SAFE public projects git trees. - safe/jmp/linux-2.6/blob - drivers/acpi/processor_idle.c

   1 /*
   2  * processor_idle - idle state submodule to the ACPI processor driver
   3  *
   4  *  Copyright (C) 2001, 2002 Andy Grover <andrew.grover@intel.com>
   5  *  Copyright (C) 2001, 2002 Paul Diefenbaugh <paul.s.diefenbaugh@intel.com>
   6  *  Copyright (C) 2004, 2005 Dominik Brodowski <linux@brodo.de>
   7  *  Copyright (C) 2004  Anil S Keshavamurthy <anil.s.keshavamurthy@intel.com>
   8  *                      - Added processor hotplug support
   9  *  Copyright (C) 2005  Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>
  10  *                      - Added support for C3 on SMP
  11  *
  12  * ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
  13  *
  14  *  This program is free software; you can redistribute it and/or modify
  15  *  it under the terms of the GNU General Public License as published by
  16  *  the Free Software Foundation; either version 2 of the License, or (at
  17  *  your option) any later version.
  18  *
  19  *  This program is distributed in the hope that it will be useful, but
  20  *  WITHOUT ANY WARRANTY; without even the implied warranty of
  21  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  22  *  General Public License for more details.
  23  *
  24  *  You should have received a copy of the GNU General Public License along
  25  *  with this program; if not, write to the Free Software Foundation, Inc.,
  26  *  59 Temple Place, Suite 330, Boston, MA 02111-1307 USA.
  27  *
  28  * ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
  29  */
  30
  31 #include <linux/kernel.h>
  32 #include <linux/module.h>
  33 #include <linux/init.h>
  34 #include <linux/cpufreq.h>
  35 #include <linux/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 int tsc_halts_in_c(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 0;
 253
 254                 /*FALL THROUGH*/
 255         default:
 256                 return state > ACPI_STATE_C1;
 257         }
 258 }
 259 #endif
 260
 261 static int acpi_processor_get_power_info_fadt(struct acpi_processor *pr)
 262 {
 263
 264         if (!pr)
 265                 return -EINVAL;
 266
 267         if (!pr->pblk)
 268                 return -ENODEV;
 269
 270         /* if info is obtained from pblk/fadt, type equals state */
 271         pr->power.states[ACPI_STATE_C2].type = ACPI_STATE_C2;
 272         pr->power.states[ACPI_STATE_C3].type = ACPI_STATE_C3;
 273
 274 #ifndef CONFIG_HOTPLUG_CPU
 275         /*
 276          * Check for P_LVL2_UP flag before entering C2 and above on
 277          * an SMP system.
 278          */
 279         if ((num_online_cpus() > 1) &&
 280             !(acpi_gbl_FADT.flags & ACPI_FADT_C2_MP_SUPPORTED))
 281                 return -ENODEV;
 282 #endif
 283
 284         /* determine C2 and C3 address from pblk */
 285         pr->power.states[ACPI_STATE_C2].address = pr->pblk + 4;
 286         pr->power.states[ACPI_STATE_C3].address = pr->pblk + 5;
 287
 288         /* determine latencies from FADT */
 289         pr->power.states[ACPI_STATE_C2].latency = acpi_gbl_FADT.C2latency;
 290         pr->power.states[ACPI_STATE_C3].latency = acpi_gbl_FADT.C3latency;
 291
 292         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 293                           "lvl2[0x%08x] lvl3[0x%08x]\n",
 294                           pr->power.states[ACPI_STATE_C2].address,
 295                           pr->power.states[ACPI_STATE_C3].address));
 296
 297         return 0;
 298 }
 299
 300 static int acpi_processor_get_power_info_default(struct acpi_processor *pr)
 301 {
 302         if (!pr->power.states[ACPI_STATE_C1].valid) {
 303                 /* set the first C-State to C1 */
 304                 /* all processors need to support C1 */
 305                 pr->power.states[ACPI_STATE_C1].type = ACPI_STATE_C1;
 306                 pr->power.states[ACPI_STATE_C1].valid = 1;
 307                 pr->power.states[ACPI_STATE_C1].entry_method = ACPI_CSTATE_HALT;
 308         }
 309         /* the C0 state only exists as a filler in our array */
 310         pr->power.states[ACPI_STATE_C0].valid = 1;
 311         return 0;
 312 }
 313
 314 static int acpi_processor_get_power_info_cst(struct acpi_processor *pr)
 315 {
 316         acpi_status status = 0;
 317         acpi_integer count;
 318         int current_count;
 319         int i;
 320         struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
 321         union acpi_object *cst;
 322
 323
 324         if (nocst)
 325                 return -ENODEV;
 326
 327         current_count = 0;
 328
