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/proc_fs.h>
  36 #include <linux/seq_file.h>
  37 #include <linux/acpi.h>
  38 #include <linux/dmi.h>
  39 #include <linux/moduleparam.h>
  40 #include <linux/sched.h>        /* need_resched() */
  41 #include <linux/pm_qos_params.h>
  42 #include <linux/clockchips.h>
  43 #include <linux/cpuidle.h>
  44 #include <linux/irqflags.h>
  45
  46 /*
  47  * Include the apic definitions for x86 to have the APIC timer related defines
  48  * available also for UP (on SMP it gets magically included via linux/smp.h).
  49  * asm/acpi.h is not an option, as it would require more include magic. Also
  50  * creating an empty asm-ia64/apic.h would just trade pest vs. cholera.
  51  */
  52 #ifdef CONFIG_X86
  53 #include <asm/apic.h>
  54 #endif
  55
  56 #include <asm/io.h>
  57 #include <asm/uaccess.h>
  58
  59 #include <acpi/acpi_bus.h>
  60 #include <acpi/processor.h>
  61 #include <asm/processor.h>
  62
  63 #define PREFIX "ACPI: "
  64
  65 #define ACPI_PROCESSOR_CLASS            "processor"
  66 #define _COMPONENT              ACPI_PROCESSOR_COMPONENT
  67 ACPI_MODULE_NAME("processor_idle");
  68 #define ACPI_PROCESSOR_FILE_POWER       "power"
  69 #define PM_TIMER_TICK_NS                (1000000000ULL/PM_TIMER_FREQUENCY)
  70 #define C2_OVERHEAD                     1       /* 1us */
  71 #define C3_OVERHEAD                     1       /* 1us */
  72 #define PM_TIMER_TICKS_TO_US(p)         (((p) * 1000)/(PM_TIMER_FREQUENCY/1000))
  73
  74 static unsigned int max_cstate __read_mostly = ACPI_PROCESSOR_MAX_POWER;
  75 module_param(max_cstate, uint, 0000);
  76 static unsigned int nocst __read_mostly;
  77 module_param(nocst, uint, 0000);
  78
  79 static unsigned int latency_factor __read_mostly = 2;
  80 module_param(latency_factor, uint, 0644);
  81
  82 static s64 us_to_pm_timer_ticks(s64 t)
  83 {
  84         return div64_u64(t * PM_TIMER_FREQUENCY, 1000000);
  85 }
  86 /*
  87  * IBM ThinkPad R40e crashes mysteriously when going into C2 or C3.
  88  * For now disable this. Probably a bug somewhere else.
  89  *
  90  * To skip this limit, boot/load with a large max_cstate limit.
  91  */
  92 static int set_max_cstate(const struct dmi_system_id *id)
  93 {
  94         if (max_cstate > ACPI_PROCESSOR_MAX_POWER)
  95                 return 0;
  96
  97         printk(KERN_NOTICE PREFIX "%s detected - limiting to C%ld max_cstate."
  98                " Override with \"processor.max_cstate=%d\"\n", id->ident,
  99                (long)id->driver_data, ACPI_PROCESSOR_MAX_POWER + 1);
 100
 101         max_cstate = (long)id->driver_data;
 102
 103         return 0;
 104 }
 105
 106 /* Actually this shouldn't be __cpuinitdata, would be better to fix the
 107    callers to only run once -AK */
 108 static struct dmi_system_id __cpuinitdata processor_power_dmi_table[] = {
 109         { set_max_cstate, "Clevo 5600D", {
 110           DMI_MATCH(DMI_BIOS_VENDOR,"Phoenix Technologies LTD"),
 111           DMI_MATCH(DMI_BIOS_VERSION,"SHE845M0.86C.0013.D.0302131307")},
 112          (void *)2},
 113         {},
 114 };
 115
 116
 117 /*
 118  * Callers should disable interrupts before the call and enable
 119  * interrupts after return.
 120  */
 121 static void acpi_safe_halt(void)
 122 {
 123         current_thread_info()->status &= ~TS_POLLING;
 124         /*
 125          * TS_POLLING-cleared state must be visible before we
 126          * test NEED_RESCHED:
 127          */
 128         smp_mb();
 129         if (!need_resched()) {
 130                 safe_halt();
 131                 local_irq_disable();
 132         }
 133         current_thread_info()->status |= TS_POLLING;
 134 }
 135
 136 #ifdef ARCH_APICTIMER_STOPS_ON_C3
 137
 138 /*
 139  * Some BIOS implementations switch to C3 in the published C2 state.
 140  * This seems to be a common problem on AMD boxen, but other vendors
 141  * are affected too. We pick the most conservative approach: we assume
 142  * that the local APIC stops in both C2 and C3.
 143  */
 144 static void lapic_timer_check_state(int state, struct acpi_processor *pr,
 145                                    struct acpi_processor_cx *cx)
 146 {
 147         struct acpi_processor_power *pwr = &pr->power;
 148         u8 type = local_apic_timer_c2_ok ? ACPI_STATE_C3 : ACPI_STATE_C2;
 149
 150         if (cpu_has(&cpu_data(pr->id), X86_FEATURE_ARAT))
 151                 return;
 152
 153         if (boot_cpu_has(X86_FEATURE_AMDC1E))
 154                 type = ACPI_STATE_C1;
 155
 156         /*
 157          * Check, if one of the previous states already marked the lapic
 158          * unstable
 159          */
 160         if (pwr->timer_broadcast_on_state < state)
 161                 return;
 162
 163         if (cx->type >= type)
 164                 pr->power.timer_broadcast_on_state = state;
 165 }
 166
 167 static void __lapic_timer_propagate_broadcast(void *arg)
 168 {
 169         struct acpi_processor *pr = (struct acpi_processor *) arg;
 170         unsigned long reason;
 171
 172         reason = pr->power.timer_broadcast_on_state < INT_MAX ?
