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 acpi_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         /*
 152          * Check, if one of the previous states already marked the lapic
 153          * unstable
 154          */
 155         if (pwr->timer_broadcast_on_state < state)
 156                 return;
 157
 158         if (cx->type >= type)
 159                 pr->power.timer_broadcast_on_state = state;
 160 }
 161
 162 static void acpi_propagate_timer_broadcast(struct acpi_processor *pr)
 163 {
 164         unsigned long reason;
 165
 166         reason = pr->power.timer_broadcast_on_state < INT_MAX ?
 167                 CLOCK_EVT_NOTIFY_BROADCAST_ON : CLOCK_EVT_NOTIFY_BROADCAST_OFF;
 168
 169         clockevents_notify(reason, &pr->id);
 170 }
 171
 172 /* Power(C) State timer broadcast control */
 173 static void acpi_state_timer_broadcast(struct acpi_processor *pr,
 174                                        struct acpi_processor_cx *cx,
 175                                        int broadcast)
 176 {
 177         int state = cx - pr->power.states;
 178
 179         if (state >= pr->power.timer_broadcast_on_state) {
 180                 unsigned long reason;
 181
 182                 reason = broadcast ?  CLOCK_EVT_NOTIFY_BROADCAST_ENTER :
 183                         CLOCK_EVT_NOTIFY_BROADCAST_EXIT;
 184                 clockevents_notify(reason, &pr->id);
 185         }
 186 }
 187
 188 #else
 189
 190 static void acpi_timer_check_state(int state, struct acpi_processor *pr,
 191                                    struct acpi_processor_cx *cstate) { }
 192 static void acpi_propagate_timer_broadcast(struct acpi_processor *pr) { }
 193 static void acpi_state_timer_broadcast(struct acpi_processor *pr,
 194                                        struct acpi_processor_cx *cx,
 195                                        int broadcast)
 196 {
 197 }
 198
 199 #endif
 200
 201 /*
 202  * Suspend / resume control
 203  */
 204 static int acpi_idle_suspend;
 205 static u32 saved_bm_rld;
 206
 207 static void acpi_idle_bm_rld_save(void)
 208 {
 209         acpi_read_bit_register(ACPI_BITREG_BUS_MASTER_RLD, &saved_bm_rld);
 210 }
 211 static void acpi_idle_bm_rld_restore(void)
 212 {
 213         u32 resumed_bm_rld;
 214
 215         acpi_read_bit_register(ACPI_BITREG_BUS_MASTER_RLD, &resumed_bm_rld);
 216
 217         if (resumed_bm_rld != saved_bm_rld)
 218                 acpi_write_bit_register(ACPI_BITREG_BUS_MASTER_RLD, saved_bm_rld);
 219 }
 220
 221 int acpi_processor_suspend(struct acpi_device * device, pm_message_t state)
 222 {
 223         if (acpi_idle_suspend == 1)
 224                 return 0;
 225
 226         acpi_idle_bm_rld_save();
 227         acpi_idle_suspend = 1;
 228         return 0;
 229 }
 230
 231 int acpi_processor_resume(struct acpi_device * device)
 232 {
 233         if (acpi_idle_suspend == 0)
 234                 return 0;
 235
 236         acpi_idle_bm_rld_restore();
 237         acpi_idle_suspend = 0;
 238         return 0;
 239 }
 240
 241 #if defined (CONFIG_GENERIC_TIME) && defined (CONFIG_X86)
 242 static void tsc_check_state(int state)
 243 {
 244         switch (boot_cpu_data.x86_vendor) {
 245         case X86_VENDOR_AMD:
 246         case X86_VENDOR_INTEL:
 247                 /*
 248                  * AMD Fam10h TSC will tick in all
 249                  * C/P/S0/S1 states when this bit is set.
