Pileus Git - ~andy/linux/blob - 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/kernel.h>
  32 #include <linux/module.h>
  33 #include <linux/init.h>
  34 #include <linux/cpufreq.h>
  35 #include <linux/slab.h>
  36 #include <linux/proc_fs.h>
  37 #include <linux/seq_file.h>
  38 #include <linux/acpi.h>
  39 #include <linux/dmi.h>
  40 #include <linux/moduleparam.h>
  41 #include <linux/sched.h>        /* need_resched() */
  42 #include <linux/pm_qos_params.h>
  43 #include <linux/clockchips.h>
  44 #include <linux/cpuidle.h>
  45 #include <linux/irqflags.h>
  46
  47 /*
  48  * Include the apic definitions for x86 to have the APIC timer related defines
  49  * available also for UP (on SMP it gets magically included via linux/smp.h).
  50  * asm/acpi.h is not an option, as it would require more include magic. Also
  51  * creating an empty asm-ia64/apic.h would just trade pest vs. cholera.
  52  */
  53 #ifdef CONFIG_X86
  54 #include <asm/apic.h>
  55 #endif
  56
  57 #include <asm/io.h>
  58 #include <asm/uaccess.h>
  59
  60 #include <acpi/acpi_bus.h>
  61 #include <acpi/processor.h>
  62 #include <asm/processor.h>
  63
  64 #define PREFIX "ACPI: "
  65
  66 #define ACPI_PROCESSOR_CLASS            "processor"
  67 #define _COMPONENT              ACPI_PROCESSOR_COMPONENT
  68 ACPI_MODULE_NAME("processor_idle");
  69 #define ACPI_PROCESSOR_FILE_POWER       "power"
  70 #define PM_TIMER_TICK_NS                (1000000000ULL/PM_TIMER_FREQUENCY)
  71 #define C2_OVERHEAD                     1       /* 1us */
  72 #define C3_OVERHEAD                     1       /* 1us */
  73 #define PM_TIMER_TICKS_TO_US(p)         (((p) * 1000)/(PM_TIMER_FREQUENCY/1000))
  74
  75 static unsigned int max_cstate __read_mostly = ACPI_PROCESSOR_MAX_POWER;
  76 module_param(max_cstate, uint, 0000);
  77 static unsigned int nocst __read_mostly;
  78 module_param(nocst, uint, 0000);
  79
  80 static unsigned int latency_factor __read_mostly = 2;
  81 module_param(latency_factor, uint, 0644);
  82
  83 #ifdef CONFIG_ACPI_PROCFS
  84 static u64 us_to_pm_timer_ticks(s64 t)
  85 {
  86         return div64_u64(t * PM_TIMER_FREQUENCY, 1000000);
  87 }
  88 #endif
  89
  90 /*
  91  * IBM ThinkPad R40e crashes mysteriously when going into C2 or C3.
  92  * For now disable this. Probably a bug somewhere else.
  93  *
  94  * To skip this limit, boot/load with a large max_cstate limit.
  95  */
  96 static int set_max_cstate(const struct dmi_system_id *id)
  97 {
  98         if (max_cstate > ACPI_PROCESSOR_MAX_POWER)
  99                 return 0;
 100
 101         printk(KERN_NOTICE PREFIX "%s detected - limiting to C%ld max_cstate."
 102                " Override with \"processor.max_cstate=%d\"\n", id->ident,
 103                (long)id->driver_data, ACPI_PROCESSOR_MAX_POWER + 1);
 104
 105         max_cstate = (long)id->driver_data;
 106
 107         return 0;
 108 }
 109
 110 /* Actually this shouldn't be __cpuinitdata, would be better to fix the
 111    callers to only run once -AK */
 112 static struct dmi_system_id __cpuinitdata processor_power_dmi_table[] = {
 113         { set_max_cstate, "Clevo 5600D", {
 114           DMI_MATCH(DMI_BIOS_VENDOR,"Phoenix Technologies LTD"),
 115           DMI_MATCH(DMI_BIOS_VERSION,"SHE845M0.86C.0013.D.0302131307")},
 116          (void *)2},
 117         { set_max_cstate, "Pavilion zv5000", {
 118           DMI_MATCH(DMI_SYS_VENDOR, "Hewlett-Packard"),
 119           DMI_MATCH(DMI_PRODUCT_NAME,"Pavilion zv5000 (DS502A#ABA)")},
 120          (void *)1},
 121         { set_max_cstate, "Asus L8400B", {
 122           DMI_MATCH(DMI_SYS_VENDOR, "ASUSTeK Computer Inc."),
 123           DMI_MATCH(DMI_PRODUCT_NAME,"L8400B series Notebook PC")},
 124          (void *)1},
 125         {},
 126 };
 127
 128
 129 /*
 130  * Callers should disable interrupts before the call and enable
 131  * interrupts after return.
 132  */
 133 static void acpi_safe_halt(void)
 134 {
 135         current_thread_info()->status &= ~TS_POLLING;
 136         /*
 137          * TS_POLLING-cleared state must be visible before we
 138          * test NEED_RESCHED:
 139          */
 140         smp_mb();
 141         if (!need_resched()) {
 142                 safe_halt();
 143                 local_irq_disable();
 144         }
 145         current_thread_info()->status |= TS_POLLING;
 146 }
 147
 148 #ifdef ARCH_APICTIMER_STOPS_ON_C3
 149
 150 /*
 151  * Some BIOS implementations switch to C3 in the published C2 state.
 152  * This seems to be a common problem on AMD boxen, but other vendors
 153  * are affected too. We pick the most conservative approach: we assume
 154  * that the local APIC stops in both C2 and C3.