 173                 CLOCK_EVT_NOTIFY_BROADCAST_ON : CLOCK_EVT_NOTIFY_BROADCAST_OFF;
 174
 175         clockevents_notify(reason, &pr->id);
 176 }
 177
 178 static void lapic_timer_propagate_broadcast(struct acpi_processor *pr)
 179 {
 180         smp_call_function_single(pr->id, __lapic_timer_propagate_broadcast,
 181                                  (void *)pr, 1);
 182 }
 183
 184 /* Power(C) State timer broadcast control */
 185 static void lapic_timer_state_broadcast(struct acpi_processor *pr,
 186                                        struct acpi_processor_cx *cx,
 187                                        int broadcast)
 188 {
 189         int state = cx - pr->power.states;
 190
 191         if (state >= pr->power.timer_broadcast_on_state) {
 192                 unsigned long reason;
 193
 194                 reason = broadcast ?  CLOCK_EVT_NOTIFY_BROADCAST_ENTER :
 195                         CLOCK_EVT_NOTIFY_BROADCAST_EXIT;
 196                 clockevents_notify(reason, &pr->id);
 197         }
 198 }
 199
 200 #else
 201
 202 static void lapic_timer_check_state(int state, struct acpi_processor *pr,
 203                                    struct acpi_processor_cx *cstate) { }
 204 static void lapic_timer_propagate_broadcast(struct acpi_processor *pr) { }
 205 static void lapic_timer_state_broadcast(struct acpi_processor *pr,
 206                                        struct acpi_processor_cx *cx,
 207                                        int broadcast)
 208 {
 209 }
 210
 211 #endif
 212
 213 /*
 214  * Suspend / resume control
 215  */
 216 static int acpi_idle_suspend;
 217 static u32 saved_bm_rld;
 218
 219 static void acpi_idle_bm_rld_save(void)
 220 {
 221         acpi_read_bit_register(ACPI_BITREG_BUS_MASTER_RLD, &saved_bm_rld);
 222 }
 223 static void acpi_idle_bm_rld_restore(void)
 224 {
 225         u32 resumed_bm_rld;
 226
 227         acpi_read_bit_register(ACPI_BITREG_BUS_MASTER_RLD, &resumed_bm_rld);
 228
 229         if (resumed_bm_rld != saved_bm_rld)
 230                 acpi_write_bit_register(ACPI_BITREG_BUS_MASTER_RLD, saved_bm_rld);
 231 }
 232
 233 int acpi_processor_suspend(struct acpi_device * device, pm_message_t state)
 234 {
 235         if (acpi_idle_suspend == 1)
 236                 return 0;
 237
 238         acpi_idle_bm_rld_save();
 239         acpi_idle_suspend = 1;
 240         return 0;
 241 }
 242
 243 int acpi_processor_resume(struct acpi_device * device)
 244 {
 245         if (acpi_idle_suspend == 0)
 246                 return 0;
 247
 248         acpi_idle_bm_rld_restore();
 249         acpi_idle_suspend = 0;
 250         return 0;
 251 }
 252
 253 #if defined (CONFIG_GENERIC_TIME) && defined (CONFIG_X86)
 254 static void tsc_check_state(int state)
 255 {
 256         switch (boot_cpu_data.x86_vendor) {
 257         case X86_VENDOR_AMD:
 258         case X86_VENDOR_INTEL:
 259                 /*
 260                  * AMD Fam10h TSC will tick in all
 261                  * C/P/S0/S1 states when this bit is set.
 262                  */
 263                 if (boot_cpu_has(X86_FEATURE_NONSTOP_TSC))
 264                         return;
 265
 266                 /*FALL THROUGH*/
 267         default:
 268                 /* TSC could halt in idle, so notify users */
 269                 if (state > ACPI_STATE_C1)
 270                         mark_tsc_unstable("TSC halts in idle");
 271         }
 272 }
 273 #else
 274 static void tsc_check_state(int state) { return; }
 275 #endif
 276
 277 static int acpi_processor_get_power_info_fadt(struct acpi_processor *pr)
 278 {
 279
 280         if (!pr)
 281                 return -EINVAL;
 282
 283         if (!pr->pblk)
 284                 return -ENODEV;
 285
 286         /* if info is obtained from pblk/fadt, type equals state */
 287         pr->power.states[ACPI_STATE_C2].type = ACPI_STATE_C2;
 288         pr->power.states[ACPI_STATE_C3].type = ACPI_STATE_C3;
 289
 290 #ifndef CONFIG_HOTPLUG_CPU
 291         /*
 292          * Check for P_LVL2_UP flag before entering C2 and above on
 293          * an SMP system.
