core.c revision 5bfd126e80dca70431aef8fdbc1cf14535f3c338
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 93void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period) 94{ 95 unsigned long delta; 96 ktime_t soft, hard, now; 97 98 for (;;) { 99 if (hrtimer_active(period_timer)) 100 break; 101 102 now = hrtimer_cb_get_time(period_timer); 103 hrtimer_forward(period_timer, now, period); 104 105 soft = hrtimer_get_softexpires(period_timer); 106 hard = hrtimer_get_expires(period_timer); 107 delta = ktime_to_ns(ktime_sub(hard, soft)); 108 __hrtimer_start_range_ns(period_timer, soft, delta, 109 HRTIMER_MODE_ABS_PINNED, 0); 110 } 111} 112 113DEFINE_MUTEX(sched_domains_mutex); 114DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues); 115 116static void update_rq_clock_task(struct rq *rq, s64 delta); 117 118void update_rq_clock(struct rq *rq) 119{ 120 s64 delta; 121 122 if (rq->skip_clock_update > 0) 123 return; 124 125 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock; 126 rq->clock += delta; 127 update_rq_clock_task(rq, delta); 128} 129 130/* 131 * Debugging: various feature bits 132 */ 133 134#define SCHED_FEAT(name, enabled) \ 135 (1UL << __SCHED_FEAT_##name) * enabled | 136 137const_debug unsigned int sysctl_sched_features = 138#include "features.h" 139 0; 140 141#undef SCHED_FEAT 142 143#ifdef CONFIG_SCHED_DEBUG 144#define SCHED_FEAT(name, enabled) \ 145 #name , 146 147static const char * const sched_feat_names[] = { 148#include "features.h" 149}; 150 151#undef SCHED_FEAT 152 153static int sched_feat_show(struct seq_file *m, void *v) 154{ 155 int i; 156 157 for (i = 0; i < __SCHED_FEAT_NR; i++) { 158 if (!(sysctl_sched_features & (1UL << i))) 159 seq_puts(m, "NO_"); 160 seq_printf(m, "%s ", sched_feat_names[i]); 161 } 162 seq_puts(m, "\n"); 163 164 return 0; 165} 166 167#ifdef HAVE_JUMP_LABEL 168 169#define jump_label_key__true STATIC_KEY_INIT_TRUE 170#define jump_label_key__false STATIC_KEY_INIT_FALSE 171 172#define SCHED_FEAT(name, enabled) \ 173 jump_label_key__##enabled , 174 175struct static_key sched_feat_keys[__SCHED_FEAT_NR] = { 176#include "features.h" 177}; 178 179#undef SCHED_FEAT 180 181static void sched_feat_disable(int i) 182{ 183 if (static_key_enabled(&sched_feat_keys[i])) 184 static_key_slow_dec(&sched_feat_keys[i]); 185} 186 187static void sched_feat_enable(int i) 188{ 189 if (!static_key_enabled(&sched_feat_keys[i])) 190 static_key_slow_inc(&sched_feat_keys[i]); 191} 192#else 193static void sched_feat_disable(int i) { }; 194static void sched_feat_enable(int i) { }; 195#endif /* HAVE_JUMP_LABEL */ 196 197static int sched_feat_set(char *cmp) 198{ 199 int i; 200 int neg = 0; 201 202 if (strncmp(cmp, "NO_", 3) == 0) { 203 neg = 1; 204 cmp += 3; 205 } 206 207 for (i = 0; i < __SCHED_FEAT_NR; i++) { 208 if (strcmp(cmp, sched_feat_names[i]) == 0) { 209 if (neg) { 210 sysctl_sched_features &= ~(1UL << i); 211 sched_feat_disable(i); 212 } else { 213 sysctl_sched_features |= (1UL << i); 214 sched_feat_enable(i); 215 } 216 break; 217 } 218 } 219 220 return i; 221} 222 223static ssize_t 224sched_feat_write(struct file *filp, const char __user *ubuf, 225 size_t cnt, loff_t *ppos) 226{ 227 char buf[64]; 228 char *cmp; 229 int i; 230 231 if (cnt > 63) 232 cnt = 63; 233 234 if (copy_from_user(&buf, ubuf, cnt)) 235 return -EFAULT; 236 237 buf[cnt] = 0; 238 cmp = strstrip(buf); 239 240 i = sched_feat_set(cmp); 241 if (i == __SCHED_FEAT_NR) 242 return -EINVAL; 243 244 *ppos += cnt; 245 246 return cnt; 247} 248 249static int sched_feat_open(struct inode *inode, struct file *filp) 250{ 251 return single_open(filp, sched_feat_show, NULL); 252} 253 254static const struct file_operations sched_feat_fops = { 255 .open = sched_feat_open, 256 .write = sched_feat_write, 257 .read = seq_read, 258 .llseek = seq_lseek, 259 .release = single_release, 260}; 261 262static __init int sched_init_debug(void) 263{ 264 debugfs_create_file("sched_features", 0644, NULL, NULL, 265 &sched_feat_fops); 266 267 return 0; 268} 269late_initcall(sched_init_debug); 270#endif /* CONFIG_SCHED_DEBUG */ 271 272/* 273 * Number of tasks to iterate in a single balance run. 274 * Limited because this is done with IRQs disabled. 275 */ 276const_debug unsigned int sysctl_sched_nr_migrate = 32; 277 278/* 279 * period over which we average the RT time consumption, measured 280 * in ms. 281 * 282 * default: 1s 283 */ 284const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC; 285 286/* 287 * period over which we measure -rt task cpu usage in us. 288 * default: 1s 289 */ 290unsigned int sysctl_sched_rt_period = 1000000; 291 292__read_mostly int scheduler_running; 293 294/* 295 * part of the period that we allow rt tasks to run in us. 296 * default: 0.95s 297 */ 298int sysctl_sched_rt_runtime = 950000; 299 300/* 301 * __task_rq_lock - lock the rq @p resides on. 302 */ 303static inline struct rq *__task_rq_lock(struct task_struct *p) 304 __acquires(rq->lock) 305{ 306 struct rq *rq; 307 308 lockdep_assert_held(&p->pi_lock); 309 310 for (;;) { 311 rq = task_rq(p); 312 raw_spin_lock(&rq->lock); 313 if (likely(rq == task_rq(p))) 314 return rq; 315 raw_spin_unlock(&rq->lock); 316 } 317} 318 319/* 320 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on. 321 */ 322static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags) 323 __acquires(p->pi_lock) 324 __acquires(rq->lock) 325{ 326 struct rq *rq; 327 328 for (;;) { 329 raw_spin_lock_irqsave(&p->pi_lock, *flags); 330 rq = task_rq(p); 331 raw_spin_lock(&rq->lock); 332 if (likely(rq == task_rq(p))) 333 return rq; 334 raw_spin_unlock(&rq->lock); 335 raw_spin_unlock_irqrestore(&p->pi_lock, *flags); 336 } 337} 338 339static void __task_rq_unlock(struct rq *rq) 340 __releases(rq->lock) 341{ 342 raw_spin_unlock(&rq->lock); 343} 344 345static inline void 346task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags) 347 __releases(rq->lock) 348 __releases(p->pi_lock) 349{ 350 raw_spin_unlock(&rq->lock); 351 raw_spin_unlock_irqrestore(&p->pi_lock, *flags); 352} 353 354/* 355 * this_rq_lock - lock this runqueue and disable interrupts. 356 */ 357static struct rq *this_rq_lock(void) 358 __acquires(rq->lock) 359{ 360 struct rq *rq; 361 362 local_irq_disable(); 363 rq = this_rq(); 364 raw_spin_lock(&rq->lock); 365 366 return rq; 367} 368 369#ifdef CONFIG_SCHED_HRTICK 370/* 371 * Use HR-timers to deliver accurate preemption points. 372 */ 373 374static void hrtick_clear(struct rq *rq) 375{ 376 if (hrtimer_active(&rq->hrtick_timer)) 377 hrtimer_cancel(&rq->hrtick_timer); 378} 379 380/* 381 * High-resolution timer tick. 382 * Runs from hardirq context with interrupts disabled. 383 */ 384static enum hrtimer_restart hrtick(struct hrtimer *timer) 385{ 386 struct rq *rq = container_of(timer, struct rq, hrtick_timer); 387 388 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id()); 389 390 raw_spin_lock(&rq->lock); 391 update_rq_clock(rq); 392 rq->curr->sched_class->task_tick(rq, rq->curr, 1); 393 raw_spin_unlock(&rq->lock); 394 395 return HRTIMER_NORESTART; 396} 397 398#ifdef CONFIG_SMP 399 400static int __hrtick_restart(struct rq *rq) 401{ 402 struct hrtimer *timer = &rq->hrtick_timer; 403 ktime_t time = hrtimer_get_softexpires(timer); 404 405 return __hrtimer_start_range_ns(timer, time, 0, HRTIMER_MODE_ABS_PINNED, 0); 406} 407 408/* 409 * called from hardirq (IPI) context 410 */ 411static void __hrtick_start(void *arg) 412{ 413 struct rq *rq = arg; 414 415 raw_spin_lock(&rq->lock); 416 __hrtick_restart(rq); 417 rq->hrtick_csd_pending = 0; 418 raw_spin_unlock(&rq->lock); 419} 420 421/* 422 * Called to set the hrtick timer state. 423 * 424 * called with rq->lock held and irqs disabled 425 */ 426void hrtick_start(struct rq *rq, u64 delay) 427{ 428 struct hrtimer *timer = &rq->hrtick_timer; 429 ktime_t time = ktime_add_ns(timer->base->get_time(), delay); 430 431 hrtimer_set_expires(timer, time); 432 433 if (rq == this_rq()) { 434 __hrtick_restart(rq); 435 } else if (!rq->hrtick_csd_pending) { 436 smp_call_function_single_async(cpu_of(rq), &rq->hrtick_csd); 437 rq->hrtick_csd_pending = 1; 438 } 439} 440 441static int 442hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu) 443{ 444 int cpu = (int)(long)hcpu; 445 446 switch (action) { 447 case CPU_UP_CANCELED: 448 case CPU_UP_CANCELED_FROZEN: 449 case CPU_DOWN_PREPARE: 450 case CPU_DOWN_PREPARE_FROZEN: 451 case CPU_DEAD: 452 case CPU_DEAD_FROZEN: 453 hrtick_clear(cpu_rq(cpu)); 454 return NOTIFY_OK; 455 } 456 457 return NOTIFY_DONE; 458} 459 460static __init void init_hrtick(void) 461{ 462 hotcpu_notifier(hotplug_hrtick, 0); 463} 464#else 465/* 466 * Called to set the hrtick timer state. 467 * 468 * called with rq->lock held and irqs disabled 469 */ 470void hrtick_start(struct rq *rq, u64 delay) 471{ 472 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0, 473 HRTIMER_MODE_REL_PINNED, 0); 474} 475 476static inline void init_hrtick(void) 477{ 478} 479#endif /* CONFIG_SMP */ 480 481static void init_rq_hrtick(struct rq *rq) 482{ 483#ifdef CONFIG_SMP 484 rq->hrtick_csd_pending = 0; 485 486 rq->hrtick_csd.flags = 0; 487 rq->hrtick_csd.func = __hrtick_start; 488 rq->hrtick_csd.info = rq; 489#endif 490 491 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); 492 rq->hrtick_timer.function = hrtick; 493} 494#else /* CONFIG_SCHED_HRTICK */ 495static inline void hrtick_clear(struct rq *rq) 496{ 497} 498 499static inline void init_rq_hrtick(struct rq *rq) 500{ 501} 502 503static inline void init_hrtick(void) 504{ 505} 506#endif /* CONFIG_SCHED_HRTICK */ 507 508/* 509 * resched_task - mark a task 'to be rescheduled now'. 510 * 511 * On UP this means the setting of the need_resched flag, on SMP it 512 * might also involve a cross-CPU call to trigger the scheduler on 513 * the target CPU. 514 */ 515void resched_task(struct task_struct *p) 516{ 517 int cpu; 518 519 lockdep_assert_held(&task_rq(p)->lock); 520 521 if (test_tsk_need_resched(p)) 522 return; 523 524 set_tsk_need_resched(p); 525 526 cpu = task_cpu(p); 527 if (cpu == smp_processor_id()) { 528 set_preempt_need_resched(); 529 return; 530 } 531 532 /* NEED_RESCHED must be visible before we test polling */ 533 smp_mb(); 534 if (!tsk_is_polling(p)) 535 smp_send_reschedule(cpu); 536} 537 538void resched_cpu(int cpu) 539{ 540 struct rq *rq = cpu_rq(cpu); 541 unsigned long flags; 542 543 if (!raw_spin_trylock_irqsave(&rq->lock, flags)) 544 return; 545 resched_task(cpu_curr(cpu)); 546 raw_spin_unlock_irqrestore(&rq->lock, flags); 547} 548 549#ifdef CONFIG_SMP 550#ifdef CONFIG_NO_HZ_COMMON 551/* 552 * In the semi idle case, use the nearest busy cpu for migrating timers 553 * from an idle cpu. This is good for power-savings. 554 * 555 * We don't do similar optimization for completely idle system, as 556 * selecting an idle cpu will add more delays to the timers than intended 557 * (as that cpu's timer base may not be uptodate wrt jiffies etc). 558 */ 559int get_nohz_timer_target(int pinned) 560{ 561 int cpu = smp_processor_id(); 562 int i; 563 struct sched_domain *sd; 564 565 if (pinned || !get_sysctl_timer_migration() || !idle_cpu(cpu)) 566 return cpu; 567 568 rcu_read_lock(); 569 for_each_domain(cpu, sd) { 570 for_each_cpu(i, sched_domain_span(sd)) { 571 if (!idle_cpu(i)) { 572 cpu = i; 573 goto unlock; 574 } 575 } 576 } 577unlock: 578 rcu_read_unlock(); 579 return cpu; 580} 581/* 582 * When add_timer_on() enqueues a timer into the timer wheel of an 583 * idle CPU then this timer might expire before the next timer event 584 * which is scheduled to wake up that CPU. In case of a completely 585 * idle system the next event might even be infinite time into the 586 * future. wake_up_idle_cpu() ensures that the CPU is woken up and 587 * leaves the inner idle loop so the newly added timer is taken into 588 * account when the CPU goes back to idle and evaluates the timer 589 * wheel for the next timer event. 590 */ 591static void wake_up_idle_cpu(int cpu) 592{ 593 struct rq *rq = cpu_rq(cpu); 594 595 if (cpu == smp_processor_id()) 596 return; 597 598 /* 599 * This is safe, as this function is called with the timer 600 * wheel base lock of (cpu) held. When the CPU is on the way 601 * to idle and has not yet set rq->curr to idle then it will 602 * be serialized on the timer wheel base lock and take the new 603 * timer into account automatically. 604 */ 605 if (rq->curr != rq->idle) 606 return; 607 608 /* 609 * We can set TIF_RESCHED on the idle task of the other CPU 610 * lockless. The worst case is that the other CPU runs the 611 * idle task through an additional NOOP schedule() 612 */ 613 set_tsk_need_resched(rq->idle); 614 615 /* NEED_RESCHED must be visible before we test polling */ 616 smp_mb(); 617 if (!tsk_is_polling(rq->idle)) 618 smp_send_reschedule(cpu); 619} 620 621static bool wake_up_full_nohz_cpu(int cpu) 622{ 623 if (tick_nohz_full_cpu(cpu)) { 624 if (cpu != smp_processor_id() || 625 tick_nohz_tick_stopped()) 626 smp_send_reschedule(cpu); 627 return true; 628 } 629 630 return false; 631} 632 633void wake_up_nohz_cpu(int cpu) 634{ 635 if (!wake_up_full_nohz_cpu(cpu)) 636 wake_up_idle_cpu(cpu); 637} 638 639static inline bool got_nohz_idle_kick(void) 640{ 641 int cpu = smp_processor_id(); 642 643 if (!test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu))) 644 return false; 645 646 if (idle_cpu(cpu) && !need_resched()) 647 return true; 648 649 /* 650 * We can't run Idle Load Balance on this CPU for this time so we 651 * cancel it and clear NOHZ_BALANCE_KICK 652 */ 653 clear_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)); 654 return false; 655} 656 657#else /* CONFIG_NO_HZ_COMMON */ 658 659static inline bool got_nohz_idle_kick(void) 660{ 661 return false; 662} 663 664#endif /* CONFIG_NO_HZ_COMMON */ 665 666#ifdef CONFIG_NO_HZ_FULL 667bool sched_can_stop_tick(void) 668{ 669 struct rq *rq; 670 671 rq = this_rq(); 672 673 /* Make sure rq->nr_running update is visible after the IPI */ 674 smp_rmb(); 675 676 /* More than one running task need preemption */ 677 if (rq->nr_running > 1) 678 return false; 679 680 return true; 681} 682#endif /* CONFIG_NO_HZ_FULL */ 683 684void sched_avg_update(struct rq *rq) 685{ 686 s64 period = sched_avg_period(); 687 688 while ((s64)(rq_clock(rq) - rq->age_stamp) > period) { 689 /* 690 * Inline assembly required to prevent the compiler 691 * optimising this loop into a divmod call. 692 * See __iter_div_u64_rem() for another example of this. 693 */ 694 asm("" : "+rm" (rq->age_stamp)); 695 rq->age_stamp += period; 696 rq->rt_avg /= 2; 697 } 698} 699 700#endif /* CONFIG_SMP */ 701 702#if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \ 703 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH))) 704/* 705 * Iterate task_group tree rooted at *from, calling @down when first entering a 706 * node and @up when leaving it for the final time. 707 * 708 * Caller must hold rcu_lock or sufficient equivalent. 709 */ 710int walk_tg_tree_from(struct task_group *from, 711 tg_visitor down, tg_visitor up, void *data) 712{ 713 struct task_group *parent, *child; 714 int ret; 715 716 parent = from; 717 718down: 719 ret = (*down)(parent, data); 720 if (ret) 721 goto out; 722 list_for_each_entry_rcu(child, &parent->children, siblings) { 723 parent = child; 724 goto down; 725 726up: 727 continue; 728 } 729 ret = (*up)(parent, data); 730 if (ret || parent == from) 731 goto out; 732 733 child = parent; 734 parent = parent->parent; 735 if (parent) 736 goto up; 737out: 738 return ret; 739} 740 741int tg_nop(struct task_group *tg, void *data) 742{ 743 return 0; 744} 745#endif 746 747static void set_load_weight(struct task_struct *p) 748{ 749 int prio = p->static_prio - MAX_RT_PRIO; 750 struct load_weight *load = &p->se.load; 751 752 /* 753 * SCHED_IDLE tasks get minimal weight: 754 */ 755 if (p->policy == SCHED_IDLE) { 756 load->weight = scale_load(WEIGHT_IDLEPRIO); 757 load->inv_weight = WMULT_IDLEPRIO; 758 return; 759 } 760 761 load->weight = scale_load(prio_to_weight[prio]); 762 load->inv_weight = prio_to_wmult[prio]; 763} 764 765static void enqueue_task(struct rq *rq, struct task_struct *p, int flags) 766{ 767 update_rq_clock(rq); 768 sched_info_queued(rq, p); 769 p->sched_class->enqueue_task(rq, p, flags); 770} 771 772static void dequeue_task(struct rq *rq, struct task_struct *p, int flags) 773{ 774 update_rq_clock(rq); 775 sched_info_dequeued(rq, p); 776 p->sched_class->dequeue_task(rq, p, flags); 777} 778 779void activate_task(struct rq *rq, struct task_struct *p, int flags) 780{ 781 if (task_contributes_to_load(p)) 782 rq->nr_uninterruptible--; 783 784 enqueue_task(rq, p, flags); 785} 786 787void deactivate_task(struct rq *rq, struct task_struct *p, int flags) 788{ 789 if (task_contributes_to_load(p)) 790 rq->nr_uninterruptible++; 791 792 dequeue_task(rq, p, flags); 793} 794 795static void update_rq_clock_task(struct rq *rq, s64 delta) 796{ 797/* 798 * In theory, the compile should just see 0 here, and optimize out the call 799 * to sched_rt_avg_update. But I don't trust it... 800 */ 801#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING) 802 s64 steal = 0, irq_delta = 0; 803#endif 804#ifdef CONFIG_IRQ_TIME_ACCOUNTING 805 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time; 806 807 /* 808 * Since irq_time is only updated on {soft,}irq_exit, we might run into 809 * this case when a previous update_rq_clock() happened inside a 810 * {soft,}irq region. 811 * 812 * When this happens, we stop ->clock_task and only update the 813 * prev_irq_time stamp to account for the part that fit, so that a next 814 * update will consume the rest. This ensures ->clock_task is 815 * monotonic. 816 * 817 * It does however cause some slight miss-attribution of {soft,}irq 818 * time, a more accurate solution would be to update the irq_time using 819 * the current rq->clock timestamp, except that would require using 820 * atomic ops. 821 */ 822 if (irq_delta > delta) 823 irq_delta = delta; 824 825 rq->prev_irq_time += irq_delta; 826 delta -= irq_delta; 827#endif 828#ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING 829 if (static_key_false((¶virt_steal_rq_enabled))) { 830 steal = paravirt_steal_clock(cpu_of(rq)); 831 steal -= rq->prev_steal_time_rq; 832 833 if (unlikely(steal > delta)) 834 steal = delta; 835 836 rq->prev_steal_time_rq += steal; 837 delta -= steal; 838 } 839#endif 840 841 rq->clock_task += delta; 842 843#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING) 844 if ((irq_delta + steal) && sched_feat(NONTASK_POWER)) 845 sched_rt_avg_update(rq, irq_delta + steal); 846#endif 847} 848 849void sched_set_stop_task(int cpu, struct task_struct *stop) 850{ 851 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 }; 852 struct task_struct *old_stop = cpu_rq(cpu)->stop; 853 854 if (stop) { 855 /* 856 * Make it appear like a SCHED_FIFO task, its something 857 * userspace knows about and won't get confused about. 858 * 859 * Also, it will make PI more or less work without too 860 * much confusion -- but then, stop work should not 861 * rely on PI working anyway. 862 */ 863 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m); 864 865 stop->sched_class = &stop_sched_class; 866 } 867 868 cpu_rq(cpu)->stop = stop; 869 870 if (old_stop) { 871 /* 872 * Reset it back to a normal scheduling class so that 873 * it can die in pieces. 874 */ 875 old_stop->sched_class = &rt_sched_class; 876 } 877} 878 879/* 880 * __normal_prio - return the priority that is based on the static prio 881 */ 882static inline int __normal_prio(struct task_struct *p) 883{ 884 return p->static_prio; 885} 886 887/* 888 * Calculate the expected normal priority: i.e. priority 889 * without taking RT-inheritance into account. Might be 890 * boosted by interactivity modifiers. Changes upon fork, 891 * setprio syscalls, and whenever the interactivity 892 * estimator recalculates. 893 */ 894static inline int normal_prio(struct task_struct *p) 895{ 896 int prio; 897 898 if (task_has_dl_policy(p)) 899 prio = MAX_DL_PRIO-1; 900 else if (task_has_rt_policy(p)) 901 prio = MAX_RT_PRIO-1 - p->rt_priority; 902 else 903 prio = __normal_prio(p); 904 return prio; 905} 906 907/* 908 * Calculate the current priority, i.e. the priority 909 * taken into account by the scheduler. This value might 910 * be boosted by RT tasks, or might be boosted by 911 * interactivity modifiers. Will be RT if the task got 912 * RT-boosted. If not then it returns p->normal_prio. 913 */ 914static int effective_prio(struct task_struct *p) 915{ 916 p->normal_prio = normal_prio(p); 917 /* 918 * If we are RT tasks or we were boosted to RT priority, 919 * keep the priority unchanged. Otherwise, update priority 920 * to the normal priority: 921 */ 922 if (!rt_prio(p->prio)) 923 return p->normal_prio; 924 return p->prio; 925} 926 927/** 928 * task_curr - is this task currently executing on a CPU? 929 * @p: the task in question. 930 * 931 * Return: 1 if the task is currently executing. 0 otherwise. 932 */ 933inline int task_curr(const struct task_struct *p) 934{ 935 return cpu_curr(task_cpu(p)) == p; 936} 937 938static inline void check_class_changed(struct rq *rq, struct task_struct *p, 939 const struct sched_class *prev_class, 940 int oldprio) 941{ 942 if (prev_class != p->sched_class) { 943 if (prev_class->switched_from) 944 prev_class->switched_from(rq, p); 945 p->sched_class->switched_to(rq, p); 946 } else if (oldprio != p->prio || dl_task(p)) 947 p->sched_class->prio_changed(rq, p, oldprio); 948} 949 950void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags) 951{ 952 const struct sched_class *class; 953 954 if (p->sched_class == rq->curr->sched_class) { 955 rq->curr->sched_class->check_preempt_curr(rq, p, flags); 956 } else { 957 for_each_class(class) { 958 if (class == rq->curr->sched_class) 959 break; 960 if (class == p->sched_class) { 961 resched_task(rq->curr); 962 break; 963 } 964 } 965 } 966 967 /* 968 * A queue event has occurred, and we're going to schedule. In 969 * this case, we can save a useless back to back clock update. 970 */ 971 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr)) 972 rq->skip_clock_update = 1; 973} 974 975#ifdef CONFIG_SMP 976void set_task_cpu(struct task_struct *p, unsigned int new_cpu) 977{ 978#ifdef CONFIG_SCHED_DEBUG 979 /* 980 * We should never call set_task_cpu() on a blocked task, 981 * ttwu() will sort out the placement. 982 */ 983 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING && 984 !(task_preempt_count(p) & PREEMPT_ACTIVE)); 985 986#ifdef CONFIG_LOCKDEP 987 /* 988 * The caller should hold either p->pi_lock or rq->lock, when changing 989 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks. 990 * 991 * sched_move_task() holds both and thus holding either pins the cgroup, 992 * see task_group(). 993 * 994 * Furthermore, all task_rq users should acquire both locks, see 995 * task_rq_lock(). 996 */ 997 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) || 998 lockdep_is_held(&task_rq(p)->lock))); 999#endif 1000#endif 1001 1002 trace_sched_migrate_task(p, new_cpu); 1003 1004 if (task_cpu(p) != new_cpu) { 1005 if (p->sched_class->migrate_task_rq) 1006 p->sched_class->migrate_task_rq(p, new_cpu); 1007 p->se.nr_migrations++; 1008 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0); 1009 } 1010 1011 __set_task_cpu(p, new_cpu); 1012} 1013 1014static void __migrate_swap_task(struct task_struct *p, int cpu) 1015{ 1016 if (p->on_rq) { 1017 struct rq *src_rq, *dst_rq; 1018 1019 src_rq = task_rq(p); 1020 dst_rq = cpu_rq(cpu); 1021 1022 deactivate_task(src_rq, p, 0); 1023 set_task_cpu(p, cpu); 1024 activate_task(dst_rq, p, 0); 1025 check_preempt_curr(dst_rq, p, 0); 1026 } else { 1027 /* 1028 * Task isn't running anymore; make it appear like we migrated 1029 * it before it went to sleep. This means on wakeup we make the 1030 * previous cpu our targer instead of where it really is. 1031 */ 1032 p->wake_cpu = cpu; 1033 } 1034} 1035 1036struct migration_swap_arg { 1037 struct task_struct *src_task, *dst_task; 1038 int src_cpu, dst_cpu; 1039}; 1040 1041static int migrate_swap_stop(void *data) 1042{ 1043 struct migration_swap_arg *arg = data; 1044 struct rq *src_rq, *dst_rq; 1045 int ret = -EAGAIN; 1046 1047 src_rq = cpu_rq(arg->src_cpu); 1048 dst_rq = cpu_rq(arg->dst_cpu); 1049 1050 double_raw_lock(&arg->src_task->pi_lock, 1051 &arg->dst_task->pi_lock); 1052 double_rq_lock(src_rq, dst_rq); 1053 if (task_cpu(arg->dst_task) != arg->dst_cpu) 1054 goto unlock; 1055 1056 if (task_cpu(arg->src_task) != arg->src_cpu) 1057 goto unlock; 1058 1059 if (!cpumask_test_cpu(arg->dst_cpu, tsk_cpus_allowed(arg->src_task))) 1060 goto unlock; 1061 1062 if (!cpumask_test_cpu(arg->src_cpu, tsk_cpus_allowed(arg->dst_task))) 1063 goto unlock; 1064 1065 __migrate_swap_task(arg->src_task, arg->dst_cpu); 1066 __migrate_swap_task(arg->dst_task, arg->src_cpu); 1067 1068 ret = 0; 1069 1070unlock: 1071 double_rq_unlock(src_rq, dst_rq); 1072 raw_spin_unlock(&arg->dst_task->pi_lock); 1073 raw_spin_unlock(&arg->src_task->pi_lock); 1074 1075 return ret; 1076} 1077 1078/* 1079 * Cross migrate two tasks 1080 */ 1081int migrate_swap(struct task_struct *cur, struct task_struct *p) 1082{ 1083 struct migration_swap_arg arg; 1084 int ret = -EINVAL; 1085 1086 arg = (struct migration_swap_arg){ 1087 .src_task = cur, 1088 .src_cpu = task_cpu(cur), 1089 .dst_task = p, 1090 .dst_cpu = task_cpu(p), 1091 }; 1092 1093 if (arg.src_cpu == arg.dst_cpu) 1094 goto out; 1095 1096 /* 1097 * These three tests are all lockless; this is OK since all of them 1098 * will be re-checked with proper locks held further down the line. 1099 */ 1100 if (!cpu_active(arg.src_cpu) || !cpu_active(arg.dst_cpu)) 1101 goto out; 1102 1103 if (!cpumask_test_cpu(arg.dst_cpu, tsk_cpus_allowed(arg.src_task))) 1104 goto out; 1105 1106 if (!cpumask_test_cpu(arg.src_cpu, tsk_cpus_allowed(arg.dst_task))) 1107 goto out; 1108 1109 trace_sched_swap_numa(cur, arg.src_cpu, p, arg.dst_cpu); 1110 ret = stop_two_cpus(arg.dst_cpu, arg.src_cpu, migrate_swap_stop, &arg); 1111 1112out: 1113 return ret; 1114} 1115 1116struct migration_arg { 1117 struct task_struct *task; 1118 int dest_cpu; 1119}; 1120 1121static int migration_cpu_stop(void *data); 1122 1123/* 1124 * wait_task_inactive - wait for a thread to unschedule. 1125 * 1126 * If @match_state is nonzero, it's the @p->state value just checked and 1127 * not expected to change. If it changes, i.e. @p might have woken up, 1128 * then return zero. When we succeed in waiting for @p to be off its CPU, 1129 * we return a positive number (its total switch count). If a second call 1130 * a short while later returns the same number, the caller can be sure that 1131 * @p has remained unscheduled the whole time. 1132 * 1133 * The caller must ensure that the task *will* unschedule sometime soon, 1134 * else this function might spin for a *long* time. This function can't 1135 * be called with interrupts off, or it may introduce deadlock with 1136 * smp_call_function() if an IPI is sent by the same process we are 1137 * waiting to become inactive. 