drivers/acpi/processor_idle.c

   1 /*
   2  * processor_idle - idle state submodule to the ACPI processor driver
   3  *
   4  *  Copyright (C) 2001, 2002 Andy Grover <andrew.grover@intel.com>
   5  *  Copyright (C) 2001, 2002 Paul Diefenbaugh <paul.s.diefenbaugh@intel.com>
   6  *  Copyright (C) 2004, 2005 Dominik Brodowski <linux@brodo.de>
   7  *  Copyright (C) 2004  Anil S Keshavamurthy <anil.s.keshavamurthy@intel.com>
   8  *                      - Added processor hotplug support
   9  *  Copyright (C) 2005  Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>
  10  *                      - Added support for C3 on SMP
  11  *
  12  * ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
  13  *
  14  *  This program is free software; you can redistribute it and/or modify
  15  *  it under the terms of the GNU General Public License as published by
  16  *  the Free Software Foundation; either version 2 of the License, or (at
  17  *  your option) any later version.
  18  *
  19  *  This program is distributed in the hope that it will be useful, but
  20  *  WITHOUT ANY WARRANTY; without even the implied warranty of
  21  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  22  *  General Public License for more details.
  23  *
  24  *  You should have received a copy of the GNU General Public License along
  25  *  with this program; if not, write to the Free Software Foundation, Inc.,
  26  *  59 Temple Place, Suite 330, Boston, MA 02111-1307 USA.
  27  *
  28  * ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
  29  */
  30
  31 #include <linux/kernel.h>
  32 #include <linux/module.h>
  33 #include <linux/init.h>
  34 #include <linux/cpufreq.h>
  35 #include <linux/slab.h>
  36 #include <linux/acpi.h>
  37 #include <linux/dmi.h>
  38 #include <linux/moduleparam.h>
  39 #include <linux/sched.h>        /* need_resched() */
  40 #include <linux/pm_qos_params.h>
  41 #include <linux/clockchips.h>
  42 #include <linux/cpuidle.h>
  43 #include <linux/irqflags.h>
  44
  45 /*
  46  * Include the apic definitions for x86 to have the APIC timer related defines
  47  * available also for UP (on SMP it gets magically included via linux/smp.h).
  48  * asm/acpi.h is not an option, as it would require more include magic. Also
  49  * creating an empty asm-ia64/apic.h would just trade pest vs. cholera.
  50  */
  51 #ifdef CONFIG_X86
  52 #include <asm/apic.h>
  53 #endif
  54
  55 #include <asm/io.h>
  56 #include <asm/uaccess.h>
  57
  58 #include <acpi/acpi_bus.h>
  59 #include <acpi/processor.h>
  60 #include <asm/processor.h>
  61
  62 #define PREFIX "ACPI: "
  63
  64 #define ACPI_PROCESSOR_CLASS            "processor"
  65 #define _COMPONENT              ACPI_PROCESSOR_COMPONENT
  66 ACPI_MODULE_NAME("processor_idle");
  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 static int bm_check_disable __read_mostly;
  77 module_param(bm_check_disable, uint, 0000);
  78
  79 static unsigned int latency_factor __read_mostly = 2;
  80 module_param(latency_factor, uint, 0644);
  81
  82 static int disabled_by_idle_boot_param(void)
  83 {
  84         return boot_option_idle_override == IDLE_POLL ||
  85                 boot_option_idle_override == IDLE_FORCE_MWAIT ||
  86                 boot_option_idle_override == IDLE_HALT;
  87 }
  88
  89 /*
  90  * IBM ThinkPad R40e crashes mysteriously when going into C2 or C3.
  91  * For now disable this. Probably a bug somewhere else.
  92  *
  93  * To skip this limit, boot/load with a large max_cstate limit.
  94  */
  95 static int set_max_cstate(const struct dmi_system_id *id)
  96 {
  97         if (max_cstate > ACPI_PROCESSOR_MAX_POWER)
  98                 return 0;
  99
 100         printk(KERN_NOTICE PREFIX "%s detected - limiting to C%ld max_cstate."
 101                " Override with \"processor.max_cstate=%d\"\n", id->ident,
 102                (long)id->driver_data, ACPI_PROCESSOR_MAX_POWER + 1);
 103
 104         max_cstate = (long)id->driver_data;
 105
 106         return 0;
 107 }
 108
 109 /* Actually this shouldn't be __cpuinitdata, would be better to fix the
 110    callers to only run once -AK */
 111 static struct dmi_system_id __cpuinitdata processor_power_dmi_table[] = {
 112         { set_max_cstate, "Clevo 5600D", {
 113           DMI_MATCH(DMI_BIOS_VENDOR,"Phoenix Technologies LTD"),
 114           DMI_MATCH(DMI_BIOS_VERSION,"SHE845M0.86C.0013.D.0302131307")},
 115          (void *)2},
 116         { set_max_cstate, "Pavilion zv5000", {
 117           DMI_MATCH(DMI_SYS_VENDOR, "Hewlett-Packard"),
 118           DMI_MATCH(DMI_PRODUCT_NAME,"Pavilion zv5000 (DS502A#ABA)")},
 119          (void *)1},
 120         { set_max_cstate, "Asus L8400B", {
 121           DMI_MATCH(DMI_SYS_VENDOR, "ASUSTeK Computer Inc."),
 122           DMI_MATCH(DMI_PRODUCT_NAME,"L8400B series Notebook PC")},
 123          (void *)1},
 124         {},
 125 };
 126
 127
 128 /*
 129  * Callers should disable interrupts before the call and enable
 130  * interrupts after return.
