained from pblk/fadt, type equals state */
 279         pr->power.states[ACPI_STATE_C2].type = ACPI_STATE_C2;
 280         pr->power.states[ACPI_STATE_C3].type = ACPI_STATE_C3;
 281
 282 #ifndef CONFIG_HOTPLUG_CPU
 283         /*
 284          * Check for P_LVL2_UP flag before entering C2 and above on
 285          * an SMP system.
 286          */
 287         if ((num_online_cpus() > 1) &&
 288             !(acpi_gbl_FADT.flags & ACPI_FADT_C2_MP_SUPPORTED))
 289                 return -ENODEV;
 290 #endif
 291
 292         /* determine C2 and C3 address from pblk */
 293         pr->power.states[ACPI_STATE_C2].address = pr->pblk + 4;
 294         pr->power.states[ACPI_STATE_C3].address = pr->pblk + 5;
 295
 296         /* determine latencies from FADT */
 297         pr->power.states[ACPI_STATE_C2].latency = acpi_gbl_FADT.c2_latency;
 298         pr->power.states[ACPI_STATE_C3].latency = acpi_gbl_FADT.c3_latency;
 299
 300         /*
 301          * FADT specified C2 latency must be less than or equal to
 302          * 100 microseconds.
 303          */
 304         if (acpi_gbl_FADT.c2_latency > ACPI_PROCESSOR_MAX_C2_LATENCY) {
 305                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 306                         "C2 latency too large [%d]\n", acpi_gbl_FADT.c2_latency));
 307                 /* invalidate C2 */
 308                 pr->power.states[ACPI_STATE_C2].address = 0;
 309         }
 310
 311         /*
 312          * FADT supplied C3 latency must be less than or equal to
 313          * 1000 microseconds.
 314          */
 315         if (acpi_gbl_FADT.c3_latency > ACPI_PROCESSOR_MAX_C3_LATENCY) {
 316                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 317                         "C3 latency too large [%d]\n", acpi_gbl_FADT.c3_latency));
 318                 /* invalidate C3 */
 319                 pr->power.states[ACPI_STATE_C3].address = 0;
 320         }
 321
 322         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 323                           "lvl2[0x%08x] lvl3[0x%08x]\n",
 324                           pr->power.states[ACPI_STATE_C2].address,
 325                           pr->power.states[ACPI_STATE_C3].address));
 326
 327         return 0;
 328 }
 329
 330 static int acpi_processor_get_power_info_default(struct acpi_processor *pr)
 331 {
 332         if (!pr->power.states[ACPI_STATE_C1].valid) {
 333                 /* set the first C-State to C1 */
 334                 /* all processors need to support C1 */
 335                 pr->power.states[ACPI_STATE_C1].type = ACPI_STATE_C1;
 336                 pr->power.states[ACPI_STATE_C1].valid = 1;
 337                 pr->power.states[ACPI_STATE_C1].entry_method = ACPI_CSTATE_HALT;
 338         }
 339         /* the C0 state only exists as a filler in our array */
 340         pr->power.states[ACPI_STATE_C0].valid = 1;
 341         return 0;
 342 }
 343
 344 static int acpi_processor_get_power_info_cst(struct acpi_processor *pr)
 345 {
 346         acpi_status status = 0;
 347         u64 count;
 348         int current_count;
 349         int i;
 350         struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
 351         union acpi_object *cst;
 352
 353
 354         if (nocst)
 355                 return -ENODEV;
 356
 357         current_count = 0;
 358
 359         status = acpi_evaluate_object(pr->handle, "_CST", NULL, &buffer);
 360         if (ACPI_FAILURE(status)) {
 361                 ACPI_DEBUG_PRINT((ACPI_DB_INFO, "No _CST, giving up\n"));
 362                 return -ENODEV;
 363         }
 364
 365         cst = buffer.pointer;
 366
 367         /* There must be at least 2 elements */
 368         if (!cst || (cst->type != ACPI_TYPE_PACKAGE) || cst->package.count < 2) {
 369                 printk(KERN_ERR PREFIX "not enough elements in _CST\n");
 370                 status = -EFAULT;
 371                 goto end;
 372         }
 373
 374         count = cst->package.elements[0].integer.value;
 375
 376         /* Validate number of power states. */
 377         if (count < 1 || count != cst->package.count - 1) {
 378                 printk(KERN_ERR PREFIX "count given by _CST is not valid\n");
 379                 status = -EFAULT;
 380                 goto end;
 381         }
 382
 383         /* Tell driver that at least _CST is supported. */
 384         pr->flags.has_cst = 1;
 385
 386         for (i = 1; i <= count; i++) {
 387                 union acpi_object *element;
 388                 union acpi_object *obj;
 389                 struct acpi_power_register *reg;
 390                 struct acpi_processor_cx cx;
 391
 392                 memset(&cx, 0, sizeof(cx));
 393
 394                 element = &(cst->package.elements[i]);
 395                 if (element->type != ACPI_TYPE_PACKAGE)
 396                         continue;
 397
 398                 if (element->package.count != 4)
 399                         continue;
 400
 401                 obj = &(element->package.elements[0]);
 402
 403                 if (obj->type != ACPI_TYPE_BUFFER)
 404                         continue;
 405
 406                 reg = (struct acpi_power_register *)obj->buffer.pointer;
 407
 408                 if (reg->space_id != ACPI_ADR_SPACE_SYSTEM_IO &&
