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.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 DEFINE_PER_CPU(struct cpuidle_device *, acpi_cpuidle_device);
  83
  84 static int disabled_by_idle_boot_param(void)
  85 {
  86         return boot_option_idle_override == IDLE_POLL ||
  87                 boot_option_idle_override == IDLE_FORCE_MWAIT ||
  88                 boot_option_idle_override == IDLE_HALT;
  89 }
  90
  91 /*
  92  * IBM ThinkPad R40e crashes mysteriously when going into C2 or C3.
  93  * For now disable this. Probably a bug somewhere else.
  94  *
  95  * To skip this limit, boot/load with a large max_cstate limit.
  96  */
  97 static int set_max_cstate(const struct dmi_system_id *id)
  98 {
  99         if (max_cstate > ACPI_PROCESSOR_MAX_POWER)
 100                 return 0;
 101
 102         printk(KERN_NOTICE PREFIX "%s detected - limiting to C%ld max_cstate."
 103                " Override with \"processor.max_cstate=%d\"\n", id->ident,
 104                (long)id->driver_data, ACPI_PROCESSOR_MAX_POWER + 1);
 105
 106         max_cstate = (long)id->driver_data;
 107
 108         return 0;
 109 }
 110
 111 /* Actually this shouldn't be __cpuinitdata, would be better to fix the
 112    callers to only run once -AK */
 113 static struct dmi_system_id __cpuinitdata processor_power_dmi_table[] = {
 114         { set_max_cstate, "Clevo 5600D", {
 115           DMI_MATCH(DMI_BIOS_VENDOR,"Phoenix Technologies LTD"),
 116           DMI_MATCH(DMI_BIOS_VERSION,"SHE845M0.86C.0013.D.0302131307")},
 117          (void *)2},
 118         { set_max_cstate, "Pavilion zv5000", {
 119           DMI_MATCH(DMI_SYS_VENDOR, "Hewlett-Packard"),
 120           DMI_MATCH(DMI_PRODUCT_NAME,"Pavilion zv5000 (DS502A#ABA)")},
 121          (void *)1},
 122         { set_max_cstate, "Asus L8400B", {
 123           DMI_MATCH(DMI_SYS_VENDOR, "ASUSTeK Computer Inc."),
 124           DMI_MATCH(DMI_PRODUCT_NAME,"L8400B series Notebook PC")},
 125          (void *)1},
 126         {},
 127 };
 128
 129
 130 /*
 131  * Callers should disable interrupts before the call and enable
 132  * interrupts after return.
 133  */
 134 static void acpi_safe_halt(void)
 135 {
 136         current_thread_info()->status &= ~TS_POLLING;
 137         /*
 138          * TS_POLLING-cleared state must be visible before we
 139          * test NEED_RESCHED:
 140          */
 141         smp_mb();
 142         if (!need_resched()) {
 143                 safe_halt();
 144                 local_irq_disable();
 145         }
 146         current_thread_info()->status |= TS_POLLING;
 147 }
 148
 149 #ifdef ARCH_APICTIMER_STOPS_ON_C3
 150
 151 /*
 152  * Some BIOS implementations switch to C3 in the published C2 state.
 153  * This seems to be a common problem on AMD boxen, but other vendors
 154  * are affected too. We pick the most conservative approach: we assume
 155  * that the local APIC stops in both C2 and C3.
 156  */
 157 static void lapic_timer_check_state(int state, struct acpi_processor *pr,
 158                                    struct acpi_processor_cx *cx)
 159 {
 160         struct acpi_processor_power *pwr = &pr->power;
 161         u8 type = local_apic_timer_c2_ok ? ACPI_STATE_C3 : ACPI_STATE_C2;
 162
 163         if (cpu_has(&cpu_data(pr->id), X86_FEATURE_ARAT))
 164                 return;
 165
 166         if (amd_e400_c1e_detected)
 167                 type = ACPI_STATE_C1;
 168
 169         /*
 170          * Check, if one of the previous states already marked the lapic
 171          * unstable
 172          */
 173         if (pwr->timer_broadcast_on_state < state)
 174                 return;
 175
 176         if (cx->type >= type)
 177                 pr->power.timer_broadcast_on_state = state;
 178 }
 179
 180 static void __lapic_timer_propagate_broadcast(void *arg)
 181 {
 182         struct acpi_processor *pr = (struct acpi_processor *) arg;
 183         unsigned long reason;
 184
 185         reason = pr->power.timer_broadcast_on_state < INT_MAX ?
