core.c revision 96f951edb1f1bdbbc99b0cd458f9808bb83d58ae
1/* 2 * kernel/sched/core.c 3 * 4 * Kernel scheduler and related syscalls 5 * 6 * Copyright (C) 1991-2002 Linus Torvalds 7 * 8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and 9 * make semaphores SMP safe 10 * 1998-11-19 Implemented schedule_timeout() and related stuff 11 * by Andrea Arcangeli 12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar: 13 * hybrid priority-list and round-robin design with 14 * an array-switch method of distributing timeslices 15 * and per-CPU runqueues. Cleanups and useful suggestions 16 * by Davide Libenzi, preemptible kernel bits by Robert Love. 17 * 2003-09-03 Interactivity tuning by Con Kolivas. 18 * 2004-04-02 Scheduler domains code by Nick Piggin 19 * 2007-04-15 Work begun on replacing all interactivity tuning with a 20 * fair scheduling design by Con Kolivas. 21 * 2007-05-05 Load balancing (smp-nice) and other improvements 22 * by Peter Williams 23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith 24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri 25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins, 26 * Thomas Gleixner, Mike Kravetz 27 */ 28 29#include <linux/mm.h> 30#include <linux/module.h> 31#include <linux/nmi.h> 32#include <linux/init.h> 33#include <linux/uaccess.h> 34#include <linux/highmem.h> 35#include <asm/mmu_context.h> 36#include <linux/interrupt.h> 37#include <linux/capability.h> 38#include <linux/completion.h> 39#include <linux/kernel_stat.h> 40#include <linux/debug_locks.h> 41#include <linux/perf_event.h> 42#include <linux/security.h> 43#include <linux/notifier.h> 44#include <linux/profile.h> 45#include <linux/freezer.h> 46#include <linux/vmalloc.h> 47#include <linux/blkdev.h> 48#include <linux/delay.h> 49#include <linux/pid_namespace.h> 50#include <linux/smp.h> 51#include <linux/threads.h> 52#include <linux/timer.h> 53#include <linux/rcupdate.h> 54#include <linux/cpu.h> 55#include <linux/cpuset.h> 56#include <linux/percpu.h> 57#include <linux/proc_fs.h> 58#include <linux/seq_file.h> 59#include <linux/sysctl.h> 60#include <linux/syscalls.h> 61#include <linux/times.h> 62#include <linux/tsacct_kern.h> 63#include <linux/kprobes.h> 64#include <linux/delayacct.h> 65#include <linux/unistd.h> 66#include <linux/pagemap.h> 67#include <linux/hrtimer.h> 68#include <linux/tick.h> 69#include <linux/debugfs.h> 70#include <linux/ctype.h> 71#include <linux/ftrace.h> 72#include <linux/slab.h> 73#include <linux/init_task.h> 74#include <linux/binfmts.h> 75 76#include <asm/switch_to.h> 77#include <asm/tlb.h> 78#include <asm/irq_regs.h> 79#include <asm/mutex.h> 80#ifdef CONFIG_PARAVIRT 81#include <asm/paravirt.h> 82#endif 83 84#include "sched.h" 85#include "../workqueue_sched.h" 86 87#define CREATE_TRACE_POINTS 88#include <trace/events/sched.h> 89 90void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period) 91{ 92 unsigned long delta; 93 ktime_t soft, hard, now; 94 95 for (;;) { 96 if (hrtimer_active(period_timer)) 97 break; 98 99 now = hrtimer_cb_get_time(period_timer); 100 hrtimer_forward(period_timer, now, period); 101 102 soft = hrtimer_get_softexpires(period_timer); 103 hard = hrtimer_get_expires(period_timer); 104 delta = ktime_to_ns(ktime_sub(hard, soft)); 105 __hrtimer_start_range_ns(period_timer, soft, delta, 106 HRTIMER_MODE_ABS_PINNED, 0); 107 } 108} 109 110DEFINE_MUTEX(sched_domains_mutex); 111DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues); 112 113static void update_rq_clock_task(struct rq *rq, s64 delta); 114 115void update_rq_clock(struct rq *rq) 116{ 117 s64 delta; 118 119 if (rq->skip_clock_update > 0) 120 return; 121 122 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock; 123 rq->clock += delta; 124 update_rq_clock_task(rq, delta); 125} 126 127/* 128 * Debugging: various feature bits 129 */ 130 131#define SCHED_FEAT(name, enabled) \ 132 (1UL << __SCHED_FEAT_##name) * enabled | 133 134const_debug unsigned int sysctl_sched_features = 135#include "features.h" 136 0; 137 138#undef SCHED_FEAT 139 140#ifdef CONFIG_SCHED_DEBUG 141#define SCHED_FEAT(name, enabled) \ 142 #name , 143 144static __read_mostly char *sched_feat_names[] = { 145#include "features.h" 146 NULL 147}; 148 149#undef SCHED_FEAT 150 151static int sched_feat_show(struct seq_file *m, void *v) 152{ 153 int i; 154 155 for (i = 0; i < __SCHED_FEAT_NR; i++) { 156 if (!(sysctl_sched_features & (1UL << i))) 157 seq_puts(m, "NO_"); 158 seq_printf(m, "%s ", sched_feat_names[i]); 159 } 160 seq_puts(m, "\n"); 161 162 return 0; 163} 164 165#ifdef HAVE_JUMP_LABEL 166 167#define jump_label_key__true STATIC_KEY_INIT_TRUE 168#define jump_label_key__false STATIC_KEY_INIT_FALSE 169 170#define SCHED_FEAT(name, enabled) \ 171 jump_label_key__##enabled , 172 173struct static_key sched_feat_keys[__SCHED_FEAT_NR] = { 174#include "features.h" 175}; 176 177#undef SCHED_FEAT 178 179static void sched_feat_disable(int i) 180{ 181 if (static_key_enabled(&sched_feat_keys[i])) 182 static_key_slow_dec(&sched_feat_keys[i]); 183} 184 185static void sched_feat_enable(int i) 186{ 187 if (!static_key_enabled(&sched_feat_keys[i])) 188 static_key_slow_inc(&sched_feat_keys[i]); 189} 190#else 191static void sched_feat_disable(int i) { }; 192static void sched_feat_enable(int i) { }; 193#endif /* HAVE_JUMP_LABEL */ 194 195static ssize_t 196sched_feat_write(struct file *filp, const char __user *ubuf, 197 size_t cnt, loff_t *ppos) 198{ 199 char buf[64]; 200 char *cmp; 201 int neg = 0; 202 int i; 203 204 if (cnt > 63) 205 cnt = 63; 206 207 if (copy_from_user(&buf, ubuf, cnt)) 208 return -EFAULT; 209 210 buf[cnt] = 0; 211 cmp = strstrip(buf); 212 213 if (strncmp(cmp, "NO_", 3) == 0) { 214 neg = 1; 215 cmp += 3; 216 } 217 218 for (i = 0; i < __SCHED_FEAT_NR; i++) { 219 if (strcmp(cmp, sched_feat_names[i]) == 0) { 220 if (neg) { 221 sysctl_sched_features &= ~(1UL << i); 222 sched_feat_disable(i); 223 } else { 224 sysctl_sched_features |= (1UL << i); 225 sched_feat_enable(i); 226 } 227 break; 228 } 229 } 230 231 if (i == __SCHED_FEAT_NR) 232 return -EINVAL; 233 234 *ppos += cnt; 235 236 return cnt; 237} 238 239static int sched_feat_open(struct inode *inode, struct file *filp) 240{ 241 return single_open(filp, sched_feat_show, NULL); 242} 243 244static const struct file_operations sched_feat_fops = { 245 .open = sched_feat_open, 246 .write = sched_feat_write, 247 .read = seq_read, 248 .llseek = seq_lseek, 249 .release = single_release, 250}; 251 252static __init int sched_init_debug(void) 253{ 254 debugfs_create_file("sched_features", 0644, NULL, NULL, 255 &sched_feat_fops); 256 257 return 0; 258} 259late_initcall(sched_init_debug); 260#endif /* CONFIG_SCHED_DEBUG */ 261 262/* 263 * Number of tasks to iterate in a single balance run. 264 * Limited because this is done with IRQs disabled. 265 */ 266const_debug unsigned int sysctl_sched_nr_migrate = 32; 267 268/* 269 * period over which we average the RT time consumption, measured 270 * in ms. 271 * 272 * default: 1s 273 */ 274const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC; 275 276/* 277 * period over which we measure -rt task cpu usage in us. 278 * default: 1s 279 */ 280unsigned int sysctl_sched_rt_period = 1000000; 281 282__read_mostly int scheduler_running; 283 284/* 285 * part of the period that we allow rt tasks to run in us. 286 * default: 0.95s 287 */ 288int sysctl_sched_rt_runtime = 950000; 289 290 291 292/* 293 * __task_rq_lock - lock the rq @p resides on. 294 */ 295static inline struct rq *__task_rq_lock(struct task_struct *p) 296 __acquires(rq->lock) 297{ 298 struct rq *rq; 299 300 lockdep_assert_held(&p->pi_lock); 301 302 for (;;) { 303 rq = task_rq(p); 304 raw_spin_lock(&rq->lock); 305 if (likely(rq == task_rq(p))) 306 return rq; 307 raw_spin_unlock(&rq->lock); 308 } 309} 310 311/* 312 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on. 313 */ 314static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags) 315 __acquires(p->pi_lock) 316 __acquires(rq->lock) 317{ 318 struct rq *rq; 319 320 for (;;) { 321 raw_spin_lock_irqsave(&p->pi_lock, *flags); 322 rq = task_rq(p); 323 raw_spin_lock(&rq->lock); 324 if (likely(rq == task_rq(p))) 325 return rq; 326 raw_spin_unlock(&rq->lock); 327 raw_spin_unlock_irqrestore(&p->pi_lock, *flags); 328 } 329} 330 331static void __task_rq_unlock(struct rq *rq) 332 __releases(rq->lock) 333{ 334 raw_spin_unlock(&rq->lock); 335} 336 337static inline void 338task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags) 339 __releases(rq->lock) 340 __releases(p->pi_lock) 341{ 342 raw_spin_unlock(&rq->lock); 343 raw_spin_unlock_irqrestore(&p->pi_lock, *flags); 344} 345 346/* 347 * this_rq_lock - lock this runqueue and disable interrupts. 348 */ 349static struct rq *this_rq_lock(void) 350 __acquires(rq->lock) 351{ 352 struct rq *rq; 353 354 local_irq_disable(); 355 rq = this_rq(); 356 raw_spin_lock(&rq->lock); 357 358 return rq; 359} 360 361#ifdef CONFIG_SCHED_HRTICK 362/* 363 * Use HR-timers to deliver accurate preemption points. 364 * 365 * Its all a bit involved since we cannot program an hrt while holding the 366 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a 367 * reschedule event. 368 * 369 * When we get rescheduled we reprogram the hrtick_timer outside of the 370 * rq->lock. 371 */ 372 373static void hrtick_clear(struct rq *rq) 374{ 375 if (hrtimer_active(&rq->hrtick_timer)) 376 hrtimer_cancel(&rq->hrtick_timer); 377} 378 379/* 380 * High-resolution timer tick. 381 * Runs from hardirq context with interrupts disabled. 382 */ 383static enum hrtimer_restart hrtick(struct hrtimer *timer) 384{ 385 struct rq *rq = container_of(timer, struct rq, hrtick_timer); 386 387 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id()); 388 389 raw_spin_lock(&rq->lock); 390 update_rq_clock(rq); 391 rq->curr->sched_class->task_tick(rq, rq->curr, 1); 392 raw_spin_unlock(&rq->lock); 393 394 return HRTIMER_NORESTART; 395} 396 397#ifdef CONFIG_SMP 398/* 399 * called from hardirq (IPI) context 400 */ 401static void __hrtick_start(void *arg) 402{ 403 struct rq *rq = arg; 404 405 raw_spin_lock(&rq->lock); 406 hrtimer_restart(&rq->hrtick_timer); 407 rq->hrtick_csd_pending = 0; 408 raw_spin_unlock(&rq->lock); 409} 410 411/* 412 * Called to set the hrtick timer state. 413 * 414 * called with rq->lock held and irqs disabled 415 */ 416void hrtick_start(struct rq *rq, u64 delay) 417{ 418 struct hrtimer *timer = &rq->hrtick_timer; 419 ktime_t time = ktime_add_ns(timer->base->get_time(), delay); 420 421 hrtimer_set_expires(timer, time); 422 423 if (rq == this_rq()) { 424 hrtimer_restart(timer); 425 } else if (!rq->hrtick_csd_pending) { 426 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0); 427 rq->hrtick_csd_pending = 1; 428 } 429} 430 431static int 432hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu) 433{ 434 int cpu = (int)(long)hcpu; 435 436 switch (action) { 437 case CPU_UP_CANCELED: 438 case CPU_UP_CANCELED_FROZEN: 439 case CPU_DOWN_PREPARE: 440 case CPU_DOWN_PREPARE_FROZEN: 441 case CPU_DEAD: 442 case CPU_DEAD_FROZEN: 443 hrtick_clear(cpu_rq(cpu)); 444 return NOTIFY_OK; 445 } 446 447 return NOTIFY_DONE; 448} 449 450static __init void init_hrtick(void) 451{ 452 hotcpu_notifier(hotplug_hrtick, 0); 453} 454#else 455/* 456 * Called to set the hrtick timer state. 457 * 458 * called with rq->lock held and irqs disabled 459 */ 460void hrtick_start(struct rq *rq, u64 delay) 461{ 462 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0, 463 HRTIMER_MODE_REL_PINNED, 0); 464} 465 466static inline void init_hrtick(void) 467{ 468} 469#endif /* CONFIG_SMP */ 470 471static void init_rq_hrtick(struct rq *rq) 472{ 473#ifdef CONFIG_SMP 474 rq->hrtick_csd_pending = 0; 475 476 rq->hrtick_csd.flags = 0; 477 rq->hrtick_csd.func = __hrtick_start; 478 rq->hrtick_csd.info = rq; 479#endif 480 481 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL); 482 rq->hrtick_timer.function = hrtick; 483} 484#else /* CONFIG_SCHED_HRTICK */ 485static inline void hrtick_clear(struct rq *rq) 486{ 487} 488 489static inline void init_rq_hrtick(struct rq *rq) 490{ 491} 492 493static inline void init_hrtick(void) 494{ 495} 496#endif /* CONFIG_SCHED_HRTICK */ 497 498/* 499 * resched_task - mark a task 'to be rescheduled now'. 500 * 501 * On UP this means the setting of the need_resched flag, on SMP it 502 * might also involve a cross-CPU call to trigger the scheduler on 503 * the target CPU. 504 */ 505#ifdef CONFIG_SMP 506 507#ifndef tsk_is_polling 508#define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG) 509#endif 510 511void resched_task(struct task_struct *p) 512{ 513 int cpu; 514 515 assert_raw_spin_locked(&task_rq(p)->lock); 516 517 if (test_tsk_need_resched(p)) 518 return; 519 520 set_tsk_need_resched(p); 521 522 cpu = task_cpu(p); 523 if (cpu == smp_processor_id()) 524 return; 525 526 /* NEED_RESCHED must be visible before we test polling */ 527 smp_mb(); 528 if (!tsk_is_polling(p)) 529 smp_send_reschedule(cpu); 530} 531 532void resched_cpu(int cpu) 533{ 534 struct rq *rq = cpu_rq(cpu); 535 unsigned long flags; 536 537 if (!raw_spin_trylock_irqsave(&rq->lock, flags)) 538 return; 539 resched_task(cpu_curr(cpu)); 540 raw_spin_unlock_irqrestore(&rq->lock, flags); 541} 542 543#ifdef CONFIG_NO_HZ 544/* 545 * In the semi idle case, use the nearest busy cpu for migrating timers 546 * from an idle cpu. This is good for power-savings. 547 * 548 * We don't do similar optimization for completely idle system, as 549 * selecting an idle cpu will add more delays to the timers than intended 550 * (as that cpu's timer base may not be uptodate wrt jiffies etc). 551 */ 552int get_nohz_timer_target(void) 553{ 554 int cpu = smp_processor_id(); 555 int i; 556 struct sched_domain *sd; 557 558 rcu_read_lock(); 559 for_each_domain(cpu, sd) { 560 for_each_cpu(i, sched_domain_span(sd)) { 561 if (!idle_cpu(i)) { 562 cpu = i; 563 goto unlock; 564 } 565 } 566 } 567unlock: 568 rcu_read_unlock(); 569 return cpu; 570} 571/* 572 * When add_timer_on() enqueues a timer into the timer wheel of an 573 * idle CPU then this timer might expire before the next timer event 574 * which is scheduled to wake up that CPU. In case of a completely 575 * idle system the next event might even be infinite time into the 576 * future. wake_up_idle_cpu() ensures that the CPU is woken up and 577 * leaves the inner idle loop so the newly added timer is taken into 578 * account when the CPU goes back to idle and evaluates the timer 579 * wheel for the next timer event. 580 */ 581void wake_up_idle_cpu(int cpu) 582{ 583 struct rq *rq = cpu_rq(cpu); 584 585 if (cpu == smp_processor_id()) 586 return; 587 588 /* 589 * This is safe, as this function is called with the timer 590 * wheel base lock of (cpu) held. When the CPU is on the way 591 * to idle and has not yet set rq->curr to idle then it will 592 * be serialized on the timer wheel base lock and take the new 593 * timer into account automatically. 594 */ 595 if (rq->curr != rq->idle) 596 return; 597 598 /* 599 * We can set TIF_RESCHED on the idle task of the other CPU 600 * lockless. The worst case is that the other CPU runs the 601 * idle task through an additional NOOP schedule() 602 */ 603 set_tsk_need_resched(rq->idle); 604 605 /* NEED_RESCHED must be visible before we test polling */ 606 smp_mb(); 607 if (!tsk_is_polling(rq->idle)) 608 smp_send_reschedule(cpu); 609} 610 611static inline bool got_nohz_idle_kick(void) 612{ 613 int cpu = smp_processor_id(); 614 return idle_cpu(cpu) && test_bit(NOHZ_BALANCE_KICK, nohz_flags(cpu)); 615} 616 617#else /* CONFIG_NO_HZ */ 618 619static inline bool got_nohz_idle_kick(void) 620{ 621 return false; 622} 623 624#endif /* CONFIG_NO_HZ */ 625 626void sched_avg_update(struct rq *rq) 627{ 628 s64 period = sched_avg_period(); 629 630 while ((s64)(rq->clock - rq->age_stamp) > period) { 631 /* 632 * Inline assembly required to prevent the compiler 633 * optimising this loop into a divmod call. 634 * See __iter_div_u64_rem() for another example of this. 635 */ 636 asm("" : "+rm" (rq->age_stamp)); 637 rq->age_stamp += period; 638 rq->rt_avg /= 2; 639 } 640} 641 642#else /* !CONFIG_SMP */ 643void resched_task(struct task_struct *p) 644{ 645 assert_raw_spin_locked(&task_rq(p)->lock); 646 set_tsk_need_resched(p); 647} 648#endif /* CONFIG_SMP */ 649 650#if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \ 651 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH))) 652/* 653 * Iterate task_group tree rooted at *from, calling @down when first entering a 654 * node and @up when leaving it for the final time. 655 * 656 * Caller must hold rcu_lock or sufficient equivalent. 657 */ 658int walk_tg_tree_from(struct task_group *from, 659 tg_visitor down, tg_visitor up, void *data) 660{ 661 struct task_group *parent, *child; 662 int ret; 663 664 parent = from; 665 666down: 667 ret = (*down)(parent, data); 668 if (ret) 669 goto out; 670 list_for_each_entry_rcu(child, &parent->children, siblings) { 671 parent = child; 672 goto down; 673 674up: 675 continue; 676 } 677 ret = (*up)(parent, data); 678 if (ret || parent == from) 679 goto out; 680 681 child = parent; 682 parent = parent->parent; 683 if (parent) 684 goto up; 685out: 686 return ret; 687} 688 689int tg_nop(struct task_group *tg, void *data) 690{ 691 return 0; 692} 693#endif 694 695void update_cpu_load(struct rq *this_rq); 696 697static void set_load_weight(struct task_struct *p) 698{ 699 int prio = p->static_prio - MAX_RT_PRIO; 700 struct load_weight *load = &p->se.load; 701 702 /* 703 * SCHED_IDLE tasks get minimal weight: 704 */ 705 if (p->policy == SCHED_IDLE) { 706 load->weight = scale_load(WEIGHT_IDLEPRIO); 707 load->inv_weight = WMULT_IDLEPRIO; 708 return; 709 } 710 711 load->weight = scale_load(prio_to_weight[prio]); 712 load->inv_weight = prio_to_wmult[prio]; 713} 714 715static void enqueue_task(struct rq *rq, struct task_struct *p, int flags) 716{ 717 update_rq_clock(rq); 718 sched_info_queued(p); 719 p->sched_class->enqueue_task(rq, p, flags); 720} 721 722static void dequeue_task(struct rq *rq, struct task_struct *p, int flags) 723{ 724 update_rq_clock(rq); 725 sched_info_dequeued(p); 726 p->sched_class->dequeue_task(rq, p, flags); 727} 728 729void activate_task(struct rq *rq, struct task_struct *p, int flags) 730{ 731 if (task_contributes_to_load(p)) 732 rq->nr_uninterruptible--; 733 734 enqueue_task(rq, p, flags); 735} 736 737void deactivate_task(struct rq *rq, struct task_struct *p, int flags) 738{ 739 if (task_contributes_to_load(p)) 740 rq->nr_uninterruptible++; 741 742 dequeue_task(rq, p, flags); 743} 744 745#ifdef CONFIG_IRQ_TIME_ACCOUNTING 746 747/* 748 * There are no locks covering percpu hardirq/softirq time. 749 * They are only modified in account_system_vtime, on corresponding CPU 750 * with interrupts disabled. So, writes are safe. 751 * They are read and saved off onto struct rq in update_rq_clock(). 752 * This may result in other CPU reading this CPU's irq time and can 753 * race with irq/account_system_vtime on this CPU. We would either get old 754 * or new value with a side effect of accounting a slice of irq time to wrong 755 * task when irq is in progress while we read rq->clock. That is a worthy 756 * compromise in place of having locks on each irq in account_system_time. 757 */ 758static DEFINE_PER_CPU(u64, cpu_hardirq_time); 759static DEFINE_PER_CPU(u64, cpu_softirq_time); 760 761static DEFINE_PER_CPU(u64, irq_start_time); 762static int sched_clock_irqtime; 763 764void enable_sched_clock_irqtime(void) 765{ 766 sched_clock_irqtime = 1; 767} 768 769void disable_sched_clock_irqtime(void) 770{ 771 sched_clock_irqtime = 0; 772} 773 774#ifndef CONFIG_64BIT 775static DEFINE_PER_CPU(seqcount_t, irq_time_seq); 776 777static inline void irq_time_write_begin(void) 778{ 779 __this_cpu_inc(irq_time_seq.sequence); 780 smp_wmb(); 781} 782 783static inline void irq_time_write_end(void) 784{ 785 smp_wmb(); 786 __this_cpu_inc(irq_time_seq.sequence); 787} 788 789static inline u64 irq_time_read(int cpu) 790{ 791 u64 irq_time; 792 unsigned seq; 793 794 do { 795 seq = read_seqcount_begin(&per_cpu(irq_time_seq, cpu)); 796 irq_time = per_cpu(cpu_softirq_time, cpu) + 797 per_cpu(cpu_hardirq_time, cpu); 798 } while (read_seqcount_retry(&per_cpu(irq_time_seq, cpu), seq)); 799 800 return irq_time; 801} 802#else /* CONFIG_64BIT */ 803static inline void irq_time_write_begin(void) 804{ 805} 806 807static inline void irq_time_write_end(void) 808{ 809} 810 811static inline u64 irq_time_read(int cpu) 812{ 813 return per_cpu(cpu_softirq_time, cpu) + per_cpu(cpu_hardirq_time, cpu); 814} 815#endif /* CONFIG_64BIT */ 816 817/* 818 * Called before incrementing preempt_count on {soft,}irq_enter 819 * and before decrementing preempt_count on {soft,}irq_exit. 820 */ 821void account_system_vtime(struct task_struct *curr) 822{ 823 unsigned long flags; 824 s64 delta; 825 int cpu; 826 827 if (!sched_clock_irqtime) 828 return; 829 830 local_irq_save(flags); 831 832 cpu = smp_processor_id(); 833 delta = sched_clock_cpu(cpu) - __this_cpu_read(irq_start_time); 834 __this_cpu_add(irq_start_time, delta); 835 836 irq_time_write_begin(); 837 /* 838 * We do not account for softirq time from ksoftirqd here. 839 * We want to continue accounting softirq time to ksoftirqd thread 840 * in that case, so as not to confuse scheduler with a special task 841 * that do not consume any time, but still wants to run. 842 */ 843 if (hardirq_count()) 844 __this_cpu_add(cpu_hardirq_time, delta); 845 else if (in_serving_softirq() && curr != this_cpu_ksoftirqd()) 846 __this_cpu_add(cpu_softirq_time, delta); 847 848 irq_time_write_end(); 849 local_irq_restore(flags); 850} 851EXPORT_SYMBOL_GPL(account_system_vtime); 852 853#endif /* CONFIG_IRQ_TIME_ACCOUNTING */ 854 855#ifdef CONFIG_PARAVIRT 856static inline u64 steal_ticks(u64 steal) 857{ 858 if (unlikely(steal > NSEC_PER_SEC)) 859 return div_u64(steal, TICK_NSEC); 860 861 return __iter_div_u64_rem(steal, TICK_NSEC, &steal); 862} 863#endif 864 865static void update_rq_clock_task(struct rq *rq, s64 delta) 866{ 867/* 868 * In theory, the compile should just see 0 here, and optimize out the call 869 * to sched_rt_avg_update. But I don't trust it... 870 */ 871#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING) 872 s64 steal = 0, irq_delta = 0; 873#endif 874#ifdef CONFIG_IRQ_TIME_ACCOUNTING 875 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time; 876 877 /* 878 * Since irq_time is only updated on {soft,}irq_exit, we might run into 879 * this case when a previous update_rq_clock() happened inside a 880 * {soft,}irq region. 881 * 882 * When this happens, we stop ->clock_task and only update the 883 * prev_irq_time stamp to account for the part that fit, so that a next 884 * update will consume the rest. This ensures ->clock_task is 885 * monotonic. 886 * 887 * It does however cause some slight miss-attribution of {soft,}irq 888 * time, a more accurate solution would be to update the irq_time using 889 * the current rq->clock timestamp, except that would require using 890 * atomic ops. 891 */ 892 if (irq_delta > delta) 893 irq_delta = delta; 894 895 rq->prev_irq_time += irq_delta; 896 delta -= irq_delta; 897#endif 898#ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING 899 if (static_key_false((¶virt_steal_rq_enabled))) { 900 u64 st; 901 902 steal = paravirt_steal_clock(cpu_of(rq)); 903 steal -= rq->prev_steal_time_rq; 904 905 if (unlikely(steal > delta)) 906 steal = delta; 907 908 st = steal_ticks(steal); 909 steal = st * TICK_NSEC; 910 911 rq->prev_steal_time_rq += steal; 912 913 delta -= steal; 914 } 915#endif 916 917 rq->clock_task += delta; 918 919#if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING) 920 if ((irq_delta + steal) && sched_feat(NONTASK_POWER)) 921 sched_rt_avg_update(rq, irq_delta + steal); 922#endif 923} 924 925#ifdef CONFIG_IRQ_TIME_ACCOUNTING 926static int irqtime_account_hi_update(void) 927{ 928 u64 *cpustat = kcpustat_this_cpu->cpustat; 929 unsigned long flags; 930 u64 latest_ns; 931 int ret = 0; 932 933 local_irq_save(flags); 934 latest_ns = this_cpu_read(cpu_hardirq_time); 935 if (nsecs_to_cputime64(latest_ns) > cpustat[CPUTIME_IRQ]) 936 ret = 1; 937 local_irq_restore(flags); 938 return ret; 939} 940 941static int irqtime_account_si_update(void) 942{ 943 u64 *cpustat = kcpustat_this_cpu->cpustat; 944 unsigned long flags; 945 u64 latest_ns; 946 int ret = 0; 947 948 local_irq_save(flags); 949 latest_ns = this_cpu_read(cpu_softirq_time); 950 if (nsecs_to_cputime64(latest_ns) > cpustat[CPUTIME_SOFTIRQ]) 951 ret = 1; 952 local_irq_restore(flags); 953 return ret; 954} 955 956#else /* CONFIG_IRQ_TIME_ACCOUNTING */ 957 958#define sched_clock_irqtime (0) 959 960#endif 961 962void sched_set_stop_task(int cpu, struct task_struct *stop) 963{ 964 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 }; 965 struct task_struct *old_stop = cpu_rq(cpu)->stop; 966 967 if (stop) { 968 /* 969 * Make it appear like a SCHED_FIFO task, its something 970 * userspace knows about and won't get confused about. 971 * 972 * Also, it will make PI more or less work without too 973 * much confusion -- but then, stop work should not 974 * rely on PI working anyway. 975 */ 976 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m); 977 978 stop->sched_class = &stop_sched_class; 979 } 980 981 cpu_rq(cpu)->stop = stop; 982 983 if (old_stop) { 984 /* 985 * Reset it back to a normal scheduling class so that 986 * it can die in pieces. 987 */ 988 old_stop->sched_class = &rt_sched_class; 989 } 990} 991 992/* 993 * __normal_prio - return the priority that is based on the static prio 994 */ 995static inline int __normal_prio(struct task_struct *p) 996{ 997 return p->static_prio; 998} 999 1000/* 1001 * Calculate the expected normal priority: i.e. priority 1002 * without taking RT-inheritance into account. Might be 1003 * boosted by interactivity modifiers. Changes upon fork, 1004 * setprio syscalls, and whenever the interactivity 1005 * estimator recalculates. 1006 */ 1007static inline int normal_prio(struct task_struct *p) 1008{ 1009 int prio; 1010 1011 if (task_has_rt_policy(p)) 1012 prio = MAX_RT_PRIO-1 - p->rt_priority; 1013 else 1014 prio = __normal_prio(p); 1015 return prio; 1016} 1017 1018/* 1019 * Calculate the current priority, i.e. the priority 1020 * taken into account by the scheduler. This value might 1021 * be boosted by RT tasks, or might be boosted by 1022 * interactivity modifiers. Will be RT if the task got 1023 * RT-boosted. If not then it returns p->normal_prio. 1024 */ 1025static int effective_prio(struct task_struct *p) 1026{ 1027 p->normal_prio = normal_prio(p); 1028 /* 1029 * If we are RT tasks or we were boosted to RT priority, 1030 * keep the priority unchanged. Otherwise, update priority 1031 * to the normal priority: 1032 */ 1033 if (!rt_prio(p->prio)) 1034 return p->normal_prio; 1035 return p->prio; 1036} 1037 1038/** 1039 * task_curr - is this task currently executing on a CPU? 1040 * @p: the task in question. 