Pileus Git - ~andy/linux/blob - 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/ACPI_PM_TIMER_FREQUENCY)
  68 #define C2_OVERHEAD                     1       /* 1us */
  69 #define C3_OVERHEAD                     1       /* 1us */
  70 #define PM_TIMER_TICKS_TO_US(p)         \
  71         (((p) * 1000)/(ACPI_PM_TIMER_FREQUENCY/1000))
  72
  73 static unsigned int max_cstate __read_mostly = ACPI_PROCESSOR_MAX_POWER;
  74 module_param(max_cstate, uint, 0000);
  75 static unsigned int nocst __read_mostly;
  76 module_param(nocst, uint, 0000);
  77 static int bm_check_disable __read_mostly;
  78 module_param(bm_check_disable, uint, 0000);
  79
  80 static unsigned int latency_factor __read_mostly = 2;
  81 module_param(latency_factor, uint, 0644);
  82
  83 static DEFINE_PER_CPU(struct cpuidle_device *, acpi_cpuidle_device);
  84
  85 static int disabled_by_idle_boot_param(void)
  86 {
  87         return boot_option_idle_override == IDLE_POLL ||
  88                 boot_option_idle_override == IDLE_FORCE_MWAIT ||
  89                 boot_option_idle_override == IDLE_HALT;
  90 }
  91
  92 /*
  93  * IBM ThinkPad R40e crashes mysteriously when going into C2 or C3.
  94  * For now disable this. Probably a bug somewhere else.
  95  *
  96  * To skip this limit, boot/load with a large max_cstate limit.
  97  */
  98 static int set_max_cstate(const struct dmi_system_id *id)
  99 {
 100         if (max_cstate > ACPI_PROCESSOR_MAX_POWER)
 101                 return 0;
 102
 103         printk(KERN_NOTICE PREFIX "%s detected - limiting to C%ld max_cstate."
 104                " Override with \"processor.max_cstate=%d\"\n", id->ident,
 105                (long)id->driver_data, ACPI_PROCESSOR_MAX_POWER + 1);
 106
 107         max_cstate = (long)id->driver_data;
 108
 109         return 0;
 110 }
 111
 112 /* Actually this shouldn't be __cpuinitdata, would be better to fix the
 113    callers to only run once -AK */
 114 static struct dmi_system_id __cpuinitdata processor_power_dmi_table[] = {
 115         { set_max_cstate, "Clevo 5600D", {
 116           DMI_MATCH(DMI_BIOS_VENDOR,"Phoenix Technologies LTD"),
 117           DMI_MATCH(DMI_BIOS_VERSION,"SHE845M0.86C.0013.D.0302131307")},
 118          (void *)2},
 119         { set_max_cstate, "Pavilion zv5000", {
 120           DMI_MATCH(DMI_SYS_VENDOR, "Hewlett-Packard"),
 121           DMI_MATCH(DMI_PRODUCT_NAME,"Pavilion zv5000 (DS502A#ABA)")},
 122          (void *)1},
 123         { set_max_cstate, "Asus L8400B", {
 124           DMI_MATCH(DMI_SYS_VENDOR, "ASUSTeK Computer Inc."),
 125           DMI_MATCH(DMI_PRODUCT_NAME,"L8400B series Notebook PC")},
 126          (void *)1},
 127         {},
 128 };
 129
 130
 131 /*
 132  * Callers should disable interrupts before the call and enable
 133  * interrupts after return.
 134  */
 135 static void acpi_safe_halt(void)
 136 {
 137         current_thread_info()->status &= ~TS_POLLING;
 138         /*
 139          * TS_POLLING-cleared state must be visible before we
 140          * test NEED_RESCHED:
 141          */
 142         smp_mb();
 143         if (!need_resched()) {
 144                 safe_halt();
 145                 local_irq_disable();
 146         }
 147         current_thread_info()->status |= TS_POLLING;
 148 }
 149
 150 #ifdef ARCH_APICTIMER_STOPS_ON_C3
 151
 152 /*
 153  * Some BIOS implementations switch to C3 in the published C2 state.
 154  * This seems to be a common problem on AMD boxen, but other vendors
 155  * are affected too. We pick the most conservative approach: we assume
 156  * that the local APIC stops in both C2 and C3.
