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
  27 #include <linux/module.h>
  28 #include <linux/acpi.h>
  29 #include <linux/dmi.h>
  30 #include <linux/sched.h>       /* need_resched() */
  31 #include <linux/sort.h>
  32 #include <linux/tick.h>
  33 #include <linux/cpuidle.h>
  34 #include <linux/syscore_ops.h>
  35 #include <acpi/processor.h>
  36
  37 /*
  38  * Include the apic definitions for x86 to have the APIC timer related defines
  39  * available also for UP (on SMP it gets magically included via linux/smp.h).
  40  * asm/acpi.h is not an option, as it would require more include magic. Also
  41  * creating an empty asm-ia64/apic.h would just trade pest vs. cholera.
  42  */
  43 #ifdef CONFIG_X86
  44 #include <asm/apic.h>
  45 #endif
  46
  47 #define PREFIX "ACPI: "
  48
  49 #define ACPI_PROCESSOR_CLASS            "processor"
  50 #define _COMPONENT              ACPI_PROCESSOR_COMPONENT
  51 ACPI_MODULE_NAME("processor_idle");
  52
  53 static unsigned int max_cstate __read_mostly = ACPI_PROCESSOR_MAX_POWER;
  54 module_param(max_cstate, uint, 0000);
  55 static unsigned int nocst __read_mostly;
  56 module_param(nocst, uint, 0000);
  57 static int bm_check_disable __read_mostly;
  58 module_param(bm_check_disable, uint, 0000);
  59
  60 static unsigned int latency_factor __read_mostly = 2;
  61 module_param(latency_factor, uint, 0644);
  62
  63 static DEFINE_PER_CPU(struct cpuidle_device *, acpi_cpuidle_device);
  64
  65 static DEFINE_PER_CPU(struct acpi_processor_cx * [CPUIDLE_STATE_MAX],
  66                                                                 acpi_cstate);
  67
  68 static int disabled_by_idle_boot_param(void)
  69 {
  70         return boot_option_idle_override == IDLE_POLL ||
  71                 boot_option_idle_override == IDLE_HALT;
  72 }
  73
  74 /*
  75  * IBM ThinkPad R40e crashes mysteriously when going into C2 or C3.
  76  * For now disable this. Probably a bug somewhere else.
  77  *
  78  * To skip this limit, boot/load with a large max_cstate limit.
  79  */
  80 static int set_max_cstate(const struct dmi_system_id *id)
  81 {
  82         if (max_cstate > ACPI_PROCESSOR_MAX_POWER)
  83                 return 0;
  84
  85         printk(KERN_NOTICE PREFIX "%s detected - limiting to C%ld max_cstate."
  86                " Override with \"processor.max_cstate=%d\"\n", id->ident,
  87                (long)id->driver_data, ACPI_PROCESSOR_MAX_POWER + 1);
  88
  89         max_cstate = (long)id->driver_data;
  90
  91         return 0;
  92 }
  93
  94 static const struct dmi_system_id processor_power_dmi_table[] = {
  95         { set_max_cstate, "Clevo 5600D", {
  96           DMI_MATCH(DMI_BIOS_VENDOR,"Phoenix Technologies LTD"),
  97           DMI_MATCH(DMI_BIOS_VERSION,"SHE845M0.86C.0013.D.0302131307")},
  98          (void *)2},
  99         { set_max_cstate, "Pavilion zv5000", {
 100           DMI_MATCH(DMI_SYS_VENDOR, "Hewlett-Packard"),
 101           DMI_MATCH(DMI_PRODUCT_NAME,"Pavilion zv5000 (DS502A#ABA)")},
 102          (void *)1},
 103         { set_max_cstate, "Asus L8400B", {
 104           DMI_MATCH(DMI_SYS_VENDOR, "ASUSTeK Computer Inc."),
 105           DMI_MATCH(DMI_PRODUCT_NAME,"L8400B series Notebook PC")},
 106          (void *)1},
 107         {},
 108 };
 109
 110
 111 /*
 112  * Callers should disable interrupts before the call and enable
 113  * interrupts after return.
 114  */
 115 static void acpi_safe_halt(void)
 116 {
 117         if (!tif_need_resched()) {
 118                 safe_halt();
 119                 local_irq_disable();
 120         }
 121 }
 122
 123 #ifdef ARCH_APICTIMER_STOPS_ON_C3
 124
 125 /*
 126  * Some BIOS implementations switch to C3 in the published C2 state.
 127  * This seems to be a common problem on AMD boxen, but other vendors
 128  * are affected too. We pick the most conservative approach: we assume
 129  * that the local APIC stops in both C2 and C3.