 294          */
 295         if ((num_online_cpus() > 1) &&
 296             !(acpi_gbl_FADT.flags & ACPI_FADT_C2_MP_SUPPORTED))
 297                 return -ENODEV;
 298 #endif
 299
 300         /* determine C2 and C3 address from pblk */
 301         pr->power.states[ACPI_STATE_C2].address = pr->pblk + 4;
 302         pr->power.states[ACPI_STATE_C3].address = pr->pblk + 5;
 303
 304         /* determine latencies from FADT */
 305         pr->power.states[ACPI_STATE_C2].latency = acpi_gbl_FADT.C2latency;
 306         pr->power.states[ACPI_STATE_C3].latency = acpi_gbl_FADT.C3latency;
 307
 308         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 309                           "lvl2[0x%08x] lvl3[0x%08x]\n",
 310                           pr->power.states[ACPI_STATE_C2].address,
 311                           pr->power.states[ACPI_STATE_C3].address));
 312
 313         return 0;
 314 }
 315
 316 static int acpi_processor_get_power_info_default(struct acpi_processor *pr)
 317 {
 318         if (!pr->power.states[ACPI_STATE_C1].valid) {
 319                 /* set the first C-State to C1 */
 320                 /* all processors need to support C1 */
 321                 pr->power.states[ACPI_STATE_C1].type = ACPI_STATE_C1;
 322                 pr->power.states[ACPI_STATE_C1].valid = 1;
 323                 pr->power.states[ACPI_STATE_C1].entry_method = ACPI_CSTATE_HALT;
 324         }
 325         /* the C0 state only exists as a filler in our array */
 326         pr->power.states[ACPI_STATE_C0].valid = 1;
 327         return 0;
 328 }
 329
 330 static int acpi_processor_get_power_info_cst(struct acpi_processor *pr)
 331 {
 332         acpi_status status = 0;
 333         acpi_integer count;
 334         int current_count;
 335         int i;
 336         struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
 337         union acpi_object *cst;
 338
 339
 340         if (nocst)
 341                 return -ENODEV;
 342
 343         current_count = 0;
 344
 345         status = acpi_evaluate_object(pr->handle, "_CST", NULL, &buffer);
 346         if (ACPI_FAILURE(status)) {
 347                 ACPI_DEBUG_PRINT((ACPI_DB_INFO, "No _CST, giving up\n"));
 348                 return -ENODEV;
 349         }
 350
 351         cst = buffer.pointer;
 352
 353         /* There must be at least 2 elements */
 354         if (!cst || (cst->type != ACPI_TYPE_PACKAGE) || cst->package.count < 2) {
 355                 printk(KERN_ERR PREFIX "not enough elements in _CST\n");
 356                 status = -EFAULT;
 357                 goto end;
 358         }
 359
 360         count = cst->package.elements[0].integer.value;
 361
 362         /* Validate number of power states. */
 363         if (count < 1 || count != cst->package.count - 1) {
 364                 printk(KERN_ERR PREFIX "count given by _CST is not valid\n");
 365                 status = -EFAULT;
 366                 goto end;
 367         }
 368
 369         /* Tell driver that at least _CST is supported. */
 370         pr->flags.has_cst = 1;
 371
 372         for (i = 1; i <= count; i++) {
 373                 union acpi_object *element;
 374                 union acpi_object *obj;
 375                 struct acpi_power_register *reg;
 376                 struct acpi_processor_cx cx;
 377
 378                 memset(&cx, 0, sizeof(cx));
 379
 380                 element = &(cst->package.elements[i]);
 381                 if (element->type != ACPI_TYPE_PACKAGE)
 382                         continue;
 383
 384                 if (element->package.count != 4)
 385                         continue;
 386
 387                 obj = &(element->package.elements[0]);
 388
 389                 if (obj->type != ACPI_TYPE_BUFFER)
 390                         continue;
 391
 392                 reg = (struct acpi_power_register *)obj->buffer.pointer;
 393
 394                 if (reg->space_id != ACPI_ADR_SPACE_SYSTEM_IO &&
 395                     (reg->space_id != ACPI_ADR_SPACE_FIXED_HARDWARE))
 396                         continue;
 397
 398                 /* There should be an easy way to extract an integer... */
 399                 obj = &(element->package.elements[1]);
 400                 if (obj->type != ACPI_TYPE_INTEGER)
 401                         continue;
 402
 403                 cx.type = obj->integer.value;
 404                 /*
 405                  * Some buggy BIOSes won't list C1 in _CST -
 406                  * Let acpi_processor_get_power_info_default() handle them later
 407                  */
 408                 if (i == 1 && cx.type != ACPI_STATE_C1)
 409                         current_count++;
 410
 411                 cx.address = reg->address;
 412                 cx.index = current_count + 1;
 413
 414                 cx.entry_method = ACPI_CSTATE_SYSTEMIO;
 415                 if (reg->space_id == ACPI_ADR_SPACE_FIXED_HARDWARE) {
 416                         if (acpi_processor_ffh_cstate_probe
 417                                         (pr->id, &cx, reg) == 0) {
 418                                 cx.entry_method = ACPI_CSTATE_FFH;
 419                         } else if (cx.type == ACPI_STATE_C1) {
 420                                 /*
 421                                  * C1 is a special case where FIXED_HARDWARE
 422                                  * can be handled in non-MWAIT way as well.
 423                                  * In that case, save this _CST entry info.
 424                                  * Otherwise, ignore this info and continue.
