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 ACPI_PROCESSOR_CLASS            "processor"
  64 #define _COMPONENT              ACPI_PROCESSOR_COMPONENT
  65 ACPI_MODULE_NAME("processor_idle");
  66 #define ACPI_PROCESSOR_FILE_POWER       "power"
  67 #define PM_TIMER_TICK_NS                (1000000000ULL/PM_TIMER_FREQUENCY)
  68 #define C2_OVERHEAD                     1       /* 1us */
  69 #define C3_OVERHEAD                     1       /* 1us */
  70 #define PM_TIMER_TICKS_TO_US(p)         (((p) * 1000)/(PM_TIMER_FREQUENCY/1000))
  71
  72 static unsigned int max_cstate __read_mostly = ACPI_PROCESSOR_MAX_POWER;
  73 module_param(max_cstate, uint, 0000);
  74 static unsigned int nocst __read_mostly;
  75 module_param(nocst, uint, 0000);
  76
  77 static unsigned int latency_factor __read_mostly = 2;
  78 module_param(latency_factor, uint, 0644);
  79
  80 static s64 us_to_pm_timer_ticks(s64 t)
  81 {
  82         return div64_u64(t * PM_TIMER_FREQUENCY, 1000000);
  83 }
  84 /*
  85  * IBM ThinkPad R40e crashes mysteriously when going into C2 or C3.
  86  * For now disable this. Probably a bug somewhere else.
  87  *
  88  * To skip this limit, boot/load with a large max_cstate limit.
  89  */
  90 static int set_max_cstate(const struct dmi_system_id *id)
  91 {
  92         if (max_cstate > ACPI_PROCESSOR_MAX_POWER)
  93                 return 0;
  94
  95         printk(KERN_NOTICE PREFIX "%s detected - limiting to C%ld max_cstate."
  96                " Override with \"processor.max_cstate=%d\"\n", id->ident,
  97                (long)id->driver_data, ACPI_PROCESSOR_MAX_POWER + 1);
  98
  99         max_cstate = (long)id->driver_data;
 100
 101         return 0;
 102 }
 103
 104 /* Actually this shouldn't be __cpuinitdata, would be better to fix the
 105    callers to only run once -AK */
 106 static struct dmi_system_id __cpuinitdata processor_power_dmi_table[] = {
 107         { set_max_cstate, "Clevo 5600D", {
 108           DMI_MATCH(DMI_BIOS_VENDOR,"Phoenix Technologies LTD"),
 109           DMI_MATCH(DMI_BIOS_VERSION,"SHE845M0.86C.0013.D.0302131307")},
 110          (void *)2},
 111         {},
 112 };
 113
 114
 115 /*
 116  * Callers should disable interrupts before the call and enable
 117  * interrupts after return.
 118  */
 119 static void acpi_safe_halt(void)
 120 {
 121         current_thread_info()->status &= ~TS_POLLING;
 122         /*
 123          * TS_POLLING-cleared state must be visible before we
 124          * test NEED_RESCHED:
 125          */
 126         smp_mb();
 127         if (!need_resched()) {
 128                 safe_halt();
 129                 local_irq_disable();
 130         }
 131         current_thread_info()->status |= TS_POLLING;
 132 }
 133
 134 #ifdef ARCH_APICTIMER_STOPS_ON_C3
 135
 136 /*
 137  * Some BIOS implementations switch to C3 in the published C2 state.
 138  * This seems to be a common problem on AMD boxen, but other vendors
 139  * are affected too. We pick the most conservative approach: we assume
 140  * that the local APIC stops in both C2 and C3.
 141  */
 142 static void lapic_timer_check_state(int state, struct acpi_processor *pr,
 143                                    struct acpi_processor_cx *cx)
 144 {
 145         struct acpi_processor_power *pwr = &pr->power;
 146         u8 type = local_apic_timer_c2_ok ? ACPI_STATE_C3 : ACPI_STATE_C2;
 147
 148         if (cpu_has(&cpu_data(pr->id), X86_FEATURE_ARAT))
 149                 return;
 150
 151         if (boot_cpu_has(X86_FEATURE_AMDC1E))
 152                 type = ACPI_STATE_C1;
 153
 154         /*
 155          * Check, if one of the previous states already marked the lapic
 156          * unstable
 157          */
 158         if (pwr->timer_broadcast_on_state < state)
 159                 return;
 160
 161         if (cx->type >= type)
 162                 pr->power.timer_broadcast_on_state = state;
 163 }
 164
 165 static void lapic_timer_propagate_broadcast(struct acpi_processor *pr)
 166 {
 167         unsigned long reason;
 168
 169         reason = pr->power.timer_broadcast_on_state < INT_MAX ?