 250                  */
 251                 if (boot_cpu_has(X86_FEATURE_NONSTOP_TSC))
 252                         return;
 253
 254                 /*FALL THROUGH*/
 255         default:
 256                 /* TSC could halt in idle, so notify users */
 257                 if (state > ACPI_STATE_C1)
 258                         mark_tsc_unstable("TSC halts in idle");
 259         }
 260 }
 261 #else
 262 static void tsc_check_state(int state) { return; }
 263 #endif
 264
 265 static int acpi_processor_get_power_info_fadt(struct acpi_processor *pr)
 266 {
 267
 268         if (!pr)
 269                 return -EINVAL;
 270
 271         if (!pr->pblk)
 272                 return -ENODEV;
 273
 274         /* if info is obtained from pblk/fadt, type equals state */
 275         pr->power.states[ACPI_STATE_C2].type = ACPI_STATE_C2;
 276         pr->power.states[ACPI_STATE_C3].type = ACPI_STATE_C3;
 277
 278 #ifndef CONFIG_HOTPLUG_CPU
 279         /*
 280          * Check for P_LVL2_UP flag before entering C2 and above on
 281          * an SMP system.
 282          */
 283         if ((num_online_cpus() > 1) &&
 284             !(acpi_gbl_FADT.flags & ACPI_FADT_C2_MP_SUPPORTED))
 285                 return -ENODEV;
 286 #endif
 287
 288         /* determine C2 and C3 address from pblk */
 289         pr->power.states[ACPI_STATE_C2].address = pr->pblk + 4;
 290         pr->power.states[ACPI_STATE_C3].address = pr->pblk + 5;
 291
 292         /* determine latencies from FADT */
 293         pr->power.states[ACPI_STATE_C2].latency = acpi_gbl_FADT.C2latency;
 294         pr->power.states[ACPI_STATE_C3].latency = acpi_gbl_FADT.C3latency;
 295
 296         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 297                           "lvl2[0x%08x] lvl3[0x%08x]\n",
 298                           pr->power.states[ACPI_STATE_C2].address,
 299                           pr->power.states[ACPI_STATE_C3].address));
 300
 301         return 0;
 302 }
 303
 304 static int acpi_processor_get_power_info_default(struct acpi_processor *pr)
 305 {
 306         if (!pr->power.states[ACPI_STATE_C1].valid) {
 307                 /* set the first C-State to C1 */
 308                 /* all processors need to support C1 */
 309                 pr->power.states[ACPI_STATE_C1].type = ACPI_STATE_C1;
 310                 pr->power.states[ACPI_STATE_C1].valid = 1;
 311                 pr->power.states[ACPI_STATE_C1].entry_method = ACPI_CSTATE_HALT;
 312         }
 313         /* the C0 state only exists as a filler in our array */
 314         pr->power.states[ACPI_STATE_C0].valid = 1;
 315         return 0;
 316 }
 317
 318 static int acpi_processor_get_power_info_cst(struct acpi_processor *pr)
 319 {
 320         acpi_status status = 0;
 321         acpi_integer count;
 322         int current_count;
 323         int i;
 324         struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
 325         union acpi_object *cst;
 326
 327
 328         if (nocst)
 329                 return -ENODEV;
 330
 331         current_count = 0;
 332
 333         status = acpi_evaluate_object(pr->handle, "_CST", NULL, &buffer);
 334         if (ACPI_FAILURE(status)) {
 335                 ACPI_DEBUG_PRINT((ACPI_DB_INFO, "No _CST, giving up\n"));
 336                 return -ENODEV;
 337         }
 338
 339         cst = buffer.pointer;
 340
 341         /* There must be at least 2 elements */
 342         if (!cst || (cst->type != ACPI_TYPE_PACKAGE) || cst->package.count < 2) {
 343                 printk(KERN_ERR PREFIX "not enough elements in _CST\n");
 344                 status = -EFAULT;
 345                 goto end;
 346         }
 347
 348         count = cst->package.elements[0].integer.value;
 349
 350         /* Validate number of power states. */
 351         if (count < 1 || count != cst->package.count - 1) {
 352                 printk(KERN_ERR PREFIX "count given by _CST is not valid\n");
 353                 status = -EFAULT;
 354                 goto end;
 355         }
 356
 357         /* Tell driver that at least _CST is supported. */
 358         pr->flags.has_cst = 1;
 359
 360         for (i = 1; i <= count; i++) {
 361                 union acpi_object *element;
 362                 union acpi_object *obj;
 363                 struct acpi_power_register *reg;
 364                 struct acpi_processor_cx cx;
 365
 366                 memset(&cx, 0, sizeof(cx));
 367
 368                 element = &(cst->package.elements[i]);
 369                 if (element->type != ACPI_TYPE_PACKAGE)
 370                         continue;
 371
 372                 if (element->package.count != 4)
 373                         continue;
 374
 375                 obj = &(element->package.elements[0]);
 376
 377                 if (obj->type != ACPI_TYPE_BUFFER)
 378                         continue;
 379
 380                 reg = (struct acpi_power_register *)obj->buffer.pointer;
 381
 382                 if (reg->space_id != ACPI_ADR_SPACE_SYSTEM_IO &&
 383                     (reg->space_id != ACPI_ADR_SPACE_FIXED_HARDWARE))
