drivers/acpi/processor_idle.c

   1 /*
   2  * processor_idle - idle state submodule to the ACPI processor driver
   3  *
   4  *  Copyright (C) 2001, 2002 Andy Grover <andrew.grover@intel.com>
   5  *  Copyright (C) 2001, 2002 Paul Diefenbaugh <paul.s.diefenbaugh@intel.com>
   6  *  Copyright (C) 2004, 2005 Dominik Brodowski <linux@brodo.de>
   7  *  Copyright (C) 2004  Anil S Keshavamurthy <anil.s.keshavamurthy@intel.com>
   8  *                      - Added processor hotplug support
   9  *  Copyright (C) 2005  Venkatesh Pallipadi <venkatesh.pallipadi@intel.com>
  10  *                      - Added support for C3 on SMP
  11  *
  12  * ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
  13  *
  14  *  This program is free software; you can redistribute it and/or modify
  15  *  it under the terms of the GNU General Public License as published by
  16  *  the Free Software Foundation; either version 2 of the License, or (at
  17  *  your option) any later version.
  18  *
  19  *  This program is distributed in the hope that it will be useful, but
  20  *  WITHOUT ANY WARRANTY; without even the implied warranty of
  21  *  MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
  22  *  General Public License for more details.
  23  *
  24  *  You should have received a copy of the GNU General Public License along
  25  *  with this program; if not, write to the Free Software Foundation, Inc.,
  26  *  59 Temple Place, Suite 330, Boston, MA 02111-1307 USA.
  27  *
  28  * ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
  29  */
  30
  31 #include <linux/kernel.h>
  32 #include <linux/module.h>
  33 #include <linux/init.h>
  34 #include <linux/cpufreq.h>
  35 #include <linux/slab.h>
  36 #include <linux/acpi.h>
  37 #include <linux/dmi.h>
  38 #include <linux/moduleparam.h>
  39 #include <linux/sched.h>        /* need_resched() */
  40 #include <linux/pm_qos.h>
  41 #include <linux/clockchips.h>
  42 #include <linux/cpuidle.h>
  43 #include <linux/irqflags.h>
  44
  45 /*
  46  * Include the apic definitions for x86 to have the APIC timer related defines
  47  * available also for UP (on SMP it gets magically included via linux/smp.h).
  48  * asm/acpi.h is not an option, as it would require more include magic. Also
  49  * creating an empty asm-ia64/apic.h would just trade pest vs. cholera.
  50  */
  51 #ifdef CONFIG_X86
  52 #include <asm/apic.h>
  53 #endif
  54
  55 #include <asm/io.h>
  56 #include <asm/uaccess.h>
  57
  58 #include <acpi/acpi_bus.h>
  59 #include <acpi/processor.h>
  60 #include <asm/processor.h>
  61
  62 #define PREFIX "ACPI: "
  63
  64 #define ACPI_PROCESSOR_CLASS            "processor"
  65 #define _COMPONENT              ACPI_PROCESSOR_COMPONENT
  66 ACPI_MODULE_NAME("processor_idle");
  67 #define PM_TIMER_TICK_NS                (1000000000ULL/PM_TIMER_FREQUENCY)
  68 #define C2_OVERHEAD                     1       /* 1us */
  69 #define C3_OVERHEAD                     1       /* 1us */
  70 #define PM_TIMER_TICKS_TO_US(p)         (((p) * 1000)/(PM_TIMER_FREQUENCY/1000))
  71
  72 static unsigned int max_cstate __read_mostly = ACPI_PROCESSOR_MAX_POWER;
  73 module_param(max_cstate, uint, 0000);
  74 static unsigned int nocst __read_mostly;
  75 module_param(nocst, uint, 0000);
  76 static int bm_check_disable __read_mostly;
  77 module_param(bm_check_disable, uint, 0000);
  78
  79 static unsigned int latency_factor __read_mostly = 2;
  80 module_param(latency_factor, uint, 0644);
  81
  82 static int disabled_by_idle_boot_param(void)
  83 {
  84         return boot_option_idle_override == IDLE_POLL ||
  85                 boot_option_idle_override == IDLE_FORCE_MWAIT ||
  86                 boot_option_idle_override == IDLE_HALT;
  87 }
  88
  89 /*
  90  * IBM ThinkPad R40e crashes mysteriously when going into C2 or C3.
  91  * For now disable this. Probably a bug somewhere else.
  92  *
  93  * To skip this limit, boot/load with a large max_cstate limit.
  94  */
  95 static int set_max_cstate(const struct dmi_system_id *id)
  96 {
  97         if (max_cstate > ACPI_PROCESSOR_MAX_POWER)
  98                 return 0;
  99
 100         printk(KERN_NOTICE PREFIX "%s detected - limiting to C%ld max_cstate."
 101                " Override with \"processor.max_cstate=%d\"\n", id->ident,
 102                (long)id->driver_data, ACPI_PROCESSOR_MAX_POWER + 1);
 103
 104         max_cstate = (long)id->driver_data;
 105
 106         return 0;
 107 }
 108
 109 /* Actually this shouldn't be __cpuinitdata, would be better to fix the
 110    callers to only run once -AK */
 111 static struct dmi_system_id __cpuinitdata processor_power_dmi_table[] = {
 112         { set_max_cstate, "Clevo 5600D", {
 113           DMI_MATCH(DMI_BIOS_VENDOR,"Phoenix Technologies LTD"),
 114           DMI_MATCH(DMI_BIOS_VERSION,"SHE845M0.86C.0013.D.0302131307")},
 115          (void *)2},
 116         { set_max_cstate, "Pavilion zv5000", {
 117           DMI_MATCH(DMI_SYS_VENDOR, "Hewlett-Packard"),
 118           DMI_MATCH(DMI_PRODUCT_NAME,"Pavilion zv5000 (DS502A#ABA)")},
 119          (void *)1},
 120         { set_max_cstate, "Asus L8400B", {
 121           DMI_MATCH(DMI_SYS_VENDOR, "ASUSTeK Computer Inc."),
 122           DMI_MATCH(DMI_PRODUCT_NAME,"L8400B series Notebook PC")},
 123          (void *)1},
 124         {},
 125 };
 126
 127
 128 /*
 129  * Callers should disable interrupts before the call and enable
 130  * interrupts after return.
