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         { set_max_cstate, "Pavilion zv5000", {
 114           DMI_MATCH(DMI_SYS_VENDOR, "Hewlett-Packard"),
 115           DMI_MATCH(DMI_PRODUCT_NAME,"Pavilion zv5000 (DS502A#ABA)")},
 116          (void *)1},
 117         { set_max_cstate, "Asus L8400B", {
 118           DMI_MATCH(DMI_SYS_VENDOR, "ASUSTeK Computer Inc."),
 119           DMI_MATCH(DMI_PRODUCT_NAME,"L8400B series Notebook PC")},
 120          (void *)1},
 121         {},
 122 };
 123
 124
 125 /*
 126  * Callers should disable interrupts before the call and enable
 127  * interrupts after return.
 128  */
 129 static void acpi_safe_halt(void)
 130 {
 131         current_thread_info()->status &= ~TS_POLLING;
 132         /*
 133          * TS_POLLING-cleared state must be visible before we
 134          * test NEED_RESCHED:
 135          */
 136         smp_mb();
 137         if (!need_resched()) {
 138                 safe_halt();
 139                 local_irq_disable();
 140         }
 141         current_thread_info()->status |= TS_POLLING;
 142 }
 143
 144 #ifdef ARCH_APICTIMER_STOPS_ON_C3
 145
 146 /*
 147  * Some BIOS implementations switch to C3 in the published C2 state.
 148  * This seems to be a common problem on AMD boxen, but other vendors
 149  * are affected too. We pick the most conservative approach: we assume
 150  * that the local APIC stops in both C2 and C3.
 151  */
 152 static void lapic_timer_check_state(int state, struct acpi_processor *pr,
 153                                    struct acpi_processor_cx *cx)
 154 {
 155         struct acpi_processor_power *pwr = &pr->power;
 156         u8 type = local_apic_timer_c2_ok ? ACPI_STATE_C3 : ACPI_STATE_C2;
 157
 158         if (cpu_has(&cpu_data(pr->id), X86_FEATURE_ARAT))
 159                 return;
 160
 161         if (boot_cpu_has(X86_FEATURE_AMDC1E))
 162                 type = ACPI_STATE_C1;
 163
 164         /*
 165          * Check, if one of the previous states already marked the lapic
 166          * unstable
 167          */
 168         if (pwr->timer_broadcast_on_state < state)
 169                 return;
 170
 171         if (cx->type >= type)
 172                 pr->power.timer_broadcast_on_state = state;
 173 }
 174
 175 static void __lapic_timer_propagate_broadcast(void *arg)
 176 {
 177         struct acpi_processor *pr = (struct acpi_processor *) arg;
 178         unsigned long reason;
 179
 180         reason = pr->power.timer_broadcast_on_state < INT_MAX ?
 181                 CLOCK_EVT_NOTIFY_BROADCAST_ON : CLOCK_EVT_NOTIFY_BROADCAST_OFF;
 182
 183         clockevents_notify(reason, &pr->id);
 184 }
 185
 186 static void lapic_timer_propagate_broadcast(struct acpi_processor *pr)
 187 {
 188         smp_call_function_single(pr->id, __lapic_timer_propagate_broadcast,
 189                                  (void *)pr, 1);
 190 }
 191
 192 /* Power(C) State timer broadcast control */
 193 static void lapic_timer_state_broadcast(struct acpi_processor *pr,
 194                                        struct acpi_processor_cx *cx,
 195                                        int broadcast)
 196 {
 197         int state = cx - pr->power.states;
 198
 199         if (state >= pr->power.timer_broadcast_on_state) {
 200                 unsigned long reason;
 201
 202                 reason = broadcast ?  CLOCK_EVT_NOTIFY_BROADCAST_ENTER :
 203                         CLOCK_EVT_NOTIFY_BROADCAST_EXIT;
 204                 clockevents_notify(reason, &pr->id);
 205         }
 206 }
 207
 208 #else
 209
 210 static void lapic_timer_check_state(int state, struct acpi_processor *pr,
 211                                    struct acpi_processor_cx *cstate) { }
 212 static void lapic_timer_propagate_broadcast(struct acpi_processor *pr) { }
 213 static void lapic_timer_state_broadcast(struct acpi_processor *pr,
 214                                        struct acpi_processor_cx *cx,
 215                                        int broadcast)
 216 {
 217 }
 218
 219 #endif
 220
 221 /*
 222  * Suspend / resume control
 223  */
 224 static int acpi_idle_suspend;
 225 static u32 saved_bm_rld;
 226
 227 static void acpi_idle_bm_rld_save(void)
 228 {
 229         acpi_read_bit_register(ACPI_BITREG_BUS_MASTER_RLD, &saved_bm_rld);
 230 }
 231 static void acpi_idle_bm_rld_restore(void)
 232 {
 233         u32 resumed_bm_rld;
 234
 235         acpi_read_bit_register(ACPI_BITREG_BUS_MASTER_RLD, &resumed_bm_rld);
 236
 237         if (resumed_bm_rld != saved_bm_rld)
 238                 acpi_write_bit_register(ACPI_BITREG_BUS_MASTER_RLD, saved_bm_rld);
 239 }
 240
 241 int acpi_processor_suspend(struct acpi_device * device, pm_message_t state)
 242 {
 243         if (acpi_idle_suspend == 1)
 244                 return 0;
 245
 246         acpi_idle_bm_rld_save();
 247         acpi_idle_suspend = 1;
 248         return 0;
 249 }
 250
 251 int acpi_processor_resume(struct acpi_device * device)
 252 {
 253         if (acpi_idle_suspend == 0)
 254                 return 0;
 255
 256         acpi_idle_bm_rld_restore();
 257         acpi_idle_suspend = 0;
 258         return 0;
 259 }
 260
 261 #if defined (CONFIG_GENERIC_TIME) && defined (CONFIG_X86)
 262 static void tsc_check_state(int state)
 263 {
 264         switch (boot_cpu_data.x86_vendor) {
 265         case X86_VENDOR_AMD:
 266         case X86_VENDOR_INTEL:
 267                 /*
 268                  * AMD Fam10h TSC will tick in all
 269                  * C/P/S0/S1 states when this bit is set.