 157  */
 158 static void lapic_timer_check_state(int state, struct acpi_processor *pr,
 159                                    struct acpi_processor_cx *cx)
 160 {
 161         struct acpi_processor_power *pwr = &pr->power;
 162         u8 type = local_apic_timer_c2_ok ? ACPI_STATE_C3 : ACPI_STATE_C2;
 163
 164         if (cpu_has(&cpu_data(pr->id), X86_FEATURE_ARAT))
 165                 return;
 166
 167         if (amd_e400_c1e_detected)
 168                 type = ACPI_STATE_C1;
 169
 170         /*
 171          * Check, if one of the previous states already marked the lapic
 172          * unstable
 173          */
 174         if (pwr->timer_broadcast_on_state < state)
 175                 return;
 176
 177         if (cx->type >= type)
 178                 pr->power.timer_broadcast_on_state = state;
 179 }
 180
 181 static void __lapic_timer_propagate_broadcast(void *arg)
 182 {
 183         struct acpi_processor *pr = (struct acpi_processor *) arg;
 184         unsigned long reason;
 185
 186         reason = pr->power.timer_broadcast_on_state < INT_MAX ?
 187                 CLOCK_EVT_NOTIFY_BROADCAST_ON : CLOCK_EVT_NOTIFY_BROADCAST_OFF;
 188
 189         clockevents_notify(reason, &pr->id);
 190 }
 191
 192 static void lapic_timer_propagate_broadcast(struct acpi_processor *pr)
 193 {
 194         smp_call_function_single(pr->id, __lapic_timer_propagate_broadcast,
 195                                  (void *)pr, 1);
 196 }
 197
 198 /* Power(C) State timer broadcast control */
 199 static void lapic_timer_state_broadcast(struct acpi_processor *pr,
 200                                        struct acpi_processor_cx *cx,
 201                                        int broadcast)
 202 {
 203         int state = cx - pr->power.states;
 204
 205         if (state >= pr->power.timer_broadcast_on_state) {
 206                 unsigned long reason;
 207
 208                 reason = broadcast ?  CLOCK_EVT_NOTIFY_BROADCAST_ENTER :
 209                         CLOCK_EVT_NOTIFY_BROADCAST_EXIT;
 210                 clockevents_notify(reason, &pr->id);
 211         }
 212 }
 213
 214 #else
 215
 216 static void lapic_timer_check_state(int state, struct acpi_processor *pr,
 217                                    struct acpi_processor_cx *cstate) { }
 218 static void lapic_timer_propagate_broadcast(struct acpi_processor *pr) { }
 219 static void lapic_timer_state_broadcast(struct acpi_processor *pr,
 220                                        struct acpi_processor_cx *cx,
 221                                        int broadcast)
 222 {
 223 }
 224
 225 #endif
 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 device *dev)
 244 {
 245         acpi_idle_bm_rld_save();
 246         return 0;
 247 }
 248
 249 int acpi_processor_resume(struct device *dev)
 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.c2_latency;
 308         pr->power.states[ACPI_STATE_C3].latency = acpi_gbl_FADT.c3_latency;
 309
 310         /*
 311          * FADT specified C2 latency must be less than or equal to
 312          * 100 microseconds.
 313          */
 314         if (acpi_gbl_FADT.c2_latency > ACPI_PROCESSOR_MAX_C2_LATENCY) {
 315                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 316                         "C2 latency too large [%d]\n", acpi_gbl_FADT.c2_latency));
 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.c3_latency > ACPI_PROCESSOR_MAX_C3_LATENCY) {
 326                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 327                         "C3 latency too large [%d]\n", acpi_gbl_FADT.c3_latency));
 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                 current_count++;
 490                 memcpy(&(pr->power.states[current_count]), &cx, sizeof(cx));
 491
 492                 /*
 493                  * We support total ACPI_PROCESSOR_MAX_POWER - 1
 494                  * (From 1 through ACPI_PROCESSOR_MAX_POWER - 1)
 495                  */
 496                 if (current_count >= (ACPI_PROCESSOR_MAX_POWER - 1)) {
 497                         printk(KERN_WARNING
 498                                "Limiting number of power states to max (%d)\n",
 499                                ACPI_PROCESSOR_MAX_POWER);
 500                         printk(KERN_WARNING
 501                                "Please increase ACPI_PROCESSOR_MAX_POWER if needed.\n");
 502                         break;
 503                 }
 504         }
 505
 506         ACPI_DEBUG_PRINT((ACPI_DB_INFO, "Found %d power states\n",
 507                           current_count));
 508
 509         /* Validate number of power states discovered */
 510         if (current_count < 2)
 511                 status = -EFAULT;
 512
 513       end:
 514         kfree(buffer.pointer);
 515
 516         return status;
 517 }
 518
 519 static void acpi_processor_power_verify_c3(struct acpi_processor *pr,
 520                                            struct acpi_processor_cx *cx)
 521 {
 522         static int bm_check_flag = -1;
 523         static int bm_control_flag = -1;
 524
 525
 526         if (!cx->address)
 527                 return;
 528
 529         /*
 530          * PIIX4 Erratum #18: We don't support C3 when Type-F (fast)
 531          * DMA transfers are used by any ISA device to avoid livelock.