 131  */
 132 static void acpi_safe_halt(void)
 133 {
 134         current_thread_info()->status &= ~TS_POLLING;
 135         /*
 136          * TS_POLLING-cleared state must be visible before we
 137          * test NEED_RESCHED:
 138          */
 139         smp_mb();
 140         if (!need_resched()) {
 141                 safe_halt();
 142                 local_irq_disable();
 143         }
 144         current_thread_info()->status |= TS_POLLING;
 145 }
 146
 147 #ifdef ARCH_APICTIMER_STOPS_ON_C3
 148
 149 /*
 150  * Some BIOS implementations switch to C3 in the published C2 state.
 151  * This seems to be a common problem on AMD boxen, but other vendors
 152  * are affected too. We pick the most conservative approach: we assume
 153  * that the local APIC stops in both C2 and C3.
 154  */
 155 static void lapic_timer_check_state(int state, struct acpi_processor *pr,
 156                                    struct acpi_processor_cx *cx)
 157 {
 158         struct acpi_processor_power *pwr = &pr->power;
 159         u8 type = local_apic_timer_c2_ok ? ACPI_STATE_C3 : ACPI_STATE_C2;
 160
 161         if (cpu_has(&cpu_data(pr->id), X86_FEATURE_ARAT))
 162                 return;
 163
 164         if (amd_e400_c1e_detected)
 165                 type = ACPI_STATE_C1;
 166
 167         /*
 168          * Check, if one of the previous states already marked the lapic
 169          * unstable
 170          */
 171         if (pwr->timer_broadcast_on_state < state)
 172                 return;
 173
 174         if (cx->type >= type)
 175                 pr->power.timer_broadcast_on_state = state;
 176 }
 177
 178 static void __lapic_timer_propagate_broadcast(void *arg)
 179 {
 180         struct acpi_processor *pr = (struct acpi_processor *) arg;
 181         unsigned long reason;
 182
 183         reason = pr->power.timer_broadcast_on_state < INT_MAX ?
 184                 CLOCK_EVT_NOTIFY_BROADCAST_ON : CLOCK_EVT_NOTIFY_BROADCAST_OFF;
 185
 186         clockevents_notify(reason, &pr->id);
 187 }
 188
 189 static void lapic_timer_propagate_broadcast(struct acpi_processor *pr)
 190 {
 191         smp_call_function_single(pr->id, __lapic_timer_propagate_broadcast,
 192                                  (void *)pr, 1);
 193 }
 194
 195 /* Power(C) State timer broadcast control */
 196 static void lapic_timer_state_broadcast(struct acpi_processor *pr,
 197                                        struct acpi_processor_cx *cx,
 198                                        int broadcast)
 199 {
 200         int state = cx - pr->power.states;
 201
 202         if (state >= pr->power.timer_broadcast_on_state) {
 203                 unsigned long reason;
 204
 205                 reason = broadcast ?  CLOCK_EVT_NOTIFY_BROADCAST_ENTER :
 206                         CLOCK_EVT_NOTIFY_BROADCAST_EXIT;
 207                 clockevents_notify(reason, &pr->id);
 208         }
 209 }
 210
 211 #else
 212
 213 static void lapic_timer_check_state(int state, struct acpi_processor *pr,
 214                                    struct acpi_processor_cx *cstate) { }
 215 static void lapic_timer_propagate_broadcast(struct acpi_processor *pr) { }
 216 static void lapic_timer_state_broadcast(struct acpi_processor *pr,
 217                                        struct acpi_processor_cx *cx,
 218                                        int broadcast)
 219 {
 220 }
 221
 222 #endif
 223
 224 /*
 225  * Suspend / resume control
 226  */
 227 static int acpi_idle_suspend;
 228 static u32 saved_bm_rld;
 229
 230 static void acpi_idle_bm_rld_save(void)
 231 {
 232         acpi_read_bit_register(ACPI_BITREG_BUS_MASTER_RLD, &saved_bm_rld);
 233 }
 234 static void acpi_idle_bm_rld_restore(void)
 235 {
 236         u32 resumed_bm_rld;
 237
 238         acpi_read_bit_register(ACPI_BITREG_BUS_MASTER_RLD, &resumed_bm_rld);
 239
 240         if (resumed_bm_rld != saved_bm_rld)
 241                 acpi_write_bit_register(ACPI_BITREG_BUS_MASTER_RLD, saved_bm_rld);
 242 }
 243
 244 int acpi_processor_suspend(struct acpi_device * device, pm_message_t state)
 245 {
 246         if (acpi_idle_suspend == 1)
 247                 return 0;
 248
 249         acpi_idle_bm_rld_save();
 250         acpi_idle_suspend = 1;
 251         return 0;
 252 }
 253
 254 int acpi_processor_resume(struct acpi_device * device)
 255 {
 256         if (acpi_idle_suspend == 0)
 257                 return 0;
 258
 259         acpi_idle_bm_rld_restore();
 260         acpi_idle_suspend = 0;
 261         return 0;
 262 }
 263
 264 #if defined(CONFIG_X86)
 265 static void tsc_check_state(int state)
 266 {
 267         switch (boot_cpu_data.x86_vendor) {
 268         case X86_VENDOR_AMD:
 269         case X86_VENDOR_INTEL:
 270                 /*
 271                  * AMD Fam10h TSC will tick in all
 272                  * C/P/S0/S1 states when this bit is set.
