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  * ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
  25  */
  26 #define pr_fmt(fmt) "ACPI: " fmt
  27
  28 #include <linux/module.h>
  29 #include <linux/acpi.h>
  30 #include <linux/dmi.h>
  31 #include <linux/sched.h>       /* need_resched() */
  32 #include <linux/tick.h>
  33 #include <linux/cpuidle.h>
  34 #include <acpi/processor.h>
  35
  36 /*
  37  * Include the apic definitions for x86 to have the APIC timer related defines
  38  * available also for UP (on SMP it gets magically included via linux/smp.h).
  39  * asm/acpi.h is not an option, as it would require more include magic. Also
  40  * creating an empty asm-ia64/apic.h would just trade pest vs. cholera.
  41  */
  42 #ifdef CONFIG_X86
  43 #include <asm/apic.h>
  44 #endif
  45
  46 #define ACPI_PROCESSOR_CLASS            "processor"
  47 #define _COMPONENT              ACPI_PROCESSOR_COMPONENT
  48 ACPI_MODULE_NAME("processor_idle");
  49
  50 static unsigned int max_cstate __read_mostly = ACPI_PROCESSOR_MAX_POWER;
  51 module_param(max_cstate, uint, 0000);
  52 static unsigned int nocst __read_mostly;
  53 module_param(nocst, uint, 0000);
  54 static int bm_check_disable __read_mostly;
  55 module_param(bm_check_disable, uint, 0000);
  56
  57 static unsigned int latency_factor __read_mostly = 2;
  58 module_param(latency_factor, uint, 0644);
  59
  60 static DEFINE_PER_CPU(struct cpuidle_device *, acpi_cpuidle_device);
  61
  62 static
  63 DEFINE_PER_CPU(struct acpi_processor_cx * [CPUIDLE_STATE_MAX], acpi_cstate);
  64
  65 static int disabled_by_idle_boot_param(void)
  66 {
  67         return boot_option_idle_override == IDLE_POLL ||
  68                 boot_option_idle_override == IDLE_HALT;
  69 }
  70
  71 /*
  72  * IBM ThinkPad R40e crashes mysteriously when going into C2 or C3.
  73  * For now disable this. Probably a bug somewhere else.
  74  *
  75  * To skip this limit, boot/load with a large max_cstate limit.
  76  */
  77 static int set_max_cstate(const struct dmi_system_id *id)
  78 {
  79         if (max_cstate > ACPI_PROCESSOR_MAX_POWER)
  80                 return 0;
  81
  82         pr_notice("%s detected - limiting to C%ld max_cstate."
  83                   " Override with \"processor.max_cstate=%d\"\n", id->ident,
  84                   (long)id->driver_data, ACPI_PROCESSOR_MAX_POWER + 1);
  85
  86         max_cstate = (long)id->driver_data;
  87
  88         return 0;
  89 }
  90
  91 static const struct dmi_system_id processor_power_dmi_table[] = {
  92         { set_max_cstate, "Clevo 5600D", {
  93           DMI_MATCH(DMI_BIOS_VENDOR,"Phoenix Technologies LTD"),
  94           DMI_MATCH(DMI_BIOS_VERSION,"SHE845M0.86C.0013.D.0302131307")},
  95          (void *)2},
  96         { set_max_cstate, "Pavilion zv5000", {
  97           DMI_MATCH(DMI_SYS_VENDOR, "Hewlett-Packard"),
  98           DMI_MATCH(DMI_PRODUCT_NAME,"Pavilion zv5000 (DS502A#ABA)")},
  99          (void *)1},
 100         { set_max_cstate, "Asus L8400B", {
 101           DMI_MATCH(DMI_SYS_VENDOR, "ASUSTeK Computer Inc."),
 102           DMI_MATCH(DMI_PRODUCT_NAME,"L8400B series Notebook PC")},
 103          (void *)1},
 104         {},
 105 };
 106
 107
 108 /*
 109  * Callers should disable interrupts before the call and enable
 110  * interrupts after return.
 111  */
 112 static void acpi_safe_halt(void)
 113 {
 114         if (!tif_need_resched()) {
 115                 safe_halt();
 116                 local_irq_disable();
 117         }
 118 }
 119
 120 #ifdef ARCH_APICTIMER_STOPS_ON_C3
 121
 122 /*
 123  * Some BIOS implementations switch to C3 in the published C2 state.
 124  * This seems to be a common problem on AMD boxen, but other vendors
 125  * are affected too. We pick the most conservative approach: we assume
 126  * that the local APIC stops in both C2 and C3.
