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