 130  */
 131 static void lapic_timer_check_state(int state, struct acpi_processor *pr,
 132                                    struct acpi_processor_cx *cx)
 133 {
 134         struct acpi_processor_power *pwr = &pr->power;
 135         u8 type = local_apic_timer_c2_ok ? ACPI_STATE_C3 : ACPI_STATE_C2;
 136
 137         if (cpu_has(&cpu_data(pr->id), X86_FEATURE_ARAT))
 138                 return;
 139
 140         if (amd_e400_c1e_detected)
 141                 type = ACPI_STATE_C1;
 142
 143         /*
 144          * Check, if one of the previous states already marked the lapic
 145          * unstable
 146          */
 147         if (pwr->timer_broadcast_on_state < state)
 148                 return;
 149
 150         if (cx->type >= type)
 151                 pr->power.timer_broadcast_on_state = state;
 152 }
 153
 154 static void __lapic_timer_propagate_broadcast(void *arg)
 155 {
 156         struct acpi_processor *pr = (struct acpi_processor *) arg;
 157
 158         if (pr->power.timer_broadcast_on_state < INT_MAX)
 159                 tick_broadcast_enable();
 160         else
 161                 tick_broadcast_disable();
 162 }
 163
 164 static void lapic_timer_propagate_broadcast(struct acpi_processor *pr)
 165 {
 166         smp_call_function_single(pr->id, __lapic_timer_propagate_broadcast,
 167                                  (void *)pr, 1);
 168 }
 169
 170 /* Power(C) State timer broadcast control */
 171 static void lapic_timer_state_broadcast(struct acpi_processor *pr,
 172                                        struct acpi_processor_cx *cx,
 173                                        int broadcast)
 174 {
 175         int state = cx - pr->power.states;
 176
 177         if (state >= pr->power.timer_broadcast_on_state) {
 178                 if (broadcast)
 179                         tick_broadcast_enter();
 180                 else
 181                         tick_broadcast_exit();
 182         }
 183 }
 184
 185 #else
 186
 187 static void lapic_timer_check_state(int state, struct acpi_processor *pr,
 188                                    struct acpi_processor_cx *cstate) { }
 189 static void lapic_timer_propagate_broadcast(struct acpi_processor *pr) { }
 190 static void lapic_timer_state_broadcast(struct acpi_processor *pr,
 191                                        struct acpi_processor_cx *cx,
 192                                        int broadcast)
 193 {
 194 }
 195
 196 #endif
 197
 198 #ifdef CONFIG_PM_SLEEP
 199 static u32 saved_bm_rld;
 200
 201 static int acpi_processor_suspend(void)
 202 {
 203         acpi_read_bit_register(ACPI_BITREG_BUS_MASTER_RLD, &saved_bm_rld);
 204         return 0;
 205 }
 206
 207 static void acpi_processor_resume(void)
 208 {
 209         u32 resumed_bm_rld = 0;
 210
 211         acpi_read_bit_register(ACPI_BITREG_BUS_MASTER_RLD, &resumed_bm_rld);
 212         if (resumed_bm_rld == saved_bm_rld)
 213                 return;
 214
 215         acpi_write_bit_register(ACPI_BITREG_BUS_MASTER_RLD, saved_bm_rld);
 216 }
 217
 218 static struct syscore_ops acpi_processor_syscore_ops = {
 219         .suspend = acpi_processor_suspend,
 220         .resume = acpi_processor_resume,
 221 };
 222
 223 void acpi_processor_syscore_init(void)
 224 {
 225         register_syscore_ops(&acpi_processor_syscore_ops);
 226 }
 227
 228 void acpi_processor_syscore_exit(void)
 229 {
 230         unregister_syscore_ops(&acpi_processor_syscore_ops);
 231 }
 232 #endif /* CONFIG_PM_SLEEP */
 233
 234 #if defined(CONFIG_X86)
 235 static void tsc_check_state(int state)
 236 {
 237         switch (boot_cpu_data.x86_vendor) {
 238         case X86_VENDOR_AMD:
 239         case X86_VENDOR_INTEL:
 240                 /*
 241                  * AMD Fam10h TSC will tick in all
 242                  * C/P/S0/S1 states when this bit is set.
 243                  */
 244                 if (boot_cpu_has(X86_FEATURE_NONSTOP_TSC))
 245                         return;
 246
 247                 /*FALL THROUGH*/
 248         default:
 249                 /* TSC could halt in idle, so notify users */
 250                 if (state > ACPI_STATE_C1)
 251                         mark_tsc_unstable("TSC halts in idle");
 252         }
 253 }
 254 #else
 255 static void tsc_check_state(int state) { return; }
 256 #endif
 257
 258 static int acpi_processor_get_power_info_fadt(struct acpi_processor *pr)
 259 {
 260
 261         if (!pr->pblk)
 262                 return -ENODEV;
 263
 264         /* if info is obtained from pblk/fadt, type equals state */
 265         pr->power.states[ACPI_STATE_C2].type = ACPI_STATE_C2;
 266         pr->power.states[ACPI_STATE_C3].type = ACPI_STATE_C3;
 267
 268 #ifndef CONFIG_HOTPLUG_CPU
 269         /*
 270          * Check for P_LVL2_UP flag before entering C2 and above on
 271          * an SMP system.
 272          */
 273         if ((num_online_cpus() > 1) &&
 274             !(acpi_gbl_FADT.flags & ACPI_FADT_C2_MP_SUPPORTED))
 275                 return -ENODEV;
 276 #endif
 277
 278         /* determine C2 and C3 address from pblk */
 279         pr->power.states[ACPI_STATE_C2].address = pr->pblk + 4;
 280         pr->power.states[ACPI_STATE_C3].address = pr->pblk + 5;
 281
 282         /* determine latencies from FADT */
 283         pr->power.states[ACPI_STATE_C2].latency = acpi_gbl_FADT.c2_latency;
 284         pr->power.states[ACPI_STATE_C3].latency = acpi_gbl_FADT.c3_latency;
 285
 286         /*
 287          * FADT specified C2 latency must be less than or equal to
 288          * 100 microseconds.
 289          */
 290         if (acpi_gbl_FADT.c2_latency > ACPI_PROCESSOR_MAX_C2_LATENCY) {
 291                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 292                         "C2 latency too large [%d]\n", acpi_gbl_FADT.c2_latency));
 293                 /* invalidate C2 */
 294                 pr->power.states[ACPI_STATE_C2].address = 0;
 295         }
 296
 297         /*
 298          * FADT supplied C3 latency must be less than or equal to
 299          * 1000 microseconds.
