core.c revision f10e00f4bf360c36edbe6bf18a6c75b171cbe012
1/* 2 * kernel/sched/core.c 3 * 4 * Kernel scheduler and related syscalls 5 * 6 * Copyright (C) 1991-2002 Linus Torvalds 7 * 8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and 9 * make semaphores SMP safe 10 * 1998-11-19 Implemented schedule_timeout() and related stuff 11 * by Andrea Arcangeli 12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar: 13 * hybrid priority-list and round-robin design with 14 * an array-switch method of distributing timeslices 15 * and per-CPU runqueues. Cleanups and useful suggestions 16 * by Davide Libenzi, preemptible kernel bits by Robert Love. 17 * 2003-09-03 Interactivity tuning by Con Kolivas. 18 * 2004-04-02 Scheduler domains code by Nick Piggin 19 * 2007-04-15 Work begun on replacing all interactivity tuning with a 20 * fair scheduling design by Con Kolivas. 21 * 2007-05-05 Load balancing (smp-nice) and other improvements 22 * by Peter Williams 23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith 24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri 25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins, 26 * Thomas Gleixner, Mike Kravetz 27 */ 28 29#include <linux/mm.h> 30#include <linux/module.h> 31#include <linux/nmi.h> 32#include <linux/init.h> 33#include <linux/uaccess.h> 34#include <linux/highmem.h> 35#include <asm/mmu_context.h> 36#include <linux/interrupt.h> 37#include <linux/capability.h> 38#include <linux/completion.h> 39#include <linux/kernel_stat.h> 40#include <linux/debug_locks.h> 41#include <linux/perf_event.h> 42#include <linux/security.h> 43#include <linux/notifier.h> 44#include <linux/profile.h> 45#include <linux/freezer.h> 46#include <linux/vmalloc.h> 47#include <linux/blkdev.h> 48#include <linux/delay.h> 49#include <linux/pid_namespace.h> 50#include <linux/smp.h> 51#include <linux/threads.h> 52#include <linux/timer.h> 53#include <linux/rcupdate.h> 54#include <linux/cpu.h> 55#include <linux/cpuset.h> 56#include <linux/percpu.h> 57#include <linux/proc_fs.h> 58#include <linux/seq_file.h> 59#include <linux/sysctl.h> 60#include <linux/syscalls.h> 61#include <linux/times.h> 62#include <linux/tsacct_kern.h> 63#include <linux/kprobes.h> 64#include <linux/delayacct.h> 65#include <linux/unistd.h> 66#include <linux/pagemap.h> 67#include <linux/hrtimer.h> 68#include <linux/tick.h> 69#include <linux/debugfs.h> 70#include <linux/ctype.h> 71#include <linux/ftrace.h> 72#include <linux/slab.h> 73#include <linux/init_task.h> 74#include <linux/binfmts.h> 75#include <linux/context_tracking.h> 76#include <linux/compiler.h> 77 78#include <asm/switch_to.h> 79#include <asm/tlb.h> 80#include <asm/irq_regs.h> 81#include <asm/mutex.h> 82#ifdef CONFIG_PARAVIRT 83#include <asm/paravirt.h> 84#endif 85 86#include "sched.h" 87#include "../workqueue_internal.h" 88#include "../smpboot.h" 89 90#define CREATE_TRACE_POINTS 91#include <trace/events/sched.h> 92 93#ifdef smp_mb__before_atomic 94void __smp_mb__before_atomic(void) 95{ 96 smp_mb__before_atomic(); 97} 98EXPORT_SYMBOL(__smp_mb__before_atomic); 99#endif 100 101#ifdef smp_mb__after_atomic 102void __smp_mb__after_atomic(void) 103{ 104 smp_mb__after_atomic(); 105} 106EXPORT_SYMBOL(__smp_mb__after_atomic); 107#endif 108 109void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period) 110{ 111 unsigned long delta; 112 ktime_t soft, hard, now; 113 114 for (;;) { 115 if (hrtimer_active(period_timer)) 116 break; 117 118 now = hrtimer_cb_get_time(period_timer); 119 hrtimer_forward(period_timer, now, period); 120 121 soft = hrtimer_get_softexpires(period_timer); 122 hard = hrtimer_get_expires(period_timer); 123 delta = ktime_to_ns(ktime_sub(hard, soft)); 124 __hrtimer_start_range_ns(period_timer, soft, delta, 125 HRTIMER_MODE_ABS_PINNED, 0); 126 } 127} 128 129DEFINE_MUTEX(sched_domains_mutex); 130DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues); 131 132static void update_rq_clock_task(struct rq *rq, s64 delta); 133 134void update_rq_clock(struct rq *rq) 135{ 136 s64 delta; 137 138 if (rq->skip_clock_update > 0) 139 return; 140 141 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock; 142 if (delta < 0) 143 return; 144 rq->clock += delta; 145 update_rq_clock_task(rq, delta); 146} 147 148/* 149 * Debugging: various feature bits 150 */ 151 152#define SCHED_FEAT(name, enabled) \ 153 (1UL << __SCHED_FEAT_##name) * enabled | 154 155const_debug unsigned int sysctl_sched_features = 156#include "features.h" 157 0; 158 159#undef SCHED_FEAT 160 161#ifdef CONFIG_SCHED_DEBUG 162#define SCHED_FEAT(name, enabled) \ 163 #name , 164 165static const char * const sched_feat_names[] = { 166#include "features.h" 167}; 168 169#undef SCHED_FEAT 170 171static int sched_feat_show(struct seq_file *m, void *v) 172{ 173 int i; 174 175 for (i = 0; i < __SCHED_FEAT_NR; i++) { 176 if (!(sysctl_sched_features & (1UL << i))) 177 seq_puts(m, "NO_"); 178 seq_printf(m, "%s ", sched_feat_names[i]); 179 } 180 seq_puts(m, "\n"); 181 182 return 0; 183} 184 185#ifdef HAVE_JUMP_LABEL 186 187#define jump_label_key__true STATIC_KEY_INIT_TRUE 188#define jump_label_key__false STATIC_KEY_INIT_FALSE 189 190#define SCHED_FEAT(name, enabled) \ 191 jump_label_key__##enabled , 192 193struct static_key sched_feat_keys[__SCHED_FEAT_NR] = { 194#include "features.h" 195}; 196 197#undef SCHED_FEAT 198 199static void sched_feat_disable(int i) 200{ 201 if (static_key_enabled(&sched_feat_keys[i])) 202 static_key_slow_dec(&sched_feat_keys[i]); 203} 204 205static void sched_feat_enable(int i) 206{ 207 if (!static_key_enabled(&sched_feat_keys[i])) 208 static_key_slow_inc(&sched_feat_keys[i]); 209} 210#else 211static void sched_feat_disable(int i) { }; 212static void sched_feat_enable(int i) { }; 213#endif /* HAVE_JUMP_LABEL */ 214 215static int sched_feat_set(char *cmp) 216{ 217 int i; 218 int neg = 0; 219 220 if (strncmp(cmp, "NO_", 3) == 0) { 221 neg = 1; 222 cmp += 3; 223 } 224 225 for (i = 0; i < __SCHED_FEAT_NR; i++) { 226 if (strcmp(cmp, sched_feat_names[i]) == 0) { 227 if (neg) { 228 sysctl_sched_features &= ~(1UL << i); 229 sched_feat_disable(i); 230 } else { 231 sysctl_sched_features |= (1UL << i); 232 sched_feat_enable(i); 233 } 234 break; 235 } 236 } 237 238 return i; 239} 240 241static ssize_t 242sched_feat_write(struct file *filp, const char __user *ubuf, 243 size_t cnt, loff_t *ppos) 244{ 245 char buf[64]; 246 char *cmp; 247 int i; 248 struct inode *inode; 249 250 if (cnt > 63) 251 cnt = 63; 252 253 if (copy_from_user(&buf, ubuf, cnt)) 254 return -EFAULT; 255 256 buf[cnt] = 0; 257 cmp = strstrip(buf); 258 259 /* Ensure the static_key remains in a consistent state */ 260 inode = file_inode(filp); 261 mutex_lock(&inode->i_mutex); 262 i = sched_feat_set(cmp); 263 mutex_unlock(&inode->i_mutex); 264 if (i == __SCHED_FEAT_NR) 265 return -EINVAL; 266 267 *ppos += cnt; 268 269 return cnt; 270} 271 272static int sched_feat_open(struct inode *inode, struct file *filp) 273{ 274 return single_open(filp, sched_feat_show, NULL); 275} 276 277static const struct file_operations sched_feat_fops = { 278 .open = sched_feat_open, 279 .write = sched_feat_write, 280 .read = seq_read, 281 .llseek = seq_lseek, 282 .release = single_release, 283}; 284 285static __init int sched_init_debug(void) 286{ 287 debugfs_create_file("sched_features", 0644, NULL, NULL, 288 &sched_feat_fops); 289 290 return 0; 291} 292late_initcall(sched_init_debug); 293#endif /* CONFIG_SCHED_DEBUG */ 294 295/* 296 * Number of tasks to iterate in a single balance run. 297 * Limited because this is done with IRQs disabled. 298 */ 299const_debug unsigned int sysctl_sched_nr_migrate = 32; 300 301/* 302 * period over which we average the RT time consumption, measured 303 * in ms. 304 * 305 * default: 1s 306 */ 307const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC; 308 309/* 310 * period over which we measure -rt task cpu usage in us. 311 * default: 1s 312 */ 313unsigned int sysctl_sched_rt_period = 1000000; 314 315__read_mostly int scheduler_running; 316 317/* 318 * part of the period that we allow rt tasks to run in us. 319 * default: 0.95s 320 */ 321int sysctl_sched_rt_runtime = 950000; 322 323/* 324 * __task_rq_lock - lock the rq @p resides on. 325 */ 326static inline struct rq *__task_rq_lock(struct task_struct *p) 327 __acquires(rq->lock) 328{ 329 struct rq *rq; 330 331 lockdep_assert_held(&p->pi_lock); 332 333 for (;;) { 334 rq = task_rq(p); 335 raw_spin_lock(&rq->lock); 336 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) 337 return rq; 338 raw_spin_unlock(&rq->lock); 339 340 while (unlikely(task_on_rq_migrating(p))) 341 cpu_relax(); 342 } 343} 344 345/* 346 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on. 347 */ 348static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags) 349 __acquires(p->pi_lock) 350 __acquires(rq->lock) 351{ 352 struct rq *rq; 353 354 for (;;) { 355 raw_spin_lock_irqsave(&p->pi_lock, *flags); 356 rq = task_rq(p); 357 raw_spin_lock(&rq->lock); 358 if (likely(rq == task_rq(p) && !task_on_rq_migrating(p))) 359 return rq; 360 raw_spin_unlock(&rq->lock); 361 raw_spin_unlock_irqrestore(&p->pi_lock, *flags); 362 363 while (unlikely(task_on_rq_migrating(p))) 364 cpu_relax(); 365 } 366} 367 368static void __task_rq_unlock(struct rq *rq) 369 __releases(rq->lock) 370{ 371 raw_spin_unlock(&rq->lock); 372} 373 374static inline void 375task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags) 376 __releases(rq->lock) 377 __releases(p->pi_lock) 378{ 379 raw_spin_unlock(&rq->lock); 380 raw_spin_unlock_irqrestore(&p->pi_lock, *flags); 381} 382 383/* 384 * this_rq_lock - lock this runqueue and disable interrupts. 385 */ 386static struct rq *this_rq_lock(void) 387 __acquires(rq->lock) 388{ 389 struct rq *rq; 390 391 local_irq_disable(); 392 rq = this_rq(); 393 raw_spin_lock(&rq->lock); 394 395 return rq; 396} 397 398#ifdef CONFIG_SCHED_HRTICK 399/* 400 * Use HR-timers to deliver accurate preemption points. 401 */ 402 403static void hrtick_clear(struct rq *rq) 404{ 405 if (hrtimer_active(&rq->hrtick_timer)) 406 hrtimer_cancel(&rq->hrtick_timer); 407} 408 409/* 410 * High-resolution timer tick. 411 * Runs from hardirq context with interrupts disabled. 412 */ 413static enum hrtimer_restart hrtick(struct hrtimer *timer) 414{ 415 struct rq *rq = container_of(timer, struct rq, hrtick_timer); 416 417 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id()); 418 419 raw_spin_lock(&rq->lock); 420 update_rq_clock(rq); 421 rq->curr->sched_class->task_tick(rq, rq->curr, 1); 422 raw_spin_unlock(&rq->lock); 423 424 return HRTIMER_NORESTART; 425} 426 427#ifdef CONFIG_SMP 428 429static int __hrtick_restart(struct rq *rq) 430{ 431 struct hrtimer *timer = &rq->hrtick_timer; 432 ktime_t time = hrtimer_get_softexpires(timer); 433 434 return __hrtimer_start_range_ns(timer, time, 0, HRTIMER_MODE_ABS_PINNED, 0); 435} 436 437/* 438 * called from hardirq (IPI) context 439 */ 440static void __hrtick_start(void *arg) 441{ 442 struct rq *rq = arg; 443 444 raw_spin_lock(&rq->lock); 445 __hrtick_restart(rq); 446 rq->hrtick_csd_pending = 0; 447 raw_spin_unlock(&rq->lock); 448} 449 450/* 451 * Called to set the hrtick timer state. 452 * 453 * called with rq->lock held and irqs disabled 454 */ 455void hrtick_start(struct rq *rq, u64 delay) 456{ 457 struct hrtimer *timer = &rq->hrtick_timer; 458 ktime_t time; 459 s64 delta; 460 461 /* 462 * Don't schedule slices shorter than 10000ns, that just 463 * doesn't make sense and can cause timer DoS. 464 */ 465 delta = max_t(s64, delay, 10000LL); 466 time = ktime_add_ns(timer->base->get_time(), delta); 467 468 hrtimer_set_expires(timer, time); 469 470 if (rq == this_rq()) { 471 __hrtick_restart(rq); 472 } else if (!rq->hrtick_csd_pending) { 473 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd); 474 rq->hrtick_csd_pending = 1; 475 } 476} 477 478static int 479hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu) 480{ 481 int cpu = (int)(long)hcpu; 482 483 switch (action) { 484 case CPU_UP_CANCELED: 485 case CPU_UP_CANCELED_FROZEN: 486 case CPU_DOWN_PREPARE: 487 case CPU_DOWN_PREPARE_FROZEN: 488 case CPU_DEAD: 489 case CPU_DEAD_FROZEN: 490 hrtick_clear(cpu_rq(cpu)); 491 return NOTIFY_OK; 492 } 493 494 return NOTIFY_DONE; 495} 496 497static __init void init_hrtick(void) 498{ 499 hotcpu_notifier(hotplug_hrtick, 0); 500} 501#else 502/* 503 * Called to set the hrtick timer state. 504 * 505 * called with rq->lock held and irqs disabled 506 */ 507void hrtick_start(struct rq *rq, u64 delay) 508{ 509 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0, 510 HRTIMER_MODE_REL_PINNED, 0); 511} 512 513static inline void init_hrtick(void) 514{ 515} 516#endif /* CONFIG_SMP */ 517 518static void init_rq_hrtick(struct rq *rq) 519{ 520#ifdef CONFIG_SMP 521 rq->hrtick_csd_pending = 0; 522 523 rq->hrtick_csd.flags = 0; 524 rq->hrtick_csd.func = __hrtick_start; 525 rq->hrtick_csd.info = rq; 526#endif 527 528 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); 529 rq->hrtick_timer.function = hrtick; 530} 531#else /* CONFIG_SCHED_HRTICK */ 532static inline void hrtick_clear(struct rq *rq) 533{ 534} 535 536static inline void init_rq_hrtick(struct rq *rq) 537{ 538} 539 540static inline void init_hrtick(void) 541{ 542} 543#endif /* CONFIG_SCHED_HRTICK */ 544 545/* 546 * cmpxchg based fetch_or, macro so it works for different integer types 547 */ 548#define fetch_or(ptr, val) \ 549({ typeof(*(ptr)) __old, __val = *(ptr); \ 550 for (;;) { \ 551 __old = cmpxchg((ptr), __val, __val | (val)); \ 552 if (__old == __val) \ 553 break; \ 554 __val = __old; \ 555 } \ 556 __old; \ 557}) 558 559#if defined(CONFIG_SMP) && defined(TIF_POLLING_NRFLAG) 560/* 561 * Atomically set TIF_NEED_RESCHED and test for TIF_POLLING_NRFLAG, 562 * this avoids any races wrt polling state changes and thereby avoids 563 * spurious IPIs. 564 */ 565static bool set_nr_and_not_polling(struct task_struct *p) 566{ 567 struct thread_info *ti = task_thread_info(p); 568 return !(fetch_or(&ti->flags, _TIF_NEED_RESCHED) & _TIF_POLLING_NRFLAG); 569} 570 571/* 572 * Atomically set TIF_NEED_RESCHED if TIF_POLLING_NRFLAG is set. 573 * 574 * If this returns true, then the idle task promises to call 575 * sched_ttwu_pending() and reschedule soon. 576 */ 577static bool set_nr_if_polling(struct task_struct *p) 578{ 579 struct thread_info *ti = task_thread_info(p); 580 typeof(ti->flags) old, val = ACCESS_ONCE(ti->flags); 581 582 for (;;) { 583 if (!(val & _TIF_POLLING_NRFLAG)) 584 return false; 585 if (val & _TIF_NEED_RESCHED) 586 return true; 587 old = cmpxchg(&ti->flags, val, val | _TIF_NEED_RESCHED); 588 if (old == val) 589 break; 590 val = old; 591 } 592 return true; 593} 594 595#else 596static bool set_nr_and_not_polling(struct task_struct *p) 597{ 598 set_tsk_need_resched(p); 599 return true; 600} 601 602#ifdef CONFIG_SMP 603static bool set_nr_if_polling(struct task_struct *p) 604{ 605 return false; 606} 607#endif 608#endif 609 610/* 611 * resched_curr - mark rq's current task 'to be rescheduled now'. 612 * 613 * On UP this means the setting of the need_resched flag, on SMP it 614 * might also involve a cross-CPU call to trigger the scheduler on 615 * the target CPU. 616 */ 617void resched_curr(struct rq *rq) 618{ 619 struct task_struct *curr = rq->curr; 620 int cpu; 621 622 lockdep_assert_held(&rq->lock); 623 624 if (test_tsk_need_resched(curr)) 625 return; 626 627 cpu = cpu_of(rq); 628 629 if (cpu == smp_processor_id()) { 630 set_tsk_need_resched(curr); 631 set_preempt_need_resched(); 632 return; 633 } 634 635 if (set_nr_and_not_polling(curr)) 636 smp_send_reschedule(cpu); 637 else 638 trace_sched_wake_idle_without_ipi(cpu); 639} 640 641void resched_cpu(int cpu) 642{ 643 struct rq *rq = cpu_rq(cpu); 644 unsigned long flags; 645 646 if (!raw_spin_trylock_irqsave(&rq->lock, flags)) 647 return; 648 resched_curr(rq); 649 raw_spin_unlock_irqrestore(&rq->lock, flags); 650} 651 652#ifdef CONFIG_SMP 653#ifdef CONFIG_NO_HZ_COMMON 654/* 655 * In the semi idle case, use the nearest busy cpu for migrating timers 656 * from an idle cpu. This is good for power-savings. 657 * 658 * We don't do similar optimization for completely idle system, as 659 * selecting an idle cpu will add more delays to the timers than intended 660 * (as that cpu's timer base may not be uptodate wrt jiffies etc). 661 */ 662int get_nohz_timer_target(int pinned) 663{ 664 int cpu = smp_processor_id(); 665 int i; 666 struct sched_domain *sd; 667 668 if (pinned || !get_sysctl_timer_migration() || !idle_cpu(cpu)) 669 return cpu; 670 671 rcu_read_lock(); 672 for_each_domain(cpu, sd) { 673 for_each_cpu(i, sched_domain_span(sd)) { 674 if (!idle_cpu(i)) { 675 cpu = i; 676 goto unlock; 677 } 678 } 679 } 680unlock: 681 rcu_read_unlock(); 682 return cpu; 683} 684/* 685 * When add_timer_on() enqueues a timer into the timer wheel of an 686 * idle CPU then this timer might expire before the next timer event 687 * which is scheduled to wake up that CPU. In case of a completely 688 * idle system the next event might even be infinite time into the 689 * future. wake_up_idle_cpu() ensures that the CPU is woken up and 690 * leaves the inner idle loop so the newly added timer is taken into 691 * account when the CPU goes back to idle and evaluates the timer 692 * wheel for the next timer event. 693 */ 694static void wake_up_idle_cpu(int cpu) 695{ 696 struct rq *rq = cpu_rq(cpu); 697 698 if (cpu == smp_processor_id()) 699 return; 700 701 if (set_nr_and_not_polling(rq->idle)) 702 smp_send_reschedule(cpu); 703 else 704 trace_sched_wake_idle_without_ipi(cpu); 705} 706 707static bool wake_up_full_nohz_cpu(int cpu) 708{ 709 /* 710 * We just need the target to call irq_exit() and re-evaluate 711 * the next tick. The nohz full kick at least implies that. 712 * If needed we can still optimize that later with an 713 * empty IRQ. 714 */ 715 if (tick_nohz_full_cpu(cpu)) { 716 if (cpu != smp_processor_id() || 717 tick_nohz_tick_stopped()) 718 tick_nohz_full_kick_cpu(cpu); 719 return true; 720 } 721 722 return false; 723} 724 725void wake_up_nohz_cpu(int cpu) 726{ 727 if (!wake_up_full_nohz_cpu(cpu)) 728 wake_up_idle_cpu(cpu); 729} 730 731static inline bool got_nohz_idle_kick(void) 732{ 733 int cpu = smp_processor_id(); 734 735 if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu))) 736 return false; 737 738 if (idle_cpu(cpu) && !need_resched()) 739 return true; 740 741 /* 742 * We can't run Idle Load Balance on this CPU for this time so we 743 * cancel it and clear NOHZ_BALANCE_KICK 744 */ 745 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)); 746 return false; 747} 748 749#else /* CONFIG_NO_HZ_COMMON */ 750 751static inline bool got_nohz_idle_kick(void) 752{ 753 return false; 754} 755 756#endif /* CONFIG_NO_HZ_COMMON */ 757 758#ifdef CONFIG_NO_HZ_FULL 759bool sched_can_stop_tick(void) 760{ 761 /* 762 * More than one running task need preemption. 763 * nr_running update is assumed to be visible 764 * after IPI is sent from wakers. 765 */ 766 if (this_rq()->nr_running > 1) 767 return false; 768 769 return true; 770} 771#endif /* CONFIG_NO_HZ_FULL */ 772 773void sched_avg_update(struct rq *rq) 774{ 775 s64 period = sched_avg_period(); 776 777 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) { 778 /* 779 * Inline assembly required to prevent the compiler 780 * optimising this loop into a divmod call. 781 * See __iter_div_u64_rem() for another example of this. 782 */ 783 asm("" : "+rm" (rq->age_stamp)); 784 rq->age_stamp += period; 785 rq->rt_avg /= 2; 786 } 787} 788 789#endif /* CONFIG_SMP */ 790 791#if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \ 792 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH))) 793/* 794 * Iterate task_group tree rooted at *from, calling @down when first entering a 795 * node and @up when leaving it for the final time. 796 * 797 * Caller must hold rcu_lock or sufficient equivalent. 798 */ 799int walk_tg_tree_from(struct task_group *from, 800 tg_visitor down, tg_visitor up, void *data) 801{ 802 struct task_group *parent, *child; 803 int ret; 804 805 parent = from; 806 807down: 808 ret = (*down)(parent, data); 809 if (ret) 810 goto out; 811 list_for_each_entry_rcu(child, &parent->children, siblings) { 812 parent = child; 813 goto down; 814 815up: 816 continue; 817 } 818 ret = (*up)(parent, data); 819 if (ret || parent == from) 820 goto out; 821 822 child = parent; 823 parent = parent->parent; 824 if (parent) 825 goto up; 826out: 827 return ret; 828} 829 830int tg_nop(struct task_group *tg, void *data) 831{ 832 return 0; 833} 834#endif 835 836static void set_load_weight(struct task_struct *p) 837{ 838 int prio = p->static_prio - MAX_RT_PRIO; 839 struct load_weight *load = &p->se.load; 840 841 /* 842 * SCHED_IDLE tasks get minimal weight: 843 */ 844 if (p->policy == SCHED_IDLE) { 845 load->weight = scale_load(WEIGHT_IDLEPRIO); 846 load->inv_weight = WMULT_IDLEPRIO; 847 return; 848 } 849 850 load->weight = scale_load(prio_to_weight[prio]); 851 load->inv_weight = prio_to_wmult[prio]; 852} 853 854static void enqueue_task(struct rq *rq, struct task_struct *p, int flags) 855{ 856 update_rq_clock(rq); 857 sched_info_queued(rq, p); 858 p->sched_class->enqueue_task(rq, p, flags); 859} 860 861static void dequeue_task(struct rq *rq, struct task_struct *p, int flags) 862{ 863 update_rq_clock(rq); 864 sched_info_dequeued(rq, p); 865 p->sched_class->dequeue_task(rq, p, flags); 866} 867 868void activate_task(struct rq *rq, struct task_struct *p, int flags) 869{ 870 if (task_contributes_to_load(p)) 871 rq->nr_uninterruptible--; 872 873 enqueue_task(rq, p, flags); 874} 875 876void deactivate_task(struct rq *rq, struct task_struct *p, int flags) 877{ 878 if (task_contributes_to_load(p)) 879 rq->nr_uninterruptible++; 880 881 dequeue_task(rq, p, flags); 882} 883 884static void update_rq_clock_task(struct rq *rq, s64 delta) 885{ 886/* 887 * In theory, the compile should just see 0 here, and optimize out the call 888 * to sched_rt_avg_update. But I don't trust it... 889 */ 890#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING) 891 s64 steal = 0, irq_delta = 0; 892#endif 893#ifdef CONFIG_IRQ_TIME_ACCOUNTING 894 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time; 895 896 /* 897 * Since irq_time is only updated on {soft,}irq_exit, we might run into 898 * this case when a previous update_rq_clock() happened inside a 899 * {soft,}irq region. 900 * 901 * When this happens, we stop ->clock_task and only update the 902 * prev_irq_time stamp to account for the part that fit, so that a next 903 * update will consume the rest. This ensures ->clock_task is 904 * monotonic. 905 * 906 * It does however cause some slight miss-attribution of {soft,}irq 907 * time, a more accurate solution would be to update the irq_time using 908 * the current rq->clock timestamp, except that would require using 909 * atomic ops. 910 */ 911 if (irq_delta > delta) 912 irq_delta = delta; 913 914 rq->prev_irq_time += irq_delta; 915 delta -= irq_delta; 916#endif 917#ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING 918 if (static_key_false((¶virt_steal_rq_enabled))) { 919 steal = paravirt_steal_clock(cpu_of(rq)); 920 steal -= rq->prev_steal_time_rq; 921 922 if (unlikely(steal > delta)) 923 steal = delta; 924 925 rq->prev_steal_time_rq += steal; 926 delta -= steal; 927 } 928#endif 929 930 rq->clock_task += delta; 931 932#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING) 933 if ((irq_delta + steal) && sched_feat(NONTASK_CAPACITY)) 934 sched_rt_avg_update(rq, irq_delta + steal); 935#endif 936} 937 938void sched_set_stop_task(int cpu, struct task_struct *stop) 939{ 940 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 }; 941 struct task_struct *old_stop = cpu_rq(cpu)->stop; 942 943 if (stop) { 944 /* 945 * Make it appear like a SCHED_FIFO task, its something 946 * userspace knows about and won't get confused about. 947 * 948 * Also, it will make PI more or less work without too 949 * much confusion -- but then, stop work should not 950 * rely on PI working anyway. 951 */ 952 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m); 953 954 stop->sched_class = &stop_sched_class; 955 } 956 957 cpu_rq(cpu)->stop = stop; 958 959 if (old_stop) { 960 /* 961 * Reset it back to a normal scheduling class so that 962 * it can die in pieces. 963 */ 964 old_stop->sched_class = &rt_sched_class; 965 } 966} 967 968/* 969 * __normal_prio - return the priority that is based on the static prio 970 */ 971static inline int __normal_prio(struct task_struct *p) 972{ 973 return p->static_prio; 974} 975 976/* 977 * Calculate the expected normal priority: i.e. priority 978 * without taking RT-inheritance into account. Might be 979 * boosted by interactivity modifiers. Changes upon fork, 980 * setprio syscalls, and whenever the interactivity 981 * estimator recalculates. 982 */ 983static inline int normal_prio(struct task_struct *p) 984{ 985 int prio; 986 987 if (task_has_dl_policy(p)) 988 prio = MAX_DL_PRIO-1; 989 else if (task_has_rt_policy(p)) 990 prio = MAX_RT_PRIO-1 - p->rt_priority; 991 else 992 prio = __normal_prio(p); 993 return prio; 994} 995 996/* 997 * Calculate the current priority, i.e. the priority 998 * taken into account by the scheduler. This value might 999 * be boosted by RT tasks, or might be boosted by 1000 * interactivity modifiers. Will be RT if the task got 1001 * RT-boosted. If not then it returns p->normal_prio. 1002 */ 1003static int effective_prio(struct task_struct *p) 1004{ 1005 p->normal_prio = normal_prio(p); 1006 /* 1007 * If we are RT tasks or we were boosted to RT priority, 1008 * keep the priority unchanged. Otherwise, update priority 1009 * to the normal priority: 1010 */ 1011 if (!rt_prio(p->prio)) 1012 return p->normal_prio; 1013 return p->prio; 1014} 1015 1016/** 1017 * task_curr - is this task currently executing on a CPU? 1018 * @p: the task in question. 1019 * 1020 * Return: 1 if the task is currently executing. 0 otherwise. 1021 */ 1022inline int task_curr(const struct task_struct *p) 1023{ 1024 return cpu_curr(task_cpu(p)) == p; 1025} 1026 1027static inline void check_class_changed(struct rq *rq, struct task_struct *p, 1028 const struct sched_class *prev_class, 1029 int oldprio) 1030{ 1031 if (prev_class != p->sched_class) { 1032 if (prev_class->switched_from) 1033 prev_class->switched_from(rq, p); 1034 p->sched_class->switched_to(rq, p); 1035 } else if (oldprio != p->prio || dl_task(p)) 1036 p->sched_class->prio_changed(rq, p, oldprio); 1037} 1038 1039void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags) 1040{ 1041 const struct sched_class *class; 1042 1043 if (p->sched_class == rq->curr->sched_class) { 1044 rq->curr->sched_class->check_preempt_curr(rq, p, flags); 1045 } else { 1046 for_each_class(class) { 1047 if (class == rq->curr->sched_class) 1048 break; 1049 if (class == p->sched_class) { 1050 resched_curr(rq); 1051 break; 1052 } 1053 } 1054 } 1055 1056 /* 1057 * A queue event has occurred, and we're going to schedule. In 1058 * this case, we can save a useless back to back clock update. 1059 */ 1060 if (task_on_rq_queued(rq->curr) && test_tsk_need_resched(rq->curr)) 1061 rq->skip_clock_update = 1; 1062} 1063 1064#ifdef CONFIG_SMP 1065void set_task_cpu(struct task_struct *p, unsigned int new_cpu) 1066{ 1067#ifdef CONFIG_SCHED_DEBUG 1068 /* 1069 * We should never call set_task_cpu() on a blocked task, 1070 * ttwu() will sort out the placement. 1071 */ 1072 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING && 1073 !(task_preempt_count(p) & PREEMPT_ACTIVE)); 1074 1075#ifdef CONFIG_LOCKDEP 1076 /* 1077 * The caller should hold either p->pi_lock or rq->lock, when changing 1078 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks. 1079 * 1080 * sched_move_task() holds both and thus holding either pins the cgroup, 1081 * see task_group(). 1082 * 1083 * Furthermore, all task_rq users should acquire both locks, see 1084 * task_rq_lock(). 