 186                 CLOCK_EVT_NOTIFY_BROADCAST_ON : CLOCK_EVT_NOTIFY_BROADCAST_OFF;
 187
 188         clockevents_notify(reason, &pr->id);
 189 }
 190
 191 static void lapic_timer_propagate_broadcast(struct acpi_processor *pr)
 192 {
 193         smp_call_function_single(pr->id, __lapic_timer_propagate_broadcast,
 194                                  (void *)pr, 1);
 195 }
 196
 197 /* Power(C) State timer broadcast control */
 198 static void lapic_timer_state_broadcast(struct acpi_processor *pr,
 199                                        struct acpi_processor_cx *cx,
 200                                        int broadcast)
 201 {
 202         int state = cx - pr->power.states;
 203
 204         if (state >= pr->power.timer_broadcast_on_state) {
 205                 unsigned long reason;
 206
 207                 reason = broadcast ?  CLOCK_EVT_NOTIFY_BROADCAST_ENTER :
 208                         CLOCK_EVT_NOTIFY_BROADCAST_EXIT;
 209                 clockevents_notify(reason, &pr->id);
 210         }
 211 }
 212
 213 #else
 214
 215 static void lapic_timer_check_state(int state, struct acpi_processor *pr,
 216                                    struct acpi_processor_cx *cstate) { }
 217 static void lapic_timer_propagate_broadcast(struct acpi_processor *pr) { }
 218 static void lapic_timer_state_broadcast(struct acpi_processor *pr,
 219                                        struct acpi_processor_cx *cx,
 220                                        int broadcast)
 221 {
 222 }
 223
 224 #endif
 225
 226 static u32 saved_bm_rld;
 227
 228 static void acpi_idle_bm_rld_save(void)
 229 {
 230         acpi_read_bit_register(ACPI_BITREG_BUS_MASTER_RLD, &saved_bm_rld);
 231 }
 232 static void acpi_idle_bm_rld_restore(void)
 233 {
 234         u32 resumed_bm_rld;
 235
 236         acpi_read_bit_register(ACPI_BITREG_BUS_MASTER_RLD, &resumed_bm_rld);
 237
 238         if (resumed_bm_rld != saved_bm_rld)
 239                 acpi_write_bit_register(ACPI_BITREG_BUS_MASTER_RLD, saved_bm_rld);
 240 }
 241
 242 int acpi_processor_suspend(struct device *dev)
 243 {
 244         acpi_idle_bm_rld_save();
 245         return 0;
 246 }
 247
 248 int acpi_processor_resume(struct device *dev)
 249 {
 250         acpi_idle_bm_rld_restore();
 251         return 0;
 252 }
 253
 254 #if defined(CONFIG_X86)
 255 static void tsc_check_state(int state)
 256 {
 257         switch (boot_cpu_data.x86_vendor) {
 258         case X86_VENDOR_AMD:
 259         case X86_VENDOR_INTEL:
 260                 /*
 261                  * AMD Fam10h TSC will tick in all
 262                  * C/P/S0/S1 states when this bit is set.
 263                  */
 264                 if (boot_cpu_has(X86_FEATURE_NONSTOP_TSC))
 265                         return;
 266
 267                 /*FALL THROUGH*/
 268         default:
 269                 /* TSC could halt in idle, so notify users */
 270                 if (state > ACPI_STATE_C1)
 271                         mark_tsc_unstable("TSC halts in idle");
 272         }
 273 }
 274 #else
 275 static void tsc_check_state(int state) { return; }
 276 #endif
 277
 278 static int acpi_processor_get_power_info_fadt(struct acpi_processor *pr)
 279 {
 280
 281         if (!pr)
 282                 return -EINVAL;
 283
 284         if (!pr->pblk)
 285                 return -ENODEV;
 286
 287         /* if info is obtained from pblk/fadt, type equals state */
 288         pr->power.states[ACPI_STATE_C2].type = ACPI_STATE_C2;
 289         pr->power.states[ACPI_STATE_C3].type = ACPI_STATE_C3;
 290
 291 #ifndef CONFIG_HOTPLUG_CPU
 292         /*
 293          * Check for P_LVL2_UP flag before entering C2 and above on
 294          * an SMP system.
 295          */
 296         if ((num_online_cpus() > 1) &&
 297             !(acpi_gbl_FADT.flags & ACPI_FADT_C2_MP_SUPPORTED))
 298                 return -ENODEV;
 299 #endif
 300
 301         /* determine C2 and C3 address from pblk */
 302         pr->power.states[ACPI_STATE_C2].address = pr->pblk + 4;
 303         pr->power.states[ACPI_STATE_C3].address = pr->pblk + 5;
 304
 305         /* determine latencies from FADT */
 306         pr->power.states[ACPI_STATE_C2].latency = acpi_gbl_FADT.c2_latency;
 307         pr->power.states[ACPI_STATE_C3].latency = acpi_gbl_FADT.c3_latency;
 308
 309         /*
 310          * FADT specified C2 latency must be less than or equal to
 311          * 100 microseconds.
 312          */
 313         if (acpi_gbl_FADT.c2_latency > ACPI_PROCESSOR_MAX_C2_LATENCY) {
 314                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 315                         "C2 latency too large [%d]\n", acpi_gbl_FADT.c2_latency));
 316                 /* invalidate C2 */
 317                 pr->power.states[ACPI_STATE_C2].address = 0;
 318         }
 319
 320         /*
 321          * FADT supplied C3 latency must be less than or equal to
 322          * 1000 microseconds.