 270                  */
 271                 if (boot_cpu_has(X86_FEATURE_NONSTOP_TSC))
 272                         return;
 273
 274                 /*FALL THROUGH*/
 275         default:
 276                 /* TSC could halt in idle, so notify users */
 277                 if (state > ACPI_STATE_C1)
 278                         mark_tsc_unstable("TSC halts in idle");
 279         }
 280 }
 281 #else
 282 static void tsc_check_state(int state) { return; }
 283 #endif
 284
 285 static int acpi_processor_get_power_info_fadt(struct acpi_processor *pr)
 286 {
 287
 288         if (!pr)
 289                 return -EINVAL;
 290
 291         if (!pr->pblk)
 292                 return -ENODEV;
 293
 294         /* if info is obtained from pblk/fadt, type equals state */
 295         pr->power.states[ACPI_STATE_C2].type = ACPI_STATE_C2;
 296         pr->power.states[ACPI_STATE_C3].type = ACPI_STATE_C3;
 297
 298 #ifndef CONFIG_HOTPLUG_CPU
 299         /*
 300          * Check for P_LVL2_UP flag before entering C2 and above on
 301          * an SMP system.
 302          */
 303         if ((num_online_cpus() > 1) &&
 304             !(acpi_gbl_FADT.flags & ACPI_FADT_C2_MP_SUPPORTED))
 305                 return -ENODEV;
 306 #endif
 307
 308         /* determine C2 and C3 address from pblk */
 309         pr->power.states[ACPI_STATE_C2].address = pr->pblk + 4;
 310         pr->power.states[ACPI_STATE_C3].address = pr->pblk + 5;
 311
 312         /* determine latencies from FADT */
 313         pr->power.states[ACPI_STATE_C2].latency = acpi_gbl_FADT.C2latency;
 314         pr->power.states[ACPI_STATE_C3].latency = acpi_gbl_FADT.C3latency;
 315
 316         /*
 317          * FADT specified C2 latency must be less than or equal to
 318          * 100 microseconds.
 319          */
 320         if (acpi_gbl_FADT.C2latency > ACPI_PROCESSOR_MAX_C2_LATENCY) {
 321                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 322                         "C2 latency too large [%d]\n", acpi_gbl_FADT.C2latency));
 323                 /* invalidate C2 */
 324                 pr->power.states[ACPI_STATE_C2].address = 0;
 325         }
 326
 327         /*
 328          * FADT supplied C3 latency must be less than or equal to
 329          * 1000 microseconds.
 330          */
 331         if (acpi_gbl_FADT.C3latency > ACPI_PROCESSOR_MAX_C3_LATENCY) {
 332                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 333                         "C3 latency too large [%d]\n", acpi_gbl_FADT.C3latency));
 334                 /* invalidate C3 */
 335                 pr->power.states[ACPI_STATE_C3].address = 0;
 336         }
 337
 338         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 339                           "lvl2[0x%08x] lvl3[0x%08x]\n",
 340                           pr->power.states[ACPI_STATE_C2].address,
 341                           pr->power.states[ACPI_STATE_C3].address));
 342
 343         return 0;
 344 }
 345
 346 static int acpi_processor_get_power_info_default(struct acpi_processor *pr)
 347 {
 348         if (!pr->power.states[ACPI_STATE_C1].valid) {
 349                 /* set the first C-State to C1 */
 350                 /* all processors need to support C1 */
 351                 pr->power.states[ACPI_STATE_C1].type = ACPI_STATE_C1;
 352                 pr->power.states[ACPI_STATE_C1].valid = 1;
 353                 pr->power.states[ACPI_STATE_C1].entry_method = ACPI_CSTATE_HALT;
 354         }
 355         /* the C0 state only exists as a filler in our array */
 356         pr->power.states[ACPI_STATE_C0].valid = 1;
 357         return 0;
 358 }
 359
 360 static int acpi_processor_get_power_info_cst(struct acpi_processor *pr)
 361 {
 362         acpi_status status = 0;
 363         acpi_integer count;
 364         int current_count;
 365         int i;
 366         struct acpi_buffer buffer = { ACPI_ALLOCATE_BUFFER, NULL };
 367         union acpi_object *cst;
 368
 369
 370         if (nocst)
 371                 return -ENODEV;
 372
 373         current_count = 0;
 374
 375         status = acpi_evaluate_object(pr->handle, "_CST", NULL, &buffer);
 376         if (ACPI_FAILURE(status)) {
 377                 ACPI_DEBUG_PRINT((ACPI_DB_INFO, "No _CST, giving up\n"));
 378                 return -ENODEV;
 379         }
 380
 381         cst = buffer.pointer;
 382
 383         /* There must be at least 2 elements */
 384         if (!cst || (cst->type != ACPI_TYPE_PACKAGE) || cst->package.count < 2) {
 385                 printk(KERN_ERR PREFIX "not enough elements in _CST\n");
 386                 status = -EFAULT;
 387                 goto end;
 388         }
 389
 390         count = cst->package.elements[0].integer.value;
 391
 392         /* Validate number of power states. */
 393         if (count < 1 || count != cst->package.count - 1) {
 394                 printk(KERN_ERR PREFIX "count given by _CST is not valid\n");
 395                 status = -EFAULT;
 396                 goto end;
 397         }
 398
 399         /* Tell driver that at least _CST is supported. */
 400         pr->flags.has_cst = 1;
 401
 402         for (i = 1; i <= count; i++) {
 403                 union acpi_object *element;
 404                 union acpi_object *obj;
 405                 struct acpi_power_register *reg;
 406                 struct acpi_processor_cx cx;
 407
 408                 memset(&cx, 0, sizeof(cx));
 409
 410                 element = &(cst->package.elements[i]);
 411                 if (element->type != ACPI_TYPE_PACKAGE)
 412                         continue;
 413
 414                 if (element->package.count != 4)
 415                         continue;
