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 <trace/events/power.h> 60 #include <linux/sched.h> 61 #include <linux/notifier.h> 62 #include <linux/cpu.h> 63 #include <linux/module.h> 64 #include <asm/cpu_device_id.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 .en_core_tk_irqen = 1, 75 }; 76 /* intel_idle.max_cstate=0 disables driver */ 77 static int max_cstate = MWAIT_MAX_NUM_CSTATES - 1; 78 79 static unsigned int mwait_substates; 80 81 #define LAPIC_TIMER_ALWAYS_RELIABLE 0xFFFFFFFF 82 /* Reliable LAPIC Timer States, bit 1 for C1 etc. */ 83 static unsigned int lapic_timer_reliable_states = (1 << 1); /* Default to only C1 */ 84 85 struct idle_cpu { 86 struct cpuidle_state *state_table; 87 88 /* 89 * Hardware C-state auto-demotion may not always be optimal. 90 * Indicate which enable bits to clear here. 91 */ 92 unsigned long auto_demotion_disable_flags; 93 }; 94 95 static const struct idle_cpu *icpu; 96 static struct cpuidle_device __percpu *intel_idle_cpuidle_devices; 97 static int intel_idle(struct cpuidle_device *dev, 98 struct cpuidle_driver *drv, int index); 99 static int intel_idle_cpu_init(int cpu); 100 101 static struct cpuidle_state *cpuidle_state_table; 102 103 /* 104 * Set this flag for states where the HW flushes the TLB for us 105 * and so we don't need cross-calls to keep it consistent. 106 * If this flag is set, SW flushes the TLB, so even if the 107 * HW doesn't do the flushing, this flag is safe to use. 108 */ 109 #define CPUIDLE_FLAG_TLB_FLUSHED 0x10000 110 111 /* 112 * States are indexed by the cstate number, 113 * which is also the index into the MWAIT hint array. 114 * Thus C0 is a dummy. 115 */ 116 static struct cpuidle_state nehalem_cstates[MWAIT_MAX_NUM_CSTATES] = { 117 { /* MWAIT C0 */ }, 118 { /* MWAIT C1 */ 119 .name = "C1-NHM", 120 .desc = "MWAIT 0x00", 121 .flags = CPUIDLE_FLAG_TIME_VALID, 122 .exit_latency = 3, 123 .target_residency = 6, 124 .enter = &intel_idle }, 125 { /* MWAIT C2 */ 126 .name = "C3-NHM", 127 .desc = "MWAIT 0x10", 128 .flags = CPUIDLE_FLAG_TIME_VALID | CPUIDLE_FLAG_TLB_FLUSHED, 129 .exit_latency = 20, 130 .target_residency = 80, 131 .enter = &intel_idle }, 132 { /* MWAIT C3 */ 133 .name = "C6-NHM", 134 .desc = "MWAIT 0x20", 135 .flags = CPUIDLE_FLAG_TIME_VALID | CPUIDLE_FLAG_TLB_FLUSHED, 136 .exit_latency = 200, 137 .target_residency = 800, 138 .enter = &intel_idle }, 139 }; 140 141 static struct cpuidle_state snb_cstates[MWAIT_MAX_NUM_CSTATES] = { 142 { /* MWAIT C0 */ }, 143 { /* MWAIT C1 */ 144 .name = "C1-SNB", 145 .desc = "MWAIT 0x00", 146 .flags = CPUIDLE_FLAG_TIME_VALID, 147 .exit_latency = 1, 148 .target_residency = 1, 149 .enter = &intel_idle }, 150 { /* MWAIT C2 */ 151 .name = "C3-SNB", 152 .desc = "MWAIT 0x10", 153 .flags = CPUIDLE_FLAG_TIME_VALID | CPUIDLE_FLAG_TLB_FLUSHED, 154 .exit_latency = 80, 155 .target_residency = 211, 156 .enter = &intel_idle }, 157 { /* MWAIT C3 */ 158 .name = "C6-SNB", 159 .desc = "MWAIT 0x20", 160 .flags = CPUIDLE_FLAG_TIME_VALID | CPUIDLE_FLAG_TLB_FLUSHED, 161 .exit_latency = 104, 162 .target_residency = 345, 163 .enter = &intel_idle }, 164 { /* MWAIT C4 */ 165 .name = "C7-SNB", 166 .desc = "MWAIT 0x30", 167 .flags = CPUIDLE_FLAG_TIME_VALID | CPUIDLE_FLAG_TLB_FLUSHED, 168 .exit_latency = 109, 169 .target_residency = 345, 170 .enter = &intel_idle }, 171 }; 172 173 static struct cpuidle_state ivb_cstates[MWAIT_MAX_NUM_CSTATES] = { 174 { /* MWAIT C0 */ }, 175 { /* MWAIT C1 */ 176 .name = "C1-IVB", 177 .desc = "MWAIT 0x00", 178 .flags = CPUIDLE_FLAG_TIME_VALID, 179 .exit_latency = 1, 180 .target_residency = 1, 181 .enter = &intel_idle }, 182 { /* MWAIT C2 */ 183 .name = "C3-IVB", 184 .desc = "MWAIT 0x10", 185 .flags = CPUIDLE_FLAG_TIME_VALID | CPUIDLE_FLAG_TLB_FLUSHED, 186 .exit_latency = 59, 187 .target_residency = 156, 188 .enter = &intel_idle }, 189 { /* MWAIT C3 */ 190 .name = "C6-IVB", 191 .desc = "MWAIT 0x20", 192 .flags = CPUIDLE_FLAG_TIME_VALID | CPUIDLE_FLAG_TLB_FLUSHED, 193 .exit_latency = 80, 194 .target_residency = 300, 195 .enter = &intel_idle }, 196 { /* MWAIT C4 */ 197 .name = "C7-IVB", 198 .desc = "MWAIT 0x30", 199 .flags = CPUIDLE_FLAG_TIME_VALID | CPUIDLE_FLAG_TLB_FLUSHED, 200 .exit_latency = 87, 201 .target_residency = 300, 202 .enter = &intel_idle }, 203 }; 204 205 static struct cpuidle_state atom_cstates[MWAIT_MAX_NUM_CSTATES] = { 206 { /* MWAIT C0 */ }, 207 { /* MWAIT C1 */ 208 .name = "C1-ATM", 209 .desc = "MWAIT 0x00", 210 .flags = CPUIDLE_FLAG_TIME_VALID, 211 .exit_latency = 1, 212 .target_residency = 4, 213 .enter = &intel_idle }, 214 { /* MWAIT C2 */ 215 .name = "C2-ATM", 216 .desc = "MWAIT 0x10", 217 .flags = CPUIDLE_FLAG_TIME_VALID, 218 .exit_latency = 20, 219 .target_residency = 80, 220 .enter = &intel_idle }, 221 { /* MWAIT C3 */ }, 222 { /* MWAIT C4 */ 223 .name = "C4-ATM", 224 .desc = "MWAIT 0x30", 225 .flags = CPUIDLE_FLAG_TIME_VALID | CPUIDLE_FLAG_TLB_FLUSHED, 226 .exit_latency = 100, 227 .target_residency = 400, 228 .enter = &intel_idle }, 229 { /* MWAIT C5 */ }, 230 { /* MWAIT C6 */ 231 .name = "C6-ATM", 232 .desc = "MWAIT 0x52", 233 .flags = CPUIDLE_FLAG_TIME_VALID | CPUIDLE_FLAG_TLB_FLUSHED, 234 .exit_latency = 140, 235 .target_residency = 560, 236 .enter = &intel_idle }, 237 }; 238 239 static long get_driver_data(int cstate) 240 { 241 int driver_data; 242 switch (cstate) { 243 244 case 1: /* MWAIT C1 */ 245 driver_data = 0x00; 246 break; 247 case 2: /* MWAIT C2 */ 248 driver_data = 0x10; 249 break; 250 case 3: /* MWAIT C3 */ 251 driver_data = 0x20; 252 break; 253 case 4: /* MWAIT C4 */ 254 driver_data = 0x30; 255 break; 256 case 5: /* MWAIT C5 */ 257 driver_data = 0x40; 258 break; 259 case 6: /* MWAIT C6 */ 260 driver_data = 0x52; 261 break; 262 default: 263 driver_data = 0x00; 264 } 265 return driver_data; 266 } 267 268 /** 269 * intel_idle 270 * @dev: cpuidle_device 271 * @drv: cpuidle driver 272 * @index: index of cpuidle state 273 * 274 * Must be called under local_irq_disable(). 275 */ 276 static int intel_idle(struct cpuidle_device *dev, 277 struct cpuidle_driver *drv, int index) 278 { 279 unsigned long ecx = 1; /* break on interrupt flag */ 280 struct cpuidle_state *state = &drv->states[index]; 281 struct cpuidle_state_usage *state_usage = &dev->states_usage[index]; 282 unsigned long eax = (unsigned long)cpuidle_get_statedata(state_usage); 283 unsigned int cstate; 284 int cpu = smp_processor_id(); 285 286 cstate = (((eax) >> MWAIT_SUBSTATE_SIZE) & MWAIT_CSTATE_MASK) + 1; 287 288 /* 289 * leave_mm() to avoid costly and often unnecessary wakeups 290 * for flushing the user TLB's associated with the active mm. 291 */ 292 if (state->flags & CPUIDLE_FLAG_TLB_FLUSHED) 293 leave_mm(cpu); 294 295 if (!