1085 */ 1086 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) || 1087 lockdep_is_held(&task_rq(p)->lock))); 1088#endif 1089#endif 1090 1091 trace_sched_migrate_task(p, new_cpu); 1092 1093 if (task_cpu(p) != new_cpu) { 1094 if (p->sched_class->migrate_task_rq) 1095 p->sched_class->migrate_task_rq(p, new_cpu); 1096 p->se.nr_migrations++; 1097 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0); 1098 } 1099 1100 __set_task_cpu(p, new_cpu); 1101} 1102 1103static void __migrate_swap_task(struct task_struct *p, int cpu) 1104{ 1105 if (task_on_rq_queued(p)) { 1106 struct rq *src_rq, *dst_rq; 1107 1108 src_rq = task_rq(p); 1109 dst_rq = cpu_rq(cpu); 1110 1111 deactivate_task(src_rq, p, 0); 1112 set_task_cpu(p, cpu); 1113 activate_task(dst_rq, p, 0); 1114 check_preempt_curr(dst_rq, p, 0); 1115 } else { 1116 /* 1117 * Task isn't running anymore; make it appear like we migrated 1118 * it before it went to sleep. This means on wakeup we make the 1119 * previous cpu our targer instead of where it really is. 1120 */ 1121 p->wake_cpu = cpu; 1122 } 1123} 1124 1125struct migration_swap_arg { 1126 struct task_struct *src_task, *dst_task; 1127 int src_cpu, dst_cpu; 1128}; 1129 1130static int migrate_swap_stop(void *data) 1131{ 1132 struct migration_swap_arg *arg = data; 1133 struct rq *src_rq, *dst_rq; 1134 int ret = -EAGAIN; 1135 1136 src_rq = cpu_rq(arg->src_cpu); 1137 dst_rq = cpu_rq(arg->dst_cpu); 1138 1139 double_raw_lock(&arg->src_task->pi_lock, 1140 &arg->dst_task->pi_lock); 1141 double_rq_lock(src_rq, dst_rq); 1142 if (task_cpu(arg->dst_task) != arg->dst_cpu) 1143 goto unlock; 1144 1145 if (task_cpu(arg->src_task) != arg->src_cpu) 1146 goto unlock; 1147 1148 if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task))) 1149 goto unlock; 1150 1151 if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task))) 1152 goto unlock; 1153 1154 __migrate_swap_task(arg->src_task, arg->dst_cpu); 1155 __migrate_swap_task(arg->dst_task, arg->src_cpu); 1156 1157 ret = 0; 1158 1159unlock: 1160 double_rq_unlock(src_rq, dst_rq); 1161 raw_spin_unlock(&arg->dst_task->pi_lock); 1162 raw_spin_unlock(&arg->src_task->pi_lock); 1163 1164 return ret; 1165} 1166 1167/* 1168 * Cross migrate two tasks 1169 */ 1170int migrate_swap(struct task_struct *cur, struct task_struct *p) 1171{ 1172 struct migration_swap_arg arg; 1173 int ret = -EINVAL; 1174 1175 arg = (struct migration_swap_arg){ 1176 .src_task = cur, 1177 .src_cpu = task_cpu(cur), 1178 .dst_task = p, 1179 .dst_cpu = task_cpu(p), 1180 }; 1181 1182 if (arg.src_cpu == arg.dst_cpu) 1183 goto out; 1184 1185 /* 1186 * These three tests are all lockless; this is OK since all of them 1187 * will be re-checked with proper locks held further down the line. 1188 */ 1189 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu)) 1190 goto out; 1191 1192 if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task))) 1193 goto out; 1194 1195 if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task))) 1196 goto out; 1197 1198 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu); 1199 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg); 1200 1201out: 1202 return ret; 1203} 1204 1205struct migration_arg { 1206 struct task_struct *task; 1207 int dest_cpu; 1208}; 1209 1210static int migration_cpu_stop(void *data); 1211 1212/* 1213 * wait_task_inactive - wait for a thread to unschedule. 1214 * 1215 * If @match_state is nonzero, it's the @p->state value just checked and 1216 * not expected to change. If it changes, i.e. @p might have woken up, 1217 * then return zero. When we succeed in waiting for @p to be off its CPU, 1218 * we return a positive number (its total switch count). If a second call 1219 * a short while later returns the same number, the caller can be sure that 1220 * @p has remained unscheduled the whole time. 1221 * 1222 * The caller must ensure that the task *will* unschedule sometime soon, 1223 * else this function might spin for a *long* time. This function can't 1224 * be called with interrupts off, or it may introduce deadlock with 1225 * smp_call_function() if an IPI is sent by the same process we are 1226 * waiting to become inactive. 1227 */ 1228unsigned long wait_task_inactive(struct task_struct *p, long match_state) 1229{ 1230 unsigned long flags; 1231 int running, queued; 1232 unsigned long ncsw; 1233 struct rq *rq; 1234 1235 for (;;) { 1236 /* 1237 * We do the initial early heuristics without holding 1238 * any task-queue locks at all. We'll only try to get 1239 * the runqueue lock when things look like they will 1240 * work out! 1241 */ 1242 rq = task_rq(p); 1243 1244 /* 1245 * If the task is actively running on another CPU 1246 * still, just relax and busy-wait without holding 1247 * any locks. 1248 * 1249 * NOTE! Since we don't hold any locks, it's not 1250 * even sure that "rq" stays as the right runqueue! 1251 * But we don't care, since "task_running()" will 1252 * return false if the runqueue has changed and p 1253 * is actually now running somewhere else! 1254 */ 1255 while (task_running(rq, p)) { 1256 if (match_state && unlikely(p->state != match_state)) 1257 return 0; 1258 cpu_relax(); 1259 } 1260 1261 /* 1262 * Ok, time to look more closely! We need the rq 1263 * lock now, to be *sure*. If we're wrong, we'll 1264 * just go back and repeat. 1265 */ 1266 rq = task_rq_lock(p, &flags); 1267 trace_sched_wait_task(p); 1268 running = task_running(rq, p); 1269 queued = task_on_rq_queued(p); 1270 ncsw = 0; 1271 if (!match_state || p->state == match_state) 1272 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */ 1273 task_rq_unlock(rq, p, &flags); 1274 1275 /* 1276 * If it changed from the expected state, bail out now. 1277 */ 1278 if (unlikely(!ncsw)) 1279 break; 1280 1281 /* 1282 * Was it really running after all now that we 1283 * checked with the proper locks actually held? 1284 * 1285 * Oops. Go back and try again.. 1286 */ 1287 if (unlikely(running)) { 1288 cpu_relax(); 1289 continue; 1290 } 1291 1292 /* 1293 * It's not enough that it's not actively running, 1294 * it must be off the runqueue _entirely_, and not 1295 * preempted! 1296 * 1297 * So if it was still runnable (but just not actively 1298 * running right now), it's preempted, and we should 1299 * yield - it could be a while. 1300 */ 1301 if (unlikely(queued)) { 1302 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ); 1303 1304 set_current_state(TASK_UNINTERRUPTIBLE); 1305 schedule_hrtimeout(&to, HRTIMER_MODE_REL); 1306 continue; 1307 } 1308 1309 /* 1310 * Ahh, all good. It wasn't running, and it wasn't 1311 * runnable, which means that it will never become 1312 * running in the future either. We're all done! 1313 */ 1314 break; 1315 } 1316 1317 return ncsw; 1318} 1319 1320/*** 1321 * kick_process - kick a running thread to enter/exit the kernel 1322 * @p: the to-be-kicked thread 1323 * 1324 * Cause a process which is running on another CPU to enter 1325 * kernel-mode, without any delay. (to get signals handled.) 1326 * 1327 * NOTE: this function doesn't have to take the runqueue lock, 1328 * because all it wants to ensure is that the remote task enters 1329 * the kernel. If the IPI races and the task has been migrated 1330 * to another CPU then no harm is done and the purpose has been 1331 * achieved as well. 1332 */ 1333void kick_process(struct task_struct *p) 1334{ 1335 int cpu; 1336 1337 preempt_disable(); 1338 cpu = task_cpu(p); 1339 if ((cpu != smp_processor_id()) && task_curr(p)) 1340 smp_send_reschedule(cpu); 1341 preempt_enable(); 1342} 1343EXPORT_SYMBOL_GPL(kick_process); 1344#endif /* CONFIG_SMP */ 1345 1346#ifdef CONFIG_SMP 1347/* 1348 * ->cpus_allowed is protected by both rq->lock and p->pi_lock 1349 */ 1350static int select_fallback_rq(int cpu, struct task_struct *p) 1351{ 1352 int nid = cpu_to_node(cpu); 1353 const struct cpumask *nodemask = NULL; 1354 enum { cpuset, possible, fail } state = cpuset; 1355 int dest_cpu; 1356 1357 /* 1358 * If the node that the cpu is on has been offlined, cpu_to_node() 1359 * will return -1. There is no cpu on the node, and we should 1360 * select the cpu on the other node. 1361 */ 1362 if (nid != -1) { 1363 nodemask = cpumask_of_node(nid); 1364 1365 /* Look for allowed, online CPU in same node. */ 1366 for_each_cpu(dest_cpu, nodemask) { 1367 if (!cpu_online(dest_cpu)) 1368 continue; 1369 if (!cpu_active(dest_cpu)) 1370 continue; 1371 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) 1372 return dest_cpu; 1373 } 1374 } 1375 1376 for (;;) { 1377 /* Any allowed, online CPU? */ 1378 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) { 1379 if (!cpu_online(dest_cpu)) 1380 continue; 1381 if (!cpu_active(dest_cpu)) 1382 continue; 1383 goto out; 1384 } 1385 1386 switch (state) { 1387 case cpuset: 1388 /* No more Mr. Nice Guy. */ 1389 cpuset_cpus_allowed_fallback(p); 1390 state = possible; 1391 break; 1392 1393 case possible: 1394 do_set_cpus_allowed(p, cpu_possible_mask); 1395 state = fail; 1396 break; 1397 1398 case fail: 1399 BUG(); 1400 break; 1401 } 1402 } 1403 1404out: 1405 if (state != cpuset) { 1406 /* 1407 * Don't tell them about moving exiting tasks or 1408 * kernel threads (both mm NULL), since they never 1409 * leave kernel. 1410 */ 1411 if (p->mm && printk_ratelimit()) { 1412 printk_deferred("process %d (%s) no longer affine to cpu%d\n", 1413 task_pid_nr(p), p->comm, cpu); 1414 } 1415 } 1416 1417 return dest_cpu; 1418} 1419 1420/* 1421 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable. 1422 */ 1423static inline 1424int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags) 1425{ 1426 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags); 1427 1428 /* 1429 * In order not to call set_task_cpu() on a blocking task we need 1430 * to rely on ttwu() to place the task on a valid ->cpus_allowed 1431 * cpu. 1432 * 1433 * Since this is common to all placement strategies, this lives here. 1434 * 1435 * [ this allows ->select_task() to simply return task_cpu(p) and 1436 * not worry about this generic constraint ] 1437 */ 1438 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) || 1439 !cpu_online(cpu))) 1440 cpu = select_fallback_rq(task_cpu(p), p); 1441 1442 return cpu; 1443} 1444 1445static void update_avg(u64 *avg, u64 sample) 1446{ 1447 s64 diff = sample - *avg; 1448 *avg += diff >> 3; 1449} 1450#endif 1451 1452static void 1453ttwu_stat(struct task_struct *p, int cpu, int wake_flags) 1454{ 1455#ifdef CONFIG_SCHEDSTATS 1456 struct rq *rq = this_rq(); 1457 1458#ifdef CONFIG_SMP 1459 int this_cpu = smp_processor_id(); 1460 1461 if (cpu == this_cpu) { 1462 schedstat_inc(rq, ttwu_local); 1463 schedstat_inc(p, se.statistics.nr_wakeups_local); 1464 } else { 1465 struct sched_domain *sd; 1466 1467 schedstat_inc(p, se.statistics.nr_wakeups_remote); 1468 rcu_read_lock(); 1469 for_each_domain(this_cpu, sd) { 1470 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) { 1471 schedstat_inc(sd, ttwu_wake_remote); 1472 break; 1473 } 1474 } 1475 rcu_read_unlock(); 1476 } 1477 1478 if (wake_flags & WF_MIGRATED) 1479 schedstat_inc(p, se.statistics.nr_wakeups_migrate); 1480 1481#endif /* CONFIG_SMP */ 1482 1483 schedstat_inc(rq, ttwu_count); 1484 schedstat_inc(p, se.statistics.nr_wakeups); 1485 1486 if (wake_flags & WF_SYNC) 1487 schedstat_inc(p, se.statistics.nr_wakeups_sync); 1488 1489#endif /* CONFIG_SCHEDSTATS */ 1490} 1491 1492static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags) 1493{ 1494 activate_task(rq, p, en_flags); 1495 p->on_rq = TASK_ON_RQ_QUEUED; 1496 1497 /* if a worker is waking up, notify workqueue */ 1498 if (p->flags & PF_WQ_WORKER) 1499 wq_worker_waking_up(p, cpu_of(rq)); 1500} 1501 1502/* 1503 * Mark the task runnable and perform wakeup-preemption. 1504 */ 1505static void 1506ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags) 1507{ 1508 check_preempt_curr(rq, p, wake_flags); 1509 trace_sched_wakeup(p, true); 1510 1511 p->state = TASK_RUNNING; 1512#ifdef CONFIG_SMP 1513 if (p->sched_class->task_woken) 1514 p->sched_class->task_woken(rq, p); 1515 1516 if (rq->idle_stamp) { 1517 u64 delta = rq_clock(rq) - rq->idle_stamp; 1518 u64 max = 2*rq->max_idle_balance_cost; 1519 1520 update_avg(&rq->avg_idle, delta); 1521 1522 if (rq->avg_idle > max) 1523 rq->avg_idle = max; 1524 1525 rq->idle_stamp = 0; 1526 } 1527#endif 1528} 1529 1530static void 1531ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags) 1532{ 1533#ifdef CONFIG_SMP 1534 if (p->sched_contributes_to_load) 1535 rq->nr_uninterruptible--; 1536#endif 1537 1538 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING); 1539 ttwu_do_wakeup(rq, p, wake_flags); 1540} 1541 1542/* 1543 * Called in case the task @p isn't fully descheduled from its runqueue, 1544 * in this case we must do a remote wakeup. Its a 'light' wakeup though, 1545 * since all we need to do is flip p->state to TASK_RUNNING, since 1546 * the task is still ->on_rq. 1547 */ 1548static int ttwu_remote(struct task_struct *p, int wake_flags) 1549{ 1550 struct rq *rq; 1551 int ret = 0; 1552 1553 rq = __task_rq_lock(p); 1554 if (task_on_rq_queued(p)) { 1555 /* check_preempt_curr() may use rq clock */ 1556 update_rq_clock(rq); 1557 ttwu_do_wakeup(rq, p, wake_flags); 1558 ret = 1; 1559 } 1560 __task_rq_unlock(rq); 1561 1562 return ret; 1563} 1564 1565#ifdef CONFIG_SMP 1566void sched_ttwu_pending(void) 1567{ 1568 struct rq *rq = this_rq(); 1569 struct llist_node *llist = llist_del_all(&rq->wake_list); 1570 struct task_struct *p; 1571 unsigned long flags; 1572 1573 if (!llist) 1574 return; 1575 1576 raw_spin_lock_irqsave(&rq->lock, flags); 1577 1578 while (llist) { 1579 p = llist_entry(llist, struct task_struct, wake_entry); 1580 llist = llist_next(llist); 1581 ttwu_do_activate(rq, p, 0); 1582 } 1583 1584 raw_spin_unlock_irqrestore(&rq->lock, flags); 1585} 1586 1587void scheduler_ipi(void) 1588{ 1589 /* 1590 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting 1591 * TIF_NEED_RESCHED remotely (for the first time) will also send 1592 * this IPI. 1593 */ 1594 preempt_fold_need_resched(); 1595 1596 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick()) 1597 return; 1598 1599 /* 1600 * Not all reschedule IPI handlers call irq_enter/irq_exit, since 1601 * traditionally all their work was done from the interrupt return 1602 * path. Now that we actually do some work, we need to make sure 1603 * we do call them. 1604 * 1605 * Some archs already do call them, luckily irq_enter/exit nest 1606 * properly. 1607 * 1608 * Arguably we should visit all archs and update all handlers, 1609 * however a fair share of IPIs are still resched only so this would 1610 * somewhat pessimize the simple resched case. 1611 */ 1612 irq_enter(); 1613 sched_ttwu_pending(); 1614 1615 /* 1616 * Check if someone kicked us for doing the nohz idle load balance. 1617 */ 1618 if (unlikely(got_nohz_idle_kick())) { 1619 this_rq()->idle_balance = 1; 1620 raise_softirq_irqoff(SCHED_SOFTIRQ); 1621 } 1622 irq_exit(); 1623} 1624 1625static void ttwu_queue_remote(struct task_struct *p, int cpu) 1626{ 1627 struct rq *rq = cpu_rq(cpu); 1628 1629 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) { 1630 if (!set_nr_if_polling(rq->idle)) 1631 smp_send_reschedule(cpu); 1632 else 1633 trace_sched_wake_idle_without_ipi(cpu); 1634 } 1635} 1636 1637void wake_up_if_idle(int cpu) 1638{ 1639 struct rq *rq = cpu_rq(cpu); 1640 unsigned long flags; 1641 1642 if (!is_idle_task(rq->curr)) 1643 return; 1644 1645 if (set_nr_if_polling(rq->idle)) { 1646 trace_sched_wake_idle_without_ipi(cpu); 1647 } else { 1648 raw_spin_lock_irqsave(&rq->lock, flags); 1649 if (is_idle_task(rq->curr)) 1650 smp_send_reschedule(cpu); 1651 /* Else cpu is not in idle, do nothing here */ 1652 raw_spin_unlock_irqrestore(&rq->lock, flags); 1653 } 1654} 1655 1656bool cpus_share_cache(int this_cpu, int that_cpu) 1657{ 1658 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu); 1659} 1660#endif /* CONFIG_SMP */ 1661 1662static void ttwu_queue(struct task_struct *p, int cpu) 1663{ 1664 struct rq *rq = cpu_rq(cpu); 1665 1666#if defined(CONFIG_SMP) 1667 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) { 1668 sched_clock_cpu(cpu); /* sync clocks x-cpu */ 1669 ttwu_queue_remote(p, cpu); 1670 return; 1671 } 1672#endif 1673 1674 raw_spin_lock(&rq->lock); 1675 ttwu_do_activate(rq, p, 0); 1676 raw_spin_unlock(&rq->lock); 1677} 1678 1679/** 1680 * try_to_wake_up - wake up a thread 1681 * @p: the thread to be awakened 1682 * @state: the mask of task states that can be woken 1683 * @wake_flags: wake modifier flags (WF_*) 1684 * 1685 * Put it on the run-queue if it's not already there. The "current" 1686 * thread is always on the run-queue (except when the actual 1687 * re-schedule is in progress), and as such you're allowed to do 1688 * the simpler "current->state = TASK_RUNNING" to mark yourself 1689 * runnable without the overhead of this. 1690 * 1691 * Return: %true if @p was woken up, %false if it was already running. 1692 * or @state didn't match @p's state. 1693 */ 1694static int 1695try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags) 1696{ 1697 unsigned long flags; 1698 int cpu, success = 0; 1699 1700 /* 1701 * If we are going to wake up a thread waiting for CONDITION we 1702 * need to ensure that CONDITION=1 done by the caller can not be 1703 * reordered with p->state check below. This pairs with mb() in 1704 * set_current_state() the waiting thread does. 1705 */ 1706 smp_mb__before_spinlock(); 1707 raw_spin_lock_irqsave(&p->pi_lock, flags); 1708 if (!(p->state & state)) 1709 goto out; 1710 1711 success = 1; /* we're going to change ->state */ 1712 cpu = task_cpu(p); 1713 1714 if (p->on_rq && ttwu_remote(p, wake_flags)) 1715 goto stat; 1716 1717#ifdef CONFIG_SMP 1718 /* 1719 * If the owning (remote) cpu is still in the middle of schedule() with 1720 * this task as prev, wait until its done referencing the task. 1721 */ 1722 while (p->on_cpu) 1723 cpu_relax(); 1724 /* 1725 * Pairs with the smp_wmb() in finish_lock_switch(). 1726 */ 1727 smp_rmb(); 1728 1729 p->sched_contributes_to_load = !!task_contributes_to_load(p); 1730 p->state = TASK_WAKING; 1731 1732 if (p->sched_class->task_waking) 1733 p->sched_class->task_waking(p); 1734 1735 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags); 1736 if (task_cpu(p) != cpu) { 1737 wake_flags |= WF_MIGRATED; 1738 set_task_cpu(p, cpu); 1739 } 1740#endif /* CONFIG_SMP */ 1741 1742 ttwu_queue(p, cpu); 1743stat: 1744 ttwu_stat(p, cpu, wake_flags); 1745out: 1746 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 1747 1748 return success; 1749} 1750 1751/** 1752 * try_to_wake_up_local - try to wake up a local task with rq lock held 1753 * @p: the thread to be awakened 1754 * 1755 * Put @p on the run-queue if it's not already there. The caller must 1756 * ensure that this_rq() is locked, @p is bound to this_rq() and not 1757 * the current task. 1758 */ 1759static void try_to_wake_up_local(struct task_struct *p) 1760{ 1761 struct rq *rq = task_rq(p); 1762 1763 if (WARN_ON_ONCE(rq != this_rq()) || 1764 WARN_ON_ONCE(p == current)) 1765 return; 1766 1767 lockdep_assert_held(&rq->lock); 1768 1769 if (!raw_spin_trylock(&p->pi_lock)) { 1770 raw_spin_unlock(&rq->lock); 1771 raw_spin_lock(&p->pi_lock); 1772 raw_spin_lock(&rq->lock); 1773 } 1774 1775 if (!(p->state & TASK_NORMAL)) 1776 goto out; 1777 1778 if (!task_on_rq_queued(p)) 1779 ttwu_activate(rq, p, ENQUEUE_WAKEUP); 1780 1781 ttwu_do_wakeup(rq, p, 0); 1782 ttwu_stat(p, smp_processor_id(), 0); 1783out: 1784 raw_spin_unlock(&p->pi_lock); 1785} 1786 1787/** 1788 * wake_up_process - Wake up a specific process 1789 * @p: The process to be woken up. 1790 * 1791 * Attempt to wake up the nominated process and move it to the set of runnable 1792 * processes. 1793 * 1794 * Return: 1 if the process was woken up, 0 if it was already running. 1795 * 1796 * It may be assumed that this function implies a write memory barrier before 1797 * changing the task state if and only if any tasks are woken up. 1798 */ 1799int wake_up_process(struct task_struct *p) 1800{ 1801 WARN_ON(task_is_stopped_or_traced(p)); 1802 return try_to_wake_up(p, TASK_NORMAL, 0); 1803} 1804EXPORT_SYMBOL(wake_up_process); 1805 1806int wake_up_state(struct task_struct *p, unsigned int state) 1807{ 1808 return try_to_wake_up(p, state, 0); 1809} 1810 1811/* 1812 * This function clears the sched_dl_entity static params. 1813 */ 1814void __dl_clear_params(struct task_struct *p) 1815{ 1816 struct sched_dl_entity *dl_se = &p->dl; 1817 1818 dl_se->dl_runtime = 0; 1819 dl_se->dl_deadline = 0; 1820 dl_se->dl_period = 0; 1821 dl_se->flags = 0; 1822 dl_se->dl_bw = 0; 1823} 1824 1825/* 1826 * Perform scheduler related setup for a newly forked process p. 1827 * p is forked by current. 1828 * 1829 * __sched_fork() is basic setup used by init_idle() too: 1830 */ 1831static void __sched_fork(unsigned long clone_flags, struct task_struct *p) 1832{ 1833 p->on_rq = 0; 1834 1835 p->se.on_rq = 0; 1836 p->se.exec_start = 0; 1837 p->se.sum_exec_runtime = 0; 1838 p->se.prev_sum_exec_runtime = 0; 1839 p->se.nr_migrations = 0; 1840 p->se.vruntime = 0; 1841 INIT_LIST_HEAD(&p->se.group_node); 1842 1843#ifdef CONFIG_SCHEDSTATS 1844 memset(&p->se.statistics, 0, sizeof(p->se.statistics)); 1845#endif 1846 1847 RB_CLEAR_NODE(&p->dl.rb_node); 1848 hrtimer_init(&p->dl.dl_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); 1849 __dl_clear_params(p); 1850 1851 INIT_LIST_HEAD(&p->rt.run_list); 1852 1853#ifdef CONFIG_PREEMPT_NOTIFIERS 1854 INIT_HLIST_HEAD(&p->preempt_notifiers); 1855#endif 1856 1857#ifdef CONFIG_NUMA_BALANCING 1858 if (p->mm && atomic_read(&p->mm->mm_users) == 1) { 1859 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay); 1860 p->mm->numa_scan_seq = 0; 1861 } 1862 1863 if (clone_flags & CLONE_VM) 1864 p->numa_preferred_nid = current->numa_preferred_nid; 1865 else 1866 p->numa_preferred_nid = -1; 1867 1868 p->node_stamp = 0ULL; 1869 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0; 1870 p->numa_scan_period = sysctl_numa_balancing_scan_delay; 1871 p->numa_work.next = &p->numa_work; 1872 p->numa_faults_memory = NULL; 1873 p->numa_faults_buffer_memory = NULL; 1874 p->last_task_numa_placement = 0; 1875 p->last_sum_exec_runtime = 0; 1876 1877 INIT_LIST_HEAD(&p->numa_entry); 1878 p->numa_group = NULL; 1879#endif /* CONFIG_NUMA_BALANCING */ 1880} 1881 1882#ifdef CONFIG_NUMA_BALANCING 1883#ifdef CONFIG_SCHED_DEBUG 1884void set_numabalancing_state(bool enabled) 1885{ 1886 if (enabled) 1887 sched_feat_set("NUMA"); 1888 else 1889 sched_feat_set("NO_NUMA"); 1890} 1891#else 1892__read_mostly bool numabalancing_enabled; 1893 1894void set_numabalancing_state(bool enabled) 1895{ 1896 numabalancing_enabled = enabled; 1897} 1898#endif /* CONFIG_SCHED_DEBUG */ 1899 1900#ifdef CONFIG_PROC_SYSCTL 1901int sysctl_numa_balancing(struct ctl_table *table, int write, 1902 void __user *buffer, size_t *lenp, loff_t *ppos) 1903{ 1904 struct ctl_table t; 1905 int err; 1906 int state = numabalancing_enabled; 1907 1908 if (write && !capable(CAP_SYS_ADMIN)) 1909 return -EPERM; 1910 1911 t = *table; 1912 t.data = &state; 1913 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos); 1914 if (err < 0) 1915 return err; 1916 if (write) 1917 set_numabalancing_state(state); 1918 return err; 1919} 1920#endif 1921#endif 1922 1923/* 1924 * fork()/clone()-time setup: 1925 */ 1926int sched_fork(unsigned long clone_flags, struct task_struct *p) 1927{ 1928 unsigned long flags; 1929 int cpu = get_cpu(); 1930 1931 __sched_fork(clone_flags, p); 1932 /* 1933 * We mark the process as running here. This guarantees that 1934 * nobody will actually run it, and a signal or other external 1935 * event cannot wake it up and insert it on the runqueue either. 1936 */ 1937 p->state = TASK_RUNNING; 1938 1939 /* 1940 * Make sure we do not leak PI boosting priority to the child. 1941 */ 1942 p->prio = current->normal_prio; 1943 1944 /* 1945 * Revert to default priority/policy on fork if requested. 1946 */ 1947 if (unlikely(p->sched_reset_on_fork)) { 1948 if (task_has_dl_policy(p) || task_has_rt_policy(p)) { 1949 p->policy = SCHED_NORMAL; 1950 p->static_prio = NICE_TO_PRIO(0); 1951 p->rt_priority = 0; 1952 } else if (PRIO_TO_NICE(p->static_prio) < 0) 1953 p->static_prio = NICE_TO_PRIO(0); 1954 1955 p->prio = p->normal_prio = __normal_prio(p); 1956 set_load_weight(p); 1957 1958 /* 1959 * We don't need the reset flag anymore after the fork. It has 1960 * fulfilled its duty: 1961 */ 1962 p->sched_reset_on_fork = 0; 1963 } 1964 1965 if (dl_prio(p->prio)) { 1966 put_cpu(); 1967 return -EAGAIN; 1968 } else if (rt_prio(p->prio)) { 1969 p->sched_class = &rt_sched_class; 1970 } else { 1971 p->sched_class = &fair_sched_class; 1972 } 1973 1974 if (p->sched_class->task_fork) 1975 p->sched_class->task_fork(p); 1976 1977 /* 1978 * The child is not yet in the pid-hash so no cgroup attach races, 1979 * and the cgroup is pinned to this child due to cgroup_fork() 1980 * is ran before sched_fork(). 1981 * 1982 * Silence PROVE_RCU. 1983 */ 1984 raw_spin_lock_irqsave(&p->pi_lock, flags); 1985 set_task_cpu(p, cpu); 1986 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 1987 1988#if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT) 1989 if (likely(sched_info_on())) 1990 memset(&p->sched_info, 0, sizeof(p->sched_info)); 1991#endif 1992#if defined(CONFIG_SMP) 1993 p->on_cpu = 0; 1994#endif 1995 init_task_preempt_count(p); 1996#ifdef CONFIG_SMP 1997 plist_node_init(&p->pushable_tasks, MAX_PRIO); 1998 RB_CLEAR_NODE(&p->pushable_dl_tasks); 1999#endif 2000 2001 put_cpu(); 2002 return 0; 2003} 2004 2005unsigned long to_ratio(u64 period, u64 runtime) 2006{ 2007 if (runtime == RUNTIME_INF) 2008 return 1ULL << 20; 2009 2010 /* 2011 * Doing this here saves a lot of checks in all 2012 * the calling paths, and returning zero seems 2013 * safe for them anyway. 2014 */ 2015 if (period == 0) 2016 return 0; 2017 2018 return div64_u64(runtime << 20, period); 2019} 2020 2021#ifdef CONFIG_SMP 2022inline struct dl_bw *dl_bw_of(int i) 2023{ 2024 rcu_lockdep_assert(rcu_read_lock_sched_held(), 2025 "sched RCU must be held"); 2026 return &cpu_rq(i)->rd->dl_bw; 2027} 2028 2029static inline int dl_bw_cpus(int i) 2030{ 2031 struct root_domain *rd = cpu_rq(i)->rd; 2032 int cpus = 0; 2033 2034 rcu_lockdep_assert(rcu_read_lock_sched_held(), 2035 "sched RCU must be held"); 2036 for_each_cpu_and(i, rd->span, cpu_active_mask) 2037 cpus++; 2038 2039 return cpus; 2040} 2041#else 2042inline struct dl_bw *dl_bw_of(int i) 2043{ 2044 return &cpu_rq(i)->dl.dl_bw; 2045} 2046 2047static inline int dl_bw_cpus(int i) 2048{ 2049 return 1; 2050} 2051#endif 2052 2053static inline 2054void __dl_clear(struct dl_bw *dl_b, u64 tsk_bw) 2055{ 2056 dl_b->total_bw -= tsk_bw; 2057} 2058 2059static inline 2060void __dl_add(struct dl_bw *dl_b, u64 tsk_bw) 2061{ 2062 dl_b->total_bw += tsk_bw; 2063} 2064 2065static inline 2066bool __dl_overflow(struct dl_bw *dl_b, int cpus, u64 old_bw, u64 new_bw) 2067{ 2068 return dl_b->bw != -1 && 2069 dl_b->bw * cpus < dl_b->total_bw - old_bw + new_bw; 2070} 2071 2072/* 2073 * We must be sure that accepting a new task (or allowing changing the 2074 * parameters of an existing one) is consistent with the bandwidth 2075 * constraints. If yes, this function also accordingly updates the currently 2076 * allocated bandwidth to reflect the new situation. 2077 * 2078 * This function is called while holding p's rq->lock. 2079 */ 2080static int dl_overflow(struct task_struct *p, int policy, 2081 const struct sched_attr *attr) 2082{ 2083 2084 struct dl_bw *dl_b = dl_bw_of(task_cpu(p)); 2085 u64 period = attr->sched_period ?: attr->sched_deadline; 2086 u64 runtime = attr->sched_runtime; 2087 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0; 2088 int cpus, err = -1; 2089 2090 if (new_bw == p->dl.dl_bw) 2091 return 0; 2092 2093 /* 2094 * Either if a task, enters, leave, or stays -deadline but changes 2095 * its parameters, we may need to update accordingly the total 2096 * allocated bandwidth of the container. 2097 */ 2098 raw_spin_lock(&dl_b->lock); 2099 cpus = dl_bw_cpus(task_cpu(p)); 2100 if (dl_policy(policy) && !task_has_dl_policy(p) && 2101 !