 127  */
 128 static void lapic_timer_check_state(int state, struct acpi_processor *pr,
 129                                    struct acpi_processor_cx *cx)
 130 {
 131         struct acpi_processor_power *pwr = &pr->power;
 132         u8 type = local_apic_timer_c2_ok ? ACPI_STATE_C3 : ACPI_STATE_C2;
 133
 134         if (cpu_has(&cpu_data(pr->id), X86_FEATURE_ARAT))
 135                 return;
 136
 137         if (amd_e400_c1e_detected)
 138                 type = ACPI_STATE_C1;
 139
 140         /*
 141          * Check, if one of the previous states already marked the lapic
 142          * unstable
 143          */
 144         if (pwr->timer_broadcast_on_state < state)
 145                 return;
 146
 147         if (cx->type >= type)
 148                 pr->power.timer_broadcast_on_state = state;
 149 }
 150
 151 static void __lapic_timer_propagate_broadcast(void *arg)
 152 {
 153         struct acpi_processor *pr = (struct acpi_processor *) arg;
 154
 155         if (pr->power.timer_broadcast_on_state < INT_MAX)
 156                 tick_broadcast_enable();
 157         else
 158                 tick_broadcast_disable();
 159 }
 160
 161 static void lapic_timer_propagate_broadcast(struct acpi_processor *pr)
 162 {
 163         smp_call_function_single(pr->id, __lapic_timer_propagate_broadcast,
 164                                  (void *)pr, 1);
 165 }
 166
 167 /* Power(C) State timer broadcast control */
 168 static void lapic_timer_state_broadcast(struct acpi_processor *pr,
 169                                        struct acpi_processor_cx *cx,
 170                                        int broadcast)
 171 {
 172         int state = cx - pr->power.states;
 173
 174         if (state >= pr->power.timer_broadcast_on_state) {
 175                 if (broadcast)
 176                         tick_broadcast_enter();
 177                 else
 178                         tick_broadcast_exit();
 179         }
 180 }
 181
 182 #else
 183
 184 static void lapic_timer_check_state(int state, struct acpi_processor *pr,
 185                                    struct acpi_processor_cx *cstate) { }
 186 static void lapic_timer_propagate_broadcast(struct acpi_processor *pr) { }
 187 static void lapic_timer_state_broadcast(struct acpi_processor *pr,
 188                                        struct acpi_processor_cx *cx,
 189                                        int broadcast)
 190 {
 191 }
 192
 193 #endif
 194
 195 #if defined(CONFIG_X86)
 196 static void tsc_check_state(int state)
 197 {
 198         switch (boot_cpu_data.x86_vendor) {
 199         case X86_VENDOR_AMD:
 200         case X86_VENDOR_INTEL:
 201                 /*
 202                  * AMD Fam10h TSC will tick in all
 203                  * C/P/S0/S1 states when this bit is set.
 204                  */
 205                 if (boot_cpu_has(X86_FEATURE_NONSTOP_TSC))
 206                         return;
 207
 208                 /*FALL THROUGH*/
 209         default:
 210                 /* TSC could halt in idle, so notify users */
 211                 if (state > ACPI_STATE_C1)
 212                         mark_tsc_unstable("TSC halts in idle");
 213         }
 214 }
 215 #else
 216 static void tsc_check_state(int state) { return; }
 217 #endif
 218
 219 static int acpi_processor_get_power_info_fadt(struct acpi_processor *pr)
 220 {
 221
 222         if (!pr->pblk)
 223                 return -ENODEV;
 224
 225         /* if info is obtained from pblk/fadt, type equals state */
 226         pr->power.states[ACPI_STATE_C2].type = ACPI_STATE_C2;
 227         pr->power.states[ACPI_STATE_C3].type = ACPI_STATE_C3;
 228
 229 #ifndef CONFIG_HOTPLUG_CPU
 230         /*
 231          * Check for P_LVL2_UP flag before entering C2 and above on
 232          * an SMP system.
 233          */
 234         if ((num_online_cpus() > 1) &&
 235             !(acpi_gbl_FADT.flags & ACPI_FADT_C2_MP_SUPPORTED))
 236                 return -ENODEV;
 237 #endif
 238
 239         /* determine C2 and C3 address from pblk */
 240         pr->power.states[ACPI_STATE_C2].address = pr->pblk + 4;
 241         pr->power.states[ACPI_STATE_C3].address = pr->pblk + 5;
 242
 243         /* determine latencies from FADT */
 244         pr->power.states[ACPI_STATE_C2].latency = acpi_gbl_FADT.c2_latency;
 245         pr->power.states[ACPI_STATE_C3].latency = acpi_gbl_FADT.c3_latency;
 246
 247         /*
 248          * FADT specified C2 latency must be less than or equal to
 249          * 100 microseconds.
 250          */
 251         if (acpi_gbl_FADT.c2_latency > ACPI_PROCESSOR_MAX_C2_LATENCY) {
 252                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 253                         "C2 latency too large [%d]\n", acpi_gbl_FADT.c2_latency));
 254                 /* invalidate C2 */
 255                 pr->power.states[ACPI_STATE_C2].address = 0;
 256         }
 257
 258         /*
 259          * FADT supplied C3 latency must be less than or equal to
 260          * 1000 microseconds.