 170                 CLOCK_EVT_NOTIFY_BROADCAST_ON : CLOCK_EVT_NOTIFY_BROADCAST_OFF;
 171
 172         clockevents_notify(reason, &pr->id);
 173 }
 174
 175 /* Power(C) State timer broadcast control */
 176 static void lapic_timer_state_broadcast(struct acpi_processor *pr,
 177                                        struct acpi_processor_cx *cx,
 178                                        int broadcast)
 179 {
 180         int state = cx - pr->power.states;
 181
 182         if (state >= pr->power.timer_broadcast_on_state) {
 183                 unsigned long reason;
 184
 185                 reason = broadcast ?  CLOCK_EVT_NOTIFY_BROADCAST_ENTER :
 186                         CLOCK_EVT_NOTIFY_BROADCAST_EXIT;
 187                 clockevents_notify(reason, &pr->id);
 188         }
 189 }
 190
 191 #else
 192
 193 static void lapic_timer_check_state(int state, struct acpi_processor *pr,
 194                                    struct acpi_processor_cx *cstate) { }
 195 static void lapic_timer_propagate_broadcast(struct acpi_processor *pr) { }
 196 static void lapic_timer_state_broadcast(struct acpi_processor *pr,
 197                                        struct acpi_processor_cx *cx,
 198                                        int broadcast)
 199 {
 200 }
 201
 202 #endif
 203
 204 /*
 205  * Suspend / resume control
 206  */
 207 static int acpi_idle_suspend;
 208 static u32 saved_bm_rld;
 209
 210 static void acpi_idle_bm_rld_save(void)
 211 {
 212         acpi_read_bit_register(ACPI_BITREG_BUS_MASTER_RLD, &saved_bm_rld);
 213 }
 214 static void acpi_idle_bm_rld_restore(void)
 215 {
 216         u32 resumed_bm_rld;
 217
 218         acpi_read_bit_register(ACPI_BITREG_BUS_MASTER_RLD, &resumed_bm_rld);
 219
 220         if (resumed_bm_rld != saved_bm_rld)
 221                 acpi_write_bit_register(ACPI_BITREG_BUS_MASTER_RLD, saved_bm_rld);
 222 }
 223
 224 int acpi_processor_suspend(struct acpi_device * device, pm_message_t state)
 225 {
 226         if (acpi_idle_suspend == 1)
 227                 return 0;
 228
 229         acpi_idle_bm_rld_save();
 230         acpi_idle_suspend = 1;
 231         return 0;
 232 }
 233
 234 int acpi_processor_resume(struct acpi_device * device)
 235 {
 236         if (acpi_idle_suspend == 0)
 237                 return 0;
 238
 239         acpi_idle_bm_rld_restore();
 240         acpi_idle_suspend = 0;
 241         return 0;
 242 }
 243
 244 #if defined (CONFIG_GENERIC_TIME) && defined (CONFIG_X86)
 245 static void tsc_check_state(int state)
 246 {
 247         switch (boot_cpu_data.x86_vendor) {
 248         case X86_VENDOR_AMD:
 249         case X86_VENDOR_INTEL:
 250                 /*
 251                  * AMD Fam10h TSC will tick in all
 252                  * C/P/S0/S1 states when this bit is set.
 253                  */
 254                 if (boot_cpu_has(X86_FEATURE_NONSTOP_TSC))
 255                         return;
 256
 257                 /*FALL THROUGH*/
 258         default:
 259                 /* TSC could halt in idle, so notify users */
 260                 if (state > ACPI_STATE_C1)
 261                         mark_tsc_unstable("TSC halts in idle");
 262         }
 263 }
 264 #else
 265 static void tsc_check_state(int state) { return; }
 266 #endif
 267
 268 static int acpi_processor_get_power_info_fadt(struct acpi_processor *pr)
 269 {
 270
 271         if (!pr)
 272                 return -EINVAL;
 273
 274         if (!pr->pblk)
 275                 return -ENODEV;
 276
 277         /* if info is obtained from pblk/fadt, type equals state */
 278         pr->power.states[ACPI_STATE_C2].type = ACPI_STATE_C2;
 279         pr->power.states[ACPI_STATE_C3].type = ACPI_STATE_C3;
 280
 281 #ifndef CONFIG_HOTPLUG_CPU
 282         /*
 283          * Check for P_LVL2_UP flag before entering C2 and above on
 284          * an SMP system.