 155  */
 156 static void lapic_timer_check_state(int state, struct acpi_processor *pr,
 157                                    struct acpi_processor_cx *cx)
 158 {
 159         struct acpi_processor_power *pwr = &pr->power;
 160         u8 type = local_apic_timer_c2_ok ? ACPI_STATE_C3 : ACPI_STATE_C2;
 161
 162         if (cpu_has(&cpu_data(pr->id), X86_FEATURE_ARAT))
 163                 return;
 164
 165         if (boot_cpu_has(X86_FEATURE_AMDC1E))
 166                 type = ACPI_STATE_C1;
 167
 168         /*
 169          * Check, if one of the previous states already marked the lapic
 170          * unstable
 171          */
 172         if (pwr->timer_broadcast_on_state < state)
 173                 return;
 174
 175         if (cx->type >= type)
 176                 pr->power.timer_broadcast_on_state = state;
 177 }
 178
 179 static void __lapic_timer_propagate_broadcast(void *arg)
 180 {
 181         struct acpi_processor *pr = (struct acpi_processor *) arg;
 182         unsigned long reason;
 183
 184         reason = pr->power.timer_broadcast_on_state < INT_MAX ?
 185                 CLOCK_EVT_NOTIFY_BROADCAST_ON : CLOCK_EVT_NOTIFY_BROADCAST_OFF;
 186
 187         clockevents_notify(reason, &pr->id);
 188 }
 189
 190 static void lapic_timer_propagate_broadcast(struct acpi_processor *pr)
 191 {
 192         smp_call_function_single(pr->id, __lapic_timer_propagate_broadcast,
 193                                  (void *)pr, 1);
 194 }
 195
 196 /* Power(C) State timer broadcast control */
 197 static void lapic_timer_state_broadcast(struct acpi_processor *pr,
 198                                        struct acpi_processor_cx *cx,
 199                                        int broadcast)
 200 {
 201         int state = cx - pr->power.states;
 202
 203         if (state >= pr->power.timer_broadcast_on_state) {
 204                 unsigned long reason;
 205
 206                 reason = broadcast ?  CLOCK_EVT_NOTIFY_BROADCAST_ENTER :
 207                         CLOCK_EVT_NOTIFY_BROADCAST_EXIT;
 208                 clockevents_notify(reason, &pr->id);
 209         }
 210 }
 211
 212 #else
 213
 214 static void lapic_timer_check_state(int state, struct acpi_processor *pr,
 215                                    struct acpi_processor_cx *cstate) { }
 216 static void lapic_timer_propagate_broadcast(struct acpi_processor *pr) { }
 217 static void lapic_timer_state_broadcast(struct acpi_processor *pr,
 218                                        struct acpi_processor_cx *cx,
 219                                        int broadcast)
 220 {
 221 }
 222
 223 #endif
 224
 225 /*
 226  * Suspend / resume control
 227  */
 228 static int acpi_idle_suspend;
 229 static u32 saved_bm_rld;
 230
 231 static void acpi_idle_bm_rld_save(void)
 232 {
 233         acpi_read_bit_register(ACPI_BITREG_BUS_MASTER_RLD, &saved_bm_rld);
 234 }
 235 static void acpi_idle_bm_rld_restore(void)
 236 {
 237         u32 resumed_bm_rld;
 238
 239         acpi_read_bit_register(ACPI_BITREG_BUS_MASTER_RLD, &resumed_bm_rld);
 240
 241         if (resumed_bm_rld != saved_bm_rld)
 242                 acpi_write_bit_register(ACPI_BITREG_BUS_MASTER_RLD, saved_bm_rld);
 243 }
 244
 245 int acpi_processor_suspend(struct acpi_device * device, pm_message_t state)
 246 {
 247         if (acpi_idle_suspend == 1)
 248                 return 0;
 249
 250         acpi_idle_bm_rld_save();
 251         acpi_idle_suspend = 1;
 252         return 0;
 253 }
 254
 255 int acpi_processor_resume(struct acpi_device * device)
 256 {
 257         if (acpi_idle_suspend == 0)
 258                 return 0;
 259
 260         acpi_idle_bm_rld_restore();
 261         acpi_idle_suspend = 0;
 262         return 0;
 263 }
 264
 265 #if defined (CONFIG_GENERIC_TIME) && defined (CONFIG_X86)
 266 static void tsc_check_state(int state)
 267 {
 268         switch (boot_cpu_data.x86_vendor) {
 269         case X86_VENDOR_AMD:
 270         case X86_VENDOR_INTEL:
 271                 /*
 272                  * AMD Fam10h TSC will tick in all
 273                  * C/P/S0/S1 states when this bit is set.
 274                  */
 275                 if (boot_cpu_has(X86_FEATURE_NONSTOP_TSC))
 276                         return;
 277
 278                 /*FALL THROUGH*/
 279         default:
 280                 /* TSC could halt in idle, so notify users */
 281                 if (state > ACPI_STATE_C1)
 282                         mark_tsc_unstable("TSC halts in idle");
 283         }
 284 }
 285 #else
 286 static void tsc_check_state(int state) { return; }
 287 #endif
 288
 289 static int acpi_processor_get_power_info_fadt(struct acpi_processor *pr)
 290 {
 291
 292         if (!pr)
 293                 return -EINVAL;
 294
 295         if (!pr->pblk)
 296                 return -ENODEV;
 297
 298         /* if info is obtained from pblk/fadt, type equals state */
 299         pr->power.states[ACPI_STATE_C2].type = ACPI_STATE_C2;
 300         pr->power.states[ACPI_STATE_C3].type = ACPI_STATE_C3;
 301
 302 #ifndef CONFIG_HOTPLUG_CPU
 303         /*
 304          * Check for P_LVL2_UP flag before entering C2 and above on
 305          * an SMP system.