 261          */
 262         if (acpi_gbl_FADT.c3_latency > ACPI_PROCESSOR_MAX_C3_LATENCY) {
 263                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 264                         "C3 latency too large [%d]\n", acpi_gbl_FADT.c3_latency));
 265                 /* invalidate C3 */
 266                 pr->power.states[ACPI_STATE_C3].address = 0;
 267         }
 268
 269         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 270                           "lvl2[0x%08x] lvl3[0x%08x]\n",
 271                           pr->power.states[ACPI_STATE_C2].address,
 272                           pr->power.states[ACPI_STATE_C3].address));
 273
 274         return 0;
 275 }
 276
 277 static int acpi_processor_get_power_info_default(struct acpi_processor *pr)
 278 {
 279         if (!pr->power.states[ACPI_STATE_C1].valid) {
 280                 /* set the first C-State to C1 */
 281                 /* all processors need to support C1 */
 282                 pr->power.states[ACPI_STATE_C1].type = ACPI_STATE_C1;
 283                 pr->power.states[ACPI_STATE_C1].valid = 1;
 284                 pr->power.states[ACPI_STATE_C1].entry_method = ACPI_CSTATE_HALT;
 285         }
 286         /* the C0 state only exists as a filler in our array */
 287         pr->power.states[ACPI_STATE_C0].valid = 1;
 288         return 0;
 289 }
 290
 291 static int acpi_processor_get_power_info_cst(struct acpi_processor *pr)
 292 {
 293         acpi_status status;
 294         u64 count;
 295         int current_count;
 296         int i, ret = 0;
 297         struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
 298         union acpi_object *cst;
 299
 300
 301         if (nocst)
 302                 return -ENODEV;
 303
 304         current_count = 0;
 305
 306         status = acpi_evaluate_object(pr->handle, "_CST", NULL, &buffer);
 307         if (ACPI_FAILURE(status)) {
 308                 ACPI_DEBUG_PRINT((ACPI_DB_INFO, "No _CST, giving up\n"));
 309                 return -ENODEV;
 310         }
 311
 312         cst = buffer.pointer;
 313
 314         /* There must be at least 2 elements */
 315         if (!cst || (cst->type != ACPI_TYPE_PACKAGE) || cst->package.count < 2) {
 316                 pr_err("not enough elements in _CST\n");
 317                 ret = -EFAULT;
 318                 goto end;
 319         }
 320
 321         count = cst->package.elements[0].integer.value;
 322
 323         /* Validate number of power states. */
 324         if (count < 1 || count != cst->package.count - 1) {
 325                 pr_err("count given by _CST is not valid\n");
 326                 ret = -EFAULT;
 327                 goto end;
 328         }
 329
 330         /* Tell driver that at least _CST is supported. */
 331         pr->flags.has_cst = 1;
 332
 333         for (i = 1; i <= count; i++) {
 334                 union acpi_object *element;
 335                 union acpi_object *obj;
 336                 struct acpi_power_register *reg;
 337                 struct acpi_processor_cx cx;
 338
 339                 memset(&cx, 0, sizeof(cx));
 340
 341                 element = &(cst->package.elements[i]);
 342                 if (element->type != ACPI_TYPE_PACKAGE)
 343                         continue;
 344
 345                 if (element->package.count != 4)
 346                         continue;
 347
 348                 obj = &(element->package.elements[0]);
 349
 350                 if (obj->type != ACPI_TYPE_BUFFER)
 351                         continue;
 352
 353                 reg = (struct acpi_power_register *)obj->buffer.pointer;
 354
 355                 if (reg->space_id != ACPI_ADR_SPACE_SYSTEM_IO &&
 356                     (reg->space_id != ACPI_ADR_SPACE_FIXED_HARDWARE))
 357                         continue;
 358
 359                 /* There should be an easy way to extract an integer... */
 360                 obj = &(element->package.elements[1]);
 361                 if (obj->type != ACPI_TYPE_INTEGER)
 362                         continue;
 363
 364                 cx.type = obj->integer.value;
 365                 /*
 366                  * Some buggy BIOSes won't list C1 in _CST -
 367                  * Let acpi_processor_get_power_info_default() handle them later
 368                  */
 369                 if (i == 1 && cx.type != ACPI_STATE_C1)
 370                         current_count++;
 371
 372                 cx.address = reg->address;
 373                 cx.index = current_count + 1;
 374
 375                 cx.entry_method = ACPI_CSTATE_SYSTEMIO;
 376                 if (reg->space_id == ACPI_ADR_SPACE_FIXED_HARDWARE) {
 377                         if (acpi_processor_ffh_cstate_probe
 378                                         (pr->id, &cx, reg) == 0) {
 379                                 cx.entry_method = ACPI_CSTATE_FFH;
 380                         } else if (cx.type == ACPI_STATE_C1) {
 381                                 /*
 382                                  * C1 is a special case where FIXED_HARDWARE
 383                                  * can be handled in non-MWAIT way as well.
 384                                  * In that case, save this _CST entry info.
 385                                  * Otherwise, ignore this info and continue.
 386                                  */
 387                                 cx.entry_method = ACPI_CSTATE_HALT;
 388                                 snprintf(cx.desc, ACPI_CX_DESC_LEN, "ACPI HLT");
 389                         } else {
 390                                 continue;
 391                         }
 392                         if (cx.type == ACPI_STATE_C1 &&
 393                             (boot_option_idle_override == IDLE_NOMWAIT)) {
 394                                 /*
 395                                  * In most cases the C1 space_id obtained from
 396                                  * _CST object is FIXED_HARDWARE access mode.
 397                                  * But when the option of idle=halt is added,
 398                                  * the entry_method type should be changed from
 399                                  * CSTATE_FFH to CSTATE_HALT.
