drivers/acpi/processor_idle.c

   1 /*
   2  * processor_idle - idle state submodule to the ACPI processor driver
   3  *
   4  *  Copyright (C) 2001, 2002 Andy Grover <andrew.grover@intel.com>
   5  *  Copyright (C) 2001, 2002 Paul Diefenbaugh <paul.s.diefenbaugh@intel.com>
   6  *  Copyright (C) 2004, 2005 Dominik Brodowski <linux@brodo.de>
   7  *  Copyright (C) 2004  Anil S Keshavamurthy <anil.s.keshavamurthy@intel.com>
   8  *                      - Added processor hotplug support
   9  *  Copyright (C) 2005  Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>
  10  *                      - Added support for C3 on SMP
  11  *
  12  * ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
  13  *
  14  *  This program is free software; you can redistribute it and/or modify
  15  *  it under the terms of the GNU General Public License as published by
  16  *  the Free Software Foundation; either version 2 of the License, or (at
  17  *  your option) any later version.
  18  *
  19  *  This program is distributed in the hope that it will be useful, but
  20  *  WITHOUT ANY WARRANTY; without even the implied warranty of
  21  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  22  *  General Public License for more details.
  23  *
  24  *  You should have received a copy of the GNU General Public License along
  25  *  with this program; if not, write to the Free Software Foundation, Inc.,
  26  *  59 Temple Place, Suite 330, Boston, MA 02111-1307 USA.
  27  *
  28  * ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
  29  */
  30
  31 #include <linux/module.h>
  32 #include <linux/acpi.h>
  33 #include <linux/dmi.h>
  34 #include <linux/sched.h>       /* need_resched() */
  35 #include <linux/tick.h>
  36 #include <linux/cpuidle.h>
  37 #include <linux/syscore_ops.h>
  38 #include <acpi/processor.h>
  39
  40 /*
  41  * Include the apic definitions for x86 to have the APIC timer related defines
  42  * available also for UP (on SMP it gets magically included via linux/smp.h).
  43  * asm/acpi.h is not an option, as it would require more include magic. Also
  44  * creating an empty asm-ia64/apic.h would just trade pest vs. cholera.
  45  */
  46 #ifdef CONFIG_X86
  47 #include <asm/apic.h>
  48 #endif
  49
  50 #define PREFIX "ACPI: "
  51
  52 #define ACPI_PROCESSOR_CLASS            "processor"
  53 #define _COMPONENT              ACPI_PROCESSOR_COMPONENT
  54 ACPI_MODULE_NAME("processor_idle");
  55
  56 static unsigned int max_cstate __read_mostly = ACPI_PROCESSOR_MAX_POWER;
  57 module_param(max_cstate, uint, 0000);
  58 static unsigned int nocst __read_mostly;
  59 module_param(nocst, uint, 0000);
  60 static int bm_check_disable __read_mostly;
  61 module_param(bm_check_disable, uint, 0000);
  62
  63 static unsigned int latency_factor __read_mostly = 2;
  64 module_param(latency_factor, uint, 0644);
  65
  66 static DEFINE_PER_CPU(struct cpuidle_device *, acpi_cpuidle_device);
  67
  68 static DEFINE_PER_CPU(struct acpi_processor_cx * [CPUIDLE_STATE_MAX],
  69                                                                 acpi_cstate);
  70
  71 static int disabled_by_idle_boot_param(void)
  72 {
  73         return boot_option_idle_override == IDLE_POLL ||
  74                 boot_option_idle_override == IDLE_HALT;
  75 }
  76
  77 /*
  78  * IBM ThinkPad R40e crashes mysteriously when going into C2 or C3.
  79  * For now disable this. Probably a bug somewhere else.
  80  *
  81  * To skip this limit, boot/load with a large max_cstate limit.
  82  */
  83 static int set_max_cstate(const struct dmi_system_id *id)
  84 {
  85         if (max_cstate > ACPI_PROCESSOR_MAX_POWER)
  86                 return 0;
  87
  88         printk(KERN_NOTICE PREFIX "%s detected - limiting to C%ld max_cstate."
  89                " Override with \"processor.max_cstate=%d\"\n", id->ident,
  90                (long)id->driver_data, ACPI_PROCESSOR_MAX_POWER + 1);
  91
  92         max_cstate = (long)id->driver_data;
  93
  94         return 0;
  95 }
  96
  97 static struct dmi_system_id processor_power_dmi_table[] = {
  98         { set_max_cstate, "Clevo 5600D", {
  99           DMI_MATCH(DMI_BIOS_VENDOR,"Phoenix Technologies LTD"),
 100           DMI_MATCH(DMI_BIOS_VERSION,"SHE845M0.86C.0013.D.0302131307")},
 101          (void *)2},
 102         { set_max_cstate, "Pavilion zv5000", {
 103           DMI_MATCH(DMI_SYS_VENDOR, "Hewlett-Packard"),
 104           DMI_MATCH(DMI_PRODUCT_NAME,"Pavilion zv5000 (DS502A#ABA)")},
 105          (void *)1},
 106         { set_max_cstate, "Asus L8400B", {
 107           DMI_MATCH(DMI_SYS_VENDOR, "ASUSTeK Computer Inc."),
 108           DMI_MATCH(DMI_PRODUCT_NAME,"L8400B series Notebook PC")},
 109          (void *)1},
 110         {},
 111 };
 112
 113
 114 /*
 115  * Callers should disable interrupts before the call and enable
 116  * interrupts after return.