 306          */
 307         if ((num_online_cpus() > 1) &&
 308             !(acpi_gbl_FADT.flags & ACPI_FADT_C2_MP_SUPPORTED))
 309                 return -ENODEV;
 310 #endif
 311
 312         /* determine C2 and C3 address from pblk */
 313         pr->power.states[ACPI_STATE_C2].address = pr->pblk + 4;
 314         pr->power.states[ACPI_STATE_C3].address = pr->pblk + 5;
 315
 316         /* determine latencies from FADT */
 317         pr->power.states[ACPI_STATE_C2].latency = acpi_gbl_FADT.C2latency;
 318         pr->power.states[ACPI_STATE_C3].latency = acpi_gbl_FADT.C3latency;
 319
 320         /*
 321          * FADT specified C2 latency must be less than or equal to
 322          * 100 microseconds.
 323          */
 324         if (acpi_gbl_FADT.C2latency > ACPI_PROCESSOR_MAX_C2_LATENCY) {
 325                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 326                         "C2 latency too large [%d]\n", acpi_gbl_FADT.C2latency));
 327                 /* invalidate C2 */
 328                 pr->power.states[ACPI_STATE_C2].address = 0;
 329         }
 330
 331         /*
 332          * FADT supplied C3 latency must be less than or equal to
 333          * 1000 microseconds.
 334          */
 335         if (acpi_gbl_FADT.C3latency > ACPI_PROCESSOR_MAX_C3_LATENCY) {
 336                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 337                         "C3 latency too large [%d]\n", acpi_gbl_FADT.C3latency));
 338                 /* invalidate C3 */
 339                 pr->power.states[ACPI_STATE_C3].address = 0;
 340         }
 341
 342         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 343                           "lvl2[0x%08x] lvl3[0x%08x]\n",
 344                           pr->power.states[ACPI_STATE_C2].address,
 345                           pr->power.states[ACPI_STATE_C3].address));
 346
 347         return 0;
 348 }
 349
 350 static int acpi_processor_get_power_info_default(struct acpi_processor *pr)
 351 {
 352         if (!pr->power.states[ACPI_STATE_C1].valid) {
 353                 /* set the first C-State to C1 */
 354                 /* all processors need to support C1 */
 355                 pr->power.states[ACPI_STATE_C1].type = ACPI_STATE_C1;
 356                 pr->power.states[ACPI_STATE_C1].valid = 1;
 357                 pr->power.states[ACPI_STATE_C1].entry_method = ACPI_CSTATE_HALT;
 358         }
 359         /* the C0 state only exists as a filler in our array */
 360         pr->power.states[ACPI_STATE_C0].valid = 1;
 361         return 0;
 362 }
 363
 364 static int acpi_processor_get_power_info_cst(struct acpi_processor *pr)
 365 {
 366         acpi_status status = 0;
 367         u64 count;
 368         int current_count;
 369         int i;
 370         struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
 371         union acpi_object *cst;
 372
 373
 374         if (nocst)
 375                 return -ENODEV;
 376
 377         current_count = 0;
 378
 379         status = acpi_evaluate_object(pr->handle, "_CST", NULL, &buffer);
 380         if (ACPI_FAILURE(status)) {
 381                 ACPI_DEBUG_PRINT((ACPI_DB_INFO, "No _CST, giving up\n"));
 382                 return -ENODEV;
 383         }
 384
 385         cst = buffer.pointer;
 386
 387         /* There must be at least 2 elements */
 388         if (!cst || (cst->type != ACPI_TYPE_PACKAGE) || cst->package.count < 2) {
 389                 printk(KERN_ERR PREFIX "not enough elements in _CST\n");
 390                 status = -EFAULT;
 391                 goto end;
 392         }
 393
 394         count = cst->package.elements[0].integer.value;
 395
 396         /* Validate number of power states. */
 397         if (count < 1 || count != cst->package.count - 1) {
 398                 printk(KERN_ERR PREFIX "count given by _CST is not valid\n");
 399                 status = -EFAULT;
 400                 goto end;
 401         }
 402
 403         /* Tell driver that at least _CST is supported. */
 404         pr->flags.has_cst = 1;
 405
 406         for (i = 1; i <= count; i++) {
 407                 union acpi_object *element;
 408                 union acpi_object *obj;
 409                 struct acpi_power_register *reg;
 410                 struct acpi_processor_cx cx;
 411
 412                 memset(&cx, 0, sizeof(cx));
 413
 414                 element = &(cst->package.elements[i]);
 415                 if (element->type != ACPI_TYPE_PACKAGE)
 416                         continue;
 417
 418                 if (element->package.count != 4)
 419                         continue;
 420
 421                 obj = &(element->package.elements[0]);
 422
 423                 if (obj->type != ACPI_TYPE_BUFFER)
 424                         continue;
 425
 426                 reg = (struct acpi_power_register *)obj->buffer.pointer;
 427
 428                 if (reg->space_id != ACPI_ADR_SPACE_SYSTEM_IO &&
 429                     (reg->space_id != ACPI_ADR_SPACE_FIXED_HARDWARE))
 430                         continue;
 431
 432                 /* There should be an easy way to extract an integer... */
 433                 obj = &(element->package.elements[1]);
 434                 if (obj->type != ACPI_TYPE_INTEGER)
 435                         continue;
 436
 437                 cx.type = obj->integer.value;
 438                 /*
 439                  * Some buggy BIOSes won't list C1 in _CST -
 440                  * Let acpi_processor_get_power_info_default() handle them later
 441                  */
 442                 if (i == 1 && cx.type != ACPI_STATE_C1)
 443                         current_count++;
 444
 445                 cx.address = reg->address;
 446                 cx.index = current_count + 1;
 447
 448                 cx.entry_method = ACPI_CSTATE_SYSTEMIO;
 449                 if (reg->space_id == ACPI_ADR_SPACE_FIXED_HARDWARE) {
 450                         if (acpi_processor_ffh_cstate_probe
 451                                         (pr->id, &cx, reg) == 0) {
 452                                 cx.entry_method = ACPI_CSTATE_FFH;
 453                         } else if (cx.type == ACPI_STATE_C1) {
 454                                 /*
 455                                  * C1 is a special case where FIXED_HARDWARE
 456                                  * can be handled in non-MWAIT way as well.