 416
 417                 obj = &(element->package.elements[0]);
 418
 419                 if (obj->type != ACPI_TYPE_BUFFER)
 420                         continue;
 421
 422                 reg = (struct acpi_power_register *)obj->buffer.pointer;
 423
 424                 if (reg->space_id != ACPI_ADR_SPACE_SYSTEM_IO &&
 425                     (reg->space_id != ACPI_ADR_SPACE_FIXED_HARDWARE))
 426                         continue;
 427
 428                 /* There should be an easy way to extract an integer... */
 429                 obj = &(element->package.elements[1]);
 430                 if (obj->type != ACPI_TYPE_INTEGER)
 431                         continue;
 432
 433                 cx.type = obj->integer.value;
 434                 /*
 435                  * Some buggy BIOSes won't list C1 in _CST -
 436                  * Let acpi_processor_get_power_info_default() handle them later
 437                  */
 438                 if (i == 1 && cx.type != ACPI_STATE_C1)
 439                         current_count++;
 440
 441                 cx.address = reg->address;
 442                 cx.index = current_count + 1;
 443
 444                 cx.entry_method = ACPI_CSTATE_SYSTEMIO;
 445                 if (reg->space_id == ACPI_ADR_SPACE_FIXED_HARDWARE) {
 446                         if (acpi_processor_ffh_cstate_probe
 447                                         (pr->id, &cx, reg) == 0) {
 448                                 cx.entry_method = ACPI_CSTATE_FFH;
 449                         } else if (cx.type == ACPI_STATE_C1) {
 450                                 /*
 451                                  * C1 is a special case where FIXED_HARDWARE
 452                                  * can be handled in non-MWAIT way as well.
 453                                  * In that case, save this _CST entry info.
 454                                  * Otherwise, ignore this info and continue.
 455                                  */
 456                                 cx.entry_method = ACPI_CSTATE_HALT;
 457                                 snprintf(cx.desc, ACPI_CX_DESC_LEN, "ACPI HLT");
 458                         } else {
 459                                 continue;
 460                         }
 461                         if (cx.type == ACPI_STATE_C1 &&
 462                                         (idle_halt || idle_nomwait)) {
 463                                 /*
 464                                  * In most cases the C1 space_id obtained from
 465                                  * _CST object is FIXED_HARDWARE access mode.
 466                                  * But when the option of idle=halt is added,
 467                                  * the entry_method type should be changed from
 468                                  * CSTATE_FFH to CSTATE_HALT.
 469                                  * When the option of idle=nomwait is added,
 470                                  * the C1 entry_method type should be
 471                                  * CSTATE_HALT.
 472                                  */
 473                                 cx.entry_method = ACPI_CSTATE_HALT;
 474                                 snprintf(cx.desc, ACPI_CX_DESC_LEN, "ACPI HLT");
 475                         }
 476                 } else {
 477                         snprintf(cx.desc, ACPI_CX_DESC_LEN, "ACPI IOPORT 0x%x",
 478                                  cx.address);
 479                 }
 480
 481                 if (cx.type == ACPI_STATE_C1) {
 482                         cx.valid = 1;
 483                 }
 484
 485                 obj = &(element->package.elements[2]);
 486                 if (obj->type != ACPI_TYPE_INTEGER)
 487                         continue;
 488
 489                 cx.latency = obj->integer.value;
 490
 491                 obj = &(element->package.elements[3]);
 492                 if (obj->type != ACPI_TYPE_INTEGER)
 493                         continue;
 494
 495                 cx.power = obj->integer.value;
 496
 497                 current_count++;
 498                 memcpy(&(pr->power.states[current_count]), &cx, sizeof(cx));
 499
 500                 /*
 501                  * We support total ACPI_PROCESSOR_MAX_POWER - 1
 502                  * (From 1 through ACPI_PROCESSOR_MAX_POWER - 1)
 503                  */
 504                 if (current_count >= (ACPI_PROCESSOR_MAX_POWER - 1)) {
 505                         printk(KERN_WARNING
 506                                "Limiting number of power states to max (%d)\n",
 507                                ACPI_PROCESSOR_MAX_POWER);
 508                         printk(KERN_WARNING
 509                                "Please increase ACPI_PROCESSOR_MAX_POWER if needed.\n");
 510                         break;
 511                 }
 512         }
 513
 514         ACPI_DEBUG_PRINT((ACPI_DB_INFO, "Found %d power states\n",
 515                           current_count));
 516
 517         /* Validate number of power states discovered */
 518         if (current_count < 2)
 519                 status = -EFAULT;
 520
 521       end:
 522         kfree(buffer.pointer);
 523
 524         return status;
 525 }
 526
 527 static void acpi_processor_power_verify_c3(struct acpi_processor *pr,
 528                                            struct acpi_processor_cx *cx)
 529 {
 530         static int bm_check_flag = -1;
 531         static int bm_control_flag = -1;
 532
 533
 534         if (!cx->address)
 535                 return;
 536
 537         /*
 538          * PIIX4 Erratum #18: We don't support C3 when Type-F (fast)
 539          * DMA transfers are used by any ISA device to avoid livelock.
