core.c revision c308b56b5398779cd3da0f62ab26b0453494c3d4
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 75#include <asm/tlb.h> 76#include <asm/irq_regs.h> 77#include <asm/mutex.h> 78#ifdef CONFIG_PARAVIRT 79#include <asm/paravirt.h> 80#endif 81 82#include "sched.h" 83#include "../workqueue_sched.h" 84 85#define CREATE_TRACE_POINTS 86#include <trace/events/sched.h> 87 88void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period) 89{ 90 unsigned long delta; 91 ktime_t soft, hard, now; 92 93 for (;;) { 94 if (hrtimer_active(period_timer)) 95 break; 96 97 now = hrtimer_cb_get_time(period_timer); 98 hrtimer_forward(period_timer, now, period); 99 100 soft = hrtimer_get_softexpires(period_timer); 101 hard = hrtimer_get_expires(period_timer); 102 delta = ktime_to_ns(ktime_sub(hard, soft)); 103 __hrtimer_start_range_ns(period_timer, soft, delta, 104 HRTIMER_MODE_ABS_PINNED, 0); 105 } 106} 107 108DEFINE_MUTEX(sched_domains_mutex); 109DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues); 110 111static void update_rq_clock_task(struct rq *rq, s64 delta); 112 113void update_rq_clock(struct rq *rq) 114{ 115 s64 delta; 116 117 if (rq->skip_clock_update > 0) 118 return; 119 120 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock; 121 rq->clock += delta; 122 update_rq_clock_task(rq, delta); 123} 124 125/* 126 * Debugging: various feature bits 127 */ 128 129#define SCHED_FEAT(name, enabled) \ 130 (1UL << __SCHED_FEAT_##name) * enabled | 131 132const_debug unsigned int sysctl_sched_features = 133#include "features.h" 134 0; 135 136#undef SCHED_FEAT 137 138#ifdef CONFIG_SCHED_DEBUG 139#define SCHED_FEAT(name, enabled) \ 140 #name , 141 142static __read_mostly char *sched_feat_names[] = { 143#include "features.h" 144 NULL 145}; 146 147#undef SCHED_FEAT 148 149static int sched_feat_show(struct seq_file *m, void *v) 150{ 151 int i; 152 153 for (i = 0; i < __SCHED_FEAT_NR; i++) { 154 if (!(sysctl_sched_features & (1UL << i))) 155 seq_puts(m, "NO_"); 156 seq_printf(m, "%s ", sched_feat_names[i]); 157 } 158 seq_puts(m, "\n"); 159 160 return 0; 161} 162 163#ifdef HAVE_JUMP_LABEL 164 165#define jump_label_key__true jump_label_key_enabled 166#define jump_label_key__false jump_label_key_disabled 167 168#define SCHED_FEAT(name, enabled) \ 169 jump_label_key__##enabled , 170 171struct jump_label_key sched_feat_keys[__SCHED_FEAT_NR] = { 172#include "features.h" 173}; 174 175#undef SCHED_FEAT 176 177static void sched_feat_disable(int i) 178{ 179 if (jump_label_enabled(&sched_feat_keys[i])) 180 jump_label_dec(&sched_feat_keys[i]); 181} 182 183static void sched_feat_enable(int i) 184{ 185 if (!jump_label_enabled(&sched_feat_keys[i])) 186 jump_label_inc(&sched_feat_keys[i]); 187} 188#else 189static void sched_feat_disable(int i) { }; 190static void sched_feat_enable(int i) { }; 191#endif /* HAVE_JUMP_LABEL */ 192 193static ssize_t 194sched_feat_write(struct file *filp, const char __user *ubuf, 195 size_t cnt, loff_t *ppos) 196{ 197 char buf[64]; 198 char *cmp; 199 int neg = 0; 200 int i; 201 202 if (cnt > 63) 203 cnt = 63; 204 205 if (copy_from_user(&buf, ubuf, cnt)) 206 return -EFAULT; 207 208 buf[cnt] = 0; 209 cmp = strstrip(buf); 210 211 if (strncmp(cmp, "NO_", 3) == 0) { 212 neg = 1; 213 cmp += 3; 214 } 215 216 for (i = 0; i < __SCHED_FEAT_NR; i++) { 217 if (strcmp(cmp, sched_feat_names[i]) == 0) { 218 if (neg) { 219 sysctl_sched_features &= ~(1UL << i); 220 sched_feat_disable(i); 221 } else { 222 sysctl_sched_features |= (1UL << i); 223 sched_feat_enable(i); 224 } 225 break; 226 } 227 } 228 229 if (i == __SCHED_FEAT_NR) 230 return -EINVAL; 231 232 *ppos += cnt; 233 234 return cnt; 235} 236 237static int sched_feat_open(struct inode *inode, struct file *filp) 238{ 239 return single_open(filp, sched_feat_show, NULL); 240} 241 242static const struct file_operations sched_feat_fops = { 243 .open = sched_feat_open, 244 .write = sched_feat_write, 245 .read = seq_read, 246 .llseek = seq_lseek, 247 .release = single_release, 248}; 249 250static __init int sched_init_debug(void) 251{ 252 debugfs_create_file("sched_features", 0644, NULL, NULL, 253 &sched_feat_fops); 254 255 return 0; 256} 257late_initcall(sched_init_debug); 258#endif /* CONFIG_SCHED_DEBUG */ 259 260/* 261 * Number of tasks to iterate in a single balance run. 262 * Limited because this is done with IRQs disabled. 263 */ 264const_debug unsigned int sysctl_sched_nr_migrate = 32; 265 266/* 267 * period over which we average the RT time consumption, measured 268 * in ms. 269 * 270 * default: 1s 271 */ 272const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC; 273 274/* 275 * period over which we measure -rt task cpu usage in us. 276 * default: 1s 277 */ 278unsigned int sysctl_sched_rt_period = 1000000; 279 280__read_mostly int scheduler_running; 281 282/* 283 * part of the period that we allow rt tasks to run in us. 284 * default: 0.95s 285 */ 286int sysctl_sched_rt_runtime = 950000; 287 288 289 290/* 291 * __task_rq_lock - lock the rq @p resides on. 292 */ 293static inline struct rq *__task_rq_lock(struct task_struct *p) 294 __acquires(rq->lock) 295{ 296 struct rq *rq; 297 298 lockdep_assert_held(&p->pi_lock); 299 300 for (;;) { 301 rq = task_rq(p); 302 raw_spin_lock(&rq->lock); 303 if (likely(rq == task_rq(p))) 304 return rq; 305 raw_spin_unlock(&rq->lock); 306 } 307} 308 309/* 310 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on. 311 */ 312static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags) 313 __acquires(p->pi_lock) 314 __acquires(rq->lock) 315{ 316 struct rq *rq; 317 318 for (;;) { 319 raw_spin_lock_irqsave(&p->pi_lock, *flags); 320 rq = task_rq(p); 321 raw_spin_lock(&rq->lock); 322 if (likely(rq == task_rq(p))) 323 return rq; 324 raw_spin_unlock(&rq->lock); 325 raw_spin_unlock_irqrestore(&p->pi_lock, *flags); 326 } 327} 328 329static void __task_rq_unlock(struct rq *rq) 330 __releases(rq->lock) 331{ 332 raw_spin_unlock(&rq->lock); 333} 334 335static inline void 336task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags) 337 __releases(rq->lock) 338 __releases(p->pi_lock) 339{ 340 raw_spin_unlock(&rq->lock); 341 raw_spin_unlock_irqrestore(&p->pi_lock, *flags); 342} 343 344/* 345 * this_rq_lock - lock this runqueue and disable interrupts. 346 */ 347static struct rq *this_rq_lock(void) 348 __acquires(rq->lock) 349{ 350 struct rq *rq; 351 352 local_irq_disable(); 353 rq = this_rq(); 354 raw_spin_lock(&rq->lock); 355 356 return rq; 357} 358 359#ifdef CONFIG_SCHED_HRTICK 360/* 361 * Use HR-timers to deliver accurate preemption points. 362 * 363 * Its all a bit involved since we cannot program an hrt while holding the 364 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a 365 * reschedule event. 366 * 367 * When we get rescheduled we reprogram the hrtick_timer outside of the 368 * rq->lock. 369 */ 370 371static void hrtick_clear(struct rq *rq) 372{ 373 if (hrtimer_active(&rq->hrtick_timer)) 374 hrtimer_cancel(&rq->hrtick_timer); 375} 376 377/* 378 * High-resolution timer tick. 379 * Runs from hardirq context with interrupts disabled. 380 */ 381static enum hrtimer_restart hrtick(struct hrtimer *timer) 382{ 383 struct rq *rq = container_of(timer, struct rq, hrtick_timer); 384 385 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id()); 386 387 raw_spin_lock(&rq->lock); 388 update_rq_clock(rq); 389 rq->curr->sched_class->task_tick(rq, rq->curr, 1); 390 raw_spin_unlock(&rq->lock); 391 392 return HRTIMER_NORESTART; 393} 394 395#ifdef CONFIG_SMP 396/* 397 * called from hardirq (IPI) context 398 */ 399static void __hrtick_start(void *arg) 400{ 401 struct rq *rq = arg; 402 403 raw_spin_lock(&rq->lock); 404 hrtimer_restart(&rq->hrtick_timer); 405 rq->hrtick_csd_pending = 0; 406 raw_spin_unlock(&rq->lock); 407} 408 409/* 410 * Called to set the hrtick timer state. 411 * 412 * called with rq->lock held and irqs disabled 413 */ 414void hrtick_start(struct rq *rq, u64 delay) 415{ 416 struct hrtimer *timer = &rq->hrtick_timer; 417 ktime_t time = ktime_add_ns(timer->base->get_time(), delay); 418 419 hrtimer_set_expires(timer, time); 420 421 if (rq == this_rq()) { 422 hrtimer_restart(timer); 423 } else if (!rq->hrtick_csd_pending) { 424 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0); 425 rq->hrtick_csd_pending = 1; 426 } 427} 428 429static int 430hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu) 431{ 432 int cpu = (int)(long)hcpu; 433 434 switch (action) { 435 case CPU_UP_CANCELED: 436 case CPU_UP_CANCELED_FROZEN: 437 case CPU_DOWN_PREPARE: 438 case CPU_DOWN_PREPARE_FROZEN: 439 case CPU_DEAD: 440 case CPU_DEAD_FROZEN: 441 hrtick_clear(cpu_rq(cpu)); 442 return NOTIFY_OK; 443 } 444 445 return NOTIFY_DONE; 446} 447 448static __init void init_hrtick(void) 449{ 450 hotcpu_notifier(hotplug_hrtick, 0); 451} 452#else 453/* 454 * Called to set the hrtick timer state. 455 * 456 * called with rq->lock held and irqs disabled 457 */ 458void hrtick_start(struct rq *rq, u64 delay) 459{ 460 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0, 461 HRTIMER_MODE_REL_PINNED, 0); 462} 463 464static inline void init_hrtick(void) 465{ 466} 467#endif /* CONFIG_SMP */ 468 469static void init_rq_hrtick(struct rq *rq) 470{ 471#ifdef CONFIG_SMP 472 rq->hrtick_csd_pending = 0; 473 474 rq->hrtick_csd.flags = 0; 475 rq->hrtick_csd.func = __hrtick_start; 476 rq->hrtick_csd.info = rq; 477#endif 478 479 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); 480 rq->hrtick_timer.function = hrtick; 481} 482#else /* CONFIG_SCHED_HRTICK */ 483static inline void hrtick_clear(struct rq *rq) 484{ 485} 486 487static inline void init_rq_hrtick(struct rq *rq) 488{ 489} 490 491static inline void init_hrtick(void) 492{ 493} 494#endif /* CONFIG_SCHED_HRTICK */ 495 496/* 497 * resched_task - mark a task 'to be rescheduled now'. 498 * 499 * On UP this means the setting of the need_resched flag, on SMP it 500 * might also involve a cross-CPU call to trigger the scheduler on 501 * the target CPU. 502 */ 503#ifdef CONFIG_SMP 504 505#ifndef tsk_is_polling 506#define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG) 507#endif 508 509void resched_task(struct task_struct *p) 510{ 511 int cpu; 512 513 assert_raw_spin_locked(&task_rq(p)->lock); 514 515 if (test_tsk_need_resched(p)) 516 return; 517 518 set_tsk_need_resched(p); 519 520 cpu = task_cpu(p); 521 if (cpu == smp_processor_id()) 522 return; 523 524 /* NEED_RESCHED must be visible before we test polling */ 525 smp_mb(); 526 if (!tsk_is_polling(p)) 527 smp_send_reschedule(cpu); 528} 529 530void resched_cpu(int cpu) 531{ 532 struct rq *rq = cpu_rq(cpu); 533 unsigned long flags; 534 535 if (!raw_spin_trylock_irqsave(&rq->lock, flags)) 536 return; 537 resched_task(cpu_curr(cpu)); 538 raw_spin_unlock_irqrestore(&rq->lock, flags); 539} 540 541#ifdef CONFIG_NO_HZ 542/* 543 * In the semi idle case, use the nearest busy cpu for migrating timers 544 * from an idle cpu. This is good for power-savings. 545 * 546 * We don't do similar optimization for completely idle system, as 547 * selecting an idle cpu will add more delays to the timers than intended 548 * (as that cpu's timer base may not be uptodate wrt jiffies etc). 549 */ 550int get_nohz_timer_target(void) 551{ 552 int cpu = smp_processor_id(); 553 int i; 554 struct sched_domain *sd; 555 556 rcu_read_lock(); 557 for_each_domain(cpu, sd) { 558 for_each_cpu(i, sched_domain_span(sd)) { 559 if (!idle_cpu(i)) { 560 cpu = i; 561 goto unlock; 562 } 563 } 564 } 565unlock: 566 rcu_read_unlock(); 567 return cpu; 568} 569/* 570 * When add_timer_on() enqueues a timer into the timer wheel of an 571 * idle CPU then this timer might expire before the next timer event 572 * which is scheduled to wake up that CPU. In case of a completely 573 * idle system the next event might even be infinite time into the 574 * future. wake_up_idle_cpu() ensures that the CPU is woken up and 575 * leaves the inner idle loop so the newly added timer is taken into 576 * account when the CPU goes back to idle and evaluates the timer 577 * wheel for the next timer event. 578 */ 579void wake_up_idle_cpu(int cpu) 580{ 581 struct rq *rq = cpu_rq(cpu); 582 583 if (cpu == smp_processor_id()) 584 return; 585 586 /* 587 * This is safe, as this function is called with the timer 588 * wheel base lock of (cpu) held. When the CPU is on the way 589 * to idle and has not yet set rq->curr to idle then it will 590 * be serialized on the timer wheel base lock and take the new 591 * timer into account automatically. 592 */ 593 if (rq->curr != rq->idle) 594 return; 595 596 /* 597 * We can set TIF_RESCHED on the idle task of the other CPU 598 * lockless. The worst case is that the other CPU runs the 599 * idle task through an additional NOOP schedule() 600 */ 601 set_tsk_need_resched(rq->idle); 602 603 /* NEED_RESCHED must be visible before we test polling */ 604 smp_mb(); 605 if (!tsk_is_polling(rq->idle)) 606 smp_send_reschedule(cpu); 607} 608 609static inline bool got_nohz_idle_kick(void) 610{ 611 int cpu = smp_processor_id(); 612 return idle_cpu(cpu) && test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)); 613} 614 615#else /* CONFIG_NO_HZ */ 616 617static inline bool got_nohz_idle_kick(void) 618{ 619 return false; 620} 621 622#endif /* CONFIG_NO_HZ */ 623 624void sched_avg_update(struct rq *rq) 625{ 626 s64 period = sched_avg_period(); 627 628 while ((s64)(rq->clock - rq->age_stamp) > period) { 629 /* 630 * Inline assembly required to prevent the compiler 631 * optimising this loop into a divmod call. 632 * See __iter_div_u64_rem() for another example of this. 633 */ 634 asm("" : "+rm" (rq->age_stamp)); 635 rq->age_stamp += period; 636 rq->rt_avg /= 2; 637 } 638} 639 640#else /* !CONFIG_SMP */ 641void resched_task(struct task_struct *p) 642{ 643 assert_raw_spin_locked(&task_rq(p)->lock); 644 set_tsk_need_resched(p); 645} 646#endif /* CONFIG_SMP */ 647 648#if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \ 649 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH))) 650/* 651 * Iterate task_group tree rooted at *from, calling @down when first entering a 652 * node and @up when leaving it for the final time. 653 * 654 * Caller must hold rcu_lock or sufficient equivalent. 655 */ 656int walk_tg_tree_from(struct task_group *from, 657 tg_visitor down, tg_visitor up, void *data) 658{ 659 struct task_group *parent, *child; 660 int ret; 661 662 parent = from; 663 664down: 665 ret = (*down)(parent, data); 666 if (ret) 667 goto out; 668 list_for_each_entry_rcu(child, &parent->children, siblings) { 669 parent = child; 670 goto down; 671 672up: 673 continue; 674 } 675 ret = (*up)(parent, data); 676 if (ret || parent == from) 677 goto out; 678 679 child = parent; 680 parent = parent->parent; 681 if (parent) 682 goto up; 683out: 684 return ret; 685} 686 687int tg_nop(struct task_group *tg, void *data) 688{ 689 return 0; 690} 691#endif 692 693void update_cpu_load(struct rq *this_rq); 694 695static void set_load_weight(struct task_struct *p) 696{ 697 int prio = p->static_prio - MAX_RT_PRIO; 698 struct load_weight *load = &p->se.load; 699 700 /* 701 * SCHED_IDLE tasks get minimal weight: 702 */ 703 if (p->policy == SCHED_IDLE) { 704 load->weight = scale_load(WEIGHT_IDLEPRIO); 705 load->inv_weight = WMULT_IDLEPRIO; 706 return; 707 } 708 709 load->weight = scale_load(prio_to_weight[prio]); 710 load->inv_weight = prio_to_wmult[prio]; 711} 712 713static void enqueue_task(struct rq *rq, struct task_struct *p, int flags) 714{ 715 update_rq_clock(rq); 716 sched_info_queued(p); 717 p->sched_class->enqueue_task(rq, p, flags); 718} 719 720static void dequeue_task(struct rq *rq, struct task_struct *p, int flags) 721{ 722 update_rq_clock(rq); 723 sched_info_dequeued(p); 724 p->sched_class->dequeue_task(rq, p, flags); 725} 726 727void activate_task(struct rq *rq, struct task_struct *p, int flags) 728{ 729 if (task_contributes_to_load(p)) 730 rq->nr_uninterruptible--; 731 732 enqueue_task(rq, p, flags); 733} 734 735void deactivate_task(struct rq *rq, struct task_struct *p, int flags) 736{ 737 if (task_contributes_to_load(p)) 738 rq->nr_uninterruptible++; 739 740 dequeue_task(rq, p, flags); 741} 742 743#ifdef CONFIG_IRQ_TIME_ACCOUNTING 744 745/* 746 * There are no locks covering percpu hardirq/softirq time. 747 * They are only modified in account_system_vtime, on corresponding CPU 748 * with interrupts disabled. So, writes are safe. 749 * They are read and saved off onto struct rq in update_rq_clock(). 750 * This may result in other CPU reading this CPU's irq time and can 751 * race with irq/account_system_vtime on this CPU. We would either get old 752 * or new value with a side effect of accounting a slice of irq time to wrong 753 * task when irq is in progress while we read rq->clock. That is a worthy 754 * compromise in place of having locks on each irq in account_system_time. 755 */ 756static DEFINE_PER_CPU(u64, cpu_hardirq_time); 757static DEFINE_PER_CPU(u64, cpu_softirq_time); 758 759static DEFINE_PER_CPU(u64, irq_start_time); 760static int sched_clock_irqtime; 761 762void enable_sched_clock_irqtime(void) 763{ 764 sched_clock_irqtime = 1; 765} 766 767void disable_sched_clock_irqtime(void) 768{ 769 sched_clock_irqtime = 0; 770} 771 772#ifndef CONFIG_64BIT 773static DEFINE_PER_CPU(seqcount_t, irq_time_seq); 774 775static inline void irq_time_write_begin(void) 776{ 777 __this_cpu_inc(irq_time_seq.sequence); 778 smp_wmb(); 779} 780 781static inline void irq_time_write_end(void) 782{ 783 smp_wmb(); 784 __this_cpu_inc(irq_time_seq.sequence); 785} 786 787static inline u64 irq_time_read(int cpu) 788{ 789 u64 irq_time; 790 unsigned seq; 791 792 do { 793 seq = read_seqcount_begin(&per_cpu(irq_time_seq, cpu)); 794 irq_time = per_cpu(cpu_softirq_time, cpu) + 795 per_cpu(cpu_hardirq_time, cpu); 796 } while (read_seqcount_retry(&per_cpu(irq_time_seq, cpu), seq)); 797 798 return irq_time; 799} 800#else /* CONFIG_64BIT */ 801static inline void irq_time_write_begin(void) 802{ 803} 804 805static inline void irq_time_write_end(void) 806{ 807} 808 809static inline u64 irq_time_read(int cpu) 810{ 811 return per_cpu(cpu_softirq_time, cpu) + per_cpu(cpu_hardirq_time, cpu); 812} 813#endif /* CONFIG_64BIT */ 814 815/* 816 * Called before incrementing preempt_count on {soft,}irq_enter 817 * and before decrementing preempt_count on {soft,}irq_exit. 818 */ 819void account_system_vtime(struct task_struct *curr) 820{ 821 unsigned long flags; 822 s64 delta; 823 int cpu; 824 825 if (!sched_clock_irqtime) 826 return; 827 828 local_irq_save(flags); 829 830 cpu = smp_processor_id(); 831 delta = sched_clock_cpu(cpu) - __this_cpu_read(irq_start_time); 832 __this_cpu_add(irq_start_time, delta); 833 834 irq_time_write_begin(); 835 /* 836 * We do not account for softirq time from ksoftirqd here. 837 * We want to continue accounting softirq time to ksoftirqd thread 838 * in that case, so as not to confuse scheduler with a special task 839 * that do not consume any time, but still wants to run. 840 */ 841 if (hardirq_count()) 842 __this_cpu_add(cpu_hardirq_time, delta); 843 else if (in_serving_softirq() && curr != this_cpu_ksoftirqd()) 844 __this_cpu_add(cpu_softirq_time, delta); 845 846 irq_time_write_end(); 847 local_irq_restore(flags); 848} 849EXPORT_SYMBOL_GPL(account_system_vtime); 850 851#endif /* CONFIG_IRQ_TIME_ACCOUNTING */ 852 853#ifdef CONFIG_PARAVIRT 854static inline u64 steal_ticks(u64 steal) 855{ 856 if (unlikely(steal > NSEC_PER_SEC)) 857 return div_u64(steal, TICK_NSEC); 858 859 return __iter_div_u64_rem(steal, TICK_NSEC, &steal); 860} 861#endif 862 863static void update_rq_clock_task(struct rq *rq, s64 delta) 864{ 865/* 866 * In theory, the compile should just see 0 here, and optimize out the call 867 * to sched_rt_avg_update. But I don't trust it... 868 */ 869#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING) 870 s64 steal = 0, irq_delta = 0; 871#endif 872#ifdef CONFIG_IRQ_TIME_ACCOUNTING 873 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time; 874 875 /* 876 * Since irq_time is only updated on {soft,}irq_exit, we might run into 877 * this case when a previous update_rq_clock() happened inside a 878 * {soft,}irq region. 879 * 880 * When this happens, we stop ->clock_task and only update the 881 * prev_irq_time stamp to account for the part that fit, so that a next 882 * update will consume the rest. This ensures ->clock_task is 883 * monotonic. 884 * 885 * It does however cause some slight miss-attribution of {soft,}irq 886 * time, a more accurate solution would be to update the irq_time using 887 * the current rq->clock timestamp, except that would require using 888 * atomic ops. 889 */ 890 if (irq_delta > delta) 891 irq_delta = delta; 892 893 rq->prev_irq_time += irq_delta; 894 delta -= irq_delta; 895#endif 896#ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING 897 if (static_branch((¶virt_steal_rq_enabled))) { 898 u64 st; 899 900 steal = paravirt_steal_clock(cpu_of(rq)); 901 steal -= rq->prev_steal_time_rq; 902 903 if (unlikely(steal > delta)) 904 steal = delta; 905 906 st = steal_ticks(steal); 907 steal = st * TICK_NSEC; 908 909 rq->prev_steal_time_rq += steal; 910 911 delta -= steal; 912 } 913#endif 914 915 rq->clock_task += delta; 916 917#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING) 918 if ((irq_delta + steal) && sched_feat(NONTASK_POWER)) 919 sched_rt_avg_update(rq, irq_delta + steal); 920#endif 921} 922 923#ifdef CONFIG_IRQ_TIME_ACCOUNTING 924static int irqtime_account_hi_update(void) 925{ 926 u64 *cpustat = kcpustat_this_cpu->cpustat; 927 unsigned long flags; 928 u64 latest_ns; 929 int ret = 0; 930 931 local_irq_save(flags); 932 latest_ns = this_cpu_read(cpu_hardirq_time); 933 if (nsecs_to_cputime64(latest_ns) > cpustat[CPUTIME_IRQ]) 934 ret = 1; 935 local_irq_restore(flags); 936 return ret; 937} 938 939static int irqtime_account_si_update(void) 940{ 941 u64 *cpustat = kcpustat_this_cpu->cpustat; 942 unsigned long flags; 943 u64 latest_ns; 944 int ret = 0; 945 946 local_irq_save(flags); 947 latest_ns = this_cpu_read(cpu_softirq_time); 948 if (nsecs_to_cputime64(latest_ns) > cpustat[CPUTIME_SOFTIRQ]) 949 ret = 1; 950 local_irq_restore(flags); 951 return ret; 952} 953 954#else /* CONFIG_IRQ_TIME_ACCOUNTING */ 955 956#define sched_clock_irqtime (0) 957 958#endif 959 960void sched_set_stop_task(int cpu, struct task_struct *stop) 961{ 962 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 }; 963 struct task_struct *old_stop = cpu_rq(cpu)->stop; 964 965 if (stop) { 966 /* 967 * Make it appear like a SCHED_FIFO task, its something 968 * userspace knows about and won't get confused about. 969 * 970 * Also, it will make PI more or less work without too 971 * much confusion -- but then, stop work should not 972 * rely on PI working anyway. 973 */ 974 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m); 975 976 stop->sched_class = &stop_sched_class; 977 } 978 979 cpu_rq(cpu)->stop = stop; 980 981 if (old_stop) { 982 /* 983 * Reset it back to a normal scheduling class so that 984 * it can die in pieces. 985 */ 986 old_stop->sched_class = &rt_sched_class; 987 } 988} 989 990/* 991 * __normal_prio - return the priority that is based on the static prio 992 */ 993static inline int __normal_prio(struct task_struct *p) 994{ 995 return p->static_prio; 996} 997 998/* 999 * Calculate the expected normal priority: i.e. priority 1000 * without taking RT-inheritance into account. Might be 1001 * boosted by interactivity modifiers. Changes upon fork, 1002 * setprio syscalls, and whenever the interactivity 1003 * estimator recalculates. 1004 */ 1005static inline int normal_prio(struct task_struct *p) 1006{ 1007 int prio; 1008 1009 if (task_has_rt_policy(p)) 1010 prio = MAX_RT_PRIO-1 - p->rt_priority; 1011 else 1012 prio = __normal_prio(p); 1013 return prio; 1014} 1015 1016/* 1017 * Calculate the current priority, i.e. the priority 1018 * taken into account by the scheduler. This value might 1019 * be boosted by RT tasks, or might be boosted by 1020 * interactivity modifiers. Will be RT if the task got 1021 * RT-boosted. If not then it returns p->normal_prio. 1022 */ 1023static int effective_prio(struct task_struct *p) 1024{ 1025 p->normal_prio = normal_prio(p); 1026 /* 1027 * If we are RT tasks or we were boosted to RT priority, 1028 * keep the priority unchanged. Otherwise, update priority 1029 * to the normal priority: 1030 */ 1031 if (!rt_prio(p->prio)) 1032 return p->normal_prio; 1033 return p->prio; 1034} 1035 1036/** 1037 * task_curr - is this task currently executing on a CPU? 1038 * @p: the task in question. 1039 */ 1040inline int task_curr(const struct task_struct *p) 1041{ 1042 return cpu_curr(task_cpu(p)) == p; 1043} 1044 1045static inline void check_class_changed(struct rq *rq, struct task_struct *p, 1046 const struct sched_class *prev_class, 1047 int oldprio) 1048{ 1049 if (prev_class != p->sched_class) { 1050 if (prev_class->switched_from) 1051 prev_class->switched_from(rq, p); 1052 p->sched_class->switched_to(rq, p); 1053 } else if (oldprio != p->prio) 1054 p->sched_class->prio_changed(rq, p, oldprio); 1055} 1056 1057void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags) 1058{ 1059 const struct sched_class *class; 1060 1061 if (p->sched_class == rq->curr->sched_class) { 1062 rq->curr->sched_class->check_preempt_curr(rq, p, flags); 1063 } else { 1064 for_each_class(class) { 1065 if (class == rq->curr->sched_class) 1066 break; 1067 if (class == p->sched_class) { 1068 resched_task(rq->curr); 1069 break; 1070 } 1071 } 1072 } 1073 1074 /* 1075 * A queue event has occurred, and we're going to schedule. In 1076 * this case, we can save a useless back to back clock update. 1077 */ 1078 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr)) 1079 rq->skip_clock_update = 1; 1080} 1081 1082#ifdef CONFIG_SMP 1083void set_task_cpu(struct task_struct *p, unsigned int new_cpu) 1084{ 1085#ifdef CONFIG_SCHED_DEBUG 1086 /* 1087 * We should never call set_task_cpu() on a blocked task, 1088 * ttwu() will sort out the placement. 1089 */ 1090 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING && 1091 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE)); 1092 1093#ifdef CONFIG_LOCKDEP 1094 /* 1095 * The caller should hold either p->pi_lock or rq->lock, when changing 1096 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks. 1097 * 1098 * sched_move_task() holds both and thus holding either pins the cgroup, 1099 * see set_task_rq(). 1100 * 1101 * Furthermore, all task_rq users should acquire both locks, see 1102 * task_rq_lock(). 1103 */ 1104 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) || 1105 lockdep_is_held(&task_rq(p)->lock))); 1106#endif 1107#endif 1108 1109 trace_sched_migrate_task(p, new_cpu); 1110 1111 if (task_cpu(p) != new_cpu) { 1112 p->se.nr_migrations++; 1113 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0); 1114 } 1115 1116 __set_task_cpu(p, new_cpu); 1117} 1118 1119struct migration_arg { 1120 struct task_struct *task; 1121 int dest_cpu; 1122}; 1123 1124static int migration_cpu_stop(void *data); 1125 1126/* 1127 * wait_task_inactive - wait for a thread to unschedule. 1128 * 1129 * If @match_state is nonzero, it's the @p->state value just checked and 1130 * not expected to change. If it changes, i.e. @p might have woken up, 1131 * then return zero. When we succeed in waiting for @p to be off its CPU, 1132 * we return a positive number (its total switch count). If a second call 1133 * a short while later returns the same number, the caller can be sure that 1134 * @p has remained unscheduled the whole time. 1135 * 1136 * The caller must ensure that the task *will* unschedule sometime soon, 1137 * else this function might spin for a *long* time. This function can't 1138 * be called with interrupts off, or it may introduce deadlock with 1139 * smp_call_function() if an IPI is sent by the same process we are 1140 * waiting to become inactive. 1141 */ 1142unsigned long wait_task_inactive(struct task_struct *p, long match_state) 1143{ 1144 unsigned long flags; 1145 int running, on_rq; 1146 unsigned long ncsw; 1147 struct rq *rq; 1148 1149 for (;;) { 1150 /* 1151 * We do the initial early heuristics without holding 1152 * any task-queue locks at all. We'll only try to get 1153 * the runqueue lock when things look like they will 1154 * work out! 1155 */ 1156 rq = task_rq(p); 1157 1158 /* 1159 * If the task is actively running on another CPU 1160 * still, just relax and busy-wait without holding 1161 * any locks. 