SAFE public projects git trees. - safe/jmp/linux-2.6/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/proc_fs.h>
  36 #include <linux/seq_file.h>
  37 #include <linux/acpi.h>
  38 #include <linux/dmi.h>
  39 #include <linux/moduleparam.h>
  40 #include <linux/sched.h>        /* need_resched() */
  41 #include <linux/pm_qos_params.h>
  42 #include <linux/clockchips.h>
  43 #include <linux/cpuidle.h>
  44 #include <linux/irqflags.h>
  45
  46 /*
  47  * Include the apic definitions for x86 to have the APIC timer related defines
  48  * available also for UP (on SMP it gets magically included via linux/smp.h).
  49  * asm/acpi.h is not an option, as it would require more include magic. Also
  50  * creating an empty asm-ia64/apic.h would just trade pest vs. cholera.
  51  */
  52 #ifdef CONFIG_X86
  53 #include <asm/apic.h>
  54 #endif
  55
  56 #include <asm/io.h>
  57 #include <asm/uaccess.h>
  58
  59 #include <acpi/acpi_bus.h>
  60 #include <acpi/processor.h>
  61 #include <asm/processor.h>
  62
  63 #define PREFIX "ACPI: "
  64
  65 #define ACPI_PROCESSOR_CLASS            "processor"
  66 #define _COMPONENT              ACPI_PROCESSOR_COMPONENT
  67 ACPI_MODULE_NAME("processor_idle");
  68 #define ACPI_PROCESSOR_FILE_POWER       "power"
  69 #define PM_TIMER_TICK_NS                (1000000000ULL/PM_TIMER_FREQUENCY)
  70 #define C2_OVERHEAD                     1       /* 1us */
  71 #define C3_OVERHEAD                     1       /* 1us */
  72 #define PM_TIMER_TICKS_TO_US(p)         (((p) * 1000)/(PM_TIMER_FREQUENCY/1000))
  73
  74 static unsigned int max_cstate __read_mostly = ACPI_PROCESSOR_MAX_POWER;
  75 module_param(max_cstate, uint, 0000);
  76 static unsigned int nocst __read_mostly;
  77 module_param(nocst, uint, 0000);
  78
  79 static unsigned int latency_factor __read_mostly = 2;
  80 module_param(latency_factor, uint, 0644);
  81
  82 static s64 us_to_pm_timer_ticks(s64 t)
  83 {
  84         return div64_u64(t * PM_TIMER_FREQUENCY, 1000000);
  85 }
  86 /*
  87  * IBM ThinkPad R40e crashes mysteriously when going into C2 or C3.
  88  * For now disable this. Probably a bug somewhere else.
  89  *
  90  * To skip this limit, boot/load with a large max_cstate limit.
  91  */
  92 static int set_max_cstate(const struct dmi_system_id *id)
  93 {
  94         if (max_cstate > ACPI_PROCESSOR_MAX_POWER)
  95                 return 0;
  96
  97         printk(KERN_NOTICE PREFIX "%s detected - limiting to C%ld max_cstate."
  98                " Override with \"processor.max_cstate=%d\"\n", id->ident,
  99                (long)id->driver_data, ACPI_PROCESSOR_MAX_POWER + 1);
 100
 101         max_cstate = (long)id->driver_data;
 102
 103         return 0;
 104 }
 105
 106 /* Actually this shouldn't be __cpuinitdata, would be better to fix the
 107    callers to only run once -AK */
 108 static struct dmi_system_id __cpuinitdata processor_power_dmi_table[] = {
 109         { set_max_cstate, "Clevo 5600D", {
 110           DMI_MATCH(DMI_BIOS_VENDOR,"Phoenix Technologies LTD"),
 111           DMI_MATCH(DMI_BIOS_VERSION,"SHE845M0.86C.0013.D.0302131307")},
 112          (void *)2},
 113         {},
 114 };
 115
 116
 117 /*
 118  * Callers should disable interrupts before the call and enable
 119  * interrupts after return.
 120  */
 121 static void acpi_safe_halt(void)
 122 {
 123         current_thread_info()->status &= ~TS_POLLING;
 124         /*
 125          * TS_POLLING-cleared state must be visible before we
 126          * test NEED_RESCHED:
 127          */
 128         smp_mb();
 129         if (!need_resched()) {
 130                 safe_halt();
 131                 local_irq_disable();
 132         }
 133         current_thread_info()->status |= TS_POLLING;
 134 }
 135
 136 #ifdef ARCH_APICTIMER_STOPS_ON_C3
 137
 138 /*
 139  * Some BIOS implementations switch to C3 in the published C2 state.
 140  * This seems to be a common problem on AMD boxen, but other vendors
 141  * are affected too. We pick the most conservative approach: we assume
 142  * that the local APIC stops in both C2 and C3.
 143  */
 144 static void lapic_timer_check_state(int state, struct acpi_processor *pr,
 145                                    struct acpi_processor_cx *cx)
 146 {
 147         struct acpi_processor_power *pwr = &pr->power;
 148         u8 type = local_apic_timer_c2_ok ? ACPI_STATE_C3 : ACPI_STATE_C2;
 149
 150         if (cpu_has(&cpu_data(pr->id), X86_FEATURE_ARAT))
 151                 return;
 152
 153         if (boot_cpu_has(X86_FEATURE_AMDC1E))
 154                 type = ACPI_STATE_C1;
 155
 156         /*
 157          * Check, if one of the previous states already marked the lapic
 158          * unstable
 159          */
 160         if (pwr->timer_broadcast_on_state < state)
 161                 return;
 162
 163         if (cx->type >= type)
 164                 pr->power.timer_broadcast_on_state = state;
 165 }
 166
 167 static void __lapic_timer_propagate_broadcast(void *arg)
 168 {
 169         struct acpi_processor *pr = (struct acpi_processor *) arg;
 170         unsigned long reason;
 171
 172         reason = pr->power.timer_broadcast_on_state < INT_MAX ?