(lapic_timer_reliable_states & (1 << (cstate)))) 296 clockevents_notify(CLOCK_EVT_NOTIFY_BROADCAST_ENTER, &cpu); 297 298 stop_critical_timings(); 299 if (!need_resched()) { 300 301 __monitor((void *)¤t_thread_info()->flags, 0, 0); 302 smp_mb(); 303 if (!need_resched()) 304 __mwait(eax, ecx); 305 } 306 307 start_critical_timings(); 308 309 if (!(lapic_timer_reliable_states & (1 << (cstate)))) 310 clockevents_notify(CLOCK_EVT_NOTIFY_BROADCAST_EXIT, &cpu); 311 312 return index; 313 } 314 315 static void __setup_broadcast_timer(void *arg) 316 { 317 unsigned long reason = (unsigned long)arg; 318 int cpu = smp_processor_id(); 319 320 reason = reason ? 321 CLOCK_EVT_NOTIFY_BROADCAST_ON : CLOCK_EVT_NOTIFY_BROADCAST_OFF; 322 323 clockevents_notify(reason, &cpu); 324 } 325 326 static int cpu_hotplug_notify(struct notifier_block *n, 327 unsigned long action, void *hcpu) 328 { 329 int hotcpu = (unsigned long)hcpu; 330 struct cpuidle_device *dev; 331 332 switch (action & 0xf) { 333 case CPU_ONLINE: 334 335 if (lapic_timer_reliable_states != LAPIC_TIMER_ALWAYS_RELIABLE) 336 smp_call_function_single(hotcpu, __setup_broadcast_timer, 337 (void *)true, 1); 338 339 /* 340 * Some systems can hotplug a cpu at runtime after 341 * the kernel has booted, we have to initialize the 342 * driver in this case 343 */ 344 dev = per_cpu_ptr(intel_idle_cpuidle_devices, hotcpu); 345 if (!dev->registered) 346 intel_idle_cpu_init(hotcpu); 347 348 break; 349 } 350 return NOTIFY_OK; 351 } 352 353 static struct notifier_block cpu_hotplug_notifier = { 354 .notifier_call = cpu_hotplug_notify, 355 }; 356 357 static void auto_demotion_disable(void *dummy) 358 { 359 unsigned long long msr_bits; 360 361 rdmsrl(MSR_NHM_SNB_PKG_CST_CFG_CTL, msr_bits); 362 msr_bits &= ~(icpu->auto_demotion_disable_flags); 363 wrmsrl(MSR_NHM_SNB_PKG_CST_CFG_CTL, msr_bits); 364 } 365 366 static const struct idle_cpu idle_cpu_nehalem = { 367 .state_table = nehalem_cstates, 368 .auto_demotion_disable_flags = NHM_C1_AUTO_DEMOTE | NHM_C3_AUTO_DEMOTE, 369 }; 370 371 static const struct idle_cpu idle_cpu_atom = { 372 .state_table = atom_cstates, 373 }; 374 375 static const struct idle_cpu idle_cpu_lincroft = { 376 .state_table = atom_cstates, 377 .auto_demotion_disable_flags = ATM_LNC_C6_AUTO_DEMOTE, 378 }; 379 380 static const struct idle_cpu idle_cpu_snb = { 381 .state_table = snb_cstates, 382 }; 383 384 static const struct idle_cpu idle_cpu_ivb = { 385 .state_table = ivb_cstates, 386 }; 387 388 #define ICPU(model, cpu) \ 389 { X86_VENDOR_INTEL, 6, model, X86_FEATURE_MWAIT, (unsigned long)&cpu } 390 391 static const struct x86_cpu_id intel_idle_ids[] = { 392 ICPU(0x1a, idle_cpu_nehalem), 393 ICPU(0x1e, idle_cpu_nehalem), 394 ICPU(0x1f, idle_cpu_nehalem), 395 ICPU(0x25, idle_cpu_nehalem), 396 ICPU(0x2c, idle_cpu_nehalem), 397 ICPU(0x2e, idle_cpu_nehalem), 398 