 457                                  * In that case, save this _CST entry info.
 458                                  * Otherwise, ignore this info and continue.
 459                                  */
 460                                 cx.entry_method = ACPI_CSTATE_HALT;
 461                                 snprintf(cx.desc, ACPI_CX_DESC_LEN, "ACPI HLT");
 462                         } else {
 463                                 continue;
 464                         }
 465                         if (cx.type == ACPI_STATE_C1 &&
 466                                         (idle_halt || idle_nomwait)) {
 467                                 /*
 468                                  * In most cases the C1 space_id obtained from
 469                                  * _CST object is FIXED_HARDWARE access mode.
 470                                  * But when the option of idle=halt is added,
 471                                  * the entry_method type should be changed from
 472                                  * CSTATE_FFH to CSTATE_HALT.
 473                                  * When the option of idle=nomwait is added,
 474                                  * the C1 entry_method type should be
 475                                  * CSTATE_HALT.
 476                                  */
 477                                 cx.entry_method = ACPI_CSTATE_HALT;
 478                                 snprintf(cx.desc, ACPI_CX_DESC_LEN, "ACPI HLT");
 479                         }
 480                 } else {
 481                         snprintf(cx.desc, ACPI_CX_DESC_LEN, "ACPI IOPORT 0x%x",
 482                                  cx.address);
 483                 }
 484
 485                 if (cx.type == ACPI_STATE_C1) {
 486                         cx.valid = 1;
 487                 }
 488
 489                 obj = &(element->package.elements[2]);
 490                 if (obj->type != ACPI_TYPE_INTEGER)
 491                         continue;
 492
 493                 cx.latency = obj->integer.value;
 494
 495                 obj = &(element->package.elements[3]);
 496                 if (obj->type != ACPI_TYPE_INTEGER)
 497                         continue;
 498
 499                 cx.power = obj->integer.value;
 500
 501                 current_count++;
 502                 memcpy(&(pr->power.states[current_count]), &cx, sizeof(cx));
 503
 504                 /*
 505                  * We support total ACPI_PROCESSOR_MAX_POWER - 1
 506                  * (From 1 through ACPI_PROCESSOR_MAX_POWER - 1)
 507                  */
 508                 if (current_count >= (ACPI_PROCESSOR_MAX_POWER - 1)) {
 509                         printk(KERN_WARNING
 510                                "Limiting number of power states to max (%d)\n",
 511                                ACPI_PROCESSOR_MAX_POWER);
 512                         printk(KERN_WARNING
 513                                "Please increase ACPI_PROCESSOR_MAX_POWER if needed.\n");
 514                         break;
 515                 }
 516         }
 517
 518         ACPI_DEBUG_PRINT((ACPI_DB_INFO, "Found %d power states\n",
 519                           current_count));
 520
 521         /* Validate number of power states discovered */
 522         if (current_count < 2)
 523                 status = -EFAULT;
 524
 525       end:
 526         kfree(buffer.pointer);
 527
 528         return status;
 529 }
 530
 531 static void acpi_processor_power_verify_c3(struct acpi_processor *pr,
 532                                            struct acpi_processor_cx *cx)
 533 {
 534         static int bm_check_flag = -1;
 535         static int bm_control_flag = -1;
 536
 537
 538         if (!cx->address)
 539                 return;
 540
 541         /*
 542          * PIIX4 Erratum #18: We don't support C3 when Type-F (fast)
 543          * DMA transfers are used by any ISA device to avoid livelock.
 544          * Note that we could disable Type-F DMA (as recommended by
 545          * the erratum), but this is known to disrupt certain ISA
 546          * devices thus we take the conservative approach.
 547          */
 548         else if (errata.piix4.fdma) {
 549                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 550                                   "C3 not supported on PIIX4 with Type-F DMA\n"));
 551                 return;
 552         }
 553
 554         /* All the logic here assumes flags.bm_check is same across all CPUs */
 555         if (bm_check_flag == -1) {
 556                 /* Determine whether bm_check is needed based on CPU  */
 557                 acpi_processor_power_init_bm_check(&(pr->flags), pr->id);
 558                 bm_check_flag = pr->flags.bm_check;
 559                 bm_control_flag = pr->flags.bm_control;
 560         } else {
 561                 pr->flags.bm_check = bm_check_flag;
 562                 pr->flags.bm_control = bm_control_flag;
 563         }
 564
 565         if (pr->flags.bm_check) {
 566                 if (!pr->flags.bm_control) {
 567                         if (pr->flags.has_cst != 1) {
 568                                 /* bus mastering control is necessary */
 569                                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 570                                         "C3 support requires BM control\n"));
 571                                 return;
 572                         } else {
 573                                 /* Here we enter C3 without bus mastering */
 574                                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 575                                         "C3 support without BM control\n"));
 576                         }
 577                 }
 578         } else {
 579                 /*
 580                  * WBINVD should be set in fadt, for C3 state to be
 581                  * supported on when bm_check is not required.
 582                  */
 583                 if (!(acpi_gbl_FADT.flags & ACPI_FADT_WBINVD)) {
 584                         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 585                                           "Cache invalidation should work properly"
 586                                           " for C3 to be enabled on SMP systems\n"));
 587                         return;
 588                 }
 589         }
 590
 591         /*
 592          * Otherwise we've met all of our C3 requirements.
 593          * Normalize the C3 latency to expidite policy.  Enable
 594          * checking of bus mastering status (bm_check) so we can
 595          * use this in our C3 policy
 596          */
 597         cx->valid = 1;
 598
 599         cx->latency_ticks = cx->latency;
 600         /*
 601          * On older chipsets, BM_RLD needs to be set
 602          * in order for Bus Master activity to wake the
 603          * system from C3.  Newer chipsets handle DMA
 604          * during C3 automatically and BM_RLD is a NOP.