 173                 CLOCK_EVT_NOTIFY_BROADCAST_ON : CLOCK_EVT_NOTIFY_BROADCAST_OFF;
 174
 175         clockevents_notify(reason, &pr->id);
 176 }
 177
 178 static void lapic_timer_propagate_broadcast(struct acpi_processor *pr)
 179 {
 180         smp_call_function_single(pr->id, __lapic_timer_propagate_broadcast,
 181                                  (void *)pr, 1);
 182 }
 183
 184 /* Power(C) State timer broadcast control */
 185 static void lapic_timer_state_broadcast(struct acpi_processor *pr,
 186                                        struct acpi_processor_cx *cx,
 187                                        int broadcast)
 188 {
 189         int state = cx - pr->power.states;
 190
 191         if (state >= pr->power.timer_broadcast_on_state) {
 192                 unsigned long reason;
 193
 194                 reason = broadcast ?  CLOCK_EVT_NOTIFY_BROADCAST_ENTER :
 195                         CLOCK_EVT_NOTIFY_BROADCAST_EXIT;
 196                 clockevents_notify(reason, &pr->id);
 197         }
 198 }
 199
 200 #else
 201
 202 static void lapic_timer_check_state(int state, struct acpi_processor *pr,
 203                                    struct acpi_processor_cx *cstate) { }
 204 static void lapic_timer_propagate_broadcast(struct acpi_processor *pr) { }
 205 static void lapic_timer_state_broadcast(struct acpi_processor *pr,
 206                                        struct acpi_processor_cx *cx,
 207                                        int broadcast)
 208 {
 209 }
 210
 211 #endif
 212
 213 /*
 214  * Suspend / resume control
 215  */
 216 static int acpi_idle_suspend;
 217 static u32 saved_bm_rld;
 218
 219 static void acpi_idle_bm_rld_save(void)
 220 {
 221         acpi_read_bit_register(ACPI_BITREG_BUS_MASTER_RLD, &saved_bm_rld);
 222 }
 223 static void acpi_idle_bm_rld_restore(void)
 224 {
 225         u32 resumed_bm_rld;
 226
 227         acpi_read_bit_register(ACPI_BITREG_BUS_MASTER_RLD, &resumed_bm_rld);
 228
 229         if (resumed_bm_rld != saved_bm_rld)
 230                 acpi_write_bit_register(ACPI_BITREG_BUS_MASTER_RLD, saved_bm_rld);
 231 }
 232
 233 int acpi_processor_suspend(struct acpi_device * device, pm_message_t state)
 234 {
 235         if (acpi_idle_suspend == 1)
 236                 return 0;
 237
 238         acpi_idle_bm_rld_save();
 239         acpi_idle_suspend = 1;
 240         return 0;
 241 }
 242
 243 int acpi_processor_resume(struct acpi_device * device)
 244 {
 245         if (acpi_idle_suspend == 0)
 246                 return 0;
 247
 248         acpi_idle_bm_rld_restore();
 249         acpi_idle_suspend = 0;
 250         return 0;
 251 }
 252
 253 #if defined (CONFIG_GENERIC_TIME) && defined (CONFIG_X86)
 254 static void tsc_check_state(int state)
 255 {
 256         switch (boot_cpu_data.x86_vendor) {
 257         case X86_VENDOR_AMD:
 258         case X86_VENDOR_INTEL:
 259                 /*
 260                  * AMD Fam10h TSC will tick in all
 261                  * C/P/S0/S1 states when this bit is set.
 262                  */
 263                 if (boot_cpu_has(X86_FEATURE_NONSTOP_TSC))
 264                         return;
 265
 266                 /*FALL THROUGH*/
 267         default:
 268                 /* TSC could halt in idle, so notify users */
 269                 if (state > ACPI_STATE_C1)
 270                         mark_tsc_unstable("TSC halts in idle");
 271         }
 272 }
 273 #else
 274 static void tsc_check_state(int state) { return; }
 275 #endif
 276
 277 static int acpi_processor_get_power_info_fadt(struct acpi_processor *pr)
 278 {
 279
 280         if (!pr)
 281                 return -EINVAL;
 282
 283         if (!pr->pblk)
 284                 return -ENODEV;
 285
 286         /* if info is obtained from pblk/fadt, type equals state */
 287         pr->power.states[ACPI_STATE_C2].type = ACPI_STATE_C2;
 288         pr->power.states[ACPI_STATE_C3].type = ACPI_STATE_C3;
 289
 290 #ifndef CONFIG_HOTPLUG_CPU
 291         /*
 292          * Check for P_LVL2_UP flag before entering C2 and above on
 293          * an SMP system.
 294          */
 295         if ((num_online_cpus() > 1) &&
 296             !(acpi_gbl_FADT.flags & ACPI_FADT_C2_MP_SUPPORTED))
 297                 return -ENODEV;
 298 #endif
 299
 300         /* determine C2 and C3 address from pblk */
 301         pr->power.states[ACPI_STATE_C2].address = pr->pblk + 4;
 302         pr->power.states[ACPI_STATE_C3].address = pr->pblk + 5;
 303
 304         /* determine latencies from FADT */
 305         pr->power.states[ACPI_STATE_C2].latency = acpi_gbl_FADT.C2latency;
 306         pr->power.states[ACPI_STATE_C3].latency = acpi_gbl_FADT.C3latency;
 307
 308         /*
 309          * FADT specified C2 latency must be less than or equal to
 310          * 100 microseconds.