1138 */ 1139unsigned long wait_task_inactive(struct task_struct *p, long match_state) 1140{ 1141 unsigned long flags; 1142 int running, on_rq; 1143 unsigned long ncsw; 1144 struct rq *rq; 1145 1146 for (;;) { 1147 /* 1148 * We do the initial early heuristics without holding 1149 * any task-queue locks at all. We'll only try to get 1150 * the runqueue lock when things look like they will 1151 * work out! 1152 */ 1153 rq = task_rq(p); 1154 1155 /* 1156 * If the task is actively running on another CPU 1157 * still, just relax and busy-wait without holding 1158 * any locks. 1159 * 1160 * NOTE! Since we don't hold any locks, it's not 1161 * even sure that "rq" stays as the right runqueue! 1162 * But we don't care, since "task_running()" will 1163 * return false if the runqueue has changed and p 1164 * is actually now running somewhere else! 1165 */ 1166 while (task_running(rq, p)) { 1167 if (match_state && unlikely(p->state != match_state)) 1168 return 0; 1169 cpu_relax(); 1170 } 1171 1172 /* 1173 * Ok, time to look more closely! We need the rq 1174 * lock now, to be *sure*. If we're wrong, we'll 1175 * just go back and repeat. 1176 */ 1177 rq = task_rq_lock(p, &flags); 1178 trace_sched_wait_task(p); 1179 running = task_running(rq, p); 1180 on_rq = p->on_rq; 1181 ncsw = 0; 1182 if (!match_state || p->state == match_state) 1183 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */ 1184 task_rq_unlock(rq, p, &flags); 1185 1186 /* 1187 * If it changed from the expected state, bail out now. 1188 */ 1189 if (unlikely(!ncsw)) 1190 break; 1191 1192 /* 1193 * Was it really running after all now that we 1194 * checked with the proper locks actually held? 1195 * 1196 * Oops. Go back and try again.. 1197 */ 1198 if (unlikely(running)) { 1199 cpu_relax(); 1200 continue; 1201 } 1202 1203 /* 1204 * It's not enough that it's not actively running, 1205 * it must be off the runqueue _entirely_, and not 1206 * preempted! 1207 * 1208 * So if it was still runnable (but just not actively 1209 * running right now), it's preempted, and we should 1210 * yield - it could be a while. 1211 */ 1212 if (unlikely(on_rq)) { 1213 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ); 1214 1215 set_current_state(TASK_UNINTERRUPTIBLE); 1216 schedule_hrtimeout(&to, HRTIMER_MODE_REL); 1217 continue; 1218 } 1219 1220 /* 1221 * Ahh, all good. It wasn't running, and it wasn't 1222 * runnable, which means that it will never become 1223 * running in the future either. We're all done! 1224 */ 1225 break; 1226 } 1227 1228 return ncsw; 1229} 1230 1231/*** 1232 * kick_process - kick a running thread to enter/exit the kernel 1233 * @p: the to-be-kicked thread 1234 * 1235 * Cause a process which is running on another CPU to enter 1236 * kernel-mode, without any delay. (to get signals handled.) 1237 * 1238 * NOTE: this function doesn't have to take the runqueue lock, 1239 * because all it wants to ensure is that the remote task enters 1240 * the kernel. If the IPI races and the task has been migrated 1241 * to another CPU then no harm is done and the purpose has been 1242 * achieved as well. 1243 */ 1244void kick_process(struct task_struct *p) 1245{ 1246 int cpu; 1247 1248 preempt_disable(); 1249 cpu = task_cpu(p); 1250 if ((cpu != smp_processor_id()) && task_curr(p)) 1251 smp_send_reschedule(cpu); 1252 preempt_enable(); 1253} 1254EXPORT_SYMBOL_GPL(kick_process); 1255#endif /* CONFIG_SMP */ 1256 1257#ifdef CONFIG_SMP 1258/* 1259 * ->cpus_allowed is protected by both rq->lock and p->pi_lock 1260 */ 1261static int select_fallback_rq(int cpu, struct task_struct *p) 1262{ 1263 int nid = cpu_to_node(cpu); 1264 const struct cpumask *nodemask = NULL; 1265 enum { cpuset, possible, fail } state = cpuset; 1266 int dest_cpu; 1267 1268 /* 1269 * If the node that the cpu is on has been offlined, cpu_to_node() 1270 * will return -1. There is no cpu on the node, and we should 1271 * select the cpu on the other node. 1272 */ 1273 if (nid != -1) { 1274 nodemask = cpumask_of_node(nid); 1275 1276 /* Look for allowed, online CPU in same node. */ 1277 for_each_cpu(dest_cpu, nodemask) { 1278 if (!cpu_online(dest_cpu)) 1279 continue; 1280 if (!cpu_active(dest_cpu)) 1281 continue; 1282 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) 1283 return dest_cpu; 1284 } 1285 } 1286 1287 for (;;) { 1288 /* Any allowed, online CPU? */ 1289 for_each_cpu(dest_cpu, tsk_cpus_allowed(p)) { 1290 if (!cpu_online(dest_cpu)) 1291 continue; 1292 if (!cpu_active(dest_cpu)) 1293 continue; 1294 goto out; 1295 } 1296 1297 switch (state) { 1298 case cpuset: 1299 /* No more Mr. Nice Guy. */ 1300 cpuset_cpus_allowed_fallback(p); 1301 state = possible; 1302 break; 1303 1304 case possible: 1305 do_set_cpus_allowed(p, cpu_possible_mask); 1306 state = fail; 1307 break; 1308 1309 case fail: 1310 BUG(); 1311 break; 1312 } 1313 } 1314 1315out: 1316 if (state != cpuset) { 1317 /* 1318 * Don't tell them about moving exiting tasks or 1319 * kernel threads (both mm NULL), since they never 1320 * leave kernel. 1321 */ 1322 if (p->mm && printk_ratelimit()) { 1323 printk_sched("process %d (%s) no longer affine to cpu%d\n", 1324 task_pid_nr(p), p->comm, cpu); 1325 } 1326 } 1327 1328 return dest_cpu; 1329} 1330 1331/* 1332 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable. 1333 */ 1334static inline 1335int select_task_rq(struct task_struct *p, int cpu, int sd_flags, int wake_flags) 1336{ 1337 cpu = p->sched_class->select_task_rq(p, cpu, sd_flags, wake_flags); 1338 1339 /* 1340 * In order not to call set_task_cpu() on a blocking task we need 1341 * to rely on ttwu() to place the task on a valid ->cpus_allowed 1342 * cpu. 1343 * 1344 * Since this is common to all placement strategies, this lives here. 1345 * 1346 * [ this allows ->select_task() to simply return task_cpu(p) and 1347 * not worry about this generic constraint ] 1348 */ 1349 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) || 1350 !cpu_online(cpu))) 1351 cpu = select_fallback_rq(task_cpu(p), p); 1352 1353 return cpu; 1354} 1355 1356static void update_avg(u64 *avg, u64 sample) 1357{ 1358 s64 diff = sample - *avg; 1359 *avg += diff >> 3; 1360} 1361#endif 1362 1363static void 1364ttwu_stat(struct task_struct *p, int cpu, int wake_flags) 1365{ 1366#ifdef CONFIG_SCHEDSTATS 1367 struct rq *rq = this_rq(); 1368 1369#ifdef CONFIG_SMP 1370 int this_cpu = smp_processor_id(); 1371 1372 if (cpu == this_cpu) { 1373 schedstat_inc(rq, ttwu_local); 1374 schedstat_inc(p, se.statistics.nr_wakeups_local); 1375 } else { 1376 struct sched_domain *sd; 1377 1378 schedstat_inc(p, se.statistics.nr_wakeups_remote); 1379 rcu_read_lock(); 1380 for_each_domain(this_cpu, sd) { 1381 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) { 1382 schedstat_inc(sd, ttwu_wake_remote); 1383 break; 1384 } 1385 } 1386 rcu_read_unlock(); 1387 } 1388 1389 if (wake_flags & WF_MIGRATED) 1390 schedstat_inc(p, se.statistics.nr_wakeups_migrate); 1391 1392#endif /* CONFIG_SMP */ 1393 1394 schedstat_inc(rq, ttwu_count); 1395 schedstat_inc(p, se.statistics.nr_wakeups); 1396 1397 if (wake_flags & WF_SYNC) 1398 schedstat_inc(p, se.statistics.nr_wakeups_sync); 1399 1400#endif /* CONFIG_SCHEDSTATS */ 1401} 1402 1403static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags) 1404{ 1405 activate_task(rq, p, en_flags); 1406 p->on_rq = 1; 1407 1408 /* if a worker is waking up, notify workqueue */ 1409 if (p->flags & PF_WQ_WORKER) 1410 wq_worker_waking_up(p, cpu_of(rq)); 1411} 1412 1413/* 1414 * Mark the task runnable and perform wakeup-preemption. 1415 */ 1416static void 1417ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags) 1418{ 1419 check_preempt_curr(rq, p, wake_flags); 1420 trace_sched_wakeup(p, true); 1421 1422 p->state = TASK_RUNNING; 1423#ifdef CONFIG_SMP 1424 if (p->sched_class->task_woken) 1425 p->sched_class->task_woken(rq, p); 1426 1427 if (rq->idle_stamp) { 1428 u64 delta = rq_clock(rq) - rq->idle_stamp; 1429 u64 max = 2*rq->max_idle_balance_cost; 1430 1431 update_avg(&rq->avg_idle, delta); 1432 1433 if (rq->avg_idle > max) 1434 rq->avg_idle = max; 1435 1436 rq->idle_stamp = 0; 1437 } 1438#endif 1439} 1440 1441static void 1442ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags) 1443{ 1444#ifdef CONFIG_SMP 1445 if (p->sched_contributes_to_load) 1446 rq->nr_uninterruptible--; 1447#endif 1448 1449 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING); 1450 ttwu_do_wakeup(rq, p, wake_flags); 1451} 1452 1453/* 1454 * Called in case the task @p isn't fully descheduled from its runqueue, 1455 * in this case we must do a remote wakeup. Its a 'light' wakeup though, 1456 * since all we need to do is flip p->state to TASK_RUNNING, since 1457 * the task is still ->on_rq. 1458 */ 1459static int ttwu_remote(struct task_struct *p, int wake_flags) 1460{ 1461 struct rq *rq; 1462 int ret = 0; 1463 1464 rq = __task_rq_lock(p); 1465 if (p->on_rq) { 1466 /* check_preempt_curr() may use rq clock */ 1467 update_rq_clock(rq); 1468 ttwu_do_wakeup(rq, p, wake_flags); 1469 ret = 1; 1470 } 1471 __task_rq_unlock(rq); 1472 1473 return ret; 1474} 1475 1476#ifdef CONFIG_SMP 1477static void sched_ttwu_pending(void) 1478{ 1479 struct rq *rq = this_rq(); 1480 struct llist_node *llist = llist_del_all(&rq->wake_list); 1481 struct task_struct *p; 1482 1483 raw_spin_lock(&rq->lock); 1484 1485 while (llist) { 1486 p = llist_entry(llist, struct task_struct, wake_entry); 1487 llist = llist_next(llist); 1488 ttwu_do_activate(rq, p, 0); 1489 } 1490 1491 raw_spin_unlock(&rq->lock); 1492} 1493 1494void scheduler_ipi(void) 1495{ 1496 /* 1497 * Fold TIF_NEED_RESCHED into the preempt_count; anybody setting 1498 * TIF_NEED_RESCHED remotely (for the first time) will also send 1499 * this IPI. 1500 */ 1501 preempt_fold_need_resched(); 1502 1503 if (llist_empty(&this_rq()->wake_list) 1504 && !tick_nohz_full_cpu(smp_processor_id()) 1505 && !got_nohz_idle_kick()) 1506 return; 1507 1508 /* 1509 * Not all reschedule IPI handlers call irq_enter/irq_exit, since 1510 * traditionally all their work was done from the interrupt return 1511 * path. Now that we actually do some work, we need to make sure 1512 * we do call them. 1513 * 1514 * Some archs already do call them, luckily irq_enter/exit nest 1515 * properly. 1516 * 1517 * Arguably we should visit all archs and update all handlers, 1518 * however a fair share of IPIs are still resched only so this would 1519 * somewhat pessimize the simple resched case. 1520 */ 1521 irq_enter(); 1522 tick_nohz_full_check(); 1523 sched_ttwu_pending(); 1524 1525 /* 1526 * Check if someone kicked us for doing the nohz idle load balance. 1527 */ 1528 if (unlikely(got_nohz_idle_kick())) { 1529 this_rq()->idle_balance = 1; 1530 raise_softirq_irqoff(SCHED_SOFTIRQ); 1531 } 1532 irq_exit(); 1533} 1534 1535static void ttwu_queue_remote(struct task_struct *p, int cpu) 1536{ 1537 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) 1538 smp_send_reschedule(cpu); 1539} 1540 1541bool cpus_share_cache(int this_cpu, int that_cpu) 1542{ 1543 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu); 1544} 1545#endif /* CONFIG_SMP */ 1546 1547static void ttwu_queue(struct task_struct *p, int cpu) 1548{ 1549 struct rq *rq = cpu_rq(cpu); 1550 1551#if defined(CONFIG_SMP) 1552 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) { 1553 sched_clock_cpu(cpu); /* sync clocks x-cpu */ 1554 ttwu_queue_remote(p, cpu); 1555 return; 1556 } 1557#endif 1558 1559 raw_spin_lock(&rq->lock); 1560 ttwu_do_activate(rq, p, 0); 1561 raw_spin_unlock(&rq->lock); 1562} 1563 1564/** 1565 * try_to_wake_up - wake up a thread 1566 * @p: the thread to be awakened 1567 * @state: the mask of task states that can be woken 1568 * @wake_flags: wake modifier flags (WF_*) 1569 * 1570 * Put it on the run-queue if it's not already there. The "current" 1571 * thread is always on the run-queue (except when the actual 1572 * re-schedule is in progress), and as such you're allowed to do 1573 * the simpler "current->state = TASK_RUNNING" to mark yourself 1574 * runnable without the overhead of this. 1575 * 1576 * Return: %true if @p was woken up, %false if it was already running. 1577 * or @state didn't match @p's state. 1578 */ 1579static int 1580try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags) 1581{ 1582 unsigned long flags; 1583 int cpu, success = 0; 1584 1585 /* 1586 * If we are going to wake up a thread waiting for CONDITION we 1587 * need to ensure that CONDITION=1 done by the caller can not be 1588 * reordered with p->state check below. This pairs with mb() in 1589 * set_current_state() the waiting thread does. 1590 */ 1591 smp_mb__before_spinlock(); 1592 raw_spin_lock_irqsave(&p->pi_lock, flags); 1593 if (!(p->state & state)) 1594 goto out; 1595 1596 success = 1; /* we're going to change ->state */ 1597 cpu = task_cpu(p); 1598 1599 if (p->on_rq && ttwu_remote(p, wake_flags)) 1600 goto stat; 1601 1602#ifdef CONFIG_SMP 1603 /* 1604 * If the owning (remote) cpu is still in the middle of schedule() with 1605 * this task as prev, wait until its done referencing the task. 1606 */ 1607 while (p->on_cpu) 1608 cpu_relax(); 1609 /* 1610 * Pairs with the smp_wmb() in finish_lock_switch(). 1611 */ 1612 smp_rmb(); 1613 1614 p->sched_contributes_to_load = !!task_contributes_to_load(p); 1615 p->state = TASK_WAKING; 1616 1617 if (p->sched_class->task_waking) 1618 p->sched_class->task_waking(p); 1619 1620 cpu = select_task_rq(p, p->wake_cpu, SD_BALANCE_WAKE, wake_flags); 1621 if (task_cpu(p) != cpu) { 1622 wake_flags |= WF_MIGRATED; 1623 set_task_cpu(p, cpu); 1624 } 1625#endif /* CONFIG_SMP */ 1626 1627 ttwu_queue(p, cpu); 1628stat: 1629 ttwu_stat(p, cpu, wake_flags); 1630out: 1631 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 1632 1633 return success; 1634} 1635 1636/** 1637 * try_to_wake_up_local - try to wake up a local task with rq lock held 1638 * @p: the thread to be awakened 1639 * 1640 * Put @p on the run-queue if it's not already there. The caller must 1641 * ensure that this_rq() is locked, @p is bound to this_rq() and not 1642 * the current task. 1643 */ 1644static void try_to_wake_up_local(struct task_struct *p) 1645{ 1646 struct rq *rq = task_rq(p); 1647 1648 if (WARN_ON_ONCE(rq != this_rq()) || 1649 WARN_ON_ONCE(p == current)) 1650 return; 1651 1652 lockdep_assert_held(&rq->lock); 1653 1654 if (!raw_spin_trylock(&p->pi_lock)) { 1655 raw_spin_unlock(&rq->lock); 1656 raw_spin_lock(&p->pi_lock); 1657 raw_spin_lock(&rq->lock); 1658 } 1659 1660 if (!(p->state & TASK_NORMAL)) 1661 goto out; 1662 1663 if (!p->on_rq) 1664 ttwu_activate(rq, p, ENQUEUE_WAKEUP); 1665 1666 ttwu_do_wakeup(rq, p, 0); 1667 ttwu_stat(p, smp_processor_id(), 0); 1668out: 1669 raw_spin_unlock(&p->pi_lock); 1670} 1671 1672/** 1673 * wake_up_process - Wake up a specific process 1674 * @p: The process to be woken up. 1675 * 1676 * Attempt to wake up the nominated process and move it to the set of runnable 1677 * processes. 1678 * 1679 * Return: 1 if the process was woken up, 0 if it was already running. 1680 * 1681 * It may be assumed that this function implies a write memory barrier before 1682 * changing the task state if and only if any tasks are woken up. 1683 */ 1684int wake_up_process(struct task_struct *p) 1685{ 1686 WARN_ON(task_is_stopped_or_traced(p)); 1687 return try_to_wake_up(p, TASK_NORMAL, 0); 1688} 1689EXPORT_SYMBOL(wake_up_process); 1690 1691int wake_up_state(struct task_struct *p, unsigned int state) 1692{ 1693 return try_to_wake_up(p, state, 0); 1694} 1695 1696/* 1697 * Perform scheduler related setup for a newly forked process p. 1698 * p is forked by current. 1699 * 1700 * __sched_fork() is basic setup used by init_idle() too: 1701 */ 1702static void __sched_fork(unsigned long clone_flags, struct task_struct *p) 1703{ 1704 p->on_rq = 0; 1705 1706 p->se.on_rq = 0; 1707 p->se.exec_start = 0; 1708 p->se.sum_exec_runtime = 0; 1709 p->se.prev_sum_exec_runtime = 0; 1710 p->se.nr_migrations = 0; 1711 p->se.vruntime = 0; 1712 INIT_LIST_HEAD(&p->se.group_node); 1713 1714#ifdef CONFIG_SCHEDSTATS 1715 memset(&p->se.statistics, 0, sizeof(p->se.statistics)); 1716#endif 1717 1718 RB_CLEAR_NODE(&p->dl.rb_node); 1719 hrtimer_init(&p->dl.dl_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); 1720 p->dl.dl_runtime = p->dl.runtime = 0; 1721 p->dl.dl_deadline = p->dl.deadline = 0; 1722 p->dl.dl_period = 0; 1723 p->dl.flags = 0; 1724 1725 INIT_LIST_HEAD(&p->rt.run_list); 1726 1727#ifdef CONFIG_PREEMPT_NOTIFIERS 1728 INIT_HLIST_HEAD(&p->preempt_notifiers); 1729#endif 1730 1731#ifdef CONFIG_NUMA_BALANCING 1732 if (p->mm && atomic_read(&p->mm->mm_users) == 1) { 1733 p->mm->numa_next_scan = jiffies + msecs_to_jiffies(sysctl_numa_balancing_scan_delay); 1734 p->mm->numa_scan_seq = 0; 1735 } 1736 1737 if (clone_flags & CLONE_VM) 1738 p->numa_preferred_nid = current->numa_preferred_nid; 1739 else 1740 p->numa_preferred_nid = -1; 1741 1742 p->node_stamp = 0ULL; 1743 p->numa_scan_seq = p->mm ? p->mm->numa_scan_seq : 0; 1744 p->numa_scan_period = sysctl_numa_balancing_scan_delay; 1745 p->numa_work.next = &p->numa_work; 1746 p->numa_faults_memory = NULL; 1747 p->numa_faults_buffer_memory = NULL; 1748 p->last_task_numa_placement = 0; 1749 p->last_sum_exec_runtime = 0; 1750 1751 INIT_LIST_HEAD(&p->numa_entry); 1752 p->numa_group = NULL; 1753#endif /* CONFIG_NUMA_BALANCING */ 1754} 1755 1756#ifdef CONFIG_NUMA_BALANCING 1757#ifdef CONFIG_SCHED_DEBUG 1758void set_numabalancing_state(bool enabled) 1759{ 1760 if (enabled) 1761 sched_feat_set("NUMA"); 1762 else 1763 sched_feat_set("NO_NUMA"); 1764} 1765#else 1766__read_mostly bool numabalancing_enabled; 1767 1768void set_numabalancing_state(bool enabled) 1769{ 1770 numabalancing_enabled = enabled; 1771} 1772#endif /* CONFIG_SCHED_DEBUG */ 1773 1774#ifdef CONFIG_PROC_SYSCTL 1775int sysctl_numa_balancing(struct ctl_table *table, int write, 1776 void __user *buffer, size_t *lenp, loff_t *ppos) 1777{ 1778 struct ctl_table t; 1779 int err; 1780 int state = numabalancing_enabled; 1781 1782 if (write && !capable(CAP_SYS_ADMIN)) 1783 return -EPERM; 1784 1785 t = *table; 1786 t.data = &state; 1787 err = proc_dointvec_minmax(&t, write, buffer, lenp, ppos); 1788 if (err < 0) 1789 return err; 1790 if (write) 1791 set_numabalancing_state(state); 1792 return err; 1793} 1794#endif 1795#endif 1796 1797/* 1798 * fork()/clone()-time setup: 1799 */ 1800int sched_fork(unsigned long clone_flags, struct task_struct *p) 1801{ 1802 unsigned long flags; 1803 int cpu = get_cpu(); 1804 1805 __sched_fork(clone_flags, p); 1806 /* 1807 * We mark the process as running here. This guarantees that 1808 * nobody will actually run it, and a signal or other external 1809 * event cannot wake it up and insert it on the runqueue either. 1810 */ 1811 p->state = TASK_RUNNING; 1812 1813 /* 1814 * Make sure we do not leak PI boosting priority to the child. 1815 */ 1816 p->prio = current->normal_prio; 1817 1818 /* 1819 * Revert to default priority/policy on fork if requested. 1820 */ 1821 if (unlikely(p->sched_reset_on_fork)) { 1822 if (task_has_dl_policy(p) || task_has_rt_policy(p)) { 1823 p->policy = SCHED_NORMAL; 1824 p->static_prio = NICE_TO_PRIO(0); 1825 p->rt_priority = 0; 1826 } else if (PRIO_TO_NICE(p->static_prio) < 0) 1827 p->static_prio = NICE_TO_PRIO(0); 1828 1829 p->prio = p->normal_prio = __normal_prio(p); 1830 set_load_weight(p); 1831 1832 /* 1833 * We don't need the reset flag anymore after the fork. It has 1834 * fulfilled its duty: 1835 */ 1836 p->sched_reset_on_fork = 0; 1837 } 1838 1839 if (dl_prio(p->prio)) { 1840 put_cpu(); 1841 return -EAGAIN; 1842 } else if (rt_prio(p->prio)) { 1843 p->sched_class = &rt_sched_class; 1844 } else { 1845 p->sched_class = &fair_sched_class; 1846 } 1847 1848 if (p->sched_class->task_fork) 1849 p->sched_class->task_fork(p); 1850 1851 /* 1852 * The child is not yet in the pid-hash so no cgroup attach races, 1853 * and the cgroup is pinned to this child due to cgroup_fork() 1854 * is ran before sched_fork(). 1855 * 1856 * Silence PROVE_RCU. 1857 */ 1858 raw_spin_lock_irqsave(&p->pi_lock, flags); 1859 set_task_cpu(p, cpu); 1860 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 1861 1862#if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT) 1863 if (likely(sched_info_on())) 1864 memset(&p->sched_info, 0, sizeof(p->sched_info)); 1865#endif 1866#if defined(CONFIG_SMP) 1867 p->on_cpu = 0; 1868#endif 1869 init_task_preempt_count(p); 1870#ifdef CONFIG_SMP 1871 plist_node_init(&p->pushable_tasks, MAX_PRIO); 1872 RB_CLEAR_NODE(&p->pushable_dl_tasks); 1873#endif 1874 1875 put_cpu(); 1876 return 0; 1877} 1878 1879unsigned long to_ratio(u64 period, u64 runtime) 1880{ 1881 if (runtime == RUNTIME_INF) 1882 return 1ULL << 20; 1883 1884 /* 1885 * Doing this here saves a lot of checks in all 1886 * the calling paths, and returning zero seems 1887 * safe for them anyway. 1888 */ 1889 if (period == 0) 1890 return 0; 1891 1892 return div64_u64(runtime << 20, period); 1893} 1894 1895#ifdef CONFIG_SMP 1896inline struct dl_bw *dl_bw_of(int i) 1897{ 1898 return &cpu_rq(i)->rd->dl_bw; 1899} 1900 1901static inline int dl_bw_cpus(int i) 1902{ 1903 struct root_domain *rd = cpu_rq(i)->rd; 1904 int cpus = 0; 1905 1906 for_each_cpu_and(i, rd->span, cpu_active_mask) 1907 cpus++; 1908 1909 return cpus; 1910} 1911#else 1912inline struct dl_bw *dl_bw_of(int i) 1913{ 1914 return &cpu_rq(i)->dl.dl_bw; 1915} 1916 1917static inline int dl_bw_cpus(int i) 1918{ 1919 return 1; 1920} 1921#endif 1922 1923static inline 1924void __dl_clear(struct dl_bw *dl_b, u64 tsk_bw) 1925{ 1926 dl_b->total_bw -= tsk_bw; 1927} 1928 1929static inline 1930void __dl_add(struct dl_bw *dl_b, u64 tsk_bw) 1931{ 1932 dl_b->total_bw += tsk_bw; 1933} 1934 1935static inline 1936bool __dl_overflow(struct dl_bw *dl_b, int cpus, u64 old_bw, u64 new_bw) 1937{ 1938 return dl_b->bw != -1 && 1939 dl_b->bw * cpus < dl_b->total_bw - old_bw + new_bw; 1940} 1941 1942/* 1943 * We must be sure that accepting a new task (or allowing changing the 1944 * parameters of an existing one) is consistent with the bandwidth 1945 * constraints. If yes, this function also accordingly updates the currently 1946 * allocated bandwidth to reflect the new situation. 1947 * 1948 * This function is called while holding p's rq->lock. 1949 */ 1950static int dl_overflow(struct task_struct *p, int policy, 1951 const struct sched_attr *attr) 1952{ 1953 1954 struct dl_bw *dl_b = dl_bw_of(task_cpu(p)); 1955 u64 period = attr->sched_period ?: attr->sched_deadline; 1956 u64 runtime = attr->sched_runtime; 1957 u64 new_bw = dl_policy(policy) ? to_ratio(period, runtime) : 0; 1958 int cpus, err = -1; 1959 1960 if (new_bw == p->dl.dl_bw) 1961 return 0; 1962 1963 /* 1964 * Either if a task, enters, leave, or stays -deadline but changes 1965 * its parameters, we may need to update accordingly the total 1966 * allocated bandwidth of the container. 1967 */ 1968 raw_spin_lock(&dl_b->lock); 1969 cpus = dl_bw_cpus(task_cpu(p)); 1970 if (dl_policy(policy) && !task_has_dl_policy(p) && 1971 !__dl_overflow(dl_b, cpus, 0, new_bw)) { 1972 __dl_add(dl_b, new_bw); 1973 err = 0; 1974 } else if (dl_policy(policy) && task_has_dl_policy(p) && 1975 !__dl_overflow(dl_b, cpus, p->dl.dl_bw, new_bw)) { 1976 __dl_clear(dl_b, p->dl.dl_bw); 1977 __dl_add(dl_b, new_bw); 1978 err = 0; 1979 } else if (!dl_policy(policy) && task_has_dl_policy(p)) { 1980 __dl_clear(dl_b, p->dl.dl_bw); 1981 err = 0; 1982 } 1983 raw_spin_unlock(&dl_b->lock); 1984 1985 return err; 1986} 1987 1988extern void init_dl_bw(struct dl_bw *dl_b); 1989 1990/* 1991 * wake_up_new_task - wake up a newly created task for the first time. 1992 * 1993 * This function will do some initial scheduler statistics housekeeping 1994 * that must be done for every newly created context, then puts the task 1995 * on the runqueue and wakes it. 1996 */ 1997void wake_up_new_task(struct task_struct *p) 1998{ 1999 unsigned long flags; 2000 struct rq *rq; 2001 2002 raw_spin_lock_irqsave(&p->pi_lock, flags); 2003#ifdef CONFIG_SMP 2004 /* 2005 * Fork balancing, do it here and not earlier because: 2006 * - cpus_allowed can change in the fork path 2007 * - any previously selected cpu might disappear through hotplug 2008 */ 2009 set_task_cpu(p, select_task_rq(p, task_cpu(p), SD_BALANCE_FORK, 0)); 2010#endif 2011 2012 /* Initialize new task's runnable average */ 2013 init_task_runnable_average(p); 2014 rq = __task_rq_lock(p); 2015 activate_task(rq, p, 0); 2016 p->on_rq = 1; 2017 trace_sched_wakeup_new(p, true); 2018 check_preempt_curr(rq, p, WF_FORK); 2019#ifdef CONFIG_SMP 2020 if (p->sched_class->task_woken) 2021 p->sched_class->task_woken(rq, p); 2022#endif 2023 task_rq_unlock(rq, p, &flags); 2024} 2025 2026#ifdef CONFIG_PREEMPT_NOTIFIERS 2027 2028/** 2029 * preempt_notifier_register - tell me when current is being preempted & rescheduled 2030 * @notifier: notifier struct to register 2031 */ 2032void preempt_notifier_register(struct preempt_notifier *notifier) 2033{ 2034 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers); 2035} 2036EXPORT_SYMBOL_GPL(preempt_notifier_register); 2037 2038/** 2039 * preempt_notifier_unregister - no longer interested in preemption notifications 2040 * @notifier: notifier struct to unregister 2041 * 2042 * This is safe to call from within a preemption notifier. 2043 */ 2044void preempt_notifier_unregister(struct preempt_notifier *notifier) 2045{ 2046 hlist_del(¬ifier->link); 2047} 2048EXPORT_SYMBOL_GPL(preempt_notifier_unregister); 2049 2050static void fire_sched_in_preempt_notifiers(struct task_struct *curr) 2051{ 2052 struct preempt_notifier *notifier; 2053 2054 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link) 2055 notifier->ops->sched_in(notifier, raw_smp_processor_id()); 2056} 2057 2058static void 2059fire_sched_out_preempt_notifiers(struct task_struct *curr, 2060 struct task_struct *next) 2061{ 2062 struct preempt_notifier *notifier; 2063 2064 hlist_for_each_entry(notifier, &curr->preempt_notifiers, link) 2065 notifier->ops->sched_out(notifier, next); 2066} 2067 2068#else /* !CONFIG_PREEMPT_NOTIFIERS */ 2069 2070static void fire_sched_in_preempt_notifiers(struct task_struct *curr) 2071{ 2072} 2073 2074static void 2075fire_sched_out_preempt_notifiers(struct task_struct *curr, 2076 struct task_struct *next) 2077{ 2078} 2079 2080#endif /* CONFIG_PREEMPT_NOTIFIERS */ 2081 2082/** 2083 * prepare_task_switch - prepare to switch tasks 2084 * @rq: the runqueue preparing to switch 2085 * @prev: the current task that is being switched out 2086 * @next: the task we are going to switch to. 2087 * 2088 * This is called with the rq lock held and interrupts off. It must 2089 * be paired with a subsequent finish_task_switch after the context 2090 * switch. 2091 * 2092 * prepare_task_switch sets up locking and calls architecture specific 2093 * hooks. 2094 */ 2095static inline void 2096prepare_task_switch(struct rq *rq, struct task_struct *prev, 2097 struct task_struct *next) 2098{ 2099 trace_sched_switch(prev, next); 2100 sched_info_switch(rq, prev, next); 2101 perf_event_task_sched_out(prev, next); 2102 fire_sched_out_preempt_notifiers(prev, next); 2103 prepare_lock_switch(rq, next); 2104 prepare_arch_switch(next); 2105} 2106 2107/** 2108 * finish_task_switch - clean up after a task-switch 2109 * @rq: runqueue associated with task-switch 2110 * @prev: the thread we just switched away from. 2111 * 2112 * finish_task_switch must be called after the context switch, paired 2113 * with a prepare_task_switch call before the context switch. 