 409                     (reg->space_id != ACPI_ADR_SPACE_FIXED_HARDWARE))
 410                         continue;
 411
 412                 /* There should be an easy way to extract an integer... */
 413                 obj = &(element->package.elements[1]);
 414                 if (obj->type != ACPI_TYPE_INTEGER)
 415                         continue;
 416
 417                 cx.type = obj->integer.value;
 418                 /*
 419                  * Some buggy BIOSes won't list C1 in _CST -
 420                  * Let acpi_processor_get_power_info_default() handle them later
 421                  */
 422                 if (i == 1 && cx.type != ACPI_STATE_C1)
 423                         current_count++;
 424
 425                 cx.address = reg->address;
 426                 cx.index = current_count + 1;
 427
 428                 cx.entry_method = ACPI_CSTATE_SYSTEMIO;
 429                 if (reg->space_id == ACPI_ADR_SPACE_FIXED_HARDWARE) {
 430                         if (acpi_processor_ffh_cstate_probe
 431                                         (pr->id, &cx, reg) == 0) {
 432                                 cx.entry_method = ACPI_CSTATE_FFH;
 433                         } else if (cx.type == ACPI_STATE_C1) {
 434                                 /*
 435                                  * C1 is a special case where FIXED_HARDWARE
 436                                  * can be handled in non-MWAIT way as well.
 437                                  * In that case, save this _CST entry info.
 438                                  * Otherwise, ignore this info and continue.
 439                                  */
 440                                 cx.entry_method = ACPI_CSTATE_HALT;
 441                                 snprintf(cx.desc, ACPI_CX_DESC_LEN, "ACPI HLT");
 442                         } else {
 443                                 continue;
 444                         }
 445                         if (cx.type == ACPI_STATE_C1 &&
 446                             (boot_option_idle_override == IDLE_NOMWAIT)) {
 447                                 /*
 448                                  * In most cases the C1 space_id obtained from
 449                                  * _CST object is FIXED_HARDWARE access mode.
 450                                  * But when the option of idle=halt is added,
 451                                  * the entry_method type should be changed from
 452                                  * CSTATE_FFH to CSTATE_HALT.
 453                                  * When the option of idle=nomwait is added,
 454                                  * the C1 entry_method type should be
 455                                  * CSTATE_HALT.
 456                                  */
 457                                 cx.entry_method = ACPI_CSTATE_HALT;
 458                                 snprintf(cx.desc, ACPI_CX_DESC_LEN, "ACPI HLT");
 459                         }
 460                 } else {
 461                         snprintf(cx.desc, ACPI_CX_DESC_LEN, "ACPI IOPORT 0x%x",
 462                                  cx.address);
 463                 }
 464
 465                 if (cx.type == ACPI_STATE_C1) {
 466                         cx.valid = 1;
 467                 }
 468
 469                 obj = &(element->package.elements[2]);
 470                 if (obj->type != ACPI_TYPE_INTEGER)
 471                         continue;
 472
 473                 cx.latency = obj->integer.value;
 474
 475                 obj = &(element->package.elements[3]);
 476                 if (obj->type != ACPI_TYPE_INTEGER)
 477                         continue;
 478
 479                 current_count++;
 480                 memcpy(&(pr->power.states[current_count]), &cx, sizeof(cx));
 481
 482                 /*
 483                  * We support total ACPI_PROCESSOR_MAX_POWER - 1
 484                  * (From 1 through ACPI_PROCESSOR_MAX_POWER - 1)
 485                  */
 486                 if (current_count >= (ACPI_PROCESSOR_MAX_POWER - 1)) {
 487                         printk(KERN_WARNING
 488                                "Limiting number of power states to max (%d)\n",
 489                                ACPI_PROCESSOR_MAX_POWER);
 490                         printk(KERN_WARNING
 491                                "Please increase ACPI_PROCESSOR_MAX_POWER if needed.\n");
 492                         break;
 493                 }
 494         }
 495
 496         ACPI_DEBUG_PRINT((ACPI_DB_INFO, "Found %d power states\n",
 497                           current_count));
 498
 499         /* Validate number of power states discovered */
 500         if (current_count < 2)
 501                 status = -EFAULT;
 502
 503       end:
 504         kfree(buffer.pointer);
 505
 506         return status;
 507 }
 508
 509 static void acpi_processor_power_verify_c3(struct acpi_processor *pr,
 510                                            struct acpi_processor_cx *cx)
 511 {
 512         static int bm_check_flag = -1;
 513         static int bm_control_flag = -1;
 514
 515
 516         if (!cx->address)
 517                 return;
 518
 519         /*
 520          * PIIX4 Erratum #18: We don't support C3 when Type-F (fast)
 521          * DMA transfers are used by any ISA device to avoid livelock.