 384                         continue;
 385
 386                 /* There should be an easy way to extract an integer... */
 387                 obj = &(element->package.elements[1]);
 388                 if (obj->type != ACPI_TYPE_INTEGER)
 389                         continue;
 390
 391                 cx.type = obj->integer.value;
 392                 /*
 393                  * Some buggy BIOSes won't list C1 in _CST -
 394                  * Let acpi_processor_get_power_info_default() handle them later
 395                  */
 396                 if (i == 1 && cx.type != ACPI_STATE_C1)
 397                         current_count++;
 398
 399                 cx.address = reg->address;
 400                 cx.index = current_count + 1;
 401
 402                 cx.entry_method = ACPI_CSTATE_SYSTEMIO;
 403                 if (reg->space_id == ACPI_ADR_SPACE_FIXED_HARDWARE) {
 404                         if (acpi_processor_ffh_cstate_probe
 405                                         (pr->id, &cx, reg) == 0) {
 406                                 cx.entry_method = ACPI_CSTATE_FFH;
 407                         } else if (cx.type == ACPI_STATE_C1) {
 408                                 /*
 409                                  * C1 is a special case where FIXED_HARDWARE
 410                                  * can be handled in non-MWAIT way as well.
 411                                  * In that case, save this _CST entry info.
 412                                  * Otherwise, ignore this info and continue.
 413                                  */
 414                                 cx.entry_method = ACPI_CSTATE_HALT;
 415                                 snprintf(cx.desc, ACPI_CX_DESC_LEN, "ACPI HLT");
 416                         } else {
 417                                 continue;
 418                         }
 419                         if (cx.type == ACPI_STATE_C1 &&
 420                                         (idle_halt || idle_nomwait)) {
 421                                 /*
 422                                  * In most cases the C1 space_id obtained from
 423                                  * _CST object is FIXED_HARDWARE access mode.
 424                                  * But when the option of idle=halt is added,
 425                                  * the entry_method type should be changed from
 426                                  * CSTATE_FFH to CSTATE_HALT.
 427                                  * When the option of idle=nomwait is added,
 428                                  * the C1 entry_method type should be
 429                                  * CSTATE_HALT.
 430                                  */
 431                                 cx.entry_method = ACPI_CSTATE_HALT;
 432                                 snprintf(cx.desc, ACPI_CX_DESC_LEN, "ACPI HLT");
 433                         }
 434                 } else {
 435                         snprintf(cx.desc, ACPI_CX_DESC_LEN, "ACPI IOPORT 0x%x",
 436                                  cx.address);
 437                 }
 438
 439                 if (cx.type == ACPI_STATE_C1) {
 440                         cx.valid = 1;
 441                 }
 442
 443                 obj = &(element->package.elements[2]);
 444                 if (obj->type != ACPI_TYPE_INTEGER)
 445                         continue;
 446
 447                 cx.latency = obj->integer.value;
 448
 449                 obj = &(element->package.elements[3]);
 450                 if (obj->type != ACPI_TYPE_INTEGER)
 451                         continue;
 452
 453                 cx.power = obj->integer.value;
 454
 455                 current_count++;
 456                 memcpy(&(pr->power.states[current_count]), &cx, sizeof(cx));
 457
 458                 /*
 459                  * We support total ACPI_PROCESSOR_MAX_POWER - 1
 460                  * (From 1 through ACPI_PROCESSOR_MAX_POWER - 1)
 461                  */
 462                 if (current_count >= (ACPI_PROCESSOR_MAX_POWER - 1)) {
 463                         printk(KERN_WARNING
 464                                "Limiting number of power states to max (%d)\n",
 465                                ACPI_PROCESSOR_MAX_POWER);
 466                         printk(KERN_WARNING
 467                                "Please increase ACPI_PROCESSOR_MAX_POWER if needed.\n");
 468                         break;
 469                 }
 470         }
 471
 472         ACPI_DEBUG_PRINT((ACPI_DB_INFO, "Found %d power states\n",
 473                           current_count));
 474
 475         /* Validate number of power states discovered */
 476         if (current_count < 2)
 477                 status = -EFAULT;
 478
 479       end:
 480         kfree(buffer.pointer);
 481
 482         return status;
 483 }
 484
 485 static void acpi_processor_power_verify_c2(struct acpi_processor_cx *cx)
 486 {
 487
 488         if (!cx->address)
 489                 return;
 490
 491         /*
 492          * C2 latency must be less than or equal to 100
 493          * microseconds.