 605          * In either case, the proper way to
 606          * handle BM_RLD is to set it and leave it set.
 607          */
 608         acpi_write_bit_register(ACPI_BITREG_BUS_MASTER_RLD, 1);
 609
 610         return;
 611 }
 612
 613 static int acpi_processor_power_verify(struct acpi_processor *pr)
 614 {
 615         unsigned int i;
 616         unsigned int working = 0;
 617
 618         pr->power.timer_broadcast_on_state = INT_MAX;
 619
 620         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER && i <= max_cstate; i++) {
 621                 struct acpi_processor_cx *cx = &pr->power.states[i];
 622
 623                 switch (cx->type) {
 624                 case ACPI_STATE_C1:
 625                         cx->valid = 1;
 626                         break;
 627
 628                 case ACPI_STATE_C2:
 629                         if (!cx->address)
 630                                 break;
 631                         cx->valid = 1;
 632                         cx->latency_ticks = cx->latency; /* Normalize latency */
 633                         break;
 634
 635                 case ACPI_STATE_C3:
 636                         acpi_processor_power_verify_c3(pr, cx);
 637                         break;
 638                 }
 639                 if (!cx->valid)
 640                         continue;
 641
 642                 lapic_timer_check_state(i, pr, cx);
 643                 tsc_check_state(cx->type);
 644                 working++;
 645         }
 646
 647         lapic_timer_propagate_broadcast(pr);
 648
 649         return (working);
 650 }
 651
 652 static int acpi_processor_get_power_info(struct acpi_processor *pr)
 653 {
 654         unsigned int i;
 655         int result;
 656
 657
 658         /* NOTE: the idle thread may not be running while calling
 659          * this function */
 660
 661         /* Zero initialize all the C-states info. */
 662         memset(pr->power.states, 0, sizeof(pr->power.states));
 663
 664         result = acpi_processor_get_power_info_cst(pr);
 665         if (result == -ENODEV)
 666                 result = acpi_processor_get_power_info_fadt(pr);
 667
 668         if (result)
 669                 return result;
 670
 671         acpi_processor_get_power_info_default(pr);
 672
 673         pr->power.count = acpi_processor_power_verify(pr);
 674
 675         /*
 676          * if one state of type C2 or C3 is available, mark this
 677          * CPU as being "idle manageable"
 678          */
 679         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
 680                 if (pr->power.states[i].valid) {
 681                         pr->power.count = i;
 682                         if (pr->power.states[i].type >= ACPI_STATE_C2)
 683                                 pr->flags.power = 1;
 684                 }
 685         }
 686
 687         return 0;
 688 }
 689
 690 #ifdef CONFIG_ACPI_PROCFS
 691 static int acpi_processor_power_seq_show(struct seq_file *seq, void *offset)
 692 {
 693         struct acpi_processor *pr = seq->private;
 694         unsigned int i;
 695
 696
 697         if (!pr)
 698                 goto end;
 699
 700         seq_printf(seq, "active state:            C%zd\n"
 701                    "max_cstate:              C%d\n"
 702                    "maximum allowed latency: %d usec\n",
 703                    pr->power.state ? pr->power.state - pr->power.states : 0,
 704                    max_cstate, pm_qos_request(PM_QOS_CPU_DMA_LATENCY));
 705
 706         seq_puts(seq, "states:\n");
 707
 708         for (i = 1; i <= pr->power.count; i++) {
 709                 seq_printf(seq, "   %cC%d:                  ",
 710                            (&pr->power.states[i] ==
 711                             pr->power.state ? '*' : ' '), i);
 712
 713                 if (!pr->power.states[i].valid) {
 714                         seq_puts(seq, "<not supported>\n");
 715                         continue;
 716                 }
 717
 718                 switch (pr->power.states[i].type) {
 719                 case ACPI_STATE_C1:
 720                         seq_printf(seq, "type[C1] ");
 721                         break;
 722                 case ACPI_STATE_C2:
 723                         seq_printf(seq, "type[C2] ");
 724                         break;
 725                 case ACPI_STATE_C3:
 726                         seq_printf(seq, "type[C3] ");
 727                         break;
 728                 default:
 729                         seq_printf(seq, "type[--] ");
 730                         break;
 731                 }
 732
 733                 seq_puts(seq, "promotion[--] ");
 734
 735                 seq_puts(seq, "demotion[--] ");
 736
 737                 seq_printf(seq, "latency[%03d] usage[%08d] duration[%020Lu]\n",
 738                            pr->power.states[i].latency,
 739                            pr->power.states[i].usage,
 740                            us_to_pm_timer_ticks(pr->power.states[i].time));
 741         }
 742
 743       end:
 744         return 0;
 745 }
 746
 747 static int acpi_processor_power_open_fs(struct inode *inode, struct file *file)
 748 {
 749         return single_open(file, acpi_processor_power_seq_show,
 750                            PDE(inode)->data);
 751 }
 752
 753 static const struct file_operations acpi_processor_power_fops = {
 754         .owner = THIS_MODULE,
 755         .open = acpi_processor_power_open_fs,
 756         .read = seq_read,
 757         .llseek = seq_lseek,
 758         .release = single_release,
 759 };
 760 #endif
 761
 762 /**
 763  * acpi_idle_bm_check - checks if bus master activity was detected
 764  */
 765 static int acpi_idle_bm_check(void)
 766 {
 767         u32 bm_status = 0;
 768
 769         acpi_read_bit_register(ACPI_BITREG_BUS_MASTER_STATUS, &bm_status);
 770         if (bm_status)
 771                 acpi_write_bit_register(ACPI_BITREG_BUS_MASTER_STATUS, 1);
 772         /*
 773          * PIIX4 Erratum #18: Note that BM_STS doesn't always reflect
 774          * the true state of bus mastering activity; forcing us to
 775          * manually check the BMIDEA bit of each IDE channel.