 425                                  */
 426                                 cx.entry_method = ACPI_CSTATE_HALT;
 427                                 snprintf(cx.desc, ACPI_CX_DESC_LEN, "ACPI HLT");
 428                         } else {
 429                                 continue;
 430                         }
 431                         if (cx.type == ACPI_STATE_C1 &&
 432                                         (idle_halt || idle_nomwait)) {
 433                                 /*
 434                                  * In most cases the C1 space_id obtained from
 435                                  * _CST object is FIXED_HARDWARE access mode.
 436                                  * But when the option of idle=halt is added,
 437                                  * the entry_method type should be changed from
 438                                  * CSTATE_FFH to CSTATE_HALT.
 439                                  * When the option of idle=nomwait is added,
 440                                  * the C1 entry_method type should be
 441                                  * CSTATE_HALT.
 442                                  */
 443                                 cx.entry_method = ACPI_CSTATE_HALT;
 444                                 snprintf(cx.desc, ACPI_CX_DESC_LEN, "ACPI HLT");
 445                         }
 446                 } else {
 447                         snprintf(cx.desc, ACPI_CX_DESC_LEN, "ACPI IOPORT 0x%x",
 448                                  cx.address);
 449                 }
 450
 451                 if (cx.type == ACPI_STATE_C1) {
 452                         cx.valid = 1;
 453                 }
 454
 455                 obj = &(element->package.elements[2]);
 456                 if (obj->type != ACPI_TYPE_INTEGER)
 457                         continue;
 458
 459                 cx.latency = obj->integer.value;
 460
 461                 obj = &(element->package.elements[3]);
 462                 if (obj->type != ACPI_TYPE_INTEGER)
 463                         continue;
 464
 465                 cx.power = obj->integer.value;
 466
 467                 current_count++;
 468                 memcpy(&(pr->power.states[current_count]), &cx, sizeof(cx));
 469
 470                 /*
 471                  * We support total ACPI_PROCESSOR_MAX_POWER - 1
 472                  * (From 1 through ACPI_PROCESSOR_MAX_POWER - 1)
 473                  */
 474                 if (current_count >= (ACPI_PROCESSOR_MAX_POWER - 1)) {
 475                         printk(KERN_WARNING
 476                                "Limiting number of power states to max (%d)\n",
 477                                ACPI_PROCESSOR_MAX_POWER);
 478                         printk(KERN_WARNING
 479                                "Please increase ACPI_PROCESSOR_MAX_POWER if needed.\n");
 480                         break;
 481                 }
 482         }
 483
 484         ACPI_DEBUG_PRINT((ACPI_DB_INFO, "Found %d power states\n",
 485                           current_count));
 486
 487         /* Validate number of power states discovered */
 488         if (current_count < 2)
 489                 status = -EFAULT;
 490
 491       end:
 492         kfree(buffer.pointer);
 493
 494         return status;
 495 }
 496
 497 static void acpi_processor_power_verify_c2(struct acpi_processor_cx *cx)
 498 {
 499
 500         if (!cx->address)
 501                 return;
 502
 503         /*
 504          * C2 latency must be less than or equal to 100
 505          * microseconds.
 506          */
 507         else if (cx->latency > ACPI_PROCESSOR_MAX_C2_LATENCY) {
 508                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 509                                   "latency too large [%d]\n", cx->latency));
 510                 return;
 511         }
 512
 513         /*
 514          * Otherwise we've met all of our C2 requirements.
 515          * Normalize the C2 latency to expidite policy
 516          */
 517         cx->valid = 1;
 518
 519         cx->latency_ticks = cx->latency;
 520
 521         return;
 522 }
 523
 524 static void acpi_processor_power_verify_c3(struct acpi_processor *pr,
 525                                            struct acpi_processor_cx *cx)
 526 {
 527         static int bm_check_flag = -1;
 528         static int bm_control_flag = -1;
 529
 530
 531         if (!cx->address)
 532                 return;
 533
 534         /*
 535          * C3 latency must be less than or equal to 1000
 536          * microseconds.
 537          */
 538         else if (cx->latency > ACPI_PROCESSOR_MAX_C3_LATENCY) {
 539                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 540                                   "latency too large [%d]\n", cx->latency));
 541                 return;
 542         }
 543
 544         /*
 545          * PIIX4 Erratum #18: We don't support C3 when Type-F (fast)
 546          * DMA transfers are used by any ISA device to avoid livelock.
 547          * Note that we could disable Type-F DMA (as recommended by
 548          * the erratum), but this is known to disrupt certain ISA
 549          * devices thus we take the conservative approach.
 550          */
 551         else if (errata.piix4.fdma) {
 552                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 553                                   "C3 not supported on PIIX4 with Type-F DMA\n"));
 554                 return;
 555         }
 556
 557         /* All the logic here assumes flags.bm_check is same across all CPUs */
 558         if (bm_check_flag == -1) {
 559                 /* Determine whether bm_check is needed based on CPU  */
 560                 acpi_processor_power_init_bm_check(&(pr->flags), pr->id);
 561                 bm_check_flag = pr->flags.bm_check;
 562                 bm_control_flag = pr->flags.bm_control;
 563         } else {
 564                 pr->flags.bm_check = bm_check_flag;
 565                 pr->flags.bm_control = bm_control_flag;
 566         }
 567
 568         if (pr->flags.bm_check) {
 569                 if (!pr->flags.bm_control) {
 570                         if (pr->flags.has_cst != 1) {
 571                                 /* bus mastering control is necessary */
 572                                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 573                                         "C3 support requires BM control\n"));
 574                                 return;
 575                         } else {
 576                                 /* Here we enter C3 without bus mastering */
 577                                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 578                                         "C3 support without BM control\n"));
 579                         }
 580                 }
 581         } else {
 582                 /*
 583                  * WBINVD should be set in fadt, for C3 state to be
 584                  * supported on when bm_check is not required.