 400                                  * When the option of idle=nomwait is added,
 401                                  * the C1 entry_method type should be
 402                                  * CSTATE_HALT.
 403                                  */
 404                                 cx.entry_method = ACPI_CSTATE_HALT;
 405                                 snprintf(cx.desc, ACPI_CX_DESC_LEN, "ACPI HLT");
 406                         }
 407                 } else {
 408                         snprintf(cx.desc, ACPI_CX_DESC_LEN, "ACPI IOPORT 0x%x",
 409                                  cx.address);
 410                 }
 411
 412                 if (cx.type == ACPI_STATE_C1) {
 413                         cx.valid = 1;
 414                 }
 415
 416                 obj = &(element->package.elements[2]);
 417                 if (obj->type != ACPI_TYPE_INTEGER)
 418                         continue;
 419
 420                 cx.latency = obj->integer.value;
 421
 422                 obj = &(element->package.elements[3]);
 423                 if (obj->type != ACPI_TYPE_INTEGER)
 424                         continue;
 425
 426                 current_count++;
 427                 memcpy(&(pr->power.states[current_count]), &cx, sizeof(cx));
 428
 429                 /*
 430                  * We support total ACPI_PROCESSOR_MAX_POWER - 1
 431                  * (From 1 through ACPI_PROCESSOR_MAX_POWER - 1)
 432                  */
 433                 if (current_count >= (ACPI_PROCESSOR_MAX_POWER - 1)) {
 434                         pr_warn("Limiting number of power states to max (%d)\n",
 435                                 ACPI_PROCESSOR_MAX_POWER);
 436                         pr_warn("Please increase ACPI_PROCESSOR_MAX_POWER if needed.\n");
 437                         break;
 438                 }
 439         }
 440
 441         ACPI_DEBUG_PRINT((ACPI_DB_INFO, "Found %d power states\n",
 442                           current_count));
 443
 444         /* Validate number of power states discovered */
 445         if (current_count < 2)
 446                 ret = -EFAULT;
 447
 448       end:
 449         kfree(buffer.pointer);
 450
 451         return ret;
 452 }
 453
 454 static void acpi_processor_power_verify_c3(struct acpi_processor *pr,
 455                                            struct acpi_processor_cx *cx)
 456 {
 457         static int bm_check_flag = -1;
 458         static int bm_control_flag = -1;
 459
 460
 461         if (!cx->address)
 462                 return;
 463
 464         /*
 465          * PIIX4 Erratum #18: We don't support C3 when Type-F (fast)
 466          * DMA transfers are used by any ISA device to avoid livelock.
 467          * Note that we could disable Type-F DMA (as recommended by
 468          * the erratum), but this is known to disrupt certain ISA
 469          * devices thus we take the conservative approach.
 470          */
 471         else if (errata.piix4.fdma) {
 472                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 473                                   "C3 not supported on PIIX4 with Type-F DMA\n"));
 474                 return;
 475         }
 476
 477         /* All the logic here assumes flags.bm_check is same across all CPUs */
 478         if (bm_check_flag == -1) {
 479                 /* Determine whether bm_check is needed based on CPU  */
 480                 acpi_processor_power_init_bm_check(&(pr->flags), pr->id);
 481                 bm_check_flag = pr->flags.bm_check;
 482                 bm_control_flag = pr->flags.bm_control;
 483         } else {
 484                 pr->flags.bm_check = bm_check_flag;
 485                 pr->flags.bm_control = bm_control_flag;
 486         }
 487
 488         if (pr->flags.bm_check) {
 489                 if (!pr->flags.bm_control) {
 490                         if (pr->flags.has_cst != 1) {
 491                                 /* bus mastering control is necessary */
 492                                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 493                                         "C3 support requires BM control\n"));
 494                                 return;
 495                         } else {
 496                                 /* Here we enter C3 without bus mastering */
 497                                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 498                                         "C3 support without BM control\n"));
 499                         }
 500                 }
 501         } else {
 502                 /*
 503                  * WBINVD should be set in fadt, for C3 state to be
 504                  * supported on when bm_check is not required.
 505                  */
 506                 if (!(acpi_gbl_FADT.flags & ACPI_FADT_WBINVD)) {
 507                         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 508                                           "Cache invalidation should work properly"
 509                                           " for C3 to be enabled on SMP systems\n"));
 510                         return;
 511                 }
 512         }
 513
 514         /*
 515          * Otherwise we've met all of our C3 requirements.
 516          * Normalize the C3 latency to expidite policy.  Enable
 517          * checking of bus mastering status (bm_check) so we can
 518          * use this in our C3 policy
 519          */
 520         cx->valid = 1;
 521
 522         /*
 523          * On older chipsets, BM_RLD needs to be set
 524          * in order for Bus Master activity to wake the
 525          * system from C3.  Newer chipsets handle DMA
 526          * during C3 automatically and BM_RLD is a NOP.
 527          * In either case, the proper way to
 528          * handle BM_RLD is to set it and leave it set.