 311          */
 312         if (acpi_gbl_FADT.C2latency > ACPI_PROCESSOR_MAX_C2_LATENCY) {
 313                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 314                         "C2 latency too large [%d]\n", acpi_gbl_FADT.C2latency));
 315                 /* invalidate C2 */
 316                 pr->power.states[ACPI_STATE_C2].address = 0;
 317         }
 318
 319         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 320                           "lvl2[0x%08x] lvl3[0x%08x]\n",
 321                           pr->power.states[ACPI_STATE_C2].address,
 322                           pr->power.states[ACPI_STATE_C3].address));
 323
 324         return 0;
 325 }
 326
 327 static int acpi_processor_get_power_info_default(struct acpi_processor *pr)
 328 {
 329         if (!pr->power.states[ACPI_STATE_C1].valid) {
 330                 /* set the first C-State to C1 */
 331                 /* all processors need to support C1 */
 332                 pr->power.states[ACPI_STATE_C1].type = ACPI_STATE_C1;
 333                 pr->power.states[ACPI_STATE_C1].valid = 1;
 334                 pr->power.states[ACPI_STATE_C1].entry_method = ACPI_CSTATE_HALT;
 335         }
 336         /* the C0 state only exists as a filler in our array */
 337         pr->power.states[ACPI_STATE_C0].valid = 1;
 338         return 0;
 339 }
 340
 341 static int acpi_processor_get_power_info_cst(struct acpi_processor *pr)
 342 {
 343         acpi_status status = 0;
 344         acpi_integer count;
 345         int current_count;
 346         int i;
 347         struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
 348         union acpi_object *cst;
 349
 350
 351         if (nocst)
 352                 return -ENODEV;
 353
 354         current_count = 0;
 355
 356         status = acpi_evaluate_object(pr->handle, "_CST", NULL, &buffer);
 357         if (ACPI_FAILURE(status)) {
 358                 ACPI_DEBUG_PRINT((ACPI_DB_INFO, "No _CST, giving up\n"));
 359                 return -ENODEV;
 360         }
 361
 362         cst = buffer.pointer;
 363
 364         /* There must be at least 2 elements */
 365         if (!cst || (cst->type != ACPI_TYPE_PACKAGE) || cst->package.count < 2) {
 366                 printk(KERN_ERR PREFIX "not enough elements in _CST\n");
 367                 status = -EFAULT;
 368                 goto end;
 369         }
 370
 371         count = cst->package.elements[0].integer.value;
 372
 373         /* Validate number of power states. */
 374         if (count < 1 || count != cst->package.count - 1) {
 375                 printk(KERN_ERR PREFIX "count given by _CST is not valid\n");
 376                 status = -EFAULT;
 377                 goto end;
 378         }
 379
 380         /* Tell driver that at least _CST is supported. */
 381         pr->flags.has_cst = 1;
 382
 383         for (i = 1; i <= count; i++) {
 384                 union acpi_object *element;
 385                 union acpi_object *obj;
 386                 struct acpi_power_register *reg;
 387                 struct acpi_processor_cx cx;
 388
 389                 memset(&cx, 0, sizeof(cx));
 390
 391                 element = &(cst->package.elements[i]);
 392                 if (element->type != ACPI_TYPE_PACKAGE)
 393                         continue;
 394
 395                 if (element->package.count != 4)
 396                         continue;
 397
 398                 obj = &(element->package.elements[0]);
 399
 400                 if (obj->type != ACPI_TYPE_BUFFER)
 401                         continue;
 402
 403                 reg = (struct acpi_power_register *)obj->buffer.pointer;
 404
 405                 if (reg->space_id != ACPI_ADR_SPACE_SYSTEM_IO &&
 406                     (reg->space_id != ACPI_ADR_SPACE_FIXED_HARDWARE))
 407                         continue;
 408
 409                 /* There should be an easy way to extract an integer... */
 410                 obj = &(element->package.elements[1]);
 411                 if (obj->type != ACPI_TYPE_INTEGER)
 412                         continue;
 413
 414                 cx.type = obj->integer.value;
 415                 /*
 416                  * Some buggy BIOSes won't list C1 in _CST -
 417                  * Let acpi_processor_get_power_info_default() handle them later
 418                  */
 419                 if (i == 1 && cx.type != ACPI_STATE_C1)
 420                         current_count++;
 421
 422                 cx.address = reg->address;
 423                 cx.index = current_count + 1;
 424
 425                 cx.entry_method = ACPI_CSTATE_SYSTEMIO;
 426                 if (reg->space_id == ACPI_ADR_SPACE_FIXED_HARDWARE) {
 427                         if (acpi_processor_ffh_cstate_probe
 428                                         (pr->id, &cx, reg) == 0) {
 429                                 cx.entry_method = ACPI_CSTATE_FFH;
 430                         } else if (cx.type == ACPI_STATE_C1) {
 431                                 /*
 432                                  * C1 is a special case where FIXED_HARDWARE
 433                                  * can be handled in non-MWAIT way as well.
 434                                  * In that case, save this _CST entry info.
 435                                  * Otherwise, ignore this info and continue.
 436                                  */
 437                                 cx.entry_method = ACPI_CSTATE_HALT;
 438                                 snprintf(cx.desc, ACPI_CX_DESC_LEN, "ACPI HLT");
 439                         } else {
 440                                 continue;
 441                         }
 442                         if (cx.type == ACPI_STATE_C1 &&
 443                                         (idle_halt || idle_nomwait)) {
 444                                 /*
 445                                  * In most cases the C1 space_id obtained from
 446                                  * _CST object is FIXED_HARDWARE access mode.
 447                                  * But when the option of idle=halt is added,
 448                                  * the entry_method type should be changed from
 449                                  * CSTATE_FFH to CSTATE_HALT.
 450                                  * When the option of idle=nomwait is added,
 451                                  * the C1 entry_method type should be
 452                                  * CSTATE_HALT.