 329         status = acpi_evaluate_object(pr->handle, "_CST", NULL, &buffer);
 330         if (ACPI_FAILURE(status)) {
 331                 ACPI_DEBUG_PRINT((ACPI_DB_INFO, "No _CST, giving up\n"));
 332                 return -ENODEV;
 333         }
 334
 335         cst = buffer.pointer;
 336
 337         /* There must be at least 2 elements */
 338         if (!cst || (cst->type != ACPI_TYPE_PACKAGE) || cst->package.count < 2) {
 339                 printk(KERN_ERR PREFIX "not enough elements in _CST\n");
 340                 status = -EFAULT;
 341                 goto end;
 342         }
 343
 344         count = cst->package.elements[0].integer.value;
 345
 346         /* Validate number of power states. */
 347         if (count < 1 || count != cst->package.count - 1) {
 348                 printk(KERN_ERR PREFIX "count given by _CST is not valid\n");
 349                 status = -EFAULT;
 350                 goto end;
 351         }
 352
 353         /* Tell driver that at least _CST is supported. */
 354         pr->flags.has_cst = 1;
 355
 356         for (i = 1; i <= count; i++) {
 357                 union acpi_object *element;
 358                 union acpi_object *obj;
 359                 struct acpi_power_register *reg;
 360                 struct acpi_processor_cx cx;
 361
 362                 memset(&cx, 0, sizeof(cx));
 363
 364                 element = &(cst->package.elements[i]);
 365                 if (element->type != ACPI_TYPE_PACKAGE)
 366                         continue;
 367
 368                 if (element->package.count != 4)
 369                         continue;
 370
 371                 obj = &(element->package.elements[0]);
 372
 373                 if (obj->type != ACPI_TYPE_BUFFER)
 374                         continue;
 375
 376                 reg = (struct acpi_power_register *)obj->buffer.pointer;
 377
 378                 if (reg->space_id != ACPI_ADR_SPACE_SYSTEM_IO &&
 379                     (reg->space_id != ACPI_ADR_SPACE_FIXED_HARDWARE))
 380                         continue;
 381
 382                 /* There should be an easy way to extract an integer... */
 383                 obj = &(element->package.elements[1]);
 384                 if (obj->type != ACPI_TYPE_INTEGER)
 385                         continue;
 386
 387                 cx.type = obj->integer.value;
 388                 /*
 389                  * Some buggy BIOSes won't list C1 in _CST -
 390                  * Let acpi_processor_get_power_info_default() handle them later
 391                  */
 392                 if (i == 1 && cx.type != ACPI_STATE_C1)
 393                         current_count++;
 394
 395                 cx.address = reg->address;
 396                 cx.index = current_count + 1;
 397
 398                 cx.entry_method = ACPI_CSTATE_SYSTEMIO;
 399                 if (reg->space_id == ACPI_ADR_SPACE_FIXED_HARDWARE) {
 400                         if (acpi_processor_ffh_cstate_probe
 401                                         (pr->id, &cx, reg) == 0) {
 402                                 cx.entry_method = ACPI_CSTATE_FFH;
 403                         } else if (cx.type == ACPI_STATE_C1) {
 404                                 /*
 405                                  * C1 is a special case where FIXED_HARDWARE
 406                                  * can be handled in non-MWAIT way as well.
 407                                  * In that case, save this _CST entry info.
 408                                  * Otherwise, ignore this info and continue.
 409                                  */
 410                                 cx.entry_method = ACPI_CSTATE_HALT;
 411                                 snprintf(cx.desc, ACPI_CX_DESC_LEN, "ACPI HLT");
 412                         } else {
 413                                 continue;
 414                         }
 415                         if (cx.type == ACPI_STATE_C1 &&
 416                                         (idle_halt || idle_nomwait)) {
 417                                 /*
 418                                  * In most cases the C1 space_id obtained from
 419                                  * _CST object is FIXED_HARDWARE access mode.
 420                                  * But when the option of idle=halt is added,
 421                                  * the entry_method type should be changed from
 422                                  * CSTATE_FFH to CSTATE_HALT.
 423                                  * When the option of idle=nomwait is added,
 424                                  * the C1 entry_method type should be
 425                                  * CSTATE_HALT.