 273                  */
 274                 if (boot_cpu_has(X86_FEATURE_NONSTOP_TSC))
 275                         return;
 276
 277                 /*FALL THROUGH*/
 278         default:
 279                 /* TSC could halt in idle, so notify users */
 280                 if (state > ACPI_STATE_C1)
 281                         mark_tsc_unstable("TSC halts in idle");
 282         }
 283 }
 284 #else
 285 static void tsc_check_state(int state) { return; }
 286 #endif
 287
 288 static int acpi_processor_get_power_info_fadt(struct acpi_processor *pr)
 289 {
 290
 291         if (!pr)
 292                 return -EINVAL;
 293
 294         if (!pr->pblk)
 295                 return -ENODEV;
 296
 297         /* if info is obtained from pblk/fadt, type equals state */
 298         pr->power.states[ACPI_STATE_C2].type = ACPI_STATE_C2;
 299         pr->power.states[ACPI_STATE_C3].type = ACPI_STATE_C3;
 300
 301 #ifndef CONFIG_HOTPLUG_CPU
 302         /*
 303          * Check for P_LVL2_UP flag before entering C2 and above on
 304          * an SMP system.
 305          */
 306         if ((num_online_cpus() > 1) &&
 307             !(acpi_gbl_FADT.flags & ACPI_FADT_C2_MP_SUPPORTED))
 308                 return -ENODEV;
 309 #endif
 310
 311         /* determine C2 and C3 address from pblk */
 312         pr->power.states[ACPI_STATE_C2].address = pr->pblk + 4;
 313         pr->power.states[ACPI_STATE_C3].address = pr->pblk + 5;
 314
 315         /* determine latencies from FADT */
 316         pr->power.states[ACPI_STATE_C2].latency = acpi_gbl_FADT.C2latency;
 317         pr->power.states[ACPI_STATE_C3].latency = acpi_gbl_FADT.C3latency;
 318
 319         /*
 320          * FADT specified C2 latency must be less than or equal to
 321          * 100 microseconds.
 322          */
 323         if (acpi_gbl_FADT.C2latency > ACPI_PROCESSOR_MAX_C2_LATENCY) {
 324                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 325                         "C2 latency too large [%d]\n", acpi_gbl_FADT.C2latency));
 326                 /* invalidate C2 */
 327                 pr->power.states[ACPI_STATE_C2].address = 0;
 328         }
 329
 330         /*
 331          * FADT supplied C3 latency must be less than or equal to
 332          * 1000 microseconds.
 333          */
 334         if (acpi_gbl_FADT.C3latency > ACPI_PROCESSOR_MAX_C3_LATENCY) {
 335                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 336                         "C3 latency too large [%d]\n", acpi_gbl_FADT.C3latency));
 337                 /* invalidate C3 */
 338                 pr->power.states[ACPI_STATE_C3].address = 0;
 339         }
 340
 341         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 342                           "lvl2[0x%08x] lvl3[0x%08x]\n",
 343                           pr->power.states[ACPI_STATE_C2].address,
 344                           pr->power.states[ACPI_STATE_C3].address));
 345
 346         return 0;
 347 }
 348
 349 static int acpi_processor_get_power_info_default(struct acpi_processor *pr)
 350 {
 351         if (!pr->power.states[ACPI_STATE_C1].valid) {
 352                 /* set the first C-State to C1 */
 353                 /* all processors need to support C1 */
 354                 pr->power.states[ACPI_STATE_C1].type = ACPI_STATE_C1;
 355                 pr->power.states[ACPI_STATE_C1].valid = 1;
 356                 pr->power.states[ACPI_STATE_C1].entry_method = ACPI_CSTATE_HALT;
 357         }
 358         /* the C0 state only exists as a filler in our array */
 359         pr->power.states[ACPI_STATE_C0].valid = 1;
 360         return 0;
 361 }
 362
 363 static int acpi_processor_get_power_info_cst(struct acpi_processor *pr)
 364 {
 365         acpi_status status = 0;
 366         u64 count;
 367         int current_count;
 368         int i;
 369         struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
 370         union acpi_object *cst;
 371
 372
 373         if (nocst)
 374                 return -ENODEV;
 375
 376         current_count = 0;
 377
 378         status = acpi_evaluate_object(pr->handle, "_CST", NULL, &buffer);
 379         if (ACPI_FAILURE(status)) {
 380                 ACPI_DEBUG_PRINT((ACPI_DB_INFO, "No _CST, giving up\n"));
 381                 return -ENODEV;
 382         }
 383
 384         cst = buffer.pointer;
 385
 386         /* There must be at least 2 elements */
 387         if (!cst || (cst->type != ACPI_TYPE_PACKAGE) || cst->package.count < 2) {
 388                 printk(KERN_ERR PREFIX "not enough elements in _CST\n");
 389                 status = -EFAULT;
 390                 goto end;
 391         }
 392
 393         count = cst->package.elements[0].integer.value;
 394
 395         /* Validate number of power states. */
 396         if (count < 1 || count != cst->package.count - 1) {
 397                 printk(KERN_ERR PREFIX "count given by _CST is not valid\n");
 398                 status = -EFAULT;
 399                 goto end;
 400         }
 401
 402         /* Tell driver that at least _CST is supported. */
 403         pr->flags.has_cst = 1;
 404
 405         for (i = 1; i <= count; i++) {
 406                 union acpi_object *element;
 407                 union acpi_object *obj;
 408                 struct acpi_power_register *reg;
 409                 struct acpi_processor_cx cx;
 410
 411                 memset(&cx, 0, sizeof(cx));
 412
 413                 element = &(cst->package.elements[i]);
 414                 if (element->type != ACPI_TYPE_PACKAGE)
 415                         continue;
 416
 417                 if (element->package.count != 4)
 418                         continue;
 419
 420                 obj = &(element->package.elements[0]);
 421
 422                 if (obj->type != ACPI_TYPE_BUFFER)