 529          */
 530         acpi_write_bit_register(ACPI_BITREG_BUS_MASTER_RLD, 1);
 531
 532         return;
 533 }
 534
 535 static int acpi_processor_power_verify(struct acpi_processor *pr)
 536 {
 537         unsigned int i;
 538         unsigned int working = 0;
 539
 540         pr->power.timer_broadcast_on_state = INT_MAX;
 541
 542         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER && i <= max_cstate; i++) {
 543                 struct acpi_processor_cx *cx = &pr->power.states[i];
 544
 545                 switch (cx->type) {
 546                 case ACPI_STATE_C1:
 547                         cx->valid = 1;
 548                         break;
 549
 550                 case ACPI_STATE_C2:
 551                         if (!cx->address)
 552                                 break;
 553                         cx->valid = 1;
 554                         break;
 555
 556                 case ACPI_STATE_C3:
 557                         acpi_processor_power_verify_c3(pr, cx);
 558                         break;
 559                 }
 560                 if (!cx->valid)
 561                         continue;
 562
 563                 lapic_timer_check_state(i, pr, cx);
 564                 tsc_check_state(cx->type);
 565                 working++;
 566         }
 567
 568         lapic_timer_propagate_broadcast(pr);
 569
 570         return (working);
 571 }
 572
 573 static int acpi_processor_get_power_info(struct acpi_processor *pr)
 574 {
 575         unsigned int i;
 576         int result;
 577
 578
 579         /* NOTE: the idle thread may not be running while calling
 580          * this function */
 581
 582         /* Zero initialize all the C-states info. */
 583         memset(pr->power.states, 0, sizeof(pr->power.states));
 584
 585         result = acpi_processor_get_power_info_cst(pr);
 586         if (result == -ENODEV)
 587                 result = acpi_processor_get_power_info_fadt(pr);
 588
 589         if (result)
 590                 return result;
 591
 592         acpi_processor_get_power_info_default(pr);
 593
 594         pr->power.count = acpi_processor_power_verify(pr);
 595
 596         /*
 597          * if one state of type C2 or C3 is available, mark this
 598          * CPU as being "idle manageable"
 599          */
 600         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
 601                 if (pr->power.states[i].valid) {
 602                         pr->power.count = i;
 603                         if (pr->power.states[i].type >= ACPI_STATE_C2)
 604                                 pr->flags.power = 1;
 605                 }
 606         }
 607
 608         return 0;
 609 }
 610
 611 /**
 612  * acpi_idle_bm_check - checks if bus master activity was detected
 613  */
 614 static int acpi_idle_bm_check(void)
 615 {
 616         u32 bm_status = 0;
 617
 618         if (bm_check_disable)
 619                 return 0;
 620
 621         acpi_read_bit_register(ACPI_BITREG_BUS_MASTER_STATUS, &bm_status);
 622         if (bm_status)
 623                 acpi_write_bit_register(ACPI_BITREG_BUS_MASTER_STATUS, 1);
 624         /*
 625          * PIIX4 Erratum #18: Note that BM_STS doesn't always reflect
 626          * the true state of bus mastering activity; forcing us to
 627          * manually check the BMIDEA bit of each IDE channel.
 628          */
 629         else if (errata.piix4.bmisx) {
 630                 if ((inb_p(errata.piix4.bmisx + 0x02) & 0x01)
 631                     || (inb_p(errata.piix4.bmisx + 0x0A) & 0x01))
 632                         bm_status = 1;
 633         }
 634         return bm_status;
 635 }
 636
 637 /**
 638  * acpi_idle_do_entry - enter idle state using the appropriate method
 639  * @cx: cstate data
 640  *
 641  * Caller disables interrupt before call and enables interrupt after return.
 642  */
 643 static void acpi_idle_do_entry(struct acpi_processor_cx *cx)
 644 {
 645         if (cx->entry_method == ACPI_CSTATE_FFH) {
 646                 /* Call into architectural FFH based C-state */
 647                 acpi_processor_ffh_cstate_enter(cx);
 648         } else if (cx->entry_method == ACPI_CSTATE_HALT) {
 649                 acpi_safe_halt();
 650         } else {
 651                 /* IO port based C-state */
 652                 inb(cx->address);
 653                 /* Dummy wait op - must do something useless after P_LVL2 read
 654                    because chipsets cannot guarantee that STPCLK# signal
 655                    gets asserted in time to freeze execution properly. */
 656                 inl(acpi_gbl_FADT.xpm_timer_block.address);
 657         }
 658 }
 659
 660 /**
 661  * acpi_idle_play_dead - enters an ACPI state for long-term idle (i.e. off-lining)
 662  * @dev: the target CPU
 663  * @index: the index of suggested state
 664  */
 665 static int acpi_idle_play_dead(struct cpuidle_device *dev, int index)
 666 {
 667         struct acpi_processor_cx *cx = per_cpu(acpi_cstate[index], dev->cpu);
 668
 669         ACPI_FLUSH_CPU_CACHE();
 670
 671         while (1) {
 672
 673                 if (cx->entry_method == ACPI_CSTATE_HALT)
 674                         safe_halt();
 675                 else if (cx->entry_method == ACPI_CSTATE_SYSTEMIO) {
 676                         inb(cx->address);
 677                         /* See comment in acpi_idle_do_entry() */
 678                         inl(acpi_gbl_FADT.xpm_timer_block.address);
 679                 } else
 680                         return -ENODEV;
 681         }
 682
 683         /* Never reached */
 684         return 0;
 685 }
 686
 687 static bool acpi_idle_fallback_to_c1(struct acpi_processor *pr)
 688 {
 689         return IS_ENABLED(CONFIG_HOTPLUG_CPU) && !pr->flags.has_cst &&
 690                 !(acpi_gbl_FADT.flags & ACPI_FADT_C2_MP_SUPPORTED);
 691 }
 692
 693 static int c3_cpu_count;
 694 static DEFINE_RAW_SPINLOCK(c3_lock);
 695
 696 /**
 697  * acpi_idle_enter_bm - enters C3 with proper BM handling
 698  * @pr: Target processor
 699  * @cx: Target state context
 700  * @timer_bc: Whether or not to change timer mode to broadcast
 701  */
 702 static void acpi_idle_enter_bm(struct acpi_processor *pr,
 703                                struct acpi_processor_cx *cx, bool timer_bc)
 704 {
 705         acpi_unlazy_tlb(smp_processor_id());
 706
 707         /*
 708          * Must be done before busmaster disable as we might need to
 709          * access HPET !