 426                                  */
 427                                 cx.entry_method = ACPI_CSTATE_HALT;
 428                                 snprintf(cx.desc, ACPI_CX_DESC_LEN, "ACPI HLT");
 429                         }
 430                 } else {
 431                         snprintf(cx.desc, ACPI_CX_DESC_LEN, "ACPI IOPORT 0x%x",
 432                                  cx.address);
 433                 }
 434
 435                 if (cx.type == ACPI_STATE_C1) {
 436                         cx.valid = 1;
 437                 }
 438
 439                 obj = &(element->package.elements[2]);
 440                 if (obj->type != ACPI_TYPE_INTEGER)
 441                         continue;
 442
 443                 cx.latency = obj->integer.value;
 444
 445                 obj = &(element->package.elements[3]);
 446                 if (obj->type != ACPI_TYPE_INTEGER)
 447                         continue;
 448
 449                 cx.power = obj->integer.value;
 450
 451                 current_count++;
 452                 memcpy(&(pr->power.states[current_count]), &cx, sizeof(cx));
 453
 454                 /*
 455                  * We support total ACPI_PROCESSOR_MAX_POWER - 1
 456                  * (From 1 through ACPI_PROCESSOR_MAX_POWER - 1)
 457                  */
 458                 if (current_count >= (ACPI_PROCESSOR_MAX_POWER - 1)) {
 459                         printk(KERN_WARNING
 460                                "Limiting number of power states to max (%d)\n",
 461                                ACPI_PROCESSOR_MAX_POWER);
 462                         printk(KERN_WARNING
 463                                "Please increase ACPI_PROCESSOR_MAX_POWER if needed.\n");
 464                         break;
 465                 }
 466         }
 467
 468         ACPI_DEBUG_PRINT((ACPI_DB_INFO, "Found %d power states\n",
 469                           current_count));
 470
 471         /* Validate number of power states discovered */
 472         if (current_count < 2)
 473                 status = -EFAULT;
 474
 475       end:
 476         kfree(buffer.pointer);
 477
 478         return status;
 479 }
 480
 481 static void acpi_processor_power_verify_c2(struct acpi_processor_cx *cx)
 482 {
 483
 484         if (!cx->address)
 485                 return;
 486
 487         /*
 488          * C2 latency must be less than or equal to 100
 489          * microseconds.
 490          */
 491         else if (cx->latency > ACPI_PROCESSOR_MAX_C2_LATENCY) {
 492                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 493                                   "latency too large [%d]\n", cx->latency));
 494                 return;
 495         }
 496
 497         /*
 498          * Otherwise we've met all of our C2 requirements.
 499          * Normalize the C2 latency to expidite policy
 500          */
 501         cx->valid = 1;
 502
 503         cx->latency_ticks = cx->latency;
 504
 505         return;
 506 }
 507
 508 static void acpi_processor_power_verify_c3(struct acpi_processor *pr,
 509                                            struct acpi_processor_cx *cx)
 510 {
 511         static int bm_check_flag;
 512
 513
 514         if (!cx->address)
 515                 return;
 516
 517         /*
 518          * C3 latency must be less than or equal to 1000
 519          * microseconds.
 520          */
 521         else if (cx->latency > ACPI_PROCESSOR_MAX_C3_LATENCY) {
 522                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 523                                   "latency too large [%d]\n", cx->latency));
 524                 return;
 525         }
 526
 527         /*
 528          * PIIX4 Erratum #18: We don't support C3 when Type-F (fast)
 529          * DMA transfers are used by any ISA device to avoid livelock.
 530          * Note that we could disable Type-F DMA (as recommended by
 531          * the erratum), but this is known to disrupt certain ISA
 532          * devices thus we take the conservative approach.
 533          */
 534         else if (errata.piix4.fdma) {
 535                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 536                                   "C3 not supported on PIIX4 with Type-F DMA\n"));
 537                 return;
 538         }
 539
 540         /* All the logic here assumes flags.bm_check is same across all CPUs */
 541         if (!bm_check_flag) {
 542                 /* Determine whether bm_check is needed based on CPU  */
 543                 acpi_processor_power_init_bm_check(&(pr->flags), pr->id);
 544                 bm_check_flag = pr->flags.bm_check;
 545         } else {
 546                 pr->flags.bm_check = bm_check_flag;
 547         }
 548
 549         if (pr->flags.bm_check) {
 550                 if (!pr->flags.bm_control) {
 551                         if (pr->flags.has_cst != 1) {
 552                                 /* bus mastering control is necessary */
 553                                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 554                                         "C3 support requires BM control\n"));
 555                                 return;
 556                         } else {
 557                                 /* Here we enter C3 without bus mastering */
 558                                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 559                                         "C3 support without BM control\n"));
 560                         }
 561                 }
 562         } else {
 563                 /*
 564                  * WBINVD should be set in fadt, for C3 state to be
 565                  * supported on when bm_check is not required.
 566                  */
 567                 if (!(acpi_gbl_FADT.flags & ACPI_FADT_WBINVD)) {
 568                         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 569                                           "Cache invalidation should work properly"
 570                                           " for C3 to be enabled on SMP systems\n"));
 571                         return;
 572                 }
 573         }
 574
 575         /*
 576          * Otherwise we've met all of our C3 requirements.