 423                         continue;
 424
 425                 reg = (struct acpi_power_register *)obj->buffer.pointer;
 426
 427                 if (reg->space_id != ACPI_ADR_SPACE_SYSTEM_IO &&
 428                     (reg->space_id != ACPI_ADR_SPACE_FIXED_HARDWARE))
 429                         continue;
 430
 431                 /* There should be an easy way to extract an integer... */
 432                 obj = &(element->package.elements[1]);
 433                 if (obj->type != ACPI_TYPE_INTEGER)
 434                         continue;
 435
 436                 cx.type = obj->integer.value;
 437                 /*
 438                  * Some buggy BIOSes won't list C1 in _CST -
 439                  * Let acpi_processor_get_power_info_default() handle them later
 440                  */
 441                 if (i == 1 && cx.type != ACPI_STATE_C1)
 442                         current_count++;
 443
 444                 cx.address = reg->address;
 445                 cx.index = current_count + 1;
 446
 447                 cx.entry_method = ACPI_CSTATE_SYSTEMIO;
 448                 if (reg->space_id == ACPI_ADR_SPACE_FIXED_HARDWARE) {
 449                         if (acpi_processor_ffh_cstate_probe
 450                                         (pr->id, &cx, reg) == 0) {
 451                                 cx.entry_method = ACPI_CSTATE_FFH;
 452                         } else if (cx.type == ACPI_STATE_C1) {
 453                                 /*
 454                                  * C1 is a special case where FIXED_HARDWARE
 455                                  * can be handled in non-MWAIT way as well.
 456                                  * In that case, save this _CST entry info.
 457                                  * Otherwise, ignore this info and continue.
 458                                  */
 459                                 cx.entry_method = ACPI_CSTATE_HALT;
 460                                 snprintf(cx.desc, ACPI_CX_DESC_LEN, "ACPI HLT");
 461                         } else {
 462                                 continue;
 463                         }
 464                         if (cx.type == ACPI_STATE_C1 &&
 465                             (boot_option_idle_override == IDLE_NOMWAIT)) {
 466                                 /*
 467                                  * In most cases the C1 space_id obtained from
 468                                  * _CST object is FIXED_HARDWARE access mode.
 469                                  * But when the option of idle=halt is added,
 470                                  * the entry_method type should be changed from
 471                                  * CSTATE_FFH to CSTATE_HALT.
 472                                  * When the option of idle=nomwait is added,
 473                                  * the C1 entry_method type should be
 474                                  * CSTATE_HALT.
 475                                  */
 476                                 cx.entry_method = ACPI_CSTATE_HALT;
 477                                 snprintf(cx.desc, ACPI_CX_DESC_LEN, "ACPI HLT");
 478                         }
 479                 } else {
 480                         snprintf(cx.desc, ACPI_CX_DESC_LEN, "ACPI IOPORT 0x%x",
 481                                  cx.address);
 482                 }
 483
 484                 if (cx.type == ACPI_STATE_C1) {
 485                         cx.valid = 1;
 486                 }
 487
 488                 obj = &(element->package.elements[2]);
 489                 if (obj->type != ACPI_TYPE_INTEGER)
 490                         continue;
 491
 492                 cx.latency = obj->integer.value;
 493
 494                 obj = &(element->package.elements[3]);
 495                 if (obj->type != ACPI_TYPE_INTEGER)
 496                         continue;
 497
 498                 cx.power = obj->integer.value;
 499
 500                 current_count++;
 501                 memcpy(&(pr->power.states[current_count]), &cx, sizeof(cx));
 502
 503                 /*
 504                  * We support total ACPI_PROCESSOR_MAX_POWER - 1
 505                  * (From 1 through ACPI_PROCESSOR_MAX_POWER - 1)
 506                  */
 507                 if (current_count >= (ACPI_PROCESSOR_MAX_POWER - 1)) {
 508                         printk(KERN_WARNING
 509                                "Limiting number of power states to max (%d)\n",
 510                                ACPI_PROCESSOR_MAX_POWER);
 511                         printk(KERN_WARNING
 512                                "Please increase ACPI_PROCESSOR_MAX_POWER if needed.\n");
 513                         break;
 514                 }
 515         }
 516
 517         ACPI_DEBUG_PRINT((ACPI_DB_INFO, "Found %d power states\n",
 518                           current_count));
 519
 520         /* Validate number of power states discovered */
 521         if (current_count < 2)
 522                 status = -EFAULT;
 523
 524       end:
 525         kfree(buffer.pointer);
 526
 527         return status;
 528 }
 529
 530 static void acpi_processor_power_verify_c3(struct acpi_processor *pr,
 531                                            struct acpi_processor_cx *cx)
 532 {
 533         static int bm_check_flag = -1;
 534         static int bm_control_flag = -1;
 535
 536
 537         if (!cx->address)
 538                 return;
 539
 540         /*
 541          * PIIX4 Erratum #18: We don't support C3 when Type-F (fast)
 542          * DMA transfers are used by any ISA device to avoid livelock.
 543          * Note that we could disable Type-F DMA (as recommended by
 544          * the erratum), but this is known to disrupt certain ISA
 545          * devices thus we take the conservative approach.