 131  */
 132 static void acpi_safe_halt(void)
 133 {
 134         current_thread_info()->status &= ~TS_POLLING;
 135         /*
 136          * TS_POLLING-cleared state must be visible before we
 137          * test NEED_RESCHED:
 138          */
 139         smp_mb();
 140         if (!need_resched()) {
 141                 safe_halt();
 142                 local_irq_disable();
 143         }
 144         current_thread_info()->status |= TS_POLLING;
 145 }
 146
 147 #ifdef ARCH_APICTIMER_STOPS_ON_C3
 148
 149 /*
 150  * Some BIOS implementations switch to C3 in the published C2 state.
 151  * This seems to be a common problem on AMD boxen, but other vendors
 152  * are affected too. We pick the most conservative approach: we assume
 153  * that the local APIC stops in both C2 and C3.
 154  */
 155 static void lapic_timer_check_state(int state, struct acpi_processor *pr,
 156                                    struct acpi_processor_cx *cx)
 157 {
 158         struct acpi_processor_power *pwr = &pr->power;
 159         u8 type = local_apic_timer_c2_ok ? ACPI_STATE_C3 : ACPI_STATE_C2;
 160
 161         if (cpu_has(&cpu_data(pr->id), X86_FEATURE_ARAT))
 162                 return;
 163
 164         if (amd_e400_c1e_detected)
 165                 type = ACPI_STATE_C1;
 166
 167         /*
 168          * Check, if one of the previous states already marked the lapic
 169          * unstable
 170          */
 171         if (pwr->timer_broadcast_on_state < state)
 172                 return;
 173
 174         if (cx->type >= type)
 175                 pr->power.timer_broadcast_on_state = state;
 176 }
 177
 178 static void __lapic_timer_propagate_broadcast(void *arg)
 179 {
 180         struct acpi_processor *pr = (struct acpi_processor *) arg;
 181         unsigned long reason;
 182
 183         reason = pr->power.timer_broadcast_on_state < INT_MAX ?
 184                 CLOCK_EVT_NOTIFY_BROADCAST_ON : CLOCK_EVT_NOTIFY_BROADCAST_OFF;
 185
 186         clockevents_notify(reason, &pr->id);
 187 }
 188
 189 static void lapic_timer_propagate_broadcast(struct acpi_processor *pr)
 190 {
 191         smp_call_function_single(pr->id, __lapic_timer_propagate_broadcast,
 192                                  (void *)pr, 1);
 193 }
 194
 195 /* Power(C) State timer broadcast control */
 196 static void lapic_timer_state_broadcast(struct acpi_processor *pr,
 197                                        struct acpi_processor_cx *cx,
 198                                        int broadcast)
 199 {
 200         int state = cx - pr->power.states;
 201
 202         if (state >= pr->power.timer_broadcast_on_state) {
 203                 unsigned long reason;
 204
 205                 reason = broadcast ?  CLOCK_EVT_NOTIFY_BROADCAST_ENTER :
 206                         CLOCK_EVT_NOTIFY_BROADCAST_EXIT;
 207                 clockevents_notify(reason, &pr->id);
 208         }
 209 }
 210
 211 #else
 212
 213 static void lapic_timer_check_state(int state, struct acpi_processor *pr,
 214                                    struct acpi_processor_cx *cstate) { }
 215 static void lapic_timer_propagate_broadcast(struct acpi_processor *pr) { }
 216 static void lapic_timer_state_broadcast(struct acpi_processor *pr,
 217                                        struct acpi_processor_cx *cx,
 218                                        int broadcast)
 219 {
 220 }
 221
 222 #endif
 223
 224 /*
 225  * Suspend / resume control
 226  */
 227 static u32 saved_bm_rld;
 228
 229 static void acpi_idle_bm_rld_save(void)
 230 {
 231         acpi_read_bit_register(ACPI_BITREG_BUS_MASTER_RLD, &saved_bm_rld);
 232 }
 233 static void acpi_idle_bm_rld_restore(void)
 234 {
 235         u32 resumed_bm_rld;
 236
 237         acpi_read_bit_register(ACPI_BITREG_BUS_MASTER_RLD, &resumed_bm_rld);
 238
 239         if (resumed_bm_rld != saved_bm_rld)
 240                 acpi_write_bit_register(ACPI_BITREG_BUS_MASTER_RLD, saved_bm_rld);
 241 }
 242
 243 int acpi_processor_suspend(struct acpi_device * device, pm_message_t state)
 244 {
 245         acpi_idle_bm_rld_save();
 246         return 0;
 247 }
 248
 249 int acpi_processor_resume(struct acpi_device * device)
 250 {
 251         acpi_idle_bm_rld_restore();
 252         return 0;
 253 }
 254
 255 #if defined(CONFIG_X86)
 256 static void tsc_check_state(int state)
 257 {
 258         switch (boot_cpu_data.x86_vendor) {
 259         case X86_VENDOR_AMD:
 260         case X86_VENDOR_INTEL:
 261                 /*
 262                  * AMD Fam10h TSC will tick in all
 263                  * C/P/S0/S1 states when this bit is set.
 264                  */
 265                 if (boot_cpu_has(X86_FEATURE_NONSTOP_TSC))
 266                         return;
 267
 268                 /*FALL THROUGH*/
 269         default:
 270                 /* TSC could halt in idle, so notify users */
 271                 if (state > ACPI_STATE_C1)
 272                         mark_tsc_unstable("TSC halts in idle");
 273         }
 274 }
 275 #else
 276 static void tsc_check_state(int state) { return; }
 277 #endif
 278
 279 static int acpi_processor_get_power_info_fadt(struct acpi_processor *pr)
 280 {
 281
 282         if (!pr)
 283                 return -EINVAL;
 284
 285         if (!pr->pblk)
 286                 return -ENODEV;
 287
 288         /* if info is obtained from pblk/fadt, type equals state */
 289         pr->power.states[ACPI_STATE_C2].type = ACPI_STATE_C2;
 290         pr->power.states[ACPI_STATE_C3].type = ACPI_STATE_C3;
 291
 292 #ifndef CONFIG_HOTPLUG_CPU
 293         /*
 294          * Check for P_LVL2_UP flag before entering C2 and above on
 295          * an SMP system.
