drivers/idle/intel_idle.c

   1 /*
   2  * intel_idle.c - native hardware idle loop for modern Intel processors
   3  *
   4  * Copyright (c) 2010, Intel Corporation.
   5  * Len Brown <len.brown@intel.com>
   6  *
   7  * This program is free software; you can redistribute it and/or modify it
   8  * under the terms and conditions of the GNU General Public License,
   9  * version 2, as published by the Free Software Foundation.
  10  *
  11  * This program is distributed in the hope it will be useful, but WITHOUT
  12  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  13  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License for
  14  * more details.
  15  *
  16  * You should have received a copy of the GNU General Public License along with
  17  * this program; if not, write to the Free Software Foundation, Inc.,
  18  * 51 Franklin St - Fifth Floor, Boston, MA 02110-1301 USA.
  19  */
  20
  21 /*
  22  * intel_idle is a cpuidle driver that loads on specific Intel processors
  23  * in lieu of the legacy ACPI processor_idle driver.  The intent is to
  24  * make Linux more efficient on these processors, as intel_idle knows
  25  * more than ACPI, as well as make Linux more immune to ACPI BIOS bugs.
  26  */
  27
  28 /*
  29  * Design Assumptions
  30  *
  31  * All CPUs have same idle states as boot CPU
  32  *
  33  * Chipset BM_STS (bus master status) bit is a NOP
  34  *      for preventing entry into deep C-stats
  35  */
  36
  37 /*
  38  * Known limitations
  39  *
  40  * The driver currently initializes for_each_online_cpu() upon modprobe.
  41  * It it unaware of subsequent processors hot-added to the system.
  42  * This means that if you boot with maxcpus=n and later online
  43  * processors above n, those processors will use C1 only.
  44  *
  45  * ACPI has a .suspend hack to turn off deep c-statees during suspend
  46  * to avoid complications with the lapic timer workaround.
  47  * Have not seen issues with suspend, but may need same workaround here.
  48  *
  49  * There is currently no kernel-based automatic probing/loading mechanism
  50  * if the driver is built as a module.
  51  */
  52
  53 /* un-comment DEBUG to enable pr_debug() statements */
  54 #define DEBUG
  55
  56 #include <linux/kernel.h>
  57 #include <linux/cpuidle.h>
  58 #include <linux/clockchips.h>
  59 #include <linux/hrtimer.h>      /* ktime_get_real() */
  60 #include <trace/events/power.h>
  61 #include <linux/sched.h>
  62 #include <linux/notifier.h>
  63 #include <linux/cpu.h>
  64 #include <linux/module.h>
  65 #include <asm/mwait.h>
  66 #include <asm/msr.h>
  67
  68 #define INTEL_IDLE_VERSION "0.4"
  69 #define PREFIX "intel_idle: "
  70
  71 static struct cpuidle_driver intel_idle_driver = {
  72         .name = "intel_idle",
  73         .owner = THIS_MODULE,
  74 };
  75 /* intel_idle.max_cstate=0 disables driver */
  76 static int max_cstate = MWAIT_MAX_NUM_CSTATES - 1;
  77
  78 static unsigned int mwait_substates;
  79
  80 #define LAPIC_TIMER_ALWAYS_RELIABLE 0xFFFFFFFF
  81 /* Reliable LAPIC Timer States, bit 1 for C1 etc.  */
  82 static unsigned int lapic_timer_reliable_states = (1 << 1);      /* Default to only C1 */
  83
  84 static struct cpuidle_device __percpu *intel_idle_cpuidle_devices;
  85 static int intel_idle(struct cpuidle_device *dev,
  86                         struct cpuidle_driver *drv, int index);
  87
  88 static struct cpuidle_state *cpuidle_state_table;
  89
  90 /*
  91  * Hardware C-state auto-demotion may not always be optimal.
  92  * Indicate which enable bits to clear here.
  93  */
  94 static unsigned long long auto_demotion_disable_flags;
  95
  96 /*
  97  * Set this flag for states where the HW flushes the TLB for us
  98  * and so we don't need cross-calls to keep it consistent.