 285          */
 286         if ((num_online_cpus() > 1) &&
 287             !(acpi_gbl_FADT.flags & ACPI_FADT_C2_MP_SUPPORTED))
 288                 return -ENODEV;
 289 #endif
 290
 291         /* determine C2 and C3 address from pblk */
 292         pr->power.states[ACPI_STATE_C2].address = pr->pblk + 4;
 293         pr->power.states[ACPI_STATE_C3].address = pr->pblk + 5;
 294
 295         /* determine latencies from FADT */
 296         pr->power.states[ACPI_STATE_C2].latency = acpi_gbl_FADT.C2latency;
 297         pr->power.states[ACPI_STATE_C3].latency = acpi_gbl_FADT.C3latency;
 298
 299         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 300                           "lvl2[0x%08x] lvl3[0x%08x]\n",
 301                           pr->power.states[ACPI_STATE_C2].address,
 302                           pr->power.states[ACPI_STATE_C3].address));
 303
 304         return 0;
 305 }
 306
 307 static int acpi_processor_get_power_info_default(struct acpi_processor *pr)
 308 {
 309         if (!pr->power.states[ACPI_STATE_C1].valid) {
 310                 /* set the first C-State to C1 */
 311                 /* all processors need to support C1 */
 312                 pr->power.states[ACPI_STATE_C1].type = ACPI_STATE_C1;
 313                 pr->power.states[ACPI_STATE_C1].valid = 1;
 314                 pr->power.states[ACPI_STATE_C1].entry_method = ACPI_CSTATE_HALT;
 315         }
 316         /* the C0 state only exists as a filler in our array */
 317         pr->power.states[ACPI_STATE_C0].valid = 1;
 318         return 0;
 319 }
 320
 321 static int acpi_processor_get_power_info_cst(struct acpi_processor *pr)
 322 {
 323         acpi_status status = 0;
 324         acpi_integer count;
 325         int current_count;
 326         int i;
 327         struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
 328         union acpi_object *cst;
 329
 330
 331         if (nocst)
 332                 return -ENODEV;
 333
 334         current_count = 0;
 335
 336         status = acpi_evaluate_object(pr->handle, "_CST", NULL, &buffer);
 337         if (ACPI_FAILURE(status)) {
 338                 ACPI_DEBUG_PRINT((ACPI_DB_INFO, "No _CST, giving up\n"));
 339                 return -ENODEV;
 340         }
 341
 342         cst = buffer.pointer;
 343
 344         /* There must be at least 2 elements */
 345         if (!cst || (cst->type != ACPI_TYPE_PACKAGE) || cst->package.count < 2) {
 346                 printk(KERN_ERR PREFIX "not enough elements in _CST\n");
 347                 status = -EFAULT;
 348                 goto end;
 349         }
 350
 351         count = cst->package.elements[0].integer.value;
 352
 353         /* Validate number of power states. */
 354         if (count < 1 || count != cst->package.count - 1) {
 355                 printk(KERN_ERR PREFIX "count given by _CST is not valid\n");
 356                 status = -EFAULT;
 357                 goto end;
 358         }
 359
 360         /* Tell driver that at least _CST is supported. */
 361         pr->flags.has_cst = 1;
 362
 363         for (i = 1; i <= count; i++) {
 364                 union acpi_object *element;
 365                 union acpi_object *obj;
 366                 struct acpi_power_register *reg;
 367                 struct acpi_processor_cx cx;
 368
 369                 memset(&cx, 0, sizeof(cx));
 370
 371                 element = &(cst->package.elements[i]);
 372                 if (element->type != ACPI_TYPE_PACKAGE)
 373                         continue;
 374
 375                 if (element->package.count != 4)
 376                         continue;
 377
 378                 obj = &(element->package.elements[0]);
 379
 380                 if (obj->type != ACPI_TYPE_BUFFER)
 381                         continue;
 382
 383                 reg = (struct acpi_power_register *)obj->buffer.pointer;
 384
 385                 if (reg->space_id != ACPI_ADR_SPACE_SYSTEM_IO &&
 386                     (reg->space_id != ACPI_ADR_SPACE_FIXED_HARDWARE))
 387                         continue;
 388
 389                 /* There should be an easy way to extract an integer... */
 390                 obj = &(element->package.elements[1]);
 391                 if (obj->type != ACPI_TYPE_INTEGER)
 392                         continue;
 393
 394                 cx.type = obj->integer.value;
 395                 /*
 396                  * Some buggy BIOSes won't list C1 in _CST -
 397                  * Let acpi_processor_get_power_info_default() handle them later
 398                  */
 399                 if (i == 1 && cx.type != ACPI_STATE_C1)
 400                         current_count++;
 401
 402                 cx.address = reg->address;
 403                 cx.index = current_count + 1;
 404
 405                 cx.entry_method = ACPI_CSTATE_SYSTEMIO;
 406                 if (reg->space_id == ACPI_ADR_SPACE_FIXED_HARDWARE) {
 407                         if (acpi_processor_ffh_cstate_probe
 408                                         (pr->id, &cx, reg) == 0) {
 409                                 cx.entry_method = ACPI_CSTATE_FFH;
 410                         } else if (cx.type == ACPI_STATE_C1) {
 411                                 /*
 412                                  * C1 is a special case where FIXED_HARDWARE
 413                                  * can be handled in non-MWAIT way as well.
 414                                  * In that case, save this _CST entry info.
 415                                  * Otherwise, ignore this info and continue.
 416                                  */
 417                                 cx.entry_method = ACPI_CSTATE_HALT;
 418                                 snprintf(cx.desc, ACPI_CX_DESC_LEN, "ACPI HLT");
 419                         } else {
 420                                 continue;
 421                         }
 422                         if (cx.type == ACPI_STATE_C1 &&
 423                                         (idle_halt || idle_nomwait)) {
 424                                 /*
 425                                  * In most cases the C1 space_id obtained from
 426                                  * _CST object is FIXED_HARDWARE access mode.