 522          * Note that we could disable Type-F DMA (as recommended by
 523          * the erratum), but this is known to disrupt certain ISA
 524          * devices thus we take the conservative approach.
 525          */
 526         else if (errata.piix4.fdma) {
 527                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 528                                   "C3 not supported on PIIX4 with Type-F DMA\n"));
 529                 return;
 530         }
 531
 532         /* All the logic here assumes flags.bm_check is same across all CPUs */
 533         if (bm_check_flag == -1) {
 534                 /* Determine whether bm_check is needed based on CPU  */
 535                 acpi_processor_power_init_bm_check(&(pr->flags), pr->id);
 536                 bm_check_flag = pr->flags.bm_check;
 537                 bm_control_flag = pr->flags.bm_control;
 538         } else {
 539                 pr->flags.bm_check = bm_check_flag;
 540                 pr->flags.bm_control = bm_control_flag;
 541         }
 542
 543         if (pr->flags.bm_check) {
 544                 if (!pr->flags.bm_control) {
 545                         if (pr->flags.has_cst != 1) {
 546                                 /* bus mastering control is necessary */
 547                                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 548                                         "C3 support requires BM control\n"));
 549                                 return;
 550                         } else {
 551                                 /* Here we enter C3 without bus mastering */
 552                                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 553                                         "C3 support without BM control\n"));
 554                         }
 555                 }
 556         } else {
 557                 /*
 558                  * WBINVD should be set in fadt, for C3 state to be
 559                  * supported on when bm_check is not required.
 560                  */
 561                 if (!(acpi_gbl_FADT.flags & ACPI_FADT_WBINVD)) {
 562                         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 563                                           "Cache invalidation should work properly"
 564                                           " for C3 to be enabled on SMP systems\n"));
 565                         return;
 566                 }
 567         }
 568
 569         /*
 570          * Otherwise we've met all of our C3 requirements.
 571          * Normalize the C3 latency to expidite policy.  Enable
 572          * checking of bus mastering status (bm_check) so we can
 573          * use this in our C3 policy
 574          */
 575         cx->valid = 1;
 576
 577         /*
 578          * On older chipsets, BM_RLD needs to be set
 579          * in order for Bus Master activity to wake the
 580          * system from C3.  Newer chipsets handle DMA
 581          * during C3 automatically and BM_RLD is a NOP.
 582          * In either case, the proper way to
 583          * handle BM_RLD is to set it and leave it set.