 296          */
 297         if ((num_online_cpus() > 1) &&
 298             !(acpi_gbl_FADT.flags & ACPI_FADT_C2_MP_SUPPORTED))
 299                 return -ENODEV;
 300 #endif
 301
 302         /* determine C2 and C3 address from pblk */
 303         pr->power.states[ACPI_STATE_C2].address = pr->pblk + 4;
 304         pr->power.states[ACPI_STATE_C3].address = pr->pblk + 5;
 305
 306         /* determine latencies from FADT */
 307         pr->power.states[ACPI_STATE_C2].latency = acpi_gbl_FADT.C2latency;
 308         pr->power.states[ACPI_STATE_C3].latency = acpi_gbl_FADT.C3latency;
 309
 310         /*
 311          * FADT specified C2 latency must be less than or equal to
 312          * 100 microseconds.
 313          */
 314         if (acpi_gbl_FADT.C2latency > ACPI_PROCESSOR_MAX_C2_LATENCY) {
 315                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 316                         "C2 latency too large [%d]\n", acpi_gbl_FADT.C2latency));
 317                 /* invalidate C2 */
 318                 pr->power.states[ACPI_STATE_C2].address = 0;
 319         }
 320
 321         /*
 322          * FADT supplied C3 latency must be less than or equal to
 323          * 1000 microseconds.
 324          */
 325         if (acpi_gbl_FADT.C3latency > ACPI_PROCESSOR_MAX_C3_LATENCY) {
 326                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 327                         "C3 latency too large [%d]\n", acpi_gbl_FADT.C3latency));
 328                 /* invalidate C3 */
 329                 pr->power.states[ACPI_STATE_C3].address = 0;
 330         }
 331
 332         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 333                           "lvl2[0x%08x] lvl3[0x%08x]\n",
 334                           pr->power.states[ACPI_STATE_C2].address,
 335                           pr->power.states[ACPI_STATE_C3].address));
 336
 337         return 0;
 338 }
 339
 340 static int acpi_processor_get_power_info_default(struct acpi_processor *pr)
 341 {
 342         if (!pr->power.states[ACPI_STATE_C1].valid) {
 343                 /* set the first C-State to C1 */
 344                 /* all processors need to support C1 */
 345                 pr->power.states[ACPI_STATE_C1].type = ACPI_STATE_C1;
 346                 pr->power.states[ACPI_STATE_C1].valid = 1;
 347                 pr->power.states[ACPI_STATE_C1].entry_method = ACPI_CSTATE_HALT;
 348         }
 349         /* the C0 state only exists as a filler in our array */
 350         pr->power.states[ACPI_STATE_C0].valid = 1;
 351         return 0;
 352 }
 353
 354 static int acpi_processor_get_power_info_cst(struct acpi_processor *pr)
 355 {
 356         acpi_status status = 0;
 357         u64 count;
 358         int current_count;
 359         int i;
 360         struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
 361         union acpi_object *cst;
 362
 363
 364         if (nocst)
 365                 return -ENODEV;
 366
 367         current_count = 0;
 368
 369         status = acpi_evaluate_object(pr->handle, "_CST", NULL, &buffer);
 370         if (ACPI_FAILURE(status)) {
 371                 ACPI_DEBUG_PRINT((ACPI_DB_INFO, "No _CST, giving up\n"));
 372                 return -ENODEV;
 373         }
 374
 375         cst = buffer.pointer;
 376
 377         /* There must be at least 2 elements */
 378         if (!cst || (cst->type != ACPI_TYPE_PACKAGE) || cst->package.count < 2) {
 379                 printk(KERN_ERR PREFIX "not enough elements in _CST\n");
 380                 status = -EFAULT;
 381                 goto end;
 382         }
 383
 384         count = cst->package.elements[0].integer.value;
 385
 386         /* Validate number of power states. */
 387         if (count < 1 || count != cst->package.count - 1) {
 388                 printk(KERN_ERR PREFIX "count given by _CST is not valid\n");
 389                 status = -EFAULT;
 390                 goto end;
 391         }
 392
 393         /* Tell driver that at least _CST is supported. */
 394         pr->flags.has_cst = 1;
 395
 396         for (i = 1; i <= count; i++) {
 397                 union acpi_object *element;
 398                 union acpi_object *obj;
 399                 struct acpi_power_register *reg;
 400                 struct acpi_processor_cx cx;
 401
 402                 memset(&cx, 0, sizeof(cx));
 403
 404                 element = &(cst->package.elements[i]);
 405                 if (element->type != ACPI_TYPE_PACKAGE)
 406                         continue;
 407
 408                 if (element->package.count != 4)
 409                         continue;
 410
 411                 obj = &(element->package.elements[0]);
 412
 413                 if (obj->type != ACPI_TYPE_BUFFER)
 414                         continue;
 415
 416                 reg = (struct acpi_power_register *)obj->buffer.pointer;
 417
 418                 if (reg->space_id != ACPI_ADR_SPACE_SYSTEM_IO &&
 419                     (reg->space_id != ACPI_ADR_SPACE_FIXED_HARDWARE))
 420                         continue;
 421
 422                 /* There should be an easy way to extract an integer... */
 423                 obj = &(element->package.elements[1]);
 424                 if (obj->type != ACPI_TYPE_INTEGER)
 425                         continue;
 426
 427                 cx.type = obj->integer.value;
 428                 /*
 429                  * Some buggy BIOSes won't list C1 in _CST -
 430                  * Let acpi_processor_get_power_info_default() handle them later
 431                  */
 432                 if (i == 1 && cx.type != ACPI_STATE_C1)
 433                         current_count++;
 434
 435                 cx.address = reg->address;
 436                 cx.index = current_count + 1;
 437
 438                 cx.entry_method = ACPI_CSTATE_SYSTEMIO;
 439                 if (reg->space_id == ACPI_ADR_SPACE_FIXED_HARDWARE) {
 440                         if (acpi_processor_ffh_cstate_probe
 441                                         (pr->id, &cx, reg) == 0) {
 442                                 cx.entry_method = ACPI_CSTATE_FFH;
 443                         } else if (cx.type == ACPI_STATE_C1) {
 444                                 /*
 445                                  * C1 is a special case where FIXED_HARDWARE
 446                                  * can be handled in non-MWAIT way as well.