 585                  */
 586                 if (!(acpi_gbl_FADT.flags & ACPI_FADT_WBINVD)) {
 587                         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 588                                           "Cache invalidation should work properly"
 589                                           " for C3 to be enabled on SMP systems\n"));
 590                         return;
 591                 }
 592         }
 593
 594         /*
 595          * Otherwise we've met all of our C3 requirements.
 596          * Normalize the C3 latency to expidite policy.  Enable
 597          * checking of bus mastering status (bm_check) so we can
 598          * use this in our C3 policy
 599          */
 600         cx->valid = 1;
 601
 602         cx->latency_ticks = cx->latency;
 603         /*
 604          * On older chipsets, BM_RLD needs to be set
 605          * in order for Bus Master activity to wake the
 606          * system from C3.  Newer chipsets handle DMA
 607          * during C3 automatically and BM_RLD is a NOP.
 608          * In either case, the proper way to
 609          * handle BM_RLD is to set it and leave it set.
 610          */
 611         acpi_write_bit_register(ACPI_BITREG_BUS_MASTER_RLD, 1);
 612
 613         return;
 614 }
 615
 616 static int acpi_processor_power_verify(struct acpi_processor *pr)
 617 {
 618         unsigned int i;
 619         unsigned int working = 0;
 620
 621         pr->power.timer_broadcast_on_state = INT_MAX;
 622
 623         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER && i <= max_cstate; i++) {
 624                 struct acpi_processor_cx *cx = &pr->power.states[i];
 625
 626                 switch (cx->type) {
 627                 case ACPI_STATE_C1:
 628                         cx->valid = 1;
 629                         break;
 630
 631                 case ACPI_STATE_C2:
 632                         acpi_processor_power_verify_c2(cx);
 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_requirement(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                 if (pr->power.states[i].promotion.state)
 734                         seq_printf(seq, "promotion[C%zd] ",
 735                                    (pr->power.states[i].promotion.state -
 736                                     pr->power.states));
 737                 else
 738                         seq_puts(seq, "promotion[--] ");
 739
 740                 if (pr->power.states[i].demotion.state)
 741                         seq_printf(seq, "demotion[C%zd] ",
 742                                    (pr->power.states[i].demotion.state -
 743                                     pr->power.states));
 744                 else
 745                         seq_puts(seq, "demotion[--] ");
 746
 747                 seq_printf(seq, "latency[%03d] usage[%08d] duration[%020llu]\n",
 748                            pr->power.states[i].latency,
 749                            pr->power.states[i].usage,
 750                            (unsigned long long)pr->power.states[i].time);
 751         }
 752
 753       end:
 754         return 0;
 755 }
 756
 757 static int acpi_processor_power_open_fs(struct inode *inode, struct file *file)
 758 {
 759         return single_open(file, acpi_processor_power_seq_show,
 760                            PDE(inode)->data);
 761 }
 762
 763 static const struct file_operations acpi_processor_power_fops = {
 764         .owner = THIS_MODULE,
 765         .open = acpi_processor_power_open_fs,
 766         .read = seq_read,
 767         .llseek = seq_lseek,
 768         .release = single_release,
 769 };
 770 #endif
 771
 772 /**
 773  * acpi_idle_bm_check - checks if bus master activity was detected
 774  */
 775 static int acpi_idle_bm_check(void)
 776 {
 777         u32 bm_status = 0;
 778
 779         acpi_read_bit_register(ACPI_BITREG_BUS_MASTER_STATUS, &bm_status);
 780         if (bm_status)
 781                 acpi_write_bit_register(ACPI_BITREG_BUS_MASTER_STATUS, 1);
 782         /*
 783          * PIIX4 Erratum #18: Note that BM_STS doesn't always reflect
 784          * the true state of bus mastering activity; forcing us to
 785          * manually check the BMIDEA bit of each IDE channel.
 786          */
 787         else if (errata.piix4.bmisx) {
 788                 if ((inb_p(errata.piix4.bmisx + 0x02) & 0x01)
 789                     || (inb_p(errata.piix4.bmisx + 0x0A) & 0x01))
 790                         bm_status = 1;
 791         }
 792         return bm_status;
 793 }
 794
 795 /**
 796  * acpi_idle_do_entry - a helper function that does C2 and C3 type entry
 797  * @cx: cstate data
 798  *
 799  * Caller disables interrupt before call and enables interrupt after return.
 800  */
 801 static inline void acpi_idle_do_entry(struct acpi_processor_cx *cx)
 802 {
 803         /* Don't trace irqs off for idle */
 804         stop_critical_timings();
 805         if (cx->entry_method == ACPI_CSTATE_FFH) {
 806                 /* Call into architectural FFH based C-state */
 807                 acpi_processor_ffh_cstate_enter(cx);
 808         } else if (cx->entry_method == ACPI_CSTATE_HALT) {
 809                 acpi_safe_halt();
 810         } else {
 811                 int unused;
 812                 /* IO port based C-state */
 813                 inb(cx->address);
 814                 /* Dummy wait op - must do something useless after P_LVL2 read
 815                    because chipsets cannot guarantee that STPCLK# signal
 816                    gets asserted in time to freeze execution properly. */
 817                 unused = inl(acpi_gbl_FADT.xpm_timer_block.address);
 818         }
 819         start_critical_timings();
 820 }
 821
 822 /**
 823  * acpi_idle_enter_c1 - enters an ACPI C1 state-type
 824  * @dev: the target CPU
 825  * @state: the state data
 826  *
 827  * This is equivalent to the HALT instruction.