 584          */
 585         acpi_write_bit_register(ACPI_BITREG_BUS_MASTER_RLD, 1);
 586
 587         return;
 588 }
 589
 590 static int acpi_processor_power_verify(struct acpi_processor *pr)
 591 {
 592         unsigned int i;
 593         unsigned int working = 0;
 594
 595         pr->power.timer_broadcast_on_state = INT_MAX;
 596
 597         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER && i <= max_cstate; i++) {
 598                 struct acpi_processor_cx *cx = &pr->power.states[i];
 599
 600                 switch (cx->type) {
 601                 case ACPI_STATE_C1:
 602                         cx->valid = 1;
 603                         break;
 604
 605                 case ACPI_STATE_C2:
 606                         if (!cx->address)
 607                                 break;
 608                         cx->valid = 1;
 609                         break;
 610
 611                 case ACPI_STATE_C3:
 612                         acpi_processor_power_verify_c3(pr, cx);
 613                         break;
 614                 }
 615                 if (!cx->valid)
 616                         continue;
 617
 618                 lapic_timer_check_state(i, pr, cx);
 619                 tsc_check_state(cx->type);
 620                 working++;
 621         }
 622
 623         lapic_timer_propagate_broadcast(pr);
 624
 625         return (working);
 626 }
 627
 628 static int acpi_processor_get_power_info(struct acpi_processor *pr)
 629 {
 630         unsigned int i;
 631         int result;
 632
 633
 634         /* NOTE: the idle thread may not be running while calling
 635          * this function */
 636
 637         /* Zero initialize all the C-states info. */
 638         memset(pr->power.states, 0, sizeof(pr->power.states));
 639
 640         result = acpi_processor_get_power_info_cst(pr);
 641         if (result == -ENODEV)
 642                 result = acpi_processor_get_power_info_fadt(pr);
 643
 644         if (result)
 645                 return result;
 646
 647         acpi_processor_get_power_info_default(pr);
 648
 649         pr->power.count = acpi_processor_power_verify(pr);
 650
 651         /*
 652          * if one state of type C2 or C3 is available, mark this
 653          * CPU as being "idle manageable"
 654          */
 655         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
 656                 if (pr->power.states[i].valid) {
 657                         pr->power.count = i;
 658                         if (pr->power.states[i].type >= ACPI_STATE_C2)
 659                                 pr->flags.power = 1;
 660                 }
 661         }
 662
 663         return 0;
 664 }
 665
 666 /**
 667  * acpi_idle_bm_check - checks if bus master activity was detected
 668  */
 669 static int acpi_idle_bm_check(void)
 670 {
 671         u32 bm_status = 0;
 672
 673         if (bm_check_disable)
 674                 return 0;
 675
 676         acpi_read_bit_register(ACPI_BITREG_BUS_MASTER_STATUS, &bm_status);
 677         if (bm_status)
 678                 acpi_write_bit_register(ACPI_BITREG_BUS_MASTER_STATUS, 1);
 679         /*
 680          * PIIX4 Erratum #18: Note that BM_STS doesn't always reflect
 681          * the true state of bus mastering activity; forcing us to
 682          * manually check the BMIDEA bit of each IDE channel.
 683          */
 684         else if (errata.piix4.bmisx) {
 685                 if ((inb_p(errata.piix4.bmisx + 0x02) & 0x01)
 686                     || (inb_p(errata.piix4.bmisx + 0x0A) & 0x01))
 687                         bm_status = 1;
 688         }
 689         return bm_status;
 690 }
 691
 692 /**
 693  * acpi_idle_do_entry - a helper function that does C2 and C3 type entry
 694  * @cx: cstate data
 695  *
 696  * Caller disables interrupt before call and enables interrupt after return.
 697  */
 698 static inline void acpi_idle_do_entry(struct acpi_processor_cx *cx)
 699 {
 700         /* Don't trace irqs off for idle */
 701         stop_critical_timings();
 702         if (cx->entry_method == ACPI_CSTATE_FFH) {
 703                 /* Call into architectural FFH based C-state */
 704                 acpi_processor_ffh_cstate_enter(cx);
 705         } else if (cx->entry_method == ACPI_CSTATE_HALT) {
 706                 acpi_safe_halt();
 707         } else {
 708                 /* IO port based C-state */
 709                 inb(cx->address);
 710                 /* Dummy wait op - must do something useless after P_LVL2 read
 711                    because chipsets cannot guarantee that STPCLK# signal
 712                    gets asserted in time to freeze execution properly. */
 713                 inl(acpi_gbl_FADT.xpm_timer_block.address);
 714         }
 715         start_critical_timings();
 716 }
 717
 718 /**
 719  * acpi_idle_enter_c1 - enters an ACPI C1 state-type
 720  * @dev: the target CPU
 721  * @drv: cpuidle driver containing cpuidle state info
 722  * @index: index of target state
 723  *
 724  * This is equivalent to the HALT instruction.