 577          * Normalize the C3 latency to expidite policy.  Enable
 578          * checking of bus mastering status (bm_check) so we can
 579          * use this in our C3 policy
 580          */
 581         cx->valid = 1;
 582
 583         cx->latency_ticks = cx->latency;
 584         /*
 585          * On older chipsets, BM_RLD needs to be set
 586          * in order for Bus Master activity to wake the
 587          * system from C3.  Newer chipsets handle DMA
 588          * during C3 automatically and BM_RLD is a NOP.
 589          * In either case, the proper way to
 590          * handle BM_RLD is to set it and leave it set.
 591          */
 592         acpi_write_bit_register(ACPI_BITREG_BUS_MASTER_RLD, 1);
 593
 594         return;
 595 }
 596
 597 static int acpi_processor_power_verify(struct acpi_processor *pr)
 598 {
 599         unsigned int i;
 600         unsigned int working = 0;
 601
 602         pr->power.timer_broadcast_on_state = INT_MAX;
 603
 604         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
 605                 struct acpi_processor_cx *cx = &pr->power.states[i];
 606
 607 #if defined (CONFIG_GENERIC_TIME) && defined (CONFIG_X86)
 608                 /* TSC could halt in idle, so notify users */
 609                 if (tsc_halts_in_c(cx->type))
 610                         mark_tsc_unstable("TSC halts in idle");;
 611 #endif
 612                 switch (cx->type) {
 613                 case ACPI_STATE_C1:
 614                         cx->valid = 1;
 615                         break;
 616
 617                 case ACPI_STATE_C2:
 618                         acpi_processor_power_verify_c2(cx);
 619                         if (cx->valid)
 620                                 acpi_timer_check_state(i, pr, cx);
 621                         break;
 622
 623                 case ACPI_STATE_C3:
 624                         acpi_processor_power_verify_c3(pr, cx);
 625                         if (cx->valid)
 626                                 acpi_timer_check_state(i, pr, cx);
 627                         break;
 628                 }
 629
 630                 if (cx->valid)
 631                         working++;
 632         }
 633
 634         acpi_propagate_timer_broadcast(pr);
 635
 636         return (working);
 637 }
 638
 639 static int acpi_processor_get_power_info(struct acpi_processor *pr)
 640 {
 641         unsigned int i;
 642         int result;
 643
 644
 645         /* NOTE: the idle thread may not be running while calling
 646          * this function */
 647
 648         /* Zero initialize all the C-states info. */
 649         memset(pr->power.states, 0, sizeof(pr->power.states));
 650
 651         result = acpi_processor_get_power_info_cst(pr);
 652         if (result == -ENODEV)
 653                 result = acpi_processor_get_power_info_fadt(pr);
 654
 655         if (result)
 656                 return result;
 657
 658         acpi_processor_get_power_info_default(pr);
 659
 660         pr->power.count = acpi_processor_power_verify(pr);
 661
 662         /*
 663          * if one state of type C2 or C3 is available, mark this
 664          * CPU as being "idle manageable"
 665          */
 666         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
 667                 if (pr->power.states[i].valid) {
 668                         pr->power.count = i;
 669                         if (pr->power.states[i].type >= ACPI_STATE_C2)
 670                                 pr->flags.power = 1;
 671                 }
 672         }
 673
 674         return 0;
 675 }
 676
 677 static int acpi_processor_power_seq_show(struct seq_file *seq, void *offset)
 678 {
 679         struct acpi_processor *pr = seq->private;
 680         unsigned int i;
 681
 682
 683         if (!pr)
 684                 goto end;
 685
 686         seq_printf(seq, "active state:            C%zd\n"
 687                    "max_cstate:              C%d\n"
 688                    "maximum allowed latency: %d usec\n",
 689                    pr->power.state ? pr->power.state - pr->power.states : 0,
 690                    max_cstate, pm_qos_requirement(PM_QOS_CPU_DMA_LATENCY));
 691
 692         seq_puts(seq, "states:\n");
 693
 694         for (i = 1; i <= pr->power.count; i++) {
 695                 seq_printf(seq, "   %cC%d:                  ",
 696                            (&pr->power.states[i] ==
 697                             pr->power.state ? '*' : ' '), i);
 698
 699                 if (!pr->power.states[i].valid) {