1041 */ 1042inline int task_curr(const struct task_struct *p) 1043{ 1044 return cpu_curr(task_cpu(p)) == p; 1045} 1046 1047static inline void check_class_changed(struct rq *rq, struct task_struct *p, 1048 const struct sched_class *prev_class, 1049 int oldprio) 1050{ 1051 if (prev_class != p->sched_class) { 1052 if (prev_class->switched_from) 1053 prev_class->switched_from(rq, p); 1054 p->sched_class->switched_to(rq, p); 1055 } else if (oldprio != p->prio) 1056 p->sched_class->prio_changed(rq, p, oldprio); 1057} 1058 1059void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags) 1060{ 1061 const struct sched_class *class; 1062 1063 if (p->sched_class == rq->curr->sched_class) { 1064 rq->curr->sched_class->check_preempt_curr(rq, p, flags); 1065 } else { 1066 for_each_class(class) { 1067 if (class == rq->curr->sched_class) 1068 break; 1069 if (class == p->sched_class) { 1070 resched_task(rq->curr); 1071 break; 1072 } 1073 } 1074 } 1075 1076 /* 1077 * A queue event has occurred, and we're going to schedule. In 1078 * this case, we can save a useless back to back clock update. 1079 */ 1080 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr)) 1081 rq->skip_clock_update = 1; 1082} 1083 1084#ifdef CONFIG_SMP 1085void set_task_cpu(struct task_struct *p, unsigned int new_cpu) 1086{ 1087#ifdef CONFIG_SCHED_DEBUG 1088 /* 1089 * We should never call set_task_cpu() on a blocked task, 1090 * ttwu() will sort out the placement. 1091 */ 1092 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING && 1093 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE)); 1094 1095#ifdef CONFIG_LOCKDEP 1096 /* 1097 * The caller should hold either p->pi_lock or rq->lock, when changing 1098 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks. 1099 * 1100 * sched_move_task() holds both and thus holding either pins the cgroup, 1101 * see set_task_rq(). 1102 * 1103 * Furthermore, all task_rq users should acquire both locks, see 1104 * task_rq_lock(). 1105 */ 1106 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) || 1107 lockdep_is_held(&task_rq(p)->lock))); 1108#endif 1109#endif 1110 1111 trace_sched_migrate_task(p, new_cpu); 1112 1113 if (task_cpu(p) != new_cpu) { 1114 p->se.nr_migrations++; 1115 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0); 1116 } 1117 1118 __set_task_cpu(p, new_cpu); 1119} 1120 1121struct migration_arg { 1122 struct task_struct *task; 1123 int dest_cpu; 1124}; 1125 1126static int migration_cpu_stop(void *data); 1127 1128/* 1129 * wait_task_inactive - wait for a thread to unschedule. 1130 * 1131 * If @match_state is nonzero, it's the @p->state value just checked and 1132 * not expected to change. If it changes, i.e. @p might have woken up, 1133 * then return zero. When we succeed in waiting for @p to be off its CPU, 1134 * we return a positive number (its total switch count). If a second call 1135 * a short while later returns the same number, the caller can be sure that 1136 * @p has remained unscheduled the whole time. 1137 * 1138 * The caller must ensure that the task *will* unschedule sometime soon, 1139 * else this function might spin for a *long* time. This function can't 1140 * be called with interrupts off, or it may introduce deadlock with 1141 * smp_call_function() if an IPI is sent by the same process we are 1142 * waiting to become inactive. 1143 */ 1144unsigned long wait_task_inactive(struct task_struct *p, long match_state) 1145{ 1146 unsigned long flags; 1147 int running, on_rq; 1148 unsigned long ncsw; 1149 struct rq *rq; 1150 1151 for (;;) { 1152 /* 1153 * We do the initial early heuristics without holding 1154 * any task-queue locks at all. We'll only try to get 1155 * the runqueue lock when things look like they will 1156 * work out! 1157 */ 1158 rq = task_rq(p); 1159 1160 /* 1161 * If the task is actively running on another CPU 1162 * still, just relax and busy-wait without holding 1163 * any locks. 1164 * 1165 * NOTE! Since we don't hold any locks, it's not 1166 * even sure that "rq" stays as the right runqueue! 1167 * But we don't care, since "task_running()" will 1168 * return false if the runqueue has changed and p 1169 * is actually now running somewhere else! 1170 */ 1171 while (task_running(rq, p)) { 1172 if (match_state && unlikely(p->state != match_state)) 1173 return 0; 1174 cpu_relax(); 1175 } 1176 1177 /* 1178 * Ok, time to look more closely! We need the rq 1179 * lock now, to be *sure*. If we're wrong, we'll 1180 * just go back and repeat. 1181 */ 1182 rq = task_rq_lock(p, &flags); 1183 trace_sched_wait_task(p); 1184 running = task_running(rq, p); 1185 on_rq = p->on_rq; 1186 ncsw = 0; 1187 if (!match_state || p->state == match_state) 1188 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */ 1189 task_rq_unlock(rq, p, &flags); 1190 1191 /* 1192 * If it changed from the expected state, bail out now. 1193 */ 1194 if (unlikely(!ncsw)) 1195 break; 1196 1197 /* 1198 * Was it really running after all now that we 1199 * checked with the proper locks actually held? 1200 * 1201 * Oops. Go back and try again.. 1202 */ 1203 if (unlikely(running)) { 1204 cpu_relax(); 1205 continue; 1206 } 1207 1208 /* 1209 * It's not enough that it's not actively running, 1210 * it must be off the runqueue _entirely_, and not 1211 * preempted! 1212 * 1213 * So if it was still runnable (but just not actively 1214 * running right now), it's preempted, and we should 1215 * yield - it could be a while. 1216 */ 1217 if (unlikely(on_rq)) { 1218 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ); 1219 1220 set_current_state(TASK_UNINTERRUPTIBLE); 1221 schedule_hrtimeout(&to, HRTIMER_MODE_REL); 1222 continue; 1223 } 1224 1225 /* 1226 * Ahh, all good. It wasn't running, and it wasn't 1227 * runnable, which means that it will never become 1228 * running in the future either. We're all done! 1229 */ 1230 break; 1231 } 1232 1233 return ncsw; 1234} 1235 1236/*** 1237 * kick_process - kick a running thread to enter/exit the kernel 1238 * @p: the to-be-kicked thread 1239 * 1240 * Cause a process which is running on another CPU to enter 1241 * kernel-mode, without any delay. (to get signals handled.) 1242 * 1243 * NOTE: this function doesn't have to take the runqueue lock, 1244 * because all it wants to ensure is that the remote task enters 1245 * the kernel. If the IPI races and the task has been migrated 1246 * to another CPU then no harm is done and the purpose has been 1247 * achieved as well. 1248 */ 1249void kick_process(struct task_struct *p) 1250{ 1251 int cpu; 1252 1253 preempt_disable(); 1254 cpu = task_cpu(p); 1255 if ((cpu != smp_processor_id()) && task_curr(p)) 1256 smp_send_reschedule(cpu); 1257 preempt_enable(); 1258} 1259EXPORT_SYMBOL_GPL(kick_process); 1260#endif /* CONFIG_SMP */ 1261 1262#ifdef CONFIG_SMP 1263/* 1264 * ->cpus_allowed is protected by both rq->lock and p->pi_lock 1265 */ 1266static int select_fallback_rq(int cpu, struct task_struct *p) 1267{ 1268 int dest_cpu; 1269 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu)); 1270 1271 /* Look for allowed, online CPU in same node. */ 1272 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask) 1273 if (cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) 1274 return dest_cpu; 1275 1276 /* Any allowed, online CPU? */ 1277 dest_cpu = cpumask_any_and(tsk_cpus_allowed(p), cpu_active_mask); 1278 if (dest_cpu < nr_cpu_ids) 1279 return dest_cpu; 1280 1281 /* No more Mr. Nice Guy. */ 1282 dest_cpu = cpuset_cpus_allowed_fallback(p); 1283 /* 1284 * Don't tell them about moving exiting tasks or 1285 * kernel threads (both mm NULL), since they never 1286 * leave kernel. 1287 */ 1288 if (p->mm && printk_ratelimit()) { 1289 printk_sched("process %d (%s) no longer affine to cpu%d\n", 1290 task_pid_nr(p), p->comm, cpu); 1291 } 1292 1293 return dest_cpu; 1294} 1295 1296/* 1297 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable. 1298 */ 1299static inline 1300int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags) 1301{ 1302 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags); 1303 1304 /* 1305 * In order not to call set_task_cpu() on a blocking task we need 1306 * to rely on ttwu() to place the task on a valid ->cpus_allowed 1307 * cpu. 1308 * 1309 * Since this is common to all placement strategies, this lives here. 1310 * 1311 * [ this allows ->select_task() to simply return task_cpu(p) and 1312 * not worry about this generic constraint ] 1313 */ 1314 if (unlikely(!cpumask_test_cpu(cpu, tsk_cpus_allowed(p)) || 1315 !cpu_online(cpu))) 1316 cpu = select_fallback_rq(task_cpu(p), p); 1317 1318 return cpu; 1319} 1320 1321static void update_avg(u64 *avg, u64 sample) 1322{ 1323 s64 diff = sample - *avg; 1324 *avg += diff >> 3; 1325} 1326#endif 1327 1328static void 1329ttwu_stat(struct task_struct *p, int cpu, int wake_flags) 1330{ 1331#ifdef CONFIG_SCHEDSTATS 1332 struct rq *rq = this_rq(); 1333 1334#ifdef CONFIG_SMP 1335 int this_cpu = smp_processor_id(); 1336 1337 if (cpu == this_cpu) { 1338 schedstat_inc(rq, ttwu_local); 1339 schedstat_inc(p, se.statistics.nr_wakeups_local); 1340 } else { 1341 struct sched_domain *sd; 1342 1343 schedstat_inc(p, se.statistics.nr_wakeups_remote); 1344 rcu_read_lock(); 1345 for_each_domain(this_cpu, sd) { 1346 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) { 1347 schedstat_inc(sd, ttwu_wake_remote); 1348 break; 1349 } 1350 } 1351 rcu_read_unlock(); 1352 } 1353 1354 if (wake_flags & WF_MIGRATED) 1355 schedstat_inc(p, se.statistics.nr_wakeups_migrate); 1356 1357#endif /* CONFIG_SMP */ 1358 1359 schedstat_inc(rq, ttwu_count); 1360 schedstat_inc(p, se.statistics.nr_wakeups); 1361 1362 if (wake_flags & WF_SYNC) 1363 schedstat_inc(p, se.statistics.nr_wakeups_sync); 1364 1365#endif /* CONFIG_SCHEDSTATS */ 1366} 1367 1368static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags) 1369{ 1370 activate_task(rq, p, en_flags); 1371 p->on_rq = 1; 1372 1373 /* if a worker is waking up, notify workqueue */ 1374 if (p->flags & PF_WQ_WORKER) 1375 wq_worker_waking_up(p, cpu_of(rq)); 1376} 1377 1378/* 1379 * Mark the task runnable and perform wakeup-preemption. 1380 */ 1381static void 1382ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags) 1383{ 1384 trace_sched_wakeup(p, true); 1385 check_preempt_curr(rq, p, wake_flags); 1386 1387 p->state = TASK_RUNNING; 1388#ifdef CONFIG_SMP 1389 if (p->sched_class->task_woken) 1390 p->sched_class->task_woken(rq, p); 1391 1392 if (rq->idle_stamp) { 1393 u64 delta = rq->clock - rq->idle_stamp; 1394 u64 max = 2*sysctl_sched_migration_cost; 1395 1396 if (delta > max) 1397 rq->avg_idle = max; 1398 else 1399 update_avg(&rq->avg_idle, delta); 1400 rq->idle_stamp = 0; 1401 } 1402#endif 1403} 1404 1405static void 1406ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags) 1407{ 1408#ifdef CONFIG_SMP 1409 if (p->sched_contributes_to_load) 1410 rq->nr_uninterruptible--; 1411#endif 1412 1413 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING); 1414 ttwu_do_wakeup(rq, p, wake_flags); 1415} 1416 1417/* 1418 * Called in case the task @p isn't fully descheduled from its runqueue, 1419 * in this case we must do a remote wakeup. Its a 'light' wakeup though, 1420 * since all we need to do is flip p->state to TASK_RUNNING, since 1421 * the task is still ->on_rq. 1422 */ 1423static int ttwu_remote(struct task_struct *p, int wake_flags) 1424{ 1425 struct rq *rq; 1426 int ret = 0; 1427 1428 rq = __task_rq_lock(p); 1429 if (p->on_rq) { 1430 ttwu_do_wakeup(rq, p, wake_flags); 1431 ret = 1; 1432 } 1433 __task_rq_unlock(rq); 1434 1435 return ret; 1436} 1437 1438#ifdef CONFIG_SMP 1439static void sched_ttwu_pending(void) 1440{ 1441 struct rq *rq = this_rq(); 1442 struct llist_node *llist = llist_del_all(&rq->wake_list); 1443 struct task_struct *p; 1444 1445 raw_spin_lock(&rq->lock); 1446 1447 while (llist) { 1448 p = llist_entry(llist, struct task_struct, wake_entry); 1449 llist = llist_next(llist); 1450 ttwu_do_activate(rq, p, 0); 1451 } 1452 1453 raw_spin_unlock(&rq->lock); 1454} 1455 1456void scheduler_ipi(void) 1457{ 1458 if (llist_empty(&this_rq()->wake_list) && !got_nohz_idle_kick()) 1459 return; 1460 1461 /* 1462 * Not all reschedule IPI handlers call irq_enter/irq_exit, since 1463 * traditionally all their work was done from the interrupt return 1464 * path. Now that we actually do some work, we need to make sure 1465 * we do call them. 1466 * 1467 * Some archs already do call them, luckily irq_enter/exit nest 1468 * properly. 1469 * 1470 * Arguably we should visit all archs and update all handlers, 1471 * however a fair share of IPIs are still resched only so this would 1472 * somewhat pessimize the simple resched case. 1473 */ 1474 irq_enter(); 1475 sched_ttwu_pending(); 1476 1477 /* 1478 * Check if someone kicked us for doing the nohz idle load balance. 1479 */ 1480 if (unlikely(got_nohz_idle_kick() && !need_resched())) { 1481 this_rq()->idle_balance = 1; 1482 raise_softirq_irqoff(SCHED_SOFTIRQ); 1483 } 1484 irq_exit(); 1485} 1486 1487static void ttwu_queue_remote(struct task_struct *p, int cpu) 1488{ 1489 if (llist_add(&p->wake_entry, &cpu_rq(cpu)->wake_list)) 1490 smp_send_reschedule(cpu); 1491} 1492 1493#ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW 1494static int ttwu_activate_remote(struct task_struct *p, int wake_flags) 1495{ 1496 struct rq *rq; 1497 int ret = 0; 1498 1499 rq = __task_rq_lock(p); 1500 if (p->on_cpu) { 1501 ttwu_activate(rq, p, ENQUEUE_WAKEUP); 1502 ttwu_do_wakeup(rq, p, wake_flags); 1503 ret = 1; 1504 } 1505 __task_rq_unlock(rq); 1506 1507 return ret; 1508 1509} 1510#endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */ 1511 1512bool cpus_share_cache(int this_cpu, int that_cpu) 1513{ 1514 return per_cpu(sd_llc_id, this_cpu) == per_cpu(sd_llc_id, that_cpu); 1515} 1516#endif /* CONFIG_SMP */ 1517 1518static void ttwu_queue(struct task_struct *p, int cpu) 1519{ 1520 struct rq *rq = cpu_rq(cpu); 1521 1522#if defined(CONFIG_SMP) 1523 if (sched_feat(TTWU_QUEUE) && !cpus_share_cache(smp_processor_id(), cpu)) { 1524 sched_clock_cpu(cpu); /* sync clocks x-cpu */ 1525 ttwu_queue_remote(p, cpu); 1526 return; 1527 } 1528#endif 1529 1530 raw_spin_lock(&rq->lock); 1531 ttwu_do_activate(rq, p, 0); 1532 raw_spin_unlock(&rq->lock); 1533} 1534 1535/** 1536 * try_to_wake_up - wake up a thread 1537 * @p: the thread to be awakened 1538 * @state: the mask of task states that can be woken 1539 * @wake_flags: wake modifier flags (WF_*) 1540 * 1541 * Put it on the run-queue if it's not already there. The "current" 1542 * thread is always on the run-queue (except when the actual 1543 * re-schedule is in progress), and as such you're allowed to do 1544 * the simpler "current->state = TASK_RUNNING" to mark yourself 1545 * runnable without the overhead of this. 1546 * 1547 * Returns %true if @p was woken up, %false if it was already running 1548 * or @state didn't match @p's state. 1549 */ 1550static int 1551try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags) 1552{ 1553 unsigned long flags; 1554 int cpu, success = 0; 1555 1556 smp_wmb(); 1557 raw_spin_lock_irqsave(&p->pi_lock, flags); 1558 if (!(p->state & state)) 1559 goto out; 1560 1561 success = 1; /* we're going to change ->state */ 1562 cpu = task_cpu(p); 1563 1564 if (p->on_rq && ttwu_remote(p, wake_flags)) 1565 goto stat; 1566 1567#ifdef CONFIG_SMP 1568 /* 1569 * If the owning (remote) cpu is still in the middle of schedule() with 1570 * this task as prev, wait until its done referencing the task. 1571 */ 1572 while (p->on_cpu) { 1573#ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW 1574 /* 1575 * In case the architecture enables interrupts in 1576 * context_switch(), we cannot busy wait, since that 1577 * would lead to deadlocks when an interrupt hits and 1578 * tries to wake up @prev. So bail and do a complete 1579 * remote wakeup. 1580 */ 1581 if (ttwu_activate_remote(p, wake_flags)) 1582 goto stat; 1583#else 1584 cpu_relax(); 1585#endif 1586 } 1587 /* 1588 * Pairs with the smp_wmb() in finish_lock_switch(). 1589 */ 1590 smp_rmb(); 1591 1592 p->sched_contributes_to_load = !!task_contributes_to_load(p); 1593 p->state = TASK_WAKING; 1594 1595 if (p->sched_class->task_waking) 1596 p->sched_class->task_waking(p); 1597 1598 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags); 1599 if (task_cpu(p) != cpu) { 1600 wake_flags |= WF_MIGRATED; 1601 set_task_cpu(p, cpu); 1602 } 1603#endif /* CONFIG_SMP */ 1604 1605 ttwu_queue(p, cpu); 1606stat: 1607 ttwu_stat(p, cpu, wake_flags); 1608out: 1609 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 1610 1611 return success; 1612} 1613 1614/** 1615 * try_to_wake_up_local - try to wake up a local task with rq lock held 1616 * @p: the thread to be awakened 1617 * 1618 * Put @p on the run-queue if it's not already there. The caller must 1619 * ensure that this_rq() is locked, @p is bound to this_rq() and not 1620 * the current task. 1621 */ 1622static void try_to_wake_up_local(struct task_struct *p) 1623{ 1624 struct rq *rq = task_rq(p); 1625 1626 BUG_ON(rq != this_rq()); 1627 BUG_ON(p == current); 1628 lockdep_assert_held(&rq->lock); 1629 1630 if (!raw_spin_trylock(&p->pi_lock)) { 1631 raw_spin_unlock(&rq->lock); 1632 raw_spin_lock(&p->pi_lock); 1633 raw_spin_lock(&rq->lock); 1634 } 1635 1636 if (!(p->state & TASK_NORMAL)) 1637 goto out; 1638 1639 if (!p->on_rq) 1640 ttwu_activate(rq, p, ENQUEUE_WAKEUP); 1641 1642 ttwu_do_wakeup(rq, p, 0); 1643 ttwu_stat(p, smp_processor_id(), 0); 1644out: 1645 raw_spin_unlock(&p->pi_lock); 1646} 1647 1648/** 1649 * wake_up_process - Wake up a specific process 1650 * @p: The process to be woken up. 1651 * 1652 * Attempt to wake up the nominated process and move it to the set of runnable 1653 * processes. Returns 1 if the process was woken up, 0 if it was already 1654 * running. 1655 * 1656 * It may be assumed that this function implies a write memory barrier before 1657 * changing the task state if and only if any tasks are woken up. 1658 */ 1659int wake_up_process(struct task_struct *p) 1660{ 1661 return try_to_wake_up(p, TASK_ALL, 0); 1662} 1663EXPORT_SYMBOL(wake_up_process); 1664 1665int wake_up_state(struct task_struct *p, unsigned int state) 1666{ 1667 return try_to_wake_up(p, state, 0); 1668} 1669 1670/* 1671 * Perform scheduler related setup for a newly forked process p. 1672 * p is forked by current. 1673 * 1674 * __sched_fork() is basic setup used by init_idle() too: 1675 */ 1676static void __sched_fork(struct task_struct *p) 1677{ 1678 p->on_rq = 0; 1679 1680 p->se.on_rq = 0; 1681 p->se.exec_start = 0; 1682 p->se.sum_exec_runtime = 0; 1683 p->se.prev_sum_exec_runtime = 0; 1684 p->se.nr_migrations = 0; 1685 p->se.vruntime = 0; 1686 INIT_LIST_HEAD(&p->se.group_node); 1687 1688#ifdef CONFIG_SCHEDSTATS 1689 memset(&p->se.statistics, 0, sizeof(p->se.statistics)); 1690#endif 1691 1692 INIT_LIST_HEAD(&p->rt.run_list); 1693 1694#ifdef CONFIG_PREEMPT_NOTIFIERS 1695 INIT_HLIST_HEAD(&p->preempt_notifiers); 1696#endif 1697} 1698 1699/* 1700 * fork()/clone()-time setup: 1701 */ 1702void sched_fork(struct task_struct *p) 1703{ 1704 unsigned long flags; 1705 int cpu = get_cpu(); 1706 1707 __sched_fork(p); 1708 /* 1709 * We mark the process as running here. This guarantees that 1710 * nobody will actually run it, and a signal or other external 1711 * event cannot wake it up and insert it on the runqueue either. 1712 */ 1713 p->state = TASK_RUNNING; 1714 1715 /* 1716 * Make sure we do not leak PI boosting priority to the child. 1717 */ 1718 p->prio = current->normal_prio; 1719 1720 /* 1721 * Revert to default priority/policy on fork if requested. 1722 */ 1723 if (unlikely(p->sched_reset_on_fork)) { 1724 if (task_has_rt_policy(p)) { 1725 p->policy = SCHED_NORMAL; 1726 p->static_prio = NICE_TO_PRIO(0); 1727 p->rt_priority = 0; 1728 } else if (PRIO_TO_NICE(p->static_prio) < 0) 1729 p->static_prio = NICE_TO_PRIO(0); 1730 1731 p->prio = p->normal_prio = __normal_prio(p); 1732 set_load_weight(p); 1733 1734 /* 1735 * We don't need the reset flag anymore after the fork. It has 1736 * fulfilled its duty: 1737 */ 1738 p->sched_reset_on_fork = 0; 1739 } 1740 1741 if (!rt_prio(p->prio)) 1742 p->sched_class = &fair_sched_class; 1743 1744 if (p->sched_class->task_fork) 1745 p->sched_class->task_fork(p); 1746 1747 /* 1748 * The child is not yet in the pid-hash so no cgroup attach races, 1749 * and the cgroup is pinned to this child due to cgroup_fork() 1750 * is ran before sched_fork(). 1751 * 1752 * Silence PROVE_RCU. 1753 */ 1754 raw_spin_lock_irqsave(&p->pi_lock, flags); 1755 set_task_cpu(p, cpu); 1756 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 1757 1758#if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT) 1759 if (likely(sched_info_on())) 1760 memset(&p->sched_info, 0, sizeof(p->sched_info)); 1761#endif 1762#if defined(CONFIG_SMP) 1763 p->on_cpu = 0; 1764#endif 1765#ifdef CONFIG_PREEMPT_COUNT 1766 /* Want to start with kernel preemption disabled. */ 1767 task_thread_info(p)->preempt_count = 1; 1768#endif 1769#ifdef CONFIG_SMP 1770 plist_node_init(&p->pushable_tasks, MAX_PRIO); 1771#endif 1772 1773 put_cpu(); 1774} 1775 1776/* 1777 * wake_up_new_task - wake up a newly created task for the first time. 1778 * 1779 * This function will do some initial scheduler statistics housekeeping 1780 * that must be done for every newly created context, then puts the task 1781 * on the runqueue and wakes it. 1782 */ 1783void wake_up_new_task(struct task_struct *p) 1784{ 1785 unsigned long flags; 1786 struct rq *rq; 1787 1788 raw_spin_lock_irqsave(&p->pi_lock, flags); 1789#ifdef CONFIG_SMP 1790 /* 1791 * Fork balancing, do it here and not earlier because: 1792 * - cpus_allowed can change in the fork path 1793 * - any previously selected cpu might disappear through hotplug 1794 */ 1795 set_task_cpu(p, select_task_rq(p, SD_BALANCE_FORK, 0)); 1796#endif 1797 1798 rq = __task_rq_lock(p); 1799 activate_task(rq, p, 0); 1800 p->on_rq = 1; 1801 trace_sched_wakeup_new(p, true); 1802 check_preempt_curr(rq, p, WF_FORK); 1803#ifdef CONFIG_SMP 1804 if (p->sched_class->task_woken) 1805 p->sched_class->task_woken(rq, p); 1806#endif 1807 task_rq_unlock(rq, p, &flags); 1808} 1809 1810#ifdef CONFIG_PREEMPT_NOTIFIERS 1811 1812/** 1813 * preempt_notifier_register - tell me when current is being preempted & rescheduled 1814 * @notifier: notifier struct to register 1815 */ 1816void preempt_notifier_register(struct preempt_notifier *notifier) 1817{ 1818 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers); 1819} 1820EXPORT_SYMBOL_GPL(preempt_notifier_register); 1821 1822/** 1823 * preempt_notifier_unregister - no longer interested in preemption notifications 1824 * @notifier: notifier struct to unregister 1825 * 1826 * This is safe to call from within a preemption notifier. 1827 */ 1828void preempt_notifier_unregister(struct preempt_notifier *notifier) 1829{ 1830 hlist_del(¬ifier->link); 1831} 1832EXPORT_SYMBOL_GPL(preempt_notifier_unregister); 1833 1834static void fire_sched_in_preempt_notifiers(struct task_struct *curr) 1835{ 1836 struct preempt_notifier *notifier; 1837 struct hlist_node *node; 1838 1839 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link) 1840 notifier->ops->sched_in(notifier, raw_smp_processor_id()); 1841} 1842 1843static void 1844fire_sched_out_preempt_notifiers(struct task_struct *curr, 1845 struct task_struct *next) 1846{ 1847 struct preempt_notifier *notifier; 1848 struct hlist_node *node; 1849 1850 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link) 1851 notifier->ops->sched_out(notifier, next); 1852} 1853 1854#else /* !CONFIG_PREEMPT_NOTIFIERS */ 1855 1856static void fire_sched_in_preempt_notifiers(struct task_struct *curr) 1857{ 1858} 1859 1860static void 1861fire_sched_out_preempt_notifiers(struct task_struct *curr, 1862 struct task_struct *next) 1863{ 1864} 1865 1866#endif /* CONFIG_PREEMPT_NOTIFIERS */ 1867 1868/** 1869 * prepare_task_switch - prepare to switch tasks 1870 * @rq: the runqueue preparing to switch 1871 * @prev: the current task that is being switched out 1872 * @next: the task we are going to switch to. 1873 * 1874 * This is called with the rq lock held and interrupts off. It must 1875 * be paired with a subsequent finish_task_switch after the context 1876 * switch. 1877 * 1878 * prepare_task_switch sets up locking and calls architecture specific 1879 * hooks. 1880 */ 1881static inline void 1882prepare_task_switch(struct rq *rq, struct task_struct *prev, 1883 struct task_struct *next) 1884{ 1885 sched_info_switch(prev, next); 1886 perf_event_task_sched_out(prev, next); 1887 fire_sched_out_preempt_notifiers(prev, next); 1888 prepare_lock_switch(rq, next); 1889 prepare_arch_switch(next); 1890 trace_sched_switch(prev, next); 1891} 1892 1893/** 1894 * finish_task_switch - clean up after a task-switch 1895 * @rq: runqueue associated with task-switch 1896 * @prev: the thread we just switched away from. 1897 * 1898 * finish_task_switch must be called after the context switch, paired 1899 * with a prepare_task_switch call before the context switch. 1900 * finish_task_switch will reconcile locking set up by prepare_task_switch, 1901 * and do any other architecture-specific cleanup actions. 1902 * 1903 * Note that we may have delayed dropping an mm in context_switch(). If 1904 * so, we finish that here outside of the runqueue lock. (Doing it 1905 * with the lock held can cause deadlocks; see schedule() for 1906 * details.) 1907 */ 1908static void finish_task_switch(struct rq *rq, struct task_struct *prev) 1909 __releases(rq->lock) 1910{ 1911 struct mm_struct *mm = rq->prev_mm; 1912 long prev_state; 1913 1914 rq->prev_mm = NULL; 1915 1916 /* 1917 * A task struct has one reference for the use as "current". 1918 * If a task dies, then it sets TASK_DEAD in tsk->state and calls 1919 * schedule one last time. The schedule call will never return, and 1920 * the scheduled task must drop that reference. 1921 * The test for TASK_DEAD must occur while the runqueue locks are 1922 * still held, otherwise prev could be scheduled on another cpu, die 1923 * there before we look at prev->state, and then the reference would 1924 * be dropped twice. 1925 * Manfred Spraul <manfred@colorfullife.com> 1926 */ 1927 prev_state = prev->state; 1928 finish_arch_switch(prev); 1929#ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW 1930 local_irq_disable(); 1931#endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */ 1932 perf_event_task_sched_in(prev, current); 1933#ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW 1934 local_irq_enable(); 1935#endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */ 1936 finish_lock_switch(rq, prev); 1937 1938 fire_sched_in_preempt_notifiers(current); 1939 if (mm) 1940 mmdrop(mm); 1941 if (unlikely(prev_state == TASK_DEAD)) { 1942 /* 1943 * Remove function-return probe instances associated with this 1944 * task and put them back on the free list. 1945 */ 1946 kprobe_flush_task(prev); 1947 put_task_struct(prev); 1948 } 1949} 1950 1951#ifdef CONFIG_SMP 1952 1953/* assumes rq->lock is held */ 1954static inline void pre_schedule(struct rq *rq, struct task_struct *prev) 1955{ 1956 if (prev->sched_class->pre_schedule) 1957 prev->sched_class->pre_schedule(rq, prev); 1958} 1959 1960/* rq->lock is NOT held, but preemption is disabled */ 1961static inline void post_schedule(struct rq *rq) 1962{ 1963 if (rq->post_schedule) { 1964 unsigned long flags; 1965 1966 raw_spin_lock_irqsave(&rq->lock, flags); 1967 if (rq->curr->sched_class->post_schedule) 1968 rq->curr->sched_class->post_schedule(rq); 1969 raw_spin_unlock_irqrestore(&rq->lock, flags); 1970 1971 rq->post_schedule = 0; 1972 } 1973} 1974 1975#else 1976 1977static inline void pre_schedule(struct rq *rq, struct task_struct *p) 1978{ 1979} 1980 1981static inline void post_schedule(struct rq *rq) 1982{ 1983} 1984 1985#endif 1986 1987/** 1988 * schedule_tail - first thing a freshly forked thread must call. 1989 * @prev: the thread we just switched away from. 1990 */ 1991asmlinkage void schedule_tail(struct task_struct *prev) 1992 __releases(rq->lock) 1993{ 1994 struct rq *rq = this_rq(); 1995 1996 finish_task_switch(rq, prev); 1997 1998 /* 1999 * FIXME: do we need to worry about rq being invalidated by the 2000 * task_switch? 