__dl_overflow(dl_b, cpus, 0, new_bw)) { 2102 __dl_add(dl_b, new_bw); 2103 err = 0; 2104 } else if (dl_policy(policy) && task_has_dl_policy(p) && 2105 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) { 2106 __dl_clear(dl_b, p->dl.dl_bw); 2107 __dl_add(dl_b, new_bw); 2108 err = 0; 2109 } else if (!dl_policy(policy) && task_has_dl_policy(p)) { 2110 __dl_clear(dl_b, p->dl.dl_bw); 2111 err = 0; 2112 } 2113 raw_spin_unlock(&dl_b->lock); 2114 2115 return err; 2116} 2117 2118extern void init_dl_bw(struct dl_bw *dl_b); 2119 2120/* 2121 * wake_up_new_task - wake up a newly created task for the first time. 2122 * 2123 * This function will do some initial scheduler statistics housekeeping 2124 * that must be done for every newly created context, then puts the task 2125 * on the runqueue and wakes it. 2126 */ 2127void wake_up_new_task(struct task_struct *p) 2128{ 2129 unsigned long flags; 2130 struct rq *rq; 2131 2132 raw_spin_lock_irqsave(&p->pi_lock, flags); 2133#ifdef CONFIG_SMP 2134 /* 2135 * Fork balancing, do it here and not earlier because: 2136 * - cpus_allowed can change in the fork path 2137 * - any previously selected cpu might disappear through hotplug 2138 */ 2139 set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0)); 2140#endif 2141 2142 /* Initialize new task's runnable average */ 2143 init_task_runnable_average(p); 2144 rq = __task_rq_lock(p); 2145 activate_task(rq, p, 0); 2146 p->on_rq = TASK_ON_RQ_QUEUED; 2147 trace_sched_wakeup_new(p, true); 2148 check_preempt_curr(rq, p, WF_FORK); 2149#ifdef CONFIG_SMP 2150 if (p->sched_class->task_woken) 2151 p->sched_class->task_woken(rq, p); 2152#endif 2153 task_rq_unlock(rq, p, &flags); 2154} 2155 2156#ifdef CONFIG_PREEMPT_NOTIFIERS 2157 2158/** 2159 * preempt_notifier_register - tell me when current is being preempted & rescheduled 2160 * @notifier: notifier struct to register 2161 */ 2162void preempt_notifier_register(struct preempt_notifier *notifier) 2163{ 2164 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers); 2165} 2166EXPORT_SYMBOL_GPL(preempt_notifier_register); 2167 2168/** 2169 * preempt_notifier_unregister - no longer interested in preemption notifications 2170 * @notifier: notifier struct to unregister 2171 * 2172 * This is safe to call from within a preemption notifier. 2173 */ 2174void preempt_notifier_unregister(struct preempt_notifier *notifier) 2175{ 2176 hlist_del(¬ifier->link); 2177} 2178EXPORT_SYMBOL_GPL(preempt_notifier_unregister); 2179 2180static void fire_sched_in_preempt_notifiers(struct task_struct *curr) 2181{ 2182 struct preempt_notifier *notifier; 2183 2184 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link) 2185 notifier->ops->sched_in(notifier, raw_smp_processor_id()); 2186} 2187 2188static void 2189fire_sched_out_preempt_notifiers(struct task_struct *curr, 2190 struct task_struct *next) 2191{ 2192 struct preempt_notifier *notifier; 2193 2194 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link) 2195 notifier->ops->sched_out(notifier, next); 2196} 2197 2198#else /* !CONFIG_PREEMPT_NOTIFIERS */ 2199 2200static void fire_sched_in_preempt_notifiers(struct task_struct *curr) 2201{ 2202} 2203 2204static void 2205fire_sched_out_preempt_notifiers(struct task_struct *curr, 2206 struct task_struct *next) 2207{ 2208} 2209 2210#endif /* CONFIG_PREEMPT_NOTIFIERS */ 2211 2212/** 2213 * prepare_task_switch - prepare to switch tasks 2214 * @rq: the runqueue preparing to switch 2215 * @prev: the current task that is being switched out 2216 * @next: the task we are going to switch to. 2217 * 2218 * This is called with the rq lock held and interrupts off. It must 2219 * be paired with a subsequent finish_task_switch after the context 2220 * switch. 2221 * 2222 * prepare_task_switch sets up locking and calls architecture specific 2223 * hooks. 2224 */ 2225static inline void 2226prepare_task_switch(struct rq *rq, struct task_struct *prev, 2227 struct task_struct *next) 2228{ 2229 trace_sched_switch(prev, next); 2230 sched_info_switch(rq, prev, next); 2231 perf_event_task_sched_out(prev, next); 2232 fire_sched_out_preempt_notifiers(prev, next); 2233 prepare_lock_switch(rq, next); 2234 prepare_arch_switch(next); 2235} 2236 2237/** 2238 * finish_task_switch - clean up after a task-switch 2239 * @rq: runqueue associated with task-switch 2240 * @prev: the thread we just switched away from. 2241 * 2242 * finish_task_switch must be called after the context switch, paired 2243 * with a prepare_task_switch call before the context switch. 2244 * finish_task_switch will reconcile locking set up by prepare_task_switch, 2245 * and do any other architecture-specific cleanup actions. 2246 * 2247 * Note that we may have delayed dropping an mm in context_switch(). If 2248 * so, we finish that here outside of the runqueue lock. (Doing it 2249 * with the lock held can cause deadlocks; see schedule() for 2250 * details.) 2251 */ 2252static void finish_task_switch(struct rq *rq, struct task_struct *prev) 2253 __releases(rq->lock) 2254{ 2255 struct mm_struct *mm = rq->prev_mm; 2256 long prev_state; 2257 2258 rq->prev_mm = NULL; 2259 2260 /* 2261 * A task struct has one reference for the use as "current". 2262 * If a task dies, then it sets TASK_DEAD in tsk->state and calls 2263 * schedule one last time. The schedule call will never return, and 2264 * the scheduled task must drop that reference. 2265 * The test for TASK_DEAD must occur while the runqueue locks are 2266 * still held, otherwise prev could be scheduled on another cpu, die 2267 * there before we look at prev->state, and then the reference would 2268 * be dropped twice. 2269 * Manfred Spraul <manfred@colorfullife.com> 2270 */ 2271 prev_state = prev->state; 2272 vtime_task_switch(prev); 2273 finish_arch_switch(prev); 2274 perf_event_task_sched_in(prev, current); 2275 finish_lock_switch(rq, prev); 2276 finish_arch_post_lock_switch(); 2277 2278 fire_sched_in_preempt_notifiers(current); 2279 if (mm) 2280 mmdrop(mm); 2281 if (unlikely(prev_state == TASK_DEAD)) { 2282 if (prev->sched_class->task_dead) 2283 prev->sched_class->task_dead(prev); 2284 2285 /* 2286 * Remove function-return probe instances associated with this 2287 * task and put them back on the free list. 2288 */ 2289 kprobe_flush_task(prev); 2290 put_task_struct(prev); 2291 } 2292 2293 tick_nohz_task_switch(current); 2294} 2295 2296#ifdef CONFIG_SMP 2297 2298/* rq->lock is NOT held, but preemption is disabled */ 2299static inline void post_schedule(struct rq *rq) 2300{ 2301 if (rq->post_schedule) { 2302 unsigned long flags; 2303 2304 raw_spin_lock_irqsave(&rq->lock, flags); 2305 if (rq->curr->sched_class->post_schedule) 2306 rq->curr->sched_class->post_schedule(rq); 2307 raw_spin_unlock_irqrestore(&rq->lock, flags); 2308 2309 rq->post_schedule = 0; 2310 } 2311} 2312 2313#else 2314 2315static inline void post_schedule(struct rq *rq) 2316{ 2317} 2318 2319#endif 2320 2321/** 2322 * schedule_tail - first thing a freshly forked thread must call. 2323 * @prev: the thread we just switched away from. 2324 */ 2325asmlinkage __visible void schedule_tail(struct task_struct *prev) 2326 __releases(rq->lock) 2327{ 2328 struct rq *rq = this_rq(); 2329 2330 finish_task_switch(rq, prev); 2331 2332 /* 2333 * FIXME: do we need to worry about rq being invalidated by the 2334 * task_switch? 2335 */ 2336 post_schedule(rq); 2337 2338 if (current->set_child_tid) 2339 put_user(task_pid_vnr(current), current->set_child_tid); 2340} 2341 2342/* 2343 * context_switch - switch to the new MM and the new 2344 * thread's register state. 2345 */ 2346static inline void 2347context_switch(struct rq *rq, struct task_struct *prev, 2348 struct task_struct *next) 2349{ 2350 struct mm_struct *mm, *oldmm; 2351 2352 prepare_task_switch(rq, prev, next); 2353 2354 mm = next->mm; 2355 oldmm = prev->active_mm; 2356 /* 2357 * For paravirt, this is coupled with an exit in switch_to to 2358 * combine the page table reload and the switch backend into 2359 * one hypercall. 2360 */ 2361 arch_start_context_switch(prev); 2362 2363 if (!mm) { 2364 next->active_mm = oldmm; 2365 atomic_inc(&oldmm->mm_count); 2366 enter_lazy_tlb(oldmm, next); 2367 } else 2368 switch_mm(oldmm, mm, next); 2369 2370 if (!prev->mm) { 2371 prev->active_mm = NULL; 2372 rq->prev_mm = oldmm; 2373 } 2374 /* 2375 * Since the runqueue lock will be released by the next 2376 * task (which is an invalid locking op but in the case 2377 * of the scheduler it's an obvious special-case), so we 2378 * do an early lockdep release here: 2379 */ 2380 spin_release(&rq->lock.dep_map, 1, _THIS_IP_); 2381 2382 context_tracking_task_switch(prev, next); 2383 /* Here we just switch the register state and the stack. */ 2384 switch_to(prev, next, prev); 2385 2386 barrier(); 2387 /* 2388 * this_rq must be evaluated again because prev may have moved 2389 * CPUs since it called schedule(), thus the 'rq' on its stack 2390 * frame will be invalid. 2391 */ 2392 finish_task_switch(this_rq(), prev); 2393} 2394 2395/* 2396 * nr_running and nr_context_switches: 2397 * 2398 * externally visible scheduler statistics: current number of runnable 2399 * threads, total number of context switches performed since bootup. 2400 */ 2401unsigned long nr_running(void) 2402{ 2403 unsigned long i, sum = 0; 2404 2405 for_each_online_cpu(i) 2406 sum += cpu_rq(i)->nr_running; 2407 2408 return sum; 2409} 2410 2411unsigned long long nr_context_switches(void) 2412{ 2413 int i; 2414 unsigned long long sum = 0; 2415 2416 for_each_possible_cpu(i) 2417 sum += cpu_rq(i)->nr_switches; 2418 2419 return sum; 2420} 2421 2422unsigned long nr_iowait(void) 2423{ 2424 unsigned long i, sum = 0; 2425 2426 for_each_possible_cpu(i) 2427 sum += atomic_read(&cpu_rq(i)->nr_iowait); 2428 2429 return sum; 2430} 2431 2432unsigned long nr_iowait_cpu(int cpu) 2433{ 2434 struct rq *this = cpu_rq(cpu); 2435 return atomic_read(&this->nr_iowait); 2436} 2437 2438void get_iowait_load(unsigned long *nr_waiters, unsigned long *load) 2439{ 2440 struct rq *this = this_rq(); 2441 *nr_waiters = atomic_read(&this->nr_iowait); 2442 *load = this->cpu_load[0]; 2443} 2444 2445#ifdef CONFIG_SMP 2446 2447/* 2448 * sched_exec - execve() is a valuable balancing opportunity, because at 2449 * this point the task has the smallest effective memory and cache footprint. 2450 */ 2451void sched_exec(void) 2452{ 2453 struct task_struct *p = current; 2454 unsigned long flags; 2455 int dest_cpu; 2456 2457 raw_spin_lock_irqsave(&p->pi_lock, flags); 2458 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0); 2459 if (dest_cpu == smp_processor_id()) 2460 goto unlock; 2461 2462 if (likely(cpu_active(dest_cpu))) { 2463 struct migration_arg arg = { p, dest_cpu }; 2464 2465 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 2466 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg); 2467 return; 2468 } 2469unlock: 2470 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 2471} 2472 2473#endif 2474 2475DEFINE_PER_CPU(struct kernel_stat, kstat); 2476DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat); 2477 2478EXPORT_PER_CPU_SYMBOL(kstat); 2479EXPORT_PER_CPU_SYMBOL(kernel_cpustat); 2480 2481/* 2482 * Return any ns on the sched_clock that have not yet been accounted in 2483 * @p in case that task is currently running. 2484 * 2485 * Called with task_rq_lock() held on @rq. 2486 */ 2487static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq) 2488{ 2489 u64 ns = 0; 2490 2491 /* 2492 * Must be ->curr _and_ ->on_rq. If dequeued, we would 2493 * project cycles that may never be accounted to this 2494 * thread, breaking clock_gettime(). 2495 */ 2496 if (task_current(rq, p) && task_on_rq_queued(p)) { 2497 update_rq_clock(rq); 2498 ns = rq_clock_task(rq) - p->se.exec_start; 2499 if ((s64)ns < 0) 2500 ns = 0; 2501 } 2502 2503 return ns; 2504} 2505 2506unsigned long long task_delta_exec(struct task_struct *p) 2507{ 2508 unsigned long flags; 2509 struct rq *rq; 2510 u64 ns = 0; 2511 2512 rq = task_rq_lock(p, &flags); 2513 ns = do_task_delta_exec(p, rq); 2514 task_rq_unlock(rq, p, &flags); 2515 2516 return ns; 2517} 2518 2519/* 2520 * Return accounted runtime for the task. 2521 * In case the task is currently running, return the runtime plus current's 2522 * pending runtime that have not been accounted yet. 2523 */ 2524unsigned long long task_sched_runtime(struct task_struct *p) 2525{ 2526 unsigned long flags; 2527 struct rq *rq; 2528 u64 ns = 0; 2529 2530#if defined(CONFIG_64BIT) && defined(CONFIG_SMP) 2531 /* 2532 * 64-bit doesn't need locks to atomically read a 64bit value. 2533 * So we have a optimization chance when the task's delta_exec is 0. 2534 * Reading ->on_cpu is racy, but this is ok. 2535 * 2536 * If we race with it leaving cpu, we'll take a lock. So we're correct. 2537 * If we race with it entering cpu, unaccounted time is 0. This is 2538 * indistinguishable from the read occurring a few cycles earlier. 2539 * If we see ->on_cpu without ->on_rq, the task is leaving, and has 2540 * been accounted, so we're correct here as well. 2541 */ 2542 if (!p->on_cpu || !task_on_rq_queued(p)) 2543 return p->se.sum_exec_runtime; 2544#endif 2545 2546 rq = task_rq_lock(p, &flags); 2547 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq); 2548 task_rq_unlock(rq, p, &flags); 2549 2550 return ns; 2551} 2552 2553/* 2554 * This function gets called by the timer code, with HZ frequency. 2555 * We call it with interrupts disabled. 2556 */ 2557void scheduler_tick(void) 2558{ 2559 int cpu = smp_processor_id(); 2560 struct rq *rq = cpu_rq(cpu); 2561 struct task_struct *curr = rq->curr; 2562 2563 sched_clock_tick(); 2564 2565 raw_spin_lock(&rq->lock); 2566 update_rq_clock(rq); 2567 curr->sched_class->task_tick(rq, curr, 0); 2568 update_cpu_load_active(rq); 2569 raw_spin_unlock(&rq->lock); 2570 2571 perf_event_task_tick(); 2572 2573#ifdef CONFIG_SMP 2574 rq->idle_balance = idle_cpu(cpu); 2575 trigger_load_balance(rq); 2576#endif 2577 rq_last_tick_reset(rq); 2578} 2579 2580#ifdef CONFIG_NO_HZ_FULL 2581/** 2582 * scheduler_tick_max_deferment 2583 * 2584 * Keep at least one tick per second when a single 2585 * active task is running because the scheduler doesn't 2586 * yet completely support full dynticks environment. 2587 * 2588 * This makes sure that uptime, CFS vruntime, load 2589 * balancing, etc... continue to move forward, even 2590 * with a very low granularity. 2591 * 2592 * Return: Maximum deferment in nanoseconds. 2593 */ 2594u64 scheduler_tick_max_deferment(void) 2595{ 2596 struct rq *rq = this_rq(); 2597 unsigned long next, now = ACCESS_ONCE(jiffies); 2598 2599 next = rq->last_sched_tick + HZ; 2600 2601 if (time_before_eq(next, now)) 2602 return 0; 2603 2604 return jiffies_to_nsecs(next - now); 2605} 2606#endif 2607 2608notrace unsigned long get_parent_ip(unsigned long addr) 2609{ 2610 if (in_lock_functions(addr)) { 2611 addr = CALLER_ADDR2; 2612 if (in_lock_functions(addr)) 2613 addr = CALLER_ADDR3; 2614 } 2615 return addr; 2616} 2617 2618#if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \ 2619 defined(CONFIG_PREEMPT_TRACER)) 2620 2621void preempt_count_add(int val) 2622{ 2623#ifdef CONFIG_DEBUG_PREEMPT 2624 /* 2625 * Underflow? 2626 */ 2627 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0))) 2628 return; 2629#endif 2630 __preempt_count_add(val); 2631#ifdef CONFIG_DEBUG_PREEMPT 2632 /* 2633 * Spinlock count overflowing soon? 2634 */ 2635 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >= 2636 PREEMPT_MASK - 10); 2637#endif 2638 if (preempt_count() == val) { 2639 unsigned long ip = get_parent_ip(CALLER_ADDR1); 2640#ifdef CONFIG_DEBUG_PREEMPT 2641 current->preempt_disable_ip = ip; 2642#endif 2643 trace_preempt_off(CALLER_ADDR0, ip); 2644 } 2645} 2646EXPORT_SYMBOL(preempt_count_add); 2647NOKPROBE_SYMBOL(preempt_count_add); 2648 2649void preempt_count_sub(int val) 2650{ 2651#ifdef CONFIG_DEBUG_PREEMPT 2652 /* 2653 * Underflow? 2654 */ 2655 if (DEBUG_LOCKS_WARN_ON(val > preempt_count())) 2656 return; 2657 /* 2658 * Is the spinlock portion underflowing? 2659 */ 2660 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) && 2661 !(preempt_count() & PREEMPT_MASK))) 2662 return; 2663#endif 2664 2665 if (preempt_count() == val) 2666 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1)); 2667 __preempt_count_sub(val); 2668} 2669EXPORT_SYMBOL(preempt_count_sub); 2670NOKPROBE_SYMBOL(preempt_count_sub); 2671 2672#endif 2673 2674/* 2675 * Print scheduling while atomic bug: 2676 */ 2677static noinline void __schedule_bug(struct task_struct *prev) 2678{ 2679 if (oops_in_progress) 2680 return; 2681 2682 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n", 2683 prev->comm, prev->pid, preempt_count()); 2684 2685 debug_show_held_locks(prev); 2686 print_modules(); 2687 if (irqs_disabled()) 2688 print_irqtrace_events(prev); 2689#ifdef CONFIG_DEBUG_PREEMPT 2690 if (in_atomic_preempt_off()) { 2691 pr_err("Preemption disabled at:"); 2692 print_ip_sym(current->preempt_disable_ip); 2693 pr_cont("\n"); 2694 } 2695#endif 2696 dump_stack(); 2697 add_taint(TAINT_WARN, LOCKDEP_STILL_OK); 2698} 2699 2700/* 2701 * Various schedule()-time debugging checks and statistics: 2702 */ 2703static inline void schedule_debug(struct task_struct *prev) 2704{ 2705#ifdef CONFIG_SCHED_STACK_END_CHECK 2706 BUG_ON(unlikely(task_stack_end_corrupted(prev))); 2707#endif 2708 /* 2709 * Test if we are atomic. Since do_exit() needs to call into 2710 * schedule() atomically, we ignore that path. Otherwise whine 2711 * if we are scheduling when we should not. 2712 */ 2713 if (unlikely(in_atomic_preempt_off() && prev->state != TASK_DEAD)) 2714 __schedule_bug(prev); 2715 rcu_sleep_check(); 2716 2717 profile_hit(SCHED_PROFILING, __builtin_return_address(0)); 2718 2719 schedstat_inc(this_rq(), sched_count); 2720} 2721 2722/* 2723 * Pick up the highest-prio task: 2724 */ 2725static inline struct task_struct * 2726pick_next_task(struct rq *rq, struct task_struct *prev) 2727{ 2728 const struct sched_class *class = &fair_sched_class; 2729 struct task_struct *p; 2730 2731 /* 2732 * Optimization: we know that if all tasks are in 2733 * the fair class we can call that function directly: 2734 */ 2735 if (likely(prev->sched_class == class && 2736 rq->nr_running == rq->cfs.h_nr_running)) { 2737 p = fair_sched_class.pick_next_task(rq, prev); 2738 if (unlikely(p == RETRY_TASK)) 2739 goto again; 2740 2741 /* assumes fair_sched_class->next == idle_sched_class */ 2742 if (unlikely(!p)) 2743 p = idle_sched_class.pick_next_task(rq, prev); 2744 2745 return p; 2746 } 2747 2748again: 2749 for_each_class(class) { 2750 p = class->pick_next_task(rq, prev); 2751 if (p) { 2752 if (unlikely(p == RETRY_TASK)) 2753 goto again; 2754 return p; 2755 } 2756 } 2757 2758 BUG(); /* the idle class will always have a runnable task */ 2759} 2760 2761/* 2762 * __schedule() is the main scheduler function. 2763 * 2764 * The main means of driving the scheduler and thus entering this function are: 2765 * 2766 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc. 2767 * 2768 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return 2769 * paths. For example, see arch/x86/entry_64.S. 2770 * 2771 * To drive preemption between tasks, the scheduler sets the flag in timer 2772 * interrupt handler scheduler_tick(). 2773 * 2774 * 3. Wakeups don't really cause entry into schedule(). They add a 2775 * task to the run-queue and that's it. 2776 * 2777 * Now, if the new task added to the run-queue preempts the current 2778 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets 2779 * called on the nearest possible occasion: 2780 * 2781 * - If the kernel is preemptible (CONFIG_PREEMPT=y): 2782 * 2783 * - in syscall or exception context, at the next outmost 2784 * preempt_enable(). (this might be as soon as the wake_up()'s 2785 * spin_unlock()!) 2786 * 2787 * - in IRQ context, return from interrupt-handler to 2788 * preemptible context 2789 * 2790 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set) 2791 * then at the next: 2792 * 2793 * - cond_resched() call 2794 * - explicit schedule() call 2795 * - return from syscall or exception to user-space 2796 * - return from interrupt-handler to user-space 2797 */ 2798static void __sched __schedule(void) 2799{ 2800 struct task_struct *prev, *next; 2801 unsigned long *switch_count; 2802 struct rq *rq; 2803 int cpu; 2804 2805need_resched: 2806 preempt_disable(); 2807 cpu = smp_processor_id(); 2808 rq = cpu_rq(cpu); 2809 rcu_note_context_switch(cpu); 2810 prev = rq->curr; 2811 2812 schedule_debug(prev); 2813 2814 if (sched_feat(HRTICK)) 2815 hrtick_clear(rq); 2816 2817 /* 2818 * Make sure that signal_pending_state()->signal_pending() below 2819 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE) 2820 * done by the caller to avoid the race with signal_wake_up(). 2821 */ 2822 smp_mb__before_spinlock(); 2823 raw_spin_lock_irq(&rq->lock); 2824 2825 switch_count = &prev->nivcsw; 2826 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) { 2827 if (unlikely(signal_pending_state(prev->state, prev))) { 2828 prev->state = TASK_RUNNING; 2829 } else { 2830 deactivate_task(rq, prev, DEQUEUE_SLEEP); 2831 prev->on_rq = 0; 2832 2833 /* 2834 * If a worker went to sleep, notify and ask workqueue 2835 * whether it wants to wake up a task to maintain 2836 * concurrency. 2837 */ 2838 if (prev->flags & PF_WQ_WORKER) { 2839 struct task_struct *to_wakeup; 2840 2841 to_wakeup = wq_worker_sleeping(prev, cpu); 2842 if (to_wakeup) 2843 try_to_wake_up_local(to_wakeup); 2844 } 2845 } 2846 switch_count = &prev->nvcsw; 2847 } 2848 2849 if (task_on_rq_queued(prev) || rq->skip_clock_update < 0) 2850 update_rq_clock(rq); 2851 2852 next = pick_next_task(rq, prev); 2853 clear_tsk_need_resched(prev); 2854 clear_preempt_need_resched(); 2855 rq->skip_clock_update = 0; 2856 2857 if (likely(prev != next)) { 2858 rq->nr_switches++; 2859 rq->curr = next; 2860 ++*switch_count; 2861 2862 context_switch(rq, prev, next); /* unlocks the rq */ 2863 /* 2864 * The context switch have flipped the stack from under us 2865 * and restored the local variables which were saved when 2866 * this task called schedule() in the past. prev == current 2867 * is still correct, but it can be moved to another cpu/rq. 2868 */ 2869 cpu = smp_processor_id(); 2870 rq = cpu_rq(cpu); 2871 } else 2872 raw_spin_unlock_irq(&rq->lock); 2873 2874 post_schedule(rq); 2875 2876 sched_preempt_enable_no_resched(); 2877 if (need_resched()) 2878 goto need_resched; 2879} 2880 2881static inline void sched_submit_work(struct task_struct *tsk) 2882{ 2883 if (!tsk->state || tsk_is_pi_blocked(tsk)) 2884 return; 2885 /* 2886 * If we are going to sleep and we have plugged IO queued, 2887 * make sure to submit it to avoid deadlocks. 2888 */ 2889 if (blk_needs_flush_plug(tsk)) 2890 blk_schedule_flush_plug(tsk); 2891} 2892 2893asmlinkage __visible void __sched schedule(void) 2894{ 2895 struct task_struct *tsk = current; 2896 2897 sched_submit_work(tsk); 2898 __schedule(); 2899} 2900EXPORT_SYMBOL(schedule); 2901 2902#ifdef CONFIG_CONTEXT_TRACKING 2903asmlinkage __visible void __sched schedule_user(void) 2904{ 2905 /* 2906 * If we come here after a random call to set_need_resched(), 2907 * or we have been woken up remotely but the IPI has not yet arrived, 2908 * we haven't yet exited the RCU idle mode. Do it here manually until 2909 * we find a better solution. 2910 */ 2911 user_exit(); 2912 schedule(); 2913 user_enter(); 2914} 2915#endif 2916 2917/** 2918 * schedule_preempt_disabled - called with preemption disabled 2919 * 2920 * Returns with preemption disabled. Note: preempt_count must be 1 2921 */ 2922void __sched schedule_preempt_disabled(void) 2923{ 2924 sched_preempt_enable_no_resched(); 2925 schedule(); 2926 preempt_disable(); 2927} 2928 2929#ifdef CONFIG_PREEMPT 2930/* 2931 * this is the entry point to schedule() from in-kernel preemption 2932 * off of preempt_enable. Kernel preemptions off return from interrupt 2933 * occur there and call schedule directly. 2934 */ 2935asmlinkage __visible void __sched notrace preempt_schedule(void) 2936{ 2937 /* 2938 * If there is a non-zero preempt_count or interrupts are disabled, 2939 * we do not want to preempt the current task. Just return.. 2940 */ 2941 if (likely(!preemptible())) 2942 return; 2943 2944 do { 2945 __preempt_count_add(PREEMPT_ACTIVE); 2946 __schedule(); 2947 __preempt_count_sub(PREEMPT_ACTIVE); 2948 2949 /* 2950 * Check again in case we missed a preemption opportunity 2951 * between schedule and now. 2952 */ 2953 barrier(); 2954 } while (need_resched()); 2955} 2956NOKPROBE_SYMBOL(preempt_schedule); 2957EXPORT_SYMBOL(preempt_schedule); 2958#endif /* CONFIG_PREEMPT */ 2959 2960/* 2961 * this is the entry point to schedule() from kernel preemption 2962 * off of irq context. 2963 * Note, that this is called and return with irqs disabled. This will 2964 * protect us against recursive calling from irq. 2965 */ 2966asmlinkage __visible void __sched preempt_schedule_irq(void) 2967{ 2968 enum ctx_state prev_state; 2969 2970 /* Catch callers which need to be fixed */ 2971 BUG_ON(preempt_count() || !irqs_disabled()); 2972 2973 prev_state = exception_enter(); 2974 2975 do { 2976 __preempt_count_add(PREEMPT_ACTIVE); 2977 local_irq_enable(); 2978 __schedule(); 2979 local_irq_disable(); 2980 __preempt_count_sub(PREEMPT_ACTIVE); 2981 2982 /* 2983 * Check again in case we missed a preemption opportunity 2984 * between schedule and now. 2985 */ 2986 barrier(); 2987 } while (need_resched()); 2988 2989 exception_exit(prev_state); 2990} 2991 2992int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags, 2993 void *key) 2994{ 2995 return try_to_wake_up(curr->private, mode, wake_flags); 2996} 2997EXPORT_SYMBOL(default_wake_function); 2998 2999#ifdef CONFIG_RT_MUTEXES 3000 3001/* 3002 * rt_mutex_setprio - set the current priority of a task 3003 * @p: task 3004 * @prio: prio value (kernel-internal form) 3005 * 3006 * This function changes the 'effective' priority of a task. It does 3007 * not touch ->normal_prio like __setscheduler(). 3008 * 3009 * Used by the rt_mutex code to implement priority inheritance 3010 * logic. Call site only calls if the priority of the task changed. 3011 */ 3012void rt_mutex_setprio(struct task_struct *p, int prio) 3013{ 3014 int oldprio, queued, running, enqueue_flag = 0; 3015 struct rq *rq; 3016 const struct sched_class *prev_class; 3017 3018 BUG_ON(prio > MAX_PRIO); 3019 3020 rq = __task_rq_lock(p); 3021 3022 /* 3023 * Idle task boosting is a nono in general. There is one 3024 * exception, when PREEMPT_RT and NOHZ is active: 3025 * 3026 * The idle task calls get_next_timer_interrupt() and holds 3027 * the timer wheel base->lock on the CPU and another CPU wants 3028 * to access the timer (probably to cancel it). We can safely 3029 * ignore the boosting request, as the idle CPU runs this code 3030 * with interrupts disabled and will complete the lock 3031 * protected section without being interrupted. So there is no 3032 * real need to boost. 3033 */ 3034 if (unlikely(p == rq->idle)) { 3035 WARN_ON(p != rq->curr); 3036 WARN_ON(p->pi_blocked_on); 3037 goto out_unlock; 3038 } 3039 3040 trace_sched_pi_setprio(p, prio); 3041 oldprio = p->prio; 3042 prev_class = p->sched_class; 3043 queued = task_on_rq_queued(p); 3044 running = task_current(rq, p); 3045 if (queued) 3046 dequeue_task(rq, p, 0); 3047 if (running) 3048 put_prev_task(rq, p); 3049 3050 /* 3051 * Boosting condition are: 3052 * 1. -rt task is running and holds mutex A 3053 * --> -dl task blocks on mutex A 3054 * 3055 * 2. -dl task is running and holds mutex A 3056 * --> -dl task blocks on mutex A and could preempt the 3057 * running task 3058 */ 3059 if (dl_prio(prio)) { 3060 struct task_struct *pi_task = rt_mutex_get_top_task(p); 3061 if (!dl_prio(p->normal_prio) || 3062 (pi_task && dl_entity_preempt(&pi_task->dl, &p->dl))) { 3063 p->dl.dl_boosted = 1; 3064 p->dl.dl_throttled = 0; 3065 enqueue_flag = ENQUEUE_REPLENISH; 3066 } else 3067 p->dl.dl_boosted = 0; 3068 p->sched_class = &dl_sched_class; 3069 } else if (rt_prio(prio)) { 3070 if (dl_prio(oldprio)) 3071 p->dl.dl_boosted = 0; 3072 if (oldprio < prio) 3073 enqueue_flag = ENQUEUE_HEAD; 3074 p->sched_class = &rt_sched_class; 3075 } else { 3076 if (dl_prio(oldprio)) 3077 p->dl.dl_boosted = 0; 3078 p->sched_class = &fair_sched_class; 3079 } 3080 3081 p->prio = prio; 3082 3083 if (running) 3084 p->sched_class->set_curr_task(rq); 3085 if (queued) 3086 enqueue_task(rq, p, enqueue_flag); 3087 3088 check_class_changed(rq, p, prev_class, oldprio); 3089out_unlock: 3090 __task_rq_unlock(rq); 3091} 3092#endif 3093 3094void set_user_nice(struct task_struct *p, long nice) 3095{ 3096 int old_prio, delta, queued; 3097 unsigned long flags; 3098 struct rq *rq; 3099 3100 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE) 3101 return; 3102 /* 3103 * We have to be careful, if called from sys_setpriority(), 3104 * the task might be in the middle of scheduling on another CPU. 3105 */ 3106 rq = task_rq_lock(p, &flags); 3107 /* 3108 * The RT priorities are set via sched_setscheduler(), but we still 3109 * allow the 'normal' nice value to be set - but as expected 3110 * it wont have any effect on scheduling until the task is 3111 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR: 3112 */ 3113 if (task_has_dl_policy(p) || task_has_rt_policy(p)) { 3114 p->static_prio = NICE_TO_PRIO(nice); 3115 goto out_unlock; 3116 } 3117 queued = task_on_rq_queued(p); 3118 if (queued) 3119 dequeue_task(rq, p, 0); 3120 3121 p->static_prio = NICE_TO_PRIO(nice); 3122 set_load_weight(p); 3123 old_prio = p->prio; 3124 p->prio = effective_prio(p); 3125 delta = p->prio - old_prio; 3126 3127 if (queued) { 3128 enqueue_task(rq, p, 0); 3129 /* 3130 * If the task increased its priority or is running and 3131 * lowered its priority, then reschedule its CPU: 3132 */ 3133 if (delta < 0 || (delta > 0 && task_running(rq, p))) 3134 resched_curr(rq); 3135 } 3136out_unlock: 3137 task_rq_unlock(rq, p, &flags); 3138} 3139EXPORT_SYMBOL(set_user_nice); 3140 3141/* 3142 * can_nice - check if a task can reduce its nice value 3143 * @p: task 3144 * @nice: nice value 3145 */ 3146int can_nice(const struct task_struct *p, const int nice) 3147{ 3148 /* convert nice value [19,-20] to rlimit style value [1,40] */ 3149 int nice_rlim = nice_to_rlimit(nice); 3150 3151 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) || 3152 capable(CAP_SYS_NICE)); 3153} 3154 3155#ifdef __ARCH_WANT_SYS_NICE 3156 3157/* 3158 * sys_nice - change the priority of the current process. 