 532          * Note that we could disable Type-F DMA (as recommended by
 533          * the erratum), but this is known to disrupt certain ISA
 534          * devices thus we take the conservative approach.
 535          */
 536         else if (errata.piix4.fdma) {
 537                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 538                                   "C3 not supported on PIIX4 with Type-F DMA\n"));
 539                 return;
 540         }
 541
 542         /* All the logic here assumes flags.bm_check is same across all CPUs */
 543         if (bm_check_flag == -1) {
 544                 /* Determine whether bm_check is needed based on CPU  */
 545                 acpi_processor_power_init_bm_check(&(pr->flags), pr->id);
 546                 bm_check_flag = pr->flags.bm_check;
 547                 bm_control_flag = pr->flags.bm_control;
 548         } else {
 549                 pr->flags.bm_check = bm_check_flag;
 550                 pr->flags.bm_control = bm_control_flag;
 551         }
 552
 553         if (pr->flags.bm_check) {
 554                 if (!pr->flags.bm_control) {
 555                         if (pr->flags.has_cst != 1) {
 556                                 /* bus mastering control is necessary */
 557                                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 558                                         "C3 support requires BM control\n"));
 559                                 return;
 560                         } else {
 561                                 /* Here we enter C3 without bus mastering */
 562                                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 563                                         "C3 support without BM control\n"));
 564                         }
 565                 }
 566         } else {
 567                 /*
 568                  * WBINVD should be set in fadt, for C3 state to be
 569                  * supported on when bm_check is not required.
 570                  */
 571                 if (!(acpi_gbl_FADT.flags & ACPI_FADT_WBINVD)) {
 572                         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 573                                           "Cache invalidation should work properly"
 574                                           " for C3 to be enabled on SMP systems\n"));
 575                         return;
 576                 }
 577         }
 578
 579         /*
 580          * Otherwise we've met all of our C3 requirements.
 581          * Normalize the C3 latency to expidite policy.  Enable
 582          * checking of bus mastering status (bm_check) so we can
 583          * use this in our C3 policy
 584          */
 585         cx->valid = 1;
 586
 587         /*
 588          * On older chipsets, BM_RLD needs to be set
 589          * in order for Bus Master activity to wake the
 590          * system from C3.  Newer chipsets handle DMA
 591          * during C3 automatically and BM_RLD is a NOP.
 592          * In either case, the proper way to
 593          * handle BM_RLD is to set it and leave it set.
 594          */
 595         acpi_write_bit_register(ACPI_BITREG_BUS_MASTER_RLD, 1);
 596
 597         return;
 598 }
 599
 600 static int acpi_processor_power_verify(struct acpi_processor *pr)
 601 {
 602         unsigned int i;
 603         unsigned int working = 0;
 604
 605         pr->power.timer_broadcast_on_state = INT_MAX;
 606
 607         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER && i <= max_cstate; i++) {
 608                 struct acpi_processor_cx *cx = &pr->power.states[i];
 609
 610                 switch (cx->type) {
 611                 case ACPI_STATE_C1:
 612                         cx->valid = 1;
 613                         break;
 614
 615                 case ACPI_STATE_C2:
 616                         if (!cx->address)
 617                                 break;
 618                         cx->valid = 1;
 619                         break;
 620
 621                 case ACPI_STATE_C3:
 622                         acpi_processor_power_verify_c3(pr, cx);
 623                         break;
 624                 }
 625                 if (!cx->valid)
 626                         continue;
 627
 628                 lapic_timer_check_state(i, pr, cx);
 629                 tsc_check_state(cx->type);
 630                 working++;
 631         }
 632
 633         lapic_timer_propagate_broadcast(pr);
 634
 635         return (working);
 636 }
 637
 638 static int acpi_processor_get_power_info(struct acpi_processor *pr)
 639 {
 640         unsigned int i;
 641         int result;
 642
 643
 644         /* NOTE: the idle thread may not be running while calling
 645          * this function */
 646
 647         /* Zero initialize all the C-states info. */
 648         memset(pr->power.states, 0, sizeof(pr->power.states));
 649
 650         result = acpi_processor_get_power_info_cst(pr);
 651         if (result == -ENODEV)
 652                 result = acpi_processor_get_power_info_fadt(pr);
 653
 654         if (result)
 655                 return result;
 656
 657         acpi_processor_get_power_info_default(pr);
 658
 659         pr->power.count = acpi_processor_power_verify(pr);
 660
 661         /*
 662          * if one state of type C2 or C3 is available, mark this
 663          * CPU as being "idle manageable"
 664          */
 665         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
 666                 if (pr->power.states[i].valid) {
 667                         pr->power.count = i;
 668                         if (pr->power.states[i].type >= ACPI_STATE_C2)
 669                                 pr->flags.power = 1;
 670                 }
 671         }
 672
 673         return 0;
 674 }
 675
 676 /**
 677  * acpi_idle_bm_check - checks if bus master activity was detected
 678  */
 679 static int acpi_idle_bm_check(void)
 680 {
 681         u32 bm_status = 0;
 682
 683         if (bm_check_disable)
 684                 return 0;
 685
 686         acpi_read_bit_register(ACPI_BITREG_BUS_MASTER_STATUS, &bm_status);
 687         if (bm_status)
 688                 acpi_write_bit_register(ACPI_BITREG_BUS_MASTER_STATUS, 1);
 689         /*
 690          * PIIX4 Erratum #18: Note that BM_STS doesn't always reflect
 691          * the true state of bus mastering activity; forcing us to
 692          * manually check the BMIDEA bit of each IDE channel.