 427                                  * But when the option of idle=halt is added,
 428                                  * the entry_method type should be changed from
 429                                  * CSTATE_FFH to CSTATE_HALT.
 430                                  * When the option of idle=nomwait is added,
 431                                  * the C1 entry_method type should be
 432                                  * CSTATE_HALT.
 433                                  */
 434                                 cx.entry_method = ACPI_CSTATE_HALT;
 435                                 snprintf(cx.desc, ACPI_CX_DESC_LEN, "ACPI HLT");
 436                         }
 437                 } else {
 438                         snprintf(cx.desc, ACPI_CX_DESC_LEN, "ACPI IOPORT 0x%x",
 439                                  cx.address);
 440                 }
 441
 442                 if (cx.type == ACPI_STATE_C1) {
 443                         cx.valid = 1;
 444                 }
 445
 446                 obj = &(element->package.elements[2]);
 447                 if (obj->type != ACPI_TYPE_INTEGER)
 448                         continue;
 449
 450                 cx.latency = obj->integer.value;
 451
 452                 obj = &(element->package.elements[3]);
 453                 if (obj->type != ACPI_TYPE_INTEGER)
 454                         continue;
 455
 456                 cx.power = obj->integer.value;
 457
 458                 current_count++;
 459                 memcpy(&(pr->power.states[current_count]), &cx, sizeof(cx));
 460
 461                 /*
 462                  * We support total ACPI_PROCESSOR_MAX_POWER - 1
 463                  * (From 1 through ACPI_PROCESSOR_MAX_POWER - 1)
 464                  */
 465                 if (current_count >= (ACPI_PROCESSOR_MAX_POWER - 1)) {
 466                         printk(KERN_WARNING
 467                                "Limiting number of power states to max (%d)\n",
 468                                ACPI_PROCESSOR_MAX_POWER);
 469                         printk(KERN_WARNING
 470                                "Please increase ACPI_PROCESSOR_MAX_POWER if needed.\n");
 471                         break;
 472                 }
 473         }
 474
 475         ACPI_DEBUG_PRINT((ACPI_DB_INFO, "Found %d power states\n",
 476                           current_count));
 477
 478         /* Validate number of power states discovered */
 479         if (current_count < 2)
 480                 status = -EFAULT;
 481
 482       end:
 483         kfree(buffer.pointer);
 484
 485         return status;
 486 }
 487
 488 static void acpi_processor_power_verify_c2(struct acpi_processor_cx *cx)
 489 {
 490
 491         if (!cx->address)
 492                 return;
 493
 494         /*
 495          * C2 latency must be less than or equal to 100
 496          * microseconds.
 497          */
 498         else if (cx->latency > ACPI_PROCESSOR_MAX_C2_LATENCY) {
 499                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 500                                   "latency too large [%d]\n", cx->latency));
 501                 return;
 502         }
 503
 504         /*
 505          * Otherwise we've met all of our C2 requirements.
 506          * Normalize the C2 latency to expidite policy
 507          */
 508         cx->valid = 1;
 509
 510         cx->latency_ticks = cx->latency;
 511
 512         return;
 513 }
 514
 515 static void acpi_processor_power_verify_c3(struct acpi_processor *pr,
 516                                            struct acpi_processor_cx *cx)
 517 {
 518         static int bm_check_flag = -1;
 519         static int bm_control_flag = -1;
 520
 521
 522         if (!cx->address)
 523                 return;
 524
 525         /*
 526          * C3 latency must be less than or equal to 1000
 527          * microseconds.
 528          */
 529         else if (cx->latency > ACPI_PROCESSOR_MAX_C3_LATENCY) {
 530                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 531                                   "latency too large [%d]\n", cx->latency));
 532                 return;
 533         }
 534
 535         /*
 536          * PIIX4 Erratum #18: We don't support C3 when Type-F (fast)
 537          * DMA transfers are used by any ISA device to avoid livelock.
 538          * Note that we could disable Type-F DMA (as recommended by
 539          * the erratum), but this is known to disrupt certain ISA
 540          * devices thus we take the conservative approach.
 541          */
 542         else if (errata.piix4.fdma) {
 543                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 544                                   "C3 not supported on PIIX4 with Type-F DMA\n"));
 545                 return;
 546         }
 547
 548         /* All the logic here assumes flags.bm_check is same across all CPUs */
 549         if (bm_check_flag == -1) {
 550                 /* Determine whether bm_check is needed based on CPU  */
 551                 acpi_processor_power_init_bm_check(&(pr->flags), pr->id);
 552                 bm_check_flag = pr->flags.bm_check;
 553                 bm_control_flag = pr->flags.bm_control;
 554         } else {
 555                 pr->flags.bm_check = bm_check_flag;
 556                 pr->flags.bm_control = bm_control_flag;
 557         }
 558
 559         if (pr->flags.bm_check) {
 560                 if (!pr->flags.bm_control) {
 561                         if (pr->flags.has_cst != 1) {
 562                                 /* bus mastering control is necessary */
 563                                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 564                                         "C3 support requires BM control\n"));
 565                                 return;
 566                         } else {
 567                                 /* Here we enter C3 without bus mastering */
 568                                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 569                                         "C3 support without BM control\n"));
 570                         }
 571                 }
 572         } else {
 573                 /*
 574                  * WBINVD should be set in fadt, for C3 state to be
 575                  * supported on when bm_check is not required.
