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