 447                                  * In that case, save this _CST entry info.
 448                                  * Otherwise, ignore this info and continue.
 449                                  */
 450                                 cx.entry_method = ACPI_CSTATE_HALT;
 451                                 snprintf(cx.desc, ACPI_CX_DESC_LEN, "ACPI HLT");
 452                         } else {
 453                                 continue;
 454                         }
 455                         if (cx.type == ACPI_STATE_C1 &&
 456                             (boot_option_idle_override == IDLE_NOMWAIT)) {
 457                                 /*
 458                                  * In most cases the C1 space_id obtained from
 459                                  * _CST object is FIXED_HARDWARE access mode.
 460                                  * But when the option of idle=halt is added,
 461                                  * the entry_method type should be changed from
 462                                  * CSTATE_FFH to CSTATE_HALT.
 463                                  * When the option of idle=nomwait is added,
 464                                  * the C1 entry_method type should be
 465                                  * CSTATE_HALT.
 466                                  */
 467                                 cx.entry_method = ACPI_CSTATE_HALT;
 468                                 snprintf(cx.desc, ACPI_CX_DESC_LEN, "ACPI HLT");
 469                         }
 470                 } else {
 471                         snprintf(cx.desc, ACPI_CX_DESC_LEN, "ACPI IOPORT 0x%x",
 472                                  cx.address);
 473                 }
 474
 475                 if (cx.type == ACPI_STATE_C1) {
 476                         cx.valid = 1;
 477                 }
 478
 479                 obj = &(element->package.elements[2]);
 480                 if (obj->type != ACPI_TYPE_INTEGER)
 481                         continue;
 482
 483                 cx.latency = obj->integer.value;
 484
 485                 obj = &(element->package.elements[3]);
 486                 if (obj->type != ACPI_TYPE_INTEGER)
 487                         continue;
 488
 489                 cx.power = obj->integer.value;
 490
 491                 current_count++;
 492                 memcpy(&(pr->power.states[current_count]), &cx, sizeof(cx));
 493
 494                 /*
 495                  * We support total ACPI_PROCESSOR_MAX_POWER - 1
 496                  * (From 1 through ACPI_PROCESSOR_MAX_POWER - 1)
 497                  */
 498                 if (current_count >= (ACPI_PROCESSOR_MAX_POWER - 1)) {
 499                         printk(KERN_WARNING
 500                                "Limiting number of power states to max (%d)\n",
 501                                ACPI_PROCESSOR_MAX_POWER);
 502                         printk(KERN_WARNING
 503                                "Please increase ACPI_PROCESSOR_MAX_POWER if needed.\n");
 504                         break;
 505                 }
 506         }
 507
 508         ACPI_DEBUG_PRINT((ACPI_DB_INFO, "Found %d power states\n",
 509                           current_count));
 510
 511         /* Validate number of power states discovered */
 512         if (current_count < 2)
 513                 status = -EFAULT;
 514
 515       end:
 516         kfree(buffer.pointer);
 517
 518         return status;
 519 }
 520
 521 static void acpi_processor_power_verify_c3(struct acpi_processor *pr,
 522                                            struct acpi_processor_cx *cx)
 523 {
 524         static int bm_check_flag = -1;
 525         static int bm_control_flag = -1;
 526
 527
 528         if (!cx->address)
 529                 return;
 530
 531         /*
 532          * PIIX4 Erratum #18: We don't support C3 when Type-F (fast)
 533          * DMA transfers are used by any ISA device to avoid livelock.
 534          * Note that we could disable Type-F DMA (as recommended by
 535          * the erratum), but this is known to disrupt certain ISA
 536          * devices thus we take the conservative approach.
 537          */
 538         else if (errata.piix4.fdma) {
 539                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 540                                   "C3 not supported on PIIX4 with Type-F DMA\n"));
 541                 return;
 542         }
 543
 544         /* All the logic here assumes flags.bm_check is same across all CPUs */
 545         if (bm_check_flag == -1) {
 546                 /* Determine whether bm_check is needed based on CPU  */
 547                 acpi_processor_power_init_bm_check(&(pr->flags), pr->id);
 548                 bm_check_flag = pr->flags.bm_check;
 549                 bm_control_flag = pr->flags.bm_control;
 550         } else {
 551                 pr->flags.bm_check = bm_check_flag;
 552                 pr->flags.bm_control = bm_control_flag;
 553         }
 554
 555         if (pr->flags.bm_check) {
 556                 if (!pr->flags.bm_control) {
 557                         if (pr->flags.has_cst != 1) {
 558                                 /* bus mastering control is necessary */
 559                                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 560                                         "C3 support requires BM control\n"));
 561                                 return;
 562                         } else {
 563                                 /* Here we enter C3 without bus mastering */
 564                                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 565                                         "C3 support without BM control\n"));
 566                         }
 567                 }
 568         } else {
 569                 /*
 570                  * WBINVD should be set in fadt, for C3 state to be
 571                  * supported on when bm_check is not required.
 572                  */
 573                 if (!(acpi_gbl_FADT.flags & ACPI_FADT_WBINVD)) {
 574                         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 575                                           "Cache invalidation should work properly"
 576                                           " for C3 to be enabled on SMP systems\n"));
 577                         return;
 578                 }
 579         }
 580
 581         /*
 582          * Otherwise we've met all of our C3 requirements.
 583          * Normalize the C3 latency to expidite policy.  Enable
 584          * checking of bus mastering status (bm_check) so we can
 585          * use this in our C3 policy
 586          */
 587         cx->valid = 1;
 588
 589         cx->latency_ticks = cx->latency;
 590         /*
 591          * On older chipsets, BM_RLD needs to be set
 592          * in order for Bus Master activity to wake the
 593          * system from C3.  Newer chipsets handle DMA
 594          * during C3 automatically and BM_RLD is a NOP.