 693          */
 694         else if (errata.piix4.bmisx) {
 695                 if ((inb_p(errata.piix4.bmisx + 0x02) & 0x01)
 696                     || (inb_p(errata.piix4.bmisx + 0x0A) & 0x01))
 697                         bm_status = 1;
 698         }
 699         return bm_status;
 700 }
 701
 702 /**
 703  * acpi_idle_do_entry - a helper function that does C2 and C3 type entry
 704  * @cx: cstate data
 705  *
 706  * Caller disables interrupt before call and enables interrupt after return.
 707  */
 708 static inline void acpi_idle_do_entry(struct acpi_processor_cx *cx)
 709 {
 710         /* Don't trace irqs off for idle */
 711         stop_critical_timings();
 712         if (cx->entry_method == ACPI_CSTATE_FFH) {
 713                 /* Call into architectural FFH based C-state */
 714                 acpi_processor_ffh_cstate_enter(cx);
 715         } else if (cx->entry_method == ACPI_CSTATE_HALT) {
 716                 acpi_safe_halt();
 717         } else {
 718                 /* IO port based C-state */
 719                 inb(cx->address);
 720                 /* Dummy wait op - must do something useless after P_LVL2 read
 721                    because chipsets cannot guarantee that STPCLK# signal
 722                    gets asserted in time to freeze execution properly. */
 723                 inl(acpi_gbl_FADT.xpm_timer_block.address);
 724         }
 725         start_critical_timings();
 726 }
 727
 728 /**
 729  * acpi_idle_enter_c1 - enters an ACPI C1 state-type
 730  * @dev: the target CPU
 731  * @drv: cpuidle driver containing cpuidle state info
 732  * @index: index of target state
 733  *
 734  * This is equivalent to the HALT instruction.
 735  */
 736 static int acpi_idle_enter_c1(struct cpuidle_device *dev,
 737                 struct cpuidle_driver *drv, int index)
 738 {
 739         struct acpi_processor *pr;
 740         struct cpuidle_state_usage *state_usage = &dev->states_usage[index];
 741         struct acpi_processor_cx *cx = cpuidle_get_statedata(state_usage);
 742
 743         pr = __this_cpu_read(processors);
 744
 745         if (unlikely(!pr))
 746                 return -EINVAL;
 747
 748         lapic_timer_state_broadcast(pr, cx, 1);
 749         acpi_idle_do_entry(cx);
 750
 751         lapic_timer_state_broadcast(pr, cx, 0);
 752
 753         return index;
 754 }
 755
 756
 757 /**
 758  * acpi_idle_play_dead - enters an ACPI state for long-term idle (i.e. off-lining)
 759  * @dev: the target CPU
 760  * @index: the index of suggested state
 761  */
 762 static int acpi_idle_play_dead(struct cpuidle_device *dev, int index)
 763 {
 764         struct cpuidle_state_usage *state_usage = &dev->states_usage[index];
 765         struct acpi_processor_cx *cx = cpuidle_get_statedata(state_usage);
 766
 767         ACPI_FLUSH_CPU_CACHE();
 768
 769         while (1) {
 770
 771                 if (cx->entry_method == ACPI_CSTATE_HALT)
 772                         safe_halt();
 773                 else if (cx->entry_method == ACPI_CSTATE_SYSTEMIO) {
 774                         inb(cx->address);
 775                         /* See comment in acpi_idle_do_entry() */
 776                         inl(acpi_gbl_FADT.xpm_timer_block.address);
 777                 } else
 778                         return -ENODEV;
 779         }
 780
 781         /* Never reached */
 782         return 0;
 783 }
 784
 785 /**
 786  * acpi_idle_enter_simple - enters an ACPI state without BM handling
 787  * @dev: the target CPU
 788  * @drv: cpuidle driver with cpuidle state information
 789  * @index: the index of suggested state
 790  */
 791 static int acpi_idle_enter_simple(struct cpuidle_device *dev,
 792                 struct cpuidle_driver *drv, int index)