 700                         seq_puts(seq, "<not supported>\n");
 701                         continue;
 702                 }
 703
 704                 switch (pr->power.states[i].type) {
 705                 case ACPI_STATE_C1:
 706                         seq_printf(seq, "type[C1] ");
 707                         break;
 708                 case ACPI_STATE_C2:
 709                         seq_printf(seq, "type[C2] ");
 710                         break;
 711                 case ACPI_STATE_C3:
 712                         seq_printf(seq, "type[C3] ");
 713                         break;
 714                 default:
 715                         seq_printf(seq, "type[--] ");
 716                         break;
 717                 }
 718
 719                 if (pr->power.states[i].promotion.state)
 720                         seq_printf(seq, "promotion[C%zd] ",
 721                                    (pr->power.states[i].promotion.state -
 722                                     pr->power.states));
 723                 else
 724                         seq_puts(seq, "promotion[--] ");
 725
 726                 if (pr->power.states[i].demotion.state)
 727                         seq_printf(seq, "demotion[C%zd] ",
 728                                    (pr->power.states[i].demotion.state -
 729                                     pr->power.states));
 730                 else
 731                         seq_puts(seq, "demotion[--] ");
 732
 733                 seq_printf(seq, "latency[%03d] usage[%08d] duration[%020llu]\n",
 734                            pr->power.states[i].latency,
 735                            pr->power.states[i].usage,
 736                            (unsigned long long)pr->power.states[i].time);
 737         }
 738
 739       end:
 740         return 0;
 741 }
 742
 743 static int acpi_processor_power_open_fs(struct inode *inode, struct file *file)
 744 {
 745         return single_open(file, acpi_processor_power_seq_show,
 746                            PDE(inode)->data);
 747 }
 748
 749 static const struct file_operations acpi_processor_power_fops = {
 750         .owner = THIS_MODULE,
 751         .open = acpi_processor_power_open_fs,
 752         .read = seq_read,
 753         .llseek = seq_lseek,
 754         .release = single_release,
 755 };
 756
 757
 758 /**
 759  * acpi_idle_bm_check - checks if bus master activity was detected
 760  */
 761 static int acpi_idle_bm_check(void)
 762 {
 763         u32 bm_status = 0;
 764
 765         acpi_read_bit_register(ACPI_BITREG_BUS_MASTER_STATUS, &bm_status);
 766         if (bm_status)
 767                 acpi_write_bit_register(ACPI_BITREG_BUS_MASTER_STATUS, 1);
 768         /*
 769          * PIIX4 Erratum #18: Note that BM_STS doesn't always reflect
 770          * the true state of bus mastering activity; forcing us to
 771          * manually check the BMIDEA bit of each IDE channel.
 772          */
 773         else if (errata.piix4.bmisx) {
 774                 if ((inb_p(errata.piix4.bmisx + 0x02) & 0x01)
 775                     || (inb_p(errata.piix4.bmisx + 0x0A) & 0x01))
 776                         bm_status = 1;
 777         }
 778         return bm_status;
 779 }
 780
 781 /**
 782  * acpi_idle_do_entry - a helper function that does C2 and C3 type entry
 783  * @cx: cstate data
 784  *
 785  * Caller disables interrupt before call and enables interrupt after return.
 786  */
 787 static inline void acpi_idle_do_entry(struct acpi_processor_cx *cx)
 788 {
 789         /* Don't trace irqs off for idle */
 790         stop_critical_timings();
 791         if (cx->entry_method == ACPI_CSTATE_FFH) {
 792                 /* Call into architectural FFH based C-state */
 793                 acpi_processor_ffh_cstate_enter(cx);
 794         } else if (cx->entry_method == ACPI_CSTATE_HALT) {
 795                 acpi_safe_halt();
 796         } else {
 797                 int unused;
 798                 /* IO port based C-state */
 799                 inb(cx->address);
 800                 /* Dummy wait op - must do something useless after P_LVL2 read
 801                    because chipsets cannot guarantee that STPCLK# signal
 802                    gets asserted in time to freeze execution properly. */
 803                 unused = inl(acpi_gbl_FADT.xpm_timer_block.address);
 804         }
 805         start_critical_timings();
 806 }
 807
 808 /**
 809  * acpi_idle_enter_c1 - enters an ACPI C1 state-type
 810  * @dev: the target CPU
 811  * @state: the state data
 812  *
 813  * This is equivalent to the HALT instruction.