 546          */
 547         else if (errata.piix4.fdma) {
 548                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 549                                   "C3 not supported on PIIX4 with Type-F DMA\n"));
 550                 return;
 551         }
 552
 553         /* All the logic here assumes flags.bm_check is same across all CPUs */
 554         if (bm_check_flag == -1) {
 555                 /* Determine whether bm_check is needed based on CPU  */
 556                 acpi_processor_power_init_bm_check(&(pr->flags), pr->id);
 557                 bm_check_flag = pr->flags.bm_check;
 558                 bm_control_flag = pr->flags.bm_control;
 559         } else {
 560                 pr->flags.bm_check = bm_check_flag;
 561                 pr->flags.bm_control = bm_control_flag;
 562         }
 563
 564         if (pr->flags.bm_check) {
 565                 if (!pr->flags.bm_control) {
 566                         if (pr->flags.has_cst != 1) {
 567                                 /* bus mastering control is necessary */
 568                                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 569                                         "C3 support requires BM control\n"));
 570                                 return;
 571                         } else {
 572                                 /* Here we enter C3 without bus mastering */
 573                                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 574                                         "C3 support without BM control\n"));
 575                         }
 576                 }
 577         } else {
 578                 /*
 579                  * WBINVD should be set in fadt, for C3 state to be
 580                  * supported on when bm_check is not required.
 581                  */
 582                 if (!(acpi_gbl_FADT.flags & ACPI_FADT_WBINVD)) {
 583                         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 584                                           "Cache invalidation should work properly"
 585                                           " for C3 to be enabled on SMP systems\n"));
 586                         return;
 587                 }
 588         }
 589
 590         /*
 591          * Otherwise we've met all of our C3 requirements.
 592          * Normalize the C3 latency to expidite policy.  Enable
 593          * checking of bus mastering status (bm_check) so we can
 594          * use this in our C3 policy
 595          */
 596         cx->valid = 1;
 597
 598         cx->latency_ticks = cx->latency;
 599         /*
 600          * On older chipsets, BM_RLD needs to be set
 601          * in order for Bus Master activity to wake the
 602          * system from C3.  Newer chipsets handle DMA
 603          * during C3 automatically and BM_RLD is a NOP.
 604          * In either case, the proper way to
 605          * handle BM_RLD is to set it and leave it set.
 606          */
 607         acpi_write_bit_register(ACPI_BITREG_BUS_MASTER_RLD, 1);
 608
 609         return;
 610 }
 611
 612 static int acpi_processor_power_verify(struct acpi_processor *pr)
 613 {
 614         unsigned int i;
 615         unsigned int working = 0;
 616
 617         pr->power.timer_broadcast_on_state = INT_MAX;
 618
 619         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER && i <= max_cstate; i++) {
 620                 struct acpi_processor_cx *cx = &pr->power.states[i];
 621
 622                 switch (cx->type) {
 623                 case ACPI_STATE_C1:
 624                         cx->valid = 1;
 625                         break;
 626
 627                 case ACPI_STATE_C2:
 628                         if (!cx->address)
 629                                 break;
 630                         cx->valid = 1;
 631                         cx->latency_ticks = cx->latency; /* Normalize latency */
 632                         break;
 633
 634                 case ACPI_STATE_C3:
 635                         acpi_processor_power_verify_c3(pr, cx);
 636                         break;
 637                 }
 638                 if (!cx->valid)
 639                         continue;
 640
 641                 lapic_timer_check_state(i, pr, cx);
 642                 tsc_check_state(cx->type);
 643                 working++;
 644         }
 645
 646         lapic_timer_propagate_broadcast(pr);
 647
 648         return (working);
 649 }
 650
 651 static int acpi_processor_get_power_info(struct acpi_processor *pr)
 652 {
 653         unsigned int i;
 654         int result;
 655
 656
 657         /* NOTE: the idle thread may not be running while calling
 658          * this function */
 659
 660         /* Zero initialize all the C-states info. */
 661         memset(pr->power.states, 0, sizeof(pr->power.states));
 662
 663         result = acpi_processor_get_power_info_cst(pr);
 664         if (result == -ENODEV)
 665                 result = acpi_processor_get_power_info_fadt(pr);
 666
 667         if (result)
 668                 return result;
 669
 670         acpi_processor_get_power_info_default(pr);
 671
 672         pr->power.count = acpi_processor_power_verify(pr);
 673
 674         /*
 675          * if one state of type C2 or C3 is available, mark this
 676          * CPU as being "idle manageable"
 677          */
 678         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
 679                 if (pr->power.states[i].valid) {
 680                         pr->power.count = i;
 681                         if (pr->power.states[i].type >= ACPI_STATE_C2)
 682                                 pr->flags.power = 1;
 683                 }
 684         }
 685
 686         return 0;
 687 }
 688
 689 /**
 690  * acpi_idle_bm_check - checks if bus master activity was detected
 691  */
 692 static int acpi_idle_bm_check(void)
 693 {
 694         u32 bm_status = 0;
 695
 696         if (bm_check_disable)
 697                 return 0;
 698
 699         acpi_read_bit_register(ACPI_BITREG_BUS_MASTER_STATUS, &bm_status);
 700         if (bm_status)
 701                 acpi_write_bit_register(ACPI_BITREG_BUS_MASTER_STATUS, 1);
 702         /*
 703          * PIIX4 Erratum #18: Note that BM_STS doesn't always reflect
 704          * the true state of bus mastering activity; forcing us to
 705          * manually check the BMIDEA bit of each IDE channel.