1162 * 1163 * NOTE! Since we don't hold any locks, it's not 1164 * even sure that "rq" stays as the right runqueue! 1165 * But we don't care, since "task_running()" will 1166 * return false if the runqueue has changed and p 1167 * is actually now running somewhere else! 1168 */ 1169 while (task_running(rq, p)) { 1170 if (match_state && unlikely(p->state != match_state)) 1171 return 0; 1172 cpu_relax(); 1173 } 1174 1175 /* 1176 * Ok, time to look more closely! We need the rq 1177 * lock now, to be *sure*. If we're wrong, we'll 1178 * just go back and repeat. 1179 */ 1180 rq = task_rq_lock(p, &flags); 1181 trace_sched_wait_task(p); 1182 running = task_running(rq, p); 1183 on_rq = p->on_rq; 1184 ncsw = 0; 1185 if (!match_state || p->state == match_state) 1186 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */ 1187 task_rq_unlock(rq, p, &flags); 1188 1189 /* 1190 * If it changed from the expected state, bail out now. 1191 */ 1192 if (unlikely(!ncsw)) 1193 break; 1194 1195 /* 1196 * Was it really running after all now that we 1197 * checked with the proper locks actually held? 1198 * 1199 * Oops. Go back and try again.. 1200 */ 1201 if (unlikely(running)) { 1202 cpu_relax(); 1203 continue; 1204 } 1205 1206 /* 1207 * It's not enough that it's not actively running, 1208 * it must be off the runqueue _entirely_, and not 1209 * preempted! 1210 * 1211 * So if it was still runnable (but just not actively 1212 * running right now), it's preempted, and we should 1213 * yield - it could be a while. 1214 */ 1215 if (unlikely(on_rq)) { 1216 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ); 1217 1218 set_current_state(TASK_UNINTERRUPTIBLE); 1219 schedule_hrtimeout(&to, HRTIMER_MODE_REL); 1220 continue; 1221 } 1222 1223 /* 1224 * Ahh, all good. It wasn't running, and it wasn't 1225 * runnable, which means that it will never become 1226 * running in the future either. We're all done! 1227 */ 1228 break; 1229 } 1230 1231 return ncsw; 1232} 1233 1234/*** 1235 * kick_process - kick a running thread to enter/exit the kernel 1236 * @p: the to-be-kicked thread 1237 * 1238 * Cause a process which is running on another CPU to enter 1239 * kernel-mode, without any delay. (to get signals handled.) 1240 * 1241 * NOTE: this function doesn't have to take the runqueue lock, 1242 * because all it wants to ensure is that the remote task enters 1243 * the kernel. If the IPI races and the task has been migrated 1244 * to another CPU then no harm is done and the purpose has been 1245 * achieved as well. 1246 */ 1247void kick_process(struct task_struct *p) 1248{ 1249 int cpu; 1250 1251 preempt_disable(); 1252 cpu = task_cpu(p); 1253 if ((cpu != smp_processor_id()) && task_curr(p)) 1254 smp_send_reschedule(cpu); 1255 preempt_enable(); 1256} 1257EXPORT_SYMBOL_GPL(kick_process); 1258#endif /* CONFIG_SMP */ 1259 1260#ifdef CONFIG_SMP 1261/* 1262 * ->cpus_allowed is protected by both rq->lock and p->pi_lock 1263 */ 1264static int select_fallback_rq(int cpu, struct task_struct *p) 1265{ 1266 int dest_cpu; 1267 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu)); 1268 1269 /* Look for allowed, online CPU in same node. */ 1270 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask) 1271 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) 1272 return dest_cpu; 1273 1274 /* Any allowed, online CPU? */ 1275 dest_cpu = cpumask_any_and(tsk_cpus_allowed(p), cpu_active_mask); 1276 if (dest_cpu < nr_cpu_ids) 1277 return dest_cpu; 1278 1279 /* No more Mr. Nice Guy. */ 1280 dest_cpu = cpuset_cpus_allowed_fallback(p); 1281 /* 1282 * Don't tell them about moving exiting tasks or 1283 * kernel threads (both mm NULL), since they never 1284 * leave kernel. 1285 */ 1286 if (p->mm && printk_ratelimit()) { 1287 printk_sched("process %d (%s) no longer affine to cpu%d\n", 1288 task_pid_nr(p), p->comm, cpu); 1289 } 1290 1291 return dest_cpu; 1292} 1293 1294/* 1295 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable. 1296 */ 1297static inline 1298int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags) 1299{ 1300 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags); 1301 1302 /* 1303 * In order not to call set_task_cpu() on a blocking task we need 1304 * to rely on ttwu() to place the task on a valid ->cpus_allowed 1305 * cpu. 1306 * 1307 * Since this is common to all placement strategies, this lives here. 1308 * 1309 * [ this allows ->select_task() to simply return task_cpu(p) and 1310 * not worry about this generic constraint ] 1311 */ 1312 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) || 1313 !cpu_online(cpu))) 1314 cpu = select_fallback_rq(task_cpu(p), p); 1315 1316 return cpu; 1317} 1318 1319static void update_avg(u64 *avg, u64 sample) 1320{ 1321 s64 diff = sample - *avg; 1322 *avg += diff >> 3; 1323} 1324#endif 1325 1326static void 1327ttwu_stat(struct task_struct *p, int cpu, int wake_flags) 1328{ 1329#ifdef CONFIG_SCHEDSTATS 1330 struct rq *rq = this_rq(); 1331 1332#ifdef CONFIG_SMP 1333 int this_cpu = smp_processor_id(); 1334 1335 if (cpu == this_cpu) { 1336 schedstat_inc(rq, ttwu_local); 1337 schedstat_inc(p, se.statistics.nr_wakeups_local); 1338 } else { 1339 struct sched_domain *sd; 1340 1341 schedstat_inc(p, se.statistics.nr_wakeups_remote); 1342 rcu_read_lock(); 1343 for_each_domain(this_cpu, sd) { 1344 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) { 1345 schedstat_inc(sd, ttwu_wake_remote); 1346 break; 1347 } 1348 } 1349 rcu_read_unlock(); 1350 } 1351 1352 if (wake_flags & WF_MIGRATED) 1353 schedstat_inc(p, se.statistics.nr_wakeups_migrate); 1354 1355#endif /* CONFIG_SMP */ 1356 1357 schedstat_inc(rq, ttwu_count); 1358 schedstat_inc(p, se.statistics.nr_wakeups); 1359 1360 if (wake_flags & WF_SYNC) 1361 schedstat_inc(p, se.statistics.nr_wakeups_sync); 1362 1363#endif /* CONFIG_SCHEDSTATS */ 1364} 1365 1366static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags) 1367{ 1368 activate_task(rq, p, en_flags); 1369 p->on_rq = 1; 1370 1371 /* if a worker is waking up, notify workqueue */ 1372 if (p->flags & PF_WQ_WORKER) 1373 wq_worker_waking_up(p, cpu_of(rq)); 1374} 1375 1376/* 1377 * Mark the task runnable and perform wakeup-preemption. 1378 */ 1379static void 1380ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags) 1381{ 1382 trace_sched_wakeup(p, true); 1383 check_preempt_curr(rq, p, wake_flags); 1384 1385 p->state = TASK_RUNNING; 1386#ifdef CONFIG_SMP 1387 if (p->sched_class->task_woken) 1388 p->sched_class->task_woken(rq, p); 1389 1390 if (rq->idle_stamp) { 1391 u64 delta = rq->clock - rq->idle_stamp; 1392 u64 max = 2*sysctl_sched_migration_cost; 1393 1394 if (delta > max) 1395 rq->avg_idle = max; 1396 else 1397 update_avg(&rq->avg_idle, delta); 1398 rq->idle_stamp = 0; 1399 } 1400#endif 1401} 1402 1403static void 1404ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags) 1405{ 1406#ifdef CONFIG_SMP 1407 if (p->sched_contributes_to_load) 1408 rq->nr_uninterruptible--; 1409#endif 1410 1411 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING); 1412 ttwu_do_wakeup(rq, p, wake_flags); 1413} 1414 1415/* 1416 * Called in case the task @p isn't fully descheduled from its runqueue, 1417 * in this case we must do a remote wakeup. Its a 'light' wakeup though, 1418 * since all we need to do is flip p->state to TASK_RUNNING, since 1419 * the task is still ->on_rq. 1420 */ 1421static int ttwu_remote(struct task_struct *p, int wake_flags) 1422{ 1423 struct rq *rq; 1424 int ret = 0; 1425 1426 rq = __task_rq_lock(p); 1427 if (p->on_rq) { 1428 ttwu_do_wakeup(rq, p, wake_flags); 1429 ret = 1; 1430 } 1431 __task_rq_unlock(rq); 1432 1433 return ret; 1434} 1435 1436#ifdef CONFIG_SMP 1437static void sched_ttwu_pending(void) 1438{ 1439 struct rq *rq = this_rq(); 1440 struct llist_node *llist = llist_del_all(&rq->wake_list); 1441 struct task_struct *p; 1442 1443 raw_spin_lock(&rq->lock); 1444 1445 while (llist) { 1446 p = llist_entry(llist, struct task_struct, wake_entry); 1447 llist = llist_next(llist); 1448 ttwu_do_activate(rq, p, 0); 1449 } 1450 1451 raw_spin_unlock(&rq->lock); 1452} 1453 1454void scheduler_ipi(void) 1455{ 1456 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick()) 1457 return; 1458 1459 /* 1460 * Not all reschedule IPI handlers call irq_enter/irq_exit, since 1461 * traditionally all their work was done from the interrupt return 1462 * path. Now that we actually do some work, we need to make sure 1463 * we do call them. 1464 * 1465 * Some archs already do call them, luckily irq_enter/exit nest 1466 * properly. 1467 * 1468 * Arguably we should visit all archs and update all handlers, 1469 * however a fair share of IPIs are still resched only so this would 1470 * somewhat pessimize the simple resched case. 1471 */ 1472 irq_enter(); 1473 sched_ttwu_pending(); 1474 1475 /* 1476 * Check if someone kicked us for doing the nohz idle load balance. 1477 */ 1478 if (unlikely(got_nohz_idle_kick() && !need_resched())) { 1479 this_rq()->idle_balance = 1; 1480 raise_softirq_irqoff(SCHED_SOFTIRQ); 1481 } 1482 irq_exit(); 1483} 1484 1485static void ttwu_queue_remote(struct task_struct *p, int cpu) 1486{ 1487 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) 1488 smp_send_reschedule(cpu); 1489} 1490 1491#ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW 1492static int ttwu_activate_remote(struct task_struct *p, int wake_flags) 1493{ 1494 struct rq *rq; 1495 int ret = 0; 1496 1497 rq = __task_rq_lock(p); 1498 if (p->on_cpu) { 1499 ttwu_activate(rq, p, ENQUEUE_WAKEUP); 1500 ttwu_do_wakeup(rq, p, wake_flags); 1501 ret = 1; 1502 } 1503 __task_rq_unlock(rq); 1504 1505 return ret; 1506 1507} 1508#endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */ 1509 1510bool cpus_share_cache(int this_cpu, int that_cpu) 1511{ 1512 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu); 1513} 1514#endif /* CONFIG_SMP */ 1515 1516static void ttwu_queue(struct task_struct *p, int cpu) 1517{ 1518 struct rq *rq = cpu_rq(cpu); 1519 1520#if defined(CONFIG_SMP) 1521 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) { 1522 sched_clock_cpu(cpu); /* sync clocks x-cpu */ 1523 ttwu_queue_remote(p, cpu); 1524 return; 1525 } 1526#endif 1527 1528 raw_spin_lock(&rq->lock); 1529 ttwu_do_activate(rq, p, 0); 1530 raw_spin_unlock(&rq->lock); 1531} 1532 1533/** 1534 * try_to_wake_up - wake up a thread 1535 * @p: the thread to be awakened 1536 * @state: the mask of task states that can be woken 1537 * @wake_flags: wake modifier flags (WF_*) 1538 * 1539 * Put it on the run-queue if it's not already there. The "current" 1540 * thread is always on the run-queue (except when the actual 1541 * re-schedule is in progress), and as such you're allowed to do 1542 * the simpler "current->state = TASK_RUNNING" to mark yourself 1543 * runnable without the overhead of this. 1544 * 1545 * Returns %true if @p was woken up, %false if it was already running 1546 * or @state didn't match @p's state. 1547 */ 1548static int 1549try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags) 1550{ 1551 unsigned long flags; 1552 int cpu, success = 0; 1553 1554 smp_wmb(); 1555 raw_spin_lock_irqsave(&p->pi_lock, flags); 1556 if (!(p->state & state)) 1557 goto out; 1558 1559 success = 1; /* we're going to change ->state */ 1560 cpu = task_cpu(p); 1561 1562 if (p->on_rq && ttwu_remote(p, wake_flags)) 1563 goto stat; 1564 1565#ifdef CONFIG_SMP 1566 /* 1567 * If the owning (remote) cpu is still in the middle of schedule() with 1568 * this task as prev, wait until its done referencing the task. 1569 */ 1570 while (p->on_cpu) { 1571#ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW 1572 /* 1573 * In case the architecture enables interrupts in 1574 * context_switch(), we cannot busy wait, since that 1575 * would lead to deadlocks when an interrupt hits and 1576 * tries to wake up @prev. So bail and do a complete 1577 * remote wakeup. 1578 */ 1579 if (ttwu_activate_remote(p, wake_flags)) 1580 goto stat; 1581#else 1582 cpu_relax(); 1583#endif 1584 } 1585 /* 1586 * Pairs with the smp_wmb() in finish_lock_switch(). 1587 */ 1588 smp_rmb(); 1589 1590 p->sched_contributes_to_load = !!task_contributes_to_load(p); 1591 p->state = TASK_WAKING; 1592 1593 if (p->sched_class->task_waking) 1594 p->sched_class->task_waking(p); 1595 1596 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags); 1597 if (task_cpu(p) != cpu) { 1598 wake_flags |= WF_MIGRATED; 1599 set_task_cpu(p, cpu); 1600 } 1601#endif /* CONFIG_SMP */ 1602 1603 ttwu_queue(p, cpu); 1604stat: 1605 ttwu_stat(p, cpu, wake_flags); 1606out: 1607 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 1608 1609 return success; 1610} 1611 1612/** 1613 * try_to_wake_up_local - try to wake up a local task with rq lock held 1614 * @p: the thread to be awakened 1615 * 1616 * Put @p on the run-queue if it's not already there. The caller must 1617 * ensure that this_rq() is locked, @p is bound to this_rq() and not 1618 * the current task. 1619 */ 1620static void try_to_wake_up_local(struct task_struct *p) 1621{ 1622 struct rq *rq = task_rq(p); 1623 1624 BUG_ON(rq != this_rq()); 1625 BUG_ON(p == current); 1626 lockdep_assert_held(&rq->lock); 1627 1628 if (!raw_spin_trylock(&p->pi_lock)) { 1629 raw_spin_unlock(&rq->lock); 1630 raw_spin_lock(&p->pi_lock); 1631 raw_spin_lock(&rq->lock); 1632 } 1633 1634 if (!(p->state & TASK_NORMAL)) 1635 goto out; 1636 1637 if (!p->on_rq) 1638 ttwu_activate(rq, p, ENQUEUE_WAKEUP); 1639 1640 ttwu_do_wakeup(rq, p, 0); 1641 ttwu_stat(p, smp_processor_id(), 0); 1642out: 1643 raw_spin_unlock(&p->pi_lock); 1644} 1645 1646/** 1647 * wake_up_process - Wake up a specific process 1648 * @p: The process to be woken up. 1649 * 1650 * Attempt to wake up the nominated process and move it to the set of runnable 1651 * processes. Returns 1 if the process was woken up, 0 if it was already 1652 * running. 1653 * 1654 * It may be assumed that this function implies a write memory barrier before 1655 * changing the task state if and only if any tasks are woken up. 1656 */ 1657int wake_up_process(struct task_struct *p) 1658{ 1659 return try_to_wake_up(p, TASK_ALL, 0); 1660} 1661EXPORT_SYMBOL(wake_up_process); 1662 1663int wake_up_state(struct task_struct *p, unsigned int state) 1664{ 1665 return try_to_wake_up(p, state, 0); 1666} 1667 1668/* 1669 * Perform scheduler related setup for a newly forked process p. 1670 * p is forked by current. 1671 * 1672 * __sched_fork() is basic setup used by init_idle() too: 1673 */ 1674static void __sched_fork(struct task_struct *p) 1675{ 1676 p->on_rq = 0; 1677 1678 p->se.on_rq = 0; 1679 p->se.exec_start = 0; 1680 p->se.sum_exec_runtime = 0; 1681 p->se.prev_sum_exec_runtime = 0; 1682 p->se.nr_migrations = 0; 1683 p->se.vruntime = 0; 1684 INIT_LIST_HEAD(&p->se.group_node); 1685 1686#ifdef CONFIG_SCHEDSTATS 1687 memset(&p->se.statistics, 0, sizeof(p->se.statistics)); 1688#endif 1689 1690 INIT_LIST_HEAD(&p->rt.run_list); 1691 1692#ifdef CONFIG_PREEMPT_NOTIFIERS 1693 INIT_HLIST_HEAD(&p->preempt_notifiers); 1694#endif 1695} 1696 1697/* 1698 * fork()/clone()-time setup: 1699 */ 1700void sched_fork(struct task_struct *p) 1701{ 1702 unsigned long flags; 1703 int cpu = get_cpu(); 1704 1705 __sched_fork(p); 1706 /* 1707 * We mark the process as running here. This guarantees that 1708 * nobody will actually run it, and a signal or other external 1709 * event cannot wake it up and insert it on the runqueue either. 1710 */ 1711 p->state = TASK_RUNNING; 1712 1713 /* 1714 * Make sure we do not leak PI boosting priority to the child. 1715 */ 1716 p->prio = current->normal_prio; 1717 1718 /* 1719 * Revert to default priority/policy on fork if requested. 1720 */ 1721 if (unlikely(p->sched_reset_on_fork)) { 1722 if (task_has_rt_policy(p)) { 1723 p->policy = SCHED_NORMAL; 1724 p->static_prio = NICE_TO_PRIO(0); 1725 p->rt_priority = 0; 1726 } else if (PRIO_TO_NICE(p->static_prio) < 0) 1727 p->static_prio = NICE_TO_PRIO(0); 1728 1729 p->prio = p->normal_prio = __normal_prio(p); 1730 set_load_weight(p); 1731 1732 /* 1733 * We don't need the reset flag anymore after the fork. It has 1734 * fulfilled its duty: 1735 */ 1736 p->sched_reset_on_fork = 0; 1737 } 1738 1739 if (!rt_prio(p->prio)) 1740 p->sched_class = &fair_sched_class; 1741 1742 if (p->sched_class->task_fork) 1743 p->sched_class->task_fork(p); 1744 1745 /* 1746 * The child is not yet in the pid-hash so no cgroup attach races, 1747 * and the cgroup is pinned to this child due to cgroup_fork() 1748 * is ran before sched_fork(). 1749 * 1750 * Silence PROVE_RCU. 1751 */ 1752 raw_spin_lock_irqsave(&p->pi_lock, flags); 1753 set_task_cpu(p, cpu); 1754 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 1755 1756#if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT) 1757 if (likely(sched_info_on())) 1758 memset(&p->sched_info, 0, sizeof(p->sched_info)); 1759#endif 1760#if defined(CONFIG_SMP) 1761 p->on_cpu = 0; 1762#endif 1763#ifdef CONFIG_PREEMPT_COUNT 1764 /* Want to start with kernel preemption disabled. */ 1765 task_thread_info(p)->preempt_count = 1; 1766#endif 1767#ifdef CONFIG_SMP 1768 plist_node_init(&p->pushable_tasks, MAX_PRIO); 1769#endif 1770 1771 put_cpu(); 1772} 1773 1774/* 1775 * wake_up_new_task - wake up a newly created task for the first time. 1776 * 1777 * This function will do some initial scheduler statistics housekeeping 1778 * that must be done for every newly created context, then puts the task 1779 * on the runqueue and wakes it. 1780 */ 1781void wake_up_new_task(struct task_struct *p) 1782{ 1783 unsigned long flags; 1784 struct rq *rq; 1785 1786 raw_spin_lock_irqsave(&p->pi_lock, flags); 1787#ifdef CONFIG_SMP 1788 /* 1789 * Fork balancing, do it here and not earlier because: 1790 * - cpus_allowed can change in the fork path 1791 * - any previously selected cpu might disappear through hotplug 1792 */ 1793 set_task_cpu(p, select_task_rq(p, SD_BALANCE_FORK, 0)); 1794#endif 1795 1796 rq = __task_rq_lock(p); 1797 activate_task(rq, p, 0); 1798 p->on_rq = 1; 1799 trace_sched_wakeup_new(p, true); 1800 check_preempt_curr(rq, p, WF_FORK); 1801#ifdef CONFIG_SMP 1802 if (p->sched_class->task_woken) 1803 p->sched_class->task_woken(rq, p); 1804#endif 1805 task_rq_unlock(rq, p, &flags); 1806} 1807 1808#ifdef CONFIG_PREEMPT_NOTIFIERS 1809 1810/** 1811 * preempt_notifier_register - tell me when current is being preempted & rescheduled 1812 * @notifier: notifier struct to register 1813 */ 1814void preempt_notifier_register(struct preempt_notifier *notifier) 1815{ 1816 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers); 1817} 1818EXPORT_SYMBOL_GPL(preempt_notifier_register); 1819 1820/** 1821 * preempt_notifier_unregister - no longer interested in preemption notifications 1822 * @notifier: notifier struct to unregister 1823 * 1824 * This is safe to call from within a preemption notifier. 1825 */ 1826void preempt_notifier_unregister(struct preempt_notifier *notifier) 1827{ 1828 hlist_del(¬ifier->link); 1829} 1830EXPORT_SYMBOL_GPL(preempt_notifier_unregister); 1831 1832static void fire_sched_in_preempt_notifiers(struct task_struct *curr) 1833{ 1834 struct preempt_notifier *notifier; 1835 struct hlist_node *node; 1836 1837 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link) 1838 notifier->ops->sched_in(notifier, raw_smp_processor_id()); 1839} 1840 1841static void 1842fire_sched_out_preempt_notifiers(struct task_struct *curr, 1843 struct task_struct *next) 1844{ 1845 struct preempt_notifier *notifier; 1846 struct hlist_node *node; 1847 1848 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link) 1849 notifier->ops->sched_out(notifier, next); 1850} 1851 1852#else /* !CONFIG_PREEMPT_NOTIFIERS */ 1853 1854static void fire_sched_in_preempt_notifiers(struct task_struct *curr) 1855{ 1856} 1857 1858static void 1859fire_sched_out_preempt_notifiers(struct task_struct *curr, 1860 struct task_struct *next) 1861{ 1862} 1863 1864#endif /* CONFIG_PREEMPT_NOTIFIERS */ 1865 1866/** 1867 * prepare_task_switch - prepare to switch tasks 1868 * @rq: the runqueue preparing to switch 1869 * @prev: the current task that is being switched out 1870 * @next: the task we are going to switch to. 1871 * 1872 * This is called with the rq lock held and interrupts off. It must 1873 * be paired with a subsequent finish_task_switch after the context 1874 * switch. 1875 * 1876 * prepare_task_switch sets up locking and calls architecture specific 1877 * hooks. 1878 */ 1879static inline void 1880prepare_task_switch(struct rq *rq, struct task_struct *prev, 1881 struct task_struct *next) 1882{ 1883 sched_info_switch(prev, next); 1884 perf_event_task_sched_out(prev, next); 1885 fire_sched_out_preempt_notifiers(prev, next); 1886 prepare_lock_switch(rq, next); 1887 prepare_arch_switch(next); 1888 trace_sched_switch(prev, next); 1889} 1890 1891/** 1892 * finish_task_switch - clean up after a task-switch 1893 * @rq: runqueue associated with task-switch 1894 * @prev: the thread we just switched away from. 1895 * 1896 * finish_task_switch must be called after the context switch, paired 1897 * with a prepare_task_switch call before the context switch. 1898 * finish_task_switch will reconcile locking set up by prepare_task_switch, 1899 * and do any other architecture-specific cleanup actions. 1900 * 1901 * Note that we may have delayed dropping an mm in context_switch(). If 1902 * so, we finish that here outside of the runqueue lock. (Doing it 1903 * with the lock held can cause deadlocks; see schedule() for 1904 * details.) 1905 */ 1906static void finish_task_switch(struct rq *rq, struct task_struct *prev) 1907 __releases(rq->lock) 1908{ 1909 struct mm_struct *mm = rq->prev_mm; 1910 long prev_state; 1911 1912 rq->prev_mm = NULL; 1913 1914 /* 1915 * A task struct has one reference for the use as "current". 1916 * If a task dies, then it sets TASK_DEAD in tsk->state and calls 1917 * schedule one last time. The schedule call will never return, and 1918 * the scheduled task must drop that reference. 1919 * The test for TASK_DEAD must occur while the runqueue locks are 1920 * still held, otherwise prev could be scheduled on another cpu, die 1921 * there before we look at prev->state, and then the reference would 1922 * be dropped twice. 1923 * Manfred Spraul <manfred@colorfullife.com> 1924 */ 1925 prev_state = prev->state; 1926 finish_arch_switch(prev); 1927#ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW 1928 local_irq_disable(); 1929#endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */ 1930 perf_event_task_sched_in(prev, current); 1931#ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW 1932 local_irq_enable(); 1933#endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */ 1934 finish_lock_switch(rq, prev); 1935 1936 fire_sched_in_preempt_notifiers(current); 1937 if (mm) 1938 mmdrop(mm); 1939 if (unlikely(prev_state == TASK_DEAD)) { 1940 /* 1941 * Remove function-return probe instances associated with this 1942 * task and put them back on the free list. 1943 */ 1944 kprobe_flush_task(prev); 1945 put_task_struct(prev); 1946 } 1947} 1948 1949#ifdef CONFIG_SMP 1950 1951/* assumes rq->lock is held */ 1952static inline void pre_schedule(struct rq *rq, struct task_struct *prev) 1953{ 1954 if (prev->sched_class->pre_schedule) 1955 prev->sched_class->pre_schedule(rq, prev); 1956} 1957 1958/* rq->lock is NOT held, but preemption is disabled */ 1959static inline void post_schedule(struct rq *rq) 1960{ 1961 if (rq->post_schedule) { 1962 unsigned long flags; 1963 1964 raw_spin_lock_irqsave(&rq->lock, flags); 1965 if (rq->curr->sched_class->post_schedule) 1966 rq->curr->sched_class->post_schedule(rq); 1967 raw_spin_unlock_irqrestore(&rq->lock, flags); 1968 1969 rq->post_schedule = 0; 1970 } 1971} 1972 1973#else 1974 1975static inline void pre_schedule(struct rq *rq, struct task_struct *p) 1976{ 1977} 1978 1979static inline void post_schedule(struct rq *rq) 1980{ 1981} 1982 1983#endif 1984 1985/** 1986 * schedule_tail - first thing a freshly forked thread must call. 1987 * @prev: the thread we just switched away from. 1988 */ 1989asmlinkage void schedule_tail(struct task_struct *prev) 1990 __releases(rq->lock) 1991{ 1992 struct rq *rq = this_rq(); 1993 1994 finish_task_switch(rq, prev); 1995 1996 /* 1997 * FIXME: do we need to worry about rq being invalidated by the 1998 * task_switch? 1999 */ 2000 post_schedule(rq); 2001 2002#ifdef __ARCH_WANT_UNLOCKED_CTXSW 2003 /* In this case, finish_task_switch does not reenable preemption */ 2004 preempt_enable(); 2005#endif 2006 if (current->set_child_tid) 2007 put_user(task_pid_vnr(current), current->set_child_tid); 2008} 2009 2010/* 2011 * context_switch - switch to the new MM and the new 2012 * thread's register state. 2013 */ 2014static inline void 2015context_switch(struct rq *rq, struct task_struct *prev, 2016 struct task_struct *next) 2017{ 2018 struct mm_struct *mm, *oldmm; 2019 2020 prepare_task_switch(rq, prev, next); 2021 2022 mm = next->mm; 2023 oldmm = prev->active_mm; 2024 /* 2025 * For paravirt, this is coupled with an exit in switch_to to 2026 * combine the page table reload and the switch backend into 2027 * one hypercall. 2028 */ 2029 arch_start_context_switch(prev); 2030 2031 if (!mm) { 2032 next->active_mm = oldmm; 2033 atomic_inc(&oldmm->mm_count); 2034 enter_lazy_tlb(oldmm, next); 2035 } else 2036 switch_mm(oldmm, mm, next); 2037 2038 if (!prev->mm) { 2039 prev->active_mm = NULL; 2040 rq->prev_mm = oldmm; 2041 } 2042 /* 2043 * Since the runqueue lock will be released by the next 2044 * task (which is an invalid locking op but in the case 2045 * of the scheduler it's an obvious special-case), so we 2046 * do an early lockdep release here: 2047 */ 2048#ifndef __ARCH_WANT_UNLOCKED_CTXSW 2049 spin_release(&rq->lock.dep_map, 1, _THIS_IP_); 2050#endif 2051 2052 /* Here we just switch the register state and the stack. */ 2053 switch_to(prev, next, prev); 2054 2055 barrier(); 2056 /* 2057 * this_rq must be evaluated again because prev may have moved 2058 * CPUs since it called schedule(), thus the 'rq' on its stack 2059 * frame will be invalid. 2060 */ 2061 finish_task_switch(this_rq(), prev); 2062} 2063 2064/* 2065 * nr_running, nr_uninterruptible and nr_context_switches: 2066 * 2067 * externally visible scheduler statistics: current number of runnable 2068 * threads, current number of uninterruptible-sleeping threads, total 2069 * number of context switches performed since bootup. 2070 */ 2071unsigned long nr_running(void) 2072{ 2073 unsigned long i, sum = 0; 2074 2075 for_each_online_cpu(i) 2076 sum += cpu_rq(i)->nr_running; 2077 2078 return sum; 2079} 2080 2081unsigned long nr_uninterruptible(void) 2082{ 2083 unsigned long i, sum = 0; 2084 2085 for_each_possible_cpu(i) 2086 sum += cpu_rq(i)->nr_uninterruptible; 2087 2088 /* 2089 * Since we read the counters lockless, it might be slightly 2090 * inaccurate. Do not allow it to go below zero though: 2091 */ 2092 if (unlikely((long)sum < 0)) 2093 sum = 0; 2094 2095 return sum; 2096} 2097 2098unsigned long long nr_context_switches(void) 2099{ 2100 int i; 2101 unsigned long long sum = 0; 2102 2103 for_each_possible_cpu(i) 2104 sum += cpu_rq(i)->nr_switches; 2105 2106 return sum; 2107} 2108 2109unsigned long nr_iowait(void) 2110{ 2111 unsigned long i, sum = 0; 2112 2113 for_each_possible_cpu(i) 2114 sum += atomic_read(&cpu_rq(i)->nr_iowait); 2115 2116 return sum; 2117} 2118 2119unsigned long nr_iowait_cpu(int cpu) 2120{ 2121 struct rq *this = cpu_rq(cpu); 2122 return atomic_read(&this->nr_iowait); 2123} 2124 2125unsigned long this_cpu_load(void) 2126{ 2127 struct rq *this = this_rq(); 2128 return this->cpu_load[0]; 2129} 2130 2131 2132/* Variables and functions for calc_load */ 2133static atomic_long_t calc_load_tasks; 2134static unsigned long calc_load_update; 2135unsigned long avenrun[3]; 2136EXPORT_SYMBOL(avenrun); 2137 2138static long calc_load_fold_active(struct rq *this_rq) 2139{ 2140 long nr_active, delta = 0; 2141 2142 nr_active = this_rq->nr_running; 2143 nr_active += (long) this_rq->nr_uninterruptible; 2144 2145 if (nr_active != this_rq->calc_load_active) { 2146 delta = nr_active - this_rq->calc_load_active; 2147 this_rq->calc_load_active = nr_active; 2148 } 2149 2150 return delta; 2151} 2152 2153static unsigned long 2154calc_load(unsigned long load, unsigned long exp, unsigned long active) 2155{ 2156 load *= exp; 2157 load += active * (FIXED_1 - exp); 2158 load += 1UL << (FSHIFT - 1); 2159 return load >> FSHIFT; 2160} 2161 2162#ifdef CONFIG_NO_HZ 2163/* 2164 * For NO_HZ we delay the active fold to the next LOAD_FREQ update. 