 595          * In either case, the proper way to
 596          * handle BM_RLD is to set it and leave it set.
 597          */
 598         acpi_write_bit_register(ACPI_BITREG_BUS_MASTER_RLD, 1);
 599
 600         return;
 601 }
 602
 603 static int acpi_processor_power_verify(struct acpi_processor *pr)
 604 {
 605         unsigned int i;
 606         unsigned int working = 0;
 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                         cx->latency_ticks = cx->latency; /* Normalize latency */
 623                         break;
 624
 625                 case ACPI_STATE_C3:
 626                         acpi_processor_power_verify_c3(pr, cx);
 627                         break;
 628                 }
 629                 if (!cx->valid)
 630                         continue;
 631
 632                 lapic_timer_check_state(i, pr, cx);
 633                 tsc_check_state(cx->type);
 634                 working++;
 635         }
 636
 637         lapic_timer_propagate_broadcast(pr);
 638
 639         return (working);
 640 }
 641
 642 static int acpi_processor_get_power_info(struct acpi_processor *pr)
 643 {
 644         unsigned int i;
 645         int result;
 646
 647
 648         /* NOTE: the idle thread may not be running while calling
 649          * this function */
 650
 651         /* Zero initialize all the C-states info. */
 652         memset(pr->power.states, 0, sizeof(pr->power.states));
 653
 654         result = acpi_processor_get_power_info_cst(pr);
 655         if (result == -ENODEV)
 656                 result = acpi_processor_get_power_info_fadt(pr);
 657
 658         if (result)
 659                 return result;
 660
 661         acpi_processor_get_power_info_default(pr);
 662
 663         pr->power.count = acpi_processor_power_verify(pr);
 664
 665         /*
 666          * if one state of type C2 or C3 is available, mark this
 667          * CPU as being "idle manageable"
 668          */
 669         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
 670                 if (pr->power.states[i].valid) {
 671                         pr->power.count = i;
 672                         if (pr->power.states[i].type >= ACPI_STATE_C2)
 673                                 pr->flags.power = 1;
 674                 }
 675         }
 676
 677         return 0;
 678 }
 679
 680 /**
 681  * acpi_idle_bm_check - checks if bus master activity was detected
 682  */
 683 static int acpi_idle_bm_check(void)
 684 {
 685         u32 bm_status = 0;
 686
 687         if (bm_check_disable)
 688                 return 0;
 689
 690         acpi_read_bit_register(ACPI_BITREG_BUS_MASTER_STATUS, &bm_status);
 691         if (bm_status)
 692                 acpi_write_bit_register(ACPI_BITREG_BUS_MASTER_STATUS, 1);
 693         /*
 694          * PIIX4 Erratum #18: Note that BM_STS doesn't always reflect
 695          * the true state of bus mastering activity; forcing us to
 696          * manually check the BMIDEA bit of each IDE channel.
 697          */
 698         else if (errata.piix4.bmisx) {
 699                 if ((inb_p(errata.piix4.bmisx + 0x02) & 0x01)
 700                     || (inb_p(errata.piix4.bmisx + 0x0A) & 0x01))
 701                         bm_status = 1;
 702         }
 703         return bm_status;
 704 }
 705
 706 /**
 707  * acpi_idle_do_entry - a helper function that does C2 and C3 type entry
 708  * @cx: cstate data
 709  *
 710  * Caller disables interrupt before call and enables interrupt after return.
 711  */
 712 static inline void acpi_idle_do_entry(struct acpi_processor_cx *cx)
 713 {
 714         /* Don't trace irqs off for idle */
 715         stop_critical_timings();
 716         if (cx->entry_method == ACPI_CSTATE_FFH) {
 717                 /* Call into architectural FFH based C-state */
 718                 acpi_processor_ffh_cstate_enter(cx);
 719         } else if (cx->entry_method == ACPI_CSTATE_HALT) {
 720                 acpi_safe_halt();
 721         } else {
 722                 /* IO port based C-state */
 723                 inb(cx->address);
 724                 /* Dummy wait op - must do something useless after P_LVL2 read
 725                    because chipsets cannot guarantee that STPCLK# signal
 726                    gets asserted in time to freeze execution properly. */
 727                 inl(acpi_gbl_FADT.xpm_timer_block.address);
 728         }
 729         start_critical_timings();
 730 }
 731
 732 /**
 733  * acpi_idle_enter_c1 - enters an ACPI C1 state-type
 734  * @dev: the target CPU
 735  * @drv: cpuidle driver containing cpuidle state info
 736  * @index: index of target state
 737  *
 738  * This is equivalent to the HALT instruction.
 739  */
 740 static int acpi_idle_enter_c1(struct cpuidle_device *dev,
 741                 struct cpuidle_driver *drv, int index)
 742 {
 743         ktime_t  kt1, kt2;
 744         s64 idle_time;
 745         struct acpi_processor *pr;
 746         struct cpuidle_state_usage *state_usage = &dev->states_usage[index];
 747         struct acpi_processor_cx *cx = cpuidle_get_statedata(state_usage);
 748
 749         pr = __this_cpu_read(processors);
 750         dev->last_residency = 0;
 751
 752         if (unlikely(!pr))
 753                 return -EINVAL;
 754
 755         local_irq_disable();
 756
 757         lapic_timer_state_broadcast(pr, cx, 1);
 758         kt1 = ktime_get_real();
 759         acpi_idle_do_entry(cx);
 760         kt2 = ktime_get_real();
 761         idle_time =  ktime_to_us(ktime_sub(kt2, kt1));
 762
 763         /* Update device last_residency*/
 764         dev->last_residency = (int)idle_time;
 765
 766         local_irq_enable();
 767         cx->usage++;
 768         lapic_timer_state_broadcast(pr, cx, 0);
 769
 770         return index;
 771 }
 772
 773 /**
 774  * acpi_idle_enter_simple - enters an ACPI state without BM handling
 775  * @dev: the target CPU
 776  * @drv: cpuidle driver with cpuidle state information
 777  * @index: the index of suggested state
 778  */
 779 static int acpi_idle_enter_simple(struct cpuidle_device *dev,
 780                 struct cpuidle_driver *drv, int index)
 781 {
 782         struct acpi_processor *pr;
 783         struct cpuidle_state_usage *state_usage = &dev->states_usage[index];
 784         struct acpi_processor_cx *cx = cpuidle_get_statedata(state_usage);
 785         ktime_t  kt1, kt2;
 786         s64 idle_time_ns;
 787         s64 idle_time;
 788
 789         pr = __this_cpu_read(processors);
 790         dev->last_residency = 0;
 791
 792         if (unlikely(!pr))
 793                 return -EINVAL;
 794
 795         local_irq_disable();
 796
 797         if (cx->entry_method != ACPI_CSTATE_FFH) {
 798                 current_thread_info()->status &= ~TS_POLLING;
 799                 /*
 800                  * TS_POLLING-cleared state must be visible before we test
 801                  * NEED_RESCHED:
 802                  */
 803                 smp_mb();
 804
 805                 if (unlikely(need_resched())) {
 806                         current_thread_info()->status |= TS_POLLING;
 807                         local_irq_enable();
 808                         return -EINVAL;
 809                 }
 810         }
 811
 812         /*
 813          * Must be done before busmaster disable as we might need to
 814          * access HPET !