2001 */ 2002 post_schedule(rq); 2003 2004#ifdef __ARCH_WANT_UNLOCKED_CTXSW 2005 /* In this case, finish_task_switch does not reenable preemption */ 2006 preempt_enable(); 2007#endif 2008 if (current->set_child_tid) 2009 put_user(task_pid_vnr(current), current->set_child_tid); 2010} 2011 2012/* 2013 * context_switch - switch to the new MM and the new 2014 * thread's register state. 2015 */ 2016static inline void 2017context_switch(struct rq *rq, struct task_struct *prev, 2018 struct task_struct *next) 2019{ 2020 struct mm_struct *mm, *oldmm; 2021 2022 prepare_task_switch(rq, prev, next); 2023 2024 mm = next->mm; 2025 oldmm = prev->active_mm; 2026 /* 2027 * For paravirt, this is coupled with an exit in switch_to to 2028 * combine the page table reload and the switch backend into 2029 * one hypercall. 2030 */ 2031 arch_start_context_switch(prev); 2032 2033 if (!mm) { 2034 next->active_mm = oldmm; 2035 atomic_inc(&oldmm->mm_count); 2036 enter_lazy_tlb(oldmm, next); 2037 } else 2038 switch_mm(oldmm, mm, next); 2039 2040 if (!prev->mm) { 2041 prev->active_mm = NULL; 2042 rq->prev_mm = oldmm; 2043 } 2044 /* 2045 * Since the runqueue lock will be released by the next 2046 * task (which is an invalid locking op but in the case 2047 * of the scheduler it's an obvious special-case), so we 2048 * do an early lockdep release here: 2049 */ 2050#ifndef __ARCH_WANT_UNLOCKED_CTXSW 2051 spin_release(&rq->lock.dep_map, 1, _THIS_IP_); 2052#endif 2053 2054 /* Here we just switch the register state and the stack. */ 2055 switch_to(prev, next, prev); 2056 2057 barrier(); 2058 /* 2059 * this_rq must be evaluated again because prev may have moved 2060 * CPUs since it called schedule(), thus the 'rq' on its stack 2061 * frame will be invalid. 2062 */ 2063 finish_task_switch(this_rq(), prev); 2064} 2065 2066/* 2067 * nr_running, nr_uninterruptible and nr_context_switches: 2068 * 2069 * externally visible scheduler statistics: current number of runnable 2070 * threads, current number of uninterruptible-sleeping threads, total 2071 * number of context switches performed since bootup. 2072 */ 2073unsigned long nr_running(void) 2074{ 2075 unsigned long i, sum = 0; 2076 2077 for_each_online_cpu(i) 2078 sum += cpu_rq(i)->nr_running; 2079 2080 return sum; 2081} 2082 2083unsigned long nr_uninterruptible(void) 2084{ 2085 unsigned long i, sum = 0; 2086 2087 for_each_possible_cpu(i) 2088 sum += cpu_rq(i)->nr_uninterruptible; 2089 2090 /* 2091 * Since we read the counters lockless, it might be slightly 2092 * inaccurate. Do not allow it to go below zero though: 2093 */ 2094 if (unlikely((long)sum < 0)) 2095 sum = 0; 2096 2097 return sum; 2098} 2099 2100unsigned long long nr_context_switches(void) 2101{ 2102 int i; 2103 unsigned long long sum = 0; 2104 2105 for_each_possible_cpu(i) 2106 sum += cpu_rq(i)->nr_switches; 2107 2108 return sum; 2109} 2110 2111unsigned long nr_iowait(void) 2112{ 2113 unsigned long i, sum = 0; 2114 2115 for_each_possible_cpu(i) 2116 sum += atomic_read(&cpu_rq(i)->nr_iowait); 2117 2118 return sum; 2119} 2120 2121unsigned long nr_iowait_cpu(int cpu) 2122{ 2123 struct rq *this = cpu_rq(cpu); 2124 return atomic_read(&this->nr_iowait); 2125} 2126 2127unsigned long this_cpu_load(void) 2128{ 2129 struct rq *this = this_rq(); 2130 return this->cpu_load[0]; 2131} 2132 2133 2134/* Variables and functions for calc_load */ 2135static atomic_long_t calc_load_tasks; 2136static unsigned long calc_load_update; 2137unsigned long avenrun[3]; 2138EXPORT_SYMBOL(avenrun); 2139 2140static long calc_load_fold_active(struct rq *this_rq) 2141{ 2142 long nr_active, delta = 0; 2143 2144 nr_active = this_rq->nr_running; 2145 nr_active += (long) this_rq->nr_uninterruptible; 2146 2147 if (nr_active != this_rq->calc_load_active) { 2148 delta = nr_active - this_rq->calc_load_active; 2149 this_rq->calc_load_active = nr_active; 2150 } 2151 2152 return delta; 2153} 2154 2155static unsigned long 2156calc_load(unsigned long load, unsigned long exp, unsigned long active) 2157{ 2158 load *= exp; 2159 load += active * (FIXED_1 - exp); 2160 load += 1UL << (FSHIFT - 1); 2161 return load >> FSHIFT; 2162} 2163 2164#ifdef CONFIG_NO_HZ 2165/* 2166 * For NO_HZ we delay the active fold to the next LOAD_FREQ update. 2167 * 2168 * When making the ILB scale, we should try to pull this in as well. 2169 */ 2170static atomic_long_t calc_load_tasks_idle; 2171 2172void calc_load_account_idle(struct rq *this_rq) 2173{ 2174 long delta; 2175 2176 delta = calc_load_fold_active(this_rq); 2177 if (delta) 2178 atomic_long_add(delta, &calc_load_tasks_idle); 2179} 2180 2181static long calc_load_fold_idle(void) 2182{ 2183 long delta = 0; 2184 2185 /* 2186 * Its got a race, we don't care... 2187 */ 2188 if (atomic_long_read(&calc_load_tasks_idle)) 2189 delta = atomic_long_xchg(&calc_load_tasks_idle, 0); 2190 2191 return delta; 2192} 2193 2194/** 2195 * fixed_power_int - compute: x^n, in O(log n) time 2196 * 2197 * @x: base of the power 2198 * @frac_bits: fractional bits of @x 2199 * @n: power to raise @x to. 2200 * 2201 * By exploiting the relation between the definition of the natural power 2202 * function: x^n := x*x*...*x (x multiplied by itself for n times), and 2203 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i, 2204 * (where: n_i \elem {0, 1}, the binary vector representing n), 2205 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is 2206 * of course trivially computable in O(log_2 n), the length of our binary 2207 * vector. 2208 */ 2209static unsigned long 2210fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n) 2211{ 2212 unsigned long result = 1UL << frac_bits; 2213 2214 if (n) for (;;) { 2215 if (n & 1) { 2216 result *= x; 2217 result += 1UL << (frac_bits - 1); 2218 result >>= frac_bits; 2219 } 2220 n >>= 1; 2221 if (!n) 2222 break; 2223 x *= x; 2224 x += 1UL << (frac_bits - 1); 2225 x >>= frac_bits; 2226 } 2227 2228 return result; 2229} 2230 2231/* 2232 * a1 = a0 * e + a * (1 - e) 2233 * 2234 * a2 = a1 * e + a * (1 - e) 2235 * = (a0 * e + a * (1 - e)) * e + a * (1 - e) 2236 * = a0 * e^2 + a * (1 - e) * (1 + e) 2237 * 2238 * a3 = a2 * e + a * (1 - e) 2239 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e) 2240 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2) 2241 * 2242 * ... 2243 * 2244 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1] 2245 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e) 2246 * = a0 * e^n + a * (1 - e^n) 2247 * 2248 * [1] application of the geometric series: 2249 * 2250 * n 1 - x^(n+1) 2251 * S_n := \Sum x^i = ------------- 2252 * i=0 1 - x 2253 */ 2254static unsigned long 2255calc_load_n(unsigned long load, unsigned long exp, 2256 unsigned long active, unsigned int n) 2257{ 2258 2259 return calc_load(load, fixed_power_int(exp, FSHIFT, n), active); 2260} 2261 2262/* 2263 * NO_HZ can leave us missing all per-cpu ticks calling 2264 * calc_load_account_active(), but since an idle CPU folds its delta into 2265 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold 2266 * in the pending idle delta if our idle period crossed a load cycle boundary. 2267 * 2268 * Once we've updated the global active value, we need to apply the exponential 2269 * weights adjusted to the number of cycles missed. 2270 */ 2271static void calc_global_nohz(void) 2272{ 2273 long delta, active, n; 2274 2275 /* 2276 * If we crossed a calc_load_update boundary, make sure to fold 2277 * any pending idle changes, the respective CPUs might have 2278 * missed the tick driven calc_load_account_active() update 2279 * due to NO_HZ. 2280 */ 2281 delta = calc_load_fold_idle(); 2282 if (delta) 2283 atomic_long_add(delta, &calc_load_tasks); 2284 2285 /* 2286 * It could be the one fold was all it took, we done! 2287 */ 2288 if (time_before(jiffies, calc_load_update + 10)) 2289 return; 2290 2291 /* 2292 * Catch-up, fold however many we are behind still 2293 */ 2294 delta = jiffies - calc_load_update - 10; 2295 n = 1 + (delta / LOAD_FREQ); 2296 2297 active = atomic_long_read(&calc_load_tasks); 2298 active = active > 0 ? active * FIXED_1 : 0; 2299 2300 avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n); 2301 avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n); 2302 avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n); 2303 2304 calc_load_update += n * LOAD_FREQ; 2305} 2306#else 2307void calc_load_account_idle(struct rq *this_rq) 2308{ 2309} 2310 2311static inline long calc_load_fold_idle(void) 2312{ 2313 return 0; 2314} 2315 2316static void calc_global_nohz(void) 2317{ 2318} 2319#endif 2320 2321/** 2322 * get_avenrun - get the load average array 2323 * @loads: pointer to dest load array 2324 * @offset: offset to add 2325 * @shift: shift count to shift the result left 2326 * 2327 * These values are estimates at best, so no need for locking. 2328 */ 2329void get_avenrun(unsigned long *loads, unsigned long offset, int shift) 2330{ 2331 loads[0] = (avenrun[0] + offset) << shift; 2332 loads[1] = (avenrun[1] + offset) << shift; 2333 loads[2] = (avenrun[2] + offset) << shift; 2334} 2335 2336/* 2337 * calc_load - update the avenrun load estimates 10 ticks after the 2338 * CPUs have updated calc_load_tasks. 2339 */ 2340void calc_global_load(unsigned long ticks) 2341{ 2342 long active; 2343 2344 if (time_before(jiffies, calc_load_update + 10)) 2345 return; 2346 2347 active = atomic_long_read(&calc_load_tasks); 2348 active = active > 0 ? active * FIXED_1 : 0; 2349 2350 avenrun[0] = calc_load(avenrun[0], EXP_1, active); 2351 avenrun[1] = calc_load(avenrun[1], EXP_5, active); 2352 avenrun[2] = calc_load(avenrun[2], EXP_15, active); 2353 2354 calc_load_update += LOAD_FREQ; 2355 2356 /* 2357 * Account one period with whatever state we found before 2358 * folding in the nohz state and ageing the entire idle period. 2359 * 2360 * This avoids loosing a sample when we go idle between 2361 * calc_load_account_active() (10 ticks ago) and now and thus 2362 * under-accounting. 2363 */ 2364 calc_global_nohz(); 2365} 2366 2367/* 2368 * Called from update_cpu_load() to periodically update this CPU's 2369 * active count. 2370 */ 2371static void calc_load_account_active(struct rq *this_rq) 2372{ 2373 long delta; 2374 2375 if (time_before(jiffies, this_rq->calc_load_update)) 2376 return; 2377 2378 delta = calc_load_fold_active(this_rq); 2379 delta += calc_load_fold_idle(); 2380 if (delta) 2381 atomic_long_add(delta, &calc_load_tasks); 2382 2383 this_rq->calc_load_update += LOAD_FREQ; 2384} 2385 2386/* 2387 * The exact cpuload at various idx values, calculated at every tick would be 2388 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load 2389 * 2390 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called 2391 * on nth tick when cpu may be busy, then we have: 2392 * load = ((2^idx - 1) / 2^idx)^(n-1) * load 2393 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load 2394 * 2395 * decay_load_missed() below does efficient calculation of 2396 * load = ((2^idx - 1) / 2^idx)^(n-1) * load 2397 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load 2398 * 2399 * The calculation is approximated on a 128 point scale. 2400 * degrade_zero_ticks is the number of ticks after which load at any 2401 * particular idx is approximated to be zero. 2402 * degrade_factor is a precomputed table, a row for each load idx. 2403 * Each column corresponds to degradation factor for a power of two ticks, 2404 * based on 128 point scale. 2405 * Example: 2406 * row 2, col 3 (=12) says that the degradation at load idx 2 after 2407 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8). 2408 * 2409 * With this power of 2 load factors, we can degrade the load n times 2410 * by looking at 1 bits in n and doing as many mult/shift instead of 2411 * n mult/shifts needed by the exact degradation. 2412 */ 2413#define DEGRADE_SHIFT 7 2414static const unsigned char 2415 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128}; 2416static const unsigned char 2417 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = { 2418 {0, 0, 0, 0, 0, 0, 0, 0}, 2419 {64, 32, 8, 0, 0, 0, 0, 0}, 2420 {96, 72, 40, 12, 1, 0, 0}, 2421 {112, 98, 75, 43, 15, 1, 0}, 2422 {120, 112, 98, 76, 45, 16, 2} }; 2423 2424/* 2425 * Update cpu_load for any missed ticks, due to tickless idle. The backlog 2426 * would be when CPU is idle and so we just decay the old load without 2427 * adding any new load. 2428 */ 2429static unsigned long 2430decay_load_missed(unsigned long load, unsigned long missed_updates, int idx) 2431{ 2432 int j = 0; 2433 2434 if (!missed_updates) 2435 return load; 2436 2437 if (missed_updates >= degrade_zero_ticks[idx]) 2438 return 0; 2439 2440 if (idx == 1) 2441 return load >> missed_updates; 2442 2443 while (missed_updates) { 2444 if (missed_updates % 2) 2445 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT; 2446 2447 missed_updates >>= 1; 2448 j++; 2449 } 2450 return load; 2451} 2452 2453/* 2454 * Update rq->cpu_load[] statistics. This function is usually called every 2455 * scheduler tick (TICK_NSEC). With tickless idle this will not be called 2456 * every tick. We fix it up based on jiffies. 2457 */ 2458void update_cpu_load(struct rq *this_rq) 2459{ 2460 unsigned long this_load = this_rq->load.weight; 2461 unsigned long curr_jiffies = jiffies; 2462 unsigned long pending_updates; 2463 int i, scale; 2464 2465 this_rq->nr_load_updates++; 2466 2467 /* Avoid repeated calls on same jiffy, when moving in and out of idle */ 2468 if (curr_jiffies == this_rq->last_load_update_tick) 2469 return; 2470 2471 pending_updates = curr_jiffies - this_rq->last_load_update_tick; 2472 this_rq->last_load_update_tick = curr_jiffies; 2473 2474 /* Update our load: */ 2475 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */ 2476 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) { 2477 unsigned long old_load, new_load; 2478 2479 /* scale is effectively 1 << i now, and >> i divides by scale */ 2480 2481 old_load = this_rq->cpu_load[i]; 2482 old_load = decay_load_missed(old_load, pending_updates - 1, i); 2483 new_load = this_load; 2484 /* 2485 * Round up the averaging division if load is increasing. This 2486 * prevents us from getting stuck on 9 if the load is 10, for 2487 * example. 2488 */ 2489 if (new_load > old_load) 2490 new_load += scale - 1; 2491 2492 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i; 2493 } 2494 2495 sched_avg_update(this_rq); 2496} 2497 2498static void update_cpu_load_active(struct rq *this_rq) 2499{ 2500 update_cpu_load(this_rq); 2501 2502 calc_load_account_active(this_rq); 2503} 2504 2505#ifdef CONFIG_SMP 2506 2507/* 2508 * sched_exec - execve() is a valuable balancing opportunity, because at 2509 * this point the task has the smallest effective memory and cache footprint. 2510 */ 2511void sched_exec(void) 2512{ 2513 struct task_struct *p = current; 2514 unsigned long flags; 2515 int dest_cpu; 2516 2517 raw_spin_lock_irqsave(&p->pi_lock, flags); 2518 dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0); 2519 if (dest_cpu == smp_processor_id()) 2520 goto unlock; 2521 2522 if (likely(cpu_active(dest_cpu))) { 2523 struct migration_arg arg = { p, dest_cpu }; 2524 2525 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 2526 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg); 2527 return; 2528 } 2529unlock: 2530 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 2531} 2532 2533#endif 2534 2535DEFINE_PER_CPU(struct kernel_stat, kstat); 2536DEFINE_PER_CPU(struct kernel_cpustat, kernel_cpustat); 2537 2538EXPORT_PER_CPU_SYMBOL(kstat); 2539EXPORT_PER_CPU_SYMBOL(kernel_cpustat); 2540 2541/* 2542 * Return any ns on the sched_clock that have not yet been accounted in 2543 * @p in case that task is currently running. 2544 * 2545 * Called with task_rq_lock() held on @rq. 2546 */ 2547static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq) 2548{ 2549 u64 ns = 0; 2550 2551 if (task_current(rq, p)) { 2552 update_rq_clock(rq); 2553 ns = rq->clock_task - p->se.exec_start; 2554 if ((s64)ns < 0) 2555 ns = 0; 2556 } 2557 2558 return ns; 2559} 2560 2561unsigned long long task_delta_exec(struct task_struct *p) 2562{ 2563 unsigned long flags; 2564 struct rq *rq; 2565 u64 ns = 0; 2566 2567 rq = task_rq_lock(p, &flags); 2568 ns = do_task_delta_exec(p, rq); 2569 task_rq_unlock(rq, p, &flags); 2570 2571 return ns; 2572} 2573 2574/* 2575 * Return accounted runtime for the task. 2576 * In case the task is currently running, return the runtime plus current's 2577 * pending runtime that have not been accounted yet. 2578 */ 2579unsigned long long task_sched_runtime(struct task_struct *p) 2580{ 2581 unsigned long flags; 2582 struct rq *rq; 2583 u64 ns = 0; 2584 2585 rq = task_rq_lock(p, &flags); 2586 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq); 2587 task_rq_unlock(rq, p, &flags); 2588 2589 return ns; 2590} 2591 2592#ifdef CONFIG_CGROUP_CPUACCT 2593struct cgroup_subsys cpuacct_subsys; 2594struct cpuacct root_cpuacct; 2595#endif 2596 2597static inline void task_group_account_field(struct task_struct *p, int index, 2598 u64 tmp) 2599{ 2600#ifdef CONFIG_CGROUP_CPUACCT 2601 struct kernel_cpustat *kcpustat; 2602 struct cpuacct *ca; 2603#endif 2604 /* 2605 * Since all updates are sure to touch the root cgroup, we 2606 * get ourselves ahead and touch it first. If the root cgroup 2607 * is the only cgroup, then nothing else should be necessary. 2608 * 2609 */ 2610 __get_cpu_var(kernel_cpustat).cpustat[index] += tmp; 2611 2612#ifdef CONFIG_CGROUP_CPUACCT 2613 if (unlikely(!cpuacct_subsys.active)) 2614 return; 2615 2616 rcu_read_lock(); 2617 ca = task_ca(p); 2618 while (ca && (ca != &root_cpuacct)) { 2619 kcpustat = this_cpu_ptr(ca->cpustat); 2620 kcpustat->cpustat[index] += tmp; 2621 ca = parent_ca(ca); 2622 } 2623 rcu_read_unlock(); 2624#endif 2625} 2626 2627 2628/* 2629 * Account user cpu time to a process. 2630 * @p: the process that the cpu time gets accounted to 2631 * @cputime: the cpu time spent in user space since the last update 2632 * @cputime_scaled: cputime scaled by cpu frequency 2633 */ 2634void account_user_time(struct task_struct *p, cputime_t cputime, 2635 cputime_t cputime_scaled) 2636{ 2637 int index; 2638 2639 /* Add user time to process. */ 2640 p->utime += cputime; 2641 p->utimescaled += cputime_scaled; 2642 account_group_user_time(p, cputime); 2643 2644 index = (TASK_NICE(p) > 0) ? CPUTIME_NICE : CPUTIME_USER; 2645 2646 /* Add user time to cpustat. */ 2647 task_group_account_field(p, index, (__force u64) cputime); 2648 2649 /* Account for user time used */ 2650 acct_update_integrals(p); 2651} 2652 2653/* 2654 * Account guest cpu time to a process. 2655 * @p: the process that the cpu time gets accounted to 2656 * @cputime: the cpu time spent in virtual machine since the last update 2657 * @cputime_scaled: cputime scaled by cpu frequency 2658 */ 2659static void account_guest_time(struct task_struct *p, cputime_t cputime, 2660 cputime_t cputime_scaled) 2661{ 2662 u64 *cpustat = kcpustat_this_cpu->cpustat; 2663 2664 /* Add guest time to process. */ 2665 p->utime += cputime; 2666 p->utimescaled += cputime_scaled; 2667 account_group_user_time(p, cputime); 2668 p->gtime += cputime; 2669 2670 /* Add guest time to cpustat. */ 2671 if (TASK_NICE(p) > 0) { 2672 cpustat[CPUTIME_NICE] += (__force u64) cputime; 2673 cpustat[CPUTIME_GUEST_NICE] += (__force u64) cputime; 2674 } else { 2675 cpustat[CPUTIME_USER] += (__force u64) cputime; 2676 cpustat[CPUTIME_GUEST] += (__force u64) cputime; 2677 } 2678} 2679 2680/* 2681 * Account system cpu time to a process and desired cpustat field 2682 * @p: the process that the cpu time gets accounted to 2683 * @cputime: the cpu time spent in kernel space since the last update 2684 * @cputime_scaled: cputime scaled by cpu frequency 2685 * @target_cputime64: pointer to cpustat field that has to be updated 2686 */ 2687static inline 2688void __account_system_time(struct task_struct *p, cputime_t cputime, 2689 cputime_t cputime_scaled, int index) 2690{ 2691 /* Add system time to process. */ 2692 p->stime += cputime; 2693 p->stimescaled += cputime_scaled; 2694 account_group_system_time(p, cputime); 2695 2696 /* Add system time to cpustat. */ 2697 task_group_account_field(p, index, (__force u64) cputime); 2698 2699 /* Account for system time used */ 2700 acct_update_integrals(p); 2701} 2702 2703/* 2704 * Account system cpu time to a process. 2705 * @p: the process that the cpu time gets accounted to 2706 * @hardirq_offset: the offset to subtract from hardirq_count() 2707 * @cputime: the cpu time spent in kernel space since the last update 2708 * @cputime_scaled: cputime scaled by cpu frequency 2709 */ 2710void account_system_time(struct task_struct *p, int hardirq_offset, 2711 cputime_t cputime, cputime_t cputime_scaled) 2712{ 2713 int index; 2714 2715 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) { 2716 account_guest_time(p, cputime, cputime_scaled); 2717 return; 2718 } 2719 2720 if (hardirq_count() - hardirq_offset) 2721 index = CPUTIME_IRQ; 2722 else if (in_serving_softirq()) 2723 index = CPUTIME_SOFTIRQ; 2724 else 2725 index = CPUTIME_SYSTEM; 2726 2727 __account_system_time(p, cputime, cputime_scaled, index); 2728} 2729 2730/* 2731 * Account for involuntary wait time. 2732 * @cputime: the cpu time spent in involuntary wait 2733 */ 2734void account_steal_time(cputime_t cputime) 2735{ 2736 u64 *cpustat = kcpustat_this_cpu->cpustat; 2737 2738 cpustat[CPUTIME_STEAL] += (__force u64) cputime; 2739} 2740 2741/* 2742 * Account for idle time. 2743 * @cputime: the cpu time spent in idle wait 2744 */ 2745void account_idle_time(cputime_t cputime) 2746{ 2747 u64 *cpustat = kcpustat_this_cpu->cpustat; 2748 struct rq *rq = this_rq(); 2749 2750 if (atomic_read(&rq->nr_iowait) > 0) 2751 cpustat[CPUTIME_IOWAIT] += (__force u64) cputime; 2752 else 2753 cpustat[CPUTIME_IDLE] += (__force u64) cputime; 2754} 2755 2756static __always_inline bool steal_account_process_tick(void) 2757{ 2758#ifdef CONFIG_PARAVIRT 2759 if (static_key_false(¶virt_steal_enabled)) { 2760 u64 steal, st = 0; 2761 2762 steal = paravirt_steal_clock(smp_processor_id()); 2763 steal -= this_rq()->prev_steal_time; 2764 2765 st = steal_ticks(steal); 2766 this_rq()->prev_steal_time += st * TICK_NSEC; 2767 2768 account_steal_time(st); 2769 return st; 2770 } 2771#endif 2772 return false; 2773} 2774 2775#ifndef CONFIG_VIRT_CPU_ACCOUNTING 2776 2777#ifdef CONFIG_IRQ_TIME_ACCOUNTING 2778/* 2779 * Account a tick to a process and cpustat 2780 * @p: the process that the cpu time gets accounted to 2781 * @user_tick: is the tick from userspace 2782 * @rq: the pointer to rq 2783 * 2784 * Tick demultiplexing follows the order 2785 * - pending hardirq update 2786 * - pending softirq update 2787 * - user_time 2788 * - idle_time 2789 * - system time 2790 * - check for guest_time 2791 * - else account as system_time 2792 * 2793 * Check for hardirq is done both for system and user time as there is 2794 * no timer going off while we are on hardirq and hence we may never get an 2795 * opportunity to update it solely in system time. 2796 * p->stime and friends are only updated on system time and not on irq 2797 * softirq as those do not count in task exec_runtime any more. 2798 */ 2799static void irqtime_account_process_tick(struct task_struct *p, int user_tick, 2800 struct rq *rq) 2801{ 2802 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy); 2803 u64 *cpustat = kcpustat_this_cpu->cpustat; 2804 2805 if (steal_account_process_tick()) 2806 return; 2807 2808 if (irqtime_account_hi_update()) { 2809 cpustat[CPUTIME_IRQ] += (__force u64) cputime_one_jiffy; 2810 } else if (irqtime_account_si_update()) { 2811 cpustat[CPUTIME_SOFTIRQ] += (__force u64) cputime_one_jiffy; 2812 } else if (this_cpu_ksoftirqd() == p) { 2813 /* 2814 * ksoftirqd time do not get accounted in cpu_softirq_time. 2815 * So, we have to handle it separately here. 2816 * Also, p->stime needs to be updated for ksoftirqd. 2817 */ 2818 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled, 2819 CPUTIME_SOFTIRQ); 2820 } else if (user_tick) { 2821 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled); 2822 } else if (p == rq->idle) { 2823 account_idle_time(cputime_one_jiffy); 2824 } else if (p->flags & PF_VCPU) { /* System time or guest time */ 2825 account_guest_time(p, cputime_one_jiffy, one_jiffy_scaled); 2826 } else { 2827 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled, 2828 CPUTIME_SYSTEM); 2829 } 2830} 2831 2832static void irqtime_account_idle_ticks(int ticks) 2833{ 2834 int i; 2835 struct rq *rq = this_rq(); 2836 2837 for (i = 0; i < ticks; i++) 2838 irqtime_account_process_tick(current, 0, rq); 2839} 2840#else /* CONFIG_IRQ_TIME_ACCOUNTING */ 2841static void irqtime_account_idle_ticks(int ticks) {} 2842static void irqtime_account_process_tick(struct task_struct *p, int user_tick, 2843 struct rq *rq) {} 2844#endif /* CONFIG_IRQ_TIME_ACCOUNTING */ 2845 2846/* 2847 * Account a single tick of cpu time. 2848 * @p: the process that the cpu time gets accounted to 2849 * @user_tick: indicates if the tick is a user or a system tick 2850 */ 2851void account_process_tick(struct task_struct *p, int user_tick) 2852{ 2853 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy); 2854 struct rq *rq = this_rq(); 2855 2856 if (sched_clock_irqtime) { 2857 irqtime_account_process_tick(p, user_tick, rq); 2858 return; 2859 } 2860 2861 if (steal_account_process_tick()) 2862 return; 2863 2864 if (user_tick) 2865 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled); 2866 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET)) 2867 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy, 2868 one_jiffy_scaled); 2869 else 2870 account_idle_time(cputime_one_jiffy); 2871} 2872 2873/* 2874 * Account multiple ticks of steal time. 2875 * @p: the process from which the cpu time has been stolen 2876 * @ticks: number of stolen ticks 2877 */ 2878void account_steal_ticks(unsigned long ticks) 2879{ 2880 account_steal_time(jiffies_to_cputime(ticks)); 2881} 2882 2883/* 2884 * Account multiple ticks of idle time. 2885 * @ticks: number of stolen ticks 2886 */ 2887void account_idle_ticks(unsigned long ticks) 2888{ 2889 2890 if (sched_clock_irqtime) { 2891 irqtime_account_idle_ticks(ticks); 2892 return; 2893 } 2894 2895 account_idle_time(jiffies_to_cputime(ticks)); 2896} 2897 2898#endif 2899 2900/* 2901 * Use precise platform statistics if available: 2902 */ 2903#ifdef CONFIG_VIRT_CPU_ACCOUNTING 2904void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st) 2905{ 2906 *ut = p->utime; 2907 *st = p->stime; 2908} 2909 2910void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st) 2911{ 2912 struct task_cputime cputime; 2913 2914 thread_group_cputime(p, &cputime); 2915 2916 *ut = cputime.utime; 2917 *st = cputime.stime; 2918} 2919#else 2920 2921#ifndef nsecs_to_cputime 2922# define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs) 2923#endif 2924 2925void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st) 2926{ 2927 cputime_t rtime, utime = p->utime, total = utime + p->stime; 2928 2929 /* 2930 * Use CFS's precise accounting: 2931 */ 2932 rtime = nsecs_to_cputime(p->se.sum_exec_runtime); 2933 2934 if (total) { 2935 u64 temp = (__force u64) rtime; 2936 2937 temp *= (__force u64) utime; 2938 do_div(temp, (__force u32) total); 2939 utime = (__force cputime_t) temp; 2940 } else 2941 utime = rtime; 2942 2943 /* 2944 * Compare with previous values, to keep monotonicity: 2945 */ 2946 p->prev_utime = max(p->prev_utime, utime); 2947 p->prev_stime = max(p->prev_stime, rtime - p->prev_utime); 2948 2949 *ut = p->prev_utime; 2950 *st = p->prev_stime; 2951} 2952 2953/* 2954 * Must be called with siglock held. 2955 */ 2956void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st) 2957{ 2958 struct signal_struct *sig = p->signal; 2959 struct task_cputime cputime; 2960 cputime_t rtime, utime, total; 2961 2962 thread_group_cputime(p, &cputime); 2963 2964 total = cputime.utime + cputime.stime; 2965 rtime = nsecs_to_cputime(cputime.sum_exec_runtime); 2966 2967 if (total) { 2968 u64 temp = (__force u64) rtime; 2969 2970 temp *= (__force u64) cputime.utime; 2971 do_div(temp, (__force u32) total); 2972 utime = (__force cputime_t) temp; 2973 } else 2974 utime = rtime; 2975 2976 sig->prev_utime = max(sig->prev_utime, utime); 2977 sig->prev_stime = max(sig->prev_stime, rtime - sig->prev_utime); 2978 2979 *ut = sig->prev_utime; 2980 *st = sig->prev_stime; 2981} 2982#endif 2983 2984/* 2985 * This function gets called by the timer code, with HZ frequency. 