2165 * 2166 * When making the ILB scale, we should try to pull this in as well. 2167 */ 2168static atomic_long_t calc_load_tasks_idle; 2169 2170void calc_load_account_idle(struct rq *this_rq) 2171{ 2172 long delta; 2173 2174 delta = calc_load_fold_active(this_rq); 2175 if (delta) 2176 atomic_long_add(delta, &calc_load_tasks_idle); 2177} 2178 2179static long calc_load_fold_idle(void) 2180{ 2181 long delta = 0; 2182 2183 /* 2184 * Its got a race, we don't care... 2185 */ 2186 if (atomic_long_read(&calc_load_tasks_idle)) 2187 delta = atomic_long_xchg(&calc_load_tasks_idle, 0); 2188 2189 return delta; 2190} 2191 2192/** 2193 * fixed_power_int - compute: x^n, in O(log n) time 2194 * 2195 * @x: base of the power 2196 * @frac_bits: fractional bits of @x 2197 * @n: power to raise @x to. 2198 * 2199 * By exploiting the relation between the definition of the natural power 2200 * function: x^n := x*x*...*x (x multiplied by itself for n times), and 2201 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i, 2202 * (where: n_i \elem {0, 1}, the binary vector representing n), 2203 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is 2204 * of course trivially computable in O(log_2 n), the length of our binary 2205 * vector. 2206 */ 2207static unsigned long 2208fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n) 2209{ 2210 unsigned long result = 1UL << frac_bits; 2211 2212 if (n) for (;;) { 2213 if (n & 1) { 2214 result *= x; 2215 result += 1UL << (frac_bits - 1); 2216 result >>= frac_bits; 2217 } 2218 n >>= 1; 2219 if (!n) 2220 break; 2221 x *= x; 2222 x += 1UL << (frac_bits - 1); 2223 x >>= frac_bits; 2224 } 2225 2226 return result; 2227} 2228 2229/* 2230 * a1 = a0 * e + a * (1 - e) 2231 * 2232 * a2 = a1 * e + a * (1 - e) 2233 * = (a0 * e + a * (1 - e)) * e + a * (1 - e) 2234 * = a0 * e^2 + a * (1 - e) * (1 + e) 2235 * 2236 * a3 = a2 * e + a * (1 - e) 2237 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e) 2238 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2) 2239 * 2240 * ... 2241 * 2242 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1] 2243 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e) 2244 * = a0 * e^n + a * (1 - e^n) 2245 * 2246 * [1] application of the geometric series: 2247 * 2248 * n 1 - x^(n+1) 2249 * S_n := \Sum x^i = ------------- 2250 * i=0 1 - x 2251 */ 2252static unsigned long 2253calc_load_n(unsigned long load, unsigned long exp, 2254 unsigned long active, unsigned int n) 2255{ 2256 2257 return calc_load(load, fixed_power_int(exp, FSHIFT, n), active); 2258} 2259 2260/* 2261 * NO_HZ can leave us missing all per-cpu ticks calling 2262 * calc_load_account_active(), but since an idle CPU folds its delta into 2263 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold 2264 * in the pending idle delta if our idle period crossed a load cycle boundary. 2265 * 2266 * Once we've updated the global active value, we need to apply the exponential 2267 * weights adjusted to the number of cycles missed. 2268 */ 2269static void calc_global_nohz(void) 2270{ 2271 long delta, active, n; 2272 2273 /* 2274 * If we crossed a calc_load_update boundary, make sure to fold 2275 * any pending idle changes, the respective CPUs might have 2276 * missed the tick driven calc_load_account_active() update 2277 * due to NO_HZ. 2278 */ 2279 delta = calc_load_fold_idle(); 2280 if (delta) 2281 atomic_long_add(delta, &calc_load_tasks); 2282 2283 /* 2284 * It could be the one fold was all it took, we done! 2285 */ 2286 if (time_before(jiffies, calc_load_update + 10)) 2287 return; 2288 2289 /* 2290 * Catch-up, fold however many we are behind still 2291 */ 2292 delta = jiffies - calc_load_update - 10; 2293 n = 1 + (delta / LOAD_FREQ); 2294 2295 active = atomic_long_read(&calc_load_tasks); 2296 active = active > 0 ? active * FIXED_1 : 0; 2297 2298 avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n); 2299 avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n); 2300 avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n); 2301 2302 calc_load_update += n * LOAD_FREQ; 2303} 2304#else 2305void calc_load_account_idle(struct rq *this_rq) 2306{ 2307} 2308 2309static inline long calc_load_fold_idle(void) 2310{ 2311 return 0; 2312} 2313 2314static void calc_global_nohz(void) 2315{ 2316} 2317#endif 2318 2319/** 2320 * get_avenrun - get the load average array 2321 * @loads: pointer to dest load array 2322 * @offset: offset to add 2323 * @shift: shift count to shift the result left 2324 * 2325 * These values are estimates at best, so no need for locking. 2326 */ 2327void get_avenrun(unsigned long *loads, unsigned long offset, int shift) 2328{ 2329 loads[0] = (avenrun[0] + offset) << shift; 2330 loads[1] = (avenrun[1] + offset) << shift; 2331 loads[2] = (avenrun[2] + offset) << shift; 2332} 2333 2334/* 2335 * calc_load - update the avenrun load estimates 10 ticks after the 2336 * CPUs have updated calc_load_tasks. 2337 */ 2338void calc_global_load(unsigned long ticks) 2339{ 2340 long active; 2341 2342 if (time_before(jiffies, calc_load_update + 10)) 2343 return; 2344 2345 active = atomic_long_read(&calc_load_tasks); 2346 active = active > 0 ? active * FIXED_1 : 0; 2347 2348 avenrun[0] = calc_load(avenrun[0], EXP_1, active); 2349 avenrun[1] = calc_load(avenrun[1], EXP_5, active); 2350 avenrun[2] = calc_load(avenrun[2], EXP_15, active); 2351 2352 calc_load_update += LOAD_FREQ; 2353 2354 /* 2355 * Account one period with whatever state we found before 2356 * folding in the nohz state and ageing the entire idle period. 2357 * 2358 * This avoids loosing a sample when we go idle between 2359 * calc_load_account_active() (10 ticks ago) and now and thus 2360 * under-accounting. 2361 */ 2362 calc_global_nohz(); 2363} 2364 2365/* 2366 * Called from update_cpu_load() to periodically update this CPU's 2367 * active count. 2368 */ 2369static void calc_load_account_active(struct rq *this_rq) 2370{ 2371 long delta; 2372 2373 if (time_before(jiffies, this_rq->calc_load_update)) 2374 return; 2375 2376 delta = calc_load_fold_active(this_rq); 2377 delta += calc_load_fold_idle(); 2378 if (delta) 2379 atomic_long_add(delta, &calc_load_tasks); 2380 2381 this_rq->calc_load_update += LOAD_FREQ; 2382} 2383 2384/* 2385 * The exact cpuload at various idx values, calculated at every tick would be 2386 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load 2387 * 2388 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called 2389 * on nth tick when cpu may be busy, then we have: 2390 * load = ((2^idx - 1) / 2^idx)^(n-1) * load 2391 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load 2392 * 2393 * decay_load_missed() below does efficient calculation of 2394 * load = ((2^idx - 1) / 2^idx)^(n-1) * load 2395 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load 2396 * 2397 * The calculation is approximated on a 128 point scale. 2398 * degrade_zero_ticks is the number of ticks after which load at any 2399 * particular idx is approximated to be zero. 2400 * degrade_factor is a precomputed table, a row for each load idx. 2401 * Each column corresponds to degradation factor for a power of two ticks, 2402 * based on 128 point scale. 2403 * Example: 2404 * row 2, col 3 (=12) says that the degradation at load idx 2 after 2405 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8). 2406 * 2407 * With this power of 2 load factors, we can degrade the load n times 2408 * by looking at 1 bits in n and doing as many mult/shift instead of 2409 * n mult/shifts needed by the exact degradation. 2410 */ 2411#define DEGRADE_SHIFT 7 2412static const unsigned char 2413 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128}; 2414static const unsigned char 2415 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = { 2416 {0, 0, 0, 0, 0, 0, 0, 0}, 2417 {64, 32, 8, 0, 0, 0, 0, 0}, 2418 {96, 72, 40, 12, 1, 0, 0}, 2419 {112, 98, 75, 43, 15, 1, 0}, 2420 {120, 112, 98, 76, 45, 16, 2} }; 2421 2422/* 2423 * Update cpu_load for any missed ticks, due to tickless idle. The backlog 2424 * would be when CPU is idle and so we just decay the old load without 2425 * adding any new load. 2426 */ 2427static unsigned long 2428decay_load_missed(unsigned long load, unsigned long missed_updates, int idx) 2429{ 2430 int j = 0; 2431 2432 if (!missed_updates) 2433 return load; 2434 2435 if (missed_updates >= degrade_zero_ticks[idx]) 2436 return 0; 2437 2438 if (idx == 1) 2439 return load >> missed_updates; 2440 2441 while (missed_updates) { 2442 if (missed_updates % 2) 2443 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT; 2444 2445 missed_updates >>= 1; 2446 j++; 2447 } 2448 return load; 2449} 2450 2451/* 2452 * Update rq->cpu_load[] statistics. This function is usually called every 2453 * scheduler tick (TICK_NSEC). With tickless idle this will not be called 2454 * every tick. We fix it up based on jiffies. 2455 */ 2456void update_cpu_load(struct rq *this_rq) 2457{ 2458 unsigned long this_load = this_rq->load.weight; 2459 unsigned long curr_jiffies = jiffies; 2460 unsigned long pending_updates; 2461 int i, scale; 2462 2463 this_rq->nr_load_updates++; 2464 2465 /* Avoid repeated calls on same jiffy, when moving in and out of idle */ 2466 if (curr_jiffies == this_rq->last_load_update_tick) 2467 return; 2468 2469 pending_updates = curr_jiffies - this_rq->last_load_update_tick; 2470 this_rq->last_load_update_tick = curr_jiffies; 2471 2472 /* Update our load: */ 2473 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */ 2474 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) { 2475 unsigned long old_load, new_load; 2476 2477 /* scale is effectively 1 << i now, and >> i divides by scale */ 2478 2479 old_load = this_rq->cpu_load[i]; 2480 old_load = decay_load_missed(old_load, pending_updates - 1, i); 2481 new_load = this_load; 2482 /* 2483 * Round up the averaging division if load is increasing. This 2484 * prevents us from getting stuck on 9 if the load is 10, for 2485 * example. 2486 */ 2487 if (new_load > old_load) 2488 new_load += scale - 1; 2489 2490 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i; 2491 } 2492 2493 sched_avg_update(this_rq); 2494} 2495 2496static void update_cpu_load_active(struct rq *this_rq) 2497{ 2498 update_cpu_load(this_rq); 2499 2500 calc_load_account_active(this_rq); 2501} 2502 2503#ifdef CONFIG_SMP 2504 2505/* 2506 * sched_exec - execve() is a valuable balancing opportunity, because at 2507 * this point the task has the smallest effective memory and cache footprint. 2508 */ 2509void sched_exec(void) 2510{ 2511 struct task_struct *p = current; 2512 unsigned long flags; 2513 int dest_cpu; 2514 2515 raw_spin_lock_irqsave(&p->pi_lock, flags); 2516 dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0); 2517 if (dest_cpu == smp_processor_id()) 2518 goto unlock; 2519 2520 if (likely(cpu_active(dest_cpu))) { 2521 struct migration_arg arg = { p, dest_cpu }; 2522 2523 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 2524 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg); 2525 return; 2526 } 2527unlock: 2528 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 2529} 2530 2531#endif 2532 2533DEFINE_PER_CPU(struct kernel_stat, kstat); 2534DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat); 2535 2536EXPORT_PER_CPU_SYMBOL(kstat); 2537EXPORT_PER_CPU_SYMBOL(kernel_cpustat); 2538 2539/* 2540 * Return any ns on the sched_clock that have not yet been accounted in 2541 * @p in case that task is currently running. 2542 * 2543 * Called with task_rq_lock() held on @rq. 2544 */ 2545static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq) 2546{ 2547 u64 ns = 0; 2548 2549 if (task_current(rq, p)) { 2550 update_rq_clock(rq); 2551 ns = rq->clock_task - p->se.exec_start; 2552 if ((s64)ns < 0) 2553 ns = 0; 2554 } 2555 2556 return ns; 2557} 2558 2559unsigned long long task_delta_exec(struct task_struct *p) 2560{ 2561 unsigned long flags; 2562 struct rq *rq; 2563 u64 ns = 0; 2564 2565 rq = task_rq_lock(p, &flags); 2566 ns = do_task_delta_exec(p, rq); 2567 task_rq_unlock(rq, p, &flags); 2568 2569 return ns; 2570} 2571 2572/* 2573 * Return accounted runtime for the task. 2574 * In case the task is currently running, return the runtime plus current's 2575 * pending runtime that have not been accounted yet. 2576 */ 2577unsigned long long task_sched_runtime(struct task_struct *p) 2578{ 2579 unsigned long flags; 2580 struct rq *rq; 2581 u64 ns = 0; 2582 2583 rq = task_rq_lock(p, &flags); 2584 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq); 2585 task_rq_unlock(rq, p, &flags); 2586 2587 return ns; 2588} 2589 2590#ifdef CONFIG_CGROUP_CPUACCT 2591struct cgroup_subsys cpuacct_subsys; 2592struct cpuacct root_cpuacct; 2593#endif 2594 2595static inline void task_group_account_field(struct task_struct *p, int index, 2596 u64 tmp) 2597{ 2598#ifdef CONFIG_CGROUP_CPUACCT 2599 struct kernel_cpustat *kcpustat; 2600 struct cpuacct *ca; 2601#endif 2602 /* 2603 * Since all updates are sure to touch the root cgroup, we 2604 * get ourselves ahead and touch it first. If the root cgroup 2605 * is the only cgroup, then nothing else should be necessary. 2606 * 2607 */ 2608 __get_cpu_var(kernel_cpustat).cpustat[index] += tmp; 2609 2610#ifdef CONFIG_CGROUP_CPUACCT 2611 if (unlikely(!cpuacct_subsys.active)) 2612 return; 2613 2614 rcu_read_lock(); 2615 ca = task_ca(p); 2616 while (ca && (ca != &root_cpuacct)) { 2617 kcpustat = this_cpu_ptr(ca->cpustat); 2618 kcpustat->cpustat[index] += tmp; 2619 ca = parent_ca(ca); 2620 } 2621 rcu_read_unlock(); 2622#endif 2623} 2624 2625 2626/* 2627 * Account user cpu time to a process. 2628 * @p: the process that the cpu time gets accounted to 2629 * @cputime: the cpu time spent in user space since the last update 2630 * @cputime_scaled: cputime scaled by cpu frequency 2631 */ 2632void account_user_time(struct task_struct *p, cputime_t cputime, 2633 cputime_t cputime_scaled) 2634{ 2635 int index; 2636 2637 /* Add user time to process. */ 2638 p->utime += cputime; 2639 p->utimescaled += cputime_scaled; 2640 account_group_user_time(p, cputime); 2641 2642 index = (TASK_NICE(p) > 0) ? CPUTIME_NICE : CPUTIME_USER; 2643 2644 /* Add user time to cpustat. */ 2645 task_group_account_field(p, index, (__force u64) cputime); 2646 2647 /* Account for user time used */ 2648 acct_update_integrals(p); 2649} 2650 2651/* 2652 * Account guest cpu time to a process. 2653 * @p: the process that the cpu time gets accounted to 2654 * @cputime: the cpu time spent in virtual machine since the last update 2655 * @cputime_scaled: cputime scaled by cpu frequency 2656 */ 2657static void account_guest_time(struct task_struct *p, cputime_t cputime, 2658 cputime_t cputime_scaled) 2659{ 2660 u64 *cpustat = kcpustat_this_cpu->cpustat; 2661 2662 /* Add guest time to process. */ 2663 p->utime += cputime; 2664 p->utimescaled += cputime_scaled; 2665 account_group_user_time(p, cputime); 2666 p->gtime += cputime; 2667 2668 /* Add guest time to cpustat. */ 2669 if (TASK_NICE(p) > 0) { 2670 cpustat[CPUTIME_NICE] += (__force u64) cputime; 2671 cpustat[CPUTIME_GUEST_NICE] += (__force u64) cputime; 2672 } else { 2673 cpustat[CPUTIME_USER] += (__force u64) cputime; 2674 cpustat[CPUTIME_GUEST] += (__force u64) cputime; 2675 } 2676} 2677 2678/* 2679 * Account system cpu time to a process and desired cpustat field 2680 * @p: the process that the cpu time gets accounted to 2681 * @cputime: the cpu time spent in kernel space since the last update 2682 * @cputime_scaled: cputime scaled by cpu frequency 2683 * @target_cputime64: pointer to cpustat field that has to be updated 2684 */ 2685static inline 2686void __account_system_time(struct task_struct *p, cputime_t cputime, 2687 cputime_t cputime_scaled, int index) 2688{ 2689 /* Add system time to process. */ 2690 p->stime += cputime; 2691 p->stimescaled += cputime_scaled; 2692 account_group_system_time(p, cputime); 2693 2694 /* Add system time to cpustat. */ 2695 task_group_account_field(p, index, (__force u64) cputime); 2696 2697 /* Account for system time used */ 2698 acct_update_integrals(p); 2699} 2700 2701/* 2702 * Account system cpu time to a process. 2703 * @p: the process that the cpu time gets accounted to 2704 * @hardirq_offset: the offset to subtract from hardirq_count() 2705 * @cputime: the cpu time spent in kernel space since the last update 2706 * @cputime_scaled: cputime scaled by cpu frequency 2707 */ 2708void account_system_time(struct task_struct *p, int hardirq_offset, 2709 cputime_t cputime, cputime_t cputime_scaled) 2710{ 2711 int index; 2712 2713 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) { 2714 account_guest_time(p, cputime, cputime_scaled); 2715 return; 2716 } 2717 2718 if (hardirq_count() - hardirq_offset) 2719 index = CPUTIME_IRQ; 2720 else if (in_serving_softirq()) 2721 index = CPUTIME_SOFTIRQ; 2722 else 2723 index = CPUTIME_SYSTEM; 2724 2725 __account_system_time(p, cputime, cputime_scaled, index); 2726} 2727 2728/* 2729 * Account for involuntary wait time. 2730 * @cputime: the cpu time spent in involuntary wait 2731 */ 2732void account_steal_time(cputime_t cputime) 2733{ 2734 u64 *cpustat = kcpustat_this_cpu->cpustat; 2735 2736 cpustat[CPUTIME_STEAL] += (__force u64) cputime; 2737} 2738 2739/* 2740 * Account for idle time. 2741 * @cputime: the cpu time spent in idle wait 2742 */ 2743void account_idle_time(cputime_t cputime) 2744{ 2745 u64 *cpustat = kcpustat_this_cpu->cpustat; 2746 struct rq *rq = this_rq(); 2747 2748 if (atomic_read(&rq->nr_iowait) > 0) 2749 cpustat[CPUTIME_IOWAIT] += (__force u64) cputime; 2750 else 2751 cpustat[CPUTIME_IDLE] += (__force u64) cputime; 2752} 2753 2754static __always_inline bool steal_account_process_tick(void) 2755{ 2756#ifdef CONFIG_PARAVIRT 2757 if (static_branch(¶virt_steal_enabled)) { 2758 u64 steal, st = 0; 2759 2760 steal = paravirt_steal_clock(smp_processor_id()); 2761 steal -= this_rq()->prev_steal_time; 2762 2763 st = steal_ticks(steal); 2764 this_rq()->prev_steal_time += st * TICK_NSEC; 2765 2766 account_steal_time(st); 2767 return st; 2768 } 2769#endif 2770 return false; 2771} 2772 2773#ifndef CONFIG_VIRT_CPU_ACCOUNTING 2774 2775#ifdef CONFIG_IRQ_TIME_ACCOUNTING 2776/* 2777 * Account a tick to a process and cpustat 2778 * @p: the process that the cpu time gets accounted to 2779 * @user_tick: is the tick from userspace 2780 * @rq: the pointer to rq 2781 * 2782 * Tick demultiplexing follows the order 2783 * - pending hardirq update 2784 * - pending softirq update 2785 * - user_time 2786 * - idle_time 2787 * - system time 2788 * - check for guest_time 2789 * - else account as system_time 2790 * 2791 * Check for hardirq is done both for system and user time as there is 2792 * no timer going off while we are on hardirq and hence we may never get an 2793 * opportunity to update it solely in system time. 2794 * p->stime and friends are only updated on system time and not on irq 2795 * softirq as those do not count in task exec_runtime any more. 2796 */ 2797static void irqtime_account_process_tick(struct task_struct *p, int user_tick, 2798 struct rq *rq) 2799{ 2800 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy); 2801 u64 *cpustat = kcpustat_this_cpu->cpustat; 2802 2803 if (steal_account_process_tick()) 2804 return; 2805 2806 if (irqtime_account_hi_update()) { 2807 cpustat[CPUTIME_IRQ] += (__force u64) cputime_one_jiffy; 2808 } else if (irqtime_account_si_update()) { 2809 cpustat[CPUTIME_SOFTIRQ] += (__force u64) cputime_one_jiffy; 2810 } else if (this_cpu_ksoftirqd() == p) { 2811 /* 2812 * ksoftirqd time do not get accounted in cpu_softirq_time. 2813 * So, we have to handle it separately here. 2814 * Also, p->stime needs to be updated for ksoftirqd. 2815 */ 2816 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled, 2817 CPUTIME_SOFTIRQ); 2818 } else if (user_tick) { 2819 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled); 2820 } else if (p == rq->idle) { 2821 account_idle_time(cputime_one_jiffy); 2822 } else if (p->flags & PF_VCPU) { /* System time or guest time */ 2823 account_guest_time(p, cputime_one_jiffy, one_jiffy_scaled); 2824 } else { 2825 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled, 2826 CPUTIME_SYSTEM); 2827 } 2828} 2829 2830static void irqtime_account_idle_ticks(int ticks) 2831{ 2832 int i; 2833 struct rq *rq = this_rq(); 2834 2835 for (i = 0; i < ticks; i++) 2836 irqtime_account_process_tick(current, 0, rq); 2837} 2838#else /* CONFIG_IRQ_TIME_ACCOUNTING */ 2839static void irqtime_account_idle_ticks(int ticks) {} 2840static void irqtime_account_process_tick(struct task_struct *p, int user_tick, 2841 struct rq *rq) {} 2842#endif /* CONFIG_IRQ_TIME_ACCOUNTING */ 2843 2844/* 2845 * Account a single tick of cpu time. 2846 * @p: the process that the cpu time gets accounted to 2847 * @user_tick: indicates if the tick is a user or a system tick 2848 */ 2849void account_process_tick(struct task_struct *p, int user_tick) 2850{ 2851 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy); 2852 struct rq *rq = this_rq(); 2853 2854 if (sched_clock_irqtime) { 2855 irqtime_account_process_tick(p, user_tick, rq); 2856 return; 2857 } 2858 2859 if (steal_account_process_tick()) 2860 return; 2861 2862 if (user_tick) 2863 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled); 2864 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET)) 2865 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy, 2866 one_jiffy_scaled); 2867 else 2868 account_idle_time(cputime_one_jiffy); 2869} 2870 2871/* 2872 * Account multiple ticks of steal time. 2873 * @p: the process from which the cpu time has been stolen 2874 * @ticks: number of stolen ticks 2875 */ 2876void account_steal_ticks(unsigned long ticks) 2877{ 2878 account_steal_time(jiffies_to_cputime(ticks)); 2879} 2880 2881/* 2882 * Account multiple ticks of idle time. 2883 * @ticks: number of stolen ticks 2884 */ 2885void account_idle_ticks(unsigned long ticks) 2886{ 2887 2888 if (sched_clock_irqtime) { 2889 irqtime_account_idle_ticks(ticks); 2890 return; 2891 } 2892 2893 account_idle_time(jiffies_to_cputime(ticks)); 2894} 2895 2896#endif 2897 2898/* 2899 * Use precise platform statistics if available: 2900 */ 2901#ifdef CONFIG_VIRT_CPU_ACCOUNTING 2902void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st) 2903{ 2904 *ut = p->utime; 2905 *st = p->stime; 2906} 2907 2908void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st) 2909{ 2910 struct task_cputime cputime; 2911 2912 thread_group_cputime(p, &cputime); 2913 2914 *ut = cputime.utime; 2915 *st = cputime.stime; 2916} 2917#else 2918 2919#ifndef nsecs_to_cputime 2920# define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs) 2921#endif 2922 2923void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st) 2924{ 2925 cputime_t rtime, utime = p->utime, total = utime + p->stime; 2926 2927 /* 2928 * Use CFS's precise accounting: 2929 */ 2930 rtime = nsecs_to_cputime(p->se.sum_exec_runtime); 2931 2932 if (total) { 2933 u64 temp = (__force u64) rtime; 2934 2935 temp *= (__force u64) utime; 2936 do_div(temp, (__force u32) total); 2937 utime = (__force cputime_t) temp; 2938 } else 2939 utime = rtime; 2940 2941 /* 2942 * Compare with previous values, to keep monotonicity: 2943 */ 2944 p->prev_utime = max(p->prev_utime, utime); 2945 p->prev_stime = max(p->prev_stime, rtime - p->prev_utime); 2946 2947 *ut = p->prev_utime; 2948 *st = p->prev_stime; 2949} 2950 2951/* 2952 * Must be called with siglock held. 2953 */ 2954void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st) 2955{ 2956 struct signal_struct *sig = p->signal; 2957 struct task_cputime cputime; 2958 cputime_t rtime, utime, total; 2959 2960 thread_group_cputime(p, &cputime); 2961 2962 total = cputime.utime + cputime.stime; 2963 rtime = nsecs_to_cputime(cputime.sum_exec_runtime); 2964 2965 if (total) { 2966 u64 temp = (__force u64) rtime; 2967 2968 temp *= (__force u64) cputime.utime; 2969 do_div(temp, (__force u32) total); 2970 utime = (__force cputime_t) temp; 2971 } else 2972 utime = rtime; 2973 2974 sig->prev_utime = max(sig->prev_utime, utime); 2975 sig->prev_stime = max(sig->prev_stime, rtime - sig->prev_utime); 2976 2977 *ut = sig->prev_utime; 2978 *st = sig->prev_stime; 2979} 2980#endif 2981 2982/* 2983 * This function gets called by the timer code, with HZ frequency. 2984 * We call it with interrupts disabled. 2985 */ 2986void scheduler_tick(void) 2987{ 2988 int cpu = smp_processor_id(); 2989 struct rq *rq = cpu_rq(cpu); 2990 struct task_struct *curr = rq->curr; 2991 2992 sched_clock_tick(); 2993 2994 raw_spin_lock(&rq->lock); 2995 update_rq_clock(rq); 2996 update_cpu_load_active(rq); 2997 curr->sched_class->task_tick(rq, curr, 0); 2998 raw_spin_unlock(&rq->lock); 2999 3000 perf_event_task_tick(); 3001 3002#ifdef CONFIG_SMP 3003 rq->idle_balance = idle_cpu(cpu); 3004 trigger_load_balance(rq, cpu); 3005#endif 3006} 3007 3008notrace unsigned long get_parent_ip(unsigned long addr) 3009{ 3010 if (in_lock_functions(addr)) { 3011 addr = CALLER_ADDR2; 3012 if (in_lock_functions(addr)) 3013 addr = CALLER_ADDR3; 3014 } 3015 return addr; 3016} 3017 3018#if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \ 3019 defined(CONFIG_PREEMPT_TRACER)) 3020 3021void __kprobes add_preempt_count(int val) 3022{ 3023#ifdef CONFIG_DEBUG_PREEMPT 3024 /* 3025 * Underflow? 3026 */ 3027 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0))) 3028 return; 3029#endif 3030 preempt_count() += val; 3031#ifdef CONFIG_DEBUG_PREEMPT 3032 /* 3033 * Spinlock count overflowing soon? 3034 */ 3035 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >= 3036 PREEMPT_MASK - 10); 3037#endif 3038 if (preempt_count() == val) 3039 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1)); 3040} 3041EXPORT_SYMBOL(add_preempt_count); 3042 3043void __kprobes sub_preempt_count(int val) 3044{ 3045#ifdef CONFIG_DEBUG_PREEMPT 3046 /* 3047 * Underflow? 3048 */ 3049 if (DEBUG_LOCKS_WARN_ON(val > preempt_count())) 3050 return; 3051 /* 3052 * Is the spinlock portion underflowing? 3053 */ 3054 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) && 3055 !(preempt_count() & PREEMPT_MASK))) 3056 return; 3057#endif 3058 3059 if (preempt_count() == val) 3060 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1)); 3061 preempt_count() -= val; 3062} 3063EXPORT_SYMBOL(sub_preempt_count); 3064 3065#endif 3066 3067/* 3068 * Print scheduling while atomic bug: 3069 */ 3070static noinline void __schedule_bug(struct task_struct *prev) 3071{ 3072 struct pt_regs *regs = get_irq_regs(); 3073 3074 if (oops_in_progress) 3075 return; 3076 3077 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n", 3078 prev->comm, prev->pid, preempt_count()); 3079 3080 debug_show_held_locks(prev); 3081 print_modules(); 3082 if (irqs_disabled()) 3083 print_irqtrace_events(prev); 3084 3085 if (regs) 3086 show_regs(regs); 3087 else 3088 dump_stack(); 3089} 3090 3091/* 3092 * Various schedule()-time debugging checks and statistics: 3093 */ 3094static inline void schedule_debug(struct task_struct *prev) 3095{ 3096 /* 3097 * Test if we are atomic. Since do_exit() needs to call into 3098 * schedule() atomically, we ignore that path for now. 3099 * Otherwise, whine if we are scheduling when we should not be. 3100 */ 3101 if (unlikely(in_atomic_preempt_off() && !prev->exit_state)) 3102 __schedule_bug(prev); 3103 rcu_sleep_check(); 3104 3105 profile_hit(SCHED_PROFILING, __builtin_return_address(0)); 3106 3107 schedstat_inc(this_rq(), sched_count); 3108} 3109 3110static void put_prev_task(struct rq *rq, struct task_struct *prev) 3111{ 3112 if (prev->on_rq || rq->skip_clock_update < 0) 3113 update_rq_clock(rq); 3114 prev->sched_class->put_prev_task(rq, prev); 3115} 3116 3117/* 3118 * Pick up the highest-prio task: 3119 */ 3120static inline struct task_struct * 3121pick_next_task(struct rq *rq) 3122{ 3123 const struct sched_class *class; 3124 struct task_struct *p; 3125 3126 /* 3127 * Optimization: we know that if all tasks are in 3128 * the fair class we can call that function directly: 3129 */ 3130 if (likely(rq->nr_running == rq->cfs.h_nr_running)) { 3131 p = fair_sched_class.pick_next_task(rq); 3132 if (likely(p)) 3133 return p; 3134 } 3135 3136 for_each_class(class) { 3137 p = class->pick_next_task(rq); 3138 if (p) 3139 return p; 3140 } 3141 3142 BUG(); /* the idle class will always have a runnable task */ 3143} 3144 3145/* 3146 * __schedule() is the main scheduler function. 