 815          */
 816         lapic_timer_state_broadcast(pr, cx, 1);
 817
 818         if (cx->type == ACPI_STATE_C3)
 819                 ACPI_FLUSH_CPU_CACHE();
 820
 821         kt1 = ktime_get_real();
 822         /* Tell the scheduler that we are going deep-idle: */
 823         sched_clock_idle_sleep_event();
 824         acpi_idle_do_entry(cx);
 825         kt2 = ktime_get_real();
 826         idle_time_ns = ktime_to_ns(ktime_sub(kt2, kt1));
 827         idle_time = idle_time_ns;
 828         do_div(idle_time, NSEC_PER_USEC);
 829
 830         /* Update device last_residency*/
 831         dev->last_residency = (int)idle_time;
 832
 833         /* Tell the scheduler how much we idled: */
 834         sched_clock_idle_wakeup_event(idle_time_ns);
 835
 836         local_irq_enable();
 837         if (cx->entry_method != ACPI_CSTATE_FFH)
 838                 current_thread_info()->status |= TS_POLLING;
 839
 840         cx->usage++;
 841
 842         lapic_timer_state_broadcast(pr, cx, 0);
 843         cx->time += idle_time;
 844         return index;
 845 }
 846
 847 static int c3_cpu_count;
 848 static DEFINE_RAW_SPINLOCK(c3_lock);
 849
 850 /**
 851  * acpi_idle_enter_bm - enters C3 with proper BM handling
 852  * @dev: the target CPU
 853  * @drv: cpuidle driver containing state data
 854  * @index: the index of suggested state
 855  *
 856  * If BM is detected, the deepest non-C3 idle state is entered instead.
 857  */
 858 static int acpi_idle_enter_bm(struct cpuidle_device *dev,
 859                 struct cpuidle_driver *drv, int index)
 860 {
 861         struct acpi_processor *pr;
 862         struct cpuidle_state_usage *state_usage = &dev->states_usage[index];
 863         struct acpi_processor_cx *cx = cpuidle_get_statedata(state_usage);
 864         ktime_t  kt1, kt2;
 865         s64 idle_time_ns;
 866         s64 idle_time;
 867
 868
 869         pr = __this_cpu_read(processors);
 870         dev->last_residency = 0;
 871
 872         if (unlikely(!pr))
 873                 return -EINVAL;
 874
 875         if (!cx->bm_sts_skip && acpi_idle_bm_check()) {
 876                 if (drv->safe_state_index >= 0) {
 877                         return drv->states[drv->safe_state_index].enter(dev,
 878                                                 drv, drv->safe_state_index);
 879                 } else {
 880                         local_irq_disable();
 881                         acpi_safe_halt();
 882                         local_irq_enable();
 883                         return -EINVAL;
 884                 }
 885         }
 886
 887         local_irq_disable();
 888
 889         if (cx->entry_method != ACPI_CSTATE_FFH) {
 890                 current_thread_info()->status &= ~TS_POLLING;
 891                 /*
 892                  * TS_POLLING-cleared state must be visible before we test
 893                  * NEED_RESCHED:
 894                  */
 895                 smp_mb();
 896
 897                 if (unlikely(need_resched())) {
 898                         current_thread_info()->status |= TS_POLLING;
 899                         local_irq_enable();
 900                         return -EINVAL;
 901                 }
 902         }
 903
 904         acpi_unlazy_tlb(smp_processor_id());
 905
 906         /* Tell the scheduler that we are going deep-idle: */
 907         sched_clock_idle_sleep_event();
 908         /*
 909          * Must be done before busmaster disable as we might need to
 910          * access HPET !
 911          */
 912         lapic_timer_state_broadcast(pr, cx, 1);
 913
 914         kt1 = ktime_get_real();
 915         /*
 916          * disable bus master
 917          * bm_check implies we need ARB_DIS
 918          * !bm_check implies we need cache flush
 919          * bm_control implies whether we can do ARB_DIS
 920          *
 921          * That leaves a case where bm_check is set and bm_control is
 922          * not set. In that case we cannot do much, we enter C3
 923          * without doing anything.