3159 * @increment: priority increment 3160 * 3161 * sys_setpriority is a more generic, but much slower function that 3162 * does similar things. 3163 */ 3164SYSCALL_DEFINE1(nice, int, increment) 3165{ 3166 long nice, retval; 3167 3168 /* 3169 * Setpriority might change our priority at the same moment. 3170 * We don't have to worry. Conceptually one call occurs first 3171 * and we have a single winner. 3172 */ 3173 increment = clamp(increment, -NICE_WIDTH, NICE_WIDTH); 3174 nice = task_nice(current) + increment; 3175 3176 nice = clamp_val(nice, MIN_NICE, MAX_NICE); 3177 if (increment < 0 && !can_nice(current, nice)) 3178 return -EPERM; 3179 3180 retval = security_task_setnice(current, nice); 3181 if (retval) 3182 return retval; 3183 3184 set_user_nice(current, nice); 3185 return 0; 3186} 3187 3188#endif 3189 3190/** 3191 * task_prio - return the priority value of a given task. 3192 * @p: the task in question. 3193 * 3194 * Return: The priority value as seen by users in /proc. 3195 * RT tasks are offset by -200. Normal tasks are centered 3196 * around 0, value goes from -16 to +15. 3197 */ 3198int task_prio(const struct task_struct *p) 3199{ 3200 return p->prio - MAX_RT_PRIO; 3201} 3202 3203/** 3204 * idle_cpu - is a given cpu idle currently? 3205 * @cpu: the processor in question. 3206 * 3207 * Return: 1 if the CPU is currently idle. 0 otherwise. 3208 */ 3209int idle_cpu(int cpu) 3210{ 3211 struct rq *rq = cpu_rq(cpu); 3212 3213 if (rq->curr != rq->idle) 3214 return 0; 3215 3216 if (rq->nr_running) 3217 return 0; 3218 3219#ifdef CONFIG_SMP 3220 if (!llist_empty(&rq->wake_list)) 3221 return 0; 3222#endif 3223 3224 return 1; 3225} 3226 3227/** 3228 * idle_task - return the idle task for a given cpu. 3229 * @cpu: the processor in question. 3230 * 3231 * Return: The idle task for the cpu @cpu. 3232 */ 3233struct task_struct *idle_task(int cpu) 3234{ 3235 return cpu_rq(cpu)->idle; 3236} 3237 3238/** 3239 * find_process_by_pid - find a process with a matching PID value. 3240 * @pid: the pid in question. 3241 * 3242 * The task of @pid, if found. %NULL otherwise. 3243 */ 3244static struct task_struct *find_process_by_pid(pid_t pid) 3245{ 3246 return pid ? find_task_by_vpid(pid) : current; 3247} 3248 3249/* 3250 * This function initializes the sched_dl_entity of a newly becoming 3251 * SCHED_DEADLINE task. 3252 * 3253 * Only the static values are considered here, the actual runtime and the 3254 * absolute deadline will be properly calculated when the task is enqueued 3255 * for the first time with its new policy. 3256 */ 3257static void 3258__setparam_dl(struct task_struct *p, const struct sched_attr *attr) 3259{ 3260 struct sched_dl_entity *dl_se = &p->dl; 3261 3262 init_dl_task_timer(dl_se); 3263 dl_se->dl_runtime = attr->sched_runtime; 3264 dl_se->dl_deadline = attr->sched_deadline; 3265 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline; 3266 dl_se->flags = attr->sched_flags; 3267 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime); 3268 dl_se->dl_throttled = 0; 3269 dl_se->dl_new = 1; 3270 dl_se->dl_yielded = 0; 3271} 3272 3273/* 3274 * sched_setparam() passes in -1 for its policy, to let the functions 3275 * it calls know not to change it. 3276 */ 3277#define SETPARAM_POLICY -1 3278 3279static void __setscheduler_params(struct task_struct *p, 3280 const struct sched_attr *attr) 3281{ 3282 int policy = attr->sched_policy; 3283 3284 if (policy == SETPARAM_POLICY) 3285 policy = p->policy; 3286 3287 p->policy = policy; 3288 3289 if (dl_policy(policy)) 3290 __setparam_dl(p, attr); 3291 else if (fair_policy(policy)) 3292 p->static_prio = NICE_TO_PRIO(attr->sched_nice); 3293 3294 /* 3295 * __sched_setscheduler() ensures attr->sched_priority == 0 when 3296 * !rt_policy. Always setting this ensures that things like 3297 * getparam()/getattr() don't report silly values for !rt tasks. 3298 */ 3299 p->rt_priority = attr->sched_priority; 3300 p->normal_prio = normal_prio(p); 3301 set_load_weight(p); 3302} 3303 3304/* Actually do priority change: must hold pi & rq lock. */ 3305static void __setscheduler(struct rq *rq, struct task_struct *p, 3306 const struct sched_attr *attr) 3307{ 3308 __setscheduler_params(p, attr); 3309 3310 /* 3311 * If we get here, there was no pi waiters boosting the 3312 * task. It is safe to use the normal prio. 3313 */ 3314 p->prio = normal_prio(p); 3315 3316 if (dl_prio(p->prio)) 3317 p->sched_class = &dl_sched_class; 3318 else if (rt_prio(p->prio)) 3319 p->sched_class = &rt_sched_class; 3320 else 3321 p->sched_class = &fair_sched_class; 3322} 3323 3324static void 3325__getparam_dl(struct task_struct *p, struct sched_attr *attr) 3326{ 3327 struct sched_dl_entity *dl_se = &p->dl; 3328 3329 attr->sched_priority = p->rt_priority; 3330 attr->sched_runtime = dl_se->dl_runtime; 3331 attr->sched_deadline = dl_se->dl_deadline; 3332 attr->sched_period = dl_se->dl_period; 3333 attr->sched_flags = dl_se->flags; 3334} 3335 3336/* 3337 * This function validates the new parameters of a -deadline task. 3338 * We ask for the deadline not being zero, and greater or equal 3339 * than the runtime, as well as the period of being zero or 3340 * greater than deadline. Furthermore, we have to be sure that 3341 * user parameters are above the internal resolution of 1us (we 3342 * check sched_runtime only since it is always the smaller one) and 3343 * below 2^63 ns (we have to check both sched_deadline and 3344 * sched_period, as the latter can be zero). 3345 */ 3346static bool 3347__checkparam_dl(const struct sched_attr *attr) 3348{ 3349 /* deadline != 0 */ 3350 if (attr->sched_deadline == 0) 3351 return false; 3352 3353 /* 3354 * Since we truncate DL_SCALE bits, make sure we're at least 3355 * that big. 3356 */ 3357 if (attr->sched_runtime < (1ULL << DL_SCALE)) 3358 return false; 3359 3360 /* 3361 * Since we use the MSB for wrap-around and sign issues, make 3362 * sure it's not set (mind that period can be equal to zero). 3363 */ 3364 if (attr->sched_deadline & (1ULL << 63) || 3365 attr->sched_period & (1ULL << 63)) 3366 return false; 3367 3368 /* runtime <= deadline <= period (if period != 0) */ 3369 if ((attr->sched_period != 0 && 3370 attr->sched_period < attr->sched_deadline) || 3371 attr->sched_deadline < attr->sched_runtime) 3372 return false; 3373 3374 return true; 3375} 3376 3377/* 3378 * check the target process has a UID that matches the current process's 3379 */ 3380static bool check_same_owner(struct task_struct *p) 3381{ 3382 const struct cred *cred = current_cred(), *pcred; 3383 bool match; 3384 3385 rcu_read_lock(); 3386 pcred = __task_cred(p); 3387 match = (uid_eq(cred->euid, pcred->euid) || 3388 uid_eq(cred->euid, pcred->uid)); 3389 rcu_read_unlock(); 3390 return match; 3391} 3392 3393static int __sched_setscheduler(struct task_struct *p, 3394 const struct sched_attr *attr, 3395 bool user) 3396{ 3397 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 : 3398 MAX_RT_PRIO - 1 - attr->sched_priority; 3399 int retval, oldprio, oldpolicy = -1, queued, running; 3400 int policy = attr->sched_policy; 3401 unsigned long flags; 3402 const struct sched_class *prev_class; 3403 struct rq *rq; 3404 int reset_on_fork; 3405 3406 /* may grab non-irq protected spin_locks */ 3407 BUG_ON(in_interrupt()); 3408recheck: 3409 /* double check policy once rq lock held */ 3410 if (policy < 0) { 3411 reset_on_fork = p->sched_reset_on_fork; 3412 policy = oldpolicy = p->policy; 3413 } else { 3414 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK); 3415 3416 if (policy != SCHED_DEADLINE && 3417 policy != SCHED_FIFO && policy != SCHED_RR && 3418 policy != SCHED_NORMAL && policy != SCHED_BATCH && 3419 policy != SCHED_IDLE) 3420 return -EINVAL; 3421 } 3422 3423 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK)) 3424 return -EINVAL; 3425 3426 /* 3427 * Valid priorities for SCHED_FIFO and SCHED_RR are 3428 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL, 3429 * SCHED_BATCH and SCHED_IDLE is 0. 3430 */ 3431 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) || 3432 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1)) 3433 return -EINVAL; 3434 if ((dl_policy(policy) && !__checkparam_dl(attr)) || 3435 (rt_policy(policy) != (attr->sched_priority != 0))) 3436 return -EINVAL; 3437 3438 /* 3439 * Allow unprivileged RT tasks to decrease priority: 3440 */ 3441 if (user && !capable(CAP_SYS_NICE)) { 3442 if (fair_policy(policy)) { 3443 if (attr->sched_nice < task_nice(p) && 3444 !can_nice(p, attr->sched_nice)) 3445 return -EPERM; 3446 } 3447 3448 if (rt_policy(policy)) { 3449 unsigned long rlim_rtprio = 3450 task_rlimit(p, RLIMIT_RTPRIO); 3451 3452 /* can't set/change the rt policy */ 3453 if (policy != p->policy && !rlim_rtprio) 3454 return -EPERM; 3455 3456 /* can't increase priority */ 3457 if (attr->sched_priority > p->rt_priority && 3458 attr->sched_priority > rlim_rtprio) 3459 return -EPERM; 3460 } 3461 3462 /* 3463 * Can't set/change SCHED_DEADLINE policy at all for now 3464 * (safest behavior); in the future we would like to allow 3465 * unprivileged DL tasks to increase their relative deadline 3466 * or reduce their runtime (both ways reducing utilization) 3467 */ 3468 if (dl_policy(policy)) 3469 return -EPERM; 3470 3471 /* 3472 * Treat SCHED_IDLE as nice 20. Only allow a switch to 3473 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it. 3474 */ 3475 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) { 3476 if (!can_nice(p, task_nice(p))) 3477 return -EPERM; 3478 } 3479 3480 /* can't change other user's priorities */ 3481 if (!check_same_owner(p)) 3482 return -EPERM; 3483 3484 /* Normal users shall not reset the sched_reset_on_fork flag */ 3485 if (p->sched_reset_on_fork && !reset_on_fork) 3486 return -EPERM; 3487 } 3488 3489 if (user) { 3490 retval = security_task_setscheduler(p); 3491 if (retval) 3492 return retval; 3493 } 3494 3495 /* 3496 * make sure no PI-waiters arrive (or leave) while we are 3497 * changing the priority of the task: 3498 * 3499 * To be able to change p->policy safely, the appropriate 3500 * runqueue lock must be held. 3501 */ 3502 rq = task_rq_lock(p, &flags); 3503 3504 /* 3505 * Changing the policy of the stop threads its a very bad idea 3506 */ 3507 if (p == rq->stop) { 3508 task_rq_unlock(rq, p, &flags); 3509 return -EINVAL; 3510 } 3511 3512 /* 3513 * If not changing anything there's no need to proceed further, 3514 * but store a possible modification of reset_on_fork. 3515 */ 3516 if (unlikely(policy == p->policy)) { 3517 if (fair_policy(policy) && attr->sched_nice != task_nice(p)) 3518 goto change; 3519 if (rt_policy(policy) && attr->sched_priority != p->rt_priority) 3520 goto change; 3521 if (dl_policy(policy)) 3522 goto change; 3523 3524 p->sched_reset_on_fork = reset_on_fork; 3525 task_rq_unlock(rq, p, &flags); 3526 return 0; 3527 } 3528change: 3529 3530 if (user) { 3531#ifdef CONFIG_RT_GROUP_SCHED 3532 /* 3533 * Do not allow realtime tasks into groups that have no runtime 3534 * assigned. 3535 */ 3536 if (rt_bandwidth_enabled() && rt_policy(policy) && 3537 task_group(p)->rt_bandwidth.rt_runtime == 0 && 3538 !task_group_is_autogroup(task_group(p))) { 3539 task_rq_unlock(rq, p, &flags); 3540 return -EPERM; 3541 } 3542#endif 3543#ifdef CONFIG_SMP 3544 if (dl_bandwidth_enabled() && dl_policy(policy)) { 3545 cpumask_t *span = rq->rd->span; 3546 3547 /* 3548 * Don't allow tasks with an affinity mask smaller than 3549 * the entire root_domain to become SCHED_DEADLINE. We 3550 * will also fail if there's no bandwidth available. 3551 */ 3552 if (!cpumask_subset(span, &p->cpus_allowed) || 3553 rq->rd->dl_bw.bw == 0) { 3554 task_rq_unlock(rq, p, &flags); 3555 return -EPERM; 3556 } 3557 } 3558#endif 3559 } 3560 3561 /* recheck policy now with rq lock held */ 3562 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) { 3563 policy = oldpolicy = -1; 3564 task_rq_unlock(rq, p, &flags); 3565 goto recheck; 3566 } 3567 3568 /* 3569 * If setscheduling to SCHED_DEADLINE (or changing the parameters 3570 * of a SCHED_DEADLINE task) we need to check if enough bandwidth 3571 * is available. 3572 */ 3573 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) { 3574 task_rq_unlock(rq, p, &flags); 3575 return -EBUSY; 3576 } 3577 3578 p->sched_reset_on_fork = reset_on_fork; 3579 oldprio = p->prio; 3580 3581 /* 3582 * Special case for priority boosted tasks. 3583 * 3584 * If the new priority is lower or equal (user space view) 3585 * than the current (boosted) priority, we just store the new 3586 * normal parameters and do not touch the scheduler class and 3587 * the runqueue. This will be done when the task deboost 3588 * itself. 3589 */ 3590 if (rt_mutex_check_prio(p, newprio)) { 3591 __setscheduler_params(p, attr); 3592 task_rq_unlock(rq, p, &flags); 3593 return 0; 3594 } 3595 3596 queued = task_on_rq_queued(p); 3597 running = task_current(rq, p); 3598 if (queued) 3599 dequeue_task(rq, p, 0); 3600 if (running) 3601 put_prev_task(rq, p); 3602 3603 prev_class = p->sched_class; 3604 __setscheduler(rq, p, attr); 3605 3606 if (running) 3607 p->sched_class->set_curr_task(rq); 3608 if (queued) { 3609 /* 3610 * We enqueue to tail when the priority of a task is 3611 * increased (user space view). 3612 */ 3613 enqueue_task(rq, p, oldprio <= p->prio ? ENQUEUE_HEAD : 0); 3614 } 3615 3616 check_class_changed(rq, p, prev_class, oldprio); 3617 task_rq_unlock(rq, p, &flags); 3618 3619 rt_mutex_adjust_pi(p); 3620 3621 return 0; 3622} 3623 3624static int _sched_setscheduler(struct task_struct *p, int policy, 3625 const struct sched_param *param, bool check) 3626{ 3627 struct sched_attr attr = { 3628 .sched_policy = policy, 3629 .sched_priority = param->sched_priority, 3630 .sched_nice = PRIO_TO_NICE(p->static_prio), 3631 }; 3632 3633 /* Fixup the legacy SCHED_RESET_ON_FORK hack. */ 3634 if ((policy != SETPARAM_POLICY) && (policy & SCHED_RESET_ON_FORK)) { 3635 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK; 3636 policy &= ~SCHED_RESET_ON_FORK; 3637 attr.sched_policy = policy; 3638 } 3639 3640 return __sched_setscheduler(p, &attr, check); 3641} 3642/** 3643 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread. 3644 * @p: the task in question. 3645 * @policy: new policy. 3646 * @param: structure containing the new RT priority. 3647 * 3648 * Return: 0 on success. An error code otherwise. 3649 * 3650 * NOTE that the task may be already dead. 3651 */ 3652int sched_setscheduler(struct task_struct *p, int policy, 3653 const struct sched_param *param) 3654{ 3655 return _sched_setscheduler(p, policy, param, true); 3656} 3657EXPORT_SYMBOL_GPL(sched_setscheduler); 3658 3659int sched_setattr(struct task_struct *p, const struct sched_attr *attr) 3660{ 3661 return __sched_setscheduler(p, attr, true); 3662} 3663EXPORT_SYMBOL_GPL(sched_setattr); 3664 3665/** 3666 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace. 3667 * @p: the task in question. 3668 * @policy: new policy. 3669 * @param: structure containing the new RT priority. 3670 * 3671 * Just like sched_setscheduler, only don't bother checking if the 3672 * current context has permission. For example, this is needed in 3673 * stop_machine(): we create temporary high priority worker threads, 3674 * but our caller might not have that capability. 3675 * 3676 * Return: 0 on success. An error code otherwise. 3677 */ 3678int sched_setscheduler_nocheck(struct task_struct *p, int policy, 3679 const struct sched_param *param) 3680{ 3681 return _sched_setscheduler(p, policy, param, false); 3682} 3683 3684static int 3685do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param) 3686{ 3687 struct sched_param lparam; 3688 struct task_struct *p; 3689 int retval; 3690 3691 if (!param || pid < 0) 3692 return -EINVAL; 3693 if (copy_from_user(&lparam, param, sizeof(struct sched_param))) 3694 return -EFAULT; 3695 3696 rcu_read_lock(); 3697 retval = -ESRCH; 3698 p = find_process_by_pid(pid); 3699 if (p != NULL) 3700 retval = sched_setscheduler(p, policy, &lparam); 3701 rcu_read_unlock(); 3702 3703 return retval; 3704} 3705 3706/* 3707 * Mimics kernel/events/core.c perf_copy_attr(). 3708 */ 3709static int sched_copy_attr(struct sched_attr __user *uattr, 3710 struct sched_attr *attr) 3711{ 3712 u32 size; 3713 int ret; 3714 3715 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0)) 3716 return -EFAULT; 3717 3718 /* 3719 * zero the full structure, so that a short copy will be nice. 3720 */ 3721 memset(attr, 0, sizeof(*attr)); 3722 3723 ret = get_user(size, &uattr->size); 3724 if (ret) 3725 return ret; 3726 3727 if (size > PAGE_SIZE) /* silly large */ 3728 goto err_size; 3729 3730 if (!size) /* abi compat */ 3731 size = SCHED_ATTR_SIZE_VER0; 3732 3733 if (size < SCHED_ATTR_SIZE_VER0) 3734 goto err_size; 3735 3736 /* 3737 * If we're handed a bigger struct than we know of, 3738 * ensure all the unknown bits are 0 - i.e. new 3739 * user-space does not rely on any kernel feature 3740 * extensions we dont know about yet. 3741 */ 3742 if (size > sizeof(*attr)) { 3743 unsigned char __user *addr; 3744 unsigned char __user *end; 3745 unsigned char val; 3746 3747 addr = (void __user *)uattr + sizeof(*attr); 3748 end = (void __user *)uattr + size; 3749 3750 for (; addr < end; addr++) { 3751 ret = get_user(val, addr); 3752 if (ret) 3753 return ret; 3754 if (val) 3755 goto err_size; 3756 } 3757 size = sizeof(*attr); 3758 } 3759 3760 ret = copy_from_user(attr, uattr, size); 3761 if (ret) 3762 return -EFAULT; 3763 3764 /* 3765 * XXX: do we want to be lenient like existing syscalls; or do we want 3766 * to be strict and return an error on out-of-bounds values? 3767 */ 3768 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE); 3769 3770 return 0; 3771 3772err_size: 3773 put_user(sizeof(*attr), &uattr->size); 3774 return -E2BIG; 3775} 3776 3777/** 3778 * sys_sched_setscheduler - set/change the scheduler policy and RT priority 3779 * @pid: the pid in question. 3780 * @policy: new policy. 3781 * @param: structure containing the new RT priority. 3782 * 3783 * Return: 0 on success. An error code otherwise. 3784 */ 3785SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, 3786 struct sched_param __user *, param) 3787{ 3788 /* negative values for policy are not valid */ 3789 if (policy < 0) 3790 return -EINVAL; 3791 3792 return do_sched_setscheduler(pid, policy, param); 3793} 3794 3795/** 3796 * sys_sched_setparam - set/change the RT priority of a thread 3797 * @pid: the pid in question. 3798 * @param: structure containing the new RT priority. 3799 * 3800 * Return: 0 on success. An error code otherwise. 3801 */ 3802SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param) 3803{ 3804 return do_sched_setscheduler(pid, SETPARAM_POLICY, param); 3805} 3806 3807/** 3808 * sys_sched_setattr - same as above, but with extended sched_attr 3809 * @pid: the pid in question. 3810 * @uattr: structure containing the extended parameters. 3811 * @flags: for future extension. 3812 */ 3813SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr, 3814 unsigned int, flags) 3815{ 3816 struct sched_attr attr; 3817 struct task_struct *p; 3818 int retval; 3819 3820 if (!uattr || pid < 0 || flags) 3821 return -EINVAL; 3822 3823 retval = sched_copy_attr(uattr, &attr); 3824 if (retval) 3825 return retval; 3826 3827 if ((int)attr.sched_policy < 0) 3828 return -EINVAL; 3829 3830 rcu_read_lock(); 3831 retval = -ESRCH; 3832 p = find_process_by_pid(pid); 3833 if (p != NULL) 3834 retval = sched_setattr(p, &attr); 3835 rcu_read_unlock(); 3836 3837 return retval; 3838} 3839 3840/** 3841 * sys_sched_getscheduler - get the policy (scheduling class) of a thread 3842 * @pid: the pid in question. 3843 * 3844 * Return: On success, the policy of the thread. Otherwise, a negative error 3845 * code. 3846 */ 3847SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid) 3848{ 3849 struct task_struct *p; 3850 int retval; 3851 3852 if (pid < 0) 3853 return -EINVAL; 3854 3855 retval = -ESRCH; 3856 rcu_read_lock(); 3857 p = find_process_by_pid(pid); 3858 if (p) { 3859 retval = security_task_getscheduler(p); 3860 if (!retval) 3861 retval = p->policy 3862 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0); 3863 } 3864 rcu_read_unlock(); 3865 return retval; 3866} 3867 3868/** 3869 * sys_sched_getparam - get the RT priority of a thread 3870 * @pid: the pid in question. 3871 * @param: structure containing the RT priority. 3872 * 3873 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error 3874 * code. 3875 */ 3876SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param) 3877{ 3878 struct sched_param lp = { .sched_priority = 0 }; 3879 struct task_struct *p; 3880 int retval; 3881 3882 if (!param || pid < 0) 3883 return -EINVAL; 3884 3885 rcu_read_lock(); 3886 p = find_process_by_pid(pid); 3887 retval = -ESRCH; 3888 if (!p) 3889 goto out_unlock; 3890 3891 retval = security_task_getscheduler(p); 3892 if (retval) 3893 goto out_unlock; 3894 3895 if (task_has_rt_policy(p)) 3896 lp.sched_priority = p->rt_priority; 3897 rcu_read_unlock(); 3898 3899 /* 3900 * This one might sleep, we cannot do it with a spinlock held ... 3901 */ 3902 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0; 3903 3904 return retval; 3905 3906out_unlock: 3907 rcu_read_unlock(); 3908 return retval; 3909} 3910 3911static int sched_read_attr(struct sched_attr __user *uattr, 3912 struct sched_attr *attr, 3913 unsigned int usize) 3914{ 3915 int ret; 3916 3917 if (!access_ok(VERIFY_WRITE, uattr, usize)) 3918 return -EFAULT; 3919 3920 /* 3921 * If we're handed a smaller struct than we know of, 3922 * ensure all the unknown bits are 0 - i.e. old 3923 * user-space does not get uncomplete information. 3924 */ 3925 if (usize < sizeof(*attr)) { 3926 unsigned char *addr; 3927 unsigned char *end; 3928 3929 addr = (void *)attr + usize; 3930 end = (void *)attr + sizeof(*attr); 3931 3932 for (; addr < end; addr++) { 3933 if (*addr) 3934 return -EFBIG; 3935 } 3936 3937 attr->size = usize; 3938 } 3939 3940 ret = copy_to_user(uattr, attr, attr->size); 3941 if (ret) 3942 return -EFAULT; 3943 3944 return 0; 3945} 3946 3947/** 3948 * sys_sched_getattr - similar to sched_getparam, but with sched_attr 3949 * @pid: the pid in question. 3950 * @uattr: structure containing the extended parameters. 3951 * @size: sizeof(attr) for fwd/bwd comp. 3952 * @flags: for future extension. 3953 */ 3954SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr, 3955 unsigned int, size, unsigned int, flags) 3956{ 3957 struct sched_attr attr = { 3958 .size = sizeof(struct sched_attr), 3959 }; 3960 struct task_struct *p; 3961 int retval; 3962 3963 if (!uattr || pid < 0 || size > PAGE_SIZE || 3964 size < SCHED_ATTR_SIZE_VER0 || flags) 3965 return -EINVAL; 3966 3967 rcu_read_lock(); 3968 p = find_process_by_pid(pid); 3969 retval = -ESRCH; 3970 if (!p) 3971 goto out_unlock; 3972 3973 retval = security_task_getscheduler(p); 3974 if (retval) 3975 goto out_unlock; 3976 3977 attr.sched_policy = p->policy; 3978 if (p->sched_reset_on_fork) 3979 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK; 3980 if (task_has_dl_policy(p)) 3981 __getparam_dl(p, &attr); 3982 else if (task_has_rt_policy(p)) 3983 attr.sched_priority = p->rt_priority; 3984 else 3985 attr.sched_nice = task_nice(p); 3986 3987 rcu_read_unlock(); 3988 3989 retval = sched_read_attr(uattr, &attr, size); 3990 return retval; 3991 3992out_unlock: 3993 rcu_read_unlock(); 3994 return retval; 3995} 3996 3997long sched_setaffinity(pid_t pid, const struct cpumask *in_mask) 3998{ 3999 cpumask_var_t cpus_allowed, new_mask; 4000 struct task_struct *p; 4001 int retval; 4002 4003 rcu_read_lock(); 4004 4005 p = find_process_by_pid(pid); 4006 if (!p) { 4007 rcu_read_unlock(); 4008 return -ESRCH; 4009 } 4010 4011 /* Prevent p going away */ 4012 get_task_struct(p); 4013 rcu_read_unlock(); 4014 4015 if (p->flags & PF_NO_SETAFFINITY) { 4016 retval = -EINVAL; 4017 goto out_put_task; 4018 } 4019 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) { 4020 retval = -ENOMEM; 4021 goto out_put_task; 4022 } 4023 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) { 4024 retval = -ENOMEM; 4025 goto out_free_cpus_allowed; 4026 } 4027 retval = -EPERM; 4028 if (!check_same_owner(p)) { 4029 rcu_read_lock(); 4030 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) { 4031 rcu_read_unlock(); 4032 goto out_free_new_mask; 4033 } 4034 rcu_read_unlock(); 4035 } 4036 4037 retval = security_task_setscheduler(p); 4038 if (retval) 4039 goto out_free_new_mask; 4040 4041 4042 cpuset_cpus_allowed(p, cpus_allowed); 4043 cpumask_and(new_mask, in_mask, cpus_allowed); 4044 4045 /* 4046 * Since bandwidth control happens on root_domain basis, 4047 * if admission test is enabled, we only admit -deadline 4048 * tasks allowed to run on all the CPUs in the task's 4049 * root_domain. 4050 */ 4051#ifdef CONFIG_SMP 4052 if (task_has_dl_policy(p) && dl_bandwidth_enabled()) { 4053 rcu_read_lock(); 4054 if (!cpumask_subset(task_rq(p)->rd->span, new_mask)) { 4055 retval = -EBUSY; 4056 rcu_read_unlock(); 4057 goto out_free_new_mask; 4058 } 4059 rcu_read_unlock(); 4060 } 4061#endif 4062again: 4063 retval = set_cpus_allowed_ptr(p, new_mask); 4064 4065 if (!retval) { 4066 cpuset_cpus_allowed(p, cpus_allowed); 4067 if (!cpumask_subset(new_mask, cpus_allowed)) { 4068 /* 4069 * We must have raced with a concurrent cpuset 4070 * update. Just reset the cpus_allowed to the 4071 * cpuset's cpus_allowed 4072 */ 4073 cpumask_copy(new_mask, cpus_allowed); 4074 goto again; 4075 } 4076 } 4077out_free_new_mask: 4078 free_cpumask_var(new_mask); 4079out_free_cpus_allowed: 4080 free_cpumask_var(cpus_allowed); 4081out_put_task: 4082 put_task_struct(p); 4083 return retval; 4084} 4085 4086static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len, 4087 struct cpumask *new_mask) 4088{ 4089 if (len < cpumask_size()) 4090 cpumask_clear(new_mask); 4091 else if (len > cpumask_size()) 4092 len = cpumask_size(); 4093 4094 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0; 4095} 4096 4097/** 4098 * sys_sched_setaffinity - set the cpu affinity of a process 4099 * @pid: pid of the process 4100 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 4101 * @user_mask_ptr: user-space pointer to the new cpu mask 4102 * 4103 * Return: 0 on success. An error code otherwise. 