3147 */ 3148static void __sched __schedule(void) 3149{ 3150 struct task_struct *prev, *next; 3151 unsigned long *switch_count; 3152 struct rq *rq; 3153 int cpu; 3154 3155need_resched: 3156 preempt_disable(); 3157 cpu = smp_processor_id(); 3158 rq = cpu_rq(cpu); 3159 rcu_note_context_switch(cpu); 3160 prev = rq->curr; 3161 3162 schedule_debug(prev); 3163 3164 if (sched_feat(HRTICK)) 3165 hrtick_clear(rq); 3166 3167 raw_spin_lock_irq(&rq->lock); 3168 3169 switch_count = &prev->nivcsw; 3170 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) { 3171 if (unlikely(signal_pending_state(prev->state, prev))) { 3172 prev->state = TASK_RUNNING; 3173 } else { 3174 deactivate_task(rq, prev, DEQUEUE_SLEEP); 3175 prev->on_rq = 0; 3176 3177 /* 3178 * If a worker went to sleep, notify and ask workqueue 3179 * whether it wants to wake up a task to maintain 3180 * concurrency. 3181 */ 3182 if (prev->flags & PF_WQ_WORKER) { 3183 struct task_struct *to_wakeup; 3184 3185 to_wakeup = wq_worker_sleeping(prev, cpu); 3186 if (to_wakeup) 3187 try_to_wake_up_local(to_wakeup); 3188 } 3189 } 3190 switch_count = &prev->nvcsw; 3191 } 3192 3193 pre_schedule(rq, prev); 3194 3195 if (unlikely(!rq->nr_running)) 3196 idle_balance(cpu, rq); 3197 3198 put_prev_task(rq, prev); 3199 next = pick_next_task(rq); 3200 clear_tsk_need_resched(prev); 3201 rq->skip_clock_update = 0; 3202 3203 if (likely(prev != next)) { 3204 rq->nr_switches++; 3205 rq->curr = next; 3206 ++*switch_count; 3207 3208 context_switch(rq, prev, next); /* unlocks the rq */ 3209 /* 3210 * The context switch have flipped the stack from under us 3211 * and restored the local variables which were saved when 3212 * this task called schedule() in the past. prev == current 3213 * is still correct, but it can be moved to another cpu/rq. 3214 */ 3215 cpu = smp_processor_id(); 3216 rq = cpu_rq(cpu); 3217 } else 3218 raw_spin_unlock_irq(&rq->lock); 3219 3220 post_schedule(rq); 3221 3222 sched_preempt_enable_no_resched(); 3223 if (need_resched()) 3224 goto need_resched; 3225} 3226 3227static inline void sched_submit_work(struct task_struct *tsk) 3228{ 3229 if (!tsk->state || tsk_is_pi_blocked(tsk)) 3230 return; 3231 /* 3232 * If we are going to sleep and we have plugged IO queued, 3233 * make sure to submit it to avoid deadlocks. 3234 */ 3235 if (blk_needs_flush_plug(tsk)) 3236 blk_schedule_flush_plug(tsk); 3237} 3238 3239asmlinkage void __sched schedule(void) 3240{ 3241 struct task_struct *tsk = current; 3242 3243 sched_submit_work(tsk); 3244 __schedule(); 3245} 3246EXPORT_SYMBOL(schedule); 3247 3248/** 3249 * schedule_preempt_disabled - called with preemption disabled 3250 * 3251 * Returns with preemption disabled. Note: preempt_count must be 1 3252 */ 3253void __sched schedule_preempt_disabled(void) 3254{ 3255 sched_preempt_enable_no_resched(); 3256 schedule(); 3257 preempt_disable(); 3258} 3259 3260#ifdef CONFIG_MUTEX_SPIN_ON_OWNER 3261 3262static inline bool owner_running(struct mutex *lock, struct task_struct *owner) 3263{ 3264 if (lock->owner != owner) 3265 return false; 3266 3267 /* 3268 * Ensure we emit the owner->on_cpu, dereference _after_ checking 3269 * lock->owner still matches owner, if that fails, owner might 3270 * point to free()d memory, if it still matches, the rcu_read_lock() 3271 * ensures the memory stays valid. 3272 */ 3273 barrier(); 3274 3275 return owner->on_cpu; 3276} 3277 3278/* 3279 * Look out! "owner" is an entirely speculative pointer 3280 * access and not reliable. 3281 */ 3282int mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner) 3283{ 3284 if (!sched_feat(OWNER_SPIN)) 3285 return 0; 3286 3287 rcu_read_lock(); 3288 while (owner_running(lock, owner)) { 3289 if (need_resched()) 3290 break; 3291 3292 arch_mutex_cpu_relax(); 3293 } 3294 rcu_read_unlock(); 3295 3296 /* 3297 * We break out the loop above on need_resched() and when the 3298 * owner changed, which is a sign for heavy contention. Return 3299 * success only when lock->owner is NULL. 3300 */ 3301 return lock->owner == NULL; 3302} 3303#endif 3304 3305#ifdef CONFIG_PREEMPT 3306/* 3307 * this is the entry point to schedule() from in-kernel preemption 3308 * off of preempt_enable. Kernel preemptions off return from interrupt 3309 * occur there and call schedule directly. 3310 */ 3311asmlinkage void __sched notrace preempt_schedule(void) 3312{ 3313 struct thread_info *ti = current_thread_info(); 3314 3315 /* 3316 * If there is a non-zero preempt_count or interrupts are disabled, 3317 * we do not want to preempt the current task. Just return.. 3318 */ 3319 if (likely(ti->preempt_count || irqs_disabled())) 3320 return; 3321 3322 do { 3323 add_preempt_count_notrace(PREEMPT_ACTIVE); 3324 __schedule(); 3325 sub_preempt_count_notrace(PREEMPT_ACTIVE); 3326 3327 /* 3328 * Check again in case we missed a preemption opportunity 3329 * between schedule and now. 3330 */ 3331 barrier(); 3332 } while (need_resched()); 3333} 3334EXPORT_SYMBOL(preempt_schedule); 3335 3336/* 3337 * this is the entry point to schedule() from kernel preemption 3338 * off of irq context. 3339 * Note, that this is called and return with irqs disabled. This will 3340 * protect us against recursive calling from irq. 3341 */ 3342asmlinkage void __sched preempt_schedule_irq(void) 3343{ 3344 struct thread_info *ti = current_thread_info(); 3345 3346 /* Catch callers which need to be fixed */ 3347 BUG_ON(ti->preempt_count || !irqs_disabled()); 3348 3349 do { 3350 add_preempt_count(PREEMPT_ACTIVE); 3351 local_irq_enable(); 3352 __schedule(); 3353 local_irq_disable(); 3354 sub_preempt_count(PREEMPT_ACTIVE); 3355 3356 /* 3357 * Check again in case we missed a preemption opportunity 3358 * between schedule and now. 3359 */ 3360 barrier(); 3361 } while (need_resched()); 3362} 3363 3364#endif /* CONFIG_PREEMPT */ 3365 3366int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags, 3367 void *key) 3368{ 3369 return try_to_wake_up(curr->private, mode, wake_flags); 3370} 3371EXPORT_SYMBOL(default_wake_function); 3372 3373/* 3374 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just 3375 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve 3376 * number) then we wake all the non-exclusive tasks and one exclusive task. 3377 * 3378 * There are circumstances in which we can try to wake a task which has already 3379 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns 3380 * zero in this (rare) case, and we handle it by continuing to scan the queue. 3381 */ 3382static void __wake_up_common(wait_queue_head_t *q, unsigned int mode, 3383 int nr_exclusive, int wake_flags, void *key) 3384{ 3385 wait_queue_t *curr, *next; 3386 3387 list_for_each_entry_safe(curr, next, &q->task_list, task_list) { 3388 unsigned flags = curr->flags; 3389 3390 if (curr->func(curr, mode, wake_flags, key) && 3391 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive) 3392 break; 3393 } 3394} 3395 3396/** 3397 * __wake_up - wake up threads blocked on a waitqueue. 3398 * @q: the waitqueue 3399 * @mode: which threads 3400 * @nr_exclusive: how many wake-one or wake-many threads to wake up 3401 * @key: is directly passed to the wakeup function 3402 * 3403 * It may be assumed that this function implies a write memory barrier before 3404 * changing the task state if and only if any tasks are woken up. 3405 */ 3406void __wake_up(wait_queue_head_t *q, unsigned int mode, 3407 int nr_exclusive, void *key) 3408{ 3409 unsigned long flags; 3410 3411 spin_lock_irqsave(&q->lock, flags); 3412 __wake_up_common(q, mode, nr_exclusive, 0, key); 3413 spin_unlock_irqrestore(&q->lock, flags); 3414} 3415EXPORT_SYMBOL(__wake_up); 3416 3417/* 3418 * Same as __wake_up but called with the spinlock in wait_queue_head_t held. 3419 */ 3420void __wake_up_locked(wait_queue_head_t *q, unsigned int mode, int nr) 3421{ 3422 __wake_up_common(q, mode, nr, 0, NULL); 3423} 3424EXPORT_SYMBOL_GPL(__wake_up_locked); 3425 3426void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key) 3427{ 3428 __wake_up_common(q, mode, 1, 0, key); 3429} 3430EXPORT_SYMBOL_GPL(__wake_up_locked_key); 3431 3432/** 3433 * __wake_up_sync_key - wake up threads blocked on a waitqueue. 3434 * @q: the waitqueue 3435 * @mode: which threads 3436 * @nr_exclusive: how many wake-one or wake-many threads to wake up 3437 * @key: opaque value to be passed to wakeup targets 3438 * 3439 * The sync wakeup differs that the waker knows that it will schedule 3440 * away soon, so while the target thread will be woken up, it will not 3441 * be migrated to another CPU - ie. the two threads are 'synchronized' 3442 * with each other. This can prevent needless bouncing between CPUs. 3443 * 3444 * On UP it can prevent extra preemption. 3445 * 3446 * It may be assumed that this function implies a write memory barrier before 3447 * changing the task state if and only if any tasks are woken up. 3448 */ 3449void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode, 3450 int nr_exclusive, void *key) 3451{ 3452 unsigned long flags; 3453 int wake_flags = WF_SYNC; 3454 3455 if (unlikely(!q)) 3456 return; 3457 3458 if (unlikely(!nr_exclusive)) 3459 wake_flags = 0; 3460 3461 spin_lock_irqsave(&q->lock, flags); 3462 __wake_up_common(q, mode, nr_exclusive, wake_flags, key); 3463 spin_unlock_irqrestore(&q->lock, flags); 3464} 3465EXPORT_SYMBOL_GPL(__wake_up_sync_key); 3466 3467/* 3468 * __wake_up_sync - see __wake_up_sync_key() 3469 */ 3470void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive) 3471{ 3472 __wake_up_sync_key(q, mode, nr_exclusive, NULL); 3473} 3474EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */ 3475 3476/** 3477 * complete: - signals a single thread waiting on this completion 3478 * @x: holds the state of this particular completion 3479 * 3480 * This will wake up a single thread waiting on this completion. Threads will be 3481 * awakened in the same order in which they were queued. 3482 * 3483 * See also complete_all(), wait_for_completion() and related routines. 3484 * 3485 * It may be assumed that this function implies a write memory barrier before 3486 * changing the task state if and only if any tasks are woken up. 3487 */ 3488void complete(struct completion *x) 3489{ 3490 unsigned long flags; 3491 3492 spin_lock_irqsave(&x->wait.lock, flags); 3493 x->done++; 3494 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL); 3495 spin_unlock_irqrestore(&x->wait.lock, flags); 3496} 3497EXPORT_SYMBOL(complete); 3498 3499/** 3500 * complete_all: - signals all threads waiting on this completion 3501 * @x: holds the state of this particular completion 3502 * 3503 * This will wake up all threads waiting on this particular completion event. 3504 * 3505 * It may be assumed that this function implies a write memory barrier before 3506 * changing the task state if and only if any tasks are woken up. 3507 */ 3508void complete_all(struct completion *x) 3509{ 3510 unsigned long flags; 3511 3512 spin_lock_irqsave(&x->wait.lock, flags); 3513 x->done += UINT_MAX/2; 3514 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL); 3515 spin_unlock_irqrestore(&x->wait.lock, flags); 3516} 3517EXPORT_SYMBOL(complete_all); 3518 3519static inline long __sched 3520do_wait_for_common(struct completion *x, long timeout, int state) 3521{ 3522 if (!x->done) { 3523 DECLARE_WAITQUEUE(wait, current); 3524 3525 __add_wait_queue_tail_exclusive(&x->wait, &wait); 3526 do { 3527 if (signal_pending_state(state, current)) { 3528 timeout = -ERESTARTSYS; 3529 break; 3530 } 3531 __set_current_state(state); 3532 spin_unlock_irq(&x->wait.lock); 3533 timeout = schedule_timeout(timeout); 3534 spin_lock_irq(&x->wait.lock); 3535 } while (!x->done && timeout); 3536 __remove_wait_queue(&x->wait, &wait); 3537 if (!x->done) 3538 return timeout; 3539 } 3540 x->done--; 3541 return timeout ?: 1; 3542} 3543 3544static long __sched 3545wait_for_common(struct completion *x, long timeout, int state) 3546{ 3547 might_sleep(); 3548 3549 spin_lock_irq(&x->wait.lock); 3550 timeout = do_wait_for_common(x, timeout, state); 3551 spin_unlock_irq(&x->wait.lock); 3552 return timeout; 3553} 3554 3555/** 3556 * wait_for_completion: - waits for completion of a task 3557 * @x: holds the state of this particular completion 3558 * 3559 * This waits to be signaled for completion of a specific task. It is NOT 3560 * interruptible and there is no timeout. 3561 * 3562 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout 3563 * and interrupt capability. Also see complete(). 3564 */ 3565void __sched wait_for_completion(struct completion *x) 3566{ 3567 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE); 3568} 3569EXPORT_SYMBOL(wait_for_completion); 3570 3571/** 3572 * wait_for_completion_timeout: - waits for completion of a task (w/timeout) 3573 * @x: holds the state of this particular completion 3574 * @timeout: timeout value in jiffies 3575 * 3576 * This waits for either a completion of a specific task to be signaled or for a 3577 * specified timeout to expire. The timeout is in jiffies. It is not 3578 * interruptible. 3579 * 3580 * The return value is 0 if timed out, and positive (at least 1, or number of 3581 * jiffies left till timeout) if completed. 3582 */ 3583unsigned long __sched 3584wait_for_completion_timeout(struct completion *x, unsigned long timeout) 3585{ 3586 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE); 3587} 3588EXPORT_SYMBOL(wait_for_completion_timeout); 3589 3590/** 3591 * wait_for_completion_interruptible: - waits for completion of a task (w/intr) 3592 * @x: holds the state of this particular completion 3593 * 3594 * This waits for completion of a specific task to be signaled. It is 3595 * interruptible. 3596 * 3597 * The return value is -ERESTARTSYS if interrupted, 0 if completed. 3598 */ 3599int __sched wait_for_completion_interruptible(struct completion *x) 3600{ 3601 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE); 3602 if (t == -ERESTARTSYS) 3603 return t; 3604 return 0; 3605} 3606EXPORT_SYMBOL(wait_for_completion_interruptible); 3607 3608/** 3609 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr)) 3610 * @x: holds the state of this particular completion 3611 * @timeout: timeout value in jiffies 3612 * 3613 * This waits for either a completion of a specific task to be signaled or for a 3614 * specified timeout to expire. It is interruptible. The timeout is in jiffies. 3615 * 3616 * The return value is -ERESTARTSYS if interrupted, 0 if timed out, 3617 * positive (at least 1, or number of jiffies left till timeout) if completed. 3618 */ 3619long __sched 3620wait_for_completion_interruptible_timeout(struct completion *x, 3621 unsigned long timeout) 3622{ 3623 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE); 3624} 3625EXPORT_SYMBOL(wait_for_completion_interruptible_timeout); 3626 3627/** 3628 * wait_for_completion_killable: - waits for completion of a task (killable) 3629 * @x: holds the state of this particular completion 3630 * 3631 * This waits to be signaled for completion of a specific task. It can be 3632 * interrupted by a kill signal. 3633 * 3634 * The return value is -ERESTARTSYS if interrupted, 0 if completed. 3635 */ 3636int __sched wait_for_completion_killable(struct completion *x) 3637{ 3638 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE); 3639 if (t == -ERESTARTSYS) 3640 return t; 3641 return 0; 3642} 3643EXPORT_SYMBOL(wait_for_completion_killable); 3644 3645/** 3646 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable)) 3647 * @x: holds the state of this particular completion 3648 * @timeout: timeout value in jiffies 3649 * 3650 * This waits for either a completion of a specific task to be 3651 * signaled or for a specified timeout to expire. It can be 3652 * interrupted by a kill signal. The timeout is in jiffies. 3653 * 3654 * The return value is -ERESTARTSYS if interrupted, 0 if timed out, 3655 * positive (at least 1, or number of jiffies left till timeout) if completed. 3656 */ 3657long __sched 3658wait_for_completion_killable_timeout(struct completion *x, 3659 unsigned long timeout) 3660{ 3661 return wait_for_common(x, timeout, TASK_KILLABLE); 3662} 3663EXPORT_SYMBOL(wait_for_completion_killable_timeout); 3664 3665/** 3666 * try_wait_for_completion - try to decrement a completion without blocking 3667 * @x: completion structure 3668 * 3669 * Returns: 0 if a decrement cannot be done without blocking 3670 * 1 if a decrement succeeded. 3671 * 3672 * If a completion is being used as a counting completion, 3673 * attempt to decrement the counter without blocking. This 3674 * enables us to avoid waiting if the resource the completion 3675 * is protecting is not available. 3676 */ 3677bool try_wait_for_completion(struct completion *x) 3678{ 3679 unsigned long flags; 3680 int ret = 1; 3681 3682 spin_lock_irqsave(&x->wait.lock, flags); 3683 if (!x->done) 3684 ret = 0; 3685 else 3686 x->done--; 3687 spin_unlock_irqrestore(&x->wait.lock, flags); 3688 return ret; 3689} 3690EXPORT_SYMBOL(try_wait_for_completion); 3691 3692/** 3693 * completion_done - Test to see if a completion has any waiters 3694 * @x: completion structure 3695 * 3696 * Returns: 0 if there are waiters (wait_for_completion() in progress) 3697 * 1 if there are no waiters. 3698 * 3699 */ 3700bool completion_done(struct completion *x) 3701{ 3702 unsigned long flags; 3703 int ret = 1; 3704 3705 spin_lock_irqsave(&x->wait.lock, flags); 3706 if (!x->done) 3707 ret = 0; 3708 spin_unlock_irqrestore(&x->wait.lock, flags); 3709 return ret; 3710} 3711EXPORT_SYMBOL(completion_done); 3712 3713static long __sched 3714sleep_on_common(wait_queue_head_t *q, int state, long timeout) 3715{ 3716 unsigned long flags; 3717 wait_queue_t wait; 3718 3719 init_waitqueue_entry(&wait, current); 3720 3721 __set_current_state(state); 3722 3723 spin_lock_irqsave(&q->lock, flags); 3724 __add_wait_queue(q, &wait); 3725 spin_unlock(&q->lock); 3726 timeout = schedule_timeout(timeout); 3727 spin_lock_irq(&q->lock); 3728 __remove_wait_queue(q, &wait); 3729 spin_unlock_irqrestore(&q->lock, flags); 3730 3731 return timeout; 3732} 3733 3734void __sched interruptible_sleep_on(wait_queue_head_t *q) 3735{ 3736 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT); 3737} 3738EXPORT_SYMBOL(interruptible_sleep_on); 3739 3740long __sched 3741interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout) 3742{ 3743 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout); 3744} 3745EXPORT_SYMBOL(interruptible_sleep_on_timeout); 3746 3747void __sched sleep_on(wait_queue_head_t *q) 3748{ 3749 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT); 3750} 3751EXPORT_SYMBOL(sleep_on); 3752 3753long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout) 3754{ 3755 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout); 3756} 3757EXPORT_SYMBOL(sleep_on_timeout); 3758 3759#ifdef CONFIG_RT_MUTEXES 3760 3761/* 3762 * rt_mutex_setprio - set the current priority of a task 3763 * @p: task 3764 * @prio: prio value (kernel-internal form) 3765 * 3766 * This function changes the 'effective' priority of a task. It does 3767 * not touch ->normal_prio like __setscheduler(). 3768 * 3769 * Used by the rt_mutex code to implement priority inheritance logic. 3770 */ 3771void rt_mutex_setprio(struct task_struct *p, int prio) 3772{ 3773 int oldprio, on_rq, running; 3774 struct rq *rq; 3775 const struct sched_class *prev_class; 3776 3777 BUG_ON(prio < 0 || prio > MAX_PRIO); 3778 3779 rq = __task_rq_lock(p); 3780 3781 /* 3782 * Idle task boosting is a nono in general. There is one 3783 * exception, when PREEMPT_RT and NOHZ is active: 3784 * 3785 * The idle task calls get_next_timer_interrupt() and holds 3786 * the timer wheel base->lock on the CPU and another CPU wants 3787 * to access the timer (probably to cancel it). We can safely 3788 * ignore the boosting request, as the idle CPU runs this code 3789 * with interrupts disabled and will complete the lock 3790 * protected section without being interrupted. So there is no 3791 * real need to boost. 3792 */ 3793 if (unlikely(p == rq->idle)) { 3794 WARN_ON(p != rq->curr); 3795 WARN_ON(p->pi_blocked_on); 3796 goto out_unlock; 3797 } 3798 3799 trace_sched_pi_setprio(p, prio); 3800 oldprio = p->prio; 3801 prev_class = p->sched_class; 3802 on_rq = p->on_rq; 3803 running = task_current(rq, p); 3804 if (on_rq) 3805 dequeue_task(rq, p, 0); 3806 if (running) 3807 p->sched_class->put_prev_task(rq, p); 3808 3809 if (rt_prio(prio)) 3810 p->sched_class = &rt_sched_class; 3811 else 3812 p->sched_class = &fair_sched_class; 3813 3814 p->prio = prio; 3815 3816 if (running) 3817 p->sched_class->set_curr_task(rq); 3818 if (on_rq) 3819 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0); 3820 3821 check_class_changed(rq, p, prev_class, oldprio); 3822out_unlock: 3823 __task_rq_unlock(rq); 3824} 3825#endif 3826void set_user_nice(struct task_struct *p, long nice) 3827{ 3828 int old_prio, delta, on_rq; 3829 unsigned long flags; 3830 struct rq *rq; 3831 3832 if (TASK_NICE(p) == nice || nice < -20 || nice > 19) 3833 return; 3834 /* 3835 * We have to be careful, if called from sys_setpriority(), 3836 * the task might be in the middle of scheduling on another CPU. 3837 */ 3838 rq = task_rq_lock(p, &flags); 3839 /* 3840 * The RT priorities are set via sched_setscheduler(), but we still 3841 * allow the 'normal' nice value to be set - but as expected 3842 * it wont have any effect on scheduling until the task is 3843 * SCHED_FIFO/SCHED_RR: 3844 */ 3845 if (task_has_rt_policy(p)) { 3846 p->static_prio = NICE_TO_PRIO(nice); 3847 goto out_unlock; 3848 } 3849 on_rq = p->on_rq; 3850 if (on_rq) 3851 dequeue_task(rq, p, 0); 3852 3853 p->static_prio = NICE_TO_PRIO(nice); 3854 set_load_weight(p); 3855 old_prio = p->prio; 3856 p->prio = effective_prio(p); 3857 delta = p->prio - old_prio; 3858 3859 if (on_rq) { 3860 enqueue_task(rq, p, 0); 3861 /* 3862 * If the task increased its priority or is running and 3863 * lowered its priority, then reschedule its CPU: 3864 */ 3865 if (delta < 0 || (delta > 0 && task_running(rq, p))) 3866 resched_task(rq->curr); 3867 } 3868out_unlock: 3869 task_rq_unlock(rq, p, &flags); 3870} 3871EXPORT_SYMBOL(set_user_nice); 3872 3873/* 3874 * can_nice - check if a task can reduce its nice value 3875 * @p: task 3876 * @nice: nice value 3877 */ 3878int can_nice(const struct task_struct *p, const int nice) 3879{ 3880 /* convert nice value [19,-20] to rlimit style value [1,40] */ 3881 int nice_rlim = 20 - nice; 3882 3883 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) || 3884 capable(CAP_SYS_NICE)); 3885} 3886 3887#ifdef __ARCH_WANT_SYS_NICE 3888 3889/* 3890 * sys_nice - change the priority of the current process. 3891 * @increment: priority increment 3892 * 3893 * sys_setpriority is a more generic, but much slower function that 3894 * does similar things. 3895 */ 3896SYSCALL_DEFINE1(nice, int, increment) 3897{ 3898 long nice, retval; 3899 3900 /* 3901 * Setpriority might change our priority at the same moment. 3902 * We don't have to worry. Conceptually one call occurs first 3903 * and we have a single winner. 3904 */ 3905 if (increment < -40) 3906 increment = -40; 3907 if (increment > 40) 3908 increment = 40; 3909 3910 nice = TASK_NICE(current) + increment; 3911 if (nice < -20) 3912 nice = -20; 3913 if (nice > 19) 3914 nice = 19; 3915 3916 if (increment < 0 && !can_nice(current, nice)) 3917 return -EPERM; 3918 3919 retval = security_task_setnice(current, nice); 3920 if (retval) 3921 return retval; 3922 3923 set_user_nice(current, nice); 3924 return 0; 3925} 3926 3927#endif 3928 3929/** 3930 * task_prio - return the priority value of a given task. 3931 * @p: the task in question. 3932 * 3933 * This is the priority value as seen by users in /proc. 3934 * RT tasks are offset by -200. Normal tasks are centered 3935 * around 0, value goes from -16 to +15. 3936 */ 3937int task_prio(const struct task_struct *p) 3938{ 3939 return p->prio - MAX_RT_PRIO; 3940} 3941 3942/** 3943 * task_nice - return the nice value of a given task. 3944 * @p: the task in question. 3945 */ 3946int task_nice(const struct task_struct *p) 3947{ 3948 return TASK_NICE(p); 3949} 3950EXPORT_SYMBOL(task_nice); 3951 3952/** 3953 * idle_cpu - is a given cpu idle currently? 3954 * @cpu: the processor in question. 3955 */ 3956int idle_cpu(int cpu) 3957{ 3958 struct rq *rq = cpu_rq(cpu); 3959 3960 if (rq->curr != rq->idle) 3961 return 0; 3962 3963 if (rq->nr_running) 3964 return 0; 3965 3966#ifdef CONFIG_SMP 3967 if (!llist_empty(&rq->wake_list)) 3968 return 0; 3969#endif 3970 3971 return 1; 3972} 3973 3974/** 3975 * idle_task - return the idle task for a given cpu. 3976 * @cpu: the processor in question. 3977 */ 3978struct task_struct *idle_task(int cpu) 3979{ 3980 return cpu_rq(cpu)->idle; 3981} 3982 3983/** 3984 * find_process_by_pid - find a process with a matching PID value. 3985 * @pid: the pid in question. 3986 */ 3987static struct task_struct *find_process_by_pid(pid_t pid) 3988{ 3989 return pid ? find_task_by_vpid(pid) : current; 3990} 3991 3992/* Actually do priority change: must hold rq lock. */ 3993static void 3994__setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio) 3995{ 3996 p->policy = policy; 3997 p->rt_priority = prio; 3998 p->normal_prio = normal_prio(p); 3999 /* we are holding p->pi_lock already */ 4000 p->prio = rt_mutex_getprio(p); 4001 if (rt_prio(p->prio)) 4002 p->sched_class = &rt_sched_class; 4003 else 4004 p->sched_class = &fair_sched_class; 4005 set_load_weight(p); 4006} 4007 4008/* 4009 * check the target process has a UID that matches the current process's 4010 */ 4011static bool check_same_owner(struct task_struct *p) 4012{ 4013 const struct cred *cred = current_cred(), *pcred; 4014 bool match; 4015 4016 rcu_read_lock(); 4017 pcred = __task_cred(p); 4018 if (cred->user->user_ns == pcred->user->user_ns) 4019 match = (cred->euid == pcred->euid || 4020 cred->euid == pcred->uid); 4021 else 4022 match = false; 4023 rcu_read_unlock(); 4024 return match; 4025} 4026 4027static int __sched_setscheduler(struct task_struct *p, int policy, 4028 const struct sched_param *param, bool user) 4029{ 4030 int retval, oldprio, oldpolicy = -1, on_rq, running; 4031 unsigned long flags; 4032 const struct sched_class *prev_class; 4033 struct rq *rq; 4034 int reset_on_fork; 4035 4036 /* may grab non-irq protected spin_locks */ 4037 BUG_ON(in_interrupt()); 4038recheck: 4039 /* double check policy once rq lock held */ 4040 if (policy < 0) { 4041 reset_on_fork = p->sched_reset_on_fork; 4042 policy = oldpolicy = p->policy; 4043 } else { 4044 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK); 4045 policy &= ~SCHED_RESET_ON_FORK; 4046 4047 if (policy != SCHED_FIFO && policy != SCHED_RR && 4048 policy != SCHED_NORMAL && policy != SCHED_BATCH && 4049 policy != SCHED_IDLE) 4050 return -EINVAL; 4051 } 4052 4053 /* 4054 * Valid priorities for SCHED_FIFO and SCHED_RR are 4055 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL, 4056 * SCHED_BATCH and SCHED_IDLE is 0. 