2114 * finish_task_switch will reconcile locking set up by prepare_task_switch, 2115 * and do any other architecture-specific cleanup actions. 2116 * 2117 * Note that we may have delayed dropping an mm in context_switch(). If 2118 * so, we finish that here outside of the runqueue lock. (Doing it 2119 * with the lock held can cause deadlocks; see schedule() for 2120 * details.) 2121 */ 2122static void finish_task_switch(struct rq *rq, struct task_struct *prev) 2123 __releases(rq->lock) 2124{ 2125 struct mm_struct *mm = rq->prev_mm; 2126 long prev_state; 2127 2128 rq->prev_mm = NULL; 2129 2130 /* 2131 * A task struct has one reference for the use as "current". 2132 * If a task dies, then it sets TASK_DEAD in tsk->state and calls 2133 * schedule one last time. The schedule call will never return, and 2134 * the scheduled task must drop that reference. 2135 * The test for TASK_DEAD must occur while the runqueue locks are 2136 * still held, otherwise prev could be scheduled on another cpu, die 2137 * there before we look at prev->state, and then the reference would 2138 * be dropped twice. 2139 * Manfred Spraul <manfred@colorfullife.com> 2140 */ 2141 prev_state = prev->state; 2142 vtime_task_switch(prev); 2143 finish_arch_switch(prev); 2144 perf_event_task_sched_in(prev, current); 2145 finish_lock_switch(rq, prev); 2146 finish_arch_post_lock_switch(); 2147 2148 fire_sched_in_preempt_notifiers(current); 2149 if (mm) 2150 mmdrop(mm); 2151 if (unlikely(prev_state == TASK_DEAD)) { 2152 if (prev->sched_class->task_dead) 2153 prev->sched_class->task_dead(prev); 2154 2155 /* 2156 * Remove function-return probe instances associated with this 2157 * task and put them back on the free list. 2158 */ 2159 kprobe_flush_task(prev); 2160 put_task_struct(prev); 2161 } 2162 2163 tick_nohz_task_switch(current); 2164} 2165 2166#ifdef CONFIG_SMP 2167 2168/* rq->lock is NOT held, but preemption is disabled */ 2169static inline void post_schedule(struct rq *rq) 2170{ 2171 if (rq->post_schedule) { 2172 unsigned long flags; 2173 2174 raw_spin_lock_irqsave(&rq->lock, flags); 2175 if (rq->curr->sched_class->post_schedule) 2176 rq->curr->sched_class->post_schedule(rq); 2177 raw_spin_unlock_irqrestore(&rq->lock, flags); 2178 2179 rq->post_schedule = 0; 2180 } 2181} 2182 2183#else 2184 2185static inline void post_schedule(struct rq *rq) 2186{ 2187} 2188 2189#endif 2190 2191/** 2192 * schedule_tail - first thing a freshly forked thread must call. 2193 * @prev: the thread we just switched away from. 2194 */ 2195asmlinkage void schedule_tail(struct task_struct *prev) 2196 __releases(rq->lock) 2197{ 2198 struct rq *rq = this_rq(); 2199 2200 finish_task_switch(rq, prev); 2201 2202 /* 2203 * FIXME: do we need to worry about rq being invalidated by the 2204 * task_switch? 2205 */ 2206 post_schedule(rq); 2207 2208#ifdef __ARCH_WANT_UNLOCKED_CTXSW 2209 /* In this case, finish_task_switch does not reenable preemption */ 2210 preempt_enable(); 2211#endif 2212 if (current->set_child_tid) 2213 put_user(task_pid_vnr(current), current->set_child_tid); 2214} 2215 2216/* 2217 * context_switch - switch to the new MM and the new 2218 * thread's register state. 2219 */ 2220static inline void 2221context_switch(struct rq *rq, struct task_struct *prev, 2222 struct task_struct *next) 2223{ 2224 struct mm_struct *mm, *oldmm; 2225 2226 prepare_task_switch(rq, prev, next); 2227 2228 mm = next->mm; 2229 oldmm = prev->active_mm; 2230 /* 2231 * For paravirt, this is coupled with an exit in switch_to to 2232 * combine the page table reload and the switch backend into 2233 * one hypercall. 2234 */ 2235 arch_start_context_switch(prev); 2236 2237 if (!mm) { 2238 next->active_mm = oldmm; 2239 atomic_inc(&oldmm->mm_count); 2240 enter_lazy_tlb(oldmm, next); 2241 } else 2242 switch_mm(oldmm, mm, next); 2243 2244 if (!prev->mm) { 2245 prev->active_mm = NULL; 2246 rq->prev_mm = oldmm; 2247 } 2248 /* 2249 * Since the runqueue lock will be released by the next 2250 * task (which is an invalid locking op but in the case 2251 * of the scheduler it's an obvious special-case), so we 2252 * do an early lockdep release here: 2253 */ 2254#ifndef __ARCH_WANT_UNLOCKED_CTXSW 2255 spin_release(&rq->lock.dep_map, 1, _THIS_IP_); 2256#endif 2257 2258 context_tracking_task_switch(prev, next); 2259 /* Here we just switch the register state and the stack. */ 2260 switch_to(prev, next, prev); 2261 2262 barrier(); 2263 /* 2264 * this_rq must be evaluated again because prev may have moved 2265 * CPUs since it called schedule(), thus the 'rq' on its stack 2266 * frame will be invalid. 2267 */ 2268 finish_task_switch(this_rq(), prev); 2269} 2270 2271/* 2272 * nr_running and nr_context_switches: 2273 * 2274 * externally visible scheduler statistics: current number of runnable 2275 * threads, total number of context switches performed since bootup. 2276 */ 2277unsigned long nr_running(void) 2278{ 2279 unsigned long i, sum = 0; 2280 2281 for_each_online_cpu(i) 2282 sum += cpu_rq(i)->nr_running; 2283 2284 return sum; 2285} 2286 2287unsigned long long nr_context_switches(void) 2288{ 2289 int i; 2290 unsigned long long sum = 0; 2291 2292 for_each_possible_cpu(i) 2293 sum += cpu_rq(i)->nr_switches; 2294 2295 return sum; 2296} 2297 2298unsigned long nr_iowait(void) 2299{ 2300 unsigned long i, sum = 0; 2301 2302 for_each_possible_cpu(i) 2303 sum += atomic_read(&cpu_rq(i)->nr_iowait); 2304 2305 return sum; 2306} 2307 2308unsigned long nr_iowait_cpu(int cpu) 2309{ 2310 struct rq *this = cpu_rq(cpu); 2311 return atomic_read(&this->nr_iowait); 2312} 2313 2314#ifdef CONFIG_SMP 2315 2316/* 2317 * sched_exec - execve() is a valuable balancing opportunity, because at 2318 * this point the task has the smallest effective memory and cache footprint. 2319 */ 2320void sched_exec(void) 2321{ 2322 struct task_struct *p = current; 2323 unsigned long flags; 2324 int dest_cpu; 2325 2326 raw_spin_lock_irqsave(&p->pi_lock, flags); 2327 dest_cpu = p->sched_class->select_task_rq(p, task_cpu(p), SD_BALANCE_EXEC, 0); 2328 if (dest_cpu == smp_processor_id()) 2329 goto unlock; 2330 2331 if (likely(cpu_active(dest_cpu))) { 2332 struct migration_arg arg = { p, dest_cpu }; 2333 2334 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 2335 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg); 2336 return; 2337 } 2338unlock: 2339 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 2340} 2341 2342#endif 2343 2344DEFINE_PER_CPU(struct kernel_stat, kstat); 2345DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat); 2346 2347EXPORT_PER_CPU_SYMBOL(kstat); 2348EXPORT_PER_CPU_SYMBOL(kernel_cpustat); 2349 2350/* 2351 * Return any ns on the sched_clock that have not yet been accounted in 2352 * @p in case that task is currently running. 2353 * 2354 * Called with task_rq_lock() held on @rq. 2355 */ 2356static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq) 2357{ 2358 u64 ns = 0; 2359 2360 if (task_current(rq, p)) { 2361 update_rq_clock(rq); 2362 ns = rq_clock_task(rq) - p->se.exec_start; 2363 if ((s64)ns < 0) 2364 ns = 0; 2365 } 2366 2367 return ns; 2368} 2369 2370unsigned long long task_delta_exec(struct task_struct *p) 2371{ 2372 unsigned long flags; 2373 struct rq *rq; 2374 u64 ns = 0; 2375 2376 rq = task_rq_lock(p, &flags); 2377 ns = do_task_delta_exec(p, rq); 2378 task_rq_unlock(rq, p, &flags); 2379 2380 return ns; 2381} 2382 2383/* 2384 * Return accounted runtime for the task. 2385 * In case the task is currently running, return the runtime plus current's 2386 * pending runtime that have not been accounted yet. 2387 */ 2388unsigned long long task_sched_runtime(struct task_struct *p) 2389{ 2390 unsigned long flags; 2391 struct rq *rq; 2392 u64 ns = 0; 2393 2394#if defined(CONFIG_64BIT) && defined(CONFIG_SMP) 2395 /* 2396 * 64-bit doesn't need locks to atomically read a 64bit value. 2397 * So we have a optimization chance when the task's delta_exec is 0. 2398 * Reading ->on_cpu is racy, but this is ok. 2399 * 2400 * If we race with it leaving cpu, we'll take a lock. So we're correct. 2401 * If we race with it entering cpu, unaccounted time is 0. This is 2402 * indistinguishable from the read occurring a few cycles earlier. 2403 */ 2404 if (!p->on_cpu) 2405 return p->se.sum_exec_runtime; 2406#endif 2407 2408 rq = task_rq_lock(p, &flags); 2409 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq); 2410 task_rq_unlock(rq, p, &flags); 2411 2412 return ns; 2413} 2414 2415/* 2416 * This function gets called by the timer code, with HZ frequency. 2417 * We call it with interrupts disabled. 2418 */ 2419void scheduler_tick(void) 2420{ 2421 int cpu = smp_processor_id(); 2422 struct rq *rq = cpu_rq(cpu); 2423 struct task_struct *curr = rq->curr; 2424 2425 sched_clock_tick(); 2426 2427 raw_spin_lock(&rq->lock); 2428 update_rq_clock(rq); 2429 curr->sched_class->task_tick(rq, curr, 0); 2430 update_cpu_load_active(rq); 2431 raw_spin_unlock(&rq->lock); 2432 2433 perf_event_task_tick(); 2434 2435#ifdef CONFIG_SMP 2436 rq->idle_balance = idle_cpu(cpu); 2437 trigger_load_balance(rq); 2438#endif 2439 rq_last_tick_reset(rq); 2440} 2441 2442#ifdef CONFIG_NO_HZ_FULL 2443/** 2444 * scheduler_tick_max_deferment 2445 * 2446 * Keep at least one tick per second when a single 2447 * active task is running because the scheduler doesn't 2448 * yet completely support full dynticks environment. 2449 * 2450 * This makes sure that uptime, CFS vruntime, load 2451 * balancing, etc... continue to move forward, even 2452 * with a very low granularity. 2453 * 2454 * Return: Maximum deferment in nanoseconds. 2455 */ 2456u64 scheduler_tick_max_deferment(void) 2457{ 2458 struct rq *rq = this_rq(); 2459 unsigned long next, now = ACCESS_ONCE(jiffies); 2460 2461 next = rq->last_sched_tick + HZ; 2462 2463 if (time_before_eq(next, now)) 2464 return 0; 2465 2466 return jiffies_to_nsecs(next - now); 2467} 2468#endif 2469 2470notrace unsigned long get_parent_ip(unsigned long addr) 2471{ 2472 if (in_lock_functions(addr)) { 2473 addr = CALLER_ADDR2; 2474 if (in_lock_functions(addr)) 2475 addr = CALLER_ADDR3; 2476 } 2477 return addr; 2478} 2479 2480#if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \ 2481 defined(CONFIG_PREEMPT_TRACER)) 2482 2483void __kprobes preempt_count_add(int val) 2484{ 2485#ifdef CONFIG_DEBUG_PREEMPT 2486 /* 2487 * Underflow? 2488 */ 2489 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0))) 2490 return; 2491#endif 2492 __preempt_count_add(val); 2493#ifdef CONFIG_DEBUG_PREEMPT 2494 /* 2495 * Spinlock count overflowing soon? 2496 */ 2497 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >= 2498 PREEMPT_MASK - 10); 2499#endif 2500 if (preempt_count() == val) { 2501 unsigned long ip = get_parent_ip(CALLER_ADDR1); 2502#ifdef CONFIG_DEBUG_PREEMPT 2503 current->preempt_disable_ip = ip; 2504#endif 2505 trace_preempt_off(CALLER_ADDR0, ip); 2506 } 2507} 2508EXPORT_SYMBOL(preempt_count_add); 2509 2510void __kprobes preempt_count_sub(int val) 2511{ 2512#ifdef CONFIG_DEBUG_PREEMPT 2513 /* 2514 * Underflow? 2515 */ 2516 if (DEBUG_LOCKS_WARN_ON(val > preempt_count())) 2517 return; 2518 /* 2519 * Is the spinlock portion underflowing? 2520 */ 2521 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) && 2522 !(preempt_count() & PREEMPT_MASK))) 2523 return; 2524#endif 2525 2526 if (preempt_count() == val) 2527 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1)); 2528 __preempt_count_sub(val); 2529} 2530EXPORT_SYMBOL(preempt_count_sub); 2531 2532#endif 2533 2534/* 2535 * Print scheduling while atomic bug: 2536 */ 2537static noinline void __schedule_bug(struct task_struct *prev) 2538{ 2539 if (oops_in_progress) 2540 return; 2541 2542 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n", 2543 prev->comm, prev->pid, preempt_count()); 2544 2545 debug_show_held_locks(prev); 2546 print_modules(); 2547 if (irqs_disabled()) 2548 print_irqtrace_events(prev); 2549#ifdef CONFIG_DEBUG_PREEMPT 2550 if (in_atomic_preempt_off()) { 2551 pr_err("Preemption disabled at:"); 2552 print_ip_sym(current->preempt_disable_ip); 2553 pr_cont("\n"); 2554 } 2555#endif 2556 dump_stack(); 2557 add_taint(TAINT_WARN, LOCKDEP_STILL_OK); 2558} 2559 2560/* 2561 * Various schedule()-time debugging checks and statistics: 2562 */ 2563static inline void schedule_debug(struct task_struct *prev) 2564{ 2565 /* 2566 * Test if we are atomic. Since do_exit() needs to call into 2567 * schedule() atomically, we ignore that path. Otherwise whine 2568 * if we are scheduling when we should not. 2569 */ 2570 if (unlikely(in_atomic_preempt_off() && prev->state != TASK_DEAD)) 2571 __schedule_bug(prev); 2572 rcu_sleep_check(); 2573 2574 profile_hit(SCHED_PROFILING, __builtin_return_address(0)); 2575 2576 schedstat_inc(this_rq(), sched_count); 2577} 2578 2579/* 2580 * Pick up the highest-prio task: 2581 */ 2582static inline struct task_struct * 2583pick_next_task(struct rq *rq, struct task_struct *prev) 2584{ 2585 const struct sched_class *class = &fair_sched_class; 2586 struct task_struct *p; 2587 2588 /* 2589 * Optimization: we know that if all tasks are in 2590 * the fair class we can call that function directly: 2591 */ 2592 if (likely(prev->sched_class == class && 2593 rq->nr_running == rq->cfs.h_nr_running)) { 2594 p = fair_sched_class.pick_next_task(rq, prev); 2595 if (likely(p && p != RETRY_TASK)) 2596 return p; 2597 } 2598 2599again: 2600 for_each_class(class) { 2601 p = class->pick_next_task(rq, prev); 2602 if (p) { 2603 if (unlikely(p == RETRY_TASK)) 2604 goto again; 2605 return p; 2606 } 2607 } 2608 2609 BUG(); /* the idle class will always have a runnable task */ 2610} 2611 2612/* 2613 * __schedule() is the main scheduler function. 2614 * 2615 * The main means of driving the scheduler and thus entering this function are: 2616 * 2617 * 1. Explicit blocking: mutex, semaphore, waitqueue, etc. 2618 * 2619 * 2. TIF_NEED_RESCHED flag is checked on interrupt and userspace return 2620 * paths. For example, see arch/x86/entry_64.S. 2621 * 2622 * To drive preemption between tasks, the scheduler sets the flag in timer 2623 * interrupt handler scheduler_tick(). 2624 * 2625 * 3. Wakeups don't really cause entry into schedule(). They add a 2626 * task to the run-queue and that's it. 2627 * 2628 * Now, if the new task added to the run-queue preempts the current 2629 * task, then the wakeup sets TIF_NEED_RESCHED and schedule() gets 2630 * called on the nearest possible occasion: 2631 * 2632 * - If the kernel is preemptible (CONFIG_PREEMPT=y): 2633 * 2634 * - in syscall or exception context, at the next outmost 2635 * preempt_enable(). (this might be as soon as the wake_up()'s 2636 * spin_unlock()!) 2637 * 2638 * - in IRQ context, return from interrupt-handler to 2639 * preemptible context 2640 * 2641 * - If the kernel is not preemptible (CONFIG_PREEMPT is not set) 2642 * then at the next: 2643 * 2644 * - cond_resched() call 2645 * - explicit schedule() call 2646 * - return from syscall or exception to user-space 2647 * - return from interrupt-handler to user-space 2648 */ 2649static void __sched __schedule(void) 2650{ 2651 struct task_struct *prev, *next; 2652 unsigned long *switch_count; 2653 struct rq *rq; 2654 int cpu; 2655 2656need_resched: 2657 preempt_disable(); 2658 cpu = smp_processor_id(); 2659 rq = cpu_rq(cpu); 2660 rcu_note_context_switch(cpu); 2661 prev = rq->curr; 2662 2663 schedule_debug(prev); 2664 2665 if (sched_feat(HRTICK)) 2666 hrtick_clear(rq); 2667 2668 /* 2669 * Make sure that signal_pending_state()->signal_pending() below 2670 * can't be reordered with __set_current_state(TASK_INTERRUPTIBLE) 2671 * done by the caller to avoid the race with signal_wake_up(). 2672 */ 2673 smp_mb__before_spinlock(); 2674 raw_spin_lock_irq(&rq->lock); 2675 2676 switch_count = &prev->nivcsw; 2677 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) { 2678 if (unlikely(signal_pending_state(prev->state, prev))) { 2679 prev->state = TASK_RUNNING; 2680 } else { 2681 deactivate_task(rq, prev, DEQUEUE_SLEEP); 2682 prev->on_rq = 0; 2683 2684 /* 2685 * If a worker went to sleep, notify and ask workqueue 2686 * whether it wants to wake up a task to maintain 2687 * concurrency. 2688 */ 2689 if (prev->flags & PF_WQ_WORKER) { 2690 struct task_struct *to_wakeup; 2691 2692 to_wakeup = wq_worker_sleeping(prev, cpu); 2693 if (to_wakeup) 2694 try_to_wake_up_local(to_wakeup); 2695 } 2696 } 2697 switch_count = &prev->nvcsw; 2698 } 2699 2700 if (prev->on_rq || rq->skip_clock_update < 0) 2701 update_rq_clock(rq); 2702 2703 next = pick_next_task(rq, prev); 2704 clear_tsk_need_resched(prev); 2705 clear_preempt_need_resched(); 2706 rq->skip_clock_update = 0; 2707 2708 if (likely(prev != next)) { 2709 rq->nr_switches++; 2710 rq->curr = next; 2711 ++*switch_count; 2712 2713 context_switch(rq, prev, next); /* unlocks the rq */ 2714 /* 2715 * The context switch have flipped the stack from under us 2716 * and restored the local variables which were saved when 2717 * this task called schedule() in the past. prev == current 2718 * is still correct, but it can be moved to another cpu/rq. 2719 */ 2720 cpu = smp_processor_id(); 2721 rq = cpu_rq(cpu); 2722 } else 2723 raw_spin_unlock_irq(&rq->lock); 2724 2725 post_schedule(rq); 2726 2727 sched_preempt_enable_no_resched(); 2728 if (need_resched()) 2729 goto need_resched; 2730} 2731 2732static inline void sched_submit_work(struct task_struct *tsk) 2733{ 2734 if (!tsk->state || tsk_is_pi_blocked(tsk)) 2735 return; 2736 /* 2737 * If we are going to sleep and we have plugged IO queued, 2738 * make sure to submit it to avoid deadlocks. 2739 */ 2740 if (blk_needs_flush_plug(tsk)) 2741 blk_schedule_flush_plug(tsk); 2742} 2743 2744asmlinkage void __sched schedule(void) 2745{ 2746 struct task_struct *tsk = current; 2747 2748 sched_submit_work(tsk); 2749 __schedule(); 2750} 2751EXPORT_SYMBOL(schedule); 2752 2753#ifdef CONFIG_CONTEXT_TRACKING 2754asmlinkage void __sched schedule_user(void) 2755{ 2756 /* 2757 * If we come here after a random call to set_need_resched(), 2758 * or we have been woken up remotely but the IPI has not yet arrived, 2759 * we haven't yet exited the RCU idle mode. Do it here manually until 2760 * we find a better solution. 2761 */ 2762 user_exit(); 2763 schedule(); 2764 user_enter(); 2765} 2766#endif 2767 2768/** 2769 * schedule_preempt_disabled - called with preemption disabled 2770 * 2771 * Returns with preemption disabled. Note: preempt_count must be 1 2772 */ 2773void __sched schedule_preempt_disabled(void) 2774{ 2775 sched_preempt_enable_no_resched(); 2776 schedule(); 2777 preempt_disable(); 2778} 2779 2780#ifdef CONFIG_PREEMPT 2781/* 2782 * this is the entry point to schedule() from in-kernel preemption 2783 * off of preempt_enable. Kernel preemptions off return from interrupt 2784 * occur there and call schedule directly. 2785 */ 2786asmlinkage void __sched notrace preempt_schedule(void) 2787{ 2788 /* 2789 * If there is a non-zero preempt_count or interrupts are disabled, 2790 * we do not want to preempt the current task. Just return.. 2791 */ 2792 if (likely(!preemptible())) 2793 return; 2794 2795 do { 2796 __preempt_count_add(PREEMPT_ACTIVE); 2797 __schedule(); 2798 __preempt_count_sub(PREEMPT_ACTIVE); 2799 2800 /* 2801 * Check again in case we missed a preemption opportunity 2802 * between schedule and now. 2803 */ 2804 barrier(); 2805 } while (need_resched()); 2806} 2807EXPORT_SYMBOL(preempt_schedule); 2808#endif /* CONFIG_PREEMPT */ 2809 2810/* 2811 * this is the entry point to schedule() from kernel preemption 2812 * off of irq context. 2813 * Note, that this is called and return with irqs disabled. This will 2814 * protect us against recursive calling from irq. 2815 */ 2816asmlinkage void __sched preempt_schedule_irq(void) 2817{ 2818 enum ctx_state prev_state; 2819 2820 /* Catch callers which need to be fixed */ 2821 BUG_ON(preempt_count() || !irqs_disabled()); 2822 2823 prev_state = exception_enter(); 2824 2825 do { 2826 __preempt_count_add(PREEMPT_ACTIVE); 2827 local_irq_enable(); 2828 __schedule(); 2829 local_irq_disable(); 2830 __preempt_count_sub(PREEMPT_ACTIVE); 2831 2832 /* 2833 * Check again in case we missed a preemption opportunity 2834 * between schedule and now. 2835 */ 2836 barrier(); 2837 } while (need_resched()); 2838 2839 exception_exit(prev_state); 2840} 2841 2842int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags, 2843 void *key) 2844{ 2845 return try_to_wake_up(curr->private, mode, wake_flags); 2846} 2847EXPORT_SYMBOL(default_wake_function); 2848 2849#ifdef CONFIG_RT_MUTEXES 2850 2851/* 2852 * rt_mutex_setprio - set the current priority of a task 2853 * @p: task 2854 * @prio: prio value (kernel-internal form) 2855 * 2856 * This function changes the 'effective' priority of a task. It does 2857 * not touch ->normal_prio like __setscheduler(). 2858 * 2859 * Used by the rt_mutex code to implement priority inheritance 2860 * logic. Call site only calls if the priority of the task changed. 2861 */ 2862void rt_mutex_setprio(struct task_struct *p, int prio) 2863{ 2864 int oldprio, on_rq, running, enqueue_flag = 0; 2865 struct rq *rq; 2866 const struct sched_class *prev_class; 2867 2868 BUG_ON(prio > MAX_PRIO); 2869 2870 rq = __task_rq_lock(p); 2871 2872 /* 2873 * Idle task boosting is a nono in general. There is one 2874 * exception, when PREEMPT_RT and NOHZ is active: 2875 * 2876 * The idle task calls get_next_timer_interrupt() and holds 2877 * the timer wheel base->lock on the CPU and another CPU wants 2878 * to access the timer (probably to cancel it). We can safely 2879 * ignore the boosting request, as the idle CPU runs this code 2880 * with interrupts disabled and will complete the lock 2881 * protected section without being interrupted. So there is no 2882 * real need to boost. 2883 */ 2884 if (unlikely(p == rq->idle)) { 2885 WARN_ON(p != rq->curr); 2886 WARN_ON(p->pi_blocked_on); 2887 goto out_unlock; 2888 } 2889 2890 trace_sched_pi_setprio(p, prio); 2891 p->pi_top_task = rt_mutex_get_top_task(p); 2892 oldprio = p->prio; 2893 prev_class = p->sched_class; 2894 on_rq = p->on_rq; 2895 running = task_current(rq, p); 2896 if (on_rq) 2897 dequeue_task(rq, p, 0); 2898 if (running) 2899 p->sched_class->put_prev_task(rq, p); 2900 2901 /* 2902 * Boosting condition are: 2903 * 1. -rt task is running and holds mutex A 2904 * --> -dl task blocks on mutex A 2905 * 2906 * 2. -dl task is running and holds mutex A 2907 * --> -dl task blocks on mutex A and could preempt the 2908 * running task 2909 */ 2910 if (dl_prio(prio)) { 2911 if (!dl_prio(p->normal_prio) || (p->pi_top_task && 2912 dl_entity_preempt(&p->pi_top_task->dl, &p->dl))) { 2913 p->dl.dl_boosted = 1; 2914 p->dl.dl_throttled = 0; 2915 enqueue_flag = ENQUEUE_REPLENISH; 2916 } else 2917 p->dl.dl_boosted = 0; 2918 p->sched_class = &dl_sched_class; 2919 } else if (rt_prio(prio)) { 2920 if (dl_prio(oldprio)) 2921 p->dl.dl_boosted = 0; 2922 if (oldprio < prio) 2923 enqueue_flag = ENQUEUE_HEAD; 2924 p->sched_class = &rt_sched_class; 2925 } else { 2926 if (dl_prio(oldprio)) 2927 p->dl.dl_boosted = 0; 2928 p->sched_class = &fair_sched_class; 2929 } 2930 2931 p->prio = prio; 2932 2933 if (running) 2934 p->sched_class->set_curr_task(rq); 2935 if (on_rq) 2936 enqueue_task(rq, p, enqueue_flag); 2937 2938 check_class_changed(rq, p, prev_class, oldprio); 2939out_unlock: 2940 __task_rq_unlock(rq); 2941} 2942#endif 2943 2944void set_user_nice(struct task_struct *p, long nice) 2945{ 2946 int old_prio, delta, on_rq; 2947 unsigned long flags; 2948 struct rq *rq; 2949 2950 if (task_nice(p) == nice || nice < MIN_NICE || nice > MAX_NICE) 2951 return; 2952 /* 2953 * We have to be careful, if called from sys_setpriority(), 2954 * the task might be in the middle of scheduling on another CPU. 2955 */ 2956 rq = task_rq_lock(p, &flags); 2957 /* 2958 * The RT priorities are set via sched_setscheduler(), but we still 2959 * allow the 'normal' nice value to be set - but as expected 2960 * it wont have any effect on scheduling until the task is 2961 * SCHED_DEADLINE, SCHED_FIFO or SCHED_RR: 2962 */ 2963 if (task_has_dl_policy(p) || task_has_rt_policy(p)) { 2964 p->static_prio = NICE_TO_PRIO(nice); 2965 goto out_unlock; 2966 } 2967 on_rq = p->on_rq; 2968 if (on_rq) 2969 dequeue_task(rq, p, 0); 2970 2971 p->static_prio = NICE_TO_PRIO(nice); 2972 set_load_weight(p); 2973 old_prio = p->prio; 2974 p->prio = effective_prio(p); 2975 delta = p->prio - old_prio; 2976 2977 if (on_rq) { 2978 enqueue_task(rq, p, 0); 2979 /* 2980 * If the task increased its priority or is running and 2981 * lowered its priority, then reschedule its CPU: 2982 */ 2983 if (delta < 0 || (delta > 0 && task_running(rq, p))) 2984 resched_task(rq->curr); 2985 } 2986out_unlock: 2987 task_rq_unlock(rq, p, &flags); 2988} 2989EXPORT_SYMBOL(set_user_nice); 2990 2991/* 2992 * can_nice - check if a task can reduce its nice value 2993 * @p: task 2994 * @nice: nice value 2995 */ 2996int can_nice(const struct task_struct *p, const int nice) 2997{ 2998 /* convert nice value [19,-20] to rlimit style value [1,40] */ 2999 int nice_rlim = 20 - nice; 3000 3001 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) || 3002 capable(CAP_SYS_NICE)); 3003} 3004 3005#ifdef __ARCH_WANT_SYS_NICE 3006 3007/* 3008 * sys_nice - change the priority of the current process. 3009 * @increment: priority increment 3010 * 3011 * sys_setpriority is a more generic, but much slower function that 3012 * does similar things. 3013 */ 3014SYSCALL_DEFINE1(nice, int, increment) 3015{ 3016 long nice, retval; 3017 3018 /* 3019 * Setpriority might change our priority at the same moment. 3020 * We don't have to worry. Conceptually one call occurs first 3021 * and we have a single winner. 3022 */ 3023 if (increment < -40) 3024 increment = -40; 3025 if (increment > 40) 3026 increment = 40; 3027 3028 nice = task_nice(current) + increment; 3029 if (nice < MIN_NICE) 3030 nice = MIN_NICE; 3031 if (nice > MAX_NICE) 3032 nice = MAX_NICE; 3033 3034 if (increment < 0 && !can_nice(current, nice)) 3035 return -EPERM; 3036 3037 retval = security_task_setnice(current, nice); 3038 if (retval) 3039 return retval; 3040 3041 set_user_nice(current, nice); 3042 return 0; 3043} 3044 3045#endif 3046 3047/** 3048 * task_prio - return the priority value of a given task. 3049 * @p: the task in question. 3050 * 3051 * Return: The priority value as seen by users in /proc. 3052 * RT tasks are offset by -200. Normal tasks are centered 3053 * around 0, value goes from -16 to +15. 3054 */ 3055int task_prio(const struct task_struct *p) 3056{ 3057 return p->prio - MAX_RT_PRIO; 3058} 3059 3060/** 3061 * idle_cpu - is a given cpu idle currently? 3062 * @cpu: the processor in question. 3063 * 3064 * Return: 1 if the CPU is currently idle. 0 otherwise. 3065 */ 3066int idle_cpu(int cpu) 3067{ 3068 struct rq *rq = cpu_rq(cpu); 3069 3070 if (rq->curr != rq->idle) 3071 return 0; 3072 3073 if (rq->nr_running) 3074 return 0; 3075 3076#ifdef CONFIG_SMP 3077 if (!llist_empty(&rq->wake_list)) 3078 return 0; 3079#endif 3080 3081 return 1; 3082} 3083 3084/** 3085 * idle_task - return the idle task for a given cpu. 3086 * @cpu: the processor in question. 