 576                  */
 577                 if (!(acpi_gbl_FADT.flags & ACPI_FADT_WBINVD)) {
 578                         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 579                                           "Cache invalidation should work properly"
 580                                           " for C3 to be enabled on SMP systems\n"));
 581                         return;
 582                 }
 583         }
 584
 585         /*
 586          * Otherwise we've met all of our C3 requirements.
 587          * Normalize the C3 latency to expidite policy.  Enable
 588          * checking of bus mastering status (bm_check) so we can
 589          * use this in our C3 policy
 590          */
 591         cx->valid = 1;
 592
 593         cx->latency_ticks = cx->latency;
 594         /*
 595          * On older chipsets, BM_RLD needs to be set
 596          * in order for Bus Master activity to wake the
 597          * system from C3.  Newer chipsets handle DMA
 598          * during C3 automatically and BM_RLD is a NOP.
 599          * In either case, the proper way to
 600          * handle BM_RLD is to set it and leave it set.
 601          */
 602         acpi_write_bit_register(ACPI_BITREG_BUS_MASTER_RLD, 1);
 603
 604         return;
 605 }
 606
 607 static int acpi_processor_power_verify(struct acpi_processor *pr)
 608 {
 609         unsigned int i;
 610         unsigned int working = 0;
 611
 612         pr->power.timer_broadcast_on_state = INT_MAX;
 613
 614         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER && i <= max_cstate; i++) {
 615                 struct acpi_processor_cx *cx = &pr->power.states[i];
 616
 617                 switch (cx->type) {
 618                 case ACPI_STATE_C1:
 619                         cx->valid = 1;
 620                         break;
 621
 622                 case ACPI_STATE_C2:
 623                         acpi_processor_power_verify_c2(cx);
 624                         break;
 625
 626                 case ACPI_STATE_C3:
 627                         acpi_processor_power_verify_c3(pr, cx);
 628                         break;
 629                 }
 630                 if (!cx->valid)
 631                         continue;
 632
 633                 lapic_timer_check_state(i, pr, cx);
 634                 tsc_check_state(cx->type);
 635                 working++;
 636         }
 637
 638         lapic_timer_propagate_broadcast(pr);
 639
 640         return (working);
 641 }
 642
 643 static int acpi_processor_get_power_info(struct acpi_processor *pr)
 644 {
 645         unsigned int i;
 646         int result;
 647
 648
 649         /* NOTE: the idle thread may not be running while calling
 650          * this function */
 651
 652         /* Zero initialize all the C-states info. */
 653         memset(pr->power.states, 0, sizeof(pr->power.states));
 654
 655         result = acpi_processor_get_power_info_cst(pr);
 656         if (result == -ENODEV)
 657                 result = acpi_processor_get_power_info_fadt(pr);
 658
 659         if (result)
 660                 return result;
 661
 662         acpi_processor_get_power_info_default(pr);
 663
 664         pr->power.count = acpi_processor_power_verify(pr);
 665
 666         /*
 667          * if one state of type C2 or C3 is available, mark this
 668          * CPU as being "idle manageable"
 669          */
 670         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
 671                 if (pr->power.states[i].valid) {
 672                         pr->power.count = i;
 673                         if (pr->power.states[i].type >= ACPI_STATE_C2)
 674                                 pr->flags.power = 1;
 675                 }
 676         }
 677
 678         return 0;
 679 }
 680
 681 static int acpi_processor_power_seq_show(struct seq_file *seq, void *offset)
 682 {
 683         struct acpi_processor *pr = seq->private;
 684         unsigned int i;
 685
 686
 687         if (!pr)
 688                 goto end;
 689
 690         seq_printf(seq, "active state:            C%zd\n"
 691                    "max_cstate:              C%d\n"
 692                    "maximum allowed latency: %d usec\n",
 693                    pr->power.state ? pr->power.state - pr->power.states : 0,
 694                    max_cstate, pm_qos_requirement(PM_QOS_CPU_DMA_LATENCY));
 695
 696         seq_puts(seq, "states:\n");
 697
 698         for (i = 1; i <= pr->power.count; i++) {
 699                 seq_printf(seq, "   %cC%d:                  ",
 700                            (&pr->power.states[i] ==
 701                             pr->power.state ? '*' : ' '), i);
 702
 703                 if (!pr->power.states[i].valid) {
 704                         seq_puts(seq, "<not supported>\n");
 705                         continue;
 706                 }
 707
 708                 switch (pr->power.states[i].type) {
 709                 case ACPI_STATE_C1:
 710                         seq_printf(seq, "type[C1] ");
 711                         break;
 712                 case ACPI_STATE_C2:
 713                         seq_printf(seq, "type[C2] ");
 714                         break;
 715                 case ACPI_STATE_C3:
 716                         seq_printf(seq, "type[C3] ");
 717                         break;
 718                 default:
 719                         seq_printf(seq, "type[--] ");
 720                         break;
 721                 }
 722
 723                 if (pr->power.states[i].promotion.state)
 724                         seq_printf(seq, "promotion[C%zd] ",
 725                                    (pr->power.states[i].promotion.state -
 726                                     pr->power.states));
 727                 else
 728                         seq_puts(seq, "promotion[--] ");
 729
 730                 if (pr->power.states[i].demotion.state)
 731                         seq_printf(seq, "demotion[C%zd] ",
 732                                    (pr->power.states[i].demotion.state -
 733                                     pr->power.states));
 734                 else
 735                         seq_puts(seq, "demotion[--] ");
 736
 737                 seq_printf(seq, "latency[%03d] usage[%08d] duration[%020llu]\n",
 738                            pr->power.states[i].latency,
 739                            pr->power.states[i].usage,
 740                            (unsigned long long)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
 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;
 866         s64 sleep_ticks = 0;
 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         current_thread_info()->status &= ~TS_POLLING;
 878         /*
 879          * TS_POLLING-cleared state must be visible before we test
 880          * NEED_RESCHED:
 881          */
 882         smp_mb();
 883
 884         if (unlikely(need_resched())) {
 885                 current_thread_info()->status |= TS_POLLING;
 886                 local_irq_enable();
 887                 return 0;
 888         }
 889
 890         /*
 891          * Must be done before busmaster disable as we might need to
 892          * access HPET !
 893          */
 894         lapic_timer_state_broadcast(pr, cx, 1);
 895
 896         if (cx->type == ACPI_STATE_C3)
 897                 ACPI_FLUSH_CPU_CACHE();
 898
 899         kt1 = ktime_get_real();
 900         /* Tell the scheduler that we are going deep-idle: */
 901         sched_clock_idle_sleep_event();
 902         acpi_idle_do_entry(cx);
 903         kt2 = ktime_get_real();
 904         idle_time =  ktime_to_us(ktime_sub(kt2, kt1));
 905
 906         sleep_ticks = us_to_pm_timer_ticks(idle_time);
 907
 908         /* Tell the scheduler how much we idled: */
 909         sched_clock_idle_wakeup_event(sleep_ticks*PM_TIMER_TICK_NS);
 910
 911         local_irq_enable();
 912         current_thread_info()->status |= TS_POLLING;
 913
 914         cx->usage++;
 915
 916         lapic_timer_state_broadcast(pr, cx, 0);
 917         cx->time += sleep_ticks;
 918         return idle_time;
 919 }
 920
 921 static int c3_cpu_count;
 922 static DEFINE_SPINLOCK(c3_lock);
 923
 924 /**
 925  * acpi_idle_enter_bm - enters C3 with proper BM handling
 926  * @dev: the target CPU
 927  * @state: the state data
 928  *
 929  * If BM is detected, the deepest non-C3 idle state is entered instead.
 930  */
 931 static int acpi_idle_enter_bm(struct cpuidle_device *dev,
 932                               struct cpuidle_state *state)
 933 {
 934         struct acpi_processor *pr;
 935         struct acpi_processor_cx *cx = cpuidle_get_statedata(state);
 936         ktime_t  kt1, kt2;
 937         s64 idle_time;
 938         s64 sleep_ticks = 0;
 939
 940
 941         pr = __get_cpu_var(processors);
 942
 943         if (unlikely(!pr))
 944                 return 0;
 945
 946         if (acpi_idle_suspend)
 947                 return(acpi_idle_enter_c1(dev, state));
 948
 949         if (acpi_idle_bm_check()) {
 950                 if (dev->safe_state) {
 951                         dev->last_state = dev->safe_state;
 952                         return dev->safe_state->enter(dev, dev->safe_state);
 953                 } else {
 954                         local_irq_disable();
 955                         acpi_safe_halt();
 956                         local_irq_enable();
 957                         return 0;
 958                 }
 959         }
 960
 961         local_irq_disable();
 962         current_thread_info()->status &= ~TS_POLLING;
 963         /*
 964          * TS_POLLING-cleared state must be visible before we test
 965          * NEED_RESCHED:
 966          */
 967         smp_mb();
 968
 969         if (unlikely(need_resched())) {
 970                 current_thread_info()->status |= TS_POLLING;
 971                 local_irq_enable();
 972                 return 0;
 973         }
 974
 975         acpi_unlazy_tlb(smp_processor_id());
 976
 977         /* Tell the scheduler that we are going deep-idle: */
 978         sched_clock_idle_sleep_event();
 979         /*
 980          * Must be done before busmaster disable as we might need to
 981          * access HPET !