 828  */
 829 static int acpi_idle_enter_c1(struct cpuidle_device *dev,
 830                               struct cpuidle_state *state)
 831 {
 832         ktime_t  kt1, kt2;
 833         s64 idle_time;
 834         struct acpi_processor *pr;
 835         struct acpi_processor_cx *cx = cpuidle_get_statedata(state);
 836
 837         pr = __get_cpu_var(processors);
 838
 839         if (unlikely(!pr))
 840                 return 0;
 841
 842         local_irq_disable();
 843
 844         /* Do not access any ACPI IO ports in suspend path */
 845         if (acpi_idle_suspend) {
 846                 local_irq_enable();
 847                 cpu_relax();
 848                 return 0;
 849         }
 850
 851         lapic_timer_state_broadcast(pr, cx, 1);
 852         kt1 = ktime_get_real();
 853         acpi_idle_do_entry(cx);
 854         kt2 = ktime_get_real();
 855         idle_time =  ktime_to_us(ktime_sub(kt2, kt1));
 856
 857         local_irq_enable();
 858         cx->usage++;
 859         lapic_timer_state_broadcast(pr, cx, 0);
 860
 861         return idle_time;
 862 }
 863
 864 /**
 865  * acpi_idle_enter_simple - enters an ACPI state without BM handling
 866  * @dev: the target CPU
 867  * @state: the state data
 868  */
 869 static int acpi_idle_enter_simple(struct cpuidle_device *dev,
 870                                   struct cpuidle_state *state)
 871 {
 872         struct acpi_processor *pr;
 873         struct acpi_processor_cx *cx = cpuidle_get_statedata(state);
 874         ktime_t  kt1, kt2;
 875         s64 idle_time;
 876         s64 sleep_ticks = 0;
 877
 878         pr = __get_cpu_var(processors);
 879
 880         if (unlikely(!pr))
 881                 return 0;
 882
 883         if (acpi_idle_suspend)
 884                 return(acpi_idle_enter_c1(dev, state));
 885
 886         local_irq_disable();
 887         current_thread_info()->status &= ~TS_POLLING;
 888         /*
 889          * TS_POLLING-cleared state must be visible before we test
 890          * NEED_RESCHED:
 891          */
 892         smp_mb();
 893
 894         if (unlikely(need_resched())) {
 895                 current_thread_info()->status |= TS_POLLING;
 896                 local_irq_enable();
 897                 return 0;
 898         }
 899
 900         /*
 901          * Must be done before busmaster disable as we might need to
 902          * access HPET !
 903          */
 904         lapic_timer_state_broadcast(pr, cx, 1);
 905
 906         if (cx->type == ACPI_STATE_C3)
 907                 ACPI_FLUSH_CPU_CACHE();
 908
 909         kt1 = ktime_get_real();
 910         /* Tell the scheduler that we are going deep-idle: */
 911         sched_clock_idle_sleep_event();
 912         acpi_idle_do_entry(cx);
 913         kt2 = ktime_get_real();
 914         idle_time =  ktime_to_us(ktime_sub(kt2, kt1));
 915
 916         sleep_ticks = us_to_pm_timer_ticks(idle_time);
 917
 918         /* Tell the scheduler how much we idled: */
 919         sched_clock_idle_wakeup_event(sleep_ticks*PM_TIMER_TICK_NS);
 920
 921         local_irq_enable();
 922         current_thread_info()->status |= TS_POLLING;
 923
 924         cx->usage++;
 925
 926         lapic_timer_state_broadcast(pr, cx, 0);
 927         cx->time += sleep_ticks;
 928         return idle_time;
 929 }
 930
 931 static int c3_cpu_count;
 932 static DEFINE_SPINLOCK(c3_lock);
 933
 934 /**
 935  * acpi_idle_enter_bm - enters C3 with proper BM handling
 936  * @dev: the target CPU
 937  * @state: the state data
 938  *
 939  * If BM is detected, the deepest non-C3 idle state is entered instead.