4057 */ 4058 if (param->sched_priority < 0 || 4059 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) || 4060 (!p->mm && param->sched_priority > MAX_RT_PRIO-1)) 4061 return -EINVAL; 4062 if (rt_policy(policy) != (param->sched_priority != 0)) 4063 return -EINVAL; 4064 4065 /* 4066 * Allow unprivileged RT tasks to decrease priority: 4067 */ 4068 if (user && !capable(CAP_SYS_NICE)) { 4069 if (rt_policy(policy)) { 4070 unsigned long rlim_rtprio = 4071 task_rlimit(p, RLIMIT_RTPRIO); 4072 4073 /* can't set/change the rt policy */ 4074 if (policy != p->policy && !rlim_rtprio) 4075 return -EPERM; 4076 4077 /* can't increase priority */ 4078 if (param->sched_priority > p->rt_priority && 4079 param->sched_priority > rlim_rtprio) 4080 return -EPERM; 4081 } 4082 4083 /* 4084 * Treat SCHED_IDLE as nice 20. Only allow a switch to 4085 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it. 4086 */ 4087 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) { 4088 if (!can_nice(p, TASK_NICE(p))) 4089 return -EPERM; 4090 } 4091 4092 /* can't change other user's priorities */ 4093 if (!check_same_owner(p)) 4094 return -EPERM; 4095 4096 /* Normal users shall not reset the sched_reset_on_fork flag */ 4097 if (p->sched_reset_on_fork && !reset_on_fork) 4098 return -EPERM; 4099 } 4100 4101 if (user) { 4102 retval = security_task_setscheduler(p); 4103 if (retval) 4104 return retval; 4105 } 4106 4107 /* 4108 * make sure no PI-waiters arrive (or leave) while we are 4109 * changing the priority of the task: 4110 * 4111 * To be able to change p->policy safely, the appropriate 4112 * runqueue lock must be held. 4113 */ 4114 rq = task_rq_lock(p, &flags); 4115 4116 /* 4117 * Changing the policy of the stop threads its a very bad idea 4118 */ 4119 if (p == rq->stop) { 4120 task_rq_unlock(rq, p, &flags); 4121 return -EINVAL; 4122 } 4123 4124 /* 4125 * If not changing anything there's no need to proceed further: 4126 */ 4127 if (unlikely(policy == p->policy && (!rt_policy(policy) || 4128 param->sched_priority == p->rt_priority))) { 4129 4130 __task_rq_unlock(rq); 4131 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 4132 return 0; 4133 } 4134 4135#ifdef CONFIG_RT_GROUP_SCHED 4136 if (user) { 4137 /* 4138 * Do not allow realtime tasks into groups that have no runtime 4139 * assigned. 4140 */ 4141 if (rt_bandwidth_enabled() && rt_policy(policy) && 4142 task_group(p)->rt_bandwidth.rt_runtime == 0 && 4143 !task_group_is_autogroup(task_group(p))) { 4144 task_rq_unlock(rq, p, &flags); 4145 return -EPERM; 4146 } 4147 } 4148#endif 4149 4150 /* recheck policy now with rq lock held */ 4151 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) { 4152 policy = oldpolicy = -1; 4153 task_rq_unlock(rq, p, &flags); 4154 goto recheck; 4155 } 4156 on_rq = p->on_rq; 4157 running = task_current(rq, p); 4158 if (on_rq) 4159 dequeue_task(rq, p, 0); 4160 if (running) 4161 p->sched_class->put_prev_task(rq, p); 4162 4163 p->sched_reset_on_fork = reset_on_fork; 4164 4165 oldprio = p->prio; 4166 prev_class = p->sched_class; 4167 __setscheduler(rq, p, policy, param->sched_priority); 4168 4169 if (running) 4170 p->sched_class->set_curr_task(rq); 4171 if (on_rq) 4172 enqueue_task(rq, p, 0); 4173 4174 check_class_changed(rq, p, prev_class, oldprio); 4175 task_rq_unlock(rq, p, &flags); 4176 4177 rt_mutex_adjust_pi(p); 4178 4179 return 0; 4180} 4181 4182/** 4183 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread. 4184 * @p: the task in question. 4185 * @policy: new policy. 4186 * @param: structure containing the new RT priority. 4187 * 4188 * NOTE that the task may be already dead. 4189 */ 4190int sched_setscheduler(struct task_struct *p, int policy, 4191 const struct sched_param *param) 4192{ 4193 return __sched_setscheduler(p, policy, param, true); 4194} 4195EXPORT_SYMBOL_GPL(sched_setscheduler); 4196 4197/** 4198 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace. 4199 * @p: the task in question. 4200 * @policy: new policy. 4201 * @param: structure containing the new RT priority. 4202 * 4203 * Just like sched_setscheduler, only don't bother checking if the 4204 * current context has permission. For example, this is needed in 4205 * stop_machine(): we create temporary high priority worker threads, 4206 * but our caller might not have that capability. 4207 */ 4208int sched_setscheduler_nocheck(struct task_struct *p, int policy, 4209 const struct sched_param *param) 4210{ 4211 return __sched_setscheduler(p, policy, param, false); 4212} 4213 4214static int 4215do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param) 4216{ 4217 struct sched_param lparam; 4218 struct task_struct *p; 4219 int retval; 4220 4221 if (!param || pid < 0) 4222 return -EINVAL; 4223 if (copy_from_user(&lparam, param, sizeof(struct sched_param))) 4224 return -EFAULT; 4225 4226 rcu_read_lock(); 4227 retval = -ESRCH; 4228 p = find_process_by_pid(pid); 4229 if (p != NULL) 4230 retval = sched_setscheduler(p, policy, &lparam); 4231 rcu_read_unlock(); 4232 4233 return retval; 4234} 4235 4236/** 4237 * sys_sched_setscheduler - set/change the scheduler policy and RT priority 4238 * @pid: the pid in question. 4239 * @policy: new policy. 4240 * @param: structure containing the new RT priority. 4241 */ 4242SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, 4243 struct sched_param __user *, param) 4244{ 4245 /* negative values for policy are not valid */ 4246 if (policy < 0) 4247 return -EINVAL; 4248 4249 return do_sched_setscheduler(pid, policy, param); 4250} 4251 4252/** 4253 * sys_sched_setparam - set/change the RT priority of a thread 4254 * @pid: the pid in question. 4255 * @param: structure containing the new RT priority. 4256 */ 4257SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param) 4258{ 4259 return do_sched_setscheduler(pid, -1, param); 4260} 4261 4262/** 4263 * sys_sched_getscheduler - get the policy (scheduling class) of a thread 4264 * @pid: the pid in question. 4265 */ 4266SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid) 4267{ 4268 struct task_struct *p; 4269 int retval; 4270 4271 if (pid < 0) 4272 return -EINVAL; 4273 4274 retval = -ESRCH; 4275 rcu_read_lock(); 4276 p = find_process_by_pid(pid); 4277 if (p) { 4278 retval = security_task_getscheduler(p); 4279 if (!retval) 4280 retval = p->policy 4281 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0); 4282 } 4283 rcu_read_unlock(); 4284 return retval; 4285} 4286 4287/** 4288 * sys_sched_getparam - get the RT priority of a thread 4289 * @pid: the pid in question. 4290 * @param: structure containing the RT priority. 4291 */ 4292SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param) 4293{ 4294 struct sched_param lp; 4295 struct task_struct *p; 4296 int retval; 4297 4298 if (!param || pid < 0) 4299 return -EINVAL; 4300 4301 rcu_read_lock(); 4302 p = find_process_by_pid(pid); 4303 retval = -ESRCH; 4304 if (!p) 4305 goto out_unlock; 4306 4307 retval = security_task_getscheduler(p); 4308 if (retval) 4309 goto out_unlock; 4310 4311 lp.sched_priority = p->rt_priority; 4312 rcu_read_unlock(); 4313 4314 /* 4315 * This one might sleep, we cannot do it with a spinlock held ... 4316 */ 4317 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0; 4318 4319 return retval; 4320 4321out_unlock: 4322 rcu_read_unlock(); 4323 return retval; 4324} 4325 4326long sched_setaffinity(pid_t pid, const struct cpumask *in_mask) 4327{ 4328 cpumask_var_t cpus_allowed, new_mask; 4329 struct task_struct *p; 4330 int retval; 4331 4332 get_online_cpus(); 4333 rcu_read_lock(); 4334 4335 p = find_process_by_pid(pid); 4336 if (!p) { 4337 rcu_read_unlock(); 4338 put_online_cpus(); 4339 return -ESRCH; 4340 } 4341 4342 /* Prevent p going away */ 4343 get_task_struct(p); 4344 rcu_read_unlock(); 4345 4346 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) { 4347 retval = -ENOMEM; 4348 goto out_put_task; 4349 } 4350 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) { 4351 retval = -ENOMEM; 4352 goto out_free_cpus_allowed; 4353 } 4354 retval = -EPERM; 4355 if (!check_same_owner(p) && !ns_capable(task_user_ns(p), CAP_SYS_NICE)) 4356 goto out_unlock; 4357 4358 retval = security_task_setscheduler(p); 4359 if (retval) 4360 goto out_unlock; 4361 4362 cpuset_cpus_allowed(p, cpus_allowed); 4363 cpumask_and(new_mask, in_mask, cpus_allowed); 4364again: 4365 retval = set_cpus_allowed_ptr(p, new_mask); 4366 4367 if (!retval) { 4368 cpuset_cpus_allowed(p, cpus_allowed); 4369 if (!cpumask_subset(new_mask, cpus_allowed)) { 4370 /* 4371 * We must have raced with a concurrent cpuset 4372 * update. Just reset the cpus_allowed to the 4373 * cpuset's cpus_allowed 4374 */ 4375 cpumask_copy(new_mask, cpus_allowed); 4376 goto again; 4377 } 4378 } 4379out_unlock: 4380 free_cpumask_var(new_mask); 4381out_free_cpus_allowed: 4382 free_cpumask_var(cpus_allowed); 4383out_put_task: 4384 put_task_struct(p); 4385 put_online_cpus(); 4386 return retval; 4387} 4388 4389static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len, 4390 struct cpumask *new_mask) 4391{ 4392 if (len < cpumask_size()) 4393 cpumask_clear(new_mask); 4394 else if (len > cpumask_size()) 4395 len = cpumask_size(); 4396 4397 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0; 4398} 4399 4400/** 4401 * sys_sched_setaffinity - set the cpu affinity of a process 4402 * @pid: pid of the process 4403 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 4404 * @user_mask_ptr: user-space pointer to the new cpu mask 4405 */ 4406SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len, 4407 unsigned long __user *, user_mask_ptr) 4408{ 4409 cpumask_var_t new_mask; 4410 int retval; 4411 4412 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) 4413 return -ENOMEM; 4414 4415 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask); 4416 if (retval == 0) 4417 retval = sched_setaffinity(pid, new_mask); 4418 free_cpumask_var(new_mask); 4419 return retval; 4420} 4421 4422long sched_getaffinity(pid_t pid, struct cpumask *mask) 4423{ 4424 struct task_struct *p; 4425 unsigned long flags; 4426 int retval; 4427 4428 get_online_cpus(); 4429 rcu_read_lock(); 4430 4431 retval = -ESRCH; 4432 p = find_process_by_pid(pid); 4433 if (!p) 4434 goto out_unlock; 4435 4436 retval = security_task_getscheduler(p); 4437 if (retval) 4438 goto out_unlock; 4439 4440 raw_spin_lock_irqsave(&p->pi_lock, flags); 4441 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask); 4442 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 4443 4444out_unlock: 4445 rcu_read_unlock(); 4446 put_online_cpus(); 4447 4448 return retval; 4449} 4450 4451/** 4452 * sys_sched_getaffinity - get the cpu affinity of a process 4453 * @pid: pid of the process 4454 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 4455 * @user_mask_ptr: user-space pointer to hold the current cpu mask 4456 */ 4457SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len, 4458 unsigned long __user *, user_mask_ptr) 4459{ 4460 int ret; 4461 cpumask_var_t mask; 4462 4463 if ((len * BITS_PER_BYTE) < nr_cpu_ids) 4464 return -EINVAL; 4465 if (len & (sizeof(unsigned long)-1)) 4466 return -EINVAL; 4467 4468 if (!alloc_cpumask_var(&mask, GFP_KERNEL)) 4469 return -ENOMEM; 4470 4471 ret = sched_getaffinity(pid, mask); 4472 if (ret == 0) { 4473 size_t retlen = min_t(size_t, len, cpumask_size()); 4474 4475 if (copy_to_user(user_mask_ptr, mask, retlen)) 4476 ret = -EFAULT; 4477 else 4478 ret = retlen; 4479 } 4480 free_cpumask_var(mask); 4481 4482 return ret; 4483} 4484 4485/** 4486 * sys_sched_yield - yield the current processor to other threads. 4487 * 4488 * This function yields the current CPU to other tasks. If there are no 4489 * other threads running on this CPU then this function will return. 4490 */ 4491SYSCALL_DEFINE0(sched_yield) 4492{ 4493 struct rq *rq = this_rq_lock(); 4494 4495 schedstat_inc(rq, yld_count); 4496 current->sched_class->yield_task(rq); 4497 4498 /* 4499 * Since we are going to call schedule() anyway, there's 4500 * no need to preempt or enable interrupts: 4501 */ 4502 __release(rq->lock); 4503 spin_release(&rq->lock.dep_map, 1, _THIS_IP_); 4504 do_raw_spin_unlock(&rq->lock); 4505 sched_preempt_enable_no_resched(); 4506 4507 schedule(); 4508 4509 return 0; 4510} 4511 4512static inline int should_resched(void) 4513{ 4514 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE); 4515} 4516 4517static void __cond_resched(void) 4518{ 4519 add_preempt_count(PREEMPT_ACTIVE); 4520 __schedule(); 4521 sub_preempt_count(PREEMPT_ACTIVE); 4522} 4523 4524int __sched _cond_resched(void) 4525{ 4526 if (should_resched()) { 4527 __cond_resched(); 4528 return 1; 4529 } 4530 return 0; 4531} 4532EXPORT_SYMBOL(_cond_resched); 4533 4534/* 4535 * __cond_resched_lock() - if a reschedule is pending, drop the given lock, 4536 * call schedule, and on return reacquire the lock. 4537 * 4538 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level 4539 * operations here to prevent schedule() from being called twice (once via 4540 * spin_unlock(), once by hand). 4541 */ 4542int __cond_resched_lock(spinlock_t *lock) 4543{ 4544 int resched = should_resched(); 4545 int ret = 0; 4546 4547 lockdep_assert_held(lock); 4548 4549 if (spin_needbreak(lock) || resched) { 4550 spin_unlock(lock); 4551 if (resched) 4552 __cond_resched(); 4553 else 4554 cpu_relax(); 4555 ret = 1; 4556 spin_lock(lock); 4557 } 4558 return ret; 4559} 4560EXPORT_SYMBOL(__cond_resched_lock); 4561 4562int __sched __cond_resched_softirq(void) 4563{ 4564 BUG_ON(!in_softirq()); 4565 4566 if (should_resched()) { 4567 local_bh_enable(); 4568 __cond_resched(); 4569 local_bh_disable(); 4570 return 1; 4571 } 4572 return 0; 4573} 4574EXPORT_SYMBOL(__cond_resched_softirq); 4575 4576/** 4577 * yield - yield the current processor to other threads. 4578 * 4579 * Do not ever use this function, there's a 99% chance you're doing it wrong. 4580 * 4581 * The scheduler is at all times free to pick the calling task as the most 4582 * eligible task to run, if removing the yield() call from your code breaks 4583 * it, its already broken. 4584 * 4585 * Typical broken usage is: 4586 * 4587 * while (!event) 4588 * yield(); 4589 * 4590 * where one assumes that yield() will let 'the other' process run that will 4591 * make event true. If the current task is a SCHED_FIFO task that will never 4592 * happen. Never use yield() as a progress guarantee!! 4593 * 4594 * If you want to use yield() to wait for something, use wait_event(). 4595 * If you want to use yield() to be 'nice' for others, use cond_resched(). 4596 * If you still want to use yield(), do not! 4597 */ 4598void __sched yield(void) 4599{ 4600 set_current_state(TASK_RUNNING); 4601 sys_sched_yield(); 4602} 4603EXPORT_SYMBOL(yield); 4604 4605/** 4606 * yield_to - yield the current processor to another thread in 4607 * your thread group, or accelerate that thread toward the 4608 * processor it's on. 4609 * @p: target task 4610 * @preempt: whether task preemption is allowed or not 4611 * 4612 * It's the caller's job to ensure that the target task struct 4613 * can't go away on us before we can do any checks. 4614 * 4615 * Returns true if we indeed boosted the target task. 4616 */ 4617bool __sched yield_to(struct task_struct *p, bool preempt) 4618{ 4619 struct task_struct *curr = current; 4620 struct rq *rq, *p_rq; 4621 unsigned long flags; 4622 bool yielded = 0; 4623 4624 local_irq_save(flags); 4625 rq = this_rq(); 4626 4627again: 4628 p_rq = task_rq(p); 4629 double_rq_lock(rq, p_rq); 4630 while (task_rq(p) != p_rq) { 4631 double_rq_unlock(rq, p_rq); 4632 goto again; 4633 } 4634 4635 if (!curr->sched_class->yield_to_task) 4636 goto out; 4637 4638 if (curr->sched_class != p->sched_class) 4639 goto out; 4640 4641 if (task_running(p_rq, p) || p->state) 4642 goto out; 4643 4644 yielded = curr->sched_class->yield_to_task(rq, p, preempt); 4645 if (yielded) { 4646 schedstat_inc(rq, yld_count); 4647 /* 4648 * Make p's CPU reschedule; pick_next_entity takes care of 4649 * fairness. 4650 */ 4651 if (preempt && rq != p_rq) 4652 resched_task(p_rq->curr); 4653 } else { 4654 /* 4655 * We might have set it in task_yield_fair(), but are 4656 * not going to schedule(), so don't want to skip 4657 * the next update. 4658 */ 4659 rq->skip_clock_update = 0; 4660 } 4661 4662out: 4663 double_rq_unlock(rq, p_rq); 4664 local_irq_restore(flags); 4665 4666 if (yielded) 4667 schedule(); 4668 4669 return yielded; 4670} 4671EXPORT_SYMBOL_GPL(yield_to); 4672 4673/* 4674 * This task is about to go to sleep on IO. Increment rq->nr_iowait so 4675 * that process accounting knows that this is a task in IO wait state. 4676 */ 4677void __sched io_schedule(void) 4678{ 4679 struct rq *rq = raw_rq(); 4680 4681 delayacct_blkio_start(); 4682 atomic_inc(&rq->nr_iowait); 4683 blk_flush_plug(current); 4684 current->in_iowait = 1; 4685 schedule(); 4686 current->in_iowait = 0; 4687 atomic_dec(&rq->nr_iowait); 4688 delayacct_blkio_end(); 4689} 4690EXPORT_SYMBOL(io_schedule); 4691 4692long __sched io_schedule_timeout(long timeout) 4693{ 4694 struct rq *rq = raw_rq(); 4695 long ret; 4696 4697 delayacct_blkio_start(); 4698 atomic_inc(&rq->nr_iowait); 4699 blk_flush_plug(current); 4700 current->in_iowait = 1; 4701 ret = schedule_timeout(timeout); 4702 current->in_iowait = 0; 4703 atomic_dec(&rq->nr_iowait); 4704 delayacct_blkio_end(); 4705 return ret; 4706} 4707 4708/** 4709 * sys_sched_get_priority_max - return maximum RT priority. 4710 * @policy: scheduling class. 4711 * 4712 * this syscall returns the maximum rt_priority that can be used 4713 * by a given scheduling class. 4714 */ 4715SYSCALL_DEFINE1(sched_get_priority_max, int, policy) 4716{ 4717 int ret = -EINVAL; 4718 4719 switch (policy) { 4720 case SCHED_FIFO: 4721 case SCHED_RR: 4722 ret = MAX_USER_RT_PRIO-1; 4723 break; 4724 case SCHED_NORMAL: 4725 case SCHED_BATCH: 4726 case SCHED_IDLE: 4727 ret = 0; 4728 break; 4729 } 4730 return ret; 4731} 4732 4733/** 4734 * sys_sched_get_priority_min - return minimum RT priority. 4735 * @policy: scheduling class. 4736 * 4737 * this syscall returns the minimum rt_priority that can be used 4738 * by a given scheduling class. 4739 */ 4740SYSCALL_DEFINE1(sched_get_priority_min, int, policy) 4741{ 4742 int ret = -EINVAL; 4743 4744 switch (policy) { 4745 case SCHED_FIFO: 4746 case SCHED_RR: 4747 ret = 1; 4748 break; 4749 case SCHED_NORMAL: 4750 case SCHED_BATCH: 4751 case SCHED_IDLE: 4752 ret = 0; 4753 } 4754 return ret; 4755} 4756 4757/** 4758 * sys_sched_rr_get_interval - return the default timeslice of a process. 4759 * @pid: pid of the process. 4760 * @interval: userspace pointer to the timeslice value. 4761 * 4762 * this syscall writes the default timeslice value of a given process 4763 * into the user-space timespec buffer. A value of '0' means infinity. 4764 */ 4765SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid, 4766 struct timespec __user *, interval) 4767{ 4768 struct task_struct *p; 4769 unsigned int time_slice; 4770 unsigned long flags; 4771 struct rq *rq; 4772 int retval; 4773 struct timespec t; 4774 4775 if (pid < 0) 4776 return -EINVAL; 4777 4778 retval = -ESRCH; 4779 rcu_read_lock(); 4780 p = find_process_by_pid(pid); 4781 if (!p) 4782 goto out_unlock; 4783 4784 retval = security_task_getscheduler(p); 4785 if (retval) 4786 goto out_unlock; 4787 4788 rq = task_rq_lock(p, &flags); 4789 time_slice = p->sched_class->get_rr_interval(rq, p); 4790 task_rq_unlock(rq, p, &flags); 4791 4792 rcu_read_unlock(); 4793 jiffies_to_timespec(time_slice, &t); 4794 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0; 4795 return retval; 4796 4797out_unlock: 4798 rcu_read_unlock(); 4799 return retval; 4800} 4801 4802static const char stat_nam[] = TASK_STATE_TO_CHAR_STR; 4803 4804void sched_show_task(struct task_struct *p) 4805{ 4806 unsigned long free = 0; 4807 unsigned state; 4808 4809 state = p->state ? __ffs(p->state) + 1 : 0; 4810 printk(KERN_INFO "%-15.15s %c", p->comm, 4811 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?'); 4812#if BITS_PER_LONG == 32 4813 if (state == TASK_RUNNING) 4814 printk(KERN_CONT " running "); 4815 else 4816 printk(KERN_CONT " %08lx ", thread_saved_pc(p)); 4817#else 4818 if (state == TASK_RUNNING) 4819 printk(KERN_CONT " running task "); 4820 else 4821 printk(KERN_CONT " %016lx ", thread_saved_pc(p)); 4822#endif 4823#ifdef CONFIG_DEBUG_STACK_USAGE 4824 free = stack_not_used(p); 4825#endif 4826 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free, 4827 task_pid_nr(p), task_pid_nr(rcu_dereference(p->real_parent)), 4828 (unsigned long)task_thread_info(p)->flags); 4829 4830 show_stack(p, NULL); 4831} 4832 4833void show_state_filter(unsigned long state_filter) 4834{ 4835 struct task_struct *g, *p; 4836 4837#if BITS_PER_LONG == 32 4838 printk(KERN_INFO 4839 " task PC stack pid father\n"); 4840#else 4841 printk(KERN_INFO 4842 " task PC stack pid father\n"); 4843#endif 4844 rcu_read_lock(); 4845 do_each_thread(g, p) { 4846 /* 4847 * reset the NMI-timeout, listing all files on a slow 4848 * console might take a lot of time: 4849 */ 4850 touch_nmi_watchdog(); 4851 if (!state_filter || (p->state & state_filter)) 4852 sched_show_task(p); 4853 } while_each_thread(g, p); 4854 4855 touch_all_softlockup_watchdogs(); 4856 4857#ifdef CONFIG_SCHED_DEBUG 4858 sysrq_sched_debug_show(); 4859#endif 4860 rcu_read_unlock(); 4861 /* 4862 * Only show locks if all tasks are dumped: 4863 */ 4864 if (!state_filter) 4865 debug_show_all_locks(); 4866} 4867 4868void __cpuinit init_idle_bootup_task(struct task_struct *idle) 4869{ 4870 idle->sched_class = &idle_sched_class; 4871} 4872 4873/** 4874 * init_idle - set up an idle thread for a given CPU 4875 * @idle: task in question 4876 * @cpu: cpu the idle task belongs to 4877 * 4878 * NOTE: this function does not set the idle thread's NEED_RESCHED 4879 * flag, to make booting more robust. 4880 */ 4881void __cpuinit init_idle(struct task_struct *idle, int cpu) 4882{ 4883 struct rq *rq = cpu_rq(cpu); 4884 unsigned long flags; 4885 4886 raw_spin_lock_irqsave(&rq->lock, flags); 4887 4888 __sched_fork(idle); 4889 idle->state = TASK_RUNNING; 4890 idle->se.exec_start = sched_clock(); 4891 4892 do_set_cpus_allowed(idle, cpumask_of(cpu)); 4893 /* 4894 * We're having a chicken and egg problem, even though we are 4895 * holding rq->lock, the cpu isn't yet set to this cpu so the 4896 * lockdep check in task_group() will fail. 4897 * 4898 * Similar case to sched_fork(). / Alternatively we could 4899 * use task_rq_lock() here and obtain the other rq->lock. 4900 * 4901 * Silence PROVE_RCU 4902 */ 4903 rcu_read_lock(); 4904 __set_task_cpu(idle, cpu); 4905 rcu_read_unlock(); 4906 4907 rq->curr = rq->idle = idle; 4908#if defined(CONFIG_SMP) 4909 idle->on_cpu = 1; 4910#endif 4911 raw_spin_unlock_irqrestore(&rq->lock, flags); 4912 4913 /* Set the preempt count _outside_ the spinlocks! */ 4914 task_thread_info(idle)->preempt_count = 0; 4915 4916 /* 4917 * The idle tasks have their own, simple scheduling class: 4918 */ 4919 idle->sched_class = &idle_sched_class; 4920 ftrace_graph_init_idle_task(idle, cpu); 4921#if defined(CONFIG_SMP) 4922 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu); 4923#endif 4924} 4925 4926#ifdef CONFIG_SMP 4927void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask) 4928{ 4929 if (p->sched_class && p->sched_class->set_cpus_allowed) 4930 p->sched_class->set_cpus_allowed(p, new_mask); 4931 4932 cpumask_copy(&p->cpus_allowed, new_mask); 4933 p->rt.nr_cpus_allowed = cpumask_weight(new_mask); 4934} 4935 4936/* 4937 * This is how migration works: 4938 * 4939 * 1) we invoke migration_cpu_stop() on the target CPU using 4940 * stop_one_cpu(). 4941 * 2) stopper starts to run (implicitly forcing the migrated thread 4942 * off the CPU) 4943 * 3) it checks whether the migrated task is still in the wrong runqueue. 4944 * 4) if it's in the wrong runqueue then the migration thread removes 4945 * it and puts it into the right queue. 4946 * 5) stopper completes and stop_one_cpu() returns and the migration 4947 * is done. 4948 */ 4949 4950/* 4951 * Change a given task's CPU affinity. Migrate the thread to a 4952 * proper CPU and schedule it away if the CPU it's executing on 4953 * is removed from the allowed bitmask. 4954 * 4955 * NOTE: the caller must have a valid reference to the task, the 4956 * task must not exit() & deallocate itself prematurely. The 4957 * call is not atomic; no spinlocks may be held. 4958 */ 4959int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask) 4960{ 4961 unsigned long flags; 4962 struct rq *rq; 4963 unsigned int dest_cpu; 4964 int ret = 0; 4965 4966 rq = task_rq_lock(p, &flags); 4967 4968 if (cpumask_equal(&p->cpus_allowed, new_mask)) 4969 goto out; 4970 4971 if (!cpumask_intersects(new_mask, cpu_active_mask)) { 4972 ret = -EINVAL; 4973 goto out; 4974 } 4975 4976 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current)) { 4977 ret = -EINVAL; 4978 goto out; 4979 } 4980 4981 do_set_cpus_allowed(p, new_mask); 4982 4983 /* Can the task run on the task's current CPU? If so, we're done */ 4984 if (cpumask_test_cpu(task_cpu(p), new_mask)) 4985 goto out; 4986 4987 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask); 4988 if (p->on_rq) { 4989 struct migration_arg arg = { p, dest_cpu }; 4990 /* Need help from migration thread: drop lock and wait. */ 4991 task_rq_unlock(rq, p, &flags); 4992 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg); 4993 tlb_migrate_finish(p->mm); 4994 return 0; 4995 } 4996out: 4997 task_rq_unlock(rq, p, &flags); 4998 4999 return ret; 5000} 5001EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr); 5002 5003/* 5004 * Move (not current) task off this cpu, onto dest cpu. We're doing 5005 * this because either it can't run here any more (set_cpus_allowed() 5006 * away from this CPU, or CPU going down), or because we're 5007 * attempting to rebalance this task on exec (sched_exec). 5008 * 5009 * So we race with normal scheduler movements, but that's OK, as long 5010 * as the task is no longer on this CPU. 5011 * 5012 * Returns non-zero if task was successfully migrated. 5013 */ 5014static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu) 5015{ 5016 struct rq *rq_dest, *rq_src; 5017 int ret = 0; 5018 5019 if (unlikely(!cpu_active(dest_cpu))) 5020 return ret; 5021 5022 rq_src = cpu_rq(src_cpu); 5023 rq_dest = cpu_rq(dest_cpu); 5024 5025 raw_spin_lock(&p->pi_lock); 5026 double_rq_lock(rq_src, rq_dest); 5027 /* Already moved. */ 5028 if (task_cpu(p) != src_cpu) 5029 goto done; 5030 /* Affinity changed (again). */ 5031 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) 5032 goto fail; 5033 5034 /* 5035 * If we're not on a rq, the next wake-up will ensure we're 5036 * placed properly. 5037 */ 5038 if (p->on_rq) { 5039 dequeue_task(rq_src, p, 0); 5040 set_task_cpu(p, dest_cpu); 5041 enqueue_task(rq_dest, p, 0); 5042 check_preempt_curr(rq_dest, p, 0); 5043 } 5044done: 5045 ret = 1; 5046fail: 5047 double_rq_unlock(rq_src, rq_dest); 5048 raw_spin_unlock(&p->pi_lock); 5049 return ret; 5050} 5051 5052/* 5053 * migration_cpu_stop - this will be executed by a highprio stopper thread 5054 * and performs thread migration by bumping thread off CPU then 5055 * 'pushing' onto another runqueue. 5056 */ 5057static int migration_cpu_stop(void *data) 5058{ 5059 struct migration_arg *arg = data; 5060 5061 /* 5062 * The original target cpu might have gone down and we might 5063 * be on another cpu but it doesn't matter. 5064 */ 5065 local_irq_disable(); 5066 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu); 5067 local_irq_enable(); 5068 return 0; 5069} 5070 5071#ifdef CONFIG_HOTPLUG_CPU 5072 5073/* 5074 * Ensures that the idle task is using init_mm right before its cpu goes 5075 * offline. 5076 */ 5077void idle_task_exit(void) 5078{ 5079 struct mm_struct *mm = current->active_mm; 5080 5081 BUG_ON(cpu_online(smp_processor_id())); 5082 5083 if (mm != &init_mm) 5084 switch_mm(mm, &init_mm, current); 5085 mmdrop(mm); 5086} 5087 5088/* 5089 * While a dead CPU has no uninterruptible tasks queued at this point, 5090 * it might still have a nonzero ->nr_uninterruptible counter, because 5091 * for performance reasons the counter is not stricly tracking tasks to 5092 * their home CPUs. So we just add the counter to another CPU's counter, 5093 * to keep the global sum constant after CPU-down: 5094 */ 5095static void migrate_nr_uninterruptible(struct rq *rq_src) 5096{ 5097 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask)); 5098 5099 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible; 5100 rq_src->nr_uninterruptible = 0; 5101} 5102 5103/* 5104 * remove the tasks which were accounted by rq from calc_load_tasks. 