 710          */
 711         if (timer_bc)
 712                 lapic_timer_state_broadcast(pr, cx, 1);
 713
 714         /*
 715          * disable bus master
 716          * bm_check implies we need ARB_DIS
 717          * bm_control implies whether we can do ARB_DIS
 718          *
 719          * That leaves a case where bm_check is set and bm_control is
 720          * not set. In that case we cannot do much, we enter C3
 721          * without doing anything.
 722          */
 723         if (pr->flags.bm_control) {
 724                 raw_spin_lock(&c3_lock);
 725                 c3_cpu_count++;
 726                 /* Disable bus master arbitration when all CPUs are in C3 */
 727                 if (c3_cpu_count == num_online_cpus())
 728                         acpi_write_bit_register(ACPI_BITREG_ARB_DISABLE, 1);
 729                 raw_spin_unlock(&c3_lock);
 730         }
 731
 732         acpi_idle_do_entry(cx);
 733
 734         /* Re-enable bus master arbitration */
 735         if (pr->flags.bm_control) {
 736                 raw_spin_lock(&c3_lock);
 737                 acpi_write_bit_register(ACPI_BITREG_ARB_DISABLE, 0);
 738                 c3_cpu_count--;
 739                 raw_spin_unlock(&c3_lock);
 740         }
 741
 742         if (timer_bc)
 743                 lapic_timer_state_broadcast(pr, cx, 0);
 744 }
 745
 746 static int acpi_idle_enter(struct cpuidle_device *dev,
 747                            struct cpuidle_driver *drv, int index)
 748 {
 749         struct acpi_processor_cx *cx = per_cpu(acpi_cstate[index], dev->cpu);
 750         struct acpi_processor *pr;
 751
 752         pr = __this_cpu_read(processors);
 753         if (unlikely(!pr))
 754                 return -EINVAL;
 755
 756         if (cx->type != ACPI_STATE_C1) {
 757                 if (acpi_idle_fallback_to_c1(pr) && num_online_cpus() > 1) {
 758                         index = CPUIDLE_DRIVER_STATE_START;
 759                         cx = per_cpu(acpi_cstate[index], dev->cpu);
 760                 } else if (cx->type == ACPI_STATE_C3 && pr->flags.bm_check) {
 761                         if (cx->bm_sts_skip || !acpi_idle_bm_check()) {
 762                                 acpi_idle_enter_bm(pr, cx, true);
 763                                 return index;
 764                         } else if (drv->safe_state_index >= 0) {
 765                                 index = drv->safe_state_index;
 766                                 cx = per_cpu(acpi_cstate[index], dev->cpu);
 767                         } else {
 768                                 acpi_safe_halt();
 769                                 return -EBUSY;
 770                         }
 771                 }
 772         }
 773
 774         lapic_timer_state_broadcast(pr, cx, 1);
 775
 776         if (cx->type == ACPI_STATE_C3)
 777                 ACPI_FLUSH_CPU_CACHE();
 778
 779         acpi_idle_do_entry(cx);
 780
 781         lapic_timer_state_broadcast(pr, cx, 0);
 782
 783         return index;
 784 }
 785
 786 static void acpi_idle_enter_freeze(struct cpuidle_device *dev,
 787                                    struct cpuidle_driver *drv, int index)
 788 {
 789         struct acpi_processor_cx *cx = per_cpu(acpi_cstate[index], dev->cpu);
 790
 791         if (cx->type == ACPI_STATE_C3) {
 792                 struct acpi_processor *pr = __this_cpu_read(processors);
 793
 794                 if (unlikely(!pr))
 795                         return;
 796
 797                 if (pr->flags.bm_check) {
 798                         acpi_idle_enter_bm(pr, cx, false);
 799                         return;
 800                 } else {
 801                         ACPI_FLUSH_CPU_CACHE();
 802                 }
 803         }
 804         acpi_idle_do_entry(cx);
 805 }
 806
 807 struct cpuidle_driver acpi_idle_driver = {
 808         .name =         "acpi_idle",
 809         .owner =        THIS_MODULE,
 810 };
 811
 812 /**
 813  * acpi_processor_setup_cpuidle_cx - prepares and configures CPUIDLE
 814  * device i.e. per-cpu data
 815  *
 816  * @pr: the ACPI processor
 817  * @dev : the cpuidle device
 818  */
 819 static int acpi_processor_setup_cpuidle_cx(struct acpi_processor *pr,
 820                                            struct cpuidle_device *dev)
 821 {
 822         int i, count = CPUIDLE_DRIVER_STATE_START;
 823         struct acpi_processor_cx *cx;
 824
 825         if (!pr->flags.power_setup_done)
 826                 return -EINVAL;
 827
 828         if (pr->flags.power == 0) {
 829                 return -EINVAL;
 830         }
 831
 832         if (!dev)
 833                 return -EINVAL;
 834
 835         dev->cpu = pr->id;
 836
 837         if (max_cstate == 0)
 838                 max_cstate = 1;