 117  */
 118 static void acpi_safe_halt(void)
 119 {
 120         if (!tif_need_resched()) {
 121                 safe_halt();
 122                 local_irq_disable();
 123         }
 124 }
 125
 126 #ifdef ARCH_APICTIMER_STOPS_ON_C3
 127
 128 /*
 129  * Some BIOS implementations switch to C3 in the published C2 state.
 130  * This seems to be a common problem on AMD boxen, but other vendors
 131  * are affected too. We pick the most conservative approach: we assume
 132  * that the local APIC stops in both C2 and C3.
 133  */
 134 static void lapic_timer_check_state(int state, struct acpi_processor *pr,
 135                                    struct acpi_processor_cx *cx)
 136 {
 137         struct acpi_processor_power *pwr = &pr->power;
 138         u8 type = local_apic_timer_c2_ok ? ACPI_STATE_C3 : ACPI_STATE_C2;
 139
 140         if (cpu_has(&cpu_data(pr->id), X86_FEATURE_ARAT))
 141                 return;
 142
 143         if (amd_e400_c1e_detected)
 144                 type = ACPI_STATE_C1;
 145
 146         /*
 147          * Check, if one of the previous states already marked the lapic
 148          * unstable
 149          */
 150         if (pwr->timer_broadcast_on_state < state)
 151                 return;
 152
 153         if (cx->type >= type)
 154                 pr->power.timer_broadcast_on_state = state;
 155 }
 156
 157 static void __lapic_timer_propagate_broadcast(void *arg)
 158 {
 159         struct acpi_processor *pr = (struct acpi_processor *) arg;
 160
 161         if (pr->power.timer_broadcast_on_state < INT_MAX)
 162                 tick_broadcast_enable();
 163         else
 164                 tick_broadcast_disable();
 165 }
 166
 167 static void lapic_timer_propagate_broadcast(struct acpi_processor *pr)
 168 {
 169         smp_call_function_single(pr->id, __lapic_timer_propagate_broadcast,
 170                                  (void *)pr, 1);
 171 }
 172
 173 /* Power(C) State timer broadcast control */
 174 static void lapic_timer_state_broadcast(struct acpi_processor *pr,
 175                                        struct acpi_processor_cx *cx,
 176                                        int broadcast)
 177 {
 178         int state = cx - pr->power.states;
 179
 180         if (state >= pr->power.timer_broadcast_on_state) {
 181                 if (broadcast)
 182                         tick_broadcast_enter();
 183                 else
 184                         tick_broadcast_exit();
 185         }
 186 }
 187
 188 #else
 189
 190 static void lapic_timer_check_state(int state, struct acpi_processor *pr,
 191                                    struct acpi_processor_cx *cstate) { }
 192 static void lapic_timer_propagate_broadcast(struct acpi_processor *pr) { }
 193 static void lapic_timer_state_broadcast(struct acpi_processor *pr,
 194                                        struct acpi_processor_cx *cx,
 195                                        int broadcast)
 196 {
 197 }
 198
 199 #endif
 200
 201 #ifdef CONFIG_PM_SLEEP
 202 static u32 saved_bm_rld;
 203
 204 static int acpi_processor_suspend(void)
 205 {
 206         acpi_read_bit_register(ACPI_BITREG_BUS_MASTER_RLD, &saved_bm_rld);
 207         return 0;
 208 }
 209
 210 static void acpi_processor_resume(void)
 211 {
 212         u32 resumed_bm_rld = 0;
 213
 214         acpi_read_bit_register(ACPI_BITREG_BUS_MASTER_RLD, &resumed_bm_rld);
 215         if (resumed_bm_rld == saved_bm_rld)
 216                 return;
 217
 218         acpi_write_bit_register(ACPI_BITREG_BUS_MASTER_RLD, saved_bm_rld);
 219 }
 220
 221 static struct syscore_ops acpi_processor_syscore_ops = {
 222         .suspend = acpi_processor_suspend,
 223         .resume = acpi_processor_resume,
 224 };
 225
 226 void acpi_processor_syscore_init(void)
 227 {
 228         register_syscore_ops(&acpi_processor_syscore_ops);
 229 }
 230
 231 void acpi_processor_syscore_exit(void)
 232 {
 233         unregister_syscore_ops(&acpi_processor_syscore_ops);
 234 }
 235 #endif /* CONFIG_PM_SLEEP */
 236
 237 #if defined(CONFIG_X86)
 238 static void tsc_check_state(int state)
 239 {
 240         switch (boot_cpu_data.x86_vendor) {
 241         case X86_VENDOR_AMD:
 242         case X86_VENDOR_INTEL:
 243                 /*
 244                  * AMD Fam10h TSC will tick in all
 245                  * C/P/S0/S1 states when this bit is set.
 246                  */
 247                 if (boot_cpu_has(X86_FEATURE_NONSTOP_TSC))
 248                         return;
 249
 250                 /*FALL THROUGH*/
 251         default:
 252                 /* TSC could halt in idle, so notify users */
 253                 if (state > ACPI_STATE_C1)
 254                         mark_tsc_unstable("TSC halts in idle");
 255         }
 256 }
 257 #else
 258 static void tsc_check_state(int state) { return; }
 259 #endif
 260
 261 static int acpi_processor_get_power_info_fadt(struct acpi_processor *pr)
 262 {
 263
 264         if (!pr->pblk)
 265                 return -ENODEV;
 266
 267         /* if info is obtained from pblk/fadt, type equals state */
 268         pr->power.states[ACPI_STATE_C2].type = ACPI_STATE_C2;
 269         pr->power.states[ACPI_STATE_C3].type = ACPI_STATE_C3;
 270
 271 #ifndef CONFIG_HOTPLUG_CPU
 272         /*
 273          * Check for P_LVL2_UP flag before entering C2 and above on
 274          * an SMP system.