 776          */
 777         else if (errata.piix4.bmisx) {
 778                 if ((inb_p(errata.piix4.bmisx + 0x02) & 0x01)
 779                     || (inb_p(errata.piix4.bmisx + 0x0A) & 0x01))
 780                         bm_status = 1;
 781         }
 782         return bm_status;
 783 }
 784
 785 /**
 786  * acpi_idle_do_entry - a helper function that does C2 and C3 type entry
 787  * @cx: cstate data
 788  *
 789  * Caller disables interrupt before call and enables interrupt after return.
 790  */
 791 static inline void acpi_idle_do_entry(struct acpi_processor_cx *cx)
 792 {
 793         /* Don't trace irqs off for idle */
 794         stop_critical_timings();
 795         if (cx->entry_method == ACPI_CSTATE_FFH) {
 796                 /* Call into architectural FFH based C-state */
 797                 acpi_processor_ffh_cstate_enter(cx);
 798         } else if (cx->entry_method == ACPI_CSTATE_HALT) {
 799                 acpi_safe_halt();
 800         } else {
 801                 int unused;
 802                 /* IO port based C-state */
 803                 inb(cx->address);
 804                 /* Dummy wait op - must do something useless after P_LVL2 read
 805                    because chipsets cannot guarantee that STPCLK# signal
 806                    gets asserted in time to freeze execution properly. */
 807                 unused = inl(acpi_gbl_FADT.xpm_timer_block.address);
 808         }
 809         start_critical_timings();
 810 }
 811
 812 /**
 813  * acpi_idle_enter_c1 - enters an ACPI C1 state-type
 814  * @dev: the target CPU
 815  * @state: the state data
 816  *
 817  * This is equivalent to the HALT instruction.
 818  */
 819 static int acpi_idle_enter_c1(struct cpuidle_device *dev,
 820                               struct cpuidle_state *state)
 821 {
 822         ktime_t  kt1, kt2;
 823         s64 idle_time;
 824         struct acpi_processor *pr;
 825         struct acpi_processor_cx *cx = cpuidle_get_statedata(state);
 826
 827         pr = __get_cpu_var(processors);
 828
 829         if (unlikely(!pr))
 830                 return 0;
 831
 832         local_irq_disable();
 833
 834         /* Do not access any ACPI IO ports in suspend path */
 835         if (acpi_idle_suspend) {
 836                 local_irq_enable();
 837                 cpu_relax();
 838                 return 0;
 839         }
 840
 841         lapic_timer_state_broadcast(pr, cx, 1);
 842         kt1 = ktime_get_real();
 843         acpi_idle_do_entry(cx);
 844         kt2 = ktime_get_real();
 845         idle_time =  ktime_to_us(ktime_sub(kt2, kt1));
 846
 847         local_irq_enable();
 848         cx->usage++;
 849         lapic_timer_state_broadcast(pr, cx, 0);
 850
 851         return idle_time;
 852 }
 853
 854 /**
 855  * acpi_idle_enter_simple - enters an ACPI state without BM handling
 856  * @dev: the target CPU
 857  * @state: the state data
 858  */
 859 static int acpi_idle_enter_simple(struct cpuidle_device *dev,
 860                                   struct cpuidle_state *state)
 861 {
 862         struct acpi_processor *pr;
 863         struct acpi_processor_cx *cx = cpuidle_get_statedata(state);
 864         ktime_t  kt1, kt2;
 865         s64 idle_time_ns;
 866         s64 idle_time;
 867
 868         pr = __get_cpu_var(processors);
 869
 870         if (unlikely(!pr))
 871                 return 0;
 872
 873         if (acpi_idle_suspend)
 874                 return(acpi_idle_enter_c1(dev, state));
 875
 876         local_irq_disable();
 877
 878         if (cx->entry_method != ACPI_CSTATE_FFH) {
 879                 current_thread_info()->status &= ~TS_POLLING;
 880                 /*
 881                  * TS_POLLING-cleared state must be visible before we test
 882                  * NEED_RESCHED:
 883                  */
 884                 smp_mb();
 885
 886                 if (unlikely(need_resched())) {
 887                         current_thread_info()->status |= TS_POLLING;
 888                         local_irq_enable();
 889                         return 0;
 890                 }
 891         }
 892
 893         /*
 894          * Must be done before busmaster disable as we might need to
 895          * access HPET !
 896          */
 897         lapic_timer_state_broadcast(pr, cx, 1);
 898
 899         if (cx->type == ACPI_STATE_C3)
 900                 ACPI_FLUSH_CPU_CACHE();
 901
 902         kt1 = ktime_get_real();
 903         /* Tell the scheduler that we are going deep-idle: */
 904         sched_clock_idle_sleep_event();
 905         acpi_idle_do_entry(cx);
 906         kt2 = ktime_get_real();
 907         idle_time_ns = ktime_to_ns(ktime_sub(kt2, kt1));
 908         idle_time = idle_time_ns;
 909         do_div(idle_time, NSEC_PER_USEC);
 910
 911         /* Tell the scheduler how much we idled: */
 912         sched_clock_idle_wakeup_event(idle_time_ns);
 913
 914         local_irq_enable();
 915         if (cx->entry_method != ACPI_CSTATE_FFH)
 916                 current_thread_info()->status |= TS_POLLING;
 917
 918         cx->usage++;
 919
 920         lapic_timer_state_broadcast(pr, cx, 0);
 921         cx->time += idle_time;
 922         return idle_time;
 923 }
 924
 925 static int c3_cpu_count;
 926 static DEFINE_SPINLOCK(c3_lock);
 927
 928 /**
 929  * acpi_idle_enter_bm - enters C3 with proper BM handling
 930  * @dev: the target CPU
 931  * @state: the state data
 932  *
 933  * If BM is detected, the deepest non-C3 idle state is entered instead.