 839
 840         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER && i <= max_cstate; i++) {
 841                 cx = &pr->power.states[i];
 842
 843                 if (!cx->valid)
 844                         continue;
 845
 846                 per_cpu(acpi_cstate[count], dev->cpu) = cx;
 847
 848                 count++;
 849                 if (count == CPUIDLE_STATE_MAX)
 850                         break;
 851         }
 852
 853         if (!count)
 854                 return -EINVAL;
 855
 856         return 0;
 857 }
 858
 859 /**
 860  * acpi_processor_setup_cpuidle states- prepares and configures cpuidle
 861  * global state data i.e. idle routines
 862  *
 863  * @pr: the ACPI processor
 864  */
 865 static int acpi_processor_setup_cpuidle_states(struct acpi_processor *pr)
 866 {
 867         int i, count = CPUIDLE_DRIVER_STATE_START;
 868         struct acpi_processor_cx *cx;
 869         struct cpuidle_state *state;
 870         struct cpuidle_driver *drv = &acpi_idle_driver;
 871
 872         if (!pr->flags.power_setup_done)
 873                 return -EINVAL;
 874
 875         if (pr->flags.power == 0)
 876                 return -EINVAL;
 877
 878         drv->safe_state_index = -1;
 879         for (i = CPUIDLE_DRIVER_STATE_START; i < CPUIDLE_STATE_MAX; i++) {
 880                 drv->states[i].name[0] = '\0';
 881                 drv->states[i].desc[0] = '\0';
 882         }
 883
 884         if (max_cstate == 0)
 885                 max_cstate = 1;
 886
 887         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER && i <= max_cstate; i++) {
 888                 cx = &pr->power.states[i];
 889
 890                 if (!cx->valid)
 891                         continue;
 892
 893                 state = &drv->states[count];
 894                 snprintf(state->name, CPUIDLE_NAME_LEN, "C%d", i);
 895                 strncpy(state->desc, cx->desc, CPUIDLE_DESC_LEN);
 896                 state->exit_latency = cx->latency;
 897                 state->target_residency = cx->latency * latency_factor;
 898                 state->enter = acpi_idle_enter;
 899
 900                 state->flags = 0;
 901                 if (cx->type == ACPI_STATE_C1 || cx->type == ACPI_STATE_C2) {
 902                         state->enter_dead = acpi_idle_play_dead;
 903                         drv->safe_state_index = count;
 904                 }
 905                 /*
 906                  * Halt-induced C1 is not good for ->enter_freeze, because it
 907                  * re-enables interrupts on exit.  Moreover, C1 is generally not
 908                  * particularly interesting from the suspend-to-idle angle, so
 909                  * avoid C1 and the situations in which we may need to fall back
 910                  * to it altogether.
 911                  */
 912                 if (cx->type != ACPI_STATE_C1 && !acpi_idle_fallback_to_c1(pr))
 913                         state->enter_freeze = acpi_idle_enter_freeze;
 914
 915                 count++;
 916                 if (count == CPUIDLE_STATE_MAX)
 917                         break;
 918         }
 919
 920         drv->state_count = count;
 921
 922         if (!count)
 923                 return -EINVAL;
 924
 925         return 0;
 926 }
 927
 928 int acpi_processor_hotplug(struct acpi_processor *pr)
 929 {
 930         int ret = 0;
 931         struct cpuidle_device *dev;
 932
 933         if (disabled_by_idle_boot_param())
 934                 return 0;
 935
 936         if (nocst)
 937                 return -ENODEV;
 938
 939         if (!pr->flags.power_setup_done)
 940                 return -ENODEV;
 941
 942         dev = per_cpu(acpi_cpuidle_device, pr->id);
 943         cpuidle_pause_and_lock();
 944         cpuidle_disable_device(dev);
 945         acpi_processor_get_power_info(pr);
 946         if (pr->flags.power) {
 947                 acpi_processor_setup_cpuidle_cx(pr, dev);
 948                 ret = cpuidle_enable_device(dev);
 949         }
 950         cpuidle_resume_and_unlock();
 951
 952         return ret;
 953 }
 954
 955 int acpi_processor_cst_has_changed(struct acpi_processor *pr)
 956 {
 957         int cpu;
 958         struct acpi_processor *_pr;
 959         struct cpuidle_device *dev;
 960
 961         if (disabled_by_idle_boot_param())
 962                 return 0;
 963
 964         if (nocst)
 965                 return -ENODEV;
 966
 967         if (!pr->flags.power_setup_done)
 968                 return -ENODEV;
 969
 970         /*
 971          * FIXME:  Design the ACPI notification to make it once per
 972          * system instead of once per-cpu.  This condition is a hack
 973          * to make the code that updates C-States be called once.
