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