 323          */
 324         if (acpi_gbl_FADT.c3_latency > ACPI_PROCESSOR_MAX_C3_LATENCY) {
 325                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 326                         "C3 latency too large [%d]\n", acpi_gbl_FADT.c3_latency));
 327                 /* invalidate C3 */
 328                 pr->power.states[ACPI_STATE_C3].address = 0;
 329         }
 330
 331         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 332                           "lvl2[0x%08x] lvl3[0x%08x]\n",
 333                           pr->power.states[ACPI_STATE_C2].address,
 334                           pr->power.states[ACPI_STATE_C3].address));
 335
 336         return 0;
 337 }
 338
 339 static int acpi_processor_get_power_info_default(struct acpi_processor *pr)
 340 {
 341         if (!pr->power.states[ACPI_STATE_C1].valid) {
 342                 /* set the first C-State to C1 */
 343                 /* all processors need to support C1 */
 344                 pr->power.states[ACPI_STATE_C1].type = ACPI_STATE_C1;
 345                 pr->power.states[ACPI_STATE_C1].valid = 1;
 346                 pr->power.states[ACPI_STATE_C1].entry_method = ACPI_CSTATE_HALT;
 347         }
 348         /* the C0 state only exists as a filler in our array */
 349         pr->power.states[ACPI_STATE_C0].valid = 1;
 350         return 0;
 351 }
 352
 353 static int acpi_processor_get_power_info_cst(struct acpi_processor *pr)
 354 {
 355         acpi_status status = 0;
 356         u64 count;
 357         int current_count;
 358         int i;
 359         struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
 360         union acpi_object *cst;
 361
 362
 363         if (nocst)
 364                 return -ENODEV;
 365
 366         current_count = 0;
 367
 368         status = acpi_evaluate_object(pr->handle, "_CST", NULL, &buffer);
 369         if (ACPI_FAILURE(status)) {
 370                 ACPI_DEBUG_PRINT((ACPI_DB_INFO, "No _CST, giving up\n"));
 371                 return -ENODEV;
 372         }
 373
 374         cst = buffer.pointer;
 375
 376         /* There must be at least 2 elements */
 377         if (!cst || (cst->type != ACPI_TYPE_PACKAGE) || cst->package.count < 2) {
 378                 printk(KERN_ERR PREFIX "not enough elements in _CST\n");
 379                 status = -EFAULT;
 380                 goto end;
 381         }
 382
 383         count = cst->package.elements[0].integer.value;
 384
 385         /* Validate number of power states. */
 386         if (count < 1 || count != cst->package.count - 1) {
 387                 printk(KERN_ERR PREFIX "count given by _CST is not valid\n");
 388                 status = -EFAULT;
 389                 goto end;
 390         }
 391
 392         /* Tell driver that at least _CST is supported. */
 393         pr->flags.has_cst = 1;
 394
 395         for (i = 1; i <= count; i++) {
 396                 union acpi_object *element;
 397                 union acpi_object *obj;
 398                 struct acpi_power_register *reg;
 399                 struct acpi_processor_cx cx;
 400
 401                 memset(&cx, 0, sizeof(cx));
 402
 403                 element = &(cst->package.elements[i]);
 404                 if (element->type != ACPI_TYPE_PACKAGE)
 405                         continue;
 406
 407                 if (element->package.count != 4)
 408                         continue;
 409
 410                 obj = &(element->package.elements[0]);
 411
 412                 if (obj->type != ACPI_TYPE_BUFFER)
 413                         continue;
 414
 415                 reg = (struct acpi_power_register *)obj->buffer.pointer;
 416
 417                 if (reg->space_id != ACPI_ADR_SPACE_SYSTEM_IO &&
 418                     (reg->space_id != ACPI_ADR_SPACE_FIXED_HARDWARE))
 419                         continue;
 420
 421                 /* There should be an easy way to extract an integer... */
 422                 obj = &(element->package.elements[1]);
 423                 if (obj->type != ACPI_TYPE_INTEGER)
 424                         continue;
 425
 426                 cx.type = obj->integer.value;
 427                 /*
 428                  * Some buggy BIOSes won't list C1 in _CST -
 429                  * Let acpi_processor_get_power_info_default() handle them later
 430                  */
 431                 if (i == 1 && cx.type != ACPI_STATE_C1)
 432                         current_count++;
 433
 434                 cx.address = reg->address;
 435                 cx.index = current_count + 1;
 436
 437                 cx.entry_method = ACPI_CSTATE_SYSTEMIO;
 438                 if (reg->space_id == ACPI_ADR_SPACE_FIXED_HARDWARE) {
 439                         if (acpi_processor_ffh_cstate_probe
 440                                         (pr->id, &cx, reg) == 0) {
 441                                 cx.entry_method = ACPI_CSTATE_FFH;
 442                         } else if (cx.type == ACPI_STATE_C1) {
 443                                 /*
 444                                  * C1 is a special case where FIXED_HARDWARE
 445                                  * can be handled in non-MWAIT way as well.
 446                                  * In that case, save this _CST entry info.
 447                                  * Otherwise, ignore this info and continue.
 448                                  */
 449                                 cx.entry_method = ACPI_CSTATE_HALT;
 450                                 snprintf(cx.desc, ACPI_CX_DESC_LEN, "ACPI HLT");
 451                         } else {
 452                                 continue;
 453                         }
 454                         if (cx.type == ACPI_STATE_C1 &&
 455                             (boot_option_idle_override == IDLE_NOMWAIT)) {
 456                                 /*
 457                                  * In most cases the C1 space_id obtained from
 458                                  * _CST object is FIXED_HARDWARE access mode.
 459                                  * But when the option of idle=halt is added,
 460                                  * the entry_method type should be changed from
 461                                  * CSTATE_FFH to CSTATE_HALT.
 462                                  * When the option of idle=nomwait is added,
 463                                  * the C1 entry_method type should be
 464                                  * CSTATE_HALT.