4104 */ 4105SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len, 4106 unsigned long __user *, user_mask_ptr) 4107{ 4108 cpumask_var_t new_mask; 4109 int retval; 4110 4111 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) 4112 return -ENOMEM; 4113 4114 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask); 4115 if (retval == 0) 4116 retval = sched_setaffinity(pid, new_mask); 4117 free_cpumask_var(new_mask); 4118 return retval; 4119} 4120 4121long sched_getaffinity(pid_t pid, struct cpumask *mask) 4122{ 4123 struct task_struct *p; 4124 unsigned long flags; 4125 int retval; 4126 4127 rcu_read_lock(); 4128 4129 retval = -ESRCH; 4130 p = find_process_by_pid(pid); 4131 if (!p) 4132 goto out_unlock; 4133 4134 retval = security_task_getscheduler(p); 4135 if (retval) 4136 goto out_unlock; 4137 4138 raw_spin_lock_irqsave(&p->pi_lock, flags); 4139 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask); 4140 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 4141 4142out_unlock: 4143 rcu_read_unlock(); 4144 4145 return retval; 4146} 4147 4148/** 4149 * sys_sched_getaffinity - get the cpu affinity of a process 4150 * @pid: pid of the process 4151 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 4152 * @user_mask_ptr: user-space pointer to hold the current cpu mask 4153 * 4154 * Return: 0 on success. An error code otherwise. 4155 */ 4156SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len, 4157 unsigned long __user *, user_mask_ptr) 4158{ 4159 int ret; 4160 cpumask_var_t mask; 4161 4162 if ((len * BITS_PER_BYTE) < nr_cpu_ids) 4163 return -EINVAL; 4164 if (len & (sizeof(unsigned long)-1)) 4165 return -EINVAL; 4166 4167 if (!alloc_cpumask_var(&mask, GFP_KERNEL)) 4168 return -ENOMEM; 4169 4170 ret = sched_getaffinity(pid, mask); 4171 if (ret == 0) { 4172 size_t retlen = min_t(size_t, len, cpumask_size()); 4173 4174 if (copy_to_user(user_mask_ptr, mask, retlen)) 4175 ret = -EFAULT; 4176 else 4177 ret = retlen; 4178 } 4179 free_cpumask_var(mask); 4180 4181 return ret; 4182} 4183 4184/** 4185 * sys_sched_yield - yield the current processor to other threads. 4186 * 4187 * This function yields the current CPU to other tasks. If there are no 4188 * other threads running on this CPU then this function will return. 4189 * 4190 * Return: 0. 4191 */ 4192SYSCALL_DEFINE0(sched_yield) 4193{ 4194 struct rq *rq = this_rq_lock(); 4195 4196 schedstat_inc(rq, yld_count); 4197 current->sched_class->yield_task(rq); 4198 4199 /* 4200 * Since we are going to call schedule() anyway, there's 4201 * no need to preempt or enable interrupts: 4202 */ 4203 __release(rq->lock); 4204 spin_release(&rq->lock.dep_map, 1, _THIS_IP_); 4205 do_raw_spin_unlock(&rq->lock); 4206 sched_preempt_enable_no_resched(); 4207 4208 schedule(); 4209 4210 return 0; 4211} 4212 4213static void __cond_resched(void) 4214{ 4215 __preempt_count_add(PREEMPT_ACTIVE); 4216 __schedule(); 4217 __preempt_count_sub(PREEMPT_ACTIVE); 4218} 4219 4220int __sched _cond_resched(void) 4221{ 4222 if (should_resched()) { 4223 __cond_resched(); 4224 return 1; 4225 } 4226 return 0; 4227} 4228EXPORT_SYMBOL(_cond_resched); 4229 4230/* 4231 * __cond_resched_lock() - if a reschedule is pending, drop the given lock, 4232 * call schedule, and on return reacquire the lock. 4233 * 4234 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level 4235 * operations here to prevent schedule() from being called twice (once via 4236 * spin_unlock(), once by hand). 4237 */ 4238int __cond_resched_lock(spinlock_t *lock) 4239{ 4240 int resched = should_resched(); 4241 int ret = 0; 4242 4243 lockdep_assert_held(lock); 4244 4245 if (spin_needbreak(lock) || resched) { 4246 spin_unlock(lock); 4247 if (resched) 4248 __cond_resched(); 4249 else 4250 cpu_relax(); 4251 ret = 1; 4252 spin_lock(lock); 4253 } 4254 return ret; 4255} 4256EXPORT_SYMBOL(__cond_resched_lock); 4257 4258int __sched __cond_resched_softirq(void) 4259{ 4260 BUG_ON(!in_softirq()); 4261 4262 if (should_resched()) { 4263 local_bh_enable(); 4264 __cond_resched(); 4265 local_bh_disable(); 4266 return 1; 4267 } 4268 return 0; 4269} 4270EXPORT_SYMBOL(__cond_resched_softirq); 4271 4272/** 4273 * yield - yield the current processor to other threads. 4274 * 4275 * Do not ever use this function, there's a 99% chance you're doing it wrong. 4276 * 4277 * The scheduler is at all times free to pick the calling task as the most 4278 * eligible task to run, if removing the yield() call from your code breaks 4279 * it, its already broken. 4280 * 4281 * Typical broken usage is: 4282 * 4283 * while (!event) 4284 * yield(); 4285 * 4286 * where one assumes that yield() will let 'the other' process run that will 4287 * make event true. If the current task is a SCHED_FIFO task that will never 4288 * happen. Never use yield() as a progress guarantee!! 4289 * 4290 * If you want to use yield() to wait for something, use wait_event(). 4291 * If you want to use yield() to be 'nice' for others, use cond_resched(). 4292 * If you still want to use yield(), do not! 4293 */ 4294void __sched yield(void) 4295{ 4296 set_current_state(TASK_RUNNING); 4297 sys_sched_yield(); 4298} 4299EXPORT_SYMBOL(yield); 4300 4301/** 4302 * yield_to - yield the current processor to another thread in 4303 * your thread group, or accelerate that thread toward the 4304 * processor it's on. 4305 * @p: target task 4306 * @preempt: whether task preemption is allowed or not 4307 * 4308 * It's the caller's job to ensure that the target task struct 4309 * can't go away on us before we can do any checks. 4310 * 4311 * Return: 4312 * true (>0) if we indeed boosted the target task. 4313 * false (0) if we failed to boost the target. 4314 * -ESRCH if there's no task to yield to. 4315 */ 4316int __sched yield_to(struct task_struct *p, bool preempt) 4317{ 4318 struct task_struct *curr = current; 4319 struct rq *rq, *p_rq; 4320 unsigned long flags; 4321 int yielded = 0; 4322 4323 local_irq_save(flags); 4324 rq = this_rq(); 4325 4326again: 4327 p_rq = task_rq(p); 4328 /* 4329 * If we're the only runnable task on the rq and target rq also 4330 * has only one task, there's absolutely no point in yielding. 4331 */ 4332 if (rq->nr_running == 1 && p_rq->nr_running == 1) { 4333 yielded = -ESRCH; 4334 goto out_irq; 4335 } 4336 4337 double_rq_lock(rq, p_rq); 4338 if (task_rq(p) != p_rq) { 4339 double_rq_unlock(rq, p_rq); 4340 goto again; 4341 } 4342 4343 if (!curr->sched_class->yield_to_task) 4344 goto out_unlock; 4345 4346 if (curr->sched_class != p->sched_class) 4347 goto out_unlock; 4348 4349 if (task_running(p_rq, p) || p->state) 4350 goto out_unlock; 4351 4352 yielded = curr->sched_class->yield_to_task(rq, p, preempt); 4353 if (yielded) { 4354 schedstat_inc(rq, yld_count); 4355 /* 4356 * Make p's CPU reschedule; pick_next_entity takes care of 4357 * fairness. 4358 */ 4359 if (preempt && rq != p_rq) 4360 resched_curr(p_rq); 4361 } 4362 4363out_unlock: 4364 double_rq_unlock(rq, p_rq); 4365out_irq: 4366 local_irq_restore(flags); 4367 4368 if (yielded > 0) 4369 schedule(); 4370 4371 return yielded; 4372} 4373EXPORT_SYMBOL_GPL(yield_to); 4374 4375/* 4376 * This task is about to go to sleep on IO. Increment rq->nr_iowait so 4377 * that process accounting knows that this is a task in IO wait state. 4378 */ 4379void __sched io_schedule(void) 4380{ 4381 struct rq *rq = raw_rq(); 4382 4383 delayacct_blkio_start(); 4384 atomic_inc(&rq->nr_iowait); 4385 blk_flush_plug(current); 4386 current->in_iowait = 1; 4387 schedule(); 4388 current->in_iowait = 0; 4389 atomic_dec(&rq->nr_iowait); 4390 delayacct_blkio_end(); 4391} 4392EXPORT_SYMBOL(io_schedule); 4393 4394long __sched io_schedule_timeout(long timeout) 4395{ 4396 struct rq *rq = raw_rq(); 4397 long ret; 4398 4399 delayacct_blkio_start(); 4400 atomic_inc(&rq->nr_iowait); 4401 blk_flush_plug(current); 4402 current->in_iowait = 1; 4403 ret = schedule_timeout(timeout); 4404 current->in_iowait = 0; 4405 atomic_dec(&rq->nr_iowait); 4406 delayacct_blkio_end(); 4407 return ret; 4408} 4409 4410/** 4411 * sys_sched_get_priority_max - return maximum RT priority. 4412 * @policy: scheduling class. 4413 * 4414 * Return: On success, this syscall returns the maximum 4415 * rt_priority that can be used by a given scheduling class. 4416 * On failure, a negative error code is returned. 4417 */ 4418SYSCALL_DEFINE1(sched_get_priority_max, int, policy) 4419{ 4420 int ret = -EINVAL; 4421 4422 switch (policy) { 4423 case SCHED_FIFO: 4424 case SCHED_RR: 4425 ret = MAX_USER_RT_PRIO-1; 4426 break; 4427 case SCHED_DEADLINE: 4428 case SCHED_NORMAL: 4429 case SCHED_BATCH: 4430 case SCHED_IDLE: 4431 ret = 0; 4432 break; 4433 } 4434 return ret; 4435} 4436 4437/** 4438 * sys_sched_get_priority_min - return minimum RT priority. 4439 * @policy: scheduling class. 4440 * 4441 * Return: On success, this syscall returns the minimum 4442 * rt_priority that can be used by a given scheduling class. 4443 * On failure, a negative error code is returned. 4444 */ 4445SYSCALL_DEFINE1(sched_get_priority_min, int, policy) 4446{ 4447 int ret = -EINVAL; 4448 4449 switch (policy) { 4450 case SCHED_FIFO: 4451 case SCHED_RR: 4452 ret = 1; 4453 break; 4454 case SCHED_DEADLINE: 4455 case SCHED_NORMAL: 4456 case SCHED_BATCH: 4457 case SCHED_IDLE: 4458 ret = 0; 4459 } 4460 return ret; 4461} 4462 4463/** 4464 * sys_sched_rr_get_interval - return the default timeslice of a process. 4465 * @pid: pid of the process. 4466 * @interval: userspace pointer to the timeslice value. 4467 * 4468 * this syscall writes the default timeslice value of a given process 4469 * into the user-space timespec buffer. A value of '0' means infinity. 4470 * 4471 * Return: On success, 0 and the timeslice is in @interval. Otherwise, 4472 * an error code. 4473 */ 4474SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid, 4475 struct timespec __user *, interval) 4476{ 4477 struct task_struct *p; 4478 unsigned int time_slice; 4479 unsigned long flags; 4480 struct rq *rq; 4481 int retval; 4482 struct timespec t; 4483 4484 if (pid < 0) 4485 return -EINVAL; 4486 4487 retval = -ESRCH; 4488 rcu_read_lock(); 4489 p = find_process_by_pid(pid); 4490 if (!p) 4491 goto out_unlock; 4492 4493 retval = security_task_getscheduler(p); 4494 if (retval) 4495 goto out_unlock; 4496 4497 rq = task_rq_lock(p, &flags); 4498 time_slice = 0; 4499 if (p->sched_class->get_rr_interval) 4500 time_slice = p->sched_class->get_rr_interval(rq, p); 4501 task_rq_unlock(rq, p, &flags); 4502 4503 rcu_read_unlock(); 4504 jiffies_to_timespec(time_slice, &t); 4505 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0; 4506 return retval; 4507 4508out_unlock: 4509 rcu_read_unlock(); 4510 return retval; 4511} 4512 4513static const char stat_nam[] = TASK_STATE_TO_CHAR_STR; 4514 4515void sched_show_task(struct task_struct *p) 4516{ 4517 unsigned long free = 0; 4518 int ppid; 4519 unsigned state; 4520 4521 state = p->state ? __ffs(p->state) + 1 : 0; 4522 printk(KERN_INFO "%-15.15s %c", p->comm, 4523 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?'); 4524#if BITS_PER_LONG == 32 4525 if (state == TASK_RUNNING) 4526 printk(KERN_CONT " running "); 4527 else 4528 printk(KERN_CONT " %08lx ", thread_saved_pc(p)); 4529#else 4530 if (state == TASK_RUNNING) 4531 printk(KERN_CONT " running task "); 4532 else 4533 printk(KERN_CONT " %016lx ", thread_saved_pc(p)); 4534#endif 4535#ifdef CONFIG_DEBUG_STACK_USAGE 4536 free = stack_not_used(p); 4537#endif 4538 rcu_read_lock(); 4539 ppid = task_pid_nr(rcu_dereference(p->real_parent)); 4540 rcu_read_unlock(); 4541 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free, 4542 task_pid_nr(p), ppid, 4543 (unsigned long)task_thread_info(p)->flags); 4544 4545 print_worker_info(KERN_INFO, p); 4546 show_stack(p, NULL); 4547} 4548 4549void show_state_filter(unsigned long state_filter) 4550{ 4551 struct task_struct *g, *p; 4552 4553#if BITS_PER_LONG == 32 4554 printk(KERN_INFO 4555 " task PC stack pid father\n"); 4556#else 4557 printk(KERN_INFO 4558 " task PC stack pid father\n"); 4559#endif 4560 rcu_read_lock(); 4561 for_each_process_thread(g, p) { 4562 /* 4563 * reset the NMI-timeout, listing all files on a slow 4564 * console might take a lot of time: 4565 */ 4566 touch_nmi_watchdog(); 4567 if (!state_filter || (p->state & state_filter)) 4568 sched_show_task(p); 4569 } 4570 4571 touch_all_softlockup_watchdogs(); 4572 4573#ifdef CONFIG_SCHED_DEBUG 4574 sysrq_sched_debug_show(); 4575#endif 4576 rcu_read_unlock(); 4577 /* 4578 * Only show locks if all tasks are dumped: 4579 */ 4580 if (!state_filter) 4581 debug_show_all_locks(); 4582} 4583 4584void init_idle_bootup_task(struct task_struct *idle) 4585{ 4586 idle->sched_class = &idle_sched_class; 4587} 4588 4589/** 4590 * init_idle - set up an idle thread for a given CPU 4591 * @idle: task in question 4592 * @cpu: cpu the idle task belongs to 4593 * 4594 * NOTE: this function does not set the idle thread's NEED_RESCHED 4595 * flag, to make booting more robust. 4596 */ 4597void init_idle(struct task_struct *idle, int cpu) 4598{ 4599 struct rq *rq = cpu_rq(cpu); 4600 unsigned long flags; 4601 4602 raw_spin_lock_irqsave(&rq->lock, flags); 4603 4604 __sched_fork(0, idle); 4605 idle->state = TASK_RUNNING; 4606 idle->se.exec_start = sched_clock(); 4607 4608 do_set_cpus_allowed(idle, cpumask_of(cpu)); 4609 /* 4610 * We're having a chicken and egg problem, even though we are 4611 * holding rq->lock, the cpu isn't yet set to this cpu so the 4612 * lockdep check in task_group() will fail. 4613 * 4614 * Similar case to sched_fork(). / Alternatively we could 4615 * use task_rq_lock() here and obtain the other rq->lock. 4616 * 4617 * Silence PROVE_RCU 4618 */ 4619 rcu_read_lock(); 4620 __set_task_cpu(idle, cpu); 4621 rcu_read_unlock(); 4622 4623 rq->curr = rq->idle = idle; 4624 idle->on_rq = TASK_ON_RQ_QUEUED; 4625#if defined(CONFIG_SMP) 4626 idle->on_cpu = 1; 4627#endif 4628 raw_spin_unlock_irqrestore(&rq->lock, flags); 4629 4630 /* Set the preempt count _outside_ the spinlocks! */ 4631 init_idle_preempt_count(idle, cpu); 4632 4633 /* 4634 * The idle tasks have their own, simple scheduling class: 4635 */ 4636 idle->sched_class = &idle_sched_class; 4637 ftrace_graph_init_idle_task(idle, cpu); 4638 vtime_init_idle(idle, cpu); 4639#if defined(CONFIG_SMP) 4640 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu); 4641#endif 4642} 4643 4644#ifdef CONFIG_SMP 4645/* 4646 * move_queued_task - move a queued task to new rq. 4647 * 4648 * Returns (locked) new rq. Old rq's lock is released. 4649 */ 4650static struct rq *move_queued_task(struct task_struct *p, int new_cpu) 4651{ 4652 struct rq *rq = task_rq(p); 4653 4654 lockdep_assert_held(&rq->lock); 4655 4656 dequeue_task(rq, p, 0); 4657 p->on_rq = TASK_ON_RQ_MIGRATING; 4658 set_task_cpu(p, new_cpu); 4659 raw_spin_unlock(&rq->lock); 4660 4661 rq = cpu_rq(new_cpu); 4662 4663 raw_spin_lock(&rq->lock); 4664 BUG_ON(task_cpu(p) != new_cpu); 4665 p->on_rq = TASK_ON_RQ_QUEUED; 4666 enqueue_task(rq, p, 0); 4667 check_preempt_curr(rq, p, 0); 4668 4669 return rq; 4670} 4671 4672void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask) 4673{ 4674 if (p->sched_class && p->sched_class->set_cpus_allowed) 4675 p->sched_class->set_cpus_allowed(p, new_mask); 4676 4677 cpumask_copy(&p->cpus_allowed, new_mask); 4678 p->nr_cpus_allowed = cpumask_weight(new_mask); 4679} 4680 4681/* 4682 * This is how migration works: 4683 * 4684 * 1) we invoke migration_cpu_stop() on the target CPU using 4685 * stop_one_cpu(). 4686 * 2) stopper starts to run (implicitly forcing the migrated thread 4687 * off the CPU) 4688 * 3) it checks whether the migrated task is still in the wrong runqueue. 4689 * 4) if it's in the wrong runqueue then the migration thread removes 4690 * it and puts it into the right queue. 4691 * 5) stopper completes and stop_one_cpu() returns and the migration 4692 * is done. 4693 */ 4694 4695/* 4696 * Change a given task's CPU affinity. Migrate the thread to a 4697 * proper CPU and schedule it away if the CPU it's executing on 4698 * is removed from the allowed bitmask. 4699 * 4700 * NOTE: the caller must have a valid reference to the task, the 4701 * task must not exit() & deallocate itself prematurely. The 4702 * call is not atomic; no spinlocks may be held. 4703 */ 4704int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask) 4705{ 4706 unsigned long flags; 4707 struct rq *rq; 4708 unsigned int dest_cpu; 4709 int ret = 0; 4710 4711 rq = task_rq_lock(p, &flags); 4712 4713 if (cpumask_equal(&p->cpus_allowed, new_mask)) 4714 goto out; 4715 4716 if (!cpumask_intersects(new_mask, cpu_active_mask)) { 4717 ret = -EINVAL; 4718 goto out; 4719 } 4720 4721 do_set_cpus_allowed(p, new_mask); 4722 4723 /* Can the task run on the task's current CPU? If so, we're done */ 4724 if (cpumask_test_cpu(task_cpu(p), new_mask)) 4725 goto out; 4726 4727 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask); 4728 if (task_running(rq, p) || p->state == TASK_WAKING) { 4729 struct migration_arg arg = { p, dest_cpu }; 4730 /* Need help from migration thread: drop lock and wait. */ 4731 task_rq_unlock(rq, p, &flags); 4732 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg); 4733 tlb_migrate_finish(p->mm); 4734 return 0; 4735 } else if (task_on_rq_queued(p)) 4736 rq = move_queued_task(p, dest_cpu); 4737out: 4738 task_rq_unlock(rq, p, &flags); 4739 4740 return ret; 4741} 4742EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr); 4743 4744/* 4745 * Move (not current) task off this cpu, onto dest cpu. We're doing 4746 * this because either it can't run here any more (set_cpus_allowed() 4747 * away from this CPU, or CPU going down), or because we're 4748 * attempting to rebalance this task on exec (sched_exec). 4749 * 4750 * So we race with normal scheduler movements, but that's OK, as long 4751 * as the task is no longer on this CPU. 4752 * 4753 * Returns non-zero if task was successfully migrated. 4754 */ 4755static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu) 4756{ 4757 struct rq *rq; 4758 int ret = 0; 4759 4760 if (unlikely(!cpu_active(dest_cpu))) 4761 return ret; 4762 4763 rq = cpu_rq(src_cpu); 4764 4765 raw_spin_lock(&p->pi_lock); 4766 raw_spin_lock(&rq->lock); 4767 /* Already moved. */ 4768 if (task_cpu(p) != src_cpu) 4769 goto done; 4770 4771 /* Affinity changed (again). */ 4772 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) 4773 goto fail; 4774 4775 /* 4776 * If we're not on a rq, the next wake-up will ensure we're 4777 * placed properly. 4778 */ 4779 if (task_on_rq_queued(p)) 4780 rq = move_queued_task(p, dest_cpu); 4781done: 4782 ret = 1; 4783fail: 4784 raw_spin_unlock(&rq->lock); 4785 raw_spin_unlock(&p->pi_lock); 4786 return ret; 4787} 4788 4789#ifdef CONFIG_NUMA_BALANCING 4790/* Migrate current task p to target_cpu */ 4791int migrate_task_to(struct task_struct *p, int target_cpu) 4792{ 4793 struct migration_arg arg = { p, target_cpu }; 4794 int curr_cpu = task_cpu(p); 4795 4796 if (curr_cpu == target_cpu) 4797 return 0; 4798 4799 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p))) 4800 return -EINVAL; 4801 4802 /* TODO: This is not properly updating schedstats */ 4803 4804 trace_sched_move_numa(p, curr_cpu, target_cpu); 4805 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg); 4806} 4807 4808/* 4809 * Requeue a task on a given node and accurately track the number of NUMA 4810 * tasks on the runqueues 4811 */ 4812void sched_setnuma(struct task_struct *p, int nid) 4813{ 4814 struct rq *rq; 4815 unsigned long flags; 4816 bool queued, running; 4817 4818 rq = task_rq_lock(p, &flags); 4819 queued = task_on_rq_queued(p); 4820 running = task_current(rq, p); 4821 4822 if (queued) 4823 dequeue_task(rq, p, 0); 4824 if (running) 4825 put_prev_task(rq, p); 4826 4827 p->numa_preferred_nid = nid; 4828 4829 if (running) 4830 p->sched_class->set_curr_task(rq); 4831 if (queued) 4832 enqueue_task(rq, p, 0); 4833 task_rq_unlock(rq, p, &flags); 4834} 4835#endif 4836 4837/* 4838 * migration_cpu_stop - this will be executed by a highprio stopper thread 4839 * and performs thread migration by bumping thread off CPU then 4840 * 'pushing' onto another runqueue. 4841 */ 4842static int migration_cpu_stop(void *data) 4843{ 4844 struct migration_arg *arg = data; 4845 4846 /* 4847 * The original target cpu might have gone down and we might 4848 * be on another cpu but it doesn't matter. 4849 */ 4850 local_irq_disable(); 4851 /* 4852 * We need to explicitly wake pending tasks before running 4853 * __migrate_task() such that we will not miss enforcing cpus_allowed 4854 * during wakeups, see set_cpus_allowed_ptr()'s TASK_WAKING test. 4855 */ 4856 sched_ttwu_pending(); 4857 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu); 4858 local_irq_enable(); 4859 return 0; 4860} 4861 4862#ifdef CONFIG_HOTPLUG_CPU 4863 4864/* 4865 * Ensures that the idle task is using init_mm right before its cpu goes 4866 * offline. 4867 */ 4868void idle_task_exit(void) 4869{ 4870 struct mm_struct *mm = current->active_mm; 4871 4872 BUG_ON(cpu_online(smp_processor_id())); 4873 4874 if (mm != &init_mm) { 4875 switch_mm(mm, &init_mm, current); 4876 finish_arch_post_lock_switch(); 4877 } 4878 mmdrop(mm); 4879} 4880 4881/* 4882 * Since this CPU is going 'away' for a while, fold any nr_active delta 4883 * we might have. Assumes we're called after migrate_tasks() so that the 4884 * nr_active count is stable. 4885 * 4886 * Also see the comment "Global load-average calculations". 4887 */ 4888static void calc_load_migrate(struct rq *rq) 4889{ 4890 long delta = calc_load_fold_active(rq); 4891 if (delta) 4892 atomic_long_add(delta, &calc_load_tasks); 4893} 4894 4895static void put_prev_task_fake(struct rq *rq, struct task_struct *prev) 4896{ 4897} 4898 4899static const struct sched_class fake_sched_class = { 4900 .put_prev_task = put_prev_task_fake, 4901}; 4902 4903static struct task_struct fake_task = { 4904 /* 4905 * Avoid pull_{rt,dl}_task() 4906 */ 4907 .prio = MAX_PRIO + 1, 4908 .sched_class = &fake_sched_class, 4909}; 4910 4911/* 4912 * Migrate all tasks from the rq, sleeping tasks will be migrated by 4913 * try_to_wake_up()->select_task_rq(). 4914 * 4915 * Called with rq->lock held even though we'er in stop_machine() and 4916 * there's no concurrency possible, we hold the required locks anyway 4917 * because of lock validation efforts. 4918 */ 4919static void migrate_tasks(unsigned int dead_cpu) 4920{ 4921 struct rq *rq = cpu_rq(dead_cpu); 4922 struct task_struct *next, *stop = rq->stop; 4923 int dest_cpu; 4924 4925 /* 4926 * Fudge the rq selection such that the below task selection loop 4927 * doesn't get stuck on the currently eligible stop task. 4928 * 4929 * We're currently inside stop_machine() and the rq is either stuck 4930 * in the stop_machine_cpu_stop() loop, or we're executing this code, 4931 * either way we should never end up calling schedule() until we're 4932 * done here. 4933 */ 4934 rq->stop = NULL; 4935 4936 /* 4937 * put_prev_task() and pick_next_task() sched 4938 * class method both need to have an up-to-date 4939 * value of rq->clock[_task] 4940 */ 4941 update_rq_clock(rq); 4942 4943 for ( ; ; ) { 4944 /* 4945 * There's this thread running, bail when that's the only 4946 * remaining thread. 4947 */ 4948 if (rq->nr_running == 1) 4949 break; 4950 4951 next = pick_next_task(rq, &fake_task); 4952 BUG_ON(!next); 4953 next->sched_class->put_prev_task(rq, next); 4954 4955 /* Find suitable destination for @next, with force if needed. */ 4956 dest_cpu = select_fallback_rq(dead_cpu, next); 4957 raw_spin_unlock(&rq->lock); 4958 4959 __migrate_task(next, dead_cpu, dest_cpu); 4960 4961 raw_spin_lock(&rq->lock); 4962 } 4963 4964 rq->stop = stop; 4965} 4966 4967#endif /* CONFIG_HOTPLUG_CPU */ 4968 4969#if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL) 4970 4971static struct ctl_table sd_ctl_dir[] = { 4972 { 4973 .procname = "sched_domain", 4974 .mode = 0555, 4975 }, 4976 {} 4977}; 4978 4979static struct ctl_table sd_ctl_root[] = { 4980 { 4981 .procname = "kernel", 4982 .mode = 0555, 4983 .child = sd_ctl_dir, 4984 }, 4985 {} 4986}; 4987 4988static struct ctl_table *sd_alloc_ctl_entry(int n) 4989{ 4990 struct ctl_table *entry = 4991 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL); 4992 4993 return entry; 4994} 4995 4996static void sd_free_ctl_entry(struct ctl_table **tablep) 4997{ 4998 struct ctl_table *entry; 4999 5000 /* 5001 * In the intermediate directories, both the child directory and 5002 * procname are dynamically allocated and could fail but the mode 5003 * will always be set. In the lowest directory the names are 5004 * static strings and all have proc handlers. 5005 */ 5006 for (entry = *tablep; entry->mode; entry++) { 5007 if (entry->child) 5008 sd_free_ctl_entry(&entry->child); 5009 if (entry->proc_handler == NULL) 5010 kfree(entry->procname); 5011 } 5012 5013 kfree(*tablep); 5014 *tablep = NULL; 5015} 5016 5017static int min_load_idx = 0; 5018static int max_load_idx = CPU_LOAD_IDX_MAX-1; 5019 5020static void 5021set_table_entry(struct ctl_table *entry, 5022 const char *procname, void *data, int maxlen, 5023 umode_t mode, proc_handler *proc_handler, 5024 bool load_idx) 5025{ 5026 entry->procname = procname; 5027 entry->data = data; 5028 entry->maxlen = maxlen; 5029 entry->mode = mode; 5030 entry->proc_handler = proc_handler; 5031 5032 if (load_idx) { 5033 entry->extra1 = &min_load_idx; 5034 entry->extra2 = &max_load_idx; 5035 } 5036} 5037 5038static struct ctl_table * 5039sd_alloc_ctl_domain_table(struct sched_domain *sd) 5040{ 5041 struct ctl_table *table = sd_alloc_ctl_entry(14); 5042 5043 if (table == NULL) 5044 return NULL; 5045 5046 set_table_entry(&table[0], "min_interval", &sd->min_interval, 5047 sizeof(long), 0644, proc_doulongvec_minmax, false); 5048 set_table_entry(&table[1], "max_interval", &sd->max_interval, 5049 sizeof(long), 0644, proc_doulongvec_minmax, false); 5050 set_table_entry(&table[2], "busy_idx", &sd->busy_idx, 5051 sizeof(int), 0644, proc_dointvec_minmax, true); 5052 set_table_entry(&table[3], "idle_idx", &sd->idle_idx, 5053 sizeof(int), 0644, proc_dointvec_minmax, true); 5054 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx, 5055 sizeof(int), 0644, proc_dointvec_minmax, true); 5056 set_table_entry(&table[5], "wake_idx", &sd->wake_idx, 5057 sizeof(int), 0644, proc_dointvec_minmax, true); 5058 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx, 5059 sizeof(int), 0644, proc_dointvec_minmax, true); 5060 set_table_entry(&table[7], "busy_factor", &sd->busy_factor, 5061 sizeof(int), 0644, proc_dointvec_minmax, false); 5062 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct, 5063 sizeof(int), 0644, proc_dointvec_minmax, false); 5064 set_table_entry(&table[9], "cache_nice_tries", 5065 &sd->cache_nice_tries, 5066 sizeof(int), 0644, proc_dointvec_minmax, false); 5067 set_table_entry(&table[10], "flags", &sd->flags, 5068 sizeof(int), 0644, proc_dointvec_minmax, false); 5069 set_table_entry(&table[11], "max_newidle_lb_cost", 5070 &sd->max_newidle_lb_cost, 5071 sizeof(long), 0644, proc_doulongvec_minmax, false); 5072 set_table_entry(&table[12], "name", sd->name, 5073 CORENAME_MAX_SIZE, 0444, proc_dostring, false); 5074 /* &table[13] is terminator */ 5075 5076 return table; 5077} 5078 5079static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu) 5080{ 5081 struct ctl_table *entry, *table; 5082 struct sched_domain *sd; 5083 int domain_num = 0, i; 5084 char buf[32]; 5085 5086 for_each_domain(cpu, sd) 5087 domain_num++; 5088 entry = table = sd_alloc_ctl_entry(domain_num + 1); 5089 if (table == NULL) 5090 return NULL; 5091 5092 i = 0; 5093 for_each_domain(cpu, sd) { 5094 snprintf(buf, 32, "domain%d", i); 5095 entry->procname = kstrdup(buf, GFP_KERNEL); 5096 entry->mode = 0555; 5097 entry->child = sd_alloc_ctl_domain_table(sd); 5098 entry++; 5099 i++; 5100 } 5101 return table; 5102} 5103 5104static struct ctl_table_header *sd_sysctl_header; 5105static void register_sched_domain_sysctl(void) 5106{ 5107 int i, cpu_num = num_possible_cpus(); 5108 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1); 5109 char buf[32]; 5110 5111 WARN_ON(sd_ctl_dir[0].child); 5112 sd_ctl_dir[0].child = entry; 5113 5114 if (entry == NULL) 5115 return; 5116 5117 for_each_possible_cpu(i) { 5118 snprintf(buf, 32, "cpu%d", i); 5119 entry->procname = kstrdup(buf, GFP_KERNEL); 5120 entry->mode = 0555; 5121 entry->child = sd_alloc_ctl_cpu_table(i); 5122 entry++; 5123 } 5124 5125 WARN_ON(sd_sysctl_header); 5126 sd_sysctl_header = register_sysctl_table(sd_ctl_root); 5127} 5128 5129/* may be called multiple times per register */ 5130static void unregister_sched_domain_sysctl(void) 5131{ 5132 if (sd_sysctl_header) 5133 unregister_sysctl_table(sd_sysctl_header); 5134 sd_sysctl_header = NULL; 5135 if (sd_ctl_dir[0].child) 5136 sd_free_ctl_entry(&sd_ctl_dir[0].child); 5137} 5138#else 5139static void register_sched_domain_sysctl(void) 5140{ 5141} 5142static void unregister_sched_domain_sysctl(void) 5143{ 5144} 5145#endif 5146 5147static void set_rq_online(struct rq *rq) 5148{ 5149 if (!rq->online) { 5150 const struct sched_class *class; 5151 5152 cpumask_set_cpu(rq->cpu, rq->rd->online); 5153 rq->online = 1; 5154 5155 for_each_class(class) { 5156 if (class->rq_online) 5157 class->rq_online(rq); 5158 } 5159 } 5160} 5161 5162static void set_rq_offline(struct rq *rq) 5163{ 5164 if (rq->online) { 5165 const struct sched_class *class; 5166 5167 for_each_class(class) { 5168 if (class->rq_offline) 5169 class->rq_offline(rq); 5170 } 5171 5172 cpumask_clear_cpu(rq->cpu, rq->rd->online); 5173 rq->online = 0; 5174 } 5175} 5176 5177/* 5178 * migration_call - callback that gets triggered when a CPU is added. 