 725  */
 726 static int acpi_idle_enter_c1(struct cpuidle_device *dev,
 727                 struct cpuidle_driver *drv, int index)
 728 {
 729         struct acpi_processor *pr;
 730         struct acpi_processor_cx *cx = per_cpu(acpi_cstate[index], dev->cpu);
 731
 732         pr = __this_cpu_read(processors);
 733
 734         if (unlikely(!pr))
 735                 return -EINVAL;
 736
 737         lapic_timer_state_broadcast(pr, cx, 1);
 738         acpi_idle_do_entry(cx);
 739
 740         lapic_timer_state_broadcast(pr, cx, 0);
 741
 742         return index;
 743 }
 744
 745
 746 /**
 747  * acpi_idle_play_dead - enters an ACPI state for long-term idle (i.e. off-lining)
 748  * @dev: the target CPU
 749  * @index: the index of suggested state
 750  */
 751 static int acpi_idle_play_dead(struct cpuidle_device *dev, int index)
 752 {
 753         struct acpi_processor_cx *cx = per_cpu(acpi_cstate[index], dev->cpu);
 754
 755         ACPI_FLUSH_CPU_CACHE();
 756
 757         while (1) {
 758
 759                 if (cx->entry_method == ACPI_CSTATE_HALT)
 760                         safe_halt();
 761                 else if (cx->entry_method == ACPI_CSTATE_SYSTEMIO) {
 762                         inb(cx->address);
 763                         /* See comment in acpi_idle_do_entry() */
 764                         inl(acpi_gbl_FADT.xpm_timer_block.address);
 765                 } else
 766                         return -ENODEV;
 767         }
 768
 769         /* Never reached */
 770         return 0;
 771 }
 772
 773 /**
 774  * acpi_idle_enter_simple - enters an ACPI state without BM handling
 775  * @dev: the target CPU
 776  * @drv: cpuidle driver with cpuidle state information
 777  * @index: the index of suggested state
 778  */
 779 static int acpi_idle_enter_simple(struct cpuidle_device *dev,
 780                 struct cpuidle_driver *drv, int index)
 781 {
 782         struct acpi_processor *pr;
 783         struct acpi_processor_cx *cx = per_cpu(acpi_cstate[index], dev->cpu);
 784
 785         pr = __this_cpu_read(processors);
 786
 787         if (unlikely(!pr))
 788                 return -EINVAL;
 789
 790         if (cx->entry_method != ACPI_CSTATE_FFH) {
 791                 current_thread_info()->status &= ~TS_POLLING;
 792                 /*
 793                  * TS_POLLING-cleared state must be visible before we test
 794                  * NEED_RESCHED:
 795                  */
 796                 smp_mb();
 797
 798                 if (unlikely(need_resched())) {
 799                         current_thread_info()->status |= TS_POLLING;
 800                         return -EINVAL;
 801                 }
 802         }
 803
 804         /*
 805          * Must be done before busmaster disable as we might need to
 806          * access HPET !
 807          */
 808         lapic_timer_state_broadcast(pr, cx, 1);
 809
 810         if (cx->type == ACPI_STATE_C3)
 811                 ACPI_FLUSH_CPU_CACHE();
 812
 813         /* Tell the scheduler that we are going deep-idle: */
 814         sched_clock_idle_sleep_event();
 815         acpi_idle_do_entry(cx);
 816
 817         sched_clock_idle_wakeup_event(0);
 818
 819         if (cx->entry_method != ACPI_CSTATE_FFH)
 820                 current_thread_info()->status |= TS_POLLING;
 821
 822         lapic_timer_state_broadcast(pr, cx, 0);
 823         return index;
 824 }
 825
 826 static int c3_cpu_count;
 827 static DEFINE_RAW_SPINLOCK(c3_lock);
 828
 829 /**
 830  * acpi_idle_enter_bm - enters C3 with proper BM handling
 831  * @dev: the target CPU
 832  * @drv: cpuidle driver containing state data
 833  * @index: the index of suggested state
 834  *
 835  * If BM is detected, the deepest non-C3 idle state is entered instead.
 836  */
 837 static int acpi_idle_enter_bm(struct cpuidle_device *dev,
 838                 struct cpuidle_driver *drv, int index)
 839 {
 840         struct acpi_processor *pr;
 841         struct acpi_processor_cx *cx = per_cpu(acpi_cstate[index], dev->cpu);
 842
 843         pr = __this_cpu_read(processors);
 844
 845         if (unlikely(!pr))
 846                 return -EINVAL;
 847
 848         if (!cx->bm_sts_skip && acpi_idle_bm_check()) {
 849                 if (drv->safe_state_index >= 0) {
 850                         return drv->states[drv->safe_state_index].enter(dev,
 851                                                 drv, drv->safe_state_index);
 852                 } else {
 853                         acpi_safe_halt();
 854                         return -EBUSY;
 855                 }
 856         }
 857
 858         if (cx->entry_method != ACPI_CSTATE_FFH) {
 859                 current_thread_info()->status &= ~TS_POLLING;
 860                 /*
 861                  * TS_POLLING-cleared state must be visible before we test
 862                  * NEED_RESCHED:
 863                  */
 864                 smp_mb();
 865
 866                 if (unlikely(need_resched())) {
 867                         current_thread_info()->status |= TS_POLLING;
 868                         return -EINVAL;
 869                 }
 870         }
 871
 872         acpi_unlazy_tlb(smp_processor_id());
 873
 874         /* Tell the scheduler that we are going deep-idle: */
 875         sched_clock_idle_sleep_event();
 876         /*
 877          * Must be done before busmaster disable as we might need to
 878          * access HPET !