 465                                  */
 466                                 cx.entry_method = ACPI_CSTATE_HALT;
 467                                 snprintf(cx.desc, ACPI_CX_DESC_LEN, "ACPI HLT");
 468                         }
 469                 } else {
 470                         snprintf(cx.desc, ACPI_CX_DESC_LEN, "ACPI IOPORT 0x%x",
 471                                  cx.address);
 472                 }
 473
 474                 if (cx.type == ACPI_STATE_C1) {
 475                         cx.valid = 1;
 476                 }
 477
 478                 obj = &(element->package.elements[2]);
 479                 if (obj->type != ACPI_TYPE_INTEGER)
 480                         continue;
 481
 482                 cx.latency = obj->integer.value;
 483
 484                 obj = &(element->package.elements[3]);
 485                 if (obj->type != ACPI_TYPE_INTEGER)
 486                         continue;
 487
 488                 current_count++;
 489                 memcpy(&(pr->power.states[current_count]), &cx, sizeof(cx));
 490
 491                 /*
 492                  * We support total ACPI_PROCESSOR_MAX_POWER - 1
 493                  * (From 1 through ACPI_PROCESSOR_MAX_POWER - 1)
 494                  */
 495                 if (current_count >= (ACPI_PROCESSOR_MAX_POWER - 1)) {
 496                         printk(KERN_WARNING
 497                                "Limiting number of power states to max (%d)\n",
 498                                ACPI_PROCESSOR_MAX_POWER);
 499                         printk(KERN_WARNING
 500                                "Please increase ACPI_PROCESSOR_MAX_POWER if needed.\n");
 501                         break;
 502                 }
 503         }
 504
 505         ACPI_DEBUG_PRINT((ACPI_DB_INFO, "Found %d power states\n",
 506                           current_count));
 507
 508         /* Validate number of power states discovered */
 509         if (current_count < 2)
 510                 status = -EFAULT;
 511
 512       end:
 513         kfree(buffer.pointer);
 514
 515         return status;
 516 }
 517
 518 static void acpi_processor_power_verify_c3(struct acpi_processor *pr,
 519                                            struct acpi_processor_cx *cx)
 520 {
 521         static int bm_check_flag = -1;
 522         static int bm_control_flag = -1;
 523
 524
 525         if (!cx->address)
 526                 return;
 527
 528         /*
 529          * PIIX4 Erratum #18: We don't support C3 when Type-F (fast)
 530          * DMA transfers are used by any ISA device to avoid livelock.
 531          * Note that we could disable Type-F DMA (as recommended by
 532          * the erratum), but this is known to disrupt certain ISA
 533          * devices thus we take the conservative approach.
 534          */
 535         else if (errata.piix4.fdma) {
 536                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 537                                   "C3 not supported on PIIX4 with Type-F DMA\n"));
 538                 return;
 539         }
 540
 541         /* All the logic here assumes flags.bm_check is same across all CPUs */
 542         if (bm_check_flag == -1) {
 543                 /* Determine whether bm_check is needed based on CPU  */
 544                 acpi_processor_power_init_bm_check(&(pr->flags), pr->id);
 545                 bm_check_flag = pr->flags.bm_check;
 546                 bm_control_flag = pr->flags.bm_control;
 547         } else {
 548                 pr->flags.bm_check = bm_check_flag;
 549                 pr->flags.bm_control = bm_control_flag;
 550         }
 551
 552         if (pr->flags.bm_check) {
 553                 if (!pr->flags.bm_control) {
 554                         if (pr->flags.has_cst != 1) {
 555                                 /* bus mastering control is necessary */
 556                                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 557                                         "C3 support requires BM control\n"));
 558                                 return;
 559                         } else {
 560                                 /* Here we enter C3 without bus mastering */
 561                                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 562                                         "C3 support without BM control\n"));
 563                         }
 564                 }
 565         } else {
 566                 /*
 567                  * WBINVD should be set in fadt, for C3 state to be
 568                  * supported on when bm_check is not required.
 569                  */
 570                 if (!(acpi_gbl_FADT.flags & ACPI_FADT_WBINVD)) {
 571                         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 572                                           "Cache invalidation should work properly"
 573                                           " for C3 to be enabled on SMP systems\n"));
 574                         return;
 575                 }
 576         }
 577
 578         /*
 579          * Otherwise we've met all of our C3 requirements.
 580          * Normalize the C3 latency to expidite policy.  Enable
 581          * checking of bus mastering status (bm_check) so we can
 582          * use this in our C3 policy
 583          */
 584         cx->valid = 1;
 585
 586         /*
 587          * On older chipsets, BM_RLD needs to be set
 588          * in order for Bus Master activity to wake the
 589          * system from C3.  Newer chipsets handle DMA
 590          * during C3 automatically and BM_RLD is a NOP.
 591          * In either case, the proper way to
 592          * handle BM_RLD is to set it and leave it set.
 593          */
 594         acpi_write_bit_register(ACPI_BITREG_BUS_MASTER_RLD, 1);
 595
 596         return;
 597 }
 598
 599 static int acpi_processor_power_verify(struct acpi_processor *pr)
 600 {
 601         unsigned int i;
 602         unsigned int working = 0;
 603
 604         pr->power.timer_broadcast_on_state = INT_MAX;
 605
 606         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER && i <= max_cstate; i++) {
 607                 struct acpi_processor_cx *cx = &pr->power.states[i];
 608
 609                 switch (cx->type) {
 610                 case ACPI_STATE_C1:
 611                         cx->valid = 1;
 612                         break;
 613
 614                 case ACPI_STATE_C2:
 615                         if (!cx->address)
 616                                 break;
 617                         cx->valid = 1;
 618                         break;
 619
 620                 case ACPI_STATE_C3:
 621                         acpi_processor_power_verify_c3(pr, cx);
 622                         break;
 623                 }
 624                 if (!cx->valid)
 625                         continue;
 626
 627                 lapic_timer_check_state(i, pr, cx);
 628                 tsc_check_state(cx->type);
 629                 working++;
 630         }
 631
 632         lapic_timer_propagate_broadcast(pr);
 633
 634         return (working);
 635 }
 636
 637 static int acpi_processor_get_power_info(struct acpi_processor *pr)
 638 {
 639         unsigned int i;
 640         int result;
 641
 642
 643         /* NOTE: the idle thread may not be running while calling
 644          * this function */
 645
 646         /* Zero initialize all the C-states info. */
 647         memset(pr->power.states, 0, sizeof(pr->power.states));
 648
 649         result = acpi_processor_get_power_info_cst(pr);
 650         if (result == -ENODEV)
 651                 result = acpi_processor_get_power_info_fadt(pr);
 652
 653         if (result)
 654                 return result;
 655
 656         acpi_processor_get_power_info_default(pr);
 657
 658         pr->power.count = acpi_processor_power_verify(pr);
 659
 660         /*
 661          * if one state of type C2 or C3 is available, mark this
 662          * CPU as being "idle manageable"
 663          */
 664         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
 665                 if (pr->power.states[i].valid) {
 666                         pr->power.count = i;
 667                         if (pr->power.states[i].type >= ACPI_STATE_C2)
 668                                 pr->flags.power = 1;
 669                 }
 670         }
 671
 672         return 0;
 673 }
 674
 675 /**
 676  * acpi_idle_bm_check - checks if bus master activity was detected
 677  */
 678 static int acpi_idle_bm_check(void)
 679 {
 680         u32 bm_status = 0;
 681
 682         if (bm_check_disable)
 683                 return 0;
 684
 685         acpi_read_bit_register(ACPI_BITREG_BUS_MASTER_STATUS, &bm_status);
 686         if (bm_status)
 687                 acpi_write_bit_register(ACPI_BITREG_BUS_MASTER_STATUS, 1);
 688         /*
 689          * PIIX4 Erratum #18: Note that BM_STS doesn't always reflect
 690          * the true state of bus mastering activity; forcing us to
 691          * manually check the BMIDEA bit of each IDE channel.