 453                                  */
 454                                 cx.entry_method = ACPI_CSTATE_HALT;
 455                                 snprintf(cx.desc, ACPI_CX_DESC_LEN, "ACPI HLT");
 456                         }
 457                 } else {
 458                         snprintf(cx.desc, ACPI_CX_DESC_LEN, "ACPI IOPORT 0x%x",
 459                                  cx.address);
 460                 }
 461
 462                 if (cx.type == ACPI_STATE_C1) {
 463                         cx.valid = 1;
 464                 }
 465
 466                 obj = &(element->package.elements[2]);
 467                 if (obj->type != ACPI_TYPE_INTEGER)
 468                         continue;
 469
 470                 cx.latency = obj->integer.value;
 471
 472                 obj = &(element->package.elements[3]);
 473                 if (obj->type != ACPI_TYPE_INTEGER)
 474                         continue;
 475
 476                 cx.power = obj->integer.value;
 477
 478                 current_count++;
 479                 memcpy(&(pr->power.states[current_count]), &cx, sizeof(cx));
 480
 481                 /*
 482                  * We support total ACPI_PROCESSOR_MAX_POWER - 1
 483                  * (From 1 through ACPI_PROCESSOR_MAX_POWER - 1)
 484                  */
 485                 if (current_count >= (ACPI_PROCESSOR_MAX_POWER - 1)) {
 486                         printk(KERN_WARNING
 487                                "Limiting number of power states to max (%d)\n",
 488                                ACPI_PROCESSOR_MAX_POWER);
 489                         printk(KERN_WARNING
 490                                "Please increase ACPI_PROCESSOR_MAX_POWER if needed.\n");
 491                         break;
 492                 }
 493         }
 494
 495         ACPI_DEBUG_PRINT((ACPI_DB_INFO, "Found %d power states\n",
 496                           current_count));
 497
 498         /* Validate number of power states discovered */
 499         if (current_count < 2)
 500                 status = -EFAULT;
 501
 502       end:
 503         kfree(buffer.pointer);
 504
 505         return status;
 506 }
 507
 508 static void acpi_processor_power_verify_c2(struct acpi_processor_cx *cx)
 509 {
 510
 511         if (!cx->address)
 512                 return;
 513
 514         /*
 515          * Otherwise we've met all of our C2 requirements.
 516          * Normalize the C2 latency to expidite policy
 517          */
 518         cx->valid = 1;
 519
 520         cx->latency_ticks = cx->latency;
 521
 522         return;
 523 }
 524
 525 static void acpi_processor_power_verify_c3(struct acpi_processor *pr,
 526                                            struct acpi_processor_cx *cx)
 527 {
 528         static int bm_check_flag = -1;
 529         static int bm_control_flag = -1;
 530
 531
 532         if (!cx->address)
 533                 return;
 534
 535         /*
 536          * C3 latency must be less than or equal to 1000
 537          * microseconds.
 538          */
 539         else if (cx->latency > ACPI_PROCESSOR_MAX_C3_LATENCY) {
 540                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 541                                   "latency too large [%d]\n", cx->latency));
 542                 return;
 543         }
 544
 545         /*
 546          * PIIX4 Erratum #18: We don't support C3 when Type-F (fast)
 547          * DMA transfers are used by any ISA device to avoid livelock.
 548          * Note that we could disable Type-F DMA (as recommended by
 549          * the erratum), but this is known to disrupt certain ISA
 550          * devices thus we take the conservative approach.
 551          */
 552         else if (errata.piix4.fdma) {
 553                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 554                                   "C3 not supported on PIIX4 with Type-F DMA\n"));
 555                 return;
 556         }
 557
 558         /* All the logic here assumes flags.bm_check is same across all CPUs */
 559         if (bm_check_flag == -1) {
 560                 /* Determine whether bm_check is needed based on CPU  */
 561                 acpi_processor_power_init_bm_check(&(pr->flags), pr->id);
 562                 bm_check_flag = pr->flags.bm_check;
 563                 bm_control_flag = pr->flags.bm_control;
 564         } else {
 565                 pr->flags.bm_check = bm_check_flag;
 566                 pr->flags.bm_control = bm_control_flag;
 567         }
 568
 569         if (pr->flags.bm_check) {
 570                 if (!pr->flags.bm_control) {
 571                         if (pr->flags.has_cst != 1) {
 572                                 /* bus mastering control is necessary */
 573                                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 574                                         "C3 support requires BM control\n"));
 575                                 return;
 576                         } else {
 577                                 /* Here we enter C3 without bus mastering */
 578                                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 579                                         "C3 support without BM control\n"));
 580                         }
 581                 }
 582         } else {
 583                 /*
 584                  * WBINVD should be set in fadt, for C3 state to be
 585                  * supported on when bm_check is not required.
 586                  */
 587                 if (!(acpi_gbl_FADT.flags & ACPI_FADT_WBINVD)) {
 588                         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 589                                           "Cache invalidation should work properly"
 590                                           " for C3 to be enabled on SMP systems\n"));
 591                         return;
 592                 }
 593         }
 594
 595         /*
 596          * Otherwise we've met all of our C3 requirements.
 597          * Normalize the C3 latency to expidite policy.  Enable
 598          * checking of bus mastering status (bm_check) so we can
 599          * use this in our C3 policy
 600          */
 601         cx->valid = 1;
 602
 603         cx->latency_ticks = cx->latency;
 604         /*
 605          * On older chipsets, BM_RLD needs to be set
 606          * in order for Bus Master activity to wake the
 607          * system from C3.  Newer chipsets handle DMA
 608          * during C3 automatically and BM_RLD is a NOP.
 609          * In either case, the proper way to
 610          * handle BM_RLD is to set it and leave it set.