  99  * If this flag is set, SW flushes the TLB, so even if the
 100  * HW doesn't do the flushing, this flag is safe to use.
 101  */
 102 #define CPUIDLE_FLAG_TLB_FLUSHED        0x10000
 103
 104 /*
 105  * States are indexed by the cstate number,
 106  * which is also the index into the MWAIT hint array.
 107  * Thus C0 is a dummy.
 108  */
 109 static struct cpuidle_state nehalem_cstates[MWAIT_MAX_NUM_CSTATES] = {
 110         { /* MWAIT C0 */ },
 111         { /* MWAIT C1 */
 112                 .name = "C1-NHM",
 113                 .desc = "MWAIT 0x00",
 114                 .flags = CPUIDLE_FLAG_TIME_VALID,
 115                 .exit_latency = 3,
 116                 .target_residency = 6,
 117                 .enter = &intel_idle },
 118         { /* MWAIT C2 */
 119                 .name = "C3-NHM",
 120                 .desc = "MWAIT 0x10",
 121                 .flags = CPUIDLE_FLAG_TIME_VALID | CPUIDLE_FLAG_TLB_FLUSHED,
 122                 .exit_latency = 20,
 123                 .target_residency = 80,
 124                 .enter = &intel_idle },
 125         { /* MWAIT C3 */
 126                 .name = "C6-NHM",
 127                 .desc = "MWAIT 0x20",
 128                 .flags = CPUIDLE_FLAG_TIME_VALID | CPUIDLE_FLAG_TLB_FLUSHED,
 129                 .exit_latency = 200,
 130                 .target_residency = 800,
 131                 .enter = &intel_idle },
 132 };
 133
 134 static struct cpuidle_state snb_cstates[MWAIT_MAX_NUM_CSTATES] = {
 135         { /* MWAIT C0 */ },
 136         { /* MWAIT C1 */
 137                 .name = "C1-SNB",
 138                 .desc = "MWAIT 0x00",
 139                 .flags = CPUIDLE_FLAG_TIME_VALID,
 140                 .exit_latency = 1,
 141                 .target_residency = 1,
 142                 .enter = &intel_idle },
 143         { /* MWAIT C2 */
 144                 .name = "C3-SNB",
 145                 .desc = "MWAIT 0x10",
 146                 .flags = CPUIDLE_FLAG_TIME_VALID | CPUIDLE_FLAG_TLB_FLUSHED,
 147                 .exit_latency = 80,
 148                 .target_residency = 211,
 149                 .enter = &intel_idle },
 150         { /* MWAIT C3 */
 151                 .name = "C6-SNB",
 152                 .desc = "MWAIT 0x20",
 153                 .flags = CPUIDLE_FLAG_TIME_VALID | CPUIDLE_FLAG_TLB_FLUSHED,
 154                 .exit_latency = 104,
 155                 .target_residency = 345,
 156                 .enter = &intel_idle },
 157         { /* MWAIT C4 */
 158                 .name = "C7-SNB",
 159                 .desc = "MWAIT 0x30",
 160                 .flags = CPUIDLE_FLAG_TIME_VALID | CPUIDLE_FLAG_TLB_FLUSHED,
 161                 .exit_latency = 109,
 162                 .target_residency = 345,
 163                 .enter = &intel_idle },
 164 };
 165
 166 static struct cpuidle_state atom_cstates[MWAIT_MAX_NUM_CSTATES] = {
 167         { /* MWAIT C0 */ },
 168         { /* MWAIT C1 */
 169                 .name = "C1-ATM",
 170                 .desc = "MWAIT 0x00",
 171                 .flags = CPUIDLE_FLAG_TIME_VALID,
 172                 .exit_latency = 1,
 173                 .target_residency = 4,
 174                 .enter = &intel_idle },
 175         { /* MWAIT C2 */
 176                 .name = "C2-ATM",
 177                 .desc = "MWAIT 0x10",
 178                 .flags = CPUIDLE_FLAG_TIME_VALID,
 179                 .exit_latency = 20,
 180                 .target_residency = 80,
 181                 .enter = &intel_idle },
 182         { /* MWAIT C3 */ },
 183         { /* MWAIT C4 */
 184                 .name = "C4-ATM",
 185                 .desc = "MWAIT 0x30",
 186                 .flags = CPUIDLE_FLAG_TIME_VALID | CPUIDLE_FLAG_TLB_FLUSHED,