5105 */ 5106static void calc_global_load_remove(struct rq *rq) 5107{ 5108 atomic_long_sub(rq->calc_load_active, &calc_load_tasks); 5109 rq->calc_load_active = 0; 5110} 5111 5112/* 5113 * Migrate all tasks from the rq, sleeping tasks will be migrated by 5114 * try_to_wake_up()->select_task_rq(). 5115 * 5116 * Called with rq->lock held even though we'er in stop_machine() and 5117 * there's no concurrency possible, we hold the required locks anyway 5118 * because of lock validation efforts. 5119 */ 5120static void migrate_tasks(unsigned int dead_cpu) 5121{ 5122 struct rq *rq = cpu_rq(dead_cpu); 5123 struct task_struct *next, *stop = rq->stop; 5124 int dest_cpu; 5125 5126 /* 5127 * Fudge the rq selection such that the below task selection loop 5128 * doesn't get stuck on the currently eligible stop task. 5129 * 5130 * We're currently inside stop_machine() and the rq is either stuck 5131 * in the stop_machine_cpu_stop() loop, or we're executing this code, 5132 * either way we should never end up calling schedule() until we're 5133 * done here. 5134 */ 5135 rq->stop = NULL; 5136 5137 /* Ensure any throttled groups are reachable by pick_next_task */ 5138 unthrottle_offline_cfs_rqs(rq); 5139 5140 for ( ; ; ) { 5141 /* 5142 * There's this thread running, bail when that's the only 5143 * remaining thread. 5144 */ 5145 if (rq->nr_running == 1) 5146 break; 5147 5148 next = pick_next_task(rq); 5149 BUG_ON(!next); 5150 next->sched_class->put_prev_task(rq, next); 5151 5152 /* Find suitable destination for @next, with force if needed. */ 5153 dest_cpu = select_fallback_rq(dead_cpu, next); 5154 raw_spin_unlock(&rq->lock); 5155 5156 __migrate_task(next, dead_cpu, dest_cpu); 5157 5158 raw_spin_lock(&rq->lock); 5159 } 5160 5161 rq->stop = stop; 5162} 5163 5164#endif /* CONFIG_HOTPLUG_CPU */ 5165 5166#if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL) 5167 5168static struct ctl_table sd_ctl_dir[] = { 5169 { 5170 .procname = "sched_domain", 5171 .mode = 0555, 5172 }, 5173 {} 5174}; 5175 5176static struct ctl_table sd_ctl_root[] = { 5177 { 5178 .procname = "kernel", 5179 .mode = 0555, 5180 .child = sd_ctl_dir, 5181 }, 5182 {} 5183}; 5184 5185static struct ctl_table *sd_alloc_ctl_entry(int n) 5186{ 5187 struct ctl_table *entry = 5188 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL); 5189 5190 return entry; 5191} 5192 5193static void sd_free_ctl_entry(struct ctl_table **tablep) 5194{ 5195 struct ctl_table *entry; 5196 5197 /* 5198 * In the intermediate directories, both the child directory and 5199 * procname are dynamically allocated and could fail but the mode 5200 * will always be set. In the lowest directory the names are 5201 * static strings and all have proc handlers. 5202 */ 5203 for (entry = *tablep; entry->mode; entry++) { 5204 if (entry->child) 5205 sd_free_ctl_entry(&entry->child); 5206 if (entry->proc_handler == NULL) 5207 kfree(entry->procname); 5208 } 5209 5210 kfree(*tablep); 5211 *tablep = NULL; 5212} 5213 5214static void 5215set_table_entry(struct ctl_table *entry, 5216 const char *procname, void *data, int maxlen, 5217 umode_t mode, proc_handler *proc_handler) 5218{ 5219 entry->procname = procname; 5220 entry->data = data; 5221 entry->maxlen = maxlen; 5222 entry->mode = mode; 5223 entry->proc_handler = proc_handler; 5224} 5225 5226static struct ctl_table * 5227sd_alloc_ctl_domain_table(struct sched_domain *sd) 5228{ 5229 struct ctl_table *table = sd_alloc_ctl_entry(13); 5230 5231 if (table == NULL) 5232 return NULL; 5233 5234 set_table_entry(&table[0], "min_interval", &sd->min_interval, 5235 sizeof(long), 0644, proc_doulongvec_minmax); 5236 set_table_entry(&table[1], "max_interval", &sd->max_interval, 5237 sizeof(long), 0644, proc_doulongvec_minmax); 5238 set_table_entry(&table[2], "busy_idx", &sd->busy_idx, 5239 sizeof(int), 0644, proc_dointvec_minmax); 5240 set_table_entry(&table[3], "idle_idx", &sd->idle_idx, 5241 sizeof(int), 0644, proc_dointvec_minmax); 5242 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx, 5243 sizeof(int), 0644, proc_dointvec_minmax); 5244 set_table_entry(&table[5], "wake_idx", &sd->wake_idx, 5245 sizeof(int), 0644, proc_dointvec_minmax); 5246 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx, 5247 sizeof(int), 0644, proc_dointvec_minmax); 5248 set_table_entry(&table[7], "busy_factor", &sd->busy_factor, 5249 sizeof(int), 0644, proc_dointvec_minmax); 5250 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct, 5251 sizeof(int), 0644, proc_dointvec_minmax); 5252 set_table_entry(&table[9], "cache_nice_tries", 5253 &sd->cache_nice_tries, 5254 sizeof(int), 0644, proc_dointvec_minmax); 5255 set_table_entry(&table[10], "flags", &sd->flags, 5256 sizeof(int), 0644, proc_dointvec_minmax); 5257 set_table_entry(&table[11], "name", sd->name, 5258 CORENAME_MAX_SIZE, 0444, proc_dostring); 5259 /* &table[12] is terminator */ 5260 5261 return table; 5262} 5263 5264static ctl_table *sd_alloc_ctl_cpu_table(int cpu) 5265{ 5266 struct ctl_table *entry, *table; 5267 struct sched_domain *sd; 5268 int domain_num = 0, i; 5269 char buf[32]; 5270 5271 for_each_domain(cpu, sd) 5272 domain_num++; 5273 entry = table = sd_alloc_ctl_entry(domain_num + 1); 5274 if (table == NULL) 5275 return NULL; 5276 5277 i = 0; 5278 for_each_domain(cpu, sd) { 5279 snprintf(buf, 32, "domain%d", i); 5280 entry->procname = kstrdup(buf, GFP_KERNEL); 5281 entry->mode = 0555; 5282 entry->child = sd_alloc_ctl_domain_table(sd); 5283 entry++; 5284 i++; 5285 } 5286 return table; 5287} 5288 5289static struct ctl_table_header *sd_sysctl_header; 5290static void register_sched_domain_sysctl(void) 5291{ 5292 int i, cpu_num = num_possible_cpus(); 5293 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1); 5294 char buf[32]; 5295 5296 WARN_ON(sd_ctl_dir[0].child); 5297 sd_ctl_dir[0].child = entry; 5298 5299 if (entry == NULL) 5300 return; 5301 5302 for_each_possible_cpu(i) { 5303 snprintf(buf, 32, "cpu%d", i); 5304 entry->procname = kstrdup(buf, GFP_KERNEL); 5305 entry->mode = 0555; 5306 entry->child = sd_alloc_ctl_cpu_table(i); 5307 entry++; 5308 } 5309 5310 WARN_ON(sd_sysctl_header); 5311 sd_sysctl_header = register_sysctl_table(sd_ctl_root); 5312} 5313 5314/* may be called multiple times per register */ 5315static void unregister_sched_domain_sysctl(void) 5316{ 5317 if (sd_sysctl_header) 5318 unregister_sysctl_table(sd_sysctl_header); 5319 sd_sysctl_header = NULL; 5320 if (sd_ctl_dir[0].child) 5321 sd_free_ctl_entry(&sd_ctl_dir[0].child); 5322} 5323#else 5324static void register_sched_domain_sysctl(void) 5325{ 5326} 5327static void unregister_sched_domain_sysctl(void) 5328{ 5329} 5330#endif 5331 5332static void set_rq_online(struct rq *rq) 5333{ 5334 if (!rq->online) { 5335 const struct sched_class *class; 5336 5337 cpumask_set_cpu(rq->cpu, rq->rd->online); 5338 rq->online = 1; 5339 5340 for_each_class(class) { 5341 if (class->rq_online) 5342 class->rq_online(rq); 5343 } 5344 } 5345} 5346 5347static void set_rq_offline(struct rq *rq) 5348{ 5349 if (rq->online) { 5350 const struct sched_class *class; 5351 5352 for_each_class(class) { 5353 if (class->rq_offline) 5354 class->rq_offline(rq); 5355 } 5356 5357 cpumask_clear_cpu(rq->cpu, rq->rd->online); 5358 rq->online = 0; 5359 } 5360} 5361 5362/* 5363 * migration_call - callback that gets triggered when a CPU is added. 5364 * Here we can start up the necessary migration thread for the new CPU. 5365 */ 5366static int __cpuinit 5367migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu) 5368{ 5369 int cpu = (long)hcpu; 5370 unsigned long flags; 5371 struct rq *rq = cpu_rq(cpu); 5372 5373 switch (action & ~CPU_TASKS_FROZEN) { 5374 5375 case CPU_UP_PREPARE: 5376 rq->calc_load_update = calc_load_update; 5377 break; 5378 5379 case CPU_ONLINE: 5380 /* Update our root-domain */ 5381 raw_spin_lock_irqsave(&rq->lock, flags); 5382 if (rq->rd) { 5383 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 5384 5385 set_rq_online(rq); 5386 } 5387 raw_spin_unlock_irqrestore(&rq->lock, flags); 5388 break; 5389 5390#ifdef CONFIG_HOTPLUG_CPU 5391 case CPU_DYING: 5392 sched_ttwu_pending(); 5393 /* Update our root-domain */ 5394 raw_spin_lock_irqsave(&rq->lock, flags); 5395 if (rq->rd) { 5396 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 5397 set_rq_offline(rq); 5398 } 5399 migrate_tasks(cpu); 5400 BUG_ON(rq->nr_running != 1); /* the migration thread */ 5401 raw_spin_unlock_irqrestore(&rq->lock, flags); 5402 5403 migrate_nr_uninterruptible(rq); 5404 calc_global_load_remove(rq); 5405 break; 5406#endif 5407 } 5408 5409 update_max_interval(); 5410 5411 return NOTIFY_OK; 5412} 5413 5414/* 5415 * Register at high priority so that task migration (migrate_all_tasks) 5416 * happens before everything else. This has to be lower priority than 5417 * the notifier in the perf_event subsystem, though. 5418 */ 5419static struct notifier_block __cpuinitdata migration_notifier = { 5420 .notifier_call = migration_call, 5421 .priority = CPU_PRI_MIGRATION, 5422}; 5423 5424static int __cpuinit sched_cpu_active(struct notifier_block *nfb, 5425 unsigned long action, void *hcpu) 5426{ 5427 switch (action & ~CPU_TASKS_FROZEN) { 5428 case CPU_STARTING: 5429 case CPU_DOWN_FAILED: 5430 set_cpu_active((long)hcpu, true); 5431 return NOTIFY_OK; 5432 default: 5433 return NOTIFY_DONE; 5434 } 5435} 5436 5437static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb, 5438 unsigned long action, void *hcpu) 5439{ 5440 switch (action & ~CPU_TASKS_FROZEN) { 5441 case CPU_DOWN_PREPARE: 5442 set_cpu_active((long)hcpu, false); 5443 return NOTIFY_OK; 5444 default: 5445 return NOTIFY_DONE; 5446 } 5447} 5448 5449static int __init migration_init(void) 5450{ 5451 void *cpu = (void *)(long)smp_processor_id(); 5452 int err; 5453 5454 /* Initialize migration for the boot CPU */ 5455 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu); 5456 BUG_ON(err == NOTIFY_BAD); 5457 migration_call(&migration_notifier, CPU_ONLINE, cpu); 5458 register_cpu_notifier(&migration_notifier); 5459 5460 /* Register cpu active notifiers */ 5461 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE); 5462 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE); 5463 5464 return 0; 5465} 5466early_initcall(migration_init); 5467#endif 5468 5469#ifdef CONFIG_SMP 5470 5471static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */ 5472 5473#ifdef CONFIG_SCHED_DEBUG 5474 5475static __read_mostly int sched_domain_debug_enabled; 5476 5477static int __init sched_domain_debug_setup(char *str) 5478{ 5479 sched_domain_debug_enabled = 1; 5480 5481 return 0; 5482} 5483early_param("sched_debug", sched_domain_debug_setup); 5484 5485static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level, 5486 struct cpumask *groupmask) 5487{ 5488 struct sched_group *group = sd->groups; 5489 char str[256]; 5490 5491 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd)); 5492 cpumask_clear(groupmask); 5493 5494 printk(KERN_DEBUG "%*s domain %d: ", level, "", level); 5495 5496 if (!(sd->flags & SD_LOAD_BALANCE)) { 5497 printk("does not load-balance\n"); 5498 if (sd->parent) 5499 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain" 5500 " has parent"); 5501 return -1; 5502 } 5503 5504 printk(KERN_CONT "span %s level %s\n", str, sd->name); 5505 5506 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) { 5507 printk(KERN_ERR "ERROR: domain->span does not contain " 5508 "CPU%d\n", cpu); 5509 } 5510 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) { 5511 printk(KERN_ERR "ERROR: domain->groups does not contain" 5512 " CPU%d\n", cpu); 5513 } 5514 5515 printk(KERN_DEBUG "%*s groups:", level + 1, ""); 5516 do { 5517 if (!group) { 5518 printk("\n"); 5519 printk(KERN_ERR "ERROR: group is NULL\n"); 5520 break; 5521 } 5522 5523 if (!group->sgp->power) { 5524 printk(KERN_CONT "\n"); 5525 printk(KERN_ERR "ERROR: domain->cpu_power not " 5526 "set\n"); 5527 break; 5528 } 5529 5530 if (!cpumask_weight(sched_group_cpus(group))) { 5531 printk(KERN_CONT "\n"); 5532 printk(KERN_ERR "ERROR: empty group\n"); 5533 break; 5534 } 5535 5536 if (cpumask_intersects(groupmask, sched_group_cpus(group))) { 5537 printk(KERN_CONT "\n"); 5538 printk(KERN_ERR "ERROR: repeated CPUs\n"); 5539 break; 5540 } 5541 5542 cpumask_or(groupmask, groupmask, sched_group_cpus(group)); 5543 5544 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group)); 5545 5546 printk(KERN_CONT " %s", str); 5547 if (group->sgp->power != SCHED_POWER_SCALE) { 5548 printk(KERN_CONT " (cpu_power = %d)", 5549 group->sgp->power); 5550 } 5551 5552 group = group->next; 5553 } while (group != sd->groups); 5554 printk(KERN_CONT "\n"); 5555 5556 if (!cpumask_equal(sched_domain_span(sd), groupmask)) 5557 printk(KERN_ERR "ERROR: groups don't span domain->span\n"); 5558 5559 if (sd->parent && 5560 !cpumask_subset(groupmask, sched_domain_span(sd->parent))) 5561 printk(KERN_ERR "ERROR: parent span is not a superset " 5562 "of domain->span\n"); 5563 return 0; 5564} 5565 5566static void sched_domain_debug(struct sched_domain *sd, int cpu) 5567{ 5568 int level = 0; 5569 5570 if (!sched_domain_debug_enabled) 5571 return; 5572 5573 if (!sd) { 5574 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu); 5575 return; 5576 } 5577 5578 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu); 5579 5580 for (;;) { 5581 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask)) 5582 break; 5583 level++; 5584 sd = sd->parent; 5585 if (!sd) 5586 break; 5587 } 5588} 5589#else /* !CONFIG_SCHED_DEBUG */ 5590# define sched_domain_debug(sd, cpu) do { } while (0) 5591#endif /* CONFIG_SCHED_DEBUG */ 5592 5593static int sd_degenerate(struct sched_domain *sd) 5594{ 5595 if (cpumask_weight(sched_domain_span(sd)) == 1) 5596 return 1; 5597 5598 /* Following flags need at least 2 groups */ 5599 if (sd->flags & (SD_LOAD_BALANCE | 5600 SD_BALANCE_NEWIDLE | 5601 SD_BALANCE_FORK | 5602 SD_BALANCE_EXEC | 5603 SD_SHARE_CPUPOWER | 5604 SD_SHARE_PKG_RESOURCES)) { 5605 if (sd->groups != sd->groups->next) 5606 return 0; 5607 } 5608 5609 /* Following flags don't use groups */ 5610 if (sd->flags & (SD_WAKE_AFFINE)) 5611 return 0; 5612 5613 return 1; 5614} 5615 5616static int 5617sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent) 5618{ 5619 unsigned long cflags = sd->flags, pflags = parent->flags; 5620 5621 if (sd_degenerate(parent)) 5622 return 1; 5623 5624 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent))) 5625 return 0; 5626 5627 /* Flags needing groups don't count if only 1 group in parent */ 5628 if (parent->groups == parent->groups->next) { 5629 pflags &= ~(SD_LOAD_BALANCE | 5630 SD_BALANCE_NEWIDLE | 5631 SD_BALANCE_FORK | 5632 SD_BALANCE_EXEC | 5633 SD_SHARE_CPUPOWER | 5634 SD_SHARE_PKG_RESOURCES); 5635 if (nr_node_ids == 1) 5636 pflags &= ~SD_SERIALIZE; 5637 } 5638 if (~cflags & pflags) 5639 return 0; 5640 5641 return 1; 5642} 5643 5644static void free_rootdomain(struct rcu_head *rcu) 5645{ 5646 struct root_domain *rd = container_of(rcu, struct root_domain, rcu); 5647 5648 cpupri_cleanup(&rd->cpupri); 5649 free_cpumask_var(rd->rto_mask); 5650 free_cpumask_var(rd->online); 5651 free_cpumask_var(rd->span); 5652 kfree(rd); 5653} 5654 5655static void rq_attach_root(struct rq *rq, struct root_domain *rd) 5656{ 5657 struct root_domain *old_rd = NULL; 5658 unsigned long flags; 5659 5660 raw_spin_lock_irqsave(&rq->lock, flags); 5661 5662 if (rq->rd) { 5663 old_rd = rq->rd; 5664 5665 if (cpumask_test_cpu(rq->cpu, old_rd->online)) 5666 set_rq_offline(rq); 5667 5668 cpumask_clear_cpu(rq->cpu, old_rd->span); 5669 5670 /* 5671 * If we dont want to free the old_rt yet then 5672 * set old_rd to NULL to skip the freeing later 5673 * in this function: 5674 */ 5675 if (!atomic_dec_and_test(&old_rd->refcount)) 5676 old_rd = NULL; 5677 } 5678 5679 atomic_inc(&rd->refcount); 5680 rq->rd = rd; 5681 5682 cpumask_set_cpu(rq->cpu, rd->span); 5683 if (cpumask_test_cpu(rq->cpu, cpu_active_mask)) 5684 set_rq_online(rq); 5685 5686 raw_spin_unlock_irqrestore(&rq->lock, flags); 5687 5688 if (old_rd) 5689 call_rcu_sched(&old_rd->rcu, free_rootdomain); 5690} 5691 5692static int init_rootdomain(struct root_domain *rd) 5693{ 5694 memset(rd, 0, sizeof(*rd)); 5695 5696 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL)) 5697 goto out; 5698 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL)) 5699 goto free_span; 5700 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL)) 5701 goto free_online; 5702 5703 if (cpupri_init(&rd->cpupri) != 0) 5704 goto free_rto_mask; 5705 return 0; 5706 5707free_rto_mask: 5708 free_cpumask_var(rd->rto_mask); 5709free_online: 5710 free_cpumask_var(rd->online); 5711free_span: 5712 free_cpumask_var(rd->span); 5713out: 5714 return -ENOMEM; 5715} 5716 5717/* 5718 * By default the system creates a single root-domain with all cpus as 5719 * members (mimicking the global state we have today). 5720 */ 5721struct root_domain def_root_domain; 5722 5723static void init_defrootdomain(void) 5724{ 5725 init_rootdomain(&def_root_domain); 5726 5727 atomic_set(&def_root_domain.refcount, 1); 5728} 5729 5730static struct root_domain *alloc_rootdomain(void) 5731{ 5732 struct root_domain *rd; 5733 5734 rd = kmalloc(sizeof(*rd), GFP_KERNEL); 5735 if (!rd) 5736 return NULL; 5737 5738 if (init_rootdomain(rd) != 0) { 5739 kfree(rd); 5740 return NULL; 5741 } 5742 5743 return rd; 5744} 5745 5746static void free_sched_groups(struct sched_group *sg, int free_sgp) 5747{ 5748 struct sched_group *tmp, *first; 5749 5750 if (!sg) 5751 return; 5752 5753 first = sg; 5754 do { 5755 tmp = sg->next; 5756 5757 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref)) 5758 kfree(sg->sgp); 5759 5760 kfree(sg); 5761 sg = tmp; 5762 } while (sg != first); 5763} 5764 5765static void free_sched_domain(struct rcu_head *rcu) 5766{ 5767 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu); 5768 5769 /* 5770 * If its an overlapping domain it has private groups, iterate and 5771 * nuke them all. 5772 */ 5773 if (sd->flags & SD_OVERLAP) { 5774 free_sched_groups(sd->groups, 1); 5775 } else if (atomic_dec_and_test(&sd->groups->ref)) { 5776 kfree(sd->groups->sgp); 5777 kfree(sd->groups); 5778 } 5779 kfree(sd); 5780} 5781 5782static void destroy_sched_domain(struct sched_domain *sd, int cpu) 5783{ 5784 call_rcu(&sd->rcu, free_sched_domain); 5785} 5786 5787static void destroy_sched_domains(struct sched_domain *sd, int cpu) 5788{ 5789 for (; sd; sd = sd->parent) 5790 destroy_sched_domain(sd, cpu); 5791} 5792 5793/* 5794 * Keep a special pointer to the highest sched_domain that has 5795 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this 5796 * allows us to avoid some pointer chasing select_idle_sibling(). 5797 * 5798 * Also keep a unique ID per domain (we use the first cpu number in 5799 * the cpumask of the domain), this allows us to quickly tell if 5800 * two cpus are in the same cache domain, see cpus_share_cache(). 5801 */ 5802DEFINE_PER_CPU(struct sched_domain *, sd_llc); 5803DEFINE_PER_CPU(int, sd_llc_id); 5804 5805static void update_top_cache_domain(int cpu) 5806{ 5807 struct sched_domain *sd; 5808 int id = cpu; 5809 5810 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES); 5811 if (sd) 5812 id = cpumask_first(sched_domain_span(sd)); 5813 5814 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd); 5815 per_cpu(sd_llc_id, cpu) = id; 5816} 5817 5818/* 5819 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must 5820 * hold the hotplug lock. 5821 */ 5822static void 5823cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu) 5824{ 5825 struct rq *rq = cpu_rq(cpu); 5826 struct sched_domain *tmp; 5827 5828 /* Remove the sched domains which do not contribute to scheduling. */ 5829 for (tmp = sd; tmp; ) { 5830 struct sched_domain *parent = tmp->parent; 5831 if (!parent) 5832 break; 5833 5834 if (sd_parent_degenerate(tmp, parent)) { 5835 tmp->parent = parent->parent; 5836 if (parent->parent) 5837 parent->parent->child = tmp; 5838 destroy_sched_domain(parent, cpu); 5839 } else 5840 tmp = tmp->parent; 5841 } 5842 5843 if (sd && sd_degenerate(sd)) { 5844 tmp = sd; 5845 sd = sd->parent; 5846 destroy_sched_domain(tmp, cpu); 5847 if (sd) 5848 sd->child = NULL; 5849 } 5850 5851 sched_domain_debug(sd, cpu); 5852 5853 rq_attach_root(rq, rd); 5854 tmp = rq->sd; 5855 rcu_assign_pointer(rq->sd, sd); 5856 destroy_sched_domains(tmp, cpu); 5857 5858 update_top_cache_domain(cpu); 5859} 5860 5861/* cpus with isolated domains */ 5862static cpumask_var_t cpu_isolated_map; 5863 5864/* Setup the mask of cpus configured for isolated domains */ 5865static int __init isolated_cpu_setup(char *str) 5866{ 5867 alloc_bootmem_cpumask_var(&cpu_isolated_map); 5868 cpulist_parse(str, cpu_isolated_map); 5869 return 1; 5870} 5871 5872__setup("isolcpus=", isolated_cpu_setup); 5873 5874#ifdef CONFIG_NUMA 5875 5876/** 5877 * find_next_best_node - find the next node to include in a sched_domain 5878 * @node: node whose sched_domain we're building 5879 * @used_nodes: nodes already in the sched_domain 5880 * 5881 * Find the next node to include in a given scheduling domain. Simply 5882 * finds the closest node not already in the @used_nodes map. 5883 * 5884 * Should use nodemask_t. 5885 */ 5886static int find_next_best_node(int node, nodemask_t *used_nodes) 5887{ 5888 int i, n, val, min_val, best_node = -1; 5889 5890 min_val = INT_MAX; 5891 5892 for (i = 0; i < nr_node_ids; i++) { 5893 /* Start at @node */ 5894 n = (node + i) % nr_node_ids; 5895 5896 if (!nr_cpus_node(n)) 5897 continue; 5898 5899 /* Skip already used nodes */ 5900 if (node_isset(n, *used_nodes)) 5901 continue; 5902 5903 /* Simple min distance search */ 5904 val = node_distance(node, n); 5905 5906 if (val < min_val) { 5907 min_val = val; 5908 best_node = n; 5909 } 5910 } 5911 5912 if (best_node != -1) 5913 node_set(best_node, *used_nodes); 5914 return best_node; 5915} 5916 5917/** 5918 * sched_domain_node_span - get a cpumask for a node's sched_domain 5919 * @node: node whose cpumask we're constructing 5920 * @span: resulting cpumask 5921 * 5922 * Given a node, construct a good cpumask for its sched_domain to span. It 5923 * should be one that prevents unnecessary balancing, but also spreads tasks 5924 * out optimally. 5925 */ 5926static void sched_domain_node_span(int node, struct cpumask *span) 5927{ 5928 nodemask_t used_nodes; 5929 int i; 5930 5931 cpumask_clear(span); 5932 nodes_clear(used_nodes); 5933 5934 cpumask_or(span, span, cpumask_of_node(node)); 5935 node_set(node, used_nodes); 5936 5937 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) { 5938 int next_node = find_next_best_node(node, &used_nodes); 5939 if (next_node < 0) 5940 break; 5941 cpumask_or(span, span, cpumask_of_node(next_node)); 5942 } 5943} 5944 5945static const struct cpumask *cpu_node_mask(int cpu) 5946{ 5947 lockdep_assert_held(&sched_domains_mutex); 5948 5949 sched_domain_node_span(cpu_to_node(cpu), sched_domains_tmpmask); 5950 5951 return sched_domains_tmpmask; 5952} 5953 5954static const struct cpumask *cpu_allnodes_mask(int cpu) 5955{ 5956 return cpu_possible_mask; 5957} 5958#endif /* CONFIG_NUMA */ 5959 5960static const struct cpumask *cpu_cpu_mask(int cpu) 5961{ 5962 return cpumask_of_node(cpu_to_node(cpu)); 5963} 5964 5965int sched_smt_power_savings = 0, sched_mc_power_savings = 0; 5966 5967struct sd_data { 5968 struct sched_domain **__percpu sd; 5969 struct sched_group **__percpu sg; 5970 struct sched_group_power **__percpu sgp; 5971}; 5972 5973struct s_data { 5974 struct sched_domain ** __percpu sd; 5975 struct root_domain *rd; 5976}; 5977 5978enum s_alloc { 5979 sa_rootdomain, 5980 sa_sd, 5981 sa_sd_storage, 5982 sa_none, 5983}; 5984 5985struct sched_domain_topology_level; 5986 5987typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu); 5988typedef const struct cpumask *(*sched_domain_mask_f)(int cpu); 5989 5990#define SDTL_OVERLAP 0x01 5991 5992struct sched_domain_topology_level { 5993 sched_domain_init_f init; 5994 sched_domain_mask_f mask; 5995 int flags; 5996 struct sd_data data; 5997}; 5998 5999static int 6000build_overlap_sched_groups(struct sched_domain *sd, int cpu) 6001{ 6002 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg; 6003 const struct cpumask *span = sched_domain_span(sd); 6004 struct cpumask *covered = sched_domains_tmpmask; 6005 struct sd_data *sdd = sd->private; 6006 struct sched_domain *child; 6007 int i; 6008 6009 cpumask_clear(covered); 6010 6011 for_each_cpu(i, span) { 6012 struct cpumask *sg_span; 6013 6014 if (cpumask_test_cpu(i, covered)) 6015 continue; 6016 6017 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 6018 GFP_KERNEL, cpu_to_node(cpu)); 6019 6020 if (!sg) 6021 goto fail; 6022 6023 sg_span = sched_group_cpus(sg); 6024 6025 child = *per_cpu_ptr(sdd->sd, i); 6026 if (child->child) { 6027 child = child->child; 6028 cpumask_copy(sg_span, sched_domain_span(child)); 6029 } else 6030 cpumask_set_cpu(i, sg_span); 6031 6032 cpumask_or(covered, covered, sg_span); 6033 6034 sg->sgp = *per_cpu_ptr(sdd->sgp, cpumask_first(sg_span)); 6035 atomic_inc(&sg->sgp->ref); 6036 6037 if (cpumask_test_cpu(cpu, sg_span)) 6038 groups = sg; 6039 6040 if (!first) 6041 first = sg; 6042 if (last) 6043 last->next = sg; 6044 last = sg; 6045 last->next = first; 6046 } 6047 sd->groups = groups; 6048 6049 return 0; 6050 6051fail: 6052 free_sched_groups(first, 0); 6053 6054 return -ENOMEM; 6055} 6056 6057static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg) 6058{ 6059 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu); 6060 struct sched_domain *child = sd->child; 6061 6062 if (child) 6063 cpu = cpumask_first(sched_domain_span(child)); 6064 6065 if (sg) { 6066 *sg = *per_cpu_ptr(sdd->sg, cpu); 6067 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu); 6068 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */ 6069 } 6070 6071 return cpu; 6072} 6073 6074/* 6075 * build_sched_groups will build a circular linked list of the groups 6076 * covered by the given span, and will set each group's ->cpumask correctly, 6077 * and ->cpu_power to 0. 6078 * 6079 * Assumes the sched_domain tree is fully constructed 6080 */ 6081static int 6082build_sched_groups(struct sched_domain *sd, int cpu) 6083{ 6084 struct sched_group *first = NULL, *last = NULL; 6085 struct sd_data *sdd = sd->private; 6086 const struct cpumask *span = sched_domain_span(sd); 6087 struct cpumask *covered; 6088 int i; 6089 6090 get_group(cpu, sdd, &sd->groups); 6091 atomic_inc(&sd->groups->ref); 6092 6093 if (cpu != cpumask_first(sched_domain_span(sd))) 6094 return 0; 6095 6096 lockdep_assert_held(&sched_domains_mutex); 6097 covered = sched_domains_tmpmask; 6098 6099 cpumask_clear(covered); 6100 6101 for_each_cpu(i, span) { 6102 struct sched_group *sg; 6103 int group = get_group(i, sdd, &sg); 6104 int j; 6105 6106 if (cpumask_test_cpu(i, covered)) 6107 continue; 6108 6109 cpumask_clear(sched_group_cpus(sg)); 6110 sg->sgp->power = 0; 6111 6112 for_each_cpu(j, span) { 6113 if (get_group(j, sdd, NULL) != group) 6114 continue; 6115 6116 cpumask_set_cpu(j, covered); 6117 cpumask_set_cpu(j, sched_group_cpus(sg)); 6118 } 6119 6120 if (!first) 6121 first = sg; 6122 if (last) 6123 last->next = sg; 6124 last = sg; 6125 } 6126 last->next = first; 6127 6128 return 0; 6129} 6130 6131/* 6132 * Initialize sched groups cpu_power. 6133 * 6134 * cpu_power indicates the capacity of sched group, which is used while 6135 * distributing the load between different sched groups in a sched domain. 6136 * Typically cpu_power for all the groups in a sched domain will be same unless 6137 * there are asymmetries in the topology. If there are asymmetries, group 6138 * having more cpu_power will pickup more load compared to the group having 6139 * less cpu_power. 