 611          */
 612         acpi_write_bit_register(ACPI_BITREG_BUS_MASTER_RLD, 1);
 613
 614         return;
 615 }
 616
 617 static int acpi_processor_power_verify(struct acpi_processor *pr)
 618 {
 619         unsigned int i;
 620         unsigned int working = 0;
 621
 622         pr->power.timer_broadcast_on_state = INT_MAX;
 623
 624         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER && i <= max_cstate; i++) {
 625                 struct acpi_processor_cx *cx = &pr->power.states[i];
 626
 627                 switch (cx->type) {
 628                 case ACPI_STATE_C1:
 629                         cx->valid = 1;
 630                         break;
 631
 632                 case ACPI_STATE_C2:
 633                         acpi_processor_power_verify_c2(cx);
 634                         break;
 635
 636                 case ACPI_STATE_C3:
 637                         acpi_processor_power_verify_c3(pr, cx);
 638                         break;
 639                 }
 640                 if (!cx->valid)
 641                         continue;
 642
 643                 lapic_timer_check_state(i, pr, cx);
 644                 tsc_check_state(cx->type);
 645                 working++;
 646         }
 647
 648         lapic_timer_propagate_broadcast(pr);
 649
 650         return (working);
 651 }
 652
 653 static int acpi_processor_get_power_info(struct acpi_processor *pr)
 654 {
 655         unsigned int i;
 656         int result;
 657
 658
 659         /* NOTE: the idle thread may not be running while calling
 660          * this function */
 661
 662         /* Zero initialize all the C-states info. */
 663         memset(pr->power.states, 0, sizeof(pr->power.states));
 664
 665         result = acpi_processor_get_power_info_cst(pr);
 666         if (result == -ENODEV)
 667                 result = acpi_processor_get_power_info_fadt(pr);
 668
 669         if (result)
 670                 return result;
 671
 672         acpi_processor_get_power_info_default(pr);
 673
 674         pr->power.count = acpi_processor_power_verify(pr);
 675
 676         /*
 677          * if one state of type C2 or C3 is available, mark this
 678          * CPU as being "idle manageable"
 679          */
 680         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
 681                 if (pr->power.states[i].valid) {
 682                         pr->power.count = i;
 683                         if (pr->power.states[i].type >= ACPI_STATE_C2)
 684                                 pr->flags.power = 1;
 685                 }
 686         }
 687
 688         return 0;
 689 }
 690
 691 #ifdef CONFIG_ACPI_PROCFS
 692 static int acpi_processor_power_seq_show(struct seq_file *seq, void *offset)
 693 {
 694         struct acpi_processor *pr = seq->private;
 695         unsigned int i;
 696
 697
 698         if (!pr)
 699                 goto end;
 700
 701         seq_printf(seq, "active state:            C%zd\n"
 702                    "max_cstate:              C%d\n"
 703                    "maximum allowed latency: %d usec\n",
 704                    pr->power.state ? pr->power.state - pr->power.states : 0,
 705                    max_cstate, pm_qos_requirement(PM_QOS_CPU_DMA_LATENCY));
 706
 707         seq_puts(seq, "states:\n");
 708
 709         for (i = 1; i <= pr->power.count; i++) {
 710                 seq_printf(seq, "   %cC%d:                  ",
 711                            (&pr->power.states[i] ==
 712                             pr->power.state ? '*' : ' '), i);
 713
 714                 if (!pr->power.states[i].valid) {
 715                         seq_puts(seq, "<not supported>\n");
 716                         continue;
 717                 }
 718
 719                 switch (pr->power.states[i].type) {
 720                 case ACPI_STATE_C1:
 721                         seq_printf(seq, "type[C1] ");
 722                         break;
 723                 case ACPI_STATE_C2:
 724                         seq_printf(seq, "type[C2] ");
 725                         break;
 726                 case ACPI_STATE_C3:
 727                         seq_printf(seq, "type[C3] ");
 728                         break;
 729                 default:
 730                         seq_printf(seq, "type[--] ");
 731                         break;
 732                 }
 733
 734                 if (pr->power.states[i].promotion.state)
 735                         seq_printf(seq, "promotion[C%zd] ",
 736                                    (pr->power.states[i].promotion.state -
 737                                     pr->power.states));
 738                 else
 739                         seq_puts(seq, "promotion[--] ");
 740
 741                 if (pr->power.states[i].demotion.state)
 742                         seq_printf(seq, "demotion[C%zd] ",
 743                                    (pr->power.states[i].demotion.state -
 744                                     pr->power.states));
 745                 else
 746                         seq_puts(seq, "demotion[--] ");
 747
 748                 seq_printf(seq, "latency[%03d] usage[%08d] duration[%020llu]\n",
 749                            pr->power.states[i].latency,
 750                            pr->power.states[i].usage,
 751                            (unsigned long long)pr->power.states[i].time);
 752         }
 753
 754       end:
 755         return 0;
 756 }
 757
 758 static int acpi_processor_power_open_fs(struct inode *inode, struct file *file)
 759 {
 760         return single_open(file, acpi_processor_power_seq_show,
 761                            PDE(inode)->data);
 762 }
 763
 764 static const struct file_operations acpi_processor_power_fops = {
 765         .owner = THIS_MODULE,
 766         .open = acpi_processor_power_open_fs,
 767         .read = seq_read,
 768         .llseek = seq_lseek,
 769         .release = single_release,
 770 };
 771 #endif
 772
 773 /**
 774  * acpi_idle_bm_check - checks if bus master activity was detected
 775  */
 776 static int acpi_idle_bm_check(void)
 777 {
 778         u32 bm_status = 0;
 779
 780         acpi_read_bit_register(ACPI_BITREG_BUS_MASTER_STATUS, &bm_status);
 781         if (bm_status)
 782                 acpi_write_bit_register(ACPI_BITREG_BUS_MASTER_STATUS, 1);
 783         /*
 784          * PIIX4 Erratum #18: Note that BM_STS doesn't always reflect
 785          * the true state of bus mastering activity; forcing us to
 786          * manually check the BMIDEA bit of each IDE channel.