 494          */
 495         else if (cx->latency > ACPI_PROCESSOR_MAX_C2_LATENCY) {
 496                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 497                                   "latency too large [%d]\n", cx->latency));
 498                 return;
 499         }
 500
 501         /*
 502          * Otherwise we've met all of our C2 requirements.
 503          * Normalize the C2 latency to expidite policy
 504          */
 505         cx->valid = 1;
 506
 507         cx->latency_ticks = cx->latency;
 508
 509         return;
 510 }
 511
 512 static void acpi_processor_power_verify_c3(struct acpi_processor *pr,
 513                                            struct acpi_processor_cx *cx)
 514 {
 515         static int bm_check_flag = -1;
 516         static int bm_control_flag = -1;
 517
 518
 519         if (!cx->address)
 520                 return;
 521
 522         /*
 523          * C3 latency must be less than or equal to 1000
 524          * microseconds.
 525          */
 526         else if (cx->latency > ACPI_PROCESSOR_MAX_C3_LATENCY) {
 527                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 528                                   "latency too large [%d]\n", cx->latency));
 529                 return;
 530         }
 531
 532         /*
 533          * PIIX4 Erratum #18: We don't support C3 when Type-F (fast)
 534          * DMA transfers are used by any ISA device to avoid livelock.
 535          * Note that we could disable Type-F DMA (as recommended by
 536          * the erratum), but this is known to disrupt certain ISA
 537          * devices thus we take the conservative approach.
 538          */
 539         else if (errata.piix4.fdma) {
 540                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 541                                   "C3 not supported on PIIX4 with Type-F DMA\n"));
 542                 return;
 543         }
 544
 545         /* All the logic here assumes flags.bm_check is same across all CPUs */
 546         if (bm_check_flag == -1) {
 547                 /* Determine whether bm_check is needed based on CPU  */
 548                 acpi_processor_power_init_bm_check(&(pr->flags), pr->id);
 549                 bm_check_flag = pr->flags.bm_check;
 550                 bm_control_flag = pr->flags.bm_control;
 551         } else {
 552                 pr->flags.bm_check = bm_check_flag;
 553                 pr->flags.bm_control = bm_control_flag;
 554         }
 555
 556         if (pr->flags.bm_check) {
 557                 if (!pr->flags.bm_control) {
 558                         if (pr->flags.has_cst != 1) {
 559                                 /* bus mastering control is necessary */
 560                                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 561                                         "C3 support requires BM control\n"));
 562                                 return;
 563                         } else {
 564                                 /* Here we enter C3 without bus mastering */
 565                                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 566                                         "C3 support without BM control\n"));
 567                         }
 568                 }
 569         } else {
 570                 /*
 571                  * WBINVD should be set in fadt, for C3 state to be
 572                  * supported on when bm_check is not required.
 573                  */
 574                 if (!(acpi_gbl_FADT.flags & ACPI_FADT_WBINVD)) {
 575                         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 576                                           "Cache invalidation should work properly"
 577                                           " for C3 to be enabled on SMP systems\n"));
 578                         return;
 579                 }
 580         }
 581
 582         /*
 583          * Otherwise we've met all of our C3 requirements.