2986 * We call it with interrupts disabled. 2987 */ 2988void scheduler_tick(void) 2989{ 2990 int cpu = smp_processor_id(); 2991 struct rq *rq = cpu_rq(cpu); 2992 struct task_struct *curr = rq->curr; 2993 2994 sched_clock_tick(); 2995 2996 raw_spin_lock(&rq->lock); 2997 update_rq_clock(rq); 2998 update_cpu_load_active(rq); 2999 curr->sched_class->task_tick(rq, curr, 0); 3000 raw_spin_unlock(&rq->lock); 3001 3002 perf_event_task_tick(); 3003 3004#ifdef CONFIG_SMP 3005 rq->idle_balance = idle_cpu(cpu); 3006 trigger_load_balance(rq, cpu); 3007#endif 3008} 3009 3010notrace unsigned long get_parent_ip(unsigned long addr) 3011{ 3012 if (in_lock_functions(addr)) { 3013 addr = CALLER_ADDR2; 3014 if (in_lock_functions(addr)) 3015 addr = CALLER_ADDR3; 3016 } 3017 return addr; 3018} 3019 3020#if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \ 3021 defined(CONFIG_PREEMPT_TRACER)) 3022 3023void __kprobes add_preempt_count(int val) 3024{ 3025#ifdef CONFIG_DEBUG_PREEMPT 3026 /* 3027 * Underflow? 3028 */ 3029 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0))) 3030 return; 3031#endif 3032 preempt_count() += val; 3033#ifdef CONFIG_DEBUG_PREEMPT 3034 /* 3035 * Spinlock count overflowing soon? 3036 */ 3037 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >= 3038 PREEMPT_MASK - 10); 3039#endif 3040 if (preempt_count() == val) 3041 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1)); 3042} 3043EXPORT_SYMBOL(add_preempt_count); 3044 3045void __kprobes sub_preempt_count(int val) 3046{ 3047#ifdef CONFIG_DEBUG_PREEMPT 3048 /* 3049 * Underflow? 3050 */ 3051 if (DEBUG_LOCKS_WARN_ON(val > preempt_count())) 3052 return; 3053 /* 3054 * Is the spinlock portion underflowing? 3055 */ 3056 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) && 3057 !(preempt_count() & PREEMPT_MASK))) 3058 return; 3059#endif 3060 3061 if (preempt_count() == val) 3062 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1)); 3063 preempt_count() -= val; 3064} 3065EXPORT_SYMBOL(sub_preempt_count); 3066 3067#endif 3068 3069/* 3070 * Print scheduling while atomic bug: 3071 */ 3072static noinline void __schedule_bug(struct task_struct *prev) 3073{ 3074 struct pt_regs *regs = get_irq_regs(); 3075 3076 if (oops_in_progress) 3077 return; 3078 3079 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n", 3080 prev->comm, prev->pid, preempt_count()); 3081 3082 debug_show_held_locks(prev); 3083 print_modules(); 3084 if (irqs_disabled()) 3085 print_irqtrace_events(prev); 3086 3087 if (regs) 3088 show_regs(regs); 3089 else 3090 dump_stack(); 3091} 3092 3093/* 3094 * Various schedule()-time debugging checks and statistics: 3095 */ 3096static inline void schedule_debug(struct task_struct *prev) 3097{ 3098 /* 3099 * Test if we are atomic. Since do_exit() needs to call into 3100 * schedule() atomically, we ignore that path for now. 3101 * Otherwise, whine if we are scheduling when we should not be. 3102 */ 3103 if (unlikely(in_atomic_preempt_off() && !prev->exit_state)) 3104 __schedule_bug(prev); 3105 rcu_sleep_check(); 3106 3107 profile_hit(SCHED_PROFILING, __builtin_return_address(0)); 3108 3109 schedstat_inc(this_rq(), sched_count); 3110} 3111 3112static void put_prev_task(struct rq *rq, struct task_struct *prev) 3113{ 3114 if (prev->on_rq || rq->skip_clock_update < 0) 3115 update_rq_clock(rq); 3116 prev->sched_class->put_prev_task(rq, prev); 3117} 3118 3119/* 3120 * Pick up the highest-prio task: 3121 */ 3122static inline struct task_struct * 3123pick_next_task(struct rq *rq) 3124{ 3125 const struct sched_class *class; 3126 struct task_struct *p; 3127 3128 /* 3129 * Optimization: we know that if all tasks are in 3130 * the fair class we can call that function directly: 3131 */ 3132 if (likely(rq->nr_running == rq->cfs.h_nr_running)) { 3133 p = fair_sched_class.pick_next_task(rq); 3134 if (likely(p)) 3135 return p; 3136 } 3137 3138 for_each_class(class) { 3139 p = class->pick_next_task(rq); 3140 if (p) 3141 return p; 3142 } 3143 3144 BUG(); /* the idle class will always have a runnable task */ 3145} 3146 3147/* 3148 * __schedule() is the main scheduler function. 3149 */ 3150static void __sched __schedule(void) 3151{ 3152 struct task_struct *prev, *next; 3153 unsigned long *switch_count; 3154 struct rq *rq; 3155 int cpu; 3156 3157need_resched: 3158 preempt_disable(); 3159 cpu = smp_processor_id(); 3160 rq = cpu_rq(cpu); 3161 rcu_note_context_switch(cpu); 3162 prev = rq->curr; 3163 3164 schedule_debug(prev); 3165 3166 if (sched_feat(HRTICK)) 3167 hrtick_clear(rq); 3168 3169 raw_spin_lock_irq(&rq->lock); 3170 3171 switch_count = &prev->nivcsw; 3172 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) { 3173 if (unlikely(signal_pending_state(prev->state, prev))) { 3174 prev->state = TASK_RUNNING; 3175 } else { 3176 deactivate_task(rq, prev, DEQUEUE_SLEEP); 3177 prev->on_rq = 0; 3178 3179 /* 3180 * If a worker went to sleep, notify and ask workqueue 3181 * whether it wants to wake up a task to maintain 3182 * concurrency. 3183 */ 3184 if (prev->flags & PF_WQ_WORKER) { 3185 struct task_struct *to_wakeup; 3186 3187 to_wakeup = wq_worker_sleeping(prev, cpu); 3188 if (to_wakeup) 3189 try_to_wake_up_local(to_wakeup); 3190 } 3191 } 3192 switch_count = &prev->nvcsw; 3193 } 3194 3195 pre_schedule(rq, prev); 3196 3197 if (unlikely(!rq->nr_running)) 3198 idle_balance(cpu, rq); 3199 3200 put_prev_task(rq, prev); 3201 next = pick_next_task(rq); 3202 clear_tsk_need_resched(prev); 3203 rq->skip_clock_update = 0; 3204 3205 if (likely(prev != next)) { 3206 rq->nr_switches++; 3207 rq->curr = next; 3208 ++*switch_count; 3209 3210 context_switch(rq, prev, next); /* unlocks the rq */ 3211 /* 3212 * The context switch have flipped the stack from under us 3213 * and restored the local variables which were saved when 3214 * this task called schedule() in the past. prev == current 3215 * is still correct, but it can be moved to another cpu/rq. 3216 */ 3217 cpu = smp_processor_id(); 3218 rq = cpu_rq(cpu); 3219 } else 3220 raw_spin_unlock_irq(&rq->lock); 3221 3222 post_schedule(rq); 3223 3224 sched_preempt_enable_no_resched(); 3225 if (need_resched()) 3226 goto need_resched; 3227} 3228 3229static inline void sched_submit_work(struct task_struct *tsk) 3230{ 3231 if (!tsk->state || tsk_is_pi_blocked(tsk)) 3232 return; 3233 /* 3234 * If we are going to sleep and we have plugged IO queued, 3235 * make sure to submit it to avoid deadlocks. 3236 */ 3237 if (blk_needs_flush_plug(tsk)) 3238 blk_schedule_flush_plug(tsk); 3239} 3240 3241asmlinkage void __sched schedule(void) 3242{ 3243 struct task_struct *tsk = current; 3244 3245 sched_submit_work(tsk); 3246 __schedule(); 3247} 3248EXPORT_SYMBOL(schedule); 3249 3250/** 3251 * schedule_preempt_disabled - called with preemption disabled 3252 * 3253 * Returns with preemption disabled. Note: preempt_count must be 1 3254 */ 3255void __sched schedule_preempt_disabled(void) 3256{ 3257 sched_preempt_enable_no_resched(); 3258 schedule(); 3259 preempt_disable(); 3260} 3261 3262#ifdef CONFIG_MUTEX_SPIN_ON_OWNER 3263 3264static inline bool owner_running(struct mutex *lock, struct task_struct *owner) 3265{ 3266 if (lock->owner != owner) 3267 return false; 3268 3269 /* 3270 * Ensure we emit the owner->on_cpu, dereference _after_ checking 3271 * lock->owner still matches owner, if that fails, owner might 3272 * point to free()d memory, if it still matches, the rcu_read_lock() 3273 * ensures the memory stays valid. 3274 */ 3275 barrier(); 3276 3277 return owner->on_cpu; 3278} 3279 3280/* 3281 * Look out! "owner" is an entirely speculative pointer 3282 * access and not reliable. 3283 */ 3284int mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner) 3285{ 3286 if (!sched_feat(OWNER_SPIN)) 3287 return 0; 3288 3289 rcu_read_lock(); 3290 while (owner_running(lock, owner)) { 3291 if (need_resched()) 3292 break; 3293 3294 arch_mutex_cpu_relax(); 3295 } 3296 rcu_read_unlock(); 3297 3298 /* 3299 * We break out the loop above on need_resched() and when the 3300 * owner changed, which is a sign for heavy contention. Return 3301 * success only when lock->owner is NULL. 3302 */ 3303 return lock->owner == NULL; 3304} 3305#endif 3306 3307#ifdef CONFIG_PREEMPT 3308/* 3309 * this is the entry point to schedule() from in-kernel preemption 3310 * off of preempt_enable. Kernel preemptions off return from interrupt 3311 * occur there and call schedule directly. 3312 */ 3313asmlinkage void __sched notrace preempt_schedule(void) 3314{ 3315 struct thread_info *ti = current_thread_info(); 3316 3317 /* 3318 * If there is a non-zero preempt_count or interrupts are disabled, 3319 * we do not want to preempt the current task. Just return.. 3320 */ 3321 if (likely(ti->preempt_count || irqs_disabled())) 3322 return; 3323 3324 do { 3325 add_preempt_count_notrace(PREEMPT_ACTIVE); 3326 __schedule(); 3327 sub_preempt_count_notrace(PREEMPT_ACTIVE); 3328 3329 /* 3330 * Check again in case we missed a preemption opportunity 3331 * between schedule and now. 3332 */ 3333 barrier(); 3334 } while (need_resched()); 3335} 3336EXPORT_SYMBOL(preempt_schedule); 3337 3338/* 3339 * this is the entry point to schedule() from kernel preemption 3340 * off of irq context. 3341 * Note, that this is called and return with irqs disabled. This will 3342 * protect us against recursive calling from irq. 3343 */ 3344asmlinkage void __sched preempt_schedule_irq(void) 3345{ 3346 struct thread_info *ti = current_thread_info(); 3347 3348 /* Catch callers which need to be fixed */ 3349 BUG_ON(ti->preempt_count || !irqs_disabled()); 3350 3351 do { 3352 add_preempt_count(PREEMPT_ACTIVE); 3353 local_irq_enable(); 3354 __schedule(); 3355 local_irq_disable(); 3356 sub_preempt_count(PREEMPT_ACTIVE); 3357 3358 /* 3359 * Check again in case we missed a preemption opportunity 3360 * between schedule and now. 3361 */ 3362 barrier(); 3363 } while (need_resched()); 3364} 3365 3366#endif /* CONFIG_PREEMPT */ 3367 3368int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags, 3369 void *key) 3370{ 3371 return try_to_wake_up(curr->private, mode, wake_flags); 3372} 3373EXPORT_SYMBOL(default_wake_function); 3374 3375/* 3376 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just 3377 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve 3378 * number) then we wake all the non-exclusive tasks and one exclusive task. 3379 * 3380 * There are circumstances in which we can try to wake a task which has already 3381 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns 3382 * zero in this (rare) case, and we handle it by continuing to scan the queue. 3383 */ 3384static void __wake_up_common(wait_queue_head_t *q, unsigned int mode, 3385 int nr_exclusive, int wake_flags, void *key) 3386{ 3387 wait_queue_t *curr, *next; 3388 3389 list_for_each_entry_safe(curr, next, &q->task_list, task_list) { 3390 unsigned flags = curr->flags; 3391 3392 if (curr->func(curr, mode, wake_flags, key) && 3393 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive) 3394 break; 3395 } 3396} 3397 3398/** 3399 * __wake_up - wake up threads blocked on a waitqueue. 3400 * @q: the waitqueue 3401 * @mode: which threads 3402 * @nr_exclusive: how many wake-one or wake-many threads to wake up 3403 * @key: is directly passed to the wakeup function 3404 * 3405 * It may be assumed that this function implies a write memory barrier before 3406 * changing the task state if and only if any tasks are woken up. 3407 */ 3408void __wake_up(wait_queue_head_t *q, unsigned int mode, 3409 int nr_exclusive, void *key) 3410{ 3411 unsigned long flags; 3412 3413 spin_lock_irqsave(&q->lock, flags); 3414 __wake_up_common(q, mode, nr_exclusive, 0, key); 3415 spin_unlock_irqrestore(&q->lock, flags); 3416} 3417EXPORT_SYMBOL(__wake_up); 3418 3419/* 3420 * Same as __wake_up but called with the spinlock in wait_queue_head_t held. 3421 */ 3422void __wake_up_locked(wait_queue_head_t *q, unsigned int mode, int nr) 3423{ 3424 __wake_up_common(q, mode, nr, 0, NULL); 3425} 3426EXPORT_SYMBOL_GPL(__wake_up_locked); 3427 3428void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key) 3429{ 3430 __wake_up_common(q, mode, 1, 0, key); 3431} 3432EXPORT_SYMBOL_GPL(__wake_up_locked_key); 3433 3434/** 3435 * __wake_up_sync_key - wake up threads blocked on a waitqueue. 3436 * @q: the waitqueue 3437 * @mode: which threads 3438 * @nr_exclusive: how many wake-one or wake-many threads to wake up 3439 * @key: opaque value to be passed to wakeup targets 3440 * 3441 * The sync wakeup differs that the waker knows that it will schedule 3442 * away soon, so while the target thread will be woken up, it will not 3443 * be migrated to another CPU - ie. the two threads are 'synchronized' 3444 * with each other. This can prevent needless bouncing between CPUs. 3445 * 3446 * On UP it can prevent extra preemption. 3447 * 3448 * It may be assumed that this function implies a write memory barrier before 3449 * changing the task state if and only if any tasks are woken up. 3450 */ 3451void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode, 3452 int nr_exclusive, void *key) 3453{ 3454 unsigned long flags; 3455 int wake_flags = WF_SYNC; 3456 3457 if (unlikely(!q)) 3458 return; 3459 3460 if (unlikely(!nr_exclusive)) 3461 wake_flags = 0; 3462 3463 spin_lock_irqsave(&q->lock, flags); 3464 __wake_up_common(q, mode, nr_exclusive, wake_flags, key); 3465 spin_unlock_irqrestore(&q->lock, flags); 3466} 3467EXPORT_SYMBOL_GPL(__wake_up_sync_key); 3468 3469/* 3470 * __wake_up_sync - see __wake_up_sync_key() 3471 */ 3472void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive) 3473{ 3474 __wake_up_sync_key(q, mode, nr_exclusive, NULL); 3475} 3476EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */ 3477 3478/** 3479 * complete: - signals a single thread waiting on this completion 3480 * @x: holds the state of this particular completion 3481 * 3482 * This will wake up a single thread waiting on this completion. Threads will be 3483 * awakened in the same order in which they were queued. 3484 * 3485 * See also complete_all(), wait_for_completion() and related routines. 3486 * 3487 * It may be assumed that this function implies a write memory barrier before 3488 * changing the task state if and only if any tasks are woken up. 3489 */ 3490void complete(struct completion *x) 3491{ 3492 unsigned long flags; 3493 3494 spin_lock_irqsave(&x->wait.lock, flags); 3495 x->done++; 3496 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL); 3497 spin_unlock_irqrestore(&x->wait.lock, flags); 3498} 3499EXPORT_SYMBOL(complete); 3500 3501/** 3502 * complete_all: - signals all threads waiting on this completion 3503 * @x: holds the state of this particular completion 3504 * 3505 * This will wake up all threads waiting on this particular completion event. 3506 * 3507 * It may be assumed that this function implies a write memory barrier before 3508 * changing the task state if and only if any tasks are woken up. 3509 */ 3510void complete_all(struct completion *x) 3511{ 3512 unsigned long flags; 3513 3514 spin_lock_irqsave(&x->wait.lock, flags); 3515 x->done += UINT_MAX/2; 3516 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL); 3517 spin_unlock_irqrestore(&x->wait.lock, flags); 3518} 3519EXPORT_SYMBOL(complete_all); 3520 3521static inline long __sched 3522do_wait_for_common(struct completion *x, long timeout, int state) 3523{ 3524 if (!x->done) { 3525 DECLARE_WAITQUEUE(wait, current); 3526 3527 __add_wait_queue_tail_exclusive(&x->wait, &wait); 3528 do { 3529 if (signal_pending_state(state, current)) { 3530 timeout = -ERESTARTSYS; 3531 break; 3532 } 3533 __set_current_state(state); 3534 spin_unlock_irq(&x->wait.lock); 3535 timeout = schedule_timeout(timeout); 3536 spin_lock_irq(&x->wait.lock); 3537 } while (!x->done && timeout); 3538 __remove_wait_queue(&x->wait, &wait); 3539 if (!x->done) 3540 return timeout; 3541 } 3542 x->done--; 3543 return timeout ?: 1; 3544} 3545 3546static long __sched 3547wait_for_common(struct completion *x, long timeout, int state) 3548{ 3549 might_sleep(); 3550 3551 spin_lock_irq(&x->wait.lock); 3552 timeout = do_wait_for_common(x, timeout, state); 3553 spin_unlock_irq(&x->wait.lock); 3554 return timeout; 3555} 3556 3557/** 3558 * wait_for_completion: - waits for completion of a task 3559 * @x: holds the state of this particular completion 3560 * 3561 * This waits to be signaled for completion of a specific task. It is NOT 3562 * interruptible and there is no timeout. 3563 * 3564 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout 3565 * and interrupt capability. Also see complete(). 3566 */ 3567void __sched wait_for_completion(struct completion *x) 3568{ 3569 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE); 3570} 3571EXPORT_SYMBOL(wait_for_completion); 3572 3573/** 3574 * wait_for_completion_timeout: - waits for completion of a task (w/timeout) 3575 * @x: holds the state of this particular completion 3576 * @timeout: timeout value in jiffies 3577 * 3578 * This waits for either a completion of a specific task to be signaled or for a 3579 * specified timeout to expire. The timeout is in jiffies. It is not 3580 * interruptible. 3581 * 3582 * The return value is 0 if timed out, and positive (at least 1, or number of 3583 * jiffies left till timeout) if completed. 3584 */ 3585unsigned long __sched 3586wait_for_completion_timeout(struct completion *x, unsigned long timeout) 3587{ 3588 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE); 3589} 3590EXPORT_SYMBOL(wait_for_completion_timeout); 3591 3592/** 3593 * wait_for_completion_interruptible: - waits for completion of a task (w/intr) 3594 * @x: holds the state of this particular completion 3595 * 3596 * This waits for completion of a specific task to be signaled. It is 3597 * interruptible. 3598 * 3599 * The return value is -ERESTARTSYS if interrupted, 0 if completed. 3600 */ 3601int __sched wait_for_completion_interruptible(struct completion *x) 3602{ 3603 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE); 3604 if (t == -ERESTARTSYS) 3605 return t; 3606 return 0; 3607} 3608EXPORT_SYMBOL(wait_for_completion_interruptible); 3609 3610/** 3611 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr)) 3612 * @x: holds the state of this particular completion 3613 * @timeout: timeout value in jiffies 3614 * 3615 * This waits for either a completion of a specific task to be signaled or for a 3616 * specified timeout to expire. It is interruptible. The timeout is in jiffies. 3617 * 3618 * The return value is -ERESTARTSYS if interrupted, 0 if timed out, 3619 * positive (at least 1, or number of jiffies left till timeout) if completed. 3620 */ 3621long __sched 3622wait_for_completion_interruptible_timeout(struct completion *x, 3623 unsigned long timeout) 3624{ 3625 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE); 3626} 3627EXPORT_SYMBOL(wait_for_completion_interruptible_timeout); 3628 3629/** 3630 * wait_for_completion_killable: - waits for completion of a task (killable) 3631 * @x: holds the state of this particular completion 3632 * 3633 * This waits to be signaled for completion of a specific task. It can be 3634 * interrupted by a kill signal. 3635 * 3636 * The return value is -ERESTARTSYS if interrupted, 0 if completed. 3637 */ 3638int __sched wait_for_completion_killable(struct completion *x) 3639{ 3640 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE); 3641 if (t == -ERESTARTSYS) 3642 return t; 3643 return 0; 3644} 3645EXPORT_SYMBOL(wait_for_completion_killable); 3646 3647/** 3648 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable)) 3649 * @x: holds the state of this particular completion 3650 * @timeout: timeout value in jiffies 3651 * 3652 * This waits for either a completion of a specific task to be 3653 * signaled or for a specified timeout to expire. It can be 3654 * interrupted by a kill signal. The timeout is in jiffies. 3655 * 3656 * The return value is -ERESTARTSYS if interrupted, 0 if timed out, 3657 * positive (at least 1, or number of jiffies left till timeout) if completed. 3658 */ 3659long __sched 3660wait_for_completion_killable_timeout(struct completion *x, 3661 unsigned long timeout) 3662{ 3663 return wait_for_common(x, timeout, TASK_KILLABLE); 3664} 3665EXPORT_SYMBOL(wait_for_completion_killable_timeout); 3666 3667/** 3668 * try_wait_for_completion - try to decrement a completion without blocking 3669 * @x: completion structure 3670 * 3671 * Returns: 0 if a decrement cannot be done without blocking 3672 * 1 if a decrement succeeded. 3673 * 3674 * If a completion is being used as a counting completion, 3675 * attempt to decrement the counter without blocking. This 3676 * enables us to avoid waiting if the resource the completion 3677 * is protecting is not available. 3678 */ 3679bool try_wait_for_completion(struct completion *x) 3680{ 3681 unsigned long flags; 3682 int ret = 1; 3683 3684 spin_lock_irqsave(&x->wait.lock, flags); 3685 if (!x->done) 3686 ret = 0; 3687 else 3688 x->done--; 3689 spin_unlock_irqrestore(&x->wait.lock, flags); 3690 return ret; 3691} 3692EXPORT_SYMBOL(try_wait_for_completion); 3693 3694/** 3695 * completion_done - Test to see if a completion has any waiters 3696 * @x: completion structure 3697 * 3698 * Returns: 0 if there are waiters (wait_for_completion() in progress) 3699 * 1 if there are no waiters. 3700 * 3701 */ 3702bool completion_done(struct completion *x) 3703{ 3704 unsigned long flags; 3705 int ret = 1; 3706 3707 spin_lock_irqsave(&x->wait.lock, flags); 3708 if (!x->done) 3709 ret = 0; 3710 spin_unlock_irqrestore(&x->wait.lock, flags); 3711 return ret; 3712} 3713EXPORT_SYMBOL(completion_done); 3714 3715static long __sched 3716sleep_on_common(wait_queue_head_t *q, int state, long timeout) 3717{ 3718 unsigned long flags; 3719 wait_queue_t wait; 3720 3721 init_waitqueue_entry(&wait, current); 3722 3723 __set_current_state(state); 3724 3725 spin_lock_irqsave(&q->lock, flags); 3726 __add_wait_queue(q, &wait); 3727 spin_unlock(&q->lock); 3728 timeout = schedule_timeout(timeout); 3729 spin_lock_irq(&q->lock); 3730 __remove_wait_queue(q, &wait); 3731 spin_unlock_irqrestore(&q->lock, flags); 3732 3733 return timeout; 3734} 3735 3736void __sched interruptible_sleep_on(wait_queue_head_t *q) 3737{ 3738 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT); 3739} 3740EXPORT_SYMBOL(interruptible_sleep_on); 3741 3742long __sched 3743interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout) 3744{ 3745 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout); 3746} 3747EXPORT_SYMBOL(interruptible_sleep_on_timeout); 3748 3749void __sched sleep_on(wait_queue_head_t *q) 3750{ 3751 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT); 3752} 3753EXPORT_SYMBOL(sleep_on); 3754 3755long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout) 3756{ 3757 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout); 3758} 3759EXPORT_SYMBOL(sleep_on_timeout); 3760 3761#ifdef CONFIG_RT_MUTEXES 3762 3763/* 3764 * rt_mutex_setprio - set the current priority of a task 3765 * @p: task 3766 * @prio: prio value (kernel-internal form) 3767 * 3768 * This function changes the 'effective' priority of a task. It does 3769 * not touch ->normal_prio like __setscheduler(). 3770 * 3771 * Used by the rt_mutex code to implement priority inheritance logic. 3772 */ 3773void rt_mutex_setprio(struct task_struct *p, int prio) 3774{ 3775 int oldprio, on_rq, running; 3776 struct rq *rq; 3777 const struct sched_class *prev_class; 3778 3779 BUG_ON(prio < 0 || prio > MAX_PRIO); 3780 3781 rq = __task_rq_lock(p); 3782 3783 /* 3784 * Idle task boosting is a nono in general. There is one 3785 * exception, when PREEMPT_RT and NOHZ is active: 3786 * 3787 * The idle task calls get_next_timer_interrupt() and holds 3788 * the timer wheel base->lock on the CPU and another CPU wants 3789 * to access the timer (probably to cancel it). We can safely 3790 * ignore the boosting request, as the idle CPU runs this code 3791 * with interrupts disabled and will complete the lock 3792 * protected section without being interrupted. So there is no 3793 * real need to boost. 