5179 * Here we can start up the necessary migration thread for the new CPU. 5180 */ 5181static int 5182migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu) 5183{ 5184 int cpu = (long)hcpu; 5185 unsigned long flags; 5186 struct rq *rq = cpu_rq(cpu); 5187 5188 switch (action & ~CPU_TASKS_FROZEN) { 5189 5190 case CPU_UP_PREPARE: 5191 rq->calc_load_update = calc_load_update; 5192 break; 5193 5194 case CPU_ONLINE: 5195 /* Update our root-domain */ 5196 raw_spin_lock_irqsave(&rq->lock, flags); 5197 if (rq->rd) { 5198 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 5199 5200 set_rq_online(rq); 5201 } 5202 raw_spin_unlock_irqrestore(&rq->lock, flags); 5203 break; 5204 5205#ifdef CONFIG_HOTPLUG_CPU 5206 case CPU_DYING: 5207 sched_ttwu_pending(); 5208 /* Update our root-domain */ 5209 raw_spin_lock_irqsave(&rq->lock, flags); 5210 if (rq->rd) { 5211 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 5212 set_rq_offline(rq); 5213 } 5214 migrate_tasks(cpu); 5215 BUG_ON(rq->nr_running != 1); /* the migration thread */ 5216 raw_spin_unlock_irqrestore(&rq->lock, flags); 5217 break; 5218 5219 case CPU_DEAD: 5220 calc_load_migrate(rq); 5221 break; 5222#endif 5223 } 5224 5225 update_max_interval(); 5226 5227 return NOTIFY_OK; 5228} 5229 5230/* 5231 * Register at high priority so that task migration (migrate_all_tasks) 5232 * happens before everything else. This has to be lower priority than 5233 * the notifier in the perf_event subsystem, though. 5234 */ 5235static struct notifier_block migration_notifier = { 5236 .notifier_call = migration_call, 5237 .priority = CPU_PRI_MIGRATION, 5238}; 5239 5240static void __cpuinit set_cpu_rq_start_time(void) 5241{ 5242 int cpu = smp_processor_id(); 5243 struct rq *rq = cpu_rq(cpu); 5244 rq->age_stamp = sched_clock_cpu(cpu); 5245} 5246 5247static int sched_cpu_active(struct notifier_block *nfb, 5248 unsigned long action, void *hcpu) 5249{ 5250 switch (action & ~CPU_TASKS_FROZEN) { 5251 case CPU_STARTING: 5252 set_cpu_rq_start_time(); 5253 return NOTIFY_OK; 5254 case CPU_DOWN_FAILED: 5255 set_cpu_active((long)hcpu, true); 5256 return NOTIFY_OK; 5257 default: 5258 return NOTIFY_DONE; 5259 } 5260} 5261 5262static int sched_cpu_inactive(struct notifier_block *nfb, 5263 unsigned long action, void *hcpu) 5264{ 5265 unsigned long flags; 5266 long cpu = (long)hcpu; 5267 struct dl_bw *dl_b; 5268 5269 switch (action & ~CPU_TASKS_FROZEN) { 5270 case CPU_DOWN_PREPARE: 5271 set_cpu_active(cpu, false); 5272 5273 /* explicitly allow suspend */ 5274 if (!(action & CPU_TASKS_FROZEN)) { 5275 bool overflow; 5276 int cpus; 5277 5278 rcu_read_lock_sched(); 5279 dl_b = dl_bw_of(cpu); 5280 5281 raw_spin_lock_irqsave(&dl_b->lock, flags); 5282 cpus = dl_bw_cpus(cpu); 5283 overflow = __dl_overflow(dl_b, cpus, 0, 0); 5284 raw_spin_unlock_irqrestore(&dl_b->lock, flags); 5285 5286 rcu_read_unlock_sched(); 5287 5288 if (overflow) 5289 return notifier_from_errno(-EBUSY); 5290 } 5291 return NOTIFY_OK; 5292 } 5293 5294 return NOTIFY_DONE; 5295} 5296 5297static int __init migration_init(void) 5298{ 5299 void *cpu = (void *)(long)smp_processor_id(); 5300 int err; 5301 5302 /* Initialize migration for the boot CPU */ 5303 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu); 5304 BUG_ON(err == NOTIFY_BAD); 5305 migration_call(&migration_notifier, CPU_ONLINE, cpu); 5306 register_cpu_notifier(&migration_notifier); 5307 5308 /* Register cpu active notifiers */ 5309 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE); 5310 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE); 5311 5312 return 0; 5313} 5314early_initcall(migration_init); 5315#endif 5316 5317#ifdef CONFIG_SMP 5318 5319static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */ 5320 5321#ifdef CONFIG_SCHED_DEBUG 5322 5323static __read_mostly int sched_debug_enabled; 5324 5325static int __init sched_debug_setup(char *str) 5326{ 5327 sched_debug_enabled = 1; 5328 5329 return 0; 5330} 5331early_param("sched_debug", sched_debug_setup); 5332 5333static inline bool sched_debug(void) 5334{ 5335 return sched_debug_enabled; 5336} 5337 5338static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level, 5339 struct cpumask *groupmask) 5340{ 5341 struct sched_group *group = sd->groups; 5342 char str[256]; 5343 5344 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd)); 5345 cpumask_clear(groupmask); 5346 5347 printk(KERN_DEBUG "%*s domain %d: ", level, "", level); 5348 5349 if (!(sd->flags & SD_LOAD_BALANCE)) { 5350 printk("does not load-balance\n"); 5351 if (sd->parent) 5352 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain" 5353 " has parent"); 5354 return -1; 5355 } 5356 5357 printk(KERN_CONT "span %s level %s\n", str, sd->name); 5358 5359 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) { 5360 printk(KERN_ERR "ERROR: domain->span does not contain " 5361 "CPU%d\n", cpu); 5362 } 5363 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) { 5364 printk(KERN_ERR "ERROR: domain->groups does not contain" 5365 " CPU%d\n", cpu); 5366 } 5367 5368 printk(KERN_DEBUG "%*s groups:", level + 1, ""); 5369 do { 5370 if (!group) { 5371 printk("\n"); 5372 printk(KERN_ERR "ERROR: group is NULL\n"); 5373 break; 5374 } 5375 5376 /* 5377 * Even though we initialize ->capacity to something semi-sane, 5378 * we leave capacity_orig unset. This allows us to detect if 5379 * domain iteration is still funny without causing /0 traps. 5380 */ 5381 if (!group->sgc->capacity_orig) { 5382 printk(KERN_CONT "\n"); 5383 printk(KERN_ERR "ERROR: domain->cpu_capacity not set\n"); 5384 break; 5385 } 5386 5387 if (!cpumask_weight(sched_group_cpus(group))) { 5388 printk(KERN_CONT "\n"); 5389 printk(KERN_ERR "ERROR: empty group\n"); 5390 break; 5391 } 5392 5393 if (!(sd->flags & SD_OVERLAP) && 5394 cpumask_intersects(groupmask, sched_group_cpus(group))) { 5395 printk(KERN_CONT "\n"); 5396 printk(KERN_ERR "ERROR: repeated CPUs\n"); 5397 break; 5398 } 5399 5400 cpumask_or(groupmask, groupmask, sched_group_cpus(group)); 5401 5402 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group)); 5403 5404 printk(KERN_CONT " %s", str); 5405 if (group->sgc->capacity != SCHED_CAPACITY_SCALE) { 5406 printk(KERN_CONT " (cpu_capacity = %d)", 5407 group->sgc->capacity); 5408 } 5409 5410 group = group->next; 5411 } while (group != sd->groups); 5412 printk(KERN_CONT "\n"); 5413 5414 if (!cpumask_equal(sched_domain_span(sd), groupmask)) 5415 printk(KERN_ERR "ERROR: groups don't span domain->span\n"); 5416 5417 if (sd->parent && 5418 !cpumask_subset(groupmask, sched_domain_span(sd->parent))) 5419 printk(KERN_ERR "ERROR: parent span is not a superset " 5420 "of domain->span\n"); 5421 return 0; 5422} 5423 5424static void sched_domain_debug(struct sched_domain *sd, int cpu) 5425{ 5426 int level = 0; 5427 5428 if (!sched_debug_enabled) 5429 return; 5430 5431 if (!sd) { 5432 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu); 5433 return; 5434 } 5435 5436 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu); 5437 5438 for (;;) { 5439 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask)) 5440 break; 5441 level++; 5442 sd = sd->parent; 5443 if (!sd) 5444 break; 5445 } 5446} 5447#else /* !CONFIG_SCHED_DEBUG */ 5448# define sched_domain_debug(sd, cpu) do { } while (0) 5449static inline bool sched_debug(void) 5450{ 5451 return false; 5452} 5453#endif /* CONFIG_SCHED_DEBUG */ 5454 5455static int sd_degenerate(struct sched_domain *sd) 5456{ 5457 if (cpumask_weight(sched_domain_span(sd)) == 1) 5458 return 1; 5459 5460 /* Following flags need at least 2 groups */ 5461 if (sd->flags & (SD_LOAD_BALANCE | 5462 SD_BALANCE_NEWIDLE | 5463 SD_BALANCE_FORK | 5464 SD_BALANCE_EXEC | 5465 SD_SHARE_CPUCAPACITY | 5466 SD_SHARE_PKG_RESOURCES | 5467 SD_SHARE_POWERDOMAIN)) { 5468 if (sd->groups != sd->groups->next) 5469 return 0; 5470 } 5471 5472 /* Following flags don't use groups */ 5473 if (sd->flags & (SD_WAKE_AFFINE)) 5474 return 0; 5475 5476 return 1; 5477} 5478 5479static int 5480sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent) 5481{ 5482 unsigned long cflags = sd->flags, pflags = parent->flags; 5483 5484 if (sd_degenerate(parent)) 5485 return 1; 5486 5487 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent))) 5488 return 0; 5489 5490 /* Flags needing groups don't count if only 1 group in parent */ 5491 if (parent->groups == parent->groups->next) { 5492 pflags &= ~(SD_LOAD_BALANCE | 5493 SD_BALANCE_NEWIDLE | 5494 SD_BALANCE_FORK | 5495 SD_BALANCE_EXEC | 5496 SD_SHARE_CPUCAPACITY | 5497 SD_SHARE_PKG_RESOURCES | 5498 SD_PREFER_SIBLING | 5499 SD_SHARE_POWERDOMAIN); 5500 if (nr_node_ids == 1) 5501 pflags &= ~SD_SERIALIZE; 5502 } 5503 if (~cflags & pflags) 5504 return 0; 5505 5506 return 1; 5507} 5508 5509static void free_rootdomain(struct rcu_head *rcu) 5510{ 5511 struct root_domain *rd = container_of(rcu, struct root_domain, rcu); 5512 5513 cpupri_cleanup(&rd->cpupri); 5514 cpudl_cleanup(&rd->cpudl); 5515 free_cpumask_var(rd->dlo_mask); 5516 free_cpumask_var(rd->rto_mask); 5517 free_cpumask_var(rd->online); 5518 free_cpumask_var(rd->span); 5519 kfree(rd); 5520} 5521 5522static void rq_attach_root(struct rq *rq, struct root_domain *rd) 5523{ 5524 struct root_domain *old_rd = NULL; 5525 unsigned long flags; 5526 5527 raw_spin_lock_irqsave(&rq->lock, flags); 5528 5529 if (rq->rd) { 5530 old_rd = rq->rd; 5531 5532 if (cpumask_test_cpu(rq->cpu, old_rd->online)) 5533 set_rq_offline(rq); 5534 5535 cpumask_clear_cpu(rq->cpu, old_rd->span); 5536 5537 /* 5538 * If we dont want to free the old_rd yet then 5539 * set old_rd to NULL to skip the freeing later 5540 * in this function: 5541 */ 5542 if (!atomic_dec_and_test(&old_rd->refcount)) 5543 old_rd = NULL; 5544 } 5545 5546 atomic_inc(&rd->refcount); 5547 rq->rd = rd; 5548 5549 cpumask_set_cpu(rq->cpu, rd->span); 5550 if (cpumask_test_cpu(rq->cpu, cpu_active_mask)) 5551 set_rq_online(rq); 5552 5553 raw_spin_unlock_irqrestore(&rq->lock, flags); 5554 5555 if (old_rd) 5556 call_rcu_sched(&old_rd->rcu, free_rootdomain); 5557} 5558 5559static int init_rootdomain(struct root_domain *rd) 5560{ 5561 memset(rd, 0, sizeof(*rd)); 5562 5563 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL)) 5564 goto out; 5565 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL)) 5566 goto free_span; 5567 if (!alloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL)) 5568 goto free_online; 5569 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL)) 5570 goto free_dlo_mask; 5571 5572 init_dl_bw(&rd->dl_bw); 5573 if (cpudl_init(&rd->cpudl) != 0) 5574 goto free_dlo_mask; 5575 5576 if (cpupri_init(&rd->cpupri) != 0) 5577 goto free_rto_mask; 5578 return 0; 5579 5580free_rto_mask: 5581 free_cpumask_var(rd->rto_mask); 5582free_dlo_mask: 5583 free_cpumask_var(rd->dlo_mask); 5584free_online: 5585 free_cpumask_var(rd->online); 5586free_span: 5587 free_cpumask_var(rd->span); 5588out: 5589 return -ENOMEM; 5590} 5591 5592/* 5593 * By default the system creates a single root-domain with all cpus as 5594 * members (mimicking the global state we have today). 5595 */ 5596struct root_domain def_root_domain; 5597 5598static void init_defrootdomain(void) 5599{ 5600 init_rootdomain(&def_root_domain); 5601 5602 atomic_set(&def_root_domain.refcount, 1); 5603} 5604 5605static struct root_domain *alloc_rootdomain(void) 5606{ 5607 struct root_domain *rd; 5608 5609 rd = kmalloc(sizeof(*rd), GFP_KERNEL); 5610 if (!rd) 5611 return NULL; 5612 5613 if (init_rootdomain(rd) != 0) { 5614 kfree(rd); 5615 return NULL; 5616 } 5617 5618 return rd; 5619} 5620 5621static void free_sched_groups(struct sched_group *sg, int free_sgc) 5622{ 5623 struct sched_group *tmp, *first; 5624 5625 if (!sg) 5626 return; 5627 5628 first = sg; 5629 do { 5630 tmp = sg->next; 5631 5632 if (free_sgc && atomic_dec_and_test(&sg->sgc->ref)) 5633 kfree(sg->sgc); 5634 5635 kfree(sg); 5636 sg = tmp; 5637 } while (sg != first); 5638} 5639 5640static void free_sched_domain(struct rcu_head *rcu) 5641{ 5642 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu); 5643 5644 /* 5645 * If its an overlapping domain it has private groups, iterate and 5646 * nuke them all. 5647 */ 5648 if (sd->flags & SD_OVERLAP) { 5649 free_sched_groups(sd->groups, 1); 5650 } else if (atomic_dec_and_test(&sd->groups->ref)) { 5651 kfree(sd->groups->sgc); 5652 kfree(sd->groups); 5653 } 5654 kfree(sd); 5655} 5656 5657static void destroy_sched_domain(struct sched_domain *sd, int cpu) 5658{ 5659 call_rcu(&sd->rcu, free_sched_domain); 5660} 5661 5662static void destroy_sched_domains(struct sched_domain *sd, int cpu) 5663{ 5664 for (; sd; sd = sd->parent) 5665 destroy_sched_domain(sd, cpu); 5666} 5667 5668/* 5669 * Keep a special pointer to the highest sched_domain that has 5670 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this 5671 * allows us to avoid some pointer chasing select_idle_sibling(). 5672 * 5673 * Also keep a unique ID per domain (we use the first cpu number in 5674 * the cpumask of the domain), this allows us to quickly tell if 5675 * two cpus are in the same cache domain, see cpus_share_cache(). 5676 */ 5677DEFINE_PER_CPU(struct sched_domain *, sd_llc); 5678DEFINE_PER_CPU(int, sd_llc_size); 5679DEFINE_PER_CPU(int, sd_llc_id); 5680DEFINE_PER_CPU(struct sched_domain *, sd_numa); 5681DEFINE_PER_CPU(struct sched_domain *, sd_busy); 5682DEFINE_PER_CPU(struct sched_domain *, sd_asym); 5683 5684static void update_top_cache_domain(int cpu) 5685{ 5686 struct sched_domain *sd; 5687 struct sched_domain *busy_sd = NULL; 5688 int id = cpu; 5689 int size = 1; 5690 5691 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES); 5692 if (sd) { 5693 id = cpumask_first(sched_domain_span(sd)); 5694 size = cpumask_weight(sched_domain_span(sd)); 5695 busy_sd = sd->parent; /* sd_busy */ 5696 } 5697 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd); 5698 5699 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd); 5700 per_cpu(sd_llc_size, cpu) = size; 5701 per_cpu(sd_llc_id, cpu) = id; 5702 5703 sd = lowest_flag_domain(cpu, SD_NUMA); 5704 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd); 5705 5706 sd = highest_flag_domain(cpu, SD_ASYM_PACKING); 5707 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd); 5708} 5709 5710/* 5711 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must 5712 * hold the hotplug lock. 5713 */ 5714static void 5715cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu) 5716{ 5717 struct rq *rq = cpu_rq(cpu); 5718 struct sched_domain *tmp; 5719 5720 /* Remove the sched domains which do not contribute to scheduling. */ 5721 for (tmp = sd; tmp; ) { 5722 struct sched_domain *parent = tmp->parent; 5723 if (!parent) 5724 break; 5725 5726 if (sd_parent_degenerate(tmp, parent)) { 5727 tmp->parent = parent->parent; 5728 if (parent->parent) 5729 parent->parent->child = tmp; 5730 /* 5731 * Transfer SD_PREFER_SIBLING down in case of a 5732 * degenerate parent; the spans match for this 5733 * so the property transfers. 5734 */ 5735 if (parent->flags & SD_PREFER_SIBLING) 5736 tmp->flags |= SD_PREFER_SIBLING; 5737 destroy_sched_domain(parent, cpu); 5738 } else 5739 tmp = tmp->parent; 5740 } 5741 5742 if (sd && sd_degenerate(sd)) { 5743 tmp = sd; 5744 sd = sd->parent; 5745 destroy_sched_domain(tmp, cpu); 5746 if (sd) 5747 sd->child = NULL; 5748 } 5749 5750 sched_domain_debug(sd, cpu); 5751 5752 rq_attach_root(rq, rd); 5753 tmp = rq->sd; 5754 rcu_assign_pointer(rq->sd, sd); 5755 destroy_sched_domains(tmp, cpu); 5756 5757 update_top_cache_domain(cpu); 5758} 5759 5760/* cpus with isolated domains */ 5761static cpumask_var_t cpu_isolated_map; 5762 5763/* Setup the mask of cpus configured for isolated domains */ 5764static int __init isolated_cpu_setup(char *str) 5765{ 5766 alloc_bootmem_cpumask_var(&cpu_isolated_map); 5767 cpulist_parse(str, cpu_isolated_map); 5768 return 1; 5769} 5770 5771__setup("isolcpus=", isolated_cpu_setup); 5772 5773struct s_data { 5774 struct sched_domain ** __percpu sd; 5775 struct root_domain *rd; 5776}; 5777 5778enum s_alloc { 5779 sa_rootdomain, 5780 sa_sd, 5781 sa_sd_storage, 5782 sa_none, 5783}; 5784 5785/* 5786 * Build an iteration mask that can exclude certain CPUs from the upwards 5787 * domain traversal. 5788 * 5789 * Asymmetric node setups can result in situations where the domain tree is of 5790 * unequal depth, make sure to skip domains that already cover the entire 5791 * range. 5792 * 5793 * In that case build_sched_domains() will have terminated the iteration early 5794 * and our sibling sd spans will be empty. Domains should always include the 5795 * cpu they're built on, so check that. 5796 * 5797 */ 5798static void build_group_mask(struct sched_domain *sd, struct sched_group *sg) 5799{ 5800 const struct cpumask *span = sched_domain_span(sd); 5801 struct sd_data *sdd = sd->private; 5802 struct sched_domain *sibling; 5803 int i; 5804 5805 for_each_cpu(i, span) { 5806 sibling = *per_cpu_ptr(sdd->sd, i); 5807 if (!cpumask_test_cpu(i, sched_domain_span(sibling))) 5808 continue; 5809 5810 cpumask_set_cpu(i, sched_group_mask(sg)); 5811 } 5812} 5813 5814/* 5815 * Return the canonical balance cpu for this group, this is the first cpu 5816 * of this group that's also in the iteration mask. 5817 */ 5818int group_balance_cpu(struct sched_group *sg) 5819{ 5820 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg)); 5821} 5822 5823static int 5824build_overlap_sched_groups(struct sched_domain *sd, int cpu) 5825{ 5826 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg; 5827 const struct cpumask *span = sched_domain_span(sd); 5828 struct cpumask *covered = sched_domains_tmpmask; 5829 struct sd_data *sdd = sd->private; 5830 struct sched_domain *sibling; 5831 int i; 5832 5833 cpumask_clear(covered); 5834 5835 for_each_cpu(i, span) { 5836 struct cpumask *sg_span; 5837 5838 if (cpumask_test_cpu(i, covered)) 5839 continue; 5840 5841 sibling = *per_cpu_ptr(sdd->sd, i); 5842 5843 /* See the comment near build_group_mask(). */ 5844 if (!cpumask_test_cpu(i, sched_domain_span(sibling))) 5845 continue; 5846 5847 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 5848 GFP_KERNEL, cpu_to_node(cpu)); 5849 5850 if (!sg) 5851 goto fail; 5852 5853 sg_span = sched_group_cpus(sg); 5854 if (sibling->child) 5855 cpumask_copy(sg_span, sched_domain_span(sibling->child)); 5856 else 5857 cpumask_set_cpu(i, sg_span); 5858 5859 cpumask_or(covered, covered, sg_span); 5860 5861 sg->sgc = *per_cpu_ptr(sdd->sgc, i); 5862 if (atomic_inc_return(&sg->sgc->ref) == 1) 5863 build_group_mask(sd, sg); 5864 5865 /* 5866 * Initialize sgc->capacity such that even if we mess up the 5867 * domains and no possible iteration will get us here, we won't 5868 * die on a /0 trap. 5869 */ 5870 sg->sgc->capacity = SCHED_CAPACITY_SCALE * cpumask_weight(sg_span); 5871 sg->sgc->capacity_orig = sg->sgc->capacity; 5872 5873 /* 5874 * Make sure the first group of this domain contains the 5875 * canonical balance cpu. Otherwise the sched_domain iteration 5876 * breaks. See update_sg_lb_stats(). 5877 */ 5878 if ((!groups && cpumask_test_cpu(cpu, sg_span)) || 5879 group_balance_cpu(sg) == cpu) 5880 groups = sg; 5881 5882 if (!first) 5883 first = sg; 5884 if (last) 5885 last->next = sg; 5886 last = sg; 5887 last->next = first; 5888 } 5889 sd->groups = groups; 5890 5891 return 0; 5892 5893fail: 5894 free_sched_groups(first, 0); 5895 5896 return -ENOMEM; 5897} 5898 5899static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg) 5900{ 5901 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu); 5902 struct sched_domain *child = sd->child; 5903 5904 if (child) 5905 cpu = cpumask_first(sched_domain_span(child)); 5906 5907 if (sg) { 5908 *sg = *per_cpu_ptr(sdd->sg, cpu); 5909 (*sg)->sgc = *per_cpu_ptr(sdd->sgc, cpu); 5910 atomic_set(&(*sg)->sgc->ref, 1); /* for claim_allocations */ 5911 } 5912 5913 return cpu; 5914} 5915 5916/* 5917 * build_sched_groups will build a circular linked list of the groups 5918 * covered by the given span, and will set each group's ->cpumask correctly, 5919 * and ->cpu_capacity to 0. 5920 * 5921 * Assumes the sched_domain tree is fully constructed 5922 */ 5923static int 5924build_sched_groups(struct sched_domain *sd, int cpu) 5925{ 5926 struct sched_group *first = NULL, *last = NULL; 5927 struct sd_data *sdd = sd->private; 5928 const struct cpumask *span = sched_domain_span(sd); 5929 struct cpumask *covered; 5930 int i; 5931 5932 get_group(cpu, sdd, &sd->groups); 5933 atomic_inc(&sd->groups->ref); 5934 5935 if (cpu != cpumask_first(span)) 5936 return 0; 5937 5938 lockdep_assert_held(&sched_domains_mutex); 5939 covered = sched_domains_tmpmask; 5940 5941 cpumask_clear(covered); 5942 5943 for_each_cpu(i, span) { 5944 struct sched_group *sg; 5945 int group, j; 5946 5947 if (cpumask_test_cpu(i, covered)) 5948 continue; 5949 5950 group = get_group(i, sdd, &sg); 5951 cpumask_setall(sched_group_mask(sg)); 5952 5953 for_each_cpu(j, span) { 5954 if (get_group(j, sdd, NULL) != group) 5955 continue; 5956 5957 cpumask_set_cpu(j, covered); 5958 cpumask_set_cpu(j, sched_group_cpus(sg)); 5959 } 5960 5961 if (!first) 5962 first = sg; 5963 if (last) 5964 last->next = sg; 5965 last = sg; 5966 } 5967 last->next = first; 5968 5969 return 0; 5970} 5971 5972/* 5973 * Initialize sched groups cpu_capacity. 5974 * 5975 * cpu_capacity indicates the capacity of sched group, which is used while 5976 * distributing the load between different sched groups in a sched domain. 5977 * Typically cpu_capacity for all the groups in a sched domain will be same 5978 * unless there are asymmetries in the topology. If there are asymmetries, 5979 * group having more cpu_capacity will pickup more load compared to the 5980 * group having less cpu_capacity. 5981 */ 5982static void init_sched_groups_capacity(int cpu, struct sched_domain *sd) 5983{ 5984 struct sched_group *sg = sd->groups; 5985 5986 WARN_ON(!sg); 5987 5988 do { 5989 sg->group_weight = cpumask_weight(sched_group_cpus(sg)); 5990 sg = sg->next; 5991 } while (sg != sd->groups); 5992 5993 if (cpu != group_balance_cpu(sg)) 5994 return; 5995 5996 update_group_capacity(sd, cpu); 5997 atomic_set(&sg->sgc->nr_busy_cpus, sg->group_weight); 5998} 5999 6000/* 6001 * Initializers for schedule domains 6002 * Non-inlined to reduce accumulated stack pressure in build_sched_domains() 6003 */ 6004 6005static int default_relax_domain_level = -1; 6006int sched_domain_level_max; 6007 6008static int __init setup_relax_domain_level(char *str) 6009{ 6010 if (kstrtoint(str, 0, &default_relax_domain_level)) 6011 pr_warn("Unable to set relax_domain_level\n"); 6012 6013 return 1; 6014} 6015__setup("relax_domain_level=", setup_relax_domain_level); 6016 6017static void set_domain_attribute(struct sched_domain *sd, 6018 struct sched_domain_attr *attr) 6019{ 6020 int request; 6021 6022 if (!attr || attr->relax_domain_level < 0) { 6023 if (default_relax_domain_level < 0) 6024 return; 6025 else 6026 request = default_relax_domain_level; 6027 } else 6028 request = attr->relax_domain_level; 6029 if (request < sd->level) { 6030 /* turn off idle balance on this domain */ 6031 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 6032 } else { 6033 /* turn on idle balance on this domain */ 6034 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 6035 } 6036} 6037 6038static void __sdt_free(const struct cpumask *cpu_map); 6039static int __sdt_alloc(const struct cpumask *cpu_map); 6040 6041static void __free_domain_allocs(struct s_data *d, enum s_alloc what, 6042 const struct cpumask *cpu_map) 6043{ 6044 switch (what) { 6045 case sa_rootdomain: 6046 if (!atomic_read(&d->rd->refcount)) 6047 free_rootdomain(&d->rd->rcu); /* fall through */ 6048 case sa_sd: 6049 free_percpu(d->sd); /* fall through */ 6050 case sa_sd_storage: 6051 __sdt_free(cpu_map); /* fall through */ 6052 case sa_none: 6053 break; 6054 } 6055} 6056 6057static enum s_alloc __visit_domain_allocation_hell(struct s_data *d, 6058 const struct cpumask *cpu_map) 6059{ 6060 memset(d, 0, sizeof(*d)); 6061 6062 if (__sdt_alloc(cpu_map)) 6063 return sa_sd_storage; 6064 d->sd = alloc_percpu(struct sched_domain *); 6065 if (!d->sd) 6066 return sa_sd_storage; 6067 d->rd = alloc_rootdomain(); 6068 if (!d->rd) 6069 return sa_sd; 6070 return sa_rootdomain; 6071} 6072 6073/* 6074 * NULL the sd_data elements we've used to build the sched_domain and 6075 * sched_group structure so that the subsequent __free_domain_allocs() 6076 * will not free the data we're using. 6077 */ 6078static void claim_allocations(int cpu, struct sched_domain *sd) 6079{ 6080 struct sd_data *sdd = sd->private; 6081 6082 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd); 6083 *per_cpu_ptr(sdd->sd, cpu) = NULL; 6084 6085 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref)) 6086 *per_cpu_ptr(sdd->sg, cpu) = NULL; 6087 6088 if (atomic_read(&(*per_cpu_ptr(sdd->sgc, cpu))->ref)) 6089 *per_cpu_ptr(sdd->sgc, cpu) = NULL; 6090} 6091 6092#ifdef CONFIG_NUMA 6093static int sched_domains_numa_levels; 6094static int *sched_domains_numa_distance; 6095static struct cpumask ***sched_domains_numa_masks; 6096static int sched_domains_curr_level; 6097#endif 6098 6099/* 6100 * SD_flags allowed in topology descriptions. 6101 * 6102 * SD_SHARE_CPUCAPACITY - describes SMT topologies 6103 * SD_SHARE_PKG_RESOURCES - describes shared caches 6104 * SD_NUMA - describes NUMA topologies 6105 * SD_SHARE_POWERDOMAIN - describes shared power domain 6106 * 6107 * Odd one out: 6108 * SD_ASYM_PACKING - describes SMT quirks 6109 */ 6110#define TOPOLOGY_SD_FLAGS \ 6111 (SD_SHARE_CPUCAPACITY | \ 6112 SD_SHARE_PKG_RESOURCES | \ 6113 SD_NUMA | \ 6114 SD_ASYM_PACKING | \ 6115 SD_SHARE_POWERDOMAIN) 6116 6117static struct sched_domain * 6118sd_init(struct sched_domain_topology_level *tl, int cpu) 6119{ 6120 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); 6121 int sd_weight, sd_flags = 0; 6122 6123#ifdef CONFIG_NUMA 6124 /* 6125 * Ugly hack to pass state to sd_numa_mask()... 6126 */ 6127 sched_domains_curr_level = tl->numa_level; 6128#endif 6129 6130 sd_weight = cpumask_weight(tl->mask(cpu)); 6131 6132 if (tl->sd_flags) 6133 sd_flags = (*tl->sd_flags)(); 6134 if (WARN_ONCE(sd_flags & ~TOPOLOGY_SD_FLAGS, 6135 "wrong sd_flags in topology description\n")) 6136 sd_flags &= ~TOPOLOGY_SD_FLAGS; 6137 6138 *sd = (struct sched_domain){ 6139 .min_interval = sd_weight, 6140 .max_interval = 2*sd_weight, 6141 .busy_factor = 32, 6142 .imbalance_pct = 125, 6143 6144 .cache_nice_tries = 0, 6145 .busy_idx = 0, 6146 .idle_idx = 0, 6147 .newidle_idx = 0, 6148 .wake_idx = 0, 6149 .forkexec_idx = 0, 6150 6151 .flags = 1*SD_LOAD_BALANCE 6152 | 1*SD_BALANCE_NEWIDLE 6153 | 1*SD_BALANCE_EXEC 6154 | 1*SD_BALANCE_FORK 6155 | 0*SD_BALANCE_WAKE 6156 | 1*SD_WAKE_AFFINE 6157 | 0*SD_SHARE_CPUCAPACITY 6158 | 0*SD_SHARE_PKG_RESOURCES 6159 | 0*SD_SERIALIZE 6160 | 0*SD_PREFER_SIBLING 6161 | 0*SD_NUMA 6162 | sd_flags 6163 , 6164 6165 .last_balance = jiffies, 6166 .balance_interval = sd_weight, 6167 .smt_gain = 0, 6168 .max_newidle_lb_cost = 0, 6169 .next_decay_max_lb_cost = jiffies, 6170#ifdef CONFIG_SCHED_DEBUG 6171 .name = tl->name, 6172#endif 6173 }; 6174 6175 /* 6176 * Convert topological properties into behaviour. 