 692          */
 693         else if (errata.piix4.bmisx) {
 694                 if ((inb_p(errata.piix4.bmisx + 0x02) & 0x01)
 695                     || (inb_p(errata.piix4.bmisx + 0x0A) & 0x01))
 696                         bm_status = 1;
 697         }
 698         return bm_status;
 699 }
 700
 701 /**
 702  * acpi_idle_do_entry - a helper function that does C2 and C3 type entry
 703  * @cx: cstate data
 704  *
 705  * Caller disables interrupt before call and enables interrupt after return.
 706  */
 707 static inline void acpi_idle_do_entry(struct acpi_processor_cx *cx)
 708 {
 709         /* Don't trace irqs off for idle */
 710         stop_critical_timings();
 711         if (cx->entry_method == ACPI_CSTATE_FFH) {
 712                 /* Call into architectural FFH based C-state */
 713                 acpi_processor_ffh_cstate_enter(cx);
 714         } else if (cx->entry_method == ACPI_CSTATE_HALT) {
 715                 acpi_safe_halt();
 716         } else {
 717                 /* IO port based C-state */
 718                 inb(cx->address);
 719                 /* Dummy wait op - must do something useless after P_LVL2 read
 720                    because chipsets cannot guarantee that STPCLK# signal
 721                    gets asserted in time to freeze execution properly. */
 722                 inl(acpi_gbl_FADT.xpm_timer_block.address);
 723         }
 724         start_critical_timings();
 725 }
 726
 727 /**
 728  * acpi_idle_enter_c1 - enters an ACPI C1 state-type
 729  * @dev: the target CPU
 730  * @drv: cpuidle driver containing cpuidle state info
 731  * @index: index of target state
 732  *
 733  * This is equivalent to the HALT instruction.
 734  */
 735 static int acpi_idle_enter_c1(struct cpuidle_device *dev,
 736                 struct cpuidle_driver *drv, int index)
 737 {
 738         struct acpi_processor *pr;
 739         struct cpuidle_state_usage *state_usage = &dev->states_usage[index];
 740         struct acpi_processor_cx *cx = cpuidle_get_statedata(state_usage);
 741
 742         pr = __this_cpu_read(processors);
 743
 744         if (unlikely(!pr))
 745                 return -EINVAL;
 746
 747         lapic_timer_state_broadcast(pr, cx, 1);
 748         acpi_idle_do_entry(cx);
 749
 750         lapic_timer_state_broadcast(pr, cx, 0);
 751
 752         return index;
 753 }
 754
 755
 756 /**
 757  * acpi_idle_play_dead - enters an ACPI state for long-term idle (i.e. off-lining)
 758  * @dev: the target CPU
 759  * @index: the index of suggested state
 760  */
 761 static int acpi_idle_play_dead(struct cpuidle_device *dev, int index)
 762 {
 763         struct cpuidle_state_usage *state_usage = &dev->states_usage[index];
 764         struct acpi_processor_cx *cx = cpuidle_get_statedata(state_usage);
 765
 766         ACPI_FLUSH_CPU_CACHE();
 767
 768         while (1) {
 769
 770                 if (cx->entry_method == ACPI_CSTATE_HALT)
 771                         safe_halt();
 772                 else if (cx->entry_method == ACPI_CSTATE_SYSTEMIO) {
 773                         inb(cx->address);
 774                         /* See comment in acpi_idle_do_entry() */
 775                         inl(acpi_gbl_FADT.xpm_timer_block.address);
 776                 } else
 777                         return -ENODEV;
 778         }
 779
 780         /* Never reached */
 781         return 0;
 782 }
 783
 784 /**
 785  * acpi_idle_enter_simple - enters an ACPI state without BM handling
 786  * @dev: the target CPU
 787  * @drv: cpuidle driver with cpuidle state information
 788  * @index: the index of suggested state
 789  */
 790 static int acpi_idle_enter_simple(struct cpuidle_device *dev,
 791                 struct cpuidle_driver *drv, int index)
 792 {
 793         struct acpi_processor *pr;
 794         struct cpuidle_state_usage *state_usage = &dev->states_usage[index];
 795         struct acpi_processor_cx *cx = cpuidle_get_statedata(state_usage);
 796
 797         pr = __this_cpu_read(processors);
 798
 799         if (unlikely(!pr))
 800                 return -EINVAL;
 801
 802         if (cx->entry_method != ACPI_CSTATE_FFH) {
 803                 current_thread_info()->status &= ~TS_POLLING;
 804                 /*
 805                  * TS_POLLING-cleared state must be visible before we test
 806                  * NEED_RESCHED:
 807                  */
 808                 smp_mb();
 809
 810                 if (unlikely(need_resched())) {
 811                         current_thread_info()->status |= TS_POLLING;
 812                         return -EINVAL;
 813                 }
 814         }
 815
 816         /*
 817          * Must be done before busmaster disable as we might need to
 818          * access HPET !