 706          */
 707         else if (errata.piix4.bmisx) {
 708                 if ((inb_p(errata.piix4.bmisx + 0x02) & 0x01)
 709                     || (inb_p(errata.piix4.bmisx + 0x0A) & 0x01))
 710                         bm_status = 1;
 711         }
 712         return bm_status;
 713 }
 714
 715 /**
 716  * acpi_idle_do_entry - a helper function that does C2 and C3 type entry
 717  * @cx: cstate data
 718  *
 719  * Caller disables interrupt before call and enables interrupt after return.
 720  */
 721 static inline void acpi_idle_do_entry(struct acpi_processor_cx *cx)
 722 {
 723         /* Don't trace irqs off for idle */
 724         stop_critical_timings();
 725         if (cx->entry_method == ACPI_CSTATE_FFH) {
 726                 /* Call into architectural FFH based C-state */
 727                 acpi_processor_ffh_cstate_enter(cx);
 728         } else if (cx->entry_method == ACPI_CSTATE_HALT) {
 729                 acpi_safe_halt();
 730         } else {
 731                 /* IO port based C-state */
 732                 inb(cx->address);
 733                 /* Dummy wait op - must do something useless after P_LVL2 read
 734                    because chipsets cannot guarantee that STPCLK# signal
 735                    gets asserted in time to freeze execution properly. */
 736                 inl(acpi_gbl_FADT.xpm_timer_block.address);
 737         }
 738         start_critical_timings();
 739 }
 740
 741 /**
 742  * acpi_idle_enter_c1 - enters an ACPI C1 state-type
 743  * @dev: the target CPU
 744  * @index: index of target state
 745  *
 746  * This is equivalent to the HALT instruction.
 747  */
 748 static int acpi_idle_enter_c1(struct cpuidle_device *dev,
 749                                 int index)
 750 {
 751         ktime_t  kt1, kt2;
 752         s64 idle_time;
 753         struct acpi_processor *pr;
 754         struct cpuidle_state *state = &dev->states[index];
 755         struct acpi_processor_cx *cx = cpuidle_get_statedata(state);
 756
 757         pr = __this_cpu_read(processors);
 758         dev->last_residency = 0;
 759
 760         if (unlikely(!pr))
 761                 return -EINVAL;
 762
 763         local_irq_disable();
 764
 765         /* Do not access any ACPI IO ports in suspend path */
 766         if (acpi_idle_suspend) {
 767                 local_irq_enable();
 768                 cpu_relax();
 769                 return -EINVAL;
 770         }
 771
 772         lapic_timer_state_broadcast(pr, cx, 1);
 773         kt1 = ktime_get_real();
 774         acpi_idle_do_entry(cx);
 775         kt2 = ktime_get_real();
 776         idle_time =  ktime_to_us(ktime_sub(kt2, kt1));
 777
 778         /* Update device last_residency*/
 779         dev->last_residency = (int)idle_time;
 780
 781         local_irq_enable();
 782         cx->usage++;
 783         lapic_timer_state_broadcast(pr, cx, 0);
 784
 785         return index;
 786 }
 787
 788 /**
 789  * acpi_idle_enter_simple - enters an ACPI state without BM handling
 790  * @dev: the target CPU
 791  * @index: the index of suggested state
 792  */
 793 static int acpi_idle_enter_simple(struct cpuidle_device *dev,
 794                                 int index)
 795 {
 796         struct acpi_processor *pr;
 797         struct cpuidle_state *state = &dev->states[index];
 798         struct acpi_processor_cx *cx = cpuidle_get_statedata(state);
 799         ktime_t  kt1, kt2;
 800         s64 idle_time_ns;
 801         s64 idle_time;
 802
 803         pr = __this_cpu_read(processors);
 804         dev->last_residency = 0;
 805
 806         if (unlikely(!pr))
 807                 return -EINVAL;
 808
 809         local_irq_disable();
 810
 811         if (acpi_idle_suspend) {
 812                 local_irq_enable();
 813                 cpu_relax();
 814                 return -EINVAL;
 815         }
 816
 817
 818         if (cx->entry_method != ACPI_CSTATE_FFH) {
 819                 current_thread_info()->status &= ~TS_POLLING;
 820                 /*
 821                  * TS_POLLING-cleared state must be visible before we test
 822                  * NEED_RESCHED:
 823                  */
 824                 smp_mb();
 825
 826                 if (unlikely(need_resched())) {
 827                         current_thread_info()->status |= TS_POLLING;
 828                         local_irq_enable();
 829                         return -EINVAL;
 830                 }
 831         }
 832
 833         /*
 834          * Must be done before busmaster disable as we might need to
 835          * access HPET !
 836          */
 837         lapic_timer_state_broadcast(pr, cx, 1);
 838
 839         if (cx->type == ACPI_STATE_C3)
 840                 ACPI_FLUSH_CPU_CACHE();
 841
 842         kt1 = ktime_get_real();
 843         /* Tell the scheduler that we are going deep-idle: */
 844         sched_clock_idle_sleep_event();
 845         acpi_idle_do_entry(cx);
 846         kt2 = ktime_get_real();
 847         idle_time_ns = ktime_to_ns(ktime_sub(kt2, kt1));
 848         idle_time = idle_time_ns;
 849         do_div(idle_time, NSEC_PER_USEC);
 850
 851         /* Update device last_residency*/
 852         dev->last_residency = (int)idle_time;
 853
 854         /* Tell the scheduler how much we idled: */
 855         sched_clock_idle_wakeup_event(idle_time_ns);
 856
 857         local_irq_enable();
 858         if (cx->entry_method != ACPI_CSTATE_FFH)
 859                 current_thread_info()->status |= TS_POLLING;
 860
 861         cx->usage++;
 862
 863         lapic_timer_state_broadcast(pr, cx, 0);
 864         cx->time += idle_time;
 865         return index;
 866 }
 867
 868 static int c3_cpu_count;
 869 static DEFINE_SPINLOCK(c3_lock);
 870
 871 /**
 872  * acpi_idle_enter_bm - enters C3 with proper BM handling
 873  * @dev: the target CPU
 874  * @index: the index of suggested state
 875  *
 876  * If BM is detected, the deepest non-C3 idle state is entered instead.