 275          */
 276         if ((num_online_cpus() > 1) &&
 277             !(acpi_gbl_FADT.flags & ACPI_FADT_C2_MP_SUPPORTED))
 278                 return -ENODEV;
 279 #endif
 280
 281         /* determine C2 and C3 address from pblk */
 282         pr->power.states[ACPI_STATE_C2].address = pr->pblk + 4;
 283         pr->power.states[ACPI_STATE_C3].address = pr->pblk + 5;
 284
 285         /* determine latencies from FADT */
 286         pr->power.states[ACPI_STATE_C2].latency = acpi_gbl_FADT.c2_latency;
 287         pr->power.states[ACPI_STATE_C3].latency = acpi_gbl_FADT.c3_latency;
 288
 289         /*
 290          * FADT specified C2 latency must be less than or equal to
 291          * 100 microseconds.
 292          */
 293         if (acpi_gbl_FADT.c2_latency > ACPI_PROCESSOR_MAX_C2_LATENCY) {
 294                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 295                         "C2 latency too large [%d]\n", acpi_gbl_FADT.c2_latency));
 296                 /* invalidate C2 */
 297                 pr->power.states[ACPI_STATE_C2].address = 0;
 298         }
 299
 300         /*
 301          * FADT supplied C3 latency must be less than or equal to
 302          * 1000 microseconds.
 303          */
 304         if (acpi_gbl_FADT.c3_latency > ACPI_PROCESSOR_MAX_C3_LATENCY) {
 305                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 306                         "C3 latency too large [%d]\n", acpi_gbl_FADT.c3_latency));
 307                 /* invalidate C3 */
 308                 pr->power.states[ACPI_STATE_C3].address = 0;
 309         }
 310
 311         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 312                           "lvl2[0x%08x] lvl3[0x%08x]\n",
 313                           pr->power.states[ACPI_STATE_C2].address,
 314                           pr->power.states[ACPI_STATE_C3].address));
 315
 316         return 0;
 317 }
 318
 319 static int acpi_processor_get_power_info_default(struct acpi_processor *pr)
 320 {
 321         if (!pr->power.states[ACPI_STATE_C1].valid) {
 322                 /* set the first C-State to C1 */
 323                 /* all processors need to support C1 */
 324                 pr->power.states[ACPI_STATE_C1].type = ACPI_STATE_C1;
 325                 pr->power.states[ACPI_STATE_C1].valid = 1;
 326                 pr->power.states[ACPI_STATE_C1].entry_method = ACPI_CSTATE_HALT;
 327         }
 328         /* the C0 state only exists as a filler in our array */
 329         pr->power.states[ACPI_STATE_C0].valid = 1;
 330         return 0;
 331 }
 332
 333 static int acpi_processor_get_power_info_cst(struct acpi_processor *pr)
 334 {
 335         acpi_status status;
 336         u64 count;
 337         int current_count;
 338         int i, ret = 0;
 339         struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
 340         union acpi_object *cst;
 341
 342
 343         if (nocst)
 344                 return -ENODEV;
 345
 346         current_count = 0;
 347
 348         status = acpi_evaluate_object(pr->handle, "_CST", NULL, &buffer);
 349         if (ACPI_FAILURE(status)) {
 350                 ACPI_DEBUG_PRINT((ACPI_DB_INFO, "No _CST, giving up\n"));
 351                 return -ENODEV;
 352         }
 353
 354         cst = buffer.pointer;
 355
 356         /* There must be at least 2 elements */
 357         if (!cst || (cst->type != ACPI_TYPE_PACKAGE) || cst->package.count < 2) {
 358                 printk(KERN_ERR PREFIX "not enough elements in _CST\n");
 359                 ret = -EFAULT;
 360                 goto end;
 361         }
 362
 363         count = cst->package.elements[0].integer.value;
 364
 365         /* Validate number of power states. */
 366         if (count < 1 || count != cst->package.count - 1) {
 367                 printk(KERN_ERR PREFIX "count given by _CST is not valid\n");
 368                 ret = -EFAULT;
 369                 goto end;
 370         }
 371
 372         /* Tell driver that at least _CST is supported. */
 373         pr->flags.has_cst = 1;
 374
 375         for (i = 1; i <= count; i++) {
 376                 union acpi_object *element;
 377                 union acpi_object *obj;
 378                 struct acpi_power_register *reg;
 379                 struct acpi_processor_cx cx;
 380
 381                 memset(&cx, 0, sizeof(cx));
 382
 383                 element = &(cst->package.elements[i]);
 384                 if (element->type != ACPI_TYPE_PACKAGE)
 385                         continue;
 386
 387                 if (element->package.count != 4)
 388                         continue;
 389
 390                 obj = &(element->package.elements[0]);
 391
 392                 if (obj->type != ACPI_TYPE_BUFFER)
 393                         continue;
 394
 395                 reg = (struct acpi_power_register *)obj->buffer.pointer;
 396
 397                 if (reg->space_id != ACPI_ADR_SPACE_SYSTEM_IO &&
 398                     (reg->space_id != ACPI_ADR_SPACE_FIXED_HARDWARE))
 399                         continue;
 400
 401                 /* There should be an easy way to extract an integer... */
 402                 obj = &(element->package.elements[1]);
 403                 if (obj->type != ACPI_TYPE_INTEGER)
 404                         continue;
 405
 406                 cx.type = obj->integer.value;
 407                 /*
 408                  * Some buggy BIOSes won't list C1 in _CST -
 409                  * Let acpi_processor_get_power_info_default() handle them later
 410                  */
 411                 if (i == 1 && cx.type != ACPI_STATE_C1)
 412                         current_count++;
 413
 414                 cx.address = reg->address;
 415                 cx.index = current_count + 1;
 416
 417                 cx.entry_method = ACPI_CSTATE_SYSTEMIO;
 418                 if (reg->space_id == ACPI_ADR_SPACE_FIXED_HARDWARE) {
 419                         if (acpi_processor_ffh_cstate_probe
 420                                         (pr->id, &cx, reg) == 0) {
 421                                 cx.entry_method = ACPI_CSTATE_FFH;
 422                         } else if (cx.type == ACPI_STATE_C1) {
 423                                 /*
 424                                  * C1 is a special case where FIXED_HARDWARE
 425                                  * can be handled in non-MWAIT way as well.