3794 */ 3795 if (unlikely(p == rq->idle)) { 3796 WARN_ON(p != rq->curr); 3797 WARN_ON(p->pi_blocked_on); 3798 goto out_unlock; 3799 } 3800 3801 trace_sched_pi_setprio(p, prio); 3802 oldprio = p->prio; 3803 prev_class = p->sched_class; 3804 on_rq = p->on_rq; 3805 running = task_current(rq, p); 3806 if (on_rq) 3807 dequeue_task(rq, p, 0); 3808 if (running) 3809 p->sched_class->put_prev_task(rq, p); 3810 3811 if (rt_prio(prio)) 3812 p->sched_class = &rt_sched_class; 3813 else 3814 p->sched_class = &fair_sched_class; 3815 3816 p->prio = prio; 3817 3818 if (running) 3819 p->sched_class->set_curr_task(rq); 3820 if (on_rq) 3821 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0); 3822 3823 check_class_changed(rq, p, prev_class, oldprio); 3824out_unlock: 3825 __task_rq_unlock(rq); 3826} 3827#endif 3828void set_user_nice(struct task_struct *p, long nice) 3829{ 3830 int old_prio, delta, on_rq; 3831 unsigned long flags; 3832 struct rq *rq; 3833 3834 if (TASK_NICE(p) == nice || nice < -20 || nice > 19) 3835 return; 3836 /* 3837 * We have to be careful, if called from sys_setpriority(), 3838 * the task might be in the middle of scheduling on another CPU. 3839 */ 3840 rq = task_rq_lock(p, &flags); 3841 /* 3842 * The RT priorities are set via sched_setscheduler(), but we still 3843 * allow the 'normal' nice value to be set - but as expected 3844 * it wont have any effect on scheduling until the task is 3845 * SCHED_FIFO/SCHED_RR: 3846 */ 3847 if (task_has_rt_policy(p)) { 3848 p->static_prio = NICE_TO_PRIO(nice); 3849 goto out_unlock; 3850 } 3851 on_rq = p->on_rq; 3852 if (on_rq) 3853 dequeue_task(rq, p, 0); 3854 3855 p->static_prio = NICE_TO_PRIO(nice); 3856 set_load_weight(p); 3857 old_prio = p->prio; 3858 p->prio = effective_prio(p); 3859 delta = p->prio - old_prio; 3860 3861 if (on_rq) { 3862 enqueue_task(rq, p, 0); 3863 /* 3864 * If the task increased its priority or is running and 3865 * lowered its priority, then reschedule its CPU: 3866 */ 3867 if (delta < 0 || (delta > 0 && task_running(rq, p))) 3868 resched_task(rq->curr); 3869 } 3870out_unlock: 3871 task_rq_unlock(rq, p, &flags); 3872} 3873EXPORT_SYMBOL(set_user_nice); 3874 3875/* 3876 * can_nice - check if a task can reduce its nice value 3877 * @p: task 3878 * @nice: nice value 3879 */ 3880int can_nice(const struct task_struct *p, const int nice) 3881{ 3882 /* convert nice value [19,-20] to rlimit style value [1,40] */ 3883 int nice_rlim = 20 - nice; 3884 3885 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) || 3886 capable(CAP_SYS_NICE)); 3887} 3888 3889#ifdef __ARCH_WANT_SYS_NICE 3890 3891/* 3892 * sys_nice - change the priority of the current process. 3893 * @increment: priority increment 3894 * 3895 * sys_setpriority is a more generic, but much slower function that 3896 * does similar things. 3897 */ 3898SYSCALL_DEFINE1(nice, int, increment) 3899{ 3900 long nice, retval; 3901 3902 /* 3903 * Setpriority might change our priority at the same moment. 3904 * We don't have to worry. Conceptually one call occurs first 3905 * and we have a single winner. 3906 */ 3907 if (increment < -40) 3908 increment = -40; 3909 if (increment > 40) 3910 increment = 40; 3911 3912 nice = TASK_NICE(current) + increment; 3913 if (nice < -20) 3914 nice = -20; 3915 if (nice > 19) 3916 nice = 19; 3917 3918 if (increment < 0 && !can_nice(current, nice)) 3919 return -EPERM; 3920 3921 retval = security_task_setnice(current, nice); 3922 if (retval) 3923 return retval; 3924 3925 set_user_nice(current, nice); 3926 return 0; 3927} 3928 3929#endif 3930 3931/** 3932 * task_prio - return the priority value of a given task. 3933 * @p: the task in question. 3934 * 3935 * This is the priority value as seen by users in /proc. 3936 * RT tasks are offset by -200. Normal tasks are centered 3937 * around 0, value goes from -16 to +15. 3938 */ 3939int task_prio(const struct task_struct *p) 3940{ 3941 return p->prio - MAX_RT_PRIO; 3942} 3943 3944/** 3945 * task_nice - return the nice value of a given task. 3946 * @p: the task in question. 3947 */ 3948int task_nice(const struct task_struct *p) 3949{ 3950 return TASK_NICE(p); 3951} 3952EXPORT_SYMBOL(task_nice); 3953 3954/** 3955 * idle_cpu - is a given cpu idle currently? 3956 * @cpu: the processor in question. 3957 */ 3958int idle_cpu(int cpu) 3959{ 3960 struct rq *rq = cpu_rq(cpu); 3961 3962 if (rq->curr != rq->idle) 3963 return 0; 3964 3965 if (rq->nr_running) 3966 return 0; 3967 3968#ifdef CONFIG_SMP 3969 if (!llist_empty(&rq->wake_list)) 3970 return 0; 3971#endif 3972 3973 return 1; 3974} 3975 3976/** 3977 * idle_task - return the idle task for a given cpu. 3978 * @cpu: the processor in question. 3979 */ 3980struct task_struct *idle_task(int cpu) 3981{ 3982 return cpu_rq(cpu)->idle; 3983} 3984 3985/** 3986 * find_process_by_pid - find a process with a matching PID value. 3987 * @pid: the pid in question. 3988 */ 3989static struct task_struct *find_process_by_pid(pid_t pid) 3990{ 3991 return pid ? find_task_by_vpid(pid) : current; 3992} 3993 3994/* Actually do priority change: must hold rq lock. */ 3995static void 3996__setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio) 3997{ 3998 p->policy = policy; 3999 p->rt_priority = prio; 4000 p->normal_prio = normal_prio(p); 4001 /* we are holding p->pi_lock already */ 4002 p->prio = rt_mutex_getprio(p); 4003 if (rt_prio(p->prio)) 4004 p->sched_class = &rt_sched_class; 4005 else 4006 p->sched_class = &fair_sched_class; 4007 set_load_weight(p); 4008} 4009 4010/* 4011 * check the target process has a UID that matches the current process's 4012 */ 4013static bool check_same_owner(struct task_struct *p) 4014{ 4015 const struct cred *cred = current_cred(), *pcred; 4016 bool match; 4017 4018 rcu_read_lock(); 4019 pcred = __task_cred(p); 4020 if (cred->user->user_ns == pcred->user->user_ns) 4021 match = (cred->euid == pcred->euid || 4022 cred->euid == pcred->uid); 4023 else 4024 match = false; 4025 rcu_read_unlock(); 4026 return match; 4027} 4028 4029static int __sched_setscheduler(struct task_struct *p, int policy, 4030 const struct sched_param *param, bool user) 4031{ 4032 int retval, oldprio, oldpolicy = -1, on_rq, running; 4033 unsigned long flags; 4034 const struct sched_class *prev_class; 4035 struct rq *rq; 4036 int reset_on_fork; 4037 4038 /* may grab non-irq protected spin_locks */ 4039 BUG_ON(in_interrupt()); 4040recheck: 4041 /* double check policy once rq lock held */ 4042 if (policy < 0) { 4043 reset_on_fork = p->sched_reset_on_fork; 4044 policy = oldpolicy = p->policy; 4045 } else { 4046 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK); 4047 policy &= ~SCHED_RESET_ON_FORK; 4048 4049 if (policy != SCHED_FIFO && policy != SCHED_RR && 4050 policy != SCHED_NORMAL && policy != SCHED_BATCH && 4051 policy != SCHED_IDLE) 4052 return -EINVAL; 4053 } 4054 4055 /* 4056 * Valid priorities for SCHED_FIFO and SCHED_RR are 4057 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL, 4058 * SCHED_BATCH and SCHED_IDLE is 0. 4059 */ 4060 if (param->sched_priority < 0 || 4061 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) || 4062 (!p->mm && param->sched_priority > MAX_RT_PRIO-1)) 4063 return -EINVAL; 4064 if (rt_policy(policy) != (param->sched_priority != 0)) 4065 return -EINVAL; 4066 4067 /* 4068 * Allow unprivileged RT tasks to decrease priority: 4069 */ 4070 if (user && !capable(CAP_SYS_NICE)) { 4071 if (rt_policy(policy)) { 4072 unsigned long rlim_rtprio = 4073 task_rlimit(p, RLIMIT_RTPRIO); 4074 4075 /* can't set/change the rt policy */ 4076 if (policy != p->policy && !rlim_rtprio) 4077 return -EPERM; 4078 4079 /* can't increase priority */ 4080 if (param->sched_priority > p->rt_priority && 4081 param->sched_priority > rlim_rtprio) 4082 return -EPERM; 4083 } 4084 4085 /* 4086 * Treat SCHED_IDLE as nice 20. Only allow a switch to 4087 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it. 4088 */ 4089 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) { 4090 if (!can_nice(p, TASK_NICE(p))) 4091 return -EPERM; 4092 } 4093 4094 /* can't change other user's priorities */ 4095 if (!check_same_owner(p)) 4096 return -EPERM; 4097 4098 /* Normal users shall not reset the sched_reset_on_fork flag */ 4099 if (p->sched_reset_on_fork && !reset_on_fork) 4100 return -EPERM; 4101 } 4102 4103 if (user) { 4104 retval = security_task_setscheduler(p); 4105 if (retval) 4106 return retval; 4107 } 4108 4109 /* 4110 * make sure no PI-waiters arrive (or leave) while we are 4111 * changing the priority of the task: 4112 * 4113 * To be able to change p->policy safely, the appropriate 4114 * runqueue lock must be held. 4115 */ 4116 rq = task_rq_lock(p, &flags); 4117 4118 /* 4119 * Changing the policy of the stop threads its a very bad idea 4120 */ 4121 if (p == rq->stop) { 4122 task_rq_unlock(rq, p, &flags); 4123 return -EINVAL; 4124 } 4125 4126 /* 4127 * If not changing anything there's no need to proceed further: 4128 */ 4129 if (unlikely(policy == p->policy && (!rt_policy(policy) || 4130 param->sched_priority == p->rt_priority))) { 4131 4132 __task_rq_unlock(rq); 4133 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 4134 return 0; 4135 } 4136 4137#ifdef CONFIG_RT_GROUP_SCHED 4138 if (user) { 4139 /* 4140 * Do not allow realtime tasks into groups that have no runtime 4141 * assigned. 4142 */ 4143 if (rt_bandwidth_enabled() && rt_policy(policy) && 4144 task_group(p)->rt_bandwidth.rt_runtime == 0 && 4145 !task_group_is_autogroup(task_group(p))) { 4146 task_rq_unlock(rq, p, &flags); 4147 return -EPERM; 4148 } 4149 } 4150#endif 4151 4152 /* recheck policy now with rq lock held */ 4153 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) { 4154 policy = oldpolicy = -1; 4155 task_rq_unlock(rq, p, &flags); 4156 goto recheck; 4157 } 4158 on_rq = p->on_rq; 4159 running = task_current(rq, p); 4160 if (on_rq) 4161 dequeue_task(rq, p, 0); 4162 if (running) 4163 p->sched_class->put_prev_task(rq, p); 4164 4165 p->sched_reset_on_fork = reset_on_fork; 4166 4167 oldprio = p->prio; 4168 prev_class = p->sched_class; 4169 __setscheduler(rq, p, policy, param->sched_priority); 4170 4171 if (running) 4172 p->sched_class->set_curr_task(rq); 4173 if (on_rq) 4174 enqueue_task(rq, p, 0); 4175 4176 check_class_changed(rq, p, prev_class, oldprio); 4177 task_rq_unlock(rq, p, &flags); 4178 4179 rt_mutex_adjust_pi(p); 4180 4181 return 0; 4182} 4183 4184/** 4185 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread. 4186 * @p: the task in question. 4187 * @policy: new policy. 4188 * @param: structure containing the new RT priority. 4189 * 4190 * NOTE that the task may be already dead. 4191 */ 4192int sched_setscheduler(struct task_struct *p, int policy, 4193 const struct sched_param *param) 4194{ 4195 return __sched_setscheduler(p, policy, param, true); 4196} 4197EXPORT_SYMBOL_GPL(sched_setscheduler); 4198 4199/** 4200 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace. 4201 * @p: the task in question. 4202 * @policy: new policy. 4203 * @param: structure containing the new RT priority. 4204 * 4205 * Just like sched_setscheduler, only don't bother checking if the 4206 * current context has permission. For example, this is needed in 4207 * stop_machine(): we create temporary high priority worker threads, 4208 * but our caller might not have that capability. 4209 */ 4210int sched_setscheduler_nocheck(struct task_struct *p, int policy, 4211 const struct sched_param *param) 4212{ 4213 return __sched_setscheduler(p, policy, param, false); 4214} 4215 4216static int 4217do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param) 4218{ 4219 struct sched_param lparam; 4220 struct task_struct *p; 4221 int retval; 4222 4223 if (!param || pid < 0) 4224 return -EINVAL; 4225 if (copy_from_user(&lparam, param, sizeof(struct sched_param))) 4226 return -EFAULT; 4227 4228 rcu_read_lock(); 4229 retval = -ESRCH; 4230 p = find_process_by_pid(pid); 4231 if (p != NULL) 4232 retval = sched_setscheduler(p, policy, &lparam); 4233 rcu_read_unlock(); 4234 4235 return retval; 4236} 4237 4238/** 4239 * sys_sched_setscheduler - set/change the scheduler policy and RT priority 4240 * @pid: the pid in question. 4241 * @policy: new policy. 4242 * @param: structure containing the new RT priority. 4243 */ 4244SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy, 4245 struct sched_param __user *, param) 4246{ 4247 /* negative values for policy are not valid */ 4248 if (policy < 0) 4249 return -EINVAL; 4250 4251 return do_sched_setscheduler(pid, policy, param); 4252} 4253 4254/** 4255 * sys_sched_setparam - set/change the RT priority of a thread 4256 * @pid: the pid in question. 4257 * @param: structure containing the new RT priority. 4258 */ 4259SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param) 4260{ 4261 return do_sched_setscheduler(pid, -1, param); 4262} 4263 4264/** 4265 * sys_sched_getscheduler - get the policy (scheduling class) of a thread 4266 * @pid: the pid in question. 4267 */ 4268SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid) 4269{ 4270 struct task_struct *p; 4271 int retval; 4272 4273 if (pid < 0) 4274 return -EINVAL; 4275 4276 retval = -ESRCH; 4277 rcu_read_lock(); 4278 p = find_process_by_pid(pid); 4279 if (p) { 4280 retval = security_task_getscheduler(p); 4281 if (!retval) 4282 retval = p->policy 4283 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0); 4284 } 4285 rcu_read_unlock(); 4286 return retval; 4287} 4288 4289/** 4290 * sys_sched_getparam - get the RT priority of a thread 4291 * @pid: the pid in question. 4292 * @param: structure containing the RT priority. 4293 */ 4294SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param) 4295{ 4296 struct sched_param lp; 4297 struct task_struct *p; 4298 int retval; 4299 4300 if (!param || pid < 0) 4301 return -EINVAL; 4302 4303 rcu_read_lock(); 4304 p = find_process_by_pid(pid); 4305 retval = -ESRCH; 4306 if (!p) 4307 goto out_unlock; 4308 4309 retval = security_task_getscheduler(p); 4310 if (retval) 4311 goto out_unlock; 4312 4313 lp.sched_priority = p->rt_priority; 4314 rcu_read_unlock(); 4315 4316 /* 4317 * This one might sleep, we cannot do it with a spinlock held ... 4318 */ 4319 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0; 4320 4321 return retval; 4322 4323out_unlock: 4324 rcu_read_unlock(); 4325 return retval; 4326} 4327 4328long sched_setaffinity(pid_t pid, const struct cpumask *in_mask) 4329{ 4330 cpumask_var_t cpus_allowed, new_mask; 4331 struct task_struct *p; 4332 int retval; 4333 4334 get_online_cpus(); 4335 rcu_read_lock(); 4336 4337 p = find_process_by_pid(pid); 4338 if (!p) { 4339 rcu_read_unlock(); 4340 put_online_cpus(); 4341 return -ESRCH; 4342 } 4343 4344 /* Prevent p going away */ 4345 get_task_struct(p); 4346 rcu_read_unlock(); 4347 4348 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) { 4349 retval = -ENOMEM; 4350 goto out_put_task; 4351 } 4352 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) { 4353 retval = -ENOMEM; 4354 goto out_free_cpus_allowed; 4355 } 4356 retval = -EPERM; 4357 if (!check_same_owner(p) && !ns_capable(task_user_ns(p), CAP_SYS_NICE)) 4358 goto out_unlock; 4359 4360 retval = security_task_setscheduler(p); 4361 if (retval) 4362 goto out_unlock; 4363 4364 cpuset_cpus_allowed(p, cpus_allowed); 4365 cpumask_and(new_mask, in_mask, cpus_allowed); 4366again: 4367 retval = set_cpus_allowed_ptr(p, new_mask); 4368 4369 if (!retval) { 4370 cpuset_cpus_allowed(p, cpus_allowed); 4371 if (!cpumask_subset(new_mask, cpus_allowed)) { 4372 /* 4373 * We must have raced with a concurrent cpuset 4374 * update. Just reset the cpus_allowed to the 4375 * cpuset's cpus_allowed 4376 */ 4377 cpumask_copy(new_mask, cpus_allowed); 4378 goto again; 4379 } 4380 } 4381out_unlock: 4382 free_cpumask_var(new_mask); 4383out_free_cpus_allowed: 4384 free_cpumask_var(cpus_allowed); 4385out_put_task: 4386 put_task_struct(p); 4387 put_online_cpus(); 4388 return retval; 4389} 4390 4391static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len, 4392 struct cpumask *new_mask) 4393{ 4394 if (len < cpumask_size()) 4395 cpumask_clear(new_mask); 4396 else if (len > cpumask_size()) 4397 len = cpumask_size(); 4398 4399 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0; 4400} 4401 4402/** 4403 * sys_sched_setaffinity - set the cpu affinity of a process 4404 * @pid: pid of the process 4405 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 4406 * @user_mask_ptr: user-space pointer to the new cpu mask 4407 */ 4408SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len, 4409 unsigned long __user *, user_mask_ptr) 4410{ 4411 cpumask_var_t new_mask; 4412 int retval; 4413 4414 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) 4415 return -ENOMEM; 4416 4417 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask); 4418 if (retval == 0) 4419 retval = sched_setaffinity(pid, new_mask); 4420 free_cpumask_var(new_mask); 4421 return retval; 4422} 4423 4424long sched_getaffinity(pid_t pid, struct cpumask *mask) 4425{ 4426 struct task_struct *p; 4427 unsigned long flags; 4428 int retval; 4429 4430 get_online_cpus(); 4431 rcu_read_lock(); 4432 4433 retval = -ESRCH; 4434 p = find_process_by_pid(pid); 4435 if (!p) 4436 goto out_unlock; 4437 4438 retval = security_task_getscheduler(p); 4439 if (retval) 4440 goto out_unlock; 4441 4442 raw_spin_lock_irqsave(&p->pi_lock, flags); 4443 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask); 4444 raw_spin_unlock_irqrestore(&p->pi_lock, flags); 4445 4446out_unlock: 4447 rcu_read_unlock(); 4448 put_online_cpus(); 4449 4450 return retval; 4451} 4452 4453/** 4454 * sys_sched_getaffinity - get the cpu affinity of a process 4455 * @pid: pid of the process 4456 * @len: length in bytes of the bitmask pointed to by user_mask_ptr 4457 * @user_mask_ptr: user-space pointer to hold the current cpu mask 4458 */ 4459SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len, 4460 unsigned long __user *, user_mask_ptr) 4461{ 4462 int ret; 4463 cpumask_var_t mask; 4464 4465 if ((len * BITS_PER_BYTE) < nr_cpu_ids) 4466 return -EINVAL; 4467 if (len & (sizeof(unsigned long)-1)) 4468 return -EINVAL; 4469 4470 if (!alloc_cpumask_var(&mask, GFP_KERNEL)) 4471 return -ENOMEM; 4472 4473 ret = sched_getaffinity(pid, mask); 4474 if (ret == 0) { 4475 size_t retlen = min_t(size_t, len, cpumask_size()); 4476 4477 if (copy_to_user(user_mask_ptr, mask, retlen)) 4478 ret = -EFAULT; 4479 else 4480 ret = retlen; 4481 } 4482 free_cpumask_var(mask); 4483 4484 return ret; 4485} 4486 4487/** 4488 * sys_sched_yield - yield the current processor to other threads. 4489 * 4490 * This function yields the current CPU to other tasks. If there are no 4491 * other threads running on this CPU then this function will return. 4492 */ 4493SYSCALL_DEFINE0(sched_yield) 4494{ 4495 struct rq *rq = this_rq_lock(); 4496 4497 schedstat_inc(rq, yld_count); 4498 current->sched_class->yield_task(rq); 4499 4500 /* 4501 * Since we are going to call schedule() anyway, there's 4502 * no need to preempt or enable interrupts: 4503 */ 4504 __release(rq->lock); 4505 spin_release(&rq->lock.dep_map, 1, _THIS_IP_); 4506 do_raw_spin_unlock(&rq->lock); 4507 sched_preempt_enable_no_resched(); 4508 4509 schedule(); 4510 4511 return 0; 4512} 4513 4514static inline int should_resched(void) 4515{ 4516 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE); 4517} 4518 4519static void __cond_resched(void) 4520{ 4521 add_preempt_count(PREEMPT_ACTIVE); 4522 __schedule(); 4523 sub_preempt_count(PREEMPT_ACTIVE); 4524} 4525 4526int __sched _cond_resched(void) 4527{ 4528 if (should_resched()) { 4529 __cond_resched(); 4530 return 1; 4531 } 4532 return 0; 4533} 4534EXPORT_SYMBOL(_cond_resched); 4535 4536/* 4537 * __cond_resched_lock() - if a reschedule is pending, drop the given lock, 4538 * call schedule, and on return reacquire the lock. 4539 * 4540 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level 4541 * operations here to prevent schedule() from being called twice (once via 4542 * spin_unlock(), once by hand). 4543 */ 4544int __cond_resched_lock(spinlock_t *lock) 4545{ 4546 int resched = should_resched(); 4547 int ret = 0; 4548 4549 lockdep_assert_held(lock); 4550 4551 if (spin_needbreak(lock) || resched) { 4552 spin_unlock(lock); 4553 if (resched) 4554 __cond_resched(); 4555 else 4556 cpu_relax(); 4557 ret = 1; 4558 spin_lock(lock); 4559 } 4560 return ret; 4561} 4562EXPORT_SYMBOL(__cond_resched_lock); 4563 4564int __sched __cond_resched_softirq(void) 4565{ 4566 BUG_ON(!in_softirq()); 4567 4568 if (should_resched()) { 4569 local_bh_enable(); 4570 __cond_resched(); 4571 local_bh_disable(); 4572 return 1; 4573 } 4574 return 0; 4575} 4576EXPORT_SYMBOL(__cond_resched_softirq); 4577 4578/** 4579 * yield - yield the current processor to other threads. 4580 * 4581 * Do not ever use this function, there's a 99% chance you're doing it wrong. 4582 * 4583 * The scheduler is at all times free to pick the calling task as the most 4584 * eligible task to run, if removing the yield() call from your code breaks 4585 * it, its already broken. 4586 * 4587 * Typical broken usage is: 4588 * 4589 * while (!event) 4590 * yield(); 4591 * 4592 * where one assumes that yield() will let 'the other' process run that will 4593 * make event true. If the current task is a SCHED_FIFO task that will never 4594 * happen. Never use yield() as a progress guarantee!! 4595 * 4596 * If you want to use yield() to wait for something, use wait_event(). 4597 * If you want to use yield() to be 'nice' for others, use cond_resched(). 4598 * If you still want to use yield(), do not! 4599 */ 4600void __sched yield(void) 4601{ 4602 set_current_state(TASK_RUNNING); 4603 sys_sched_yield(); 4604} 4605EXPORT_SYMBOL(yield); 4606 4607/** 4608 * yield_to - yield the current processor to another thread in 4609 * your thread group, or accelerate that thread toward the 4610 * processor it's on. 4611 * @p: target task 4612 * @preempt: whether task preemption is allowed or not 4613 * 4614 * It's the caller's job to ensure that the target task struct 4615 * can't go away on us before we can do any checks. 4616 * 4617 * Returns true if we indeed boosted the target task. 4618 */ 4619bool __sched yield_to(struct task_struct *p, bool preempt) 4620{ 4621 struct task_struct *curr = current; 4622 struct rq *rq, *p_rq; 4623 unsigned long flags; 4624 bool yielded = 0; 4625 4626 local_irq_save(flags); 4627 rq = this_rq(); 4628 4629again: 4630 p_rq = task_rq(p); 4631 double_rq_lock(rq, p_rq); 4632 while (task_rq(p) != p_rq) { 4633 double_rq_unlock(rq, p_rq); 4634 goto again; 4635 } 4636 4637 if (!curr->sched_class->yield_to_task) 4638 goto out; 4639 4640 if (curr->sched_class != p->sched_class) 4641 goto out; 4642 4643 if (task_running(p_rq, p) || p->state) 4644 goto out; 4645 4646 yielded = curr->sched_class->yield_to_task(rq, p, preempt); 4647 if (yielded) { 4648 schedstat_inc(rq, yld_count); 4649 /* 4650 * Make p's CPU reschedule; pick_next_entity takes care of 4651 * fairness. 4652 */ 4653 if (preempt && rq != p_rq) 4654 resched_task(p_rq->curr); 4655 } else { 4656 /* 4657 * We might have set it in task_yield_fair(), but are 4658 * not going to schedule(), so don't want to skip 4659 * the next update. 4660 */ 4661 rq->skip_clock_update = 0; 4662 } 4663 4664out: 4665 double_rq_unlock(rq, p_rq); 4666 local_irq_restore(flags); 4667 4668 if (yielded) 4669 schedule(); 4670 4671 return yielded; 4672} 4673EXPORT_SYMBOL_GPL(yield_to); 4674 4675/* 4676 * This task is about to go to sleep on IO. Increment rq->nr_iowait so 4677 * that process accounting knows that this is a task in IO wait state. 4678 */ 4679void __sched io_schedule(void) 4680{ 4681 struct rq *rq = raw_rq(); 4682 4683 delayacct_blkio_start(); 4684 atomic_inc(&rq->nr_iowait); 4685 blk_flush_plug(current); 4686 current->in_iowait = 1; 4687 schedule(); 4688 current->in_iowait = 0; 4689 atomic_dec(&rq->nr_iowait); 4690 delayacct_blkio_end(); 4691} 4692EXPORT_SYMBOL(io_schedule); 4693 4694long __sched io_schedule_timeout(long timeout) 4695{ 4696 struct rq *rq = raw_rq(); 4697 long ret; 4698 4699 delayacct_blkio_start(); 4700 atomic_inc(&rq->nr_iowait); 4701 blk_flush_plug(current); 4702 current->in_iowait = 1; 4703 ret = schedule_timeout(timeout); 4704 current->in_iowait = 0; 4705 atomic_dec(&rq->nr_iowait); 4706 delayacct_blkio_end(); 4707 return ret; 4708} 4709 4710/** 4711 * sys_sched_get_priority_max - return maximum RT priority. 4712 * @policy: scheduling class. 4713 * 4714 * this syscall returns the maximum rt_priority that can be used 4715 * by a given scheduling class. 4716 */ 4717SYSCALL_DEFINE1(sched_get_priority_max, int, policy) 4718{ 4719 int ret = -EINVAL; 4720 4721 switch (policy) { 4722 case SCHED_FIFO: 4723 case SCHED_RR: 4724 ret = MAX_USER_RT_PRIO-1; 4725 break; 4726 case SCHED_NORMAL: 4727 case SCHED_BATCH: 4728 case SCHED_IDLE: 4729 ret = 0; 4730 break; 4731 } 4732 return ret; 4733} 4734 4735/** 4736 * sys_sched_get_priority_min - return minimum RT priority. 4737 * @policy: scheduling class. 4738 * 4739 * this syscall returns the minimum rt_priority that can be used 4740 * by a given scheduling class. 4741 */ 4742SYSCALL_DEFINE1(sched_get_priority_min, int, policy) 4743{ 4744 int ret = -EINVAL; 4745 4746 switch (policy) { 4747 case SCHED_FIFO: 4748 case SCHED_RR: 4749 ret = 1; 4750 break; 4751 case SCHED_NORMAL: 4752 case SCHED_BATCH: 4753 case SCHED_IDLE: 4754 ret = 0; 4755 } 4756 return ret; 4757} 4758 4759/** 4760 * sys_sched_rr_get_interval - return the default timeslice of a process. 4761 * @pid: pid of the process. 4762 * @interval: userspace pointer to the timeslice value. 4763 * 4764 * this syscall writes the default timeslice value of a given process 4765 * into the user-space timespec buffer. A value of '0' means infinity. 4766 */ 4767SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid, 4768 struct timespec __user *, interval) 4769{ 4770 struct task_struct *p; 4771 unsigned int time_slice; 4772 unsigned long flags; 4773 struct rq *rq; 4774 int retval; 4775 struct timespec t; 4776 4777 if (pid < 0) 4778 return -EINVAL; 4779 4780 retval = -ESRCH; 4781 rcu_read_lock(); 4782 p = find_process_by_pid(pid); 4783 if (!p) 4784 goto out_unlock; 4785 4786 retval = security_task_getscheduler(p); 4787 if (retval) 4788 goto out_unlock; 4789 4790 rq = task_rq_lock(p, &flags); 4791 time_slice = p->sched_class->get_rr_interval(rq, p); 4792 task_rq_unlock(rq, p, &flags); 4793 4794 rcu_read_unlock(); 4795 jiffies_to_timespec(time_slice, &t); 4796 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0; 4797 return retval; 4798 4799out_unlock: 4800 rcu_read_unlock(); 4801 return retval; 4802} 4803 4804static const char stat_nam[] = TASK_STATE_TO_CHAR_STR; 4805 4806void sched_show_task(struct task_struct *p) 4807{ 4808 unsigned long free = 0; 4809 unsigned state; 4810 4811 state = p->state ? __ffs(p->state) + 1 : 0; 4812 printk(KERN_INFO "%-15.15s %c", p->comm, 4813 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?'); 4814#if BITS_PER_LONG == 32 4815 if (state == TASK_RUNNING) 4816 printk(KERN_CONT " running "); 4817 else 4818 printk(KERN_CONT " %08lx ", thread_saved_pc(p)); 4819#else 4820 if (state == TASK_RUNNING) 4821 printk(KERN_CONT " running task "); 4822 else 4823 printk(KERN_CONT " %016lx ", thread_saved_pc(p)); 4824#endif 4825#ifdef CONFIG_DEBUG_STACK_USAGE 4826 free = stack_not_used(p); 4827#endif 4828 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free, 4829 task_pid_nr(p), task_pid_nr(rcu_dereference(p->real_parent)), 4830 (unsigned long)task_thread_info(p)->flags); 4831 4832 show_stack(p, NULL); 4833} 4834 4835void show_state_filter(unsigned long state_filter) 4836{ 4837 struct task_struct *g, *p; 4838 4839#if BITS_PER_LONG == 32 4840 printk(KERN_INFO 4841 " task PC stack pid father\n"); 4842#else 4843 printk(KERN_INFO 4844 " task PC stack pid father\n"); 4845#endif 4846 rcu_read_lock(); 4847 do_each_thread(g, p) { 4848 /* 4849 * reset the NMI-timeout, listing all files on a slow 4850 * console might take a lot of time: 4851 */ 4852 touch_nmi_watchdog(); 4853 if (!state_filter || (p->state & state_filter)) 4854 sched_show_task(p); 4855 } while_each_thread(g, p); 4856 4857 touch_all_softlockup_watchdogs(); 4858 4859#ifdef CONFIG_SCHED_DEBUG 4860 sysrq_sched_debug_show(); 4861#endif 4862 rcu_read_unlock(); 4863 /* 4864 * Only show locks if all tasks are dumped: 4865 */ 4866 if (!state_filter) 4867 debug_show_all_locks(); 4868} 4869 4870void __cpuinit init_idle_bootup_task(struct task_struct *idle) 4871{ 4872 idle->sched_class = &idle_sched_class; 4873} 4874 4875/** 4876 * init_idle - set up an idle thread for a given CPU 4877 * @idle: task in question 4878 * @cpu: cpu the idle task belongs to 4879 * 4880 * NOTE: this function does not set the idle thread's NEED_RESCHED 4881 * flag, to make booting more robust. 