 940  */
 941 static int acpi_idle_enter_bm(struct cpuidle_device *dev,
 942                               struct cpuidle_state *state)
 943 {
 944         struct acpi_processor *pr;
 945         struct acpi_processor_cx *cx = cpuidle_get_statedata(state);
 946         ktime_t  kt1, kt2;
 947         s64 idle_time;
 948         s64 sleep_ticks = 0;
 949
 950
 951         pr = __get_cpu_var(processors);
 952
 953         if (unlikely(!pr))
 954                 return 0;
 955
 956         if (acpi_idle_suspend)
 957                 return(acpi_idle_enter_c1(dev, state));
 958
 959         if (acpi_idle_bm_check()) {
 960                 if (dev->safe_state) {
 961                         dev->last_state = dev->safe_state;
 962                         return dev->safe_state->enter(dev, dev->safe_state);
 963                 } else {
 964                         local_irq_disable();
 965                         acpi_safe_halt();
 966                         local_irq_enable();
 967                         return 0;
 968                 }
 969         }
 970
 971         local_irq_disable();
 972         current_thread_info()->status &= ~TS_POLLING;
 973         /*
 974          * TS_POLLING-cleared state must be visible before we test
 975          * NEED_RESCHED:
 976          */
 977         smp_mb();
 978
 979         if (unlikely(need_resched())) {
 980                 current_thread_info()->status |= TS_POLLING;
 981                 local_irq_enable();
 982                 return 0;
 983         }
 984
 985         acpi_unlazy_tlb(smp_processor_id());
 986
 987         /* Tell the scheduler that we are going deep-idle: */
 988         sched_clock_idle_sleep_event();
 989         /*
 990          * Must be done before busmaster disable as we might need to
 991          * access HPET !
 992          */
 993         lapic_timer_state_broadcast(pr, cx, 1);
 994
 995         kt1 = ktime_get_real();
 996         /*
 997          * disable bus master
 998          * bm_check implies we need ARB_DIS
 999          * !bm_check implies we need cache flush
1000          * bm_control implies whether we can do ARB_DIS
1001          *
1002          * That leaves a case where bm_check is set and bm_control is
1003          * not set. In that case we cannot do much, we enter C3
1004          * without doing anything.
1005          */
1006         if (pr->flags.bm_check && pr->flags.bm_control) {
1007                 spin_lock(&c3_lock);
1008                 c3_cpu_count++;
1009                 /* Disable bus master arbitration when all CPUs are in C3 */
1010                 if (c3_cpu_count == num_online_cpus())
1011                         acpi_write_bit_register(ACPI_BITREG_ARB_DISABLE, 1);
1012                 spin_unlock(&c3_lock);
1013         } else if (!pr->flags.bm_check) {
1014                 ACPI_FLUSH_CPU_CACHE();
1015         }
1016
1017         acpi_idle_do_entry(cx);
1018
1019         /* Re-enable bus master arbitration */
1020         if (pr->flags.bm_check && pr->flags.bm_control) {
1021                 spin_lock(&c3_lock);
1022                 acpi_write_bit_register(ACPI_BITREG_ARB_DISABLE, 0);
1023                 c3_cpu_count--;
1024                 spin_unlock(&c3_lock);
1025         }
1026         kt2 = ktime_get_real();
1027         idle_time =  ktime_to_us(ktime_sub(kt2, kt1));
1028
1029         sleep_ticks = us_to_pm_timer_ticks(idle_time);
1030         /* Tell the scheduler how much we idled: */
1031         sched_clock_idle_wakeup_event(sleep_ticks*PM_TIMER_TICK_NS);
1032
1033         local_irq_enable();
1034         current_thread_info()->status |= TS_POLLING;
1035
1036         cx->usage++;
1037
1038         lapic_timer_state_broadcast(pr, cx, 0);
1039         cx->time += sleep_ticks;
1040         return idle_time;
1041 }
1042
1043 struct cpuidle_driver acpi_idle_driver = {
1044         .name =         "acpi_idle",
1045         .owner =        THIS_MODULE,
1046 };
1047
1048 /**
1049  * acpi_processor_setup_cpuidle - prepares and configures CPUIDLE
1050  * @pr: the ACPI processor
1051  */
1052 static int acpi_processor_setup_cpuidle(struct acpi_processor *pr)
1053 {
1054         int i, count = CPUIDLE_DRIVER_STATE_START;
1055         struct acpi_processor_cx *cx;
1056         struct cpuidle_state *state;
1057         struct cpuidle_device *dev = &pr->power.dev;
1058
1059         if (!pr->flags.power_setup_done)
1060                 return -EINVAL;
1061
1062         if (pr->flags.power == 0) {
1063                 return -EINVAL;
1064         }
1065
1066         dev->cpu = pr->id;
1067         for (i = 0; i < CPUIDLE_STATE_MAX; i++) {
1068                 dev->states[i].name[0] = '\0';
1069                 dev->states[i].desc[0] = '\0';
1070         }
1071
1072         if (max_cstate == 0)
1073                 max_cstate = 1;
1074
1075         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER && i <= max_cstate; i++) {
1076                 cx = &pr->power.states[i];
1077                 state = &dev->states[count];
1078
1079                 if (!cx->valid)
1080                         continue;
1081
1082 #ifdef CONFIG_HOTPLUG_CPU
1083                 if ((cx->type != ACPI_STATE_C1) && (num_online_cpus() > 1) &&
1084                     !pr->flags.has_cst &&
1085                     !(acpi_gbl_FADT.flags & ACPI_FADT_C2_MP_SUPPORTED))
1086                         continue;
1087 #endif
1088                 cpuidle_set_statedata(state, cx);
1089
1090                 snprintf(state->name, CPUIDLE_NAME_LEN, "C%d", i);
1091                 strncpy(state->desc, cx->desc, CPUIDLE_DESC_LEN);
1092                 state->exit_latency = cx->latency;
1093                 state->target_residency = cx->latency * latency_factor;
1094                 state->power_usage = cx->power;
1095
1096                 state->flags = 0;
1097                 switch (cx->type) {
1098                         case ACPI_STATE_C1:
1099                         state->flags |= CPUIDLE_FLAG_SHALLOW;
1100                         if (cx->entry_method == ACPI_CSTATE_FFH)
1101                                 state->flags |= CPUIDLE_FLAG_TIME_VALID;
1102
1103                         state->enter = acpi_idle_enter_c1;
1104                         dev->safe_state = state;
1105                         break;
1106
1107                         case ACPI_STATE_C2:
1108                         state->flags |= CPUIDLE_FLAG_BALANCED;
1109                         state->flags |= CPUIDLE_FLAG_TIME_VALID;
1110                         state->enter = acpi_idle_enter_simple;
1111                         dev->safe_state = state;
1112                         break;
1113
1114                         case ACPI_STATE_C3:
1115                         state->flags |= CPUIDLE_FLAG_DEEP;
1116                         state->flags |= CPUIDLE_FLAG_TIME_VALID;
1117                         state->flags |= CPUIDLE_FLAG_CHECK_BM;
1118                         state->enter = pr->flags.bm_check ?