6140 */ 6141static void init_sched_groups_power(int cpu, struct sched_domain *sd) 6142{ 6143 struct sched_group *sg = sd->groups; 6144 6145 WARN_ON(!sd || !sg); 6146 6147 do { 6148 sg->group_weight = cpumask_weight(sched_group_cpus(sg)); 6149 sg = sg->next; 6150 } while (sg != sd->groups); 6151 6152 if (cpu != group_first_cpu(sg)) 6153 return; 6154 6155 update_group_power(sd, cpu); 6156 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight); 6157} 6158 6159int __weak arch_sd_sibling_asym_packing(void) 6160{ 6161 return 0*SD_ASYM_PACKING; 6162} 6163 6164/* 6165 * Initializers for schedule domains 6166 * Non-inlined to reduce accumulated stack pressure in build_sched_domains() 6167 */ 6168 6169#ifdef CONFIG_SCHED_DEBUG 6170# define SD_INIT_NAME(sd, type) sd->name = #type 6171#else 6172# define SD_INIT_NAME(sd, type) do { } while (0) 6173#endif 6174 6175#define SD_INIT_FUNC(type) \ 6176static noinline struct sched_domain * \ 6177sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \ 6178{ \ 6179 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \ 6180 *sd = SD_##type##_INIT; \ 6181 SD_INIT_NAME(sd, type); \ 6182 sd->private = &tl->data; \ 6183 return sd; \ 6184} 6185 6186SD_INIT_FUNC(CPU) 6187#ifdef CONFIG_NUMA 6188 SD_INIT_FUNC(ALLNODES) 6189 SD_INIT_FUNC(NODE) 6190#endif 6191#ifdef CONFIG_SCHED_SMT 6192 SD_INIT_FUNC(SIBLING) 6193#endif 6194#ifdef CONFIG_SCHED_MC 6195 SD_INIT_FUNC(MC) 6196#endif 6197#ifdef CONFIG_SCHED_BOOK 6198 SD_INIT_FUNC(BOOK) 6199#endif 6200 6201static int default_relax_domain_level = -1; 6202int sched_domain_level_max; 6203 6204static int __init setup_relax_domain_level(char *str) 6205{ 6206 unsigned long val; 6207 6208 val = simple_strtoul(str, NULL, 0); 6209 if (val < sched_domain_level_max) 6210 default_relax_domain_level = val; 6211 6212 return 1; 6213} 6214__setup("relax_domain_level=", setup_relax_domain_level); 6215 6216static void set_domain_attribute(struct sched_domain *sd, 6217 struct sched_domain_attr *attr) 6218{ 6219 int request; 6220 6221 if (!attr || attr->relax_domain_level < 0) { 6222 if (default_relax_domain_level < 0) 6223 return; 6224 else 6225 request = default_relax_domain_level; 6226 } else 6227 request = attr->relax_domain_level; 6228 if (request < sd->level) { 6229 /* turn off idle balance on this domain */ 6230 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 6231 } else { 6232 /* turn on idle balance on this domain */ 6233 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 6234 } 6235} 6236 6237static void __sdt_free(const struct cpumask *cpu_map); 6238static int __sdt_alloc(const struct cpumask *cpu_map); 6239 6240static void __free_domain_allocs(struct s_data *d, enum s_alloc what, 6241 const struct cpumask *cpu_map) 6242{ 6243 switch (what) { 6244 case sa_rootdomain: 6245 if (!atomic_read(&d->rd->refcount)) 6246 free_rootdomain(&d->rd->rcu); /* fall through */ 6247 case sa_sd: 6248 free_percpu(d->sd); /* fall through */ 6249 case sa_sd_storage: 6250 __sdt_free(cpu_map); /* fall through */ 6251 case sa_none: 6252 break; 6253 } 6254} 6255 6256static enum s_alloc __visit_domain_allocation_hell(struct s_data *d, 6257 const struct cpumask *cpu_map) 6258{ 6259 memset(d, 0, sizeof(*d)); 6260 6261 if (__sdt_alloc(cpu_map)) 6262 return sa_sd_storage; 6263 d->sd = alloc_percpu(struct sched_domain *); 6264 if (!d->sd) 6265 return sa_sd_storage; 6266 d->rd = alloc_rootdomain(); 6267 if (!d->rd) 6268 return sa_sd; 6269 return sa_rootdomain; 6270} 6271 6272/* 6273 * NULL the sd_data elements we've used to build the sched_domain and 6274 * sched_group structure so that the subsequent __free_domain_allocs() 6275 * will not free the data we're using. 6276 */ 6277static void claim_allocations(int cpu, struct sched_domain *sd) 6278{ 6279 struct sd_data *sdd = sd->private; 6280 6281 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd); 6282 *per_cpu_ptr(sdd->sd, cpu) = NULL; 6283 6284 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref)) 6285 *per_cpu_ptr(sdd->sg, cpu) = NULL; 6286 6287 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref)) 6288 *per_cpu_ptr(sdd->sgp, cpu) = NULL; 6289} 6290 6291#ifdef CONFIG_SCHED_SMT 6292static const struct cpumask *cpu_smt_mask(int cpu) 6293{ 6294 return topology_thread_cpumask(cpu); 6295} 6296#endif 6297 6298/* 6299 * Topology list, bottom-up. 6300 */ 6301static struct sched_domain_topology_level default_topology[] = { 6302#ifdef CONFIG_SCHED_SMT 6303 { sd_init_SIBLING, cpu_smt_mask, }, 6304#endif 6305#ifdef CONFIG_SCHED_MC 6306 { sd_init_MC, cpu_coregroup_mask, }, 6307#endif 6308#ifdef CONFIG_SCHED_BOOK 6309 { sd_init_BOOK, cpu_book_mask, }, 6310#endif 6311 { sd_init_CPU, cpu_cpu_mask, }, 6312#ifdef CONFIG_NUMA 6313 { sd_init_NODE, cpu_node_mask, SDTL_OVERLAP, }, 6314 { sd_init_ALLNODES, cpu_allnodes_mask, }, 6315#endif 6316 { NULL, }, 6317}; 6318 6319static struct sched_domain_topology_level *sched_domain_topology = default_topology; 6320 6321static int __sdt_alloc(const struct cpumask *cpu_map) 6322{ 6323 struct sched_domain_topology_level *tl; 6324 int j; 6325 6326 for (tl = sched_domain_topology; tl->init; tl++) { 6327 struct sd_data *sdd = &tl->data; 6328 6329 sdd->sd = alloc_percpu(struct sched_domain *); 6330 if (!sdd->sd) 6331 return -ENOMEM; 6332 6333 sdd->sg = alloc_percpu(struct sched_group *); 6334 if (!sdd->sg) 6335 return -ENOMEM; 6336 6337 sdd->sgp = alloc_percpu(struct sched_group_power *); 6338 if (!sdd->sgp) 6339 return -ENOMEM; 6340 6341 for_each_cpu(j, cpu_map) { 6342 struct sched_domain *sd; 6343 struct sched_group *sg; 6344 struct sched_group_power *sgp; 6345 6346 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(), 6347 GFP_KERNEL, cpu_to_node(j)); 6348 if (!sd) 6349 return -ENOMEM; 6350 6351 *per_cpu_ptr(sdd->sd, j) = sd; 6352 6353 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 6354 GFP_KERNEL, cpu_to_node(j)); 6355 if (!sg) 6356 return -ENOMEM; 6357 6358 *per_cpu_ptr(sdd->sg, j) = sg; 6359 6360 sgp = kzalloc_node(sizeof(struct sched_group_power), 6361 GFP_KERNEL, cpu_to_node(j)); 6362 if (!sgp) 6363 return -ENOMEM; 6364 6365 *per_cpu_ptr(sdd->sgp, j) = sgp; 6366 } 6367 } 6368 6369 return 0; 6370} 6371 6372static void __sdt_free(const struct cpumask *cpu_map) 6373{ 6374 struct sched_domain_topology_level *tl; 6375 int j; 6376 6377 for (tl = sched_domain_topology; tl->init; tl++) { 6378 struct sd_data *sdd = &tl->data; 6379 6380 for_each_cpu(j, cpu_map) { 6381 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, j); 6382 if (sd && (sd->flags & SD_OVERLAP)) 6383 free_sched_groups(sd->groups, 0); 6384 kfree(*per_cpu_ptr(sdd->sd, j)); 6385 kfree(*per_cpu_ptr(sdd->sg, j)); 6386 kfree(*per_cpu_ptr(sdd->sgp, j)); 6387 } 6388 free_percpu(sdd->sd); 6389 free_percpu(sdd->sg); 6390 free_percpu(sdd->sgp); 6391 } 6392} 6393 6394struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl, 6395 struct s_data *d, const struct cpumask *cpu_map, 6396 struct sched_domain_attr *attr, struct sched_domain *child, 6397 int cpu) 6398{ 6399 struct sched_domain *sd = tl->init(tl, cpu); 6400 if (!sd) 6401 return child; 6402 6403 set_domain_attribute(sd, attr); 6404 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu)); 6405 if (child) { 6406 sd->level = child->level + 1; 6407 sched_domain_level_max = max(sched_domain_level_max, sd->level); 6408 child->parent = sd; 6409 } 6410 sd->child = child; 6411 6412 return sd; 6413} 6414 6415/* 6416 * Build sched domains for a given set of cpus and attach the sched domains 6417 * to the individual cpus 6418 */ 6419static int build_sched_domains(const struct cpumask *cpu_map, 6420 struct sched_domain_attr *attr) 6421{ 6422 enum s_alloc alloc_state = sa_none; 6423 struct sched_domain *sd; 6424 struct s_data d; 6425 int i, ret = -ENOMEM; 6426 6427 alloc_state = __visit_domain_allocation_hell(&d, cpu_map); 6428 if (alloc_state != sa_rootdomain) 6429 goto error; 6430 6431 /* Set up domains for cpus specified by the cpu_map. */ 6432 for_each_cpu(i, cpu_map) { 6433 struct sched_domain_topology_level *tl; 6434 6435 sd = NULL; 6436 for (tl = sched_domain_topology; tl->init; tl++) { 6437 sd = build_sched_domain(tl, &d, cpu_map, attr, sd, i); 6438 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP)) 6439 sd->flags |= SD_OVERLAP; 6440 if (cpumask_equal(cpu_map, sched_domain_span(sd))) 6441 break; 6442 } 6443 6444 while (sd->child) 6445 sd = sd->child; 6446 6447 *per_cpu_ptr(d.sd, i) = sd; 6448 } 6449 6450 /* Build the groups for the domains */ 6451 for_each_cpu(i, cpu_map) { 6452 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 6453 sd->span_weight = cpumask_weight(sched_domain_span(sd)); 6454 if (sd->flags & SD_OVERLAP) { 6455 if (build_overlap_sched_groups(sd, i)) 6456 goto error; 6457 } else { 6458 if (build_sched_groups(sd, i)) 6459 goto error; 6460 } 6461 } 6462 } 6463 6464 /* Calculate CPU power for physical packages and nodes */ 6465 for (i = nr_cpumask_bits-1; i >= 0; i--) { 6466 if (!cpumask_test_cpu(i, cpu_map)) 6467 continue; 6468 6469 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 6470 claim_allocations(i, sd); 6471 init_sched_groups_power(i, sd); 6472 } 6473 } 6474 6475 /* Attach the domains */ 6476 rcu_read_lock(); 6477 for_each_cpu(i, cpu_map) { 6478 sd = *per_cpu_ptr(d.sd, i); 6479 cpu_attach_domain(sd, d.rd, i); 6480 } 6481 rcu_read_unlock(); 6482 6483 ret = 0; 6484error: 6485 __free_domain_allocs(&d, alloc_state, cpu_map); 6486 return ret; 6487} 6488 6489static cpumask_var_t *doms_cur; /* current sched domains */ 6490static int ndoms_cur; /* number of sched domains in 'doms_cur' */ 6491static struct sched_domain_attr *dattr_cur; 6492 /* attribues of custom domains in 'doms_cur' */ 6493 6494/* 6495 * Special case: If a kmalloc of a doms_cur partition (array of 6496 * cpumask) fails, then fallback to a single sched domain, 6497 * as determined by the single cpumask fallback_doms. 6498 */ 6499static cpumask_var_t fallback_doms; 6500 6501/* 6502 * arch_update_cpu_topology lets virtualized architectures update the 6503 * cpu core maps. It is supposed to return 1 if the topology changed 6504 * or 0 if it stayed the same. 6505 */ 6506int __attribute__((weak)) arch_update_cpu_topology(void) 6507{ 6508 return 0; 6509} 6510 6511cpumask_var_t *alloc_sched_domains(unsigned int ndoms) 6512{ 6513 int i; 6514 cpumask_var_t *doms; 6515 6516 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL); 6517 if (!doms) 6518 return NULL; 6519 for (i = 0; i < ndoms; i++) { 6520 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) { 6521 free_sched_domains(doms, i); 6522 return NULL; 6523 } 6524 } 6525 return doms; 6526} 6527 6528void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms) 6529{ 6530 unsigned int i; 6531 for (i = 0; i < ndoms; i++) 6532 free_cpumask_var(doms[i]); 6533 kfree(doms); 6534} 6535 6536/* 6537 * Set up scheduler domains and groups. Callers must hold the hotplug lock. 6538 * For now this just excludes isolated cpus, but could be used to 6539 * exclude other special cases in the future. 6540 */ 6541static int init_sched_domains(const struct cpumask *cpu_map) 6542{ 6543 int err; 6544 6545 arch_update_cpu_topology(); 6546 ndoms_cur = 1; 6547 doms_cur = alloc_sched_domains(ndoms_cur); 6548 if (!doms_cur) 6549 doms_cur = &fallback_doms; 6550 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map); 6551 dattr_cur = NULL; 6552 err = build_sched_domains(doms_cur[0], NULL); 6553 register_sched_domain_sysctl(); 6554 6555 return err; 6556} 6557 6558/* 6559 * Detach sched domains from a group of cpus specified in cpu_map 6560 * These cpus will now be attached to the NULL domain 6561 */ 6562static void detach_destroy_domains(const struct cpumask *cpu_map) 6563{ 6564 int i; 6565 6566 rcu_read_lock(); 6567 for_each_cpu(i, cpu_map) 6568 cpu_attach_domain(NULL, &def_root_domain, i); 6569 rcu_read_unlock(); 6570} 6571 6572/* handle null as "default" */ 6573static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur, 6574 struct sched_domain_attr *new, int idx_new) 6575{ 6576 struct sched_domain_attr tmp; 6577 6578 /* fast path */ 6579 if (!new && !cur) 6580 return 1; 6581 6582 tmp = SD_ATTR_INIT; 6583 return !memcmp(cur ? (cur + idx_cur) : &tmp, 6584 new ? (new + idx_new) : &tmp, 6585 sizeof(struct sched_domain_attr)); 6586} 6587 6588/* 6589 * Partition sched domains as specified by the 'ndoms_new' 6590 * cpumasks in the array doms_new[] of cpumasks. This compares 6591 * doms_new[] to the current sched domain partitioning, doms_cur[]. 6592 * It destroys each deleted domain and builds each new domain. 6593 * 6594 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'. 6595 * The masks don't intersect (don't overlap.) We should setup one 6596 * sched domain for each mask. CPUs not in any of the cpumasks will 6597 * not be load balanced. If the same cpumask appears both in the 6598 * current 'doms_cur' domains and in the new 'doms_new', we can leave 6599 * it as it is. 6600 * 6601 * The passed in 'doms_new' should be allocated using 6602 * alloc_sched_domains. This routine takes ownership of it and will 6603 * free_sched_domains it when done with it. If the caller failed the 6604 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1, 6605 * and partition_sched_domains() will fallback to the single partition 6606 * 'fallback_doms', it also forces the domains to be rebuilt. 6607 * 6608 * If doms_new == NULL it will be replaced with cpu_online_mask. 6609 * ndoms_new == 0 is a special case for destroying existing domains, 6610 * and it will not create the default domain. 6611 * 6612 * Call with hotplug lock held 6613 */ 6614void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[], 6615 struct sched_domain_attr *dattr_new) 6616{ 6617 int i, j, n; 6618 int new_topology; 6619 6620 mutex_lock(&sched_domains_mutex); 6621 6622 /* always unregister in case we don't destroy any domains */ 6623 unregister_sched_domain_sysctl(); 6624 6625 /* Let architecture update cpu core mappings. */ 6626 new_topology = arch_update_cpu_topology(); 6627 6628 n = doms_new ? ndoms_new : 0; 6629 6630 /* Destroy deleted domains */ 6631 for (i = 0; i < ndoms_cur; i++) { 6632 for (j = 0; j < n && !new_topology; j++) { 6633 if (cpumask_equal(doms_cur[i], doms_new[j]) 6634 && dattrs_equal(dattr_cur, i, dattr_new, j)) 6635 goto match1; 6636 } 6637 /* no match - a current sched domain not in new doms_new[] */ 6638 detach_destroy_domains(doms_cur[i]); 6639match1: 6640 ; 6641 } 6642 6643 if (doms_new == NULL) { 6644 ndoms_cur = 0; 6645 doms_new = &fallback_doms; 6646 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map); 6647 WARN_ON_ONCE(dattr_new); 6648 } 6649 6650 /* Build new domains */ 6651 for (i = 0; i < ndoms_new; i++) { 6652 for (j = 0; j < ndoms_cur && !new_topology; j++) { 6653 if (cpumask_equal(doms_new[i], doms_cur[j]) 6654 && dattrs_equal(dattr_new, i, dattr_cur, j)) 6655 goto match2; 6656 } 6657 /* no match - add a new doms_new */ 6658 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL); 6659match2: 6660 ; 6661 } 6662 6663 /* Remember the new sched domains */ 6664 if (doms_cur != &fallback_doms) 6665 free_sched_domains(doms_cur, ndoms_cur); 6666 kfree(dattr_cur); /* kfree(NULL) is safe */ 6667 doms_cur = doms_new; 6668 dattr_cur = dattr_new; 6669 ndoms_cur = ndoms_new; 6670 6671 register_sched_domain_sysctl(); 6672 6673 mutex_unlock(&sched_domains_mutex); 6674} 6675 6676#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT) 6677static void reinit_sched_domains(void) 6678{ 6679 get_online_cpus(); 6680 6681 /* Destroy domains first to force the rebuild */ 6682 partition_sched_domains(0, NULL, NULL); 6683 6684 rebuild_sched_domains(); 6685 put_online_cpus(); 6686} 6687 6688static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt) 6689{ 6690 unsigned int level = 0; 6691 6692 if (sscanf(buf, "%u", &level) != 1) 6693 return -EINVAL; 6694 6695 /* 6696 * level is always be positive so don't check for 6697 * level < POWERSAVINGS_BALANCE_NONE which is 0 6698 * What happens on 0 or 1 byte write, 6699 * need to check for count as well? 6700 */ 6701 6702 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS) 6703 return -EINVAL; 6704 6705 if (smt) 6706 sched_smt_power_savings = level; 6707 else 6708 sched_mc_power_savings = level; 6709 6710 reinit_sched_domains(); 6711 6712 return count; 6713} 6714 6715#ifdef CONFIG_SCHED_MC 6716static ssize_t sched_mc_power_savings_show(struct device *dev, 6717 struct device_attribute *attr, 6718 char *buf) 6719{ 6720 return sprintf(buf, "%u\n", sched_mc_power_savings); 6721} 6722static ssize_t sched_mc_power_savings_store(struct device *dev, 6723 struct device_attribute *attr, 6724 const char *buf, size_t count) 6725{ 6726 return sched_power_savings_store(buf, count, 0); 6727} 6728static DEVICE_ATTR(sched_mc_power_savings, 0644, 6729 sched_mc_power_savings_show, 6730 sched_mc_power_savings_store); 6731#endif 6732 6733#ifdef CONFIG_SCHED_SMT 6734static ssize_t sched_smt_power_savings_show(struct device *dev, 6735 struct device_attribute *attr, 6736 char *buf) 6737{ 6738 return sprintf(buf, "%u\n", sched_smt_power_savings); 6739} 6740static ssize_t sched_smt_power_savings_store(struct device *dev, 6741 struct device_attribute *attr, 6742 const char *buf, size_t count) 6743{ 6744 return sched_power_savings_store(buf, count, 1); 6745} 6746static DEVICE_ATTR(sched_smt_power_savings, 0644, 6747 sched_smt_power_savings_show, 6748 sched_smt_power_savings_store); 6749#endif 6750 6751int __init sched_create_sysfs_power_savings_entries(struct device *dev) 6752{ 6753 int err = 0; 6754 6755#ifdef CONFIG_SCHED_SMT 6756 if (smt_capable()) 6757 err = device_create_file(dev, &dev_attr_sched_smt_power_savings); 6758#endif 6759#ifdef CONFIG_SCHED_MC 6760 if (!err && mc_capable()) 6761 err = device_create_file(dev, &dev_attr_sched_mc_power_savings); 6762#endif 6763 return err; 6764} 6765#endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */ 6766 6767/* 6768 * Update cpusets according to cpu_active mask. If cpusets are 6769 * disabled, cpuset_update_active_cpus() becomes a simple wrapper 6770 * around partition_sched_domains(). 6771 */ 6772static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action, 6773 void *hcpu) 6774{ 6775 switch (action & ~CPU_TASKS_FROZEN) { 6776 case CPU_ONLINE: 6777 case CPU_DOWN_FAILED: 6778 cpuset_update_active_cpus(); 6779 return NOTIFY_OK; 6780 default: 6781 return NOTIFY_DONE; 6782 } 6783} 6784 6785static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action, 6786 void *hcpu) 6787{ 6788 switch (action & ~CPU_TASKS_FROZEN) { 6789 case CPU_DOWN_PREPARE: 6790 cpuset_update_active_cpus(); 6791 return NOTIFY_OK; 6792 default: 6793 return NOTIFY_DONE; 6794 } 6795} 6796 6797void __init sched_init_smp(void) 6798{ 6799 cpumask_var_t non_isolated_cpus; 6800 6801 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL); 6802 alloc_cpumask_var(&fallback_doms, GFP_KERNEL); 6803 6804 get_online_cpus(); 6805 mutex_lock(&sched_domains_mutex); 6806 init_sched_domains(cpu_active_mask); 6807 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map); 6808 if (cpumask_empty(non_isolated_cpus)) 6809 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus); 6810 mutex_unlock(&sched_domains_mutex); 6811 put_online_cpus(); 6812 6813 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE); 6814 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE); 6815 6816 /* RT runtime code needs to handle some hotplug events */ 6817 hotcpu_notifier(update_runtime, 0); 6818 6819 init_hrtick(); 6820 6821 /* Move init over to a non-isolated CPU */ 6822 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0) 6823 BUG(); 6824 sched_init_granularity(); 6825 free_cpumask_var(non_isolated_cpus); 6826 6827 init_sched_rt_class(); 6828} 6829#else 6830void __init sched_init_smp(void) 6831{ 6832 sched_init_granularity(); 6833} 6834#endif /* CONFIG_SMP */ 6835 6836const_debug unsigned int sysctl_timer_migration = 1; 6837 6838int in_sched_functions(unsigned long addr) 6839{ 6840 return in_lock_functions(addr) || 6841 (addr >= (unsigned long)__sched_text_start 6842 && addr < (unsigned long)__sched_text_end); 6843} 6844 6845#ifdef CONFIG_CGROUP_SCHED 6846struct task_group root_task_group; 6847#endif 6848 6849DECLARE_PER_CPU(cpumask_var_t, load_balance_tmpmask); 6850 6851void __init sched_init(void) 6852{ 6853 int i, j; 6854 unsigned long alloc_size = 0, ptr; 6855 6856#ifdef CONFIG_FAIR_GROUP_SCHED 6857 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 6858#endif 6859#ifdef CONFIG_RT_GROUP_SCHED 6860 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 6861#endif 6862#ifdef CONFIG_CPUMASK_OFFSTACK 6863 alloc_size += num_possible_cpus() * cpumask_size(); 6864#endif 6865 if (alloc_size) { 6866 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT); 6867 6868#ifdef CONFIG_FAIR_GROUP_SCHED 6869 root_task_group.se = (struct sched_entity **)ptr; 6870 ptr += nr_cpu_ids * sizeof(void **); 6871 6872 root_task_group.cfs_rq = (struct cfs_rq **)ptr; 6873 ptr += nr_cpu_ids * sizeof(void **); 6874 6875#endif /* CONFIG_FAIR_GROUP_SCHED */ 6876#ifdef CONFIG_RT_GROUP_SCHED 6877 root_task_group.rt_se = (struct sched_rt_entity **)ptr; 6878 ptr += nr_cpu_ids * sizeof(void **); 6879 6880 root_task_group.rt_rq = (struct rt_rq **)ptr; 6881 ptr += nr_cpu_ids * sizeof(void **); 6882 6883#endif /* CONFIG_RT_GROUP_SCHED */ 6884#ifdef CONFIG_CPUMASK_OFFSTACK 6885 for_each_possible_cpu(i) { 6886 per_cpu(load_balance_tmpmask, i) = (void *)ptr; 6887 ptr += cpumask_size(); 6888 } 6889#endif /* CONFIG_CPUMASK_OFFSTACK */ 6890 } 6891 6892#ifdef CONFIG_SMP 6893 init_defrootdomain(); 6894#endif 6895 6896 init_rt_bandwidth(&def_rt_bandwidth, 6897 global_rt_period(), global_rt_runtime()); 6898 6899#ifdef CONFIG_RT_GROUP_SCHED 6900 init_rt_bandwidth(&root_task_group.rt_bandwidth, 6901 global_rt_period(), global_rt_runtime()); 6902#endif /* CONFIG_RT_GROUP_SCHED */ 6903 6904#ifdef CONFIG_CGROUP_SCHED 6905 list_add(&root_task_group.list, &task_groups); 6906 INIT_LIST_HEAD(&root_task_group.children); 6907 INIT_LIST_HEAD(&root_task_group.siblings); 6908 autogroup_init(&init_task); 6909 6910#endif /* CONFIG_CGROUP_SCHED */ 6911 6912#ifdef CONFIG_CGROUP_CPUACCT 6913 root_cpuacct.cpustat = &kernel_cpustat; 6914 root_cpuacct.cpuusage = alloc_percpu(u64); 6915 /* Too early, not expected to fail */ 6916 BUG_ON(!root_cpuacct.cpuusage); 6917#endif 6918 for_each_possible_cpu(i) { 6919 struct rq *rq; 6920 6921 rq = cpu_rq(i); 6922 raw_spin_lock_init(&rq->lock); 6923 rq->nr_running = 0; 6924 rq->calc_load_active = 0; 6925 rq->calc_load_update = jiffies + LOAD_FREQ; 6926 init_cfs_rq(&rq->cfs); 6927 init_rt_rq(&rq->rt, rq); 6928#ifdef CONFIG_FAIR_GROUP_SCHED 6929 root_task_group.shares = ROOT_TASK_GROUP_LOAD; 6930 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list); 6931 /* 6932 * How much cpu bandwidth does root_task_group get? 6933 * 6934 * In case of task-groups formed thr' the cgroup filesystem, it 6935 * gets 100% of the cpu resources in the system. This overall 6936 * system cpu resource is divided among the tasks of 6937 * root_task_group and its child task-groups in a fair manner, 6938 * based on each entity's (task or task-group's) weight 6939 * (se->load.weight). 6940 * 6941 * In other words, if root_task_group has 10 tasks of weight 6942 * 1024) and two child groups A0 and A1 (of weight 1024 each), 6943 * then A0's share of the cpu resource is: 6944 * 6945 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33% 6946 * 6947 * We achieve this by letting root_task_group's tasks sit 6948 * directly in rq->cfs (i.e root_task_group->se[] = NULL). 6949 */ 6950 init_cfs_bandwidth(&root_task_group.cfs_bandwidth); 6951 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL); 6952#endif /* CONFIG_FAIR_GROUP_SCHED */ 6953 6954 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime; 6955#ifdef CONFIG_RT_GROUP_SCHED 6956 INIT_LIST_HEAD(&rq->leaf_rt_rq_list); 6957 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL); 6958#endif 6959 6960 for (j = 0; j < CPU_LOAD_IDX_MAX; j++) 6961 rq->cpu_load[j] = 0; 6962 6963 rq->last_load_update_tick = jiffies; 6964 6965#ifdef CONFIG_SMP 6966 rq->sd = NULL; 6967 rq->rd = NULL; 6968 rq->cpu_power = SCHED_POWER_SCALE; 6969 rq->post_schedule = 0; 6970 rq->active_balance = 0; 6971 rq->next_balance = jiffies; 6972 rq->push_cpu = 0; 6973 rq->cpu = i; 6974 rq->online = 0; 6975 rq->idle_stamp = 0; 6976 rq->avg_idle = 2*sysctl_sched_migration_cost; 6977 6978 INIT_LIST_HEAD(&rq->cfs_tasks); 6979 6980 rq_attach_root(rq, &def_root_domain); 6981#ifdef CONFIG_NO_HZ 6982 rq->nohz_flags = 0; 6983#endif 6984#endif 6985 init_rq_hrtick(rq); 6986 atomic_set(&rq->nr_iowait, 0); 6987 } 6988 6989 set_load_weight(&init_task); 6990 6991#ifdef CONFIG_PREEMPT_NOTIFIERS 6992 INIT_HLIST_HEAD(&init_task.preempt_notifiers); 6993#endif 6994 6995#ifdef CONFIG_RT_MUTEXES 6996 plist_head_init(&init_task.pi_waiters); 6997#endif 6998 6999 /* 7000 * The boot idle thread does lazy MMU switching as well: 7001 */ 7002 atomic_inc(&init_mm.mm_count); 7003 enter_lazy_tlb(&init_mm, current); 7004 7005 /* 7006 * Make us the idle thread. Technically, schedule() should not be 7007 * called from this thread, however somewhere below it might be, 7008 * but because we are the idle thread, we just pick up running again 7009 * when this runqueue becomes "idle". 7010 */ 7011 init_idle(current, smp_processor_id()); 7012 7013 calc_load_update = jiffies + LOAD_FREQ; 7014 7015 /* 7016 * During early bootup we pretend to be a normal task: 7017 */ 7018 current->sched_class = &fair_sched_class; 7019 7020#ifdef CONFIG_SMP 7021 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT); 7022 /* May be allocated at isolcpus cmdline parse time */ 7023 if (cpu_isolated_map == NULL) 7024 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT); 7025#endif 7026 init_sched_fair_class(); 7027 7028 scheduler_running = 1; 7029} 7030 7031#ifdef CONFIG_DEBUG_ATOMIC_SLEEP 7032static inline int preempt_count_equals(int preempt_offset) 7033{ 7034 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth(); 7035 7036 return (nested == preempt_offset); 7037} 7038 7039void __might_sleep(const char *file, int line, int preempt_offset) 7040{ 7041 static unsigned long prev_jiffy; /* ratelimiting */ 7042 7043 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */ 7044 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) || 7045 system_state != SYSTEM_RUNNING || oops_in_progress) 7046 return; 7047 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy) 7048 return; 7049 prev_jiffy = jiffies; 7050 7051 printk(KERN_ERR 7052 "BUG: sleeping function called from invalid context at %s:%d\n", 7053 file, line); 7054 printk(KERN_ERR 7055 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n", 7056 in_atomic(), irqs_disabled(), 7057 current->pid, current->comm); 7058 7059 debug_show_held_locks(current); 7060 if (irqs_disabled()) 7061 print_irqtrace_events(current); 7062 dump_stack(); 7063} 7064EXPORT_SYMBOL(__might_sleep); 7065#endif 7066 7067#ifdef CONFIG_MAGIC_SYSRQ 7068static void normalize_task(struct rq *rq, struct task_struct *p) 7069{ 7070 const struct sched_class *prev_class = p->sched_class; 7071 int old_prio = p->prio; 7072 int on_rq; 7073 7074 on_rq = p->on_rq; 7075 if (on_rq) 7076 dequeue_task(rq, p, 0); 7077 __setscheduler(rq, p, SCHED_NORMAL, 0); 7078 if (on_rq) { 7079 enqueue_task(rq, p, 0); 7080 resched_task(rq->curr); 7081 } 7082 7083 check_class_changed(rq, p, prev_class, old_prio); 7084} 7085 7086void normalize_rt_tasks(void) 7087{ 7088 struct task_struct *g, *p; 7089 unsigned long flags; 7090 struct rq *rq; 7091 7092 read_lock_irqsave(&tasklist_lock, flags); 7093 do_each_thread(g, p) { 7094 /* 7095 * Only normalize user tasks: 7096 */ 7097 if (!p->mm) 7098 continue; 7099 7100 p->se.exec_start = 0; 7101#ifdef CONFIG_SCHEDSTATS 7102 p->se.statistics.wait_start = 0; 7103 p->se.statistics.sleep_start = 0; 7104 p->se.statistics.block_start = 0; 7105#endif 7106 7107 if (!rt_task(p)) { 7108 /* 7109 * Renice negative nice level userspace 7110 * tasks back to 0: 7111 */ 7112 if (TASK_NICE(p) < 0 && p->mm) 7113 set_user_nice(p, 0); 7114 continue; 7115 } 7116 7117 raw_spin_lock(&p->pi_lock); 7118 rq = __task_rq_lock(p); 7119 7120 normalize_task(rq, p); 7121 7122 __task_rq_unlock(rq); 7123 raw_spin_unlock(&p->pi_lock); 7124 } while_each_thread(g, p); 7125 7126 read_unlock_irqrestore(&tasklist_lock, flags); 7127} 7128 7129#endif /* CONFIG_MAGIC_SYSRQ */ 7130 7131#if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) 7132/* 7133 * These functions are only useful for the IA64 MCA handling, or kdb. 7134 * 7135 * They can only be called when the whole system has been 7136 * stopped - every CPU needs to be quiescent, and no scheduling 7137 * activity can take place. Using them for anything else would 7138 * be a serious bug, and as a result, they aren't even visible 7139 * under any other configuration. 