 814  */
 815 static int acpi_idle_enter_c1(struct cpuidle_device *dev,
 816                               struct cpuidle_state *state)
 817 {
 818         ktime_t  kt1, kt2;
 819         s64 idle_time;
 820         struct acpi_processor *pr;
 821         struct acpi_processor_cx *cx = cpuidle_get_statedata(state);
 822
 823         pr = __get_cpu_var(processors);
 824
 825         if (unlikely(!pr))
 826                 return 0;
 827
 828         local_irq_disable();
 829
 830         /* Do not access any ACPI IO ports in suspend path */
 831         if (acpi_idle_suspend) {
 832                 acpi_safe_halt();
 833                 local_irq_enable();
 834                 return 0;
 835         }
 836
 837         kt1 = ktime_get_real();
 838         acpi_idle_do_entry(cx);
 839         kt2 = ktime_get_real();
 840         idle_time =  ktime_to_us(ktime_sub(kt2, kt1));
 841
 842         local_irq_enable();
 843         cx->usage++;
 844
 845         return idle_time;
 846 }
 847
 848 /**
 849  * acpi_idle_enter_simple - enters an ACPI state without BM handling
 850  * @dev: the target CPU
 851  * @state: the state data
 852  */
 853 static int acpi_idle_enter_simple(struct cpuidle_device *dev,
 854                                   struct cpuidle_state *state)
 855 {
 856         struct acpi_processor *pr;
 857         struct acpi_processor_cx *cx = cpuidle_get_statedata(state);
 858         ktime_t  kt1, kt2;
 859         s64 idle_time;
 860         s64 sleep_ticks = 0;
 861
 862         pr = __get_cpu_var(processors);
 863
 864         if (unlikely(!pr))
 865                 return 0;
 866
 867         if (acpi_idle_suspend)
 868                 return(acpi_idle_enter_c1(dev, state));
 869
 870         local_irq_disable();
 871         current_thread_info()->status &= ~TS_POLLING;
 872         /*
 873          * TS_POLLING-cleared state must be visible before we test
 874          * NEED_RESCHED:
 875          */
 876         smp_mb();
 877
 878         if (unlikely(need_resched())) {
 879                 current_thread_info()->status |= TS_POLLING;
 880                 local_irq_enable();
 881                 return 0;
 882         }
 883
 884         /*
 885          * Must be done before busmaster disable as we might need to
 886          * access HPET !
 887          */
 888         acpi_state_timer_broadcast(pr, cx, 1);
 889
 890         if (cx->type == ACPI_STATE_C3)
 891                 ACPI_FLUSH_CPU_CACHE();
 892
 893         kt1 = ktime_get_real();
 894         /* Tell the scheduler that we are going deep-idle: */
 895         sched_clock_idle_sleep_event();
 896         acpi_idle_do_entry(cx);
 897         kt2 = ktime_get_real();
 898         idle_time =  ktime_to_us(ktime_sub(kt2, kt1));
 899
 900         sleep_ticks = us_to_pm_timer_ticks(idle_time);
 901
 902         /* Tell the scheduler how much we idled: */
 903         sched_clock_idle_wakeup_event(sleep_ticks*PM_TIMER_TICK_NS);
 904
 905         local_irq_enable();
 906         current_thread_info()->status |= TS_POLLING;
 907
 908         cx->usage++;
 909
 910         acpi_state_timer_broadcast(pr, cx, 0);
 911         cx->time += sleep_ticks;
 912         return idle_time;
 913 }
 914
 915 static int c3_cpu_count;
 916 static DEFINE_SPINLOCK(c3_lock);
 917
 918 /**
 919  * acpi_idle_enter_bm - enters C3 with proper BM handling
 920  * @dev: the target CPU
 921  * @state: the state data
 922  *
 923  * If BM is detected, the deepest non-C3 idle state is entered instead.
 924  */
 925 static int acpi_idle_enter_bm(struct cpuidle_device *dev,
 926                               struct cpuidle_state *state)
 927 {
 928         struct acpi_processor *pr;
 929         struct acpi_processor_cx *cx = cpuidle_get_statedata(state);
 930         ktime_t  kt1, kt2;
 931         s64 idle_time;
 932         s64 sleep_ticks = 0;
 933
 934
 935         pr = __get_cpu_var(processors);
 936
 937         if (unlikely(!pr))
 938                 return 0;
 939
 940         if (acpi_idle_suspend)
 941                 return(acpi_idle_enter_c1(dev, state));
 942
 943         if (acpi_idle_bm_check()) {
 944                 if (dev->safe_state) {
 945                         dev->last_state = dev->safe_state;
 946                         return dev->safe_state->enter(dev, dev->safe_state);
 947                 } else {
 948                         local_irq_disable();
 949                         acpi_safe_halt();
 950                         local_irq_enable();
 951                         return 0;
 952                 }
 953         }
 954
 955         local_irq_disable();
 956         current_thread_info()->status &= ~TS_POLLING;
 957         /*
 958          * TS_POLLING-cleared state must be visible before we test
 959          * NEED_RESCHED:
 960          */
 961         smp_mb();
 962
 963         if (unlikely(need_resched())) {
 964                 current_thread_info()->status |= TS_POLLING;
 965                 local_irq_enable();
 966                 return 0;
 967         }
 968
 969         acpi_unlazy_tlb(smp_processor_id());
 970
 971         /* Tell the scheduler that we are going deep-idle: */
 972         sched_clock_idle_sleep_event();
 973         /*
 974          * Must be done before busmaster disable as we might need to
 975          * access HPET !
 976          */
 977         acpi_state_timer_broadcast(pr, cx, 1);
 978
 979         kt1 = ktime_get_real();
 980         /*
 981          * disable bus master
 982          * bm_check implies we need ARB_DIS
 983          * !bm_check implies we need cache flush
 984          * bm_control implies whether we can do ARB_DIS
 985          *
 986          * That leaves a case where bm_check is set and bm_control is
 987          * not set. In that case we cannot do much, we enter C3
 988          * without doing anything.