 787          */
 788         else if (errata.piix4.bmisx) {
 789                 if ((inb_p(errata.piix4.bmisx + 0x02) & 0x01)
 790                     || (inb_p(errata.piix4.bmisx + 0x0A) & 0x01))
 791                         bm_status = 1;
 792         }
 793         return bm_status;
 794 }
 795
 796 /**
 797  * acpi_idle_do_entry - a helper function that does C2 and C3 type entry
 798  * @cx: cstate data
 799  *
 800  * Caller disables interrupt before call and enables interrupt after return.
 801  */
 802 static inline void acpi_idle_do_entry(struct acpi_processor_cx *cx)
 803 {
 804         /* Don't trace irqs off for idle */
 805         stop_critical_timings();
 806         if (cx->entry_method == ACPI_CSTATE_FFH) {
 807                 /* Call into architectural FFH based C-state */
 808                 acpi_processor_ffh_cstate_enter(cx);
 809         } else if (cx->entry_method == ACPI_CSTATE_HALT) {
 810                 acpi_safe_halt();
 811         } else {
 812                 int unused;
 813                 /* IO port based C-state */
 814                 inb(cx->address);
 815                 /* Dummy wait op - must do something useless after P_LVL2 read
 816                    because chipsets cannot guarantee that STPCLK# signal
 817                    gets asserted in time to freeze execution properly. */
 818                 unused = inl(acpi_gbl_FADT.xpm_timer_block.address);
 819         }
 820         start_critical_timings();
 821 }
 822
 823 /**
 824  * acpi_idle_enter_c1 - enters an ACPI C1 state-type
 825  * @dev: the target CPU
 826  * @state: the state data
 827  *
 828  * This is equivalent to the HALT instruction.
 829  */
 830 static int acpi_idle_enter_c1(struct cpuidle_device *dev,
 831                               struct cpuidle_state *state)
 832 {
 833         ktime_t  kt1, kt2;
 834         s64 idle_time;
 835         struct acpi_processor *pr;
 836         struct acpi_processor_cx *cx = cpuidle_get_statedata(state);
 837
 838         pr = __get_cpu_var(processors);
 839
 840         if (unlikely(!pr))
 841                 return 0;
 842
 843         local_irq_disable();
 844
 845         /* Do not access any ACPI IO ports in suspend path */
 846         if (acpi_idle_suspend) {
 847                 local_irq_enable();
 848                 cpu_relax();
 849                 return 0;
 850         }
 851
 852         lapic_timer_state_broadcast(pr, cx, 1);
 853         kt1 = ktime_get_real();
 854         acpi_idle_do_entry(cx);
 855         kt2 = ktime_get_real();
 856         idle_time =  ktime_to_us(ktime_sub(kt2, kt1));
 857
 858         local_irq_enable();
 859         cx->usage++;
 860         lapic_timer_state_broadcast(pr, cx, 0);
 861
 862         return idle_time;
 863 }
 864
 865 /**
 866  * acpi_idle_enter_simple - enters an ACPI state without BM handling
 867  * @dev: the target CPU
 868  * @state: the state data
 869  */
 870 static int acpi_idle_enter_simple(struct cpuidle_device *dev,
 871                                   struct cpuidle_state *state)
 872 {
 873         struct acpi_processor *pr;
 874         struct acpi_processor_cx *cx = cpuidle_get_statedata(state);
 875         ktime_t  kt1, kt2;
 876         s64 idle_time;
 877         s64 sleep_ticks = 0;
 878
 879         pr = __get_cpu_var(processors);
 880
 881         if (unlikely(!pr))
 882                 return 0;
 883
 884         if (acpi_idle_suspend)
 885                 return(acpi_idle_enter_c1(dev, state));
 886
 887         local_irq_disable();
 888         current_thread_info()->status &= ~TS_POLLING;
 889         /*
 890          * TS_POLLING-cleared state must be visible before we test
 891          * NEED_RESCHED:
 892          */
 893         smp_mb();
 894
 895         if (unlikely(need_resched())) {
 896                 current_thread_info()->status |= TS_POLLING;
 897                 local_irq_enable();
 898                 return 0;
 899         }
 900
 901         /*
 902          * Must be done before busmaster disable as we might need to
 903          * access HPET !
 904          */
 905         lapic_timer_state_broadcast(pr, cx, 1);
 906
 907         if (cx->type == ACPI_STATE_C3)
 908                 ACPI_FLUSH_CPU_CACHE();
 909
 910         kt1 = ktime_get_real();
 911         /* Tell the scheduler that we are going deep-idle: */
 912         sched_clock_idle_sleep_event();
 913         acpi_idle_do_entry(cx);
 914         kt2 = ktime_get_real();
 915         idle_time =  ktime_to_us(ktime_sub(kt2, kt1));
 916
 917         sleep_ticks = us_to_pm_timer_ticks(idle_time);
 918
 919         /* Tell the scheduler how much we idled: */
 920         sched_clock_idle_wakeup_event(sleep_ticks*PM_TIMER_TICK_NS);
 921
 922         local_irq_enable();
 923         current_thread_info()->status |= TS_POLLING;
 924
 925         cx->usage++;
 926
 927         lapic_timer_state_broadcast(pr, cx, 0);
 928         cx->time += sleep_ticks;
 929         return idle_time;
 930 }
 931
 932 static int c3_cpu_count;
 933 static DEFINE_SPINLOCK(c3_lock);
 934
 935 /**
 936  * acpi_idle_enter_bm - enters C3 with proper BM handling
 937  * @dev: the target CPU
 938  * @state: the state data
 939  *
 940  * If BM is detected, the deepest non-C3 idle state is entered instead.