4882 */ 4883void __cpuinit init_idle(struct task_struct *idle, int cpu) 4884{ 4885 struct rq *rq = cpu_rq(cpu); 4886 unsigned long flags; 4887 4888 raw_spin_lock_irqsave(&rq->lock, flags); 4889 4890 __sched_fork(idle); 4891 idle->state = TASK_RUNNING; 4892 idle->se.exec_start = sched_clock(); 4893 4894 do_set_cpus_allowed(idle, cpumask_of(cpu)); 4895 /* 4896 * We're having a chicken and egg problem, even though we are 4897 * holding rq->lock, the cpu isn't yet set to this cpu so the 4898 * lockdep check in task_group() will fail. 4899 * 4900 * Similar case to sched_fork(). / Alternatively we could 4901 * use task_rq_lock() here and obtain the other rq->lock. 4902 * 4903 * Silence PROVE_RCU 4904 */ 4905 rcu_read_lock(); 4906 __set_task_cpu(idle, cpu); 4907 rcu_read_unlock(); 4908 4909 rq->curr = rq->idle = idle; 4910#if defined(CONFIG_SMP) 4911 idle->on_cpu = 1; 4912#endif 4913 raw_spin_unlock_irqrestore(&rq->lock, flags); 4914 4915 /* Set the preempt count _outside_ the spinlocks! */ 4916 task_thread_info(idle)->preempt_count = 0; 4917 4918 /* 4919 * The idle tasks have their own, simple scheduling class: 4920 */ 4921 idle->sched_class = &idle_sched_class; 4922 ftrace_graph_init_idle_task(idle, cpu); 4923#if defined(CONFIG_SMP) 4924 sprintf(idle->comm, "%s/%d", INIT_TASK_COMM, cpu); 4925#endif 4926} 4927 4928#ifdef CONFIG_SMP 4929void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask) 4930{ 4931 if (p->sched_class && p->sched_class->set_cpus_allowed) 4932 p->sched_class->set_cpus_allowed(p, new_mask); 4933 4934 cpumask_copy(&p->cpus_allowed, new_mask); 4935 p->rt.nr_cpus_allowed = cpumask_weight(new_mask); 4936} 4937 4938/* 4939 * This is how migration works: 4940 * 4941 * 1) we invoke migration_cpu_stop() on the target CPU using 4942 * stop_one_cpu(). 4943 * 2) stopper starts to run (implicitly forcing the migrated thread 4944 * off the CPU) 4945 * 3) it checks whether the migrated task is still in the wrong runqueue. 4946 * 4) if it's in the wrong runqueue then the migration thread removes 4947 * it and puts it into the right queue. 4948 * 5) stopper completes and stop_one_cpu() returns and the migration 4949 * is done. 4950 */ 4951 4952/* 4953 * Change a given task's CPU affinity. Migrate the thread to a 4954 * proper CPU and schedule it away if the CPU it's executing on 4955 * is removed from the allowed bitmask. 4956 * 4957 * NOTE: the caller must have a valid reference to the task, the 4958 * task must not exit() & deallocate itself prematurely. The 4959 * call is not atomic; no spinlocks may be held. 4960 */ 4961int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask) 4962{ 4963 unsigned long flags; 4964 struct rq *rq; 4965 unsigned int dest_cpu; 4966 int ret = 0; 4967 4968 rq = task_rq_lock(p, &flags); 4969 4970 if (cpumask_equal(&p->cpus_allowed, new_mask)) 4971 goto out; 4972 4973 if (!cpumask_intersects(new_mask, cpu_active_mask)) { 4974 ret = -EINVAL; 4975 goto out; 4976 } 4977 4978 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current)) { 4979 ret = -EINVAL; 4980 goto out; 4981 } 4982 4983 do_set_cpus_allowed(p, new_mask); 4984 4985 /* Can the task run on the task's current CPU? If so, we're done */ 4986 if (cpumask_test_cpu(task_cpu(p), new_mask)) 4987 goto out; 4988 4989 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask); 4990 if (p->on_rq) { 4991 struct migration_arg arg = { p, dest_cpu }; 4992 /* Need help from migration thread: drop lock and wait. */ 4993 task_rq_unlock(rq, p, &flags); 4994 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg); 4995 tlb_migrate_finish(p->mm); 4996 return 0; 4997 } 4998out: 4999 task_rq_unlock(rq, p, &flags); 5000 5001 return ret; 5002} 5003EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr); 5004 5005/* 5006 * Move (not current) task off this cpu, onto dest cpu. We're doing 5007 * this because either it can't run here any more (set_cpus_allowed() 5008 * away from this CPU, or CPU going down), or because we're 5009 * attempting to rebalance this task on exec (sched_exec). 5010 * 5011 * So we race with normal scheduler movements, but that's OK, as long 5012 * as the task is no longer on this CPU. 5013 * 5014 * Returns non-zero if task was successfully migrated. 5015 */ 5016static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu) 5017{ 5018 struct rq *rq_dest, *rq_src; 5019 int ret = 0; 5020 5021 if (unlikely(!cpu_active(dest_cpu))) 5022 return ret; 5023 5024 rq_src = cpu_rq(src_cpu); 5025 rq_dest = cpu_rq(dest_cpu); 5026 5027 raw_spin_lock(&p->pi_lock); 5028 double_rq_lock(rq_src, rq_dest); 5029 /* Already moved. */ 5030 if (task_cpu(p) != src_cpu) 5031 goto done; 5032 /* Affinity changed (again). */ 5033 if (!cpumask_test_cpu(dest_cpu, tsk_cpus_allowed(p))) 5034 goto fail; 5035 5036 /* 5037 * If we're not on a rq, the next wake-up will ensure we're 5038 * placed properly. 5039 */ 5040 if (p->on_rq) { 5041 dequeue_task(rq_src, p, 0); 5042 set_task_cpu(p, dest_cpu); 5043 enqueue_task(rq_dest, p, 0); 5044 check_preempt_curr(rq_dest, p, 0); 5045 } 5046done: 5047 ret = 1; 5048fail: 5049 double_rq_unlock(rq_src, rq_dest); 5050 raw_spin_unlock(&p->pi_lock); 5051 return ret; 5052} 5053 5054/* 5055 * migration_cpu_stop - this will be executed by a highprio stopper thread 5056 * and performs thread migration by bumping thread off CPU then 5057 * 'pushing' onto another runqueue. 5058 */ 5059static int migration_cpu_stop(void *data) 5060{ 5061 struct migration_arg *arg = data; 5062 5063 /* 5064 * The original target cpu might have gone down and we might 5065 * be on another cpu but it doesn't matter. 5066 */ 5067 local_irq_disable(); 5068 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu); 5069 local_irq_enable(); 5070 return 0; 5071} 5072 5073#ifdef CONFIG_HOTPLUG_CPU 5074 5075/* 5076 * Ensures that the idle task is using init_mm right before its cpu goes 5077 * offline. 5078 */ 5079void idle_task_exit(void) 5080{ 5081 struct mm_struct *mm = current->active_mm; 5082 5083 BUG_ON(cpu_online(smp_processor_id())); 5084 5085 if (mm != &init_mm) 5086 switch_mm(mm, &init_mm, current); 5087 mmdrop(mm); 5088} 5089 5090/* 5091 * While a dead CPU has no uninterruptible tasks queued at this point, 5092 * it might still have a nonzero ->nr_uninterruptible counter, because 5093 * for performance reasons the counter is not stricly tracking tasks to 5094 * their home CPUs. So we just add the counter to another CPU's counter, 5095 * to keep the global sum constant after CPU-down: 5096 */ 5097static void migrate_nr_uninterruptible(struct rq *rq_src) 5098{ 5099 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask)); 5100 5101 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible; 5102 rq_src->nr_uninterruptible = 0; 5103} 5104 5105/* 5106 * remove the tasks which were accounted by rq from calc_load_tasks. 5107 */ 5108static void calc_global_load_remove(struct rq *rq) 5109{ 5110 atomic_long_sub(rq->calc_load_active, &calc_load_tasks); 5111 rq->calc_load_active = 0; 5112} 5113 5114/* 5115 * Migrate all tasks from the rq, sleeping tasks will be migrated by 5116 * try_to_wake_up()->select_task_rq(). 5117 * 5118 * Called with rq->lock held even though we'er in stop_machine() and 5119 * there's no concurrency possible, we hold the required locks anyway 5120 * because of lock validation efforts. 5121 */ 5122static void migrate_tasks(unsigned int dead_cpu) 5123{ 5124 struct rq *rq = cpu_rq(dead_cpu); 5125 struct task_struct *next, *stop = rq->stop; 5126 int dest_cpu; 5127 5128 /* 5129 * Fudge the rq selection such that the below task selection loop 5130 * doesn't get stuck on the currently eligible stop task. 5131 * 5132 * We're currently inside stop_machine() and the rq is either stuck 5133 * in the stop_machine_cpu_stop() loop, or we're executing this code, 5134 * either way we should never end up calling schedule() until we're 5135 * done here. 5136 */ 5137 rq->stop = NULL; 5138 5139 /* Ensure any throttled groups are reachable by pick_next_task */ 5140 unthrottle_offline_cfs_rqs(rq); 5141 5142 for ( ; ; ) { 5143 /* 5144 * There's this thread running, bail when that's the only 5145 * remaining thread. 5146 */ 5147 if (rq->nr_running == 1) 5148 break; 5149 5150 next = pick_next_task(rq); 5151 BUG_ON(!next); 5152 next->sched_class->put_prev_task(rq, next); 5153 5154 /* Find suitable destination for @next, with force if needed. */ 5155 dest_cpu = select_fallback_rq(dead_cpu, next); 5156 raw_spin_unlock(&rq->lock); 5157 5158 __migrate_task(next, dead_cpu, dest_cpu); 5159 5160 raw_spin_lock(&rq->lock); 5161 } 5162 5163 rq->stop = stop; 5164} 5165 5166#endif /* CONFIG_HOTPLUG_CPU */ 5167 5168#if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL) 5169 5170static struct ctl_table sd_ctl_dir[] = { 5171 { 5172 .procname = "sched_domain", 5173 .mode = 0555, 5174 }, 5175 {} 5176}; 5177 5178static struct ctl_table sd_ctl_root[] = { 5179 { 5180 .procname = "kernel", 5181 .mode = 0555, 5182 .child = sd_ctl_dir, 5183 }, 5184 {} 5185}; 5186 5187static struct ctl_table *sd_alloc_ctl_entry(int n) 5188{ 5189 struct ctl_table *entry = 5190 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL); 5191 5192 return entry; 5193} 5194 5195static void sd_free_ctl_entry(struct ctl_table **tablep) 5196{ 5197 struct ctl_table *entry; 5198 5199 /* 5200 * In the intermediate directories, both the child directory and 5201 * procname are dynamically allocated and could fail but the mode 5202 * will always be set. In the lowest directory the names are 5203 * static strings and all have proc handlers. 5204 */ 5205 for (entry = *tablep; entry->mode; entry++) { 5206 if (entry->child) 5207 sd_free_ctl_entry(&entry->child); 5208 if (entry->proc_handler == NULL) 5209 kfree(entry->procname); 5210 } 5211 5212 kfree(*tablep); 5213 *tablep = NULL; 5214} 5215 5216static void 5217set_table_entry(struct ctl_table *entry, 5218 const char *procname, void *data, int maxlen, 5219 umode_t mode, proc_handler *proc_handler) 5220{ 5221 entry->procname = procname; 5222 entry->data = data; 5223 entry->maxlen = maxlen; 5224 entry->mode = mode; 5225 entry->proc_handler = proc_handler; 5226} 5227 5228static struct ctl_table * 5229sd_alloc_ctl_domain_table(struct sched_domain *sd) 5230{ 5231 struct ctl_table *table = sd_alloc_ctl_entry(13); 5232 5233 if (table == NULL) 5234 return NULL; 5235 5236 set_table_entry(&table[0], "min_interval", &sd->min_interval, 5237 sizeof(long), 0644, proc_doulongvec_minmax); 5238 set_table_entry(&table[1], "max_interval", &sd->max_interval, 5239 sizeof(long), 0644, proc_doulongvec_minmax); 5240 set_table_entry(&table[2], "busy_idx", &sd->busy_idx, 5241 sizeof(int), 0644, proc_dointvec_minmax); 5242 set_table_entry(&table[3], "idle_idx", &sd->idle_idx, 5243 sizeof(int), 0644, proc_dointvec_minmax); 5244 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx, 5245 sizeof(int), 0644, proc_dointvec_minmax); 5246 set_table_entry(&table[5], "wake_idx", &sd->wake_idx, 5247 sizeof(int), 0644, proc_dointvec_minmax); 5248 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx, 5249 sizeof(int), 0644, proc_dointvec_minmax); 5250 set_table_entry(&table[7], "busy_factor", &sd->busy_factor, 5251 sizeof(int), 0644, proc_dointvec_minmax); 5252 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct, 5253 sizeof(int), 0644, proc_dointvec_minmax); 5254 set_table_entry(&table[9], "cache_nice_tries", 5255 &sd->cache_nice_tries, 5256 sizeof(int), 0644, proc_dointvec_minmax); 5257 set_table_entry(&table[10], "flags", &sd->flags, 5258 sizeof(int), 0644, proc_dointvec_minmax); 5259 set_table_entry(&table[11], "name", sd->name, 5260 CORENAME_MAX_SIZE, 0444, proc_dostring); 5261 /* &table[12] is terminator */ 5262 5263 return table; 5264} 5265 5266static ctl_table *sd_alloc_ctl_cpu_table(int cpu) 5267{ 5268 struct ctl_table *entry, *table; 5269 struct sched_domain *sd; 5270 int domain_num = 0, i; 5271 char buf[32]; 5272 5273 for_each_domain(cpu, sd) 5274 domain_num++; 5275 entry = table = sd_alloc_ctl_entry(domain_num + 1); 5276 if (table == NULL) 5277 return NULL; 5278 5279 i = 0; 5280 for_each_domain(cpu, sd) { 5281 snprintf(buf, 32, "domain%d", i); 5282 entry->procname = kstrdup(buf, GFP_KERNEL); 5283 entry->mode = 0555; 5284 entry->child = sd_alloc_ctl_domain_table(sd); 5285 entry++; 5286 i++; 5287 } 5288 return table; 5289} 5290 5291static struct ctl_table_header *sd_sysctl_header; 5292static void register_sched_domain_sysctl(void) 5293{ 5294 int i, cpu_num = num_possible_cpus(); 5295 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1); 5296 char buf[32]; 5297 5298 WARN_ON(sd_ctl_dir[0].child); 5299 sd_ctl_dir[0].child = entry; 5300 5301 if (entry == NULL) 5302 return; 5303 5304 for_each_possible_cpu(i) { 5305 snprintf(buf, 32, "cpu%d", i); 5306 entry->procname = kstrdup(buf, GFP_KERNEL); 5307 entry->mode = 0555; 5308 entry->child = sd_alloc_ctl_cpu_table(i); 5309 entry++; 5310 } 5311 5312 WARN_ON(sd_sysctl_header); 5313 sd_sysctl_header = register_sysctl_table(sd_ctl_root); 5314} 5315 5316/* may be called multiple times per register */ 5317static void unregister_sched_domain_sysctl(void) 5318{ 5319 if (sd_sysctl_header) 5320 unregister_sysctl_table(sd_sysctl_header); 5321 sd_sysctl_header = NULL; 5322 if (sd_ctl_dir[0].child) 5323 sd_free_ctl_entry(&sd_ctl_dir[0].child); 5324} 5325#else 5326static void register_sched_domain_sysctl(void) 5327{ 5328} 5329static void unregister_sched_domain_sysctl(void) 5330{ 5331} 5332#endif 5333 5334static void set_rq_online(struct rq *rq) 5335{ 5336 if (!rq->online) { 5337 const struct sched_class *class; 5338 5339 cpumask_set_cpu(rq->cpu, rq->rd->online); 5340 rq->online = 1; 5341 5342 for_each_class(class) { 5343 if (class->rq_online) 5344 class->rq_online(rq); 5345 } 5346 } 5347} 5348 5349static void set_rq_offline(struct rq *rq) 5350{ 5351 if (rq->online) { 5352 const struct sched_class *class; 5353 5354 for_each_class(class) { 5355 if (class->rq_offline) 5356 class->rq_offline(rq); 5357 } 5358 5359 cpumask_clear_cpu(rq->cpu, rq->rd->online); 5360 rq->online = 0; 5361 } 5362} 5363 5364/* 5365 * migration_call - callback that gets triggered when a CPU is added. 5366 * Here we can start up the necessary migration thread for the new CPU. 5367 */ 5368static int __cpuinit 5369migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu) 5370{ 5371 int cpu = (long)hcpu; 5372 unsigned long flags; 5373 struct rq *rq = cpu_rq(cpu); 5374 5375 switch (action & ~CPU_TASKS_FROZEN) { 5376 5377 case CPU_UP_PREPARE: 5378 rq->calc_load_update = calc_load_update; 5379 break; 5380 5381 case CPU_ONLINE: 5382 /* Update our root-domain */ 5383 raw_spin_lock_irqsave(&rq->lock, flags); 5384 if (rq->rd) { 5385 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 5386 5387 set_rq_online(rq); 5388 } 5389 raw_spin_unlock_irqrestore(&rq->lock, flags); 5390 break; 5391 5392#ifdef CONFIG_HOTPLUG_CPU 5393 case CPU_DYING: 5394 sched_ttwu_pending(); 5395 /* Update our root-domain */ 5396 raw_spin_lock_irqsave(&rq->lock, flags); 5397 if (rq->rd) { 5398 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span)); 5399 set_rq_offline(rq); 5400 } 5401 migrate_tasks(cpu); 5402 BUG_ON(rq->nr_running != 1); /* the migration thread */ 5403 raw_spin_unlock_irqrestore(&rq->lock, flags); 5404 5405 migrate_nr_uninterruptible(rq); 5406 calc_global_load_remove(rq); 5407 break; 5408#endif 5409 } 5410 5411 update_max_interval(); 5412 5413 return NOTIFY_OK; 5414} 5415 5416/* 5417 * Register at high priority so that task migration (migrate_all_tasks) 5418 * happens before everything else. This has to be lower priority than 5419 * the notifier in the perf_event subsystem, though. 5420 */ 5421static struct notifier_block __cpuinitdata migration_notifier = { 5422 .notifier_call = migration_call, 5423 .priority = CPU_PRI_MIGRATION, 5424}; 5425 5426static int __cpuinit sched_cpu_active(struct notifier_block *nfb, 5427 unsigned long action, void *hcpu) 5428{ 5429 switch (action & ~CPU_TASKS_FROZEN) { 5430 case CPU_STARTING: 5431 case CPU_DOWN_FAILED: 5432 set_cpu_active((long)hcpu, true); 5433 return NOTIFY_OK; 5434 default: 5435 return NOTIFY_DONE; 5436 } 5437} 5438 5439static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb, 5440 unsigned long action, void *hcpu) 5441{ 5442 switch (action & ~CPU_TASKS_FROZEN) { 5443 case CPU_DOWN_PREPARE: 5444 set_cpu_active((long)hcpu, false); 5445 return NOTIFY_OK; 5446 default: 5447 return NOTIFY_DONE; 5448 } 5449} 5450 5451static int __init migration_init(void) 5452{ 5453 void *cpu = (void *)(long)smp_processor_id(); 5454 int err; 5455 5456 /* Initialize migration for the boot CPU */ 5457 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu); 5458 BUG_ON(err == NOTIFY_BAD); 5459 migration_call(&migration_notifier, CPU_ONLINE, cpu); 5460 register_cpu_notifier(&migration_notifier); 5461 5462 /* Register cpu active notifiers */ 5463 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE); 5464 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE); 5465 5466 return 0; 5467} 5468early_initcall(migration_init); 5469#endif 5470 5471#ifdef CONFIG_SMP 5472 5473static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */ 5474 5475#ifdef CONFIG_SCHED_DEBUG 5476 5477static __read_mostly int sched_domain_debug_enabled; 5478 5479static int __init sched_domain_debug_setup(char *str) 5480{ 5481 sched_domain_debug_enabled = 1; 5482 5483 return 0; 5484} 5485early_param("sched_debug", sched_domain_debug_setup); 5486 5487static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level, 5488 struct cpumask *groupmask) 5489{ 5490 struct sched_group *group = sd->groups; 5491 char str[256]; 5492 5493 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd)); 5494 cpumask_clear(groupmask); 5495 5496 printk(KERN_DEBUG "%*s domain %d: ", level, "", level); 5497 5498 if (!(sd->flags & SD_LOAD_BALANCE)) { 5499 printk("does not load-balance\n"); 5500 if (sd->parent) 5501 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain" 5502 " has parent"); 5503 return -1; 5504 } 5505 5506 printk(KERN_CONT "span %s level %s\n", str, sd->name); 5507 5508 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) { 5509 printk(KERN_ERR "ERROR: domain->span does not contain " 5510 "CPU%d\n", cpu); 5511 } 5512 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) { 5513 printk(KERN_ERR "ERROR: domain->groups does not contain" 5514 " CPU%d\n", cpu); 5515 } 5516 5517 printk(KERN_DEBUG "%*s groups:", level + 1, ""); 5518 do { 5519 if (!group) { 5520 printk("\n"); 5521 printk(KERN_ERR "ERROR: group is NULL\n"); 5522 break; 5523 } 5524 5525 if (!group->sgp->power) { 5526 printk(KERN_CONT "\n"); 5527 printk(KERN_ERR "ERROR: domain->cpu_power not " 5528 "set\n"); 5529 break; 5530 } 5531 5532 if (!cpumask_weight(sched_group_cpus(group))) { 5533 printk(KERN_CONT "\n"); 5534 printk(KERN_ERR "ERROR: empty group\n"); 5535 break; 5536 } 5537 5538 if (cpumask_intersects(groupmask, sched_group_cpus(group))) { 5539 printk(KERN_CONT "\n"); 5540 printk(KERN_ERR "ERROR: repeated CPUs\n"); 5541 break; 5542 } 5543 5544 cpumask_or(groupmask, groupmask, sched_group_cpus(group)); 5545 5546 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group)); 5547 5548 printk(KERN_CONT " %s", str); 5549 if (group->sgp->power != SCHED_POWER_SCALE) { 5550 printk(KERN_CONT " (cpu_power = %d)", 5551 group->sgp->power); 5552 } 5553 5554 group = group->next; 5555 } while (group != sd->groups); 5556 printk(KERN_CONT "\n"); 5557 5558 if (!cpumask_equal(sched_domain_span(sd), groupmask)) 5559 printk(KERN_ERR "ERROR: groups don't span domain->span\n"); 5560 5561 if (sd->parent && 5562 !cpumask_subset(groupmask, sched_domain_span(sd->parent))) 5563 printk(KERN_ERR "ERROR: parent span is not a superset " 5564 "of domain->span\n"); 5565 return 0; 5566} 5567 5568static void sched_domain_debug(struct sched_domain *sd, int cpu) 5569{ 5570 int level = 0; 5571 5572 if (!sched_domain_debug_enabled) 5573 return; 5574 5575 if (!sd) { 5576 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu); 5577 return; 5578 } 5579 5580 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu); 5581 5582 for (;;) { 5583 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask)) 5584 break; 5585 level++; 5586 sd = sd->parent; 5587 if (!sd) 5588 break; 5589 } 5590} 5591#else /* !CONFIG_SCHED_DEBUG */ 5592# define sched_domain_debug(sd, cpu) do { } while (0) 5593#endif /* CONFIG_SCHED_DEBUG */ 5594 5595static int sd_degenerate(struct sched_domain *sd) 5596{ 5597 if (cpumask_weight(sched_domain_span(sd)) == 1) 5598 return 1; 5599 5600 /* Following flags need at least 2 groups */ 5601 if (sd->flags & (SD_LOAD_BALANCE | 5602 SD_BALANCE_NEWIDLE | 5603 SD_BALANCE_FORK | 5604 SD_BALANCE_EXEC | 5605 SD_SHARE_CPUPOWER | 5606 SD_SHARE_PKG_RESOURCES)) { 5607 if (sd->groups != sd->groups->next) 5608 return 0; 5609 } 5610 5611 /* Following flags don't use groups */ 5612 if (sd->flags & (SD_WAKE_AFFINE)) 5613 return 0; 5614 5615 return 1; 5616} 5617 5618static int 5619sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent) 5620{ 5621 unsigned long cflags = sd->flags, pflags = parent->flags; 5622 5623 if (sd_degenerate(parent)) 5624 return 1; 5625 5626 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent))) 5627 return 0; 5628 5629 /* Flags needing groups don't count if only 1 group in parent */ 5630 if (parent->groups == parent->groups->next) { 5631 pflags &= ~(SD_LOAD_BALANCE | 5632 SD_BALANCE_NEWIDLE | 5633 SD_BALANCE_FORK | 5634 SD_BALANCE_EXEC | 5635 SD_SHARE_CPUPOWER | 5636 SD_SHARE_PKG_RESOURCES); 5637 if (nr_node_ids == 1) 5638 pflags &= ~SD_SERIALIZE; 5639 } 5640 if (~cflags & pflags) 5641 return 0; 5642 5643 return 1; 5644} 5645 5646static void free_rootdomain(struct rcu_head *rcu) 5647{ 5648 struct root_domain *rd = container_of(rcu, struct root_domain, rcu); 5649 5650 cpupri_cleanup(&rd->cpupri); 5651 free_cpumask_var(rd->rto_mask); 5652 free_cpumask_var(rd->online); 5653 free_cpumask_var(rd->span); 5654 kfree(rd); 5655} 5656 5657static void rq_attach_root(struct rq *rq, struct root_domain *rd) 5658{ 5659 struct root_domain *old_rd = NULL; 5660 unsigned long flags; 5661 5662 raw_spin_lock_irqsave(&rq->lock, flags); 5663 5664 if (rq->rd) { 5665 old_rd = rq->rd; 5666 5667 if (cpumask_test_cpu(rq->cpu, old_rd->online)) 5668 set_rq_offline(rq); 5669 5670 cpumask_clear_cpu(rq->cpu, old_rd->span); 5671 5672 /* 5673 * If we dont want to free the old_rt yet then 5674 * set old_rd to NULL to skip the freeing later 5675 * in this function: 5676 */ 5677 if (!atomic_dec_and_test(&old_rd->refcount)) 5678 old_rd = NULL; 5679 } 5680 5681 atomic_inc(&rd->refcount); 5682 rq->rd = rd; 5683 5684 cpumask_set_cpu(rq->cpu, rd->span); 5685 if (cpumask_test_cpu(rq->cpu, cpu_active_mask)) 5686 set_rq_online(rq); 5687 5688 raw_spin_unlock_irqrestore(&rq->lock, flags); 5689 5690 if (old_rd) 5691 call_rcu_sched(&old_rd->rcu, free_rootdomain); 5692} 5693 5694static int init_rootdomain(struct root_domain *rd) 5695{ 5696 memset(rd, 0, sizeof(*rd)); 5697 5698 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL)) 5699 goto out; 5700 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL)) 5701 goto free_span; 5702 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL)) 5703 goto free_online; 5704 5705 if (cpupri_init(&rd->cpupri) != 0) 5706 goto free_rto_mask; 5707 return 0; 5708 5709free_rto_mask: 5710 free_cpumask_var(rd->rto_mask); 5711free_online: 5712 free_cpumask_var(rd->online); 5713free_span: 5714 free_cpumask_var(rd->span); 5715out: 5716 return -ENOMEM; 5717} 5718 5719/* 5720 * By default the system creates a single root-domain with all cpus as 5721 * members (mimicking the global state we have today). 5722 */ 5723struct root_domain def_root_domain; 5724 5725static void init_defrootdomain(void) 5726{ 5727 init_rootdomain(&def_root_domain); 5728 5729 atomic_set(&def_root_domain.refcount, 1); 5730} 5731 5732static struct root_domain *alloc_rootdomain(void) 5733{ 5734 struct root_domain *rd; 5735 5736 rd = kmalloc(sizeof(*rd), GFP_KERNEL); 5737 if (!rd) 5738 return NULL; 5739 5740 if (init_rootdomain(rd) != 0) { 5741 kfree(rd); 5742 return NULL; 5743 } 5744 5745 return rd; 5746} 5747 5748static void free_sched_groups(struct sched_group *sg, int free_sgp) 5749{ 5750 struct sched_group *tmp, *first; 5751 5752 if (!sg) 5753 return; 5754 5755 first = sg; 5756 do { 5757 tmp = sg->next; 5758 5759 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref)) 5760 kfree(sg->sgp); 5761 5762 kfree(sg); 5763 sg = tmp; 5764 } while (sg != first); 5765} 5766 5767static void free_sched_domain(struct rcu_head *rcu) 5768{ 5769 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu); 5770 5771 /* 5772 * If its an overlapping domain it has private groups, iterate and 5773 * nuke them all. 5774 */ 5775 if (sd->flags & SD_OVERLAP) { 5776 free_sched_groups(sd->groups, 1); 5777 } else if (atomic_dec_and_test(&sd->groups->ref)) { 5778 kfree(sd->groups->sgp); 5779 kfree(sd->groups); 5780 } 5781 kfree(sd); 5782} 5783 5784static void destroy_sched_domain(struct sched_domain *sd, int cpu) 5785{ 5786 call_rcu(&sd->rcu, free_sched_domain); 5787} 5788 5789static void destroy_sched_domains(struct sched_domain *sd, int cpu) 5790{ 5791 for (; sd; sd = sd->parent) 5792 destroy_sched_domain(sd, cpu); 5793} 5794 5795/* 5796 * Keep a special pointer to the highest sched_domain that has 5797 * SD_SHARE_PKG_RESOURCE set (Last Level Cache Domain) for this 5798 * allows us to avoid some pointer chasing select_idle_sibling(). 5799 * 5800 * Also keep a unique ID per domain (we use the first cpu number in 5801 * the cpumask of the domain), this allows us to quickly tell if 5802 * two cpus are in the same cache domain, see cpus_share_cache(). 