 426                                  * In that case, save this _CST entry info.
 427                                  * Otherwise, ignore this info and continue.
 428                                  */
 429                                 cx.entry_method = ACPI_CSTATE_HALT;
 430                                 snprintf(cx.desc, ACPI_CX_DESC_LEN, "ACPI HLT");
 431                         } else {
 432                                 continue;
 433                         }
 434                         if (cx.type == ACPI_STATE_C1 &&
 435                             (boot_option_idle_override == IDLE_NOMWAIT)) {
 436                                 /*
 437                                  * In most cases the C1 space_id obtained from
 438                                  * _CST object is FIXED_HARDWARE access mode.
 439                                  * But when the option of idle=halt is added,
 440                                  * the entry_method type should be changed from
 441                                  * CSTATE_FFH to CSTATE_HALT.
 442                                  * When the option of idle=nomwait is added,
 443                                  * the C1 entry_method type should be
 444                                  * CSTATE_HALT.
 445                                  */
 446                                 cx.entry_method = ACPI_CSTATE_HALT;
 447                                 snprintf(cx.desc, ACPI_CX_DESC_LEN, "ACPI HLT");
 448                         }
 449                 } else {
 450                         snprintf(cx.desc, ACPI_CX_DESC_LEN, "ACPI IOPORT 0x%x",
 451                                  cx.address);
 452                 }
 453
 454                 if (cx.type == ACPI_STATE_C1) {
 455                         cx.valid = 1;
 456                 }
 457
 458                 obj = &(element->package.elements[2]);
 459                 if (obj->type != ACPI_TYPE_INTEGER)
 460                         continue;
 461
 462                 cx.latency = obj->integer.value;
 463
 464                 obj = &(element->package.elements[3]);
 465                 if (obj->type != ACPI_TYPE_INTEGER)
 466                         continue;
 467
 468                 current_count++;
 469                 memcpy(&(pr->power.states[current_count]), &cx, sizeof(cx));
 470
 471                 /*
 472                  * We support total ACPI_PROCESSOR_MAX_POWER - 1
 473                  * (From 1 through ACPI_PROCESSOR_MAX_POWER - 1)
 474                  */
 475                 if (current_count >= (ACPI_PROCESSOR_MAX_POWER - 1)) {
 476                         printk(KERN_WARNING
 477                                "Limiting number of power states to max (%d)\n",
 478                                ACPI_PROCESSOR_MAX_POWER);
 479                         printk(KERN_WARNING
 480                                "Please increase ACPI_PROCESSOR_MAX_POWER if needed.\n");
 481                         break;
 482                 }
 483         }
 484
 485         ACPI_DEBUG_PRINT((ACPI_DB_INFO, "Found %d power states\n",
 486                           current_count));
 487
 488         /* Validate number of power states discovered */
 489         if (current_count < 2)
 490                 ret = -EFAULT;
 491
 492       end:
 493         kfree(buffer.pointer);
 494
 495         return ret;
 496 }
 497
 498 static void acpi_processor_power_verify_c3(struct acpi_processor *pr,
 499                                            struct acpi_processor_cx *cx)
 500 {
 501         static int bm_check_flag = -1;
 502         static int bm_control_flag = -1;
 503
 504
 505         if (!cx->address)
 506                 return;
 507
 508         /*
 509          * PIIX4 Erratum #18: We don't support C3 when Type-F (fast)
 510          * DMA transfers are used by any ISA device to avoid livelock.
 511          * Note that we could disable Type-F DMA (as recommended by
 512          * the erratum), but this is known to disrupt certain ISA
 513          * devices thus we take the conservative approach.
 514          */
 515         else if (errata.piix4.fdma) {
 516                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 517                                   "C3 not supported on PIIX4 with Type-F DMA\n"));
 518                 return;
 519         }
 520
 521         /* All the logic here assumes flags.bm_check is same across all CPUs */
 522         if (bm_check_flag == -1) {
 523                 /* Determine whether bm_check is needed based on CPU  */
 524                 acpi_processor_power_init_bm_check(&(pr->flags), pr->id);
 525                 bm_check_flag = pr->flags.bm_check;
 526                 bm_control_flag = pr->flags.bm_control;
 527         } else {
 528                 pr->flags.bm_check = bm_check_flag;
 529                 pr->flags.bm_control = bm_control_flag;
 530         }
 531
 532         if (pr->flags.bm_check) {
 533                 if (!pr->flags.bm_control) {
 534                         if (pr->flags.has_cst != 1) {
 535                                 /* bus mastering control is necessary */
 536                                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 537                                         "C3 support requires BM control\n"));
 538                                 return;
 539                         } else {
 540                                 /* Here we enter C3 without bus mastering */
 541                                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 542                                         "C3 support without BM control\n"));
 543                         }
 544                 }
 545         } else {
 546                 /*
 547                  * WBINVD should be set in fadt, for C3 state to be
 548                  * supported on when bm_check is not required.