 879          */
 880         lapic_timer_state_broadcast(pr, cx, 1);
 881
 882         /*
 883          * disable bus master
 884          * bm_check implies we need ARB_DIS
 885          * !bm_check implies we need cache flush
 886          * bm_control implies whether we can do ARB_DIS
 887          *
 888          * That leaves a case where bm_check is set and bm_control is
 889          * not set. In that case we cannot do much, we enter C3
 890          * without doing anything.
 891          */
 892         if (pr->flags.bm_check && pr->flags.bm_control) {
 893                 raw_spin_lock(&c3_lock);
 894                 c3_cpu_count++;
 895                 /* Disable bus master arbitration when all CPUs are in C3 */
 896                 if (c3_cpu_count == num_online_cpus())
 897                         acpi_write_bit_register(ACPI_BITREG_ARB_DISABLE, 1);
 898                 raw_spin_unlock(&c3_lock);
 899         } else if (!pr->flags.bm_check) {
 900                 ACPI_FLUSH_CPU_CACHE();
 901         }
 902
 903         acpi_idle_do_entry(cx);
 904
 905         /* Re-enable bus master arbitration */
 906         if (pr->flags.bm_check && pr->flags.bm_control) {
 907                 raw_spin_lock(&c3_lock);
 908                 acpi_write_bit_register(ACPI_BITREG_ARB_DISABLE, 0);
 909                 c3_cpu_count--;
 910                 raw_spin_unlock(&c3_lock);
 911         }
 912
 913         sched_clock_idle_wakeup_event(0);
 914
 915         if (cx->entry_method != ACPI_CSTATE_FFH)
 916                 current_thread_info()->status |= TS_POLLING;
 917
 918         lapic_timer_state_broadcast(pr, cx, 0);
 919         return index;
 920 }
 921
 922 struct cpuidle_driver acpi_idle_driver = {
 923         .name =         "acpi_idle",
 924         .owner =        THIS_MODULE,
 925 };
 926
 927 /**
 928  * acpi_processor_setup_cpuidle_cx - prepares and configures CPUIDLE
 929  * device i.e. per-cpu data
 930  *
 931  * @pr: the ACPI processor
 932  * @dev : the cpuidle device
 933  */
 934 static int acpi_processor_setup_cpuidle_cx(struct acpi_processor *pr,
 935                                            struct cpuidle_device *dev)
 936 {
 937         int i, count = CPUIDLE_DRIVER_STATE_START;
 938         struct acpi_processor_cx *cx;
 939
 940         if (!pr->flags.power_setup_done)
 941                 return -EINVAL;
 942
 943         if (pr->flags.power == 0) {
 944                 return -EINVAL;
 945         }
 946
 947         if (!dev)
 948                 return -EINVAL;
 949
 950         dev->cpu = pr->id;
 951
 952         if (max_cstate == 0)
 953                 max_cstate = 1;
 954
 955         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER && i <= max_cstate; i++) {
 956                 cx = &pr->power.states[i];
 957
 958                 if (!cx->valid)
 959                         continue;
 960
 961 #ifdef CONFIG_HOTPLUG_CPU
 962                 if ((cx->type != ACPI_STATE_C1) && (num_online_cpus() > 1) &&
 963                     !pr->flags.has_cst &&
 964                     !(acpi_gbl_FADT.flags & ACPI_FADT_C2_MP_SUPPORTED))
 965                         continue;
 966 #endif
 967                 per_cpu(acpi_cstate[count], dev->cpu) = cx;
 968
 969                 count++;
 970                 if (count == CPUIDLE_STATE_MAX)
 971                         break;
 972         }
 973
 974         dev->state_count = count;
 975
 976         if (!count)
 977                 return -EINVAL;
 978
 979         return 0;
 980 }
 981
 982 /**
 983  * acpi_processor_setup_cpuidle states- prepares and configures cpuidle
 984  * global state data i.e. idle routines
 985  *
 986  * @pr: the ACPI processor
 987  */
 988 static int acpi_processor_setup_cpuidle_states(struct acpi_processor *pr)
 989 {
 990         int i, count = CPUIDLE_DRIVER_STATE_START;
 991         struct acpi_processor_cx *cx;
 992         struct cpuidle_state *state;
 993         struct cpuidle_driver *drv = &acpi_idle_driver;
 994
 995         if (!pr->flags.power_setup_done)
 996                 return -EINVAL;
 997
 998         if (pr->flags.power == 0)
 999                 return -EINVAL;