 540          * Note that we could disable Type-F DMA (as recommended by
 541          * the erratum), but this is known to disrupt certain ISA
 542          * devices thus we take the conservative approach.
 543          */
 544         else if (errata.piix4.fdma) {
 545                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 546                                   "C3 not supported on PIIX4 with Type-F DMA\n"));
 547                 return;
 548         }
 549
 550         /* All the logic here assumes flags.bm_check is same across all CPUs */
 551         if (bm_check_flag == -1) {
 552                 /* Determine whether bm_check is needed based on CPU  */
 553                 acpi_processor_power_init_bm_check(&(pr->flags), pr->id);
 554                 bm_check_flag = pr->flags.bm_check;
 555                 bm_control_flag = pr->flags.bm_control;
 556         } else {
 557                 pr->flags.bm_check = bm_check_flag;
 558                 pr->flags.bm_control = bm_control_flag;
 559         }
 560
 561         if (pr->flags.bm_check) {
 562                 if (!pr->flags.bm_control) {
 563                         if (pr->flags.has_cst != 1) {
 564                                 /* bus mastering control is necessary */
 565                                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 566                                         "C3 support requires BM control\n"));
 567                                 return;
 568                         } else {
 569                                 /* Here we enter C3 without bus mastering */
 570                                 ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 571                                         "C3 support without BM control\n"));
 572                         }
 573                 }
 574         } else {
 575                 /*
 576                  * WBINVD should be set in fadt, for C3 state to be
 577                  * supported on when bm_check is not required.
 578                  */
 579                 if (!(acpi_gbl_FADT.flags & ACPI_FADT_WBINVD)) {
 580                         ACPI_DEBUG_PRINT((ACPI_DB_INFO,
 581                                           "Cache invalidation should work properly"
 582                                           " for C3 to be enabled on SMP systems\n"));
 583                         return;
 584                 }
 585         }
 586
 587         /*
 588          * Otherwise we've met all of our C3 requirements.
 589          * Normalize the C3 latency to expidite policy.  Enable
 590          * checking of bus mastering status (bm_check) so we can
 591          * use this in our C3 policy
 592          */
 593         cx->valid = 1;
 594
 595         cx->latency_ticks = cx->latency;
 596         /*
 597          * On older chipsets, BM_RLD needs to be set
 598          * in order for Bus Master activity to wake the
 599          * system from C3.  Newer chipsets handle DMA
 600          * during C3 automatically and BM_RLD is a NOP.
 601          * In either case, the proper way to
 602          * handle BM_RLD is to set it and leave it set.
 603          */
 604         acpi_write_bit_register(ACPI_BITREG_BUS_MASTER_RLD, 1);
 605
 606         return;
 607 }
 608
 609 static int acpi_processor_power_verify(struct acpi_processor *pr)
 610 {
 611         unsigned int i;
 612         unsigned int working = 0;
 613
 614         pr->power.timer_broadcast_on_state = INT_MAX;
 615
 616         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER && i <= max_cstate; i++) {
 617                 struct acpi_processor_cx *cx = &pr->power.states[i];
 618
 619                 switch (cx->type) {
 620                 case ACPI_STATE_C1:
 621                         cx->valid = 1;
 622                         break;
 623
 624                 case ACPI_STATE_C2:
 625                         if (!cx->address)
 626                                 break;
 627                         cx->valid = 1;
 628                         cx->latency_ticks = cx->latency; /* Normalize latency */
 629                         break;
 630
 631                 case ACPI_STATE_C3:
 632                         acpi_processor_power_verify_c3(pr, cx);
 633                         break;
 634                 }
 635                 if (!cx->valid)
 636                         continue;
 637
 638                 lapic_timer_check_state(i, pr, cx);
 639                 tsc_check_state(cx->type);
 640                 working++;
 641         }
 642
 643         lapic_timer_propagate_broadcast(pr);
 644
 645         return (working);
 646 }
 647
 648 static int acpi_processor_get_power_info(struct acpi_processor *pr)
 649 {
 650         unsigned int i;
 651         int result;
 652
 653
 654         /* NOTE: the idle thread may not be running while calling
 655          * this function */
 656
 657         /* Zero initialize all the C-states info. */
 658         memset(pr->power.states, 0, sizeof(pr->power.states));
 659
 660         result = acpi_processor_get_power_info_cst(pr);
 661         if (result == -ENODEV)
 662                 result = acpi_processor_get_power_info_fadt(pr);