 924          */
 925         if (pr->flags.bm_check && pr->flags.bm_control) {
 926                 raw_spin_lock(&c3_lock);
 927                 c3_cpu_count++;
 928                 /* Disable bus master arbitration when all CPUs are in C3 */
 929                 if (c3_cpu_count == num_online_cpus())
 930                         acpi_write_bit_register(ACPI_BITREG_ARB_DISABLE, 1);
 931                 raw_spin_unlock(&c3_lock);
 932         } else if (!pr->flags.bm_check) {
 933                 ACPI_FLUSH_CPU_CACHE();
 934         }
 935
 936         acpi_idle_do_entry(cx);
 937
 938         /* Re-enable bus master arbitration */
 939         if (pr->flags.bm_check && pr->flags.bm_control) {
 940                 raw_spin_lock(&c3_lock);
 941                 acpi_write_bit_register(ACPI_BITREG_ARB_DISABLE, 0);
 942                 c3_cpu_count--;
 943                 raw_spin_unlock(&c3_lock);
 944         }
 945         kt2 = ktime_get_real();
 946         idle_time_ns = ktime_to_ns(ktime_sub(kt2, kt1));
 947         idle_time = idle_time_ns;
 948         do_div(idle_time, NSEC_PER_USEC);
 949
 950         /* Update device last_residency*/
 951         dev->last_residency = (int)idle_time;
 952
 953         /* Tell the scheduler how much we idled: */
 954         sched_clock_idle_wakeup_event(idle_time_ns);
 955
 956         local_irq_enable();
 957         if (cx->entry_method != ACPI_CSTATE_FFH)
 958                 current_thread_info()->status |= TS_POLLING;
 959
 960         cx->usage++;
 961
 962         lapic_timer_state_broadcast(pr, cx, 0);
 963         cx->time += idle_time;
 964         return index;
 965 }
 966
 967 struct cpuidle_driver acpi_idle_driver = {
 968         .name =         "acpi_idle",
 969         .owner =        THIS_MODULE,
 970 };
 971
 972 /**
 973  * acpi_processor_setup_cpuidle_cx - prepares and configures CPUIDLE
 974  * device i.e. per-cpu data
 975  *
 976  * @pr: the ACPI processor
 977  */
 978 static int acpi_processor_setup_cpuidle_cx(struct acpi_processor *pr)
 979 {
 980         int i, count = CPUIDLE_DRIVER_STATE_START;
 981         struct acpi_processor_cx *cx;
 982         struct cpuidle_state_usage *state_usage;
 983         struct cpuidle_device *dev = &pr->power.dev;
 984
 985         if (!pr->flags.power_setup_done)
 986                 return -EINVAL;
 987
 988         if (pr->flags.power == 0) {
 989                 return -EINVAL;
 990         }
 991
 992         if (!dev)
 993                 return -EINVAL;
 994
 995         dev->cpu = pr->id;
 996
 997         if (max_cstate == 0)
 998                 max_cstate = 1;
 999
1000         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER && i <= max_cstate; i++) {
1001                 cx = &pr->power.states[i];
1002                 state_usage = &dev->states_usage[count];
1003
1004                 if (!cx->valid)
1005                         continue;
1006
1007 #ifdef CONFIG_HOTPLUG_CPU
1008                 if ((cx->type != ACPI_STATE_C1) && (num_online_cpus() > 1) &&
1009                     !pr->flags.has_cst &&
1010                     !(acpi_gbl_FADT.flags & ACPI_FADT_C2_MP_SUPPORTED))
1011                         continue;
1012 #endif
1013
1014                 cpuidle_set_statedata(state_usage, cx);
1015
1016                 count++;
1017                 if (count == CPUIDLE_STATE_MAX)
1018                         break;
1019         }
1020
1021         dev->state_count = count;
1022
1023         if (!count)
1024                 return -EINVAL;
1025
1026         return 0;
1027 }
1028
1029 /**
1030  * acpi_processor_setup_cpuidle states- prepares and configures cpuidle
1031  * global state data i.e. idle routines
1032  *
1033  * @pr: the ACPI processor
1034  */
1035 static int acpi_processor_setup_cpuidle_states(struct acpi_processor *pr)
1036 {
1037         int i, count = CPUIDLE_DRIVER_STATE_START;
1038         struct acpi_processor_cx *cx;
1039         struct cpuidle_state *state;
1040         struct cpuidle_driver *drv = &acpi_idle_driver;
1041
1042         if (!pr->flags.power_setup_done)
1043                 return -EINVAL;
1044
1045         if (pr->flags.power == 0)
1046                 return -EINVAL;
1047
1048         drv->safe_state_index = -1;
1049         for (i = 0; i < CPUIDLE_STATE_MAX; i++) {
1050                 drv->states[i].name[0] = '\0';
1051                 drv->states[i].desc[0] = '\0';
1052         }
1053
1054         if (max_cstate == 0)
1055                 max_cstate = 1;
1056
1057         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER && i <= max_cstate; i++) {
1058                 cx = &pr->power.states[i];
1059
1060                 if (!cx->valid)
1061                         continue;
1062
1063 #ifdef CONFIG_HOTPLUG_CPU
1064                 if ((cx->type != ACPI_STATE_C1) && (num_online_cpus() > 1) &&
1065                     !pr->flags.has_cst &&
1066                     !(acpi_gbl_FADT.flags & ACPI_FADT_C2_MP_SUPPORTED))
1067                         continue;
1068 #endif
1069
1070                 state = &drv->states[count];
1071                 snprintf(state->name, CPUIDLE_NAME_LEN, "C%d", i);
1072                 strncpy(state->desc, cx->desc, CPUIDLE_DESC_LEN);
1073                 state->exit_latency = cx->latency;
1074                 state->target_residency = cx->latency * latency_factor;
1075
1076                 state->flags = 0;
1077                 switch (cx->type) {
1078                         case ACPI_STATE_C1:
1079                         if (cx->entry_method == ACPI_CSTATE_FFH)
1080                                 state->flags |= CPUIDLE_FLAG_TIME_VALID;
1081
1082                         state->enter = acpi_idle_enter_c1;
1083                         drv->safe_state_index = count;
1084                         break;
1085
1086                         case ACPI_STATE_C2:
1087                         state->flags |= CPUIDLE_FLAG_TIME_VALID;
1088                         state->enter = acpi_idle_enter_simple;
1089                         drv->safe_state_index = count;
1090                         break;
1091
1092                         case ACPI_STATE_C3:
1093                         state->flags |= CPUIDLE_FLAG_TIME_VALID;
1094                         state->enter = pr->flags.bm_check ?