1000
1001         drv->safe_state_index = -1;
1002         for (i = 0; i < CPUIDLE_STATE_MAX; i++) {
1003                 drv->states[i].name[0] = '\0';
1004                 drv->states[i].desc[0] = '\0';
1005         }
1006
1007         if (max_cstate == 0)
1008                 max_cstate = 1;
1009
1010         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER && i <= max_cstate; i++) {
1011                 cx = &pr->power.states[i];
1012
1013                 if (!cx->valid)
1014                         continue;
1015
1016 #ifdef CONFIG_HOTPLUG_CPU
1017                 if ((cx->type != ACPI_STATE_C1) && (num_online_cpus() > 1) &&
1018                     !pr->flags.has_cst &&
1019                     !(acpi_gbl_FADT.flags & ACPI_FADT_C2_MP_SUPPORTED))
1020                         continue;
1021 #endif
1022
1023                 state = &drv->states[count];
1024                 snprintf(state->name, CPUIDLE_NAME_LEN, "C%d", i);
1025                 strncpy(state->desc, cx->desc, CPUIDLE_DESC_LEN);
1026                 state->exit_latency = cx->latency;
1027                 state->target_residency = cx->latency * latency_factor;
1028
1029                 state->flags = 0;
1030                 switch (cx->type) {
1031                         case ACPI_STATE_C1:
1032                         if (cx->entry_method == ACPI_CSTATE_FFH)
1033                                 state->flags |= CPUIDLE_FLAG_TIME_VALID;
1034
1035                         state->enter = acpi_idle_enter_c1;
1036                         state->enter_dead = acpi_idle_play_dead;
1037                         drv->safe_state_index = count;
1038                         break;
1039
1040                         case ACPI_STATE_C2:
1041                         state->flags |= CPUIDLE_FLAG_TIME_VALID;
1042                         state->enter = acpi_idle_enter_simple;
1043                         state->enter_dead = acpi_idle_play_dead;
1044                         drv->safe_state_index = count;
1045                         break;
1046
1047                         case ACPI_STATE_C3:
1048                         state->flags |= CPUIDLE_FLAG_TIME_VALID;
1049                         state->enter = pr->flags.bm_check ?
1050                                         acpi_idle_enter_bm :
1051                                         acpi_idle_enter_simple;
1052                         break;
1053                 }
1054
1055                 count++;
1056                 if (count == CPUIDLE_STATE_MAX)
1057                         break;
1058         }
1059
1060         drv->state_count = count;
1061
1062         if (!count)
1063                 return -EINVAL;
1064
1065         return 0;
1066 }
1067
1068 int acpi_processor_hotplug(struct acpi_processor *pr)
1069 {
1070         int ret = 0;
1071         struct cpuidle_device *dev;
1072
1073         if (disabled_by_idle_boot_param())
1074                 return 0;
1075
1076         if (nocst)
1077                 return -ENODEV;
1078
1079         if (!pr->flags.power_setup_done)
1080                 return -ENODEV;
1081
1082         dev = per_cpu(acpi_cpuidle_device, pr->id);
1083         cpuidle_pause_and_lock();
1084         cpuidle_disable_device(dev);
1085         acpi_processor_get_power_info(pr);
1086         if (pr->flags.power) {
1087                 acpi_processor_setup_cpuidle_cx(pr, dev);
1088                 ret = cpuidle_enable_device(dev);
1089         }
1090         cpuidle_resume_and_unlock();
1091
1092         return ret;
1093 }
1094
1095 int acpi_processor_cst_has_changed(struct acpi_processor *pr)
1096 {
1097         int cpu;
1098         struct acpi_processor *_pr;
1099         struct cpuidle_device *dev;
1100
1101         if (disabled_by_idle_boot_param())
1102                 return 0;
1103
1104         if (nocst)
1105                 return -ENODEV;
1106
1107         if (!pr->flags.power_setup_done)
1108                 return -ENODEV;
1109
1110         /*
1111          * FIXME:  Design the ACPI notification to make it once per
1112          * system instead of once per-cpu.  This condition is a hack
1113          * to make the code that updates C-States be called once.