 982          */
 983         lapic_timer_state_broadcast(pr, cx, 1);
 984
 985         kt1 = ktime_get_real();
 986         /*
 987          * disable bus master
 988          * bm_check implies we need ARB_DIS
 989          * !bm_check implies we need cache flush
 990          * bm_control implies whether we can do ARB_DIS
 991          *
 992          * That leaves a case where bm_check is set and bm_control is
 993          * not set. In that case we cannot do much, we enter C3
 994          * without doing anything.
 995          */
 996         if (pr->flags.bm_check && pr->flags.bm_control) {
 997                 spin_lock(&c3_lock);
 998                 c3_cpu_count++;
 999                 /* Disable bus master arbitration when all CPUs are in C3 */
1000                 if (c3_cpu_count == num_online_cpus())
1001                         acpi_write_bit_register(ACPI_BITREG_ARB_DISABLE, 1);
1002                 spin_unlock(&c3_lock);
1003         } else if (!pr->flags.bm_check) {
1004                 ACPI_FLUSH_CPU_CACHE();
1005         }
1006
1007         acpi_idle_do_entry(cx);
1008
1009         /* Re-enable bus master arbitration */
1010         if (pr->flags.bm_check && pr->flags.bm_control) {
1011                 spin_lock(&c3_lock);
1012                 acpi_write_bit_register(ACPI_BITREG_ARB_DISABLE, 0);
1013                 c3_cpu_count--;
1014                 spin_unlock(&c3_lock);
1015         }
1016         kt2 = ktime_get_real();
1017         idle_time =  ktime_to_us(ktime_sub(kt2, kt1));
1018
1019         sleep_ticks = us_to_pm_timer_ticks(idle_time);
1020         /* Tell the scheduler how much we idled: */
1021         sched_clock_idle_wakeup_event(sleep_ticks*PM_TIMER_TICK_NS);
1022
1023         local_irq_enable();
1024         current_thread_info()->status |= TS_POLLING;
1025
1026         cx->usage++;
1027
1028         lapic_timer_state_broadcast(pr, cx, 0);
1029         cx->time += sleep_ticks;
1030         return idle_time;
1031 }
1032
1033 struct cpuidle_driver acpi_idle_driver = {
1034         .name =         "acpi_idle",
1035         .owner =        THIS_MODULE,
1036 };
1037
1038 /**
1039  * acpi_processor_setup_cpuidle - prepares and configures CPUIDLE
1040  * @pr: the ACPI processor
1041  */
1042 static int acpi_processor_setup_cpuidle(struct acpi_processor *pr)
1043 {
1044         int i, count = CPUIDLE_DRIVER_STATE_START;
1045         struct acpi_processor_cx *cx;
1046         struct cpuidle_state *state;
1047         struct cpuidle_device *dev = &pr->power.dev;
1048
1049         if (!pr->flags.power_setup_done)
1050                 return -EINVAL;
1051
1052         if (pr->flags.power == 0) {
1053                 return -EINVAL;
1054         }
1055
1056         dev->cpu = pr->id;
1057         for (i = 0; i < CPUIDLE_STATE_MAX; i++) {
1058                 dev->states[i].name[0] = '\0';
1059                 dev->states[i].desc[0] = '\0';
1060         }
1061
1062         if (max_cstate == 0)
1063                 max_cstate = 1;
1064
1065         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER && i <= max_cstate; i++) {
1066                 cx = &pr->power.states[i];
1067                 state = &dev->states[count];
1068
1069                 if (!cx->valid)
1070                         continue;
1071
1072 #ifdef CONFIG_HOTPLUG_CPU
1073                 if ((cx->type != ACPI_STATE_C1) && (num_online_cpus() > 1) &&
1074                     !pr->flags.has_cst &&
1075                     !(acpi_gbl_FADT.flags & ACPI_FADT_C2_MP_SUPPORTED))
1076                         continue;
1077 #endif
1078                 cpuidle_set_statedata(state, cx);
1079
1080                 snprintf(state->name, CPUIDLE_NAME_LEN, "C%d", i);
1081                 strncpy(state->desc, cx->desc, CPUIDLE_DESC_LEN);
1082                 state->exit_latency = cx->latency;
1083                 state->target_residency = cx->latency * latency_factor;
1084                 state->power_usage = cx->power;
1085
1086                 state->flags = 0;
1087                 switch (cx->type) {
1088                         case ACPI_STATE_C1:
1089                         state->flags |= CPUIDLE_FLAG_SHALLOW;
1090                         if (cx->entry_method == ACPI_CSTATE_FFH)
1091                                 state->flags |= CPUIDLE_FLAG_TIME_VALID;
1092
1093                         state->enter = acpi_idle_enter_c1;
1094                         dev->safe_state = state;
1095                         break;
1096
1097                         case ACPI_STATE_C2:
1098                         state->flags |= CPUIDLE_FLAG_BALANCED;
1099                         state->flags |= CPUIDLE_FLAG_TIME_VALID;
1100                         state->enter = acpi_idle_enter_simple;
1101                         dev->safe_state = state;
1102                         break;
1103
1104                         case ACPI_STATE_C3:
1105                         state->flags |= CPUIDLE_FLAG_DEEP;
1106                         state->flags |= CPUIDLE_FLAG_TIME_VALID;
1107                         state->flags |= CPUIDLE_FLAG_CHECK_BM;
1108                         state->enter = pr->flags.bm_check ?