 549                  */
 550                 if (!(acpi_gbl_FADT.flags & ACPI_FADT_WBINVD)) {
 551                         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 552                                           "Cache invalidation should work properly"
 553                                           " for C3 to be enabled on SMP systems\n"));
 554                         return;
 555                 }
 556         }
 557
 558         /*
 559          * Otherwise we've met all of our C3 requirements.
 560          * Normalize the C3 latency to expidite policy.  Enable
 561          * checking of bus mastering status (bm_check) so we can
 562          * use this in our C3 policy
 563          */
 564         cx->valid = 1;
 565
 566         /*
 567          * On older chipsets, BM_RLD needs to be set
 568          * in order for Bus Master activity to wake the
 569          * system from C3.  Newer chipsets handle DMA
 570          * during C3 automatically and BM_RLD is a NOP.
 571          * In either case, the proper way to
 572          * handle BM_RLD is to set it and leave it set.
 573          */
 574         acpi_write_bit_register(ACPI_BITREG_BUS_MASTER_RLD, 1);
 575
 576         return;
 577 }
 578
 579 static int acpi_processor_power_verify(struct acpi_processor *pr)
 580 {
 581         unsigned int i;
 582         unsigned int working = 0;
 583
 584         pr->power.timer_broadcast_on_state = INT_MAX;
 585
 586         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER && i <= max_cstate; i++) {
 587                 struct acpi_processor_cx *cx = &pr->power.states[i];
 588
 589                 switch (cx->type) {
 590                 case ACPI_STATE_C1:
 591                         cx->valid = 1;
 592                         break;
 593
 594                 case ACPI_STATE_C2:
 595                         if (!cx->address)
 596                                 break;
 597                         cx->valid = 1;
 598                         break;
 599
 600                 case ACPI_STATE_C3:
 601                         acpi_processor_power_verify_c3(pr, cx);
 602                         break;
 603                 }
 604                 if (!cx->valid)
 605                         continue;
 606
 607                 lapic_timer_check_state(i, pr, cx);
 608                 tsc_check_state(cx->type);
 609                 working++;
 610         }
 611
 612         lapic_timer_propagate_broadcast(pr);
 613
 614         return (working);
 615 }
 616
 617 static int acpi_processor_get_power_info(struct acpi_processor *pr)
 618 {
 619         unsigned int i;
 620         int result;
 621
 622
 623         /* NOTE: the idle thread may not be running while calling
 624          * this function */
 625
 626         /* Zero initialize all the C-states info. */
 627         memset(pr->power.states, 0, sizeof(pr->power.states));
 628
 629         result = acpi_processor_get_power_info_cst(pr);
 630         if (result == -ENODEV)
 631                 result = acpi_processor_get_power_info_fadt(pr);
 632
 633         if (result)
 634                 return result;
 635
 636         acpi_processor_get_power_info_default(pr);
 637
 638         pr->power.count = acpi_processor_power_verify(pr);
 639
 640         /*
 641          * if one state of type C2 or C3 is available, mark this
 642          * CPU as being "idle manageable"
 643          */
 644         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
 645                 if (pr->power.states[i].valid) {
 646                         pr->power.count = i;
 647                         if (pr->power.states[i].type >= ACPI_STATE_C2)
 648                                 pr->flags.power = 1;
 649                 }
 650         }
 651
 652         return 0;
 653 }
 654
 655 /**
 656  * acpi_idle_bm_check - checks if bus master activity was detected
 657  */
 658 static int acpi_idle_bm_check(void)
 659 {
 660         u32 bm_status = 0;
 661
 662         if (bm_check_disable)
 663                 return 0;
 664
 665         acpi_read_bit_register(ACPI_BITREG_BUS_MASTER_STATUS, &bm_status);
 666         if (bm_status)
 667                 acpi_write_bit_register(ACPI_BITREG_BUS_MASTER_STATUS, 1);
 668         /*
 669          * PIIX4 Erratum #18: Note that BM_STS doesn't always reflect
 670          * the true state of bus mastering activity; forcing us to
 671          * manually check the BMIDEA bit of each IDE channel.
 672          */
 673         else if (errata.piix4.bmisx) {
 674                 if ((inb_p(errata.piix4.bmisx + 0x02) & 0x01)
 675                     || (inb_p(errata.piix4.bmisx + 0x0A) & 0x01))
 676                         bm_status = 1;
 677         }
 678         return bm_status;
 679 }
 680
 681 /**
 682  * acpi_idle_do_entry - enter idle state using the appropriate method
 683  * @cx: cstate data
 684  *
 685  * Caller disables interrupt before call and enables interrupt after return.