 663
 664         if (result)
 665                 return result;
 666
 667         acpi_processor_get_power_info_default(pr);
 668
 669         pr->power.count = acpi_processor_power_verify(pr);
 670
 671         /*
 672          * if one state of type C2 or C3 is available, mark this
 673          * CPU as being "idle manageable"
 674          */
 675         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER; i++) {
 676                 if (pr->power.states[i].valid) {
 677                         pr->power.count = i;
 678                         if (pr->power.states[i].type >= ACPI_STATE_C2)
 679                                 pr->flags.power = 1;
 680                 }
 681         }
 682
 683         return 0;
 684 }
 685
 686 #ifdef CONFIG_ACPI_PROCFS
 687 static int acpi_processor_power_seq_show(struct seq_file *seq, void *offset)
 688 {
 689         struct acpi_processor *pr = seq->private;
 690         unsigned int i;
 691
 692
 693         if (!pr)
 694                 goto end;
 695
 696         seq_printf(seq, "active state:            C%zd\n"
 697                    "max_cstate:              C%d\n"
 698                    "maximum allowed latency: %d usec\n",
 699                    pr->power.state ? pr->power.state - pr->power.states : 0,
 700                    max_cstate, pm_qos_requirement(PM_QOS_CPU_DMA_LATENCY));
 701
 702         seq_puts(seq, "states:\n");
 703
 704         for (i = 1; i <= pr->power.count; i++) {
 705                 seq_printf(seq, "   %cC%d:                  ",
 706                            (&pr->power.states[i] ==
 707                             pr->power.state ? '*' : ' '), i);
 708
 709                 if (!pr->power.states[i].valid) {
 710                         seq_puts(seq, "<not supported>\n");
 711                         continue;
 712                 }
 713
 714                 switch (pr->power.states[i].type) {
 715                 case ACPI_STATE_C1:
 716                         seq_printf(seq, "type[C1] ");
 717                         break;
 718                 case ACPI_STATE_C2:
 719                         seq_printf(seq, "type[C2] ");
 720                         break;
 721                 case ACPI_STATE_C3:
 722                         seq_printf(seq, "type[C3] ");
 723                         break;
 724                 default:
 725                         seq_printf(seq, "type[--] ");
 726                         break;
 727                 }
 728
 729                 if (pr->power.states[i].promotion.state)
 730                         seq_printf(seq, "promotion[C%zd] ",
 731                                    (pr->power.states[i].promotion.state -
 732                                     pr->power.states));
 733                 else
 734                         seq_puts(seq, "promotion[--] ");
 735
 736                 if (pr->power.states[i].demotion.state)
 737                         seq_printf(seq, "demotion[C%zd] ",
 738                                    (pr->power.states[i].demotion.state -
 739                                     pr->power.states));
 740                 else
 741                         seq_puts(seq, "demotion[--] ");
 742
 743                 seq_printf(seq, "latency[%03d] usage[%08d] duration[%020llu]\n",
 744                            pr->power.states[i].latency,
 745                            pr->power.states[i].usage,
 746                            (unsigned long long)pr->power.states[i].time);
 747         }
 748
 749       end:
 750         return 0;
 751 }
 752
 753 static int acpi_processor_power_open_fs(struct inode *inode, struct file *file)
 754 {
 755         return single_open(file, acpi_processor_power_seq_show,
 756                            PDE(inode)->data);
 757 }
 758
 759 static const struct file_operations acpi_processor_power_fops = {
 760         .owner = THIS_MODULE,
 761         .open = acpi_processor_power_open_fs,
 762         .read = seq_read,
 763         .llseek = seq_lseek,
 764         .release = single_release,
 765 };
 766 #endif
 767
 768 /**
 769  * acpi_idle_bm_check - checks if bus master activity was detected
 770  */
 771 static int acpi_idle_bm_check(void)
 772 {
 773         u32 bm_status = 0;
 774
 775         acpi_read_bit_register(ACPI_BITREG_BUS_MASTER_STATUS, &bm_status);
 776         if (bm_status)
 777                 acpi_write_bit_register(ACPI_BITREG_BUS_MASTER_STATUS, 1);
 778         /*
 779          * PIIX4 Erratum #18: Note that BM_STS doesn't always reflect
 780          * the true state of bus mastering activity; forcing us to
 781          * manually check the BMIDEA bit of each IDE channel.
 782          */
 783         else if (errata.piix4.bmisx) {
 784                 if ((inb_p(errata.piix4.bmisx + 0x02) & 0x01)
 785                     || (inb_p(errata.piix4.bmisx + 0x0A) & 0x01))
 786                         bm_status = 1;
 787         }
 788         return bm_status;
 789 }
 790
 791 /**
 792  * acpi_idle_do_entry - a helper function that does C2 and C3 type entry
 793  * @cx: cstate data
 794  *
 795  * Caller disables interrupt before call and enables interrupt after return.