 793 {
 794         struct acpi_processor *pr;
 795         struct cpuidle_state_usage *state_usage = &dev->states_usage[index];
 796         struct acpi_processor_cx *cx = cpuidle_get_statedata(state_usage);
 797
 798         pr = __this_cpu_read(processors);
 799
 800         if (unlikely(!pr))
 801                 return -EINVAL;
 802
 803         if (cx->entry_method != ACPI_CSTATE_FFH) {
 804                 current_thread_info()->status &= ~TS_POLLING;
 805                 /*
 806                  * TS_POLLING-cleared state must be visible before we test
 807                  * NEED_RESCHED:
 808                  */
 809                 smp_mb();
 810
 811                 if (unlikely(need_resched())) {
 812                         current_thread_info()->status |= TS_POLLING;
 813                         return -EINVAL;
 814                 }
 815         }
 816
 817         /*
 818          * Must be done before busmaster disable as we might need to
 819          * access HPET !
 820          */
 821         lapic_timer_state_broadcast(pr, cx, 1);
 822
 823         if (cx->type == ACPI_STATE_C3)
 824                 ACPI_FLUSH_CPU_CACHE();
 825
 826         /* Tell the scheduler that we are going deep-idle: */
 827         sched_clock_idle_sleep_event();
 828         acpi_idle_do_entry(cx);
 829
 830         sched_clock_idle_wakeup_event(0);
 831
 832         if (cx->entry_method != ACPI_CSTATE_FFH)
 833                 current_thread_info()->status |= TS_POLLING;
 834
 835         lapic_timer_state_broadcast(pr, cx, 0);
 836         return index;
 837 }
 838
 839 static int c3_cpu_count;
 840 static DEFINE_RAW_SPINLOCK(c3_lock);
 841
 842 /**
 843  * acpi_idle_enter_bm - enters C3 with proper BM handling
 844  * @dev: the target CPU
 845  * @drv: cpuidle driver containing state data
 846  * @index: the index of suggested state
 847  *
 848  * If BM is detected, the deepest non-C3 idle state is entered instead.
 849  */
 850 static int acpi_idle_enter_bm(struct cpuidle_device *dev,
 851                 struct cpuidle_driver *drv, int index)
 852 {
 853         struct acpi_processor *pr;
 854         struct cpuidle_state_usage *state_usage = &dev->states_usage[index];
 855         struct acpi_processor_cx *cx = cpuidle_get_statedata(state_usage);
 856
 857         pr = __this_cpu_read(processors);
 858
 859         if (unlikely(!pr))
 860                 return -EINVAL;
 861
 862         if (!cx->bm_sts_skip && acpi_idle_bm_check()) {
 863                 if (drv->safe_state_index >= 0) {
 864                         return drv->states[drv->safe_state_index].enter(dev,
 865                                                 drv, drv->safe_state_index);
 866                 } else {
 867                         acpi_safe_halt();
 868                         return -EBUSY;
 869                 }
 870         }
 871
 872         if (cx->entry_method != ACPI_CSTATE_FFH) {
 873                 current_thread_info()->status &= ~TS_POLLING;
 874                 /*
 875                  * TS_POLLING-cleared state must be visible before we test
 876                  * NEED_RESCHED:
 877                  */
 878                 smp_mb();
 879
 880                 if (unlikely(need_resched())) {
 881                         current_thread_info()->status |= TS_POLLING;
 882                         return -EINVAL;
 883                 }
 884         }
 885
 886         acpi_unlazy_tlb(smp_processor_id());
 887
 888         /* Tell the scheduler that we are going deep-idle: */
 889         sched_clock_idle_sleep_event();
 890         /*
 891          * Must be done before busmaster disable as we might need to
 892          * access HPET !