 819          */
 820         lapic_timer_state_broadcast(pr, cx, 1);
 821
 822         if (cx->type == ACPI_STATE_C3)
 823                 ACPI_FLUSH_CPU_CACHE();
 824
 825         /* Tell the scheduler that we are going deep-idle: */
 826         sched_clock_idle_sleep_event();
 827         acpi_idle_do_entry(cx);
 828
 829         sched_clock_idle_wakeup_event(0);
 830
 831         if (cx->entry_method != ACPI_CSTATE_FFH)
 832                 current_thread_info()->status |= TS_POLLING;
 833
 834         lapic_timer_state_broadcast(pr, cx, 0);
 835         return index;
 836 }
 837
 838 static int c3_cpu_count;
 839 static DEFINE_RAW_SPINLOCK(c3_lock);
 840
 841 /**
 842  * acpi_idle_enter_bm - enters C3 with proper BM handling
 843  * @dev: the target CPU
 844  * @drv: cpuidle driver containing state data
 845  * @index: the index of suggested state
 846  *
 847  * If BM is detected, the deepest non-C3 idle state is entered instead.
 848  */
 849 static int acpi_idle_enter_bm(struct cpuidle_device *dev,
 850                 struct cpuidle_driver *drv, int index)
 851 {
 852         struct acpi_processor *pr;
 853         struct cpuidle_state_usage *state_usage = &dev->states_usage[index];
 854         struct acpi_processor_cx *cx = cpuidle_get_statedata(state_usage);
 855
 856         pr = __this_cpu_read(processors);
 857
 858         if (unlikely(!pr))
 859                 return -EINVAL;
 860
 861         if (!cx->bm_sts_skip && acpi_idle_bm_check()) {
 862                 if (drv->safe_state_index >= 0) {
 863                         return drv->states[drv->safe_state_index].enter(dev,
 864                                                 drv, drv->safe_state_index);
 865                 } else {
 866                         acpi_safe_halt();
 867                         return -EBUSY;
 868                 }
 869         }
 870
 871         if (cx->entry_method != ACPI_CSTATE_FFH) {
 872                 current_thread_info()->status &= ~TS_POLLING;
 873                 /*
 874                  * TS_POLLING-cleared state must be visible before we test
 875                  * NEED_RESCHED:
 876                  */
 877                 smp_mb();
 878
 879                 if (unlikely(need_resched())) {
 880                         current_thread_info()->status |= TS_POLLING;
 881                         return -EINVAL;
 882                 }
 883         }
 884
 885         acpi_unlazy_tlb(smp_processor_id());
 886
 887         /* Tell the scheduler that we are going deep-idle: */
 888         sched_clock_idle_sleep_event();
 889         /*
 890          * Must be done before busmaster disable as we might need to
 891          * access HPET !
 892          */
 893         lapic_timer_state_broadcast(pr, cx, 1);
 894
 895         /*
 896          * disable bus master
 897          * bm_check implies we need ARB_DIS
 898          * !bm_check implies we need cache flush
 899          * bm_control implies whether we can do ARB_DIS
 900          *
 901          * That leaves a case where bm_check is set and bm_control is
 902          * not set. In that case we cannot do much, we enter C3
 903          * without doing anything.