 941  */
 942 static int acpi_idle_enter_bm(struct cpuidle_device *dev,
 943                               struct cpuidle_state *state)
 944 {
 945         struct acpi_processor *pr;
 946         struct acpi_processor_cx *cx = cpuidle_get_statedata(state);
 947         ktime_t  kt1, kt2;
 948         s64 idle_time;
 949         s64 sleep_ticks = 0;
 950
 951
 952         pr = __get_cpu_var(processors);
 953
 954         if (unlikely(!pr))
 955                 return 0;
 956
 957         if (acpi_idle_suspend)
 958                 return(acpi_idle_enter_c1(dev, state));
 959
 960         if (acpi_idle_bm_check()) {
 961                 if (dev->safe_state) {
 962                         dev->last_state = dev->safe_state;
 963                         return dev->safe_state->enter(dev, dev->safe_state);
 964                 } else {
 965                         local_irq_disable();
 966                         acpi_safe_halt();
 967                         local_irq_enable();
 968                         return 0;
 969                 }
 970         }
 971
 972         local_irq_disable();
 973         current_thread_info()->status &= ~TS_POLLING;
 974         /*
 975          * TS_POLLING-cleared state must be visible before we test
 976          * NEED_RESCHED:
 977          */
 978         smp_mb();
 979
 980         if (unlikely(need_resched())) {
 981                 current_thread_info()->status |= TS_POLLING;
 982                 local_irq_enable();
 983                 return 0;
 984         }
 985
 986         acpi_unlazy_tlb(smp_processor_id());
 987
 988         /* Tell the scheduler that we are going deep-idle: */
 989         sched_clock_idle_sleep_event();
 990         /*
 991          * Must be done before busmaster disable as we might need to
 992          * access HPET !
 993          */
 994         lapic_timer_state_broadcast(pr, cx, 1);
 995
 996         kt1 = ktime_get_real();
 997         /*
 998          * disable bus master
 999          * bm_check implies we need ARB_DIS
1000          * !bm_check implies we need cache flush
1001          * bm_control implies whether we can do ARB_DIS
1002          *
1003          * That leaves a case where bm_check is set and bm_control is
1004          * not set. In that case we cannot do much, we enter C3
1005          * without doing anything.
1006          */
1007         if (pr->flags.bm_check && pr->flags.bm_control) {
1008                 spin_lock(&c3_lock);
1009                 c3_cpu_count++;
1010                 /* Disable bus master arbitration when all CPUs are in C3 */
1011                 if (c3_cpu_count == num_online_cpus())
1012                         acpi_write_bit_register(ACPI_BITREG_ARB_DISABLE, 1);
1013                 spin_unlock(&c3_lock);
1014         } else if (!pr->flags.bm_check) {
1015                 ACPI_FLUSH_CPU_CACHE();
1016         }
1017
1018         acpi_idle_do_entry(cx);
1019
1020         /* Re-enable bus master arbitration */
1021         if (pr->flags.bm_check && pr->flags.bm_control) {
1022                 spin_lock(&c3_lock);
1023                 acpi_write_bit_register(ACPI_BITREG_ARB_DISABLE, 0);
1024                 c3_cpu_count--;
1025                 spin_unlock(&c3_lock);
1026         }
1027         kt2 = ktime_get_real();
1028         idle_time =  ktime_to_us(ktime_sub(kt2, kt1));
1029
1030         sleep_ticks = us_to_pm_timer_ticks(idle_time);
1031         /* Tell the scheduler how much we idled: */
1032         sched_clock_idle_wakeup_event(sleep_ticks*PM_TIMER_TICK_NS);
1033
1034         local_irq_enable();
1035         current_thread_info()->status |= TS_POLLING;
1036
1037         cx->usage++;
1038
1039         lapic_timer_state_broadcast(pr, cx, 0);
1040         cx->time += sleep_ticks;
1041         return idle_time;
1042 }
1043
1044 struct cpuidle_driver acpi_idle_driver = {
1045         .name =         "acpi_idle",
1046         .owner =        THIS_MODULE,
1047 };
1048
1049 /**
1050  * acpi_processor_setup_cpuidle - prepares and configures CPUIDLE
1051  * @pr: the ACPI processor
1052  */
1053 static int acpi_processor_setup_cpuidle(struct acpi_processor *pr)
1054 {
1055         int i, count = CPUIDLE_DRIVER_STATE_START;
1056         struct acpi_processor_cx *cx;
1057         struct cpuidle_state *state;
1058         struct cpuidle_device *dev = &pr->power.dev;
1059
1060         if (!pr->flags.power_setup_done)
1061                 return -EINVAL;
1062
1063         if (pr->flags.power == 0) {
1064                 return -EINVAL;
1065         }
1066
1067         dev->cpu = pr->id;
1068         for (i = 0; i < CPUIDLE_STATE_MAX; i++) {
1069                 dev->states[i].name[0] = '\0';
1070                 dev->states[i].desc[0] = '\0';
1071         }
1072
1073         if (max_cstate == 0)
1074                 max_cstate = 1;
1075
1076         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER && i <= max_cstate; i++) {
1077                 cx = &pr->power.states[i];
1078                 state = &dev->states[count];
1079
1080                 if (!cx->valid)
1081                         continue;
1082
1083 #ifdef CONFIG_HOTPLUG_CPU
1084                 if ((cx->type != ACPI_STATE_C1) && (num_online_cpus() > 1) &&
1085                     !pr->flags.has_cst &&
1086                     !(acpi_gbl_FADT.flags & ACPI_FADT_C2_MP_SUPPORTED))
1087                         continue;
1088 #endif
1089                 cpuidle_set_statedata(state, cx);
1090
1091                 snprintf(state->name, CPUIDLE_NAME_LEN, "C%d", i);
1092                 strncpy(state->desc, cx->desc, CPUIDLE_DESC_LEN);
1093                 state->exit_latency = cx->latency;
1094                 state->target_residency = cx->latency * latency_factor;
1095                 state->power_usage = cx->power;
1096
1097                 state->flags = 0;
1098                 switch (cx->type) {
1099                         case ACPI_STATE_C1:
1100                         state->flags |= CPUIDLE_FLAG_SHALLOW;
1101                         if (cx->entry_method == ACPI_CSTATE_FFH)
1102                                 state->flags |= CPUIDLE_FLAG_TIME_VALID;
1103
1104                         state->enter = acpi_idle_enter_c1;
1105                         dev->safe_state = state;
1106                         break;
1107
1108                         case ACPI_STATE_C2:
1109                         state->flags |= CPUIDLE_FLAG_BALANCED;
1110                         state->flags |= CPUIDLE_FLAG_TIME_VALID;
1111                         state->enter = acpi_idle_enter_simple;
1112                         dev->safe_state = state;
1113                         break;
1114
1115                         case ACPI_STATE_C3:
1116                         state->flags |= CPUIDLE_FLAG_DEEP;
1117                         state->flags |= CPUIDLE_FLAG_TIME_VALID;
1118                         state->flags |= CPUIDLE_FLAG_CHECK_BM;
1119                         state->enter = pr->flags.bm_check ?