 893          */
 894         lapic_timer_state_broadcast(pr, cx, 1);
 895
 896         /*
 897          * disable bus master
 898          * bm_check implies we need ARB_DIS
 899          * !bm_check implies we need cache flush
 900          * bm_control implies whether we can do ARB_DIS
 901          *
 902          * That leaves a case where bm_check is set and bm_control is
 903          * not set. In that case we cannot do much, we enter C3
 904          * without doing anything.
 905          */
 906         if (pr->flags.bm_check && pr->flags.bm_control) {
 907                 raw_spin_lock(&c3_lock);
 908                 c3_cpu_count++;
 909                 /* Disable bus master arbitration when all CPUs are in C3 */
 910                 if (c3_cpu_count == num_online_cpus())
 911                         acpi_write_bit_register(ACPI_BITREG_ARB_DISABLE, 1);
 912                 raw_spin_unlock(&c3_lock);
 913         } else if (!pr->flags.bm_check) {
 914                 ACPI_FLUSH_CPU_CACHE();
 915         }
 916
 917         acpi_idle_do_entry(cx);
 918
 919         /* Re-enable bus master arbitration */
 920         if (pr->flags.bm_check && pr->flags.bm_control) {
 921                 raw_spin_lock(&c3_lock);
 922                 acpi_write_bit_register(ACPI_BITREG_ARB_DISABLE, 0);
 923                 c3_cpu_count--;
 924                 raw_spin_unlock(&c3_lock);
 925         }
 926
 927         sched_clock_idle_wakeup_event(0);
 928
 929         if (cx->entry_method != ACPI_CSTATE_FFH)
 930                 current_thread_info()->status |= TS_POLLING;
 931
 932         lapic_timer_state_broadcast(pr, cx, 0);
 933         return index;
 934 }
 935
 936 struct cpuidle_driver acpi_idle_driver = {
 937         .name =         "acpi_idle",
 938         .owner =        THIS_MODULE,
 939         .en_core_tk_irqen = 1,
 940 };
 941
 942 /**
 943  * acpi_processor_setup_cpuidle_cx - prepares and configures CPUIDLE
 944  * device i.e. per-cpu data
 945  *
 946  * @pr: the ACPI processor
 947  */
 948 static int acpi_processor_setup_cpuidle_cx(struct acpi_processor *pr)
 949 {
 950         int i, count = CPUIDLE_DRIVER_STATE_START;
 951         struct acpi_processor_cx *cx;
 952         struct cpuidle_state_usage *state_usage;
 953         struct cpuidle_device *dev = per_cpu(acpi_cpuidle_device, pr->id);
 954
 955         if (!pr->flags.power_setup_done)
 956                 return -EINVAL;
 957
 958         if (pr->flags.power == 0) {
 959                 return -EINVAL;
 960         }
 961
 962         if (!dev)
 963                 return -EINVAL;
 964
 965         dev->cpu = pr->id;
 966
 967         if (max_cstate == 0)
 968                 max_cstate = 1;
 969
 970         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER && i <= max_cstate; i++) {
 971                 cx = &pr->power.states[i];
 972                 state_usage = &dev->states_usage[count];
 973
 974                 if (!cx->valid)
 975                         continue;
 976
 977 #ifdef CONFIG_HOTPLUG_CPU
 978                 if ((cx->type != ACPI_STATE_C1) && (num_online_cpus() > 1) &&
 979                     !pr->flags.has_cst &&
 980                     !(acpi_gbl_FADT.flags & ACPI_FADT_C2_MP_SUPPORTED))
 981                         continue;
 982 #endif
 983
 984                 cpuidle_set_statedata(state_usage, cx);
 985
 986                 count++;
 987                 if (count == CPUIDLE_STATE_MAX)
 988                         break;
 989         }
 990
 991         dev->state_count = count;
 992
 993         if (!count)
 994                 return -EINVAL;
 995
 996         return 0;
 997 }
 998
 999 /**
1000  * acpi_processor_setup_cpuidle states- prepares and configures cpuidle
1001  * global state data i.e. idle routines
1002  *
1003  * @pr: the ACPI processor
1004  */
1005 static int acpi_processor_setup_cpuidle_states(struct acpi_processor *pr)
1006 {
1007         int i, count = CPUIDLE_DRIVER_STATE_START;
1008         struct acpi_processor_cx *cx;
1009         struct cpuidle_state *state;
1010         struct cpuidle_driver *drv = &acpi_idle_driver;
1011
1012         if (!pr->flags.power_setup_done)
1013                 return -EINVAL;
1014
1015         if (pr->flags.power == 0)
1016                 return -EINVAL;
1017