 584          * Normalize the C3 latency to expidite policy.  Enable
 585          * checking of bus mastering status (bm_check) so we can
 586          * use this in our C3 policy
 587          */
 588         cx->valid = 1;
 589
 590         cx->latency_ticks = cx->latency;
 591         /*
 592          * On older chipsets, BM_RLD needs to be set
 593          * in order for Bus Master activity to wake the
 594          * system from C3.  Newer chipsets handle DMA
 595          * during C3 automatically and BM_RLD is a NOP.
 596          * In either case, the proper way to
 597          * handle BM_RLD is to set it and leave it set.
 598          */
 599         acpi_write_bit_register(ACPI_BITREG_BUS_MASTER_RLD, 1);
 600
 601         return;
 602 }
 603
 604 static int acpi_processor_power_verify(struct acpi_processor *pr)
 605 {
 606         unsigned int i;
 607         unsigned int working = 0;
 608
 609         pr->power.timer_broadcast_on_state = INT_MAX;
 610
 611         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER && i <= max_cstate; i++) {
 612                 struct acpi_processor_cx *cx = &pr->power.states[i];
 613
 614                 switch (cx->type) {
 615                 case ACPI_STATE_C1:
 616                         cx->valid = 1;
 617                         break;
 618
 619                 case ACPI_STATE_C2:
 620                         acpi_processor_power_verify_c2(cx);
 621                         if (cx->valid)
 622                                 acpi_timer_check_state(i, pr, cx);
 623                         break;
 624
 625                 case ACPI_STATE_C3:
 626                         acpi_processor_power_verify_c3(pr, cx);
 627                         if (cx->valid)
 628                                 acpi_timer_check_state(i, pr, cx);
 629                         break;
 630                 }
 631                 if (cx->valid)
 632                         tsc_check_state(cx->type);
 633
 634                 if (cx->valid)
 635                         working++;
 636         }
 637
 638         acpi_propagate_timer_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                 acpi_safe_halt();
 837                 local_irq_enable();
 838                 return 0;
 839         }
 840
 841         kt1 = ktime_get_real();
 842         acpi_idle_do_entry(cx);
 843         kt2 = ktime_get_real();
 844         idle_time =  ktime_to_us(ktime_sub(kt2, kt1));
 845
 846         local_irq_enable();
 847         cx->usage++;
 848
 849         return idle_time;
 850 }
 851
 852 /**
 853  * acpi_idle_enter_simple - enters an ACPI state without BM handling
 854  * @dev: the target CPU
 855  * @state: the state data
 856  */
 857 static int acpi_idle_enter_simple(struct cpuidle_device *dev,
 858                                   struct cpuidle_state *state)
 859 {
 860         struct acpi_processor *pr;
 861         struct acpi_processor_cx *cx = cpuidle_get_statedata(state);
 862         ktime_t  kt1, kt2;
 863         s64 idle_time;
 864         s64 sleep_ticks = 0;
 865
 866         pr = __get_cpu_var(processors);
 867
 868         if (unlikely(!pr))
 869                 return 0;
 870
 871         if (acpi_idle_suspend)
 872                 return(acpi_idle_enter_c1(dev, state));
 873
 874         local_irq_disable();
 875         current_thread_info()->status &= ~TS_POLLING;
 876         /*
 877          * TS_POLLING-cleared state must be visible before we test
 878          * NEED_RESCHED:
 879          */
 880         smp_mb();
 881
 882         if (unlikely(need_resched())) {
 883                 current_thread_info()->status |= TS_POLLING;
 884                 local_irq_enable();
 885                 return 0;
 886         }
 887
 888         /*
 889          * Must be done before busmaster disable as we might need to
 890          * access HPET !