ICPU(0x1c, idle_cpu_atom), 399 ICPU(0x26, idle_cpu_lincroft), 400 ICPU(0x2f, idle_cpu_nehalem), 401 ICPU(0x2a, idle_cpu_snb), 402 ICPU(0x2d, idle_cpu_snb), 403 ICPU(0x3a, idle_cpu_ivb), 404 ICPU(0x3e, idle_cpu_ivb), 405 {} 406 }; 407 MODULE_DEVICE_TABLE(x86cpu, intel_idle_ids); 408 409 /* 410 * intel_idle_probe() 411 */ 412 static int intel_idle_probe(void) 413 { 414 unsigned int eax, ebx, ecx; 415 const struct x86_cpu_id *id; 416 417 if (max_cstate == 0) { 418 pr_debug(PREFIX "disabled\n"); 419 return -EPERM; 420 } 421 422 id = x86_match_cpu(intel_idle_ids); 423 if (!id) { 424 if (boot_cpu_data.x86_vendor == X86_VENDOR_INTEL && 425 boot_cpu_data.x86 == 6) 426 pr_debug(PREFIX "does not run on family %d model %d\n", 427 boot_cpu_data.x86, boot_cpu_data.x86_model); 428 return -ENODEV; 429 } 430 431 if (boot_cpu_data.cpuid_level < CPUID_MWAIT_LEAF) 432 return -ENODEV; 433 434 cpuid(CPUID_MWAIT_LEAF, &eax, &ebx, &ecx, &mwait_substates); 435 436 if (!(ecx & CPUID5_ECX_EXTENSIONS_SUPPORTED) || 437 !(ecx & CPUID5_ECX_INTERRUPT_BREAK) || 438 !mwait_substates) 439 return -ENODEV; 440 441 pr_debug(PREFIX "MWAIT substates: 0x%x\n", mwait_substates); 442 443 icpu = (const struct idle_cpu *)id->driver_data; 444 cpuidle_state_table = icpu->state_table; 445 446 if (boot_cpu_has(X86_FEATURE_ARAT)) /* Always Reliable APIC Timer */ 447 lapic_timer_reliable_states = LAPIC_TIMER_ALWAYS_RELIABLE; 448 else 449 on_each_cpu(__setup_broadcast_timer, (void *)true, 1); 450 451 pr_debug(PREFIX "v" INTEL_IDLE_VERSION 452 " model 0x%X\n", boot_cpu_data.x86_model); 453 454 pr_debug(PREFIX "lapic_timer_reliable_states 0x%x\n", 455 lapic_timer_reliable_states); 456 return 0; 457 } 458 459 /* 460 * intel_idle_cpuidle_devices_uninit() 461 * unregister, free cpuidle_devices 462 */ 463 static void intel_idle_cpuidle_devices_uninit(void) 464 { 465 int i; 466 struct cpuidle_device *dev; 467 468 for_each_online_cpu(i) { 469 dev = per_cpu_ptr(intel_idle_cpuidle_devices, i); 470 cpuidle_unregister_device(dev); 471 } 472 473 free_percpu(intel_idle_cpuidle_devices); 474 return; 475 } 476 /* 477 * intel_idle_cpuidle_driver_init() 478 * allocate, initialize cpuidle_states 479 */ 480 static int intel_idle_cpuidle_driver_init(void) 481 { 482 int cstate; 483 struct cpuidle_driver *drv = &intel_idle_driver; 484 485 drv->state_count = 1; 486 487 for (cstate = 1; cstate < MWAIT_MAX_NUM_CSTATES; ++cstate) { 488 int num_substates; 489 490 if (cstate > max_cstate) { 491 printk(PREFIX "max_cstate %d reached\n", 492 max_cstate); 493 break; 494 } 495 496 /* does the state exist in CPUID.MWAIT? */ 497 num_substates = (mwait_substates >> ((cstate) * 4)) 498 & MWAIT_SUBSTATE_MASK; 499 if (num_substates == 0) 500 continue; 501 /* is the state not enabled? */ 502 if (cpuidle_state_table[cstate].enter == NULL) { 503 /* does the driver not know about the state? */ 504 if (*cpuidle_state_table[cstate].name == '\0') 505 pr_debug(PREFIX "unaware of model 0x%x" 506 " MWAIT %d please" 507 " contact lenb@kernel.org\n", 508 boot_cpu_data.x86_model, cstate); 509 continue; 510 } 511 512 if ((cstate > 2) && 513 !boot_cpu_has(X86_FEATURE_NONSTOP_TSC)) 514 mark_tsc_unstable("TSC halts