1018         drv->safe_state_index = -1;
1019         for (i = 0; i < CPUIDLE_STATE_MAX; i++) {
1020                 drv->states[i].name[0] = '\0';
1021                 drv->states[i].desc[0] = '\0';
1022         }
1023
1024         if (max_cstate == 0)
1025                 max_cstate = 1;
1026
1027         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER && i <= max_cstate; i++) {
1028                 cx = &pr->power.states[i];
1029
1030                 if (!cx->valid)
1031                         continue;
1032
1033 #ifdef CONFIG_HOTPLUG_CPU
1034                 if ((cx->type != ACPI_STATE_C1) && (num_online_cpus() > 1) &&
1035                     !pr->flags.has_cst &&
1036                     !(acpi_gbl_FADT.flags & ACPI_FADT_C2_MP_SUPPORTED))
1037                         continue;
1038 #endif
1039
1040                 state = &drv->states[count];
1041                 snprintf(state->name, CPUIDLE_NAME_LEN, "C%d", i);
1042                 strncpy(state->desc, cx->desc, CPUIDLE_DESC_LEN);
1043                 state->exit_latency = cx->latency;
1044                 state->target_residency = cx->latency * latency_factor;
1045
1046                 state->flags = 0;
1047                 switch (cx->type) {
1048                         case ACPI_STATE_C1:
1049                         if (cx->entry_method == ACPI_CSTATE_FFH)
1050                                 state->flags |= CPUIDLE_FLAG_TIME_VALID;
1051
1052                         state->enter = acpi_idle_enter_c1;
1053                         state->enter_dead = acpi_idle_play_dead;
1054                         drv->safe_state_index = count;
1055                         break;
1056
1057                         case ACPI_STATE_C2:
1058                         state->flags |= CPUIDLE_FLAG_TIME_VALID;
1059                         state->enter = acpi_idle_enter_simple;
1060                         state->enter_dead = acpi_idle_play_dead;
1061                         drv->safe_state_index = count;
1062                         break;
1063
1064                         case ACPI_STATE_C3:
1065                         state->flags |= CPUIDLE_FLAG_TIME_VALID;
1066                         state->enter = pr->flags.bm_check ?
1067                                         acpi_idle_enter_bm :
1068                                         acpi_idle_enter_simple;
1069                         break;
1070                 }
1071
1072                 count++;
1073                 if (count == CPUIDLE_STATE_MAX)
1074                         break;
1075         }
1076
1077         drv->state_count = count;
1078
1079         if (!count)
1080                 return -EINVAL;
1081
1082         return 0;
1083 }
1084
1085 int acpi_processor_hotplug(struct acpi_processor *pr)
1086 {
1087         int ret = 0;
1088         struct cpuidle_device *dev;
1089
1090         if (disabled_by_idle_boot_param())
1091                 return 0;
1092
1093         if (!pr)
1094                 return -EINVAL;
1095
1096         if (nocst) {
1097                 return -ENODEV;
1098         }
1099
1100         if (!pr->flags.power_setup_done)
1101                 return -ENODEV;
1102
1103         dev = per_cpu(acpi_cpuidle_device, pr->id);
1104         cpuidle_pause_and_lock();
1105         cpuidle_disable_device(dev);
1106         acpi_processor_get_power_info(pr);
1107         if (pr->flags.power) {
1108                 acpi_processor_setup_cpuidle_cx(pr);
1109                 ret = cpuidle_enable_device(dev);
1110         }
1111         cpuidle_resume_and_unlock();
1112
1113         return ret;
1114 }
1115
1116 int acpi_processor_cst_has_changed(struct acpi_processor *pr)
1117 {
1118         int cpu;
1119         struct acpi_processor *_pr;
1120         struct cpuidle_device *dev;
1121
1122         if (disabled_by_idle_boot_param())
1123                 return 0;
1124
1125         if (!pr)
1126                 return -EINVAL;
1127
1128         if (nocst)
1129                 return -ENODEV;
1130
1131         if (!pr->flags.power_setup_done)
1132                 return -ENODEV;
1133
1134         /*
1135          * FIXME:  Design the ACPI notification to make it once per
1136          * system instead of once per-cpu.  This condition is a hack
1137          * to make the code that updates C-States be called once.