 187                 .exit_latency = 100,
 188                 .target_residency = 400,
 189                 .enter = &intel_idle },
 190         { /* MWAIT C5 */ },
 191         { /* MWAIT C6 */
 192                 .name = "C6-ATM",
 193                 .desc = "MWAIT 0x52",
 194                 .flags = CPUIDLE_FLAG_TIME_VALID | CPUIDLE_FLAG_TLB_FLUSHED,
 195                 .exit_latency = 140,
 196                 .target_residency = 560,
 197                 .enter = &intel_idle },
 198 };
 199
 200 static int get_driver_data(int cstate)
 201 {
 202         int driver_data;
 203         switch (cstate) {
 204
 205         case 1: /* MWAIT C1 */
 206                 driver_data = 0x00;
 207                 break;
 208         case 2: /* MWAIT C2 */
 209                 driver_data = 0x10;
 210                 break;
 211         case 3: /* MWAIT C3 */
 212                 driver_data = 0x20;
 213                 break;
 214         case 4: /* MWAIT C4 */
 215                 driver_data = 0x30;
 216                 break;
 217         case 5: /* MWAIT C5 */
 218                 driver_data = 0x40;
 219                 break;
 220         case 6: /* MWAIT C6 */
 221                 driver_data = 0x52;
 222                 break;
 223         default:
 224                 driver_data = 0x00;
 225         }
 226         return driver_data;
 227 }
 228
 229 /**
 230  * intel_idle
 231  * @dev: cpuidle_device
 232  * @drv: cpuidle driver
 233  * @index: index of cpuidle state
 234  *
 235  */
 236 static int intel_idle(struct cpuidle_device *dev,
 237                 struct cpuidle_driver *drv, int index)
 238 {
 239         unsigned long ecx = 1; /* break on interrupt flag */
 240         struct cpuidle_state *state = &drv->states[index];
 241         struct cpuidle_state_usage *state_usage = &dev->states_usage[index];
 242         unsigned long eax = (unsigned long)cpuidle_get_statedata(state_usage);
 243         unsigned int cstate;
 244         ktime_t kt_before, kt_after;
 245         s64 usec_delta;
 246         int cpu = smp_processor_id();
 247
 248         cstate = (((eax) >> MWAIT_SUBSTATE_SIZE) & MWAIT_CSTATE_MASK) + 1;
 249
 250         local_irq_disable();
 251
 252         /*
 253          * leave_mm() to avoid costly and often unnecessary wakeups
 254          * for flushing the user TLB's associated with the active mm.
 255          */
 256         if (state->flags & CPUIDLE_FLAG_TLB_FLUSHED)
 257                 leave_mm(cpu);
 258
 259         if (!(lapic_timer_reliable_states & (1 << (cstate))))
 260                 clockevents_notify(CLOCK_EVT_NOTIFY_BROADCAST_ENTER, &cpu);
 261
 262         kt_before = ktime_get_real();
 263
 264         stop_critical_timings();
 265         if (!need_resched()) {
 266
 267                 __monitor((void *)&current_thread_info()->flags, 0, 0);
 268                 smp_mb();
 269                 if (!need_resched())
 270                         __mwait(eax, ecx);
 271         }
 272
 273         start_critical_timings();
 274
 275         kt_after = ktime_get_real();
 276         usec_delta = ktime_to_us(ktime_sub(kt_after, kt_before));
 277
 278         local_irq_enable();
 279
 280         if (!(lapic_timer_reliable_states & (1 << (cstate))))
 281                 clockevents_notify(CLOCK_EVT_NOTIFY_BROADCAST_EXIT, &cpu);
 282
 283         /* Update cpuidle counters */
 284         dev->last_residency = (int)usec_delta;
 285
 286         return index;
 287 }
 288
 289 static void __setup_broadcast_timer(void *arg)
 290 {
 291         unsigned long reason = (unsigned long)arg;
 292         int cpu = smp_processor_id();
 293
 294         reason = reason ?