 904          */
 905         if (pr->flags.bm_check && pr->flags.bm_control) {
 906                 raw_spin_lock(&c3_lock);
 907                 c3_cpu_count++;
 908                 /* Disable bus master arbitration when all CPUs are in C3 */
 909                 if (c3_cpu_count == num_online_cpus())
 910                         acpi_write_bit_register(ACPI_BITREG_ARB_DISABLE, 1);
 911                 raw_spin_unlock(&c3_lock);
 912         } else if (!pr->flags.bm_check) {
 913                 ACPI_FLUSH_CPU_CACHE();
 914         }
 915
 916         acpi_idle_do_entry(cx);
 917
 918         /* Re-enable bus master arbitration */
 919         if (pr->flags.bm_check && pr->flags.bm_control) {
 920                 raw_spin_lock(&c3_lock);
 921                 acpi_write_bit_register(ACPI_BITREG_ARB_DISABLE, 0);
 922                 c3_cpu_count--;
 923                 raw_spin_unlock(&c3_lock);
 924         }
 925
 926         sched_clock_idle_wakeup_event(0);
 927
 928         if (cx->entry_method != ACPI_CSTATE_FFH)
 929                 current_thread_info()->status |= TS_POLLING;
 930
 931         lapic_timer_state_broadcast(pr, cx, 0);
 932         return index;
 933 }
 934
 935 struct cpuidle_driver acpi_idle_driver = {
 936         .name =         "acpi_idle",
 937         .owner =        THIS_MODULE,
 938         .en_core_tk_irqen = 1,
 939 };
 940
 941 /**
 942  * acpi_processor_setup_cpuidle_cx - prepares and configures CPUIDLE
 943  * device i.e. per-cpu data
 944  *
 945  * @pr: the ACPI processor
 946  */
 947 static int acpi_processor_setup_cpuidle_cx(struct acpi_processor *pr)
 948 {
 949         int i, count = CPUIDLE_DRIVER_STATE_START;
 950         struct acpi_processor_cx *cx;
 951         struct cpuidle_state_usage *state_usage;
 952         struct cpuidle_device *dev = per_cpu(acpi_cpuidle_device, pr->id);
 953
 954         if (!pr->flags.power_setup_done)
 955                 return -EINVAL;
 956
 957         if (pr->flags.power == 0) {
 958                 return -EINVAL;
 959         }
 960
 961         if (!dev)
 962                 return -EINVAL;
 963
 964         dev->cpu = pr->id;
 965
 966         if (max_cstate == 0)
 967                 max_cstate = 1;
 968
 969         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER && i <= max_cstate; i++) {
 970                 cx = &pr->power.states[i];
 971                 state_usage = &dev->states_usage[count];
 972
 973                 if (!cx->valid)
 974                         continue;
 975
 976 #ifdef CONFIG_HOTPLUG_CPU
 977                 if ((cx->type != ACPI_STATE_C1) && (num_online_cpus() > 1) &&
 978                     !pr->flags.has_cst &&
 979                     !(acpi_gbl_FADT.flags & ACPI_FADT_C2_MP_SUPPORTED))
 980                         continue;
 981 #endif
 982
 983                 cpuidle_set_statedata(state_usage, cx);
 984
 985                 count++;
 986                 if (count == CPUIDLE_STATE_MAX)
 987                         break;
 988         }
 989
 990         dev->state_count = count;
 991
 992         if (!count)
 993                 return -EINVAL;
 994
 995         return 0;
 996 }
 997
 998 /**
 999  * acpi_processor_setup_cpuidle states- prepares and configures cpuidle
1000  * global state data i.e. idle routines
1001  *
1002  * @pr: the ACPI processor
1003  */
1004 static int acpi_processor_setup_cpuidle_states(struct acpi_processor *pr)
1005 {
1006         int i, count = CPUIDLE_DRIVER_STATE_START;
1007         struct acpi_processor_cx *cx;
1008         struct cpuidle_state *state;
1009         struct cpuidle_driver *drv = &acpi_idle_driver;
1010
1011         if (!pr->flags.power_setup_done)
1012                 return -EINVAL;
1013
1014         if (pr->flags.power == 0)
1015                 return -EINVAL;
1016
1017         drv->safe_state_index = -1;
1018         for (i = 0; i < CPUIDLE_STATE_MAX; i++) {
1019                 drv->states[i].name[0] = '\0';
1020                 drv->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
1029                 if (!cx->valid)
1030                         continue;
1031
1032 #ifdef CONFIG_HOTPLUG_CPU
1033                 if ((cx->type != ACPI_STATE_C1) && (num_online_cpus() > 1) &&
1034                     !pr->flags.has_cst &&
1035                     !(acpi_gbl_FADT.flags & ACPI_FADT_C2_MP_SUPPORTED))
1036                         continue;
1037 #endif
1038
1039                 state = &drv->states[count];
1040                 snprintf(state->name, CPUIDLE_NAME_LEN, "C%d", i);
1041                 strncpy(state->desc, cx->desc, CPUIDLE_DESC_LEN);
1042                 state->exit_latency = cx->latency;
1043                 state->target_residency = cx->latency * latency_factor;
1044
1045                 state->flags = 0;
1046                 switch (cx->type) {
1047                         case ACPI_STATE_C1:
1048                         if (cx->entry_method == ACPI_CSTATE_FFH)
1049                                 state->flags |= CPUIDLE_FLAG_TIME_VALID;
1050
1051                         state->enter = acpi_idle_enter_c1;
1052                         state->enter_dead = acpi_idle_play_dead;
1053                         drv->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                         state->enter_dead = acpi_idle_play_dead;
1060                         drv->safe_state_index = count;
1061                         break;
1062
1063                         case ACPI_STATE_C3:
1064                         state->flags |= CPUIDLE_FLAG_TIME_VALID;
1065                         state->enter = pr->flags.bm_check ?
1066                                         acpi_idle_enter_bm :
1067                                         acpi_idle_enter_simple;
1068                         break;
1069                 }
1070
1071                 count++;
1072                 if (count == CPUIDLE_STATE_MAX)
1073                         break;
1074         }
1075
1076         drv->state_count = count;
1077
1078         if (!count)
1079                 return -EINVAL;
1080
1081         return 0;
1082 }
1083
1084 int acpi_processor_hotplug(struct acpi_processor *pr)
1085 {
1086         int ret = 0;
1087         struct cpuidle_device *dev;
1088
1089         if (disabled_by_idle_boot_param())
1090                 return 0;
1091
1092         if (!pr)
1093                 return -EINVAL;
1094
1095         if (nocst) {
1096                 return -ENODEV;
1097         }
1098
1099         if (!pr->flags.power_setup_done)
1100                 return -ENODEV;
1101
1102         dev = per_cpu(acpi_cpuidle_device, pr->id);
1103         cpuidle_pause_and_lock();
1104         cpuidle_disable_device(dev);
1105         acpi_processor_get_power_info(pr);
1106         if (pr->flags.power) {
1107                 acpi_processor_setup_cpuidle_cx(pr);
1108                 ret = cpuidle_enable_device(dev);
1109         }
1110         cpuidle_resume_and_unlock();
1111
1112         return ret;
1113 }
1114
1115 int acpi_processor_cst_has_changed(struct acpi_processor *pr)
1116 {
1117         int cpu;
1118         struct acpi_processor *_pr;
1119         struct cpuidle_device *dev;
1120
1121         if (disabled_by_idle_boot_param())
1122                 return 0;
1123
1124         if (!pr)
1125                 return -EINVAL;
1126
1127         if (nocst)
1128                 return -ENODEV;
1129
1130         if (!pr->flags.power_setup_done)
1131                 return -ENODEV;
1132
1133         /*
1134          * FIXME:  Design the ACPI notification to make it once per
1135          * system instead of once per-cpu.  This condition is a hack
1136          * to make the code that updates C-States be called once.