1095                                         acpi_idle_enter_bm :
1096                                         acpi_idle_enter_simple;
1097                         break;
1098                 }
1099
1100                 count++;
1101                 if (count == CPUIDLE_STATE_MAX)
1102                         break;
1103         }
1104
1105         drv->state_count = count;
1106
1107         if (!count)
1108                 return -EINVAL;
1109
1110         return 0;
1111 }
1112
1113 int acpi_processor_hotplug(struct acpi_processor *pr)
1114 {
1115         int ret = 0;
1116
1117         if (disabled_by_idle_boot_param())
1118                 return 0;
1119
1120         if (!pr)
1121                 return -EINVAL;
1122
1123         if (nocst) {
1124                 return -ENODEV;
1125         }
1126
1127         if (!pr->flags.power_setup_done)
1128                 return -ENODEV;
1129
1130         cpuidle_pause_and_lock();
1131         cpuidle_disable_device(&pr->power.dev);
1132         acpi_processor_get_power_info(pr);
1133         if (pr->flags.power) {
1134                 acpi_processor_setup_cpuidle_cx(pr);
1135                 ret = cpuidle_enable_device(&pr->power.dev);
1136         }
1137         cpuidle_resume_and_unlock();
1138
1139         return ret;
1140 }
1141
1142 int acpi_processor_cst_has_changed(struct acpi_processor *pr)
1143 {
1144         int cpu;
1145         struct acpi_processor *_pr;
1146
1147         if (disabled_by_idle_boot_param())
1148                 return 0;
1149
1150         if (!pr)
1151                 return -EINVAL;
1152
1153         if (nocst)
1154                 return -ENODEV;
1155
1156         if (!pr->flags.power_setup_done)
1157                 return -ENODEV;
1158
1159         /*
1160          * FIXME:  Design the ACPI notification to make it once per
1161          * system instead of once per-cpu.  This condition is a hack
1162          * to make the code that updates C-States be called once.
1163          */
1164
1165         if (smp_processor_id() == 0 &&
1166                         cpuidle_get_driver() == &acpi_idle_driver) {
1167
1168                 /* Protect against cpu-hotplug */
1169                 get_online_cpus();
1170                 cpuidle_pause_and_lock();
1171
1172                 /* Disable all cpuidle devices */
1173                 for_each_online_cpu(cpu) {
1174                         _pr = per_cpu(processors, cpu);
1175                         if (!_pr || !_pr->flags.power_setup_done)
1176                                 continue;
1177                         cpuidle_disable_device(&_pr->power.dev);
1178                 }
1179
1180                 /* Populate Updated C-state information */
1181                 acpi_processor_get_power_info(pr);
1182                 acpi_processor_setup_cpuidle_states(pr);
1183
1184                 /* Enable all cpuidle devices */
1185                 for_each_online_cpu(cpu) {
1186                         _pr = per_cpu(processors, cpu);
1187                         if (!_pr || !_pr->flags.power_setup_done)
1188                                 continue;
1189                         acpi_processor_get_power_info(_pr);
1190                         if (_pr->flags.power) {
1191                                 acpi_processor_setup_cpuidle_cx(_pr);
1192                                 cpuidle_enable_device(&_pr->power.dev);
1193                         }
1194                 }
1195                 cpuidle_resume_and_unlock();
1196                 put_online_cpus();
1197         }
1198
1199         return 0;
1200 }
1201
1202 static int acpi_processor_registered;
1203
1204 int __cpuinit acpi_processor_power_init(struct acpi_processor *pr,
1205                               struct acpi_device *device)
1206 {
1207         acpi_status status = 0;
1208         int retval;
1209         static int first_run;
1210
1211         if (disabled_by_idle_boot_param())
1212                 return 0;
1213
1214         if (!first_run) {
1215                 dmi_check_system(processor_power_dmi_table);
1216                 max_cstate = acpi_processor_cstate_check(max_cstate);
1217                 if (max_cstate < ACPI_C_STATES_MAX)
1218                         printk(KERN_NOTICE
1219                                "ACPI: processor limited to max C-state %d\n",
1220                                max_cstate);
1221                 first_run++;
1222         }
1223
1224         if (!pr)
1225                 return -EINVAL;
1226
1227         if (acpi_gbl_FADT.cst_control && !nocst) {
1228                 status =
1229                     acpi_os_write_port(acpi_gbl_FADT.smi_command, acpi_gbl_FADT.cst_control, 8);
1230                 if (ACPI_FAILURE(status)) {
1231                         ACPI_EXCEPTION((AE_INFO, status,
1232                                         "Notifying BIOS of _CST ability failed"));
1233                 }
1234         }
1235
1236         acpi_processor_get_power_info(pr);
1237         pr->flags.power_setup_done = 1;
1238
1239         /*
1240          * Install the idle handler if processor power management is supported.
1241          * Note that we use previously set idle handler will be used on
1242          * platforms that only support C1.
1243          */
1244         if (pr->flags.power) {
1245                 /* Register acpi_idle_driver if not already registered */
1246                 if (!acpi_processor_registered) {
1247                         acpi_processor_setup_cpuidle_states(pr);
1248                         retval = cpuidle_register_driver(&acpi_idle_driver);
1249                         if (retval)
1250                                 return retval;
1251                         printk(KERN_DEBUG "ACPI: %s registered with cpuidle\n",
1252                                         acpi_idle_driver.name);
1253                 }
1254                 /* Register per-cpu cpuidle_device. Cpuidle driver
1255                  * must already be registered before registering device
1256                  */
1257                 acpi_processor_setup_cpuidle_cx(pr);
1258                 retval = cpuidle_register_device(&pr->power.dev);
1259                 if (retval) {
1260                         if (acpi_processor_registered == 0)
1261                                 cpuidle_unregister_driver(&acpi_idle_driver);
1262                         return retval;
1263                 }
1264                 acpi_processor_registered++;
1265         }
1266         return 0;
1267 }
1268
1269 int acpi_processor_power_exit(struct acpi_processor *pr,
1270                               struct acpi_device *device)
1271 {
1272         if (disabled_by_idle_boot_param())
1273                 return 0;
1274
1275         if (pr->flags.power) {
1276                 cpuidle_unregister_device(&pr->power.dev);
1277                 acpi_processor_registered--;
1278                 if (acpi_processor_registered == 0)
1279                         cpuidle_unregister_driver(&acpi_idle_driver);
1280         }
1281
1282         pr->flags.power_setup_done = 0;
1283         return 0;
1284 }