 300          */
 301         if (acpi_gbl_FADT.c3_latency > ACPI_PROCESSOR_MAX_C3_LATENCY) {
 302                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 303                         "C3 latency too large [%d]\n", acpi_gbl_FADT.c3_latency));
 304                 /* invalidate C3 */
 305                 pr->power.states[ACPI_STATE_C3].address = 0;
 306         }
 307
 308         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 309                           "lvl2[0x%08x] lvl3[0x%08x]\n",
 310                           pr->power.states[ACPI_STATE_C2].address,
 311                           pr->power.states[ACPI_STATE_C3].address));
 312
 313         return 0;
 314 }
 315
 316 static int acpi_processor_get_power_info_default(struct acpi_processor *pr)
 317 {
 318         if (!pr->power.states[ACPI_STATE_C1].valid) {
 319                 /* set the first C-State to C1 */
 320                 /* all processors need to support C1 */
 321                 pr->power.states[ACPI_STATE_C1].type = ACPI_STATE_C1;
 322                 pr->power.states[ACPI_STATE_C1].valid = 1;
 323                 pr->power.states[ACPI_STATE_C1].entry_method = ACPI_CSTATE_HALT;
 324         }
 325         /* the C0 state only exists as a filler in our array */
 326         pr->power.states[ACPI_STATE_C0].valid = 1;
 327         return 0;
 328 }
 329
 330 static int acpi_processor_get_power_info_cst(struct acpi_processor *pr)
 331 {
 332         acpi_status status;
 333         u64 count;
 334         int current_count;
 335         int i, ret = 0;
 336         struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
 337         union acpi_object *cst;
 338
 339
 340         if (nocst)
 341                 return -ENODEV;
 342
 343         current_count = 0;
 344
 345         status = acpi_evaluate_object(pr->handle, "_CST", NULL, &buffer);
 346         if (ACPI_FAILURE(status)) {
 347                 ACPI_DEBUG_PRINT((ACPI_DB_INFO, "No _CST, giving up\n"));
 348                 return -ENODEV;
 349         }
 350
 351         cst = buffer.pointer;
 352
 353         /* There must be at least 2 elements */
 354         if (!cst || (cst->type != ACPI_TYPE_PACKAGE) || cst->package.count < 2) {
 355                 printk(KERN_ERR PREFIX "not enough elements in _CST\n");
 356                 ret = -EFAULT;
 357                 goto end;
 358         }
 359
 360         count = cst->package.elements[0].integer.value;
 361
 362         /* Validate number of power states. */
 363         if (count < 1 || count != cst->package.count - 1) {
 364                 printk(KERN_ERR PREFIX "count given by _CST is not valid\n");
 365                 ret = -EFAULT;
 366                 goto end;
 367         }
 368
 369         /* Tell driver that at least _CST is supported. */
 370         pr->flags.has_cst = 1;
 371
 372         for (i = 1; i <= count; i++) {
 373                 union acpi_object *element;
 374                 union acpi_object *obj;
 375                 struct acpi_power_register *reg;
 376                 struct acpi_processor_cx cx;
 377
 378                 memset(&cx, 0, sizeof(cx));
 379
 380                 element = &(cst->package.elements[i]);
 381                 if (element->type != ACPI_TYPE_PACKAGE)
 382                         continue;
 383
 384                 if (element->package.count != 4)
 385                         continue;
 386
 387                 obj = &(element->package.elements[0]);
 388
 389                 if (obj->type != ACPI_TYPE_BUFFER)
 390                         continue;
 391
 392                 reg = (struct acpi_power_register *)obj->buffer.pointer;
 393
 394                 if (reg->space_id != ACPI_ADR_SPACE_SYSTEM_IO &&
 395                     (reg->space_id != ACPI_ADR_SPACE_FIXED_HARDWARE))
 396                         continue;
 397
 398                 /* There should be an easy way to extract an integer... */
 399                 obj = &(element->package.elements[1]);
 400                 if (obj->type != ACPI_TYPE_INTEGER)
 401                         continue;
 402
 403                 cx.type = obj->integer.value;
 404                 /*
 405                  * Some buggy BIOSes won't list C1 in _CST -
 406                  * Let acpi_processor_get_power_info_default() handle them later
 407                  */
 408                 if (i == 1 && cx.type != ACPI_STATE_C1)
 409                         current_count++;
 410
 411                 cx.address = reg->address;
 412                 cx.index = current_count + 1;
 413
 414                 cx.entry_method = ACPI_CSTATE_SYSTEMIO;
 415                 if (reg->space_id == ACPI_ADR_SPACE_FIXED_HARDWARE) {
 416                         if (acpi_processor_ffh_cstate_probe
 417                                         (pr->id, &cx, reg) == 0) {
 418                                 cx.entry_method = ACPI_CSTATE_FFH;
 419                         } else if (cx.type == ACPI_STATE_C1) {
 420                                 /*
 421                                  * C1 is a special case where FIXED_HARDWARE
 422                                  * can be handled in non-MWAIT way as well.
 423                                  * In that case, save this _CST entry info.
 424                                  * Otherwise, ignore this info and continue.
 425                                  */
 426                                 cx.entry_method = ACPI_CSTATE_HALT;
 427                                 snprintf(cx.desc, ACPI_CX_DESC_LEN, "ACPI HLT");
 428                         } else {
 429                                 continue;
 430                         }
 431                         if (cx.type == ACPI_STATE_C1 &&
 432                             (boot_option_idle_override == IDLE_NOMWAIT)) {
 433                                 /*
 434                                  * In most cases the C1 space_id obtained from
 435                                  * _CST object is FIXED_HARDWARE access mode.