 891          */
 892         acpi_state_timer_broadcast(pr, cx, 1);
 893
 894         if (cx->type == ACPI_STATE_C3)
 895                 ACPI_FLUSH_CPU_CACHE();
 896
 897         kt1 = ktime_get_real();
 898         /* Tell the scheduler that we are going deep-idle: */
 899         sched_clock_idle_sleep_event();
 900         acpi_idle_do_entry(cx);
 901         kt2 = ktime_get_real();
 902         idle_time =  ktime_to_us(ktime_sub(kt2, kt1));
 903
 904         sleep_ticks = us_to_pm_timer_ticks(idle_time);
 905
 906         /* Tell the scheduler how much we idled: */
 907         sched_clock_idle_wakeup_event(sleep_ticks*PM_TIMER_TICK_NS);
 908
 909         local_irq_enable();
 910         current_thread_info()->status |= TS_POLLING;
 911
 912         cx->usage++;
 913
 914         acpi_state_timer_broadcast(pr, cx, 0);
 915         cx->time += sleep_ticks;
 916         return idle_time;
 917 }
 918
 919 static int c3_cpu_count;
 920 static DEFINE_SPINLOCK(c3_lock);
 921
 922 /**
 923  * acpi_idle_enter_bm - enters C3 with proper BM handling
 924  * @dev: the target CPU
 925  * @state: the state data
 926  *
 927  * If BM is detected, the deepest non-C3 idle state is entered instead.
 928  */
 929 static int acpi_idle_enter_bm(struct cpuidle_device *dev,
 930                               struct cpuidle_state *state)
 931 {
 932         struct acpi_processor *pr;
 933         struct acpi_processor_cx *cx = cpuidle_get_statedata(state);
 934         ktime_t  kt1, kt2;
 935         s64 idle_time;
 936         s64 sleep_ticks = 0;
 937
 938
 939         pr = __get_cpu_var(processors);
 940
 941         if (unlikely(!pr))
 942                 return 0;
 943
 944         if (acpi_idle_suspend)
 945                 return(acpi_idle_enter_c1(dev, state));
 946
 947         if (acpi_idle_bm_check()) {
 948                 if (dev->safe_state) {
 949                         dev->last_state = dev->safe_state;
 950                         return dev->safe_state->enter(dev, dev->safe_state);
 951                 } else {
 952                         local_irq_disable();
 953                         acpi_safe_halt();
 954                         local_irq_enable();
 955                         return 0;
 956                 }
 957         }
 958
 959         local_irq_disable();
 960         current_thread_info()->status &= ~TS_POLLING;
 961         /*
 962          * TS_POLLING-cleared state must be visible before we test
 963          * NEED_RESCHED:
 964          */
 965         smp_mb();
 966
 967         if (unlikely(need_resched())) {
 968                 current_thread_info()->status |= TS_POLLING;
 969                 local_irq_enable();
 970                 return 0;
 971         }
 972
 973         acpi_unlazy_tlb(smp_processor_id());
 974
 975         /* Tell the scheduler that we are going deep-idle: */
 976         sched_clock_idle_sleep_event();
 977         /*
 978          * Must be done before busmaster disable as we might need to
 979          * access HPET !
 980          */
 981         acpi_state_timer_broadcast(pr, cx, 1);
 982
 983         kt1 = ktime_get_real();
 984         /*
 985          * disable bus master
 986          * bm_check implies we need ARB_DIS
 987          * !bm_check implies we need cache flush
 988          * bm_control implies whether we can do ARB_DIS
 989          *
 990          * That leaves a case where bm_check is set and bm_control is
 991          * not set. In that case we cannot do much, we enter C3
 992          * without doing anything.