 974          */
 975
 976         if (pr->id == 0 && cpuidle_get_driver() == &acpi_idle_driver) {
 977
 978                 /* Protect against cpu-hotplug */
 979                 get_online_cpus();
 980                 cpuidle_pause_and_lock();
 981
 982                 /* Disable all cpuidle devices */
 983                 for_each_online_cpu(cpu) {
 984                         _pr = per_cpu(processors, cpu);
 985                         if (!_pr || !_pr->flags.power_setup_done)
 986                                 continue;
 987                         dev = per_cpu(acpi_cpuidle_device, cpu);
 988                         cpuidle_disable_device(dev);
 989                 }
 990
 991                 /* Populate Updated C-state information */
 992                 acpi_processor_get_power_info(pr);
 993                 acpi_processor_setup_cpuidle_states(pr);
 994
 995                 /* Enable all cpuidle devices */
 996                 for_each_online_cpu(cpu) {
 997                         _pr = per_cpu(processors, cpu);
 998                         if (!_pr || !_pr->flags.power_setup_done)
 999                                 continue;
1000                         acpi_processor_get_power_info(_pr);
1001                         if (_pr->flags.power) {
1002                                 dev = per_cpu(acpi_cpuidle_device, cpu);
1003                                 acpi_processor_setup_cpuidle_cx(_pr, dev);
1004                                 cpuidle_enable_device(dev);
1005                         }
1006                 }
1007                 cpuidle_resume_and_unlock();
1008                 put_online_cpus();
1009         }
1010
1011         return 0;
1012 }
1013
1014 static int acpi_processor_registered;
1015
1016 int acpi_processor_power_init(struct acpi_processor *pr)
1017 {
1018         acpi_status status;
1019         int retval;
1020         struct cpuidle_device *dev;
1021         static int first_run;
1022
1023         if (disabled_by_idle_boot_param())
1024                 return 0;
1025
1026         if (!first_run) {
1027                 dmi_check_system(processor_power_dmi_table);
1028                 max_cstate = acpi_processor_cstate_check(max_cstate);
1029                 if (max_cstate < ACPI_C_STATES_MAX)
1030                         printk(KERN_NOTICE
1031                                "ACPI: processor limited to max C-state %d\n",
1032                                max_cstate);
1033                 first_run++;
1034         }
1035
1036         if (acpi_gbl_FADT.cst_control && !nocst) {
1037                 status =
1038                     acpi_os_write_port(acpi_gbl_FADT.smi_command, acpi_gbl_FADT.cst_control, 8);
1039                 if (ACPI_FAILURE(status)) {
1040                         ACPI_EXCEPTION((AE_INFO, status,
1041                                         "Notifying BIOS of _CST ability failed"));
1042                 }
1043         }
1044
1045         acpi_processor_get_power_info(pr);
1046         pr->flags.power_setup_done = 1;
1047
1048         /*
1049          * Install the idle handler if processor power management is supported.
1050          * Note that we use previously set idle handler will be used on
1051          * platforms that only support C1.
1052          */
1053         if (pr->flags.power) {
1054                 /* Register acpi_idle_driver if not already registered */
1055                 if (!acpi_processor_registered) {
1056                         acpi_processor_setup_cpuidle_states(pr);
1057                         retval = cpuidle_register_driver(&acpi_idle_driver);
1058                         if (retval)
1059                                 return retval;
1060                         pr_debug("%s registered with cpuidle\n",
1061                                  acpi_idle_driver.name);
1062                 }
1063
1064                 dev = kzalloc(sizeof(*dev), GFP_KERNEL);
1065                 if (!dev)
1066                         return -ENOMEM;
1067                 per_cpu(acpi_cpuidle_device, pr->id) = dev;
1068
1069                 acpi_processor_setup_cpuidle_cx(pr, dev);
1070
1071                 /* Register per-cpu cpuidle_device. Cpuidle driver
1072                  * must already be registered before registering device
1073                  */
1074                 retval = cpuidle_register_device(dev);
1075                 if (retval) {
1076                         if (acpi_processor_registered == 0)
1077                                 cpuidle_unregister_driver(&acpi_idle_driver);
1078                         return retval;
1079                 }
1080                 acpi_processor_registered++;
1081         }
1082         return 0;
1083 }
1084
1085 int acpi_processor_power_exit(struct acpi_processor *pr)
1086 {
1087         struct cpuidle_device *dev = per_cpu(acpi_cpuidle_device, pr->id);
1088
1089         if (disabled_by_idle_boot_param())
1090                 return 0;
1091
1092         if (pr->flags.power) {
1093                 cpuidle_unregister_device(dev);
1094                 acpi_processor_registered--;
1095                 if (acpi_processor_registered == 0)
1096                         cpuidle_unregister_driver(&acpi_idle_driver);
1097         }
1098
1099         pr->flags.power_setup_done = 0;
1100         return 0;
1101 }