 436                                  * But when the option of idle=halt is added,
 437                                  * the entry_method type should be changed from
 438                                  * CSTATE_FFH to CSTATE_HALT.
 439                                  * When the option of idle=nomwait is added,
 440                                  * the C1 entry_method type should be
 441                                  * CSTATE_HALT.
 442                                  */
 443                                 cx.entry_method = ACPI_CSTATE_HALT;
 444                                 snprintf(cx.desc, ACPI_CX_DESC_LEN, "ACPI HLT");
 445                         }
 446                 } else {
 447                         snprintf(cx.desc, ACPI_CX_DESC_LEN, "ACPI IOPORT 0x%x",
 448                                  cx.address);
 449                 }
 450
 451                 if (cx.type == ACPI_STATE_C1) {
 452                         cx.valid = 1;
 453                 }
 454
 455                 obj = &(element->package.elements[2]);
 456                 if (obj->type != ACPI_TYPE_INTEGER)
 457                         continue;
 458
 459                 cx.latency = obj->integer.value;
 460
 461                 obj = &(element->package.elements[3]);
 462                 if (obj->type != ACPI_TYPE_INTEGER)
 463                         continue;
 464
 465                 current_count++;
 466                 memcpy(&(pr->power.states[current_count]), &cx, sizeof(cx));
 467
 468                 /*
 469                  * We support total ACPI_PROCESSOR_MAX_POWER - 1
 470                  * (From 1 through ACPI_PROCESSOR_MAX_POWER - 1)
 471                  */
 472                 if (current_count >= (ACPI_PROCESSOR_MAX_POWER - 1)) {
 473                         printk(KERN_WARNING
 474                                "Limiting number of power states to max (%d)\n",
 475                                ACPI_PROCESSOR_MAX_POWER);
 476                         printk(KERN_WARNING
 477                                "Please increase ACPI_PROCESSOR_MAX_POWER if needed.\n");
 478                         break;
 479                 }
 480         }
 481
 482         ACPI_DEBUG_PRINT((ACPI_DB_INFO, "Found %d power states\n",
 483                           current_count));
 484
 485         /* Validate number of power states discovered */
 486         if (current_count < 2)
 487                 ret = -EFAULT;
 488
 489       end:
 490         kfree(buffer.pointer);
 491
 492         return ret;
 493 }
 494
 495 static void acpi_processor_power_verify_c3(struct acpi_processor *pr,
 496                                            struct acpi_processor_cx *cx)
 497 {
 498         static int bm_check_flag = -1;
 499         static int bm_control_flag = -1;
 500
 501
 502         if (!cx->address)
 503                 return;
 504
 505         /*
 506          * PIIX4 Erratum #18: We don't support C3 when Type-F (fast)
 507          * DMA transfers are used by any ISA device to avoid livelock.
 508          * Note that we could disable Type-F DMA (as recommended by
 509          * the erratum), but this is known to disrupt certain ISA
 510          * devices thus we take the conservative approach.
 511          */
 512         else if (errata.piix4.fdma) {
 513                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 514                                   "C3 not supported on PIIX4 with Type-F DMA\n"));
 515                 return;
 516         }
 517
 518         /* All the logic here assumes flags.bm_check is same across all CPUs */
 519         if (bm_check_flag == -1) {
 520                 /* Determine whether bm_check is needed based on CPU  */
 521                 acpi_processor_power_init_bm_check(&(pr->flags), pr->id);
 522                 bm_check_flag = pr->flags.bm_check;
 523                 bm_control_flag = pr->flags.bm_control;
 524         } else {
 525                 pr->flags.bm_check = bm_check_flag;
 526                 pr->flags.bm_control = bm_control_flag;
 527         }
 528
 529         if (pr->flags.bm_check) {
 530                 if (!pr->flags.bm_control) {
 531                         if (pr->flags.has_cst != 1) {
 532                                 /* bus mastering control is necessary */
 533                                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 534                                         "C3 support requires BM control\n"));
 535                                 return;
 536                         } else {
 537                                 /* Here we enter C3 without bus mastering */
 538                                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 539                                         "C3 support without BM control\n"));
 540                         }
 541                 }
 542         } else {
 543                 /*
 544                  * WBINVD should be set in fadt, for C3 state to be
 545                  * supported on when bm_check is not required.
 546                  */
 547                 if (!(acpi_gbl_FADT.flags & ACPI_FADT_WBINVD)) {
 548                         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 549                                           "Cache invalidation should work properly"
 550                                           " for C3 to be enabled on SMP systems\n"));
 551                         return;
 552                 }
 553         }
 554
 555         /*
 556          * Otherwise we've met all of our C3 requirements.
 557          * Normalize the C3 latency to expidite policy.  Enable
 558          * checking of bus mastering status (bm_check) so we can
 559          * use this in our C3 policy
 560          */
 561         cx->valid = 1;
 562
 563         /*
 564          * On older chipsets, BM_RLD needs to be set
 565          * in order for Bus Master activity to wake the
 566          * system from C3.  Newer chipsets handle DMA
 567          * during C3 automatically and BM_RLD is a NOP.
 568          * In either case, the proper way to
 569          * handle BM_RLD is to set it and leave it set.