 989          */
 990         if (pr->flags.bm_check && pr->flags.bm_control) {
 991                 spin_lock(&c3_lock);
 992                 c3_cpu_count++;
 993                 /* Disable bus master arbitration when all CPUs are in C3 */
 994                 if (c3_cpu_count == num_online_cpus())
 995                         acpi_write_bit_register(ACPI_BITREG_ARB_DISABLE, 1);
 996                 spin_unlock(&c3_lock);
 997         } else if (!pr->flags.bm_check) {
 998                 ACPI_FLUSH_CPU_CACHE();
 999         }
1000
1001         acpi_idle_do_entry(cx);
1002
1003         /* Re-enable bus master arbitration */
1004         if (pr->flags.bm_check && pr->flags.bm_control) {
1005                 spin_lock(&c3_lock);
1006                 acpi_write_bit_register(ACPI_BITREG_ARB_DISABLE, 0);
1007                 c3_cpu_count--;
1008                 spin_unlock(&c3_lock);
1009         }
1010         kt2 = ktime_get_real();
1011         idle_time =  ktime_to_us(ktime_sub(kt2, kt1));
1012
1013         sleep_ticks = us_to_pm_timer_ticks(idle_time);
1014         /* Tell the scheduler how much we idled: */
1015         sched_clock_idle_wakeup_event(sleep_ticks*PM_TIMER_TICK_NS);
1016
1017         local_irq_enable();
1018         current_thread_info()->status |= TS_POLLING;
1019
1020         cx->usage++;
1021
1022         acpi_state_timer_broadcast(pr, cx, 0);
1023         cx->time += sleep_ticks;
1024         return idle_time;
1025 }
1026
1027 struct cpuidle_driver acpi_idle_driver = {
1028         .name =         "acpi_idle",
1029         .owner =        THIS_MODULE,
1030 };
1031
1032 /**
1033  * acpi_processor_setup_cpuidle - prepares and configures CPUIDLE
1034  * @pr: the ACPI processor
1035  */
1036 static int acpi_processor_setup_cpuidle(struct acpi_processor *pr)
1037 {
1038         int i, count = CPUIDLE_DRIVER_STATE_START;
1039         struct acpi_processor_cx *cx;
1040         struct cpuidle_state *state;
1041         struct cpuidle_device *dev = &pr->power.dev;
1042
1043         if (!pr->flags.power_setup_done)
1044                 return -EINVAL;
1045
1046         if (pr->flags.power == 0) {
1047                 return -EINVAL;
1048         }
1049
1050         dev->cpu = pr->id;
1051         for (i = 0; i < CPUIDLE_STATE_MAX; i++) {
1052                 dev->states[i].name[0] = '\0';
1053                 dev->states[i].desc[0] = '\0';
1054         }
1055
1056         if (max_cstate == 0)
1057                 max_cstate = 1;
1058
1059         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER && i <= max_cstate; i++) {
1060                 cx = &pr->power.states[i];
1061                 state = &dev->states[count];
1062
1063                 if (!cx->valid)
1064                         continue;
1065
1066 #ifdef CONFIG_HOTPLUG_CPU
1067                 if ((cx->type != ACPI_STATE_C1) && (num_online_cpus() > 1) &&
1068                     !pr->flags.has_cst &&
1069                     !(acpi_gbl_FADT.flags & ACPI_FADT_C2_MP_SUPPORTED))
1070                         continue;
1071 #endif
1072                 cpuidle_set_statedata(state, cx);
1073
1074                 snprintf(state->name, CPUIDLE_NAME_LEN, "C%d", i);
1075                 strncpy(state->desc, cx->desc, CPUIDLE_DESC_LEN);
1076                 state->exit_latency = cx->latency;
1077                 state->target_residency = cx->latency * latency_factor;
1078                 state->power_usage = cx->power;
1079
1080                 state->flags = 0;
1081                 switch (cx->type) {
1082                         case ACPI_STATE_C1:
1083                         state->flags |= CPUIDLE_FLAG_SHALLOW;
1084                         if (cx->entry_method == ACPI_CSTATE_FFH)
1085                                 state->flags |= CPUIDLE_FLAG_TIME_VALID;
1086
1087                         state->enter = acpi_idle_enter_c1;
1088                         dev->safe_state = state;
1089                         break;
1090
1091                         case ACPI_STATE_C2:
1092                         state->flags |= CPUIDLE_FLAG_BALANCED;
1093                         state->flags |= CPUIDLE_FLAG_TIME_VALID;
1094                         state->enter = acpi_idle_enter_simple;
1095                         dev->safe_state = state;
1096                         break;
1097
1098                         case ACPI_STATE_C3:
1099                         state->flags |= CPUIDLE_FLAG_DEEP;
1100                         state->flags |= CPUIDLE_FLAG_TIME_VALID;
1101                         state->flags |= CPUIDLE_FLAG_CHECK_BM;
1102                         state->enter = pr->flags.bm_check ?