1119                                         acpi_idle_enter_bm :
1120                                         acpi_idle_enter_simple;
1121                         break;
1122                 }
1123
1124                 count++;
1125                 if (count == CPUIDLE_STATE_MAX)
1126                         break;
1127         }
1128
1129         dev->state_count = count;
1130
1131         if (!count)
1132                 return -EINVAL;
1133
1134         return 0;
1135 }
1136
1137 int acpi_processor_cst_has_changed(struct acpi_processor *pr)
1138 {
1139         int ret = 0;
1140
1141         if (boot_option_idle_override)
1142                 return 0;
1143
1144         if (!pr)
1145                 return -EINVAL;
1146
1147         if (nocst) {
1148                 return -ENODEV;
1149         }
1150
1151         if (!pr->flags.power_setup_done)
1152                 return -ENODEV;
1153
1154         cpuidle_pause_and_lock();
1155         cpuidle_disable_device(&pr->power.dev);
1156         acpi_processor_get_power_info(pr);
1157         if (pr->flags.power) {
1158                 acpi_processor_setup_cpuidle(pr);
1159                 ret = cpuidle_enable_device(&pr->power.dev);
1160         }
1161         cpuidle_resume_and_unlock();
1162
1163         return ret;
1164 }
1165
1166 int __cpuinit acpi_processor_power_init(struct acpi_processor *pr,
1167                               struct acpi_device *device)
1168 {
1169         acpi_status status = 0;
1170         static int first_run;
1171 #ifdef CONFIG_ACPI_PROCFS
1172         struct proc_dir_entry *entry = NULL;
1173 #endif
1174
1175         if (boot_option_idle_override)
1176                 return 0;
1177
1178         if (!first_run) {
1179                 if (idle_halt) {
1180                         /*
1181                          * When the boot option of "idle=halt" is added, halt
1182                          * is used for CPU IDLE.
1183                          * In such case C2/C3 is meaningless. So the max_cstate
1184                          * is set to one.
1185                          */
1186                         max_cstate = 1;
1187                 }
1188                 dmi_check_system(processor_power_dmi_table);
1189                 max_cstate = acpi_processor_cstate_check(max_cstate);
1190                 if (max_cstate < ACPI_C_STATES_MAX)
1191                         printk(KERN_NOTICE
1192                                "ACPI: processor limited to max C-state %d\n",
1193                                max_cstate);
1194                 first_run++;
1195         }
1196
1197         if (!pr)
1198                 return -EINVAL;
1199
1200         if (acpi_gbl_FADT.cst_control && !nocst) {
1201                 status =
1202                     acpi_os_write_port(acpi_gbl_FADT.smi_command, acpi_gbl_FADT.cst_control, 8);
1203                 if (ACPI_FAILURE(status)) {
1204                         ACPI_EXCEPTION((AE_INFO, status,
1205                                         "Notifying BIOS of _CST ability failed"));
1206                 }
1207         }
1208
1209         acpi_processor_get_power_info(pr);
1210         pr->flags.power_setup_done = 1;
1211
1212         /*
1213          * Install the idle handler if processor power management is supported.
1214          * Note that we use previously set idle handler will be used on
1215          * platforms that only support C1.
1216          */
1217         if (pr->flags.power) {
1218                 acpi_processor_setup_cpuidle(pr);
1219                 if (cpuidle_register_device(&pr->power.dev))
1220                         return -EIO;
1221         }
1222 #ifdef CONFIG_ACPI_PROCFS
1223         /* 'power' [R] */
1224         entry = proc_create_data(ACPI_PROCESSOR_FILE_POWER,
1225                                  S_IRUGO, acpi_device_dir(device),
1226                                  &acpi_processor_power_fops,
1227                                  acpi_driver_data(device));
1228         if (!entry)
1229                 return -EIO;
1230 #endif
1231         return 0;
1232 }
1233
1234 int acpi_processor_power_exit(struct acpi_processor *pr,
1235                               struct acpi_device *device)
1236 {
1237         if (boot_option_idle_override)
1238                 return 0;
1239
1240         cpuidle_unregister_device(&pr->power.dev);
1241         pr->flags.power_setup_done = 0;
1242
1243 #ifdef CONFIG_ACPI_PROCFS
1244         if (acpi_device_dir(device))
1245                 remove_proc_entry(ACPI_PROCESSOR_FILE_POWER,
1246                                   acpi_device_dir(device));
1247 #endif
1248
1249         return 0;
1250 }