6177 */ 6178 6179 if (sd->flags & SD_SHARE_CPUCAPACITY) { 6180 sd->imbalance_pct = 110; 6181 sd->smt_gain = 1178; /* ~15% */ 6182 6183 } else if (sd->flags & SD_SHARE_PKG_RESOURCES) { 6184 sd->imbalance_pct = 117; 6185 sd->cache_nice_tries = 1; 6186 sd->busy_idx = 2; 6187 6188#ifdef CONFIG_NUMA 6189 } else if (sd->flags & SD_NUMA) { 6190 sd->cache_nice_tries = 2; 6191 sd->busy_idx = 3; 6192 sd->idle_idx = 2; 6193 6194 sd->flags |= SD_SERIALIZE; 6195 if (sched_domains_numa_distance[tl->numa_level] > RECLAIM_DISTANCE) { 6196 sd->flags &= ~(SD_BALANCE_EXEC | 6197 SD_BALANCE_FORK | 6198 SD_WAKE_AFFINE); 6199 } 6200 6201#endif 6202 } else { 6203 sd->flags |= SD_PREFER_SIBLING; 6204 sd->cache_nice_tries = 1; 6205 sd->busy_idx = 2; 6206 sd->idle_idx = 1; 6207 } 6208 6209 sd->private = &tl->data; 6210 6211 return sd; 6212} 6213 6214/* 6215 * Topology list, bottom-up. 6216 */ 6217static struct sched_domain_topology_level default_topology[] = { 6218#ifdef CONFIG_SCHED_SMT 6219 { cpu_smt_mask, cpu_smt_flags, SD_INIT_NAME(SMT) }, 6220#endif 6221#ifdef CONFIG_SCHED_MC 6222 { cpu_coregroup_mask, cpu_core_flags, SD_INIT_NAME(MC) }, 6223#endif 6224 { cpu_cpu_mask, SD_INIT_NAME(DIE) }, 6225 { NULL, }, 6226}; 6227 6228struct sched_domain_topology_level *sched_domain_topology = default_topology; 6229 6230#define for_each_sd_topology(tl) \ 6231 for (tl = sched_domain_topology; tl->mask; tl++) 6232 6233void set_sched_topology(struct sched_domain_topology_level *tl) 6234{ 6235 sched_domain_topology = tl; 6236} 6237 6238#ifdef CONFIG_NUMA 6239 6240static const struct cpumask *sd_numa_mask(int cpu) 6241{ 6242 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)]; 6243} 6244 6245static void sched_numa_warn(const char *str) 6246{ 6247 static int done = false; 6248 int i,j; 6249 6250 if (done) 6251 return; 6252 6253 done = true; 6254 6255 printk(KERN_WARNING "ERROR: %s\n\n", str); 6256 6257 for (i = 0; i < nr_node_ids; i++) { 6258 printk(KERN_WARNING " "); 6259 for (j = 0; j < nr_node_ids; j++) 6260 printk(KERN_CONT "%02d ", node_distance(i,j)); 6261 printk(KERN_CONT "\n"); 6262 } 6263 printk(KERN_WARNING "\n"); 6264} 6265 6266static bool find_numa_distance(int distance) 6267{ 6268 int i; 6269 6270 if (distance == node_distance(0, 0)) 6271 return true; 6272 6273 for (i = 0; i < sched_domains_numa_levels; i++) { 6274 if (sched_domains_numa_distance[i] == distance) 6275 return true; 6276 } 6277 6278 return false; 6279} 6280 6281static void sched_init_numa(void) 6282{ 6283 int next_distance, curr_distance = node_distance(0, 0); 6284 struct sched_domain_topology_level *tl; 6285 int level = 0; 6286 int i, j, k; 6287 6288 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL); 6289 if (!sched_domains_numa_distance) 6290 return; 6291 6292 /* 6293 * O(nr_nodes^2) deduplicating selection sort -- in order to find the 6294 * unique distances in the node_distance() table. 6295 * 6296 * Assumes node_distance(0,j) includes all distances in 6297 * node_distance(i,j) in order to avoid cubic time. 6298 */ 6299 next_distance = curr_distance; 6300 for (i = 0; i < nr_node_ids; i++) { 6301 for (j = 0; j < nr_node_ids; j++) { 6302 for (k = 0; k < nr_node_ids; k++) { 6303 int distance = node_distance(i, k); 6304 6305 if (distance > curr_distance && 6306 (distance < next_distance || 6307 next_distance == curr_distance)) 6308 next_distance = distance; 6309 6310 /* 6311 * While not a strong assumption it would be nice to know 6312 * about cases where if node A is connected to B, B is not 6313 * equally connected to A. 6314 */ 6315 if (sched_debug() && node_distance(k, i) != distance) 6316 sched_numa_warn("Node-distance not symmetric"); 6317 6318 if (sched_debug() && i && !find_numa_distance(distance)) 6319 sched_numa_warn("Node-0 not representative"); 6320 } 6321 if (next_distance != curr_distance) { 6322 sched_domains_numa_distance[level++] = next_distance; 6323 sched_domains_numa_levels = level; 6324 curr_distance = next_distance; 6325 } else break; 6326 } 6327 6328 /* 6329 * In case of sched_debug() we verify the above assumption. 6330 */ 6331 if (!sched_debug()) 6332 break; 6333 } 6334 /* 6335 * 'level' contains the number of unique distances, excluding the 6336 * identity distance node_distance(i,i). 6337 * 6338 * The sched_domains_numa_distance[] array includes the actual distance 6339 * numbers. 6340 */ 6341 6342 /* 6343 * Here, we should temporarily reset sched_domains_numa_levels to 0. 6344 * If it fails to allocate memory for array sched_domains_numa_masks[][], 6345 * the array will contain less then 'level' members. This could be 6346 * dangerous when we use it to iterate array sched_domains_numa_masks[][] 6347 * in other functions. 6348 * 6349 * We reset it to 'level' at the end of this function. 6350 */ 6351 sched_domains_numa_levels = 0; 6352 6353 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL); 6354 if (!sched_domains_numa_masks) 6355 return; 6356 6357 /* 6358 * Now for each level, construct a mask per node which contains all 6359 * cpus of nodes that are that many hops away from us. 6360 */ 6361 for (i = 0; i < level; i++) { 6362 sched_domains_numa_masks[i] = 6363 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL); 6364 if (!sched_domains_numa_masks[i]) 6365 return; 6366 6367 for (j = 0; j < nr_node_ids; j++) { 6368 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL); 6369 if (!mask) 6370 return; 6371 6372 sched_domains_numa_masks[i][j] = mask; 6373 6374 for (k = 0; k < nr_node_ids; k++) { 6375 if (node_distance(j, k) > sched_domains_numa_distance[i]) 6376 continue; 6377 6378 cpumask_or(mask, mask, cpumask_of_node(k)); 6379 } 6380 } 6381 } 6382 6383 /* Compute default topology size */ 6384 for (i = 0; sched_domain_topology[i].mask; i++); 6385 6386 tl = kzalloc((i + level + 1) * 6387 sizeof(struct sched_domain_topology_level), GFP_KERNEL); 6388 if (!tl) 6389 return; 6390 6391 /* 6392 * Copy the default topology bits.. 6393 */ 6394 for (i = 0; sched_domain_topology[i].mask; i++) 6395 tl[i] = sched_domain_topology[i]; 6396 6397 /* 6398 * .. and append 'j' levels of NUMA goodness. 6399 */ 6400 for (j = 0; j < level; i++, j++) { 6401 tl[i] = (struct sched_domain_topology_level){ 6402 .mask = sd_numa_mask, 6403 .sd_flags = cpu_numa_flags, 6404 .flags = SDTL_OVERLAP, 6405 .numa_level = j, 6406 SD_INIT_NAME(NUMA) 6407 }; 6408 } 6409 6410 sched_domain_topology = tl; 6411 6412 sched_domains_numa_levels = level; 6413} 6414 6415static void sched_domains_numa_masks_set(int cpu) 6416{ 6417 int i, j; 6418 int node = cpu_to_node(cpu); 6419 6420 for (i = 0; i < sched_domains_numa_levels; i++) { 6421 for (j = 0; j < nr_node_ids; j++) { 6422 if (node_distance(j, node) <= sched_domains_numa_distance[i]) 6423 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]); 6424 } 6425 } 6426} 6427 6428static void sched_domains_numa_masks_clear(int cpu) 6429{ 6430 int i, j; 6431 for (i = 0; i < sched_domains_numa_levels; i++) { 6432 for (j = 0; j < nr_node_ids; j++) 6433 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]); 6434 } 6435} 6436 6437/* 6438 * Update sched_domains_numa_masks[level][node] array when new cpus 6439 * are onlined. 6440 */ 6441static int sched_domains_numa_masks_update(struct notifier_block *nfb, 6442 unsigned long action, 6443 void *hcpu) 6444{ 6445 int cpu = (long)hcpu; 6446 6447 switch (action & ~CPU_TASKS_FROZEN) { 6448 case CPU_ONLINE: 6449 sched_domains_numa_masks_set(cpu); 6450 break; 6451 6452 case CPU_DEAD: 6453 sched_domains_numa_masks_clear(cpu); 6454 break; 6455 6456 default: 6457 return NOTIFY_DONE; 6458 } 6459 6460 return NOTIFY_OK; 6461} 6462#else 6463static inline void sched_init_numa(void) 6464{ 6465} 6466 6467static int sched_domains_numa_masks_update(struct notifier_block *nfb, 6468 unsigned long action, 6469 void *hcpu) 6470{ 6471 return 0; 6472} 6473#endif /* CONFIG_NUMA */ 6474 6475static int __sdt_alloc(const struct cpumask *cpu_map) 6476{ 6477 struct sched_domain_topology_level *tl; 6478 int j; 6479 6480 for_each_sd_topology(tl) { 6481 struct sd_data *sdd = &tl->data; 6482 6483 sdd->sd = alloc_percpu(struct sched_domain *); 6484 if (!sdd->sd) 6485 return -ENOMEM; 6486 6487 sdd->sg = alloc_percpu(struct sched_group *); 6488 if (!sdd->sg) 6489 return -ENOMEM; 6490 6491 sdd->sgc = alloc_percpu(struct sched_group_capacity *); 6492 if (!sdd->sgc) 6493 return -ENOMEM; 6494 6495 for_each_cpu(j, cpu_map) { 6496 struct sched_domain *sd; 6497 struct sched_group *sg; 6498 struct sched_group_capacity *sgc; 6499 6500 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(), 6501 GFP_KERNEL, cpu_to_node(j)); 6502 if (!sd) 6503 return -ENOMEM; 6504 6505 *per_cpu_ptr(sdd->sd, j) = sd; 6506 6507 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 6508 GFP_KERNEL, cpu_to_node(j)); 6509 if (!sg) 6510 return -ENOMEM; 6511 6512 sg->next = sg; 6513 6514 *per_cpu_ptr(sdd->sg, j) = sg; 6515 6516 sgc = kzalloc_node(sizeof(struct sched_group_capacity) + cpumask_size(), 6517 GFP_KERNEL, cpu_to_node(j)); 6518 if (!sgc) 6519 return -ENOMEM; 6520 6521 *per_cpu_ptr(sdd->sgc, j) = sgc; 6522 } 6523 } 6524 6525 return 0; 6526} 6527 6528static void __sdt_free(const struct cpumask *cpu_map) 6529{ 6530 struct sched_domain_topology_level *tl; 6531 int j; 6532 6533 for_each_sd_topology(tl) { 6534 struct sd_data *sdd = &tl->data; 6535 6536 for_each_cpu(j, cpu_map) { 6537 struct sched_domain *sd; 6538 6539 if (sdd->sd) { 6540 sd = *per_cpu_ptr(sdd->sd, j); 6541 if (sd && (sd->flags & SD_OVERLAP)) 6542 free_sched_groups(sd->groups, 0); 6543 kfree(*per_cpu_ptr(sdd->sd, j)); 6544 } 6545 6546 if (sdd->sg) 6547 kfree(*per_cpu_ptr(sdd->sg, j)); 6548 if (sdd->sgc) 6549 kfree(*per_cpu_ptr(sdd->sgc, j)); 6550 } 6551 free_percpu(sdd->sd); 6552 sdd->sd = NULL; 6553 free_percpu(sdd->sg); 6554 sdd->sg = NULL; 6555 free_percpu(sdd->sgc); 6556 sdd->sgc = NULL; 6557 } 6558} 6559 6560struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl, 6561 const struct cpumask *cpu_map, struct sched_domain_attr *attr, 6562 struct sched_domain *child, int cpu) 6563{ 6564 struct sched_domain *sd = sd_init(tl, cpu); 6565 if (!sd) 6566 return child; 6567 6568 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu)); 6569 if (child) { 6570 sd->level = child->level + 1; 6571 sched_domain_level_max = max(sched_domain_level_max, sd->level); 6572 child->parent = sd; 6573 sd->child = child; 6574 6575 if (!cpumask_subset(sched_domain_span(child), 6576 sched_domain_span(sd))) { 6577 pr_err("BUG: arch topology borken\n"); 6578#ifdef CONFIG_SCHED_DEBUG 6579 pr_err(" the %s domain not a subset of the %s domain\n", 6580 child->name, sd->name); 6581#endif 6582 /* Fixup, ensure @sd has at least @child cpus. */ 6583 cpumask_or(sched_domain_span(sd), 6584 sched_domain_span(sd), 6585 sched_domain_span(child)); 6586 } 6587 6588 } 6589 set_domain_attribute(sd, attr); 6590 6591 return sd; 6592} 6593 6594/* 6595 * Build sched domains for a given set of cpus and attach the sched domains 6596 * to the individual cpus 6597 */ 6598static int build_sched_domains(const struct cpumask *cpu_map, 6599 struct sched_domain_attr *attr) 6600{ 6601 enum s_alloc alloc_state; 6602 struct sched_domain *sd; 6603 struct s_data d; 6604 int i, ret = -ENOMEM; 6605 6606 alloc_state = __visit_domain_allocation_hell(&d, cpu_map); 6607 if (alloc_state != sa_rootdomain) 6608 goto error; 6609 6610 /* Set up domains for cpus specified by the cpu_map. */ 6611 for_each_cpu(i, cpu_map) { 6612 struct sched_domain_topology_level *tl; 6613 6614 sd = NULL; 6615 for_each_sd_topology(tl) { 6616 sd = build_sched_domain(tl, cpu_map, attr, sd, i); 6617 if (tl == sched_domain_topology) 6618 *per_cpu_ptr(d.sd, i) = sd; 6619 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP)) 6620 sd->flags |= SD_OVERLAP; 6621 if (cpumask_equal(cpu_map, sched_domain_span(sd))) 6622 break; 6623 } 6624 } 6625 6626 /* Build the groups for the domains */ 6627 for_each_cpu(i, cpu_map) { 6628 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 6629 sd->span_weight = cpumask_weight(sched_domain_span(sd)); 6630 if (sd->flags & SD_OVERLAP) { 6631 if (build_overlap_sched_groups(sd, i)) 6632 goto error; 6633 } else { 6634 if (build_sched_groups(sd, i)) 6635 goto error; 6636 } 6637 } 6638 } 6639 6640 /* Calculate CPU capacity for physical packages and nodes */ 6641 for (i = nr_cpumask_bits-1; i >= 0; i--) { 6642 if (!cpumask_test_cpu(i, cpu_map)) 6643 continue; 6644 6645 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 6646 claim_allocations(i, sd); 6647 init_sched_groups_capacity(i, sd); 6648 } 6649 } 6650 6651 /* Attach the domains */ 6652 rcu_read_lock(); 6653 for_each_cpu(i, cpu_map) { 6654 sd = *per_cpu_ptr(d.sd, i); 6655 cpu_attach_domain(sd, d.rd, i); 6656 } 6657 rcu_read_unlock(); 6658 6659 ret = 0; 6660error: 6661 __free_domain_allocs(&d, alloc_state, cpu_map); 6662 return ret; 6663} 6664 6665static cpumask_var_t *doms_cur; /* current sched domains */ 6666static int ndoms_cur; /* number of sched domains in 'doms_cur' */ 6667static struct sched_domain_attr *dattr_cur; 6668 /* attribues of custom domains in 'doms_cur' */ 6669 6670/* 6671 * Special case: If a kmalloc of a doms_cur partition (array of 6672 * cpumask) fails, then fallback to a single sched domain, 6673 * as determined by the single cpumask fallback_doms. 6674 */ 6675static cpumask_var_t fallback_doms; 6676 6677/* 6678 * arch_update_cpu_topology lets virtualized architectures update the 6679 * cpu core maps. It is supposed to return 1 if the topology changed 6680 * or 0 if it stayed the same. 6681 */ 6682int __weak arch_update_cpu_topology(void) 6683{ 6684 return 0; 6685} 6686 6687cpumask_var_t *alloc_sched_domains(unsigned int ndoms) 6688{ 6689 int i; 6690 cpumask_var_t *doms; 6691 6692 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL); 6693 if (!doms) 6694 return NULL; 6695 for (i = 0; i < ndoms; i++) { 6696 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) { 6697 free_sched_domains(doms, i); 6698 return NULL; 6699 } 6700 } 6701 return doms; 6702} 6703 6704void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms) 6705{ 6706 unsigned int i; 6707 for (i = 0; i < ndoms; i++) 6708 free_cpumask_var(doms[i]); 6709 kfree(doms); 6710} 6711 6712/* 6713 * Set up scheduler domains and groups. Callers must hold the hotplug lock. 6714 * For now this just excludes isolated cpus, but could be used to 6715 * exclude other special cases in the future. 6716 */ 6717static int init_sched_domains(const struct cpumask *cpu_map) 6718{ 6719 int err; 6720 6721 arch_update_cpu_topology(); 6722 ndoms_cur = 1; 6723 doms_cur = alloc_sched_domains(ndoms_cur); 6724 if (!doms_cur) 6725 doms_cur = &fallback_doms; 6726 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map); 6727 err = build_sched_domains(doms_cur[0], NULL); 6728 register_sched_domain_sysctl(); 6729 6730 return err; 6731} 6732 6733/* 6734 * Detach sched domains from a group of cpus specified in cpu_map 6735 * These cpus will now be attached to the NULL domain 6736 */ 6737static void detach_destroy_domains(const struct cpumask *cpu_map) 6738{ 6739 int i; 6740 6741 rcu_read_lock(); 6742 for_each_cpu(i, cpu_map) 6743 cpu_attach_domain(NULL, &def_root_domain, i); 6744 rcu_read_unlock(); 6745} 6746 6747/* handle null as "default" */ 6748static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur, 6749 struct sched_domain_attr *new, int idx_new) 6750{ 6751 struct sched_domain_attr tmp; 6752 6753 /* fast path */ 6754 if (!new && !cur) 6755 return 1; 6756 6757 tmp = SD_ATTR_INIT; 6758 return !memcmp(cur ? (cur + idx_cur) : &tmp, 6759 new ? (new + idx_new) : &tmp, 6760 sizeof(struct sched_domain_attr)); 6761} 6762 6763/* 6764 * Partition sched domains as specified by the 'ndoms_new' 6765 * cpumasks in the array doms_new[] of cpumasks. This compares 6766 * doms_new[] to the current sched domain partitioning, doms_cur[]. 6767 * It destroys each deleted domain and builds each new domain. 6768 * 6769 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'. 6770 * The masks don't intersect (don't overlap.) We should setup one 6771 * sched domain for each mask. CPUs not in any of the cpumasks will 6772 * not be load balanced. If the same cpumask appears both in the 6773 * current 'doms_cur' domains and in the new 'doms_new', we can leave 6774 * it as it is. 6775 * 6776 * The passed in 'doms_new' should be allocated using 6777 * alloc_sched_domains. This routine takes ownership of it and will 6778 * free_sched_domains it when done with it. If the caller failed the 6779 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1, 6780 * and partition_sched_domains() will fallback to the single partition 6781 * 'fallback_doms', it also forces the domains to be rebuilt. 6782 * 6783 * If doms_new == NULL it will be replaced with cpu_online_mask. 6784 * ndoms_new == 0 is a special case for destroying existing domains, 6785 * and it will not create the default domain. 6786 * 6787 * Call with hotplug lock held 6788 */ 6789void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[], 6790 struct sched_domain_attr *dattr_new) 6791{ 6792 int i, j, n; 6793 int new_topology; 6794 6795 mutex_lock(&sched_domains_mutex); 6796 6797 /* always unregister in case we don't destroy any domains */ 6798 unregister_sched_domain_sysctl(); 6799 6800 /* Let architecture update cpu core mappings. */ 6801 new_topology = arch_update_cpu_topology(); 6802 6803 n = doms_new ? ndoms_new : 0; 6804 6805 /* Destroy deleted domains */ 6806 for (i = 0; i < ndoms_cur; i++) { 6807 for (j = 0; j < n && !new_topology; j++) { 6808 if (cpumask_equal(doms_cur[i], doms_new[j]) 6809 && dattrs_equal(dattr_cur, i, dattr_new, j)) 6810 goto match1; 6811 } 6812 /* no match - a current sched domain not in new doms_new[] */ 6813 detach_destroy_domains(doms_cur[i]); 6814match1: 6815 ; 6816 } 6817 6818 n = ndoms_cur; 6819 if (doms_new == NULL) { 6820 n = 0; 6821 doms_new = &fallback_doms; 6822 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map); 6823 WARN_ON_ONCE(dattr_new); 6824 } 6825 6826 /* Build new domains */ 6827 for (i = 0; i < ndoms_new; i++) { 6828 for (j = 0; j < n && !new_topology; j++) { 6829 if (cpumask_equal(doms_new[i], doms_cur[j]) 6830 && dattrs_equal(dattr_new, i, dattr_cur, j)) 6831 goto match2; 6832 } 6833 /* no match - add a new doms_new */ 6834 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL); 6835match2: 6836 ; 6837 } 6838 6839 /* Remember the new sched domains */ 6840 if (doms_cur != &fallback_doms) 6841 free_sched_domains(doms_cur, ndoms_cur); 6842 kfree(dattr_cur); /* kfree(NULL) is safe */ 6843 doms_cur = doms_new; 6844 dattr_cur = dattr_new; 6845 ndoms_cur = ndoms_new; 6846 6847 register_sched_domain_sysctl(); 6848 6849 mutex_unlock(&sched_domains_mutex); 6850} 6851 6852static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */ 6853 6854/* 6855 * Update cpusets according to cpu_active mask. If cpusets are 6856 * disabled, cpuset_update_active_cpus() becomes a simple wrapper 6857 * around partition_sched_domains(). 6858 * 6859 * If we come here as part of a suspend/resume, don't touch cpusets because we 6860 * want to restore it back to its original state upon resume anyway. 6861 */ 6862static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action, 6863 void *hcpu) 6864{ 6865 switch (action) { 6866 case CPU_ONLINE_FROZEN: 6867 case CPU_DOWN_FAILED_FROZEN: 6868 6869 /* 6870 * num_cpus_frozen tracks how many CPUs are involved in suspend 6871 * resume sequence. As long as this is not the last online 6872 * operation in the resume sequence, just build a single sched 6873 * domain, ignoring cpusets. 6874 */ 6875 num_cpus_frozen--; 6876 if (likely(num_cpus_frozen)) { 6877 partition_sched_domains(1, NULL, NULL); 6878 break; 6879 } 6880 6881 /* 6882 * This is the last CPU online operation. So fall through and 6883 * restore the original sched domains by considering the 6884 * cpuset configurations. 6885 */ 6886 6887 case CPU_ONLINE: 6888 case CPU_DOWN_FAILED: 6889 cpuset_update_active_cpus(true); 6890 break; 6891 default: 6892 return NOTIFY_DONE; 6893 } 6894 return NOTIFY_OK; 6895} 6896 6897static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action, 6898 void *hcpu) 6899{ 6900 switch (action) { 6901 case CPU_DOWN_PREPARE: 6902 cpuset_update_active_cpus(false); 6903 break; 6904 case CPU_DOWN_PREPARE_FROZEN: 6905 num_cpus_frozen++; 6906 partition_sched_domains(1, NULL, NULL); 6907 break; 6908 default: 6909 return NOTIFY_DONE; 6910 } 6911 return NOTIFY_OK; 6912} 6913 6914void __init sched_init_smp(void) 6915{ 6916 cpumask_var_t non_isolated_cpus; 6917 6918 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL); 6919 alloc_cpumask_var(&fallback_doms, GFP_KERNEL); 6920 6921 sched_init_numa(); 6922 6923 /* 6924 * There's no userspace yet to cause hotplug operations; hence all the 6925 * cpu masks are stable and all blatant races in the below code cannot 6926 * happen. 6927 */ 6928 mutex_lock(&sched_domains_mutex); 6929 init_sched_domains(cpu_active_mask); 6930 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map); 6931 if (cpumask_empty(non_isolated_cpus)) 6932 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus); 6933 mutex_unlock(&sched_domains_mutex); 6934 6935 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE); 6936 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE); 6937 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE); 6938 6939 init_hrtick(); 6940 6941 /* Move init over to a non-isolated CPU */ 6942 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0) 6943 BUG(); 6944 sched_init_granularity(); 6945 free_cpumask_var(non_isolated_cpus); 6946 6947 init_sched_rt_class(); 6948 init_sched_dl_class(); 6949} 6950#else 6951void __init sched_init_smp(void) 6952{ 6953 sched_init_granularity(); 6954} 6955#endif /* CONFIG_SMP */ 6956 6957const_debug unsigned int sysctl_timer_migration = 1; 6958 6959int in_sched_functions(unsigned long addr) 6960{ 6961 return in_lock_functions(addr) || 6962 (addr >= (unsigned long)__sched_text_start 6963 && addr < (unsigned long)__sched_text_end); 6964} 6965 6966#ifdef CONFIG_CGROUP_SCHED 6967/* 6968 * Default task group. 6969 * Every task in system belongs to this group at bootup. 6970 */ 6971struct task_group root_task_group; 6972LIST_HEAD(task_groups); 6973#endif 6974 6975DECLARE_PER_CPU(cpumask_var_t, load_balance_mask); 6976 6977void __init sched_init(void) 6978{ 6979 int i, j; 6980 unsigned long alloc_size = 0, ptr; 6981 6982#ifdef CONFIG_FAIR_GROUP_SCHED 6983 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 6984#endif 6985#ifdef CONFIG_RT_GROUP_SCHED 6986 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 6987#endif 6988#ifdef CONFIG_CPUMASK_OFFSTACK 6989 alloc_size += num_possible_cpus() * cpumask_size(); 6990#endif 6991 if (alloc_size) { 6992 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT); 6993 6994#ifdef CONFIG_FAIR_GROUP_SCHED 6995 root_task_group.se = (struct sched_entity **)ptr; 6996 ptr += nr_cpu_ids * sizeof(void **); 6997 6998 root_task_group.cfs_rq = (struct cfs_rq **)ptr; 6999 ptr += nr_cpu_ids * sizeof(void **); 7000 7001#endif /* CONFIG_FAIR_GROUP_SCHED */ 7002#ifdef CONFIG_RT_GROUP_SCHED 7003 root_task_group.rt_se = (struct sched_rt_entity **)ptr; 7004 ptr += nr_cpu_ids * sizeof(void **); 7005 7006 root_task_group.rt_rq = (struct rt_rq **)ptr; 7007 ptr += nr_cpu_ids * sizeof(void **); 7008 7009#endif /* CONFIG_RT_GROUP_SCHED */ 7010#ifdef CONFIG_CPUMASK_OFFSTACK 7011 for_each_possible_cpu(i) { 7012 per_cpu(load_balance_mask, i) = (void *)ptr; 7013 ptr += cpumask_size(); 7014 } 7015#endif /* CONFIG_CPUMASK_OFFSTACK */ 7016 } 7017 7018 init_rt_bandwidth(&def_rt_bandwidth, 7019 global_rt_period(), global_rt_runtime()); 7020 init_dl_bandwidth(&def_dl_bandwidth, 7021 global_rt_period(), global_rt_runtime()); 7022 7023#ifdef CONFIG_SMP 7024 init_defrootdomain(); 7025#endif 7026 7027#ifdef CONFIG_RT_GROUP_SCHED 7028 init_rt_bandwidth(&root_task_group.rt_bandwidth, 7029 global_rt_period(), global_rt_runtime()); 7030#endif /* CONFIG_RT_GROUP_SCHED */ 7031 7032#ifdef CONFIG_CGROUP_SCHED 7033 list_add(&root_task_group.list, &task_groups); 7034 INIT_LIST_HEAD(&root_task_group.children); 7035 INIT_LIST_HEAD(&root_task_group.siblings); 7036 autogroup_init(&init_task); 7037 7038#endif /* CONFIG_CGROUP_SCHED */ 7039 7040 for_each_possible_cpu(i) { 7041 struct rq *rq; 7042 7043 rq = cpu_rq(i); 7044 raw_spin_lock_init(&rq->lock); 7045 rq->nr_running = 0; 7046 rq->calc_load_active = 0; 7047 rq->calc_load_update = jiffies + LOAD_FREQ; 7048 init_cfs_rq(&rq->cfs); 7049 init_rt_rq(&rq->rt, rq); 7050 init_dl_rq(&rq->dl, rq); 7051#ifdef CONFIG_FAIR_GROUP_SCHED 7052 root_task_group.shares = ROOT_TASK_GROUP_LOAD; 7053 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list); 7054 /* 7055 * How much cpu bandwidth does root_task_group get? 7056 * 7057 * In case of task-groups formed thr' the cgroup filesystem, it 7058 * gets 100% of the cpu resources in the system. This overall 7059 * system cpu resource is divided among the tasks of 7060 * root_task_group and its child task-groups in a fair manner, 7061 * based on each entity's (task or task-group's) weight 7062 * (se->load.weight). 7063 * 7064 * In other words, if root_task_group has 10 tasks of weight 7065 * 1024) and two child groups A0 and A1 (of weight 1024 each), 7066 * then A0's share of the cpu resource is: 7067 * 7068 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33% 7069 * 7070 * We achieve this by letting root_task_group's tasks sit 7071 * directly in rq->cfs (i.e root_task_group->se[] = NULL). 7072 */ 7073 init_cfs_bandwidth(&root_task_group.cfs_bandwidth); 7074 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL); 7075#endif /* CONFIG_FAIR_GROUP_SCHED */ 7076 7077 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime; 7078#ifdef CONFIG_RT_GROUP_SCHED 7079 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL); 7080#endif 7081 7082 for (j = 0; j < CPU_LOAD_IDX_MAX; j++) 7083 rq->cpu_load[j] = 0; 7084 7085 rq->last_load_update_tick = jiffies; 7086 7087#ifdef CONFIG_SMP 7088 rq->sd = NULL; 7089 rq->rd = NULL; 7090 rq->cpu_capacity = SCHED_CAPACITY_SCALE; 7091 rq->post_schedule = 0; 7092 rq->active_balance = 0; 7093 rq->next_balance = jiffies; 7094 rq->push_cpu = 0; 7095 rq->cpu = i; 7096 rq->online = 0; 7097 rq->idle_stamp = 0; 7098 rq->avg_idle = 2*sysctl_sched_migration_cost; 7099 rq->max_idle_balance_cost = sysctl_sched_migration_cost; 7100 7101 INIT_LIST_HEAD(&rq->cfs_tasks); 7102 7103 rq_attach_root(rq, &def_root_domain); 7104#ifdef CONFIG_NO_HZ_COMMON 7105 rq->nohz_flags = 0; 7106#endif 7107#ifdef CONFIG_NO_HZ_FULL 7108 rq->last_sched_tick = 0; 7109#endif 7110#endif 7111 init_rq_hrtick(rq); 7112 atomic_set(&rq->nr_iowait, 0); 7113 } 7114 7115 set_load_weight(&init_task); 7116 7117#ifdef CONFIG_PREEMPT_NOTIFIERS 7118 INIT_HLIST_HEAD(&init_task.preempt_notifiers); 7119#endif 7120 7121 /* 7122 * The boot idle thread does lazy MMU switching as well: 7123 */ 7124 atomic_inc(&init_mm.mm_count); 7125 enter_lazy_tlb(&init_mm, current); 7126 7127 /* 7128 * Make us the idle thread. Technically, schedule() should not be 7129 * called from this thread, however somewhere below it might be, 7130 * but because we are the idle thread, we just pick up running again 7131 * when this runqueue becomes "idle". 