 934  */
 935 static int acpi_idle_enter_bm(struct cpuidle_device *dev,
 936                               struct cpuidle_state *state)
 937 {
 938         struct acpi_processor *pr;
 939         struct acpi_processor_cx *cx = cpuidle_get_statedata(state);
 940         ktime_t  kt1, kt2;
 941         s64 idle_time_ns;
 942         s64 idle_time;
 943
 944
 945         pr = __get_cpu_var(processors);
 946
 947         if (unlikely(!pr))
 948                 return 0;
 949
 950         if (acpi_idle_suspend)
 951                 return(acpi_idle_enter_c1(dev, state));
 952
 953         if (acpi_idle_bm_check()) {
 954                 if (dev->safe_state) {
 955                         dev->last_state = dev->safe_state;
 956                         return dev->safe_state->enter(dev, dev->safe_state);
 957                 } else {
 958                         local_irq_disable();
 959                         acpi_safe_halt();
 960                         local_irq_enable();
 961                         return 0;
 962                 }
 963         }
 964
 965         local_irq_disable();
 966
 967         if (cx->entry_method != ACPI_CSTATE_FFH) {
 968                 current_thread_info()->status &= ~TS_POLLING;
 969                 /*
 970                  * TS_POLLING-cleared state must be visible before we test
 971                  * NEED_RESCHED:
 972                  */
 973                 smp_mb();
 974
 975                 if (unlikely(need_resched())) {
 976                         current_thread_info()->status |= TS_POLLING;
 977                         local_irq_enable();
 978                         return 0;
 979                 }
 980         }
 981
 982         acpi_unlazy_tlb(smp_processor_id());
 983
 984         /* Tell the scheduler that we are going deep-idle: */
 985         sched_clock_idle_sleep_event();
 986         /*
 987          * Must be done before busmaster disable as we might need to
 988          * access HPET !
 989          */
 990         lapic_timer_state_broadcast(pr, cx, 1);
 991
 992         kt1 = ktime_get_real();
 993         /*
 994          * disable bus master
 995          * bm_check implies we need ARB_DIS
 996          * !bm_check implies we need cache flush
 997          * bm_control implies whether we can do ARB_DIS
 998          *
 999          * That leaves a case where bm_check is set and bm_control is
1000          * not set. In that case we cannot do much, we enter C3
1001          * without doing anything.
1002          */
1003         if (pr->flags.bm_check && pr->flags.bm_control) {
1004                 spin_lock(&c3_lock);
1005                 c3_cpu_count++;
1006                 /* Disable bus master arbitration when all CPUs are in C3 */
1007                 if (c3_cpu_count == num_online_cpus())
1008                         acpi_write_bit_register(ACPI_BITREG_ARB_DISABLE, 1);
1009                 spin_unlock(&c3_lock);
1010         } else if (!pr->flags.bm_check) {
1011                 ACPI_FLUSH_CPU_CACHE();
1012         }
1013
1014         acpi_idle_do_entry(cx);
1015
1016         /* Re-enable bus master arbitration */
1017         if (pr->flags.bm_check && pr->flags.bm_control) {
1018                 spin_lock(&c3_lock);
1019                 acpi_write_bit_register(ACPI_BITREG_ARB_DISABLE, 0);
1020                 c3_cpu_count--;
1021                 spin_unlock(&c3_lock);
1022         }
1023         kt2 = ktime_get_real();
1024         idle_time_ns = ktime_to_ns(ktime_sub(kt2, kt1));
1025         idle_time = idle_time_ns;
1026         do_div(idle_time, NSEC_PER_USEC);
1027
1028         /* Tell the scheduler how much we idled: */
1029         sched_clock_idle_wakeup_event(idle_time_ns);
1030
1031         local_irq_enable();
1032         if (cx->entry_method != ACPI_CSTATE_FFH)
1033                 current_thread_info()->status |= TS_POLLING;
1034
1035         cx->usage++;
1036
1037         lapic_timer_state_broadcast(pr, cx, 0);
1038         cx->time += idle_time;
1039         return idle_time;
1040 }
1041
1042 struct cpuidle_driver acpi_idle_driver = {
1043         .name =         "acpi_idle",
1044         .owner =        THIS_MODULE,
1045 };
1046
1047 /**
1048  * acpi_processor_setup_cpuidle - prepares and configures CPUIDLE
1049  * @pr: the ACPI processor
1050  */
1051 static int acpi_processor_setup_cpuidle(struct acpi_processor *pr)
1052 {
1053         int i, count = CPUIDLE_DRIVER_STATE_START;
1054         struct acpi_processor_cx *cx;
1055         struct cpuidle_state *state;
1056         struct cpuidle_device *dev = &pr->power.dev;
1057
1058         if (!pr->flags.power_setup_done)
1059                 return -EINVAL;
1060
1061         if (pr->flags.power == 0) {
1062                 return -EINVAL;
1063         }
1064
1065         dev->cpu = pr->id;
1066         for (i = 0; i < CPUIDLE_STATE_MAX; i++) {
1067                 dev->states[i].name[0] = '\0';
1068                 dev->states[i].desc[0] = '\0';
1069         }
1070
1071         if (max_cstate == 0)
1072                 max_cstate = 1;
1073
1074         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER && i <= max_cstate; i++) {
1075                 cx = &pr->power.states[i];
1076                 state = &dev->states[count];
1077
1078                 if (!cx->valid)
1079                         continue;
1080
1081 #ifdef CONFIG_HOTPLUG_CPU
1082                 if ((cx->type != ACPI_STATE_C1) && (num_online_cpus() > 1) &&
1083                     !pr->flags.has_cst &&
1084                     !(acpi_gbl_FADT.flags & ACPI_FADT_C2_MP_SUPPORTED))
1085                         continue;
1086 #endif
1087                 cpuidle_set_statedata(state, cx);
1088
1089                 snprintf(state->name, CPUIDLE_NAME_LEN, "C%d", i);
1090                 strncpy(state->desc, cx->desc, CPUIDLE_DESC_LEN);
1091                 state->exit_latency = cx->latency;
1092                 state->target_residency = cx->latency * latency_factor;
1093                 state->power_usage = cx->power;
1094
1095                 state->flags = 0;
1096                 switch (cx->type) {
1097                         case ACPI_STATE_C1:
1098                         state->flags |= CPUIDLE_FLAG_SHALLOW;
1099                         if (cx->entry_method == ACPI_CSTATE_FFH)
1100                                 state->flags |= CPUIDLE_FLAG_TIME_VALID;
1101
1102                         state->enter = acpi_idle_enter_c1;
1103                         dev->safe_state = state;
1104                         break;
1105
1106                         case ACPI_STATE_C2:
1107                         state->flags |= CPUIDLE_FLAG_BALANCED;
1108                         state->flags |= CPUIDLE_FLAG_TIME_VALID;
1109                         state->enter = acpi_idle_enter_simple;
1110                         dev->safe_state = state;
1111                         break;
1112
1113                         case ACPI_STATE_C3:
1114                         state->flags |= CPUIDLE_FLAG_DEEP;
1115                         state->flags |= CPUIDLE_FLAG_TIME_VALID;
1116                         state->flags |= CPUIDLE_FLAG_CHECK_BM;
1117                         state->enter = pr->flags.bm_check ?