1137          */
1138
1139         if (pr->id == 0 && cpuidle_get_driver() == &acpi_idle_driver) {
1140
1141                 cpuidle_pause_and_lock();
1142                 /* Protect against cpu-hotplug */
1143                 get_online_cpus();
1144
1145                 /* Disable all cpuidle devices */
1146                 for_each_online_cpu(cpu) {
1147                         _pr = per_cpu(processors, cpu);
1148                         if (!_pr || !_pr->flags.power_setup_done)
1149                                 continue;
1150                         dev = per_cpu(acpi_cpuidle_device, cpu);
1151                         cpuidle_disable_device(dev);
1152                 }
1153
1154                 /* Populate Updated C-state information */
1155                 acpi_processor_get_power_info(pr);
1156                 acpi_processor_setup_cpuidle_states(pr);
1157
1158                 /* Enable all cpuidle devices */
1159                 for_each_online_cpu(cpu) {
1160                         _pr = per_cpu(processors, cpu);
1161                         if (!_pr || !_pr->flags.power_setup_done)
1162                                 continue;
1163                         acpi_processor_get_power_info(_pr);
1164                         if (_pr->flags.power) {
1165                                 acpi_processor_setup_cpuidle_cx(_pr);
1166                                 dev = per_cpu(acpi_cpuidle_device, cpu);
1167                                 cpuidle_enable_device(dev);
1168                         }
1169                 }
1170                 put_online_cpus();
1171                 cpuidle_resume_and_unlock();
1172         }
1173
1174         return 0;
1175 }
1176
1177 static int acpi_processor_registered;
1178
1179 int __cpuinit acpi_processor_power_init(struct acpi_processor *pr)
1180 {
1181         acpi_status status = 0;
1182         int retval;
1183         struct cpuidle_device *dev;
1184         static int first_run;
1185
1186         if (disabled_by_idle_boot_param())
1187                 return 0;
1188
1189         if (!first_run) {
1190                 dmi_check_system(processor_power_dmi_table);
1191                 max_cstate = acpi_processor_cstate_check(max_cstate);
1192                 if (max_cstate < ACPI_C_STATES_MAX)
1193                         printk(KERN_NOTICE
1194                                "ACPI: processor limited to max C-state %d\n",
1195                                max_cstate);
1196                 first_run++;
1197         }
1198
1199         if (!pr)
1200                 return -EINVAL;
1201
1202         if (acpi_gbl_FADT.cst_control && !nocst) {
1203                 status =
1204                     acpi_os_write_port(acpi_gbl_FADT.smi_command, acpi_gbl_FADT.cst_control, 8);
1205                 if (ACPI_FAILURE(status)) {
1206                         ACPI_EXCEPTION((AE_INFO, status,
1207                                         "Notifying BIOS of _CST ability failed"));
1208                 }
1209         }
1210
1211         acpi_processor_get_power_info(pr);
1212         pr->flags.power_setup_done = 1;
1213
1214         /*
1215          * Install the idle handler if processor power management is supported.
1216          * Note that we use previously set idle handler will be used on
1217          * platforms that only support C1.
1218          */
1219         if (pr->flags.power) {
1220                 /* Register acpi_idle_driver if not already registered */
1221                 if (!acpi_processor_registered) {
1222                         acpi_processor_setup_cpuidle_states(pr);
1223                         retval = cpuidle_register_driver(&acpi_idle_driver);
1224                         if (retval)
1225                                 return retval;
1226                         printk(KERN_DEBUG "ACPI: %s registered with cpuidle\n",
1227                                         acpi_idle_driver.name);
1228                 }
1229
1230                 dev = kzalloc(sizeof(*dev), GFP_KERNEL);
1231                 if (!dev)
1232                         return -ENOMEM;
1233                 per_cpu(acpi_cpuidle_device, pr->id) = dev;
1234
1235                 acpi_processor_setup_cpuidle_cx(pr);
1236
1237                 /* Register per-cpu cpuidle_device. Cpuidle driver
1238                  * must already be registered before registering device
1239                  */
1240                 retval = cpuidle_register_device(dev);
1241                 if (retval) {
1242                         if (acpi_processor_registered == 0)
1243                                 cpuidle_unregister_driver(&acpi_idle_driver);
1244                         return retval;
1245                 }
1246                 acpi_processor_registered++;
1247         }
1248         return 0;
1249 }
1250
1251 int acpi_processor_power_exit(struct acpi_processor *pr)
1252 {
1253         struct cpuidle_device *dev = per_cpu(acpi_cpuidle_device, pr->id);
1254
1255         if (disabled_by_idle_boot_param())
1256                 return 0;
1257
1258         if (pr->flags.power) {
1259                 cpuidle_unregister_device(dev);
1260                 acpi_processor_registered--;
1261                 if (acpi_processor_registered == 0)
1262                         cpuidle_unregister_driver(&acpi_idle_driver);
1263         }
1264
1265         pr->flags.power_setup_done = 0;
1266         return 0;
1267 }