5803 */ 5804DEFINE_PER_CPU(struct sched_domain *, sd_llc); 5805DEFINE_PER_CPU(int, sd_llc_id); 5806 5807static void update_top_cache_domain(int cpu) 5808{ 5809 struct sched_domain *sd; 5810 int id = cpu; 5811 5812 sd = highest_flag_domain(cpu, SD_SHARE_PKG_RESOURCES); 5813 if (sd) 5814 id = cpumask_first(sched_domain_span(sd)); 5815 5816 rcu_assign_pointer(per_cpu(sd_llc, cpu), sd); 5817 per_cpu(sd_llc_id, cpu) = id; 5818} 5819 5820/* 5821 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must 5822 * hold the hotplug lock. 5823 */ 5824static void 5825cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu) 5826{ 5827 struct rq *rq = cpu_rq(cpu); 5828 struct sched_domain *tmp; 5829 5830 /* Remove the sched domains which do not contribute to scheduling. */ 5831 for (tmp = sd; tmp; ) { 5832 struct sched_domain *parent = tmp->parent; 5833 if (!parent) 5834 break; 5835 5836 if (sd_parent_degenerate(tmp, parent)) { 5837 tmp->parent = parent->parent; 5838 if (parent->parent) 5839 parent->parent->child = tmp; 5840 destroy_sched_domain(parent, cpu); 5841 } else 5842 tmp = tmp->parent; 5843 } 5844 5845 if (sd && sd_degenerate(sd)) { 5846 tmp = sd; 5847 sd = sd->parent; 5848 destroy_sched_domain(tmp, cpu); 5849 if (sd) 5850 sd->child = NULL; 5851 } 5852 5853 sched_domain_debug(sd, cpu); 5854 5855 rq_attach_root(rq, rd); 5856 tmp = rq->sd; 5857 rcu_assign_pointer(rq->sd, sd); 5858 destroy_sched_domains(tmp, cpu); 5859 5860 update_top_cache_domain(cpu); 5861} 5862 5863/* cpus with isolated domains */ 5864static cpumask_var_t cpu_isolated_map; 5865 5866/* Setup the mask of cpus configured for isolated domains */ 5867static int __init isolated_cpu_setup(char *str) 5868{ 5869 alloc_bootmem_cpumask_var(&cpu_isolated_map); 5870 cpulist_parse(str, cpu_isolated_map); 5871 return 1; 5872} 5873 5874__setup("isolcpus=", isolated_cpu_setup); 5875 5876#ifdef CONFIG_NUMA 5877 5878/** 5879 * find_next_best_node - find the next node to include in a sched_domain 5880 * @node: node whose sched_domain we're building 5881 * @used_nodes: nodes already in the sched_domain 5882 * 5883 * Find the next node to include in a given scheduling domain. Simply 5884 * finds the closest node not already in the @used_nodes map. 5885 * 5886 * Should use nodemask_t. 5887 */ 5888static int find_next_best_node(int node, nodemask_t *used_nodes) 5889{ 5890 int i, n, val, min_val, best_node = -1; 5891 5892 min_val = INT_MAX; 5893 5894 for (i = 0; i < nr_node_ids; i++) { 5895 /* Start at @node */ 5896 n = (node + i) % nr_node_ids; 5897 5898 if (!nr_cpus_node(n)) 5899 continue; 5900 5901 /* Skip already used nodes */ 5902 if (node_isset(n, *used_nodes)) 5903 continue; 5904 5905 /* Simple min distance search */ 5906 val = node_distance(node, n); 5907 5908 if (val < min_val) { 5909 min_val = val; 5910 best_node = n; 5911 } 5912 } 5913 5914 if (best_node != -1) 5915 node_set(best_node, *used_nodes); 5916 return best_node; 5917} 5918 5919/** 5920 * sched_domain_node_span - get a cpumask for a node's sched_domain 5921 * @node: node whose cpumask we're constructing 5922 * @span: resulting cpumask 5923 * 5924 * Given a node, construct a good cpumask for its sched_domain to span. It 5925 * should be one that prevents unnecessary balancing, but also spreads tasks 5926 * out optimally. 5927 */ 5928static void sched_domain_node_span(int node, struct cpumask *span) 5929{ 5930 nodemask_t used_nodes; 5931 int i; 5932 5933 cpumask_clear(span); 5934 nodes_clear(used_nodes); 5935 5936 cpumask_or(span, span, cpumask_of_node(node)); 5937 node_set(node, used_nodes); 5938 5939 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) { 5940 int next_node = find_next_best_node(node, &used_nodes); 5941 if (next_node < 0) 5942 break; 5943 cpumask_or(span, span, cpumask_of_node(next_node)); 5944 } 5945} 5946 5947static const struct cpumask *cpu_node_mask(int cpu) 5948{ 5949 lockdep_assert_held(&sched_domains_mutex); 5950 5951 sched_domain_node_span(cpu_to_node(cpu), sched_domains_tmpmask); 5952 5953 return sched_domains_tmpmask; 5954} 5955 5956static const struct cpumask *cpu_allnodes_mask(int cpu) 5957{ 5958 return cpu_possible_mask; 5959} 5960#endif /* CONFIG_NUMA */ 5961 5962static const struct cpumask *cpu_cpu_mask(int cpu) 5963{ 5964 return cpumask_of_node(cpu_to_node(cpu)); 5965} 5966 5967int sched_smt_power_savings = 0, sched_mc_power_savings = 0; 5968 5969struct sd_data { 5970 struct sched_domain **__percpu sd; 5971 struct sched_group **__percpu sg; 5972 struct sched_group_power **__percpu sgp; 5973}; 5974 5975struct s_data { 5976 struct sched_domain ** __percpu sd; 5977 struct root_domain *rd; 5978}; 5979 5980enum s_alloc { 5981 sa_rootdomain, 5982 sa_sd, 5983 sa_sd_storage, 5984 sa_none, 5985}; 5986 5987struct sched_domain_topology_level; 5988 5989typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu); 5990typedef const struct cpumask *(*sched_domain_mask_f)(int cpu); 5991 5992#define SDTL_OVERLAP 0x01 5993 5994struct sched_domain_topology_level { 5995 sched_domain_init_f init; 5996 sched_domain_mask_f mask; 5997 int flags; 5998 struct sd_data data; 5999}; 6000 6001static int 6002build_overlap_sched_groups(struct sched_domain *sd, int cpu) 6003{ 6004 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg; 6005 const struct cpumask *span = sched_domain_span(sd); 6006 struct cpumask *covered = sched_domains_tmpmask; 6007 struct sd_data *sdd = sd->private; 6008 struct sched_domain *child; 6009 int i; 6010 6011 cpumask_clear(covered); 6012 6013 for_each_cpu(i, span) { 6014 struct cpumask *sg_span; 6015 6016 if (cpumask_test_cpu(i, covered)) 6017 continue; 6018 6019 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 6020 GFP_KERNEL, cpu_to_node(cpu)); 6021 6022 if (!sg) 6023 goto fail; 6024 6025 sg_span = sched_group_cpus(sg); 6026 6027 child = *per_cpu_ptr(sdd->sd, i); 6028 if (child->child) { 6029 child = child->child; 6030 cpumask_copy(sg_span, sched_domain_span(child)); 6031 } else 6032 cpumask_set_cpu(i, sg_span); 6033 6034 cpumask_or(covered, covered, sg_span); 6035 6036 sg->sgp = *per_cpu_ptr(sdd->sgp, cpumask_first(sg_span)); 6037 atomic_inc(&sg->sgp->ref); 6038 6039 if (cpumask_test_cpu(cpu, sg_span)) 6040 groups = sg; 6041 6042 if (!first) 6043 first = sg; 6044 if (last) 6045 last->next = sg; 6046 last = sg; 6047 last->next = first; 6048 } 6049 sd->groups = groups; 6050 6051 return 0; 6052 6053fail: 6054 free_sched_groups(first, 0); 6055 6056 return -ENOMEM; 6057} 6058 6059static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg) 6060{ 6061 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu); 6062 struct sched_domain *child = sd->child; 6063 6064 if (child) 6065 cpu = cpumask_first(sched_domain_span(child)); 6066 6067 if (sg) { 6068 *sg = *per_cpu_ptr(sdd->sg, cpu); 6069 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu); 6070 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */ 6071 } 6072 6073 return cpu; 6074} 6075 6076/* 6077 * build_sched_groups will build a circular linked list of the groups 6078 * covered by the given span, and will set each group's ->cpumask correctly, 6079 * and ->cpu_power to 0. 6080 * 6081 * Assumes the sched_domain tree is fully constructed 6082 */ 6083static int 6084build_sched_groups(struct sched_domain *sd, int cpu) 6085{ 6086 struct sched_group *first = NULL, *last = NULL; 6087 struct sd_data *sdd = sd->private; 6088 const struct cpumask *span = sched_domain_span(sd); 6089 struct cpumask *covered; 6090 int i; 6091 6092 get_group(cpu, sdd, &sd->groups); 6093 atomic_inc(&sd->groups->ref); 6094 6095 if (cpu != cpumask_first(sched_domain_span(sd))) 6096 return 0; 6097 6098 lockdep_assert_held(&sched_domains_mutex); 6099 covered = sched_domains_tmpmask; 6100 6101 cpumask_clear(covered); 6102 6103 for_each_cpu(i, span) { 6104 struct sched_group *sg; 6105 int group = get_group(i, sdd, &sg); 6106 int j; 6107 6108 if (cpumask_test_cpu(i, covered)) 6109 continue; 6110 6111 cpumask_clear(sched_group_cpus(sg)); 6112 sg->sgp->power = 0; 6113 6114 for_each_cpu(j, span) { 6115 if (get_group(j, sdd, NULL) != group) 6116 continue; 6117 6118 cpumask_set_cpu(j, covered); 6119 cpumask_set_cpu(j, sched_group_cpus(sg)); 6120 } 6121 6122 if (!first) 6123 first = sg; 6124 if (last) 6125 last->next = sg; 6126 last = sg; 6127 } 6128 last->next = first; 6129 6130 return 0; 6131} 6132 6133/* 6134 * Initialize sched groups cpu_power. 6135 * 6136 * cpu_power indicates the capacity of sched group, which is used while 6137 * distributing the load between different sched groups in a sched domain. 6138 * Typically cpu_power for all the groups in a sched domain will be same unless 6139 * there are asymmetries in the topology. If there are asymmetries, group 6140 * having more cpu_power will pickup more load compared to the group having 6141 * less cpu_power. 6142 */ 6143static void init_sched_groups_power(int cpu, struct sched_domain *sd) 6144{ 6145 struct sched_group *sg = sd->groups; 6146 6147 WARN_ON(!sd || !sg); 6148 6149 do { 6150 sg->group_weight = cpumask_weight(sched_group_cpus(sg)); 6151 sg = sg->next; 6152 } while (sg != sd->groups); 6153 6154 if (cpu != group_first_cpu(sg)) 6155 return; 6156 6157 update_group_power(sd, cpu); 6158 atomic_set(&sg->sgp->nr_busy_cpus, sg->group_weight); 6159} 6160 6161int __weak arch_sd_sibling_asym_packing(void) 6162{ 6163 return 0*SD_ASYM_PACKING; 6164} 6165 6166/* 6167 * Initializers for schedule domains 6168 * Non-inlined to reduce accumulated stack pressure in build_sched_domains() 6169 */ 6170 6171#ifdef CONFIG_SCHED_DEBUG 6172# define SD_INIT_NAME(sd, type) sd->name = #type 6173#else 6174# define SD_INIT_NAME(sd, type) do { } while (0) 6175#endif 6176 6177#define SD_INIT_FUNC(type) \ 6178static noinline struct sched_domain * \ 6179sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \ 6180{ \ 6181 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \ 6182 *sd = SD_##type##_INIT; \ 6183 SD_INIT_NAME(sd, type); \ 6184 sd->private = &tl->data; \ 6185 return sd; \ 6186} 6187 6188SD_INIT_FUNC(CPU) 6189#ifdef CONFIG_NUMA 6190 SD_INIT_FUNC(ALLNODES) 6191 SD_INIT_FUNC(NODE) 6192#endif 6193#ifdef CONFIG_SCHED_SMT 6194 SD_INIT_FUNC(SIBLING) 6195#endif 6196#ifdef CONFIG_SCHED_MC 6197 SD_INIT_FUNC(MC) 6198#endif 6199#ifdef CONFIG_SCHED_BOOK 6200 SD_INIT_FUNC(BOOK) 6201#endif 6202 6203static int default_relax_domain_level = -1; 6204int sched_domain_level_max; 6205 6206static int __init setup_relax_domain_level(char *str) 6207{ 6208 unsigned long val; 6209 6210 val = simple_strtoul(str, NULL, 0); 6211 if (val < sched_domain_level_max) 6212 default_relax_domain_level = val; 6213 6214 return 1; 6215} 6216__setup("relax_domain_level=", setup_relax_domain_level); 6217 6218static void set_domain_attribute(struct sched_domain *sd, 6219 struct sched_domain_attr *attr) 6220{ 6221 int request; 6222 6223 if (!attr || attr->relax_domain_level < 0) { 6224 if (default_relax_domain_level < 0) 6225 return; 6226 else 6227 request = default_relax_domain_level; 6228 } else 6229 request = attr->relax_domain_level; 6230 if (request < sd->level) { 6231 /* turn off idle balance on this domain */ 6232 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 6233 } else { 6234 /* turn on idle balance on this domain */ 6235 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE); 6236 } 6237} 6238 6239static void __sdt_free(const struct cpumask *cpu_map); 6240static int __sdt_alloc(const struct cpumask *cpu_map); 6241 6242static void __free_domain_allocs(struct s_data *d, enum s_alloc what, 6243 const struct cpumask *cpu_map) 6244{ 6245 switch (what) { 6246 case sa_rootdomain: 6247 if (!atomic_read(&d->rd->refcount)) 6248 free_rootdomain(&d->rd->rcu); /* fall through */ 6249 case sa_sd: 6250 free_percpu(d->sd); /* fall through */ 6251 case sa_sd_storage: 6252 __sdt_free(cpu_map); /* fall through */ 6253 case sa_none: 6254 break; 6255 } 6256} 6257 6258static enum s_alloc __visit_domain_allocation_hell(struct s_data *d, 6259 const struct cpumask *cpu_map) 6260{ 6261 memset(d, 0, sizeof(*d)); 6262 6263 if (__sdt_alloc(cpu_map)) 6264 return sa_sd_storage; 6265 d->sd = alloc_percpu(struct sched_domain *); 6266 if (!d->sd) 6267 return sa_sd_storage; 6268 d->rd = alloc_rootdomain(); 6269 if (!d->rd) 6270 return sa_sd; 6271 return sa_rootdomain; 6272} 6273 6274/* 6275 * NULL the sd_data elements we've used to build the sched_domain and 6276 * sched_group structure so that the subsequent __free_domain_allocs() 6277 * will not free the data we're using. 6278 */ 6279static void claim_allocations(int cpu, struct sched_domain *sd) 6280{ 6281 struct sd_data *sdd = sd->private; 6282 6283 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd); 6284 *per_cpu_ptr(sdd->sd, cpu) = NULL; 6285 6286 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref)) 6287 *per_cpu_ptr(sdd->sg, cpu) = NULL; 6288 6289 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref)) 6290 *per_cpu_ptr(sdd->sgp, cpu) = NULL; 6291} 6292 6293#ifdef CONFIG_SCHED_SMT 6294static const struct cpumask *cpu_smt_mask(int cpu) 6295{ 6296 return topology_thread_cpumask(cpu); 6297} 6298#endif 6299 6300/* 6301 * Topology list, bottom-up. 6302 */ 6303static struct sched_domain_topology_level default_topology[] = { 6304#ifdef CONFIG_SCHED_SMT 6305 { sd_init_SIBLING, cpu_smt_mask, }, 6306#endif 6307#ifdef CONFIG_SCHED_MC 6308 { sd_init_MC, cpu_coregroup_mask, }, 6309#endif 6310#ifdef CONFIG_SCHED_BOOK 6311 { sd_init_BOOK, cpu_book_mask, }, 6312#endif 6313 { sd_init_CPU, cpu_cpu_mask, }, 6314#ifdef CONFIG_NUMA 6315 { sd_init_NODE, cpu_node_mask, SDTL_OVERLAP, }, 6316 { sd_init_ALLNODES, cpu_allnodes_mask, }, 6317#endif 6318 { NULL, }, 6319}; 6320 6321static struct sched_domain_topology_level *sched_domain_topology = default_topology; 6322 6323static int __sdt_alloc(const struct cpumask *cpu_map) 6324{ 6325 struct sched_domain_topology_level *tl; 6326 int j; 6327 6328 for (tl = sched_domain_topology; tl->init; tl++) { 6329 struct sd_data *sdd = &tl->data; 6330 6331 sdd->sd = alloc_percpu(struct sched_domain *); 6332 if (!sdd->sd) 6333 return -ENOMEM; 6334 6335 sdd->sg = alloc_percpu(struct sched_group *); 6336 if (!sdd->sg) 6337 return -ENOMEM; 6338 6339 sdd->sgp = alloc_percpu(struct sched_group_power *); 6340 if (!sdd->sgp) 6341 return -ENOMEM; 6342 6343 for_each_cpu(j, cpu_map) { 6344 struct sched_domain *sd; 6345 struct sched_group *sg; 6346 struct sched_group_power *sgp; 6347 6348 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(), 6349 GFP_KERNEL, cpu_to_node(j)); 6350 if (!sd) 6351 return -ENOMEM; 6352 6353 *per_cpu_ptr(sdd->sd, j) = sd; 6354 6355 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(), 6356 GFP_KERNEL, cpu_to_node(j)); 6357 if (!sg) 6358 return -ENOMEM; 6359 6360 *per_cpu_ptr(sdd->sg, j) = sg; 6361 6362 sgp = kzalloc_node(sizeof(struct sched_group_power), 6363 GFP_KERNEL, cpu_to_node(j)); 6364 if (!sgp) 6365 return -ENOMEM; 6366 6367 *per_cpu_ptr(sdd->sgp, j) = sgp; 6368 } 6369 } 6370 6371 return 0; 6372} 6373 6374static void __sdt_free(const struct cpumask *cpu_map) 6375{ 6376 struct sched_domain_topology_level *tl; 6377 int j; 6378 6379 for (tl = sched_domain_topology; tl->init; tl++) { 6380 struct sd_data *sdd = &tl->data; 6381 6382 for_each_cpu(j, cpu_map) { 6383 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, j); 6384 if (sd && (sd->flags & SD_OVERLAP)) 6385 free_sched_groups(sd->groups, 0); 6386 kfree(*per_cpu_ptr(sdd->sd, j)); 6387 kfree(*per_cpu_ptr(sdd->sg, j)); 6388 kfree(*per_cpu_ptr(sdd->sgp, j)); 6389 } 6390 free_percpu(sdd->sd); 6391 free_percpu(sdd->sg); 6392 free_percpu(sdd->sgp); 6393 } 6394} 6395 6396struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl, 6397 struct s_data *d, const struct cpumask *cpu_map, 6398 struct sched_domain_attr *attr, struct sched_domain *child, 6399 int cpu) 6400{ 6401 struct sched_domain *sd = tl->init(tl, cpu); 6402 if (!sd) 6403 return child; 6404 6405 set_domain_attribute(sd, attr); 6406 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu)); 6407 if (child) { 6408 sd->level = child->level + 1; 6409 sched_domain_level_max = max(sched_domain_level_max, sd->level); 6410 child->parent = sd; 6411 } 6412 sd->child = child; 6413 6414 return sd; 6415} 6416 6417/* 6418 * Build sched domains for a given set of cpus and attach the sched domains 6419 * to the individual cpus 6420 */ 6421static int build_sched_domains(const struct cpumask *cpu_map, 6422 struct sched_domain_attr *attr) 6423{ 6424 enum s_alloc alloc_state = sa_none; 6425 struct sched_domain *sd; 6426 struct s_data d; 6427 int i, ret = -ENOMEM; 6428 6429 alloc_state = __visit_domain_allocation_hell(&d, cpu_map); 6430 if (alloc_state != sa_rootdomain) 6431 goto error; 6432 6433 /* Set up domains for cpus specified by the cpu_map. */ 6434 for_each_cpu(i, cpu_map) { 6435 struct sched_domain_topology_level *tl; 6436 6437 sd = NULL; 6438 for (tl = sched_domain_topology; tl->init; tl++) { 6439 sd = build_sched_domain(tl, &d, cpu_map, attr, sd, i); 6440 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP)) 6441 sd->flags |= SD_OVERLAP; 6442 if (cpumask_equal(cpu_map, sched_domain_span(sd))) 6443 break; 6444 } 6445 6446 while (sd->child) 6447 sd = sd->child; 6448 6449 *per_cpu_ptr(d.sd, i) = sd; 6450 } 6451 6452 /* Build the groups for the domains */ 6453 for_each_cpu(i, cpu_map) { 6454 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 6455 sd->span_weight = cpumask_weight(sched_domain_span(sd)); 6456 if (sd->flags & SD_OVERLAP) { 6457 if (build_overlap_sched_groups(sd, i)) 6458 goto error; 6459 } else { 6460 if (build_sched_groups(sd, i)) 6461 goto error; 6462 } 6463 } 6464 } 6465 6466 /* Calculate CPU power for physical packages and nodes */ 6467 for (i = nr_cpumask_bits-1; i >= 0; i--) { 6468 if (!cpumask_test_cpu(i, cpu_map)) 6469 continue; 6470 6471 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) { 6472 claim_allocations(i, sd); 6473 init_sched_groups_power(i, sd); 6474 } 6475 } 6476 6477 /* Attach the domains */ 6478 rcu_read_lock(); 6479 for_each_cpu(i, cpu_map) { 6480 sd = *per_cpu_ptr(d.sd, i); 6481 cpu_attach_domain(sd, d.rd, i); 6482 } 6483 rcu_read_unlock(); 6484 6485 ret = 0; 6486error: 6487 __free_domain_allocs(&d, alloc_state, cpu_map); 6488 return ret; 6489} 6490 6491static cpumask_var_t *doms_cur; /* current sched domains */ 6492static int ndoms_cur; /* number of sched domains in 'doms_cur' */ 6493static struct sched_domain_attr *dattr_cur; 6494 /* attribues of custom domains in 'doms_cur' */ 6495 6496/* 6497 * Special case: If a kmalloc of a doms_cur partition (array of 6498 * cpumask) fails, then fallback to a single sched domain, 6499 * as determined by the single cpumask fallback_doms. 6500 */ 6501static cpumask_var_t fallback_doms; 6502 6503/* 6504 * arch_update_cpu_topology lets virtualized architectures update the 6505 * cpu core maps. It is supposed to return 1 if the topology changed 6506 * or 0 if it stayed the same. 6507 */ 6508int __attribute__((weak)) arch_update_cpu_topology(void) 6509{ 6510 return 0; 6511} 6512 6513cpumask_var_t *alloc_sched_domains(unsigned int ndoms) 6514{ 6515 int i; 6516 cpumask_var_t *doms; 6517 6518 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL); 6519 if (!doms) 6520 return NULL; 6521 for (i = 0; i < ndoms; i++) { 6522 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) { 6523 free_sched_domains(doms, i); 6524 return NULL; 6525 } 6526 } 6527 return doms; 6528} 6529 6530void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms) 6531{ 6532 unsigned int i; 6533 for (i = 0; i < ndoms; i++) 6534 free_cpumask_var(doms[i]); 6535 kfree(doms); 6536} 6537 6538/* 6539 * Set up scheduler domains and groups. Callers must hold the hotplug lock. 6540 * For now this just excludes isolated cpus, but could be used to 6541 * exclude other special cases in the future. 6542 */ 6543static int init_sched_domains(const struct cpumask *cpu_map) 6544{ 6545 int err; 6546 6547 arch_update_cpu_topology(); 6548 ndoms_cur = 1; 6549 doms_cur = alloc_sched_domains(ndoms_cur); 6550 if (!doms_cur) 6551 doms_cur = &fallback_doms; 6552 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map); 6553 dattr_cur = NULL; 6554 err = build_sched_domains(doms_cur[0], NULL); 6555 register_sched_domain_sysctl(); 6556 6557 return err; 6558} 6559 6560/* 6561 * Detach sched domains from a group of cpus specified in cpu_map 6562 * These cpus will now be attached to the NULL domain 6563 */ 6564static void detach_destroy_domains(const struct cpumask *cpu_map) 6565{ 6566 int i; 6567 6568 rcu_read_lock(); 6569 for_each_cpu(i, cpu_map) 6570 cpu_attach_domain(NULL, &def_root_domain, i); 6571 rcu_read_unlock(); 6572} 6573 6574/* handle null as "default" */ 6575static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur, 6576 struct sched_domain_attr *new, int idx_new) 6577{ 6578 struct sched_domain_attr tmp; 6579 6580 /* fast path */ 6581 if (!new && !cur) 6582 return 1; 6583 6584 tmp = SD_ATTR_INIT; 6585 return !memcmp(cur ? (cur + idx_cur) : &tmp, 6586 new ? (new + idx_new) : &tmp, 6587 sizeof(struct sched_domain_attr)); 6588} 6589 6590/* 6591 * Partition sched domains as specified by the 'ndoms_new' 6592 * cpumasks in the array doms_new[] of cpumasks. This compares 6593 * doms_new[] to the current sched domain partitioning, doms_cur[]. 6594 * It destroys each deleted domain and builds each new domain. 6595 * 6596 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'. 6597 * The masks don't intersect (don't overlap.) We should setup one 6598 * sched domain for each mask. CPUs not in any of the cpumasks will 6599 * not be load balanced. If the same cpumask appears both in the 6600 * current 'doms_cur' domains and in the new 'doms_new', we can leave 6601 * it as it is. 6602 * 6603 * The passed in 'doms_new' should be allocated using 6604 * alloc_sched_domains. This routine takes ownership of it and will 6605 * free_sched_domains it when done with it. If the caller failed the 6606 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1, 6607 * and partition_sched_domains() will fallback to the single partition 6608 * 'fallback_doms', it also forces the domains to be rebuilt. 6609 * 6610 * If doms_new == NULL it will be replaced with cpu_online_mask. 6611 * ndoms_new == 0 is a special case for destroying existing domains, 6612 * and it will not create the default domain. 6613 * 6614 * Call with hotplug lock held 6615 */ 6616void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[], 6617 struct sched_domain_attr *dattr_new) 6618{ 6619 int i, j, n; 6620 int new_topology; 6621 6622 mutex_lock(&sched_domains_mutex); 6623 6624 /* always unregister in case we don't destroy any domains */ 6625 unregister_sched_domain_sysctl(); 6626 6627 /* Let architecture update cpu core mappings. */ 6628 new_topology = arch_update_cpu_topology(); 6629 6630 n = doms_new ? ndoms_new : 0; 6631 6632 /* Destroy deleted domains */ 6633 for (i = 0; i < ndoms_cur; i++) { 6634 for (j = 0; j < n && !new_topology; j++) { 6635 if (cpumask_equal(doms_cur[i], doms_new[j]) 6636 && dattrs_equal(dattr_cur, i, dattr_new, j)) 6637 goto match1; 6638 } 6639 /* no match - a current sched domain not in new doms_new[] */ 6640 detach_destroy_domains(doms_cur[i]); 6641match1: 6642 ; 6643 } 6644 6645 if (doms_new == NULL) { 6646 ndoms_cur = 0; 6647 doms_new = &fallback_doms; 6648 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map); 6649 WARN_ON_ONCE(dattr_new); 6650 } 6651 6652 /* Build new domains */ 6653 for (i = 0; i < ndoms_new; i++) { 6654 for (j = 0; j < ndoms_cur && !new_topology; j++) { 6655 if (cpumask_equal(doms_new[i], doms_cur[j]) 6656 && dattrs_equal(dattr_new, i, dattr_cur, j)) 6657 goto match2; 6658 } 6659 /* no match - add a new doms_new */ 6660 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL); 6661match2: 6662 ; 6663 } 6664 6665 /* Remember the new sched domains */ 6666 if (doms_cur != &fallback_doms) 6667 free_sched_domains(doms_cur, ndoms_cur); 6668 kfree(dattr_cur); /* kfree(NULL) is safe */ 6669 doms_cur = doms_new; 6670 dattr_cur = dattr_new; 6671 ndoms_cur = ndoms_new; 6672 6673 register_sched_domain_sysctl(); 6674 6675 mutex_unlock(&sched_domains_mutex); 6676} 6677 6678#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT) 6679static void reinit_sched_domains(void) 6680{ 6681 get_online_cpus(); 6682 6683 /* Destroy domains first to force the rebuild */ 6684 partition_sched_domains(0, NULL, NULL); 6685 6686 rebuild_sched_domains(); 6687 put_online_cpus(); 6688} 6689 6690static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt) 6691{ 6692 unsigned int level = 0; 6693 6694 if (sscanf(buf, "%u", &level) != 1) 6695 return -EINVAL; 6696 6697 /* 6698 * level is always be positive so don't check for 6699 * level < POWERSAVINGS_BALANCE_NONE which is 0 6700 * What happens on 0 or 1 byte write, 6701 * need to check for count as well? 6702 */ 6703 6704 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS) 6705 return -EINVAL; 6706 6707 if (smt) 6708 sched_smt_power_savings = level; 6709 else 6710 sched_mc_power_savings = level; 6711 6712 reinit_sched_domains(); 6713 6714 return count; 6715} 6716 6717#ifdef CONFIG_SCHED_MC 6718static ssize_t sched_mc_power_savings_show(struct device *dev, 6719 struct device_attribute *attr, 6720 char *buf) 6721{ 6722 return sprintf(buf, "%u\n", sched_mc_power_savings); 6723} 6724static ssize_t sched_mc_power_savings_store(struct device *dev, 6725 struct device_attribute *attr, 6726 const char *buf, size_t count) 6727{ 6728 return sched_power_savings_store(buf, count, 0); 6729} 6730static DEVICE_ATTR(sched_mc_power_savings, 0644, 6731 sched_mc_power_savings_show, 6732 sched_mc_power_savings_store); 6733#endif 6734 6735#ifdef CONFIG_SCHED_SMT 6736static ssize_t sched_smt_power_savings_show(struct device *dev, 6737 struct device_attribute *attr, 6738 char *buf) 6739{ 6740 return sprintf(buf, "%u\n", sched_smt_power_savings); 6741} 6742static ssize_t sched_smt_power_savings_store(struct device *dev, 6743 struct device_attribute *attr, 6744 const char *buf, size_t count) 6745{ 6746 return sched_power_savings_store(buf, count, 1); 6747} 6748static DEVICE_ATTR(sched_smt_power_savings, 0644, 6749 sched_smt_power_savings_show, 6750 sched_smt_power_savings_store); 6751#endif 6752 6753int __init sched_create_sysfs_power_savings_entries(struct device *dev) 6754{ 6755 int err = 0; 6756 6757#ifdef CONFIG_SCHED_SMT 6758 if (smt_capable()) 6759 err = device_create_file(dev, &dev_attr_sched_smt_power_savings); 6760#endif 6761#ifdef CONFIG_SCHED_MC 6762 if (!err && mc_capable()) 6763 err = device_create_file(dev, &dev_attr_sched_mc_power_savings); 6764#endif 6765 return err; 6766} 6767#endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */ 6768 6769/* 6770 * Update cpusets according to cpu_active mask. If cpusets are 6771 * disabled, cpuset_update_active_cpus() becomes a simple wrapper 6772 * around partition_sched_domains(). 