 570          */
 571         acpi_write_bit_register(ACPI_BITREG_BUS_MASTER_RLD, 1);
 572
 573         return;
 574 }
 575
 576 static int acpi_cst_latency_cmp(const void *a, const void *b)
 577 {
 578         const struct acpi_processor_cx *x = a, *y = b;
 579
 580         if (!(x->valid && y->valid))
 581                 return 0;
 582         if (x->latency > y->latency)
 583                 return 1;
 584         if (x->latency < y->latency)
 585                 return -1;
 586         return 0;
 587 }
 588 static void acpi_cst_latency_swap(void *a, void *b, int n)
 589 {
 590         struct acpi_processor_cx *x = a, *y = b;
 591         u32 tmp;
 592
 593         if (!(x->valid && y->valid))
 594                 return;
 595         tmp = x->latency;
 596         x->latency = y->latency;
 597         y->latency = tmp;
 598 }
 599
 600 static int acpi_processor_power_verify(struct acpi_processor *pr)
 601 {
 602         unsigned int i;
 603         unsigned int working = 0;
 604         unsigned int last_latency = 0;
 605         unsigned int last_type = 0;
 606         bool buggy_latency = false;
 607
 608         pr->power.timer_broadcast_on_state = INT_MAX;
 609
 610         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER && i <= max_cstate; i++) {
 611                 struct acpi_processor_cx *cx = &pr->power.states[i];
 612
 613                 switch (cx->type) {
 614                 case ACPI_STATE_C1:
 615                         cx->valid = 1;
 616                         break;
 617
 618                 case ACPI_STATE_C2:
 619                         if (!cx->address)
 620                                 break;
 621                         cx->valid = 1;
 622                         break;
 623
 624                 case ACPI_STATE_C3:
 625                         acpi_processor_power_verify_c3(pr, cx);
 626                         break;
 627                 }
 628                 if (!cx->valid)
 629                         continue;
 630                 if (cx->type >= last_type && cx->latency < last_latency)
 631                         buggy_latency = true;
 632                 last_latency = cx->latency;
 633                 last_type = cx->type;
 634
 635                 lapic_timer_check_state(i, pr, cx);
 636                 tsc_check_state(cx->type);
 637                 working++;
 638         }
 639
 640         if (buggy_latency) {
 641                 pr_notice("FW issue: working around C-state latencies out of order\n");
 642                 sort(&pr->power.states[1], max_cstate,
 643                      sizeof(struct acpi_processor_cx),
 644                      acpi_cst_latency_cmp,
 645                      acpi_cst_latency_swap);
 646         }
 647
 648         lapic_timer_propagate_broadcast(pr);
 649
 650         return (working);
 651 }
 652
 653 static int acpi_processor_get_power_info(struct acpi_processor *pr)
 654 {
 655         unsigned int i;
 656         int result;
 657
 658
 659         /* NOTE: the idle thread may not be running while calling
 660          * this function */
 661
 662         /* Zero initialize all the C-states info. */
 663         memset(pr->power.states, 0, sizeof(pr->power.states));
 664
 665         result = acpi_processor_get_power_info_cst(pr);
 666         if (result == -ENODEV)
 667                 result = acpi_processor_get_power_info_fadt(pr);
 668
 669         if (result)
 670                 return result;
 671
 672         acpi_processor_get_power_info_default(pr);
 673
 674         pr->power.count = acpi_processor_power_verify(pr);
 675
 676         /*
 677          * if one state of type C2 or C3 is available, mark this
 678          * CPU as being "idle manageable"
 679          */
 680         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
 681                 if (pr->power.states[i].valid) {
 682                         pr->power.count = i;
 683                         if (pr->power.states[i].type >= ACPI_STATE_C2)
 684                                 pr->flags.power = 1;
 685                 }
 686         }
 687
 688         return 0;
 689 }
 690
 691 /**
 692  * acpi_idle_bm_check - checks if bus master activity was detected
 693  */
 694 static int acpi_idle_bm_check(void)
 695 {
 696         u32 bm_status = 0;
 697
 698         if (bm_check_disable)
 699                 return 0;
 700
 701         acpi_read_bit_register(ACPI_BITREG_BUS_MASTER_STATUS, &bm_status);
 702         if (bm_status)
 703                 acpi_write_bit_register(ACPI_BITREG_BUS_MASTER_STATUS, 1);
 704         /*
 705          * PIIX4 Erratum #18: Note that BM_STS doesn't always reflect
 706          * the true state of bus mastering activity; forcing us to
 707          * manually check the BMIDEA bit of each IDE channel.
 708          */
 709         else if (errata.piix4.bmisx) {
 710                 if ((inb_p(errata.piix4.bmisx + 0x02) & 0x01)
 711                     || (inb_p(errata.piix4.bmisx + 0x0A) & 0x01))
 712                         bm_status = 1;
 713         }
 714         return bm_status;
 715 }
 716
 717 /**
 718  * acpi_idle_do_entry - enter idle state using the appropriate method
 719  * @cx: cstate data
 720  *
 721  * Caller disables interrupt before call and enables interrupt after return.