7140 */ 7141 7142/** 7143 * curr_task - return the current task for a given cpu. 7144 * @cpu: the processor in question. 7145 * 7146 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 7147 */ 7148struct task_struct *curr_task(int cpu) 7149{ 7150 return cpu_curr(cpu); 7151} 7152 7153#endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */ 7154 7155#ifdef CONFIG_IA64 7156/** 7157 * set_curr_task - set the current task for a given cpu. 7158 * @cpu: the processor in question. 7159 * @p: the task pointer to set. 7160 * 7161 * Description: This function must only be used when non-maskable interrupts 7162 * are serviced on a separate stack. It allows the architecture to switch the 7163 * notion of the current task on a cpu in a non-blocking manner. This function 7164 * must be called with all CPU's synchronized, and interrupts disabled, the 7165 * and caller must save the original value of the current task (see 7166 * curr_task() above) and restore that value before reenabling interrupts and 7167 * re-starting the system. 7168 * 7169 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 7170 */ 7171void set_curr_task(int cpu, struct task_struct *p) 7172{ 7173 cpu_curr(cpu) = p; 7174} 7175 7176#endif 7177 7178#ifdef CONFIG_CGROUP_SCHED 7179/* task_group_lock serializes the addition/removal of task groups */ 7180static DEFINE_SPINLOCK(task_group_lock); 7181 7182static void free_sched_group(struct task_group *tg) 7183{ 7184 free_fair_sched_group(tg); 7185 free_rt_sched_group(tg); 7186 autogroup_free(tg); 7187 kfree(tg); 7188} 7189 7190/* allocate runqueue etc for a new task group */ 7191struct task_group *sched_create_group(struct task_group *parent) 7192{ 7193 struct task_group *tg; 7194 unsigned long flags; 7195 7196 tg = kzalloc(sizeof(*tg), GFP_KERNEL); 7197 if (!tg) 7198 return ERR_PTR(-ENOMEM); 7199 7200 if (!alloc_fair_sched_group(tg, parent)) 7201 goto err; 7202 7203 if (!alloc_rt_sched_group(tg, parent)) 7204 goto err; 7205 7206 spin_lock_irqsave(&task_group_lock, flags); 7207 list_add_rcu(&tg->list, &task_groups); 7208 7209 WARN_ON(!parent); /* root should already exist */ 7210 7211 tg->parent = parent; 7212 INIT_LIST_HEAD(&tg->children); 7213 list_add_rcu(&tg->siblings, &parent->children); 7214 spin_unlock_irqrestore(&task_group_lock, flags); 7215 7216 return tg; 7217 7218err: 7219 free_sched_group(tg); 7220 return ERR_PTR(-ENOMEM); 7221} 7222 7223/* rcu callback to free various structures associated with a task group */ 7224static void free_sched_group_rcu(struct rcu_head *rhp) 7225{ 7226 /* now it should be safe to free those cfs_rqs */ 7227 free_sched_group(container_of(rhp, struct task_group, rcu)); 7228} 7229 7230/* Destroy runqueue etc associated with a task group */ 7231void sched_destroy_group(struct task_group *tg) 7232{ 7233 unsigned long flags; 7234 int i; 7235 7236 /* end participation in shares distribution */ 7237 for_each_possible_cpu(i) 7238 unregister_fair_sched_group(tg, i); 7239 7240 spin_lock_irqsave(&task_group_lock, flags); 7241 list_del_rcu(&tg->list); 7242 list_del_rcu(&tg->siblings); 7243 spin_unlock_irqrestore(&task_group_lock, flags); 7244 7245 /* wait for possible concurrent references to cfs_rqs complete */ 7246 call_rcu(&tg->rcu, free_sched_group_rcu); 7247} 7248 7249/* change task's runqueue when it moves between groups. 7250 * The caller of this function should have put the task in its new group 7251 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to 7252 * reflect its new group. 7253 */ 7254void sched_move_task(struct task_struct *tsk) 7255{ 7256 int on_rq, running; 7257 unsigned long flags; 7258 struct rq *rq; 7259 7260 rq = task_rq_lock(tsk, &flags); 7261 7262 running = task_current(rq, tsk); 7263 on_rq = tsk->on_rq; 7264 7265 if (on_rq) 7266 dequeue_task(rq, tsk, 0); 7267 if (unlikely(running)) 7268 tsk->sched_class->put_prev_task(rq, tsk); 7269 7270#ifdef CONFIG_FAIR_GROUP_SCHED 7271 if (tsk->sched_class->task_move_group) 7272 tsk->sched_class->task_move_group(tsk, on_rq); 7273 else 7274#endif 7275 set_task_rq(tsk, task_cpu(tsk)); 7276 7277 if (unlikely(running)) 7278 tsk->sched_class->set_curr_task(rq); 7279 if (on_rq) 7280 enqueue_task(rq, tsk, 0); 7281 7282 task_rq_unlock(rq, tsk, &flags); 7283} 7284#endif /* CONFIG_CGROUP_SCHED */ 7285 7286#if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH) 7287static unsigned long to_ratio(u64 period, u64 runtime) 7288{ 7289 if (runtime == RUNTIME_INF) 7290 return 1ULL << 20; 7291 7292 return div64_u64(runtime << 20, period); 7293} 7294#endif 7295 7296#ifdef CONFIG_RT_GROUP_SCHED 7297/* 7298 * Ensure that the real time constraints are schedulable. 7299 */ 7300static DEFINE_MUTEX(rt_constraints_mutex); 7301 7302/* Must be called with tasklist_lock held */ 7303static inline int tg_has_rt_tasks(struct task_group *tg) 7304{ 7305 struct task_struct *g, *p; 7306 7307 do_each_thread(g, p) { 7308 if (rt_task(p) && task_rq(p)->rt.tg == tg) 7309 return 1; 7310 } while_each_thread(g, p); 7311 7312 return 0; 7313} 7314 7315struct rt_schedulable_data { 7316 struct task_group *tg; 7317 u64 rt_period; 7318 u64 rt_runtime; 7319}; 7320 7321static int tg_rt_schedulable(struct task_group *tg, void *data) 7322{ 7323 struct rt_schedulable_data *d = data; 7324 struct task_group *child; 7325 unsigned long total, sum = 0; 7326 u64 period, runtime; 7327 7328 period = ktime_to_ns(tg->rt_bandwidth.rt_period); 7329 runtime = tg->rt_bandwidth.rt_runtime; 7330 7331 if (tg == d->tg) { 7332 period = d->rt_period; 7333 runtime = d->rt_runtime; 7334 } 7335 7336 /* 7337 * Cannot have more runtime than the period. 7338 */ 7339 if (runtime > period && runtime != RUNTIME_INF) 7340 return -EINVAL; 7341 7342 /* 7343 * Ensure we don't starve existing RT tasks. 7344 */ 7345 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg)) 7346 return -EBUSY; 7347 7348 total = to_ratio(period, runtime); 7349 7350 /* 7351 * Nobody can have more than the global setting allows. 7352 */ 7353 if (total > to_ratio(global_rt_period(), global_rt_runtime())) 7354 return -EINVAL; 7355 7356 /* 7357 * The sum of our children's runtime should not exceed our own. 7358 */ 7359 list_for_each_entry_rcu(child, &tg->children, siblings) { 7360 period = ktime_to_ns(child->rt_bandwidth.rt_period); 7361 runtime = child->rt_bandwidth.rt_runtime; 7362 7363 if (child == d->tg) { 7364 period = d->rt_period; 7365 runtime = d->rt_runtime; 7366 } 7367 7368 sum += to_ratio(period, runtime); 7369 } 7370 7371 if (sum > total) 7372 return -EINVAL; 7373 7374 return 0; 7375} 7376 7377static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime) 7378{ 7379 int ret; 7380 7381 struct rt_schedulable_data data = { 7382 .tg = tg, 7383 .rt_period = period, 7384 .rt_runtime = runtime, 7385 }; 7386 7387 rcu_read_lock(); 7388 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data); 7389 rcu_read_unlock(); 7390 7391 return ret; 7392} 7393 7394static int tg_set_rt_bandwidth(struct task_group *tg, 7395 u64 rt_period, u64 rt_runtime) 7396{ 7397 int i, err = 0; 7398 7399 mutex_lock(&rt_constraints_mutex); 7400 read_lock(&tasklist_lock); 7401 err = __rt_schedulable(tg, rt_period, rt_runtime); 7402 if (err) 7403 goto unlock; 7404 7405 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock); 7406 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period); 7407 tg->rt_bandwidth.rt_runtime = rt_runtime; 7408 7409 for_each_possible_cpu(i) { 7410 struct rt_rq *rt_rq = tg->rt_rq[i]; 7411 7412 raw_spin_lock(&rt_rq->rt_runtime_lock); 7413 rt_rq->rt_runtime = rt_runtime; 7414 raw_spin_unlock(&rt_rq->rt_runtime_lock); 7415 } 7416 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock); 7417unlock: 7418 read_unlock(&tasklist_lock); 7419 mutex_unlock(&rt_constraints_mutex); 7420 7421 return err; 7422} 7423 7424int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us) 7425{ 7426 u64 rt_runtime, rt_period; 7427 7428 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period); 7429 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC; 7430 if (rt_runtime_us < 0) 7431 rt_runtime = RUNTIME_INF; 7432 7433 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 7434} 7435 7436long sched_group_rt_runtime(struct task_group *tg) 7437{ 7438 u64 rt_runtime_us; 7439 7440 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF) 7441 return -1; 7442 7443 rt_runtime_us = tg->rt_bandwidth.rt_runtime; 7444 do_div(rt_runtime_us, NSEC_PER_USEC); 7445 return rt_runtime_us; 7446} 7447 7448int sched_group_set_rt_period(struct task_group *tg, long rt_period_us) 7449{ 7450 u64 rt_runtime, rt_period; 7451 7452 rt_period = (u64)rt_period_us * NSEC_PER_USEC; 7453 rt_runtime = tg->rt_bandwidth.rt_runtime; 7454 7455 if (rt_period == 0) 7456 return -EINVAL; 7457 7458 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 7459} 7460 7461long sched_group_rt_period(struct task_group *tg) 7462{ 7463 u64 rt_period_us; 7464 7465 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period); 7466 do_div(rt_period_us, NSEC_PER_USEC); 7467 return rt_period_us; 7468} 7469 7470static int sched_rt_global_constraints(void) 7471{ 7472 u64 runtime, period; 7473 int ret = 0; 7474 7475 if (sysctl_sched_rt_period <= 0) 7476 return -EINVAL; 7477 7478 runtime = global_rt_runtime(); 7479 period = global_rt_period(); 7480 7481 /* 7482 * Sanity check on the sysctl variables. 7483 */ 7484 if (runtime > period && runtime != RUNTIME_INF) 7485 return -EINVAL; 7486 7487 mutex_lock(&rt_constraints_mutex); 7488 read_lock(&tasklist_lock); 7489 ret = __rt_schedulable(NULL, 0, 0); 7490 read_unlock(&tasklist_lock); 7491 mutex_unlock(&rt_constraints_mutex); 7492 7493 return ret; 7494} 7495 7496int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk) 7497{ 7498 /* Don't accept realtime tasks when there is no way for them to run */ 7499 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0) 7500 return 0; 7501 7502 return 1; 7503} 7504 7505#else /* !CONFIG_RT_GROUP_SCHED */ 7506static int sched_rt_global_constraints(void) 7507{ 7508 unsigned long flags; 7509 int i; 7510 7511 if (sysctl_sched_rt_period <= 0) 7512 return -EINVAL; 7513 7514 /* 7515 * There's always some RT tasks in the root group 7516 * -- migration, kstopmachine etc.. 7517 */ 7518 if (sysctl_sched_rt_runtime == 0) 7519 return -EBUSY; 7520 7521 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags); 7522 for_each_possible_cpu(i) { 7523 struct rt_rq *rt_rq = &cpu_rq(i)->rt; 7524 7525 raw_spin_lock(&rt_rq->rt_runtime_lock); 7526 rt_rq->rt_runtime = global_rt_runtime(); 7527 raw_spin_unlock(&rt_rq->rt_runtime_lock); 7528 } 7529 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags); 7530 7531 return 0; 7532} 7533#endif /* CONFIG_RT_GROUP_SCHED */ 7534 7535int sched_rt_handler(struct ctl_table *table, int write, 7536 void __user *buffer, size_t *lenp, 7537 loff_t *ppos) 7538{ 7539 int ret; 7540 int old_period, old_runtime; 7541 static DEFINE_MUTEX(mutex); 7542 7543 mutex_lock(&mutex); 7544 old_period = sysctl_sched_rt_period; 7545 old_runtime = sysctl_sched_rt_runtime; 7546 7547 ret = proc_dointvec(table, write, buffer, lenp, ppos); 7548 7549 if (!ret && write) { 7550 ret = sched_rt_global_constraints(); 7551 if (ret) { 7552 sysctl_sched_rt_period = old_period; 7553 sysctl_sched_rt_runtime = old_runtime; 7554 } else { 7555 def_rt_bandwidth.rt_runtime = global_rt_runtime(); 7556 def_rt_bandwidth.rt_period = 7557 ns_to_ktime(global_rt_period()); 7558 } 7559 } 7560 mutex_unlock(&mutex); 7561 7562 return ret; 7563} 7564 7565#ifdef CONFIG_CGROUP_SCHED 7566 7567/* return corresponding task_group object of a cgroup */ 7568static inline struct task_group *cgroup_tg(struct cgroup *cgrp) 7569{ 7570 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id), 7571 struct task_group, css); 7572} 7573 7574static struct cgroup_subsys_state * 7575cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp) 7576{ 7577 struct task_group *tg, *parent; 7578 7579 if (!cgrp->parent) { 7580 /* This is early initialization for the top cgroup */ 7581 return &root_task_group.css; 7582 } 7583 7584 parent = cgroup_tg(cgrp->parent); 7585 tg = sched_create_group(parent); 7586 if (IS_ERR(tg)) 7587 return ERR_PTR(-ENOMEM); 7588 7589 return &tg->css; 7590} 7591 7592static void 7593cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp) 7594{ 7595 struct task_group *tg = cgroup_tg(cgrp); 7596 7597 sched_destroy_group(tg); 7598} 7599 7600static int cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp, 7601 struct cgroup_taskset *tset) 7602{ 7603 struct task_struct *task; 7604 7605 cgroup_taskset_for_each(task, cgrp, tset) { 7606#ifdef CONFIG_RT_GROUP_SCHED 7607 if (!sched_rt_can_attach(cgroup_tg(cgrp), task)) 7608 return -EINVAL; 7609#else 7610 /* We don't support RT-tasks being in separate groups */ 7611 if (task->sched_class != &fair_sched_class) 7612 return -EINVAL; 7613#endif 7614 } 7615 return 0; 7616} 7617 7618static void cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp, 7619 struct cgroup_taskset *tset) 7620{ 7621 struct task_struct *task; 7622 7623 cgroup_taskset_for_each(task, cgrp, tset) 7624 sched_move_task(task); 7625} 7626 7627static void 7628cpu_cgroup_exit(struct cgroup_subsys *ss, struct cgroup *cgrp, 7629 struct cgroup *old_cgrp, struct task_struct *task) 7630{ 7631 /* 7632 * cgroup_exit() is called in the copy_process() failure path. 7633 * Ignore this case since the task hasn't ran yet, this avoids 7634 * trying to poke a half freed task state from generic code. 7635 */ 7636 if (!(task->flags & PF_EXITING)) 7637 return; 7638 7639 sched_move_task(task); 7640} 7641 7642#ifdef CONFIG_FAIR_GROUP_SCHED 7643static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype, 7644 u64 shareval) 7645{ 7646 return sched_group_set_shares(cgroup_tg(cgrp), scale_load(shareval)); 7647} 7648 7649static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft) 7650{ 7651 struct task_group *tg = cgroup_tg(cgrp); 7652 7653 return (u64) scale_load_down(tg->shares); 7654} 7655 7656#ifdef CONFIG_CFS_BANDWIDTH 7657static DEFINE_MUTEX(cfs_constraints_mutex); 7658 7659const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */ 7660const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */ 7661 7662static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime); 7663 7664static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota) 7665{ 7666 int i, ret = 0, runtime_enabled, runtime_was_enabled; 7667 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7668 7669 if (tg == &root_task_group) 7670 return -EINVAL; 7671 7672 /* 7673 * Ensure we have at some amount of bandwidth every period. This is 7674 * to prevent reaching a state of large arrears when throttled via 7675 * entity_tick() resulting in prolonged exit starvation. 7676 */ 7677 if (quota < min_cfs_quota_period || period < min_cfs_quota_period) 7678 return -EINVAL; 7679 7680 /* 7681 * Likewise, bound things on the otherside by preventing insane quota 7682 * periods. This also allows us to normalize in computing quota 7683 * feasibility. 7684 */ 7685 if (period > max_cfs_quota_period) 7686 return -EINVAL; 7687 7688 mutex_lock(&cfs_constraints_mutex); 7689 ret = __cfs_schedulable(tg, period, quota); 7690 if (ret) 7691 goto out_unlock; 7692 7693 runtime_enabled = quota != RUNTIME_INF; 7694 runtime_was_enabled = cfs_b->quota != RUNTIME_INF; 7695 account_cfs_bandwidth_used(runtime_enabled, runtime_was_enabled); 7696 raw_spin_lock_irq(&cfs_b->lock); 7697 cfs_b->period = ns_to_ktime(period); 7698 cfs_b->quota = quota; 7699 7700 __refill_cfs_bandwidth_runtime(cfs_b); 7701 /* restart the period timer (if active) to handle new period expiry */ 7702 if (runtime_enabled && cfs_b->timer_active) { 7703 /* force a reprogram */ 7704 cfs_b->timer_active = 0; 7705 __start_cfs_bandwidth(cfs_b); 7706 } 7707 raw_spin_unlock_irq(&cfs_b->lock); 7708 7709 for_each_possible_cpu(i) { 7710 struct cfs_rq *cfs_rq = tg->cfs_rq[i]; 7711 struct rq *rq = cfs_rq->rq; 7712 7713 raw_spin_lock_irq(&rq->lock); 7714 cfs_rq->runtime_enabled = runtime_enabled; 7715 cfs_rq->runtime_remaining = 0; 7716 7717 if (cfs_rq->throttled) 7718 unthrottle_cfs_rq(cfs_rq); 7719 raw_spin_unlock_irq(&rq->lock); 7720 } 7721out_unlock: 7722 mutex_unlock(&cfs_constraints_mutex); 7723 7724 return ret; 7725} 7726 7727int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us) 7728{ 7729 u64 quota, period; 7730 7731 period = ktime_to_ns(tg->cfs_bandwidth.period); 7732 if (cfs_quota_us < 0) 7733 quota = RUNTIME_INF; 7734 else 7735 quota = (u64)cfs_quota_us * NSEC_PER_USEC; 7736 7737 return tg_set_cfs_bandwidth(tg, period, quota); 7738} 7739 7740long tg_get_cfs_quota(struct task_group *tg) 7741{ 7742 u64 quota_us; 7743 7744 if (tg->cfs_bandwidth.quota == RUNTIME_INF) 7745 return -1; 7746 7747 quota_us = tg->cfs_bandwidth.quota; 7748 do_div(quota_us, NSEC_PER_USEC); 7749 7750 return quota_us; 7751} 7752 7753int tg_set_cfs_period(struct task_group *tg, long cfs_period_us) 7754{ 7755 u64 quota, period; 7756 7757 period = (u64)cfs_period_us * NSEC_PER_USEC; 7758 quota = tg->cfs_bandwidth.quota; 7759 7760 return tg_set_cfs_bandwidth(tg, period, quota); 7761} 7762 7763long tg_get_cfs_period(struct task_group *tg) 7764{ 7765 u64 cfs_period_us; 7766 7767 cfs_period_us = ktime_to_ns(tg->cfs_bandwidth.period); 7768 do_div(cfs_period_us, NSEC_PER_USEC); 7769 7770 return cfs_period_us; 7771} 7772 7773static s64 cpu_cfs_quota_read_s64(struct cgroup *cgrp, struct cftype *cft) 7774{ 7775 return tg_get_cfs_quota(cgroup_tg(cgrp)); 7776} 7777 7778static int cpu_cfs_quota_write_s64(struct cgroup *cgrp, struct cftype *cftype, 7779 s64 cfs_quota_us) 7780{ 7781 return tg_set_cfs_quota(cgroup_tg(cgrp), cfs_quota_us); 7782} 7783 7784static u64 cpu_cfs_period_read_u64(struct cgroup *cgrp, struct cftype *cft) 7785{ 7786 return tg_get_cfs_period(cgroup_tg(cgrp)); 7787} 7788 7789static int cpu_cfs_period_write_u64(struct cgroup *cgrp, struct cftype *cftype, 7790 u64 cfs_period_us) 7791{ 7792 return tg_set_cfs_period(cgroup_tg(cgrp), cfs_period_us); 7793} 7794 7795struct cfs_schedulable_data { 7796 struct task_group *tg; 7797 u64 period, quota; 7798}; 7799 7800/* 7801 * normalize group quota/period to be quota/max_period 7802 * note: units are usecs 7803 */ 7804static u64 normalize_cfs_quota(struct task_group *tg, 7805 struct cfs_schedulable_data *d) 7806{ 7807 u64 quota, period; 7808 7809 if (tg == d->tg) { 7810 period = d->period; 7811 quota = d->quota; 7812 } else { 7813 period = tg_get_cfs_period(tg); 7814 quota = tg_get_cfs_quota(tg); 7815 } 7816 7817 /* note: these should typically be equivalent */ 7818 if (quota == RUNTIME_INF || quota == -1) 7819 return RUNTIME_INF; 7820 7821 return to_ratio(period, quota); 7822} 7823 7824static int tg_cfs_schedulable_down(struct task_group *tg, void *data) 7825{ 7826 struct cfs_schedulable_data *d = data; 7827 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7828 s64 quota = 0, parent_quota = -1; 7829 7830 if (!tg->parent) { 7831 quota = RUNTIME_INF; 7832 } else { 7833 struct cfs_bandwidth *parent_b = &tg->parent->cfs_bandwidth; 7834 7835 quota = normalize_cfs_quota(tg, d); 7836 parent_quota = parent_b->hierarchal_quota; 7837 7838 /* 7839 * ensure max(child_quota) <= parent_quota, inherit when no 7840 * limit is set 7841 */ 7842 if (quota == RUNTIME_INF) 7843 quota = parent_quota; 7844 else if (parent_quota != RUNTIME_INF && quota > parent_quota) 7845 return -EINVAL; 7846 } 7847 cfs_b->hierarchal_quota = quota; 7848 7849 return 0; 7850} 7851 7852static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota) 7853{ 7854 int ret; 7855 struct cfs_schedulable_data data = { 7856 .tg = tg, 7857 .period = period, 7858 .quota = quota, 7859 }; 7860 7861 if (quota != RUNTIME_INF) { 7862 do_div(data.period, NSEC_PER_USEC); 7863 do_div(data.quota, NSEC_PER_USEC); 7864 } 7865 7866 rcu_read_lock(); 7867 ret = walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data); 7868 rcu_read_unlock(); 7869 7870 return ret; 7871} 7872 7873static int cpu_stats_show(struct cgroup *cgrp, struct cftype *cft, 7874 struct cgroup_map_cb *cb) 7875{ 7876 struct task_group *tg = cgroup_tg(cgrp); 7877 struct cfs_bandwidth *cfs_b = &tg->cfs_bandwidth; 7878 7879 cb->fill(cb, "nr_periods", cfs_b->nr_periods); 7880 cb->fill(cb, "nr_throttled", cfs_b->nr_throttled); 7881 cb->fill(cb, "throttled_time", cfs_b->throttled_time); 7882 7883 return 0; 7884} 7885#endif /* CONFIG_CFS_BANDWIDTH */ 7886#endif /* CONFIG_FAIR_GROUP_SCHED */ 7887 7888#ifdef CONFIG_RT_GROUP_SCHED 7889static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft, 7890 s64 val) 7891{ 7892 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val); 7893} 7894 7895static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft) 7896{ 7897 return sched_group_rt_runtime(cgroup_tg(cgrp)); 7898} 7899 7900static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype, 7901 u64 rt_period_us) 7902{ 7903 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us); 7904} 7905 7906static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft) 7907{ 7908 return sched_group_rt_period(cgroup_tg(cgrp)); 7909} 7910#endif /* CONFIG_RT_GROUP_SCHED */ 7911 7912static struct cftype cpu_files[] = { 7913#ifdef CONFIG_FAIR_GROUP_SCHED 7914 { 7915 .name = "shares", 7916 .read_u64 = cpu_shares_read_u64, 7917 .write_u64 = cpu_shares_write_u64, 7918 }, 7919#endif 7920#ifdef CONFIG_CFS_BANDWIDTH 7921 { 7922 .name = "cfs_quota_us", 7923 .read_s64 = cpu_cfs_quota_read_s64, 7924 .write_s64 = cpu_cfs_quota_write_s64, 7925 }, 7926 { 7927 .name = "cfs_period_us", 7928 .read_u64 = cpu_cfs_period_read_u64, 7929 .write_u64 = cpu_cfs_period_write_u64, 7930 }, 7931 { 7932 .name = "stat", 7933 .read_map = cpu_stats_show, 7934 }, 7935#endif 7936#ifdef CONFIG_RT_GROUP_SCHED 7937 { 7938 .name = "rt_runtime_us", 7939 .read_s64 = cpu_rt_runtime_read, 7940 .write_s64 = cpu_rt_runtime_write, 7941 }, 7942 { 7943 .name = "rt_period_us", 7944 .read_u64 = cpu_rt_period_read_uint, 7945 .write_u64 = cpu_rt_period_write_uint, 7946 }, 7947#endif 7948}; 7949 7950static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont) 7951{ 7952 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files)); 7953} 7954 7955struct cgroup_subsys cpu_cgroup_subsys = { 7956 .name = "cpu", 7957 .create = cpu_cgroup_create, 7958 .destroy = cpu_cgroup_destroy, 7959 .can_attach = cpu_cgroup_can_attach, 7960 .attach = cpu_cgroup_attach, 7961 .exit = cpu_cgroup_exit, 7962 .populate = cpu_cgroup_populate, 7963 .subsys_id = cpu_cgroup_subsys_id, 7964 .early_init = 1, 7965}; 7966 7967#endif /* CONFIG_CGROUP_SCHED */ 7968 7969#ifdef CONFIG_CGROUP_CPUACCT 7970 7971/* 7972 * CPU accounting code for task groups. 7973 * 7974 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh 7975 * (balbir@in.ibm.com). 7976 */ 7977 7978/* create a new cpu accounting group */ 7979static struct cgroup_subsys_state *cpuacct_create( 7980 struct cgroup_subsys *ss, struct cgroup *cgrp) 7981{ 7982 struct cpuacct *ca; 7983 7984 if (!cgrp->parent) 7985 return &root_cpuacct.css; 7986 7987 ca = kzalloc(sizeof(*ca), GFP_KERNEL); 7988 if (!ca) 7989 goto out; 7990 7991 ca->cpuusage = alloc_percpu(u64); 7992 if (!ca->cpuusage) 7993 goto out_free_ca; 7994 7995 ca->cpustat = alloc_percpu(struct kernel_cpustat); 7996 if (!ca->cpustat) 7997 goto out_free_cpuusage; 7998 7999 return &ca->css; 8000 8001out_free_cpuusage: 8002 free_percpu(ca->cpuusage); 8003out_free_ca: 8004 kfree(ca); 8005out: 8006 return ERR_PTR(-ENOMEM); 8007} 8008 8009/* destroy an existing cpu accounting group */ 8010static void 8011cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp) 8012{ 8013 struct cpuacct *ca = cgroup_ca(cgrp); 8014 8015 free_percpu(ca->cpustat); 8016 free_percpu(ca->cpuusage); 8017 kfree(ca); 8018} 8019 8020static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu) 8021{ 8022 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu); 8023 u64 data; 8024 8025#ifndef CONFIG_64BIT 8026 /* 8027 * Take rq->lock to make 64-bit read safe on 32-bit platforms. 8028 */ 8029 raw_spin_lock_irq(&cpu_rq(cpu)->lock); 8030 data = *cpuusage; 8031 raw_spin_unlock_irq(&cpu_rq(cpu)->lock); 8032#else 8033 data = *cpuusage; 8034#endif 8035 8036 return data; 8037} 8038 8039static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val) 8040{ 8041 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu); 8042 8043#ifndef CONFIG_64BIT 8044 /* 8045 * Take rq->lock to make 64-bit write safe on 32-bit platforms. 8046 */ 8047 raw_spin_lock_irq(&cpu_rq(cpu)->lock); 8048 *cpuusage = val; 8049 raw_spin_unlock_irq(&cpu_rq(cpu)->lock); 8050#else 8051 *cpuusage = val; 8052#endif 8053} 8054 8055/* return total cpu usage (in nanoseconds) of a group */ 8056static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft) 8057{ 8058 struct cpuacct *ca = cgroup_ca(cgrp); 8059 u64 totalcpuusage = 0; 8060 int i; 8061 8062 for_each_present_cpu(i) 8063 totalcpuusage += cpuacct_cpuusage_read(ca, i); 8064 8065 return totalcpuusage; 8066} 8067 8068static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype, 8069 u64 reset) 8070{ 8071 struct cpuacct *ca = cgroup_ca(cgrp); 8072 int err = 0; 8073 int i; 8074 8075 if (reset) { 8076 err = -EINVAL; 8077 goto out; 8078 } 8079 8080 for_each_present_cpu(i) 8081 cpuacct_cpuusage_write(ca, i, 0); 8082 8083out: 8084 return err; 8085} 8086 8087static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft, 8088 struct seq_file *m) 8089{ 8090 struct cpuacct *ca = cgroup_ca(cgroup); 8091 u64 percpu; 8092 int i; 8093 8094 for_each_present_cpu(i) { 8095 percpu = cpuacct_cpuusage_read(ca, i); 8096 seq_printf(m, "%llu ", (unsigned long long) percpu); 8097 } 8098 seq_printf(m, "\n"); 8099 return 0; 8100} 8101 8102static const char *cpuacct_stat_desc[] = { 8103 [CPUACCT_STAT_USER] = "user", 8104 [CPUACCT_STAT_SYSTEM] = "system", 8105}; 8106 8107static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft, 8108 struct cgroup_map_cb *cb) 8109{ 8110 struct cpuacct *ca = cgroup_ca(cgrp); 8111 int cpu; 8112 s64 val = 0; 8113 8114 for_each_online_cpu(cpu) { 8115 struct kernel_cpustat *kcpustat = per_cpu_ptr(ca->cpustat, cpu); 8116 val += kcpustat->cpustat[CPUTIME_USER]; 8117 val += kcpustat->cpustat[CPUTIME_NICE]; 8118 } 8119 val = cputime64_to_clock_t(val); 8120 cb->fill(cb, cpuacct_stat_desc[CPUACCT_STAT_USER], val); 8121 8122 val = 0; 8123 for_each_online_cpu(cpu) { 8124 struct kernel_cpustat *kcpustat = per_cpu_ptr(ca->cpustat, cpu); 8125 val += kcpustat->cpustat[CPUTIME_SYSTEM]; 8126 val += kcpustat->cpustat[CPUTIME_IRQ]; 8127 val += kcpustat->cpustat[CPUTIME_SOFTIRQ]; 8128 } 8129 8130 val = cputime64_to_clock_t(val); 8131 cb->fill(cb, cpuacct_stat_desc[CPUACCT_STAT_SYSTEM], val); 8132 8133 return 0; 8134} 8135 8136static struct cftype files[] = { 8137 { 8138 .name = "usage", 8139 .read_u64 = cpuusage_read, 8140 .write_u64 = cpuusage_write, 8141 }, 8142 { 8143 .name = "usage_percpu", 8144 .read_seq_string = cpuacct_percpu_seq_read, 8145 }, 8146 { 8147 .name = "stat", 8148 .read_map = cpuacct_stats_show, 8149 }, 8150}; 8151 8152static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp) 8153{ 8154 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files)); 8155} 8156 8157/* 8158 * charge this task's execution time to its accounting group. 8159 * 8160 * called with rq->lock held. 8161 */ 8162void cpuacct_charge(struct task_struct *tsk, u64 cputime) 8163{ 8164 struct cpuacct *ca; 8165 int cpu; 8166 8167 if (unlikely(!cpuacct_subsys.active)) 8168 return; 8169 8170 cpu = task_cpu(tsk); 8171 8172 rcu_read_lock(); 8173 8174 ca = task_ca(tsk); 8175 8176 for (; ca; ca = parent_ca(ca)) { 8177 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu); 8178 *cpuusage += cputime; 8179 } 8180 8181 rcu_read_unlock(); 8182} 8183 8184struct cgroup_subsys cpuacct_subsys = { 8185 .name = "cpuacct", 8186 .create = cpuacct_create, 8187 .destroy = cpuacct_destroy, 8188 .populate = cpuacct_populate, 8189 .subsys_id = cpuacct_subsys_id, 8190}; 8191#endif /* CONFIG_CGROUP_CPUACCT */ 8192