 993          */
 994         if (pr->flags.bm_check && pr->flags.bm_control) {
 995                 spin_lock(&c3_lock);
 996                 c3_cpu_count++;
 997                 /* Disable bus master arbitration when all CPUs are in C3 */
 998                 if (c3_cpu_count == num_online_cpus())
 999                         acpi_write_bit_register(ACPI_BITREG_ARB_DISABLE, 1);
1000                 spin_unlock(&c3_lock);
1001         } else if (!pr->flags.bm_check) {
1002                 ACPI_FLUSH_CPU_CACHE();
1003         }
1004
1005         acpi_idle_do_entry(cx);
1006
1007         /* Re-enable bus master arbitration */
1008         if (pr->flags.bm_check && pr->flags.bm_control) {
1009                 spin_lock(&c3_lock);
1010                 acpi_write_bit_register(ACPI_BITREG_ARB_DISABLE, 0);
1011                 c3_cpu_count--;
1012                 spin_unlock(&c3_lock);
1013         }
1014         kt2 = ktime_get_real();
1015         idle_time =  ktime_to_us(ktime_sub(kt2, kt1));
1016
1017         sleep_ticks = us_to_pm_timer_ticks(idle_time);
1018         /* Tell the scheduler how much we idled: */
1019         sched_clock_idle_wakeup_event(sleep_ticks*PM_TIMER_TICK_NS);
1020
1021         local_irq_enable();
1022         current_thread_info()->status |= TS_POLLING;
1023
1024         cx->usage++;
1025
1026         acpi_state_timer_broadcast(pr, cx, 0);
1027         cx->time += sleep_ticks;
1028         return idle_time;
1029 }
1030
1031 struct cpuidle_driver acpi_idle_driver = {
1032         .name =         "acpi_idle",
1033         .owner =        THIS_MODULE,
1034 };
1035
1036 /**
1037  * acpi_processor_setup_cpuidle - prepares and configures CPUIDLE
1038  * @pr: the ACPI processor
1039  */
1040 static int acpi_processor_setup_cpuidle(struct acpi_processor *pr)
1041 {
1042         int i, count = CPUIDLE_DRIVER_STATE_START;
1043         struct acpi_processor_cx *cx;
1044         struct cpuidle_state *state;
1045         struct cpuidle_device *dev = &pr->power.dev;
1046
1047         if (!pr->flags.power_setup_done)
1048                 return -EINVAL;
1049
1050         if (pr->flags.power == 0) {
1051                 return -EINVAL;
1052         }
1053
1054         dev->cpu = pr->id;
1055         for (i = 0; i < CPUIDLE_STATE_MAX; i++) {
1056                 dev->states[i].name[0] = '\0';
1057                 dev->states[i].desc[0] = '\0';
1058         }
1059
1060         if (max_cstate == 0)
1061                 max_cstate = 1;
1062
1063         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER && i <= max_cstate; i++) {
1064                 cx = &pr->power.states[i];
1065                 state = &dev->states[count];
1066
1067                 if (!cx->valid)
1068                         continue;
1069
1070 #ifdef CONFIG_HOTPLUG_CPU
1071                 if ((cx->type != ACPI_STATE_C1) && (num_online_cpus() > 1) &&
1072                     !pr->flags.has_cst &&
1073                     !(acpi_gbl_FADT.flags & ACPI_FADT_C2_MP_SUPPORTED))
1074                         continue;
1075 #endif
1076                 cpuidle_set_statedata(state, cx);
1077
1078                 snprintf(state->name, CPUIDLE_NAME_LEN, "C%d", i);
1079                 strncpy(state->desc, cx->desc, CPUIDLE_DESC_LEN);
1080                 state->exit_latency = cx->latency;
1081                 state->target_residency = cx->latency * latency_factor;
1082                 state->power_usage = cx->power;
1083
1084                 state->flags = 0;
1085                 switch (cx->type) {
1086                         case ACPI_STATE_C1:
1087                         state->flags |= CPUIDLE_FLAG_SHALLOW;
1088                         if (cx->entry_method == ACPI_CSTATE_FFH)
1089                                 state->flags |= CPUIDLE_FLAG_TIME_VALID;
1090
1091                         state->enter = acpi_idle_enter_c1;
1092                         dev->safe_state = state;
1093                         break;
1094
1095                         case ACPI_STATE_C2:
1096                         state->flags |= CPUIDLE_FLAG_BALANCED;
1097                         state->flags |= CPUIDLE_FLAG_TIME_VALID;
1098                         state->enter = acpi_idle_enter_simple;
1099                         dev->safe_state = state;
1100                         break;
1101
1102                         case ACPI_STATE_C3:
1103                         state->flags |= CPUIDLE_FLAG_DEEP;
1104                         state->flags |= CPUIDLE_FLAG_TIME_VALID;
1105                         state->flags |= CPUIDLE_FLAG_CHECK_BM;
1106                         state->enter = pr->flags.bm_check ?