 877  */
 878 static int acpi_idle_enter_bm(struct cpuidle_device *dev,
 879                                 int index)
 880 {
 881         struct acpi_processor *pr;
 882         struct cpuidle_state *state = &dev->states[index];
 883         struct acpi_processor_cx *cx = cpuidle_get_statedata(state);
 884         ktime_t  kt1, kt2;
 885         s64 idle_time_ns;
 886         s64 idle_time;
 887
 888
 889         pr = __this_cpu_read(processors);
 890         dev->last_residency = 0;
 891
 892         if (unlikely(!pr))
 893                 return -EINVAL;
 894
 895
 896         if (acpi_idle_suspend) {
 897                 cpu_relax();
 898                 return -EINVAL;
 899         }
 900
 901         if (!cx->bm_sts_skip && acpi_idle_bm_check()) {
 902                 if (dev->safe_state_index >= 0) {
 903                         return dev->states[dev->safe_state_index].enter(dev,
 904                                                 dev->safe_state_index);
 905                 } else {
 906                         local_irq_disable();
 907                         acpi_safe_halt();
 908                         local_irq_enable();
 909                         return -EINVAL;
 910                 }
 911         }
 912
 913         local_irq_disable();
 914
 915         if (cx->entry_method != ACPI_CSTATE_FFH) {
 916                 current_thread_info()->status &= ~TS_POLLING;
 917                 /*
 918                  * TS_POLLING-cleared state must be visible before we test
 919                  * NEED_RESCHED:
 920                  */
 921                 smp_mb();
 922
 923                 if (unlikely(need_resched())) {
 924                         current_thread_info()->status |= TS_POLLING;
 925                         local_irq_enable();
 926                         return -EINVAL;
 927                 }
 928         }
 929
 930         acpi_unlazy_tlb(smp_processor_id());
 931
 932         /* Tell the scheduler that we are going deep-idle: */
 933         sched_clock_idle_sleep_event();
 934         /*
 935          * Must be done before busmaster disable as we might need to
 936          * access HPET !
 937          */
 938         lapic_timer_state_broadcast(pr, cx, 1);
 939
 940         kt1 = ktime_get_real();
 941         /*
 942          * disable bus master
 943          * bm_check implies we need ARB_DIS
 944          * !bm_check implies we need cache flush
 945          * bm_control implies whether we can do ARB_DIS
 946          *
 947          * That leaves a case where bm_check is set and bm_control is
 948          * not set. In that case we cannot do much, we enter C3
 949          * without doing anything.
 950          */
 951         if (pr->flags.bm_check && pr->flags.bm_control) {
 952                 spin_lock(&c3_lock);
 953                 c3_cpu_count++;
 954                 /* Disable bus master arbitration when all CPUs are in C3 */
 955                 if (c3_cpu_count == num_online_cpus())
 956                         acpi_write_bit_register(ACPI_BITREG_ARB_DISABLE, 1);
 957                 spin_unlock(&c3_lock);
 958         } else if (!pr->flags.bm_check) {
 959                 ACPI_FLUSH_CPU_CACHE();
 960         }
 961
 962         acpi_idle_do_entry(cx);
 963
 964         /* Re-enable bus master arbitration */
 965         if (pr->flags.bm_check && pr->flags.bm_control) {
 966                 spin_lock(&c3_lock);
 967                 acpi_write_bit_register(ACPI_BITREG_ARB_DISABLE, 0);
 968                 c3_cpu_count--;
 969                 spin_unlock(&c3_lock);
 970         }
 971         kt2 = ktime_get_real();
 972         idle_time_ns = ktime_to_ns(ktime_sub(kt2, kt1));
 973         idle_time = idle_time_ns;
 974         do_div(idle_time, NSEC_PER_USEC);
 975
 976         /* Update device last_residency*/
 977         dev->last_residency = (int)idle_time;
 978
 979         /* Tell the scheduler how much we idled: */
 980         sched_clock_idle_wakeup_event(idle_time_ns);
 981
 982         local_irq_enable();
 983         if (cx->entry_method != ACPI_CSTATE_FFH)
 984                 current_thread_info()->status |= TS_POLLING;
 985
 986         cx->usage++;
 987
 988         lapic_timer_state_broadcast(pr, cx, 0);
 989         cx->time += idle_time;
 990         return index;
 991 }
 992
 993 struct cpuidle_driver acpi_idle_driver = {
 994         .name =         "acpi_idle",
 995         .owner =        THIS_MODULE,
 996 };
 997
 998 /**
 999  * acpi_processor_setup_cpuidle - prepares and configures CPUIDLE
1000  * @pr: the ACPI processor
1001  */
1002 static int acpi_processor_setup_cpuidle(struct acpi_processor *pr)
1003 {
1004         int i, count = CPUIDLE_DRIVER_STATE_START;
1005         struct acpi_processor_cx *cx;
1006         struct cpuidle_state *state;
1007         struct cpuidle_device *dev = &pr->power.dev;
1008
1009         if (!pr->flags.power_setup_done)
1010                 return -EINVAL;
1011
1012         if (pr->flags.power == 0) {
1013                 return -EINVAL;
1014         }
1015
1016         dev->cpu = pr->id;
1017         dev->safe_state_index = -1;
1018         for (i = 0; i < CPUIDLE_STATE_MAX; i++) {
1019                 dev->states[i].name[0] = '\0';
1020                 dev->states[i].desc[0] = '\0';
1021         }
1022
1023         if (max_cstate == 0)