1138          */
1139
1140         if (pr->id == 0 && cpuidle_get_driver() == &acpi_idle_driver) {
1141
1142                 cpuidle_pause_and_lock();
1143                 /* Protect against cpu-hotplug */
1144                 get_online_cpus();
1145
1146                 /* Disable all cpuidle devices */
1147                 for_each_online_cpu(cpu) {
1148                         _pr = per_cpu(processors, cpu);
1149                         if (!_pr || !_pr->flags.power_setup_done)
1150                                 continue;
1151                         dev = per_cpu(acpi_cpuidle_device, cpu);
1152                         cpuidle_disable_device(dev);
1153                 }
1154
1155                 /* Populate Updated C-state information */
1156                 acpi_processor_get_power_info(pr);
1157                 acpi_processor_setup_cpuidle_states(pr);
1158
1159                 /* Enable all cpuidle devices */
1160                 for_each_online_cpu(cpu) {
1161                         _pr = per_cpu(processors, cpu);
1162                         if (!_pr || !_pr->flags.power_setup_done)
1163                                 continue;
1164                         acpi_processor_get_power_info(_pr);
1165                         if (_pr->flags.power) {
1166                                 acpi_processor_setup_cpuidle_cx(_pr);
1167                                 dev = per_cpu(acpi_cpuidle_device, cpu);
1168                                 cpuidle_enable_device(dev);
1169                         }
1170                 }
1171                 put_online_cpus();
1172                 cpuidle_resume_and_unlock();
1173         }
1174
1175         return 0;
1176 }
1177
1178 static int acpi_processor_registered;
1179
1180 int __cpuinit acpi_processor_power_init(struct acpi_processor *pr)
1181 {
1182         acpi_status status = 0;
1183         int retval;
1184         struct cpuidle_device *dev;
1185         static int first_run;
1186
1187         if (disabled_by_idle_boot_param())
1188                 return 0;
1189
1190         if (!first_run) {
1191                 dmi_check_system(processor_power_dmi_table);
1192                 max_cstate = acpi_processor_cstate_check(max_cstate);
1193                 if (max_cstate < ACPI_C_STATES_MAX)
1194                         printk(KERN_NOTICE
1195                                "ACPI: processor limited to max C-state %d\n",
1196                                max_cstate);
1197                 first_run++;
1198         }
1199
1200         if (!pr)
1201                 return -EINVAL;
1202
1203         if (acpi_gbl_FADT.cst_control && !nocst) {
1204                 status =
1205                     acpi_os_write_port(acpi_gbl_FADT.smi_command, acpi_gbl_FADT.cst_control, 8);
1206                 if (ACPI_FAILURE(status)) {
1207                         ACPI_EXCEPTION((AE_INFO, status,
1208                                         "Notifying BIOS of _CST ability failed"));
1209                 }
1210         }
1211
1212         acpi_processor_get_power_info(pr);
1213         pr->flags.power_setup_done = 1;
1214
1215         /*
1216          * Install the idle handler if processor power management is supported.
1217          * Note that we use previously set idle handler will be used on
1218          * platforms that only support C1.
1219          */
1220         if (pr->flags.power) {
1221                 /* Register acpi_idle_driver if not already registered */
1222                 if (!acpi_processor_registered) {
1223                         acpi_processor_setup_cpuidle_states(pr);
1224                         retval = cpuidle_register_driver(&acpi_idle_driver);
1225                         if (retval)
1226                                 return retval;
1227                         printk(KERN_DEBUG "ACPI: %s registered with cpuidle\n",
1228                                         acpi_idle_driver.name);
1229                 }
1230
1231                 dev = kzalloc(sizeof(*dev), GFP_KERNEL);
1232                 if (!dev)
1233                         return -ENOMEM;
1234                 per_cpu(acpi_cpuidle_device, pr->id) = dev;
1235
1236                 acpi_processor_setup_cpuidle_cx(pr);
1237
1238                 /* Register per-cpu cpuidle_device. Cpuidle driver
1239                  * must already be registered before registering device
1240                  */
1241                 retval = cpuidle_register_device(dev);
1242                 if (retval) {
1243                         if (acpi_processor_registered == 0)
1244                                 cpuidle_unregister_driver(&acpi_idle_driver);
1245                         return retval;
1246                 }
1247                 acpi_processor_registered++;
1248         }
1249         return 0;
1250 }
1251
1252 int acpi_processor_power_exit(struct acpi_processor *pr)
1253 {
1254         struct cpuidle_device *dev = per_cpu(acpi_cpuidle_device, pr->id);
1255
1256         if (disabled_by_idle_boot_param())
1257                 return 0;
1258
1259         if (pr->flags.power) {
1260                 cpuidle_unregister_device(dev);
1261                 acpi_processor_registered--;
1262                 if (acpi_processor_registered == 0)
1263                         cpuidle_unregister_driver(&acpi_idle_driver);
1264         }
1265
1266         pr->flags.power_setup_done = 0;
1267         return 0;
1268 }