 686  */
 687 static void acpi_idle_do_entry(struct acpi_processor_cx *cx)
 688 {
 689         if (cx->entry_method == ACPI_CSTATE_FFH) {
 690                 /* Call into architectural FFH based C-state */
 691                 acpi_processor_ffh_cstate_enter(cx);
 692         } else if (cx->entry_method == ACPI_CSTATE_HALT) {
 693                 acpi_safe_halt();
 694         } else {
 695                 /* IO port based C-state */
 696                 inb(cx->address);
 697                 /* Dummy wait op - must do something useless after P_LVL2 read
 698                    because chipsets cannot guarantee that STPCLK# signal
 699                    gets asserted in time to freeze execution properly. */
 700                 inl(acpi_gbl_FADT.xpm_timer_block.address);
 701         }
 702 }
 703
 704 /**
 705  * acpi_idle_play_dead - enters an ACPI state for long-term idle (i.e. off-lining)
 706  * @dev: the target CPU
 707  * @index: the index of suggested state
 708  */
 709 static int acpi_idle_play_dead(struct cpuidle_device *dev, int index)
 710 {
 711         struct acpi_processor_cx *cx = per_cpu(acpi_cstate[index], dev->cpu);
 712
 713         ACPI_FLUSH_CPU_CACHE();
 714
 715         while (1) {
 716
 717                 if (cx->entry_method == ACPI_CSTATE_HALT)
 718                         safe_halt();
 719                 else if (cx->entry_method == ACPI_CSTATE_SYSTEMIO) {
 720                         inb(cx->address);
 721                         /* See comment in acpi_idle_do_entry() */
 722                         inl(acpi_gbl_FADT.xpm_timer_block.address);
 723                 } else
 724                         return -ENODEV;
 725         }
 726
 727         /* Never reached */
 728         return 0;
 729 }
 730
 731 static bool acpi_idle_fallback_to_c1(struct acpi_processor *pr)
 732 {
 733         return IS_ENABLED(CONFIG_HOTPLUG_CPU) && !pr->flags.has_cst &&
 734                 !(acpi_gbl_FADT.flags & ACPI_FADT_C2_MP_SUPPORTED);
 735 }
 736
 737 static int c3_cpu_count;
 738 static DEFINE_RAW_SPINLOCK(c3_lock);
 739
 740 /**
 741  * acpi_idle_enter_bm - enters C3 with proper BM handling
 742  * @pr: Target processor
 743  * @cx: Target state context
 744  * @timer_bc: Whether or not to change timer mode to broadcast
 745  */
 746 static void acpi_idle_enter_bm(struct acpi_processor *pr,
 747                                struct acpi_processor_cx *cx, bool timer_bc)
 748 {
 749         acpi_unlazy_tlb(smp_processor_id());
 750
 751         /*
 752          * Must be done before busmaster disable as we might need to
 753          * access HPET !
 754          */
 755         if (timer_bc)
 756                 lapic_timer_state_broadcast(pr, cx, 1);
 757
 758         /*
 759          * disable bus master
 760          * bm_check implies we need ARB_DIS
 761          * bm_control implies whether we can do ARB_DIS
 762          *
 763          * That leaves a case where bm_check is set and bm_control is
 764          * not set. In that case we cannot do much, we enter C3
 765          * without doing anything.
 766          */
 767         if (pr->flags.bm_control) {
 768                 raw_spin_lock(&c3_lock);
 769                 c3_cpu_count++;
 770                 /* Disable bus master arbitration when all CPUs are in C3 */
 771                 if (c3_cpu_count == num_online_cpus())
 772                         acpi_write_bit_register(ACPI_BITREG_ARB_DISABLE, 1);
 773                 raw_spin_unlock(&c3_lock);
 774         }
 775
 776         acpi_idle_do_entry(cx);
 777
 778         /* Re-enable bus master arbitration */
 779         if (pr->flags.bm_control) {
 780                 raw_spin_lock(&c3_lock);
 781                 acpi_write_bit_register(ACPI_BITREG_ARB_DISABLE, 0);
 782                 c3_cpu_count--;
 783                 raw_spin_unlock(&c3_lock);
 784         }
 785
 786         if (timer_bc)
 787                 lapic_timer_state_broadcast(pr, cx, 0);
 788 }
 789
 790 static int acpi_idle_enter(struct cpuidle_device *dev,
 791                            struct cpuidle_driver *drv, int index)
 792 {
 793         struct acpi_processor_cx *cx = per_cpu(acpi_cstate[index], dev->cpu);
 794         struct acpi_processor *pr;
 795
 796         pr = __this_cpu_read(processors);
 797         if (unlikely(!pr))
 798                 return -EINVAL;
 799
 800         if (cx->type != ACPI_STATE_C1) {
 801                 if (acpi_idle_fallback_to_c1(pr) && num_online_cpus() > 1) {
 802                         index = CPUIDLE_DRIVER_STATE_START;
 803                         cx = per_cpu(acpi_cstate[index], dev->cpu);
 804                 } else if (cx->type == ACPI_STATE_C3 && pr->flags.bm_check) {
 805                         if (cx->bm_sts_skip || !acpi_idle_bm_check()) {
 806                                 acpi_idle_enter_bm(pr, cx, true);
 807                                 return index;
 808                         } else if (drv->safe_state_index >= 0) {
 809                                 index = drv->safe_state_index;
 810                                 cx = per_cpu(acpi_cstate[index], dev->cpu);
 811                         } else {
 812                                 acpi_safe_halt();
 813                                 return -EBUSY;
 814                         }
 815                 }
 816         }
 817
 818         lapic_timer_state_broadcast(pr, cx, 1);
 819
 820         if (cx->type == ACPI_STATE_C3)
 821                 ACPI_FLUSH_CPU_CACHE();
 822
 823         acpi_idle_do_entry(cx);
 824
 825         lapic_timer_state_broadcast(pr, cx, 0);
 826
 827         return index;
 828 }
 829
 830 static void acpi_idle_enter_freeze(struct cpuidle_device *dev,
 831                                    struct cpuidle_driver *drv, int index)
 832 {
 833         struct acpi_processor_cx *cx = per_cpu(acpi_cstate[index], dev->cpu);
 834
 835         if (cx->type == ACPI_STATE_C3) {
 836                 struct acpi_processor *pr = __this_cpu_read(processors);
 837
 838                 if (unlikely(!pr))