1024                 max_cstate = 1;
1025
1026         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER && i <= max_cstate; i++) {
1027                 cx = &pr->power.states[i];
1028                 state = &dev->states[count];
1029
1030                 if (!cx->valid)
1031                         continue;
1032
1033 #ifdef CONFIG_HOTPLUG_CPU
1034                 if ((cx->type != ACPI_STATE_C1) && (num_online_cpus() > 1) &&
1035                     !pr->flags.has_cst &&
1036                     !(acpi_gbl_FADT.flags & ACPI_FADT_C2_MP_SUPPORTED))
1037                         continue;
1038 #endif
1039                 cpuidle_set_statedata(state, cx);
1040
1041                 snprintf(state->name, CPUIDLE_NAME_LEN, "C%d", i);
1042                 strncpy(state->desc, cx->desc, CPUIDLE_DESC_LEN);
1043                 state->exit_latency = cx->latency;
1044                 state->target_residency = cx->latency * latency_factor;
1045
1046                 state->flags = 0;
1047                 switch (cx->type) {
1048                         case ACPI_STATE_C1:
1049                         if (cx->entry_method == ACPI_CSTATE_FFH)
1050                                 state->flags |= CPUIDLE_FLAG_TIME_VALID;
1051
1052                         state->enter = acpi_idle_enter_c1;
1053                         dev->safe_state_index = count;
1054                         break;
1055
1056                         case ACPI_STATE_C2:
1057                         state->flags |= CPUIDLE_FLAG_TIME_VALID;
1058                         state->enter = acpi_idle_enter_simple;
1059                         dev->safe_state_index = count;
1060                         break;
1061
1062                         case ACPI_STATE_C3:
1063                         state->flags |= CPUIDLE_FLAG_TIME_VALID;
1064                         state->enter = pr->flags.bm_check ?
1065                                         acpi_idle_enter_bm :
1066                                         acpi_idle_enter_simple;
1067                         break;
1068                 }
1069
1070                 count++;
1071                 if (count == CPUIDLE_STATE_MAX)
1072                         break;
1073         }
1074
1075         dev->state_count = count;
1076
1077         if (!count)
1078                 return -EINVAL;
1079
1080         return 0;
1081 }
1082
1083 int acpi_processor_cst_has_changed(struct acpi_processor *pr)
1084 {
1085         int ret = 0;
1086
1087         if (disabled_by_idle_boot_param())
1088                 return 0;
1089
1090         if (!pr)
1091                 return -EINVAL;
1092
1093         if (nocst) {
1094                 return -ENODEV;
1095         }
1096
1097         if (!pr->flags.power_setup_done)
1098                 return -ENODEV;
1099
1100         cpuidle_pause_and_lock();
1101         cpuidle_disable_device(&pr->power.dev);
1102         acpi_processor_get_power_info(pr);
1103         if (pr->flags.power) {
1104                 acpi_processor_setup_cpuidle(pr);
1105                 ret = cpuidle_enable_device(&pr->power.dev);
1106         }
1107         cpuidle_resume_and_unlock();
1108
1109         return ret;
1110 }
1111
1112 int __cpuinit acpi_processor_power_init(struct acpi_processor *pr,
1113                               struct acpi_device *device)
1114 {
1115         acpi_status status = 0;
1116         static int first_run;
1117
1118         if (disabled_by_idle_boot_param())
1119                 return 0;
1120
1121         if (!first_run) {
1122                 dmi_check_system(processor_power_dmi_table);
1123                 max_cstate = acpi_processor_cstate_check(max_cstate);
1124                 if (max_cstate < ACPI_C_STATES_MAX)
1125                         printk(KERN_NOTICE
1126                                "ACPI: processor limited to max C-state %d\n",
1127                                max_cstate);
1128                 first_run++;
1129         }
1130
1131         if (!pr)
1132                 return -EINVAL;
1133
1134         if (acpi_gbl_FADT.cst_control && !nocst) {
1135                 status =
1136                     acpi_os_write_port(acpi_gbl_FADT.smi_command, acpi_gbl_FADT.cst_control, 8);
1137                 if (ACPI_FAILURE(status)) {
1138                         ACPI_EXCEPTION((AE_INFO, status,
1139                                         "Notifying BIOS of _CST ability failed"));
1140                 }
1141         }
1142
1143         acpi_processor_get_power_info(pr);
1144         pr->flags.power_setup_done = 1;
1145
1146         /*
1147          * Install the idle handler if processor power management is supported.
1148          * Note that we use previously set idle handler will be used on
1149          * platforms that only support C1.
1150          */
1151         if (pr->flags.power) {
1152                 acpi_processor_setup_cpuidle(pr);
1153                 if (cpuidle_register_device(&pr->power.dev))
1154                         return -EIO;
1155         }
1156         return 0;
1157 }
1158
1159 int acpi_processor_power_exit(struct acpi_processor *pr,
1160                               struct acpi_device *device)
1161 {
1162         if (disabled_by_idle_boot_param())
1163                 return 0;
1164
1165         cpuidle_unregister_device(&pr->power.dev);
1166         pr->flags.power_setup_done = 0;
1167
1168         return 0;
1169 }