6773 */ 6774static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action, 6775 void *hcpu) 6776{ 6777 switch (action & ~CPU_TASKS_FROZEN) { 6778 case CPU_ONLINE: 6779 case CPU_DOWN_FAILED: 6780 cpuset_update_active_cpus(); 6781 return NOTIFY_OK; 6782 default: 6783 return NOTIFY_DONE; 6784 } 6785} 6786 6787static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action, 6788 void *hcpu) 6789{ 6790 switch (action & ~CPU_TASKS_FROZEN) { 6791 case CPU_DOWN_PREPARE: 6792 cpuset_update_active_cpus(); 6793 return NOTIFY_OK; 6794 default: 6795 return NOTIFY_DONE; 6796 } 6797} 6798 6799void __init sched_init_smp(void) 6800{ 6801 cpumask_var_t non_isolated_cpus; 6802 6803 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL); 6804 alloc_cpumask_var(&fallback_doms, GFP_KERNEL); 6805 6806 get_online_cpus(); 6807 mutex_lock(&sched_domains_mutex); 6808 init_sched_domains(cpu_active_mask); 6809 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map); 6810 if (cpumask_empty(non_isolated_cpus)) 6811 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus); 6812 mutex_unlock(&sched_domains_mutex); 6813 put_online_cpus(); 6814 6815 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE); 6816 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE); 6817 6818 /* RT runtime code needs to handle some hotplug events */ 6819 hotcpu_notifier(update_runtime, 0); 6820 6821 init_hrtick(); 6822 6823 /* Move init over to a non-isolated CPU */ 6824 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0) 6825 BUG(); 6826 sched_init_granularity(); 6827 free_cpumask_var(non_isolated_cpus); 6828 6829 init_sched_rt_class(); 6830} 6831#else 6832void __init sched_init_smp(void) 6833{ 6834 sched_init_granularity(); 6835} 6836#endif /* CONFIG_SMP */ 6837 6838const_debug unsigned int sysctl_timer_migration = 1; 6839 6840int in_sched_functions(unsigned long addr) 6841{ 6842 return in_lock_functions(addr) || 6843 (addr >= (unsigned long)__sched_text_start 6844 && addr < (unsigned long)__sched_text_end); 6845} 6846 6847#ifdef CONFIG_CGROUP_SCHED 6848struct task_group root_task_group; 6849#endif 6850 6851DECLARE_PER_CPU(cpumask_var_t, load_balance_tmpmask); 6852 6853void __init sched_init(void) 6854{ 6855 int i, j; 6856 unsigned long alloc_size = 0, ptr; 6857 6858#ifdef CONFIG_FAIR_GROUP_SCHED 6859 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 6860#endif 6861#ifdef CONFIG_RT_GROUP_SCHED 6862 alloc_size += 2 * nr_cpu_ids * sizeof(void **); 6863#endif 6864#ifdef CONFIG_CPUMASK_OFFSTACK 6865 alloc_size += num_possible_cpus() * cpumask_size(); 6866#endif 6867 if (alloc_size) { 6868 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT); 6869 6870#ifdef CONFIG_FAIR_GROUP_SCHED 6871 root_task_group.se = (struct sched_entity **)ptr; 6872 ptr += nr_cpu_ids * sizeof(void **); 6873 6874 root_task_group.cfs_rq = (struct cfs_rq **)ptr; 6875 ptr += nr_cpu_ids * sizeof(void **); 6876 6877#endif /* CONFIG_FAIR_GROUP_SCHED */ 6878#ifdef CONFIG_RT_GROUP_SCHED 6879 root_task_group.rt_se = (struct sched_rt_entity **)ptr; 6880 ptr += nr_cpu_ids * sizeof(void **); 6881 6882 root_task_group.rt_rq = (struct rt_rq **)ptr; 6883 ptr += nr_cpu_ids * sizeof(void **); 6884 6885#endif /* CONFIG_RT_GROUP_SCHED */ 6886#ifdef CONFIG_CPUMASK_OFFSTACK 6887 for_each_possible_cpu(i) { 6888 per_cpu(load_balance_tmpmask, i) = (void *)ptr; 6889 ptr += cpumask_size(); 6890 } 6891#endif /* CONFIG_CPUMASK_OFFSTACK */ 6892 } 6893 6894#ifdef CONFIG_SMP 6895 init_defrootdomain(); 6896#endif 6897 6898 init_rt_bandwidth(&def_rt_bandwidth, 6899 global_rt_period(), global_rt_runtime()); 6900 6901#ifdef CONFIG_RT_GROUP_SCHED 6902 init_rt_bandwidth(&root_task_group.rt_bandwidth, 6903 global_rt_period(), global_rt_runtime()); 6904#endif /* CONFIG_RT_GROUP_SCHED */ 6905 6906#ifdef CONFIG_CGROUP_SCHED 6907 list_add(&root_task_group.list, &task_groups); 6908 INIT_LIST_HEAD(&root_task_group.children); 6909 INIT_LIST_HEAD(&root_task_group.siblings); 6910 autogroup_init(&init_task); 6911 6912#endif /* CONFIG_CGROUP_SCHED */ 6913 6914#ifdef CONFIG_CGROUP_CPUACCT 6915 root_cpuacct.cpustat = &kernel_cpustat; 6916 root_cpuacct.cpuusage = alloc_percpu(u64); 6917 /* Too early, not expected to fail */ 6918 BUG_ON(!root_cpuacct.cpuusage); 6919#endif 6920 for_each_possible_cpu(i) { 6921 struct rq *rq; 6922 6923 rq = cpu_rq(i); 6924 raw_spin_lock_init(&rq->lock); 6925 rq->nr_running = 0; 6926 rq->calc_load_active = 0; 6927 rq->calc_load_update = jiffies + LOAD_FREQ; 6928 init_cfs_rq(&rq->cfs); 6929 init_rt_rq(&rq->rt, rq); 6930#ifdef CONFIG_FAIR_GROUP_SCHED 6931 root_task_group.shares = ROOT_TASK_GROUP_LOAD; 6932 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list); 6933 /* 6934 * How much cpu bandwidth does root_task_group get? 6935 * 6936 * In case of task-groups formed thr' the cgroup filesystem, it 6937 * gets 100% of the cpu resources in the system. This overall 6938 * system cpu resource is divided among the tasks of 6939 * root_task_group and its child task-groups in a fair manner, 6940 * based on each entity's (task or task-group's) weight 6941 * (se->load.weight). 6942 * 6943 * In other words, if root_task_group has 10 tasks of weight 6944 * 1024) and two child groups A0 and A1 (of weight 1024 each), 6945 * then A0's share of the cpu resource is: 6946 * 6947 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33% 6948 * 6949 * We achieve this by letting root_task_group's tasks sit 6950 * directly in rq->cfs (i.e root_task_group->se[] = NULL). 6951 */ 6952 init_cfs_bandwidth(&root_task_group.cfs_bandwidth); 6953 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL); 6954#endif /* CONFIG_FAIR_GROUP_SCHED */ 6955 6956 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime; 6957#ifdef CONFIG_RT_GROUP_SCHED 6958 INIT_LIST_HEAD(&rq->leaf_rt_rq_list); 6959 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL); 6960#endif 6961 6962 for (j = 0; j < CPU_LOAD_IDX_MAX; j++) 6963 rq->cpu_load[j] = 0; 6964 6965 rq->last_load_update_tick = jiffies; 6966 6967#ifdef CONFIG_SMP 6968 rq->sd = NULL; 6969 rq->rd = NULL; 6970 rq->cpu_power = SCHED_POWER_SCALE; 6971 rq->post_schedule = 0; 6972 rq->active_balance = 0; 6973 rq->next_balance = jiffies; 6974 rq->push_cpu = 0; 6975 rq->cpu = i; 6976 rq->online = 0; 6977 rq->idle_stamp = 0; 6978 rq->avg_idle = 2*sysctl_sched_migration_cost; 6979 6980 INIT_LIST_HEAD(&rq->cfs_tasks); 6981 6982 rq_attach_root(rq, &def_root_domain); 6983#ifdef CONFIG_NO_HZ 6984 rq->nohz_flags = 0; 6985#endif 6986#endif 6987 init_rq_hrtick(rq); 6988 atomic_set(&rq->nr_iowait, 0); 6989 } 6990 6991 set_load_weight(&init_task); 6992 6993#ifdef CONFIG_PREEMPT_NOTIFIERS 6994 INIT_HLIST_HEAD(&init_task.preempt_notifiers); 6995#endif 6996 6997#ifdef CONFIG_RT_MUTEXES 6998 plist_head_init(&init_task.pi_waiters); 6999#endif 7000 7001 /* 7002 * The boot idle thread does lazy MMU switching as well: 7003 */ 7004 atomic_inc(&init_mm.mm_count); 7005 enter_lazy_tlb(&init_mm, current); 7006 7007 /* 7008 * Make us the idle thread. Technically, schedule() should not be 7009 * called from this thread, however somewhere below it might be, 7010 * but because we are the idle thread, we just pick up running again 7011 * when this runqueue becomes "idle". 7012 */ 7013 init_idle(current, smp_processor_id()); 7014 7015 calc_load_update = jiffies + LOAD_FREQ; 7016 7017 /* 7018 * During early bootup we pretend to be a normal task: 7019 */ 7020 current->sched_class = &fair_sched_class; 7021 7022#ifdef CONFIG_SMP 7023 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT); 7024 /* May be allocated at isolcpus cmdline parse time */ 7025 if (cpu_isolated_map == NULL) 7026 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT); 7027#endif 7028 init_sched_fair_class(); 7029 7030 scheduler_running = 1; 7031} 7032 7033#ifdef CONFIG_DEBUG_ATOMIC_SLEEP 7034static inline int preempt_count_equals(int preempt_offset) 7035{ 7036 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth(); 7037 7038 return (nested == preempt_offset); 7039} 7040 7041void __might_sleep(const char *file, int line, int preempt_offset) 7042{ 7043 static unsigned long prev_jiffy; /* ratelimiting */ 7044 7045 rcu_sleep_check(); /* WARN_ON_ONCE() by default, no rate limit reqd. */ 7046 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) || 7047 system_state != SYSTEM_RUNNING || oops_in_progress) 7048 return; 7049 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy) 7050 return; 7051 prev_jiffy = jiffies; 7052 7053 printk(KERN_ERR 7054 "BUG: sleeping function called from invalid context at %s:%d\n", 7055 file, line); 7056 printk(KERN_ERR 7057 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n", 7058 in_atomic(), irqs_disabled(), 7059 current->pid, current->comm); 7060 7061 debug_show_held_locks(current); 7062 if (irqs_disabled()) 7063 print_irqtrace_events(current); 7064 dump_stack(); 7065} 7066EXPORT_SYMBOL(__might_sleep); 7067#endif 7068 7069#ifdef CONFIG_MAGIC_SYSRQ 7070static void normalize_task(struct rq *rq, struct task_struct *p) 7071{ 7072 const struct sched_class *prev_class = p->sched_class; 7073 int old_prio = p->prio; 7074 int on_rq; 7075 7076 on_rq = p->on_rq; 7077 if (on_rq) 7078 dequeue_task(rq, p, 0); 7079 __setscheduler(rq, p, SCHED_NORMAL, 0); 7080 if (on_rq) { 7081 enqueue_task(rq, p, 0); 7082 resched_task(rq->curr); 7083 } 7084 7085 check_class_changed(rq, p, prev_class, old_prio); 7086} 7087 7088void normalize_rt_tasks(void) 7089{ 7090 struct task_struct *g, *p; 7091 unsigned long flags; 7092 struct rq *rq; 7093 7094 read_lock_irqsave(&tasklist_lock, flags); 7095 do_each_thread(g, p) { 7096 /* 7097 * Only normalize user tasks: 7098 */ 7099 if (!p->mm) 7100 continue; 7101 7102 p->se.exec_start = 0; 7103#ifdef CONFIG_SCHEDSTATS 7104 p->se.statistics.wait_start = 0; 7105 p->se.statistics.sleep_start = 0; 7106 p->se.statistics.block_start = 0; 7107#endif 7108 7109 if (!rt_task(p)) { 7110 /* 7111 * Renice negative nice level userspace 7112 * tasks back to 0: 7113 */ 7114 if (TASK_NICE(p) < 0 && p->mm) 7115 set_user_nice(p, 0); 7116 continue; 7117 } 7118 7119 raw_spin_lock(&p->pi_lock); 7120 rq = __task_rq_lock(p); 7121 7122 normalize_task(rq, p); 7123 7124 __task_rq_unlock(rq); 7125 raw_spin_unlock(&p->pi_lock); 7126 } while_each_thread(g, p); 7127 7128 read_unlock_irqrestore(&tasklist_lock, flags); 7129} 7130 7131#endif /* CONFIG_MAGIC_SYSRQ */ 7132 7133#if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) 7134/* 7135 * These functions are only useful for the IA64 MCA handling, or kdb. 7136 * 7137 * They can only be called when the whole system has been 7138 * stopped - every CPU needs to be quiescent, and no scheduling 7139 * activity can take place. Using them for anything else would 7140 * be a serious bug, and as a result, they aren't even visible 7141 * under any other configuration. 7142 */ 7143 7144/** 7145 * curr_task - return the current task for a given cpu. 7146 * @cpu: the processor in question. 7147 * 7148 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 7149 */ 7150struct task_struct *curr_task(int cpu) 7151{ 7152 return cpu_curr(cpu); 7153} 7154 7155#endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */ 7156 7157#ifdef CONFIG_IA64 7158/** 7159 * set_curr_task - set the current task for a given cpu. 7160 * @cpu: the processor in question. 7161 * @p: the task pointer to set. 7162 * 7163 * Description: This function must only be used when non-maskable interrupts 7164 * are serviced on a separate stack. It allows the architecture to switch the 7165 * notion of the current task on a cpu in a non-blocking manner. This function 7166 * must be called with all CPU's synchronized, and interrupts disabled, the 7167 * and caller must save the original value of the current task (see 7168 * curr_task() above) and restore that value before reenabling interrupts and 7169 * re-starting the system. 7170 * 7171 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED! 7172 */ 7173void set_curr_task(int cpu, struct task_struct *p) 7174{ 7175 cpu_curr(cpu) = p; 7176} 7177 7178#endif 7179 7180#ifdef CONFIG_CGROUP_SCHED 7181/* task_group_lock serializes the addition/removal of task groups */ 7182static DEFINE_SPINLOCK(task_group_lock); 7183 7184static void free_sched_group(struct task_group *tg) 7185{ 7186 free_fair_sched_group(tg); 7187 free_rt_sched_group(tg); 7188 autogroup_free(tg); 7189 kfree(tg); 7190} 7191 7192/* allocate runqueue etc for a new task group */ 7193struct task_group *sched_create_group(struct task_group *parent) 7194{ 7195 struct task_group *tg; 7196 unsigned long flags; 7197 7198 tg = kzalloc(sizeof(*tg), GFP_KERNEL); 7199 if (!tg) 7200 return ERR_PTR(-ENOMEM); 7201 7202 if (!alloc_fair_sched_group(tg, parent)) 7203 goto err; 7204 7205 if (!alloc_rt_sched_group(tg, parent)) 7206 goto err; 7207 7208 spin_lock_irqsave(&task_group_lock, flags); 7209 list_add_rcu(&tg->list, &task_groups); 7210 7211 WARN_ON(!parent); /* root should already exist */ 7212 7213 tg->parent = parent; 7214 INIT_LIST_HEAD(&tg->children); 7215 list_add_rcu(&tg->siblings, &parent->children); 7216 spin_unlock_irqrestore(&task_group_lock, flags); 7217 7218 return tg; 7219 7220err: 7221 free_sched_group(tg); 7222 return ERR_PTR(-ENOMEM); 7223} 7224 7225/* rcu callback to free various structures associated with a task group */ 7226static void free_sched_group_rcu(struct rcu_head *rhp) 7227{ 7228 /* now it should be safe to free those cfs_rqs */ 7229 free_sched_group(container_of(rhp, struct task_group, rcu)); 7230} 7231 7232/* Destroy runqueue etc associated with a task group */ 7233void sched_destroy_group(struct task_group *tg) 7234{ 7235 unsigned long flags; 7236 int i; 7237 7238 /* end participation in shares distribution */ 7239 for_each_possible_cpu(i) 7240 unregister_fair_sched_group(tg, i); 7241 7242 spin_lock_irqsave(&task_group_lock, flags); 7243 list_del_rcu(&tg->list); 7244 list_del_rcu(&tg->siblings); 7245 spin_unlock_irqrestore(&task_group_lock, flags); 7246 7247 /* wait for possible concurrent references to cfs_rqs complete */ 7248 call_rcu(&tg->rcu, free_sched_group_rcu); 7249} 7250 7251/* change task's runqueue when it moves between groups. 7252 * The caller of this function should have put the task in its new group 7253 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to 7254 * reflect its new group. 7255 */ 7256void sched_move_task(struct task_struct *tsk) 7257{ 7258 int on_rq, running; 7259 unsigned long flags; 7260 struct rq *rq; 7261 7262 rq = task_rq_lock(tsk, &flags); 7263 7264 running = task_current(rq, tsk); 7265 on_rq = tsk->on_rq; 7266 7267 if (on_rq) 7268 dequeue_task(rq, tsk, 0); 7269 if (unlikely(running)) 7270 tsk->sched_class->put_prev_task(rq, tsk); 7271 7272#ifdef CONFIG_FAIR_GROUP_SCHED 7273 if (tsk->sched_class->task_move_group) 7274 tsk->sched_class->task_move_group(tsk, on_rq); 7275 else 7276#endif 7277 set_task_rq(tsk, task_cpu(tsk)); 7278 7279 if (unlikely(running)) 7280 tsk->sched_class->set_curr_task(rq); 7281 if (on_rq) 7282 enqueue_task(rq, tsk, 0); 7283 7284 task_rq_unlock(rq, tsk, &flags); 7285} 7286#endif /* CONFIG_CGROUP_SCHED */ 7287 7288#if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH) 7289static unsigned long to_ratio(u64 period, u64 runtime) 7290{ 7291 if (runtime == RUNTIME_INF) 7292 return 1ULL << 20; 7293 7294 return div64_u64(runtime << 20, period); 7295} 7296#endif 7297 7298#ifdef CONFIG_RT_GROUP_SCHED 7299/* 7300 * Ensure that the real time constraints are schedulable. 7301 */ 7302static DEFINE_MUTEX(rt_constraints_mutex); 7303 7304/* Must be called with tasklist_lock held */ 7305static inline int tg_has_rt_tasks(struct task_group *tg) 7306{ 7307 struct task_struct *g, *p; 7308 7309 do_each_thread(g, p) { 7310 if (rt_task(p) && task_rq(p)->rt.tg == tg) 7311 return 1; 7312 } while_each_thread(g, p); 7313 7314 return 0; 7315} 7316 7317struct rt_schedulable_data { 7318 struct task_group *tg; 7319 u64 rt_period; 7320 u64 rt_runtime; 7321}; 7322 7323static int tg_rt_schedulable(struct task_group *tg, void *data) 7324{ 7325 struct rt_schedulable_data *d = data; 7326 struct task_group *child; 7327 unsigned long total, sum = 0; 7328 u64 period, runtime; 7329 7330 period = ktime_to_ns(tg->rt_bandwidth.rt_period); 7331 runtime = tg->rt_bandwidth.rt_runtime; 7332 7333 if (tg == d->tg) { 7334 period = d->rt_period; 7335 runtime = d->rt_runtime; 7336 } 7337 7338 /* 7339 * Cannot have more runtime than the period. 7340 */ 7341 if (runtime > period && runtime != RUNTIME_INF) 7342 return -EINVAL; 7343 7344 /* 7345 * Ensure we don't starve existing RT tasks. 7346 */ 7347 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg)) 7348 return -EBUSY; 7349 7350 total = to_ratio(period, runtime); 7351 7352 /* 7353 * Nobody can have more than the global setting allows. 7354 */ 7355 if (total > to_ratio(global_rt_period(), global_rt_runtime())) 7356 return -EINVAL; 7357 7358 /* 7359 * The sum of our children's runtime should not exceed our own. 7360 */ 7361 list_for_each_entry_rcu(child, &tg->children, siblings) { 7362 period = ktime_to_ns(child->rt_bandwidth.rt_period); 7363 runtime = child->rt_bandwidth.rt_runtime; 7364 7365 if (child == d->tg) { 7366 period = d->rt_period; 7367 runtime = d->rt_runtime; 7368 } 7369 7370 sum += to_ratio(period, runtime); 7371 } 7372 7373 if (sum > total) 7374 return -EINVAL; 7375 7376 return 0; 7377} 7378 7379static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime) 7380{ 7381 int ret; 7382 7383 struct rt_schedulable_data data = { 7384 .tg = tg, 7385 .rt_period = period, 7386 .rt_runtime = runtime, 7387 }; 7388 7389 rcu_read_lock(); 7390 ret = walk_tg_tree(tg_rt_schedulable, tg_nop, &data); 7391 rcu_read_unlock(); 7392 7393 return ret; 7394} 7395 7396static int tg_set_rt_bandwidth(struct task_group *tg, 7397 u64 rt_period, u64 rt_runtime) 7398{ 7399 int i, err = 0; 7400 7401 mutex_lock(&rt_constraints_mutex); 7402 read_lock(&tasklist_lock); 7403 err = __rt_schedulable(tg, rt_period, rt_runtime); 7404 if (err) 7405 goto unlock; 7406 7407 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock); 7408 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period); 7409 tg->rt_bandwidth.rt_runtime = rt_runtime; 7410 7411 for_each_possible_cpu(i) { 7412 struct rt_rq *rt_rq = tg->rt_rq[i]; 7413 7414 raw_spin_lock(&rt_rq->rt_runtime_lock); 7415 rt_rq->rt_runtime = rt_runtime; 7416 raw_spin_unlock(&rt_rq->rt_runtime_lock); 7417 } 7418 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock); 7419unlock: 7420 read_unlock(&tasklist_lock); 7421 mutex_unlock(&rt_constraints_mutex); 7422 7423 return err; 7424} 7425 7426int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us) 7427{ 7428 u64 rt_runtime, rt_period; 7429 7430 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period); 7431 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC; 7432 if (rt_runtime_us < 0) 7433 rt_runtime = RUNTIME_INF; 7434 7435 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 7436} 7437 7438long sched_group_rt_runtime(struct task_group *tg) 7439{ 7440 u64 rt_runtime_us; 7441 7442 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF) 7443 return -1; 7444 7445 rt_runtime_us = tg->rt_bandwidth.rt_runtime; 7446 do_div(rt_runtime_us, NSEC_PER_USEC); 7447 return rt_runtime_us; 7448} 7449 7450int sched_group_set_rt_period(struct task_group *tg, long rt_period_us) 7451{ 7452 u64 rt_runtime, rt_period; 7453 7454 rt_period = (u64)rt_period_us * NSEC_PER_USEC; 7455 rt_runtime = tg->rt_bandwidth.rt_runtime; 7456 7457 if (rt_period == 0) 7458 return -EINVAL; 7459 7460 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime); 7461} 7462 7463long sched_group_rt_period(struct task_group *tg) 7464{ 7465 u64 rt_period_us; 7466 7467 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period); 7468 do_div(rt_period_us, NSEC_PER_USEC); 7469 return rt_period_us; 7470} 7471 7472static int sched_rt_global_constraints(void) 7473{ 7474 u64 runtime, period; 7475 int ret = 0; 7476 7477 if (sysctl_sched_rt_period <= 0) 7478 return -EINVAL; 7479 7480 runtime = global_rt_runtime(); 7481 period = global_rt_period(); 7482 7483 /* 7484 * Sanity check on the sysctl variables. 7485 */ 7486 if (runtime > period && runtime != RUNTIME_INF) 7487 return -EINVAL; 7488 7489 mutex_lock(&rt_constraints_mutex); 7490 read_lock(&tasklist_lock); 7491 ret = __rt_schedulable(NULL, 0, 0); 7492 read_unlock(&tasklist_lock); 7493 mutex_unlock(&rt_constraints_mutex); 7494 7495 return ret; 7496} 7497 7498int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk) 7499{ 7500 /* Don't accept realtime tasks when there is no way for them to run */ 7501 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0) 7502 return 0; 7503 7504 return 1; 7505} 7506 7507#else /* !CONFIG_RT_GROUP_SCHED */ 7508static int sched_rt_global_constraints(void) 7509{ 7510 unsigned long flags; 7511 int i; 7512 7513 if (sysctl_sched_rt_period <= 0) 7514 return -EINVAL; 7515 7516 /* 7517 * There's always some RT tasks in the root group 7518 * -- migration, kstopmachine etc.. 7519 */ 7520 if (sysctl_sched_rt_runtime == 0) 7521 return -EBUSY; 7522 7523 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags); 7524 for_each_possible_cpu(i) { 7525 struct rt_rq *rt_rq = &cpu_rq(i)->rt; 7526 7527 raw_spin_lock(&rt_rq->rt_runtime_lock); 7528 rt_rq->rt_runtime = global_rt_runtime(); 7529 raw_spin_unlock(&rt_rq->rt_runtime_lock); 7530 } 7531 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags); 7532 7533 return 0; 7534} 7535#endif /* CONFIG_RT_GROUP_SCHED */ 7536 7537int sched_rt_handler(struct ctl_table *table, int write, 7538 void __user *buffer, size_t *lenp, 7539 loff_t *ppos) 7540{ 7541 int ret; 7542 int old_period, old_runtime; 7543 static DEFINE_MUTEX(mutex); 7544 7545 mutex_lock(&mutex); 7546 old_period = sysctl_sched_rt_period; 7547 old_runtime = sysctl_sched_rt_runtime; 7548 7549 ret = proc_dointvec(table, write, buffer, lenp, ppos); 7550 7551 if (!ret && write) { 7552 ret = sched_rt_global_constraints(); 7553 if (ret) { 7554 sysctl_sched_rt_period = old_period; 7555 sysctl_sched_rt_runtime = old_runtime; 7556 } else { 7557 def_rt_bandwidth.rt_runtime = global_rt_runtime(); 7558 def_rt_bandwidth.rt_period = 7559 ns_to_ktime(global_rt_period()); 7560 } 7561 } 7562 mutex_unlock(&mutex); 7563 7564 return ret; 7565} 7566 7567#ifdef CONFIG_CGROUP_SCHED 7568 7569/* return corresponding task_group object of a cgroup */ 7570static inline struct task_group *cgroup_tg(struct cgroup *cgrp) 7571{ 7572 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id), 7573 struct task_group, css); 7574} 7575 7576static struct cgroup_subsys_state *cpu_cgroup_create(struct cgroup *cgrp) 7577{ 7578 struct task_group *tg, *parent; 7579 7580 if (!cgrp->parent) { 7581 /* This is early initialization for the top cgroup */ 7582 return &root_task_group.css; 7583 } 7584 7585 parent = cgroup_tg(cgrp->parent); 7586 tg = sched_create_group(parent); 7587 if (IS_ERR(tg)) 7588 return ERR_PTR(-ENOMEM); 7589 7590 return &tg->css; 7591} 7592 7593static void cpu_cgroup_destroy(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 *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 *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 *cgrp, struct cgroup *old_cgrp, 7629 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(struct cgroup *cgrp) 7980{ 7981 struct cpuacct *ca; 7982 7983 if (!cgrp->parent) 7984 return &root_cpuacct.css; 7985 7986 ca = kzalloc(sizeof(*ca), GFP_KERNEL); 7987 if (!ca) 7988 goto out; 7989 7990 ca->cpuusage = alloc_percpu(u64); 7991 if (!ca->cpuusage) 7992 goto out_free_ca; 7993 7994 ca->cpustat = alloc_percpu(struct kernel_cpustat); 7995 if (!ca->cpustat) 7996 goto out_free_cpuusage; 7997 7998 return &ca->css; 7999 8000out_free_cpuusage: 8001 free_percpu(ca->cpuusage); 8002out_free_ca: 8003 kfree(ca); 8004out: 8005 return ERR_PTR(-ENOMEM); 8006} 8007 8008/* destroy an existing cpu accounting group */ 8009static void cpuacct_destroy(struct cgroup *cgrp) 8010{ 8011 struct cpuacct *ca = cgroup_ca(cgrp); 8012 8013 free_percpu(ca->cpustat); 8014 free_percpu(ca->cpuusage); 8015 kfree(ca); 8016} 8017 8018static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu) 8019{ 8020 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu); 8021 u64 data; 8022 8023#ifndef CONFIG_64BIT 8024 /* 8025 * Take rq->lock to make 64-bit read safe on 32-bit platforms. 8026 */ 8027 raw_spin_lock_irq(&cpu_rq(cpu)->lock); 8028 data = *cpuusage; 8029 raw_spin_unlock_irq(&cpu_rq(cpu)->lock); 8030#else 8031 data = *cpuusage; 8032#endif 8033 8034 return data; 8035} 8036 8037static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val) 8038{ 8039 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu); 8040 8041#ifndef CONFIG_64BIT 8042 /* 8043 * Take rq->lock to make 64-bit write safe on 32-bit platforms. 8044 */ 8045 raw_spin_lock_irq(&cpu_rq(cpu)->lock); 8046 *cpuusage = val; 8047 raw_spin_unlock_irq(&cpu_rq(cpu)->lock); 8048#else 8049 *cpuusage = val; 8050#endif 8051} 8052 8053/* return total cpu usage (in nanoseconds) of a group */ 8054static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft) 8055{ 8056 struct cpuacct *ca = cgroup_ca(cgrp); 8057 u64 totalcpuusage = 0; 8058 int i; 8059 8060 for_each_present_cpu(i) 8061 totalcpuusage += cpuacct_cpuusage_read(ca, i); 8062 8063 return totalcpuusage; 8064} 8065 8066static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype, 8067 u64 reset) 8068{ 8069 struct cpuacct *ca = cgroup_ca(cgrp); 8070 int err = 0; 8071 int i; 8072 8073 if (reset) { 8074 err = -EINVAL; 8075 goto out; 8076 } 8077 8078 for_each_present_cpu(i) 8079 cpuacct_cpuusage_write(ca, i, 0); 8080 8081out: 8082 return err; 8083} 8084 8085static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft, 8086 struct seq_file *m) 8087{ 8088 struct cpuacct *ca = cgroup_ca(cgroup); 8089 u64 percpu; 8090 int i; 8091 8092 for_each_present_cpu(i) { 8093 percpu = cpuacct_cpuusage_read(ca, i); 8094 seq_printf(m, "%llu ", (unsigned long long) percpu); 8095 } 8096 seq_printf(m, "\n"); 8097 return 0; 8098} 8099 8100static const char *cpuacct_stat_desc[] = { 8101 [CPUACCT_STAT_USER] = "user", 8102 [CPUACCT_STAT_SYSTEM] = "system", 8103}; 8104 8105static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft, 8106 struct cgroup_map_cb *cb) 8107{ 8108 struct cpuacct *ca = cgroup_ca(cgrp); 8109 int cpu; 8110 s64 val = 0; 8111 8112 for_each_online_cpu(cpu) { 8113 struct kernel_cpustat *kcpustat = per_cpu_ptr(ca->cpustat, cpu); 8114 val += kcpustat->cpustat[CPUTIME_USER]; 8115 val += kcpustat->cpustat[CPUTIME_NICE]; 8116 } 8117 val = cputime64_to_clock_t(val); 8118 cb->fill(cb, cpuacct_stat_desc[CPUACCT_STAT_USER], val); 8119 8120 val = 0; 8121 for_each_online_cpu(cpu) { 8122 struct kernel_cpustat *kcpustat = per_cpu_ptr(ca->cpustat, cpu); 8123 val += kcpustat->cpustat[CPUTIME_SYSTEM]; 8124 val += kcpustat->cpustat[CPUTIME_IRQ]; 8125 val += kcpustat->cpustat[CPUTIME_SOFTIRQ]; 8126 } 8127 8128 val = cputime64_to_clock_t(val); 8129 cb->fill(cb, cpuacct_stat_desc[CPUACCT_STAT_SYSTEM], val); 8130 8131 return 0; 8132} 8133 8134static struct cftype files[] = { 8135 { 8136 .name = "usage", 8137 .read_u64 = cpuusage_read, 8138 .write_u64 = cpuusage_write, 8139 }, 8140 { 8141 .name = "usage_percpu", 8142 .read_seq_string = cpuacct_percpu_seq_read, 8143 }, 8144 { 8145 .name = "stat", 8146 .read_map = cpuacct_stats_show, 8147 }, 8148}; 8149 8150static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp) 8151{ 8152 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files)); 8153} 8154 8155/* 8156 * charge this task's execution time to its accounting group. 8157 * 8158 * called with rq->lock held. 8159 */ 8160void cpuacct_charge(struct task_struct *tsk, u64 cputime) 8161{ 8162 struct cpuacct *ca; 8163 int cpu; 8164 8165 if (unlikely(!cpuacct_subsys.active)) 8166 return; 8167 8168 cpu = task_cpu(tsk); 8169 8170 rcu_read_lock(); 8171 8172 ca = task_ca(tsk); 8173 8174 for (; ca; ca = parent_ca(ca)) { 8175 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu); 8176 *cpuusage += cputime; 8177 } 8178 8179 rcu_read_unlock(); 8180} 8181 8182struct cgroup_subsys cpuacct_subsys = { 8183 .name = "cpuacct", 8184 .create = cpuacct_create, 8185 .destroy = cpuacct_destroy, 8186 .populate = cpuacct_populate, 8187 .subsys_id = cpuacct_subsys_id, 8188}; 8189#endif /* CONFIG_CGROUP_CPUACCT */ 8190