7132 */ 7133 init_idle(current, smp_processor_id()); 7134 7135 calc_load_update = jiffies + LOAD_FREQ; 7136 7137 /* 7138 * During early bootup we pretend to be a normal task: 7139 */ 7140 current->sched_class = &fair_sched_class; 7141 7142#ifdef CONFIG_SMP 7143 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT); 7144 /* May be allocated at isolcpus cmdline parse time */ 7145 if (cpu_isolated_map == NULL) 7146 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT); 7147 idle_thread_set_boot_cpu(); 7148 set_cpu_rq_start_time(); 7149#endif 7150 init_sched_fair_class(); 7151 7152 scheduler_running = 1; 7153} 7154 7155#ifdef CONFIG_DEBUG_ATOMIC_SLEEP 7156static inline int preempt_count_equals(int preempt_offset) 7157{ 7158 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth(); 7159 7160 return (nested == preempt_offset); 7161} 7162 7163void __might_sleep(const char *file, int line, int preempt_offset) 7164{ 7165 static unsigned long prev_jiffy; /* ratelimiting */ 7166 7167 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */ 7168 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() && 7169 !is_idle_task(current)) || 7170 system_state != SYSTEM_RUNNING || oops_in_progress) 7171 return; 7172 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy) 7173 return; 7174 prev_jiffy = jiffies; 7175 7176 printk(KERN_ERR 7177 "BUG: sleeping function called from invalid context at %s:%d\n", 7178 file, line); 7179 printk(KERN_ERR 7180 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n", 7181 in_atomic(), irqs_disabled(), 7182 current->pid, current->comm); 7183 7184 debug_show_held_locks(current); 7185 if (irqs_disabled()) 7186 print_irqtrace_events(current); 7187#ifdef CONFIG_DEBUG_PREEMPT 7188 if (!preempt_count_equals(preempt_offset)) { 7189 pr_err("Preemption disabled at:"); 7190 print_ip_sym(current->preempt_disable_ip); 7191 pr_cont("\n"); 7192 } 7193#endif 7194 dump_stack(); 7195} 7196EXPORT_SYMBOL(__might_sleep); 7197#endif 7198 7199#ifdef CONFIG_MAGIC_SYSRQ 7200static void normalize_task(struct rq *rq, struct task_struct *p) 7201{ 7202 const struct sched_class *prev_class = p->sched_class; 7203 struct sched_attr attr = { 7204 .sched_policy = SCHED_NORMAL, 7205 }; 7206 int old_prio = p->prio; 7207 int queued; 7208 7209 queued = task_on_rq_queued(p); 7210 if (queued) 7211 dequeue_task(rq, p, 0); 7212 __setscheduler(rq, p, &attr); 7213 if (queued) { 7214 enqueue_task(rq, p, 0); 7215 resched_curr(rq); 7216 } 7217 7218 check_class_changed(rq, p, prev_class, old_prio); 7219} 7220 7221void normalize_rt_tasks(void) 7222{ 7223 struct task_struct *g, *p; 7224 unsigned long flags; 7225 struct rq *rq; 7226 7227 read_lock(&tasklist_lock); 7228 for_each_process_thread(g, p) { 7229 /* 7230 * Only normalize user tasks: 7231 */ 7232 if (p->flags & PF_KTHREAD) 7233 continue; 7234 7235 p->se.exec_start = 0; 7236#ifdef CONFIG_SCHEDSTATS 7237 p->se.statistics.wait_start = 0; 7238 p->se.statistics.sleep_start = 0; 7239 p->se.statistics.block_start = 0; 7240#endif 7241 7242 if (!dl_task(p) && !rt_task(p)) { 7243 /* 7244 * Renice negative nice level userspace 7245 * tasks back to 0: 7246 */ 7247 if (task_nice(p) < 0) 7248 set_user_nice(p, 0); 7249 continue; 7250 } 7251 7252 rq = task_rq_lock(p, &flags); 7253 normalize_task(rq, p); 7254 task_rq_unlock(rq, p, &flags); 7255 } 7256 read_unlock(&tasklist_lock); 7257} 7258 7259#endif /* CONFIG_MAGIC_SYSRQ */ 7260 7261#if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) 7262/* 7263 * These functions are only useful for the IA64 MCA handling, or kdb. 7264 * 7265 * They can only be called when the whole system has been 7266 * stopped - every CPU needs to be quiescent, and no scheduling 7267 * activity can take place. Using them for anything else would 7268 * be a serious bug, and as a result, they aren't even visible 7269 * under any other configuration. 7270 */ 7271 7272/** 7273 * curr_task - return the current task for a given cpu. 7274 * @cpu: the processor in question. 7275 * 7276 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 7277 * 7278 * Return: The current task for @cpu. 7279 */ 7280struct task_struct *curr_task(int cpu) 7281{ 7282 return cpu_curr(cpu); 7283} 7284 7285#endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */ 7286 7287#ifdef CONFIG_IA64 7288/** 7289 * set_curr_task - set the current task for a given cpu. 7290 * @cpu: the processor in question. 7291 * @p: the task pointer to set. 7292 * 7293 * Description: This function must only be used when non-maskable interrupts 7294 * are serviced on a separate stack. It allows the architecture to switch the 7295 * notion of the current task on a cpu in a non-blocking manner. This function 7296 * must be called with all CPU's synchronized, and interrupts disabled, the 7297 * and caller must save the original value of the current task (see 7298 * curr_task() above) and restore that value before reenabling interrupts and 7299 * re-starting the system. 7300 * 7301 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 7302 */ 7303void set_curr_task(int cpu, struct task_struct *p) 7304{ 7305 cpu_curr(cpu) = p; 7306} 7307 7308#endif 7309 7310#ifdef CONFIG_CGROUP_SCHED 7311/* task_group_lock serializes the addition/removal of task groups */ 7312static DEFINE_SPINLOCK(task_group_lock); 7313 7314static void free_sched_group(struct task_group *tg) 7315{ 7316 free_fair_sched_group(tg); 7317 free_rt_sched_group(tg); 7318 autogroup_free(tg); 7319 kfree(tg); 7320} 7321 7322/* allocate runqueue etc for a new task group */ 7323struct task_group *sched_create_group(struct task_group *parent) 7324{ 7325 struct task_group *tg; 7326 7327 tg = kzalloc(sizeof(*tg), GFP_KERNEL); 7328 if (!tg) 7329 return ERR_PTR(-ENOMEM); 7330 7331 if (!alloc_fair_sched_group(tg, parent)) 7332 goto err; 7333 7334 if (!alloc_rt_sched_group(tg, parent)) 7335 goto err; 7336 7337 return tg; 7338 7339err: 7340 free_sched_group(tg); 7341 return ERR_PTR(-ENOMEM); 7342} 7343 7344void sched_online_group(struct task_group *tg, struct task_group *parent) 7345{ 7346 unsigned long flags; 7347 7348 spin_lock_irqsave(&task_group_lock, flags); 7349 list_add_rcu(&tg->list, &task_groups); 7350 7351 WARN_ON(!parent); /* root should already exist */ 7352 7353 tg->parent = parent; 7354 INIT_LIST_HEAD(&tg->children); 7355 list_add_rcu(&tg->siblings, &parent->children); 7356 spin_unlock_irqrestore(&task_group_lock, flags); 7357} 7358 7359/* rcu callback to free various structures associated with a task group */ 7360static void free_sched_group_rcu(struct rcu_head *rhp) 7361{ 7362 /* now it should be safe to free those cfs_rqs */ 7363 free_sched_group(container_of(rhp, struct task_group, rcu)); 7364} 7365 7366/* Destroy runqueue etc associated with a task group */ 7367void sched_destroy_group(struct task_group *tg) 7368{ 7369 /* wait for possible concurrent references to cfs_rqs complete */ 7370 call_rcu(&tg->rcu, free_sched_group_rcu); 7371} 7372 7373void sched_offline_group(struct task_group *tg) 7374{ 7375 unsigned long flags; 7376 int i; 7377 7378 /* end participation in shares distribution */ 7379 for_each_possible_cpu(i) 7380 unregister_fair_sched_group(tg, i); 7381 7382 spin_lock_irqsave(&task_group_lock, flags); 7383 list_del_rcu(&tg->list); 7384 list_del_rcu(&tg->siblings); 7385 spin_unlock_irqrestore(&task_group_lock, flags); 7386} 7387 7388/* change task's runqueue when it moves between groups. 7389 * The caller of this function should have put the task in its new group 7390 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to 7391 * reflect its new group. 7392 */ 7393void sched_move_task(struct task_struct *tsk) 7394{ 7395 struct task_group *tg; 7396 int queued, running; 7397 unsigned long flags; 7398 struct rq *rq; 7399 7400 rq = task_rq_lock(tsk, &flags); 7401 7402 running = task_current(rq, tsk); 7403 queued = task_on_rq_queued(tsk); 7404 7405 if (queued) 7406 dequeue_task(rq, tsk, 0); 7407 if (unlikely(running)) 7408 put_prev_task(rq, tsk); 7409 7410 tg = container_of(task_css_check(tsk, cpu_cgrp_id, 7411 lockdep_is_held(&tsk->sighand->siglock)), 7412 struct task_group, css); 7413 tg = autogroup_task_group(tsk, tg); 7414 tsk->sched_task_group = tg; 7415 7416#ifdef CONFIG_FAIR_GROUP_SCHED 7417 if (tsk->sched_class->task_move_group) 7418 tsk->sched_class->task_move_group(tsk, queued); 7419 else 7420#endif 7421 set_task_rq(tsk, task_cpu(tsk)); 7422 7423 if (unlikely(running)) 7424 tsk->sched_class->set_curr_task(rq); 7425 if (queued) 7426 enqueue_task(rq, tsk, 0); 7427 7428 task_rq_unlock(rq, tsk, &flags); 7429} 7430#endif /* CONFIG_CGROUP_SCHED */ 7431 7432#ifdef CONFIG_RT_GROUP_SCHED 7433/* 7434 * Ensure that the real time constraints are schedulable. 7435 */ 7436static DEFINE_MUTEX(rt_constraints_mutex); 7437 7438/* Must be called with tasklist_lock held */ 7439static inline int tg_has_rt_tasks(struct task_group *tg) 7440{ 7441 struct task_struct *g, *p; 7442 7443 for_each_process_thread(g, p) { 7444 if (rt_task(p) && task_group(p) == tg) 7445 return 1; 7446 } 7447 7448 return 0; 7449} 7450 7451struct rt_schedulable_data { 7452 struct task_group *tg; 7453 u64 rt_period; 7454 u64 rt_runtime; 7455}; 7456 7457static int tg_rt_schedulable(struct task_group *tg, void *data) 7458{ 7459 struct rt_schedulable_data *d = data; 7460 struct task_group *child; 7461 unsigned long total, sum = 0; 7462 u64 period, runtime; 7463 7464 period = ktime_to_ns(tg->rt_bandwidth.rt_period); 7465 runtime = tg->rt_bandwidth.rt_runtime; 7466 7467 if (tg == d->tg) { 7468 period = d->rt_period; 7469 runtime = d->rt_runtime; 7470 } 7471 7472 /* 7473 * Cannot have more runtime than the period. 7474 */ 7475 if (runtime > period && runtime != RUNTIME_INF) 7476 return -EINVAL; 7477 7478 /* 7479 * Ensure we don't starve existing RT tasks. 7480 */ 7481 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg)) 7482 return -EBUSY; 7483 7484 total = to_ratio(period, runtime); 7485 7486 /* 7487 * Nobody can have more than the global setting allows. 7488 */ 7489 if (total > to_ratio(global_rt_period(), global_rt_runtime())) 7490 return -EINVAL; 7491 7492 /* 7493 * The sum of our children's runtime should not exceed our own. 7494 */ 7495 list_for_each_entry_rcu(child, &tg->children, siblings) { 7496 period = ktime_to_ns(child->rt_bandwidth.rt_period); 7497 runtime = child->rt_bandwidth.rt_runtime; 7498 7499 if (child == d->tg) { 7500 period = d->rt_period; 7501 runtime = d->rt_runtime; 7502 } 7503 7504 sum += to_ratio(period, runtime); 7505 } 7506 7507 if (sum > total) 7508 return -EINVAL; 7509 7510 return 0; 7511} 7512 7513static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime) 7514{ 7515 int ret; 7516 7517 struct rt_schedulable_data data = { 7518 .tg = tg, 7519 .rt_period = period, 7520 .rt_runtime = runtime, 7521 }; 7522 7523 rcu_read_lock(); 7524 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data); 7525 rcu_read_unlock(); 7526 7527 return ret; 7528} 7529 7530static int tg_set_rt_bandwidth(struct task_group *tg, 7531 u64 rt_period, u64 rt_runtime) 7532{ 7533 int i, err = 0; 7534 7535 mutex_lock(&rt_constraints_mutex); 7536 read_lock(&tasklist_lock); 7537 err = __rt_schedulable(tg, rt_period, rt_runtime); 7538 if (err) 7539 goto unlock; 7540 7541 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock); 7542 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period); 7543 tg->rt_bandwidth.rt_runtime = rt_runtime; 7544 7545 for_each_possible_cpu(i) { 7546 struct rt_rq *rt_rq = tg->rt_rq[i]; 7547 7548 raw_spin_lock(&rt_rq->rt_runtime_lock); 7549 rt_rq->rt_runtime = rt_runtime; 7550 raw_spin_unlock(&rt_rq->rt_runtime_lock); 7551 } 7552 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock); 7553unlock: 7554 read_unlock(&tasklist_lock); 7555 mutex_unlock(&rt_constraints_mutex); 7556 7557 return err; 7558} 7559 7560static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us) 7561{ 7562 u64 rt_runtime, rt_period; 7563 7564 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period); 7565 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC; 7566 if (rt_runtime_us < 0) 7567 rt_runtime = RUNTIME_INF; 7568 7569 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 7570} 7571 7572static long sched_group_rt_runtime(struct task_group *tg) 7573{ 7574 u64 rt_runtime_us; 7575 7576 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF) 7577 return -1; 7578 7579 rt_runtime_us = tg->rt_bandwidth.rt_runtime; 7580 do_div(rt_runtime_us, NSEC_PER_USEC); 7581 return rt_runtime_us; 7582} 7583 7584static int sched_group_set_rt_period(struct task_group *tg, long rt_period_us) 7585{ 7586 u64 rt_runtime, rt_period; 7587 7588 rt_period = (u64)rt_period_us * NSEC_PER_USEC; 7589 rt_runtime = tg->rt_bandwidth.rt_runtime; 7590 7591 if (rt_period == 0) 7592 return -EINVAL; 7593 7594 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 7595} 7596 7597static long sched_group_rt_period(struct task_group *tg) 7598{ 7599 u64 rt_period_us; 7600 7601 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period); 7602 do_div(rt_period_us, NSEC_PER_USEC); 7603 return rt_period_us; 7604} 7605#endif /* CONFIG_RT_GROUP_SCHED */ 7606 7607#ifdef CONFIG_RT_GROUP_SCHED 7608static int sched_rt_global_constraints(void) 7609{ 7610 int ret = 0; 7611 7612 mutex_lock(&rt_constraints_mutex); 7613 read_lock(&tasklist_lock); 7614 ret = __rt_schedulable(NULL, 0, 0); 7615 read_unlock(&tasklist_lock); 7616 mutex_unlock(&rt_constraints_mutex); 7617 7618 return ret; 7619} 7620 7621static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk) 7622{ 7623 /* Don't accept realtime tasks when there is no way for them to run */ 7624 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0) 7625 return 0; 7626 7627 return 1; 7628} 7629 7630#else /* !CONFIG_RT_GROUP_SCHED */ 7631static int sched_rt_global_constraints(void) 7632{ 7633 unsigned long flags; 7634 int i, ret = 0; 7635 7636 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags); 7637 for_each_possible_cpu(i) { 7638 struct rt_rq *rt_rq = &cpu_rq(i)->rt; 7639 7640 raw_spin_lock(&rt_rq->rt_runtime_lock); 7641 rt_rq->rt_runtime = global_rt_runtime(); 7642 raw_spin_unlock(&rt_rq->rt_runtime_lock); 7643 } 7644 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags); 7645 7646 return ret; 7647} 7648#endif /* CONFIG_RT_GROUP_SCHED */ 7649 7650static int sched_dl_global_constraints(void) 7651{ 7652 u64 runtime = global_rt_runtime(); 7653 u64 period = global_rt_period(); 7654 u64 new_bw = to_ratio(period, runtime); 7655 struct dl_bw *dl_b; 7656 int cpu, ret = 0; 7657 unsigned long flags; 7658 7659 /* 7660 * Here we want to check the bandwidth not being set to some 7661 * value smaller than the currently allocated bandwidth in 7662 * any of the root_domains. 7663 * 7664 * FIXME: Cycling on all the CPUs is overdoing, but simpler than 7665 * cycling on root_domains... Discussion on different/better 7666 * solutions is welcome! 7667 */ 7668 for_each_possible_cpu(cpu) { 7669 rcu_read_lock_sched(); 7670 dl_b = dl_bw_of(cpu); 7671 7672 raw_spin_lock_irqsave(&dl_b->lock, flags); 7673 if (new_bw < dl_b->total_bw) 7674 ret = -EBUSY; 7675 raw_spin_unlock_irqrestore(&dl_b->lock, flags); 7676 7677 rcu_read_unlock_sched(); 7678 7679 if (ret) 7680 break; 7681 } 7682 7683 return ret; 7684} 7685 7686static void sched_dl_do_global(void) 7687{ 7688 u64 new_bw = -1; 7689 struct dl_bw *dl_b; 7690 int cpu; 7691 unsigned long flags; 7692 7693 def_dl_bandwidth.dl_period = global_rt_period(); 7694 def_dl_bandwidth.dl_runtime = global_rt_runtime(); 7695 7696 if (global_rt_runtime() != RUNTIME_INF) 7697 new_bw = to_ratio(global_rt_period(), global_rt_runtime()); 7698 7699 /* 7700 * FIXME: As above... 7701 */ 7702 for_each_possible_cpu(cpu) { 7703 rcu_read_lock_sched(); 7704 dl_b = dl_bw_of(cpu); 7705 7706 raw_spin_lock_irqsave(&dl_b->lock, flags); 7707 dl_b->bw = new_bw; 7708 raw_spin_unlock_irqrestore(&dl_b->lock, flags); 7709 7710 rcu_read_unlock_sched(); 7711 } 7712} 7713 7714static int sched_rt_global_validate(void) 7715{ 7716 if (sysctl_sched_rt_period <= 0) 7717 return -EINVAL; 7718 7719 if ((sysctl_sched_rt_runtime != RUNTIME_INF) && 7720 (sysctl_sched_rt_runtime > sysctl_sched_rt_period)) 7721 return -EINVAL; 7722 7723 return 0; 7724} 7725 7726static void sched_rt_do_global(void) 7727{ 7728 def_rt_bandwidth.rt_runtime = global_rt_runtime(); 7729 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period()); 7730} 7731 7732int sched_rt_handler(struct ctl_table *table, int write, 7733 void __user *buffer, size_t *lenp, 7734 loff_t *ppos) 7735{ 7736 int old_period, old_runtime; 7737 static DEFINE_MUTEX(mutex); 7738 int ret; 7739 7740 mutex_lock(&mutex); 7741 old_period = sysctl_sched_rt_period; 7742 old_runtime = sysctl_sched_rt_runtime; 7743 7744 ret = proc_dointvec(table, write, buffer, lenp, ppos); 7745 7746 if (!ret && write) { 7747 ret = sched_rt_global_validate(); 7748 if (ret) 7749 goto undo; 7750 7751 ret = sched_rt_global_constraints(); 7752 if (ret) 7753 goto undo; 7754 7755 ret = sched_dl_global_constraints(); 7756 if (ret) 7757 goto undo; 7758 7759 sched_rt_do_global(); 7760 sched_dl_do_global(); 7761 } 7762 if (0) { 7763undo: 7764 sysctl_sched_rt_period = old_period; 7765 sysctl_sched_rt_runtime = old_runtime; 7766 } 7767 mutex_unlock(&mutex); 7768 7769 return ret; 7770} 7771 7772int sched_rr_handler(struct ctl_table *table, int write, 7773 void __user *buffer, size_t *lenp, 7774 loff_t *ppos) 7775{ 7776 int ret; 7777 static DEFINE_MUTEX(mutex); 7778 7779 mutex_lock(&mutex); 7780 ret = proc_dointvec(table, write, buffer, lenp, ppos); 7781 /* make sure that internally we keep jiffies */ 7782 /* also, writing zero resets timeslice to default */ 7783 if (!ret && write) { 7784 sched_rr_timeslice = sched_rr_timeslice <= 0 ? 7785 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice); 7786 } 7787 mutex_unlock(&mutex); 7788 return ret; 7789} 7790 7791#ifdef CONFIG_CGROUP_SCHED 7792 7793static inline struct task_group *css_tg(struct cgroup_subsys_state *css) 7794{ 7795 return css ? container_of(css, struct task_group, css) : NULL; 7796} 7797 7798static struct cgroup_subsys_state * 7799cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css) 7800{ 7801 struct task_group *parent = css_tg(parent_css); 7802 struct task_group *tg; 7803 7804 if (!parent) { 7805 /* This is early initialization for the top cgroup */ 7806 return &root_task_group.css; 7807 } 7808 7809 tg = sched_create_group(parent); 7810 if (IS_ERR(tg)) 7811 return ERR_PTR(-ENOMEM); 7812 7813 return &tg->css; 7814} 7815 7816static int cpu_cgroup_css_online(struct cgroup_subsys_state *css) 7817{ 7818 struct task_group *tg = css_tg(css); 7819 struct task_group *parent = css_tg(css->parent); 7820 7821 if (parent) 7822 sched_online_group(tg, parent); 7823 return 0; 7824} 7825 7826static void cpu_cgroup_css_free(struct cgroup_subsys_state *css) 7827{ 7828 struct task_group *tg = css_tg(css); 7829 7830 sched_destroy_group(tg); 7831} 7832 7833static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css) 7834{ 7835 struct task_group *tg = css_tg(css); 7836 7837 sched_offline_group(tg); 7838} 7839 7840static int cpu_cgroup_can_attach(struct cgroup_subsys_state *css, 7841 struct cgroup_taskset *tset) 7842{ 7843 struct task_struct *task; 7844 7845 cgroup_taskset_for_each(task, tset) { 7846#ifdef CONFIG_RT_GROUP_SCHED 7847 if (!sched_rt_can_attach(css_tg(css), task)) 7848 return -EINVAL; 7849#else 7850 /* We don't support RT-tasks being in separate groups */ 7851 if (task->sched_class != &fair_sched_class) 7852 return -EINVAL; 7853#endif 7854 } 7855 return 0; 7856} 7857 7858static void cpu_cgroup_attach(struct cgroup_subsys_state *css, 7859 struct cgroup_taskset *tset) 7860{ 7861 struct task_struct *task; 7862 7863 cgroup_taskset_for_each(task, tset) 7864 sched_move_task(task); 7865} 7866 7867static void cpu_cgroup_exit(struct cgroup_subsys_state *css, 7868 struct cgroup_subsys_state *old_css, 7869 struct task_struct *task) 7870{ 7871 /* 7872 * cgroup_exit() is called in the copy_process() failure path. 7873 * Ignore this case since the task hasn't ran yet, this avoids 7874 * trying to poke a half freed task state from generic code. 7875 */ 7876 if (!(task->flags & PF_EXITING)) 7877 return; 7878 7879 sched_move_task(task); 7880} 7881 7882#ifdef CONFIG_FAIR_GROUP_SCHED 7883static int cpu_shares_write_u64(struct cgroup_subsys_state *css, 7884 struct cftype *cftype, u64 shareval) 7885{ 7886 return sched_group_set_shares(css_tg(css), scale_load(shareval)); 7887} 7888 7889static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css, 7890 struct cftype *cft) 7891{ 7892 struct task_group *tg = css_tg(css); 7893 7894 return (u64) scale_load_down(tg->shares); 7895} 7896 7897#ifdef CONFIG_CFS_BANDWIDTH 7898static DEFINE_MUTEX(cfs_constraints_mutex); 7899 7900const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */ 7901const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */ 7902 7903static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime); 7904 7905static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota) 7906{ 7907 int i, ret = 0, runtime_enabled, runtime_was_enabled; 7908 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7909 7910 if (tg == &root_task_group) 7911 return -EINVAL; 7912 7913 /* 7914 * Ensure we have at some amount of bandwidth every period. This is 7915 * to prevent reaching a state of large arrears when throttled via 7916 * entity_tick() resulting in prolonged exit starvation. 7917 */ 7918 if (quota < min_cfs_quota_period || period < min_cfs_quota_period) 7919 return -EINVAL; 7920 7921 /* 7922 * Likewise, bound things on the otherside by preventing insane quota 7923 * periods. This also allows us to normalize in computing quota 7924 * feasibility. 7925 */ 7926 if (period > max_cfs_quota_period) 7927 return -EINVAL; 7928 7929 /* 7930 * Prevent race between setting of cfs_rq->runtime_enabled and 7931 * unthrottle_offline_cfs_rqs(). 7932 */ 7933 get_online_cpus(); 7934 mutex_lock(&cfs_constraints_mutex); 7935 ret = __cfs_schedulable(tg, period, quota); 7936 if (ret) 7937 goto out_unlock; 7938 7939 runtime_enabled = quota != RUNTIME_INF; 7940 runtime_was_enabled = cfs_b->quota != RUNTIME_INF; 7941 /* 7942 * If we need to toggle cfs_bandwidth_used, off->on must occur 7943 * before making related changes, and on->off must occur afterwards 7944 */ 7945 if (runtime_enabled && !runtime_was_enabled) 7946 cfs_bandwidth_usage_inc(); 7947 raw_spin_lock_irq(&cfs_b->lock); 7948 cfs_b->period = ns_to_ktime(period); 7949 cfs_b->quota = quota; 7950 7951 __refill_cfs_bandwidth_runtime(cfs_b); 7952 /* restart the period timer (if active) to handle new period expiry */ 7953 if (runtime_enabled && cfs_b->timer_active) { 7954 /* force a reprogram */ 7955 __start_cfs_bandwidth(cfs_b, true); 7956 } 7957 raw_spin_unlock_irq(&cfs_b->lock); 7958 7959 for_each_online_cpu(i) { 7960 struct cfs_rq *cfs_rq = tg->cfs_rq[i]; 7961 struct rq *rq = cfs_rq->rq; 7962 7963 raw_spin_lock_irq(&rq->lock); 7964 cfs_rq->runtime_enabled = runtime_enabled; 7965 cfs_rq->runtime_remaining = 0; 7966 7967 if (cfs_rq->throttled) 7968 unthrottle_cfs_rq(cfs_rq); 7969 raw_spin_unlock_irq(&rq->lock); 7970 } 7971 if (runtime_was_enabled && !runtime_enabled) 7972 cfs_bandwidth_usage_dec(); 7973out_unlock: 7974 mutex_unlock(&cfs_constraints_mutex); 7975 put_online_cpus(); 7976 7977 return ret; 7978} 7979 7980int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us) 7981{ 7982 u64 quota, period; 7983 7984 period = ktime_to_ns(tg->cfs_bandwidth.period); 7985 if (cfs_quota_us < 0) 7986 quota = RUNTIME_INF; 7987 else 7988 quota = (u64)cfs_quota_us * NSEC_PER_USEC; 7989 7990 return tg_set_cfs_bandwidth(tg, period, quota); 7991} 7992 7993long tg_get_cfs_quota(struct task_group *tg) 7994{ 7995 u64 quota_us; 7996 7997 if (tg->cfs_bandwidth.quota == RUNTIME_INF) 7998 return -1; 7999 8000 quota_us = tg->cfs_bandwidth.quota; 8001 do_div(quota_us, NSEC_PER_USEC); 8002 8003 return quota_us; 8004} 8005 8006int tg_set_cfs_period(struct task_group *tg, long cfs_period_us) 8007{ 8008 u64 quota, period; 8009 8010 period = (u64)cfs_period_us * NSEC_PER_USEC; 8011 quota = tg->cfs_bandwidth.quota; 8012 8013 return tg_set_cfs_bandwidth(tg, period, quota); 8014} 8015 8016long tg_get_cfs_period(struct task_group *tg) 8017{ 8018 u64 cfs_period_us; 8019 8020 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period); 8021 do_div(cfs_period_us, NSEC_PER_USEC); 8022 8023 return cfs_period_us; 8024} 8025 8026static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css, 8027 struct cftype *cft) 8028{ 8029 return tg_get_cfs_quota(css_tg(css)); 8030} 8031 8032static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css, 8033 struct cftype *cftype, s64 cfs_quota_us) 8034{ 8035 return tg_set_cfs_quota(css_tg(css), cfs_quota_us); 8036} 8037 8038static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css, 8039 struct cftype *cft) 8040{ 8041 return tg_get_cfs_period(css_tg(css)); 8042} 8043 8044static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css, 8045 struct cftype *cftype, u64 cfs_period_us) 8046{ 8047 return tg_set_cfs_period(css_tg(css), cfs_period_us); 8048} 8049 8050struct cfs_schedulable_data { 8051 struct task_group *tg; 8052 u64 period, quota; 8053}; 8054 8055/* 8056 * normalize group quota/period to be quota/max_period 8057 * note: units are usecs 8058 */ 8059static u64 normalize_cfs_quota(struct task_group *tg, 8060 struct cfs_schedulable_data *d) 8061{ 8062 u64 quota, period; 8063 8064 if (tg == d->tg) { 8065 period = d->period; 8066 quota = d->quota; 8067 } else { 8068 period = tg_get_cfs_period(tg); 8069 quota = tg_get_cfs_quota(tg); 8070 } 8071 8072 /* note: these should typically be equivalent */ 8073 if (quota == RUNTIME_INF || quota == -1) 8074 return RUNTIME_INF; 8075 8076 return to_ratio(period, quota); 8077} 8078 8079static int tg_cfs_schedulable_down(struct task_group *tg, void *data) 8080{ 8081 struct cfs_schedulable_data *d = data; 8082 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 8083 s64 quota = 0, parent_quota = -1; 8084 8085 if (!tg->parent) { 8086 quota = RUNTIME_INF; 8087 } else { 8088 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth; 8089 8090 quota = normalize_cfs_quota(tg, d); 8091 parent_quota = parent_b->hierarchical_quota; 8092 8093 /* 8094 * ensure max(child_quota) <= parent_quota, inherit when no 8095 * limit is set 8096 */ 8097 if (quota == RUNTIME_INF) 8098 quota = parent_quota; 8099 else if (parent_quota != RUNTIME_INF && quota > parent_quota) 8100 return -EINVAL; 8101 } 8102 cfs_b->hierarchical_quota = quota; 8103 8104 return 0; 8105} 8106 8107static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota) 8108{ 8109 int ret; 8110 struct cfs_schedulable_data data = { 8111 .tg = tg, 8112 .period = period, 8113 .quota = quota, 8114 }; 8115 8116 if (quota != RUNTIME_INF) { 8117 do_div(data.period, NSEC_PER_USEC); 8118 do_div(data.quota, NSEC_PER_USEC); 8119 } 8120 8121 rcu_read_lock(); 8122 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data); 8123 rcu_read_unlock(); 8124 8125 return ret; 8126} 8127 8128static int cpu_stats_show(struct seq_file *sf, void *v) 8129{ 8130 struct task_group *tg = css_tg(seq_css(sf)); 8131 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 8132 8133 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods); 8134 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled); 8135 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time); 8136 8137 return 0; 8138} 8139#endif /* CONFIG_CFS_BANDWIDTH */ 8140#endif /* CONFIG_FAIR_GROUP_SCHED */ 8141 8142#ifdef CONFIG_RT_GROUP_SCHED 8143static int cpu_rt_runtime_write(struct cgroup_subsys_state *css, 8144 struct cftype *cft, s64 val) 8145{ 8146 return sched_group_set_rt_runtime(css_tg(css), val); 8147} 8148 8149static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css, 8150 struct cftype *cft) 8151{ 8152 return sched_group_rt_runtime(css_tg(css)); 8153} 8154 8155static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css, 8156 struct cftype *cftype, u64 rt_period_us) 8157{ 8158 return sched_group_set_rt_period(css_tg(css), rt_period_us); 8159} 8160 8161static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css, 8162 struct cftype *cft) 8163{ 8164 return sched_group_rt_period(css_tg(css)); 8165} 8166#endif /* CONFIG_RT_GROUP_SCHED */ 8167 8168static struct cftype cpu_files[] = { 8169#ifdef CONFIG_FAIR_GROUP_SCHED 8170 { 8171 .name = "shares", 8172 .read_u64 = cpu_shares_read_u64, 8173 .write_u64 = cpu_shares_write_u64, 8174 }, 8175#endif 8176#ifdef CONFIG_CFS_BANDWIDTH 8177 { 8178 .name = "cfs_quota_us", 8179 .read_s64 = cpu_cfs_quota_read_s64, 8180 .write_s64 = cpu_cfs_quota_write_s64, 8181 }, 8182 { 8183 .name = "cfs_period_us", 8184 .read_u64 = cpu_cfs_period_read_u64, 8185 .write_u64 = cpu_cfs_period_write_u64, 8186 }, 8187 { 8188 .name = "stat", 8189 .seq_show = cpu_stats_show, 8190 }, 8191#endif 8192#ifdef CONFIG_RT_GROUP_SCHED 8193 { 8194 .name = "rt_runtime_us", 8195 .read_s64 = cpu_rt_runtime_read, 8196 .write_s64 = cpu_rt_runtime_write, 8197 }, 8198 { 8199 .name = "rt_period_us", 8200 .read_u64 = cpu_rt_period_read_uint, 8201 .write_u64 = cpu_rt_period_write_uint, 8202 }, 8203#endif 8204 { } /* terminate */ 8205}; 8206 8207struct cgroup_subsys cpu_cgrp_subsys = { 8208 .css_alloc = cpu_cgroup_css_alloc, 8209 .css_free = cpu_cgroup_css_free, 8210 .css_online = cpu_cgroup_css_online, 8211 .css_offline = cpu_cgroup_css_offline, 8212 .can_attach = cpu_cgroup_can_attach, 8213 .attach = cpu_cgroup_attach, 8214 .exit = cpu_cgroup_exit, 8215 .legacy_cftypes = cpu_files, 8216 .early_init = 1, 8217}; 8218 8219#endif /* CONFIG_CGROUP_SCHED */ 8220 8221void dump_cpu_task(int cpu) 8222{ 8223 pr_info("Task dump for CPU %d:\n", cpu); 8224 sched_show_task(cpu_curr(cpu)); 8225} 8226