3087 * 3088 * Return: The idle task for the cpu @cpu. 3089 */ 3090struct task_struct *idle_task(int cpu) 3091{ 3092 return cpu_rq(cpu)->idle; 3093} 3094 3095/** 3096 * find_process_by_pid - find a process with a matching PID value. 3097 * @pid: the pid in question. 3098 * 3099 * The task of @pid, if found. %NULL otherwise. 3100 */ 3101static struct task_struct *find_process_by_pid(pid_t pid) 3102{ 3103 return pid ? find_task_by_vpid(pid) : current; 3104} 3105 3106/* 3107 * This function initializes the sched_dl_entity of a newly becoming 3108 * SCHED_DEADLINE task. 3109 * 3110 * Only the static values are considered here, the actual runtime and the 3111 * absolute deadline will be properly calculated when the task is enqueued 3112 * for the first time with its new policy. 3113 */ 3114static void 3115__setparam_dl(struct task_struct *p, const struct sched_attr *attr) 3116{ 3117 struct sched_dl_entity *dl_se = &p->dl; 3118 3119 init_dl_task_timer(dl_se); 3120 dl_se->dl_runtime = attr->sched_runtime; 3121 dl_se->dl_deadline = attr->sched_deadline; 3122 dl_se->dl_period = attr->sched_period ?: dl_se->dl_deadline; 3123 dl_se->flags = attr->sched_flags; 3124 dl_se->dl_bw = to_ratio(dl_se->dl_period, dl_se->dl_runtime); 3125 dl_se->dl_throttled = 0; 3126 dl_se->dl_new = 1; 3127 dl_se->dl_yielded = 0; 3128} 3129 3130static void __setscheduler_params(struct task_struct *p, 3131 const struct sched_attr *attr) 3132{ 3133 int policy = attr->sched_policy; 3134 3135 if (policy == -1) /* setparam */ 3136 policy = p->policy; 3137 3138 p->policy = policy; 3139 3140 if (dl_policy(policy)) 3141 __setparam_dl(p, attr); 3142 else if (fair_policy(policy)) 3143 p->static_prio = NICE_TO_PRIO(attr->sched_nice); 3144 3145 /* 3146 * __sched_setscheduler() ensures attr->sched_priority == 0 when 3147 * !rt_policy. Always setting this ensures that things like 3148 * getparam()/getattr() don't report silly values for !rt tasks. 3149 */ 3150 p->rt_priority = attr->sched_priority; 3151 p->normal_prio = normal_prio(p); 3152 set_load_weight(p); 3153} 3154 3155/* Actually do priority change: must hold pi & rq lock. */ 3156static void __setscheduler(struct rq *rq, struct task_struct *p, 3157 const struct sched_attr *attr) 3158{ 3159 __setscheduler_params(p, attr); 3160 3161 /* 3162 * If we get here, there was no pi waiters boosting the 3163 * task. It is safe to use the normal prio. 3164 */ 3165 p->prio = normal_prio(p); 3166 3167 if (dl_prio(p->prio)) 3168 p->sched_class = &dl_sched_class; 3169 else if (rt_prio(p->prio)) 3170 p->sched_class = &rt_sched_class; 3171 else 3172 p->sched_class = &fair_sched_class; 3173} 3174 3175static void 3176__getparam_dl(struct task_struct *p, struct sched_attr *attr) 3177{ 3178 struct sched_dl_entity *dl_se = &p->dl; 3179 3180 attr->sched_priority = p->rt_priority; 3181 attr->sched_runtime = dl_se->dl_runtime; 3182 attr->sched_deadline = dl_se->dl_deadline; 3183 attr->sched_period = dl_se->dl_period; 3184 attr->sched_flags = dl_se->flags; 3185} 3186 3187/* 3188 * This function validates the new parameters of a -deadline task. 3189 * We ask for the deadline not being zero, and greater or equal 3190 * than the runtime, as well as the period of being zero or 3191 * greater than deadline. Furthermore, we have to be sure that 3192 * user parameters are above the internal resolution (1us); we 3193 * check sched_runtime only since it is always the smaller one. 3194 */ 3195static bool 3196__checkparam_dl(const struct sched_attr *attr) 3197{ 3198 return attr && attr->sched_deadline != 0 && 3199 (attr->sched_period == 0 || 3200 (s64)(attr->sched_period - attr->sched_deadline) >= 0) && 3201 (s64)(attr->sched_deadline - attr->sched_runtime ) >= 0 && 3202 attr->sched_runtime >= (2 << (DL_SCALE - 1)); 3203} 3204 3205/* 3206 * check the target process has a UID that matches the current process's 3207 */ 3208static bool check_same_owner(struct task_struct *p) 3209{ 3210 const struct cred *cred = current_cred(), *pcred; 3211 bool match; 3212 3213 rcu_read_lock(); 3214 pcred = __task_cred(p); 3215 match = (uid_eq(cred->euid, pcred->euid) || 3216 uid_eq(cred->euid, pcred->uid)); 3217 rcu_read_unlock(); 3218 return match; 3219} 3220 3221static int __sched_setscheduler(struct task_struct *p, 3222 const struct sched_attr *attr, 3223 bool user) 3224{ 3225 int newprio = dl_policy(attr->sched_policy) ? MAX_DL_PRIO - 1 : 3226 MAX_RT_PRIO - 1 - attr->sched_priority; 3227 int retval, oldprio, oldpolicy = -1, on_rq, running; 3228 int policy = attr->sched_policy; 3229 unsigned long flags; 3230 const struct sched_class *prev_class; 3231 struct rq *rq; 3232 int reset_on_fork; 3233 3234 /* may grab non-irq protected spin_locks */ 3235 BUG_ON(in_interrupt()); 3236recheck: 3237 /* double check policy once rq lock held */ 3238 if (policy < 0) { 3239 reset_on_fork = p->sched_reset_on_fork; 3240 policy = oldpolicy = p->policy; 3241 } else { 3242 reset_on_fork = !!(attr->sched_flags & SCHED_FLAG_RESET_ON_FORK); 3243 3244 if (policy != SCHED_DEADLINE && 3245 policy != SCHED_FIFO && policy != SCHED_RR && 3246 policy != SCHED_NORMAL && policy != SCHED_BATCH && 3247 policy != SCHED_IDLE) 3248 return -EINVAL; 3249 } 3250 3251 if (attr->sched_flags & ~(SCHED_FLAG_RESET_ON_FORK)) 3252 return -EINVAL; 3253 3254 /* 3255 * Valid priorities for SCHED_FIFO and SCHED_RR are 3256 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL, 3257 * SCHED_BATCH and SCHED_IDLE is 0. 3258 */ 3259 if ((p->mm && attr->sched_priority > MAX_USER_RT_PRIO-1) || 3260 (!p->mm && attr->sched_priority > MAX_RT_PRIO-1)) 3261 return -EINVAL; 3262 if ((dl_policy(policy) && !__checkparam_dl(attr)) || 3263 (rt_policy(policy) != (attr->sched_priority != 0))) 3264 return -EINVAL; 3265 3266 /* 3267 * Allow unprivileged RT tasks to decrease priority: 3268 */ 3269 if (user && !capable(CAP_SYS_NICE)) { 3270 if (fair_policy(policy)) { 3271 if (attr->sched_nice < task_nice(p) && 3272 !can_nice(p, attr->sched_nice)) 3273 return -EPERM; 3274 } 3275 3276 if (rt_policy(policy)) { 3277 unsigned long rlim_rtprio = 3278 task_rlimit(p, RLIMIT_RTPRIO); 3279 3280 /* can't set/change the rt policy */ 3281 if (policy != p->policy && !rlim_rtprio) 3282 return -EPERM; 3283 3284 /* can't increase priority */ 3285 if (attr->sched_priority > p->rt_priority && 3286 attr->sched_priority > rlim_rtprio) 3287 return -EPERM; 3288 } 3289 3290 /* 3291 * Can't set/change SCHED_DEADLINE policy at all for now 3292 * (safest behavior); in the future we would like to allow 3293 * unprivileged DL tasks to increase their relative deadline 3294 * or reduce their runtime (both ways reducing utilization) 3295 */ 3296 if (dl_policy(policy)) 3297 return -EPERM; 3298 3299 /* 3300 * Treat SCHED_IDLE as nice 20. Only allow a switch to 3301 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it. 3302 */ 3303 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) { 3304 if (!can_nice(p, task_nice(p))) 3305 return -EPERM; 3306 } 3307 3308 /* can't change other user's priorities */ 3309 if (!check_same_owner(p)) 3310 return -EPERM; 3311 3312 /* Normal users shall not reset the sched_reset_on_fork flag */ 3313 if (p->sched_reset_on_fork && !reset_on_fork) 3314 return -EPERM; 3315 } 3316 3317 if (user) { 3318 retval = security_task_setscheduler(p); 3319 if (retval) 3320 return retval; 3321 } 3322 3323 /* 3324 * make sure no PI-waiters arrive (or leave) while we are 3325 * changing the priority of the task: 3326 * 3327 * To be able to change p->policy safely, the appropriate 3328 * runqueue lock must be held. 3329 */ 3330 rq = task_rq_lock(p, &flags); 3331 3332 /* 3333 * Changing the policy of the stop threads its a very bad idea 3334 */ 3335 if (p == rq->stop) { 3336 task_rq_unlock(rq, p, &flags); 3337 return -EINVAL; 3338 } 3339 3340 /* 3341 * If not changing anything there's no need to proceed further, 3342 * but store a possible modification of reset_on_fork. 3343 */ 3344 if (unlikely(policy == p->policy)) { 3345 if (fair_policy(policy) && attr->sched_nice != task_nice(p)) 3346 goto change; 3347 if (rt_policy(policy) && attr->sched_priority != p->rt_priority) 3348 goto change; 3349 if (dl_policy(policy)) 3350 goto change; 3351 3352 p->sched_reset_on_fork = reset_on_fork; 3353 task_rq_unlock(rq, p, &flags); 3354 return 0; 3355 } 3356change: 3357 3358 if (user) { 3359#ifdef CONFIG_RT_GROUP_SCHED 3360 /* 3361 * Do not allow realtime tasks into groups that have no runtime 3362 * assigned. 3363 */ 3364 if (rt_bandwidth_enabled() && rt_policy(policy) && 3365 task_group(p)->rt_bandwidth.rt_runtime == 0 && 3366 !task_group_is_autogroup(task_group(p))) { 3367 task_rq_unlock(rq, p, &flags); 3368 return -EPERM; 3369 } 3370#endif 3371#ifdef CONFIG_SMP 3372 if (dl_bandwidth_enabled() && dl_policy(policy)) { 3373 cpumask_t *span = rq->rd->span; 3374 3375 /* 3376 * Don't allow tasks with an affinity mask smaller than 3377 * the entire root_domain to become SCHED_DEADLINE. We 3378 * will also fail if there's no bandwidth available. 3379 */ 3380 if (!cpumask_subset(span, &p->cpus_allowed) || 3381 rq->rd->dl_bw.bw == 0) { 3382 task_rq_unlock(rq, p, &flags); 3383 return -EPERM; 3384 } 3385 } 3386#endif 3387 } 3388 3389 /* recheck policy now with rq lock held */ 3390 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) { 3391 policy = oldpolicy = -1; 3392 task_rq_unlock(rq, p, &flags); 3393 goto recheck; 3394 } 3395 3396 /* 3397 * If setscheduling to SCHED_DEADLINE (or changing the parameters 3398 * of a SCHED_DEADLINE task) we need to check if enough bandwidth 3399 * is available. 3400 */ 3401 if ((dl_policy(policy) || dl_task(p)) && dl_overflow(p, policy, attr)) { 3402 task_rq_unlock(rq, p, &flags); 3403 return -EBUSY; 3404 } 3405 3406 p->sched_reset_on_fork = reset_on_fork; 3407 oldprio = p->prio; 3408 3409 /* 3410 * Special case for priority boosted tasks. 3411 * 3412 * If the new priority is lower or equal (user space view) 3413 * than the current (boosted) priority, we just store the new 3414 * normal parameters and do not touch the scheduler class and 3415 * the runqueue. This will be done when the task deboost 3416 * itself. 3417 */ 3418 if (rt_mutex_check_prio(p, newprio)) { 3419 __setscheduler_params(p, attr); 3420 task_rq_unlock(rq, p, &flags); 3421 return 0; 3422 } 3423 3424 on_rq = p->on_rq; 3425 running = task_current(rq, p); 3426 if (on_rq) 3427 dequeue_task(rq, p, 0); 3428 if (running) 3429 p->sched_class->put_prev_task(rq, p); 3430 3431 prev_class = p->sched_class; 3432 __setscheduler(rq, p, attr); 3433 3434 if (running) 3435 p->sched_class->set_curr_task(rq); 3436 if (on_rq) { 3437 /* 3438 * We enqueue to tail when the priority of a task is 3439 * increased (user space view). 3440 */ 3441 enqueue_task(rq, p, oldprio <= p->prio ? ENQUEUE_HEAD : 0); 3442 } 3443 3444 check_class_changed(rq, p, prev_class, oldprio); 3445 task_rq_unlock(rq, p, &flags); 3446 3447 rt_mutex_adjust_pi(p); 3448 3449 return 0; 3450} 3451 3452static int _sched_setscheduler(struct task_struct *p, int policy, 3453 const struct sched_param *param, bool check) 3454{ 3455 struct sched_attr attr = { 3456 .sched_policy = policy, 3457 .sched_priority = param->sched_priority, 3458 .sched_nice = PRIO_TO_NICE(p->static_prio), 3459 }; 3460 3461 /* 3462 * Fixup the legacy SCHED_RESET_ON_FORK hack 3463 */ 3464 if (policy & SCHED_RESET_ON_FORK) { 3465 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK; 3466 policy &= ~SCHED_RESET_ON_FORK; 3467 attr.sched_policy = policy; 3468 } 3469 3470 return __sched_setscheduler(p, &attr, check); 3471} 3472/** 3473 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread. 3474 * @p: the task in question. 3475 * @policy: new policy. 3476 * @param: structure containing the new RT priority. 3477 * 3478 * Return: 0 on success. An error code otherwise. 3479 * 3480 * NOTE that the task may be already dead. 3481 */ 3482int sched_setscheduler(struct task_struct *p, int policy, 3483 const struct sched_param *param) 3484{ 3485 return _sched_setscheduler(p, policy, param, true); 3486} 3487EXPORT_SYMBOL_GPL(sched_setscheduler); 3488 3489int sched_setattr(struct task_struct *p, const struct sched_attr *attr) 3490{ 3491 return __sched_setscheduler(p, attr, true); 3492} 3493EXPORT_SYMBOL_GPL(sched_setattr); 3494 3495/** 3496 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace. 3497 * @p: the task in question. 3498 * @policy: new policy. 3499 * @param: structure containing the new RT priority. 3500 * 3501 * Just like sched_setscheduler, only don't bother checking if the 3502 * current context has permission. For example, this is needed in 3503 * stop_machine(): we create temporary high priority worker threads, 3504 * but our caller might not have that capability. 3505 * 3506 * Return: 0 on success. An error code otherwise. 3507 */ 3508int sched_setscheduler_nocheck(struct task_struct *p, int policy, 3509 const struct sched_param *param) 3510{ 3511 return _sched_setscheduler(p, policy, param, false); 3512} 3513 3514static int 3515do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param) 3516{ 3517 struct sched_param lparam; 3518 struct task_struct *p; 3519 int retval; 3520 3521 if (!param || pid < 0) 3522 return -EINVAL; 3523 if (copy_from_user(&lparam, param, sizeof(struct sched_param))) 3524 return -EFAULT; 3525 3526 rcu_read_lock(); 3527 retval = -ESRCH; 3528 p = find_process_by_pid(pid); 3529 if (p != NULL) 3530 retval = sched_setscheduler(p, policy, &lparam); 3531 rcu_read_unlock(); 3532 3533 return retval; 3534} 3535 3536/* 3537 * Mimics kernel/events/core.c perf_copy_attr(). 3538 */ 3539static int sched_copy_attr(struct sched_attr __user *uattr, 3540 struct sched_attr *attr) 3541{ 3542 u32 size; 3543 int ret; 3544 3545 if (!access_ok(VERIFY_WRITE, uattr, SCHED_ATTR_SIZE_VER0)) 3546 return -EFAULT; 3547 3548 /* 3549 * zero the full structure, so that a short copy will be nice. 3550 */ 3551 memset(attr, 0, sizeof(*attr)); 3552 3553 ret = get_user(size, &uattr->size); 3554 if (ret) 3555 return ret; 3556 3557 if (size > PAGE_SIZE) /* silly large */ 3558 goto err_size; 3559 3560 if (!size) /* abi compat */ 3561 size = SCHED_ATTR_SIZE_VER0; 3562 3563 if (size < SCHED_ATTR_SIZE_VER0) 3564 goto err_size; 3565 3566 /* 3567 * If we're handed a bigger struct than we know of, 3568 * ensure all the unknown bits are 0 - i.e. new 3569 * user-space does not rely on any kernel feature 3570 * extensions we dont know about yet. 3571 */ 3572 if (size > sizeof(*attr)) { 3573 unsigned char __user *addr; 3574 unsigned char __user *end; 3575 unsigned char val; 3576 3577 addr = (void __user *)uattr + sizeof(*attr); 3578 end = (void __user *)uattr + size; 3579 3580 for (; addr < end; addr++) { 3581 ret = get_user(val, addr); 3582 if (ret) 3583 return ret; 3584 if (val) 3585 goto err_size; 3586 } 3587 size = sizeof(*attr); 3588 } 3589 3590 ret = copy_from_user(attr, uattr, size); 3591 if (ret) 3592 return -EFAULT; 3593 3594 /* 3595 * XXX: do we want to be lenient like existing syscalls; or do we want 3596 * to be strict and return an error on out-of-bounds values? 3597 */ 3598 attr->sched_nice = clamp(attr->sched_nice, MIN_NICE, MAX_NICE); 3599 3600out: 3601 return ret; 3602 3603err_size: 3604 put_user(sizeof(*attr), &uattr->size); 3605 ret = -E2BIG; 3606 goto out; 3607} 3608 3609/** 3610 * sys_sched_setscheduler - set/change the scheduler policy and RT priority 3611 * @pid: the pid in question. 3612 * @policy: new policy. 3613 * @param: structure containing the new RT priority. 3614 * 3615 * Return: 0 on success. An error code otherwise. 3616 */ 3617SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, 3618 struct sched_param __user *, param) 3619{ 3620 /* negative values for policy are not valid */ 3621 if (policy < 0) 3622 return -EINVAL; 3623 3624 return do_sched_setscheduler(pid, policy, param); 3625} 3626 3627/** 3628 * sys_sched_setparam - set/change the RT priority of a thread 3629 * @pid: the pid in question. 3630 * @param: structure containing the new RT priority. 3631 * 3632 * Return: 0 on success. An error code otherwise. 3633 */ 3634SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param) 3635{ 3636 return do_sched_setscheduler(pid, -1, param); 3637} 3638 3639/** 3640 * sys_sched_setattr - same as above, but with extended sched_attr 3641 * @pid: the pid in question. 3642 * @uattr: structure containing the extended parameters. 3643 * @flags: for future extension. 3644 */ 3645SYSCALL_DEFINE3(sched_setattr, pid_t, pid, struct sched_attr __user *, uattr, 3646 unsigned int, flags) 3647{ 3648 struct sched_attr attr; 3649 struct task_struct *p; 3650 int retval; 3651 3652 if (!uattr || pid < 0 || flags) 3653 return -EINVAL; 3654 3655 if (sched_copy_attr(uattr, &attr)) 3656 return -EFAULT; 3657 3658 rcu_read_lock(); 3659 retval = -ESRCH; 3660 p = find_process_by_pid(pid); 3661 if (p != NULL) 3662 retval = sched_setattr(p, &attr); 3663 rcu_read_unlock(); 3664 3665 return retval; 3666} 3667 3668/** 3669 * sys_sched_getscheduler - get the policy (scheduling class) of a thread 3670 * @pid: the pid in question. 3671 * 3672 * Return: On success, the policy of the thread. Otherwise, a negative error 3673 * code. 3674 */ 3675SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid) 3676{ 3677 struct task_struct *p; 3678 int retval; 3679 3680 if (pid < 0) 3681 return -EINVAL; 3682 3683 retval = -ESRCH; 3684 rcu_read_lock(); 3685 p = find_process_by_pid(pid); 3686 if (p) { 3687 retval = security_task_getscheduler(p); 3688 if (!retval) 3689 retval = p->policy 3690 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0); 3691 } 3692 rcu_read_unlock(); 3693 return retval; 3694} 3695 3696/** 3697 * sys_sched_getparam - get the RT priority of a thread 3698 * @pid: the pid in question. 3699 * @param: structure containing the RT priority. 3700 * 3701 * Return: On success, 0 and the RT priority is in @param. Otherwise, an error 3702 * code. 3703 */ 3704SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param) 3705{ 3706 struct sched_param lp; 3707 struct task_struct *p; 3708 int retval; 3709 3710 if (!param || pid < 0) 3711 return -EINVAL; 3712 3713 rcu_read_lock(); 3714 p = find_process_by_pid(pid); 3715 retval = -ESRCH; 3716 if (!p) 3717 goto out_unlock; 3718 3719 retval = security_task_getscheduler(p); 3720 if (retval) 3721 goto out_unlock; 3722 3723 if (task_has_dl_policy(p)) { 3724 retval = -EINVAL; 3725 goto out_unlock; 3726 } 3727 lp.sched_priority = p->rt_priority; 3728 rcu_read_unlock(); 3729 3730 /* 3731 * This one might sleep, we cannot do it with a spinlock held ... 3732 */ 3733 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0; 3734 3735 return retval; 3736 3737out_unlock: 3738 rcu_read_unlock(); 3739 return retval; 3740} 3741 3742static int sched_read_attr(struct sched_attr __user *uattr, 3743 struct sched_attr *attr, 3744 unsigned int usize) 3745{ 3746 int ret; 3747 3748 if (!access_ok(VERIFY_WRITE, uattr, usize)) 3749 return -EFAULT; 3750 3751 /* 3752 * If we're handed a smaller struct than we know of, 3753 * ensure all the unknown bits are 0 - i.e. old 3754 * user-space does not get uncomplete information. 3755 */ 3756 if (usize < sizeof(*attr)) { 3757 unsigned char *addr; 3758 unsigned char *end; 3759 3760 addr = (void *)attr + usize; 3761 end = (void *)attr + sizeof(*attr); 3762 3763 for (; addr < end; addr++) { 3764 if (*addr) 3765 goto err_size; 3766 } 3767 3768 attr->size = usize; 3769 } 3770 3771 ret = copy_to_user(uattr, attr, attr->size); 3772 if (ret) 3773 return -EFAULT; 3774 3775out: 3776 return ret; 3777 3778err_size: 3779 ret = -E2BIG; 3780 goto out; 3781} 3782 3783/** 3784 * sys_sched_getattr - similar to sched_getparam, but with sched_attr 3785 * @pid: the pid in question. 3786 * @uattr: structure containing the extended parameters. 3787 * @size: sizeof(attr) for fwd/bwd comp. 3788 * @flags: for future extension. 3789 */ 3790SYSCALL_DEFINE4(sched_getattr, pid_t, pid, struct sched_attr __user *, uattr, 3791 unsigned int, size, unsigned int, flags) 3792{ 3793 struct sched_attr attr = { 3794 .size = sizeof(struct sched_attr), 3795 }; 3796 struct task_struct *p; 3797 int retval; 3798 3799 if (!uattr || pid < 0 || size > PAGE_SIZE || 3800 size < SCHED_ATTR_SIZE_VER0 || flags) 3801 return -EINVAL; 3802 3803 rcu_read_lock(); 3804 p = find_process_by_pid(pid); 3805 retval = -ESRCH; 3806 if (!p) 3807 goto out_unlock; 3808 3809 retval = security_task_getscheduler(p); 3810 if (retval) 3811 goto out_unlock; 3812 3813 attr.sched_policy = p->policy; 3814 if (p->sched_reset_on_fork) 3815 attr.sched_flags |= SCHED_FLAG_RESET_ON_FORK; 3816 if (task_has_dl_policy(p)) 3817 __getparam_dl(p, &attr); 3818 else if (task_has_rt_policy(p)) 3819 attr.sched_priority = p->rt_priority; 3820 else 3821 attr.sched_nice = task_nice(p); 3822 3823 rcu_read_unlock(); 3824 3825 retval = sched_read_attr(uattr, &attr, size); 3826 return retval; 3827 3828out_unlock: 3829 rcu_read_unlock(); 3830 return retval; 3831} 3832 3833long sched_setaffinity(pid_t pid, const struct cpumask *in_mask) 3834{ 3835 cpumask_var_t cpus_allowed, new_mask; 3836 struct task_struct *p; 3837 int retval; 3838 3839 rcu_read_lock(); 3840 3841 p = find_process_by_pid(pid); 3842 if (!p) { 3843 rcu_read_unlock(); 3844 return -ESRCH; 3845 } 3846 3847 /* Prevent p going away */ 3848 get_task_struct(p); 3849 rcu_read_unlock(); 3850 3851 if (p->flags & PF_NO_SETAFFINITY) { 3852 retval = -EINVAL; 3853 goto out_put_task; 3854 } 3855 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) { 3856 retval = -ENOMEM; 3857 goto out_put_task; 3858 } 3859 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) { 3860 retval = -ENOMEM; 3861 goto out_free_cpus_allowed; 3862 } 3863 retval = -EPERM; 3864 if (!check_same_owner(p)) { 3865 rcu_read_lock(); 3866 if (!ns_capable(__task_cred(p)->user_ns, CAP_SYS_NICE)) { 3867 rcu_read_unlock(); 3868 goto out_unlock; 3869 } 3870 rcu_read_unlock(); 3871 } 3872 3873 retval = security_task_setscheduler(p); 3874 if (retval) 3875 goto out_unlock; 3876 3877 3878 cpuset_cpus_allowed(p, cpus_allowed); 3879 cpumask_and(new_mask, in_mask, cpus_allowed); 3880 3881 /* 3882 * Since bandwidth control happens on root_domain basis, 3883 * if admission test is enabled, we only admit -deadline 3884 * tasks allowed to run on all the CPUs in the task's 3885 * root_domain. 3886 */ 3887#ifdef CONFIG_SMP 3888 if (task_has_dl_policy(p)) { 3889 const struct cpumask *span = task_rq(p)->rd->span; 3890 3891 if (dl_bandwidth_enabled() && !cpumask_subset(span, new_mask)) { 3892 retval = -EBUSY; 3893 goto out_unlock; 3894 } 3895 } 3896#endif 3897again: 3898 retval = set_cpus_allowed_ptr(p, new_mask); 3899 3900 if (!retval) { 3901 cpuset_cpus_allowed(p, cpus_allowed); 3902 if (!cpumask_subset(new_mask, cpus_allowed)) { 3903 /* 3904 * We must have raced with a concurrent cpuset 3905 * update. Just reset the cpus_allowed to the 3906 * cpuset's cpus_allowed 3907 */ 3908 cpumask_copy(new_mask, cpus_allowed); 3909 goto again; 3910 } 3911 } 3912out_unlock: 3913 free_cpumask_var(new_mask); 3914out_free_cpus_allowed: 3915 free_cpumask_var(cpus_allowed); 3916out_put_task: 3917 put_task_struct(p); 3918 return retval; 3919} 3920 3921static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len, 3922 struct cpumask *new_mask) 3923{ 3924 if (len < cpumask_size()) 3925 cpumask_clear(new_mask); 3926 else if (len > cpumask_size()) 3927 len = cpumask_size(); 3928 3929 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0; 3930} 3931 3932/** 3933 * sys_sched_setaffinity - set the cpu affinity of a process 3934 * @pid: pid of the process 3935 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 3936 * @user_mask_ptr: user-space pointer to the new cpu mask 3937 * 3938 * Return: 0 on success. An error code otherwise. 3939 */ 3940SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len, 3941 unsigned long __user *, user_mask_ptr) 3942{ 3943 cpumask_var_t new_mask; 3944 int retval; 3945 3946 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) 3947 return -ENOMEM; 3948 3949 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask); 3950 if (retval == 0) 3951 retval = sched_setaffinity(pid, new_mask); 3952 free_cpumask_var(new_mask); 3953 return retval; 3954} 3955 3956long sched_getaffinity(pid_t pid, struct cpumask *mask) 3957{ 3958 struct task_struct *p; 3959 unsigned long flags; 3960 int retval; 3961 3962 rcu_read_lock(); 3963 3964 retval = -ESRCH; 3965 p = find_process_by_pid(pid); 3966 if (!p) 3967 goto out_unlock; 3968 3969 retval = security_task_getscheduler(p); 3970 if (retval) 3971 goto out_unlock; 3972 3973 raw_spin_lock_irqsave(&p->pi_lock, flags); 3974 cpumask_and(mask, &p->cpus_allowed, cpu_active_mask); 3975 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 3976 3977out_unlock: 3978 rcu_read_unlock(); 3979 3980 return retval; 3981} 3982 3983/** 3984 * sys_sched_getaffinity - get the cpu affinity of a process 3985 * @pid: pid of the process 3986 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 3987 * @user_mask_ptr: user-space pointer to hold the current cpu mask 3988 * 3989 * Return: 0 on success. An error code otherwise. 3990 */ 3991SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len, 3992 unsigned long __user *, user_mask_ptr) 3993{ 3994 int ret; 3995 cpumask_var_t mask; 3996 3997 if ((len * BITS_PER_BYTE) < nr_cpu_ids) 3998 return -EINVAL; 3999 if (len & (sizeof(unsigned long)-1)) 4000 return -EINVAL; 4001 4002 if (!alloc_cpumask_var(&mask, GFP_KERNEL)) 4003 return -ENOMEM; 4004 4005 ret = sched_getaffinity(pid, mask); 4006 if (ret == 0) { 4007 size_t retlen = min_t(size_t, len, cpumask_size()); 4008 4009 if (copy_to_user(user_mask_ptr, mask, retlen)) 4010 ret = -EFAULT; 4011 else 4012 ret = retlen; 4013 } 4014 free_cpumask_var(mask); 4015 4016 return ret; 4017} 4018 4019/** 4020 * sys_sched_yield - yield the current processor to other threads. 4021 * 4022 * This function yields the current CPU to other tasks. If there are no 4023 * other threads running on this CPU then this function will return. 4024 * 4025 * Return: 0. 4026 */ 4027SYSCALL_DEFINE0(sched_yield) 4028{ 4029 struct rq *rq = this_rq_lock(); 4030 4031 schedstat_inc(rq, yld_count); 4032 current->sched_class->yield_task(rq); 4033 4034 /* 4035 * Since we are going to call schedule() anyway, there's 4036 * no need to preempt or enable interrupts: 4037 */ 4038 __release(rq->lock); 4039 spin_release(&rq->lock.dep_map, 1, _THIS_IP_); 4040 do_raw_spin_unlock(&rq->lock); 4041 sched_preempt_enable_no_resched(); 4042 4043 schedule(); 4044 4045 return 0; 4046} 4047 4048static void __cond_resched(void) 4049{ 4050 __preempt_count_add(PREEMPT_ACTIVE); 4051 __schedule(); 4052 __preempt_count_sub(PREEMPT_ACTIVE); 4053} 4054 4055int __sched _cond_resched(void) 4056{ 4057 if (should_resched()) { 4058 __cond_resched(); 4059 return 1; 4060 } 4061 return 0; 4062} 4063EXPORT_SYMBOL(_cond_resched); 4064 4065/* 4066 * __cond_resched_lock() - if a reschedule is pending, drop the given lock, 4067 * call schedule, and on return reacquire the lock. 4068 * 4069 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level 4070 * operations here to prevent schedule() from being called twice (once via 4071 * spin_unlock(), once by hand). 