1109                                         acpi_idle_enter_bm :
1110                                         acpi_idle_enter_simple;
1111                         break;
1112                 }
1113
1114                 count++;
1115                 if (count == CPUIDLE_STATE_MAX)
1116                         break;
1117         }
1118
1119         dev->state_count = count;
1120
1121         if (!count)
1122                 return -EINVAL;
1123
1124         return 0;
1125 }
1126
1127 int acpi_processor_cst_has_changed(struct acpi_processor *pr)
1128 {
1129         int ret = 0;
1130
1131         if (boot_option_idle_override)
1132                 return 0;
1133
1134         if (!pr)
1135                 return -EINVAL;
1136
1137         if (nocst) {
1138                 return -ENODEV;
1139         }
1140
1141         if (!pr->flags.power_setup_done)
1142                 return -ENODEV;
1143
1144         cpuidle_pause_and_lock();
1145         cpuidle_disable_device(&pr->power.dev);
1146         acpi_processor_get_power_info(pr);
1147         if (pr->flags.power) {
1148                 acpi_processor_setup_cpuidle(pr);
1149                 ret = cpuidle_enable_device(&pr->power.dev);
1150         }
1151         cpuidle_resume_and_unlock();
1152
1153         return ret;
1154 }
1155
1156 int __cpuinit acpi_processor_power_init(struct acpi_processor *pr,
1157                               struct acpi_device *device)
1158 {
1159         acpi_status status = 0;
1160         static int first_run;
1161         struct proc_dir_entry *entry = NULL;
1162         unsigned int i;
1163
1164         if (boot_option_idle_override)
1165                 return 0;
1166
1167         if (!first_run) {
1168                 if (idle_halt) {
1169                         /*
1170                          * When the boot option of "idle=halt" is added, halt
1171                          * is used for CPU IDLE.
1172                          * In such case C2/C3 is meaningless. So the max_cstate
1173                          * is set to one.
1174                          */
1175                         max_cstate = 1;
1176                 }
1177                 dmi_check_system(processor_power_dmi_table);
1178                 max_cstate = acpi_processor_cstate_check(max_cstate);
1179                 if (max_cstate < ACPI_C_STATES_MAX)
1180                         printk(KERN_NOTICE
1181                                "ACPI: processor limited to max C-state %d\n",
1182                                max_cstate);
1183                 first_run++;
1184         }
1185
1186         if (!pr)
1187                 return -EINVAL;
1188
1189         if (acpi_gbl_FADT.cst_control && !nocst) {
1190                 status =
1191                     acpi_os_write_port(acpi_gbl_FADT.smi_command, acpi_gbl_FADT.cst_control, 8);
1192                 if (ACPI_FAILURE(status)) {
1193                         ACPI_EXCEPTION((AE_INFO, status,
1194                                         "Notifying BIOS of _CST ability failed"));
1195                 }
1196         }
1197
1198         acpi_processor_get_power_info(pr);
1199         pr->flags.power_setup_done = 1;
1200
1201         /*
1202          * Install the idle handler if processor power management is supported.
1203          * Note that we use previously set idle handler will be used on
1204          * platforms that only support C1.
1205          */
1206         if (pr->flags.power) {
1207                 acpi_processor_setup_cpuidle(pr);
1208                 if (cpuidle_register_device(&pr->power.dev))
1209                         return -EIO;
1210
1211                 printk(KERN_INFO PREFIX "CPU%d (power states:", pr->id);
1212                 for (i = 1; i <= pr->power.count; i++)
1213                         if (pr->power.states[i].valid)
1214                                 printk(" C%d[C%d]", i,
1215                                        pr->power.states[i].type);
1216                 printk(")\n");
1217         }
1218
1219         /* 'power' [R] */
1220         entry = proc_create_data(ACPI_PROCESSOR_FILE_POWER,
1221                                  S_IRUGO, acpi_device_dir(device),
1222                                  &acpi_processor_power_fops,
1223                                  acpi_driver_data(device));
1224         if (!entry)
1225                 return -EIO;
1226         return 0;
1227 }
1228
1229 int acpi_processor_power_exit(struct acpi_processor *pr,
1230                               struct acpi_device *device)
1231 {
1232         if (boot_option_idle_override)
1233                 return 0;
1234
1235         cpuidle_unregister_device(&pr->power.dev);
1236         pr->flags.power_setup_done = 0;
1237
1238         if (acpi_device_dir(device))
1239                 remove_proc_entry(ACPI_PROCESSOR_FILE_POWER,
1240                                   acpi_device_dir(device));
1241
1242         return 0;
1243 }