 722  */
 723 static void acpi_idle_do_entry(struct acpi_processor_cx *cx)
 724 {
 725         if (cx->entry_method == ACPI_CSTATE_FFH) {
 726                 /* Call into architectural FFH based C-state */
 727                 acpi_processor_ffh_cstate_enter(cx);
 728         } else if (cx->entry_method == ACPI_CSTATE_HALT) {
 729                 acpi_safe_halt();
 730         } else {
 731                 /* IO port based C-state */
 732                 inb(cx->address);
 733                 /* Dummy wait op - must do something useless after P_LVL2 read
 734                    because chipsets cannot guarantee that STPCLK# signal
 735                    gets asserted in time to freeze execution properly. */
 736                 inl(acpi_gbl_FADT.xpm_timer_block.address);
 737         }
 738 }
 739
 740 /**
 741  * acpi_idle_play_dead - enters an ACPI state for long-term idle (i.e. off-lining)
 742  * @dev: the target CPU
 743  * @index: the index of suggested state
 744  */
 745 static int acpi_idle_play_dead(struct cpuidle_device *dev, int index)
 746 {
 747         struct acpi_processor_cx *cx = per_cpu(acpi_cstate[index], dev->cpu);
 748
 749         ACPI_FLUSH_CPU_CACHE();
 750
 751         while (1) {
 752
 753                 if (cx->entry_method == ACPI_CSTATE_HALT)
 754                         safe_halt();
 755                 else if (cx->entry_method == ACPI_CSTATE_SYSTEMIO) {
 756                         inb(cx->address);
 757                         /* See comment in acpi_idle_do_entry() */
 758                         inl(acpi_gbl_FADT.xpm_timer_block.address);
 759                 } else
 760                         return -ENODEV;
 761         }
 762
 763         /* Never reached */
 764         return 0;
 765 }
 766
 767 static bool acpi_idle_fallback_to_c1(struct acpi_processor *pr)
 768 {
 769         return IS_ENABLED(CONFIG_HOTPLUG_CPU) && !pr->flags.has_cst &&
 770                 !(acpi_gbl_FADT.flags & ACPI_FADT_C2_MP_SUPPORTED);
 771 }
 772
 773 static int c3_cpu_count;
 774 static DEFINE_RAW_SPINLOCK(c3_lock);
 775
 776 /**
 777  * acpi_idle_enter_bm - enters C3 with proper BM handling
 778  * @pr: Target processor
 779  * @cx: Target state context
 780  * @timer_bc: Whether or not to change timer mode to broadcast
 781  */
 782 static void acpi_idle_enter_bm(struct acpi_processor *pr,
 783                                struct acpi_processor_cx *cx, bool timer_bc)
 784 {
 785         acpi_unlazy_tlb(smp_processor_id());
 786
 787         /*
 788          * Must be done before busmaster disable as we might need to
 789          * access HPET !
 790          */
 791         if (timer_bc)
 792                 lapic_timer_state_broadcast(pr, cx, 1);
 793
 794         /*
 795          * disable bus master
 796          * bm_check implies we need ARB_DIS
 797          * bm_control implies whether we can do ARB_DIS
 798          *
 799          * That leaves a case where bm_check is set and bm_control is
 800          * not set. In that case we cannot do much, we enter C3
 801          * without doing anything.
 802          */
 803         if (pr->flags.bm_control) {
 804                 raw_spin_lock(&c3_lock);
 805                 c3_cpu_count++;
 806                 /* Disable bus master arbitration when all CPUs are in C3 */
 807                 if (c3_cpu_count == num_online_cpus())
 808                         acpi_write_bit_register(ACPI_BITREG_ARB_DISABLE, 1);
 809                 raw_spin_unlock(&c3_lock);
 810         }
 811
 812         acpi_idle_do_entry(cx);
 813
 814         /* Re-enable bus master arbitration */
 815         if (pr->flags.bm_control) {
 816                 raw_spin_lock(&c3_lock);
 817                 acpi_write_bit_register(ACPI_BITREG_ARB_DISABLE, 0);
 818                 c3_cpu_count--;
 819                 raw_spin_unlock(&c3_lock);
 820         }
 821
 822         if (timer_bc)
 823                 lapic_timer_state_broadcast(pr, cx, 0);
 824 }
 825
 826 static int acpi_idle_enter(struct cpuidle_device *dev,
 827                            struct cpuidle_driver *drv, int index)
 828 {
 829         struct acpi_processor_cx *cx = per_cpu(acpi_cstate[index], dev->cpu);
 830         struct acpi_processor *pr;
 831
 832         pr = __this_cpu_read(processors);
 833         if (unlikely(!pr))
 834                 return -EINVAL;
 835
 836         if (cx->type != ACPI_STATE_C1) {
 837                 if (acpi_idle_fallback_to_c1(pr) && num_online_cpus() > 1) {
 838                         index = CPUIDLE_DRIVER_STATE_START;
 839                         cx = per_cpu(acpi_cstate[index], dev->cpu);
 840                 } else if (cx->type == ACPI_STATE_C3 && pr->flags.bm_check) {
 841                         if (cx->bm_sts_skip || !acpi_idle_bm_check()) {
 842                                 acpi_idle_enter_bm(pr, cx, true);
 843                                 return index;
 844                         } else if (drv->safe_state_index >= 0) {
 845                                 index = drv->safe_state_index;
 846                                 cx = per_cpu(acpi_cstate[index], dev->cpu);
 847                         } else {
 848                                 acpi_safe_halt();
 849                                 return -EBUSY;
 850                         }
 851                 }
 852         }
 853
 854         lapic_timer_state_broadcast(pr, cx, 1);
 855
 856         if (cx->type == ACPI_STATE_C3)
 857                 ACPI_FLUSH_CPU_CACHE();
 858
 859         acpi_idle_do_entry(cx);
 860
 861         lapic_timer_state_broadcast(pr, cx, 0);
 862
 863         return index;
 864 }
 865
 866 static void acpi_idle_enter_freeze(struct cpuidle_device *dev,
 867                                    struct cpuidle_driver *drv, int index)
 868 {
 869         struct acpi_processor_cx *cx = per_cpu(acpi_cstate[index], dev->cpu);
 870
 871         if (cx->type == ACPI_STATE_C3) {
 872                 struct acpi_processor *pr = __this_cpu_read(processors);
 873
 874                 if (unlikely(!pr))