4072 */ 4073int __cond_resched_lock(spinlock_t *lock) 4074{ 4075 int resched = should_resched(); 4076 int ret = 0; 4077 4078 lockdep_assert_held(lock); 4079 4080 if (spin_needbreak(lock) || resched) { 4081 spin_unlock(lock); 4082 if (resched) 4083 __cond_resched(); 4084 else 4085 cpu_relax(); 4086 ret = 1; 4087 spin_lock(lock); 4088 } 4089 return ret; 4090} 4091EXPORT_SYMBOL(__cond_resched_lock); 4092 4093int __sched __cond_resched_softirq(void) 4094{ 4095 BUG_ON(!in_softirq()); 4096 4097 if (should_resched()) { 4098 local_bh_enable(); 4099 __cond_resched(); 4100 local_bh_disable(); 4101 return 1; 4102 } 4103 return 0; 4104} 4105EXPORT_SYMBOL(__cond_resched_softirq); 4106 4107/** 4108 * yield - yield the current processor to other threads. 4109 * 4110 * Do not ever use this function, there's a 99% chance you're doing it wrong. 4111 * 4112 * The scheduler is at all times free to pick the calling task as the most 4113 * eligible task to run, if removing the yield() call from your code breaks 4114 * it, its already broken. 4115 * 4116 * Typical broken usage is: 4117 * 4118 * while (!event) 4119 * yield(); 4120 * 4121 * where one assumes that yield() will let 'the other' process run that will 4122 * make event true. If the current task is a SCHED_FIFO task that will never 4123 * happen. Never use yield() as a progress guarantee!! 4124 * 4125 * If you want to use yield() to wait for something, use wait_event(). 4126 * If you want to use yield() to be 'nice' for others, use cond_resched(). 4127 * If you still want to use yield(), do not! 4128 */ 4129void __sched yield(void) 4130{ 4131 set_current_state(TASK_RUNNING); 4132 sys_sched_yield(); 4133} 4134EXPORT_SYMBOL(yield); 4135 4136/** 4137 * yield_to - yield the current processor to another thread in 4138 * your thread group, or accelerate that thread toward the 4139 * processor it's on. 4140 * @p: target task 4141 * @preempt: whether task preemption is allowed or not 4142 * 4143 * It's the caller's job to ensure that the target task struct 4144 * can't go away on us before we can do any checks. 4145 * 4146 * Return: 4147 * true (>0) if we indeed boosted the target task. 4148 * false (0) if we failed to boost the target. 4149 * -ESRCH if there's no task to yield to. 4150 */ 4151bool __sched yield_to(struct task_struct *p, bool preempt) 4152{ 4153 struct task_struct *curr = current; 4154 struct rq *rq, *p_rq; 4155 unsigned long flags; 4156 int yielded = 0; 4157 4158 local_irq_save(flags); 4159 rq = this_rq(); 4160 4161again: 4162 p_rq = task_rq(p); 4163 /* 4164 * If we're the only runnable task on the rq and target rq also 4165 * has only one task, there's absolutely no point in yielding. 4166 */ 4167 if (rq->nr_running == 1 && p_rq->nr_running == 1) { 4168 yielded = -ESRCH; 4169 goto out_irq; 4170 } 4171 4172 double_rq_lock(rq, p_rq); 4173 if (task_rq(p) != p_rq) { 4174 double_rq_unlock(rq, p_rq); 4175 goto again; 4176 } 4177 4178 if (!curr->sched_class->yield_to_task) 4179 goto out_unlock; 4180 4181 if (curr->sched_class != p->sched_class) 4182 goto out_unlock; 4183 4184 if (task_running(p_rq, p) || p->state) 4185 goto out_unlock; 4186 4187 yielded = curr->sched_class->yield_to_task(rq, p, preempt); 4188 if (yielded) { 4189 schedstat_inc(rq, yld_count); 4190 /* 4191 * Make p's CPU reschedule; pick_next_entity takes care of 4192 * fairness. 4193 */ 4194 if (preempt && rq != p_rq) 4195 resched_task(p_rq->curr); 4196 } 4197 4198out_unlock: 4199 double_rq_unlock(rq, p_rq); 4200out_irq: 4201 local_irq_restore(flags); 4202 4203 if (yielded > 0) 4204 schedule(); 4205 4206 return yielded; 4207} 4208EXPORT_SYMBOL_GPL(yield_to); 4209 4210/* 4211 * This task is about to go to sleep on IO. Increment rq->nr_iowait so 4212 * that process accounting knows that this is a task in IO wait state. 4213 */ 4214void __sched io_schedule(void) 4215{ 4216 struct rq *rq = raw_rq(); 4217 4218 delayacct_blkio_start(); 4219 atomic_inc(&rq->nr_iowait); 4220 blk_flush_plug(current); 4221 current->in_iowait = 1; 4222 schedule(); 4223 current->in_iowait = 0; 4224 atomic_dec(&rq->nr_iowait); 4225 delayacct_blkio_end(); 4226} 4227EXPORT_SYMBOL(io_schedule); 4228 4229long __sched io_schedule_timeout(long timeout) 4230{ 4231 struct rq *rq = raw_rq(); 4232 long ret; 4233 4234 delayacct_blkio_start(); 4235 atomic_inc(&rq->nr_iowait); 4236 blk_flush_plug(current); 4237 current->in_iowait = 1; 4238 ret = schedule_timeout(timeout); 4239 current->in_iowait = 0; 4240 atomic_dec(&rq->nr_iowait); 4241 delayacct_blkio_end(); 4242 return ret; 4243} 4244 4245/** 4246 * sys_sched_get_priority_max - return maximum RT priority. 4247 * @policy: scheduling class. 4248 * 4249 * Return: On success, this syscall returns the maximum 4250 * rt_priority that can be used by a given scheduling class. 4251 * On failure, a negative error code is returned. 4252 */ 4253SYSCALL_DEFINE1(sched_get_priority_max, int, policy) 4254{ 4255 int ret = -EINVAL; 4256 4257 switch (policy) { 4258 case SCHED_FIFO: 4259 case SCHED_RR: 4260 ret = MAX_USER_RT_PRIO-1; 4261 break; 4262 case SCHED_DEADLINE: 4263 case SCHED_NORMAL: 4264 case SCHED_BATCH: 4265 case SCHED_IDLE: 4266 ret = 0; 4267 break; 4268 } 4269 return ret; 4270} 4271 4272/** 4273 * sys_sched_get_priority_min - return minimum RT priority. 4274 * @policy: scheduling class. 4275 * 4276 * Return: On success, this syscall returns the minimum 4277 * rt_priority that can be used by a given scheduling class. 4278 * On failure, a negative error code is returned. 4279 */ 4280SYSCALL_DEFINE1(sched_get_priority_min, int, policy) 4281{ 4282 int ret = -EINVAL; 4283 4284 switch (policy) { 4285 case SCHED_FIFO: 4286 case SCHED_RR: 4287 ret = 1; 4288 break; 4289 case SCHED_DEADLINE: 4290 case SCHED_NORMAL: 4291 case SCHED_BATCH: 4292 case SCHED_IDLE: 4293 ret = 0; 4294 } 4295 return ret; 4296} 4297 4298/** 4299 * sys_sched_rr_get_interval - return the default timeslice of a process. 4300 * @pid: pid of the process. 4301 * @interval: userspace pointer to the timeslice value. 4302 * 4303 * this syscall writes the default timeslice value of a given process 4304 * into the user-space timespec buffer. A value of '0' means infinity. 4305 * 4306 * Return: On success, 0 and the timeslice is in @interval. Otherwise, 4307 * an error code. 4308 */ 4309SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid, 4310 struct timespec __user *, interval) 4311{ 4312 struct task_struct *p; 4313 unsigned int time_slice; 4314 unsigned long flags; 4315 struct rq *rq; 4316 int retval; 4317 struct timespec t; 4318 4319 if (pid < 0) 4320 return -EINVAL; 4321 4322 retval = -ESRCH; 4323 rcu_read_lock(); 4324 p = find_process_by_pid(pid); 4325 if (!p) 4326 goto out_unlock; 4327 4328 retval = security_task_getscheduler(p); 4329 if (retval) 4330 goto out_unlock; 4331 4332 rq = task_rq_lock(p, &flags); 4333 time_slice = 0; 4334 if (p->sched_class->get_rr_interval) 4335 time_slice = p->sched_class->get_rr_interval(rq, p); 4336 task_rq_unlock(rq, p, &flags); 4337 4338 rcu_read_unlock(); 4339 jiffies_to_timespec(time_slice, &t); 4340 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0; 4341 return retval; 4342 4343out_unlock: 4344 rcu_read_unlock(); 4345 return retval; 4346} 4347 4348static const char stat_nam[] = TASK_STATE_TO_CHAR_STR; 4349 4350void sched_show_task(struct task_struct *p) 4351{ 4352 unsigned long free = 0; 4353 int ppid; 4354 unsigned state; 4355 4356 state = p->state ? __ffs(p->state) + 1 : 0; 4357 printk(KERN_INFO "%-15.15s %c", p->comm, 4358 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?'); 4359#if BITS_PER_LONG == 32 4360 if (state == TASK_RUNNING) 4361 printk(KERN_CONT " running "); 4362 else 4363 printk(KERN_CONT " %08lx ", thread_saved_pc(p)); 4364#else 4365 if (state == TASK_RUNNING) 4366 printk(KERN_CONT " running task "); 4367 else 4368 printk(KERN_CONT " %016lx ", thread_saved_pc(p)); 4369#endif 4370#ifdef CONFIG_DEBUG_STACK_USAGE 4371 free = stack_not_used(p); 4372#endif 4373 rcu_read_lock(); 4374 ppid = task_pid_nr(rcu_dereference(p->real_parent)); 4375 rcu_read_unlock(); 4376 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free, 4377 task_pid_nr(p), ppid, 4378 (unsigned long)task_thread_info(p)->flags); 4379 4380 print_worker_info(KERN_INFO, p); 4381 show_stack(p, NULL); 4382} 4383 4384void show_state_filter(unsigned long state_filter) 4385{ 4386 struct task_struct *g, *p; 4387 4388#if BITS_PER_LONG == 32 4389 printk(KERN_INFO 4390 " task PC stack pid father\n"); 4391#else 4392 printk(KERN_INFO 4393 " task PC stack pid father\n"); 4394#endif 4395 rcu_read_lock(); 4396 do_each_thread(g, p) { 4397 /* 4398 * reset the NMI-timeout, listing all files on a slow 4399 * console might take a lot of time: 4400 */ 4401 touch_nmi_watchdog(); 4402 if (!state_filter || (p->state & state_filter)) 4403 sched_show_task(p); 4404 } while_each_thread(g, p); 4405 4406 touch_all_softlockup_watchdogs(); 4407 4408#ifdef CONFIG_SCHED_DEBUG 4409 sysrq_sched_debug_show(); 4410#endif 4411 rcu_read_unlock(); 4412 /* 4413 * Only show locks if all tasks are dumped: 4414 */ 4415 if (!state_filter) 4416 debug_show_all_locks(); 4417} 4418 4419void init_idle_bootup_task(struct task_struct *idle) 4420{ 4421 idle->sched_class = &idle_sched_class; 4422} 4423 4424/** 4425 * init_idle - set up an idle thread for a given CPU 4426 * @idle: task in question 4427 * @cpu: cpu the idle task belongs to 4428 * 4429 * NOTE: this function does not set the idle thread's NEED_RESCHED 4430 * flag, to make booting more robust. 4431 */ 4432void init_idle(struct task_struct *idle, int cpu) 4433{ 4434 struct rq *rq = cpu_rq(cpu); 4435 unsigned long flags; 4436 4437 raw_spin_lock_irqsave(&rq->lock, flags); 4438 4439 __sched_fork(0, idle); 4440 idle->state = TASK_RUNNING; 4441 idle->se.exec_start = sched_clock(); 4442 4443 do_set_cpus_allowed(idle, cpumask_of(cpu)); 4444 /* 4445 * We're having a chicken and egg problem, even though we are 4446 * holding rq->lock, the cpu isn't yet set to this cpu so the 4447 * lockdep check in task_group() will fail. 4448 * 4449 * Similar case to sched_fork(). / Alternatively we could 4450 * use task_rq_lock() here and obtain the other rq->lock. 4451 * 4452 * Silence PROVE_RCU 4453 */ 4454 rcu_read_lock(); 4455 __set_task_cpu(idle, cpu); 4456 rcu_read_unlock(); 4457 4458 rq->curr = rq->idle = idle; 4459 idle->on_rq = 1; 4460#if defined(CONFIG_SMP) 4461 idle->on_cpu = 1; 4462#endif 4463 raw_spin_unlock_irqrestore(&rq->lock, flags); 4464 4465 /* Set the preempt count _outside_ the spinlocks! */ 4466 init_idle_preempt_count(idle, cpu); 4467 4468 /* 4469 * The idle tasks have their own, simple scheduling class: 4470 */ 4471 idle->sched_class = &idle_sched_class; 4472 ftrace_graph_init_idle_task(idle, cpu); 4473 vtime_init_idle(idle, cpu); 4474#if defined(CONFIG_SMP) 4475 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu); 4476#endif 4477} 4478 4479#ifdef CONFIG_SMP 4480void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask) 4481{ 4482 if (p->sched_class && p->sched_class->set_cpus_allowed) 4483 p->sched_class->set_cpus_allowed(p, new_mask); 4484 4485 cpumask_copy(&p->cpus_allowed, new_mask); 4486 p->nr_cpus_allowed = cpumask_weight(new_mask); 4487} 4488 4489/* 4490 * This is how migration works: 4491 * 4492 * 1) we invoke migration_cpu_stop() on the target CPU using 4493 * stop_one_cpu(). 4494 * 2) stopper starts to run (implicitly forcing the migrated thread 4495 * off the CPU) 4496 * 3) it checks whether the migrated task is still in the wrong runqueue. 4497 * 4) if it's in the wrong runqueue then the migration thread removes 4498 * it and puts it into the right queue. 4499 * 5) stopper completes and stop_one_cpu() returns and the migration 4500 * is done. 4501 */ 4502 4503/* 4504 * Change a given task's CPU affinity. Migrate the thread to a 4505 * proper CPU and schedule it away if the CPU it's executing on 4506 * is removed from the allowed bitmask. 4507 * 4508 * NOTE: the caller must have a valid reference to the task, the 4509 * task must not exit() & deallocate itself prematurely. The 4510 * call is not atomic; no spinlocks may be held. 4511 */ 4512int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask) 4513{ 4514 unsigned long flags; 4515 struct rq *rq; 4516 unsigned int dest_cpu; 4517 int ret = 0; 4518 4519 rq = task_rq_lock(p, &flags); 4520 4521 if (cpumask_equal(&p->cpus_allowed, new_mask)) 4522 goto out; 4523 4524 if (!cpumask_intersects(new_mask, cpu_active_mask)) { 4525 ret = -EINVAL; 4526 goto out; 4527 } 4528 4529 do_set_cpus_allowed(p, new_mask); 4530 4531 /* Can the task run on the task's current CPU? If so, we're done */ 4532 if (cpumask_test_cpu(task_cpu(p), new_mask)) 4533 goto out; 4534 4535 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask); 4536 if (p->on_rq) { 4537 struct migration_arg arg = { p, dest_cpu }; 4538 /* Need help from migration thread: drop lock and wait. */ 4539 task_rq_unlock(rq, p, &flags); 4540 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg); 4541 tlb_migrate_finish(p->mm); 4542 return 0; 4543 } 4544out: 4545 task_rq_unlock(rq, p, &flags); 4546 4547 return ret; 4548} 4549EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr); 4550 4551/* 4552 * Move (not current) task off this cpu, onto dest cpu. We're doing 4553 * this because either it can't run here any more (set_cpus_allowed() 4554 * away from this CPU, or CPU going down), or because we're 4555 * attempting to rebalance this task on exec (sched_exec). 4556 * 4557 * So we race with normal scheduler movements, but that's OK, as long 4558 * as the task is no longer on this CPU. 4559 * 4560 * Returns non-zero if task was successfully migrated. 4561 */ 4562static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu) 4563{ 4564 struct rq *rq_dest, *rq_src; 4565 int ret = 0; 4566 4567 if (unlikely(!cpu_active(dest_cpu))) 4568 return ret; 4569 4570 rq_src = cpu_rq(src_cpu); 4571 rq_dest = cpu_rq(dest_cpu); 4572 4573 raw_spin_lock(&p->pi_lock); 4574 double_rq_lock(rq_src, rq_dest); 4575 /* Already moved. */ 4576 if (task_cpu(p) != src_cpu) 4577 goto done; 4578 /* Affinity changed (again). */ 4579 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) 4580 goto fail; 4581 4582 /* 4583 * If we're not on a rq, the next wake-up will ensure we're 4584 * placed properly. 4585 */ 4586 if (p->on_rq) { 4587 dequeue_task(rq_src, p, 0); 4588 set_task_cpu(p, dest_cpu); 4589 enqueue_task(rq_dest, p, 0); 4590 check_preempt_curr(rq_dest, p, 0); 4591 } 4592done: 4593 ret = 1; 4594fail: 4595 double_rq_unlock(rq_src, rq_dest); 4596 raw_spin_unlock(&p->pi_lock); 4597 return ret; 4598} 4599 4600#ifdef CONFIG_NUMA_BALANCING 4601/* Migrate current task p to target_cpu */ 4602int migrate_task_to(struct task_struct *p, int target_cpu) 4603{ 4604 struct migration_arg arg = { p, target_cpu }; 4605 int curr_cpu = task_cpu(p); 4606 4607 if (curr_cpu == target_cpu) 4608 return 0; 4609 4610 if (!cpumask_test_cpu(target_cpu, tsk_cpus_allowed(p))) 4611 return -EINVAL; 4612 4613 /* TODO: This is not properly updating schedstats */ 4614 4615 trace_sched_move_numa(p, curr_cpu, target_cpu); 4616 return stop_one_cpu(curr_cpu, migration_cpu_stop, &arg); 4617} 4618 4619/* 4620 * Requeue a task on a given node and accurately track the number of NUMA 4621 * tasks on the runqueues 4622 */ 4623void sched_setnuma(struct task_struct *p, int nid) 4624{ 4625 struct rq *rq; 4626 unsigned long flags; 4627 bool on_rq, running; 4628 4629 rq = task_rq_lock(p, &flags); 4630 on_rq = p->on_rq; 4631 running = task_current(rq, p); 4632 4633 if (on_rq) 4634 dequeue_task(rq, p, 0); 4635 if (running) 4636 p->sched_class->put_prev_task(rq, p); 4637 4638 p->numa_preferred_nid = nid; 4639 4640 if (running) 4641 p->sched_class->set_curr_task(rq); 4642 if (on_rq) 4643 enqueue_task(rq, p, 0); 4644 task_rq_unlock(rq, p, &flags); 4645} 4646#endif 4647 4648/* 4649 * migration_cpu_stop - this will be executed by a highprio stopper thread 4650 * and performs thread migration by bumping thread off CPU then 4651 * 'pushing' onto another runqueue. 4652 */ 4653static int migration_cpu_stop(void *data) 4654{ 4655 struct migration_arg *arg = data; 4656 4657 /* 4658 * The original target cpu might have gone down and we might 4659 * be on another cpu but it doesn't matter. 4660 */ 4661 local_irq_disable(); 4662 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu); 4663 local_irq_enable(); 4664 return 0; 4665} 4666 4667#ifdef CONFIG_HOTPLUG_CPU 4668 4669/* 4670 * Ensures that the idle task is using init_mm right before its cpu goes 4671 * offline. 4672 */ 4673void idle_task_exit(void) 4674{ 4675 struct mm_struct *mm = current->active_mm; 4676 4677 BUG_ON(cpu_online(smp_processor_id())); 4678 4679 if (mm != &init_mm) { 4680 switch_mm(mm, &init_mm, current); 4681 finish_arch_post_lock_switch(); 4682 } 4683 mmdrop(mm); 4684} 4685 4686/* 4687 * Since this CPU is going 'away' for a while, fold any nr_active delta 4688 * we might have. Assumes we're called after migrate_tasks() so that the 4689 * nr_active count is stable. 4690 * 4691 * Also see the comment "Global load-average calculations". 4692 */ 4693static void calc_load_migrate(struct rq *rq) 4694{ 4695 long delta = calc_load_fold_active(rq); 4696 if (delta) 4697 atomic_long_add(delta, &calc_load_tasks); 4698} 4699 4700static void put_prev_task_fake(struct rq *rq, struct task_struct *prev) 4701{ 4702} 4703 4704static const struct sched_class fake_sched_class = { 4705 .put_prev_task = put_prev_task_fake, 4706}; 4707 4708static struct task_struct fake_task = { 4709 /* 4710 * Avoid pull_{rt,dl}_task() 4711 */ 4712 .prio = MAX_PRIO + 1, 4713 .sched_class = &fake_sched_class, 4714}; 4715 4716/* 4717 * Migrate all tasks from the rq, sleeping tasks will be migrated by 4718 * try_to_wake_up()->select_task_rq(). 4719 * 4720 * Called with rq->lock held even though we'er in stop_machine() and 4721 * there's no concurrency possible, we hold the required locks anyway 4722 * because of lock validation efforts. 4723 */ 4724static void migrate_tasks(unsigned int dead_cpu) 4725{ 4726 struct rq *rq = cpu_rq(dead_cpu); 4727 struct task_struct *next, *stop = rq->stop; 4728 int dest_cpu; 4729 4730 /* 4731 * Fudge the rq selection such that the below task selection loop 4732 * doesn't get stuck on the currently eligible stop task. 4733 * 4734 * We're currently inside stop_machine() and the rq is either stuck 4735 * in the stop_machine_cpu_stop() loop, or we're executing this code, 4736 * either way we should never end up calling schedule() until we're 4737 * done here. 4738 */ 4739 rq->stop = NULL; 4740 4741 /* 4742 * put_prev_task() and pick_next_task() sched 4743 * class method both need to have an up-to-date 4744 * value of rq->clock[_task] 4745 */ 4746 update_rq_clock(rq); 4747 4748 for ( ; ; ) { 4749 /* 4750 * There's this thread running, bail when that's the only 4751 * remaining thread. 4752 */ 4753 if (rq->nr_running == 1) 4754 break; 4755 4756 next = pick_next_task(rq, &fake_task); 4757 BUG_ON(!next); 4758 next->sched_class->put_prev_task(rq, next); 4759 4760 /* Find suitable destination for @next, with force if needed. */ 4761 dest_cpu = select_fallback_rq(dead_cpu, next); 4762 raw_spin_unlock(&rq->lock); 4763 4764 __migrate_task(next, dead_cpu, dest_cpu); 4765 4766 raw_spin_lock(&rq->lock); 4767 } 4768 4769 rq->stop = stop; 4770} 4771 4772#endif /* CONFIG_HOTPLUG_CPU */ 4773 4774#if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL) 4775 4776static struct ctl_table sd_ctl_dir[] = { 4777 { 4778 .procname = "sched_domain", 4779 .mode = 0555, 4780 }, 4781 {} 4782}; 4783 4784static struct ctl_table sd_ctl_root[] = { 4785 { 4786 .procname = "kernel", 4787 .mode = 0555, 4788 .child = sd_ctl_dir, 4789 }, 4790 {} 4791}; 4792 4793static struct ctl_table *sd_alloc_ctl_entry(int n) 4794{ 4795 struct ctl_table *entry = 4796 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL); 4797 4798 return entry; 4799} 4800 4801static void sd_free_ctl_entry(struct ctl_table **tablep) 4802{ 4803 struct ctl_table *entry; 4804 4805 /* 4806 * In the intermediate directories, both the child directory and 4807 * procname are dynamically allocated and could fail but the mode 4808 * will always be set. In the lowest directory the names are 4809 * static strings and all have proc handlers. 4810 */ 4811 for (entry = *tablep; entry->mode; entry++) { 4812 if (entry->child) 4813 sd_free_ctl_entry(&entry->child); 4814 if (entry->proc_handler == NULL) 4815 kfree(entry->procname); 4816 } 4817 4818 kfree(*tablep); 4819 *tablep = NULL; 4820} 4821 4822static int min_load_idx = 0; 4823static int max_load_idx = CPU_LOAD_IDX_MAX-1; 4824 4825static void 4826set_table_entry(struct ctl_table *entry, 4827 const char *procname, void *data, int maxlen, 4828 umode_t mode, proc_handler *proc_handler, 4829 bool load_idx) 4830{ 4831 entry->procname = procname; 4832 entry->data = data; 4833 entry->maxlen = maxlen; 4834 entry->mode = mode; 4835 entry->proc_handler = proc_handler; 4836 4837 if (load_idx) { 4838 entry->extra1 = &min_load_idx; 4839 entry->extra2 = &max_load_idx; 4840 } 4841} 4842 4843static struct ctl_table * 4844sd_alloc_ctl_domain_table(struct sched_domain *sd) 4845{ 4846 struct ctl_table *table = sd_alloc_ctl_entry(14); 4847 4848 if (table == NULL) 4849 return NULL; 4850 4851 set_table_entry(&table[0], "min_interval", &sd->min_interval, 4852 sizeof(long), 0644, proc_doulongvec_minmax, false); 4853 set_table_entry(&table[1], "max_interval", &sd->max_interval, 4854 sizeof(long), 0644, proc_doulongvec_minmax, false); 4855 set_table_entry(&table[2], "busy_idx", &sd->busy_idx, 4856 sizeof(int), 0644, proc_dointvec_minmax, true); 4857 set_table_entry(&table[3], "idle_idx", &sd->idle_idx, 4858 sizeof(int), 0644, proc_dointvec_minmax, true); 4859 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx, 4860 sizeof(int), 0644, proc_dointvec_minmax, true); 4861 set_table_entry(&table[5], "wake_idx", &sd->wake_idx, 4862 sizeof(int), 0644, proc_dointvec_minmax, true); 4863 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx, 4864 sizeof(int), 0644, proc_dointvec_minmax, true); 4865 set_table_entry(&table[7], "busy_factor", &sd->busy_factor, 4866 sizeof(int), 0644, proc_dointvec_minmax, false); 4867 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct, 4868 sizeof(int), 0644, proc_dointvec_minmax, false); 4869 set_table_entry(&table[9], "cache_nice_tries", 4870 &sd->cache_nice_tries, 4871 sizeof(int), 0644, proc_dointvec_minmax, false); 4872 set_table_entry(&table[10], "flags", &sd->flags, 4873 sizeof(int), 0644, proc_dointvec_minmax, false); 4874 set_table_entry(&table[11], "max_newidle_lb_cost", 4875 &sd->max_newidle_lb_cost, 4876 sizeof(long), 0644, proc_doulongvec_minmax, false); 4877 set_table_entry(&table[12], "name", sd->name, 4878 CORENAME_MAX_SIZE, 0444, proc_dostring, false); 4879 /* &table[13] is terminator */ 4880 4881 return table; 4882} 4883 4884static struct ctl_table *sd_alloc_ctl_cpu_table(int cpu) 4885{ 4886 struct ctl_table *entry, *table; 4887 struct sched_domain *sd; 4888 int domain_num = 0, i; 4889 char buf[32]; 4890 4891 for_each_domain(cpu, sd) 4892 domain_num++; 4893 entry = table = sd_alloc_ctl_entry(domain_num + 1); 4894 if (table == NULL) 4895 return NULL; 4896 4897 i = 0; 4898 for_each_domain(cpu, sd) { 4899 snprintf(buf, 32, "domain%d", i); 4900 entry->procname = kstrdup(buf, GFP_KERNEL); 4901 entry->mode = 0555; 4902 entry->child = sd_alloc_ctl_domain_table(sd); 4903 entry++; 4904 i++; 4905 } 4906 return table; 4907} 4908 4909static struct ctl_table_header *sd_sysctl_header; 4910static void register_sched_domain_sysctl(void) 4911{ 4912 int i, cpu_num = num_possible_cpus(); 4913 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1); 4914 char buf[32]; 4915 4916 WARN_ON(sd_ctl_dir[0].child); 4917 sd_ctl_dir[0].child = entry; 4918 4919 if (entry == NULL) 4920 return; 4921 4922 for_each_possible_cpu(i) { 4923 snprintf(buf, 32, "cpu%d", i); 4924 entry->procname = kstrdup(buf, GFP_KERNEL); 4925 entry->mode = 0555; 4926 entry->child = sd_alloc_ctl_cpu_table(i); 4927 entry++; 4928 } 4929 4930 WARN_ON(sd_sysctl_header); 4931 sd_sysctl_header = register_sysctl_table(sd_ctl_root); 4932} 4933 4934/* may be called multiple times per register */ 4935static void unregister_sched_domain_sysctl(void) 4936{ 4937 if (sd_sysctl_header) 4938 unregister_sysctl_table(sd_sysctl_header); 4939 sd_sysctl_header = NULL; 4940 if (sd_ctl_dir[0].child) 4941 sd_free_ctl_entry(&sd_ctl_dir[0].child); 4942} 4943#else 4944static void register_sched_domain_sysctl(void) 4945{ 4946} 4947static void unregister_sched_domain_sysctl(void) 4948{ 4949} 4950#endif 4951 4952static void set_rq_online(struct rq *rq) 4953{ 4954 if (!rq->online) { 4955 const struct sched_class *class; 4956 4957 cpumask_set_cpu(rq->cpu, rq->rd->online); 4958 rq->online = 1; 4959 4960 for_each_class(class) { 4961 if (class->rq_online) 4962 class->rq_online(rq); 4963 } 4964 } 4965} 4966 4967static void set_rq_offline(struct rq *rq) 4968{ 4969 if (rq->online) { 4970 const struct sched_class *class; 4971 4972 for_each_class(class) { 4973 if (class->rq_offline) 4974 class->rq_offline(rq); 4975 } 4976 4977 cpumask_clear_cpu(rq->cpu, rq->rd->online); 4978 rq->online = 0; 4979 } 4980} 4981 4982/* 4983 * migration_call - callback that gets triggered when a CPU is added. 4984 * Here we can start up the necessary migration thread for the new CPU. 4985 */ 4986static int 4987migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu) 4988{ 4989 int cpu = (long)hcpu; 4990 unsigned long flags; 4991 struct rq *rq = cpu_rq(cpu); 4992 4993 switch (action & ~CPU_TASKS_FROZEN) { 4994 4995 case CPU_UP_PREPARE: 4996 rq->calc_load_update = calc_load_update; 4997 break; 4998 4999 case CPU_ONLINE: 5000 /* Update our root-domain */ 5001 raw_spin_lock_irqsave(&rq->lock, flags); 5002 if (rq->rd) { 5003 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 5004 5005 set_rq_online(rq); 5006 } 5007 raw_spin_unlock_irqrestore(&rq->lock, flags); 5008 break; 5009 5010#ifdef CONFIG_HOTPLUG_CPU 5011 case CPU_DYING: 5012 sched_ttwu_pending(); 5013 /* Update our root-domain */ 5014 raw_spin_lock_irqsave(&rq->lock, flags); 5015 if (rq->rd) { 5016 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 5017 set_rq_offline(rq); 5018 } 5019 migrate_tasks(cpu); 5020 BUG_ON(rq->nr_running != 1); /* the migration thread */ 5021 raw_spin_unlock_irqrestore(&rq->lock, flags); 5022 break; 5023 5024 case CPU_DEAD: 5025 calc_load_migrate(rq); 5026 break; 5027#endif 5028 } 5029 5030 update_max_interval(); 5031 5032 return NOTIFY_OK; 5033} 5034 5035/* 5036 * Register at high priority so that task migration (migrate_all_tasks) 5037 * happens before everything else. This has to be lower priority than 5038 * the notifier in the perf_event subsystem, though. 