1103                                         acpi_idle_enter_bm :
1104                                         acpi_idle_enter_simple;
1105                         break;
1106                 }
1107
1108                 count++;
1109                 if (count == CPUIDLE_STATE_MAX)
1110                         break;
1111         }
1112
1113         dev->state_count = count;
1114
1115         if (!count)
1116                 return -EINVAL;
1117
1118         return 0;
1119 }
1120
1121 int acpi_processor_cst_has_changed(struct acpi_processor *pr)
1122 {
1123         int ret = 0;
1124
1125         if (boot_option_idle_override)
1126                 return 0;
1127
1128         if (!pr)
1129                 return -EINVAL;
1130
1131         if (nocst) {
1132                 return -ENODEV;
1133         }
1134
1135         if (!pr->flags.power_setup_done)
1136                 return -ENODEV;
1137
1138         cpuidle_pause_and_lock();
1139         cpuidle_disable_device(&pr->power.dev);
1140         acpi_processor_get_power_info(pr);
1141         if (pr->flags.power) {
1142                 acpi_processor_setup_cpuidle(pr);
1143                 ret = cpuidle_enable_device(&pr->power.dev);
1144         }
1145         cpuidle_resume_and_unlock();
1146
1147         return ret;
1148 }
1149
1150 int __cpuinit acpi_processor_power_init(struct acpi_processor *pr,
1151                               struct acpi_device *device)
1152 {
1153         acpi_status status = 0;
1154         static int first_run;
1155         struct proc_dir_entry *entry = NULL;
1156         unsigned int i;
1157
1158         if (boot_option_idle_override)
1159                 return 0;
1160
1161         if (!first_run) {
1162                 if (idle_halt) {
1163                         /*
1164                          * When the boot option of "idle=halt" is added, halt
1165                          * is used for CPU IDLE.
1166                          * In such case C2/C3 is meaningless. So the max_cstate
1167                          * is set to one.
1168                          */
1169                         max_cstate = 1;
1170                 }
1171                 dmi_check_system(processor_power_dmi_table);
1172                 max_cstate = acpi_processor_cstate_check(max_cstate);
1173                 if (max_cstate < ACPI_C_STATES_MAX)
1174                         printk(KERN_NOTICE
1175                                "ACPI: processor limited to max C-state %d\n",
1176                                max_cstate);
1177                 first_run++;
1178         }
1179
1180         if (!pr)
1181                 return -EINVAL;
1182
1183         if (acpi_gbl_FADT.cst_control && !nocst) {
1184                 status =
1185                     acpi_os_write_port(acpi_gbl_FADT.smi_command, acpi_gbl_FADT.cst_control, 8);
1186                 if (ACPI_FAILURE(status)) {
1187                         ACPI_EXCEPTION((AE_INFO, status,
1188                                         "Notifying BIOS of _CST ability failed"));
1189                 }
1190         }
1191
1192         acpi_processor_get_power_info(pr);
1193         pr->flags.power_setup_done = 1;
1194
1195         /*
1196          * Install the idle handler if processor power management is supported.
1197          * Note that we use previously set idle handler will be used on
1198          * platforms that only support C1.
1199          */
1200         if (pr->flags.power) {
1201                 acpi_processor_setup_cpuidle(pr);
1202                 if (cpuidle_register_device(&pr->power.dev))
1203                         return -EIO;
1204
1205                 printk(KERN_INFO PREFIX "CPU%d (power states:", pr->id);
1206                 for (i = 1; i <= pr->power.count; i++)
1207                         if (pr->power.states[i].valid)
1208                                 printk(" C%d[C%d]", i,
1209                                        pr->power.states[i].type);
1210                 printk(")\n");
1211         }
1212
1213         /* 'power' [R] */
1214         entry = proc_create_data(ACPI_PROCESSOR_FILE_POWER,
1215                                  S_IRUGO, acpi_device_dir(device),
1216                                  &acpi_processor_power_fops,
1217                                  acpi_driver_data(device));
1218         if (!entry)
1219                 return -EIO;
1220         return 0;
1221 }
1222
1223 int acpi_processor_power_exit(struct acpi_processor *pr,
1224                               struct acpi_device *device)
1225 {
1226         if (boot_option_idle_override)
1227                 return 0;
1228
1229         cpuidle_unregister_device(&pr->power.dev);
1230         pr->flags.power_setup_done = 0;
1231
1232         if (acpi_device_dir(device))
1233                 remove_proc_entry(ACPI_PROCESSOR_FILE_POWER,
1234                                   acpi_device_dir(device));
1235
1236         return 0;
1237 }