1107                                         acpi_idle_enter_bm :
1108                                         acpi_idle_enter_simple;
1109                         break;
1110                 }
1111
1112                 count++;
1113                 if (count == CPUIDLE_STATE_MAX)
1114                         break;
1115         }
1116
1117         dev->state_count = count;
1118
1119         if (!count)
1120                 return -EINVAL;
1121
1122         return 0;
1123 }
1124
1125 int acpi_processor_cst_has_changed(struct acpi_processor *pr)
1126 {
1127         int ret = 0;
1128
1129         if (boot_option_idle_override)
1130                 return 0;
1131
1132         if (!pr)
1133                 return -EINVAL;
1134
1135         if (nocst) {
1136                 return -ENODEV;
1137         }
1138
1139         if (!pr->flags.power_setup_done)
1140                 return -ENODEV;
1141
1142         cpuidle_pause_and_lock();
1143         cpuidle_disable_device(&pr->power.dev);
1144         acpi_processor_get_power_info(pr);
1145         if (pr->flags.power) {
1146                 acpi_processor_setup_cpuidle(pr);
1147                 ret = cpuidle_enable_device(&pr->power.dev);
1148         }
1149         cpuidle_resume_and_unlock();
1150
1151         return ret;
1152 }
1153
1154 int __cpuinit acpi_processor_power_init(struct acpi_processor *pr,
1155                               struct acpi_device *device)
1156 {
1157         acpi_status status = 0;
1158         static int first_run;
1159         struct proc_dir_entry *entry = NULL;
1160         unsigned int i;
1161
1162         if (boot_option_idle_override)
1163                 return 0;
1164
1165         if (!first_run) {
1166                 if (idle_halt) {
1167                         /*
1168                          * When the boot option of "idle=halt" is added, halt
1169                          * is used for CPU IDLE.
1170                          * In such case C2/C3 is meaningless. So the max_cstate
1171                          * is set to one.
1172                          */
1173                         max_cstate = 1;
1174                 }
1175                 dmi_check_system(processor_power_dmi_table);
1176                 max_cstate = acpi_processor_cstate_check(max_cstate);
1177                 if (max_cstate < ACPI_C_STATES_MAX)
1178                         printk(KERN_NOTICE
1179                                "ACPI: processor limited to max C-state %d\n",
1180                                max_cstate);
1181                 first_run++;
1182         }
1183
1184         if (!pr)
1185                 return -EINVAL;
1186
1187         if (acpi_gbl_FADT.cst_control && !nocst) {
1188                 status =
1189                     acpi_os_write_port(acpi_gbl_FADT.smi_command, acpi_gbl_FADT.cst_control, 8);
1190                 if (ACPI_FAILURE(status)) {
1191                         ACPI_EXCEPTION((AE_INFO, status,
1192                                         "Notifying BIOS of _CST ability failed"));
1193                 }
1194         }
1195
1196         acpi_processor_get_power_info(pr);
1197         pr->flags.power_setup_done = 1;
1198
1199         /*
1200          * Install the idle handler if processor power management is supported.
1201          * Note that we use previously set idle handler will be used on
1202          * platforms that only support C1.
1203          */
1204         if (pr->flags.power) {
1205                 acpi_processor_setup_cpuidle(pr);
1206                 if (cpuidle_register_device(&pr->power.dev))
1207                         return -EIO;
1208
1209                 printk(KERN_INFO PREFIX "CPU%d (power states:", pr->id);
1210                 for (i = 1; i <= pr->power.count; i++)
1211                         if (pr->power.states[i].valid)
1212                                 printk(" C%d[C%d]", i,
1213                                        pr->power.states[i].type);
1214                 printk(")\n");
1215         }
1216
1217         /* 'power' [R] */
1218         entry = proc_create_data(ACPI_PROCESSOR_FILE_POWER,
1219                                  S_IRUGO, acpi_device_dir(device),
1220                                  &acpi_processor_power_fops,
1221                                  acpi_driver_data(device));
1222         if (!entry)
1223                 return -EIO;
1224         return 0;
1225 }
1226
1227 int acpi_processor_power_exit(struct acpi_processor *pr,
1228                               struct acpi_device *device)
1229 {
1230         if (boot_option_idle_override)
1231                 return 0;
1232
1233         cpuidle_unregister_device(&pr->power.dev);
1234         pr->flags.power_setup_done = 0;
1235
1236         if (acpi_device_dir(device))
1237                 remove_proc_entry(ACPI_PROCESSOR_FILE_POWER,
1238                                   acpi_device_dir(device));
1239
1240         return 0;
1241 }