5039 */ 5040static struct notifier_block migration_notifier = { 5041 .notifier_call = migration_call, 5042 .priority = CPU_PRI_MIGRATION, 5043}; 5044 5045static int sched_cpu_active(struct notifier_block *nfb, 5046 unsigned long action, void *hcpu) 5047{ 5048 switch (action & ~CPU_TASKS_FROZEN) { 5049 case CPU_STARTING: 5050 case CPU_DOWN_FAILED: 5051 set_cpu_active((long)hcpu, true); 5052 return NOTIFY_OK; 5053 default: 5054 return NOTIFY_DONE; 5055 } 5056} 5057 5058static int sched_cpu_inactive(struct notifier_block *nfb, 5059 unsigned long action, void *hcpu) 5060{ 5061 unsigned long flags; 5062 long cpu = (long)hcpu; 5063 5064 switch (action & ~CPU_TASKS_FROZEN) { 5065 case CPU_DOWN_PREPARE: 5066 set_cpu_active(cpu, false); 5067 5068 /* explicitly allow suspend */ 5069 if (!(action & CPU_TASKS_FROZEN)) { 5070 struct dl_bw *dl_b = dl_bw_of(cpu); 5071 bool overflow; 5072 int cpus; 5073 5074 raw_spin_lock_irqsave(&dl_b->lock, flags); 5075 cpus = dl_bw_cpus(cpu); 5076 overflow = __dl_overflow(dl_b, cpus, 0, 0); 5077 raw_spin_unlock_irqrestore(&dl_b->lock, flags); 5078 5079 if (overflow) 5080 return notifier_from_errno(-EBUSY); 5081 } 5082 return NOTIFY_OK; 5083 } 5084 5085 return NOTIFY_DONE; 5086} 5087 5088static int __init migration_init(void) 5089{ 5090 void *cpu = (void *)(long)smp_processor_id(); 5091 int err; 5092 5093 /* Initialize migration for the boot CPU */ 5094 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu); 5095 BUG_ON(err == NOTIFY_BAD); 5096 migration_call(&migration_notifier, CPU_ONLINE, cpu); 5097 register_cpu_notifier(&migration_notifier); 5098 5099 /* Register cpu active notifiers */ 5100 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE); 5101 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE); 5102 5103 return 0; 5104} 5105early_initcall(migration_init); 5106#endif 5107 5108#ifdef CONFIG_SMP 5109 5110static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */ 5111 5112#ifdef CONFIG_SCHED_DEBUG 5113 5114static __read_mostly int sched_debug_enabled; 5115 5116static int __init sched_debug_setup(char *str) 5117{ 5118 sched_debug_enabled = 1; 5119 5120 return 0; 5121} 5122early_param("sched_debug", sched_debug_setup); 5123 5124static inline bool sched_debug(void) 5125{ 5126 return sched_debug_enabled; 5127} 5128 5129static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level, 5130 struct cpumask *groupmask) 5131{ 5132 struct sched_group *group = sd->groups; 5133 char str[256]; 5134 5135 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd)); 5136 cpumask_clear(groupmask); 5137 5138 printk(KERN_DEBUG "%*s domain %d: ", level, "", level); 5139 5140 if (!(sd->flags & SD_LOAD_BALANCE)) { 5141 printk("does not load-balance\n"); 5142 if (sd->parent) 5143 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain" 5144 " has parent"); 5145 return -1; 5146 } 5147 5148 printk(KERN_CONT "span %s level %s\n", str, sd->name); 5149 5150 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) { 5151 printk(KERN_ERR "ERROR: domain->span does not contain " 5152 "CPU%d\n", cpu); 5153 } 5154 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) { 5155 printk(KERN_ERR "ERROR: domain->groups does not contain" 5156 " CPU%d\n", cpu); 5157 } 5158 5159 printk(KERN_DEBUG "%*s groups:", level + 1, ""); 5160 do { 5161 if (!group) { 5162 printk("\n"); 5163 printk(KERN_ERR "ERROR: group is NULL\n"); 5164 break; 5165 } 5166 5167 /* 5168 * Even though we initialize ->power to something semi-sane, 5169 * we leave power_orig unset. This allows us to detect if 5170 * domain iteration is still funny without causing /0 traps. 5171 */ 5172 if (!group->sgp->power_orig) { 5173 printk(KERN_CONT "\n"); 5174 printk(KERN_ERR "ERROR: domain->cpu_power not " 5175 "set\n"); 5176 break; 5177 } 5178 5179 if (!cpumask_weight(sched_group_cpus(group))) { 5180 printk(KERN_CONT "\n"); 5181 printk(KERN_ERR "ERROR: empty group\n"); 5182 break; 5183 } 5184 5185 if (!(sd->flags & SD_OVERLAP) && 5186 cpumask_intersects(groupmask, sched_group_cpus(group))) { 5187 printk(KERN_CONT "\n"); 5188 printk(KERN_ERR "ERROR: repeated CPUs\n"); 5189 break; 5190 } 5191 5192 cpumask_or(groupmask, groupmask, sched_group_cpus(group)); 5193 5194 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group)); 5195 5196 printk(KERN_CONT " %s", str); 5197 if (group->sgp->power != SCHED_POWER_SCALE) { 5198 printk(KERN_CONT " (cpu_power = %d)", 5199 group->sgp->power); 5200 } 5201 5202 group = group->next; 5203 } while (group != sd->groups); 5204 printk(KERN_CONT "\n"); 5205 5206 if (!cpumask_equal(sched_domain_span(sd), groupmask)) 5207 printk(KERN_ERR "ERROR: groups don't span domain->span\n"); 5208 5209 if (sd->parent && 5210 !cpumask_subset(groupmask, sched_domain_span(sd->parent))) 5211 printk(KERN_ERR "ERROR: parent span is not a superset " 5212 "of domain->span\n"); 5213 return 0; 5214} 5215 5216static void sched_domain_debug(struct sched_domain *sd, int cpu) 5217{ 5218 int level = 0; 5219 5220 if (!sched_debug_enabled) 5221 return; 5222 5223 if (!sd) { 5224 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu); 5225 return; 5226 } 5227 5228 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu); 5229 5230 for (;;) { 5231 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask)) 5232 break; 5233 level++; 5234 sd = sd->parent; 5235 if (!sd) 5236 break; 5237 } 5238} 5239#else /* !CONFIG_SCHED_DEBUG */ 5240# define sched_domain_debug(sd, cpu) do { } while (0) 5241static inline bool sched_debug(void) 5242{ 5243 return false; 5244} 5245#endif /* CONFIG_SCHED_DEBUG */ 5246 5247static int sd_degenerate(struct sched_domain *sd) 5248{ 5249 if (cpumask_weight(sched_domain_span(sd)) == 1) 5250 return 1; 5251 5252 /* Following flags need at least 2 groups */ 5253 if (sd->flags & (SD_LOAD_BALANCE | 5254 SD_BALANCE_NEWIDLE | 5255 SD_BALANCE_FORK | 5256 SD_BALANCE_EXEC | 5257 SD_SHARE_CPUPOWER | 5258 SD_SHARE_PKG_RESOURCES)) { 5259 if (sd->groups != sd->groups->next) 5260 return 0; 5261 } 5262 5263 /* Following flags don't use groups */ 5264 if (sd->flags & (SD_WAKE_AFFINE)) 5265 return 0; 5266 5267 return 1; 5268} 5269 5270static int 5271sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent) 5272{ 5273 unsigned long cflags = sd->flags, pflags = parent->flags; 5274 5275 if (sd_degenerate(parent)) 5276 return 1; 5277 5278 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent))) 5279 return 0; 5280 5281 /* Flags needing groups don't count if only 1 group in parent */ 5282 if (parent->groups == parent->groups->next) { 5283 pflags &= ~(SD_LOAD_BALANCE | 5284 SD_BALANCE_NEWIDLE | 5285 SD_BALANCE_FORK | 5286 SD_BALANCE_EXEC | 5287 SD_SHARE_CPUPOWER | 5288 SD_SHARE_PKG_RESOURCES | 5289 SD_PREFER_SIBLING); 5290 if (nr_node_ids == 1) 5291 pflags &= ~SD_SERIALIZE; 5292 } 5293 if (~cflags & pflags) 5294 return 0; 5295 5296 return 1; 5297} 5298 5299static void free_rootdomain(struct rcu_head *rcu) 5300{ 5301 struct root_domain *rd = container_of(rcu, struct root_domain, rcu); 5302 5303 cpupri_cleanup(&rd->cpupri); 5304 cpudl_cleanup(&rd->cpudl); 5305 free_cpumask_var(rd->dlo_mask); 5306 free_cpumask_var(rd->rto_mask); 5307 free_cpumask_var(rd->online); 5308 free_cpumask_var(rd->span); 5309 kfree(rd); 5310} 5311 5312static void rq_attach_root(struct rq *rq, struct root_domain *rd) 5313{ 5314 struct root_domain *old_rd = NULL; 5315 unsigned long flags; 5316 5317 raw_spin_lock_irqsave(&rq->lock, flags); 5318 5319 if (rq->rd) { 5320 old_rd = rq->rd; 5321 5322 if (cpumask_test_cpu(rq->cpu, old_rd->online)) 5323 set_rq_offline(rq); 5324 5325 cpumask_clear_cpu(rq->cpu, old_rd->span); 5326 5327 /* 5328 * If we dont want to free the old_rd yet then 5329 * set old_rd to NULL to skip the freeing later 5330 * in this function: 5331 */ 5332 if (!atomic_dec_and_test(&old_rd->refcount)) 5333 old_rd = NULL; 5334 } 5335 5336 atomic_inc(&rd->refcount); 5337 rq->rd = rd; 5338 5339 cpumask_set_cpu(rq->cpu, rd->span); 5340 if (cpumask_test_cpu(rq->cpu, cpu_active_mask)) 5341 set_rq_online(rq); 5342 5343 raw_spin_unlock_irqrestore(&rq->lock, flags); 5344 5345 if (old_rd) 5346 call_rcu_sched(&old_rd->rcu, free_rootdomain); 5347} 5348 5349static int init_rootdomain(struct root_domain *rd) 5350{ 5351 memset(rd, 0, sizeof(*rd)); 5352 5353 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL)) 5354 goto out; 5355 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL)) 5356 goto free_span; 5357 if (!alloc_cpumask_var(&rd->dlo_mask, GFP_KERNEL)) 5358 goto free_online; 5359 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL)) 5360 goto free_dlo_mask; 5361 5362 init_dl_bw(&rd->dl_bw); 5363 if (cpudl_init(&rd->cpudl) != 0) 5364 goto free_dlo_mask; 5365 5366 if (cpupri_init(&rd->cpupri) != 0) 5367 goto free_rto_mask; 5368 return 0; 5369 5370free_rto_mask: 5371 free_cpumask_var(rd->rto_mask); 5372free_dlo_mask: 5373 free_cpumask_var(rd->dlo_mask); 5374free_online: 5375 free_cpumask_var(rd->online); 5376free_span: 5377 free_cpumask_var(rd->span); 5378out: 5379 return -ENOMEM; 5380} 5381 5382/* 5383 * By default the system creates a single root-domain with all cpus as 5384 * members (mimicking the global state we have today). 5385 */ 5386struct root_domain def_root_domain; 5387 5388static void init_defrootdomain(void) 5389{ 5390 init_rootdomain(&def_root_domain); 5391 5392 atomic_set(&def_root_domain.refcount, 1); 5393} 5394 5395static struct root_domain *alloc_rootdomain(void) 5396{ 5397 struct root_domain *rd; 5398 5399 rd = kmalloc(sizeof(*rd), GFP_KERNEL); 5400 if (!rd) 5401 return NULL; 5402 5403 if (init_rootdomain(rd) != 0) { 5404 kfree(rd); 5405 return NULL; 5406 } 5407 5408 return rd; 5409} 5410 5411static void free_sched_groups(struct sched_group *sg, int free_sgp) 5412{ 5413 struct sched_group *tmp, *first; 5414 5415 if (!sg) 5416 return; 5417 5418 first = sg; 5419 do { 5420 tmp = sg->next; 5421 5422 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref)) 5423 kfree(sg->sgp); 5424 5425 kfree(sg); 5426 sg = tmp; 5427 } while (sg != first); 5428} 5429 5430static void free_sched_domain(struct rcu_head *rcu) 5431{ 5432 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu); 5433 5434 /* 5435 * If its an overlapping domain it has private groups, iterate and 5436 * nuke them all. 5437 */ 5438 if (sd->flags & SD_OVERLAP) { 5439 free_sched_groups(sd->groups, 1); 5440 } else if (atomic_dec_and_test(&sd->groups->ref)) { 5441 kfree(sd->groups->sgp); 5442 kfree(sd->groups); 5443 } 5444 kfree(sd); 5445} 5446 5447static void destroy_sched_domain(struct sched_domain *sd, int cpu) 5448{ 5449 call_rcu(&sd->rcu, free_sched_domain); 5450} 5451 5452static void destroy_sched_domains(struct sched_domain *sd, int cpu) 5453{ 5454 for (; sd; sd = sd->parent) 5455 destroy_sched_domain(sd, cpu); 5456} 5457 5458/* 5459 * Keep a special pointer to the highest sched_domain that has 5460 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this 5461 * allows us to avoid some pointer chasing select_idle_sibling(). 5462 * 5463 * Also keep a unique ID per domain (we use the first cpu number in 5464 * the cpumask of the domain), this allows us to quickly tell if 5465 * two cpus are in the same cache domain, see cpus_share_cache(). 5466 */ 5467DEFINE_PER_CPU(struct sched_domain *, sd_llc); 5468DEFINE_PER_CPU(int, sd_llc_size); 5469DEFINE_PER_CPU(int, sd_llc_id); 5470DEFINE_PER_CPU(struct sched_domain *, sd_numa); 5471DEFINE_PER_CPU(struct sched_domain *, sd_busy); 5472DEFINE_PER_CPU(struct sched_domain *, sd_asym); 5473 5474static void update_top_cache_domain(int cpu) 5475{ 5476 struct sched_domain *sd; 5477 struct sched_domain *busy_sd = NULL; 5478 int id = cpu; 5479 int size = 1; 5480 5481 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES); 5482 if (sd) { 5483 id = cpumask_first(sched_domain_span(sd)); 5484 size = cpumask_weight(sched_domain_span(sd)); 5485 busy_sd = sd->parent; /* sd_busy */ 5486 } 5487 rcu_assign_pointer(per_cpu(sd_busy, cpu), busy_sd); 5488 5489 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd); 5490 per_cpu(sd_llc_size, cpu) = size; 5491 per_cpu(sd_llc_id, cpu) = id; 5492 5493 sd = lowest_flag_domain(cpu, SD_NUMA); 5494 rcu_assign_pointer(per_cpu(sd_numa, cpu), sd); 5495 5496 sd = highest_flag_domain(cpu, SD_ASYM_PACKING); 5497 rcu_assign_pointer(per_cpu(sd_asym, cpu), sd); 5498} 5499 5500/* 5501 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must 5502 * hold the hotplug lock. 5503 */ 5504static void 5505cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu) 5506{ 5507 struct rq *rq = cpu_rq(cpu); 5508 struct sched_domain *tmp; 5509 5510 /* Remove the sched domains which do not contribute to scheduling. */ 5511 for (tmp = sd; tmp; ) { 5512 struct sched_domain *parent = tmp->parent; 5513 if (!parent) 5514 break; 5515 5516 if (sd_parent_degenerate(tmp, parent)) { 5517 tmp->parent = parent->parent; 5518 if (parent->parent) 5519 parent->parent->child = tmp; 5520 /* 5521 * Transfer SD_PREFER_SIBLING down in case of a 5522 * degenerate parent; the spans match for this 5523 * so the property transfers. 5524 */ 5525 if (parent->flags & SD_PREFER_SIBLING) 5526 tmp->flags |= SD_PREFER_SIBLING; 5527 destroy_sched_domain(parent, cpu); 5528 } else 5529 tmp = tmp->parent; 5530 } 5531 5532 if (sd && sd_degenerate(sd)) { 5533 tmp = sd; 5534 sd = sd->parent; 5535 destroy_sched_domain(tmp, cpu); 5536 if (sd) 5537 sd->child = NULL; 5538 } 5539 5540 sched_domain_debug(sd, cpu); 5541 5542 rq_attach_root(rq, rd); 5543 tmp = rq->sd; 5544 rcu_assign_pointer(rq->sd, sd); 5545 destroy_sched_domains(tmp, cpu); 5546 5547 update_top_cache_domain(cpu); 5548} 5549 5550/* cpus with isolated domains */ 5551static cpumask_var_t cpu_isolated_map; 5552 5553/* Setup the mask of cpus configured for isolated domains */ 5554static int __init isolated_cpu_setup(char *str) 5555{ 5556 alloc_bootmem_cpumask_var(&cpu_isolated_map); 5557 cpulist_parse(str, cpu_isolated_map); 5558 return 1; 5559} 5560 5561__setup("isolcpus=", isolated_cpu_setup); 5562 5563static const struct cpumask *cpu_cpu_mask(int cpu) 5564{ 5565 return cpumask_of_node(cpu_to_node(cpu)); 5566} 5567 5568struct sd_data { 5569 struct sched_domain **__percpu sd; 5570 struct sched_group **__percpu sg; 5571 struct sched_group_power **__percpu sgp; 5572}; 5573 5574struct s_data { 5575 struct sched_domain ** __percpu sd; 5576 struct root_domain *rd; 5577}; 5578 5579enum s_alloc { 5580 sa_rootdomain, 5581 sa_sd, 5582 sa_sd_storage, 5583 sa_none, 5584}; 5585 5586struct sched_domain_topology_level; 5587 5588typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu); 5589typedef const struct cpumask *(*sched_domain_mask_f)(int cpu); 5590 5591#define SDTL_OVERLAP 0x01 5592 5593struct sched_domain_topology_level { 5594 sched_domain_init_f init; 5595 sched_domain_mask_f mask; 5596 int flags; 5597 int numa_level; 5598 struct sd_data data; 5599}; 5600 5601/* 5602 * Build an iteration mask that can exclude certain CPUs from the upwards 5603 * domain traversal. 5604 * 5605 * Asymmetric node setups can result in situations where the domain tree is of 5606 * unequal depth, make sure to skip domains that already cover the entire 5607 * range. 5608 * 5609 * In that case build_sched_domains() will have terminated the iteration early 5610 * and our sibling sd spans will be empty. Domains should always include the 5611 * cpu they're built on, so check that. 5612 * 5613 */ 5614static void build_group_mask(struct sched_domain *sd, struct sched_group *sg) 5615{ 5616 const struct cpumask *span = sched_domain_span(sd); 5617 struct sd_data *sdd = sd->private; 5618 struct sched_domain *sibling; 5619 int i; 5620 5621 for_each_cpu(i, span) { 5622 sibling = *per_cpu_ptr(sdd->sd, i); 5623 if (!cpumask_test_cpu(i, sched_domain_span(sibling))) 5624 continue; 5625 5626 cpumask_set_cpu(i, sched_group_mask(sg)); 5627 } 5628} 5629 5630/* 5631 * Return the canonical balance cpu for this group, this is the first cpu 5632 * of this group that's also in the iteration mask. 5633 */ 5634int group_balance_cpu(struct sched_group *sg) 5635{ 5636 return cpumask_first_and(sched_group_cpus(sg), sched_group_mask(sg)); 5637} 5638 5639static int 5640build_overlap_sched_groups(struct sched_domain *sd, int cpu) 5641{ 5642 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg; 5643 const struct cpumask *span = sched_domain_span(sd); 5644 struct cpumask *covered = sched_domains_tmpmask; 5645 struct sd_data *sdd = sd->private; 5646 struct sched_domain *child; 5647 int i; 5648 5649 cpumask_clear(covered); 5650 5651 for_each_cpu(i, span) { 5652 struct cpumask *sg_span; 5653 5654 if (cpumask_test_cpu(i, covered)) 5655 continue; 5656 5657 child = *per_cpu_ptr(sdd->sd, i); 5658 5659 /* See the comment near build_group_mask(). */ 5660 if (!cpumask_test_cpu(i, sched_domain_span(child))) 5661 continue; 5662 5663 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 5664 GFP_KERNEL, cpu_to_node(cpu)); 5665 5666 if (!sg) 5667 goto fail; 5668 5669 sg_span = sched_group_cpus(sg); 5670 if (child->child) { 5671 child = child->child; 5672 cpumask_copy(sg_span, sched_domain_span(child)); 5673 } else 5674 cpumask_set_cpu(i, sg_span); 5675 5676 cpumask_or(covered, covered, sg_span); 5677 5678 sg->sgp = *per_cpu_ptr(sdd->sgp, i); 5679 if (atomic_inc_return(&sg->sgp->ref) == 1) 5680 build_group_mask(sd, sg); 5681 5682 /* 5683 * Initialize sgp->power such that even if we mess up the 5684 * domains and no possible iteration will get us here, we won't 5685 * die on a /0 trap. 5686 */ 5687 sg->sgp->power = SCHED_POWER_SCALE * cpumask_weight(sg_span); 5688 sg->sgp->power_orig = sg->sgp->power; 5689 5690 /* 5691 * Make sure the first group of this domain contains the 5692 * canonical balance cpu. Otherwise the sched_domain iteration 5693 * breaks. See update_sg_lb_stats(). 5694 */ 5695 if ((!groups && cpumask_test_cpu(cpu, sg_span)) || 5696 group_balance_cpu(sg) == cpu) 5697 groups = sg; 5698 5699 if (!first) 5700 first = sg; 5701 if (last) 5702 last->next = sg; 5703 last = sg; 5704 last->next = first; 5705 } 5706 sd->groups = groups; 5707 5708 return 0; 5709 5710fail: 5711 free_sched_groups(first, 0); 5712 5713 return -ENOMEM; 5714} 5715 5716static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg) 5717{ 5718 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu); 5719 struct sched_domain *child = sd->child; 5720 5721 if (child) 5722 cpu = cpumask_first(sched_domain_span(child)); 5723 5724 if (sg) { 5725 *sg = *per_cpu_ptr(sdd->sg, cpu); 5726 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu); 5727 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */ 5728 } 5729 5730 return cpu; 5731} 5732 5733/* 5734 * build_sched_groups will build a circular linked list of the groups 5735 * covered by the given span, and will set each group's ->cpumask correctly, 5736 * and ->cpu_power to 0. 5737 * 5738 * Assumes the sched_domain tree is fully constructed 5739 */ 5740static int 5741build_sched_groups(struct sched_domain *sd, int cpu) 5742{ 5743 struct sched_group *first = NULL, *last = NULL; 5744 struct sd_data *sdd = sd->private; 5745 const struct cpumask *span = sched_domain_span(sd); 5746 struct cpumask *covered; 5747 int i; 5748 5749 get_group(cpu, sdd, &sd->groups); 5750 atomic_inc(&sd->groups->ref); 5751 5752 if (cpu != cpumask_first(span)) 5753 return 0; 5754 5755 lockdep_assert_held(&sched_domains_mutex); 5756 covered = sched_domains_tmpmask; 5757 5758 cpumask_clear(covered); 5759 5760 for_each_cpu(i, span) { 5761 struct sched_group *sg; 5762 int group, j; 5763 5764 if (cpumask_test_cpu(i, covered)) 5765 continue; 5766 5767 group = get_group(i, sdd, &sg); 5768 cpumask_clear(sched_group_cpus(sg)); 5769 sg->sgp->power = 0; 5770 cpumask_setall(sched_group_mask(sg)); 5771 5772 for_each_cpu(j, span) { 5773 if (get_group(j, sdd, NULL) != group) 5774 continue; 5775 5776 cpumask_set_cpu(j, covered); 5777 cpumask_set_cpu(j, sched_group_cpus(sg)); 5778 } 5779 5780 if (!first) 5781 first = sg; 5782 if (last) 5783 last->next = sg; 5784 last = sg; 5785 } 5786 last->next = first; 5787 5788 return 0; 5789} 5790 5791/* 5792 * Initialize sched groups cpu_power. 5793 * 5794 * cpu_power indicates the capacity of sched group, which is used while 5795 * distributing the load between different sched groups in a sched domain. 5796 * Typically cpu_power for all the groups in a sched domain will be same unless 5797 * there are asymmetries in the topology. If there are asymmetries, group 5798 * having more cpu_power will pickup more load compared to the group having 5799 * less cpu_power. 5800 */ 5801static void init_sched_groups_power(int cpu, struct sched_domain *sd) 5802{ 5803 struct sched_group *sg = sd->groups; 5804 5805 WARN_ON(!sg); 5806 5807 do { 5808 sg->group_weight = cpumask_weight(sched_group_cpus(sg)); 5809 sg = sg->next; 5810 } while (sg != sd->groups); 5811 5812 if (cpu != group_balance_cpu(sg)) 5813 return; 5814 5815 update_group_power(sd, cpu); 5816 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight); 5817} 5818 5819int __weak arch_sd_sibling_asym_packing(void) 5820{ 5821 return 0*SD_ASYM_PACKING; 5822} 5823 5824/* 5825 * Initializers for schedule domains 5826 * Non-inlined to reduce accumulated stack pressure in build_sched_domains() 5827 */ 5828 5829#ifdef CONFIG_SCHED_DEBUG 5830# define SD_INIT_NAME(sd, type) sd->name = #type 5831#else 5832# define SD_INIT_NAME(sd, type) do { } while (0) 5833#endif 5834 5835#define SD_INIT_FUNC(type) \ 5836static noinline struct sched_domain * \ 5837sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \ 5838{ \ 5839 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \ 5840 *sd = SD_##type##_INIT; \ 5841 SD_INIT_NAME(sd, type); \ 5842 sd->private = &tl->data; \ 5843 return sd; \ 5844} 5845 5846SD_INIT_FUNC(CPU) 5847#ifdef CONFIG_SCHED_SMT 5848 SD_INIT_FUNC(SIBLING) 5849#endif 5850#ifdef CONFIG_SCHED_MC 5851 SD_INIT_FUNC(MC) 5852#endif 5853#ifdef CONFIG_SCHED_BOOK 5854 SD_INIT_FUNC(BOOK) 5855#endif 5856 5857static int default_relax_domain_level = -1; 5858int sched_domain_level_max; 5859 5860static int __init setup_relax_domain_level(char *str) 5861{ 5862 if (kstrtoint(str, 0, &default_relax_domain_level)) 5863 pr_warn("Unable to set relax_domain_level\n"); 5864 5865 return 1; 5866} 5867__setup("relax_domain_level=", setup_relax_domain_level); 5868 5869static void set_domain_attribute(struct sched_domain *sd, 5870 struct sched_domain_attr *attr) 5871{ 5872 int request; 5873 5874 if (!attr || attr->relax_domain_level < 0) { 5875 if (default_relax_domain_level < 0) 5876 return; 5877 else 5878 request = default_relax_domain_level; 5879 } else 5880 request = attr->relax_domain_level; 5881 if (request < sd->level) { 5882 /* turn off idle balance on this domain */ 5883 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 5884 } else { 5885 /* turn on idle balance on this domain */ 5886 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 5887 } 5888} 5889 5890static void __sdt_free(const struct cpumask *cpu_map); 5891static int __sdt_alloc(const struct cpumask *cpu_map); 5892 5893static void __free_domain_allocs(struct s_data *d, enum s_alloc what, 5894 const struct cpumask *cpu_map) 5895{ 5896 switch (what) { 5897 case sa_rootdomain: 5898 if (!atomic_read(&d->rd->refcount)) 5899 free_rootdomain(&d->rd->rcu); /* fall through */ 5900 case sa_sd: 5901 free_percpu(d->sd); /* fall through */ 5902 case sa_sd_storage: 5903 __sdt_free(cpu_map); /* fall through */ 5904 case sa_none: 5905 break; 5906 } 5907} 5908 5909static enum s_alloc __visit_domain_allocation_hell(struct s_data *d, 5910 const struct cpumask *cpu_map) 5911{ 5912 memset(d, 0, sizeof(*d)); 5913 5914 if (__sdt_alloc(cpu_map)) 5915 return sa_sd_storage; 5916 d->sd = alloc_percpu(struct sched_domain *); 5917 if (!d->sd) 5918 return sa_sd_storage; 5919 d->rd = alloc_rootdomain(); 5920 if (!d->rd) 5921 return sa_sd; 5922 return sa_rootdomain; 5923} 5924 5925/* 5926 * NULL the sd_data elements we've used to build the sched_domain and 5927 * sched_group structure so that the subsequent __free_domain_allocs() 5928 * will not free the data we're using. 5929 */ 5930static void claim_allocations(int cpu, struct sched_domain *sd) 5931{ 5932 struct sd_data *sdd = sd->private; 5933 5934 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd); 5935 *per_cpu_ptr(sdd->sd, cpu) = NULL; 5936 5937 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref)) 5938 *per_cpu_ptr(sdd->sg, cpu) = NULL; 5939 5940 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref)) 5941 *per_cpu_ptr(sdd->sgp, cpu) = NULL; 5942} 5943 5944#ifdef CONFIG_SCHED_SMT 5945static const struct cpumask *cpu_smt_mask(int cpu) 5946{ 5947 return topology_thread_cpumask(cpu); 5948} 5949#endif 5950 5951/* 5952 * Topology list, bottom-up. 5953 */ 5954static struct sched_domain_topology_level default_topology[] = { 5955#ifdef CONFIG_SCHED_SMT 5956 { sd_init_SIBLING, cpu_smt_mask, }, 5957#endif 5958#ifdef CONFIG_SCHED_MC 5959 { sd_init_MC, cpu_coregroup_mask, }, 5960#endif 5961#ifdef CONFIG_SCHED_BOOK 5962 { sd_init_BOOK, cpu_book_mask, }, 5963#endif 5964 { sd_init_CPU, cpu_cpu_mask, }, 5965 { NULL, }, 5966}; 5967 5968static struct sched_domain_topology_level *sched_domain_topology = default_topology; 5969 5970#define for_each_sd_topology(tl) \ 5971 for (tl = sched_domain_topology; tl->init; tl++) 5972 5973#ifdef CONFIG_NUMA 5974 5975static int sched_domains_numa_levels; 5976static int *sched_domains_numa_distance; 5977static struct cpumask ***sched_domains_numa_masks; 5978static int sched_domains_curr_level; 5979 5980static inline int sd_local_flags(int level) 5981{ 5982 if (sched_domains_numa_distance[level] > RECLAIM_DISTANCE) 5983 return 0; 5984 5985 return SD_BALANCE_EXEC | SD_BALANCE_FORK | SD_WAKE_AFFINE; 5986} 5987 5988static struct sched_domain * 5989sd_numa_init(struct sched_domain_topology_level *tl, int cpu) 5990{ 5991 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); 5992 int level = tl->numa_level; 5993 int sd_weight = cpumask_weight( 5994 sched_domains_numa_masks[level][cpu_to_node(cpu)]); 5995 5996 *sd = (struct sched_domain){ 5997 .min_interval = sd_weight, 5998 .max_interval = 2*sd_weight, 5999 .busy_factor = 32, 6000 .imbalance_pct = 125, 6001 .cache_nice_tries = 2, 6002 .busy_idx = 3, 6003 .idle_idx = 2, 6004 .newidle_idx = 0, 6005 .wake_idx = 0, 6006 .forkexec_idx = 0, 6007 6008 .flags = 1*SD_LOAD_BALANCE 6009 | 1*SD_BALANCE_NEWIDLE 6010 | 0*SD_BALANCE_EXEC 6011 | 0*SD_BALANCE_FORK 6012 | 0*SD_BALANCE_WAKE 6013 | 0*SD_WAKE_AFFINE 6014 | 0*SD_SHARE_CPUPOWER 6015 | 0*SD_SHARE_PKG_RESOURCES 6016 | 1*SD_SERIALIZE 6017 | 0*SD_PREFER_SIBLING 6018 | 1*SD_NUMA 6019 | sd_local_flags(level) 6020 , 6021 .last_balance = jiffies, 6022 .balance_interval = sd_weight, 6023 }; 6024 SD_INIT_NAME(sd, NUMA); 6025 sd->private = &tl->data; 6026 6027 /* 6028 * Ugly hack to pass state to sd_numa_mask()... 