 295                 CLOCK_EVT_NOTIFY_BROADCAST_ON : CLOCK_EVT_NOTIFY_BROADCAST_OFF;
 296
 297         clockevents_notify(reason, &cpu);
 298 }
 299
 300 static void auto_demotion_disable(void *dummy)
 301 {
 302         unsigned long long msr_bits;
 303
 304         rdmsrl(MSR_NHM_SNB_PKG_CST_CFG_CTL, msr_bits);
 305         msr_bits &= ~auto_demotion_disable_flags;
 306         wrmsrl(MSR_NHM_SNB_PKG_CST_CFG_CTL, msr_bits);
 307 }
 308
 309 static void __intel_idle_notify_handler(void *arg)
 310 {
 311         if (auto_demotion_disable_flags)
 312                 auto_demotion_disable(NULL);
 313
 314         if (lapic_timer_reliable_states != LAPIC_TIMER_ALWAYS_RELIABLE)
 315                 __setup_broadcast_timer((void *)true);
 316 }
 317
 318 static int setup_intelidle_cpuhp_notify(struct notifier_block *n,
 319                 unsigned long action, void *hcpu)
 320 {
 321         int hotcpu = (unsigned long)hcpu;
 322
 323         switch (action & 0xf) {
 324         case CPU_ONLINE:
 325                 smp_call_function_single(hotcpu, __intel_idle_notify_handler,
 326                         NULL, 1);
 327                 break;
 328         }
 329         return NOTIFY_OK;
 330 }
 331
 332 static struct notifier_block setup_intelidle_notifier = {
 333         .notifier_call = setup_intelidle_cpuhp_notify,
 334 };
 335
 336 /*
 337  * intel_idle_probe()
 338  */
 339 static int intel_idle_probe(void)
 340 {
 341         unsigned int eax, ebx, ecx;
 342
 343         if (max_cstate == 0) {
 344                 pr_debug(PREFIX "disabled\n");
 345                 return -EPERM;
 346         }
 347
 348         if (boot_cpu_data.x86_vendor != X86_VENDOR_INTEL)
 349                 return -ENODEV;
 350
 351         if (!boot_cpu_has(X86_FEATURE_MWAIT))
 352                 return -ENODEV;
 353
 354         if (boot_cpu_data.cpuid_level < CPUID_MWAIT_LEAF)
 355                 return -ENODEV;
 356
 357         cpuid(CPUID_MWAIT_LEAF, &eax, &ebx, &ecx, &mwait_substates);
 358
 359         if (!(ecx & CPUID5_ECX_EXTENSIONS_SUPPORTED) ||
 360                 !(ecx & CPUID5_ECX_INTERRUPT_BREAK))
 361                         return -ENODEV;
 362
 363         pr_debug(PREFIX "MWAIT substates: 0x%x\n", mwait_substates);
 364
 365
 366         if (boot_cpu_data.x86 != 6)     /* family 6 */
 367                 return -ENODEV;
 368
 369         switch (boot_cpu_data.x86_model) {
 370
 371         case 0x1A:      /* Core i7, Xeon 5500 series */
 372         case 0x1E:      /* Core i7 and i5 Processor - Lynnfield Jasper Forest */
 373         case 0x1F:      /* Core i7 and i5 Processor - Nehalem */
 374         case 0x2E:      /* Nehalem-EX Xeon */
 375         case 0x2F:      /* Westmere-EX Xeon */
 376         case 0x25:      /* Westmere */
 377         case 0x2C:      /* Westmere */
 378                 cpuidle_state_table = nehalem_cstates;
 379                 auto_demotion_disable_flags =
 380                         (NHM_C1_AUTO_DEMOTE | NHM_C3_AUTO_DEMOTE);
 381                 break;
 382
 383         case 0x1C:      /* 28 - Atom Processor */
 384                 cpuidle_state_table = atom_cstates;
 385                 break;
 386
 387         case 0x26:      /* 38 - Lincroft Atom Processor */
 388                 cpuidle_state_table = atom_cstates;
 389                 auto_demotion_disable_flags = ATM_LNC_C6_AUTO_DEMOTE;
 390                 break;
 391
 392         case 0x2A:      /* SNB */
 393         case 0x2D:      /* SNB Xeon */
 394                 cpuidle_state_table = snb_cstates;
 395                 break;
 396
 397         default:
 398                 pr_debug(PREFIX "does not run on family %d model %d\n",
 399                         boot_cpu_data.x86, boot_cpu_data.x86_model);