 839                         return;
 840
 841                 if (pr->flags.bm_check) {
 842                         acpi_idle_enter_bm(pr, cx, false);
 843                         return;
 844                 } else {
 845                         ACPI_FLUSH_CPU_CACHE();
 846                 }
 847         }
 848         acpi_idle_do_entry(cx);
 849 }
 850
 851 struct cpuidle_driver acpi_idle_driver = {
 852         .name =         "acpi_idle",
 853         .owner =        THIS_MODULE,
 854 };
 855
 856 /**
 857  * acpi_processor_setup_cpuidle_cx - prepares and configures CPUIDLE
 858  * device i.e. per-cpu data
 859  *
 860  * @pr: the ACPI processor
 861  * @dev : the cpuidle device
 862  */
 863 static int acpi_processor_setup_cpuidle_cx(struct acpi_processor *pr,
 864                                            struct cpuidle_device *dev)
 865 {
 866         int i, count = CPUIDLE_DRIVER_STATE_START;
 867         struct acpi_processor_cx *cx;
 868
 869         if (!pr->flags.power_setup_done)
 870                 return -EINVAL;
 871
 872         if (pr->flags.power == 0) {
 873                 return -EINVAL;
 874         }
 875
 876         if (!dev)
 877                 return -EINVAL;
 878
 879         dev->cpu = pr->id;
 880
 881         if (max_cstate == 0)
 882                 max_cstate = 1;
 883
 884         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER && i <= max_cstate; i++) {
 885                 cx = &pr->power.states[i];
 886
 887                 if (!cx->valid)
 888                         continue;
 889
 890                 per_cpu(acpi_cstate[count], dev->cpu) = cx;
 891
 892                 count++;
 893                 if (count == CPUIDLE_STATE_MAX)
 894                         break;
 895         }
 896
 897         if (!count)
 898                 return -EINVAL;
 899
 900         return 0;
 901 }
 902
 903 /**
 904  * acpi_processor_setup_cpuidle states- prepares and configures cpuidle
 905  * global state data i.e. idle routines
 906  *
 907  * @pr: the ACPI processor
 908  */
 909 static int acpi_processor_setup_cpuidle_states(struct acpi_processor *pr)
 910 {
 911         int i, count = CPUIDLE_DRIVER_STATE_START;
 912         struct acpi_processor_cx *cx;
 913         struct cpuidle_state *state;
 914         struct cpuidle_driver *drv = &acpi_idle_driver;
 915
 916         if (!pr->flags.power_setup_done)
 917                 return -EINVAL;
 918
 919         if (pr->flags.power == 0)
 920                 return -EINVAL;
 921
 922         drv->safe_state_index = -1;
 923         for (i = 0; i < CPUIDLE_STATE_MAX; i++) {
 924                 drv->states[i].name[0] = '\0';
 925                 drv->states[i].desc[0] = '\0';
 926         }
 927
 928         if (max_cstate == 0)
 929                 max_cstate = 1;
 930
 931         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER && i <= max_cstate; i++) {
 932                 cx = &pr->power.states[i];
 933
 934                 if (!cx->valid)
 935                         continue;
 936
 937                 state = &drv->states[count];
 938                 snprintf(state->name, CPUIDLE_NAME_LEN, "C%d", i);
 939                 strncpy(state->desc, cx->desc, CPUIDLE_DESC_LEN);
 940                 state->exit_latency = cx->latency;
 941                 state->target_residency = cx->latency * latency_factor;
 942                 state->enter = acpi_idle_enter;
 943
 944                 state->flags = 0;
 945                 if (cx->type == ACPI_STATE_C1 || cx->type == ACPI_STATE_C2) {
 946                         state->enter_dead = acpi_idle_play_dead;
 947                         drv->safe_state_index = count;
 948                 }
 949                 /*
 950                  * Halt-induced C1 is not good for ->enter_freeze, because it
 951                  * re-enables interrupts on exit.  Moreover, C1 is generally not
 952                  * particularly interesting from the suspend-to-idle angle, so
 953                  * avoid C1 and the situations in which we may need to fall back
 954                  * to it altogether.
 955                  */
 956                 if (cx->type != ACPI_STATE_C1 && !acpi_idle_fallback_to_c1(pr))
 957                         state->enter_freeze = acpi_idle_enter_freeze;
 958
 959                 count++;
 960                 if (count == CPUIDLE_STATE_MAX)
 961                         break;
 962         }
 963
 964         drv->state_count = count;
 965
 966         if (!count)
 967                 return -EINVAL;
 968
 969         return 0;
 970 }
 971
 972 int acpi_processor_hotplug(struct acpi_processor *pr)
 973 {
 974         int ret = 0;
 975         struct cpuidle_device *dev;
 976
 977         if (disabled_by_idle_boot_param())
 978                 return 0;
 979
 980         if (nocst)
 981                 return -ENODEV;
 982
 983         if (!pr->flags.power_setup_done)
 984                 return -ENODEV;
 985
 986         dev = per_cpu(acpi_cpuidle_device, pr->id);
 987         cpuidle_pause_and_lock();
 988         cpuidle_disable_device(dev);
 989         acpi_processor_get_power_info(pr);
 990         if (pr->flags.power) {
 991                 acpi_processor_setup_cpuidle_cx(pr, dev);
 992                 ret = cpuidle_enable_device(dev);
 993         }
 994         cpuidle_resume_and_unlock();
 995
 996         return ret;
 997 }
 998
 999 int acpi_processor_cst_has_changed(struct acpi_processor *pr)
1000 {
1001         int cpu;
1002         struct acpi_processor *_pr;
1003         struct cpuidle_device *dev;
1004
1005         if (disabled_by_idle_boot_param())
1006                 return 0;
1007
1008         if (nocst)
1009                 return -ENODEV;
1010
1011         if (!pr->flags.power_setup_done)
1012                 return -ENODEV;
1013
1014         /*
1015          * FIXME:  Design the ACPI notification to make it once per
1016          * system instead of once per-cpu.  This condition is a hack
1017          * to make the code that updates C-States be called once.