 875                         return;
 876
 877                 if (pr->flags.bm_check) {
 878                         acpi_idle_enter_bm(pr, cx, false);
 879                         return;
 880                 } else {
 881                         ACPI_FLUSH_CPU_CACHE();
 882                 }
 883         }
 884         acpi_idle_do_entry(cx);
 885 }
 886
 887 struct cpuidle_driver acpi_idle_driver = {
 888         .name =         "acpi_idle",
 889         .owner =        THIS_MODULE,
 890 };
 891
 892 /**
 893  * acpi_processor_setup_cpuidle_cx - prepares and configures CPUIDLE
 894  * device i.e. per-cpu data
 895  *
 896  * @pr: the ACPI processor
 897  * @dev : the cpuidle device
 898  */
 899 static int acpi_processor_setup_cpuidle_cx(struct acpi_processor *pr,
 900                                            struct cpuidle_device *dev)
 901 {
 902         int i, count = CPUIDLE_DRIVER_STATE_START;
 903         struct acpi_processor_cx *cx;
 904
 905         if (!pr->flags.power_setup_done)
 906                 return -EINVAL;
 907
 908         if (pr->flags.power == 0) {
 909                 return -EINVAL;
 910         }
 911
 912         if (!dev)
 913                 return -EINVAL;
 914
 915         dev->cpu = pr->id;
 916
 917         if (max_cstate == 0)
 918                 max_cstate = 1;
 919
 920         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER && i <= max_cstate; i++) {
 921                 cx = &pr->power.states[i];
 922
 923                 if (!cx->valid)
 924                         continue;
 925
 926                 per_cpu(acpi_cstate[count], dev->cpu) = cx;
 927
 928                 count++;
 929                 if (count == CPUIDLE_STATE_MAX)
 930                         break;
 931         }
 932
 933         if (!count)
 934                 return -EINVAL;
 935
 936         return 0;
 937 }
 938
 939 /**
 940  * acpi_processor_setup_cpuidle states- prepares and configures cpuidle
 941  * global state data i.e. idle routines
 942  *
 943  * @pr: the ACPI processor
 944  */
 945 static int acpi_processor_setup_cpuidle_states(struct acpi_processor *pr)
 946 {
 947         int i, count = CPUIDLE_DRIVER_STATE_START;
 948         struct acpi_processor_cx *cx;
 949         struct cpuidle_state *state;
 950         struct cpuidle_driver *drv = &acpi_idle_driver;
 951
 952         if (!pr->flags.power_setup_done)
 953                 return -EINVAL;
 954
 955         if (pr->flags.power == 0)
 956                 return -EINVAL;
 957
 958         drv->safe_state_index = -1;
 959         for (i = CPUIDLE_DRIVER_STATE_START; i < CPUIDLE_STATE_MAX; i++) {
 960                 drv->states[i].name[0] = '\0';
 961                 drv->states[i].desc[0] = '\0';
 962         }
 963
 964         if (max_cstate == 0)
 965                 max_cstate = 1;
 966
 967         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER && i <= max_cstate; i++) {
 968                 cx = &pr->power.states[i];
 969
 970                 if (!cx->valid)
 971                         continue;
 972
 973                 state = &drv->states[count];
 974                 snprintf(state->name, CPUIDLE_NAME_LEN, "C%d", i);
 975                 strncpy(state->desc, cx->desc, CPUIDLE_DESC_LEN);
 976                 state->exit_latency = cx->latency;
 977                 state->target_residency = cx->latency * latency_factor;
 978                 state->enter = acpi_idle_enter;
 979
 980                 state->flags = 0;
 981                 if (cx->type == ACPI_STATE_C1 || cx->type == ACPI_STATE_C2) {
 982                         state->enter_dead = acpi_idle_play_dead;
 983                         drv->safe_state_index = count;
 984                 }
 985                 /*
 986                  * Halt-induced C1 is not good for ->enter_freeze, because it
 987                  * re-enables interrupts on exit.  Moreover, C1 is generally not
 988                  * particularly interesting from the suspend-to-idle angle, so
 989                  * avoid C1 and the situations in which we may need to fall back
 990                  * to it altogether.
 991                  */
 992                 if (cx->type != ACPI_STATE_C1 && !acpi_idle_fallback_to_c1(pr))
 993                         state->enter_freeze = acpi_idle_enter_freeze;
 994
 995                 count++;
 996                 if (count == CPUIDLE_STATE_MAX)
 997                         break;
 998         }
 999
1000         drv->state_count = count;
1001
1002         if (!count)
1003                 return -EINVAL;
1004
1005         return 0;
1006 }
1007
1008 int acpi_processor_hotplug(struct acpi_processor *pr)
1009 {
1010         int ret = 0;
1011         struct cpuidle_device *dev;
1012
1013         if (disabled_by_idle_boot_param())
1014                 return 0;
1015
1016         if (nocst)
1017                 return -ENODEV;
1018
1019         if (!pr->flags.power_setup_done)
1020                 return -ENODEV;
1021
1022         dev = per_cpu(acpi_cpuidle_device, pr->id);
1023         cpuidle_pause_and_lock();
1024         cpuidle_disable_device(dev);
1025         acpi_processor_get_power_info(pr);
1026         if (pr->flags.power) {
1027                 acpi_processor_setup_cpuidle_cx(pr, dev);
1028                 ret = cpuidle_enable_device(dev);
1029         }
1030         cpuidle_resume_and_unlock();
1031
1032         return ret;
1033 }
1034
1035 int acpi_processor_cst_has_changed(struct acpi_processor *pr)
1036 {
1037         int cpu;
1038         struct acpi_processor *_pr;
1039         struct cpuidle_device *dev;
1040
1041         if (disabled_by_idle_boot_param())
1042                 return 0;
1043
1044         if (nocst)
1045                 return -ENODEV;
1046
1047         if (!pr->flags.power_setup_done)
1048                 return -ENODEV;
1049
1050         /*
1051          * FIXME:  Design the ACPI notification to make it once per
1052          * system instead of once per-cpu.  This condition is a hack
1053          * to make the code that updates C-States be called once.