6029 */ 6030 sched_domains_curr_level = tl->numa_level; 6031 6032 return sd; 6033} 6034 6035static const struct cpumask *sd_numa_mask(int cpu) 6036{ 6037 return sched_domains_numa_masks[sched_domains_curr_level][cpu_to_node(cpu)]; 6038} 6039 6040static void sched_numa_warn(const char *str) 6041{ 6042 static int done = false; 6043 int i,j; 6044 6045 if (done) 6046 return; 6047 6048 done = true; 6049 6050 printk(KERN_WARNING "ERROR: %s\n\n", str); 6051 6052 for (i = 0; i < nr_node_ids; i++) { 6053 printk(KERN_WARNING " "); 6054 for (j = 0; j < nr_node_ids; j++) 6055 printk(KERN_CONT "%02d ", node_distance(i,j)); 6056 printk(KERN_CONT "\n"); 6057 } 6058 printk(KERN_WARNING "\n"); 6059} 6060 6061static bool find_numa_distance(int distance) 6062{ 6063 int i; 6064 6065 if (distance == node_distance(0, 0)) 6066 return true; 6067 6068 for (i = 0; i < sched_domains_numa_levels; i++) { 6069 if (sched_domains_numa_distance[i] == distance) 6070 return true; 6071 } 6072 6073 return false; 6074} 6075 6076static void sched_init_numa(void) 6077{ 6078 int next_distance, curr_distance = node_distance(0, 0); 6079 struct sched_domain_topology_level *tl; 6080 int level = 0; 6081 int i, j, k; 6082 6083 sched_domains_numa_distance = kzalloc(sizeof(int) * nr_node_ids, GFP_KERNEL); 6084 if (!sched_domains_numa_distance) 6085 return; 6086 6087 /* 6088 * O(nr_nodes^2) deduplicating selection sort -- in order to find the 6089 * unique distances in the node_distance() table. 6090 * 6091 * Assumes node_distance(0,j) includes all distances in 6092 * node_distance(i,j) in order to avoid cubic time. 6093 */ 6094 next_distance = curr_distance; 6095 for (i = 0; i < nr_node_ids; i++) { 6096 for (j = 0; j < nr_node_ids; j++) { 6097 for (k = 0; k < nr_node_ids; k++) { 6098 int distance = node_distance(i, k); 6099 6100 if (distance > curr_distance && 6101 (distance < next_distance || 6102 next_distance == curr_distance)) 6103 next_distance = distance; 6104 6105 /* 6106 * While not a strong assumption it would be nice to know 6107 * about cases where if node A is connected to B, B is not 6108 * equally connected to A. 6109 */ 6110 if (sched_debug() && node_distance(k, i) != distance) 6111 sched_numa_warn("Node-distance not symmetric"); 6112 6113 if (sched_debug() && i && !find_numa_distance(distance)) 6114 sched_numa_warn("Node-0 not representative"); 6115 } 6116 if (next_distance != curr_distance) { 6117 sched_domains_numa_distance[level++] = next_distance; 6118 sched_domains_numa_levels = level; 6119 curr_distance = next_distance; 6120 } else break; 6121 } 6122 6123 /* 6124 * In case of sched_debug() we verify the above assumption. 6125 */ 6126 if (!sched_debug()) 6127 break; 6128 } 6129 /* 6130 * 'level' contains the number of unique distances, excluding the 6131 * identity distance node_distance(i,i). 6132 * 6133 * The sched_domains_numa_distance[] array includes the actual distance 6134 * numbers. 6135 */ 6136 6137 /* 6138 * Here, we should temporarily reset sched_domains_numa_levels to 0. 6139 * If it fails to allocate memory for array sched_domains_numa_masks[][], 6140 * the array will contain less then 'level' members. This could be 6141 * dangerous when we use it to iterate array sched_domains_numa_masks[][] 6142 * in other functions. 6143 * 6144 * We reset it to 'level' at the end of this function. 6145 */ 6146 sched_domains_numa_levels = 0; 6147 6148 sched_domains_numa_masks = kzalloc(sizeof(void *) * level, GFP_KERNEL); 6149 if (!sched_domains_numa_masks) 6150 return; 6151 6152 /* 6153 * Now for each level, construct a mask per node which contains all 6154 * cpus of nodes that are that many hops away from us. 6155 */ 6156 for (i = 0; i < level; i++) { 6157 sched_domains_numa_masks[i] = 6158 kzalloc(nr_node_ids * sizeof(void *), GFP_KERNEL); 6159 if (!sched_domains_numa_masks[i]) 6160 return; 6161 6162 for (j = 0; j < nr_node_ids; j++) { 6163 struct cpumask *mask = kzalloc(cpumask_size(), GFP_KERNEL); 6164 if (!mask) 6165 return; 6166 6167 sched_domains_numa_masks[i][j] = mask; 6168 6169 for (k = 0; k < nr_node_ids; k++) { 6170 if (node_distance(j, k) > sched_domains_numa_distance[i]) 6171 continue; 6172 6173 cpumask_or(mask, mask, cpumask_of_node(k)); 6174 } 6175 } 6176 } 6177 6178 tl = kzalloc((ARRAY_SIZE(default_topology) + level) * 6179 sizeof(struct sched_domain_topology_level), GFP_KERNEL); 6180 if (!tl) 6181 return; 6182 6183 /* 6184 * Copy the default topology bits.. 6185 */ 6186 for (i = 0; default_topology[i].init; i++) 6187 tl[i] = default_topology[i]; 6188 6189 /* 6190 * .. and append 'j' levels of NUMA goodness. 6191 */ 6192 for (j = 0; j < level; i++, j++) { 6193 tl[i] = (struct sched_domain_topology_level){ 6194 .init = sd_numa_init, 6195 .mask = sd_numa_mask, 6196 .flags = SDTL_OVERLAP, 6197 .numa_level = j, 6198 }; 6199 } 6200 6201 sched_domain_topology = tl; 6202 6203 sched_domains_numa_levels = level; 6204} 6205 6206static void sched_domains_numa_masks_set(int cpu) 6207{ 6208 int i, j; 6209 int node = cpu_to_node(cpu); 6210 6211 for (i = 0; i < sched_domains_numa_levels; i++) { 6212 for (j = 0; j < nr_node_ids; j++) { 6213 if (node_distance(j, node) <= sched_domains_numa_distance[i]) 6214 cpumask_set_cpu(cpu, sched_domains_numa_masks[i][j]); 6215 } 6216 } 6217} 6218 6219static void sched_domains_numa_masks_clear(int cpu) 6220{ 6221 int i, j; 6222 for (i = 0; i < sched_domains_numa_levels; i++) { 6223 for (j = 0; j < nr_node_ids; j++) 6224 cpumask_clear_cpu(cpu, sched_domains_numa_masks[i][j]); 6225 } 6226} 6227 6228/* 6229 * Update sched_domains_numa_masks[level][node] array when new cpus 6230 * are onlined. 6231 */ 6232static int sched_domains_numa_masks_update(struct notifier_block *nfb, 6233 unsigned long action, 6234 void *hcpu) 6235{ 6236 int cpu = (long)hcpu; 6237 6238 switch (action & ~CPU_TASKS_FROZEN) { 6239 case CPU_ONLINE: 6240 sched_domains_numa_masks_set(cpu); 6241 break; 6242 6243 case CPU_DEAD: 6244 sched_domains_numa_masks_clear(cpu); 6245 break; 6246 6247 default: 6248 return NOTIFY_DONE; 6249 } 6250 6251 return NOTIFY_OK; 6252} 6253#else 6254static inline void sched_init_numa(void) 6255{ 6256} 6257 6258static int sched_domains_numa_masks_update(struct notifier_block *nfb, 6259 unsigned long action, 6260 void *hcpu) 6261{ 6262 return 0; 6263} 6264#endif /* CONFIG_NUMA */ 6265 6266static int __sdt_alloc(const struct cpumask *cpu_map) 6267{ 6268 struct sched_domain_topology_level *tl; 6269 int j; 6270 6271 for_each_sd_topology(tl) { 6272 struct sd_data *sdd = &tl->data; 6273 6274 sdd->sd = alloc_percpu(struct sched_domain *); 6275 if (!sdd->sd) 6276 return -ENOMEM; 6277 6278 sdd->sg = alloc_percpu(struct sched_group *); 6279 if (!sdd->sg) 6280 return -ENOMEM; 6281 6282 sdd->sgp = alloc_percpu(struct sched_group_power *); 6283 if (!sdd->sgp) 6284 return -ENOMEM; 6285 6286 for_each_cpu(j, cpu_map) { 6287 struct sched_domain *sd; 6288 struct sched_group *sg; 6289 struct sched_group_power *sgp; 6290 6291 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(), 6292 GFP_KERNEL, cpu_to_node(j)); 6293 if (!sd) 6294 return -ENOMEM; 6295 6296 *per_cpu_ptr(sdd->sd, j) = sd; 6297 6298 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 6299 GFP_KERNEL, cpu_to_node(j)); 6300 if (!sg) 6301 return -ENOMEM; 6302 6303 sg->next = sg; 6304 6305 *per_cpu_ptr(sdd->sg, j) = sg; 6306 6307 sgp = kzalloc_node(sizeof(struct sched_group_power) + cpumask_size(), 6308 GFP_KERNEL, cpu_to_node(j)); 6309 if (!sgp) 6310 return -ENOMEM; 6311 6312 *per_cpu_ptr(sdd->sgp, j) = sgp; 6313 } 6314 } 6315 6316 return 0; 6317} 6318 6319static void __sdt_free(const struct cpumask *cpu_map) 6320{ 6321 struct sched_domain_topology_level *tl; 6322 int j; 6323 6324 for_each_sd_topology(tl) { 6325 struct sd_data *sdd = &tl->data; 6326 6327 for_each_cpu(j, cpu_map) { 6328 struct sched_domain *sd; 6329 6330 if (sdd->sd) { 6331 sd = *per_cpu_ptr(sdd->sd, j); 6332 if (sd && (sd->flags & SD_OVERLAP)) 6333 free_sched_groups(sd->groups, 0); 6334 kfree(*per_cpu_ptr(sdd->sd, j)); 6335 } 6336 6337 if (sdd->sg) 6338 kfree(*per_cpu_ptr(sdd->sg, j)); 6339 if (sdd->sgp) 6340 kfree(*per_cpu_ptr(sdd->sgp, j)); 6341 } 6342 free_percpu(sdd->sd); 6343 sdd->sd = NULL; 6344 free_percpu(sdd->sg); 6345 sdd->sg = NULL; 6346 free_percpu(sdd->sgp); 6347 sdd->sgp = NULL; 6348 } 6349} 6350 6351struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl, 6352 const struct cpumask *cpu_map, struct sched_domain_attr *attr, 6353 struct sched_domain *child, int cpu) 6354{ 6355 struct sched_domain *sd = tl->init(tl, cpu); 6356 if (!sd) 6357 return child; 6358 6359 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu)); 6360 if (child) { 6361 sd->level = child->level + 1; 6362 sched_domain_level_max = max(sched_domain_level_max, sd->level); 6363 child->parent = sd; 6364 sd->child = child; 6365 } 6366 set_domain_attribute(sd, attr); 6367 6368 return sd; 6369} 6370 6371/* 6372 * Build sched domains for a given set of cpus and attach the sched domains 6373 * to the individual cpus 6374 */ 6375static int build_sched_domains(const struct cpumask *cpu_map, 6376 struct sched_domain_attr *attr) 6377{ 6378 enum s_alloc alloc_state; 6379 struct sched_domain *sd; 6380 struct s_data d; 6381 int i, ret = -ENOMEM; 6382 6383 alloc_state = __visit_domain_allocation_hell(&d, cpu_map); 6384 if (alloc_state != sa_rootdomain) 6385 goto error; 6386 6387 /* Set up domains for cpus specified by the cpu_map. */ 6388 for_each_cpu(i, cpu_map) { 6389 struct sched_domain_topology_level *tl; 6390 6391 sd = NULL; 6392 for_each_sd_topology(tl) { 6393 sd = build_sched_domain(tl, cpu_map, attr, sd, i); 6394 if (tl == sched_domain_topology) 6395 *per_cpu_ptr(d.sd, i) = sd; 6396 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP)) 6397 sd->flags |= SD_OVERLAP; 6398 if (cpumask_equal(cpu_map, sched_domain_span(sd))) 6399 break; 6400 } 6401 } 6402 6403 /* Build the groups for the domains */ 6404 for_each_cpu(i, cpu_map) { 6405 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 6406 sd->span_weight = cpumask_weight(sched_domain_span(sd)); 6407 if (sd->flags & SD_OVERLAP) { 6408 if (build_overlap_sched_groups(sd, i)) 6409 goto error; 6410 } else { 6411 if (build_sched_groups(sd, i)) 6412 goto error; 6413 } 6414 } 6415 } 6416 6417 /* Calculate CPU power for physical packages and nodes */ 6418 for (i = nr_cpumask_bits-1; i >= 0; i--) { 6419 if (!cpumask_test_cpu(i, cpu_map)) 6420 continue; 6421 6422 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 6423 claim_allocations(i, sd); 6424 init_sched_groups_power(i, sd); 6425 } 6426 } 6427 6428 /* Attach the domains */ 6429 rcu_read_lock(); 6430 for_each_cpu(i, cpu_map) { 6431 sd = *per_cpu_ptr(d.sd, i); 6432 cpu_attach_domain(sd, d.rd, i); 6433 } 6434 rcu_read_unlock(); 6435 6436 ret = 0; 6437error: 6438 __free_domain_allocs(&d, alloc_state, cpu_map); 6439 return ret; 6440} 6441 6442static cpumask_var_t *doms_cur; /* current sched domains */ 6443static int ndoms_cur; /* number of sched domains in 'doms_cur' */ 6444static struct sched_domain_attr *dattr_cur; 6445 /* attribues of custom domains in 'doms_cur' */ 6446 6447/* 6448 * Special case: If a kmalloc of a doms_cur partition (array of 6449 * cpumask) fails, then fallback to a single sched domain, 6450 * as determined by the single cpumask fallback_doms. 6451 */ 6452static cpumask_var_t fallback_doms; 6453 6454/* 6455 * arch_update_cpu_topology lets virtualized architectures update the 6456 * cpu core maps. It is supposed to return 1 if the topology changed 6457 * or 0 if it stayed the same. 6458 */ 6459int __weak arch_update_cpu_topology(void) 6460{ 6461 return 0; 6462} 6463 6464cpumask_var_t *alloc_sched_domains(unsigned int ndoms) 6465{ 6466 int i; 6467 cpumask_var_t *doms; 6468 6469 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL); 6470 if (!doms) 6471 return NULL; 6472 for (i = 0; i < ndoms; i++) { 6473 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) { 6474 free_sched_domains(doms, i); 6475 return NULL; 6476 } 6477 } 6478 return doms; 6479} 6480 6481void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms) 6482{ 6483 unsigned int i; 6484 for (i = 0; i < ndoms; i++) 6485 free_cpumask_var(doms[i]); 6486 kfree(doms); 6487} 6488 6489/* 6490 * Set up scheduler domains and groups. Callers must hold the hotplug lock. 6491 * For now this just excludes isolated cpus, but could be used to 6492 * exclude other special cases in the future. 6493 */ 6494static int init_sched_domains(const struct cpumask *cpu_map) 6495{ 6496 int err; 6497 6498 arch_update_cpu_topology(); 6499 ndoms_cur = 1; 6500 doms_cur = alloc_sched_domains(ndoms_cur); 6501 if (!doms_cur) 6502 doms_cur = &fallback_doms; 6503 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map); 6504 err = build_sched_domains(doms_cur[0], NULL); 6505 register_sched_domain_sysctl(); 6506 6507 return err; 6508} 6509 6510/* 6511 * Detach sched domains from a group of cpus specified in cpu_map 6512 * These cpus will now be attached to the NULL domain 6513 */ 6514static void detach_destroy_domains(const struct cpumask *cpu_map) 6515{ 6516 int i; 6517 6518 rcu_read_lock(); 6519 for_each_cpu(i, cpu_map) 6520 cpu_attach_domain(NULL, &def_root_domain, i); 6521 rcu_read_unlock(); 6522} 6523 6524/* handle null as "default" */ 6525static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur, 6526 struct sched_domain_attr *new, int idx_new) 6527{ 6528 struct sched_domain_attr tmp; 6529 6530 /* fast path */ 6531 if (!new && !cur) 6532 return 1; 6533 6534 tmp = SD_ATTR_INIT; 6535 return !memcmp(cur ? (cur + idx_cur) : &tmp, 6536 new ? (new + idx_new) : &tmp, 6537 sizeof(struct sched_domain_attr)); 6538} 6539 6540/* 6541 * Partition sched domains as specified by the 'ndoms_new' 6542 * cpumasks in the array doms_new[] of cpumasks. This compares 6543 * doms_new[] to the current sched domain partitioning, doms_cur[]. 6544 * It destroys each deleted domain and builds each new domain. 6545 * 6546 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'. 6547 * The masks don't intersect (don't overlap.) We should setup one 6548 * sched domain for each mask. CPUs not in any of the cpumasks will 6549 * not be load balanced. If the same cpumask appears both in the 6550 * current 'doms_cur' domains and in the new 'doms_new', we can leave 6551 * it as it is. 6552 * 6553 * The passed in 'doms_new' should be allocated using 6554 * alloc_sched_domains. This routine takes ownership of it and will 6555 * free_sched_domains it when done with it. If the caller failed the 6556 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1, 6557 * and partition_sched_domains() will fallback to the single partition 6558 * 'fallback_doms', it also forces the domains to be rebuilt. 6559 * 6560 * If doms_new == NULL it will be replaced with cpu_online_mask. 6561 * ndoms_new == 0 is a special case for destroying existing domains, 6562 * and it will not create the default domain. 6563 * 6564 * Call with hotplug lock held 6565 */ 6566void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[], 6567 struct sched_domain_attr *dattr_new) 6568{ 6569 int i, j, n; 6570 int new_topology; 6571 6572 mutex_lock(&sched_domains_mutex); 6573 6574 /* always unregister in case we don't destroy any domains */ 6575 unregister_sched_domain_sysctl(); 6576 6577 /* Let architecture update cpu core mappings. */ 6578 new_topology = arch_update_cpu_topology(); 6579 6580 n = doms_new ? ndoms_new : 0; 6581 6582 /* Destroy deleted domains */ 6583 for (i = 0; i < ndoms_cur; i++) { 6584 for (j = 0; j < n && !new_topology; j++) { 6585 if (cpumask_equal(doms_cur[i], doms_new[j]) 6586 && dattrs_equal(dattr_cur, i, dattr_new, j)) 6587 goto match1; 6588 } 6589 /* no match - a current sched domain not in new doms_new[] */ 6590 detach_destroy_domains(doms_cur[i]); 6591match1: 6592 ; 6593 } 6594 6595 n = ndoms_cur; 6596 if (doms_new == NULL) { 6597 n = 0; 6598 doms_new = &fallback_doms; 6599 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map); 6600 WARN_ON_ONCE(dattr_new); 6601 } 6602 6603 /* Build new domains */ 6604 for (i = 0; i < ndoms_new; i++) { 6605 for (j = 0; j < n && !new_topology; j++) { 6606 if (cpumask_equal(doms_new[i], doms_cur[j]) 6607 && dattrs_equal(dattr_new, i, dattr_cur, j)) 6608 goto match2; 6609 } 6610 /* no match - add a new doms_new */ 6611 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL); 6612match2: 6613 ; 6614 } 6615 6616 /* Remember the new sched domains */ 6617 if (doms_cur != &fallback_doms) 6618 free_sched_domains(doms_cur, ndoms_cur); 6619 kfree(dattr_cur); /* kfree(NULL) is safe */ 6620 doms_cur = doms_new; 6621 dattr_cur = dattr_new; 6622 ndoms_cur = ndoms_new; 6623 6624 register_sched_domain_sysctl(); 6625 6626 mutex_unlock(&sched_domains_mutex); 6627} 6628 6629static int num_cpus_frozen; /* used to mark begin/end of suspend/resume */ 6630 6631/* 6632 * Update cpusets according to cpu_active mask. If cpusets are 6633 * disabled, cpuset_update_active_cpus() becomes a simple wrapper 6634 * around partition_sched_domains(). 6635 * 6636 * If we come here as part of a suspend/resume, don't touch cpusets because we 6637 * want to restore it back to its original state upon resume anyway. 6638 */ 6639static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action, 6640 void *hcpu) 6641{ 6642 switch (action) { 6643 case CPU_ONLINE_FROZEN: 6644 case CPU_DOWN_FAILED_FROZEN: 6645 6646 /* 6647 * num_cpus_frozen tracks how many CPUs are involved in suspend 6648 * resume sequence. As long as this is not the last online 6649 * operation in the resume sequence, just build a single sched 6650 * domain, ignoring cpusets. 6651 */ 6652 num_cpus_frozen--; 6653 if (likely(num_cpus_frozen)) { 6654 partition_sched_domains(1, NULL, NULL); 6655 break; 6656 } 6657 6658 /* 6659 * This is the last CPU online operation. So fall through and 6660 * restore the original sched domains by considering the 6661 * cpuset configurations. 6662 */ 6663 6664 case CPU_ONLINE: 6665 case CPU_DOWN_FAILED: 6666 cpuset_update_active_cpus(true); 6667 break; 6668 default: 6669 return NOTIFY_DONE; 6670 } 6671 return NOTIFY_OK; 6672} 6673 6674static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action, 6675 void *hcpu) 6676{ 6677 switch (action) { 6678 case CPU_DOWN_PREPARE: 6679 cpuset_update_active_cpus(false); 6680 break; 6681 case CPU_DOWN_PREPARE_FROZEN: 6682 num_cpus_frozen++; 6683 partition_sched_domains(1, NULL, NULL); 6684 break; 6685 default: 6686 return NOTIFY_DONE; 6687 } 6688 return NOTIFY_OK; 6689} 6690 6691void __init sched_init_smp(void) 6692{ 6693 cpumask_var_t non_isolated_cpus; 6694 6695 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL); 6696 alloc_cpumask_var(&fallback_doms, GFP_KERNEL); 6697 6698 sched_init_numa(); 6699 6700 /* 6701 * There's no userspace yet to cause hotplug operations; hence all the 6702 * cpu masks are stable and all blatant races in the below code cannot 6703 * happen. 6704 */ 6705 mutex_lock(&sched_domains_mutex); 6706 init_sched_domains(cpu_active_mask); 6707 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map); 6708 if (cpumask_empty(non_isolated_cpus)) 6709 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus); 6710 mutex_unlock(&sched_domains_mutex); 6711 6712 hotcpu_notifier(sched_domains_numa_masks_update, CPU_PRI_SCHED_ACTIVE); 6713 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE); 6714 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE); 6715 6716 init_hrtick(); 6717 6718 /* Move init over to a non-isolated CPU */ 6719 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0) 6720 BUG(); 6721 sched_init_granularity(); 6722 free_cpumask_var(non_isolated_cpus); 6723 6724 init_sched_rt_class(); 6725 init_sched_dl_class(); 6726} 6727#else 6728void __init sched_init_smp(void) 6729{ 6730 sched_init_granularity(); 6731} 6732#endif /* CONFIG_SMP */ 6733 6734const_debug unsigned int sysctl_timer_migration = 1; 6735 6736int in_sched_functions(unsigned long addr) 6737{ 6738 return in_lock_functions(addr) || 6739 (addr >= (unsigned long)__sched_text_start 6740 && addr < (unsigned long)__sched_text_end); 6741} 6742 6743#ifdef CONFIG_CGROUP_SCHED 6744/* 6745 * Default task group. 6746 * Every task in system belongs to this group at bootup. 6747 */ 6748struct task_group root_task_group; 6749LIST_HEAD(task_groups); 6750#endif 6751 6752DECLARE_PER_CPU(cpumask_var_t, load_balance_mask); 6753 6754void __init sched_init(void) 6755{ 6756 int i, j; 6757 unsigned long alloc_size = 0, ptr; 6758 6759#ifdef CONFIG_FAIR_GROUP_SCHED 6760 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 6761#endif 6762#ifdef CONFIG_RT_GROUP_SCHED 6763 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 6764#endif 6765#ifdef CONFIG_CPUMASK_OFFSTACK 6766 alloc_size += num_possible_cpus() * cpumask_size(); 6767#endif 6768 if (alloc_size) { 6769 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT); 6770 6771#ifdef CONFIG_FAIR_GROUP_SCHED 6772 root_task_group.se = (struct sched_entity **)ptr; 6773 ptr += nr_cpu_ids * sizeof(void **); 6774 6775 root_task_group.cfs_rq = (struct cfs_rq **)ptr; 6776 ptr += nr_cpu_ids * sizeof(void **); 6777 6778#endif /* CONFIG_FAIR_GROUP_SCHED */ 6779#ifdef CONFIG_RT_GROUP_SCHED 6780 root_task_group.rt_se = (struct sched_rt_entity **)ptr; 6781 ptr += nr_cpu_ids * sizeof(void **); 6782 6783 root_task_group.rt_rq = (struct rt_rq **)ptr; 6784 ptr += nr_cpu_ids * sizeof(void **); 6785 6786#endif /* CONFIG_RT_GROUP_SCHED */ 6787#ifdef CONFIG_CPUMASK_OFFSTACK 6788 for_each_possible_cpu(i) { 6789 per_cpu(load_balance_mask, i) = (void *)ptr; 6790 ptr += cpumask_size(); 6791 } 6792#endif /* CONFIG_CPUMASK_OFFSTACK */ 6793 } 6794 6795 init_rt_bandwidth(&def_rt_bandwidth, 6796 global_rt_period(), global_rt_runtime()); 6797 init_dl_bandwidth(&def_dl_bandwidth, 6798 global_rt_period(), global_rt_runtime()); 6799 6800#ifdef CONFIG_SMP 6801 init_defrootdomain(); 6802#endif 6803 6804#ifdef CONFIG_RT_GROUP_SCHED 6805 init_rt_bandwidth(&root_task_group.rt_bandwidth, 6806 global_rt_period(), global_rt_runtime()); 6807#endif /* CONFIG_RT_GROUP_SCHED */ 6808 6809#ifdef CONFIG_CGROUP_SCHED 6810 list_add(&root_task_group.list, &task_groups); 6811 INIT_LIST_HEAD(&root_task_group.children); 6812 INIT_LIST_HEAD(&root_task_group.siblings); 6813 autogroup_init(&init_task); 6814 6815#endif /* CONFIG_CGROUP_SCHED */ 6816 6817 for_each_possible_cpu(i) { 6818 struct rq *rq; 6819 6820 rq = cpu_rq(i); 6821 raw_spin_lock_init(&rq->lock); 6822 rq->nr_running = 0; 6823 rq->calc_load_active = 0; 6824 rq->calc_load_update = jiffies + LOAD_FREQ; 6825 init_cfs_rq(&rq->cfs); 6826 init_rt_rq(&rq->rt, rq); 6827 init_dl_rq(&rq->dl, rq); 6828#ifdef CONFIG_FAIR_GROUP_SCHED 6829 root_task_group.shares = ROOT_TASK_GROUP_LOAD; 6830 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list); 6831 /* 6832 * How much cpu bandwidth does root_task_group get? 6833 * 6834 * In case of task-groups formed thr' the cgroup filesystem, it 6835 * gets 100% of the cpu resources in the system. This overall 6836 * system cpu resource is divided among the tasks of 6837 * root_task_group and its child task-groups in a fair manner, 6838 * based on each entity's (task or task-group's) weight 6839 * (se->load.weight). 6840 * 6841 * In other words, if root_task_group has 10 tasks of weight 6842 * 1024) and two child groups A0 and A1 (of weight 1024 each), 6843 * then A0's share of the cpu resource is: 6844 * 6845 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33% 6846 * 6847 * We achieve this by letting root_task_group's tasks sit 6848 * directly in rq->cfs (i.e root_task_group->se[] = NULL). 6849 */ 6850 init_cfs_bandwidth(&root_task_group.cfs_bandwidth); 6851 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL); 6852#endif /* CONFIG_FAIR_GROUP_SCHED */ 6853 6854 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime; 6855#ifdef CONFIG_RT_GROUP_SCHED 6856 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL); 6857#endif 6858 6859 for (j = 0; j < CPU_LOAD_IDX_MAX; j++) 6860 rq->cpu_load[j] = 0; 6861 6862 rq->last_load_update_tick = jiffies; 6863 6864#ifdef CONFIG_SMP 6865 rq->sd = NULL; 6866 rq->rd = NULL; 6867 rq->cpu_power = SCHED_POWER_SCALE; 6868 rq->post_schedule = 0; 6869 rq->active_balance = 0; 6870 rq->next_balance = jiffies; 6871 rq->push_cpu = 0; 6872 rq->cpu = i; 6873 rq->online = 0; 6874 rq->idle_stamp = 0; 6875 rq->avg_idle = 2*sysctl_sched_migration_cost; 6876 rq->max_idle_balance_cost = sysctl_sched_migration_cost; 6877 6878 INIT_LIST_HEAD(&rq->cfs_tasks); 6879 6880 rq_attach_root(rq, &def_root_domain); 6881#ifdef CONFIG_NO_HZ_COMMON 6882 rq->nohz_flags = 0; 6883#endif 6884#ifdef CONFIG_NO_HZ_FULL 6885 rq->last_sched_tick = 0; 6886#endif 6887#endif 6888 init_rq_hrtick(rq); 6889 atomic_set(&rq->nr_iowait, 0); 6890 } 6891 6892 set_load_weight(&init_task); 6893 6894#ifdef CONFIG_PREEMPT_NOTIFIERS 6895 INIT_HLIST_HEAD(&init_task.preempt_notifiers); 6896#endif 6897 6898 /* 6899 * The boot idle thread does lazy MMU switching as well: 6900 */ 6901 atomic_inc(&init_mm.mm_count); 6902 enter_lazy_tlb(&init_mm, current); 6903 6904 /* 6905 * Make us the idle thread. Technically, schedule() should not be 6906 * called from this thread, however somewhere below it might be, 6907 * but because we are the idle thread, we just pick up running again 6908 * when this runqueue becomes "idle". 6909 */ 6910 init_idle(current, smp_processor_id()); 6911 6912 calc_load_update = jiffies + LOAD_FREQ; 6913 6914 /* 6915 * During early bootup we pretend to be a normal task: 6916 */ 6917 current->sched_class = &fair_sched_class; 6918 6919#ifdef CONFIG_SMP 6920 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT); 6921 /* May be allocated at isolcpus cmdline parse time */ 6922 if (cpu_isolated_map == NULL) 6923 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT); 6924 idle_thread_set_boot_cpu(); 6925#endif 6926 init_sched_fair_class(); 6927 6928 scheduler_running = 1; 6929} 6930 6931#ifdef CONFIG_DEBUG_ATOMIC_SLEEP 6932static inline int preempt_count_equals(int preempt_offset) 6933{ 6934 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth(); 6935 6936 return (nested == preempt_offset); 6937} 6938 6939void __might_sleep(const char *file, int line, int preempt_offset) 6940{ 6941 static unsigned long prev_jiffy; /* ratelimiting */ 6942 6943 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */ 6944 if ((preempt_count_equals(preempt_offset) && !irqs_disabled() && 6945 !is_idle_task(current)) || 6946 system_state != SYSTEM_RUNNING || oops_in_progress) 6947 return; 6948 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy) 6949 return; 6950 prev_jiffy = jiffies; 6951 6952 printk(KERN_ERR 6953 "BUG: sleeping function called from invalid context at %s:%d\n", 6954 file, line); 6955 printk(KERN_ERR 6956 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n", 6957 in_atomic(), irqs_disabled(), 6958 current->pid, current->comm); 6959 6960 debug_show_held_locks(current); 6961 if (irqs_disabled()) 6962 print_irqtrace_events(current); 6963#ifdef CONFIG_DEBUG_PREEMPT 6964 if (!preempt_count_equals(preempt_offset)) { 6965 pr_err("Preemption disabled at:"); 6966 print_ip_sym(current->preempt_disable_ip); 6967 pr_cont("\n"); 6968 } 6969#endif 6970 dump_stack(); 6971} 6972EXPORT_SYMBOL(__might_sleep); 6973#endif 6974 6975#ifdef CONFIG_MAGIC_SYSRQ 6976static void normalize_task(struct rq *rq, struct task_struct *p) 6977{ 6978 const struct sched_class *prev_class = p->sched_class; 6979 struct sched_attr attr = { 6980 .sched_policy = SCHED_NORMAL, 6981 }; 6982 int old_prio = p->prio; 6983 int on_rq; 6984 6985 on_rq = p->on_rq; 6986 if (on_rq) 6987 dequeue_task(rq, p, 0); 6988 __setscheduler(rq, p, &attr); 6989 if (on_rq) { 6990 enqueue_task(rq, p, 0); 6991 resched_task(rq->curr); 6992 } 6993 6994 check_class_changed(rq, p, prev_class, old_prio); 6995} 6996 6997void normalize_rt_tasks(void) 6998{ 6999 struct task_struct *g, *p; 7000 unsigned long flags; 7001 struct rq *rq; 7002 7003 read_lock_irqsave(&tasklist_lock, flags); 7004 do_each_thread(g, p) { 7005 /* 7006 * Only normalize user tasks: 7007 */ 7008 if (!p->mm) 7009 continue; 7010 7011 p->se.exec_start = 0; 7012#ifdef CONFIG_SCHEDSTATS 7013 p->se.statistics.wait_start = 0; 7014 p->se.statistics.sleep_start = 0; 7015 p->se.statistics.block_start = 0; 7016#endif 7017 7018 if (!dl_task(p) && !rt_task(p)) { 7019 /* 7020 * Renice negative nice level userspace 7021 * tasks back to 0: 7022 */ 7023 if (task_nice(p) < 0 && p->mm) 7024 set_user_nice(p, 0); 7025 continue; 7026 } 7027 7028 raw_spin_lock(&p->pi_lock); 7029 rq = __task_rq_lock(p); 7030 7031 normalize_task(rq, p); 7032 7033 __task_rq_unlock(rq); 7034 raw_spin_unlock(&p->pi_lock); 7035 } while_each_thread(g, p); 7036 7037 read_unlock_irqrestore(&tasklist_lock, flags); 7038} 7039 7040#endif /* CONFIG_MAGIC_SYSRQ */ 7041 7042#if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) 7043/* 7044 * These functions are only useful for the IA64 MCA handling, or kdb. 7045 * 7046 * They can only be called when the whole system has been 7047 * stopped - every CPU needs to be quiescent, and no scheduling 7048 * activity can take place. Using them for anything else would 7049 * be a serious bug, and as a result, they aren't even visible 7050 * under any other configuration. 