in idle" 515 " states deeper than C2"); 516 517 drv->states[drv->state_count] = /* structure copy */ 518 cpuidle_state_table[cstate]; 519 520 drv->state_count += 1; 521 } 522 523 if (icpu->auto_demotion_disable_flags) 524 on_each_cpu(auto_demotion_disable, NULL, 1); 525 526 return 0; 527 } 528 529 530 /* 531 * intel_idle_cpu_init() 532 * allocate, initialize, register cpuidle_devices 533 * @cpu: cpu/core to initialize 534 */ 535 static int intel_idle_cpu_init(int cpu) 536 { 537 int cstate; 538 struct cpuidle_device *dev; 539 540 dev = per_cpu_ptr(intel_idle_cpuidle_devices, cpu); 541 542 dev->state_count = 1; 543 544 for (cstate = 1; cstate < MWAIT_MAX_NUM_CSTATES; ++cstate) { 545 int num_substates; 546 547 if (cstate > max_cstate) { 548 printk(PREFIX "max_cstate %d reached\n", max_cstate); 549 break; 550 } 551 552 /* does the state exist in CPUID.MWAIT? */ 553 num_substates = (mwait_substates >> ((cstate) * 4)) 554 & MWAIT_SUBSTATE_MASK; 555 if (num_substates == 0) 556 continue; 557 /* is the state not enabled? */ 558 if (cpuidle_state_table[cstate].enter == NULL) 559 continue; 560 561 dev->states_usage[dev->state_count].driver_data = 562 (void *)get_driver_data(cstate); 563 564 dev->state_count += 1; 565 } 566 567 dev->cpu = cpu; 568 569 if (cpuidle_register_device(dev)) { 570 pr_debug(PREFIX "cpuidle_register_device %d failed!\n", cpu); 571 intel_idle_cpuidle_devices_uninit(); 572 return -EIO; 573 } 574 575 if (icpu->auto_demotion_disable_flags) 576 smp_call_function_single(cpu, auto_demotion_disable, NULL, 1); 577 578 return 0; 579 } 580 581 static int __init intel_idle_init(void) 582 { 583 int retval, i; 584 585 /* Do not load intel_idle at all for now if idle= is passed */ 586 if (boot_option_idle_override != IDLE_NO_OVERRIDE) 587 return -ENODEV; 588 589 retval = intel_idle_probe(); 590 if (retval) 591 return retval; 592 593 intel_idle_cpuidle_driver_init(); 594 retval = cpuidle_register_driver(&intel_idle_driver); 595 if (retval) { 596 struct cpuidle_driver *drv = cpuidle_get_driver(); 597 printk(KERN_DEBUG PREFIX "intel_idle yielding to %s", 598 drv ? drv->name : "none"); 599 return retval; 600 } 601 602 intel_idle_cpuidle_devices = alloc_percpu(struct cpuidle_device); 603 if (intel_idle_cpuidle_devices == NULL) 604 return -ENOMEM; 605 606 for_each_online_cpu(i) { 607 retval = intel_idle_cpu_init(i); 608 if (retval) { 609 cpuidle_unregister_driver(&intel_idle_driver); 610 return retval; 611 } 612 } 613 register_cpu_notifier(&cpu_hotplug_notifier); 614 615 return 0; 616 } 617 618 static void __exit intel_idle_exit(void) 619 { 620 intel_idle_cpuidle_devices_uninit(); 621 cpuidle_unregister_driver(&intel_idle_driver); 622 623 624 if (lapic_timer_reliable_states != LAPIC_TIMER_ALWAYS_RELIABLE) 625 on_each_cpu(__setup_broadcast_timer, (void *)false, 1); 626 unregister_cpu_notifier(&cpu_hotplug_notifier); 627 628 return; 629 } 630 631 module_init(intel_idle_init); 632 module_exit(intel_idle_exit); 633 634 module_param(max_cstate, int, 0444); 635 636 MODULE_AUTHOR("Len Brown <len.brown@intel.com>"); 637 MODULE_DESCRIPTION("Cpuidle driver for Intel Hardware v" INTEL_IDLE_VERSION); 638 MODULE_LICENSE("GPL"); 639