1118                                         acpi_idle_enter_bm :
1119                                         acpi_idle_enter_simple;
1120                         break;
1121                 }
1122
1123                 count++;
1124                 if (count == CPUIDLE_STATE_MAX)
1125                         break;
1126         }
1127
1128         dev->state_count = count;
1129
1130         if (!count)
1131                 return -EINVAL;
1132
1133         return 0;
1134 }
1135
1136 int acpi_processor_cst_has_changed(struct acpi_processor *pr)
1137 {
1138         int ret = 0;
1139
1140         if (boot_option_idle_override)
1141                 return 0;
1142
1143         if (!pr)
1144                 return -EINVAL;
1145
1146         if (nocst) {
1147                 return -ENODEV;
1148         }
1149
1150         if (!pr->flags.power_setup_done)
1151                 return -ENODEV;
1152
1153         cpuidle_pause_and_lock();
1154         cpuidle_disable_device(&pr->power.dev);
1155         acpi_processor_get_power_info(pr);
1156         if (pr->flags.power) {
1157                 acpi_processor_setup_cpuidle(pr);
1158                 ret = cpuidle_enable_device(&pr->power.dev);
1159         }
1160         cpuidle_resume_and_unlock();
1161
1162         return ret;
1163 }
1164
1165 int __cpuinit acpi_processor_power_init(struct acpi_processor *pr,
1166                               struct acpi_device *device)
1167 {
1168         acpi_status status = 0;
1169         static int first_run;
1170 #ifdef CONFIG_ACPI_PROCFS
1171         struct proc_dir_entry *entry = NULL;
1172 #endif
1173
1174         if (boot_option_idle_override)
1175                 return 0;
1176
1177         if (!first_run) {
1178                 if (idle_halt) {
1179                         /*
1180                          * When the boot option of "idle=halt" is added, halt
1181                          * is used for CPU IDLE.
1182                          * In such case C2/C3 is meaningless. So the max_cstate
1183                          * is set to one.
1184                          */
1185                         max_cstate = 1;
1186                 }
1187                 dmi_check_system(processor_power_dmi_table);
1188                 max_cstate = acpi_processor_cstate_check(max_cstate);
1189                 if (max_cstate < ACPI_C_STATES_MAX)
1190                         printk(KERN_NOTICE
1191                                "ACPI: processor limited to max C-state %d\n",
1192                                max_cstate);
1193                 first_run++;
1194         }
1195
1196         if (!pr)
1197                 return -EINVAL;
1198
1199         if (acpi_gbl_FADT.cst_control && !nocst) {
1200                 status =
1201                     acpi_os_write_port(acpi_gbl_FADT.smi_command, acpi_gbl_FADT.cst_control, 8);
1202                 if (ACPI_FAILURE(status)) {
1203                         ACPI_EXCEPTION((AE_INFO, status,
1204                                         "Notifying BIOS of _CST ability failed"));
1205                 }
1206         }
1207
1208         acpi_processor_get_power_info(pr);
1209         pr->flags.power_setup_done = 1;
1210
1211         /*
1212          * Install the idle handler if processor power management is supported.
1213          * Note that we use previously set idle handler will be used on
1214          * platforms that only support C1.
1215          */
1216         if (pr->flags.power) {
1217                 acpi_processor_setup_cpuidle(pr);
1218                 if (cpuidle_register_device(&pr->power.dev))
1219                         return -EIO;
1220         }
1221 #ifdef CONFIG_ACPI_PROCFS
1222         /* 'power' [R] */
1223         entry = proc_create_data(ACPI_PROCESSOR_FILE_POWER,
1224                                  S_IRUGO, acpi_device_dir(device),
1225                                  &acpi_processor_power_fops,
1226                                  acpi_driver_data(device));
1227         if (!entry)
1228                 return -EIO;
1229 #endif
1230         return 0;
1231 }
1232
1233 int acpi_processor_power_exit(struct acpi_processor *pr,
1234                               struct acpi_device *device)
1235 {
1236         if (boot_option_idle_override)
1237                 return 0;
1238
1239         cpuidle_unregister_device(&pr->power.dev);
1240         pr->flags.power_setup_done = 0;
1241
1242 #ifdef CONFIG_ACPI_PROCFS
1243         if (acpi_device_dir(device))
1244                 remove_proc_entry(ACPI_PROCESSOR_FILE_POWER,
1245                                   acpi_device_dir(device));
1246 #endif
1247
1248         return 0;
1249 }