1114          */
1115
1116         if (pr->id == 0 && cpuidle_get_driver() == &acpi_idle_driver) {
1117
1118                 cpuidle_pause_and_lock();
1119                 /* Protect against cpu-hotplug */
1120                 get_online_cpus();
1121
1122                 /* Disable all cpuidle devices */
1123                 for_each_online_cpu(cpu) {
1124                         _pr = per_cpu(processors, cpu);
1125                         if (!_pr || !_pr->flags.power_setup_done)
1126                                 continue;
1127                         dev = per_cpu(acpi_cpuidle_device, cpu);
1128                         cpuidle_disable_device(dev);
1129                 }
1130
1131                 /* Populate Updated C-state information */
1132                 acpi_processor_get_power_info(pr);
1133                 acpi_processor_setup_cpuidle_states(pr);
1134
1135                 /* Enable all cpuidle devices */
1136                 for_each_online_cpu(cpu) {
1137                         _pr = per_cpu(processors, cpu);
1138                         if (!_pr || !_pr->flags.power_setup_done)
1139                                 continue;
1140                         acpi_processor_get_power_info(_pr);
1141                         if (_pr->flags.power) {
1142                                 dev = per_cpu(acpi_cpuidle_device, cpu);
1143                                 acpi_processor_setup_cpuidle_cx(_pr, dev);
1144                                 cpuidle_enable_device(dev);
1145                         }
1146                 }
1147                 put_online_cpus();
1148                 cpuidle_resume_and_unlock();
1149         }
1150
1151         return 0;
1152 }
1153
1154 static int acpi_processor_registered;
1155
1156 int acpi_processor_power_init(struct acpi_processor *pr)
1157 {
1158         acpi_status status = 0;
1159         int retval;
1160         struct cpuidle_device *dev;
1161         static int first_run;
1162
1163         if (disabled_by_idle_boot_param())
1164                 return 0;
1165
1166         if (!first_run) {
1167                 dmi_check_system(processor_power_dmi_table);
1168                 max_cstate = acpi_processor_cstate_check(max_cstate);
1169                 if (max_cstate < ACPI_C_STATES_MAX)
1170                         printk(KERN_NOTICE
1171                                "ACPI: processor limited to max C-state %d\n",
1172                                max_cstate);
1173                 first_run++;
1174         }
1175
1176         if (acpi_gbl_FADT.cst_control && !nocst) {
1177                 status =
1178                     acpi_os_write_port(acpi_gbl_FADT.smi_command, acpi_gbl_FADT.cst_control, 8);
1179                 if (ACPI_FAILURE(status)) {
1180                         ACPI_EXCEPTION((AE_INFO, status,
1181                                         "Notifying BIOS of _CST ability failed"));
1182                 }
1183         }
1184
1185         acpi_processor_get_power_info(pr);
1186         pr->flags.power_setup_done = 1;
1187
1188         /*
1189          * Install the idle handler if processor power management is supported.
1190          * Note that we use previously set idle handler will be used on
1191          * platforms that only support C1.
1192          */
1193         if (pr->flags.power) {
1194                 /* Register acpi_idle_driver if not already registered */
1195                 if (!acpi_processor_registered) {
1196                         acpi_processor_setup_cpuidle_states(pr);
1197                         retval = cpuidle_register_driver(&acpi_idle_driver);
1198                         if (retval)
1199                                 return retval;
1200                         printk(KERN_DEBUG "ACPI: %s registered with cpuidle\n",
1201                                         acpi_idle_driver.name);
1202                 }
1203
1204                 dev = kzalloc(sizeof(*dev), GFP_KERNEL);
1205                 if (!dev)
1206                         return -ENOMEM;
1207                 per_cpu(acpi_cpuidle_device, pr->id) = dev;
1208
1209                 acpi_processor_setup_cpuidle_cx(pr, dev);
1210
1211                 /* Register per-cpu cpuidle_device. Cpuidle driver
1212                  * must already be registered before registering device
1213                  */
1214                 retval = cpuidle_register_device(dev);
1215                 if (retval) {
1216                         if (acpi_processor_registered == 0)
1217                                 cpuidle_unregister_driver(&acpi_idle_driver);
1218                         return retval;
1219                 }
1220                 acpi_processor_registered++;
1221         }
1222         return 0;
1223 }
1224
1225 int acpi_processor_power_exit(struct acpi_processor *pr)
1226 {
1227         struct cpuidle_device *dev = per_cpu(acpi_cpuidle_device, pr->id);
1228
1229         if (disabled_by_idle_boot_param())
1230                 return 0;
1231
1232         if (pr->flags.power) {
1233                 cpuidle_unregister_device(dev);
1234                 acpi_processor_registered--;
1235                 if (acpi_processor_registered == 0)
1236                         cpuidle_unregister_driver(&acpi_idle_driver);
1237         }
1238
1239         pr->flags.power_setup_done = 0;
1240         return 0;
1241 }