7051 */ 7052 7053/** 7054 * curr_task - return the current task for a given cpu. 7055 * @cpu: the processor in question. 7056 * 7057 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 7058 * 7059 * Return: The current task for @cpu. 7060 */ 7061struct task_struct *curr_task(int cpu) 7062{ 7063 return cpu_curr(cpu); 7064} 7065 7066#endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */ 7067 7068#ifdef CONFIG_IA64 7069/** 7070 * set_curr_task - set the current task for a given cpu. 7071 * @cpu: the processor in question. 7072 * @p: the task pointer to set. 7073 * 7074 * Description: This function must only be used when non-maskable interrupts 7075 * are serviced on a separate stack. It allows the architecture to switch the 7076 * notion of the current task on a cpu in a non-blocking manner. This function 7077 * must be called with all CPU's synchronized, and interrupts disabled, the 7078 * and caller must save the original value of the current task (see 7079 * curr_task() above) and restore that value before reenabling interrupts and 7080 * re-starting the system. 7081 * 7082 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 7083 */ 7084void set_curr_task(int cpu, struct task_struct *p) 7085{ 7086 cpu_curr(cpu) = p; 7087} 7088 7089#endif 7090 7091#ifdef CONFIG_CGROUP_SCHED 7092/* task_group_lock serializes the addition/removal of task groups */ 7093static DEFINE_SPINLOCK(task_group_lock); 7094 7095static void free_sched_group(struct task_group *tg) 7096{ 7097 free_fair_sched_group(tg); 7098 free_rt_sched_group(tg); 7099 autogroup_free(tg); 7100 kfree(tg); 7101} 7102 7103/* allocate runqueue etc for a new task group */ 7104struct task_group *sched_create_group(struct task_group *parent) 7105{ 7106 struct task_group *tg; 7107 7108 tg = kzalloc(sizeof(*tg), GFP_KERNEL); 7109 if (!tg) 7110 return ERR_PTR(-ENOMEM); 7111 7112 if (!alloc_fair_sched_group(tg, parent)) 7113 goto err; 7114 7115 if (!alloc_rt_sched_group(tg, parent)) 7116 goto err; 7117 7118 return tg; 7119 7120err: 7121 free_sched_group(tg); 7122 return ERR_PTR(-ENOMEM); 7123} 7124 7125void sched_online_group(struct task_group *tg, struct task_group *parent) 7126{ 7127 unsigned long flags; 7128 7129 spin_lock_irqsave(&task_group_lock, flags); 7130 list_add_rcu(&tg->list, &task_groups); 7131 7132 WARN_ON(!parent); /* root should already exist */ 7133 7134 tg->parent = parent; 7135 INIT_LIST_HEAD(&tg->children); 7136 list_add_rcu(&tg->siblings, &parent->children); 7137 spin_unlock_irqrestore(&task_group_lock, flags); 7138} 7139 7140/* rcu callback to free various structures associated with a task group */ 7141static void free_sched_group_rcu(struct rcu_head *rhp) 7142{ 7143 /* now it should be safe to free those cfs_rqs */ 7144 free_sched_group(container_of(rhp, struct task_group, rcu)); 7145} 7146 7147/* Destroy runqueue etc associated with a task group */ 7148void sched_destroy_group(struct task_group *tg) 7149{ 7150 /* wait for possible concurrent references to cfs_rqs complete */ 7151 call_rcu(&tg->rcu, free_sched_group_rcu); 7152} 7153 7154void sched_offline_group(struct task_group *tg) 7155{ 7156 unsigned long flags; 7157 int i; 7158 7159 /* end participation in shares distribution */ 7160 for_each_possible_cpu(i) 7161 unregister_fair_sched_group(tg, i); 7162 7163 spin_lock_irqsave(&task_group_lock, flags); 7164 list_del_rcu(&tg->list); 7165 list_del_rcu(&tg->siblings); 7166 spin_unlock_irqrestore(&task_group_lock, flags); 7167} 7168 7169/* change task's runqueue when it moves between groups. 7170 * The caller of this function should have put the task in its new group 7171 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to 7172 * reflect its new group. 7173 */ 7174void sched_move_task(struct task_struct *tsk) 7175{ 7176 struct task_group *tg; 7177 int on_rq, running; 7178 unsigned long flags; 7179 struct rq *rq; 7180 7181 rq = task_rq_lock(tsk, &flags); 7182 7183 running = task_current(rq, tsk); 7184 on_rq = tsk->on_rq; 7185 7186 if (on_rq) 7187 dequeue_task(rq, tsk, 0); 7188 if (unlikely(running)) 7189 tsk->sched_class->put_prev_task(rq, tsk); 7190 7191 tg = container_of(task_css_check(tsk, cpu_cgrp_id, 7192 lockdep_is_held(&tsk->sighand->siglock)), 7193 struct task_group, css); 7194 tg = autogroup_task_group(tsk, tg); 7195 tsk->sched_task_group = tg; 7196 7197#ifdef CONFIG_FAIR_GROUP_SCHED 7198 if (tsk->sched_class->task_move_group) 7199 tsk->sched_class->task_move_group(tsk, on_rq); 7200 else 7201#endif 7202 set_task_rq(tsk, task_cpu(tsk)); 7203 7204 if (unlikely(running)) 7205 tsk->sched_class->set_curr_task(rq); 7206 if (on_rq) 7207 enqueue_task(rq, tsk, 0); 7208 7209 task_rq_unlock(rq, tsk, &flags); 7210} 7211#endif /* CONFIG_CGROUP_SCHED */ 7212 7213#ifdef CONFIG_RT_GROUP_SCHED 7214/* 7215 * Ensure that the real time constraints are schedulable. 7216 */ 7217static DEFINE_MUTEX(rt_constraints_mutex); 7218 7219/* Must be called with tasklist_lock held */ 7220static inline int tg_has_rt_tasks(struct task_group *tg) 7221{ 7222 struct task_struct *g, *p; 7223 7224 do_each_thread(g, p) { 7225 if (rt_task(p) && task_rq(p)->rt.tg == tg) 7226 return 1; 7227 } while_each_thread(g, p); 7228 7229 return 0; 7230} 7231 7232struct rt_schedulable_data { 7233 struct task_group *tg; 7234 u64 rt_period; 7235 u64 rt_runtime; 7236}; 7237 7238static int tg_rt_schedulable(struct task_group *tg, void *data) 7239{ 7240 struct rt_schedulable_data *d = data; 7241 struct task_group *child; 7242 unsigned long total, sum = 0; 7243 u64 period, runtime; 7244 7245 period = ktime_to_ns(tg->rt_bandwidth.rt_period); 7246 runtime = tg->rt_bandwidth.rt_runtime; 7247 7248 if (tg == d->tg) { 7249 period = d->rt_period; 7250 runtime = d->rt_runtime; 7251 } 7252 7253 /* 7254 * Cannot have more runtime than the period. 7255 */ 7256 if (runtime > period && runtime != RUNTIME_INF) 7257 return -EINVAL; 7258 7259 /* 7260 * Ensure we don't starve existing RT tasks. 7261 */ 7262 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg)) 7263 return -EBUSY; 7264 7265 total = to_ratio(period, runtime); 7266 7267 /* 7268 * Nobody can have more than the global setting allows. 7269 */ 7270 if (total > to_ratio(global_rt_period(), global_rt_runtime())) 7271 return -EINVAL; 7272 7273 /* 7274 * The sum of our children's runtime should not exceed our own. 7275 */ 7276 list_for_each_entry_rcu(child, &tg->children, siblings) { 7277 period = ktime_to_ns(child->rt_bandwidth.rt_period); 7278 runtime = child->rt_bandwidth.rt_runtime; 7279 7280 if (child == d->tg) { 7281 period = d->rt_period; 7282 runtime = d->rt_runtime; 7283 } 7284 7285 sum += to_ratio(period, runtime); 7286 } 7287 7288 if (sum > total) 7289 return -EINVAL; 7290 7291 return 0; 7292} 7293 7294static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime) 7295{ 7296 int ret; 7297 7298 struct rt_schedulable_data data = { 7299 .tg = tg, 7300 .rt_period = period, 7301 .rt_runtime = runtime, 7302 }; 7303 7304 rcu_read_lock(); 7305 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data); 7306 rcu_read_unlock(); 7307 7308 return ret; 7309} 7310 7311static int tg_set_rt_bandwidth(struct task_group *tg, 7312 u64 rt_period, u64 rt_runtime) 7313{ 7314 int i, err = 0; 7315 7316 mutex_lock(&rt_constraints_mutex); 7317 read_lock(&tasklist_lock); 7318 err = __rt_schedulable(tg, rt_period, rt_runtime); 7319 if (err) 7320 goto unlock; 7321 7322 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock); 7323 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period); 7324 tg->rt_bandwidth.rt_runtime = rt_runtime; 7325 7326 for_each_possible_cpu(i) { 7327 struct rt_rq *rt_rq = tg->rt_rq[i]; 7328 7329 raw_spin_lock(&rt_rq->rt_runtime_lock); 7330 rt_rq->rt_runtime = rt_runtime; 7331 raw_spin_unlock(&rt_rq->rt_runtime_lock); 7332 } 7333 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock); 7334unlock: 7335 read_unlock(&tasklist_lock); 7336 mutex_unlock(&rt_constraints_mutex); 7337 7338 return err; 7339} 7340 7341static int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us) 7342{ 7343 u64 rt_runtime, rt_period; 7344 7345 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period); 7346 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC; 7347 if (rt_runtime_us < 0) 7348 rt_runtime = RUNTIME_INF; 7349 7350 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 7351} 7352 7353static long sched_group_rt_runtime(struct task_group *tg) 7354{ 7355 u64 rt_runtime_us; 7356 7357 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF) 7358 return -1; 7359 7360 rt_runtime_us = tg->rt_bandwidth.rt_runtime; 7361 do_div(rt_runtime_us, NSEC_PER_USEC); 7362 return rt_runtime_us; 7363} 7364 7365static int sched_group_set_rt_period(struct task_group *tg, long rt_period_us) 7366{ 7367 u64 rt_runtime, rt_period; 7368 7369 rt_period = (u64)rt_period_us * NSEC_PER_USEC; 7370 rt_runtime = tg->rt_bandwidth.rt_runtime; 7371 7372 if (rt_period == 0) 7373 return -EINVAL; 7374 7375 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 7376} 7377 7378static long sched_group_rt_period(struct task_group *tg) 7379{ 7380 u64 rt_period_us; 7381 7382 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period); 7383 do_div(rt_period_us, NSEC_PER_USEC); 7384 return rt_period_us; 7385} 7386#endif /* CONFIG_RT_GROUP_SCHED */ 7387 7388#ifdef CONFIG_RT_GROUP_SCHED 7389static int sched_rt_global_constraints(void) 7390{ 7391 int ret = 0; 7392 7393 mutex_lock(&rt_constraints_mutex); 7394 read_lock(&tasklist_lock); 7395 ret = __rt_schedulable(NULL, 0, 0); 7396 read_unlock(&tasklist_lock); 7397 mutex_unlock(&rt_constraints_mutex); 7398 7399 return ret; 7400} 7401 7402static int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk) 7403{ 7404 /* Don't accept realtime tasks when there is no way for them to run */ 7405 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0) 7406 return 0; 7407 7408 return 1; 7409} 7410 7411#else /* !CONFIG_RT_GROUP_SCHED */ 7412static int sched_rt_global_constraints(void) 7413{ 7414 unsigned long flags; 7415 int i, ret = 0; 7416 7417 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags); 7418 for_each_possible_cpu(i) { 7419 struct rt_rq *rt_rq = &cpu_rq(i)->rt; 7420 7421 raw_spin_lock(&rt_rq->rt_runtime_lock); 7422 rt_rq->rt_runtime = global_rt_runtime(); 7423 raw_spin_unlock(&rt_rq->rt_runtime_lock); 7424 } 7425 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags); 7426 7427 return ret; 7428} 7429#endif /* CONFIG_RT_GROUP_SCHED */ 7430 7431static int sched_dl_global_constraints(void) 7432{ 7433 u64 runtime = global_rt_runtime(); 7434 u64 period = global_rt_period(); 7435 u64 new_bw = to_ratio(period, runtime); 7436 int cpu, ret = 0; 7437 unsigned long flags; 7438 7439 /* 7440 * Here we want to check the bandwidth not being set to some 7441 * value smaller than the currently allocated bandwidth in 7442 * any of the root_domains. 7443 * 7444 * FIXME: Cycling on all the CPUs is overdoing, but simpler than 7445 * cycling on root_domains... Discussion on different/better 7446 * solutions is welcome! 7447 */ 7448 for_each_possible_cpu(cpu) { 7449 struct dl_bw *dl_b = dl_bw_of(cpu); 7450 7451 raw_spin_lock_irqsave(&dl_b->lock, flags); 7452 if (new_bw < dl_b->total_bw) 7453 ret = -EBUSY; 7454 raw_spin_unlock_irqrestore(&dl_b->lock, flags); 7455 7456 if (ret) 7457 break; 7458 } 7459 7460 return ret; 7461} 7462 7463static void sched_dl_do_global(void) 7464{ 7465 u64 new_bw = -1; 7466 int cpu; 7467 unsigned long flags; 7468 7469 def_dl_bandwidth.dl_period = global_rt_period(); 7470 def_dl_bandwidth.dl_runtime = global_rt_runtime(); 7471 7472 if (global_rt_runtime() != RUNTIME_INF) 7473 new_bw = to_ratio(global_rt_period(), global_rt_runtime()); 7474 7475 /* 7476 * FIXME: As above... 7477 */ 7478 for_each_possible_cpu(cpu) { 7479 struct dl_bw *dl_b = dl_bw_of(cpu); 7480 7481 raw_spin_lock_irqsave(&dl_b->lock, flags); 7482 dl_b->bw = new_bw; 7483 raw_spin_unlock_irqrestore(&dl_b->lock, flags); 7484 } 7485} 7486 7487static int sched_rt_global_validate(void) 7488{ 7489 if (sysctl_sched_rt_period <= 0) 7490 return -EINVAL; 7491 7492 if ((sysctl_sched_rt_runtime != RUNTIME_INF) && 7493 (sysctl_sched_rt_runtime > sysctl_sched_rt_period)) 7494 return -EINVAL; 7495 7496 return 0; 7497} 7498 7499static void sched_rt_do_global(void) 7500{ 7501 def_rt_bandwidth.rt_runtime = global_rt_runtime(); 7502 def_rt_bandwidth.rt_period = ns_to_ktime(global_rt_period()); 7503} 7504 7505int sched_rt_handler(struct ctl_table *table, int write, 7506 void __user *buffer, size_t *lenp, 7507 loff_t *ppos) 7508{ 7509 int old_period, old_runtime; 7510 static DEFINE_MUTEX(mutex); 7511 int ret; 7512 7513 mutex_lock(&mutex); 7514 old_period = sysctl_sched_rt_period; 7515 old_runtime = sysctl_sched_rt_runtime; 7516 7517 ret = proc_dointvec(table, write, buffer, lenp, ppos); 7518 7519 if (!ret && write) { 7520 ret = sched_rt_global_validate(); 7521 if (ret) 7522 goto undo; 7523 7524 ret = sched_rt_global_constraints(); 7525 if (ret) 7526 goto undo; 7527 7528 ret = sched_dl_global_constraints(); 7529 if (ret) 7530 goto undo; 7531 7532 sched_rt_do_global(); 7533 sched_dl_do_global(); 7534 } 7535 if (0) { 7536undo: 7537 sysctl_sched_rt_period = old_period; 7538 sysctl_sched_rt_runtime = old_runtime; 7539 } 7540 mutex_unlock(&mutex); 7541 7542 return ret; 7543} 7544 7545int sched_rr_handler(struct ctl_table *table, int write, 7546 void __user *buffer, size_t *lenp, 7547 loff_t *ppos) 7548{ 7549 int ret; 7550 static DEFINE_MUTEX(mutex); 7551 7552 mutex_lock(&mutex); 7553 ret = proc_dointvec(table, write, buffer, lenp, ppos); 7554 /* make sure that internally we keep jiffies */ 7555 /* also, writing zero resets timeslice to default */ 7556 if (!ret && write) { 7557 sched_rr_timeslice = sched_rr_timeslice <= 0 ? 7558 RR_TIMESLICE : msecs_to_jiffies(sched_rr_timeslice); 7559 } 7560 mutex_unlock(&mutex); 7561 return ret; 7562} 7563 7564#ifdef CONFIG_CGROUP_SCHED 7565 7566static inline struct task_group *css_tg(struct cgroup_subsys_state *css) 7567{ 7568 return css ? container_of(css, struct task_group, css) : NULL; 7569} 7570 7571static struct cgroup_subsys_state * 7572cpu_cgroup_css_alloc(struct cgroup_subsys_state *parent_css) 7573{ 7574 struct task_group *parent = css_tg(parent_css); 7575 struct task_group *tg; 7576 7577 if (!parent) { 7578 /* This is early initialization for the top cgroup */ 7579 return &root_task_group.css; 7580 } 7581 7582 tg = sched_create_group(parent); 7583 if (IS_ERR(tg)) 7584 return ERR_PTR(-ENOMEM); 7585 7586 return &tg->css; 7587} 7588 7589static int cpu_cgroup_css_online(struct cgroup_subsys_state *css) 7590{ 7591 struct task_group *tg = css_tg(css); 7592 struct task_group *parent = css_tg(css_parent(css)); 7593 7594 if (parent) 7595 sched_online_group(tg, parent); 7596 return 0; 7597} 7598 7599static void cpu_cgroup_css_free(struct cgroup_subsys_state *css) 7600{ 7601 struct task_group *tg = css_tg(css); 7602 7603 sched_destroy_group(tg); 7604} 7605 7606static void cpu_cgroup_css_offline(struct cgroup_subsys_state *css) 7607{ 7608 struct task_group *tg = css_tg(css); 7609 7610 sched_offline_group(tg); 7611} 7612 7613static int cpu_cgroup_can_attach(struct cgroup_subsys_state *css, 7614 struct cgroup_taskset *tset) 7615{ 7616 struct task_struct *task; 7617 7618 cgroup_taskset_for_each(task, tset) { 7619#ifdef CONFIG_RT_GROUP_SCHED 7620 if (!sched_rt_can_attach(css_tg(css), task)) 7621 return -EINVAL; 7622#else 7623 /* We don't support RT-tasks being in separate groups */ 7624 if (task->sched_class != &fair_sched_class) 7625 return -EINVAL; 7626#endif 7627 } 7628 return 0; 7629} 7630 7631static void cpu_cgroup_attach(struct cgroup_subsys_state *css, 7632 struct cgroup_taskset *tset) 7633{ 7634 struct task_struct *task; 7635 7636 cgroup_taskset_for_each(task, tset) 7637 sched_move_task(task); 7638} 7639 7640static void cpu_cgroup_exit(struct cgroup_subsys_state *css, 7641 struct cgroup_subsys_state *old_css, 7642 struct task_struct *task) 7643{ 7644 /* 7645 * cgroup_exit() is called in the copy_process() failure path. 7646 * Ignore this case since the task hasn't ran yet, this avoids 7647 * trying to poke a half freed task state from generic code. 7648 */ 7649 if (!(task->flags & PF_EXITING)) 7650 return; 7651 7652 sched_move_task(task); 7653} 7654 7655#ifdef CONFIG_FAIR_GROUP_SCHED 7656static int cpu_shares_write_u64(struct cgroup_subsys_state *css, 7657 struct cftype *cftype, u64 shareval) 7658{ 7659 return sched_group_set_shares(css_tg(css), scale_load(shareval)); 7660} 7661 7662static u64 cpu_shares_read_u64(struct cgroup_subsys_state *css, 7663 struct cftype *cft) 7664{ 7665 struct task_group *tg = css_tg(css); 7666 7667 return (u64) scale_load_down(tg->shares); 7668} 7669 7670#ifdef CONFIG_CFS_BANDWIDTH 7671static DEFINE_MUTEX(cfs_constraints_mutex); 7672 7673const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */ 7674const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */ 7675 7676static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime); 7677 7678static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota) 7679{ 7680 int i, ret = 0, runtime_enabled, runtime_was_enabled; 7681 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7682 7683 if (tg == &root_task_group) 7684 return -EINVAL; 7685 7686 /* 7687 * Ensure we have at some amount of bandwidth every period. This is 7688 * to prevent reaching a state of large arrears when throttled via 7689 * entity_tick() resulting in prolonged exit starvation. 7690 */ 7691 if (quota < min_cfs_quota_period || period < min_cfs_quota_period) 7692 return -EINVAL; 7693 7694 /* 7695 * Likewise, bound things on the otherside by preventing insane quota 7696 * periods. This also allows us to normalize in computing quota 7697 * feasibility. 7698 */ 7699 if (period > max_cfs_quota_period) 7700 return -EINVAL; 7701 7702 mutex_lock(&cfs_constraints_mutex); 7703 ret = __cfs_schedulable(tg, period, quota); 7704 if (ret) 7705 goto out_unlock; 7706 7707 runtime_enabled = quota != RUNTIME_INF; 7708 runtime_was_enabled = cfs_b->quota != RUNTIME_INF; 7709 /* 7710 * If we need to toggle cfs_bandwidth_used, off->on must occur 7711 * before making related changes, and on->off must occur afterwards 7712 */ 7713 if (runtime_enabled && !runtime_was_enabled) 7714 cfs_bandwidth_usage_inc(); 7715 raw_spin_lock_irq(&cfs_b->lock); 7716 cfs_b->period = ns_to_ktime(period); 7717 cfs_b->quota = quota; 7718 7719 __refill_cfs_bandwidth_runtime(cfs_b); 7720 /* restart the period timer (if active) to handle new period expiry */ 7721 if (runtime_enabled && cfs_b->timer_active) { 7722 /* force a reprogram */ 7723 cfs_b->timer_active = 0; 7724 __start_cfs_bandwidth(cfs_b); 7725 } 7726 raw_spin_unlock_irq(&cfs_b->lock); 7727 7728 for_each_possible_cpu(i) { 7729 struct cfs_rq *cfs_rq = tg->cfs_rq[i]; 7730 struct rq *rq = cfs_rq->rq; 7731 7732 raw_spin_lock_irq(&rq->lock); 7733 cfs_rq->runtime_enabled = runtime_enabled; 7734 cfs_rq->runtime_remaining = 0; 7735 7736 if (cfs_rq->throttled) 7737 unthrottle_cfs_rq(cfs_rq); 7738 raw_spin_unlock_irq(&rq->lock); 7739 } 7740 if (runtime_was_enabled && !runtime_enabled) 7741 cfs_bandwidth_usage_dec(); 7742out_unlock: 7743 mutex_unlock(&cfs_constraints_mutex); 7744 7745 return ret; 7746} 7747 7748int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us) 7749{ 7750 u64 quota, period; 7751 7752 period = ktime_to_ns(tg->cfs_bandwidth.period); 7753 if (cfs_quota_us < 0) 7754 quota = RUNTIME_INF; 7755 else 7756 quota = (u64)cfs_quota_us * NSEC_PER_USEC; 7757 7758 return tg_set_cfs_bandwidth(tg, period, quota); 7759} 7760 7761long tg_get_cfs_quota(struct task_group *tg) 7762{ 7763 u64 quota_us; 7764 7765 if (tg->cfs_bandwidth.quota == RUNTIME_INF) 7766 return -1; 7767 7768 quota_us = tg->cfs_bandwidth.quota; 7769 do_div(quota_us, NSEC_PER_USEC); 7770 7771 return quota_us; 7772} 7773 7774int tg_set_cfs_period(struct task_group *tg, long cfs_period_us) 7775{ 7776 u64 quota, period; 7777 7778 period = (u64)cfs_period_us * NSEC_PER_USEC; 7779 quota = tg->cfs_bandwidth.quota; 7780 7781 return tg_set_cfs_bandwidth(tg, period, quota); 7782} 7783 7784long tg_get_cfs_period(struct task_group *tg) 7785{ 7786 u64 cfs_period_us; 7787 7788 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period); 7789 do_div(cfs_period_us, NSEC_PER_USEC); 7790 7791 return cfs_period_us; 7792} 7793 7794static s64 cpu_cfs_quota_read_s64(struct cgroup_subsys_state *css, 7795 struct cftype *cft) 7796{ 7797 return tg_get_cfs_quota(css_tg(css)); 7798} 7799 7800static int cpu_cfs_quota_write_s64(struct cgroup_subsys_state *css, 7801 struct cftype *cftype, s64 cfs_quota_us) 7802{ 7803 return tg_set_cfs_quota(css_tg(css), cfs_quota_us); 7804} 7805 7806static u64 cpu_cfs_period_read_u64(struct cgroup_subsys_state *css, 7807 struct cftype *cft) 7808{ 7809 return tg_get_cfs_period(css_tg(css)); 7810} 7811 7812static int cpu_cfs_period_write_u64(struct cgroup_subsys_state *css, 7813 struct cftype *cftype, u64 cfs_period_us) 7814{ 7815 return tg_set_cfs_period(css_tg(css), cfs_period_us); 7816} 7817 7818struct cfs_schedulable_data { 7819 struct task_group *tg; 7820 u64 period, quota; 7821}; 7822 7823/* 7824 * normalize group quota/period to be quota/max_period 7825 * note: units are usecs 7826 */ 7827static u64 normalize_cfs_quota(struct task_group *tg, 7828 struct cfs_schedulable_data *d) 7829{ 7830 u64 quota, period; 7831 7832 if (tg == d->tg) { 7833 period = d->period; 7834 quota = d->quota; 7835 } else { 7836 period = tg_get_cfs_period(tg); 7837 quota = tg_get_cfs_quota(tg); 7838 } 7839 7840 /* note: these should typically be equivalent */ 7841 if (quota == RUNTIME_INF || quota == -1) 7842 return RUNTIME_INF; 7843 7844 return to_ratio(period, quota); 7845} 7846 7847static int tg_cfs_schedulable_down(struct task_group *tg, void *data) 7848{ 7849 struct cfs_schedulable_data *d = data; 7850 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7851 s64 quota = 0, parent_quota = -1; 7852 7853 if (!tg->parent) { 7854 quota = RUNTIME_INF; 7855 } else { 7856 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth; 7857 7858 quota = normalize_cfs_quota(tg, d); 7859 parent_quota = parent_b->hierarchal_quota; 7860 7861 /* 7862 * ensure max(child_quota) <= parent_quota, inherit when no 7863 * limit is set 7864 */ 7865 if (quota == RUNTIME_INF) 7866 quota = parent_quota; 7867 else if (parent_quota != RUNTIME_INF && quota > parent_quota) 7868 return -EINVAL; 7869 } 7870 cfs_b->hierarchal_quota = quota; 7871 7872 return 0; 7873} 7874 7875static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota) 7876{ 7877 int ret; 7878 struct cfs_schedulable_data data = { 7879 .tg = tg, 7880 .period = period, 7881 .quota = quota, 7882 }; 7883 7884 if (quota != RUNTIME_INF) { 7885 do_div(data.period, NSEC_PER_USEC); 7886 do_div(data.quota, NSEC_PER_USEC); 7887 } 7888 7889 rcu_read_lock(); 7890 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data); 7891 rcu_read_unlock(); 7892 7893 return ret; 7894} 7895 7896static int cpu_stats_show(struct seq_file *sf, void *v) 7897{ 7898 struct task_group *tg = css_tg(seq_css(sf)); 7899 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7900 7901 seq_printf(sf, "nr_periods %d\n", cfs_b->nr_periods); 7902 seq_printf(sf, "nr_throttled %d\n", cfs_b->nr_throttled); 7903 seq_printf(sf, "throttled_time %llu\n", cfs_b->throttled_time); 7904 7905 return 0; 7906} 7907#endif /* CONFIG_CFS_BANDWIDTH */ 7908#endif /* CONFIG_FAIR_GROUP_SCHED */ 7909 7910#ifdef CONFIG_RT_GROUP_SCHED 7911static int cpu_rt_runtime_write(struct cgroup_subsys_state *css, 7912 struct cftype *cft, s64 val) 7913{ 7914 return sched_group_set_rt_runtime(css_tg(css), val); 7915} 7916 7917static s64 cpu_rt_runtime_read(struct cgroup_subsys_state *css, 7918 struct cftype *cft) 7919{ 7920 return sched_group_rt_runtime(css_tg(css)); 7921} 7922 7923static int cpu_rt_period_write_uint(struct cgroup_subsys_state *css, 7924 struct cftype *cftype, u64 rt_period_us) 7925{ 7926 return sched_group_set_rt_period(css_tg(css), rt_period_us); 7927} 7928 7929static u64 cpu_rt_period_read_uint(struct cgroup_subsys_state *css, 7930 struct cftype *cft) 7931{ 7932 return sched_group_rt_period(css_tg(css)); 7933} 7934#endif /* CONFIG_RT_GROUP_SCHED */ 7935 7936static struct cftype cpu_files[] = { 7937#ifdef CONFIG_FAIR_GROUP_SCHED 7938 { 7939 .name = "shares", 7940 .read_u64 = cpu_shares_read_u64, 7941 .write_u64 = cpu_shares_write_u64, 7942 }, 7943#endif 7944#ifdef CONFIG_CFS_BANDWIDTH 7945 { 7946 .name = "cfs_quota_us", 7947 .read_s64 = cpu_cfs_quota_read_s64, 7948 .write_s64 = cpu_cfs_quota_write_s64, 7949 }, 7950 { 7951 .name = "cfs_period_us", 7952 .read_u64 = cpu_cfs_period_read_u64, 7953 .write_u64 = cpu_cfs_period_write_u64, 7954 }, 7955 { 7956 .name = "stat", 7957 .seq_show = cpu_stats_show, 7958 }, 7959#endif 7960#ifdef CONFIG_RT_GROUP_SCHED 7961 { 7962 .name = "rt_runtime_us", 7963 .read_s64 = cpu_rt_runtime_read, 7964 .write_s64 = cpu_rt_runtime_write, 7965 }, 7966 { 7967 .name = "rt_period_us", 7968 .read_u64 = cpu_rt_period_read_uint, 7969 .write_u64 = cpu_rt_period_write_uint, 7970 }, 7971#endif 7972 { } /* terminate */ 7973}; 7974 7975struct cgroup_subsys cpu_cgrp_subsys = { 7976 .css_alloc = cpu_cgroup_css_alloc, 7977 .css_free = cpu_cgroup_css_free, 7978 .css_online = cpu_cgroup_css_online, 7979 .css_offline = cpu_cgroup_css_offline, 7980 .can_attach = cpu_cgroup_can_attach, 7981 .attach = cpu_cgroup_attach, 7982 .exit = cpu_cgroup_exit, 7983 .base_cftypes = cpu_files, 7984 .early_init = 1, 7985}; 7986 7987#endif /* CONFIG_CGROUP_SCHED */ 7988 7989void dump_cpu_task(int cpu) 7990{ 7991 pr_info("Task dump for CPU %d:\n", cpu); 7992 sched_show_task(cpu_curr(cpu)); 7993} 7994