 796  */
 797 static inline void acpi_idle_do_entry(struct acpi_processor_cx *cx)
 798 {
 799         /* Don't trace irqs off for idle */
 800         stop_critical_timings();
 801         if (cx->entry_method == ACPI_CSTATE_FFH) {
 802                 /* Call into architectural FFH based C-state */
 803                 acpi_processor_ffh_cstate_enter(cx);
 804         } else if (cx->entry_method == ACPI_CSTATE_HALT) {
 805                 acpi_safe_halt();
 806         } else {
 807                 int unused;
 808                 /* IO port based C-state */
 809                 inb(cx->address);
 810                 /* Dummy wait op - must do something useless after P_LVL2 read
 811                    because chipsets cannot guarantee that STPCLK# signal
 812                    gets asserted in time to freeze execution properly. */
 813                 unused = inl(acpi_gbl_FADT.xpm_timer_block.address);
 814         }
 815         start_critical_timings();
 816 }
 817
 818 /**
 819  * acpi_idle_enter_c1 - enters an ACPI C1 state-type
 820  * @dev: the target CPU
 821  * @state: the state data
 822  *
 823  * This is equivalent to the HALT instruction.
 824  */
 825 static int acpi_idle_enter_c1(struct cpuidle_device *dev,
 826                               struct cpuidle_state *state)
 827 {
 828         ktime_t  kt1, kt2;
 829         s64 idle_time;
 830         struct acpi_processor *pr;
 831         struct acpi_processor_cx *cx = cpuidle_get_statedata(state);
 832
 833         pr = __get_cpu_var(processors);
 834
 835         if (unlikely(!pr))
 836                 return 0;
 837
 838         local_irq_disable();
 839
 840         /* Do not access any ACPI IO ports in suspend path */
 841         if (acpi_idle_suspend) {
 842                 local_irq_enable();
 843                 cpu_relax();
 844                 return 0;
 845         }
 846
 847         lapic_timer_state_broadcast(pr, cx, 1);
 848         kt1 = ktime_get_real();
 849         acpi_idle_do_entry(cx);
 850         kt2 = ktime_get_real();
 851         idle_time =  ktime_to_us(ktime_sub(kt2, kt1));
 852
 853         local_irq_enable();
 854         cx->usage++;
 855         lapic_timer_state_broadcast(pr, cx, 0);
 856
 857         return idle_time;
 858 }
 859
 860 /**
 861  * acpi_idle_enter_simple - enters an ACPI state without BM handling
 862  * @dev: the target CPU
 863  * @state: the state data
 864  */
 865 static int acpi_idle_enter_simple(struct cpuidle_device *dev,
 866                                   struct cpuidle_state *state)
 867 {
 868         struct acpi_processor *pr;
 869         struct acpi_processor_cx *cx = cpuidle_get_statedata(state);
 870         ktime_t  kt1, kt2;
 871         s64 idle_time;
 872         s64 sleep_ticks = 0;
 873
 874         pr = __get_cpu_var(processors);
 875
 876         if (unlikely(!pr))
 877                 return 0;
 878
 879         if (acpi_idle_suspend)
 880                 return(acpi_idle_enter_c1(dev, state));
 881
 882         local_irq_disable();
 883         current_thread_info()->status &= ~TS_POLLING;
 884         /*
 885          * TS_POLLING-cleared state must be visible before we test
 886          * NEED_RESCHED:
 887          */
 888         smp_mb();
 889
 890         if (unlikely(need_resched())) {
 891                 current_thread_info()->status |= TS_POLLING;
 892                 local_irq_enable();
 893                 return 0;
 894         }
 895
 896         /*
 897          * Must be done before busmaster disable as we might need to
 898          * access HPET !
 899          */
 900         lapic_timer_state_broadcast(pr, cx, 1);
 901
 902         if (cx->type == ACPI_STATE_C3)
 903                 ACPI_FLUSH_CPU_CACHE();
 904
 905         kt1 = ktime_get_real();
 906         /* Tell the scheduler that we are going deep-idle: */
 907         sched_clock_idle_sleep_event();
 908         acpi_idle_do_entry(cx);
 909         kt2 = ktime_get_real();
 910         idle_time =  ktime_to_us(ktime_sub(kt2, kt1));
 911
 912         sleep_ticks = us_to_pm_timer_ticks(idle_time);
 913
 914         /* Tell the scheduler how much we idled: */
 915         sched_clock_idle_wakeup_event(sleep_ticks*PM_TIMER_TICK_NS);
 916
 917         local_irq_enable();
 918         current_thread_info()->status |= TS_POLLING;
 919
 920         cx->usage++;
 921
 922         lapic_timer_state_broadcast(pr, cx, 0);
 923         cx->time += sleep_ticks;
 924         return idle_time;
 925 }
 926
 927 static int c3_cpu_count;
 928 static DEFINE_SPINLOCK(c3_lock);
 929
 930 /**
 931  * acpi_idle_enter_bm - enters C3 with proper BM handling
 932  * @dev: the target CPU
 933  * @state: the state data
 934  *
 935  * If BM is detected, the deepest non-C3 idle state is entered instead.