1054          */
1055
1056         if (pr->id == 0 && cpuidle_get_driver() == &acpi_idle_driver) {
1057
1058                 /* Protect against cpu-hotplug */
1059                 get_online_cpus();
1060                 cpuidle_pause_and_lock();
1061
1062                 /* Disable all cpuidle devices */
1063                 for_each_online_cpu(cpu) {
1064                         _pr = per_cpu(processors, cpu);
1065                         if (!_pr || !_pr->flags.power_setup_done)
1066                                 continue;
1067                         dev = per_cpu(acpi_cpuidle_device, cpu);
1068                         cpuidle_disable_device(dev);
1069                 }
1070
1071                 /* Populate Updated C-state information */
1072                 acpi_processor_get_power_info(pr);
1073                 acpi_processor_setup_cpuidle_states(pr);
1074
1075                 /* Enable all cpuidle devices */
1076                 for_each_online_cpu(cpu) {
1077                         _pr = per_cpu(processors, cpu);
1078                         if (!_pr || !_pr->flags.power_setup_done)
1079                                 continue;
1080                         acpi_processor_get_power_info(_pr);
1081                         if (_pr->flags.power) {
1082                                 dev = per_cpu(acpi_cpuidle_device, cpu);
1083                                 acpi_processor_setup_cpuidle_cx(_pr, dev);
1084                                 cpuidle_enable_device(dev);
1085                         }
1086                 }
1087                 cpuidle_resume_and_unlock();
1088                 put_online_cpus();
1089         }
1090
1091         return 0;
1092 }
1093
1094 static int acpi_processor_registered;
1095
1096 int acpi_processor_power_init(struct acpi_processor *pr)
1097 {
1098         acpi_status status;
1099         int retval;
1100         struct cpuidle_device *dev;
1101         static int first_run;
1102
1103         if (disabled_by_idle_boot_param())
1104                 return 0;
1105
1106         if (!first_run) {
1107                 dmi_check_system(processor_power_dmi_table);
1108                 max_cstate = acpi_processor_cstate_check(max_cstate);
1109                 if (max_cstate < ACPI_C_STATES_MAX)
1110                         printk(KERN_NOTICE
1111                                "ACPI: processor limited to max C-state %d\n",
1112                                max_cstate);
1113                 first_run++;
1114         }
1115
1116         if (acpi_gbl_FADT.cst_control && !nocst) {
1117                 status =
1118                     acpi_os_write_port(acpi_gbl_FADT.smi_command, acpi_gbl_FADT.cst_control, 8);
1119                 if (ACPI_FAILURE(status)) {
1120                         ACPI_EXCEPTION((AE_INFO, status,
1121                                         "Notifying BIOS of _CST ability failed"));
1122                 }
1123         }
1124
1125         acpi_processor_get_power_info(pr);
1126         pr->flags.power_setup_done = 1;
1127
1128         /*
1129          * Install the idle handler if processor power management is supported.
1130          * Note that we use previously set idle handler will be used on
1131          * platforms that only support C1.
1132          */
1133         if (pr->flags.power) {
1134                 /* Register acpi_idle_driver if not already registered */
1135                 if (!acpi_processor_registered) {
1136                         acpi_processor_setup_cpuidle_states(pr);
1137                         retval = cpuidle_register_driver(&acpi_idle_driver);
1138                         if (retval)
1139                                 return retval;
1140                         printk(KERN_DEBUG "ACPI: %s registered with cpuidle\n",
1141                                         acpi_idle_driver.name);
1142                 }
1143
1144                 dev = kzalloc(sizeof(*dev), GFP_KERNEL);
1145                 if (!dev)
1146                         return -ENOMEM;
1147                 per_cpu(acpi_cpuidle_device, pr->id) = dev;
1148
1149                 acpi_processor_setup_cpuidle_cx(pr, dev);
1150
1151                 /* Register per-cpu cpuidle_device. Cpuidle driver
1152                  * must already be registered before registering device
1153                  */
1154                 retval = cpuidle_register_device(dev);
1155                 if (retval) {
1156                         if (acpi_processor_registered == 0)
1157                                 cpuidle_unregister_driver(&acpi_idle_driver);
1158                         return retval;
1159                 }
1160                 acpi_processor_registered++;
1161         }
1162         return 0;
1163 }
1164
1165 int acpi_processor_power_exit(struct acpi_processor *pr)
1166 {
1167         struct cpuidle_device *dev = per_cpu(acpi_cpuidle_device, pr->id);
1168
1169         if (disabled_by_idle_boot_param())
1170                 return 0;
1171
1172         if (pr->flags.power) {
1173                 cpuidle_unregister_device(dev);
1174                 acpi_processor_registered--;
1175                 if (acpi_processor_registered == 0)
1176                         cpuidle_unregister_driver(&acpi_idle_driver);
1177         }
1178
1179         pr->flags.power_setup_done = 0;
1180         return 0;
1181 }