1120                                         acpi_idle_enter_bm :
1121                                         acpi_idle_enter_simple;
1122                         break;
1123                 }
1124
1125                 count++;
1126                 if (count == CPUIDLE_STATE_MAX)
1127                         break;
1128         }
1129
1130         dev->state_count = count;
1131
1132         if (!count)
1133                 return -EINVAL;
1134
1135         return 0;
1136 }
1137
1138 int acpi_processor_cst_has_changed(struct acpi_processor *pr)
1139 {
1140         int ret = 0;
1141
1142         if (boot_option_idle_override)
1143                 return 0;
1144
1145         if (!pr)
1146                 return -EINVAL;
1147
1148         if (nocst) {
1149                 return -ENODEV;
1150         }
1151
1152         if (!pr->flags.power_setup_done)
1153                 return -ENODEV;
1154
1155         cpuidle_pause_and_lock();
1156         cpuidle_disable_device(&pr->power.dev);
1157         acpi_processor_get_power_info(pr);
1158         if (pr->flags.power) {
1159                 acpi_processor_setup_cpuidle(pr);
1160                 ret = cpuidle_enable_device(&pr->power.dev);
1161         }
1162         cpuidle_resume_and_unlock();
1163
1164         return ret;
1165 }
1166
1167 int __cpuinit acpi_processor_power_init(struct acpi_processor *pr,
1168                               struct acpi_device *device)
1169 {
1170         acpi_status status = 0;
1171         static int first_run;
1172 #ifdef CONFIG_ACPI_PROCFS
1173         struct proc_dir_entry *entry = NULL;
1174 #endif
1175
1176         if (boot_option_idle_override)
1177                 return 0;
1178
1179         if (!first_run) {
1180                 if (idle_halt) {
1181                         /*
1182                          * When the boot option of "idle=halt" is added, halt
1183                          * is used for CPU IDLE.
1184                          * In such case C2/C3 is meaningless. So the max_cstate
1185                          * is set to one.
1186                          */
1187                         max_cstate = 1;
1188                 }
1189                 dmi_check_system(processor_power_dmi_table);
1190                 max_cstate = acpi_processor_cstate_check(max_cstate);
1191                 if (max_cstate < ACPI_C_STATES_MAX)
1192                         printk(KERN_NOTICE
1193                                "ACPI: processor limited to max C-state %d\n",
1194                                max_cstate);
1195                 first_run++;
1196         }
1197
1198         if (!pr)
1199                 return -EINVAL;
1200
1201         if (acpi_gbl_FADT.cst_control && !nocst) {
1202                 status =
1203                     acpi_os_write_port(acpi_gbl_FADT.smi_command, acpi_gbl_FADT.cst_control, 8);
1204                 if (ACPI_FAILURE(status)) {
1205                         ACPI_EXCEPTION((AE_INFO, status,
1206                                         "Notifying BIOS of _CST ability failed"));
1207                 }
1208         }
1209
1210         acpi_processor_get_power_info(pr);
1211         pr->flags.power_setup_done = 1;
1212
1213         /*
1214          * Install the idle handler if processor power management is supported.
1215          * Note that we use previously set idle handler will be used on
1216          * platforms that only support C1.
1217          */
1218         if (pr->flags.power) {
1219                 acpi_processor_setup_cpuidle(pr);
1220                 if (cpuidle_register_device(&pr->power.dev))
1221                         return -EIO;
1222         }
1223 #ifdef CONFIG_ACPI_PROCFS
1224         /* 'power' [R] */
1225         entry = proc_create_data(ACPI_PROCESSOR_FILE_POWER,
1226                                  S_IRUGO, acpi_device_dir(device),
1227                                  &acpi_processor_power_fops,
1228                                  acpi_driver_data(device));
1229         if (!entry)
1230                 return -EIO;
1231 #endif
1232         return 0;
1233 }
1234
1235 int acpi_processor_power_exit(struct acpi_processor *pr,
1236                               struct acpi_device *device)
1237 {
1238         if (boot_option_idle_override)
1239                 return 0;
1240
1241         cpuidle_unregister_device(&pr->power.dev);
1242         pr->flags.power_setup_done = 0;
1243
1244 #ifdef CONFIG_ACPI_PROCFS
1245         if (acpi_device_dir(device))
1246                 remove_proc_entry(ACPI_PROCESSOR_FILE_POWER,
1247                                   acpi_device_dir(device));
1248 #endif
1249
1250         return 0;
1251 }