 936  */
 937 static int acpi_idle_enter_bm(struct cpuidle_device *dev,
 938                               struct cpuidle_state *state)
 939 {
 940         struct acpi_processor *pr;
 941         struct acpi_processor_cx *cx = cpuidle_get_statedata(state);
 942         ktime_t  kt1, kt2;
 943         s64 idle_time;
 944         s64 sleep_ticks = 0;
 945
 946
 947         pr = __get_cpu_var(processors);
 948
 949         if (unlikely(!pr))
 950                 return 0;
 951
 952         if (acpi_idle_suspend)
 953                 return(acpi_idle_enter_c1(dev, state));
 954
 955         if (acpi_idle_bm_check()) {
 956                 if (dev->safe_state) {
 957                         dev->last_state = dev->safe_state;
 958                         return dev->safe_state->enter(dev, dev->safe_state);
 959                 } else {
 960                         local_irq_disable();
 961                         acpi_safe_halt();
 962                         local_irq_enable();
 963                         return 0;
 964                 }
 965         }
 966
 967         local_irq_disable();
 968         current_thread_info()->status &= ~TS_POLLING;
 969         /*
 970          * TS_POLLING-cleared state must be visible before we test
 971          * NEED_RESCHED:
 972          */
 973         smp_mb();
 974
 975         if (unlikely(need_resched())) {
 976                 current_thread_info()->status |= TS_POLLING;
 977                 local_irq_enable();
 978                 return 0;
 979         }
 980
 981         acpi_unlazy_tlb(smp_processor_id());
 982
 983         /* Tell the scheduler that we are going deep-idle: */
 984         sched_clock_idle_sleep_event();
 985         /*
 986          * Must be done before busmaster disable as we might need to
 987          * access HPET !
 988          */
 989         lapic_timer_state_broadcast(pr, cx, 1);
 990
 991         kt1 = ktime_get_real();
 992         /*
 993          * disable bus master
 994          * bm_check implies we need ARB_DIS
 995          * !bm_check implies we need cache flush
 996          * bm_control implies whether we can do ARB_DIS
 997          *
 998          * That leaves a case where bm_check is set and bm_control is
 999          * not set. In that case we cannot do much, we enter C3
1000          * without doing anything.
1001          */
1002         if (pr->flags.bm_check && pr->flags.bm_control) {
1003                 spin_lock(&c3_lock);
1004                 c3_cpu_count++;
1005                 /* Disable bus master arbitration when all CPUs are in C3 */
1006                 if (c3_cpu_count == num_online_cpus())
1007                         acpi_write_bit_register(ACPI_BITREG_ARB_DISABLE, 1);
1008                 spin_unlock(&c3_lock);
1009         } else if (!pr->flags.bm_check) {
1010                 ACPI_FLUSH_CPU_CACHE();
1011         }
1012
1013         acpi_idle_do_entry(cx);
1014
1015         /* Re-enable bus master arbitration */
1016         if (pr->flags.bm_check && pr->flags.bm_control) {
1017                 spin_lock(&c3_lock);
1018                 acpi_write_bit_register(ACPI_BITREG_ARB_DISABLE, 0);
1019                 c3_cpu_count--;
1020                 spin_unlock(&c3_lock);
1021         }
1022         kt2 = ktime_get_real();
1023         idle_time =  ktime_to_us(ktime_sub(kt2, kt1));
1024
1025         sleep_ticks = us_to_pm_timer_ticks(idle_time);
1026         /* Tell the scheduler how much we idled: */
1027         sched_clock_idle_wakeup_event(sleep_ticks*PM_TIMER_TICK_NS);
1028
1029         local_irq_enable();
1030         current_thread_info()->status |= TS_POLLING;
1031
1032         cx->usage++;
1033
1034         lapic_timer_state_broadcast(pr, cx, 0);
1035         cx->time += sleep_ticks;
1036         return idle_time;
1037 }
1038
1039 struct cpuidle_driver acpi_idle_driver = {
1040         .name =         "acpi_idle",
1041         .owner =        THIS_MODULE,
1042 };
1043
1044 /**
1045  * acpi_processor_setup_cpuidle - prepares and configures CPUIDLE
1046  * @pr: the ACPI processor
1047  */
1048 static int acpi_processor_setup_cpuidle(struct acpi_processor *pr)
1049 {
1050         int i, count = CPUIDLE_DRIVER_STATE_START;
1051         struct acpi_processor_cx *cx;
1052         struct cpuidle_state *state;
1053         struct cpuidle_device *dev = &pr->power.dev;
1054
1055         if (!pr->flags.power_setup_done)
1056                 return -EINVAL;
1057
1058         if (pr->flags.power == 0) {
1059                 return -EINVAL;
1060         }
1061
1062         dev->cpu = pr->id;
1063         for (i = 0; i < CPUIDLE_STATE_MAX; i++) {
1064                 dev->states[i].name[0] = '\0';
1065                 dev->states[i].desc[0] = '\0';
1066         }
1067
1068         if (max_cstate == 0)
1069                 max_cstate = 1;
1070
1071         for (i = 1; i < ACPI_PROCESSOR_MAX_POWER && i <= max_cstate; i++) {
1072                 cx = &pr->power.states[i];
1073                 state = &dev->states[count];
1074
1075                 if (!cx->valid)
1076                         continue;
1077
1078 #ifdef CONFIG_HOTPLUG_CPU
1079                 if ((cx->type != ACPI_STATE_C1) && (num_online_cpus() > 1) &&
1080                     !pr->flags.has_cst &&
1081                     !(acpi_gbl_FADT.flags & ACPI_FADT_C2_MP_SUPPORTED))
1082                         continue;
1083 #endif
1084                 cpuidle_set_statedata(state, cx);
1085
1086                 snprintf(state->name, CPUIDLE_NAME_LEN, "C%d", i);
1087                 strncpy(state->desc, cx->desc, CPUIDLE_DESC_LEN);
1088                 state->exit_latency = cx->latency;
1089                 state->target_residency = cx->latency * latency_factor;
1090                 state->power_usage = cx->power;
1091
1092                 state->flags = 0;
1093                 switch (cx->type) {
1094                         case ACPI_STATE_C1:
1095                         state->flags |= CPUIDLE_FLAG_SHALLOW;
1096                         if (cx->entry_method == ACPI_CSTATE_FFH)
1097                                 state->flags |= CPUIDLE_FLAG_TIME_VALID;
1098
1099                         state->enter = acpi_idle_enter_c1;
1100                         dev->safe_state = state;
1101                         break;
1102
1103                         case ACPI_STATE_C2:
1104                         state->flags |= CPUIDLE_FLAG_BALANCED;
1105                         state->flags |= CPUIDLE_FLAG_TIME_VALID;
1106                         state->enter = acpi_idle_enter_simple;
1107                         dev->safe_state = state;
1108                         break;
1109
1110                         case ACPI_STATE_C3:
1111                         state->flags |= CPUIDLE_FLAG_DEEP;
1112                         state->flags |= CPUIDLE_FLAG_TIME_VALID;
1113                         state->flags |= CPUIDLE_FLAG_CHECK_BM;
1114                         state->enter = pr->flags.bm_check ?