 400                 return -ENODEV;
 401         }
 402
 403         if (boot_cpu_has(X86_FEATURE_ARAT))     /* Always Reliable APIC Timer */
 404                 lapic_timer_reliable_states = LAPIC_TIMER_ALWAYS_RELIABLE;
 405         else
 406                 on_each_cpu(__setup_broadcast_timer, (void *)true, 1);
 407
 408         pr_debug(PREFIX "v" INTEL_IDLE_VERSION
 409                 " model 0x%X\n", boot_cpu_data.x86_model);
 410
 411         pr_debug(PREFIX "lapic_timer_reliable_states 0x%x\n",
 412                 lapic_timer_reliable_states);
 413         return 0;
 414 }
 415
 416 /*
 417  * intel_idle_cpuidle_devices_uninit()
 418  * unregister, free cpuidle_devices
 419  */
 420 static void intel_idle_cpuidle_devices_uninit(void)
 421 {
 422         int i;
 423         struct cpuidle_device *dev;
 424
 425         for_each_online_cpu(i) {
 426                 dev = per_cpu_ptr(intel_idle_cpuidle_devices, i);
 427                 cpuidle_unregister_device(dev);
 428         }
 429
 430         free_percpu(intel_idle_cpuidle_devices);
 431         return;
 432 }
 433 /*
 434  * intel_idle_cpuidle_driver_init()
 435  * allocate, initialize cpuidle_states
 436  */
 437 static int intel_idle_cpuidle_driver_init(void)
 438 {
 439         int cstate;
 440         struct cpuidle_driver *drv = &intel_idle_driver;
 441
 442         drv->state_count = 1;
 443
 444         for (cstate = 1; cstate < MWAIT_MAX_NUM_CSTATES; ++cstate) {
 445                 int num_substates;
 446
 447                 if (cstate > max_cstate) {
 448                         printk(PREFIX "max_cstate %d reached\n",
 449                                 max_cstate);
 450                         break;
 451                 }
 452
 453                 /* does the state exist in CPUID.MWAIT? */
 454                 num_substates = (mwait_substates >> ((cstate) * 4))
 455                                         & MWAIT_SUBSTATE_MASK;
 456                 if (num_substates == 0)
 457                         continue;
 458                 /* is the state not enabled? */
 459                 if (cpuidle_state_table[cstate].enter == NULL) {
 460                         /* does the driver not know about the state? */
 461                         if (*cpuidle_state_table[cstate].name == '\0')
 462                                 pr_debug(PREFIX "unaware of model 0x%x"
 463                                         " MWAIT %d please"
 464                                         " contact lenb@kernel.org",
 465                                 boot_cpu_data.x86_model, cstate);
 466                         continue;
 467                 }
 468
 469                 if ((cstate > 2) &&
 470                         !boot_cpu_has(X86_FEATURE_NONSTOP_TSC))
 471                         mark_tsc_unstable("TSC halts in idle"
 472                                         " states deeper than C2");
 473
 474                 drv->states[drv->state_count] = /* structure copy */
 475                         cpuidle_state_table[cstate];
 476
 477                 drv->state_count += 1;
 478         }
 479
 480         if (auto_demotion_disable_flags)
 481                 on_each_cpu(auto_demotion_disable, NULL, 1);
 482
 483         return 0;
 484 }
 485
 486
 487 /*
 488  * intel_idle_cpuidle_devices_init()
 489  * allocate, initialize, register cpuidle_devices
 490  */
 491 static int intel_idle_cpuidle_devices_init(void)
 492 {
 493         int i, cstate;
 494         struct cpuidle_device *dev;
 495
 496         intel_idle_cpuidle_devices = alloc_percpu(struct cpuidle_device);
 497         if (intel_idle_cpuidle_devices == NULL)
 498                 return -ENOMEM;
 499
 500         for_each_online_cpu(i) {