1018          */
1019
1020         if (pr->id == 0 && cpuidle_get_driver() == &acpi_idle_driver) {
1021
1022                 /* Protect against cpu-hotplug */
1023                 get_online_cpus();
1024                 cpuidle_pause_and_lock();
1025
1026                 /* Disable all cpuidle devices */
1027                 for_each_online_cpu(cpu) {
1028                         _pr = per_cpu(processors, cpu);
1029                         if (!_pr || !_pr->flags.power_setup_done)
1030                                 continue;
1031                         dev = per_cpu(acpi_cpuidle_device, cpu);
1032                         cpuidle_disable_device(dev);
1033                 }
1034
1035                 /* Populate Updated C-state information */
1036                 acpi_processor_get_power_info(pr);
1037                 acpi_processor_setup_cpuidle_states(pr);
1038
1039                 /* Enable all cpuidle devices */
1040                 for_each_online_cpu(cpu) {
1041                         _pr = per_cpu(processors, cpu);
1042                         if (!_pr || !_pr->flags.power_setup_done)
1043                                 continue;
1044                         acpi_processor_get_power_info(_pr);
1045                         if (_pr->flags.power) {
1046                                 dev = per_cpu(acpi_cpuidle_device, cpu);
1047                                 acpi_processor_setup_cpuidle_cx(_pr, dev);
1048                                 cpuidle_enable_device(dev);
1049                         }
1050                 }
1051                 cpuidle_resume_and_unlock();
1052                 put_online_cpus();
1053         }
1054
1055         return 0;
1056 }
1057
1058 static int acpi_processor_registered;
1059
1060 int acpi_processor_power_init(struct acpi_processor *pr)
1061 {
1062         acpi_status status;
1063         int retval;
1064         struct cpuidle_device *dev;
1065         static int first_run;
1066
1067         if (disabled_by_idle_boot_param())
1068                 return 0;
1069
1070         if (!first_run) {
1071                 dmi_check_system(processor_power_dmi_table);
1072                 max_cstate = acpi_processor_cstate_check(max_cstate);
1073                 if (max_cstate < ACPI_C_STATES_MAX)
1074                         printk(KERN_NOTICE
1075                                "ACPI: processor limited to max C-state %d\n",
1076                                max_cstate);
1077                 first_run++;
1078         }
1079
1080         if (acpi_gbl_FADT.cst_control && !nocst) {
1081                 status =
1082                     acpi_os_write_port(acpi_gbl_FADT.smi_command, acpi_gbl_FADT.cst_control, 8);
1083                 if (ACPI_FAILURE(status)) {
1084                         ACPI_EXCEPTION((AE_INFO, status,
1085                                         "Notifying BIOS of _CST ability failed"));
1086                 }
1087         }
1088
1089         acpi_processor_get_power_info(pr);
1090         pr->flags.power_setup_done = 1;
1091
1092         /*
1093          * Install the idle handler if processor power management is supported.
1094          * Note that we use previously set idle handler will be used on
1095          * platforms that only support C1.
1096          */
1097         if (pr->flags.power) {
1098                 /* Register acpi_idle_driver if not already registered */
1099                 if (!acpi_processor_registered) {
1100                         acpi_processor_setup_cpuidle_states(pr);
1101                         retval = cpuidle_register_driver(&acpi_idle_driver);
1102                         if (retval)
1103                                 return retval;
1104                         printk(KERN_DEBUG "ACPI: %s registered with cpuidle\n",
1105                                         acpi_idle_driver.name);
1106                 }
1107
1108                 dev = kzalloc(sizeof(*dev), GFP_KERNEL);
1109                 if (!dev)
1110                         return -ENOMEM;
1111                 per_cpu(acpi_cpuidle_device, pr->id) = dev;
1112
1113                 acpi_processor_setup_cpuidle_cx(pr, dev);
1114
1115                 /* Register per-cpu cpuidle_device. Cpuidle driver
1116                  * must already be registered before registering device
1117                  */
1118                 retval = cpuidle_register_device(dev);
1119                 if (retval) {
1120                         if (acpi_processor_registered == 0)
1121                                 cpuidle_unregister_driver(&acpi_idle_driver);
1122                         return retval;
1123                 }
1124                 acpi_processor_registered++;
1125         }
1126         return 0;
1127 }
1128
1129 int acpi_processor_power_exit(struct acpi_processor *pr)
1130 {
1131         struct cpuidle_device *dev = per_cpu(acpi_cpuidle_device, pr->id);
1132
1133         if (disabled_by_idle_boot_param())
1134                 return 0;
1135
1136         if (pr->flags.power) {
1137                 cpuidle_unregister_device(dev);
1138                 acpi_processor_registered--;
1139                 if (acpi_processor_registered == 0)
1140                         cpuidle_unregister_driver(&acpi_idle_driver);
1141         }
1142
1143         pr->flags.power_setup_done = 0;
1144         return 0;
1145 }