1115                                         acpi_idle_enter_bm :
1116                                         acpi_idle_enter_simple;
1117                         break;
1118                 }
1119
1120                 count++;
1121                 if (count == CPUIDLE_STATE_MAX)
1122                         break;
1123         }
1124
1125         dev->state_count = count;
1126
1127         if (!count)
1128                 return -EINVAL;
1129
1130         return 0;
1131 }
1132
1133 int acpi_processor_cst_has_changed(struct acpi_processor *pr)
1134 {
1135         int ret = 0;
1136
1137         if (boot_option_idle_override)
1138                 return 0;
1139
1140         if (!pr)
1141                 return -EINVAL;
1142
1143         if (nocst) {
1144                 return -ENODEV;
1145         }
1146
1147         if (!pr->flags.power_setup_done)
1148                 return -ENODEV;
1149
1150         cpuidle_pause_and_lock();
1151         cpuidle_disable_device(&pr->power.dev);
1152         acpi_processor_get_power_info(pr);
1153         if (pr->flags.power) {
1154                 acpi_processor_setup_cpuidle(pr);
1155                 ret = cpuidle_enable_device(&pr->power.dev);
1156         }
1157         cpuidle_resume_and_unlock();
1158
1159         return ret;
1160 }
1161
1162 int __cpuinit acpi_processor_power_init(struct acpi_processor *pr,
1163                               struct acpi_device *device)
1164 {
1165         acpi_status status = 0;
1166         static int first_run;
1167 #ifdef CONFIG_ACPI_PROCFS
1168         struct proc_dir_entry *entry = NULL;
1169 #endif
1170
1171         if (boot_option_idle_override)
1172                 return 0;
1173
1174         if (!first_run) {
1175                 if (idle_halt) {
1176                         /*
1177                          * When the boot option of "idle=halt" is added, halt
1178                          * is used for CPU IDLE.
1179                          * In such case C2/C3 is meaningless. So the max_cstate
1180                          * is set to one.
1181                          */
1182                         max_cstate = 1;
1183                 }
1184                 dmi_check_system(processor_power_dmi_table);
1185                 max_cstate = acpi_processor_cstate_check(max_cstate);
1186                 if (max_cstate < ACPI_C_STATES_MAX)
1187                         printk(KERN_NOTICE
1188                                "ACPI: processor limited to max C-state %d\n",
1189                                max_cstate);
1190                 first_run++;
1191         }
1192
1193         if (!pr)
1194                 return -EINVAL;
1195
1196         if (acpi_gbl_FADT.cst_control && !nocst) {
1197                 status =
1198                     acpi_os_write_port(acpi_gbl_FADT.smi_command, acpi_gbl_FADT.cst_control, 8);
1199                 if (ACPI_FAILURE(status)) {
1200                         ACPI_EXCEPTION((AE_INFO, status,
1201                                         "Notifying BIOS of _CST ability failed"));
1202                 }
1203         }
1204
1205         acpi_processor_get_power_info(pr);
1206         pr->flags.power_setup_done = 1;
1207
1208         /*
1209          * Install the idle handler if processor power management is supported.
1210          * Note that we use previously set idle handler will be used on
1211          * platforms that only support C1.
1212          */
1213         if (pr->flags.power) {
1214                 acpi_processor_setup_cpuidle(pr);
1215                 if (cpuidle_register_device(&pr->power.dev))
1216                         return -EIO;
1217         }
1218 #ifdef CONFIG_ACPI_PROCFS
1219         /* 'power' [R] */
1220         entry = proc_create_data(ACPI_PROCESSOR_FILE_POWER,
1221                                  S_IRUGO, acpi_device_dir(device),
1222                                  &acpi_processor_power_fops,
1223                                  acpi_driver_data(device));
1224         if (!entry)
1225                 return -EIO;
1226 #endif
1227         return 0;
1228 }
1229
1230 int acpi_processor_power_exit(struct acpi_processor *pr,
1231                               struct acpi_device *device)
1232 {
1233         if (boot_option_idle_override)
1234                 return 0;
1235
1236         cpuidle_unregister_device(&pr->power.dev);
1237         pr->flags.power_setup_done = 0;
1238
1239 #ifdef CONFIG_ACPI_PROCFS
1240         if (acpi_device_dir(device))
1241                 remove_proc_entry(ACPI_PROCESSOR_FILE_POWER,
1242                                   acpi_device_dir(device));
1243 #endif
1244
1245         return 0;
1246 }