 501                 dev = per_cpu_ptr(intel_idle_cpuidle_devices, i);
 502
 503                 dev->state_count = 1;
 504
 505                 for (cstate = 1; cstate < MWAIT_MAX_NUM_CSTATES; ++cstate) {
 506                         int num_substates;
 507
 508                         if (cstate > max_cstate) {
 509                                 printk(PREFIX "max_cstate %d reached\n",
 510                                         max_cstate);
 511                                 break;
 512                         }
 513
 514                         /* does the state exist in CPUID.MWAIT? */
 515                         num_substates = (mwait_substates >> ((cstate) * 4))
 516                                                 & MWAIT_SUBSTATE_MASK;
 517                         if (num_substates == 0)
 518                                 continue;
 519                         /* is the state not enabled? */
 520                         if (cpuidle_state_table[cstate].enter == NULL) {
 521                                 continue;
 522                         }
 523
 524                         dev->states_usage[dev->state_count].driver_data =
 525                                 (void *)get_driver_data(cstate);
 526
 527                         dev->state_count += 1;
 528                 }
 529
 530                 dev->cpu = i;
 531                 if (cpuidle_register_device(dev)) {
 532                         pr_debug(PREFIX "cpuidle_register_device %d failed!\n",
 533                                  i);
 534                         intel_idle_cpuidle_devices_uninit();
 535                         return -EIO;
 536                 }
 537         }
 538
 539         return 0;
 540 }
 541
 542
 543 static int __init intel_idle_init(void)
 544 {
 545         int retval;
 546
 547         /* Do not load intel_idle at all for now if idle= is passed */
 548         if (boot_option_idle_override != IDLE_NO_OVERRIDE)
 549                 return -ENODEV;
 550
 551         retval = intel_idle_probe();
 552         if (retval)
 553                 return retval;
 554
 555         intel_idle_cpuidle_driver_init();
 556         retval = cpuidle_register_driver(&intel_idle_driver);
 557         if (retval) {
 558                 printk(KERN_DEBUG PREFIX "intel_idle yielding to %s",
 559                         cpuidle_get_driver()->name);
 560                 return retval;
 561         }
 562
 563         retval = intel_idle_cpuidle_devices_init();
 564         if (retval) {
 565                 cpuidle_unregister_driver(&intel_idle_driver);
 566                 return retval;
 567         }
 568
 569         if (auto_demotion_disable_flags || lapic_timer_reliable_states !=
 570             LAPIC_TIMER_ALWAYS_RELIABLE)
 571                 register_cpu_notifier(&setup_intelidle_notifier);
 572
 573         return 0;
 574 }
 575
 576 static void __exit intel_idle_exit(void)
 577 {
 578         intel_idle_cpuidle_devices_uninit();
 579         cpuidle_unregister_driver(&intel_idle_driver);
 580
 581         if (lapic_timer_reliable_states != LAPIC_TIMER_ALWAYS_RELIABLE)
 582                 on_each_cpu(__setup_broadcast_timer, (void *)false, 1);
 583
 584         if (auto_demotion_disable_flags || lapic_timer_reliable_states !=
 585             LAPIC_TIMER_ALWAYS_RELIABLE)
 586                 unregister_cpu_notifier(&setup_intelidle_notifier);
 587
 588         return;
 589 }
 590
 591 module_init(intel_idle_init);
 592 module_exit(intel_idle_exit);
 593
 594 module_param(max_cstate, int, 0444);
 595
 596 MODULE_AUTHOR("Len Brown <len.brown@intel.com>");
 597 MODULE_DESCRIPTION("Cpuidle driver for Intel Hardware v" INTEL_IDLE_VERSION);
 598 MODULE_LICENSE("GPL");